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
A ances en Moni o ización P e en i a de Maquina ia
a á es de Señales de Audio y Vib ación
Tesis Doc o al p esen ada po Pa icia Hen íquez Rod íguez
Di igida po el D . D. Miguel Ángel Fe e Balles e
Codi igida po el D . D. Jesús Be na dino Alonso He nández
El Di ec o El Codi ec o La Doc o anda,
Las Palmas de G an Cana ia, a 22 de Oc ub e de 2015

A mi amilia.
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion Signals
Uni e sidad de Las Palmas de G an Cana ia
ix
ACKNOWLEDGEMENTS
This Thesis summa izes he wo k I ha e ca ied ou du ing my Ph.D. s udies wi h he
Di isión de P ocesado Digi al de la Señal (DPDS) -Digi al Signal P ocessing Di ision-
o he Ins i u o pa a el Desa ollo Tecnológico y la Inno ación en Comunicaciones
(IDeTIC)-Ins i u e o Technological De elopmen and he Inno a ion in
Communica ions- since 2009. This esea ch g oup is wi h he Depa men o Señales y
Comunicaciones o he Uni e sidad de Las Palmas de G an Cana ia. IDeTIC s a ed in
2010 om he CeTIC (Technological Cen e o he Inno a ion in Communica ions).
The CeTIC was ounded in 2006 om he h ee di e en Resea ch G oups, including
he Digi al Signal P ocessing G oup wi h was ounded in 1990. This Thesis has been
mainly suppo ed by a Ph.D. schola ship g an ed o he au ho by Gobie no de Cana ias
and Social Eu opean Fund (ESF), which co e ed he pe iod be ween Ma ch 2009 and
Ma ch 2013, and also by he Spanish go e nmen MICNN TEC2009-14123-C04 and
TEC2012-38630-C04-02 esea ch p ojec s.
Fo emos , I would like o hank my supe iso s P o . Jesús B. Alonso
He nández and P o . Miguel A. Fe e Balles e o hei guidance and suppo o e
he pas six yea s. Du ing his pe iod I ha e bene i ed om hei wise ad ices,
in elligen e o s and cou age which ha e shaped my hinking and wo king a i ude.
They also we e he esponsible o my i s s eps in esea ch wo ld when I i s ook a
hei doo o pu suing my MsC. P ojec abou la yngeal pa hology de ec ion h ough
oice signal. They in oduced me o he signal p ocessing and pa e n ecogni ion wo ld.
In he amewo k o DPDS I ha e also ecei ed suppo om P o . Ca los M. T a ieso
González.
A e inishing my MsC. P ojec , hey ga e me he g ea oppo uni y o wo king
in a na ional p ojec o homeland secu i y (Hespe ia) ounded by he o ice o he
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
x
indus ial echnological de elopmen (CDTI) (a ached o he Minis iy o Indus y and
T ade) o he Spanish go e nmen in he ame o CENIT p ojec s. In he amewo k o
his wo k, I lea ned abou condi ion moni o ing in indus ial scena ios and I could
bene i om he knownledge o many expe s in indus ial secu i y in ol ed in he
p ojec . Thanks o his wo k, I decided o ocus my PhD on condi ion moni o ing
implemen ing aul diagnosis.
Du ing my Ph. D. s udies I had he g ea oppo uni y o isi ing wo o eign
ins i u ions. The i s 4-mon h s ay was in 2011 a he Ins i u e o Sound and Vib a ion
Resea ch (ISVR) o he Uni e si y o Sou hamp on (Sou hamp on, England) wi h P o .
Paul R. Whi e. Du ing hose ou mon hs I could bene i om his mas e y in he ield o
signal p ocessing. I also had he o une o mee P o . Ling Wang om who I could
bene i om he expe ise in pump aul diagnosis.
The second 3-mon h s ay was in 2012 a he Reliabili y Resea ch Lab o he
Uni e si y o Albe a (Edmon on, Canada) wi h P o . Ming J. Zuo. I ha e me a
an as ic g oup o people he e. I would like o specially hank D . Funda Akde e
Iscioglu o he ca e du ing my isi . I ha e also would like o hank Mayank Pandey o
ou discussions abou he pump ins alla ion and condi ion moni o ing echniques.
F om such esea ch s ays he e is a huge numbe o colleagues and iends who I
wan o hank: Albe o Baldelli, Pie e Rosado, Vianey Lande os, Ki an Konde, Rami
Saba, Dao Duc Cuong, Luis Espi ia and his wi e Jessica Espi ia.
I wan o hank o my MsC. s uden Ga oé Gómez o his aluable help wi h he
pump ins alla ion and da a acquisi ion p ocess.
I also wan o hank all he wo k ma es a IDeTIC who wi h I ha e sha ed lunch
and discussions: Hima , Ja ie , Lau a, Jaime, Daniel, Ped o, Ayaya, Manolo, Víc o and
all IDeTIC human g oup. Thank you guys!
Quie o ambién da le las g acias a mis amigos de Las Palmas, los cuales han
aguan ado du an e odos es os años mis dudas, mis ausencias, mis ne ios y con los
cuales he compa ido mis log os y mis aleg ías. Muchas g acias a
Eli, Paco, Raquel, Da inia, Rosal a, Sil ia, Gema, Mónica y Oc a io. Y en especial
g acias a B iseida, po es a siemp e a mi lado.
G acias a mi amilia.
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia
x ii
LIST OF FIGURES
Figu e 1-1: S ages in a condi ion-based mon io ing sys em. ......................................... 32
Figu e 1-2: Dependence among Disse a ion chap e s. .................................................. 41
Figu e 2-1: S ages in a condi ion-based mon io ing sys em .......................................... 47
Figu e 2-2: Dis ibu ions o e e enced pape s: Audio-based and ib a ion-based aul
diagnosis dis ibu ion (uppe le ). Publishing yea s dis ibu ion (uppe
igh ).Dis ibu ion o aul s in ib a ion-based diagnosis (bo om le ) and in audio-
based diagnosis (bo om igh ). .............................................................................. 49
Figu e 3-1: Componen s o a ball bea ing ...................................................................... 76
Figu e 3-2: Bea ing ib a ion signals o bea ings wi h disc e e aul s a ou e ace, inne
ace and ball. BPFO = ball pass equency, ou e ace, BPFI = ball pass equency,
inne ace, BSF = ball spin equency, FTF: undamen al ain equency (cage
equency). Sou ce: [26]. ........................................................................................ 78
Figu e 3-3: Time bea ing ib a ion signal o 0.17 seconds (sample equency is 12kHz)
o di e en bea ing condi ions: No mal condi ion (uppe -le ), Inne ace aul
condi ion (uppe - igh ), Ou e ace aul condi ion (bo om-le ) and Ball aul
condi ion (bo om- igh ). ........................................................................................ 79
Figu e 3-4: Tes s and o he acquisi ion o bea ing ib a ion da a. Sou ce [1]. ........... 81
Figu e 3-5: Main ansmission o an UH-60 Backhawk helicop e . Sou ce: [27]. ......... 82
Figu e 3-6: Posi ion o he bea ing SB-2205 in he ansmission (le igu e) and he nal
bea ing condi ion ( igh igu e). Sou ce: [28]. ...................................................... 83
Figu e 3-7: Posi ion o he bea ings in he IMS bea ing da abase. Sou ce: [3]. ............. 84
Figu e 3-8: Le : Fluid di ec ion in a cen i ugal pump. Sou ce: [6]. Righ : eloci y and
p essu e in a cen i ugal pump. Modi ied om sou ce: [5]. The inle , ou le , olu e
and impelle a e shown. .......................................................................................... 85
Figu e 3-9: Closed impelle . Modi ied om sou ce: [6]. ............................................... 86
Figu e 3-10: Diag am wi h he componen s o he acquisi ion sys em .......................... 91
Figu e 3-11: Diag am o he ALP800 pump .................................................................. 92
Figu e 3-12: Two pho os o he ALP 800 pump. ........................................................... 94
Figu e 3-13: Pho os o he pump ALP 800. Impelle moun ed on he o o (le ). Volu e,
inle and oule pump ( igh ). ................................................................................... 94

PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
x iii
Figu e 3-14: Pho os o he pump ALP 800. Impelle moun ed on he o o (le ).
Di e en iews o he impelle (cen e and igh ). ................................................. 95
Figu e 3-15: Pho os o he pla e aul in impelle . Uppe le : impelle in no mal
condi ion. Uppe igh : one pa o he pla e emo ed. Lowe le : wo pa s o he
pla e a e emo ed. Lowe igh : he hi d pa o he impelle was emo ed. ........ 96
Figu e 3-16: Pho o o he impelle wi h leading edge aul . In his case, he sligh aul
(5mm om all ane leading edges we e emo ed) is shown. ................................ 97
Figu e 3-17: Pho o o he impelle wi h ailing edge aul . In his case, he sligh aul
(5mm om all ane ailing edges we e emo ed) is shown. ................................ 98
Figu e 3-18: Pho o o he o- ing o he impelle o he AL P800 pump......................... 99
Figu e 3-19: Pho o o he ese oi wi h he sand added o simula e s ange objec s o
pa icles in he sys em. ......................................................................................... 100
Figu e 3-20: Kind o PVC balls (6 mm o diame e ) used o simula e s ange objec s in
he sys em [29]. .................................................................................................... 100
Figu e 3-21: D awing o he oom o he expe imen al se -up wi h he measu emen s in
mm. A squema ic o he se -up is also shown. ..................................................... 101
Figu e 3-22: Senso Placemen o he Expe imen al se -up. ....................................... 102
Figu e 3-23: Spec um in ange [2.7-1000]Hz o a 8192 poin s ib a ion ame
(371.5ms) om senso RadialInle Accel. The highe peaks a e ma ked wi h ed
ci cles and he equency alue is shown. ............................................................ 106
Figu e 3-24: Spec um in ange [1000-11025]Hz o a 8192 poin s ib a ion ame
(371.5ms) om senso RadialInle Accel. The highe peaks a e ma ked wi h ed
ci cles and he equency alue is shown. ............................................................ 106
Figu e 3-25: Spec um in ange [2.7-1000]Hz o a 8192 poin s audio ame (371.5ms)
om senso Mic oInle . The highe peaks a e ma ked wi h ed ci cles and he
equency alue is shown. .................................................................................... 107
Figu e 3-26: Spec um in ange [1000-11025]Hz o a 8192 poin s audio ame
(371.5ms) om senso Mic oInle . The highe peaks a e ma ked wi h ed ci cles
and he equency alue is shown. ........................................................................ 107
Figu e 4-1: Time Vib a ion signals o bea ings wi h ou e ace aul , inne ace aul
and ball aul . Sou ce: [28]. .................................................................................. 114
Figu e 4-2: Teage -Kaise ans o med signals (TK signals) o di e en bea ing
condi ions: no mal condi ion (uppe -le ), inne ace aul condi ion (uppe - igh ),
ou e ace aul condi ion (bo om-le ) and ball condi ion (bo om- igh ). ......... 118
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia
xix
Figu e 4-3: Diag am o he expe imen a ion in bea ing aul diagnosis. The p oposal is
compa ed wi h o he me hods in he s a e o he a and an e alua ion wi h wo
classi ie s is ca ied ou . ....................................................................................... 119
Figu e 4-4: Diag am o he expe imen a ion in bea ing aul e olu ion. The p oposal is
applied o a un- o- ailu e bea ing ib a ion da abase. ......................................... 119
Figu e 4-5: Success a es using TK, T, AM and TK-AM ea u es in o de o ele ance
wi h he neu al ne wo k classi ie . ........................................................................ 126
Figu e 4-6: Success a es using TK, T, AM and TK-AM ea u es in o de o ele ance
wi h he LS-SVM classi ie . ................................................................................. 126
Figu e 4-7: E olu ion o he ea u es ex ac ed om he TK signal. ........................... 128
Figu e 4-8: E olu ion o he ea u es ex ac ed om he aw ib a ion signal (T signal).
.............................................................................................................................. 129
Figu e 4-9: E olu ion o he ea u es ex ac ed om he en elope signal o he Q
band[2500-3800Hz]. ............................................................................................. 129
Figu e 4-10: P oposed me hodology o bea ing se e i y aul assessmen . ................ 136
Figu e 4-11: Wa e o m o he Daubechies 6 wa ele mo he . .................................... 136
Figu e 4-12: Rela i e ene gies o WPT nodes o le el 3 o decomposi ion (o de ed by
equency con en ) o a ame o 4096 samples ex ac ed o no mal ib a ion
signals wi h di e en loads. .................................................................................. 139
Figu e 4-13: Rela i e ene gies o WPT nodes o le el 3 o decomposi ion (o de ed by
equency con en ) o a ame o 4096 samples ex ac ed o ib a ion signals wi h
inne ace aul (le column), ou e ace aul (cen e column) and ball aul ( igh
column) wi h di e en se e i ies and o load 0. ................................................. 139
Figu e 4-14: Rela i e ene gies o WPT nodes o le el 3 o decomposi ion (o de ed by
equency con en ) o a ame o 4096 samples ex ac ed o ib a ion signals wi h
inne ace aul (le column), ou e ace aul (cen e column) and ball aul ( igh
column) wi h di e en se e i ies and o load 1. ................................................. 140
Figu e 4-15: Rela i e ene gies o WPT nodes o le el 3 o decomposi ion (o de ed by
equency con en ) o a ame o 4096 samples ex ac ed o ib a ion signals wi h
inne ace aul (le column), ou e ace aul (cen e column) and ball aul ( igh
column) wi h di e en se e i ies and o load 2. ................................................. 140
Figu e 4-16: Rela i e ene gies o WPT nodes o le el 3 o decomposi ion (o de ed by
equency con en ) o a ame o 4096 samples ex ac ed o ib a ion signals wi h
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
xx
inne ace aul (le column), ou e ace aul (cen e column) and ball aul ( igh
column) wi h di e en se e i ies and o load 3. ................................................. 141
Figu e 4-17: E olu ion o he mean alues ob ained o each me hod: p oposed me hod
‘LZC (EmaxWPT)’ (uppe-le ), ‘LZC ( aw signal)’ (uppe - igh ), ‘Ku osis
(EmaxWPT)’ (bo om-le ) and ‘Ku osis ( aw signal)’ (bo om- igh ) o ou e
ace aul wi h di e en se e i ies and di e en loads. ........................................ 143
Figu e 4-18: E olu ion o he mean alues ob ained o each me hod: p oposed me hod
‘LZC (EmaxWPT)’ (uppe-le ), ‘LZC ( aw signal)’ (uppe - igh ), ‘Ku osis
(EmaxWPT)’ (bo om-le ) and ‘Ku osis ( aw signal)’ (bo om- igh ) o inne
ace aul wi h di e en se e i ies and di e en loads. ........................................ 144
Figu e 4-19: E olu ion o he mean alues ob ained o each me hod: p oposed me hod
‘LZC (EmaxWPT)’ (uppe-le ), ‘LZC ( aw signal)’ (uppe - igh ), ‘Ku osis
(EmaxWPT)’ (bo om-le ) and ‘Ku osis ( aw signal)’ (bo om- igh ) o ou e
ace aul wi h di e en se e i ies and di e en loads. The no mal condi ion is also
conside ed. ............................................................................................................ 144
Figu e 4-20: E olu ion o he mean alues ob ained o each me hod: p oposed me hod
‘LZC (EmaxWPT)’ (uppe-le ), ‘LZC ( aw signal)’ (uppe - igh ), ‘Ku osis
(EmaxWPT)’ (bo om-le ) and ‘Ku osis ( aw signal)’ (bo om- igh ) o inne
ace aul wi h di e en se e i ies and di e en loads. The no mal condi ion is also
conside ed. ............................................................................................................ 145
Figu e 4-21: E olu ion o he mean alues ob ained o ‘Ku osis ( aw signal)’ me hod
applied o ou e ace aul condi ion and no mal condi ion a ying he amoun o
noise added o he o iginal ib a ion signal. The esul s a e shown o load: load 0
(uppe -le ), load 1(uppe - igh ), load 2 (bo om-le ), load 3 (bo om- igh ). ..... 146
Figu e 4-22: E olu ion o he mean alues ob ained o ‘LZC ( aw signal)’ me hod
applied o ou e ace aul condi ion and no mal condi ion a ying he amoun o
noise added o he o iginal ib a ion signal. The esul s a e shown o load: load 0
(uppe -le ), load 1(uppe - igh ), load 2 (bo om-le ) and load 3 (bo om- igh ). 146
Figu e 4-23: E olu ion o he mean alues ob ained o ‘Ku osis (EmaxWPT)’ me hod
applied o ou e ace aul condi ion and no mal condi ion a ying he amoun o
noise added o he o iginal ib a ion signal. The esul s a e shown o load: load 0
(uppe -le ), load 1(uppe - igh ), load 2 (bo om-le ) and load 3 (bo om- igh ). 147
Figu e 4-24: E olu ion o he mean alues ob ained o he p oposed me hod ‘LZC
(EmaxWPT)’ applied o ou e ace aul condi ion and no mal condi ion a ying
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia
xxi
he amoun o noise added o he o iginal ib a ion signal. The esul s a e shown
o load: load 0 (uppe -le ), load 1(uppe - igh ), load 2 (bo om-le ) and load 3
(bo om- igh ). ...................................................................................................... 147
Figu e 4-25: E olu ion o he mean alues ob ained o ‘Ku osis ( aw signal)’ me hod
applied o inne ace aul condi ion and no mal condi ion a ying he amoun o
noise added o he o iginal ib a ion signal. The esul s a e shown o load: load 0
(uppe -le ), load 1(uppe - igh ), load 2 (bo om-le ) and load 3 (bo om- igh ). 148
Figu e 4-26: E olu ion o he mean alues ob ained o ‘LZC ( aw signal)’ me hod
applied o inne ace aul condi ion and no mal condi ion a ying he amoun o
noise added o he o iginal ib a ion signal. The esul s a e shown o load: load 0
(uppe -le ), load 1(uppe - igh ), load 2 (bo om-le ) and load 3 (bo om- igh ). 148
Figu e 4-27: E olu ion o he mean alues ob ained o ‘Ku osis (EmaxWPT)’ me hod
applied o inne ace aul condi ion and no mal condi ion a ying he amoun o
noise added o he o iginal ib a ion signal. The esul s a e shown o load: load 0
(uppe -le ), load 1(uppe - igh ), load 2 (bo om-le ) and load 3 (bo om- igh ). 149
Figu e 4-28: E olu ion o he mean alues ob ained o he p oposed me hod ‘LZC
(EmaxWPT)’ applied o inne ace aul condi ion and no mal condi ion a ying
he amoun o noise added o he o iginal ib a ion signal. The esul s a e shown
o load: load 0 (uppe -le ), load 1(uppe - igh ), load 2 (bo om-le ) and load 3
(bo om- igh ). ...................................................................................................... 149
Figu e 4-29: E olu ion o he ea u es using ku osis and Lempel-Zi complexi y om
he aw signal and ku osis and Lempel-Zi complexi y om he node wi h
maximal ene gy o he wa ele pake ans o m. ................................................. 150
Figu e 4-30: E olu ion o he Ku osis ex ac ed om he aw ib a ion signal o a un-
o- ailu e expe imen o he bea ing ha e en ually de eloped an ou e - ace aul .
