Advances in preventive monitoring of machinery through audio and vibration signals

UNIVERSIDAD DE LAS PALMAS DE GRAN CANARIA INSTITUTO UNIVERSITARIO DE SISTEMAS INTELIGENTES Y APLICACIONES NUMÉRICAS EN INGENIERÍA PROGRAMA DE DOCTORADO SISTEMAS INTELIGENTES Y APLICACIONES NUMÉRICAS EN INGENIERÍA TÍTULO DE LA TESIS Avances en Monitorización Preventiva de Maquinaria a tráves de Señales de Audio y Vibración Tesis Doctoral presentada por Patricia Henríquez Rodríguez Dirigida por el Dr. D. Miguel Ángel Ferrer Ballester Codirigida por el Dr. D. Jesús Bernardino Alonso Hernández El Director El Codirector La Doctoranda, Las Palmas de Gran Canaria, a 22 de Octubre de 2015 A mi familia. Advances in preventive monitoring of machinery through audio and vibration Signals Universidad de Las Palmas de Gran Canaria ix ACKNOWLEDGEMENTS This Thesis summarizes the work I have carried out during my Ph.D. studies with the División de Procesado Digital de la Señal (DPDS) -Digital Signal Processing Divisionof the Instituto para el Desarrollo Tecnológico y la Innovación en Comunicaciones (IDeTIC)-Institute for Technological Development and the Innovation in Communicationssince 2009. This research group is with the Department of Señales y Comunicaciones of the Universidad de Las Palmas de Gran Canaria. IDeTIC started in 2010 from the CeTIC (Technological Centre for the Innovation in Communications). The CeTIC was founded in 2006 from the three different Research Groups, including the Digital Signal Processing Group with was founded in 1990. This Thesis has been mainly supported by a Ph.D. scholarship granted to the author by Gobierno de Canarias and Social European Fund (ESF), which covered the period between March 2009 and March 2013, and also by the Spanish government MICNN TEC2009-14123-C04 and TEC2012-38630-C04-02 research projects. Foremost, I would like to thank my supervisors Prof. Jesús B. Alonso Hernández and Prof. Miguel A. Ferrer Ballester for their guidance and support over the past six years. During this period I have benefited from their wise advices, intelligent efforts and courage which have shaped my thinking and working attitude. They also were the responsible of my first steps in research world when I first took at their door for pursuing my MsC. Project about laryngeal pathology detection through voice signal. They introduced me to the signal processing and pattern recognition world. In the framework of DPDS I have also received support from Prof. Carlos M. Travieso González. After finishing my MsC. Project, they gave me the great opportunity of working in a national project of homeland security (Hesperia) founded by the office of the PhD Dissertation Universidad de Las Palmas de Gran Canaria x industrial technological development (CDTI) (attached to the Ministriy of Industry and Trade) of the Spanish government in the frame of CENIT projects. In the framework of this work, I learned about condition monitoring in industrial scenarios and I could benefit from the knownledge of many experts in industrial security involved in the project. Thanks to this work, I decided to focus my PhD on condition monitoring implementing fault diagnosis. During my Ph. D. studies I had the great opportunity of visiting two foreign institutions. The first 4-month stay was in 2011 at the Institute of Sound and Vibration Research (ISVR) of the University of Southampton (Southampton, England) with Prof. Paul R. White. During those four months I could benefit from his mastery in the field of signal processing. I also had the fortune to meet Prof. Ling Wang from who I could benefit from her expertise in pump fault diagnosis. The second 3-month stay was in 2012 at the Reliability Research Lab of the University of Alberta (Edmonton, Canada) with Prof. Ming J. Zuo. I have met a fantastic group of people there. I would like to specially thank Dr. Funda Akdere Iscioglu for her care during my visit. I have also would like to thank Mayank Pandey for our discussions about the pump installation and condition monitoring techniques. From such research stays there is a huge number of colleagues and friends who I want to thank: Alberto Baldelli, Pierre Rosado, Vianey Landeros, Kiran Konde, Rami Saba, Dao Duc Cuong, Luis Espitia and his wife Jessica Espitia. I want to thank to my MsC. student Garoé Gómez for his valuable help with the pump installation and data acquisition process. I also want to thank all the work mates at IDeTIC who with I have shared lunch and discussions: Himar, Javier, Laura, Jaime, Daniel, Pedro, Ayaya, Manolo, Víctor and all IDeTIC human group. Thank you guys! Quiero también darle las gracias a mis amigos de Las Palmas, los cuales han aguantado durante todos estos años mis dudas, mis ausencias, mis nervios y con los cuales he compartido mis logros y mis alegrías. Muchas gracias a Eli, Paco, Raquel, Davinia, Rosalva, Silvia, Gema, Mónica y Octavio. Y en especial gracias a Briseida, por estar siempre a mi lado. Gracias a mi familia. Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria xvii LIST OF FIGURES Figure 1-1: Stages in a condition-based montioring system. ......................................... 32 Figure 1-2: Dependence among Dissertation chapters. .................................................. 41 Figure 2-1: Stages in a condition-based montioring system .......................................... 47 Figure 2-2: Distributions of referenced papers: Audio-based and vibration-based fault diagnosis distribution (upper left). Publishing years distribution (upper right).Distribution of faults in vibration-based diagnosis (bottom left) and in audiobased diagnosis (bottom right). .............................................................................. 49 Figure 3-1: Components of a ball bearing ...................................................................... 76 Figure 3-2: Bearing vibration signals of bearings with discrete faults at outer race, inner race and ball. BPFO = ball pass frequency, outer race, BPFI = ball pass frequency, inner race, BSF = ball spin frequency, FTF: fundamental train frequency (cage frequency). Source: [26]. ........................................................................................ 78 Figure 3-3: Time bearing vibration signal of 0.17 seconds (sample frequency is 12kHz) for different bearing conditions: Normal condition (upper-left), Inner race fault condition (upper-right), Outer race fault condition (bottom-left) and Ball fault condition (bottom-right). ........................................................................................ 79 Figure 3-4: Test stand for the acquisition of bearing vibration data. Source [1]. ........... 81 Figure 3-5: Main transmission of an UH-60 Backhawk helicopter. Source: [27]. ......... 82 Figure 3-6: Position of the bearing SB-2205 in the transmission (left figure) and the fnal bearing condition (right figure). Source: [28]. ...................................................... 83 Figure 3-7: Position of the bearings in the IMS bearing database. Source: [3]. ............. 84 Figure 3-8: Left: Fluid direction in a centrifugal pump. Source: [6]. Right: velocity and pressure in a centrifugal pump. Modified from source: [5]. The inlet, outlet, volute and impeller are shown. .......................................................................................... 85 Figure 3-9: Closed impeller. Modified from source: [6]. ............................................... 86 Figure 3-10: Diagram with the components of the acquisition system .......................... 91 Figure 3-11: Diagram of the ALP800 pump .................................................................. 92 Figure 3-12: Two photos of the ALP 800 pump. ........................................................... 94 Figure 3-13: Photos of the pump ALP 800. Impeller mounted on the rotor (left). Volute, inlet and oulet pump (right). ................................................................................... 94 PhD Dissertation Universidad de Las Palmas de Gran Canaria xviii Figure 3-14: Photos of the pump ALP 800. Impeller mounted on the rotor (left). Different views of the impeller (center and right). ................................................. 95 Figure 3-15: Photos of the plate fault in impeller. Upper left: impeller in normal condition. Upper right: one part of the plate removed. Lower left: two parts of the plate are removed. Lower right: the third part of the impeller was removed. ........ 96 Figure 3-16: Photo of the impeller with leading edge fault. In this case, the slight fault (5mm from all vane leading edges were removed) is shown. ................................ 97 Figure 3-17: Photo of the impeller with trailing edge fault. In this case, the slight fault (5mm from all vane trailing edges were removed) is shown. ................................ 98 Figure 3-18: Photo of the o-ring of the impeller of the AL P800 pump......................... 99 Figure 3-19: Photo of the reservoir with the sand added to simulate strange objects or particles in the system. ......................................................................................... 100 Figure 3-20: Kind of PVC balls (6 mm of diameter) used to simulate strange objects in the system [29]. .................................................................................................... 100 Figure 3-21: Drawing of the room for the experimental set-up with the measurements in mm. A squematic of the set-up is also shown. ..................................................... 101 Figure 3-22: Sensor Placement for the Experimental set-up. ....................................... 102 Figure 3-23: Spectrum in range [2.7-1000]Hz of a 8192 points vibration frame (371.5ms) from sensor RadialInletAccel. The higher peaks are marked with red circles and the frequency value is shown. ............................................................ 106 Figure 3-24: Spectrum in range [1000-11025]Hz of a 8192 points vibration frame (371.5ms) from sensor RadialInletAccel. The higher peaks are marked with red circles and the frequency value is shown. ............................................................ 106 Figure 3-25: Spectrum in range [2.7-1000]Hz of a 8192 points audio frame (371.5ms) from sensor MicroInlet. The higher peaks are marked with red circles and the frequency value is shown. .................................................................................... 107 Figure 3-26: Spectrum in range [1000-11025]Hz of a 8192 points audio frame (371.5ms) from sensor MicroInlet. The higher peaks are marked with red circles and the frequency value is shown. ........................................................................ 107 Figure 4-1: Time Vibration signals for bearings with outer race fault, inner race fault and ball fault. Source: [28]. .................................................................................. 114 Figure 4-2: Teager-Kaiser transformed signals (TK signals) for different bearing conditions: normal condition (upper-left), inner race faul condition (upper-right), outer race fault condition (bottom-left) and ball condition (bottom-right). ......... 118 Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria xix Figure 4-3: Diagram of the experimentation in bearing fault diagnosis. The proposal is compared with other methods in the state of the art and an evaluation with two classifiers is carried out. ....................................................................................... 119 Figure 4-4: Diagram of the experimentation in bearing fault evolution. The proposal is applied to a run-to-failure bearing vibration database. ......................................... 119 Figure 4-5: Success rates using TK, T, AM and TK-AM features in order of relevance with the neural network classifier. ........................................................................ 126 Figure 4-6: Success rates using TK, T, AM and TK-AM features in order of relevance with the LS-SVM classifier. ................................................................................. 126 Figure 4-7: Evolution of the features extracted from the TK signal. ........................... 128 Figure 4-8: Evolution of the features extracted from the raw vibration signal (T signal). .............................................................................................................................. 129 Figure 4-9: Evolution of the features extracted from the envelope signal of the Q band[2500-3800Hz]. ............................................................................................. 129 Figure 4-10: Proposed methodology for bearing severity fault assessment. ................ 136 Figure 4-11: Waveform of the Daubechies 6 wavelet mother. .................................... 136 Figure 4-12: Relative energies for WPT nodes of level 3 of decomposition (ordered by frequency content) for a frame of 4096 samples extracted for normal vibration signals with different loads. .................................................................................. 139 Figure 4-13: Relative energies for WPT nodes of level 3 of decomposition (ordered by frequency content) for a frame of 4096 samples extracted for vibration signals with inner race fault (left column), outer race fault (center column) and ball fault (right column) with different severities and for load 0. ................................................. 139 Figure 4-14: Relative energies for WPT nodes of level 3 of decomposition (ordered by frequency content) for a frame of 4096 samples extracted for vibration signals with inner race fault (left column), outer race fault (center column) and ball fault (right column) with different severities and for load 1. ................................................. 140 Figure 4-15: Relative energies for WPT nodes of level 3 of decomposition (ordered by frequency content) for a frame of 4096 samples extracted for vibration signals with inner race fault (left column), outer race fault (center column) and ball fault (right column) with different severities and for load 2. ................................................. 140 Figure 4-16: Relative energies for WPT nodes of level 3 of decomposition (ordered by frequency content) for a frame of 4096 samples extracted for vibration signals with PhD Dissertation Universidad de Las Palmas de Gran Canaria xx inner race fault (left column), outer race fault (center column) and ball fault (right column) with different severities and for load 3. ................................................. 141 Figure 4-17: Evolution of the mean values obtained for each method: proposed method ‘LZC (EmaxWPT)’ (uppe-left), ‘LZC (raw signal)’ (upper-right), ‘Kurtosis (EmaxWPT)’ (bottom-left) and ‘Kurtosis (raw signal)’ (bottom-right) for outer race fault with different severities and different loads. ........................................ 143 Figure 4-18: Evolution of the mean values obtained for each method: proposed method ‘LZC (EmaxWPT)’ (uppe-left), ‘LZC (raw signal)’ (upper-right), ‘Kurtosis (EmaxWPT)’ (bottom-left) and ‘Kurtosis (raw signal)’ (bottom-right) for inner race fault with different severities and different loads. ........................................ 144 Figure 4-19: Evolution of the mean values obtained for each method: proposed method ‘LZC (EmaxWPT)’ (uppe-left), ‘LZC (raw signal)’ (upper-right), ‘Kurtosis (EmaxWPT)’ (bottom-left) and ‘Kurtosis (raw signal)’ (bottom-right) for outer race fault with different severities and different loads. The normal condition is also considered. ............................................................................................................ 144 Figure 4-20: Evolution of the mean values obtained for each method: proposed method ‘LZC (EmaxWPT)’ (uppe-left), ‘LZC (raw signal)’ (upper-right), ‘Kurtosis (EmaxWPT)’ (bottom-left) and ‘Kurtosis (raw signal)’ (bottom-right) for inner race fault with different severities and different loads. The normal condition is also considered. ............................................................................................................ 145 Figure 4-21: Evolution of the mean values obtained for ‘Kurtosis (raw signal)’ method applied for outer race fault condition and normal condition varying the amount of noise added to the original vibration signal. The results are shown for load: load 0 (upper-left), load 1(upper-right), load 2 (bottom-left), load 3 (bottom-right). ..... 146 Figure 4-22: Evolution of the mean values obtained for ‘LZC (raw signal)’ method applied for outer race fault condition and normal condition varying the amount of noise added to the original vibration signal. The results are shown for load: load 0 (upper-left), load 1(upper-right), load 2 (bottom-left) and load 3 (bottom-right). 146 Figure 4-23: Evolution of the mean values obtained for ‘Kurtosis (EmaxWPT)’ method applied for outer race fault condition and normal condition varying the amount of noise added to the original vibration signal. The results are shown for load: load 0 (upper-left), load 1(upper-right), load 2 (bottom-left) and load 3 (bottom-right). 147 Figure 4-24: Evolution of the mean values obtained for the proposed method ‘LZC (EmaxWPT)’ applied for outer race fault condition and normal condition varying Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria xxi the amount of noise added to the original vibration signal. The results are shown for load: load 0 (upper-left), load 1(upper-right), load 2 (bottom-left) and load 3 (bottom-right). ...................................................................................................... 147 Figure 4-25: Evolution of the mean values obtained for ‘Kurtosis (raw signal)’ method applied for inner race fault condition and normal condition varying the amount of noise added to the original vibration signal. The results are shown for load: load 0 (upper-left), load 1(upper-right), load 2 (bottom-left) and load 3 (bottom-right). 148 Figure 4-26: Evolution of the mean values obtained for ‘LZC (raw signal)’ method applied for inner race fault condition and normal condition varying the amount of noise added to the original vibration signal. The results are shown for load: load 0 (upper-left), load 1(upper-right), load 2 (bottom-left) and load 3 (bottom-right). 148 Figure 4-27: Evolution of the mean values obtained for ‘Kurtosis (EmaxWPT)’ method applied for inner race fault condition and normal condition varying the amount of noise added to the original vibration signal. The results are shown for load: load 0 (upper-left), load 1(upper-right), load 2 (bottom-left) and load 3 (bottom-right). 149 Figure 4-28: Evolution of the mean values obtained for the proposed method ‘LZC (EmaxWPT)’ applied for inner race fault condition and normal condition varying the amount of noise added to the original vibration signal. The results are shown for load: load 0 (upper-left), load 1(upper-right), load 2 (bottom-left) and load 3 (bottom-right). ...................................................................................................... 149 Figure 4-29: Evolution of the features using kurtosis and Lempel-Ziv complexity from the raw signal and kurtosis and Lempel-Ziv complexity from the node with maximal energy of the wavelet paket transform. ................................................. 150 Figure 4-30: Evolution of the Kurtosis extracted from the raw vibration signal of a runto-failure experiment of the bearing that eventually developed an outer-race fault. .............................................................................................................................. 152 Figure 4-31: Evolution of the Kurtosis extracted from the node of maximal energy of the WPT of a run-to-failure experiment of the bearing that eventually developed an outer-race fault. ..................................................................................................... 152 Figure 4-32: Evolution of the Lempel-Ziv complexity extracted from the raw vibration signal of a run-to-failure experiment of the bearing that eventually developed an outer-race fault. ..................................................................................................... 153 PhD Dissertation Universidad de Las Palmas de Gran Canaria xxii Figure 4-33: Evolution of the Lempel-Ziv complexity extracted from the node of maximal energy of the WPT of a run-to-failure experiment of the bearing that eventually developed an outer-race fault. ............................................................. 153 Figure 5-1: Squeme of the methodology used to quantify the ability of the features for vibration and audio based pump fault diagnosis. ................................................. 161 Figure 5-2: Time frames (8192 samples) from sensor Radial Inlet Accel (vibration) and Inlet Micro (audio) for normal (NOR) and plate (PLA) condition....................... 178 Figure 5-3: Time frames (8192 samples) from sensor Radial Inlet Accel (vibration) and Inlet Micro (audio) for leading edge damage (LED) and trailing edge damage (TED) condition. ................................................................................................... 178 Figure 5-4: Time frames (8192 samples) from sensor Radial Inlet Accel (vibration) and Inlet Micro (audio) for seal (SEA) and sand (SAN) condition............................. 179 Figure 5-5: Time frames (8192 samples) from sensor Radial Inlet Accel (vibration) and Inlet Micro (audio) for sand and paper (SAP) and pvc balls (PVC) condition. ... 179 Figure 5-6: Spectra of vibration frames (8192 samples) from sensor Radial Inlet Accel for different pump conditions: normal (NOR), plate (PLA), leading edge damage (LED), trailing edge damage (TED), seal (SEA), sand (SAN), sand and paper (SAP) and pvc balls (PVC) conditions. ................................................................ 187 Figure 5-7: Spectra of audio frames (8192 samples) from sensor Inlet Micro for different pump conditions: normal (NOR), plate (PLA), leading edge damage (LED), trailing edge damage (TED), seal (SEA), sand (SAN), sand and paper (SAP) and pvc balls (PVC) conditions. ................................................................ 188 Figure 5-8: Spectra of vibration frames (8192 samples) from sensor Radial Inlet Accel in frequency range [0-1000]Hz for different pump conditions: normal (NOR), plate (PLA), leading edge damage (LED), trailing edge damage (TED), seal (SEA), sand (SAN), sand and paper (SAP) and pvc balls (PVC) conditions. .......................... 189 Figure 5-9: Spectra of audio frames (8192 samples) from sensor Inlet Micro in frequency range [0-1000]Hz for different pump conditions: normal (NOR), plate (PLA), leading edge damage (LED), trailing edge damage (TED), seal (SEA), sand (SAN), sand and paper (SAP) and pvc balls (PVC) conditions. .......................... 189 Figure 5-10: Rectified cepstrum of an audio frame from sensor Inlet Micro in normal condition. .............................................................................................................. 193 Figure 5-11: Rectified cepstrum of an audio frame from sensor Inlet Micro in LED condition. .............................................................................................................. 193 Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria xxiii Figure 5-12: Rectified cepstrum of a vibration frame from sensor Radial Inlet Accel in normal condition. .................................................................................................. 194 Figure 5-13: Rectified cepstrum of a vibration frame from sensor Radial Inlet Accel in LED condition. ..................................................................................................... 194 Figure 5-14: Sequential evaluation of the features for each sensor in pump fault diagnosis using a neural network classifier to discriminate between 8 pump conditions. ............................................................................................................ 217 Figure 5-15: Sequential evaluation of the features for each sensor in pump fault diagnosis using a LSSVM classifier to discriminate between 8 pump conditions. .............................................................................................................................. 217 Figure 5-16: Sequential evaluation of the features for each sensor in pump fault diagnosis using a neural network classifier to discriminate between 17 pump conditions. ............................................................................................................ 218 Figure 5-17: Sequential evaluation of the features for each sensor in pump fault diagnosis using a LSSVM classifier to discriminate between 8 pump conditions. .............................................................................................................................. 218 Figure 8-1: Data distribution of each kind of voice (H: healthy voice, P: pathological voice, LP: light pathological voice, MP: moderate pathological voice, SP: severe pathological voice) for each measurement extracted from the /a/ vowel of the multiquality database (FMMI: first minimum of the mutual information function. CD: correlation dimension. CE: correlation entropy. RE1: first-order Rényi block entropy. RE2: second-order Rényi block entropy. SE: Shannon entropy). Source: [1]. ........................................................................................................................ 248 Figure 8-2: Data distribution of each kind of voice (H: healthy voice, P: pathological voice) for each measurement extracted from the MEEI database (FMMI: first minimum of the mutual information function. CD: correlation dimension. CE: correlation entropy. RE1: first-order Rényi block entropy. RE2: second-order Rényi block entropy. SE: Shannon entropy). Source: [1]. ................................... 249 Figure 8-3: Data distribution of neutral, fear and anger emotional speech for each complexity measure extracted from the Polish emotional database. MI: value of the first minimum of the mutual information function (upper left), SE: Shannon entropy (upper right), CD: Taken's estimator of the correlation dimension (middle left), CE: correlation entropy (middle right), LZC: Lempel–Ziv complexity (bottom left), H: Hurst exponent (bottom right) [40]. .......................................... 259 PhD Dissertation Universidad de Las Palmas de Gran Canaria xxiv Figure 8-4: Data distribution of neutral, fear and anger emotional speech for each complexity measure extracted from the Berlin emotional database. MI: value of the first minimum of the mutual information function (upper left), SE: Shannon entropy (upper right), CD: Taken's estimator of the correlation dimension (middle left), CE: correlation entropy (middle right), LZC: Lempel–Ziv complexity (bottom left), H: Hurst exponent (bottom right) [40]. .......................................... 260 Figure 8-5: Data distribution of neutral, fear and anger emotional speech for each complexity measure extracted from the LDC emotional database. MI: value of the first minimum of the mutual information function (upper left), SE: Shannon entropy (upper right), CD: Taken's estimator of the correlation dimension (middle left), CE: correlation entropy (middle right), LZC: Lempel–Ziv complexity (bottom left), H: Hurst exponent (bottom right) [40]. .......................................... 261 Figura 1: Etapas de un sistema de monitorización basado en la condición .................. 280 Figura 2: Esquema de dependencia entre los capítulos de la memoria. ....................... 287 Figura 3: Distribución de los artículos usados para la redacción del estado del arte: distribución de artículos basados en señales de vibración y basados en señales de audio (superior izquierda); años de publicación de los artículos (superior derecha); distribución de elementos donde se producen los fallos analizados con señales de vibración (inferior izquierda) como con señales de audio (inferior derecha)....... 289 Figura 4: Componentes de un cojinete cuyo elemento de rodamiento son bolas. ........ 295 Figura 5: Partes principales de una bomba centrífuga. Dirección del fluido en la bomba (figura izquierda). Fuente: [32]. Fluid in: entrada del fluido. Fluid out: salida del fluido. Velocidad y presión del fluido en una bomba. Figura modificada de la fuente: [34]. .......................................................................................................... 296 Figura 6: Partes de un rodete. Figura modificada de la fuente: [32]. ........................... 297 Figura 7: Esquemático de la sala donde se realiza la grabación de la base de datos junto con la posición y esquema del montaje. ............................................................... 298 Figura 8: Fotos de la bomba de agua modelo ALP 800. .............................................. 298 Figura 9: Fotos de la bomba de agua modelo ALP 800. Rodete montado en el rotor (izquierda). Voluta, entrada y salida de la bomba (derecha). ............................... 299 Figura 10: Fotos del fallo en el plato del rodete. Rodete sin fallo (superior izquierda), rodete con una parte del plato eliminada (superior derecha), rodete con dos partes de plato eliminadas (inferior izquierda) y rodete con tres partes de plato aliminadas (inferior derecha). ................................................................................................. 300 Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria xxv Figura 11: Foto del rodete con fallo en el leading edge. En la imagen se observa un recorte del borde leading edge de 5mm (fallo leve). ............................................ 301 Figura 12: Foto del rodete con fallo en el trailing edge. En la imagen se observa un recorte del borde trailing edge de 5mm (fallo leve). ............................................ 301 Figura 13: Posición de los micrófonos y de los acelerómetros. ................................... 302 Figura 14: Esquema de la experimentación en diagnóstico de fallos de cojinetes. ...... 305 Figura 15: Diagrama de la aplicación del método propuesto a la degradación de un cojinete. ................................................................................................................ 307 Figura 16: Método propuesto para la identificación de fallos en cojinetes. ................. 308 Figura 17: Forma de onda de la wavelet madre Daubechies 6. .................................... 309 Figura 18: Esquema de la metodología llevada a cabo para cuantificar la habilidad de las características extraídas de la señal de audio y vibración en diagnóstico de fallos en bombas centrífugas. .............................................................................................. 