.............................................................................................................................. 152
Figu e 4-31: E olu ion o he Ku osis ex ac ed om he node o maximal ene gy o
he WPT o a un- o- ailu e expe imen o he bea ing ha e en ually de eloped an
ou e - ace aul . ..................................................................................................... 152
Figu e 4-32: E olu ion o he Lempel-Zi complexi y ex ac ed om he aw ib a ion
signal o a un- o- ailu e expe imen o he bea ing ha e en ually de eloped an
ou e - ace aul . ..................................................................................................... 153
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
xxii
Figu e 4-33: E olu ion o he Lempel-Zi complexi y ex ac ed om he node o
maximal ene gy o he WPT o a un- o- ailu e expe imen o he bea ing ha
e en ually de eloped an ou e - ace aul . ............................................................. 153
Figu e 5-1: Squeme o he me hodology used o quan i y he abili y o he ea u es o
ib a ion and audio based pump aul diagnosis. ................................................. 161
Figu e 5-2: Time ames (8192 samples) om senso Radial Inle Accel ( ib a ion) and
Inle Mic o (audio) o no mal (NOR) and pla e (PLA) condi ion....................... 178
Figu e 5-3: Time ames (8192 samples) om senso Radial Inle Accel ( ib a ion) and
Inle Mic o (audio) o leading edge damage (LED) and ailing edge damage
(TED) condi ion. ................................................................................................... 178
Figu e 5-4: Time ames (8192 samples) om senso Radial Inle Accel ( ib a ion) and
Inle Mic o (audio) o seal (SEA) and sand (SAN) condi ion............................. 179
Figu e 5-5: Time ames (8192 samples) om senso Radial Inle Accel ( ib a ion) and
Inle Mic o (audio) o sand and pape (SAP) and p c balls (PVC) condi ion. ... 179
Figu e 5-6: Spec a o ib a ion ames (8192 samples) om senso Radial Inle Accel
o di e en pump condi ions: no mal (NOR), pla e (PLA), leading edge damage
(LED), ailing edge damage (TED), seal (SEA), sand (SAN), sand and pape
(SAP) and p c balls (PVC) condi ions. ................................................................ 187
Figu e 5-7: Spec a o audio ames (8192 samples) om senso Inle Mic o o
di e en pump condi ions: no mal (NOR), pla e (PLA), leading edge damage
(LED), ailing edge damage (TED), seal (SEA), sand (SAN), sand and pape
(SAP) and p c balls (PVC) condi ions. ................................................................ 188
Figu e 5-8: Spec a o ib a ion ames (8192 samples) om senso Radial Inle Accel
in equency ange [0-1000]Hz o di e en pump condi ions: no mal (NOR), pla e
(PLA), leading edge damage (LED), ailing edge damage (TED), seal (SEA), sand
(SAN), sand and pape (SAP) and p c balls (PVC) condi ions. .......................... 189
Figu e 5-9: Spec a o audio ames (8192 samples) om senso Inle Mic o in
equency ange [0-1000]Hz o di e en pump condi ions: no mal (NOR), pla e
(PLA), leading edge damage (LED), ailing edge damage (TED), seal (SEA), sand
(SAN), sand and pape (SAP) and p c balls (PVC) condi ions. .......................... 189
Figu e 5-10: Rec i ied ceps um o an audio ame om senso Inle Mic o in no mal
condi ion. .............................................................................................................. 193
Figu e 5-11: Rec i ied ceps um o an audio ame om senso Inle Mic o in LED
condi ion. .............................................................................................................. 193

Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia
xxiii
Figu e 5-12: Rec i ied ceps um o a ib a ion ame om senso Radial Inle Accel in
no mal condi ion. .................................................................................................. 194
Figu e 5-13: Rec i ied ceps um o a ib a ion ame om senso Radial Inle Accel in
LED condi ion. ..................................................................................................... 194
Figu e 5-14: Sequen ial e alua ion o he ea u es o each senso in pump aul
diagnosis using a neu al ne wo k classi ie o disc imina e be ween 8 pump
condi ions. ............................................................................................................ 217
Figu e 5-15: Sequen ial e alua ion o he ea u es o each senso in pump aul
diagnosis using a LSSVM classi ie o disc imina e be ween 8 pump condi ions.
.............................................................................................................................. 217
Figu e 5-16: Sequen ial e alua ion o he ea u es o each senso in pump aul
diagnosis using a neu al ne wo k classi ie o disc imina e be ween 17 pump
condi ions. ............................................................................................................ 218
Figu e 5-17: Sequen ial e alua ion o he ea u es o each senso in pump aul
diagnosis using a LSSVM classi ie o disc imina e be ween 8 pump condi ions.
.............................................................................................................................. 218
Figu e 8-1: Da a dis ibu ion o each kind o oice (H: heal hy oice, P: pa hological
oice, LP: ligh pa hological oice, MP: mode a e pa hological oice, SP: se e e
pa hological oice) o each measu emen ex ac ed om he /a/ owel o he
mul iquali y da abase (FMMI: i s minimum o he mu ual in o ma ion unc ion.
CD: co ela ion dimension. CE: co ela ion en opy. RE1: i s -o de Rényi block
en opy. RE2: second-o de Rényi block en opy. SE: Shannon en opy). Sou ce:
[1]. ........................................................................................................................ 248
Figu e 8-2: Da a dis ibu ion o each kind o oice (H: heal hy oice, P: pa hological
oice) o each measu emen ex ac ed om he MEEI da abase (FMMI: i s
minimum o he mu ual in o ma ion unc ion. CD: co ela ion dimension. CE:
co ela ion en opy. RE1: i s -o de Rényi block en opy. RE2: second-o de
Rényi block en opy. SE: Shannon en opy). Sou ce: [1]. ................................... 249
Figu e 8-3: Da a dis ibu ion o neu al, ea and ange emo ional speech o each
complexi y measu e ex ac ed om he Polish emo ional da abase. MI: alue o he
i s minimum o he mu ual in o ma ion unc ion (uppe le ), SE: Shannon
en opy (uppe igh ), CD: Taken's es ima o o he co ela ion dimension (middle
le ), CE: co ela ion en opy (middle igh ), LZC: Lempel–Zi complexi y
(bo om le ), H: Hu s exponen (bo om igh ) [40]. .......................................... 259
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
xxi
Figu e 8-4: Da a dis ibu ion o neu al, ea and ange emo ional speech o each
complexi y measu e ex ac ed om he Be lin emo ional da abase. MI: alue o he
i s minimum o he mu ual in o ma ion unc ion (uppe le ), SE: Shannon
en opy (uppe igh ), CD: Taken's es ima o o he co ela ion dimension (middle
le ), CE: co ela ion en opy (middle igh ), LZC: Lempel–Zi complexi y
(bo om le ), H: Hu s exponen (bo om igh ) [40]. .......................................... 260
Figu e 8-5: Da a dis ibu ion o neu al, ea and ange emo ional speech o each
complexi y measu e ex ac ed om he LDC emo ional da abase. MI: alue o he
i s minimum o he mu ual in o ma ion unc ion (uppe le ), SE: Shannon
en opy (uppe igh ), CD: Taken's es ima o o he co ela ion dimension (middle
le ), CE: co ela ion en opy (middle igh ), LZC: Lempel–Zi complexi y
(bo om le ), H: Hu s exponen (bo om igh ) [40]. .......................................... 261
Figu a 1: E apas de un sis ema de moni o ización basado en la condición .................. 280
Figu a 2: Esquema de dependencia en e los capí ulos de la memo ia. ....................... 287
Figu a 3: Dis ibución de los a ículos usados pa a la edacción del es ado del a e:
dis ibución de a ículos basados en señales de ib ación y basados en señales de
audio (supe io izquie da); años de publicación de los a ículos (supe io de echa);
dis ibución de elemen os donde se p oducen los allos analizados con señales de
ib ación (in e io izquie da) como con señales de audio (in e io de echa)....... 289
Figu a 4: Componen es de un cojine e cuyo elemen o de odamien o son bolas. ........ 295
Figu a 5: Pa es p incipales de una bomba cen í uga. Di ección del luido en la bomba
( igu a izquie da). Fuen e: [32]. Fluid in: en ada del luido. Fluid ou : salida del
luido. Velocidad y p esión del luido en una bomba. Figu a modi icada de la
uen e: [34]. .......................................................................................................... 296
Figu a 6: Pa es de un ode e. Figu a modi icada de la uen e: [32]. ........................... 297
Figu a 7: Esquemá ico de la sala donde se ealiza la g abación de la base de da os jun o
con la posición y esquema del mon aje. ............................................................... 298
Figu a 8: Fo os de la bomba de agua modelo ALP 800. .............................................. 298
Figu a 9: Fo os de la bomba de agua modelo ALP 800. Rode e mon ado en el o o
(izquie da). Volu a, en ada y salida de la bomba (de echa). ............................... 299
Figu a 10: Fo os del allo en el pla o del ode e. Rode e sin allo (supe io izquie da),
ode e con una pa e del pla o eliminada (supe io de echa), ode e con dos pa es
de pla o eliminadas (in e io izquie da) y ode e con es pa es de pla o aliminadas
(in e io de echa). ................................................................................................. 300
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia
xx
Figu a 11: Fo o del ode e con allo en el leading edge. En la imagen se obse a un
eco e del bo de leading edge de 5mm ( allo le e). ............................................ 301
Figu a 12: Fo o del ode e con allo en el ailing edge. En la imagen se obse a un
eco e del bo de ailing edge de 5mm ( allo le e). ............................................ 301
Figu a 13: Posición de los mic ó onos y de los acele óme os. ................................... 302
Figu a 14: Esquema de la expe imen ación en diagnós ico de allos de cojine es. ...... 305
Figu a 15: Diag ama de la aplicación del mé odo p opues o a la deg adación de un
cojine e. ................................................................................................................ 307
Figu a 16: Mé odo p opues o pa a la iden i icación de allos en cojine es. ................. 308
Figu a 17: Fo ma de onda de la wa ele mad e Daubechies 6. .................................... 309
Figu a 18: Esquema de la me odología lle ada a cabo pa a cuan i ica la habilidad de las
ca ac e ís icas ex aídas de la señal de audio y ib ación en diagnós ico de allos en
bombas cen í ugas. .............................................................................................. 311
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 33
Cen i ugal pumps a e impo an elemen s in se e al sys ems because hey
anspo luids. They a e used in se e al indus ies such as powe s a ions, chemical
indus y, wa e depu a ion sys ems, oil ex ac ion, cooling and hea ing sys ems and e en
in spas and swimming pools. So, he condi ion moni o ing o pumps is also e y
impo an o he ope a ional con inui y.
Vib a ion signals a e widely used as sou ce o in o ma ion in condi ion
moni o ing o di e en machines and componen o machines. Bea ings and cen i ugal
pumps a e no an excep ion. Vib a ion-based aul diagnosis e e s o aul diagnosis
using ib a ion as in o ma ion sou ce. Vib a ion-based aul diagnosis is a well-
es ablished ield ha includes a wide ange o echniques which ha e apidly e ol ed
du ing he las decades. In condi ion moni o ing, ib a ion based aul diagnosis
echniques ha e been widely used due o he easiness o acqui e he ib a ion o he
machine. Tha is why mos esea ch pape s in aul diagnosis li e a u e is de o ed o
iba ion aul diagnosis [1].
Audio-based o ai bo ne-based aul diagnosis e e s o aul diagnosis using
audio as in o ma ion sou ce. Audio-based aul diagnosis using mic ophones in he
audible ange (0-20kHz) is an eme ging ield wi h a g ea po en ial in aul diagnosis
since mic ophones a e non-in asi e senso s and wi h g ea e loca ion possibili ies. Fo
his eason, we hink audio-based echniques need u he esea ch e o s.