311 Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 33 Centrifugal pumps are important elements in several systems because they transport fluids. They are used in several industries such as power stations, chemical industry, water depuration systems, oil extraction, cooling and heating systems and even in spas and swimming pools. So, the condition monitoring of pumps is also very important for the operational continuity. Vibration signals are widely used as source of information in condition monitoring of different machines and component of machines. Bearings and centrifugal pumps are not an exception. Vibration-based fault diagnosis refers to fault diagnosis using vibration as information source. Vibration-based fault diagnosis is a wellestablished field that includes a wide range of techniques which have rapidly evolved during the last decades. In condition monitoring, vibration based fault diagnosis techniques have been widely used due to the easiness to acquire the vibration of the machine. That is why most research papers in fault diagnosis literature is devoted to vibation fault diagnosis [1]. Audio-based or airborne-based fault diagnosis refers to fault diagnosis using audio as information source. Audio-based fault diagnosis using microphones in the audible range (0-20kHz) is an emerging field with a great potential in fault diagnosis since microphones are non-invasive sensors and with greater location possibilities. For this reason, we think audio-based techniques need further research efforts. In short, we focus our efforts in this Ph.D. Thesis on improvements in both audio-based and vibration-based fault diagnosis in condition monitoring in bearings and centrifugal pumps. In each application area we have mainly focused on the processing stage of a condition monitoring squeme (see Figure 1-1), specifically in feature extraction. In general, time, frequency and time-frequency domains are frequenctly used for feature extraction. Recently, nonlinear techniques such as nonlinear dynamics applied to time series and complexity measures have appeared in fault diagnosis of certain machines. We think that the study of such features in condition moniroting can improve the performance of a condition monitoring system. There exist a vast literature of different methods to bearing fault diagnosis and fault identification and in bearing degradation assessment. There are also some available PhD Dissertation Universidad de Las Palmas de Gran Canaria 34 public databases of bearing vibration signals in normal (free-fault) and fault conditions. For this reason, in bearing fault diagnosis we have used only vibration signal from the public available databases. We have proposed improvements in nonlinear features aiming at detecting different bearing faults and also at the identification of faults. So, we have made contributions in the signal processing (feature extraction) stage. In pump fault diagnosis we have recorded our own database using audio and vibration signals simultenously acquired with the aim of comparing and combining vibration and audio signals as source of information because there are no public available databases of centrifugal pumps. As we stated before, most condition monitoring is carried out using vibration signals and pump condition monitoring is not an exception. Most research in centrifugal pump condition moniroting is focused in energy and statistical features extracted from time and frequency domain of vibration signals as well as pression signals. In this Thesis, we explore the use of audio signals in pump fault diagnosis, propose new features and study the combination of information from audio and vibration signals. Specifically, we have proposed three improvements not previously addressed in the literature: i) the application of features originally used with vibration signals to audio signal, ii) the proposal of new features extracted from frequency, cepstrum, time-frequency domains as well as complexity features and features related to nonlinear dynamics applied to time series in vibration and audio pump fault diagnosis and iii) a study of the combination features extracted from vibration signals and features extracted from audio signals. So, we have made contributions to the acquisition stage, only in the way there is no public audio database of centrifugal pumps available and to the processing stage (feature extraction) applying vibration features to audio signals, proposing a set of new features in pump fault diagnosis and combining information from audio and vibration signals. 1.4 Motivation of the Thesis This Thesis is focused on two aspects of condition monitoring: the signal used to extract the information about the condition of the machine and the feature extraction with signal processing techniques. The aim is fault diagnosis and fault identification (fault degradation). The two application areas are: bearings and pumps. In bearings only Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 35 vibration signal is used and in centrifugal pumps audio and vibration signals are used as source of information. The research carried out in this Thesis is motivated by the following observations from the state-of-the art: .Vibration-based monitoring is a consolidated field in CBM due to the easiness to acquire the vibration of the machine [2]-[4]. In bearing fault diagnosis, there are multiple signal processing techniques to extract features in order to discriminate between different bearing conditions and to follow the degradation of a bearing fault [5]. The early detection of a fault is very important in condition monitoring to prevent the fault develops. Moreover, nonlinear features extracted from the machine's signature can reveal new understanding of the signal under consideration and the application of nonlinear features can improve the task of fault diagnosis and fault degration [8]. For this reason, we have proposed improvements in nonlinear features aiming at detecting different bearing faults (fault diagnosis) and also at following the degradation of bearings at early stages. .Audio-based monitoring (also called airborne-based monitoring) is a less developed field than vibration-based monitoring. The main reason is that audio signal can be affected by surrounding noise. However, microphones have more possibilities of location and are not mounted on the machine [6], [7]. For this reason, we aim to explore audio-based fault diagnosis. In order to study audio signals and to compare them with vibration signals, audio and vibration signals are acquired from an experimental set of a centrifugal pump. .There is a lack of public available databases of audio signals acquired from machines for audio-based fault diagnosis. Moreover, there are only few works related to audio and vibration signals acquired together. These are more reasons for recording the database of audio and vibration signals acquired simoultaneously from an experimental set of a centrifugal pump. .As audio signals are less used in fault diagnosis literature, most features extracted from vibration signals have not previously used in audio signals. In this Thesis, we address this issue applying state-of-the-art vibration features into audio PhD Dissertation Universidad de Las Palmas de Gran Canaria 36 signals in the centrifugal pump application in order to study their discrimination ability between different condition of the centrifugal pump (normal and fault conditions). .In pump fault diagnosis literature, most features are extracted from time, frequency and time-frequency domain [9], [10]. In this Thesis, a set of new features extracted from frequency, cepstrum, time-frequency domains as well as complexity features and features based on nonlinear dynamics applied to time series are extracted from both vibration and audio signals. These features along with the state-of-the-art features are evaluated in order to study their discrimination ability between different pump conditions using two classifiers. .There are few works in literature using vibration and audio signals acquired simoulstaneously [11], [12]. For this reason, in this Thesis the fusion of both vibration and audio signals at feature level, score level and decision level is carried out in order to evaluate whether the fusion improves the performance of the system. These observations will be discussed in Chapter 2, in which the Thesis problem is analized in depth. 1.5 The Thesis The thesis developed in this Dissertation can be stated as follows: The use of audio signals as source of information and the application of nonlinear techniques improves condition monitoring performance. 1.6 Objectives of the Thesis This PhD Thesis seeks to improve the performance of condition monitoring systems in fault diagnosis and fault identification using vibration and audio signals in two applications (bearings and pumps) with special emphasis in the feature extraction stage and in the use of audio signals as source of information. Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 37 The main objetives of the PhD Thesis are as follows: 1. To review signal processing techniques and pattern recognition techniques in audiobased and vibration-based fault diagnosis. 2. To obtain vibration bearing databases of free-fault bearings and bearings with faults. 3. To study nonlinear techniques for bearing fault diagnosis and bearing fault identification. 4. To create a database of audio and vibration signals simultaneously recorded from a centrifugal circulating pump with normal (free-fault) and fault conditions. 5. To apply features extracted from vibration signals in pump fault diagnosis to audio signals obtained from a centrifugal pump. 6. To generate new features to discriminate between different pump conditions in pump fault diagnosis. 7. To compare vibration and audio features performances in the centrifucal pump application. 8. To study the combination of audio and vibration signals in the centrifugal pump application. 1.7 Methodology of the Thesis The methodology of this Thesis is divided in the following steps: Provided there are no available public audio databases, we record our own audio database acquiring at the same time vibration signals. We record simultaneously audio and vibration signals from a centrifugal pump in a closed loop. Normal and different PhD Dissertation Universidad de Las Palmas de Gran Canaria 38 faults conditions were recorded. We also collect bearing vibration databases publicly available from different internet repositories. Once the bearing vibration databases are collected and the audio and vibration database from the centrifugal pump is acquired in our laboratory, we carry out different experiments for bearing application and for pump application. For bearing application, two new proposed methods based on nonlinear features are applied to the vibration signals from the bearings for fault diagnosis and fault identification. The proposed methods are compared with methods in the literature. The ability of the proposed features for bearing fault diagnosis is quantyfied using two classifiers: a neural network classifier and a Least-Square Support Vector Machine (LSSVM) classifier. In the case of bearing identification, an index to follow the degradation of bearings with outer race fault and different severities is proposed. For centrifugal pump application, features from the state of the art of pump fault diagnosis are implemented and extracted from vibration signals of the centrifugal pump. Then, these features are applied to the audio acquired from the centrifugal pump. Features in frequency domain, cepstrum domain, time-frequency domain and nonlinear features are proposed for pump condition monitoring. Feature selection is implemented for select relevant features in pump fault diagnosis. We study the relevance of the selected features and the performance of vibration and audio signals using two standard classifiers (LS-SVM and neural networks). Finally we analyze how the fusion of audio signals to vibration signals affects the monitoring performance. 1.8 Outline of the Dissertation The Dissertation is structured according to a traditional complex type [13] with literature review and two different applications in which methods are explained and applied to experimental studies. The chapter structure is as follows: .Chapter 1 introduces the topic of condition monitoring and gives the motivation, outline, methodology and contribution of this PhD Thesis. Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 39 .Chapter 2 summarizes related works which have given rise to the motivation of the Thesis and details the motivation of this Thesis based on these previous works. The state of the art is a contribution of this Thesis. The writing of Chapter 2 is based on a paper published by the author of this Thesis. .Chapter 3 is devoted to the application areas of this Thesis, namely bearings and centrifugal pumps. A brief description of bearings basics followed by a description of the public bearing vibration databases used in this Thesis is carried out in the first part of the chapter. The second part is focused on centrifugal pump application. Pump basics and the acquisition process of the audio and vibration database are described. We contribute with the recording of an audio and vibration database of a centrifugal pump. .Chapter 4 shows the contributions in bearing fault diagnosis and fault identification using vibration signals. Two methods based on nonlinear techniques are proposed and the results of the application of each method are shown. The contributions of this Chapter are the proposed new methods based on nonlinear features for bearing fault diagnosis and identification. The writing of Chapter 4 is based on three publications by the author of this Thesis. .Chapter 5 is devoted to the contributions in the centrifugal pump application for fault diagnosis using audio and vibration signals. The features used in pump fault diagnosis literature are described as well as the features proposed in this Thesis for pump fault diagnosis. We contribuite with the application of vibration features to audio signals and with the proposal of a set of new features in frequency, cepstrum, timefrequency and nonlinear domains for pump fault diagnosis. The results of the feature evaluation with two classifiers are also shown. Part of the writing of Chapter 5 is based on two publications of the author of this Thesis. .Chapter 6 shows the results of audio and vibration fusion at feature level, score level and decision level in the pump application. We contribute with a study of the combination of audio and vibration signals in pump fault diagnosis. PhD Dissertation Universidad de Las Palmas de Gran Canaria 40 .Chapter 7 concludes the Dissertation summarizing the main results obtained and outlining future research lines. Chapters 5, 6 and 7 have an introductory part in which the methodology is explained, a second parte in which the methodology is applied to the databases and a third part with the results and conclusions. .The Appendix of the Thesis is an extra chapter that shows a summary of the work carried out in voice pathology detection and in the discrimination between emotional states in speech during the Thesis. As the background of the PhD candidate is voice characterization using nonlinear featres, during the Thesis she continued this research line. The dependence among the chapters is illustrated in Figure 1-2. For example, before reading any of the Chapters 4, 5, and 6, one should read first Chapters 3 and 41 Before Chapter 3 one should start with the introduction in Chapter 1, and it is recommended to read also Chapter 2. Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 41 Figure 1-2: Dependence among Dissertation chapters. Chapter 1: “Introduction” Chapter 2: “State of the art in machinery condition monitoring using vibration and audio signals” ” Chapter 3: “Databases: bearings and pumps” Chapter 4: “Contributions to vibration bearing fault diagnosis and fault identification” Chapter 5: “Contributions to audio and vibration pump monitoring” Chapter 6: “Results of audio and vibation fusion in pump monitoring” Chapter 7: “Conclusions” Appendix: “Contributions to voice” Preceeding block is required Preceeding block is recomended PhD Dissertation Universidad de Las Palmas de Gran Canaria 42 1.9 Research contributions The research contributions of this PhD Thesis are divided into condition monitoring contributions and voice contributions. Provided that the research background of the PhD candidate is voice characterization and she kept working on this subject during her PhD, the publications in voice are also presente here. Journal papers included in ISI JCR appear in bold. 1.9.1 Research contributions in condition monitoring 1. Literature review in audio and vibration fault diagnosis techniques focusing on feature extraction and pattern recognition techniques. Henríquez, P., Alonso, J. B., Ferrer, M., & Travieso, C. M. (2014). Review of automatic fault diagnosis systems using audio and vibration signals. Systems, Man, and Cybernetics: Systems, IEEE Transactions on, 44(5), 642-652. 2. A method based on Teager-Kaiser energy operator and statistic and energy features for bearing fault diagnosis and application of the proposal to bearing degradation in a helicopter. Henríquez, P., Alonso, J. B., Ferrer, M. A., & Travieso, C. M. (2013). Application of the Teager–Kaiser energy operator in bearing fault diagnosis.ISA transactions, 52(2), 278284. Henríquez, P., White, P., Alonso, J. B., Ferrer. M. A. (2011, October). Application of TeagerKaiser Energy Operator to the Analysis of Degradation of a Helicopter Input Pinion Bearing. In Proc. of the International Conference Surveillance 6 (pp. 265-274), University of Technology of Compiègne, France. Henríquez, P., Alonso, J. B., Ferrer, M. A., Travieso, C. M. (2011, May-June). Application of Higher Order Statistics of Teager-Kaiser Energy Transformed Vibration Signal for Bearing Fault Diagnosis. In Proc. of the 24th Int. Congress on Condition Monitoring and Diagnostics Engineering Management (pp. 265-274), Stavanger, Norway. Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 49 Figure 2-2: Distributions of referenced papers: Audio-based and vibration-based fault diagnosis distribution (upper left). Publishing years distribution (upper right).Distribution of faults in vibrationbased diagnosis (bottom left) and in audio-based diagnosis (bottom right). The remainder of this chapter is divided into five subsections. The second, third and fourth subsections focus on the steps of a CM system (acquisition, data processing and decision stage). Subsection 2.5 reports on some examples of CM systems. Finally, an analysis of the gaps of the state of the arte is carried out. 2.2 Acquisition Stage Good quality and precision of the signal in the acquisition stage is essential for posterior analysis and feature extraction. In this Thesis, we focus on vibration and audio signals acquired by accelerometers and microphones, respectively. The dynamic forces within a machine produce compression and bending waves. This vibration pattern changes when an incipient failure starts to evolve. Thus, the analysis of vibration signals is a useful tool for establishing the machine’s condition. In PhD Dissertation Universidad de Las Palmas de Gran Canaria 50 order to detect these signals, vibration sensors are mounted directly onto the machine. There are different kinds of sensors depending on the frequency range: position sensors (0Hz-10kHz), velocity sensors (10Hz-1kHz) and accelerometers (8Hz-15kHz). Piezoelectric accelerometers are popular because of their higher dynamic range of frequencies, reliability, robustness and smaller dimensions. The number and location of vibration sensors is an important issue discussed in [8]. The acoustic characteristics of a machine change when a fault evolves. Consequently, the sound of a machine carries information about its condition. Extracting the sound signature of the machine is a useful tool in fault diagnosis [1]. We focus our review on audio signals obtained with microphones. They usually acquire sounds in the 0Hz-20kHz range. Some microphones can even acquire signals above 100kHz. Microphones are not mounted directly onto the machine. As a result, they are less intrusive than vibration sensors but they are more sensitive to environmental noise. For this reason, microphones must be pointed to the machine or system under consideration and should be placed from 2 cm to 10 cm from the wanted source [1], [12]. 2.3 Processing Stage Signal processing transforms original signals into useful features to accomplish fault diagnosis. These features should be independent of the normal machine operating conditions (variations of load and speed) and extraneous noise and be sensitive only to machinery faults. This section is divided into vibration and audio signals analysis. Table 2-1 shows some of the most common vibration and audio features discussed in the literature since 2000 up to 2015. 2.3.1 Vibration signal processing The main processing techniques applied to vibration signals are based on: time analysis, frequency and cepstral analysis, time-frequency analysis and non-linear analysis. Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 51 Time domain analysis Signal processing in the time domain extracts information from the vibration signal as a function of time. Conventional techniques include several time features such as root mean square, crest factor [35], variance, skewness, kurtosis and higher order moments [68],[93]. Some features, such as crest factor, kurtosis, impulse and clearance factors, do not vary with load and speed variations and are good indicators for impulsive faults [35], [60] especially in bearings. Other conventional techniques are time averaging methods, including time synchronous average (TSA), residual signal and difference signal, all of which are powerful tools in the detection of gear faults [36], [37], [102]. TSA removes background noise and periodic events that are not synchronous with the gear of interest. The resulting signal is used for posterior advanced analysis [80]. Autoregressive (AR) modelling [30] and autoregressive moving average (ARMA) modelling [38] have also proven to be efficient tools in modelling transients in the vibration signal. Novel approaches include a modification of the time-varying AR and ARMA models in which the coefficients are updated with the incoming vibration signal [37]. These models are robust to variations in load and speed. Another novel AR approach [102] proposes the load information as an exogenous input. Frequency and cepstral domain analysis Frequency analysis gives information about the periodicity of the signal in the peaks of the frequency and detects harmonics and side-bands. Conventional frequency features, such as the mean and standard deviation of the frequency, the root mean square frequency, peak magnitude, energies and ratios of spectral energies are used in fault diagnosis in pumps, motors and gearboxes [35], [63], [67]. Another conventional technique is envelope analysis (EnA) or the amplitude demodulation technique, used especially in bearings [7], [16], [63] to identify the bearing defect characteristic frequency and also in gears [76]. EnA improves the signal to noise ratio (SNR) and makes the spectral analysis more effective. For a good review of EnA see [16]. EnA is usually applied using the Hilbert transform (HT) [16]. Recently, the skewness information wave (SIW) has been proposed to compute EnA without using the HT [103]. The skewness is computed in small regions of the signal, resulting in a skewness wave. The skewness information wave (SIW) is obtained using the Kullback-Leibler PhD Dissertation Universidad de Las Palmas de Gran Canaria 52 divergence information from the skewness wave of the reference signal (measured far from the diagnostic location) and the diagnosis signal (measured in the diagnostic location). Then the envelope SIW is obtained from their absolute values and the spectrum is computed. This method is shown to be superior to the conventional EnA and more robust to strong background noise. Another novel EnA method is the Teager Energy Operator (TEO) [50], [100]. TEO is a nonlinear operator with higher demodulation precision and needs less calculation than HT. Cepstral analysis, which gives information from the vibration signal as a function of quefrency has, since the nineties, been shown to be effective in fault diagnosis [83]. Power cepstrum gives information about the periodicity of the spectrum and detects harmonics and sideband patterns in the power spectrum. The application of Mel-Frequency Cepstral Coefficients (MFCC), a technique from speech processing, to vibration fault diagnosis [40] was proposed in 2006. MFCC contain both time and frequency information of the signal which makes them more useful for feature extraction in vibration signals. In 2007, a method called minimum variance cepstrum was proposed to detect faulty periodic impulses in bearings in noisy environments [19]. It minimizes the variance of the signal power in its cepstrum representation. Another family of techniques is based on higher order statistics (HOS) in the frequency domain: bispectrum, summed bispectrum and bicoherence have, since the late nineties, been shown to be effective in fault diagnosis in the bearings of induction motors, gearboxes and in flexible rotor systems [17], [18], [39]. These provide more information than the power spectra, in the case of non-Gaussian signals, can detect nonlinear couplings and can explain the origin of certain peaks in the power spectra. For instance, [96] proposed in 2009 the use of HOS in the cepstral domain (bicepstrum) to detect failures in gears. This technique eliminated noise and modulation effects caused in gears. Conventional frequency techniques assume stationarity and linearity and are usually applied in machines working at fixed speeds. However, most machine processes presents non-stationary components in speed-up, speed-down and in several faults. Cyclostationary analysis (a 2nd-order technique in the frequency domain), and timefrequency techniques are more appropriate for non-stationary processes. The periodic Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 53 variation of statistical moments of rotating machinery makes cyclostationarity techniques suitable for early fault detection [84], [98]. Time-Frequency domain analysis Time-frequency analysis extracts information from the vibration signal as a function of time and frequency and overcomes the problems encountered in frequency analysis when analyzing non-stationary events. Some conventional time-frequency techniques include Short time Fourier transform (STFT) (Koo et Kim, 2000) [41], Wigner-Ville distribution (WVD) (Koo et Kim, 2000; Li et Mechefske, 2006) [41], [56] and the directional Choi-Williams distribution [42]. Techniques from the late nineties include Empirical Mode Decomposition (EMD) [35], [43], the Hilbert-Huang transform (HHT) [21], [35], and the Wavelet Transform (WT) [12], [81]. The WT has adjustable window size through the choice of the mother wavelet and different approximation scales. This flexibility makes it suitable for the analysis of non-stationary signals. Multiple features, such as singularity points [20], Lipschitz exponents [13], scalogram [36], energies, statistics and entropies [44], [81], are extracted from continuous WT (CWT), discrete WT (DWT) and Wavelet Packet Transforms (WPT). These are used to detect imbalance, misalignment, spalling, pitting in gears, faulty bearings and OLTC [97]. Contrary to WT, EMD is a self-adaptive method, applied to fault diagnosis for the first time in 1998. EMD decomposes the signal into a sum of intrinsic mode functions (IMFs). The frequency components in each IMF are related to the sampling frequency and to the signal itself, whereas WT is related only to the sampling frequency. However, the number of IMFs cannot be controlled [35], [63], [91], [117]. The HHT uses the EMD to obtain IMFs [21], [35] and then EnA is applied to each IMF using the HT. Many new techniques have, in the last years, been proposed in time-frequency analysis. Most of them are related to the decomposition of the signal into monocomponent AM-FM signals (amplitude modulated and frequency modulated signals) which are then analyzed using EnA techniques, and to the improvements in WT and EMD techniques. These techniques allow a finer analysis of the signal and more accuracy to detect single and compound faults, essential in fault diagnosis. They are discussed next. PhD Dissertation Universidad de Las Palmas de Gran Canaria 54 Spectral Kurtosis (SK) is a spectral statistic reformulated in 2006 for nonstationary signals [104]. SK provides a robust way of detecting incipient faults that produce impulse-like signals, even in the presence of strong noise. SK also offers a way of designing optimal filters for filtering out the mechanical signature of faults using the kurtogram or the fast kurtogram (ways to compute the SK) as a prelude to EnA [105]. Recently, an enhanced kurtogram has been proposed for bearing fault diagnosis in combination with wavelet packet transform [114]. Local mean decomposition (LMD) and improved LMD are also proposed for fault diagnosis [110]. Contrary to the HHT, LMD does not use HT to estimate the envelope but uses the moving average. Product functions are obtained by multiplying the envelope estimates and the FM signal. Generalized demodulation time frequency (GDTF) also decomposes the signal into mono-component AM-FM signals, transforming the original signal into a new space where WT can be applied, therefore obtaining frequencies with physical meaning. GDTF was first proposed for analyzing biomedical signals. The envelope order spectrum technique, which blends GDTF and the spectrum, has also been proposed for fault diagnosis [99]. Iterated Hilbert transform (ItHT) analyzes AM-FM signals using the iterated application of the HT to a filtered version of the amplitude envelope. Amplitude envelopes and instantaneous frequencies are then extracted. ItHT has higher demodulation accuracy and lower complexity than EMD. The combination of ItHT and a smoothed instantaneous frequency estimation has been recently applied to fault diagnosis [106]. Ensemble EMD (EEMD) is a novel technique (2009) that eliminates the mode mixing problem in EMD. In the mode mixing problem the physical meaning of each IMF is unclear and EMD fails to represent the fault characteristics of a signal accurately. EEMD uses a noise-assisted technique to eliminate the mode mixing problem. EEMD and EMD are applied to a rotor fault and in a heavy oil catalytic cracking machine set [43]. Results show that EEMD can extract the fault characteristic information better than EMD. The multi-scale enveloping spectrogram (MuSEnS) [46] algorithm was developed in 2009 using time, scale, and frequency domain information contained in the signal. It decomposes the signal into different wavelet scales and the envelope signal in each scale is calculated, resulting in an ‘‘envelope spectrum’’ [46]. Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 55 TABLE 2-1: FEATURES FOR VIBRATION AND AUDIO SIGNALS Evolution Methods 2000-2006 2007-2015 Time Statistics (2004)[59] (A) (2009)[37] (V) TSA - (2010)[36] (V) AR,ARMA models (2002)[30], (2006)[38] (V) (2010)[102] (V) Frequency Statistics (2004) [59] , (2005) [58], (2006) [56] (A) (2010)[35], (2009)[54], (2006)[63], (2008)[67](V), (2007)[52] (A) EnA (HT) (2000)[26] (A) ; (2007)[77](V) (2008)[63] (V) SIW - (2010)[103] (V) TEO (2007)[100] (V) (2009)[50] (V) Cyclostationarity (2001)[84] (V) (2010)[98] (V) Polyspectrum (HOS) - (2007)[39] (V) Cepstral Cepstrum - (2007)[19] (V) MFCC (2006) [40] (V) (2009)[54] (A) TimeFrequency STFT, WVD, Choi dist. (2000)[41], (2001)[42] (V), (2006)[56] (A) (2009)[54], (2010)[86](A) WT (CWT, DWT, WPT) (2003)[12] (V, A) (2007)[13], (2007)[20], (2007)[39] (V) (2010)[36], (2009)[47](V) (2009)[57](A) SGWT (IWPT) (2007)[44] (V) (2009) [68] (V) RSGWT - (2010) [108] (V) dtcWT - (2010)[94] (V) Generalized S transform - (2011)[107] (V) MuSEnS - (2009)[46](V) EMD, HHT, EEMD (2005)[21] (V) (2010)[35], (2009) [43], (2015)[117] (V) ItHT - (2008)[106] (V) LMD - (2009)[110] (V) GDTF - (2010)[99] (V) SK - (2009)[104], (2009)[105] (2013)[114] (V) Nonlinear Phase portrait, dot pattern (2000)[10] (A), (2003)[11] (V) - - Lyapunov Exponents, CD, fractals (2001)[18], (2000)[23] (V) (2007)[24], (2007)[52], (2008)[51], (2010)[95] (V), (2005)[55] (A,V), (2007)[25] (A) ApEn, Multi-Scale PermEn (2007)[48], (2007)[49], (2015)[116](V) Multiple manifold - (2009)[109] (V) V: Vibration, A: Audio. PhD Dissertation Universidad de Las Palmas de Gran Canaria 56 Second generation WT (SGWT) [111] was proposed in 2006 and overcomes the main shortcoming of the WT (the proper selection of the mother wavelet) because SGWT is realized by a lifting scheme in the time domain (which is not based on the Fourier transform). The improved WPT (IWPT) is also based on SGWT and it is shown to be superior to WPT in extracting the fault characteristics in bearings [44], [68]. An approach that improves the SGWT is the redundant SGWT (RSGWT) proposed in 2009 [108], [111]. RSGWT is time-invariant, contrary to SGWT, which allows capturing more useful fault information. RSGWT outperforms SGWT in extracting transient components in gearbox vibration signals [111]. The dual-tree complex WT (dtcWT) was proposed in 2010 [94] in fault diagnosis. In [94], the authors show that dtcWT outperforms SGWT, fast kurtogram and DWT because dtcWT enhances noise reduction, is approximate time-invariant and can detect multiple fault features simultaneously Another application of the WT includes a novel growth index [47], insensitive to different mother wavelets and levels of decomposition. Finally, the generalized S transform was proposed in 2011 for fault diagnosis. It unifies STFT and WT so as to obtain more satisfactory time-frequency representations than other similar techniques such as STFT, WVD and the S transform. This allows more accurate detection of the bearing fault characteristic [107]. Nonlinear analysis Evidence of a complex and non-linear vibratory system has been found in stator-rotor rub, loose pedestal and unstable oil film faults [18]. Conventional non-linear methods include pseudo-phase portrait, singular spectrum analysis, correlation dimension (CD) [22], [23], fractal dimensions, approximate entropy (ApEn), information entropy [92], mutual information [25] and Lyapunov exponents [14], [109]. CD quantifies the complexity of a time series and has successfully proven to detect rotor-stator rub, loose pedestal faults [23] and bearing faults [22]. Phase portrait shows qualitative differences in a normal gear and in a gear with an early fatigue cracked tooth [10]. ApEn quantifies the regularity of a time series and can effectively indicate the condition in fans [48] and bearings with speed and load variations [49]. A new technique is proposed in [24] to select a proper fractal dimension spectrum less affected by noise. In a more recent paper (2008) [51], a modification of Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 57 the correlation integral is introduced to the real-time fault diagnosis of bearings. Multiple manifold analysis is a novel nonlinear approach [109] (2009) that extracts manifold information from the vibration signals and outperforms conventional nonlinear techniques. A paper of 2010 proposes to compute the fractal dimension using DWT [95]. Fractal features are estimated from the slope of the variances of the DWT in different scales. The multiscale permutation entropy (Multi-Scale PermEn) has recently applied to bearing fault diagnosis [116]. Multi-Scale PermEn computes the permutation entropy across different scales. 2.3.2 Audio signal processing Most research in machinery diagnosis is oriented towards the analysis of vibration signals. Audio-based CM has, however, not developed at the same rate. This is due to the contamination of the sound signal by unwanted sources such as other machines, noisy environments and the structural vibration of the machine itself [2]. This situation makes it difficult to acquire the machine’s signature. Two main options are used to improve the low SNR in audio signals: the use of partial or full enclosure using an anechoic chamber [15], [17], which is an unrealistic approach for real industrial scenarios, or the use of pre-processing de-noising methods, such as wavelet [12] or blind source separation [2], [15], [26]. These techniques can affect the feature extraction stage. The choice of certain parameters, such as the threshold in wavelet techniques, is important for the extraction of the purified signal with the smallest distortion and the highest SNR. Audio-based techniques are useful in certain cases, especially when it is impossible to access the machine. Audio measurements can be performed at a distance from the machine so the use of sensors mounted directly on the machine is avoided. The same processing techniques for vibrations are applied to extract features in audio signals obtained from machines. Statistical time domain features and energy features in the frequency domain [59] are used to detect wear in internal combustion engines and mass unbalance faults in rotary disks. EnA is used in a heavy sizing-press [26] and in the end-test of vacuum cleaner production [58], where the use of vibration and current analysis did not prove useful. As in vibration-based techniques, the audio spectrum is used when the machine is working at constant speeds [10], [56]. Some papers compare the use of Fourier analysis in audio and vibration signals, [10], [87]. PhD Dissertation Universidad de Las Palmas de Gran Canaria 58 These works claim that Fourier analysis is less effective in sound signals because of the low SNR. As in the case of vibration analysis, time-frequency techniques are more appropriate. For example, STFT is applied to identify engine fault frequencies [54]. However, WVD [86] and pseudo WVD [56] are shown to have better results in extracting non-stationary signatures in the bearings of induction motors. Continuous WT, WPT and MFCC are successfully applied to fault diagnosis of engines [15], [52], [57] working at different speeds and run-up conditions. The symmetrised dot pattern technique is based on the visualization of changes in amplitude and frequency of the audio signal. It has been applied to distinguish between normal and faulty fans [10] and in the fault identification of an internal combustion engine working at different speeds [55]. Some papers have focused on the analysis of a group of chaotic measures extracted from audio signals of different machines. Asynchronous changes in audio signals can be detected with chaotic measures [25]. The spectral entropy of audio signals has shown to be more effective than vibration signals in the diagnostic of cavitation [88]. The development of techniques to improve SNR and the use of the same tool, i.e. wavelets, in the de-noising and extraction stages can put audio-based techniques at an advantageous position in CM systems since they are less intrusive than vibrationbased techniques. Multiple features are extracted from vibration and audio signals in CM techniques. However, a unique feature capable of representing the machine condition does not exist. A good characterization requires the extraction of different features from different domains. Thus we need to find, using feature selection, the most suitable ones for the application under consideration. The selection of well-suited features providing fault-related information and the discarding or weakening of irrelevant or redundant features is an important stage in machine CM to improving system performance. Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 65 vibration-based and audio-based automatic fault diagnosis in machinery implementing the stages of a CBM system (data acquisition, signal processing and decision-making). We have paid special attention to recent advancements in signal processing techniques and classification methods. Finally, some examples of CM systems have been presented. Vibration-based monitoring is a well-established technique widely used in CBM [2], [3]. According to Figure 2-2, most research papers consulted for this state of the art are related to vibration-based fault diagnosis. This fact seems reasonable due to two aspects: the easiness to acquire the vibration signal from a machine and the transmission path between the machine or a component of the machine and the sensor is less affected by interferences than in the case of audio-based diagnosis. We can extract another conclusion from Figure 2-2. Most faults are related to bearings. In fact a large motor reliability survey [119] reports that 42% of the faults in large motors (more than 200 horsepower) are related to bearing faults. In small motor the percentage of faults related to faults is 90% [120]. For this reason, research efforts in condition monitoring are focused on bearing fault diagnosis. Most bearing fault diagnosis is done using vibration signal and current signal as source of information. In Figure 2-2 it is also observed that most referenced papers related to bearings use vibration signal for monitoring. Moreover, in the elaboration of this review, we carried out a search of public available databases. We have found three vibration databases public available. For all these reasons, in this Thesis, we focus the first part of our research in methods for bearing fault diagnosis and bearing fault identification using vibration signals. Signal processing techniques have evolved from conventional time and frequency analysis, which assume stationarity and linearity, to more developed techniques that exploit the non-stationary and non-linear nature of faulty signals and of speed-up and speed down processes. These provide a more realistic description of the real condition of the machine. For this reason, in bearing application, our research aims at fault diagnosis and fault degradation using nonlinear techniques. Audio-based monitoring has not been applied to CBM systems to the same degree as vibration-based monitoring (see Figure 2-2), even though microphones are not mounted on the machine and have greater location possibilities. The main reason is the PhD Dissertation Universidad de Las Palmas de Gran Canaria 66 difficulty of recovering the machine’s signature because the signal can be immersed in noise. Some authors have proposed to locate the microphone at a distance from the machine between 2cm and 20cm to avoid unwanted interferences [12]. The application of audio-fault diagnosis techniques in an extensive way will improve greatly the inspection of certain industrial environments in which a permanent CM system is expensive or the mounting of vibration sensors is difficult. For this reason, we think audio-based techniques need further research efforts and we have focused the second part of our research on audio-based fault diagnosis. Provided the fact that there is a lack of large publicly available databases in fault diagnosis and audio signals are not an exception, we have built an experimental set in our laboratory to acquire simultaneously vibration and audio signals from a centrifugal pump working in normal condition and in different fault conditions. In this way, we obtain vibration and audio signals for different machine conditions that we use in different fault diagnosis experiments. Moreover, this database can be made public so other researchers can benefit from it. Public databases provide common data to comparatively evaluate fault diagnosis techniques and CM systems. In this chapter, multiple features from vibration-based diagnosis are described. Most of them are not used in audio-based diagnosis. In this Thesis, we address this issue. We apply vibration features to audio signals and we make a comparison of the performance in both cases. Moreover, in the literature there are only few works related to audio and vibration signals acquired together. In this Thesis, we focus on a study of audio and vibration signals acquired simultaneously from a centrifugal pump. A set of features are extracted from both signals and classifiers are used to discriminate between different machine conditions (normal and faulty conditions). Data fusion of multiple data is a clear research line in fault diagnosis. Multiple signals can be fusioned to obtain a more reliable diagnosis. In this Thesis, fusion of audio and vibration signals from a centrifugal pump is carried out. Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 67 2.7 References [1] Henriquez, P., Alonso, J. B., Ferrer, M., & Travieso, C. M. (2014). Review of automatic fault diagnosis systems using audio and vibration signals. Systems, Man, and Cybernetics: Systems, IEEE Transactions on, 44(5), 642-652. [2] Vilela, R. M., Metrôlho, J. C., & Cardoso, J. C. (2004, May). Machine and industrial monitorization system by analysis of acoustic signatures. In Electrotechnical Conference, 2004. MELECON 2004. Proceedings of the 12th IEEE Mediterranean (Vol. 1, pp. 277-279). IEEE. [3] Basak, D., Tiwari, A., & Das, S. P. (2006, December). Fault diagnosis and condition monitoring of electrical machines-A Review. In Industrial Technology, 2006. ICIT 2006. IEEE International Conference on (pp. 3061-3066). IEEE. [4] Jardine, A. K., Lin, D., & Banjevic, D. (2006). A review on machinery diagnostics and prognostics implementing condition-based maintenance.Mechanical systems and signal processing, 20(7), 1483-1510. [5] Nandi, S., Toliyat, H., & Li, X. (2005). Condition monitoring and fault diagnosis of electrical motors-a review. Energy Conversion, IEEE Transactions on, 20(4), 719-729. [6] Vachtsevanos, G., Lewis, F. L., Roemer, M., Hess, A., and Wu, B. (2006). Intelligent Fault Diagnosis and Prognosis for Engineering Systems, 1st ed. Hoboken, New Jersey: John Wiley & Sons, Inc. [7] Tandon, N., & Choudhury, A. (1999). A review of vibration and acoustic measurement methods for the detection of defects in rolling element bearings.Tribology international, 32(8), 469-480. [8] Doebling, S. W., Farrar, C. R., & Prime, M. B. (1998). A summary review of vibration-based damage identification methods. Shock and vibration digest, 30(2), 91105. [9] Yang, H., Mathew, J., & Ma, L. (2003, Nov). Vibration feature extraction techniques for fault diagnosis of rotating machinery: a literature survey.In Asia-Pacific Vibration Conference, 2003 (pp. 277-279). [10] Shibata, K., Takahashi, A., & Shirai, T. (2000). Fault diagnosis of rotating machinery through visualisation of sound signals. Mechanical Systems and Signal Processing, 14(2), 229-241. [11] Wang, W. J., & Lin, R. M. (2003). The application of pseudo-phase portrait in machine condition monitoring. Journal of sound and vibration, 259(1), 1-16. [12] Baydar, N., & Ball, A. (2003). Detection of gear failures via vibration and acoustic signals using wavelet transform. Mechanical Systems and Signal Processing, 17(4), 787-804. [13] Shanlin, K., Peilin, P., Feng, F., & Guangbin, D. (2007, July). Feature extraction method in fault diagnosis based on wavelet fuzzy network for power system rotating machinery. In Control Conference, 2007. CCC 2007. Chinese(pp. 437-441). IEEE. [14] Müller, P. C., Bajkowski, J., & Söffker, D. (1994). Chaotic motions and fault detection in a cracked rotor. Nonlinear Dynamics, 5(2), 233-254. [15] Zhong, Z. M., Chen, J., Zhong, P., & Wu, J. B. (2006). Application of the blind source separation method to feature extraction of machine sound signals. The International Journal of Advanced Manufacturing Technology, 28(9-10), 855-862. [16] McFadden, P. D., & Smith, J. D. (1984). Vibration monitoring of rolling element bearings by the high-frequency resonance technique—a review. Tribology international, 17(1), 3-10. PhD Dissertation Universidad de Las Palmas de Gran Canaria 68 [17] Yang, D. M., Stronach, A. F., MacConnell, P., & Penman, J. (2002). Third-order spectral techniques for the diagnosis of motor bearing condition using artificial neural networks. Mechanical systems and signal processing, 16(2), 391-411. [18] Wang, W. J., Wu, Z. T., & Chen, J. (2001). Fault identification in rotating machinery using the correlation dimension and bispectra. Nonlinear Dynamics,25(4), 383-393. [19] Choi, Y. C., & Kim, Y. H. (2007). Fault detection in a ball bearing system using minimum variance cepstrum. Measurement Science and Technology, 18(5), 1433. [20] Peng, Z. K., Chu, F. L., & Peter, W. T. (2007). Singularity analysis of the vibration signals by means of wavelet modulus maximal method. Mechanical Systems and Signal Processing, 21(2), 780-794. [21] Peng, Z. K., Peter, W. T., & Chu, F. L. (2005). A comparison study of improved Hilbert–Huang transform and wavelet transform: application to fault diagnosis for rolling bearing. Mechanical systems and signal processing, 19(5), 974-988. [22] Logan, D. B., & Mathew, J. (1996). Using the correlation dimension for vibration fault diagnosis of rolling element bearings—II. Selection of experimental parameters. Mechanical Systems and Signal Processing, 10(3), 251-264. [23] Wang, W., Chen, J., & Wu, Z. (2000). The application of a correlation dimension in large rotating machinery fault diagnosis. Proceedings of the Institution of Mechanical Engineers, Part C: Journal of Mechanical Engineering Science, 214(7), 921-930. [24] Xinmin, T., Baoxiang, D., & Yong, X. (2007, August). Bearings fault diagnosis based on HMM and fractal dimensions spectrum. In Mechatronics and Automation, 2007. ICMA 2007. International Conference on (pp. 1671-1676). IEEE. [25] Henríquez, P., Alonso, J. B., Travieso, C. M., & Ferrer, M. A. (2007, August). Advances in automatic detection of failures in electric machines using audio signals. In Proc. of 11th IASTED Int. Conf. on Artificial Intell. and Soft Computing (pp. 114-119). [26] Li, L., & Qu, L. (2002, December). Machine diagnosis with independent component analysis and envelope analysis. In Industrial Technology, 2002. IEEE ICIT'02. 2002 IEEE International Conference on (Vol. 2, pp. 1360-1364). IEEE. [27] Fugate, M. L., Sohn, H., Farrar, C. R. (2001). Vibration-based damage detection using statistical process control. Mechanical Syst. and Signal Process., 15, 707-721. [28] Bunks, C., McCarthy, D., Al-Ani, T. (2000). Condition-based maintenance of machines using hidden Markov models. Mechanical Syst. and Signal Process., 14, 597612. [29] Li, Z., Wu, Z. , He, Y., Fulei, C. (2005). Hidden Markov model-based fault diagnostics method in speed-up and speed-down process for rotating machinery. Mechanical Syst. and Signal Process., 19, 329-339. [30] Wang, W., & Wong, A. K. (2002). Autoregressive Model-Based Gear Fault Diagnosis. J. of Vibration and Acoustics, 124, 172-179. [31] SAE (Society of Automotive Engineers) E-32, 2007, “Health and Usage Monitoring Metrics, Monitoring the Monitor”, SAE ARP 5783-DRAFT, Feb. 14, 2007. [32] Kurtoglu, T., Mengshoel, O., Poll, S. (2008, October). A Framework for Systematic Benchmarking of Monitoring and Diagnostic Systems. In Int. Conf. on Prognostics and Health Management, (pp. 1-13), Denver. [33] Kurtoglu, T., Narasimhan, S., Poll, S., Garcia, D., Kuhn, L., de Kleer, J., Van Gemund, A., & Feldman, A. (2009). Towards a Framework for Evaluating and Comparing Diagnosis Algorithms. In Proc. of the 20th Int. Workshop on Principles of Diagnosis (DX’09), Stockholm. Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 69 [34] Roemer, M. J., Dzakowic, J., Orsagh, R. F., Byington, C. S., & Vachtsevanos, G. (2005, March). Validation and Verification of Prognostic Health Management Technologies. In IEEE Conf. Proc.( pp. 3941-3947). [35] Lei, Y., Zuo, M. J., He, Z., Zi, Y. (2010). A multidimensional hybrid intelligent method for gear fault diagnosis. Expert Syst. with Appl., 37, 1419-1430. [36] Wang, X., Makis, V., Yang, M. (2010). A wavelet approach to fault diagnosis of a gearbox under varying load conditions. J. of Sound and Vibration, 329, 1570–1585. [37] Shao, Y., Mechefske, C. K. (2009). Gearbox vibration monitoring using extended Kalman filters and hypothesis tests. J. of Sound and Vibration, 325, 629–648. [38] Wang, F. L., Mechefske, C. K. (2006). Adaptive modelling of transient vibration signals. Mechanical Syst. and Signal Process., 20, 825-842. [39] Dou, W., Liu, Z.-S., Wang, D.-H. (2007, August). Combination Diagnosis Based on Genetic Algorithm for Rotating Machinery. In Proc. of the 3rd Int. Conf. on Natural Computation (vol. 4, pp. 307-313). IEEE. [40] Nelwamondo, F. V., & Marwala, T. (2006, October). Faults Detection Using Gaussian Mixture Models, Mel-Frequency Cepstral Coefficients and Kurtosis. In IEEE Int. Conf. on Syst., Man, and Cybern. (vol. 1, pp. 290-295), Taiwan. IEEE. [41] Koo, I. S., Kim, W. W. (2000). The development of reactor coolant pump vibration monitoring and a diagnostic system in the nuclear power plant. ISA Trans., 39, 309-316. [42] Lee, S. U., Robb, D., Besant, C. (2001) .The Direction Choi-Williams Distribution for the analysis of Rotor-Vibration Signals. Mechanical Syst. and Signal Process., 15(4), 789-811. [43] Lei, Y., He, Z., Zi, Y. (2009).Application of the EEMD method to rotor fault diagnosis of rotating machinery. Mechanical Syst. and Signal Process., 23, 1327–1338. [44] Hu, Q., He, Z., Zhang, Z., Zi, Y. (2007). Fault diagnosis of rotating machinery based on improved wavelet package transform and SVMs ensemble. Mechanical Syst. and Signal Process., 21, 688-705. [45] Yu, G., Li, C., Kamarthi, S. (2009). Machine fault diagnosis using a clusterbased wavelet feature extraction and probabilistic neural networks. Int. J. Adv. Manuf. Technol., 42, 145–151. [46] Yan, R. Gao, R. X. (2009). Multi-scale enveloping spectrogram for vibration analysis in bearing defect diagnosis. Tribology Int., 42, 293-302. [47] Wang, D., Miao, Q., Kang, R. (2009). Robust health evaluation of gearbox subject to tooth failure with wavelet decomposition. J. of Sound and Vibration, 324, 1141-1157. [48] Tan, J., Gu, J. -J., Peng, X. -Z., Qin, Z. -M. (2007, August). A fault diagnosis method based on wavelet approximate entropy for fan. In Proc. of the 6th Int. Conf. on Machine Learning and Cybern. (pp. 519-523), Hong Kong. [49] Yan, R., Gao, R. X. (2007). Approximate Entropy as a diagnostic tool for machine health monitoring. Mechanical Syst. and Signal Process., 21, 824-839. [50] Li, H., Fu, L., Zhang, Y. (2009, April). Bearing Faults Diagnosis Based on Teager Energy Operator Demodulation Technique. Int. Conf. on Meas. Technol. and Mechatronics Automation, (pp. 594-597), Zhangjiajie, China. [51] Janjarasjitt, S., Ocak, H., Loparo, K. A. (2008). Bearing condition diagnosis and prognosis using applied nonlinear dynamical analysis of machine vibration signal. J. of Sound and Vibration, 317, 112-126. [52] Yang, J., Zhang, Y., Zhu, Y. (2007). Intelligent fault diagnosis of rolling element bearing based on SVMs and fractal dimension. Mechanical Syst. and Signal Process., 21, 2012-2024. PhD Dissertation Universidad de Las Palmas de Gran Canaria 70 [53] Singh, V., Meena, N. (2009, January). Engine Fault Diagnosis using DTW, MFCC and FFT. In Proc. of the 1st Int. Conf. on Intelligent Human Computer Interaction, (pp. 83-94). [54] Mansoor, A. B., Ahmed, H., Mahmood, Z. (2009, February). Design and Development of an Automated Acoustic based Jet Engine Performance Evaluator. In Proc. of the IEEE Int. Conf. on Industrial Technol. (pp. 1-4). [55] Wu, J. -D., Chuang, C. -Q. (2005). Fault diagnosis of internal combustion engines using visual dot patterns of acoustic and vibration signals. NDT&E Int., 38, 605-614. [56] Li, W., Mechefske, C. K. (2006). Detection of Induction Motor Faults: A Comparison of Stator Current, Vibration and Acoustic Methods. J. of Vibration and Control, 12(2), 165-188. [57] Wu, J. -D., Liu, C. -H. (2009). An expert system for fault diagnosis in internal combustion engines using wavelet packet transform and neural network. Expert Syst. with Appl., 36, 4278-4286. [58] Benko, U., Petrovcic, J., Juricic, D., Tavcar, J., Rejec, J. (2005). An approach to fault diagnosis of vacuum cleaner motors based on sound analysis. Mechanical Syst. and Signal Process., 19, 427-445. [59] Li, W., Tsai, Y. P., Chiu, C. L. (2004). The experimental study of the expert system for diagnosing unbalances by ANN and acoustic signals. J. of Sound and Vibration, 272, 69–83. [60] Sakthivel, N. R., Sugumaran, V., & Nair, B. B. (2010). Comparison of decision tree-fuzzy and rough set-fuzzy methods for fault categorization of mono-block centrifugal pump. Mechanical Syst. and Signal Process., 24, 1887-1906. [61] Ballal, P., Ramani, A., Middelton, M., McMurrough, C., Athamneh, A., Lee, W., & Lewis, F. (2009, March). Mechanical Fault Diagnosis using Wireless Sensor Networks and a Two-Stage Neural Network Classifier. In IEEE Aerosp. Conf, (pp. 110). IEEE. [62] Li, Z., He, Y., Chu, F., Han, J., Hao, W. (2006). Fault recognition method for speed-up and speed-down process of rotating machinery based on independent component analysis and Factorial Hidden Markov Model. J. of Sound and Vibration, 291, 60-71. [63] Lei, Y., He, Z., Yi, Y. (2008). A new approach to intelligent fault diagnosis of rotating machinery. Expert Syst. with Appl., 35, 1593-1600. [64] Wei, L., Pu, H. (2008, July) .Wavelet Neural Network Aided On-Line Detection and Diagnosis of Rotating Machine Fault. In Chinese Control and Decision Conf., 2008 (pp.1868-1871). IEEE. [65] Kar, C., Mohanty, A. R. (2004). Application of KS test in ball bearing fault diagnosis. J. of Sound and Vibration, 269, 439-454. [66] Goumas, S. K., Zervakis, M. E., & Stavrakakis, G. S. (2002). Classification of Washing Machines Vibration Signals Using Discrete Wavelet Analysis for Feature Extraction. IEEE Trans. on Instrumentation and Meas., 51(3), 497-508. [67] Lei, Y., He, Z., Zi, Y., Chen, X. (2008). New clustering algorithm-based fault diagnosis using compensation distance evaluation technique. Mechanical Syst. and Signal Process., 22, 419-435. [68] Pan, Y., Chen, J., Guo, L. (2009). Robust bearing performance degradation assessment method based on improved wavelet packet– support vector data description. Mechanical Syst. and Signal Process., 23, 669-681. Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 71 [69] Sanz, J., Perera, R., Huerta, C. (2007). Fault diagnosis of rotating machinery based on auto-associative neural networks and wavelet transforms. J. of Sound and Vibration, 302, 981–999. [70] Shina, H. J., Eomb, D. -H., Kim, S. -S. (2005). One-class support vector machines—an application in machine fault detection and classification. Computers & Industrial Eng., 48, 395-408. [71] Case Western Reserve University Bearing Test Data Center. http://www.eecs.case.edu/laboratory/bearing/ (Last visited: Jan. 2014). [72] http://qsun.eng.ua.edu/cpw_web_new/intro.htm (Last visited: Jan. 2014) [73] MDTB Test-run Data CDs, Condition-based Maintenance Department, Applied Research Laboratory, The Pennsylvania State University. [74] Ypma A., Ligteringen, R.P.W., Duin, E.E.E., Frietman, E.E.E. (1999). Pump vibration datasets, Pattern recognition group, Delf University of Technology. [75] http://ti.arc.nasa.gov/tech/dash/pcoe/prognostic-data-repository/ (Last visited: Jan. 2012) [76] Wang, W., Kanneg, D. (2009). An integrated classifier for gear system monitoring. Mechanical Syst. and Signal Process., 23, 1298-1312. [77] Zhang, L., Yan, R., Gao, R. X., Lee, K. (2007, March). Design of a Real-time Spindle Health Monitoring and Diagnosis System Based on Open Systems Architecture. In Int. Smart Machining Syst. Conf., France (pp. 373-378). [78] Cho, S., Binsaeid, S., Asfour, S. (2010). Design of multisensor fusion-based tool condition monitoring system in end milling. Int. J. Adv. Manuf. Technol., 46, 681-694. [79] Hitchcock, L. (2006). ISO Standards for Condition Monitoring. Engineering Asset Management. Springer London, pp. 606-613. [80] Lebold, M., McClintic, K., Campbell, R., Byington, C., Maynard, K. (2000, May). Review of Vibration Analysis Methods for Gearbox Diagnostics and Prognostics. In Proc. of the 54th Meeting of the Society for Machinery Failure Prevention Technol.(Vol. 643, p. 16). [81] Rafiee, J., Rafiee, M. A., Tse, P. W. (2010). Application of mother wavelet functions for automatic gear and bearing fault diagnosis. Expert Syst. with Appl., 37, 4568-4579. [82] Farrar, C. R., Duffey, T. A., Doebling, S. W., Nix, D. A. (1999, September). A Statistical Pattern Recognition Paradigm for Vibration-Based Structural Health Monitoring. Structural Health Monitoring, 2000, 764-773. [83] Tandon, N. (1994). A comparison of some vibration parameters for the condition monitoring of rolling element bearings. Meas., 12, 285-289. [84] Bouillaut, L., Sidahmed, M. (2001).Cyclostationary approach and bilinear approach: Comparison, applications to early diagnosis for helicopter gearbox and classification method based on hocs. Mechanical Syst. and Signal Process., 15, 923943. [85] Wu, S.-j., Gebraeel, N., Lawley, M. A. & Yih, Y. (2007). A Neural Network Integrated Decision Support System for Condition-Based Optimal Predictive Maintenance Policy. IEEE Trans. On Syst., Man and Cybern.-Part A: Syst. and Humans, 37, 226-236. [86] Albarbar, A., Gu, F., Ball, A. D. (2010). Diesel engine fuel injection monitoring using acoustic measurements and independent component analysis. Meas., 43(10), 1376-1386. [87] Devi S., Siva Kumar, Dr. L., Shanker N. R., & Prabakaran, K. (2010). A Comparative Study between Vibration and Acoustic Signals in HTC Cooling Pump and Chilling Pump. Int. J. of Eng. and Technol., 2. PhD Dissertation Universidad de Las Palmas de Gran Canaria 72 [88] Al Thobiani, A., Gu, F. , & Ball, A. (2010, June). The monitoring of cavitation in centrifugal pumps based on the analysis of vibro-acoustic measurements. In CM 2010 and MFPT 2010, UK. [89] Wang, W. (2008). An Enhanced Diagnostic System for Gear System Monitoring. IEEE Trans. on Syst., Man, and Cybern.-Part B: Cybern., vol. 38(1), 102112. [90] Todd, M., McArthur, S. D. J., McDonald, J. R., & Shaw, S. J. (2007). A Semiautomatic Approach to Deriving Turbine Generator Diagnostic Knowledge. IEEE Trans. on Syst., Man, and Cybern.-Part C: Appl. and Reviews, 37(5). [91] Mu, T., & Nandi, A. K. (2009). Multiclass Classification Based on Extended Support Vector Data Description. IEEE Trans. on Syst., Man, and Cybern.-Part B: Cybern., 39(5), 1206-1216. [92] Chen, Z. (2010, May). Fault Diagnosis of Gear Box Based on Information Entropy. Control and Decision Conference (pp. 1239-1242). [93] Kankar, P. K., Sharma, S. C., Harsha, S. P. (2011). Fault diagnosis of ball bearings using machine learning methods. Expert Syst. with Appl., 38, 1876-1886. [94] Wang, Y. , He, Z., Zi, Y. (2010). Enhancement of signal denoising and multiple fault signatures detecting in rotating machinery using dual-tree complex wavelet transform. Mechanical Syst. and Signal Process., 24, 119-137. [95] Li, P., He, Q., & Kong, F. (2010, June). An Approach for Fault Diagnosis of Bearings Using Wavelet-Based Fractal Analysis. In Proc. of the IEEE Int. Conf. on Information and Automation (pp. 2338-2343). IEEE. [96] Fu, L., Li, H., Wang, Y. (2009, April). Application of Bi-cepstrum Analysis to Gear Fault Detection and Diagnosis. In 2009 Int. Conf. on Meas. Technol. and Mechatronics Automation (pp. 590-593). [97] Rivas, E., Burgos, J. C., & García-Prada, J. C. (2010). Vibration Analysis Using Envelope Wavelet for Detecting Faults in the OLTC Tap Selector. IEEE Trans. on Power Delivery, 25(3) 3,1629-1636. [98] Delvecchio, S., D’Elia, G., Mucchi, E., Dalpiaz, G. (2010). Advanced Signal Processing Tools for the Vibratory Surveillance of Assembly Faults in Diesel Engine Cold Tests. J. of Vibration and Acoustics, 132(2), 021008. [99] Cheng, J., Yang, Y., Yu, D. (2010). The envelope order spectrum based on generalized demodulation time–frequency analysis and its application to gear fault diagnosis. Mechanical Syst. and Signal Process., 24, 508-521. [100] Junsheng, C., Dejie, Y., & Yu, Y.(2007). The application of energy operator demodulation approach based on EMD in machinery fault diagnosis. Mechanical Syst. and Signal Process., 21, 668-677. [101] Sugumaran, V., Muralidharan, V., Ramachandran, K.I. (2007). Feature selection using Decision Tree and classification through Proximal Support Vector Machine for fault diagnostics of roller bearing. Mechanical Syst. and Signal Process., 21, 930-942. [102] Yang, M., Makis, V. (2010). ARX model-based gearbox fault detection and localization under varying load conditions. J. of Sound and Vibration, 329, 5209-5221. [103] Wang, H., Chen, P., Fujioka, S., Wang, S., Li, K. (2010, March). Diagnosis Method Based on Skewness Wave and Information Divergence for a Diesel Engine. In Asia-Pacific Conf. of Power and Energy Eng.( pp. 1-4), Chengdu, China. [104] Antoni, J., Randall, R. B. (2006). The spectral kurtosis: application to the vibratory surveillance and diagnostics of rotating machines. Mechanical Syst. and Signal Process., 20, 308-331. Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 73 [105] Barszcz, T., Randall, R. B. (2009). Application of spectral kurtosis for detection of a tooth crack in the planetary gear of a wind turbine. Mechanical Syst. and Signal Process., 23, 1352-1365. [106] Qin, Y., Qin, S., & Mao, Y. (2008). Research on iterated Hilbert transform and its application in mechanical fault diagnosis. Mechanical Systems and Signal Processing, 22(8), 1967-1980. [107] Li, B., Zhang, P. L., Liu, D. S., Mi, S. S., Ren, G. Q., & Tian, H. (2011). Feature extraction for rolling element bearing fault diagnosis utilizing generalized S transform and two-dimensional non-negative matrix factorization. Journal of Sound and Vibration, 330(10), 2388-2399. [108] Zhou, R., Bao, W., Li, N., Huang, X., & Yu, D. (2010). Mechanical equipment fault diagnosis based on redundant second generation wavelet packet transform. Digital signal processing, 20(1), 276-288. [109] Li, M., Xu, J., Yang, J., Yang, D., & Wang, D. (2009). Multiple manifolds analysis and its application to fault diagnosis. Mechanical systems and signal processing, 23(8), 2500-2509. [110] Wang, Y., He, Z., & Zi, Y. (2009). A demodulation method based on improved local mean decomposition and its application in rub-impact fault diagnosis.Measurement Science and Technology, 20(2), 025704. [111] Hongkai, J., Zhengjia, H., Chendong, D., & Peng, C. (2006). Gearbox fault diagnosis using adaptive redundant lifting scheme. Mechanical Systems and Signal Processing, 20(8), 1992-2006. [112] Sun, X. L. (2012, February). A new method of fault diagnosis with soft set theory. In Advanced Materials Research (Vol. 383, pp. 7470-7474). [113] Wang, W. (2008). An intelligent system for machinery condition monitoring.Fuzzy Systems, IEEE Transactions on, 16(1), 110-122. [114] Wang, D., Peter, W. T., & Tsui, K. L. (2013). An enhanced Kurtogram method for fault diagnosis of rolling element bearings. Mechanical Systems and Signal Processing, 35(1), 176-199. [115] Moosavian, A., Khazaee, M., Ahmadi, H., Khazaee, M., & Najafi, G. (2015). Fault diagnosis and classification of water pump using adaptive neuro-fuzzy inference system based on vibration signals. Structural Health Monitoring,14(5), 402-410. [116] Tiwari, R., Gupta, V. K., & Kankar, P. K. (2015). Bearing fault diagnosis based on multi-scale permutation entropy and adaptive neuro fuzzy classifier. Journal of Vibration and Control, 21(3), 461-467. [117] Ali, J. B., Fnaiech, N., Saidi, L., Chebel-Morello, B., & Fnaiech, F. (2015). Application of empirical mode decomposition and artificial neural network for automatic bearing fault diagnosis based on vibration signals. Applied Acoustics, 89, 1627. [118] Muralidharan, V., Sugumaran, V., & Indira, V. (2014). Fault diagnosis of monoblock centrifugal pump using SVM. Engineering Science and Technology, an International Journal, 17(3), 152-157. [119] Bell, R.N., McWilliams, D.W., O’Donnell, P., Singh, C., & Wells, S. J. (1985). Report of large motor reliability survey of industrial and commercial installations Part I and II. IEEE Transactions on Industry Applications, 21(4), 853-872. [120] Immovilli, F., Bellini, A., Rubini, R., & Tassoni, C. (2010). Diagnosis of bearing faults in induction machines by vibration or current signals: A critical comparison. IEEE Transactions on Industry Applications, 46(4), 1350-1359. Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 81 Faults were introduced into the drive-end 6205-2RS JEM SKF, deep groove ball bearing using the electro-discharge machining method. Faults of diameter 0.007, 0.014 and 0.021 inch (7 mills, 14 mills and 21 mills respectively) are considered, corresponding to 0.01778 cm, 0.03556 cm and 0.05334 cm respectively. The deep of the fault is 0.011 inch/ 0.02794 cm. Vibration data was collected with an accelerometer attached to the housing with magnetic bases placed at the 12 o’clock position at the drive end. The sample frequency was 12 kHz. Speed and horsepower data were collected using the torque sensor/encoder and recorded by hand. Vibration data was recorded at four different conditions: normal (N), inner race fault (IR), outer race fault (OR) and ball fault (B). Each signal is 10 seconds. Experiments were repeated for motor loads of 0 to 3 horsepower (motor speeds varying from 1797 to 1720 RPM. The higher the load is the lower the speed). The shaft rotating frequency is about 30 Hz. Data consisted of 4 vibration signals for N condition and 12 vibration signals for each fault condition (12 IR, 12 OR and 12 B). In total there are 40 vibration signals. Figure 3-4: Test stand for the acquisition of bearing vibration data. Source [1]. The Case bearing vibration database was used in this Thesis for the evaluation of the following proposals: the proposal of using Teager-Kaiser energy operator to extract PhD Dissertation Universidad de Las Palmas de Gran Canaria 82 statistics, energy and higher order statistics features in order to detect between IR, OR, B and normal condition in bearing fault diagnosis and the proposal of a new methodology based on wavelet package transform and Lempel Ziv complexity to the assessment of bearing severity and degradation (see Chapter 4). UH-60 Blackhawk Helicopter Vibration Database The UH-60 Blackhawk helicopter vibration database was recorded during a component endurance test of an UH-60 Blackhawk helicopter at Patuxent River, M.D. [2]. Figure 3-5 shows the main gearbox transmission system of the UH-60 helicopter. The gearbox transmission system translates the energy from the two engines into the main rotor and it is a complicated system. Figure 3-5: Main transmission of an UH-60 Backhawk helicopter. Source: [27]. The data sets were recorded at irregular intervals throughout the endurance tests with a sample frequency of 100kHz and only recordings at conditions within +/- 10% of full torque are used in the experiments. Vibration signals were acquired using Endevco 6259M31 accelerometers. The vibration database consists of 62 data sets of 10 seconds each. Severe degradation of the inboard roller bearing SB-2205 occurred during the endurance test. The helicopter indicators showed chip lights (a chip light is an indicator Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 83 of the presence of metal chips or particles in the module) in the experiment. The first chip light occurred in the data set number 40, 10200 minutes after the recordings had begun. The position of bearing SB-2205, which supports the combining bevel pinion in one of the input modules, is shown in Figure 3-5 (left). As it can be seen from Figure 3-6 (left), the bearing is located deep inside the gearbox. In this position, the background noise is greater and the detection of a bearing fault is more difficult. The bearing condition at the end of the test is also shown in Figure 3-6 (right). From the figure, it can be seen a fault in the rolling element of the bearing. Figure 3-6: Position of the bearing SB-2205 in the transmission (left figure) and the fnal bearing condition (right figure). Source: [28]. The UH-60 Blackhawk helicopter vibration database was used in this Thesis for the evaluation of the following proposals: the proposal of using Teager-Kaiser energy operator to extract statistics, energy and higher order statistics features in order to detect between IR, OR, B and normal condition in bearing fault diagnosis and the proposal of a new methodology based on wavelet package transform and Lempel Ziv complexity to the assessment of bearing severity and degradation (see Chapter 4). IMS Vibration Bearing Database The IMS vibration bearing database is a bearing data set provided by the Center on Intelligent Maintenance Systems (IMS) [3]. Four bearings were installed on one shaft and two accelerometers were placed in each of them to register the vibration signals in two different spatial axes (see Figure 3-7). The shaft was driven by an AC motor and coupled by rub belts. The rotation speed was kept constant at 2000 rpm and a 6000 lb. radial load was added to the shaft and bearings by a spring mechanism. Vibration data PhD Dissertation Universidad de Las Palmas de Gran Canaria 84 was collected every 10 min for 164 h with a sampling rate of 20 kHz. At the end of the test-to-failure experiment, an outer race defect was discovered on bearing 1. Captures obtained by the horizontal accelerometer of bearing 1 have been used in this Thesis for bearing degradation. According to [34], the first indication of fault is 89 hours after the beginning of the experiment. Figure 3-7: Position of the bearings in the IMS bearing database. Source: [3]. The IMS vibration bearing database was used in this Thesis for the evaluation of the following proposal: the proposal of a new methodology based on wavelet package transform and Lempel Ziv complexity to the assessment of bearing severity and degradation. 3.2 Centrifugal Pump Audio and Vibration Data A pump is defined as a mechanical device that rotates or reciprocates to move fluid from one place to another [5]. There are multiples kinds of pumps with different applications. According to the Hydraulic Institute [6] pumps can be classified according to the manner in which the pump adds energy to the pumped fluid to generate movement into: kinetic pumps and positive displacement pumps. In this Thesis we focus on centrifugal pumps, a kind of kinetic pumps. They add energy by high-speed rotating wheels or impellers. Pumps are important components in a wide range of technical processes such as power stations, chemical industry, cooling and heating systems, etc. The degradation of pump components such as impellers, bearings, seals or the presence of impurities or Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 85 strange objects in the fluid being pumped can reduce the pump performance and lead to faults. The overall reliability and safety of many systems depends on the health of pumps. Therefore, pump condition monitoring plays a key role in maintenance procedures. 3.2.1 Elements and working of a centrifugal pump A centrifugal pump consists of two main components: 1) the rotary element or impeller and 2) the stationary element or casing (called volute). The impeller is the rotating part that converts driver energy (i.e. the energy of a motor) into the kinetic energy. Rotation of the impeller forces the fluid (usually liquid) to circulate through the pump from the axial to the radial direction while energy is transferred to the fluid [7]. The volute is the stationary part that converts the kinetic energy into pressure energy. Summing up, the impeller produces fluid velocity and the volute converts velocity to pressure. Figure 3-8 shows the components of a centrifugal pump. The direction of the fluid in a centrifugal pump (left) and the zones of velocity and pressure of the fluid when passing through the centrifugal pump (right) are also shown. The fluid enters in the suction eye (attached to the inlet part or suction part of the pump) for the inlet side (suction side), then pass through the impeller and finally exits in the outlet (discharge) side of the pump. Figure 3-8: Left: Fluid direction in a centrifugal pump. Source: [6]. Right: velocity and pressure in a centrifugal pump. Modified from source: [5]. The inlet, outlet, volute and impeller are shown. PhD Dissertation Universidad de Las Palmas de Gran Canaria 86 An impeller has a number of channels or vanes delimited by curved blades. A closed impeller (see Figure 3-9) has plates on both sides called hub plate and shroud plate that totally enclose the impeller from the suction eye to its edges. The hub plate is in the front of the impeller (where the impeller is connected to the rotor) and the shroud plate is in the rear of the impeller (in the impeller eye area). The impeller also has several blades (also called vanes) to impart the centrifugal force to the fluid. The center of the impeller is called the impeller eye. The region near the impeller eye is called vane leading edge (LED). The region at the tip of the vane is called vane trailing edge (TED) [14]. Figure 3-9: Closed impeller. Modified from source: [6]. A more detailed description of how the fluid passess through a centrifugal pump is described next [5]. 1. Fluid flows through the pump by first entering the inlet side. Then the fluid enters the lowest pressure area in the pump, the impeller eye. 2. From here the fluid is picked up by the spinning impeller vanes. The fluid passes along the vanes where velocity and energy are added to it. The amount of energy given to the fluid is proportional to the velocity at the edge or vane tip of the impeller. The faster the impeller rotates or the bigger the impeller is, then the higher will be the velocity of the fluid at the vane tip and the greater the energy imparted to the fluid. 3. Through centrifugal force, the fluid is thrown to the outside tips of the impeller, against the volute and toward the discharge flange. At this point, because the fluid is confined by the volute, the velocity decreases thereby increasing the pressure. The fluid Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 87 velocity is decreasing because the volute is shaped in such a way that the impeller is not centered inside it. Rather, the impeller is offset from the center. This offset causes the impeller to volute clearance to increase from the cutwater to the discharge area. As the clearance increases the velocity decreases and the pressure increases. 4. Then the fluid moves through the inside edge of the volute to the discharge flange where it exits the pump at a higher pressure. The velocity of the fluid is converted to pressure according to Bernoulli's principle (Bernoulli's Principle states that as the speed of a moving fluid increases, the pressure within the fluid decreases). Other parts of a centrifugal pump are the pump shaft, bearings, shaft seal, the wear rings, and the inlet and outlet. The shaft seal stops the fluid from leaking out of the casing. The wear rings separate the high and low pressure areas inside the casing. The bearings make the shaft turn easier. The inlet and outlet parts of the casing connect to the fluid piping system. The sealing device is placed inside the stuffing box to control or eliminate leakage from the pump casing. 3.2.2 Vibroacoustic mechanism in a centrifugal pump The pump vibro-acoustics is generated by the following sources: hydraulics sources and mechanical sources [16], [18]. Both sources cause vibrations which make the pump structure vibrate. This vibration radiates airborne sound. Therefore, the acoustics of the pump has the same mechanism of production that the vibration mechanism [18]. Hydraulics sources are caused by the fluid-structure interaction with impeller vanes and volute (especially volute cutwater) and by flow perturbations. Mechanical sources are caused by vibration of unbalanced rotating masses and friction in bearings, seals, impeller and shaft [16]. During the working process of a centrifugal pump non-steady fluid-dynamic forces may produce either discrete or broad-band frequencies in vibration and acoustic signals [8], [15], [16], [18]. The discrete frequencies are: the rotation frequency (RF), the vane-passing frequency (rotor frequency multiplied by the number of vanes or PhD Dissertation Universidad de Las Palmas de Gran Canaria 88 blades of the impeller) and their harmonics. The rotation frequency is present due to pressure pulsation (fluctuations in the pressure being developed by the pump) caused by impeller imbalance [19] (when the impeller has an orbital motion coupled to the rotation [8]) and also by small manufacturing imperfections in the impeller [8]. The vanepassing frequency (VPF) is due to the finite thickness of the blades which causes flow disturbances associated with the passage of each blade near the cutwear or volute tongue [8]. A pressure pulse is developed as each vane passess the cutwater. Broad-band fluid-dynamic excitation is due to pressure pulsations generated by flow turbulence, viscous forces, boundary layer vortex shedding, boundary layer interaction between a higher-velocity and lower velocity regions of the process fluid, and by vortices generated in the clearances between the rotor of the centrifugal pump and the adjacent stationary part of the casing [18]. Mechanical sources such as noise from the rotation of the pump shaft and bearings also contribute to broad-band vibration content. Broad-band content is always present due to turbulence [8]. Other perturbations not related directly with the pump itself can also exist, such as an obstacle or obstruction in the pump inlet, in the pump outlet or even inside the impeller [8]. Other frequencies generated may be related to fan frequency of the motor, motor frequencies or frequencies related to other part of the system in which the pump is operating. In normal condition, the frequencies that dominate the spectrum are the discrete frequencies (RF and VPF and their harmonics). Although broadband noise is always present, in normal condition it has a minimum value. When some of the components of the centrifugal pump have a fault, other frequency components can appear, the previously mentioned frequency can increase, decrease in amplitude or even can dissapear. The broadband noise may also increase and dominate the spectra due to the presence of faults in the pump (such as cavitation for example). 3.2.3 Main faults in a centrifugal pump The main faults in pumps are associated with impeller damages [9], [10], rotor faults, seals faults, cavitation [11] and bearing faults [11], [12]. A classification of the main Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 89 faults in a centrifugal pump might be done according to the component affected by the fault into electrical faults, mechanical faults and hydraulic faults. The electrical faults are associated with the pump motor, mechanical faults are associated with bearings and shaft and hydraulic faults are associated to hydraulic parts of the centrifugal pump. In this Thesis, we have focused on hydraulic faults. The hydraulic parts of the centrifugal pump are the impeller, the blades, the volute, the inlet and the outlet. The functionality of the hydraulic parts is to convert mechanical energy from the shaft to hydraulic energy induced into the liquid pumped by the pump [17]. The hydraulics faults are produced in the impeller, the blades and the volute. The hydraulics faults associated with them are: dry running, impurities fixed on the impeller (causing inbalance), wear of the impeller (leading edge fault, trailing edge fault), blocked or partial blocked flow field inside the impeller, blocked impeller rotation, wear of the sealing ring, missing sealing ring, loss of the impeller, cavitation, instability, rotating stall and pressure pulsations [13], [18], [19], [20]. The main effects of these faults are changes in the value of pressure and the load torque generated by the impeller at a given flow. Moreover some of the faults can induce pressure oscillations. These pressure oscillations can be either harmonics of the rotational frequency, or noise like signals covering a larger frequency span. Cavitation may produce the wear of impellers and piping system. The inlet and the outlet part of the pump and the pump system can be considered apart. The faults in the inlet part of the pump are low pressure and obstruction at the inlet pump. The main effect of these faults is that the inlet pressure to the impeller becomes too low. The faults in the outlet part of the pump include leakage on the outlet pipe and obstruction of the outlet pipe. The main effects of these faults are leakages from the system and decreased pressure produced by the pump. 3.2.4 Considered Faults in a centrifugal pump in this Thesis In this Thesis, we focus on impeller-related faults (leading edge damage, trailing edge damage and plate damage), system faults and seal fault. PhD Dissertation Universidad de Las Palmas de Gran Canaria 90 Impeller-related faults were artificially created cutting the leading edge of all vanes for the leading edge fault and cutting the trailing edge of all vanes for the trailing edge fault [14]. The plate fault was created removing part of the plate. There are certain flow phenomena that occur in the fluid flow when it enters the inlet part of the pump, pass through the impeller and the volute up to the outlet part of the pump. The modification of the impeller geometry affects the flow patterns inside the centrifugal pump. When plate damage, trailing edge damage or leading edge damage occur flow pattern changes [13], [14], [19], [20]. These flow pattern changes affect mainly VPF and their harmonics [13]. They can also produce broadband noise due to the vortices generated in the clearances between the impeller and the volute [18]. The mechanism of production of flow patterns inside the pump is not very well understood yet [19] and this subject is outside the scope of this Thesis. The system faults considered in this Thesis consist in the addition of strange obtjects and impurities to the fluid such as PVC (polyvinyl chloride) balls, sand, sand and paper. This can lead to partial obstruction of the piping system, producing low flow rate and the generation of turbulence (broadband) noise [8], [15]. The addition of impurities can produce degradation of the pump itself in long term. 3.2.5 Audio and Vibration Database acquisition from a centrifugal pump In order to study audio and vibration signals simultaneously, we have recorded vibration and audio samples from a circulating centrifugal pump in an experimental test rig. This subsection describes the acquisition process of the database and the database generated. A circulating centrifugal pump is a pump designed to circulate a fluid through a closed system. A closed system is one which runs in a loop, with the pump discharge line eventually returning back to the pump suction. The pump works like any centrifugal pump, except they only need to overcome the friction of the piping system. Circulation pumps are primarily used for circulation of water in closed systems e.g. heating, cooling and air conditioning systems as well as domestic hot water systems and in applications that require chemicals to be regularly mixed into the fluid, such as pool and spa pumps. Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 97 be in the water. Figure 3-16 shows the impeller used for LED fault with 5mm of vane leading edge removed from all vanes. TABLE 3-8: SEVERITY OF LEADING EDGE FAULTS Fault Severity Length vane reduction Length of the remaining vane None 0 50 mm Slight (LED5) 10% of the total length = 5mm 50mm - 5mm = 45 mm Medium (LED10) 20% of the total length = 10mm 50mm - 10mm = 40 mm Severe (LED15) 30% of the total length = 15mm 50mm - 15mm = 35 mm Figure 3-16: Photo of the impeller with leading edge fault. In this case, the slight fault (5mm from all vane leading edges were removed) is shown. Trailing Edge Fault The trailing edge fault (TED) consists in removing part of the vane trailing edge (the region at the tip of the vane) of all vanes. Three different lenghts were used. Table 3-9 shows the original length of the vane and the milimiters of vane removed. This fault simulates the degradation of impellers along time due to particles or impurities that can be present in the water. Figure 3-17 shows the impeller used for TED fault with 5mm of vane trailing edge removed from all vanes. PhD Dissertation Universidad de Las Palmas de Gran Canaria 98 TABLE 3-9: SEVERITY OF TRAILING EDGE FAULTS Fault Severity Length vane reduction Length of the remaining vane None 0 50 mm Slight 10% of the total length = 5mm 50mm - 5mm = 45 mm Medium 20% of the total length = 10mm 50mm - 10mm = 40 mm Severe 30% of the total length = 15mm 50mm - 15mm = 35 mm Figure 3-17: Photo of the impeller with trailing edge fault. In this case, the slight fault (5mm from all vane trailing edges were removed) is shown. Seal rubber fault The seal fault was produced spontaneously while impeller 2 was being recorded in normal condition. The sealing type of the pump is RS-1. The assembly of the impeller with the rotor is configured with a ceramic "O-ring" type stationary seat and is also equipped with a "set screw collar". The RS-1 type is an "Elastomer (or Rubber) bellows seal". When we open the pump, we discovered that the ceramic "O-ring" was displaced during the pump operation. This caused the friction between the O-ring and the impeller and a screech sound was audible. In Figure 3-18 the impeller with the O-ring is shown. Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 99 Figure 3-18: Photo of the o-ring of the impeller of the AL P800 pump. System faults The ALP800 pump is designed and built for pumping water, clean and free from solids. We have added different solids to the water to cause damage and partial obstruction to the pump. This causes a decrease of water flow. The different solids added to the water are enumerated in Table 3-10. The characteristicas of the solids and their concentration in water are also shown in the table. TABLE 3-10: CHARACTERISTICS OF SOLIDS ADDED TO THE PUMP Conditions Characteristics Concentration Sand Fine sand from "Las Alcaravaneras" beach Grain diameter range: 0.0625-2 mm 2kg/50l Sand + Paper Standard paper A4 80gr 2kg/50l (sand), 2kg/50l(paper) PVC balls (no sand anymore) Polyvinyl chloride balls used for airsoft pellets Diameter: 6 mm Weight: 0.12 g 2000 balls / 50 l (240 g / 50 l) Figure 3-19 shows the reservoir with sand and Figure 3-20 shows the PVC balls. PhD Dissertation Universidad de Las Palmas de Gran Canaria 100 Figure 3-19: Photo of the reservoir with the sand added to simulate strange objects or particles in the system. Figure 3-20: Kind of PVC balls (6 mm of diameter) used to simulate strange objects in the system [29]. 3.2.5.4 Experimental set-up: sensor positions The room where the experimental set-up is placed is shown in Figure 3-21. The room has a 29 dB of acoustic isolation respect aerial noise. Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 101 Figure 3-21: Drawing of the room for the experimental set-up with the measurements in mm. A squematic of the set-up is also shown. A squematic of the experiment setup is also shown in Figure 3-21, where the pump (in green), the reservoir with water (in blue) and the pipes (in orange) are shown. The flow direction is also shown. A photograph of the experiment is shown in Figure 3-22, where the circulating pump and the sensors are shown. The elements of the experimental set-up are the following: a centrifugal pump, 4 meters of pipes of 3.81 cm (1.5 inches) of diameter connected to the inlet and outlet of the pump, a pressure gauge connected to the output of the pump, the reservoir and a trolley in which the reservoir and the pump are placed on. The reservoir is filled with 50 liters of water and a closed loop is formed with pipes. The pump is at the same height of the reservoir because is a circulating pump. Two microphones were placed in the pump inlet and in the pump outlet. OutletMicro points the pump outlet and InletMicro points the pump inlet (see Figure 3-22). The microphones are expected to obtain information about the flow changes. The distance between the microphone and the pump used in different works in centrifugal Flow direction PhD Dissertation Universidad de Las Palmas de Gran Canaria 102 pumps varies from 5cm to 100cm [16],[18],[31],[32]. These works focus on detection of cavitation. Iniatially, the distance of the microphones from the pump was varied from 2cm to 100cm (in steps of 2cm from 2cm to 20cm and in steps of 5 cm from 20cm to 100cm) and audio signals were taken. Their spectra were visually inspected for searching peaks at the discrete frequencies (rotor frequency, vane passing frequency and their harmonics). The discrete frequencies were observed in all cases but from 50cm to 100cm the amplitude of the discrete frequencies vanished progressively. Taking into account that in industrial scenarios, the background noise is usually high, microphones should be located near the source [33]. For this reason, microphones were placed at 5 cm from the pump in this Thesis. Accelerometers were placed in two orthogonal directions over the pump casing: accelerometer CESVA AC001 with sensitivity 100 ± 5 % mV/g (‘RadialInletAccel’ in Figure 3-22) and accelerometer CESVA AC006 with sensitivity 1000 ± 10 % mV/g (‘RadialAccel’ in Figure 3-22). 'RadialInletAccel' is placed on the volute in the radial plane (that is, the plane perpendicular to the rotor direction) and near the inlet, 'RadialAccel' is placed over the volute in the radial plane and perpendicular to “RadialInletAccel”. A third accelerometer was placed in axial direction (the rotor direction). However, this accelerometer is not considered in this Thesis. The accelerometers are expected to obtain information about the flows inside the volute. Figure 3-22: Sensor Placement for the Experimental set-up. OutletMicro InletMicro RadialAccel RadialInletAccel Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 103 3.2.5.5 Database Structure The files of the database are in .mat format. The content of each file in the database is described in Table 3-11. The names of the variables stored in each file of the database are enumerated in the first column of Table 3-11. In the second column the possible values of the variables are enumerated and described. The name of the files in database contains all the information to know the pump condition, the impeller used in the recording, the severity of the fault and the number of the file. The name of the files in the database has the following format: CON_ImI_SevS_XX.mat CON is the condition. See variable condition in Table 3-11. I is the number of the impeller. See variable impeller in Table 3-11. S is degree of severity in a fault. See variable severity in Table 3-11. XX is the number of the file. For example, LED_Im3_Sev1_01.mat is a file with LED condition in impeller 3 with severity 1. The number of the file is 01. PhD Dissertation Universidad de Las Palmas de Gran Canaria 104 TABLE 3-11: CONTENT OF A SAMPLE FROM THE DATABASE condition String with the name of the condition. Possible values are: .‘nor’: normal condition .‘pla’: plate fault condition. .‘sea’: plate fault condition. .‘led’: leading edge fault condition. .‘ted’: trailing edge fault condition. .‘san’: sand in water condition. .‘sap’: sand and paper in water condition. . ‘pvc’: balls of PVC in water impeller Contains the number of the impeller used in the recording. Possible values are: 1: impeller 1 2: impeller 2 3: impeller 3 4: impeller 4 multisigOri ginal Matrix that contains the audio and vibration signals of 59 seconds (with a sample frequency of 22050 Hz). Each row of the matrix corresponds to the signal acquired for a sensor. Sensorposit ion Type cell. Each cell contains the name of the sensors in the order that are stored in the matrix multisigOriginal. 'radialInletAccel' 'radialAccel' 'outletMicro' 'inletMicro' The data of 'ratialInletAccel' is stored in the first row of the multisigOriginal matrix. The data of 'radialAccel' is stored in the second row of the multisigOriginal matrix and so on. severity The severity of the fault. Possible values are: 0: no severity. 1: slight severity. 2: medium severity. 3: higher severity No severity means that the file corresponds to a normal condition or that the file corresponds to a fault with no severity associated (for example, seal fault). 1 can correspond to a plate fault with one part of plate removed, to a LED or TED fault of 5mm. 2 can correspond to a plate fault with two parts of the plate removed, to a LED or TED fault of 10 mm. 3 can correspond to a plate fault with three parts of the plate removed, to a LED or TED fault of 15 mm. dateacq String with the data of the data acquisition. Example: ‘21-Jul-2011’ File recorded 11 July 201. Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 105 3.2.5.6 Numbers of samples for each condition Table 3-12 shows the different conditions recorded for ALP800 pump. The first column of the table shows the impeller used, the second column shows the pump condition and the third column shows the number of samples acquired for each condition. Each sample has four signals of 59 seconds. Each sample was obtained in different recording sessions. TABLE 3-12: RECORDED CONDITIONS FOR EACH IMPELLER Impeller # Condition Number of samples per condition 1 Normal 20 1 1 Plate 20 1 2 Plates 20 1 3 Plates 18 2 Normal 20 2 Seal Ring 9 2 Sand 23 2 Sand and Paper 62 2 PVC balls 18 3 Normal 20 3 LED 5 mm 20 3 LED 10 mm 20 3 LED 15 mm 20 4 Normal 20 4 TED 5 mm 20 4 TED 10 mm 20 4 TED 15 mm 20 3.2.6 Pump Database Preprocessing and Baseline signals Each original file of the database is 60 seconds at a sample frequency of 44.1kHz (more than twice the bandwidth of the microphones). As the amplitude in the spectra of the audio signals drops around 15kHz and the band-width of the accelerometers is up to 10kHz, each file of the database was decimated by 2. The new sample frequency is 22050Hz. To decimate the files a linear phase FIR (finite response) filter with order 30 is used. Finally, we remove the beginning of the signal and we obtain a signal of 59 seconds. PhD Dissertation Universidad de Las Palmas de Gran Canaria 106 In the following figures, examples of spectra of vibration and audio signals from the centrifugal pump test rig in normal (free-fault) condition are shown. Figure 3-23: Spectrum in range [2.7-1000]Hz of a 8192 points vibration frame (371.5ms) from sensor RadialInletAccel. The higher peaks are marked with red circles and the frequency value is shown. Figure 3-24: Spectrum in range [1000-11025]Hz of a 8192 points vibration frame (371.5ms) from sensor RadialInletAccel. The higher peaks are marked with red circles and the frequency value is shown. 