In sho , we ocus ou e o s in his Ph.D. Thesis on imp o emen s in bo h
audio-based and ib a ion-based aul diagnosis in condi ion moni o ing in bea ings and
cen i ugal pumps. In each applica ion a ea we ha e mainly ocused on he p ocessing
s age o a condi ion moni o ing squeme (see Figu e 1-1), speci ically in ea u e
ex ac ion. In gene al, ime, equency and ime- equency domains a e equenc ly used
o ea u e ex ac ion. Recen ly, nonlinea echniques such as nonlinea dynamics
applied o ime se ies and complexi y measu es ha e appea ed in aul diagnosis o
ce ain machines. We hink ha he s udy o such ea u es in condi ion moni o ing can
imp o e he pe o mance o a condi ion moni o ing sys em.
The e exis a as li e a u e o di e en me hods o bea ing aul diagnosis and
aul iden i ica ion and in bea ing deg ada ion assessmen . The e a e also some a ailable

PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
34
public da abases o bea ing ib a ion signals in no mal ( ee- aul ) and aul condi ions.
Fo his eason, in bea ing aul diagnosis we ha e used only ib a ion signal om he
public a ailable da abases. We ha e p oposed imp o emen s in nonlinea ea u es
aiming a de ec ing di e en bea ing aul s and also a he iden i ica ion o aul s. So, we
ha e made con ibu ions in he signal p ocessing ( ea u e ex ac ion) s age.
In pump aul diagnosis we ha e eco ded ou own da abase using audio and
ib a ion signals simul enously acqui ed wi h he aim o compa ing and combining
ib a ion and audio signals as sou ce o in o ma ion because he e a e no public
a ailable da abases o cen i ugal pumps. As we s a ed be o e, mos condi ion
moni o ing is ca ied ou using ib a ion signals and pump condi ion moni o ing is no
an excep ion. Mos esea ch in cen i ugal pump condi ion moni o ing is ocused in
ene gy and s a is ical ea u es ex ac ed om ime and equency domain o ib a ion
signals as well as p ession signals. In his Thesis, we explo e he use o audio signals in
pump aul diagnosis, p opose new ea u es and s udy he combina ion o in o ma ion
om audio and ib a ion signals. Speci ically, we ha e p oposed h ee imp o emen s
no p e iously add essed in he li e a u e: i) he applica ion o ea u es o iginally used
wi h ib a ion signals o audio signal, ii) he p oposal o new ea u es ex ac ed om
equency, ceps um, ime- equency domains as well as complexi y ea u es and
ea u es ela ed o nonlinea dynamics applied o ime se ies in ib a ion and audio
pump aul diagnosis and iii) a s udy o he combina ion ea u es ex ac ed om
ib a ion signals and ea u es ex ac ed om audio signals. So, we ha e made
con ibu ions o he acquisi ion s age, only in he way he e is no public audio da abase
o cen i ugal pumps a ailable and o he p ocessing s age ( ea u e ex ac ion) applying
ib a ion ea u es o audio signals, p oposing a se o new ea u es in pump aul
diagnosis and combining in o ma ion om audio and ib a ion signals.
1.4 Mo i a ion o he Thesis
This Thesis is ocused on wo aspec s o condi ion moni o ing: he signal used o
ex ac he in o ma ion abou he condi ion o he machine and he ea u e ex ac ion
wi h signal p ocessing echniques. The aim is aul diagnosis and aul iden i ica ion
( aul deg ada ion). The wo applica ion a eas a e: bea ings and pumps. In bea ings only
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 35
ib a ion signal is used and in cen i ugal pumps audio and ib a ion signals a e used as
sou ce o in o ma ion. The esea ch ca ied ou in his Thesis is mo i a ed by he
ollowing obse a ions om he s a e-o - he a :
.Vib a ion-based moni o ing is a consolida ed ield in CBM due o he easiness
o acqui e he ib a ion o he machine [2]-[4]. In bea ing aul diagnosis, he e a e
mul iple signal p ocessing echniques o ex ac ea u es in o de o disc imina e
be ween di e en bea ing condi ions and o ollow he deg ada ion o a bea ing aul
[5]. The ea ly de ec ion o a aul is e y impo an in condi ion moni o ing o p e en
he aul de elops. Mo eo e , nonlinea ea u es ex ac ed om he machine's signa u e
can e eal new unde s anding o he signal unde conside a ion and he applica ion o
nonlinea ea u es can imp o e he ask o aul diagnosis and aul deg a ion [8]. Fo
his eason, we ha e p oposed imp o emen s in nonlinea ea u es aiming a de ec ing
di e en bea ing aul s ( aul diagnosis) and also a ollowing he deg ada ion o
bea ings a ea ly s ages.
.Audio-based moni o ing (also called ai bo ne-based moni o ing) is a less
de eloped ield han ib a ion-based moni o ing. The main eason is ha audio signal
can be a ec ed by su ounding noise. Howe e , mic ophones ha e mo e possibili ies o
loca ion and a e no moun ed on he machine [6], [7]. Fo his eason, we aim o explo e
audio-based aul diagnosis. In o de o s udy audio signals and o compa e hem wi h
ib a ion signals, audio and ib a ion signals a e acqui ed om an expe imen al se o a
cen i ugal pump.
.The e is a lack o public a ailable da abases o audio signals acqui ed om
machines o audio-based aul diagnosis. Mo eo e , he e a e only ew wo ks ela ed o
audio and ib a ion signals acqui ed oge he . These a e mo e easons o eco ding he
da abase o audio and ib a ion signals acqui ed simoul aneously om an expe imen al
se o a cen i ugal pump.
.As audio signals a e less used in aul diagnosis li e a u e, mos ea u es
ex ac ed om ib a ion signals ha e no p e iously used in audio signals. In his
Thesis, we add ess his issue applying s a e-o - he-a ib a ion ea u es in o audio
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
36
signals in he cen i ugal pump applica ion in o de o s udy hei disc imina ion abili y
be ween di e en condi ion o he cen i ugal pump (no mal and aul condi ions).
.In pump aul diagnosis li e a u e, mos ea u es a e ex ac ed om ime,
equency and ime- equency domain [9], [10]. In his Thesis, a se o new ea u es
ex ac ed om equency, ceps um, ime- equency domains as well as complexi y
ea u es and ea u es based on nonlinea dynamics applied o ime se ies a e ex ac ed
om bo h ib a ion and audio signals. These ea u es along wi h he s a e-o - he-a
ea u es a e e alua ed in o de o s udy hei disc imina ion abili y be ween di e en
pump condi ions using wo classi ie s.
.The e a e ew wo ks in li e a u e using ib a ion and audio signals acqui ed
simouls aneously [11], [12]. Fo his eason, in his Thesis he usion o bo h ib a ion
and audio signals a ea u e le el, sco e le el and decision le el is ca ied ou in o de o
e alua e whe he he usion imp o es he pe o mance o he sys em.
These obse a ions will be discussed in Chap e 2, in which he Thesis p oblem
is analized in dep h.
1.5 The Thesis
The hesis de eloped in his Disse a ion can be s a ed as ollows:
The use o audio signals as sou ce o in o ma ion and he applica ion o nonlinea
echniques imp o es condi ion moni o ing pe o mance.
1.6 Objec i es o he Thesis
This PhD Thesis seeks o imp o e he pe o mance o condi ion moni o ing sys ems in
aul diagnosis and aul iden i ica ion using ib a ion and audio signals in wo
applica ions (bea ings and pumps) wi h special emphasis in he ea u e ex ac ion s age
and in he use o audio signals as sou ce o in o ma ion.
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 37
The main obje i es o he PhD Thesis a e as ollows:
1. To e iew signal p ocessing echniques and pa e n ecogni ion echniques in audio-
based and ib a ion-based aul diagnosis.
2. To ob ain ib a ion bea ing da abases o ee- aul bea ings and bea ings wi h aul s.
3. To s udy nonlinea echniques o bea ing aul diagnosis and bea ing aul
iden i ica ion.
4. To c ea e a da abase o audio and ib a ion signals simul aneously eco ded om a
cen i ugal ci cula ing pump wi h no mal ( ee- aul ) and aul condi ions.
5. To apply ea u es ex ac ed om ib a ion signals in pump aul diagnosis o audio
signals ob ained om a cen i ugal pump.
6. To gene a e new ea u es o disc imina e be ween di e en pump condi ions in pump
aul diagnosis.
7. To compa e ib a ion and audio ea u es pe o mances in he cen i ucal pump
applica ion.
8. To s udy he combina ion o audio and ib a ion signals in he cen i ugal pump
applica ion.
1.7 Me hodology o he Thesis
The me hodology o his Thesis is di ided in he ollowing s eps:
P o ided he e a e no a ailable public audio da abases, we eco d ou own audio
da abase acqui ing a he same ime ib a ion signals. We eco d simul aneously audio
and ib a ion signals om a cen i ugal pump in a closed loop. No mal and di e en
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
38
aul s condi ions we e eco ded. We also collec bea ing ib a ion da abases publicly
a ailable om di e en in e ne eposi o ies.
Once he bea ing ib a ion da abases a e collec ed and he audio and ib a ion
da abase om he cen i ugal pump is acqui ed in ou labo a o y, we ca y ou di e en
expe imen s o bea ing applica ion and o pump applica ion.
Fo bea ing applica ion, wo new p oposed me hods based on nonlinea ea u es
a e applied o he ib a ion signals om he bea ings o aul diagnosis and aul
iden i ica ion. The p oposed me hods a e compa ed wi h me hods in he li e a u e. The
abili y o he p oposed ea u es o bea ing aul diagnosis is quan y ied using wo
classi ie s: a neu al ne wo k classi ie and a Leas -Squa e Suppo Vec o Machine (LS-
SVM) classi ie . In he case o bea ing iden i ica ion, an index o ollow he deg ada ion
o bea ings wi h ou e ace aul and di e en se e i ies is p oposed.
Fo cen i ugal pump applica ion, ea u es om he s a e o he a o pump aul
diagnosis a e implemen ed and ex ac ed om ib a ion signals o he cen i ugal pump.
Then, hese ea u es a e applied o he audio acqui ed om he cen i ugal pump.
Fea u es in equency domain, ceps um domain, ime- equency domain and nonlinea
ea u es a e p oposed o pump condi ion moni o ing. Fea u e selec ion is implemen ed
o selec ele an ea u es in pump aul diagnosis. We s udy he ele ance o he
selec ed ea u es and he pe o mance o ib a ion and audio signals using wo s anda d
classi ie s (LS-SVM and neu al ne wo ks). Finally we analyze how he usion o audio
signals o ib a ion signals a ec s he moni o ing pe o mance.
1.8 Ou line o he Disse a ion
The Disse a ion is s uc u ed acco ding o a adi ional complex ype [13] wi h
li e a u e e iew and wo di e en applica ions in which me hods a e explained and
applied o expe imen al s udies. The chap e s uc u e is as ollows:
.Chap e 1 in oduces he opic o condi ion moni o ing and gi es he mo i a ion,
ou line, me hodology and con ibu ion o his PhD Thesis.

Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 39
.Chap e 2 summa izes ela ed wo ks which ha e gi en ise o he mo i a ion o
he Thesis and de ails he mo i a ion o his Thesis based on hese p e ious wo ks. The
s a e o he a is a con ibu ion o his Thesis. The w i ing o Chap e 2 is based on a
pape published by he au ho o his Thesis.
.Chap e 3 is de o ed o he applica ion a eas o his Thesis, namely bea ings and
cen i ugal pumps. A b ie desc ip ion o bea ings basics ollowed by a desc ip ion o
he public bea ing ib a ion da abases used in his Thesis is ca ied ou in he i s pa
o he chap e . The second pa is ocused on cen i ugal pump applica ion. Pump basics
and he acquisi ion p ocess o he audio and ib a ion da abase a e desc ibed. We
con ibu e wi h he eco ding o an audio and ib a ion da abase o a cen i ugal pump.
.Chap e 4 shows he con ibu ions in bea ing aul diagnosis and aul
iden i ica ion using ib a ion signals. Two me hods based on nonlinea echniques a e
p oposed and he esul s o he applica ion o each me hod a e shown. The con ibu ions
o his Chap e a e he p oposed new me hods based on nonlinea ea u es o bea ing
aul diagnosis and iden i ica ion. The w i ing o Chap e 4 is based on h ee
publica ions by he au ho o his Thesis.
.Chap e 5 is de o ed o he con ibu ions in he cen i ugal pump applica ion o
aul diagnosis using audio and ib a ion signals. The ea u es used in pump aul
diagnosis li e a u e a e desc ibed as well as he ea u es p oposed in his Thesis o
pump aul diagnosis. We con ibui e wi h he applica ion o ib a ion ea u es o audio
signals and wi h he p oposal o a se o new ea u es in equency, ceps um, ime-
equency and nonlinea domains o pump aul diagnosis. The esul s o he ea u e
e alua ion wi h wo classi ie s a e also shown. Pa o he w i ing o Chap e 5 is based
on wo publica ions o he au ho o his Thesis.
.Chap e 6 shows he esul s o audio and ib a ion usion a ea u e le el, sco e
le el and decision le el in he pump applica ion. We con ibu e wi h a s udy o he
combina ion o audio and ib a ion signals in pump aul diagnosis.
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
40
.Chap e 7 concludes he Disse a ion summa izing he main esul s ob ained and
ou lining u u e esea ch lines.
Chap e s 5, 6 and 7 ha e an in oduc o y pa in which he me hodology is
explained, a second pa e in which he me hodology is applied o he da abases and a
hi d pa wi h he esul s and conclusions.
.The Appendix o he Thesis is an ex a chap e ha shows a summa y o he
wo k ca ied ou in oice pa hology de ec ion and in he disc imina ion be ween
emo ional s a es in speech du ing he Thesis. As he backg ound o he PhD candida e is
oice cha ac e iza ion using nonlinea ea es, du ing he Thesis she con inued his
esea ch line.
The dependence among he chap e s is illus a ed in Figu e 1-2. Fo example,
be o e eading any o he Chap e s 4, 5, and 6, one should ead i s Chap e s 3 and 41
Be o e Chap e 3 one should s a wi h he in oduc ion in Chap e 1, and i is
ecommended o ead also Chap e 2.
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 41
Figu e 1-2: Dependence among Disse a ion chap e s.
Chap e 1:
“In oduc ion”
Chap e 2:
“S a e o he a in machine y
condi ion moni o ing using ib a ion
and audio signals”
”
Chap e 3:
“Da abases: bea ings and pumps”
Chap e 4:
“Con ibu ions o
ib a ion bea ing aul
diagnosis and aul
iden i ica ion”
Chap e 5:
“Con ibu ions o audio
and ib a ion pump
moni o ing”
Chap e 6:
“Resul s o audio and
iba ion usion in
pump moni o ing”
Chap e 7:
“Conclusions”
Appendix:
“Con ibu ions o oice”
P eceeding block is equi ed
P eceeding block is ecomended
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
42
1.9 Resea ch con ibu ions
The esea ch con ibu ions o his PhD Thesis a e di ided in o condi ion moni o ing
con ibu ions and oice con ibu ions. P o ided ha he esea ch backg ound o he PhD
candida e is oice cha ac e iza ion and she kep wo king on his subjec du ing he PhD,
he publica ions in oice a e also p esen e he e. Jou nal pape s included in ISI JCR
appea in bold.
1.9.1 Resea ch con ibu ions in condi ion moni o ing
1. Li e a u e e iew in audio and ib a ion aul diagnosis echniques ocusing on
ea u e ex ac ion and pa e n ecogni ion echniques.
Hen íquez, P., Alonso, J. B., Fe e , M., & T a ieso, C. M. (2014). Re iew o au oma ic
aul diagnosis sys ems using audio and ib a ion signals. Sys ems, Man, and Cybe ne ics:
Sys ems, IEEE T ansac ions on, 44(5), 642-652.
2. A me hod based on Teage -Kaise ene gy ope a o and s a is ic and ene gy ea u es
o bea ing aul diagnosis and applica ion o he p oposal o bea ing deg ada ion in a
helicop e .
Hen íquez, P., Alonso, J. B., Fe e , M. A., & T a ieso, C. M. (2013). Applica ion o he
Teage –Kaise ene gy ope a o in bea ing aul diagnosis.ISA ansac ions, 52(2), 278-
284.
Hen íquez, P., Whi e, P., Alonso, J. B., Fe e . M. A. (2011, Oc obe ). Applica ion o Teage -
Kaise Ene gy Ope a o o he Analysis o Deg ada ion o a Helicop e Inpu Pinion Bea ing. In
P oc. o he In e na ional Con e ence Su eillance 6 (pp. 265-274), Uni e si y o Technology
o Compiègne, F ance.
Hen íquez, P., Alonso, J. B., Fe e , M. A., T a ieso, C. M. (2011, May-June). Applica ion o
Highe O de S a is ics o Teage -Kaise Ene gy T ans o med Vib a ion Signal o Bea ing
Faul Diagnosis. In P oc. o he 24 h In . Cong ess on Condi ion Moni o ing and Diagnos ics
Enginee ing Managemen (pp. 265-274), S a ange , No way.
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 49
Figu e 2-2: Dis ibu ions o e e enced pape s: Audio-based and
ib a ion-based aul diagnosis dis ibu ion (uppe le ). Publishing
yea s dis ibu ion (uppe igh ).Dis ibu ion o aul s in ib a ion-
based diagnosis (bo om le ) and in audio-based diagnosis (bo om
igh ).
The emainde o his chap e is di ided in o i e subsec ions. The second, hi d
and ou h subsec ions ocus on he s eps o a CM sys em (acquisi ion, da a p ocessing
and decision s age). Subsec ion 2.5 epo s on some examples o CM sys ems. Finally,
an analysis o he gaps o he s a e o he a e is ca ied ou .
2.2 Acquisi ion S age
Good quali y and p ecision o he signal in he acquisi ion s age is essen ial o pos e io
analysis and ea u e ex ac ion. In his Thesis, we ocus on ib a ion and audio signals
acqui ed by accele ome e s and mic ophones, espec i ely.
The dynamic o ces wi hin a machine p oduce comp ession and bending wa es.
This ib a ion pa e n changes when an incipien ailu e s a s o e ol e. Thus, he
analysis o ib a ion signals is a use ul ool o es ablishing he machine’s condi ion. In

PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
50
o de o de ec hese signals, ib a ion senso s a e moun ed di ec ly on o he machine.
The e a e di e en kinds o senso s depending on he equency ange: posi ion senso s
(0Hz-10kHz), eloci y senso s (10Hz-1kHz) and accele ome e s (8Hz-15kHz).
Piezoelec ic accele ome e s a e popula because o hei highe dynamic ange o
equencies, eliabili y, obus ness and smalle dimensions. The numbe and loca ion o
ib a ion senso s is an impo an issue discussed in [8].
The acous ic cha ac e is ics o a machine change when a aul e ol es.
Consequen ly, he sound o a machine ca ies in o ma ion abou i s condi ion.
Ex ac ing he sound signa u e o he machine is a use ul ool in aul diagnosis [1]. We
ocus ou e iew on audio signals ob ained wi h mic ophones. They usually acqui e
sounds in he 0Hz-20kHz ange. Some mic ophones can e en acqui e signals abo e
100kHz. Mic ophones a e no moun ed di ec ly on o he machine. As a esul , hey a e
less in usi e han ib a ion senso s bu hey a e mo e sensi i e o en i onmen al noise.
Fo his eason, mic ophones mus be poin ed o he machine o sys em unde
conside a ion and should be placed om 2 cm o 10 cm om he wan ed sou ce [1],
[12].
2.3 P ocessing S age
Signal p ocessing ans o ms o iginal signals in o use ul ea u es o accomplish aul
diagnosis. These ea u es should be independen o he no mal machine ope a ing
condi ions ( a ia ions o load and speed) and ex aneous noise and be sensi i e only o
machine y aul s. This sec ion is di ided in o ib a ion and audio signals analysis. Table
2-1 shows some o he mos common ib a ion and audio ea u es discussed in he
li e a u e since 2000 up o 2015.
2.3.1 Vib a ion signal p ocessing
The main p ocessing echniques applied o ib a ion signals a e based on: ime analysis,
equency and ceps al analysis, ime- equency analysis and non-linea analysis.
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 51
Time domain analysis
Signal p ocessing in he ime domain ex ac s in o ma ion om he ib a ion signal as a
unc ion o ime. Con en ional echniques include se e al ime ea u es such as oo
mean squa e, c es ac o [35], a iance, skewness, ku osis and highe o de momen s
[68],[93]. Some ea u es, such as c es ac o , ku osis, impulse and clea ance ac o s, do
no a y wi h load and speed a ia ions and a e good indica o s o impulsi e aul s
[35], [60] especially in bea ings. O he con en ional echniques a e ime a e aging
me hods, including ime synch onous a e age (TSA), esidual signal and di e ence
signal, all o which a e powe ul ools in he de ec ion o gea aul s [36], [37], [102].
TSA emo es backg ound noise and pe iodic e en s ha a e no synch onous wi h he
gea o in e es . The esul ing signal is used o pos e io ad anced analysis [80].
Au o eg essi e (AR) modelling [30] and au o eg essi e mo ing a e age (ARMA)
modelling [38] ha e also p o en o be e icien ools in modelling ansien s in he
ib a ion signal. No el app oaches include a modi ica ion o he ime- a ying AR and
ARMA models in which he coe icien s a e upda ed wi h he incoming ib a ion signal
[37]. These models a e obus o a ia ions in load and speed. Ano he no el AR
app oach [102] p oposes he load in o ma ion as an exogenous inpu .
F equency and ceps al domain analysis
F equency analysis gi es in o ma ion abou he pe iodici y o he signal in he peaks o
he equency and de ec s ha monics and side-bands. Con en ional equency ea u es,
such as he mean and s anda d de ia ion o he equency, he oo mean squa e
equency, peak magni ude, ene gies and a ios o spec al ene gies a e used in aul
diagnosis in pumps, mo o s and gea boxes [35], [63], [67]. Ano he con en ional
echnique is en elope analysis (EnA) o he ampli ude demodula ion echnique, used
especially in bea ings [7], [16], [63] o iden i y he bea ing de ec cha ac e is ic
equency and also in gea s [76]. EnA imp o es he signal o noise a io (SNR) and
makes he spec al analysis mo e e ec i e. Fo a good e iew o EnA see [16]. EnA is
usually applied using he Hilbe ans o m (HT) [16]. Recen ly, he skewness
in o ma ion wa e (SIW) has been p oposed o compu e EnA wi hou using he HT
[103]. The skewness is compu ed in small egions o he signal, esul ing in a skewness
wa e. The skewness in o ma ion wa e (SIW) is ob ained using he Kullback-Leible
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
52
di e gence in o ma ion om he skewness wa e o he e e ence signal (measu ed a
om he diagnos ic loca ion) and he diagnosis signal (measu ed in he diagnos ic
loca ion). Then he en elope SIW is ob ained om hei absolu e alues and he
spec um is compu ed. This me hod is shown o be supe io o he con en ional EnA
and mo e obus o s ong backg ound noise. Ano he no el EnA me hod is he Teage
Ene gy Ope a o (TEO) [50], [100]. TEO is a nonlinea ope a o wi h highe
demodula ion p ecision and needs less calcula ion han HT.
Ceps al analysis, which gi es in o ma ion om he ib a ion signal as a
unc ion o que ency has, since he nine ies, been shown o be e ec i e in aul
diagnosis [83]. Powe ceps um gi es in o ma ion abou he pe iodici y o he spec um
and de ec s ha monics and sideband pa e ns in he powe spec um. The applica ion o
Mel-F equency Ceps al Coe icien s (MFCC), a echnique om speech p ocessing, o
ib a ion aul diagnosis [40] was p oposed in 2006. MFCC con ain bo h ime and
equency in o ma ion o he signal which makes hem mo e use ul o ea u e
ex ac ion in ib a ion signals. In 2007, a me hod called minimum a iance ceps um
was p oposed o de ec aul y pe iodic impulses in bea ings in noisy en i onmen s [19].
I minimizes he a iance o he signal powe in i s ceps um ep esen a ion.
Ano he amily o echniques is based on highe o de s a is ics (HOS) in he
equency domain: bispec um, summed bispec um and bicohe ence ha e, since he la e
nine ies, been shown o be e ec i e in aul diagnosis in he bea ings o induc ion
mo o s, gea boxes and in lexible o o sys ems [17], [18], [39]. These p o ide mo e
in o ma ion han he powe spec a, in he case o non-Gaussian signals, can de ec
nonlinea couplings and can explain he o igin o ce ain peaks in he powe spec a. Fo
ins ance, [96] p oposed in 2009 he use o HOS in he ceps al domain (biceps um) o
de ec ailu es in gea s. This echnique elimina ed noise and modula ion e ec s caused
in gea s.
Con en ional equency echniques assume s a iona i y and linea i y and a e
usually applied in machines wo king a ixed speeds. Howe e , mos machine p ocesses
p esen s non-s a iona y componen s in speed-up, speed-down and in se e al aul s.
Cyclos a iona y analysis (a 2nd-o de echnique in he equency domain), and ime-
equency echniques a e mo e app op ia e o non-s a iona y p ocesses. The pe iodic
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 53
a ia ion o s a is ical momen s o o a ing machine y makes cyclos a iona i y
echniques sui able o ea ly aul de ec ion [84], [98].
Time-F equency domain analysis
Time- equency analysis ex ac s in o ma ion om he ib a ion signal as a unc ion o
ime and equency and o e comes he p oblems encoun e ed in equency analysis
when analyzing non-s a iona y e en s. Some con en ional ime- equency echniques
include Sho ime Fou ie ans o m (STFT) (Koo e Kim, 2000) [41], Wigne -Ville
dis ibu ion (WVD) (Koo e Kim, 2000; Li e Meche ske, 2006) [41], [56] and he
di ec ional Choi-Williams dis ibu ion [42]. Techniques om he la e nine ies include
Empi ical Mode Decomposi ion (EMD) [35], [43], he Hilbe -Huang ans o m (HHT)
[21], [35], and he Wa ele T ans o m (WT) [12], [81]. The WT has adjus able window
size h ough he choice o he mo he wa ele and di e en app oxima ion scales. This
lexibili y makes i sui able o he analysis o non-s a iona y signals. Mul iple ea u es,
such as singula i y poin s [20], Lipschi z exponen s [13], scalog am [36], ene gies,
s a is ics and en opies [44], [81], a e ex ac ed om con inuous WT (CWT), disc e e
WT (DWT) and Wa ele Packe T ans o ms (WPT). These a e used o de ec
imbalance, misalignmen , spalling, pi ing in gea s, aul y bea ings and OLTC [97].
Con a y o WT, EMD is a sel -adap i e me hod, applied o aul diagnosis o he i s
ime in 1998. EMD decomposes he signal in o a sum o in insic mode unc ions
(IMFs). The equency componen s in each IMF a e ela ed o he sampling equency
and o he signal i sel , whe eas WT is ela ed only o he sampling equency. Howe e ,
he numbe o IMFs canno be con olled [35], [63], [91], [117]. The HHT uses he
EMD o ob ain IMFs [21], [35] and hen EnA is applied o each IMF using he HT.
Many new echniques ha e, in he las yea s, been p oposed in ime- equency
analysis. Mos o hem a e ela ed o he decomposi ion o he signal in o mono-
componen AM-FM signals (ampli ude modula ed and equency modula ed signals)
which a e hen analyzed using EnA echniques, and o he imp o emen s in WT and
EMD echniques. These echniques allow a ine analysis o he signal and mo e
accu acy o de ec single and compound aul s, essen ial in aul diagnosis. They a e
discussed nex .
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
54
Spec al Ku osis (SK) is a spec al s a is ic e o mula ed in 2006 o non-
s a iona y signals [104]. SK p o ides a obus way o de ec ing incipien aul s ha
p oduce impulse-like signals, e en in he p esence o s ong noise. SK also o e s a way
o designing op imal il e s o il e ing ou he mechanical signa u e o aul s using he
ku og am o he as ku og am (ways o compu e he SK) as a p elude o EnA [105].
Recen ly, an enhanced ku og am has been p oposed o bea ing aul diagnosis in
combina ion wi h wa ele packe ans o m [114]. Local mean decomposi ion (LMD)
and imp o ed LMD a e also p oposed o aul diagnosis [110]. Con a y o he HHT,
LMD does no use HT o es ima e he en elope bu uses he mo ing a e age. P oduc
unc ions a e ob ained by mul iplying he en elope es ima es and he FM signal.
Gene alized demodula ion ime equency (GDTF) also decomposes he signal in o
mono-componen AM-FM signals, ans o ming he o iginal signal in o a new space
whe e WT can be applied, he e o e ob aining equencies wi h physical meaning.
GDTF was i s p oposed o analyzing biomedical signals. The en elope o de
spec um echnique, which blends GDTF and he spec um, has also been p oposed o
aul diagnosis [99]. I e a ed Hilbe ans o m (I HT) analyzes AM-FM signals using
he i e a ed applica ion o he HT o a il e ed e sion o he ampli ude en elope.
Ampli ude en elopes and ins an aneous equencies a e hen ex ac ed. I HT has highe
demodula ion accu acy and lowe complexi y han EMD. The combina ion o I HT and
a smoo hed ins an aneous equency es ima ion has been ecen ly applied o aul
diagnosis [106]. Ensemble EMD (EEMD) is a no el echnique (2009) ha elimina es
he mode mixing p oblem in EMD. In he mode mixing p oblem he physical meaning
o each IMF is unclea and EMD ails o ep esen he aul cha ac e is ics o a signal
accu a ely. EEMD uses a noise-assis ed echnique o elimina e he mode mixing
p oblem. EEMD and EMD a e applied o a o o aul and in a hea y oil ca aly ic
c acking machine se [43]. Resul s show ha EEMD can ex ac he aul cha ac e is ic
in o ma ion be e han EMD. The mul i-scale en eloping spec og am (MuSEnS) [46]
algo i hm was de eloped in 2009 using ime, scale, and equency domain in o ma ion
con ained in he signal. I decomposes he signal in o di e en wa ele scales and he
en elope signal in each scale is calcula ed, esul ing in an ‘‘en elope spec um’’ [46].

Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 55
TABLE 2-1: FEATURES FOR VIBRATION AND AUDIO SIGNALS
E olu ion
Me hods
2000-2006
2007-2015
Time
S a is ics
(2004)[59] (A)
(2009)[37] (V)
TSA
-
(2010)[36] (V)
AR,ARMA
models
(2002)[30], (2006)[38] (V)
(2010)[102] (V)
F equency
S a is ics
(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)
Cyclos a iona i y
(2001)[84] (V)
(2010)[98] (V)
Polyspec um
(HOS)
-
(2007)[39] (V)
Ceps al
Ceps um
-
(2007)[19] (V)
MFCC
(2006) [40] (V)
(2009)[54] (A)
Time-
F equency
STFT, WVD,
Choi dis .