0200 400 600 800 1000 0 0.5 1 Radial Inlet Accel: Normal Condition ([0-1000]Hz) Frequency (Hz) Normalized Amplitude 97 Hz 339 Hz 976 Hz 876 Hz 196 Hz 683 Hz 46 Hz 2000 4000 6000 8000 10000 0 0.5 1 Radial Inlet Accel: Normal Condition ([1000-11025]Hz) Frequency (Hz) Normalized Amplitude 1077 Hz 2719 Hz Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 113 CHAPTER 4 4. Contributions to vibration bearing fault diagnosis and fault identification This Chapter focuses on vibration features in bearing fault diagnosis and bearing degradation. The contributions of this Thesis on bearing fault diagnosis are based on nonlinear measures. The nonlinear energy operator Teager-Kaiser is proposed as a preprocessing tool for bearing fault diagnosis followed by statistical and energy based feature extraction to diagnose between normal bearing condition, inner race fault, outer race fault and ball fault conditions. A new methodology to assess the severity of a fault in bearings (i.e.: the progression of the fault from an onset fault to a developed fault) for outer race fault and inner race fault using wavelet package transform and complexity measures Lempel-Ziv complexity is also proposed. In this Chapter, we explain our proposals and the experimentation carried out. Finally, we show the results and the conclusions. The first part of the Chapter (section 4.1) is devoted to the experiment of the nonlinear energy operator Teager-Kaiser. The structure is based on papers published by the author of this Thesis [1]-[3]. The second PhD Dissertation Universidad de Las Palmas de Gran Canaria 114 part of the Chapter is devoted to the degradation experiment. The structure is based on a published conference [4] and on an ongoing paper. 4.1 Proposal of nonlinear Teager-Kaiser energy operator for bearing fault diagnosis and bearing degradation assessment As stated in Chapter 3, vibration signals generated by bearings with discrete faults located at the inner race, at outer race or at the rolling elements can be viewed as an amplitude modulated signal in which the carrier is the resonance frequency of the bearing (the frequency excited by the impacts of the discrete fault) and the fundamental frequency of the modulating signal (the envelope) is the bearing characteristic frequency of the faulty bearing. In Figure 4-1, repeated here for convenience, the vibration signals for a bearing with inner race, outer race and ball fault are shown along with their corresponding envelope signals. Figure 4-1: Time Vibration signals for bearings with outer race fault, inner race fault and ball fault. Source: [28]. Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 115 In recent decades, much attention has focused on the development of new digital signal analysis methods for modeling bearing vibration signals so as to discriminate between normal operation and different faulty conditions. Some conventional methods, after sampling the vibration signal )(tv with sample frequency s f , feature the vibration signal          s f n vnv )( with statistical measures of first, second or higher order such as variance, skewness or kurtosis and with measures such as crest factor, impulse factor and root mean square which feature the impulsive nature of the bearing vibration signal [5]. Other methods model the vibration signals generated by faulty bearings as an amplitude modulated (AM) signal, defined as: )2cos()2cos()2cos()()( nfnfAnfnAnv bbfb   [Eq. 4-1] where b f is the resonance frequency of the bearing and )(nA the AM signal or signal envelope whose bf f is the bearing fault characteristic frequency that have to be detected. The AM signal can be extracted using the high frequency resonance analysis or envelope analysis [6] which implies band-pass-filtering the )(nv bearing vibration signal. Then, Fourier transform is used to obtain bf f . The main disadvantage of this technique is the difficulty in the correct selection of the central frequency and the bandwidth of the band-pass filter. Recent methods in bearing fault diagnosis include more advanced signal processing methods such as spectral kurtosis [7], wavelet analysis and empirical mode decomposition (EMD) [8]. Spectral kurtosis offers a way of designing optimal bandpass filter for bearing fault diagnosis. Wavelet analysis and EMD are time-frequency techniques that decompose the raw vibration signal in different frequency bands. Then, several features such as entropy or statistics are extracted from the different frequency bands. PhD Dissertation Universidad de Las Palmas de Gran Canaria 116 Recently, as an alternative to obtain the AM signal from the raw vibration signal, [6], [10] proposed the nonlinear Teager-Kaiser energy operator (TKEO) [11]. The AM signal and the frequency modulated (FM) signal from a mono-component AM-FM signal can be extracted using TKEO [12]. H. Li et al. [9] proposed TKEO to extract the AM signal (AM-TK signal) from bearing vibration signal, without using a band-pass filtering process, just with some simple operations over TKEO. Liang et al. [10] proposed the application of the TKEO over the raw bearing vibration signal without computing the AM signal. Then, the Fourier transform is applied and the bf f is identified. In this Thesis, we propose to use TKEO to obtain the vibration signal in the Teager–Kaiser domain (TK signal) and then to feature it with statistical and energybased features. The objective is to show how the statistical and energy-based features extracted from the TK signal outperform the diagnosis results when extracting the same features from the raw time vibration signal (time signal), the AM signal obtained by pass-band filtering (AM signal) and the AM signal obtained by using the TKEO (TKAM signal). A comparative analysis between statistical and energy-based features extracted from TK signal, TK-AM signal, AM signal and time signal (T signal) is accomplished. Next, the conventional envelope analysis and the Teager-Kaiser energy operator are explained. From the conventional envelope analysis the AM signal is extracted. The TK signal is extracted directly from the Teager-Kaiser operator and the TK-AM signal is extracted from the Teager-Kaiser operator after doing some calcularions. The T signal is the raw vibration signal. Conventional Envelope analysis: AM signal The conventional analysis to obtain the amplitude demodulated signal ( )(nA in [Eq. 4-1]) is called envelope analysis or high frequency resonance analysis [6]. This procedure implies the following steps: (i) band-pass filtering the bearing vibration signal, (ii) application of the Hilbert transform to the band-pass filtered signal, and (iii) low-pass filtering the resulting signal. The step (i) requires a visual inspection of the Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 117 bearing vibration spectrum to estimate the central frequency and the bandwidth of the band-pass filter. Teager-Kaiser energy operator: TK-AM signal and TK signal The Teager-Kaiser energy operator was derived by Kaiser in 1990 [11] to measure the energy of the mechanical process that generated a single time-varying signal. TKEO can detect modulations in AM-FM signals by estimating the product of their time-varying amplitude and frequency. It is considered as a high-resolution energy estimator. The TKEO for continuous time signals v(t) is: )()()]([)]([ 2tvtvtvtv c   [Eq. 4-2] where dt dv tv )(  It can be shown [11] that the discrete version of the TKEO is: )1()1()()]([ 2 nvnvnvnv  [Eq. 4-3] Maragos et al. [12] developed a method to estimate the amplitude envelope (AM signal) and the instantaneous frequency (FM signal) of speech formant signals using the TKEO. The amplitude modulated signal can be extracted from TKEO (TK-AM signal) as follows [12]: )]1()1([ )]([2 )(  nvnv nv nA   [Eq. 4-4] This technique has been applied in bearing fault diagnosis in [9] to extract the TK-AM signal without using a band-pass filter. The direct application of TKEO over the raw vibration signal )(nv ([Eq. 4-3]) [10] can perform more effective bearing fault detection. In the TK domain the faulty samples can be easily discriminated because the TK signal highlights the characteristic PhD Dissertation Universidad de Las Palmas de Gran Canaria 118 impulse train which is due to the impact of the rollers with the defect. This effect can be seen in Figure 4-2 which shows an example of TK signals for normal and fault conditions. Therefore, it is expected that features obtained from the TK signal will discriminate effectively between normal and fault bearing conditions without estimating the AM envelope. The direct computation of TKEO over the vibration signal has the following advantages: (i) it does not require the use of a band-pass filter. Therefore, the appropriate estimation of the central frequency and bandwidth of the band-pass filter is avoided.; (ii) Three adjacent samples of the signal are used to compute the TKEO. This fact makes the TKEO implementation very simple and computationally efficient. Figure 4-2: Teager-Kaiser transformed signals (TK signals) for different bearing conditions: normal condition (upper-left), inner race faul condition (upper-right), outer race fault condition (bottom-left) and ball condition (bottom-right). From the AM signal, the TK-AM signal and from the TK signal a set of statistical, energy-based and entropy features are extracted and a performance comparison between the different methods are carried out. The extracted features from the AM signal, the TK-AM signal and the TK signal are: features that characterize the amplitude [1]: peak value and peak-to-peak value; statistical measures: standard deviation, skewness and kurtosis; third, features that quantify the energy in the signal [1]: root mean square (rms), squared mean root (smr); fourth, features that quantify the impulsive nature of the vibration signal [1]: crest factor (peak value/rms), L factor (peak value/smr), shape factor (rms/peak value), impulsive factor (peak value/mean) and finally the Shannon entropy that measures the disorder or complexity of a system [14]. Finally, we have the featured AM signal, the featured AM-TK signal and the featured 0 0.05 0.1 0.15 0 0.2 0.4 0.6 0.8 Time(s) Amplitude TK signal for N condition 0 0.05 0.1 0.15 0 0.5 1TK signal for IR condition Time(s) Amplitude 0 0.05 0.1 0.15 0 0.2 0.4 0.6 Time (s) Amplitude TK signal for OR condition 0 0.05 0.1 0.15 0 0.2 0.4 0.6 0.8 Time (s) Amplitude TK signal for B condition Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 119 TK signal. The proposed method is first applied to the Case Western Database to show how the proposed method outperforms the method that uses statistical features in fault diagnosis, i.e. discriminating between different bearing faults: inner race fault (IR), outer race fault (OR) and ball fault (B). The diagram of the experiment for fault diagnosis is shown in Figure 4-3. Then, the proposed method is applied to the UH-60 database. As this database is a run-to-failure database (fault degradation) the method is applied with some differences which are explained in the corresponding section. See Figure 4-4 for the diagram of the experiment for fault evolution along time. In the next subsection the singularities of the application in each database are explained and the results are shown. Figure 4-3: Diagram of the experimentation in bearing fault diagnosis. The proposal is compared with other methods in the state of the art and an evaluation with two classifiers is carried out. Figure 4-4: Diagram of the experimentation in bearing fault evolution. The proposal is applied to a run-to-failure bearing vibration database. PhD Dissertation Universidad de Las Palmas de Gran Canaria 120 4.1.1 Evaluation and results of the proposal for bearing fault diagnosis Following the diagram in Figure 4-3, the proposed methods 1 to 4 were evaluated in the Case Vibration Bearing database for fault diagnosis. To quantitative evaluate the proposed method, two classifiers are used to evaluate the discrimination ability of the features between normal condition, IR fault, OR fault and B fault: a neural network classifier and a Least Square-Support Vector Machine (LS-SVM) classifier. Moreover, the features extracted from the TK signal are sorted by relevance order with the floating forward feature selection procedure. Then, features from the different signals are evaluated (in order of relevance) with the two mentioned classifiers. A detailed explanation of the experimentation is carried out in the following paragrahps and the results are shown. Frame segmentation and feature extraction First, the mean is removed from the vibration signal of each file of the database and then the vibration signal is normalized between -1 and 1. Then, the vibration signal is divided into time frames of 0.17 seconds which are overlapped by 33%. Each frame comprises 5 revolutions of the shaft. Each frame is transformed to the TK domain using [Eq. 4-3]. Then, the mentioned features are extracted from the TK signal. We call the features extracted from the TK signal TK features. In order to compare the results, the same features are also extracted from the time vibration signal )(nv (T features), from the AM signal by means of envelope analysis (AM features) and from the AM signal by means of the Teager-Kaiser method (TK-AM signal). Relevance of the TK features In order to select the subset of TK features with gives the best success rate, the TK features are sorted by relevance order using the sequential floating forward selection (SFFS) algorithm [22] and then are evaluated incrementally with the neural network classifier and the LS-SVM classifier. SFFS is an heuristic algorithm that search the subset of features of the original set which has the best success rate. The procedure to sort the features by relevance order using the SFFS is described next. Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 121 The database is split into a training subset and a testing subset by choosing randomly three files from normal condition and nine files from each fault condition (30 files in total) for the training set and leaving the remaining files for the testing set (10 files in total). The data in the training and testing setare z-score normalized using the mean and variance of the training set. The SFFS algorithm is applied 60 times with different training and testing subsets randomly chosen. As a result, a different subset of best features is obtained each time the SFFS algorithm is applied. The classifier used by the SFFS to obtain the success rate is a 1-nearest neighbor classifier. At the beginning of the experiment, a 60(times) x 12 (features) matrix is initialized at zero values. Each time the SFFS is applied a subset of best features is obtained. Their position in the matrix is marked with ‘1’. At the end of the 60 repetitions, each column is summed so as the number of times a feature is within the subset of best features is obtained. Finally, the features are sorted by relevance in descending order: the feature that most times appears in the subset of best features is the most relevant one and the feature that least times appears in the subset of best features is the least relevant one. Classification Once the TK features are sorted by relevance order, they are incrementally evaluated using two classifiers: a neural network classifier and a LS-SVM classifier. The incremental evaluation of the features consists in the following: first, the most relevant feature is evaluated, then the most relevant is evaluated along with the second most relevant and so on. As a result, the subset of TK features which gives the best success rate can be selected. The same procedure is done with the T, AM and TK-AM features. Neural Network Classifier The features are fed into a multilayer feed forward neural network (NN) with one hidden layer trained to discriminate between normal condition, IR fault condition, OR fault condition and B fault condition. The neural network is a standard classifier and can be used to reproduce the results easily. Supervised learning is carried out using the resilient back propagation train algorithm [23]. The number of neurons in the hidden layer is selected using cross-validation technique and the output layer has 4 neurons. PhD Dissertation Universidad de Las Palmas de Gran Canaria 122 The NN input is the vector of features belonging to one frame: p T R pp ],[ 1 (T is transposition of the vector p) and R is the number of features in the vector. As the features are evaluated incrementally, when one feature is evaluated R = 1, when two features are evaluated R = 2 and when all the features are evaluated R = 12. The activation functions for the hidden layer are tansigmoid functions (hyperbolic tangents). The output layer of the NN has 4 nodes corresponding to the 4 bearing conditions: a2 T aaaa ],,,[ 2 4 2 3 2 2 2 1  , where 2 1 a is the output for normal condition, 2 2 a is the output for IR fault condition, 2 3 a is the output for OR fault condition and 2 4 a is the output for B fault condition. The activation functions for the neurons of the output layer are linear activation functions. In order to train and test the NN, the database is split into a training subset and a testing subset and the data are normalized in the same way of the feature selection procedure. The training subset is used to choose the size of the hidden layer (i.e. the number of neurons of the hidden layer S). The NN is trained with varying S from 1 to 20. For each S value, the training subset is subdivided using the 3-fold cross-validation technique, i.e. the training subset is divided in three different subsets: two of them are used to train the neural network and the remaining one (called validation subset) is used to compute the success rate. The value of S with the best success rate is chosen as the size of the hidden layer. Finally, the NN with the final configuration is trained and then evaluated with the testing subset. In the testing phase, each frame is classified according to the maximum output value of the neural network. For example, if },,,max{ 2 4 2 3 2 2 2 1 2 1aaaaa  , the frame is classified as normal condition. Then each file is classified between normal, IR, OR or B fault conditions according to the more voted rule, i.e. if the majority of the frames are classified as normal condition, the file is classified as normal condition. All the process is repeated 60 times, randomly choosing different files for training and testing set and the results were averaged. As a result, the success rate is obtained. The training process is stopped when any of these conditions occurs: the maximum number of iterations, set to 5000, is reached; the maximum amount of time, set to infinite, is exceeded; the performance of E (error function) is minimized to the Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 129 Figure 4-8: Evolution of the features extracted from the raw vibration signal (T signal). Figure 4-9: Evolution of the features extracted from the envelope signal of the Q band[2500-3800Hz]. 4.1.3 Conclusions In this Thesis we propose the use of statistics, energy-based and impulsive measures extracted from the Teager-Kaiser signal in bearing fault diagnosis. The vibration signal is transformed into the Teager-Kaiser domain using the Teager-Kaiser operator. A standard and easily reproducible neural network classifier is used to evaluate the discrimination ability of the extracted features between normal bearing, outer race fault, inner race fault and ball fault. Moreover, a LS-SVM classifier is also used to validate 10 20 30 40 50 60 0.3 0.4 10 20 30 40 50 60 -0.08 -0.04 0.02 10 20 30 40 50 60 2.65 2.75 2.85 10 20 30 40 50 60 3.8 4.2 4.6 Data Set Index Kurtosis evolution (raw signal) Skewness evolution (raw signal) RMS evolution (raw signal) Crest factor evolution (raw signal) 10 20 30 40 50 60 5 10x 10-3 10 20 30 40 50 60 2 3 10 20 30 40 50 60 10 20 10 20 30 40 50 60 6 10 Data Set Index RMS evolution (Q band envelope ) Skewness evolution (Q band envelope ) Kurtosis evolution (Q band envelope) Crest factor evolution (Q band envelope) PhD Dissertation Universidad de Las Palmas de Gran Canaria 130 the results obtained by the neural network. The results show that the use of statistical, based-energy and impulsive features extracted from the Teager-Kaiser signal in bearing fault diagnosis outperforms the results obtained by the same features extracted from the time vibration signal, from the time envelope signal (computed using TKEO) and from the time envelope computed using band-pass filter. In addition, the use of the TeagerKaiser signal avoids visual inspection of the bearing spectrum to determine the center frequency and the bandwidth of the band-pass filter. The computation of the extracted features and of the TKEO itself is also simpler and faster than band-pass filtering. The proposed method is also applied to a run-to-failure bearing vibration database obtained from an endurance test of a Black Hawk Helicopter. The results show that root mean square and kurtosis features extracted from the signal transformed to the TeagerKaiser domain are good indicator of the bearing degradation. 4.2 Proposal of Lempel-Ziv complexity measure based on wavelet package transform for bearing degradation assessment The aim of bearing fault diagnosis techniques is to detect the kind of fault presented in a bearing (i.e. outer-race fault, inner-race fault or ball fault in the case of discrete defects). Once the kind of bearing fault is diagnosed (i.e. inner race fault, outer race fault or ball fault), it is important to assess the severity of the fault (fault identification), i.e. how large the fault is. Another important task in condition monitoring is the early detection of a fault (to follow the degradation of a component from normal condition to fault condition). In this Thesis, we have developed a new methodology to assess the severity of bearings with outer race fault and inner race fault with different degrees of severity (depicted in Figure 4-10). The proposed method combines the wavelet package transform technique and the Lempel-Ziv complexity to assess the fault severity of bearings with outer race and inner race fault. We evaluate the proposed method in the Case Western Vibration Bearing database, the UH-60 helicopter database and in the IMS bearing database. Wavelet packet transform is a time-frequency technique that basically decomposes the signal into detail (low frequency components) and Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 131 approximations signals (high frequency components) in different levels of decomposition or scales. Wavelet packet transform is used as a preprocessing tool to select the node with higher energy (i.e. the node where the fault lies). Then, the LempelZiv complexity is used to measure the severity of the fault. As the Lempel-Ziv complexity values are bounded between 0 and 1, it is a good indicator of the severity of the bearing fault. In this section, a description of the methods for fault severity assessment in bearings literature is carried out. Next, the preprocessing tool (the wavelet packet transform) and the Lempel-Ziv complexity are explained. Then, the proposed methodology is shown and finally a set of experiments in bearings is carried out to show the usefulness of our proposal and a comparison with some methods in literature is also accomplished. State of the art The severity assessment in bearings in literature is mainly addressed with time-domain features including kurtosis, root mean square, crest factor and amplitudes [20], [21]. Kurtosis increases in single-point faulty bearings from a value of 3 (in the case of freefault bearing) to higher values when the bearing has a single-point defect. However, when the single-point defect is in advanced stages kurtosis values return to 3 [20]. Root mean square, crest factor, amplitudes are other techniques to follow the degradation of the fault. However, they have the same problems that kurtosis. Their values are not consistent with the evolution of the fault. There are other techniques in frequency domain such as energy features extracted from envelope analysis. The main disadvantage of these features are that thery are not bounded. Lempel-Ziv complexity was firstly investigated in bearing severity assessment Lempel-Ziv complexity in [15]. The authors proposed LZ for outer-race bearing fault deterioration. They extracted the LZ values from the raw vibration signal. They showed that the LZ value increases as the size of the outer race bearing fault increases. H. Hong et al. [16] proposed a new version of the Lempel-Ziv complexity as a bearing fault severity measure based on the continuous wavelet transform (CWT). The applications to the bearing innerand outer-race fault signals have demonstrated that the new version PhD Dissertation Universidad de Las Palmas de Gran Canaria 132 of Lempel–Ziv complexity can effectively measure the severity of both innerand outer-race faults. This technique uses the CWT to obtain the sub-band where the fault lies. Then, the signal is amplitude demodulated and the envelope signal and the carrier signal are obtained. Finally, the Lempel-Ziv complexity is computed for the envelope signal and for the carrier signal and a weighted summed is computed. They use different weights depending on the fault (IR fault or OR fault). In this Thesis, we propose a simpler way to assess the severity in bearing using wavelet package transform to obtain the sub-band where the fault lies and then applying Lempel-Ziv complexity to this subband. Methods Wavelet Packet Transform Wavelet packet transform (WPT) is an extension of the discrete wavelet that allows a finer resolution of frequencies at both high frequencies and low frequencies. WPT can simultaneously break the signal up into detail (low frequency components) and approximations signals (high frequency components) in different levels of decomposition or scales. In bearing fault diagnosis the fault information lies in high frequency. For this reason, the wavelet packet transform is a common analysis tool for bearing fault diagnosis. Mathematically, wavelet packets are a collection of functions },,),2(2{ 2/ ZkjNnktW j n j  generated recursively from the following expressions:  mnn mtWmhtW )2()(2)( 2 [Eq. 4-5]   mnn mtWmgtW )2()(2)( 12 [Eq. 4-6] Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 133 where h and g are the quadrature mirror filters, )( 0tW is the scaling function and )( 1tW is the wavelet mother function, k is a time-localization parameter, j is a scale parameter (depth) and n is an oscillation parameter. The wavelet packet coefficients for a discrete signal can be obtained using the following iterative expressions: 120,)()2()2()()( ,,2,1   j lnjnjnj nlcklhkhkckc [Eq. 4-7] 120,)()2()2()()( ,,12,1    j lnjnjnj nlcklgkgkckc [Eq. 4-8] The original signal can be reconstructed using the following iterative expression: 120,)2()2()( 12,1,2,1,    j k knj k knjnj ncklgcklhlc [Eq. 4-9] Wavelet packets are organized in trees, where the scale j defines the depth or level of the tree and n is the position or node in the tree. For each scale j, 12,,0  j n . From each node of the tree, two children nodes are obtained. One of the children node is obtained by low-pass filtering (using the impulse response h) followed by downsampling and the other children node is obtained by high-pass filtering (using the impulse response g) followed by down-sampling. It is important to mention that the frequency order of the nodes for a level is not necessary the same as the node order because of aliasing introduced by down-sampling. For example, a signal with 1kHz bandwidth decomposed using a wavelet tree of decomposition level (scale) j = 3 have eight nodes at level 3 ( 7,,0 n ). The frequency content for each node (in natural order, i.e. ) is the following: n = 0, (0-125 Hz); n = 1, (125-250 Hz); n = 2, (375-500 Hz); n = 3, (250-375 Hz); n = 4, (875-1000 Hz); n = 5, (750-875 Hz); n = 6, (500-625 Hz); n = 7, (625-750 Hz). Therefore, the frequency order is not the same as node natural order (also called Paley order). To obtain the node in frequency order (from low frequency to high frequency), the nodes need to be reordered. 7,,0 n PhD Dissertation Universidad de Las Palmas de Gran Canaria 134 Lempel-Ziv complexity Lempel and Ziv proposed a complexity measure that can characterize the degree of order or disorder and development of spatiotemporal patterns in a time series [18], [19]. The signal is transformed into binary sequences and Lempel-Ziv algorithm gives the number of distinct patterns contained in the given finite sequence. After normalization, the relative Lempel-Ziv complexity measure (LZC) reflects the rate of new pattern occurrences in the sequence. LZC values range from near 0 (deterministic sequence) to 1 (random sequence). To compute the LZC of a given sequence s, the sequence has to be coded using a finite symbol set A. Using the standard coding squeme, the sequence to be coded s is divided in α equiprobables bins, where α is the number of diferent symbols in A. An element of the symbol set A is assigned to each value of the signal according to the bin in which the value lies. Finally, a new sequence of symbols S is obtained. This sequence S is subsequently scanned looking for original subsequences of different lengths. The complexity value is related with the number of different subsequences found. Thus, if A* denotes the set of all finite length sequences over the finite symbol set A, and )(Sl denotes the length of a sequence * AS , then according to [19]. 