(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)
d cWT
-
(2010)[94] (V)
Gene alized S
ans o m
-
(2011)[107] (V)
MuSEnS
-
(2009)[46](V)
EMD, HHT,
EEMD
(2005)[21] (V)
(2010)[35], (2009) [43],
(2015)[117] (V)
I HT
-
(2008)[106] (V)
LMD
-
(2009)[110] (V)
GDTF
-
(2010)[99] (V)
SK
-
(2009)[104], (2009)[105]
(2013)[114] (V)
Non-
linea
Phase po ai ,
do pa e n
(2000)[10] (A), (2003)[11] (V)
-
-
Lyapuno
Exponen s, CD,
ac als
(2001)[18], (2000)[23] (V)
(2007)[24], (2007)[52],
(2008)[51], (2010)[95] (V),
(2005)[55] (A,V),
(2007)[25] (A)
ApEn,
Mul i-Scale
Pe mEn
(2007)[48], (2007)[49],
(2015)[116](V)
Mul iple
mani old
-
(2009)[109] (V)
V: Vib a ion, A: Audio.
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
56
Second gene a ion WT (SGWT) [111] was p oposed in 2006 and o e comes he
main sho coming o he WT ( he p ope selec ion o he mo he wa ele ) because
SGWT is ealized by a li ing scheme in he ime domain (which is no based on he
Fou ie ans o m). The imp o ed WPT (IWPT) is also based on SGWT and i is shown
o be supe io o WPT in ex ac ing he aul cha ac e is ics in bea ings [44], [68]. An
app oach ha imp o es he SGWT is he edundan SGWT (RSGWT) p oposed in 2009
[108], [111]. RSGWT is ime-in a ian , con a y o SGWT, which allows cap u ing
mo e use ul aul in o ma ion. RSGWT ou pe o ms SGWT in ex ac ing ansien
componen s in gea box ib a ion signals [111]. The dual- ee complex WT (d cWT)
was p oposed in 2010 [94] in aul diagnosis. In [94], he au ho s show ha d cWT
ou pe o ms SGWT, as ku og am and DWT because d cWT enhances noise
educ ion, is app oxima e ime-in a ian and can de ec mul iple aul ea u es
simul aneously Ano he applica ion o he WT includes a no el g ow h index [47],
insensi i e o di e en mo he wa ele s and le els o decomposi ion. Finally, he
gene alized S ans o m was p oposed in 2011 o aul diagnosis. I uni ies STFT and
WT so as o ob ain mo e sa is ac o y ime- equency ep esen a ions han o he simila
echniques such as STFT, WVD and he S ans o m. This allows mo e accu a e
de ec ion o he bea ing aul cha ac e is ic [107].
Nonlinea analysis
E idence o a complex and non-linea ib a o y sys em has been ound in s a o - o o
ub, loose pedes al and uns able oil ilm aul s [18]. Con en ional non-linea me hods
include pseudo-phase po ai , singula spec um analysis, co ela ion dimension (CD)
[22], [23], ac al dimensions, app oxima e en opy (ApEn), in o ma ion en opy [92],
mu ual in o ma ion [25] and Lyapuno exponen s [14], [109]. CD quan i ies he
complexi y o a ime se ies and has success ully p o en o de ec o o -s a o ub, loose
pedes al aul s [23] and bea ing aul s [22]. Phase po ai shows quali a i e di e ences
in a no mal gea and in a gea wi h an ea ly a igue c acked oo h [10]. ApEn quan i ies
he egula i y o a ime se ies and can e ec i ely indica e he condi ion in ans [48] and
bea ings wi h speed and load a ia ions [49].
A new echnique is p oposed in [24] o selec a p ope ac al dimension
spec um less a ec ed by noise. In a mo e ecen pape (2008) [51], a modi ica ion o
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 57
he co ela ion in eg al is in oduced o he eal- ime aul diagnosis o bea ings.
Mul iple mani old analysis is a no el nonlinea app oach [109] (2009) ha ex ac s
mani old in o ma ion om he ib a ion signals and ou pe o ms con en ional nonlinea
echniques. A pape o 2010 p oposes o compu e he ac al dimension using DWT
[95]. F ac al ea u es a e es ima ed om he slope o he a iances o he DWT in
di e en scales. The mul iscale pe mu a ion en opy (Mul i-Scale Pe mEn) has ecen ly
applied o bea ing aul diagnosis [116]. Mul i-Scale Pe mEn compu es he pe mu a ion
en opy ac oss di e en scales.
2.3.2 Audio signal p ocessing
Mos esea ch in machine y diagnosis is o ien ed owa ds he analysis o ib a ion
signals. Audio-based CM has, howe e , no de eloped a he same a e. This is due o
he con amina ion o he sound signal by unwan ed sou ces such as o he machines,
noisy en i onmen s and he s uc u al ib a ion o he machine i sel [2]. This si ua ion
makes i di icul o acqui e he machine’s signa u e. Two main op ions a e used o
imp o e he low SNR in audio signals: he use o pa ial o ull enclosu e using an
anechoic chambe [15], [17], which is an un ealis ic app oach o eal indus ial
scena ios, o he use o p e-p ocessing de-noising me hods, such as wa ele [12] o
blind sou ce sepa a ion [2], [15], [26]. These echniques can a ec he ea u e ex ac ion
s age. The choice o ce ain pa ame e s, such as he h eshold in wa ele echniques, is
impo an o he ex ac ion o he pu i ied signal wi h he smalles dis o ion and he
highes SNR. Audio-based echniques a e use ul in ce ain cases, especially when i is
impossible o access he machine. Audio measu emen s can be pe o med a a dis ance
om he machine so he use o senso s moun ed di ec ly on he machine is a oided.
The same p ocessing echniques o ib a ions a e applied o ex ac ea u es in
audio signals ob ained om machines. S a is ical ime domain ea u es and ene gy
ea u es in he equency domain [59] a e used o de ec wea in in e nal combus ion
engines and mass unbalance aul s in o a y disks. EnA is used in a hea y sizing-p ess
[26] and in he end- es o acuum cleane p oduc ion [58], whe e he use o ib a ion
and cu en analysis did no p o e use ul. As in ib a ion-based echniques, he audio
spec um is used when he machine is wo king a cons an speeds [10], [56]. Some
pape s compa e he use o Fou ie analysis in audio and ib a ion signals, [10], [87].
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
58
These wo ks claim ha Fou ie analysis is less e ec i e in sound signals because o he
low SNR. As in he case o ib a ion analysis, ime- equency echniques a e mo e
app op ia e. Fo example, STFT is applied o iden i y engine aul equencies [54].
Howe e , WVD [86] and pseudo WVD [56] a e shown o ha e be e esul s in
ex ac ing non-s a iona y signa u es in he bea ings o induc ion mo o s. Con inuous
WT, WPT and MFCC a e success ully applied o aul diagnosis o engines [15], [52],
[57] wo king a di e en speeds and un-up condi ions.
The symme ised do pa e n echnique is based on he isualiza ion o changes
in ampli ude and equency o he audio signal. I has been applied o dis inguish
be ween no mal and aul y ans [10] and in he aul iden i ica ion o an in e nal
combus ion engine wo king a di e en speeds [55]. Some pape s ha e ocused on he
analysis o a g oup o chao ic measu es ex ac ed om audio signals o di e en
machines. Asynch onous changes in audio signals can be de ec ed wi h chao ic
measu es [25]. The spec al en opy o audio signals has shown o be mo e e ec i e
han ib a ion signals in he diagnos ic o ca i a ion [88].
The de elopmen o echniques o imp o e SNR and he use o he same ool,
i.e. wa ele s, in he de-noising and ex ac ion s ages can pu audio-based echniques a
an ad an ageous posi ion in CM sys ems since hey a e less in usi e han ib a ion-
based echniques.
Mul iple ea u es a e ex ac ed om ib a ion and audio signals in CM
echniques. Howe e , a unique ea u e capable o ep esen ing he machine condi ion
does no exis . A good cha ac e iza ion equi es he ex ac ion o di e en ea u es om
di e en domains. Thus we need o ind, using ea u e selec ion, he mos sui able ones
o he applica ion unde conside a ion. The selec ion o well-sui ed ea u es p o iding
aul - ela ed in o ma ion and he disca ding o weakening o i ele an o edundan
ea u es is an impo an s age in machine CM o imp o ing sys em pe o mance.
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 65
ib a ion-based and audio-based au oma ic aul diagnosis in machine y implemen ing
he s ages o a CBM sys em (da a acquisi ion, signal p ocessing and decision-making).
We ha e paid special a en ion o ecen ad ancemen s in signal p ocessing echniques
and classi ica ion me hods. Finally, some examples o CM sys ems ha e been
p esen ed.
Vib a ion-based moni o ing is a well-es ablished echnique widely used in CBM
[2], [3]. Acco ding o Figu e 2-2, mos esea ch pape s consul ed o his s a e o he a
a e ela ed o ib a ion-based aul diagnosis. This ac seems easonable due o wo
aspec s: he easiness o acqui e he ib a ion signal om a machine and he ansmission
pa h be ween he machine o a componen o he machine and he senso is less a ec ed
by in e e ences han in he case o audio-based diagnosis. We can ex ac ano he
conclusion om Figu e 2-2. Mos aul s a e ela ed o bea ings. In ac a la ge mo o
eliabili y su ey [119] epo s ha 42% o he aul s in la ge mo o s (mo e han 200
ho sepowe ) a e ela ed o bea ing aul s. In small mo o he pe cen age o aul s ela ed
o aul s is 90% [120]. Fo his eason, esea ch e o s in condi ion moni o ing a e
ocused on bea ing aul diagnosis. Mos bea ing aul diagnosis is done using ib a ion
signal and cu en signal as sou ce o in o ma ion. In Figu e 2-2 i is also obse ed ha
mos e e enced pape s ela ed o bea ings use ib a ion signal o moni o ing.
Mo eo e , in he elabo a ion o his e iew, we ca ied ou a sea ch o public a ailable
da abases. We ha e ound h ee ib a ion da abases public a ailable. Fo all hese
easons, in his Thesis, we ocus he i s pa o ou esea ch in me hods o bea ing
aul diagnosis and bea ing aul iden i ica ion using ib a ion signals.
Signal p ocessing echniques ha e e ol ed om con en ional ime and
equency analysis, which assume s a iona i y and linea i y, o mo e de eloped
echniques ha exploi he non-s a iona y and non-linea na u e o aul y signals and o
speed-up and speed down p ocesses. These p o ide a mo e ealis ic desc ip ion o he
eal condi ion o he machine. Fo his eason, in bea ing applica ion, ou esea ch aims
a aul diagnosis and aul deg ada ion using nonlinea echniques.
Audio-based moni o ing has no been applied o CBM sys ems o he same
deg ee as ib a ion-based moni o ing (see Figu e 2-2), e en hough mic ophones a e no
moun ed on he machine and ha e g ea e loca ion possibili ies. The main eason is he

PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
66
di icul y o eco e ing he machine’s signa u e because he signal can be imme sed in
noise. Some au ho s ha e p oposed o loca e he mic ophone a a dis ance om he
machine be ween 2cm and 20cm o a oid unwan ed in e e ences [12]. The applica ion
o audio- aul diagnosis echniques in an ex ensi e way will imp o e g ea ly he
inspec ion o ce ain indus ial en i onmen s in which a pe manen CM sys em is
expensi e o he moun ing o ib a ion senso s is di icul . Fo his eason, we hink
audio-based echniques need u he esea ch e o s and we ha e ocused he second
pa o ou esea ch on audio-based aul diagnosis.
P o ided he ac ha he e is a lack o la ge publicly a ailable da abases in aul
diagnosis and audio signals a e no an excep ion, we ha e buil an expe imen al se in
ou labo a o y o acqui e simul aneously ib a ion and audio signals om a cen i ugal
pump wo king in no mal condi ion and in di e en aul condi ions. In his way, we
ob ain ib a ion and audio signals o di e en machine condi ions ha we use in
di e en aul diagnosis expe imen s. Mo eo e , his da abase can be made public so
o he esea che s can bene i om i . Public da abases p o ide common da a o
compa a i ely e alua e aul diagnosis echniques and CM sys ems.
In his chap e , mul iple ea u es om ib a ion-based diagnosis a e desc ibed.
Mos o hem a e no used in audio-based diagnosis. In his Thesis, we add ess his
issue. We apply ib a ion ea u es o audio signals and we make a compa ison o he
pe o mance in bo h cases. Mo eo e , in he li e a u e he e a e only ew wo ks ela ed
o audio and ib a ion signals acqui ed oge he . In his Thesis, we ocus on a s udy o
audio and ib a ion signals acqui ed simul aneously om a cen i ugal pump. A se o
ea u es a e ex ac ed om bo h signals and classi ie s a e used o disc imina e be ween
di e en machine condi ions (no mal and aul y condi ions).
Da a usion o mul iple da a is a clea esea ch line in aul diagnosis. Mul iple
signals can be usioned o ob ain a mo e eliable diagnosis. In his Thesis, usion o
audio and ib a ion signals om a cen i ugal pump is ca ied ou .
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 67
2.7 Re e ences
[1] Hen iquez, P., Alonso, J. B., Fe e , M., & T a ieso, C. M. (2014). Re iew o
au oma ic aul diagnosis sys ems using audio and ib a ion signals. Sys ems, Man,
and Cybe ne ics: Sys ems, IEEE T ansac ions on, 44(5), 642-652.
[2] Vilela, R. M., Me ôlho, J. C., & Ca doso, J. C. (2004, May). Machine and indus ial
moni o iza ion sys em by analysis o acous ic signa u es. In Elec o echnical
Con e ence, 2004. MELECON 2004. P oceedings o he 12 h IEEE
Medi e anean (Vol. 1, pp. 277-279). IEEE.
[3] Basak, D., Tiwa i, A., & Das, S. P. (2006, Decembe ). Faul diagnosis and condi ion
moni o ing o elec ical machines-A Re iew. In Indus ial Technology, 2006. ICIT
2006. IEEE In e na ional Con e ence on (pp. 3061-3066). IEEE.
[4] Ja dine, A. K., Lin, D., & Banje ic, D. (2006). A e iew on machine y diagnos ics
and p ognos ics implemen ing condi ion-based main enance.Mechanical sys ems and
signal p ocessing, 20(7), 1483-1510.
[5] Nandi, S., Toliya , H., & Li, X. (2005). Condi ion moni o ing and aul diagnosis o
elec ical mo o s-a e iew. Ene gy Con e sion, IEEE T ansac ions on, 20(4), 719-729.
[6] Vach se anos, G., Lewis, F. L., Roeme , M., Hess, A., and Wu, B. (2006).
In elligen Faul Diagnosis and P ognosis o Enginee ing Sys ems, 1s ed. Hoboken,
New Je sey: John Wiley & Sons, Inc.
[7] Tandon, N., & Choudhu y, A. (1999). A e iew o ib a ion and acous ic
measu emen me hods o he de ec ion o de ec s in olling elemen bea ings.T ibology
in e na ional, 32(8), 469-480.
[8] Doebling, S. W., Fa a , C. R., & P ime, M. B. (1998). A summa y e iew o
ib a ion-based damage iden i ica ion me hods. Shock and ib a ion diges , 30(2), 91-
105.
[9] Yang, H., Ma hew, J., & Ma, L. (2003, No ). Vib a ion ea u e ex ac ion
echniques o aul diagnosis o o a ing machine y: a li e a u e su ey.In Asia-Paci ic
Vib a ion Con e ence, 2003 (pp. 277-279).
[10] Shiba a, K., Takahashi, A., & Shi ai, T. (2000). Faul diagnosis o o a ing
machine y h ough isualisa ion o sound signals. Mechanical Sys ems and Signal
P ocessing, 14(2), 229-241.
[11] Wang, W. J., & Lin, R. M. (2003). The applica ion o pseudo-phase po ai in
machine condi ion moni o ing. Jou nal o sound and ib a ion, 259(1), 1-16.
[12] Bayda , N., & Ball, A. (2003). De ec ion o gea ailu es ia ib a ion and
acous ic signals using wa ele ans o m. Mechanical Sys ems and Signal
P ocessing, 17(4), 787-804.
[13] Shanlin, K., Peilin, P., Feng, F., & Guangbin, D. (2007, July). Fea u e ex ac ion
me hod in aul diagnosis based on wa ele uzzy ne wo k o powe sys em o a ing
machine y. In Con ol Con e ence, 2007. CCC 2007. Chinese(pp. 437-441). IEEE.
[14] Mülle , P. C., Bajkowski, J., & Sö ke , D. (1994). Chao ic mo ions and aul
de ec ion in a c acked o o . Nonlinea Dynamics, 5(2), 233-254.
[15] Zhong, Z. M., Chen, J., Zhong, P., & Wu, J. B. (2006). Applica ion o he blind
sou ce sepa a ion me hod o ea u e ex ac ion o machine sound signals. The
In e na ional Jou nal o Ad anced Manu ac u ing Technology, 28(9-10), 855-862.
[16] McFadden, P. D., & Smi h, J. D. (1984). Vib a ion moni o ing o olling elemen
bea ings by he high- equency esonance echnique—a e iew. T ibology
in e na ional, 17(1), 3-10.
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
68
[17] Yang, D. M., S onach, A. F., MacConnell, P., & Penman, J. (2002). Thi d-o de
spec al echniques o he diagnosis o mo o bea ing condi ion using a i icial neu al
ne wo ks. Mechanical sys ems and signal p ocessing, 16(2), 391-411.
[18] Wang, W. J., Wu, Z. T., & Chen, J. (2001). Faul iden i ica ion in o a ing
machine y using he co ela ion dimension and bispec a. Nonlinea Dynamics,25(4),
383-393.
[19] Choi, Y. C., & Kim, Y. H. (2007). Faul de ec ion in a ball bea ing sys em using
minimum a iance ceps um. Measu emen Science and Technology, 18(5), 1433.
[20] Peng, Z. K., Chu, F. L., & Pe e , W. T. (2007). Singula i y analysis o he
ib a ion signals by means o wa ele modulus maximal me hod. Mechanical Sys ems
and Signal P ocessing, 21(2), 780-794.
[21] Peng, Z. K., Pe e , W. T., & Chu, F. L. (2005). A compa ison s udy o imp o ed
Hilbe –Huang ans o m and wa ele ans o m: applica ion o aul diagnosis o
olling bea ing. Mechanical sys ems and signal p ocessing, 19(5), 974-988.
[22] Logan, D. B., & Ma hew, J. (1996). Using he co ela ion dimension o
ib a ion aul diagnosis o olling elemen bea ings—II. Selec ion o expe imen al
pa ame e s. Mechanical Sys ems and Signal P ocessing, 10(3), 251-264.
[23] Wang, W., Chen, J., & Wu, Z. (2000). The applica ion o a co ela ion
dimension in la ge o a ing machine y aul diagnosis. P oceedings o he Ins i u ion o
Mechanical Enginee s, Pa C: Jou nal o Mechanical Enginee ing Science, 214(7),
921-930.
[24] Xinmin, T., Baoxiang, D., & Yong, X. (2007, Augus ). Bea ings aul diagnosis
based on HMM and ac al dimensions spec um. In Mecha onics and Au oma ion,
2007. ICMA 2007. In e na ional Con e ence on (pp. 1671-1676). IEEE.
[25] Hen íquez, P., Alonso, J. B., T a ieso, C. M., & Fe e , M. A. (2007, Augus ).
Ad ances in au oma ic de ec ion o ailu es in elec ic machines using audio signals. In
P oc. o 11 h IASTED In . Con . on A i icial In ell. and So Compu ing (pp. 114-119).
[26] Li, L., & Qu, L. (2002, Decembe ). Machine diagnosis wi h independen
componen analysis and en elope analysis. In Indus ial Technology, 2002. IEEE
ICIT'02. 2002 IEEE In e na ional Con e ence on (Vol. 2, pp. 1360-1364). IEEE.
[27] Fuga e, M. L., Sohn, H., Fa a , C. R. (2001). Vib a ion-based damage de ec ion
using s a is ical p ocess con ol. Mechanical Sys . and Signal P ocess., 15, 707-721.
[28] Bunks, C., McCa hy, D., Al-Ani, T. (2000). Condi ion-based main enance o
machines using hidden Ma ko models. Mechanical Sys . and Signal P ocess., 14, 597-
612.
[29] Li, Z., Wu, Z. , He, Y., Fulei, C. (2005). Hidden Ma ko model-based aul
diagnos ics me hod in speed-up and speed-down p ocess o o a ing machine y.
Mechanical Sys . and Signal P ocess., 19, 329-339.
[30] Wang, W., & Wong, A. K. (2002). Au o eg essi e Model-Based Gea Faul
Diagnosis. J. o Vib a ion and Acous ics, 124, 172-179.
[31] SAE (Socie y o Au omo i e Enginee s) E-32, 2007, “Heal h and Usage
Moni o ing Me ics, Moni o ing he Moni o ”, SAE ARP 5783-DRAFT, Feb. 14, 2007.
[32] Ku oglu, T., Mengshoel, O., Poll, S. (2008, Oc obe ). A F amewo k o
Sys ema ic Benchma king o Moni o ing and Diagnos ic Sys ems. In In . Con . on
P ognos ics and Heal h Managemen , (pp. 1-13), Den e .
[33] Ku oglu, T., Na asimhan, S., Poll, S., Ga cia, D., Kuhn, L., de Klee , J., Van
Gemund, A., & Feldman, A. (2009). Towa ds a F amewo k o E alua ing and
Compa ing Diagnosis Algo i hms. In P oc. o he 20 h In . Wo kshop on P inciples o
Diagnosis (DX’09), S ockholm.
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 69
[34] Roeme , M. J., Dzakowic, J., O sagh, R. F., Bying on, C. S., & Vach se anos,
G. (2005, Ma ch). Valida ion and Ve i ica ion o P ognos ic Heal h Managemen
Technologies. In IEEE Con . P oc.( pp. 3941-3947).
[35] Lei, Y., Zuo, M. J., He, Z., Zi, Y. (2010). A mul idimensional hyb id in elligen
me hod o gea aul diagnosis. Expe Sys . wi h Appl., 37, 1419-1430.
[36] Wang, X., Makis, V., Yang, M. (2010). A wa ele app oach o aul diagnosis o
a gea box unde a ying load condi ions. J. o Sound and Vib a ion, 329, 1570–1585.
[37] Shao, Y., Meche ske, C. K. (2009). Gea box ib a ion moni o ing using
ex ended Kalman il e s and hypo hesis es s. J. o Sound and Vib a ion, 325, 629–648.
[38] Wang, F. L., Meche ske, C. K. (2006). Adap i e modelling o ansien ib a ion
signals. Mechanical Sys . and Signal P ocess., 20, 825-842.
[39] Dou, W., Liu, Z.-S., Wang, D.-H. (2007, Augus ). Combina ion Diagnosis Based
on Gene ic Algo i hm o Ro a ing Machine y. In P oc. o he 3 d In . Con . on Na u al
Compu a ion ( ol. 4, pp. 307-313). IEEE.
[40] Nelwamondo, F. V., & Ma wala, T. (2006, Oc obe ). Faul s De ec ion Using
Gaussian Mix u e Models, Mel-F equency Ceps al Coe icien s and Ku osis. In IEEE
In . Con . on Sys ., Man, and Cybe n. ( ol. 1, pp. 290-295), Taiwan. IEEE.
[41] Koo, I. S., Kim, W. W. (2000). The de elopmen o eac o coolan pump
ib a ion moni o ing and a diagnos ic sys em in he nuclea powe plan . ISA T ans.,
39, 309-316.
[42] Lee, S. U., Robb, D., Besan , C. (2001) .The Di ec ion Choi-Williams
Dis ibu ion o he analysis o Ro o -Vib a ion Signals. Mechanical Sys . and Signal
P ocess., 15(4), 789-811.
[43] Lei, Y., He, Z., Zi, Y. (2009).Applica ion o he EEMD me hod o o o aul
diagnosis o o a ing machine y. Mechanical Sys . and Signal P ocess., 23, 1327–1338.
[44] Hu, Q., He, Z., Zhang, Z., Zi, Y. (2007). Faul diagnosis o o a ing machine y
based on imp o ed wa ele package ans o m and SVMs ensemble. Mechanical Sys .
and Signal P ocess., 21, 688-705.
[45] Yu, G., Li, C., Kama hi, S. (2009). Machine aul diagnosis using a clus e -
based wa ele ea u e ex ac ion and p obabilis ic neu al ne wo ks. In . J. Ad . Manu .
Technol., 42, 145–151.
[46] Yan, R. Gao, R. X. (2009). Mul i-scale en eloping spec og am o ib a ion
analysis in bea ing de ec diagnosis. T ibology In ., 42, 293-302.
[47] Wang, D., Miao, Q., Kang, R. (2009). Robus heal h e alua ion o gea box
subjec o oo h ailu e wi h wa ele decomposi ion. J. o Sound and Vib a ion, 324,
1141-1157.
[48] Tan, J., Gu, J. -J., Peng, X. -Z., Qin, Z. -M. (2007, Augus ). A aul diagnosis
me hod based on wa ele app oxima e en opy o an. In P oc. o he 6 h In . Con . on
Machine Lea ning and Cybe n. (pp. 519-523), Hong Kong.
[49] Yan, R., Gao, R. X. (2007). App oxima e En opy as a diagnos ic ool o
machine heal h moni o ing. Mechanical Sys . and Signal P ocess., 21, 824-839.
[50] Li, H., Fu, L., Zhang, Y. (2009, Ap il). Bea ing Faul s Diagnosis Based on
Teage Ene gy Ope a o Demodula ion Technique. In . Con . on Meas. Technol. and
Mecha onics Au oma ion, (pp. 594-597), Zhangjiajie, China.
[51] Janja asji , S., Ocak, H., Lopa o, K. A. (2008). Bea ing condi ion diagnosis and
p ognosis using applied nonlinea dynamical analysis o machine ib a ion signal. J. o
Sound and Vib a ion, 317, 112-126.
[52] Yang, J., Zhang, Y., Zhu, Y. (2007). In elligen aul diagnosis o olling elemen
bea ing based on SVMs and ac al dimension. Mechanical Sys . and Signal P ocess.,
21, 2012-2024.
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
70
[53] Singh, V., Meena, N. (2009, Janua y). Engine Faul Diagnosis using DTW,
MFCC and FFT. In P oc. o he 1s In . Con . on In elligen Human Compu e
In e ac ion, (pp. 83-94).
[54] Mansoo , A. B., Ahmed, H., Mahmood, Z. (2009, Feb ua y). Design and
De elopmen o an Au oma ed Acous ic based Je Engine Pe o mance E alua o . In
P oc. o he IEEE In . Con . on Indus ial Technol. (pp. 1-4).
[55] Wu, J. -D., Chuang, C. -Q. (2005). Faul diagnosis o in e nal combus ion
engines using isual do pa e ns o acous ic and ib a ion signals. NDT&E In ., 38,
605-614.
[56] Li, W., Meche ske, C. K. (2006). De ec ion o Induc ion Mo o Faul s: A
Compa ison o S a o Cu en , Vib a ion and Acous ic Me hods. J. o Vib a ion and
Con ol, 12(2), 165-188.
[57] Wu, J. -D., Liu, C. -H. (2009). An expe sys em o aul diagnosis in in e nal
combus ion engines using wa ele packe ans o m and neu al ne wo k. Expe Sys .
wi h Appl., 36, 4278-4286.
[58] Benko, U., Pe o cic, J., Ju icic, D., Ta ca , J., Rejec, J. (2005). An app oach o
aul diagnosis o acuum cleane mo o s based on sound analysis. Mechanical Sys .
and Signal P ocess., 19, 427-445.
[59] Li, W., Tsai, Y. P., Chiu, C. L. (2004). The expe imen al s udy o he expe
sys em o diagnosing unbalances by ANN and acous ic signals. J. o Sound and
Vib a ion, 272, 69–83.
[60] Sak hi el, N. R., Suguma an, V., & Nai , B. B. (2010). Compa ison o decision
ee- uzzy and ough se - uzzy me hods o aul ca ego iza ion o mono-block
cen i ugal pump. Mechanical Sys . and Signal P ocess., 24, 1887-1906.
[61] Ballal, P., Ramani, A., Middel on, M., McMu ough, C., A hamneh, A., Lee, W.,
& Lewis, F. (2009, Ma ch). Mechanical Faul Diagnosis using Wi eless Senso
Ne wo ks and a Two-S age Neu al Ne wo k Classi ie . In IEEE Ae osp. Con , (pp. 1-
10). IEEE.
[62] Li, Z., He, Y., Chu, F., Han, J., Hao, W. (2006). Faul ecogni ion me hod o
speed-up and speed-down p ocess o o a ing machine y based on independen
componen analysis and Fac o ial Hidden Ma ko Model. J. o Sound and Vib a ion,
291, 60-71.
[63] Lei, Y., He, Z., Yi, Y. (2008). A new app oach o in elligen aul diagnosis o
o a ing machine y. Expe Sys . wi h Appl., 35, 1593-1600.
[64] Wei, L., Pu, H. (2008, July) .Wa ele Neu al Ne wo k Aided On-Line De ec ion
and Diagnosis o Ro a ing Machine Faul . In Chinese Con ol and Decision Con ., 2008
(pp.1868-1871). IEEE.
[65] Ka , C., Mohan y, A. R. (2004). Applica ion o KS es in ball bea ing aul
diagnosis. J. o Sound and Vib a ion, 269, 439-454.
[66] Goumas, S. K., Ze akis, M. E., & S a akakis, G. S. (2002). Classi ica ion o
Washing Machines Vib a ion Signals Using Disc e e Wa ele Analysis o Fea u e
Ex ac ion. IEEE T ans. on Ins umen a ion and Meas., 51(3), 497-508.
[67] Lei, Y., He, Z., Zi, Y., Chen, X. (2008). New clus e ing algo i hm-based aul
diagnosis using compensa ion dis ance e alua ion echnique. Mechanical Sys . and
Signal P ocess., 22, 419-435.
[68] Pan, Y., Chen, J., Guo, L. (2009). Robus bea ing pe o mance deg ada ion
assessmen me hod based on imp o ed wa ele packe – suppo ec o da a desc ip ion.
Mechanical Sys . and Signal P ocess., 23, 669-681.

Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 71
[69] Sanz, J., Pe e a, R., Hue a, C. (2007). Faul diagnosis o o a ing machine y
based on au o-associa i e neu al ne wo ks and wa ele ans o ms. J. o Sound and
Vib a ion, 302, 981–999.
[70] Shina, H. J., Eomb, D. -H., Kim, S. -S. (2005). One-class suppo ec o
machines—an applica ion in machine aul de ec ion and classi ica ion. Compu e s &
Indus ial Eng., 48, 395-408.
[71] Case Wes e n Rese e Uni e si y Bea ing Tes Da a Cen e .
h p://www.eecs.case.edu/labo a o y/bea ing/ (Las isi ed: Jan. 2014).
[72] h p://qsun.eng.ua.edu/cpw_web_new/in o.h m (Las isi ed: Jan. 2014)
[73] MDTB Tes - un Da a CDs, Condi ion-based Main enance Depa men , Applied
Resea ch Labo a o y, The Pennsyl ania S a e Uni e si y.
[74] Ypma A., Lig e ingen, R.P.W., Duin, E.E.E., F ie man, E.E.E. (1999). Pump
ib a ion da ase s, Pa e n ecogni ion g oup, Del Uni e si y o Technology.