0},)({ * nnSlASAn [Eq. 4-10] and for every n AS  , the Lempel-Ziv complexity can be expressed as )(log)1( )( n n nc an    [Eq. 4-11] where 0 n  if n and  is the number of different symbols in the symbol set A. The upper bound for )(nc is )(log )()(lim n n nbnc a n  [Eq. 4-12] Therefore, )(nc can be normalized by using this upper limit Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 135 1 )( )( )(0  nb nc nC [Eq. 4-13] For 2  two symbols are used to code the signal (0 and 1). For 3  there exit 3 symbols (0, 1 and 2). The LZC value depends both on the number of symbols used in the coding squeme and on the length of the sequence s. According to [16] length of sequence higher than 3600 is enough. Proposal We propose a new method for fault severity assessment in bearings (depicted in Figure 4-10). The proposed method combines the wavelet packet transform and the LempelZiv complexity to assess the fault severity of bearings with outer race and inner race. As explained before in this Chapter, the LZC over the raw vibration signal has been used in bearing literature to outer-race ault severity assessment [15]. However, this technique is affected by noise. We show that our proposal is less affected by noise. Wavelet packet decomposition (3rd level, 8 nodes) Compute Relative Energy for each Node Select Node with maximum energy Input vibration signal Output: extracted feature for the input signal Compute Lempel-Ziv complexity Reconstruct the signal for the node kmax PhD Dissertation Universidad de Las Palmas de Gran Canaria 136 Figure 4-10: Proposed methodology for bearing severity fault assessment. First, the raw vibration signal is decomposed using the wavelet packet transform with a level 3 of decomposition (8 nodes). Daubechies 6 (db6) (see Figure 4-11) was used as the mother wavelet because has an impulsive shape according to the impulsive fault bearing characteristics (see figure 4-12). The WPT is used to obtain the node where the fault lies and remove unwanted signals. Initially, mother wavelets db2 to db5 and db7 to db9 and symlet wavelet mother familiy from order 2 to 9 were used. The results obtained are very similar between wavelet mothers with orders higher than 3. Figure 4-11: Waveform of the Daubechies 6 wavelet mother. The vibration signals are broken up to 3 level ( 3j ) of decomposition, obtaining 8 frequency bands (or 8 nodes) in this level of decomposition. Signals are reconstructed from the wavelet coefficients associated to each node. The reconstructed signal for each node is a new time series. If x denotes the original signal (the vibration signal), then nj x, is the reconstructed signals for jth level of decomposition and nth denotes the frequency-band of the signal. For 3j , there are 8 frequency-band signals in level 3. Then, the relative energy of the reconstructed signals of each node is computed, i.e. the energy of each node divided by the energy of the original raw signal. Then, the node with the maximum energy is selected. Relative energy features are computed from the reconstructed signals in level 3 of decomposition ( 3j ). E denotes the energy of the original raw vibration signal )(nx : 0 2 4 6 8 10 12 -1.5 -1 -0.5 0 0.5 1 1.5 Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 137 1 ))(( 1 2     N xix E N i [Eq. 4-14] where x is the mean value of )(nx and N is the number of samples of )(nx . Then, the relative energy of the reconstructed signals in level 3 is computed as: E M xix e N i n n n1 ))(( 1 2 ,3 ,3      [Eq. 4-15] where n x,3 (n = 0,...,7) are each of the reconstructed signals of level 3, M is the number of samples of n x,3 and n x,3 is the mean value of each n x,3 . The node with the maximum relative energy corresponds to the frequency band that contains the most fault features. A bearing with a localized fault will have the maximum relative energy located in high frequencies because the bearing is modelled as an amplitude modulated signal with the frequency of the carrier being the resonance frequency of the system. Therefore, features extracted from the selected node will have information about the fault.The node with the maximum relative energy is selected and the LZC is extracted from the reconstructed signal for this node. The LZC gives information about the complexity of the reconstructed signal. This is the diagnostic feature vector extracted from the input frame. This procedure is applied for each frame and finally a feature vector is obtained. The proposed methodology is applied to the Case Western Bearing Vibration Database, to the UH-60 helicopter database and to the IMS database. In order to compare with classical method in the literature, we extract kurtosis, Lempel-Ziv complexity from raw vibration data and kurtosis from wavelet package transform. PhD Dissertation Universidad de Las Palmas de Gran Canaria 138 4.2.1 Evaluation with the Case Bearing Database The proposed methodology is applied to the Case Bearing Database. Each sample of the database is divided in frames of 4096 samples each and for each frame the procedure shown in figure 4-11 is applied. As the sample frequency is 12000 s f Hz, the bandwitdth of the signal is 6000 N f Hz. Therefore, x has a frequency interval (0,6000]Hz, 0,3 x is the reconstructed signal in level 3 with a frequency range of (0, 750]Hz and 7,3 x is the reconstructed signal in level 3 with a frequency range of (5250, 6000]Hz. The nodes are reordered in frequency order. In Figure 4-12-Figure 4-16 the relative energies for each node of the WPT (in level 3) for a frame is shown for the normal condition and for fault conditions IR, OR and B with different severities and different loads. In the figures a title is added to each subfigure: IRXXYY means inner race fault with XX mils of diameter and load YY. XX can be '07' for a severity of 7 mils in diameter, '14' for a severity of 14 mils in diameter or '21' for a severity of 21 mills in diameter. YY can be '00' for load 0, '01' for load 1, '02' for load 2 or '03' for load 3. The nodes of the wavelet packet were obtained in natural order (Paley order) and then were reordered in frequency order. We can observe from the figures that the energy is concentrated in high frequency bands, where the fault characteristic lies. In the case of normal condition, the energy is concentrated in low frequency bands. Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 145 Figure 4-20: Evolution of the mean values obtained for each method: proposed method ‘LZC (EmaxWPT)’ (uppe-left), ‘LZC (raw signal)’ (upper-right), ‘Kurtosis (EmaxWPT)’ (bottom-left) and ‘Kurtosis (raw signal)’ (bottom-right) for inner race fault with different severities and different loads. The normal condition is also considered. The robustness of the algorithm to Gaussian noise is also studied. Gaussian noise was added to the original vibration signal with different signal to noise ratios (SNR): 20dB, 15dB, 10dB, 5dB and 0dB. Then, the proposed method is applied (with α = 4) for each SNR and for the original signal without noise added. Kurtosis of the raw signal, kurtosis of the maximum energy node and the LZC of the raw signal are also computed for each SNR and for the original signal. The motivation of this experiment is to show how the proposed method is more robuts to noise than computing the LZC from the raw vibration signal. The LZC measures the complexity in a signal. For random signals, the LZC is 1. Therefore, if the signals are contaminated with undesired Gaussian noise, the LZC will increase its value. If LZC is extracted from the raw vibration signal, it will be difficult to determine if the increasing or decreasing of the LZC is due to a fault or to noise. However, if the LZC is extracted from the node of the maximal energy associated to impulsive fault, the Gaussian noise will not affect in the same degree the signal. The following figures (Figure 4-21-Figure 4-28) show the results for each method and confirm our thoughts. N IR07 IR14 IR21 0.3 0.4 0.5 LZC (EmaxWPT) load 0 load 1 load 2 load 3 N IR07 IR14 IR21 0,5 0.7 0,9 LZC (raw signal) N IR07 IR14 IR21 0 20 35 Kurtosis (Emax WPT) N IR07 IR14 IR21 0 10 25 Kurtosis (raw signal) PhD Dissertation Universidad de Las Palmas de Gran Canaria 146 Figure 4-21: Evolution of the mean values obtained for ‘Kurtosis (raw signal)’ method applied for outer race fault condition and normal condition varying the amount of noise added to the original vibration signal. The results are shown for load: load 0 (upper-left), load 1(upper-right), load 2 (bottom-left), load 3 (bottom-right). Figure 4-22: Evolution of the mean values obtained for ‘LZC (raw signal)’ method applied for outer race fault condition and normal condition varying the amount of noise added to the original vibration signal. The results are shown for load: load 0 (upper-left), load 1(upper-right), load 2 (bottom-left) and load 3 (bottom-right). N OR07 OR14 OR21 0 10 20 30 Original 20dB 15dB 10dB 5dB 0dB N OR07 OR14 OR21 0 10 20 30 N OR07 OR14 OR21 0 10 20 30 N OR07 OR14 OR21 0 10 20 30 Kurtosis (raw signal) load 0 load 1 load 2 load 3 SNR N OR07 OR14 OR21 0.4 0.6 0.8 1 Original 20dB 15dB 10dB 5dB 0dB N OR07 OR14 OR21 0.7 0.8 0.9 1 N OR07 OR14 OR21 0.7 0.8 0.9 1 N OR07 OR14 OR21 0.7 0.8 0.9 1 LZC (raw signal) load 0 load 1 load 2 load 3 SNR Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 147 Figure 4-23: Evolution of the mean values obtained for ‘Kurtosis (EmaxWPT)’ method applied for outer race fault condition and normal condition varying the amount of noise added to the original vibration signal. The results are shown for load: load 0 (upper-left), load 1(upper-right), load 2 (bottom-left) and load 3 (bottom-right). Figure 4-24: Evolution of the mean values obtained for the proposed method ‘LZC (EmaxWPT)’ applied for outer race fault condition and normal condition varying the amount of noise added to the original vibration signal. The results are shown for load: load 0 (upper-left), load 1(upper-right), load 2 (bottom-left) and load 3 (bottom-right). N OR07 OR14 OR21 0 20 40 Original 20dB 15dB 10dB 5dB 0dB N OR07 OR14 OR21 0 20 40 N OR07 OR14 OR21 0 20 40 N OR07 OR14 OR21 0 20 40 Kurtosis (Emax WPT) SNR load 0 load 1 load 2 load 3 N OR07 OR14 OR21 0,4 0,5 0,6 Original 20dB 15dB 10dB 5dB 0dB N OR07 OR14 OR21 0,2 0,4 0,6 N OR07 OR14 OR21 0,2 0,4 0,6 N OR07 OR14 OR21 0.2 0.4 LZC (Emax WPT) SNR load 0 load 1 load 2 load 3 PhD Dissertation Universidad de Las Palmas de Gran Canaria 148 Figure 4-25: Evolution of the mean values obtained for ‘Kurtosis (raw signal)’ method applied for inner race fault condition and normal condition varying the amount of noise added to the original vibration signal. The results are shown for load: load 0 (upper-left), load 1(upper-right), load 2 (bottom-left) and load 3 (bottom-right). Figure 4-26: Evolution of the mean values obtained for ‘LZC (raw signal)’ method applied for inner race fault condition and normal condition varying the amount of noise added to the original vibration signal. The results are shown for load: load 0 (upper-left), load 1(upper-right), load 2 (bottom-left) and load 3 (bottom-right). N IR07 IR14 IR21 0 10 20 N IR07 IR14 IR21 0 10 20 30 N IR07 IR14 IR21 0 10 20 30 N IR07 IR14 IR21 0 10 20 Kurtosis (raw signal) Original 20dB 15dB 10dB 5dB 0dB load 0 SNR load 2 load 1 load 3 N IR07 IR14 IR21 0.4 0.6 0.8 1 N IR07 IR14 IR21 0.7 0.8 0.9 1 N IR07 IR14 IR21 0.7 0.8 0.9 1 N IR07 IR14 IR21 0.7 0.8 0.9 1 LZC (raw signal) Original 20dB 15dB 10dB 5dB 0dB SNR load 0 load 1 load 2 load 3 Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 149 Figure 4-27: Evolution of the mean values obtained for ‘Kurtosis (EmaxWPT)’ method applied for inner race fault condition and normal condition varying the amount of noise added to the original vibration signal. The results are shown for load: load 0 (upper-left), load 1(upper-right), load 2 (bottom-left) and load 3 (bottom-right). Figure 4-28: Evolution of the mean values obtained for the proposed method ‘LZC (EmaxWPT)’ applied for inner race fault condition and normal condition varying the amount of noise added to the original vibration signal. The results are shown for load: load 0 (upper-left), load 1(upper-right), load 2 (bottom-left) and load 3 (bottom-right). N IR07 IR14 IR21 0 20 40 N IR07 IR14 IR21 0 20 40 N IR07 IR14 IR21 0 20 40 N IR07 IR14 IR21 0 20 40 Kurtosis (Emax WPT) Original 20dB 15dB 10dB 5dB 0dB SNR load 0 load 1 load 2 load 3 N IR07 IR14 IR21 0.45 0.5 0.55 N IR07 IR14 IR21 0.3 0.4 0.5 0.6 N IR07 IR14 IR21 0.2 0.4 N IR07 IR14 IR21 0.2 0.4 LZC (Emax WPT) Original 20dB 15dB 10dB 5dB 0dB load 3 load 2 load 1 load 0 SNR PhD Dissertation Universidad de Las Palmas de Gran Canaria 150 4.2.2 Evaluation with the UH-60 Helicopter Database for bearing degradation assessment The proposed method of the Lempel-Ziv complexity based on wavelet packet transform is also evaluated with UH-60 Black Hawk Helicopter database where a rolling element of the bearing was damaged during an endurance test. Although our proposal does not follow monotonically a fault in a rolling element, according to the results with the Case Western database it should detect a ball fault. For this reason, we apply the proposal to the UH-60 helicopter database. As each dataset in the UH-60 database is 10 seconds (with a sample frequency of 100kHz) and the computation of the LZC is very slow with such amount of data, each dataset is divided in frames of 1 second each. The results are averaged per dataset. The results are shown in Figure 4-29 for ‘Kurtosis (raw signal)’, ‘LZC (raw signal)’, ‘Kurtosis (EmaxWPT)’ and the proposed method ‘LZC (EmaxWPT)’. As it is observed, none of the features can detect the fault. The ‘LZC (raw signal)’ and our proposal ‘LZC (EmaxWPT)’ shows an increase in their values in dataset 48, long after the fault. Figure 4-29: Evolution of the features using kurtosis and Lempel-Ziv complexity from the raw signal and kurtosis and Lempel-Ziv complexity from the node with maximal energy of the wavelet paket transform. 010 20 30 40 50 60 2.6 2.8 3Kurtosis (raw signal) 010 20 30 40 50 60 0.35 0.4 0.45 LZC (raw signal) 010 20 30 40 50 60 2.6 2.8 3Kurtosis (EmaxWPT) 010 20 30 40 50 60 0.25 0.3 0.35 LZC (EmaxWPT) Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 151 4.2.3 Evaluation with the IMS database The IMS database is a run-to-failure experiment of 164 hours in where an outer race fault is developed in a bearing. Previous paper indicates that the start of the fault is 89 hours after the beginning of the experiment [29]. In this case, the proposed method is applied to each sample of the database (each sample has 20000 data points corresponding to 1 second of signal). As the sample frequency is 20000 s f Hz, the bandwidth of the signal is 10000 N f Hz. Therefore, x has a frequency interval (0,10000]Hz, 0,3 x is the reconstructed signal in level 3 with a frequency range of (0, 1250]Hz and 7,3 x is the reconstructed signal in level 3 with a frequency range of (8750, 10000]Hz. Kurtosis from the raw signal, kurtosis from the node of maximal energy of the WPT and the LZC from the raw signal are also extracted and the results compared. Figure 4-30, Figure 4-31, Figure 4-32 and Figure 4-33 show the evolution of the features extracted by the ‘Kurtosis (raw signal)’ method, by the ‘Kurtosis (EmaxWPT)’ method, by the ‘LZC (raw signal)’ method and by the proposed method ‘LZC (EmaxWPT)’ respectively. The figures reveal that the kurtosis extracted from the raw vibration signal increase and then decrease again. Moreover, the first indication of fault is at 117 hours from the beginning of the run-to-failure experiment. The kurtosis extracted with the method KurtosisEmaxWPT shows a more consistent tendence that the kurtosis extracted from the raw vibration signal and the first indication of fault is at 89 hours from the beginning of the experiment. The LZCraw shows values near 1 in the normal condition. The reason can be the noise and signals of the rest of the gear that interfere with the bearing signal. 89 hours after the beginning of the experiment, the LZC values start decreasing. Then, the LZC values change with not clear tendency. When the LZCEmaxWPT method (proposed method) is applied, the LZC increase their values 89 hours after the beginning of the experiment (when the fault starts developing) and remains around the same values of complexity until the end of the experiment. PhD Dissertation Universidad de Las Palmas de Gran Canaria 152 Figure 4-30: Evolution of the Kurtosis extracted from the raw vibration signal of a run-to-failure experiment of the bearing that eventually developed an outer-race fault. Figure 4-31: Evolution of the Kurtosis extracted from the node of maximal energy of the WPT of a run-to-failure experiment of the bearing that eventually developed an outer-race fault. 0 20 40 60 80 100 120 140 160 180 0 5 10 15 20 Time (hours) Kurtosis (raw signal) 0 20 40 60 80 100 120 140 160 180 0 2 4 6 8 10 12 Time (hours) Kurtosis (Emax WPT) Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 153 Figure 4-32: Evolution of the Lempel-Ziv complexity extracted from the raw vibration signal of a run-to-failure experiment of the bearing that eventually developed an outer-race fault. Figure 4-33: Evolution of the Lempel-Ziv complexity extracted from the node of maximal energy of the WPT of a run-to-failure experiment of the bearing that eventually developed an outer-race fault. 4.2.4 Conclusions We have proposed a method to assess the fault severity, also called fault identification (how large the fault is), using the wavelet packet transform and the Lempel-Ziv complexity for inner race fault and for outer race fault. In the case of inner-race fault, the values of Lempel-Ziv complexity decrease when the fault evolves and in the case of the outer-race fault the values of Lempel-Ziv complexity increase when the fault evolves. The results obtained with the proposed method are in concordance with the 0 20 40 60 80 100 120 140 160 180 0.7 0.75 0.8 0.85 0.9 0.95 1 Time (hours) LZC (raw signal) 0 20 40 60 80 100 120 140 160 180 0.2 0.3 0.4 0.5 Time (hours) LZC (Emax WPT) PhD Dissertation Universidad de Las Palmas de Gran Canaria 154 results in literature [15], [16]. In [15] the Lempel-Ziv complexity was applied to raw bearing vibration signals for normal and outer-race fault conditions. The Lempel-Ziv complexity increases from normal to outer-race fault with high level of severity. However, as we have shown in our experiments, when the raw bearing vibration signal is contaminated with Gaussian noise, the Lempel-Ziv complexity is affected, leading to incorrect interpretation of the Lempel-Ziv complexity values. In [16], the authors propose the use of the Lempel-Ziv complexity and the continuous wavelet transform in outer-race fault and inner-race fault severity assessment. A method to select the best scale in the continuous wavelet transform based on energy and kurtosis was proposed. Then, the coefficients in the best scale are recovered and amplitude demodulated. The final value of the Lempel-Ziv complexity is a weighted summed of the Lempel-Ziv complexity extracted over the coefficients at the best scale and the Lempel-Ziv complexity extracted over the amplitude demodulated coefficients at the best scale. Using this method, they also found that the Lempel-Ziv complexity increase with the outer-race fault severity and decrease with the inner-race fault severity. In our proposal, it is not necessary to compute the amplitude demodulated signal of the signal in the node with maximal energy and therefore we only compute Lempel-Ziv complexity one time. One of the main advantages of using Lempel-Ziv complexity is that it is bounded between 0 and 1. With the use of the wavelet packet transform and the selection of the node with maximal energy (the node in which the fault lies), the effects of gaussian noise contamination is reduced. In the case of outer-race fault, this method can follow the fault degradation from normal condition to fault condition monotonically. It is worth to mention that the results obtained with the ball fault are not monotonically increasing or decreasing with the fault severity. However, the ball faut can be detected with the proposed method. The method was applied to two run-to-failure experiments: UH-60 helicopter database and IMS database. In the case of the UH-60 helicopter the proposed method does not detect the fault in early stages. A possible reason is that the wavelet packet is not detecting the correct node where the fault lies. More research is needed in this case. Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 257 of the correlation dimension of each emotion are below this quantity (see Figure 8-3) where the data distributions of the correlation dimension estimated for each database are shown). The delay (τ) and the minimum embedding dimension (m) are estimated for each frame using the first minimum of the mutual information function technique and the false neighbours technique respectively. Then, six complexity measures are extracted for each frame: value of the first minimum of mutual information function (MI), TakenTheiler estimator of the correlation dimension (CD), Shannon entropy (SE), correlation entropy (CE), Lempel-Ziv complexity (LZC) and Hurst exponent (H). Finally, for each measure, four statistical are computed: mean (μ), standard deviation (σ), skewness (sk) and kurtosis (k). Therefore, 24 features are extracted for each frame (mean of MI: μMI, mean of CD: μCD and so on). Feature selection and evaluation of the selected features A set of features are selected using a feature selection technique proposed by the author of this Thesis. In this Chapter, the proposed feature selection procedure will not be explained (see [39]). Once the selected features are identified, their discrimination ability between the different emotions is evaluated using a neural network. The database is split into a training subset and a test subset with 70% and 30% of each kind of emotional speech recordings, respectively. The experiments are repeated 25 times, each time using different training and test sets randomly chosen and the global success rate is computed as an average of the success rates in each iteration. 8.2.3 Results The results obtained in the experiments are shown and discussed. First, we show a data distribution analysis of the complexity measures (MI, SE, CD, CE, LZC and H). Then, we show the results of the feature selection procedure. The feature selection procedure is performed in the Polish emotional speech database. Finally, we show the success rates for the three emotional databases: the Polish emotional speech database, the Berlin emotional speech database and the LDC emotional speech database using the GOFS previously identified. PhD Dissertation Universidad de Las Palmas de Gran Canaria 258 Analysis of complexity measures The complexity measures extracted from the three databases are analyzed. Figure 8-3, Figure 8-4 and Figure 8-5 show the distribution of the six complexity measures (MI, SE, CD, CE, LZC and H) for the Polish emotional speech database, for the Berlin emotional speech database and for the LDC emotional speech database respectively using boxplots. The boxes have lines at the lower quartile, median (center line) und upper quartile values. The whiskers are lines extending from each end of the boxes to show the extent of the rest of the data. Boxes whose notches do not overlap indicate that the medians of the two groups differ at the 5% significance level. In Figure 8-3, Figure 8-4 and Figure 8-5 the upper left illustration corresponds to the MI, the upper right illustration corresponds to the SE, the middle left illustration corresponds to the CD, the middle right illustration corresponds to the CE, the bottom left illustration corresponds to the LZC and finally the bottom right illustration corresponds to the H. Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 259 Figure 8-3: Data distribution of neutral, fear and anger emotional speech for each complexity measure extracted from the Polish emotional database. MI: value of the first minimum of the mutual information function (upper left), SE: Shannon entropy (upper right), CD: Taken's estimator of the correlation dimension (middle left), CE: correlation entropy (middle right), LZC: Lempel–Ziv complexity (bottom left), H: Hurst exponent (bottom right) [40]. PhD Dissertation Universidad de Las Palmas de Gran Canaria 260 Figure 8-4: Data distribution of neutral, fear and anger emotional speech for each complexity measure extracted from the Berlin emotional database. MI: value of the first minimum of the mutual information function (upper left), SE: Shannon entropy (upper right), CD: Taken's estimator of the correlation dimension (middle left), CE: correlation entropy (middle right), LZC: Lempel–Ziv complexity (bottom left), H: Hurst exponent (bottom right) [40]. Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 261 Figure 8-5: Data distribution of neutral, fear and anger emotional speech for each complexity measure extracted from the LDC emotional database. MI: value of the first minimum of the mutual information function (upper left), SE: Shannon entropy (upper right), CD: Taken's estimator of the correlation dimension (middle left), CE: correlation entropy (middle right), LZC: Lempel–Ziv complexity (bottom left), H: Hurst exponent (bottom right) [40]. According to the Figure 8-3 the median of the values of the first minimum of the mutual information function (MI) between a signal and its delayed version is higher in neutral speech than in fear emotional speech and in anger emotional speech. This means that in the time of maximum difference (i.e. when the first minimum of the mutual information occurs) of a signal with its delayed version, this difference is lower in neutral speech than in fear or anger emotional speech. This tendency is the same in the three databases. In the case of the Shannon entropy distributions, there are not clear differences between neutral, fear and anger emotional speech. The median of the distributions are overlapped in the case of the LDC database. However, they are more clearly separated in the case of the Berlin emotional speech database and in the case of the Polish emotional speech database. PhD Dissertation Universidad de Las Palmas de Gran Canaria 262 The distribution of the Taken's estimator of the correlation dimension (CD) also shows a similar behavior in the three databases. The median values of the CD are higher in fear emotional speech and in anger emotional speech than in neutral speech. This is an indicator of a more complex geometrical structure in anger and fear emotional speech. However, there are not clear differences in the values of the CD between anger and fear emotional states in the Berlin database. The median of the correlation entropy (CE) distribution for neutral speech shows lower values than the median for anger and fear emotional speech. This is an indicator of a more complex structure in anger and fear speech than in neutral speech. The CE measure shows to be discriminative between neutral and fear and anger emotional speech. However, CE is less discriminative between fear and anger emotional speech. The Lempel-Ziv complexity (LZC) distribution shows values more near to 1 for fear and anger emotional speech than in the case of neutral speech. This means that anger and fear emotional speech records show more complexity than neutral speech records. Finally, the Hurst exponent (H) shows that the median value of neutral speech is higher than median of fear and anger emotional speech. Values of H for fear and anger emotional speech are closer to 0.5, showing that this signals has more randomness components. According to the data distributions, the six complexity measures are discriminative between neutral speech and negative emotional speech (anger emotional speech and fear emotional speech). However, in the case of the CE and CD the discrimination between fear emotional speech and anger emotional speech are less clear. Moreover, the data distributions are very similar in the three databases. This means that the discriminative ability of the complexity measures is independent from the language. Results of feature selection The following features were selected in the feature selection procedure: μMI, μH, sSE, sLZC. Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 263 Results of the databases evaluation with the selected features The selected features (μMI, μH, sSE and sLZC) were evaluated with a neural network classifier (with 5 neurons in the hidden layer) for the three emotional speech databases (Polish, Berlin and LDC databases). The global success rates for the different emotional databases are shown in Table 8-5 with the standard deviation (σ). According to the results, the selected features show a good discrimination ability between three emotional speech states (neutral state, fear emotional state and anger emotional state) in the three databases. TABLE 8-5: GLOBAL SUCCESS RATES OF THE SELECTED FEATURES IN THREE EMOTIONAL DATABASES Polish database Berlin database LDC database Success rates (%) 72.78 (σ = 5.13) 75.40 (σ = 3.86) 80.75 (σ = 3.75) Table 8-6, Table 8-7 and Table 8-8 show the confusion matrix of the selected features in the three emotional speech databases with the mean and standard deviation (σ) values obtained averaging the results for each individual experiment. According to the results, the selected features show good discrimination ability in the three databases between neutral, fear emotional state and anger emotional state. The discrimination ability between neutral and negative emotional states are higher than between fear and anger emotional speech in the case of the Polish emotional speech database and the Berlin emotional speech database. However, the LDC database shows good discrimination ability in anger emotional state against fear and neutral states. The similar results in the three databases show that the selected features are independent of the language. TABLE 8-6: CONFUSION MATRIX OF THE SELECTED FEATURES IN POLISH EMOTIONAL SPEECH DATABASE Classifier decision (%) Actual emotional state Neutral Fear Anger Neutral 88.00 (σ = 11.04) 8.00 (σ = 10.34 ) 4.00 (σ = 4.88) Fear 17.00 (σ = 10.06) 64.00 (σ = 14.77) 19.00 (σ = 14.54) Anger 7.33 (σ = 8.44) 26.33 (σ = 13.32) 66.33 (σ = 15.86) PhD Dissertation Universidad de Las Palmas de Gran Canaria 264 TABLE 8-7: CONFUSION MATRIX OF THE SELECTED FEATURES IN BERLIN EMOTIONAL SPEECH DATABASE Classifier decision (%) Actual emotional state Neutral Fear Anger Neutral 78.80 (σ = 7.81) 15.00 (σ = 6.45) 6.20 (σ = 3.89) Fear 16.00 (σ = 10.10) 63.20 (σ = 11.35) 20.80 (σ = 10.07) Anger 4.20 (σ = 5.34) 11.60 (σ = 8.38) 84.20 (σ = 8.74) TABLE 8-8: FUSION MATRIX OF THE SELECTED FEATURES IN LDC EMOTIONAL SPEECH DATABASE Classifier decision (%) Actual emotional state Neutral Fear Anger Neutral 77.57 (σ = 8.76) 16.70 (σ = 7.50) 5.74 (σ = 4.65) Fear 19.30 (σ = 9.57) 72.17 (σ = 8.33) 8.52 (σ = 4.78) Anger 3.30 (σ = 3.82) 4.17 (σ = 3.19) 92.52 (σ = 4.44) 8.2.4 Conclusions The usefulness of complexity features in discriminating between neutral state, fear emotional state and anger emotional state is evaluated. Six complexity measures including the value of first minimum of mutual information function, the Shannon entropy, the Takens estimator of the correlation dimension, the correlation entropy, the Lempel-Ziv complexity and the Hurst exponent are extracted from three emotional databases (the Polish emotional speech database, the Berlin emotional speech database and the LCD database). Then, the mean, standard deviation, skewness and kurtosis are applied to the six complexity measures and 24 features are obtained. Feature selection is accomplished to select a reduced number of features over the Polish emotional database. Finally, the selected features are evaluated in the Berlin emotional speech database and in the LDC emotional database using a neural network classifier. A qualitative analysis of the six complexity measures extracted from the three emotional databases is accomplished with the observation of the data distribution of the complexity measures. From this analysis, the following conclusions are extracted. According to the data distributions analysis, the six complexity measures are discriminative between neutral speech, fear emotional speech and anger emotional speech. In general, fear and anger emotional speech records show more complexity than neutral speech records. The reason can be that in fear and anger speech records, people Advances in preventive monitoring of machinery through audio and vibration signals Universidad de Las Palmas de Gran Canaria 265 tend to use more fricative sounds than in neutral state. Fricative sounds are noisier than voiced sounds. Moreover, the behavior of the data distributions of the six complexity measures is the same in the three databases. The databases consist of emotional speech records in Polish, German and English languages. We can conclude, therefore, that the complexity measures are independent from the language. The four selected features (mean of the value of the first minimum of the mutual information function, the mean of the Hurst exponent, the standard deviation of the Shannon entropy and the standard deviation of the Lempel-Ziv complexity) were evaluated with a neural network classifier in the three databases. Global success rates of 72.28%, 75.4% and 80.75%, were obtained for the Polish emotional speech database, the Berlin emotional speech database and the LDC emotional speech database respectively in the discrimination between neutral, fear and anger emotional states. Possibly applications of an automatic recognition system that discriminate between neutral emotion and negative emotions such as fear and anger can be applied in call centers in order to detect problems in costumer-system interaction and in security applications in order to detect security threats. 8.3 Contributions The contributions of this Chapter are the studies of nonlinear and complexity measures in pathological voice detection and in emotional voice detection. PhD Dissertation Universidad de Las Palmas de Gran Canaria 266 8.4 References [1] Henríquez, P., Alonso, J. B., Ferrer, M., Travieso, C. M., Godino-Llorente, J., & Díaz-de-María, F. (2009). Characterization of healthy and pathological voice through measures based on nonlinear dynamics. Audio, Speech, and Language Processing, IEEE Transactions on, 17(6), 1186-1195. [2] Hirano, M. Clinical Examination of Voice. New York: Springer, Verlag, 1981. [3] Nawka, T., Anders, L. C., & Wendler, J. (1994). Die auditive Beurteilung heiserer Stimmen nach dem RBH System. Sprache Stimme Gehör, 18, 130-133. [4] Boyanov, B., Hadjotodorov, S., Teston, B., & Doskov, D. (1997). Robust hybrid pitch detector for pathological voice analysis. In Larynx 97 (pp. 55–58). [5] Boyanov, B., Ivanov, T., & Chollet, G. (1993). Robust hybrid pitch detector. Electron. Lett., 29(22), 1924–1926. [6] H. Kasuya, Y. Endo, and S. Saliu, “Novel acoustic measurements of jitter and shimmer characteristics from pathologic voice,” in Proc. Eurospeech’ 93, Berlin, Germany, 1993, pp. 1973–1976. [7] C. Ludlow, C. Bassich, N. Connor, D. Coulter, and Y. Lee, “The validity of using phonatory jitter and shimmer to detect laryngeal pathology,” in Laryngeal Function in Phonation and Respiration. Boston, MA: Brown, 1987, pp. 492–508. [8] Yumoto, E., Gould, W. J., & Baer, T. (1982). Harmonics to noise ratio as an index of the degree of hoarseness. The journal of the Acoustical Society of America, 71(6), 1544-1550. [9] M. Yunik and B. Boyanov, “Method for evaluation of the noise-to-harmoniccomponent ratios in pathological and normal voices,” Acustica, vol. 70, pp. 89–91, 1990. [10] Kasuya, H., Ogawa, S., Mashima, K., & Ebihara, S. (1986). Normalized noise energy as an acoustic measure to evaluate pathologic voice. The Journal of the Acoustical Society of America, 80(5), 1329-1334. [11] M. Frohlich, D. Michaelis, and H. W. Srube, “Acoustic ‘Breathiness Measures’ in the description of pathologic voices,” in Proc. Acoust., Speech, Signal Process. ICASSP’98, Seattle, WA, 1998, vol. 2, pp. 937–940. [12] L. Gu, J. G. Harris, R. Shrivastav, and C. Sapienza, “Disordered speech assessment using automatic methods based on quantitative measures,” EURASIP J. Appl. Signal Process., vol. 9, pp. 1400–1409, 2005. [13] Boyanov, B., & Hadjitodorov, S. (1997). Acoustic analysis of pathological voices. A voice analysis system for the screening of laryngeal diseases.Engineering in Medicine and Biology Magazine, IEEE, 16(4), 74-82. [14] Wallen, E. J., & Hansen, J. H. (1996, October). A screening test for speech pathology assessment using objective quality measures. In Spoken Language, 1996. ICSLP 96. Proceedings., Fourth International Conference on (Vol. 2, pp. 776-779). IEEE. [15] De Oliveira Rosa, M., Pereira, J. C., & Carvalho, A. C. (1998, December). Evaluation of neural classifiers using statistic methods for identification of Laryngeal pathologies. In Proc. 5th Brazilian Symp. Neural Netw. (vol. 1, pp. 220– 225). [16] Hadjitodorov, S., & Mitev, P. (2002). A computer system for acoustic analysis of pathological voices and laryngeal diseases screening. Medical engineering & physics, 24(6), 419-429. 1El presente documento ha sido realizado a partir de la financiación del Programa de Formación del Personal Investigador de la Agencia Canaria de Investigación, Innovación y Sociedad de la Información del Gobierno de Canarias con una tasa de confinanciación del 85% del Fondo Social Europeo y de los proyectos TEC2009-14123-C04 y TEC2012-38630-C04-02 del Ministerio de Economía y Competitividad del Gobierno de España. UNIVERSIDAD DE LAS PALMAS DE GRAN CANARIA INSTITUTO UNIVERSITARIO DE SISTEMAS INTELIGENTES Y APLICACIONES NUMÉRICAS EN INGENIERÍA PROGRAMA DE DOCTORADO SISTEMAS INTELIGENTES Y APLICACIONES NUMÉRICAS EN INGENIERÍA Resumen de la Tesis Doctoral Avances en Monitorización Preventiva de Maquinaria a tráves de Señales de Audio y Vibración1 Autora: Patricia Henríquez Rodríguez Directores: Dr. D. Miguel Ángel Ferrer Ballester Dr. D. Jesús Bernardino Alonso Hernández Las Palmas de Gran Canaria, 22 de Octubre de 2015 Avances en monitorización preventiva de maquinaria a través de señales de audio y vibración Universidad de Las Palmas de Gran Canaria 275 A. Resumen de la Tesis 1 Introducción La maquinaria en general y los equipos industriales en particular se deterioran a lo largo del tiempo por el estrés que sufren durante su vida operativa. El fallo inesperado en una máquina trabajando en un entorno industrial puede tener consecuencias graves tanto desde el punto de vida humano, por el posible accidente que pueda causar, como desde el punto de vista de costes de productividad. Es por ello que el mantenimiento de equipos industriales se ha convertido en un aspecto estratégico en muchas empresas. La palabra mantenimiento se emplea para designar las técnicas usadas para asegurar el uso correcto y continuo de maquinaria, equipos e instalaciones. Monitorizar de forma continua el estado de funcionamiento (o condición) de un activo físico (una máquina, parte de una máquina o un sistema compuesto de varias máquinas) es de vital importancia para la detección temprana de fallos y tiene una gran influencia en la continuidad operacional de muchos procesos industriales. Ayuda a reducir costes de Tesis Doctoral Universidad de Las Palmas de Gran Canaria 276 mantenimiento e incrementa la seguridad y confiabilidad de los equipos industriales. En la siguiente tabla se muestran en porcentajes las ventajas del uso de un sistema de monitorización de la maquinaria [1]. TABLA 1: VENTAJAS DE UN SISTEMA DE MONITORIZACIÓN [1] Costes de mantenimiento Reducción del 50% al 80% Daños en los equipos Reducción del 50% al 60% Gastos en horas extras Reducción del 20% al 50% Esperanza de vida de la máquina Incremento del 50% al 60% Productividad total Incremento del 20% al 30% Los costes de mantenimiento se reducen debido al hecho de que se detectan fallos incipientes, evitando que dichos fallos crezcan y se conviertan en un problema grave y caro. Se reduce la probabilidad de aparición de fallos destructivos que dañen a los equipos y afecten a la seguridad de las personan. Se reducen las actividades de reparación y por lo tanto las horas extras dedicadas a ello. La esperanza de vida de la máquina se incrementa así como la productividad total puesto que se evitan paradas innecesarias de la maquinaria. El interés creciente en las técnicas de monitorización del estado de la maquinaria para la detección, diagnóstico y degradación de fallos tanto en el campo de la investigación como en el de la industria es evidente por la gran cantidad de artículos publicados en el campo, por los esfuerzos de las organizaciones de estandarización (ISO, SAE, etc) y por la organización de diferentes conferencias en el campo del diagnóstico de fallos, como por ejemplo la conferencia COMADEM (Condition Monitoring and Diagnostic Engineering Management). Avances en monitorización preventiva de maquinaria a través de señales de audio y vibración Universidad de Las Palmas de Gran Canaria 277 Técnicas de mantenimiento Desde tiempos antiguos, la humanidad ha usado diferentes técnicas de mantenimiento. Los hombres primitivos afilaban sus enseres y armas y cosían sus pieles. Durante la revolución industrial, se empezaron a usar relés de protección de sobre-corriente y protección de fallo de tierra mientras que los últimos desarrollos incluyen técnicas de procesado de la señal y reconocimiento de patrones. Generalmente, las técnicas de mantenimiento se dividen en dos: mantenimiento correctivo y mantenimiento preventivo. Este último se introdujo en los años 1950 y se divide a su vez en mantenimiento preventivo predeterminado y en mantenimiento basado en la condición, conocido como CBM en sus siglas en inglés (condition-based maintenance). El mantenimiento basado en la condición también se denomina mantenimiento predictivo. En el mantenimiento correctivo se toman acciones después de que el fallo haya ocurrido. Estas acciones van dirigidas a arreglar el fallo o a postponer su reparación de acuerdo al criterio de personal cualificado. En el mantenimiento preventivo predeterminado, se realizan actividades de mantenimiento planificadas a intervalos periódicos para evitar que los componentes se degraden hasta el punto de que la máquina deje de funcionar. La máquina o parte de la máquina se repara o se cambia antes de que ocurra un fallo. Tesis Doctoral Universidad de Las Palmas de Gran Canaria 278 Las técnicas correctivas y de mantenimiento preventivo predeterminado han demostrado ser bastante costosas puesto que muchas veces se realizan cambios de pieza innecesarios, al parar la máquina para realizar el cambio de pieza se para la producción con lo cual se aumentan los costes de producción. Además las actividades planificadas de mantenimiento sueles ser costosas. Por estas razones, algunas industrias empezaron a realizar mantenimiento basado en la condición en los años 1980. El mantenimiento preventivo basado en la condición (CBM) o mantenimiento predictivo se refiere a la monitorización del estado de la máquina en la cual se obtiene de forma continua información de parámetros que indican el estado de la máquina. La desviación de los parámetros de la condición normal indica el desarrollo de un fallo. La CBM lleva a cabo acciones de mantenimiento sólo cuando hay evidencia de comportamiento anormal. La CBM reduce el número de actividades planificadas reduciendo por tanto costes [1]. Etapas de un sistema de monitorización basado en la condición En la figura 1 se muestran las etapas típicas de un sistema de monitorización que implementa mantenimiento basado en la condición. Estas etapas incluyen la adquisición de datos, el procesado de los datos y un sistema de ayuda a la decisión. La salida del sistema de monitorización será el diagnóstico del fallo (detectar el fallo y saber dónde se ha producido) y posiblemente también la identificación del fallo (qué grado de severidad presenta el fallo) para así determinar qué tiempo de vida útil le queda al elemento que está siendo supervisado. Avances en monitorización preventiva de maquinaria a través de señales de audio y vibración Universidad de Las Palmas de Gran Canaria 279 A continuación se detallan cada una de las etapas de un sistema de monitorización de la condición. Puesto que la presente Tesis se centra en diversos aspectos de las distintas etapas de un sistema de monitorización de la condición, se explicará en cada etapa en qué se centra la presente Tesis. Adquisición de datos: esta etapa consiste en la obtención de información relevante sobre el estado de la máquina. Los datos adquiridos pueden variar según la clase de máquina o la naturaleza del fallo. La información adquirida puede ser de muchos tipos [2]: datos de tipo valor como presión, temperatura, datos de análisis de aceite; datos de forma de onda (es decir, señales) como señales de vibración, de audio, señales de emisión acústica; y datos multidimensionales como imágenes. El conjunto de datos adquiridos se denomina firma de la máquina. La presente Tesis se centra en señales de vibración y en señales de audio como fuente de información. Procesado de datos: los datos obtenidos en la etapa anterior se analizan para obtener información que permita detectar un posible fallo o diagnosticarlo. En esta Tesis se usan técnicas de procesado de la señal para analizar los datos y extraer información valiosa para la monitorización de la condición. Este proceso se denomina extracción de características. Sistema de ayuda a la decisión o la clasificación de los datos previamente analizados en diferentes estados de la condición. En esta Tesis, se usan técnicas de reconocimiento de patrones, en concreto dos clasificadores que permiten, una vez entrenados, diagnosticar de forma automática el estado de la máquina. Tesis Doctoral Universidad de Las Palmas de Gran Canaria 280 Diagnóstico de fallos/Predicción de fallo. El objetivo de un sistema de monitorización de la condición es el diagnóstico e identificación de fallos. El diagnóstico de fallo consiste en la detección del fallo, que responde a la pregunta “¿hay un fallo?” y al aislamiento del fallo, que responde a la pregunta “¿dónde está el fallo?” usando las etapas mencionadas. La identificación del fallo se refiere a la determinación de la severidad o el tamaño del fallo y a veces también a la determinación del tiempo de comienzo del fallo (o predicción del fallo). En esta Tesis, se realiza diagnóstico de fallo e identificación de fallo. Figura 1: Etapas de un sistema de monitorización basado en la condición Diagnóstico de fallos basado en vibración y audio: cojinetes y bombas centrífugas Tal y como se ha explicado en el apartado anterior en las etapas de un sistema de monitorización basado en la condición, a la hora de adquirir datos de la máquina se pueden usar multitud de datos diferentes. En esta Tesis nos centramos en señales de vibración y de audio (en el espectro audible 0-20kHz). Es por eso que en este subapartado se introduce el diagnóstico de fallos basados en estas dos señales. Asimismo, la investigación de la Tesis se centra en dos aplicaciones: cojinetes y bombas centrífugas. Los cojinetes son elementos esenciales en las máquinas rotativas puesto que soportan la estructura de la máquina permitiendo y facilitando su rotación. Diagnóstico/ Predicción Adquisición de datos (Firma de la máquina) Procesado de datos Sistema de ayuda a la decisión Avances en monitorización preventiva de maquinaria a través de señales de audio y vibración Universidad de Las Palmas de Gran Canaria 281 Un fallo no detectado en un cojinete puede causar una avería catastrófica (contactos indeseados entre partes fijas y móviles de la máquina, bloqueo del motor, etc.). Las máquinas rotativas están muy extendidas en la industria. Ejemplos de máquinas rotativas son los motores y los generadores. Debido a la gran utilización de este tipo de máquinas, la aplicación de técnicas encaminadas a la vigilancia y control del estado de los cojinetes adquiere suma importancia. Aunque el coste de los cojinetes en comparación con la máquina en sí es muy bajo, el hecho de que haya que desmontar la máquina casi en su totalidad para cambiar un cojinete hace que las técnicas de monitorización basadas en la condición sean muy útiles para permitir que estos elementos funcionen hasta el máximo de su vida útil. Por otra parte, las bombas centrífugas son máquinas que forman parte importante de muchos sistemas, puesto que permiten el movimiento de fluidos entre dos puntos del sistema. Se usan en la industria eléctrica, en la química, en la industria de extracción de petróleo, en sistemas de refrigeración, en spas y piscinas, etc. Por lo tanto, la monitorización de esta clase de máquinas es muy importante para el correcto funcionamiento de muchos sistemas industriales. Tanto en los cojinetes como en las bombas centrífugas, así como en muchas otras clases de máquinas su monitorización se hace atendiendo normalmente a señales de vibración. Por supuesto también se usan otro tipo de datos como la corriente, voltaje o temperatura en el caso de los cojinetes y la presión en el caso de las bombas centrífugas. El diagnóstico de fallos basado en vibración (vibration-based fault diagnosis) se refiere al diagnóstico de fallos usando la señal de vibración como fuente de información. El diagnóstico de fallos basado en vibración es un área de estudio muy Tesis Doctoral Universidad de Las Palmas de Gran Canaria 282 desarrollada que incluye un amplio rango de técnicas que han evolucionado de forma rápida durante las últimas décadas. En la monitorización de la condición, las técnicas de diagnóstico de fallos basadas en vibración han sido ampliamente usadas debido a la facilidad de adquirir la señal de vibración de la máquina. Por esta razón, la mayoría de los artículos de investigación en la literatura del diagnóstico de fallos están enfocados al diagnóstico de fallos basado en vibración [2],[3],[46]. El diagnóstico de fallos basado en audio se refiere al diagnóstico de fallos usando la señal de audio como fuente de información (audio-based fault diagnosis o también llamado airborne fault diagnosis). El diagnóstico de fallos basado en audio usando micrófonos en el rango audible (0-20kHz) es un campo emergente con un gran potencial en el campo del diagnóstico de fallos puesto que los micrófonos son sensores no invasivos (no van montados encima de la máquina) y tienen mayores posibilidades de localización que los acelerómetros. Además, el uso de señales de audio podría mejorar la inspección de ciertos entornos industriales en los cuales un sistema de monitorización fijo es caro o el montaje de los sensores de vibración es complicado como por ejemplo en la industria de extracción de petróleo. Por estas razones, pensamos que el diagnóstico basado en audio necesita más esfuerzos de investigación. De hecho, la monitorización de la condición con señales de audio presenta menos estudios científicos que la vibración [2],[3],[46]. Por lo tanto, esta Tesis se enfoca en dos aplicaciones: cojinetes y bombas centrífugas. En el caso de los cojinetes se usan diferentes bases de datos de vibración públicas y nos centramos tanto en el diagnóstico de fallo como en la identificación del fallo (severidad del fallo). La investigación se ha enfocado en la etapa de procesado de Avances en monitorización preventiva de maquinaria a través de señales de audio y vibración Universidad de Las Palmas de Gran Canaria 289 Figura 3: Distribución de los artículos usados para la redacción del estado del arte: distribución de artículos basados en señales de vibración y basados en señales de audio (superior izquierda); años de publicación de los artículos (superior derecha); distribución de elementos donde se producen los fallos analizados con señales de vibración (inferior izquierda) como con señales de audio (inferior derecha). De la figura 3 se puede observar también que una gran cantidad de fallos se produce en los cojinetes. De hecho, de acuerdo con una encuesta sobre confiabilidad de motores [6] el 42% de los fallos de motores de más de 200 caballos de potencia se deben a fallos en los cojinetes. Motores de menor caballaje presentan aún más porcentaje de fallos en los cojinetes, llegando a ser del 90% [7]. Como ya se ha comentado en la introducción, las técnicas de monitorización de la condición son muy útiles para detectar si un cojinete presenta algún tipo de fallo. También en la figura 3 se puede observar que la mayor parte de la monitorización de cojinetes se realiza usando Tesis Doctoral Universidad de Las Palmas de Gran Canaria 290 señales de vibración. Por otra parte, a lo largo de la elaboración del estado del arte se realizó una búsqueda exhaustiva de bases de datos públicas, de libre acceso y gratuitas. Fruto de esa búsqueda se obtuvieron tres bases de datos de señales de vibración de cojinetes [30]-[32]. Basándonos en estas observaciones derivadas del estado del arte, la primera parte de esta Tesis se centra en el estudio de señales de vibración de cojinetes para el diagnóstico e identificación de fallos. En la literatura de diagnóstico de fallos de cojinetes se encuentra una amplia variedad de técnicas de procesado de la señal usadas para extraer características de la señal de vibración en diferentes dominios de representación: dominio del tiempo [37], de la frecuencia [9], cepstral [10], tiempo-frecuencia [11]-[15], así como técnicas no lineales basadas en medidas de dinámica no lineal y medidas de complejidad y predictibilidad [28], [43]. Las técnicas no lineales permiten extraer características de la firma de la máquina que pueden revelar un entendimiento más preciso de la señal. De hecho, se han encontrado indicios de no linealidad en el funcionamiento de cojinetes [28]. Por este motivo, en la fase de procesado de la señal (en donde se extraen las características), nos hemos centrado en el estudio de técnicas no lineales dentro de la aplicación de los cojinetes tanto para diagnóstico como para identificación de fallos. Si bien la monitorización usando señales de vibración es mucho más usada en comparación con el audio, en los últimos años se ha incrementado la investigación en torno a la monitorización basada en señales de audio [5],[19]-[27]. Además, si bien la localización de los micrófonos debe de ser lo más cercana posible a la máquina para evitar en la medida de lo posible interferencias con otros elementos, estos tienen más posibilidades de localización que los sensores de vibración [5]. Además, no tienen que Avances en monitorización preventiva de maquinaria a través de señales de audio y vibración Universidad de Las Palmas de Gran Canaria 291 ser montados en la máquina, por lo que son elementos no invasivos en absoluto. Basándonos en estas observaciones, la segunda parte de la presente Tesis se centra en explorar la monitorización del estado de la máquina usando señales de audio. Puesto que hay una carencia de bases de datos públicas de señales de audio en diagnóstico de fallos en maquinaria, en la presente Tesis se opta por grabar una base de datos propia. Dicha base de datos consta de señales de audio y de señales de vibración adquiridas simultáneamente de una bomba centrífuga de agua. Se adquieren también señales de vibración para comparar resultados. Como ya hemos explicado, las señales de audio son menos usadas que las de vibración en el campo de diagnóstico de fallos, y el caso de las bombas centrífugas no es una excepción. Es por esto que la mayoría de características usadas en diagnóstico de fallos en bombas de agua son extraídas de señales de vibración. Por este motivo, en esta Tesis se realiza un estudio de las medidas usadas en señales de vibración para diagnóstico de fallos en bombas centrífugas y estas medidas se extraen también de las señales de audio. De igual forma se pretende, en base al estudio de las señales de audio y vibración de la bomba centrífuga proponer otras medidas que sean capaces de discriminar entre diferentes estados de normalidad y fallo de la bomba centrífuga. Otra conclusión que se puede extraer al analizar el estado del arte es que hay escasos estudios en los que se trabaje conjuntamente con señales de audio y vibración [5], [22], [23]. Es por este motivo que en la presente Tesis se realiza un estudio de la combinación de señales de audio y vibración. Tesis Doctoral Universidad de Las Palmas de Gran Canaria 292 3 Objetivos El objetivo de la presente Tesis es la mejora de los sistemas de monitorización del estado de la maquinaria en diagnóstico e identificación de fallo usando señales de vibración y de audio en dos aplicaciones (cojinetes y bombas de agua) con especial énfasis en la etapa de extracción de características y en la utilización del audio como fuente de información audio. El objetivo de la presente Tesis puede ser dividido en objetivos parciales que deben ser alcanzados para lograr el objetivo general. Seguidamente se exponen los objetivos parciales: 1. Realizar una revisión general de las técnicas de procesado de señal usadas para la extracción de diferentes características así como de las técnicas reconocimiento de patrones usadas en el diagnóstico de fallos de maquinaria usando señales de vibración y de audio. 2. Realizar una búsqueda de bases de datos de vibración de cojinetes sin fallo y con diferentes tipos de fallos para generar así un repositorio de bases de datos con las que trabajar. 3. Desarrollar métodos no lineales que ayuden al diagnóstico y a la identificación de fallos en cojinetes. Avances en monitorización preventiva de maquinaria a través de señales de audio y vibración Universidad de Las Palmas de Gran Canaria 293 4. Generar una base de datos de audio y vibración de una bomba centrífuga en diferentes estados de funcionamiento (sin fallo y con diferentes tipos de fallos) para poder realizar una evaluación de la capacidad de las señales de audio en el diagnóstico de fallos y compararlas con las señales de vibración. 5. Aplicar características del estado del arte usadas comúnmente en señales de vibración de bombas centrífugas a las señales de audio captadas de una bomba centrífuga y cuantificar la capacidad de las medidas en la discriminación entre varios estados de funcionamiento de la bomba centrífuga (estado normal y estados de fallo). 6. Buscar nuevas medidas para diagnóstico de fallos en bombas centrífugas y cuantificar su capacidad de discriminación entre varios estados de funcionamiento de la bomba centrífuga. 7. Comparar los resultados obtenidos con señales de vibración y de audio. 8. Realizar un estudio de utilización conjunta de señales de audio y vibración en la monitorización de una bomba centrífuga Tesis Doctoral Universidad de Las Palmas de Gran Canaria 294 4 Metodología En el presente apartado se expone la metodología seguida en la presente Tesis para cumplir los objetivos propuestos. Estado del arte En primer lugar se realiza un estudio exhaustivo del estado de la técnica o estado del arte en el ámbito de la monitorización de maquinaria con señales de audio y vibración. Nos centramos en el estudio de técnicas de procesado de la señal para extraer características y en técnicas de reconocimiento de patrones usadas para discriminar entre diferentes estados de funcionamiento de la máquina. En el capítulo 2 de la Tesis se encuentra el estado del arte junto con un análisis crítico del mismo. Obtención y generación de bases de datos Se realiza una búsqueda de bases de datos que comprendan muestras de cojinetes en estado de funcionamiento normal (sin fallo) y funcionando con diversos fallos. Las bases de datos obtenidas son de vibración y están disponibles en internet [30]-[32]. Una de las bases de datos [30] comprende señales de vibración de cojinetes sin fallos y de cojinetes con fallos puntuales en el anillo externo (outer-race fault), en el anillo interno (inner-race fault) y con fallo en los elementos de rodamiento (ball fault). A modo de ilustración, en la figura 4 se muestran los diferentes elementos que componen un cojinete. Las otras dos bases de datos [31], [32] son test-to-failure. Es decir, se graba la señal de vibración hasta que el cojinete o cojinetes tienen un fallo. Avances en monitorización preventiva de maquinaria a través de señales de audio y vibración Universidad de Las Palmas de Gran Canaria 295 Figura 4: Componentes de un cojinete cuyo elemento de rodamiento son bolas. Por otra parte, dado que no hay bases de datos públicas con señales de audio proveniente de maquinaria, se decide grabar una base de datos propia de señales de audio y vibración adquiridas de forma simultánea de una bomba de agua centrífuga de circulación (modelo ALP800) en un circuito cerrado formado por la bomba en sí y un tanque con 50 litros de agua. Antes de proseguir con la base de datos generada se explicará brevemente los principales elementos de una bomba centrífuga. Una bomba centrífuga consta de dos elementos fundamentales, el rodete y la voluta. El rodete es el elemento rotatorio y la voluta el elemento estacionario. El rodete convierte la energía suministrada por el motor en energía cinética. La rotación del rodete fuerza al líquido a circular a través de la bomba desde la dirección axial hasta la dirección radial mientras se transfiere energía al líquido bombeado. La voluta convierte la energía cinética en energía de presión. En resumen, el rodete produce velocidad en el líquido y la voluta convierte dicha velocidad en presión. A modo de ilustración en la figura 5 se muestran las diferentes partes de una bomba de agua. Se indica también la dirección de fluido o líquido bombeado. Tesis Doctoral Universidad de Las Palmas de Gran Canaria 296 Figura 5: Partes principales de una bomba centrífuga. Dirección del fluido en la bomba (figura izquierda). Fuente: [32]. Fluid in: entrada del fluido. Fluid out: salida del fluido. Velocidad y presión del fluido en una bomba. Figura modificada de la fuente: [34]. En la figura 6 se muestra un rodete y se indican las diferentes partes que lo conforman. El rodete de una bomba centrífuga presente una serie de vanos delimitados por palas curvadas. Dichas palas también se denominan álabes. Los álabes son los que imparten fuerza centrífuga al fluido. Un rodete cerrado como el que se muestra en la figura 6 tiene platos en ambos lados que encierran completamente el rodete desde el ojo de succión (u ojo de rodete) a sus bordes. El ojo de rodete es la parte central del rodete. La región del álabe más cercana al ojo del rodete se denomina leading edge y la región en el borde del álabe se denomina trailing edge. Avances en monitorización preventiva de maquinaria a través de señales de audio y vibración Universidad de Las Palmas de Gran Canaria 297 Figura 6: Partes de un rodete. Figura modificada de la fuente: [32]. En la figura 7 se muestra un plano de la sala donde se graba la base de datos de la bomba centrífuga. En el plano se observa un esquema del montaje realizado con la bomba (en verde), el tanque con 50 litros de agua (en celeste) y las tuberías que forman un circuito cerrado (en naranja). También se indica en la figura 7 la dirección de flujo del agua. La sala donde se graba la base de datos presenta 29dB de aislamiento acústico respecto a ruido aéreo. En el montaje se usan 4 metros de tuberías de 3.81 cm de diámetro. La bomba usada es una bomba centrífuga circuladora de agua modelo ALP800 de 0.5 caballos de potencia y con 2925 RPM (revoluciones por minuto), equivalente a 48.75 Hz. El rodete de la bomba presenta 7 álabes. Tesis Doctoral Universidad de Las Palmas de Gran Canaria 298 Figura 7: Esquemático de la sala donde se realiza la grabación de la base de datos junto con la posición y esquema del montaje. En las figuras 8 y 9 se observan fotos de la bomba centrífuga modelo ALP800 utilizada para generar la base de datos de la presente Tesis. En la figura 7 se muestra la bomba de agua sin desmontar y en la figura 8 se muestra el rodete (foto de la izquierda) y la voluta. Figura 8: Fotos de la bomba de agua modelo ALP 800. Dirección del flujo Avances en monitorización preventiva de maquinaria a través de señales de audio y vibración Universidad de Las Palmas de Gran Canaria 305 métodos del estado del arte. En la figura 14 se observan el método propuesto (método 1) y los demás métodos del estado del arte que se evalúan: en el método 2 se extraen de la señal de vibración en el dominio del tiempo (señal T) las características estadísticas y de energía [8], en el método 3 se extraen de la señal de vibración demodulada en amplitud (señal AM) usando la técnica de análisis de envolvente comúnmente utilizada en cojinetes [39] y en el método 4 se extraen de la señal demodulada en amplitud obtenida a partir del operador Teager-Kaiser (señal TK-AM) [41]. Figura 14: Esquema de la experimentación en diagnóstico de fallos de cojinetes. Una vez obtenidas las características en cada uno de los métodos, se evalúa la habilidad de dichas características para discriminar entre normalidad y diferentes tipos de fallos de cojinetes. Para ello se realiza un estudio de relevancia de las características usando un método de selección de características denominado sequential floating forward selection (SFFS) [44]. El método de selección de características usado selecciona un subgrupo de características que mejores resultados dan a la hora de discriminar entre las diferentes clases del problema (en este caso entre normal, fallo en Tesis Doctoral Universidad de Las Palmas de Gran Canaria 306 cojinete externo, fallo en cojinete interno y fallo en rodamiento). La selección de características se aplica varias veces para obtener finalmente las características en orden de relevancia. Finalmente se usan dos clasificadores, uno de redes neuronales y otro de máquinas de soporte vectorial de mínimos cuadrados (LS-SVM) para evaluar las medidas ordenas por relevancia. Los resultados obtenidos con el método propuesto mejoran los obtenidos con los demás métodos comparados [47]. Es importante recalcar que a la hora de evaluar las características con los clasificadores se divide la base de datos en un conjunto de entrenamiento formado por el 70% de las muestras de cada clase y en un conjunto de test formado por el 30% de las muestras restantes. Con el conjunto de entrenamiento se ajustan los parámetros de los clasificadores usando la técnica de validación cruzada k-fold con k = 3. En el caso de las redes neuronales, donde el clasificador usado tiene la capa de entrada, la capa de salida y una capa oculta, se ajusta el número de neuronas de la capa oculta. En el caso del clasificador LS-SVM que ha sido programado para que utilice como kernel funciones de base radial Gaussiana, se ajusta el parámetro de regularización y el ancho de banda de la función de base radial Gaussiana. Una vez aplicada la técnica propuesta a la base de datos de cojinetes que presenta fallos en diferentes partes de los mismos (base de datos Case Western) [30] se aplica la misma técnica a la base de datos de degradación del helicóptero [31] en la que un cojinete situado en una parte bastante inaccesible de uno de los trenes principales de transmisión de un helicóptero Black-Hawk tiene un fallo en uno de sus elementos rodantes al final de un test de resistencia. En la figura 15 se muestra el esquema de la propuesta aplicado a la degradación del cojinete. Los resultados usando el método Avances en monitorización preventiva de maquinaria a través de señales de audio y vibración Universidad de Las Palmas de Gran Canaria 307 propuesto mejoran resultados obtenidos con esa misma base de datos a la hora de detectar el comienzo del fallo [49]. Figura 15: Diagrama de la aplicación del método propuesto a la degradación de un cojinete. En la primera parte del capítulo 4 se encuentra el desarrollo del método propuesto y de la experimentación realizada junto con los resultados obtenidos. Metodología para detección e identificación de fallo en cojinetes Para desarrollar un índice que sirva para identificar la severidad del fallo en cojinete se realiza primeramente un estudio del estado de la técnica en técnicas de identificación de fallos (es decir, técnicas que son capaces de determinar la severidad del fallo en cuestión). Basándonos en ese estudio se propone la aplicación de la transformada wavelet packet a la señal de vibración seguida de la complejidad de Lempel-Ziv para la identificación de fallos en la cara externa y en la cara interna de cojinetes. El método es capaz de detectar fallo en los elementos de rodamiento, pero no de determinar la severidad del mismo. Además, el método propuesto es capaz de seguir de forma Tesis Doctoral Universidad de Las Palmas de Gran Canaria 308 monótona la evolución de fallos en cara externa. Se compara el método propuesto con otros tres métodos usados en la literatura: kurtosis, complejidad Lempel-Ziv aplicada a la señal de vibración sin preprocesar, kurtosis aplicado al nodo de máxima energía de la wavelet packet [37], [43]. También se realiza un estudio de la robustez del método propuesto frente a ruido blanco Gaussiano. La figura 16 muestra un diagrama del método propuesto para la identificación de fallos en la cara externa e interna de cojinetes. Figura 16: Método propuesto para la identificación de fallos en cojinetes. Primero la señal de vibración se descompone usando la transformada wavelet packet con nivel de descomposición 3 (donde hay 8 nodos). La wavelet madre usada para la descomposición es la Daubechies 6 (db6) (ver figura 17) porque tiene una forma Descomposición “Wavelet packet” (3er nivel, 8 nodos) Cálculo de la energía relativa por nodo Selección del nodo con mayor energía Entrada: Señal de vibración Salida: característica extraída de la señal de entrada Cálculo de la complejidad LempelZiv Reconstrucción de la señal del nodo kmax Avances en monitorización preventiva de maquinaria a través de señales de audio y vibración Universidad de Las Palmas de Gran Canaria 309 de onda impulsiva, igual que la forma de onda de las señales de vibración de cojinetes con fallos puntuales. Una vez descompuesta la señal de vibración se reconstruye cada una de las señales de los nodos. Se calcula luego la energía de esas 8 señales con respecto a la energía de la señal de vibración original. Finalmente se selecciona el nodo con mayor energía relativa. Puesto que se ha usado una wavelet madre con forma impulsiva, el nodo con mayor energía será aquel donde se encuentre el fallo si lo hubiese. Así la transformada wavelet elimina partes de la señal indeseada y se centra en el fallo. Una vez obtenido el nodo con mayor energía se calcula la complejidad de Lempel-Ziv sobre la señal reconstruida del nodo de mayor energía. Con la complejidad de Lempel-Ziv se evalúa, como su propio nombre indica, la complejidad de la señal obtenida. Figura 17: Forma de onda de la wavelet madre Daubechies 6. El método propuesto se aplica a la base de datos Case Western [30] puesto que esta base de datos presenta señales de vibración de cojinete con fallos de diferente severidad (severidad leve, media y alta) en cara externa, cara interna y rodamiento. Asimismo también se aplican los otros tres métodos de la literatura mencionados anterioremente [42]. Los resultados obtenidos demuestran que el método propuesto da mejores resultados [52] sobre todo cuando se añade ruido blanco Gaussiano a las señales para comprobar lo robusto del algoritmo propuesto frente al ruido. 