[75] h p:// i.a c.nasa.go / ech/dash/pcoe/p ognos ic-da a- eposi o y/ (Las isi ed:
Jan. 2012)
[76] Wang, W., Kanneg, D. (2009). An in eg a ed classi ie o gea sys em
moni o ing. Mechanical Sys . and Signal P ocess., 23, 1298-1312.
[77] Zhang, L., Yan, R., Gao, R. X., Lee, K. (2007, Ma ch). Design o a Real- ime
Spindle Heal h Moni o ing and Diagnosis Sys em Based on Open Sys ems A chi ec u e.
In In . Sma Machining Sys . Con ., F ance (pp. 373-378).
[78] Cho, S., Binsaeid, S., As ou , S. (2010). Design o mul isenso usion-based ool
condi ion moni o ing sys em in end milling. In . J. Ad . Manu . Technol., 46, 681-694.
[79] Hi chcock, L. (2006). ISO S anda ds o Condi ion Moni o ing. Enginee ing
Asse Managemen . Sp inge London, pp. 606-613.
[80] Lebold, M., McClin ic, K., Campbell, R., Bying on, C., Mayna d, K. (2000,
May). Re iew o Vib a ion Analysis Me hods o Gea box Diagnos ics and P ognos ics.
In P oc. o he 54 h Mee ing o he Socie y o Machine y Failu e P e en ion
Technol.(Vol. 643, p. 16).
[81] Ra iee, J., Ra iee, M. A., Tse, P. W. (2010). Applica ion o mo he wa ele
unc ions o au oma ic gea and bea ing aul diagnosis. Expe Sys . wi h Appl., 37,
4568-4579.
[82] Fa a , C. R., Du ey, T. A., Doebling, S. W., Nix, D. A. (1999, Sep embe ). A
S a is ical Pa e n Recogni ion Pa adigm o Vib a ion-Based S uc u al Heal h
Moni o ing. S uc u al Heal h Moni o ing, 2000, 764-773.
[83] Tandon, N. (1994). A compa ison o some ib a ion pa ame e s o he condi ion
moni o ing o olling elemen bea ings. Meas., 12, 285-289.
[84] Bouillau , L., Sidahmed, M. (2001).Cyclos a iona y app oach and bilinea
app oach: Compa ison, applica ions o ea ly diagnosis o helicop e gea box and
classi ica ion me hod based on hocs. Mechanical Sys . and Signal P ocess., 15, 923-
943.
[85] Wu, S.-j., Geb aeel, N., Lawley, M. A. & Yih, Y. (2007). A Neu al Ne wo k
In eg a ed Decision Suppo Sys em o Condi ion-Based Op imal P edic i e
Main enance Policy. IEEE T ans. On Sys ., Man and Cybe n.-Pa A: Sys . and
Humans, 37, 226-236.
[86] Alba ba , A., Gu, F., Ball, A. D. (2010). Diesel engine uel injec ion moni o ing
using acous ic measu emen s and independen componen analysis. Meas., 43(10),
1376-1386.
[87] De i S., Si a Kuma , D . L., Shanke N. R., & P abaka an, K. (2010). A
Compa a i e S udy be ween Vib a ion and Acous ic Signals in HTC Cooling Pump and
Chilling Pump. In . J. o Eng. and Technol., 2.
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
72
[88] Al Thobiani, A., Gu, F. , & Ball, A. (2010, June). The moni o ing o ca i a ion
in cen i ugal pumps based on he analysis o ib o-acous ic measu emen s. In CM 2010
and MFPT 2010, UK.
[89] Wang, W. (2008). An Enhanced Diagnos ic Sys em o Gea Sys em
Moni o ing. IEEE T ans. on Sys ., Man, and Cybe n.-Pa B: Cybe n., ol. 38(1), 102-
112.
[90] Todd, M., McA hu , S. D. J., McDonald, J. R., & Shaw, S. J. (2007). A
Semiau oma ic App oach o De i ing Tu bine Gene a o Diagnos ic Knowledge. IEEE
T ans. on Sys ., Man, and Cybe n.-Pa C: Appl. and Re iews, 37(5).
[91] Mu, T., & Nandi, A. K. (2009). Mul iclass Classi ica ion Based on Ex ended
Suppo Vec o Da a Desc ip ion. IEEE T ans. on Sys ., Man, and Cybe n.-Pa B:
Cybe n., 39(5), 1206-1216.
[92] Chen, Z. (2010, May). Faul Diagnosis o Gea Box Based on In o ma ion
En opy. Con ol and Decision Con e ence (pp. 1239-1242).
[93] Kanka , P. K., Sha ma, S. C., Ha sha, S. P. (2011). Faul diagnosis o ball
bea ings using machine lea ning me hods. Expe Sys . wi h Appl., 38, 1876-1886.
[94] Wang, Y. , He, Z., Zi, Y. (2010). Enhancemen o signal denoising and mul iple
aul signa u es de ec ing in o a ing machine y using dual- ee complex wa ele
ans o m. Mechanical Sys . and Signal P ocess., 24, 119-137.
[95] Li, P., He, Q., & Kong, F. (2010, June). An App oach o Faul Diagnosis o
Bea ings Using Wa ele -Based F ac al Analysis. In P oc. o he IEEE In . Con . on
In o ma ion and Au oma ion (pp. 2338-2343). IEEE.
[96] Fu, L., Li, H., Wang, Y. (2009, Ap il). Applica ion o Bi-ceps um Analysis o
Gea Faul De ec ion and Diagnosis. In 2009 In . Con . on Meas. Technol. and
Mecha onics Au oma ion (pp. 590-593).
[97] Ri as, E., Bu gos, J. C., & Ga cía-P ada, J. C. (2010). Vib a ion Analysis Using
En elope Wa ele o De ec ing Faul s in he OLTC Tap Selec o . IEEE T ans. on
Powe Deli e y, 25(3) 3,1629-1636.
[98] Del ecchio, S., D’Elia, G., Mucchi, E., Dalpiaz, G. (2010). Ad anced Signal
P ocessing Tools o he Vib a o y Su eillance o Assembly Faul s in Diesel Engine
Cold Tes s. J. o Vib a ion and Acous ics, 132(2), 021008.
[99] Cheng, J., Yang, Y., Yu, D. (2010). The en elope o de spec um based on
gene alized demodula ion ime– equency analysis and i s applica ion o gea aul
diagnosis. Mechanical Sys . and Signal P ocess., 24, 508-521.
[100] Junsheng, C., Dejie, Y., & Yu, Y.(2007). The applica ion o ene gy ope a o
demodula ion app oach based on EMD in machine y aul diagnosis. Mechanical Sys .
and Signal P ocess., 21, 668-677.
[101] Suguma an, V., Mu alidha an, V., Ramachand an, K.I. (2007). Fea u e selec ion
using Decision T ee and classi ica ion h ough P oximal Suppo Vec o Machine o
aul diagnos ics o olle bea ing. Mechanical Sys . and Signal P ocess., 21, 930-942.
[102] Yang, M., Makis, V. (2010). ARX model-based gea box aul de ec ion and
localiza ion unde a ying load condi ions. J. o Sound and Vib a ion, 329, 5209-5221.
[103] Wang, H., Chen, P., Fujioka, S., Wang, S., Li, K. (2010, Ma ch). Diagnosis
Me hod Based on Skewness Wa e and In o ma ion Di e gence o a Diesel Engine.
In Asia-Paci ic Con . o Powe and Ene gy Eng.( pp. 1-4), Chengdu, China.
[104] An oni, J., Randall, R. B. (2006). The spec al ku osis: applica ion o he
ib a o y su eillance and diagnos ics o o a ing machines. Mechanical Sys . and
Signal P ocess., 20, 308-331.
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 73
[105] Ba szcz, T., Randall, R. B. (2009). Applica ion o spec al ku osis o de ec ion
o a oo h c ack in he plane a y gea o a wind u bine. Mechanical Sys . and Signal
P ocess., 23, 1352-1365.
[106] Qin, Y., Qin, S., & Mao, Y. (2008). Resea ch on i e a ed Hilbe ans o m and
i s applica ion in mechanical aul diagnosis. Mechanical Sys ems and Signal
P ocessing, 22(8), 1967-1980.
[107] Li, B., Zhang, P. L., Liu, D. S., Mi, S. S., Ren, G. Q., & Tian, H. (2011). Fea u e
ex ac ion o olling elemen bea ing aul diagnosis u ilizing gene alized S ans o m
and wo-dimensional non-nega i e ma ix ac o iza ion. Jou nal o Sound and
Vib a ion, 330(10), 2388-2399.
[108] Zhou, R., Bao, W., Li, N., Huang, X., & Yu, D. (2010). Mechanical equipmen
aul diagnosis based on edundan second gene a ion wa ele packe ans o m. Digi al
signal p ocessing, 20(1), 276-288.
[109] Li, M., Xu, J., Yang, J., Yang, D., & Wang, D. (2009). Mul iple mani olds
analysis and i s applica ion o aul diagnosis. Mechanical sys ems and signal
p ocessing, 23(8), 2500-2509.
[110] Wang, Y., He, Z., & Zi, Y. (2009). A demodula ion me hod based on imp o ed
local mean decomposi ion and i s applica ion in ub-impac aul
diagnosis.Measu emen Science and Technology, 20(2), 025704.
[111] Hongkai, J., Zhengjia, H., Chendong, D., & Peng, C. (2006). Gea box aul
diagnosis using adap i e edundan li ing scheme. Mechanical Sys ems and Signal
P ocessing, 20(8), 1992-2006.
[112] Sun, X. L. (2012, Feb ua y). A new me hod o aul diagnosis wi h so se
heo y. In Ad anced Ma e ials Resea ch (Vol. 383, pp. 7470-7474).
[113] Wang, W. (2008). An in elligen sys em o machine y condi ion
moni o ing.Fuzzy Sys ems, IEEE T ansac ions on, 16(1), 110-122.
[114] Wang, D., Pe e , W. T., & Tsui, K. L. (2013). An enhanced Ku og am me hod
o aul diagnosis o olling elemen bea ings. Mechanical Sys ems and Signal
P ocessing, 35(1), 176-199.
[115] Moosa ian, A., Khazaee, M., Ahmadi, H., Khazaee, M., & Naja i, G. (2015).
Faul diagnosis and classi ica ion o wa e pump using adap i e neu o- uzzy in e ence
sys em based on ib a ion signals. S uc u al Heal h Moni o ing,14(5), 402-410.
[116] Tiwa i, R., Gup a, V. K., & Kanka , P. K. (2015). Bea ing aul diagnosis based
on mul i-scale pe mu a ion en opy and adap i e neu o uzzy classi ie . Jou nal o
Vib a ion and Con ol, 21(3), 461-467.
[117] Ali, J. B., Fnaiech, N., Saidi, L., Chebel-Mo ello, B., & Fnaiech, F. (2015).
Applica ion o empi ical mode decomposi ion and a i icial neu al ne wo k o
au oma ic bea ing aul diagnosis based on ib a ion signals. Applied Acous ics, 89, 16-
27.
[118] Mu alidha an, V., Suguma an, V., & Indi a, V. (2014). Faul diagnosis o
monoblock cen i ugal pump using SVM. Enginee ing Science and Technology, an
In e na ional Jou nal, 17(3), 152-157.
[119] Bell, R.N., McWilliams, D.W., O’Donnell, P., Singh, C., & Wells, S. J. (1985).
Repo o la ge mo o eliabili y su ey o indus ial and comme cial ins alla ions Pa I
and II. IEEE T ansac ions on Indus y Applica ions, 21(4), 853-872.
[120] Immo illi, F., Bellini, A., Rubini, R., & Tassoni, C. (2010). Diagnosis o bea ing
aul s in induc ion machines by ib a ion o cu en signals: A c i ical
compa ison. IEEE T ansac ions on Indus y Applica ions, 46(4), 1350-1359.
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 81
Faul s we e in oduced in o he d i e-end 6205-2RS JEM SKF, deep g oo e ball
bea ing using he elec o-discha ge machining me hod. Faul s o diame e 0.007, 0.014
and 0.021 inch (7 mills, 14 mills and 21 mills espec i ely) a e conside ed,
co esponding o 0.01778 cm, 0.03556 cm and 0.05334 cm espec i ely. The deep o
he aul is 0.011 inch/ 0.02794 cm.
Vib a ion da a was collec ed wi h an accele ome e a ached o he housing wi h
magne ic bases placed a he 12 o’clock posi ion a he d i e end. The sample equency
was 12 kHz. Speed and ho sepowe da a we e collec ed using he o que senso /encode
and eco ded by hand. Vib a ion da a was eco ded a ou di e en condi ions: no mal
(N), inne ace aul (IR), ou e ace aul (OR) and ball aul (B). Each signal is 10
seconds. Expe imen s we e epea ed o mo o loads o 0 o 3 ho sepowe (mo o speeds
a ying om 1797 o 1720 RPM. The highe he load is he lowe he speed). The sha
o a ing equency is abou 30 Hz. Da a consis ed o 4 ib a ion signals o N condi ion
and 12 ib a ion signals o each aul condi ion (12 IR, 12 OR and 12 B). In o al he e
a e 40 ib a ion signals.
Figu e 3-4: Tes s and o he acquisi ion o bea ing ib a ion da a.
Sou ce [1].
The Case bea ing ib a ion da abase was used in his Thesis o he e alua ion o
he ollowing p oposals: he p oposal o using Teage -Kaise ene gy ope a o o ex ac

PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
82
s a is ics, ene gy and highe o de s a is ics ea u es in o de o de ec be ween IR, OR,
B and no mal condi ion in bea ing aul diagnosis and he p oposal o a new
me hodology based on wa ele package ans o m and Lempel Zi complexi y o he
assessmen o bea ing se e i y and deg ada ion (see Chap e 4).
UH-60 Blackhawk Helicop e Vib a ion Da abase
The UH-60 Blackhawk helicop e ib a ion da abase was eco ded du ing a componen
endu ance es o an UH-60 Blackhawk helicop e a Pa uxen Ri e , M.D. [2]. Figu e
3-5 shows he main gea box ansmission sys em o he UH-60 helicop e . The gea box
ansmission sys em ansla es he ene gy om he wo engines in o he main o o and
i is a complica ed sys em.
Figu e 3-5: Main ansmission o an UH-60 Backhawk helicop e .
Sou ce: [27].
The da a se s we e eco ded a i egula in e als h oughou he endu ance es s
wi h a sample equency o 100kHz and only eco dings a condi ions wi hin +/- 10% o
ull o que a e used in he expe imen s. Vib a ion signals we e acqui ed using Ende co
6259M31 accele ome e s. The ib a ion da abase consis s o 62 da a se s o 10 seconds
each.
Se e e deg ada ion o he inboa d olle bea ing SB-2205 occu ed du ing he
endu ance es . The helicop e indica o s showed chip ligh s (a chip ligh is an indica o
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 83
o he p esence o me al chips o pa icles in he module) in he expe imen . The i s
chip ligh occu ed in he da a se numbe 40, 10200 minu es a e he eco dings had
begun. The posi ion o bea ing SB-2205, which suppo s he combining be el pinion in
one o he inpu modules, is shown in Figu e 3-5 (le ). As i can be seen om Figu e
3-6 (le ), he bea ing is loca ed deep inside he gea box. In his posi ion, he
backg ound noise is g ea e and he de ec ion o a bea ing aul is mo e di icul . The
bea ing condi ion a he end o he es is also shown in Figu e 3-6 ( igh ). F om he
igu e, i can be seen a aul in he olling elemen o he bea ing.