0 2 4 6 8 10 12 -1.5 -1 -0.5 0 0.5 1 1.5 Tesis Doctoral Universidad de Las Palmas de Gran Canaria 310 El método propuesto también se aplica a las otras dos bases de datos de cojinetes, las base de datos de degradación de cojinetes del IMS [32] y del helicóptero UH-60 (Black Hawk) [31]. En la base de datos del IMS el cojinete presenta un fallo de cara externa a lo largo del test run-to-failure. El método propuesto obtiene mejores resultados que los otros tres métodos con los que se ha comparado. En el caso de la base de datos del helicóptero, donde el fallo se produce en el elemento rodante del cojinete, los resultados no son tan satisfactorios y el método propuesto se comporta de forma similar el método de extraer la complejidad de Lempel-Ziv de la señal de vibración sin pre-procesar. En la segunda parte del capítulo 4 se encuentra el desarrollo del método propuesto para la identificación de fallos en la cara interna y la cara externa de cojinetes así como la experimentación realizada junto con los resultados obtenidos. Metodología usada para el diagnóstico de fallos en bomba de agua usando señales de audio y vibración Una vez generada la base de datos de audio y vibración de la bomba centrífuga, nos centramos en el diagnóstico de los fallos generados en la bomba. El objetivo principal en este caso es determinar si con el audio como fuente de información se es capaz de distinguir entre los diferentes fallos de la bomba centrífuga. De igual forma se pretende comparar los resultados con los obtenidos usando señales de vibración. La metodología usada para el diagnóstico de fallos en bomba de agua usando señales de audio y vibración se explica con la ayuda de la siguiente figura (figura 18). Avances en monitorización preventiva de maquinaria a través de señales de audio y vibración Universidad de Las Palmas de Gran Canaria 311 Figura 18: Esquema de la metodología llevada a cabo para cuantificar la habilidad de las características extraídas de la señal de audio y vibración en diagnóstico de fallos en bombas centrífugas. Extracción de características En el bloque de Extracción de características se extrae un conjunto de características tanto de la señal de vibración como de la señal de audio. En realidad, esto se hacer por sensor. Puesto que hay 4 sensores (2 de audio y 2 de vibración) se extraen características de las señales provenientes de los 4 sensores. Las características que se extraen son características usadas en la literatura del diagnóstico de fallos en bombas centrífugas usando señales de vibración y de audio. Para identificar las características primero se debe realizar un estudio de las técnicas de procesado de la señal usadas en bombas centrífugas para el diagnóstico de fallos. Es necesario mencionar que puesto que la mayoría del diagnóstico de fallo en bombas centrífugas se realiza con señales de vibración así como con señales de presión, se encuentran escasas características que se hayan extraído usando el audio. Por ello las mismas características que se extraen de la señal de vibración se extraen también de la señal de audio. Basándose en la inspección visual de las señales obtenidas en la base de datos generada, se proponen una serie de nuevas medidas no utilizadas anteriormente en el diagnóstico de fallos de bombas centrífugas. Extracción de características Clasificación Selección de características selection Señal Diagnóstico de clasificación Tesis Doctoral Universidad de Las Palmas de Gran Canaria 312 Una vez identificadas las características del estado del arte y propuestas nuevas medidas, éstas se deben implementar. Para extraer las características cada observación (o muestra) de la base de datos generada se divide en segmentos, también llamados tramas, de 8192 muestras cada uno, equivalentes a 371.5 ms. Las tramas se toman una cada tres por lo que si cada observación dura 59 segundos, tenemos 53 tramas por observación. Por cada una de las tramas se extrae el conjunto de características implementado. Antes de extraer las características a cada una de las observaciones se le quita la media y se normaliza entre -1 y 1. Una vez extraídas las características por observación, se calcula el promedio de cada una de las características por observación. Selección de características Una vez que se extraen las características por observación de la base de datos, el objetivo es identificar aquellas más apropiadas por sensor que mejor discriminen entre los diferentes estados de la bomba centrífuga. El proceso de seleccionar características de un conjunto de características se denomina selección de características. Este proceso es importante puesto que la selección de características que aporten información relacionada con el fallo y el descarte de características que no aporten información mejora la tasa de éxito en el diagnóstico. Así pues, en el segundo bloque de la figura 18 se realiza la selección de características y un análisis de relevancia, el cual se explicó anteriormente cuando se explicó la metodología para el diagnóstico de fallos en cojinetes. Simplemente recordar que la técnica utilizada para la selección de Avances en monitorización preventiva de maquinaria a través de señales de audio y vibración Universidad de Las Palmas de Gran Canaria 313 características se denomina sequential floating forward selection (SFFS) [44]. La razón por la que se ha elegido esta técnica de selección de características es porque es un método determinista y por lo tanto se obtendrán los mismos resultados cuando se repita. De igual forma el coste computacional de la técnica es bajo. Después de aplicar la selección de características y de realizar el análisis de relevancia se obtendrá un conjunto de características por sensor. Evaluación de las características: Clasificación El bloque de clasificación de la figura 18 se utiliza para evaluar la capacidad de las características a la hora de discriminar entre los diferentes estados de la bomba centrífuga. Así que el clasificador nos dará una tasa de éxito en la clasificación de los diferentes estados de funcionamiento de la máquina. El clasificador clasifica las observaciones en diferentes unidades de clasificación también llamadas clases con la información que obtiene las características. Las unidades de clasificación corresponden con los diferentes estados o condiciones de la máquina que se quieran clasificar. En la presente Tesis, se ha decidido implementar dos configuraciones diferentes, así que las unidades de clasificación consideradas son las siguientes (entre paréntesis está el acrónimo de la clase):  8 unidades de clasificación: estado normal (NOR), fallo en el plato del rodete (PLA), fallo en el leading edge (LED), fallo en el trailing edge (TED), fallo en Tesis Doctoral Universidad de Las Palmas de Gran Canaria 314 la goma del sellado (SEA), arena añadida al agua (SAN), arena y papel añadidos conjuntamente al agua (SAP) y bolas de PVC añadidas al agua (PVC).  17 unidades de clasificación. En este caso se tienen en cuenta las severidades en los casos de rotura del plato, del fallo leading edge y del fallo trailing edge. Las 17 unidades de clasificación son: estado normal (NOR), fallo en el plato del rodete con una porción de plato eliminada (PLA1), con dos porciones de plato eliminadas (PLA2) y con tres porciones de plato eliminadas (PLA3), fallo en el leading edge (LED1) de 5 mm de tamaño, de 10 mm de tamaño (LED2) y de 15 mm de tamaño (LED3), fallo en el trailing edge de 5 mm de tamaño (TED1), de 10 mm de tamaño (TED2) y de 15 mm de tamaño (TED3), fallo en la goma del sellado (SEA), arena añadida al agua (SAN), arena y papel añadidos conjuntamente al agua (SAP) y bolas de PVC añadidas al agua (PVC). Al igual que en el caso del diagnóstico de fallos en cojinete se usan dos clasificadores: un clasificador de red neuronal y otro de máquinas de soporte vectorial de mínimos cuadrados (LS-SVM). El clasificador de red neuronal tiene estructura feedforward y un algoritmo de resilient backpropagation. El clasificador LS-SVM usa funciones Gaussianas de base radial. Al igual que en el caso de diagnóstico de fallos en cojinetes, la metodología para evaluar los diferentes casos (8 unidades de clasificación y 17 unidades de clasificación) consiste en dividir la base de datos en un conjunto de entrenamiento formado por el 70% de las observaciones de cada una de las clases de la base de datos y en un conjunto de test formado por el 30%. El conjunto de entrenamiento a su vez se usa para ajustar los parámetros de los clasificadores. El procedimiento se repite 20 veces y se promedian los resultados. Avances en monitorización preventiva de maquinaria a través de señales de audio y vibración Universidad de Las Palmas de Gran Canaria 321 ruido frecuencial (2 características), energía de ruido frecuencial (1 característica), medidas no lineales (7 características). Parte de estas contribuciones han sido publicadas en conferencias [51], [52]. En el capítulo 5 se detallan las medidas aportadas. Aportación original 7: evaluación de las medidas en señales de audio para el diagnóstico de fallos en bomba centrífuga La capacidad para discriminar entre diferentes estados de la maquinaria (sin fallo y con diferentes tipos de fallos) de las características extraídas se ha evaluado usando dos clasificadores, uno basado en redes neuronales y otro basado en máquinas de soporte vectorial de mínimos cuadrados (LS-SVM). Los resultados de la evaluación de las características usando la señal de audio son muy satisfactorios, llegándose a tasas de acierto del 96.53% en la discriminación entre 8 estados (estado normal de funcionamiento y 7 estados de fallo) usando sólo 7 características seleccionadas y de 88.29% en la discriminación entre 17 estados (estado normal y 10 estados de fallo en los que se separan fallos por severidades) usando 16 características seleccionadas. También en el capítulo 5 se describe la evaluación de las medidas y los resultados obtenidos en dicha evaluación tanto para las señales de audio como para las señales de vibración por sensor individual. Aportación original 8: estudio de la fusión de señales de audio y vibración en el diagnóstico de fallos en bomba centrífuga Finalmente se ha realizado un estudio de la fusión de las señales de audio y vibración para determinar si su unión mejora el diagnóstico de fallos en la aplicación e bomba Tesis Doctoral Universidad de Las Palmas de Gran Canaria 322 centrífuga. Los resultados obtenidos son satisfactorios llegándose a aumentar la tasa de acierto en el diagnóstico del fallo hasta un 100%. En el capítulo 6 se explican las técnicas de fusión usadas y se muestran los resultados obtenidos. Por último, y como ya se comentó en la introducción, durante el transcurso de la presente Tesis la doctoranda siguió con la investigación en señales de voz. En esta área se realizan las siguientes aportaciones: Aportación original 9 (voz): propuesta de características basadas en dinámica no lineal para la detección de patologías laríngeas del sistema fonador Se proponen una serie de características basadas en dinámica no lineal (teoría del caos) para la detección de patologías del sistema fonador usando la señal de voz. Fruto de este estudio se genera un artículo en revista de índica de impacto [53]. Aportación original 10 (voz): propuesta de características basadas en dinámica no lineal y de medidas de complejidad para la detección de emociones a través de la señal de voz Se proponen una serie de características basadas en dinámica no lineal, así como medidas de complejidad para la discriminación de diferentes estados de emoción usando la señal de voz. Se genera de este trabajo dos artículos en revistas con índice de impacto [54], [55]. En el anexo de la Tesis se hace un resumen de ambos trabajos en voz, explicando la metodología usada, las características extraídas y los resultados obtenidos. Se debe mencionar que parte de las medidas propuestas en voz también han sido propuestas en esta Tesis para el diagnóstico de fallos en bombas centrífugas. Avances en monitorización preventiva de maquinaria a través de señales de audio y vibración Universidad de Las Palmas de Gran Canaria 323 6 Conclusiones El objetivo de la presente Tesis es mejorar la tasa de acierto en los sistemas de monitorización del estado de la máquina en diagnóstico e identificación (severidad del fallo) de fallos usando señales de audio y vibración y centrándonos en dos aplicaciones (cojinetes y bombas centrífugas) con especial énfasis en la etapa de extracción de características y en el uso de señales de audio como fuente de información. Basándonos en la investigación realizada en las dos áreas de aplicación de la Tesis, cojinetes y bombas centrífugas, y en los resultados obtenidos podemos concluir que tanto el uso de audio como fuente de información como el uso de técnicas no lineales a la hora de extraer características mejora la tasa de éxito en las dos aplicaciones en las que se ha centrado la Tesis: cojinetes y bombas centrífugas. En la aplicación de cojinetes se han usado señales de vibración como fuente de información puesto que se han usado bases de datos públicas disponibles en Internet con muestras de cojinetes funcionando en estado normal y con fallos puntuales en las diferentes partes de los cojinetes. Se han propuesto dos métodos basados en técnicas no lineales para el diagnóstico de fallos en la cara externa, en la cara interna y en los elementos rodante de cojinetes y para la identificación de fallos en la cara externa y en la interna. Con el segundo método propuesto se ha generado un índice que sigue de forma monótona la degradación de un cojinete con fallo normal a fallo en la cara externa. Los métodos propuestos se han comparado con métodos de la literatura y los resultados obtenidos han sido satisfactorios, mejorando los resultados. Además, en el segundo método propuesto se ha realizado un estudio de la robustez del algoritmo frente a ruido Gaussiana. El estudio demuestra que el algoritmo propuesto presenta más Tesis Doctoral Universidad de Las Palmas de Gran Canaria 324 robustez al ruido blanco Gaussiano que los métodos de la literatura con los que se ha comparado. Por lo tanto se puede concluir que el uso de técnicas no lineales para el diagnóstico e identificación de fallos en cojinetes es una línea de investigación prometedora. La aplicación de la bomba centrífuga se ha usado para explorar el uso de la señal de audio en el diagnóstico de fallos de bombas centrífugas. La monitorización de maquinaria en general, y en particular la referida a bombas de agua y a cojinetes, utiliza en la mayoría de los casos la señal de vibración como fuente de información. El uso de audio (rango de frecuencias [0-20kHz]) ha sido menos explorado, por esto en la presente Tesis nos hemos centrado en su estudio. Para ello se necesita obtener una base de datos de audio obtenida de una bomba centrífuga. Puesto que no hay ninguna disponible, se ha generado una base de datos propia en la que se adquieren señales de audio y vibración de forma simultánea de una bomba centrífuga funcionando con normalidad y a la que se generan diferentes tipos de fallos. Un conjunto de características del estado del arte de diagnóstico de fallos en bombas centrífuga se extraen de las señales de vibración obtenidas y se aplican también a las señales de audio. También se proponen un conjunto de 31 nuevas medidas extraídas en el dominio de la frecuencia, en el dominio de los cepstrum y en el dominio no lineal. Se realiza una selección de características y luego se evalúan por sensor con dos clasificadores. Los resultados muestran tasas de acierto altas y bastante similares en sensores de audio y de vibración. Se obtienen tasas de éxito del 99.82% y del 99.59% para los sensores de audio y del 99.91% y 99.59% para los sensores de vibración en el caso de discriminar entre 8 estados de la bomba centrífuga. En el caso de discriminar entre 17 estados se obtienen tasas de éxito del 87.48% y 94.23% para los sensores de audio y del 86.71% y Avances en monitorización preventiva de maquinaria a través de señales de audio y vibración Universidad de Las Palmas de Gran Canaria 325 87.30% para los sensores de vibración. De los resultados obtenidos, se puede concluir que el audio puede ser usado como fuente de información en el diagnóstico de fallos de la bomba centrífuga utilizada. De todas formas, es necesario realizar más investigación al respecto con diferentes bombas centrífugas. En la Tesis también se ha realizado un estudio de la combinación de las señales de audio y vibración obtenidas en la bomba centrífuga para determinar si la fusión de la información de ambas señales mejora la tasa de acierto. Las tasas de éxito se incrementaron significativamente a la hora de fusionar un sensor de audio y otro de vibración en el caso de discriminar entre 17 estados de la bomba centrífuga. En este caso se logró una tasa de acierto del 99.55%, bastante superior a los 94.93% obtenidos por uno de los sensores de audio. En el caso de discriminar entre 8 condiciones de la bomba centrífuga, se llegaron a obtener tasas del 100% a la hora de fusionar un sensor de audio y otro de vibración. De los resultados del estudio de la fusión se puede concluir que si bien los resultados individuales por sensor son buenos, la fusión aumenta la tasa de acierto. Este aumento es especialmente evidente en el caso del diagnóstico de 17 condiciones de la bomba centrífuga. Tesis Doctoral Universidad de Las Palmas de Gran Canaria 326 7 Referencias [1] Vilela, R.M., Metrôlho, J. C., & Cardoso, J. C. (2004, May). Machine and insdustrial monitorization system by analysis of acoustic signatures. In Electrotechnical Conference, 2004. MELECON 2004. Proceedings of the 12th IEEE Mediterranean (Vol. 1, pp. 277-279). IEEE. [2] Jardine, A. K., Lin, D., & Banjevic, D. (2006). A review on machinery diagnostics and prognostics implementing condition-based maintenance. Mechanical systems and signal processing, 20(7), 1483-1510. [3] Vachtsevanos, G., Lewis, F. L., Roemer, M., Hess, A. & Wu, B. Intelligent Fault Diagnosis and Prognosis for Engineering Systems, 1st ed. Hoboken, New Jersey: John Wiley & Sons, Inc., 2006. [4] Paltridge, B. (2002). Thesis and dissertation writing: an examination of published advice and actual practice. English for Scientific Purposes, 21(2), 125–143. [5] Baydar, N., & Ball, A. (2003). Detection of gear failures via vibration and acoustic signals using wavelet transform. Mechanical Systems and Signal Processing, 17(4), 787-804. [6] Bell, R.N., McWilliams, D.W., O’Donnell, P., Singh, C., & Wells, S. J. (1985). Report of large motor reliability survey of industrial and commercial installations Part I and II. IEEE Transactions on Industry Applications, 21(4), 853-872. [7] Immovilli, F., Bellini, A., Rubini, R., & Tassoni, C. (2010). Diagnosis of bearing faults in induction machines by vibration or current signals: A critical comparison. IEEE Transactions on Industry Applications, 46(4), 1350-1359. [8] Samanta, B., Al-Balushi, K. R., & Al-Araimi, S. A. (2006). Artificial neural networks and genetic algorithm for bearing fault detection. Soft Computing, 10(3), 264-271. [9] McFadden, P. D., & Smith, J. D. (1984). Vibration monitoring of rolling element bearings by the high-frequency resonance technique—a review. Tribology international, 17(1), 3-10. [10] Choi, Y. C., & Kim, Y. H. (2007). Fault detection in a ball bearing system using minimum variance cepstrum. Measurement Science and Technology, 18(5), 1433. [11] Zhang, Y., & Randall, R. B. (2009). Rolling element bearing fault diagnosis based on the combination of genetic algorithms and fast kurtogram.Mechanical Systems and Signal Processing, 23(5), 1509-1517. [12] Antoni, J., Randall, R. B. (2006). The spectral kurtosis: application to the vibratory surveillance and diagnostics of rotating machines. Mechanical Syst. and Signal Process., 20, 308-331. [13] Wang, D., Peter, W. T., & Tsui, K. L. (2013). An enhanced Kurtogram method for fault diagnosis of rolling element bearings. Mechanical Systems and Signal Processing, 35(1), 176-199. [14] Peng, Z. K., Peter, W. T., & Chu, F. L. (2005). A comparison study of improved Hilbert–Huang transform and wavelet transform: application to fault diagnosis for rolling bearing. Mechanical systems and signal processing, 19(5), 974-988. [15] Ali, J. B., Fnaiech, N., Saidi, L., Chebel-Morello, B., & Fnaiech, F. (2015). Application of empirical mode decomposition and artificial neural network for automatic bearing fault diagnosis based on vibration signals. Applied Acoustics, 89, 16-27. Avances en monitorización preventiva de maquinaria a través de señales de audio y vibración Universidad de Las Palmas de Gran Canaria 327 [16] Yang, D. M., Stronach, A. F., MacConnell, P., & Penman, J. (2002). Third-order spectral techniques for the diagnosis of motor bearing condition using artificial neural networks. Mechanical systems and signal processing, 16(2), 391-411. [17] Wang, W. J., Wu, Z. T., & Chen, J. (2001). Fault identification in rotating machinery using the correlation dimension and bispectra. Nonlinear Dynamics, 25(4), 383-393. [18] Shibata, K., Takahashi, A., & Shirai, T. (2000). Fault diagnosis of rotating machinery through visualisation of sound signals. Mechanical Systems and Signal Processing, 14(2), 229-241. [19] Cudina, M. (2003). Detection of cavitation phenomenon in a centrifugal pump using audible sound. Mechanical Systems and Signal Processing 17(6), 1335–1347. [20] Singh, V., Meena, N. (2009, January). Engine Fault Diagnosis using DTW, MFCC and FFT. In Proc. of the 1st Int. Conf. on Intelligent Human Computer Interaction, (pp. 83-94). [21] Mansoor, A. B., Ahmed, H., Mahmood, Z. (2009, February). Design and Development of an Automated Acoustic based Jet Engine Performance Evaluator. In Proc. of the IEEE Int. Conf. on Industrial Technol. (pp. 1-4). [22] Wu, J. -D., Chuang, C. -Q. (2005). Fault diagnosis of internal combustion engines using visual dot patterns of acoustic and vibration signals. NDT&E Int., 38, 605-614. [23] Al Thobiani, A., Gu, F., & Ball, A. (2010 June 22-24). The monitoring of cavitation in centrifugal pumps based on the analysis of vibro-acoustic measurements. In CM 2010 and MFPT 2010, UK. [24] Al-Hashmi, S.A. (2012, September 23-26). Spectrum Analysis of Acoustic Signals for Cavitation Detection. In 2012 IEEE Symposium on Industrial Electronics and Applications (ISIEA2012), Bandung, Indonesia. [25] Saeid Farokhzad, H.A., (2013). Acoustic Based Cavitation Detection of Centrifugal Pump by Neural Network. Journal of Mechanical Engineering and Technology, 1(1), 1-5. [26] Al-Hashmi, S.A. (2013, October). Statistical analysis of acoustic signal for cavitation detection. International Journal of Emerging Technology and Advanced Engineering, 3(4). [27] Benko, U., Petrovcic, J., Juricic, D., Tavcar, J., Rejec, J. (2005). An approach to fault diagnosis of vacuum cleaner motors based on sound analysis. Mechanical Syst. and Signal Process., 19, 427-445. [28] Janjarasjitt, S., Ocak, H., Loparo, K. A. (2008). Bearing condition diagnosis and prognosis using applied nonlinear dynamical analysis of machine vibration signal. J. of Sound and Vibration, 317, 112-126. [29] Devi, S., Siva, L., Shanker N. R., & Prabakaran, K. (2010). A Comparative Study between Vibration and Acoustic Signals in HTC Cooling Pump and Chilling Pump. Int. J. of Eng. and Technol., 2. [30] Case Western Bearing Data Center, http://csegroups.case.edu/bearingdatacenter/home. Última visita: Noviembre 2014. [31] http://qsun.eng.ua.edu/cpw_web_new/intro.htm. Última visita: Enero 2014 [32] Lee, J., Qiu, H., Yu, G., Lin, J., & Rexnord Technical Services (2007). Bearing Data Set. IMS, University of Cincinnati. NASA Ames Prognostics Data Repository, [http://ti.arc.nasa.gov/project/prognostic-data-repository], NASA Ames, Moffett Field, CA." Tesis Doctoral Universidad de Las Palmas de Gran Canaria 328 [33] Webpage Explore the World of Piping http://www.wermac.org/equipment/pumps_general.html. Last visited: 19 November 2014 [34] "Pump Principles Manual: global training education and development program" eds. Chesterton. [35] Zhao, X., Zuo, M. J., & Patel, T. H. (2012). Generating an indicator for pump impeller damage using half and full spectra, fuzzy preference-based rough sets and PCA. Measurement Science and Technology, 23(4), 045607. [36] Zhao, X. M., Hu, Q. H., Lei, Y. G., & Zuo, M. J. (2010). Vibration-based fault diagnosis of slurry pump impellers using neighbourhood rough set models.Proceedings of the Institution of Mechanical Engineers, Part C: Journal of Mechanical Engineering Science, 224(4), 995-1006. [37] Samanta, B., Al-Balushi, K. R., & Al-Araimi, S. A. (2006). Artificial neural networks and genetic algorithm for bearing fault detection. Soft Computing, 10(3), 264-271. [38] Randall, R. B. (2004). State of the art in monitoring rotating machinery-part 1. Sound and vibration, 38(3), 14-21. [39] McFadden, P. D., & Smith, J. D. (1984). Vibration monitoring of rolling element bearings by the high-frequency resonance technique—a review. Tribology international, 17(1), 3-10. [40] Kaiser, JF., On a simple algorithm to calculate the energy of a signal. In: International conference on acoustics, speech, and signal processing; 1990. p. 381– 384. [41] Li, H., Fu, L., & Zhang, Y. (2009, April). Bearing faults diagnosis based on teager energy operator demodulation technique. In IEEE International Conference on Measuring Technology and Mechatronics Automation (ICMTMA'09). (Vol. 1, pp. 594-597). [42] Al-Ghamd, A. M., & Mba, D. (2006). A comparative experimental study on the use of acoustic emission and vibration analysis for bearing defect identification and estimation of defect size. Mechanical systems and signal processing, 20(7), 15371571. [43] Yan, R., & Gao, R. X. (2004). Complexity as a measure for machine health evaluation. Instrumentation and Measurement, IEEE Transactions on, 53(4), 13271334. [44] Pudil, P., Novovičová, J., & Kittler, J. (1994). Floating search methods in feature selection. Pattern Recognition Letters, 15(11), 1119-1125. [45] Kuncheva, L. I. (2004). Combining pattern classifiers: methods and algorithms. John Wiley & Sons. [46] Henríquez, P., Alonso, J. B., Ferrer, M. A., & Travieso, C. M. (2014). Review of automatic fault diagnosis systems using audio and vibration signals. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 44(5), 642-652. [47] Henríquez, P., Alonso, J. B., Ferrer, M. A., & Travieso, C. M. (2013). Application of the Teager–Kaiser energy operator in bearing fault diagnosis. ISA transactions, 52(2), 278-284. [48] Henríquez, P., Alonso, J. B., Ferrer, M. A., Travieso, C. M. (2011, May-June). Application of Higher Order Statistics of Teager-Kaiser Energy Transformed Vibration Signal for Bearing Fault Diagnosis. In Proc. of the 24th Int. Congress on Condition Monitoring and Diagnostics Engineering Management (pp. 265-274), Stavanger, Norway. Avances en monitorización preventiva de maquinaria a través de señales de audio y vibración Universidad de Las Palmas de Gran Canaria 329 [49] Henríquez, P., White, P., Alonso, J. B., Ferrer. M. A. (2011, October). Application of Teager-Kaiser Energy Operator to the Analysis of Degradation of a Helicopter Input Pinion Bearing. In Proc. of the International Conference Surveillance 6 (pp. 265-274), University of Technology of Compiègne, France. [50] Henríquez, P., Alonso, J. B., Ferrer, M. A., Travieso, C. M. (2014, September). Degradation assessment in bearings with outer race fault. In 27th International Congress of Condition Monitoring and Diagnostic Engineering Management, Brisbane, Australia. [51] Henríquez, P.; Alonso, J. B.; Ferrer, M. A.; Travieso, C. M.; Gómez, G. (2012). Fault diagnosis using audio and vibration signals in a circulating pump. Journal of Physics: Conference Series 25th Int. Congress on Condition Monitoring and Diagnostic Engineering, 364, 012135. [52] Henríquez, P.; Alonso, J. B.; Ferrer, M. A.; Travieso, C. M. (2014). Application of complexity measures to pump fault diagnosis. In 27th Int. Congress on Condition Monitoring and Diagnostic Engineering, Brisbane, Australia. [53] Henríquez, P., Alonso, J. B., Ferrer, M., Travieso, C. M., Godino-Llorente, J., & Díaz-de-María, F. (2009). Characterization of healthy and pathological voice through measures based on nonlinear dynamics. IEEE Transactions on Audio, Speech, and Language Processing, 17(6), 1186-1195. [54] Henríquez, P., Alonso, J. B., Ferrer, M. A., Travieso, C. M., & OrozcoArroyave, J. R. (2013). Global Selection of Features for Nonlinear Dynamics Characterization of Emotional Speech. Cognitive Computation, 5(4), 517-525. [55] Henríquez, P., Alonso, J. B., Ferrer, M. A., Travieso, C. M., Orozo-Arroyave, J. R. (2014). Nonlinear Dynamics Characterization of Emotional Speech. Neurocomputing, 132, 126-135.