Figu e 3-6: Posi ion o he bea ing SB-2205 in he ansmission (le
igu e) and he nal bea ing condi ion ( igh igu e). Sou ce: [28].
The UH-60 Blackhawk helicop e ib a ion da abase was used in his Thesis o
he e alua ion o he ollowing p oposals: he p oposal o using Teage -Kaise ene gy
ope a o o ex ac s a is ics, ene gy and highe o de s a is ics ea u es in o de o de ec
be ween IR, OR, B and no mal condi ion in bea ing aul diagnosis and he p oposal o
a new me hodology based on wa ele package ans o m and Lempel Zi complexi y o
he assessmen o bea ing se e i y and deg ada ion (see Chap e 4).
IMS Vib a ion Bea ing Da abase
The IMS ib a ion bea ing da abase is a bea ing da a se p o ided by he Cen e on
In elligen Main enance Sys ems (IMS) [3]. Fou bea ings we e ins alled on one sha
and wo accele ome e s we e placed in each o hem o egis e he ib a ion signals in
wo di e en spa ial axes (see Figu e 3-7). The sha was d i en by an AC mo o and
coupled by ub bel s. The o a ion speed was kep cons an a 2000 pm and a 6000 lb.
adial load was added o he sha and bea ings by a sp ing mechanism. Vib a ion da a
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
84
was collec ed e e y 10 min o 164 h wi h a sampling a e o 20 kHz. A he end o he
es - o- ailu e expe imen , an ou e ace de ec was disco e ed on bea ing 1. Cap u es
ob ained by he ho izon al accele ome e o bea ing 1 ha e been used in his Thesis o
bea ing deg ada ion. Acco ding o [34], he i s indica ion o aul is 89 hou s a e he
beginning o he expe imen .
Figu e 3-7: Posi ion o he bea ings in he IMS bea ing da abase.
Sou ce: [3].
The IMS ib a ion bea ing da abase was used in his Thesis o he e alua ion o
he ollowing p oposal: he p oposal o a new me hodology based on wa ele package
ans o m and Lempel Zi complexi y o he assessmen o bea ing se e i y and
deg ada ion.
3.2 Cen i ugal Pump Audio and Vib a ion Da a
A pump is de ined as a mechanical de ice ha o a es o ecip oca es o mo e luid
om one place o ano he [5]. The e a e mul iples kinds o pumps wi h di e en
applica ions. Acco ding o he Hyd aulic Ins i u e [6] pumps can be classi ied acco ding
o he manne in which he pump adds ene gy o he pumped luid o gene a e
mo emen in o: kine ic pumps and posi i e displacemen pumps. In his Thesis we
ocus on cen i ugal pumps, a kind o kine ic pumps. They add ene gy by high-speed
o a ing wheels o impelle s.
Pumps a e impo an componen s in a wide ange o echnical p ocesses such as
powe s a ions, chemical indus y, cooling and hea ing sys ems, e c. The deg ada ion o
pump componen s such as impelle s, bea ings, seals o he p esence o impu i ies o
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 85
s ange objec s in he luid being pumped can educe he pump pe o mance and lead o
aul s. The o e all eliabili y and sa e y o many sys ems depends on he heal h o
pumps. The e o e, pump condi ion moni o ing plays a key ole in main enance
p ocedu es.
3.2.1 Elemen s and wo king o a cen i ugal pump
A cen i ugal pump consis s o wo main componen s: 1) he o a y elemen o impelle
and 2) he s a iona y elemen o casing (called olu e). The impelle is he o a ing pa
ha con e s d i e ene gy (i.e. he ene gy o a mo o ) in o he kine ic ene gy. Ro a ion
o he impelle o ces he luid (usually liquid) o ci cula e h ough he pump om he
axial o he adial di ec ion while ene gy is ans e ed o he luid [7]. The olu e is he
s a iona y pa ha con e s he kine ic ene gy in o p essu e ene gy. Summing up, he
impelle p oduces luid eloci y and he olu e con e s eloci y o p essu e.
Figu e 3-8 shows he componen s o a cen i ugal pump. The di ec ion o he
luid in a cen i ugal pump (le ) and he zones o eloci y and p essu e o he luid
when passing h ough he cen i ugal pump ( igh ) a e also shown. The luid en e s in
he suc ion eye (a ached o he inle pa o suc ion pa o he pump) o he inle side
(suc ion side), hen pass h ough he impelle and inally exi s in he ou le (discha ge)
side o he pump.
Figu e 3-8: Le : Fluid di ec ion in a cen i ugal pump. Sou ce: [6].
Righ : eloci y and p essu e in a cen i ugal pump. Modi ied om
sou ce: [5]. The inle , ou le , olu e and impelle a e shown.
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
86
An impelle has a numbe o channels o anes delimi ed by cu ed blades. A
closed impelle (see Figu e 3-9) has pla es on bo h sides called hub pla e and sh oud
pla e ha o ally enclose he impelle om he suc ion eye o i s edges. The hub pla e is
in he on o he impelle (whe e he impelle is connec ed o he o o ) and he sh oud
pla e is in he ea o he impelle (in he impelle eye a ea). The impelle also has
se e al blades (also called anes) o impa he cen i ugal o ce o he luid. The cen e
o he impelle is called he impelle eye. The egion nea he impelle eye is called ane
leading edge (LED). The egion a he ip o he ane is called ane ailing edge (TED)
[14].
Figu e 3-9: Closed impelle . Modi ied om sou ce: [6].
A mo e de ailed desc ip ion o how he luid passess h ough a cen i ugal pump
is desc ibed nex [5].
1. Fluid lows h ough he pump by i s en e ing he inle side. Then he luid en e s he
lowes p essu e a ea in he pump, he impelle eye.
2. F om he e he luid is picked up by he spinning impelle anes. The luid passes
along he anes whe e eloci y and ene gy a e added o i . The amoun o ene gy gi en
o he luid is p opo ional o he eloci y a he edge o ane ip o he impelle . The
as e he impelle o a es o he bigge he impelle is, hen he highe will be he
eloci y o he luid a he ane ip and he g ea e he ene gy impa ed o he luid.
3. Th ough cen i ugal o ce, he luid is h own o he ou side ips o he impelle ,
agains he olu e and owa d he discha ge lange. A his poin , because he luid is
con ined by he olu e, he eloci y dec eases he eby inc easing he p essu e. The luid

Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 87
eloci y is dec easing because he olu e is shaped in such a way ha he impelle is no
cen e ed inside i . Ra he , he impelle is o se om he cen e . This o se causes he
impelle o olu e clea ance o inc ease om he cu wa e o he discha ge a ea. As he
clea ance inc eases he eloci y dec eases and he p essu e inc eases.
4. Then he luid mo es h ough he inside edge o he olu e o he discha ge lange
whe e i exi s he pump a a highe p essu e. The eloci y o he luid is con e ed o
p essu e acco ding o Be noulli's p inciple (Be noulli's P inciple s a es ha as he speed
o a mo ing luid inc eases, he p essu e wi hin he luid dec eases).
O he pa s o a cen i ugal pump a e he pump sha , bea ings, sha seal, he
wea ings, and he inle and ou le . The sha seal s ops he luid om leaking ou o he
casing. The wea ings sepa a e he high and low p essu e a eas inside he casing. The
bea ings make he sha u n easie . The inle and ou le pa s o he casing connec o
he luid piping sys em. The sealing de ice is placed inside he s u ing box o con ol o
elimina e leakage om he pump casing.
3.2.2 Vib oacous ic mechanism in a cen i ugal pump
The pump ib o-acous ics is gene a ed by he ollowing sou ces: hyd aulics sou ces and
mechanical sou ces [16], [18]. Bo h sou ces cause ib a ions which make he pump
s uc u e ib a e. This ib a ion adia es ai bo ne sound. The e o e, he acous ics o he
pump has he same mechanism o p oduc ion ha he ib a ion mechanism [18].
Hyd aulics sou ces a e caused by he luid-s uc u e in e ac ion wi h impelle anes
and olu e (especially olu e cu wa e ) and by low pe u ba ions.
Mechanical sou ces a e caused by ib a ion o unbalanced o a ing masses and ic ion
in bea ings, seals, impelle and sha [16].
Du ing he wo king p ocess o a cen i ugal pump non-s eady luid-dynamic
o ces may p oduce ei he disc e e o b oad-band equencies in ib a ion and acous ic
signals [8], [15], [16], [18]. The disc e e equencies a e: he o a ion equency (RF),
he ane-passing equency ( o o equency mul iplied by he numbe o anes o
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
88
blades o he impelle ) and hei ha monics. The o a ion equency is p esen due o
p essu e pulsa ion ( luc ua ions in he p essu e being de eloped by he pump) caused by
impelle imbalance [19] (when he impelle has an o bi al mo ion coupled o he o a ion
[8]) and also by small manu ac u ing impe ec ions in he impelle [8]. The ane-
passing equency (VPF) is due o he ini e hickness o he blades which causes low
dis u bances associa ed wi h he passage o each blade nea he cu wea o olu e
ongue [8]. A p essu e pulse is de eloped as each ane passess he cu wa e .
B oad-band luid-dynamic exci a ion is due o p essu e pulsa ions gene a ed by
low u bulence, iscous o ces, bounda y laye o ex shedding, bounda y laye
in e ac ion be ween a highe - eloci y and lowe eloci y egions o he p ocess luid,
and by o ices gene a ed in he clea ances be ween he o o o he cen i ugal pump
and he adjacen s a iona y pa o he casing [18]. Mechanical sou ces such as noise
om he o a ion o he pump sha and bea ings also con ibu e o b oad-band ib a ion
con en . B oad-band con en is always p esen due o u bulence [8].
O he pe u ba ions no ela ed di ec ly wi h he pump i sel can also exis , such
as an obs acle o obs uc ion in he pump inle , in he pump ou le o e en inside he
impelle [8]. O he equencies gene a ed may be ela ed o an equency o he mo o ,
mo o equencies o equencies ela ed o o he pa o he sys em in which he pump
is ope a ing.
In no mal condi ion, he equencies ha domina e he spec um a e he disc e e
equencies (RF and VPF and hei ha monics). Al hough b oadband noise is always
p esen , in no mal condi ion i has a minimum alue. When some o he componen s o he
cen i ugal pump ha e a aul , o he equency componen s can appea , he p e iously
men ioned equency can inc ease, dec ease in ampli ude o e en can dissapea . The b oadband
noise may also inc ease and domina e he spec a due o he p esence o aul s in he pump (such
as ca i a ion o example).
3.2.3 Main aul s in a cen i ugal pump
The main aul s in pumps a e associa ed wi h impelle damages [9], [10], o o aul s,
seals aul s, ca i a ion [11] and bea ing aul s [11], [12]. A classi ica ion o he main
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 89
aul s in a cen i ugal pump migh be done acco ding o he componen a ec ed by he
aul in o elec ical aul s, mechanical aul s and hyd aulic aul s. The elec ical aul s
a e associa ed wi h he pump mo o , mechanical aul s a e associa ed wi h bea ings and
sha and hyd aulic aul s a e associa ed o hyd aulic pa s o he cen i ugal pump. In
his Thesis, we ha e ocused on hyd aulic aul s. The hyd aulic pa s o he cen i ugal
pump a e he impelle , he blades, he olu e, he inle and he ou le . The unc ionali y
o he hyd aulic pa s is o con e mechanical ene gy om he sha o hyd aulic ene gy
induced in o he liquid pumped by he pump [17].
The hyd aulics aul s a e p oduced in he impelle , he blades and he olu e.
The hyd aulics aul s associa ed wi h hem a e: d y unning, impu i ies ixed on he
impelle (causing inbalance), wea o he impelle (leading edge aul , ailing edge
aul ), blocked o pa ial blocked low ield inside he impelle , blocked impelle
o a ion, wea o he sealing ing, missing sealing ing, loss o he impelle , ca i a ion,
ins abili y, o a ing s all and p essu e pulsa ions [13], [18], [19], [20]. The main e ec s
o hese aul s a e changes in he alue o p essu e and he load o que gene a ed by he
impelle a a gi en low. Mo eo e some o he aul s can induce p essu e oscilla ions.
These p essu e oscilla ions can be ei he ha monics o he o a ional equency, o noise
like signals co e ing a la ge equency span. Ca i a ion may p oduce he wea o
impelle s and piping sys em.
The inle and he ou le pa o he pump and he pump sys em can be conside ed
apa . The aul s in he inle pa o he pump a e low p essu e and obs uc ion a he
inle pump. The main e ec o hese aul s is ha he inle p essu e o he impelle
becomes oo low. The aul s in he ou le pa o he pump include leakage on he ou le
pipe and obs uc ion o he ou le pipe. The main e ec s o hese aul s a e leakages
om he sys em and dec eased p essu e p oduced by he pump.
3.2.4 Conside ed Faul s in a cen i ugal pump in his Thesis
In his Thesis, we ocus on impelle - ela ed aul s (leading edge damage, ailing edge
damage and pla e damage), sys em aul s and seal aul .
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
90
Impelle - ela ed aul s we e a i icially c ea ed cu ing he leading edge o all
anes o he leading edge aul and cu ing he ailing edge o all anes o he ailing
edge aul [14]. The pla e aul was c ea ed emo ing pa o he pla e. The e a e ce ain
low phenomena ha occu in he luid low when i en e s he inle pa o he pump,
pass h ough he impelle and he olu e up o he ou le pa o he pump. The
modi ica ion o he impelle geome y a ec s he low pa e ns inside he cen i ugal
pump. When pla e damage, ailing edge damage o leading edge damage occu low
pa e n changes [13], [14], [19], [20]. These low pa e n changes a ec mainly VPF
and hei ha monics [13]. They can also p oduce b oadband noise due o he o ices
gene a ed in he clea ances be ween he impelle and he olu e [18]. The mechanism o
p oduc ion o low pa e ns inside he pump is no e y well unde s ood ye [19] and
his subjec is ou side he scope o his Thesis.
The sys em aul s conside ed in his Thesis consis in he addi ion o s ange
ob jec s and impu i ies o he luid such as PVC (poly inyl chlo ide) balls, sand, sand
and pape . This can lead o pa ial obs uc ion o he piping sys em, p oducing low low
a e and he gene a ion o u bulence (b oadband) noise [8], [15]. The addi ion o
impu i ies can p oduce deg ada ion o he pump i sel in long e m.
3.2.5 Audio and Vib a ion Da abase acquisi ion om a cen i ugal pump
In o de o s udy audio and ib a ion signals simul aneously, we ha e eco ded ib a ion
and audio samples om a ci cula ing cen i ugal pump in an expe imen al es ig. This
subsec ion desc ibes he acquisi ion p ocess o he da abase and he da abase gene a ed.
A ci cula ing cen i ugal pump is a pump designed o ci cula e a luid h ough a
closed sys em. A closed sys em is one which uns in a loop, wi h he pump discha ge
line e en ually e u ning back o he pump suc ion. The pump wo ks like any
cen i ugal pump, excep hey only need o o e come he ic ion o he piping sys em.
Ci cula ion pumps a e p ima ily used o ci cula ion o wa e in closed sys ems e.g.
hea ing, cooling and ai condi ioning sys ems as well as domes ic ho wa e sys ems and
in applica ions ha equi e chemicals o be egula ly mixed in o he luid, such as pool
and spa pumps.
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 97
be in he wa e . Figu e 3-16 shows he impelle used o LED aul wi h 5mm o ane
leading edge emo ed om all anes.
TABLE 3-8: SEVERITY OF LEADING EDGE FAULTS
Faul Se e i y
Leng h ane educ ion
Leng h o he emaining ane
None
0
50 mm
Sligh
(LED5)
10% o he o al leng h = 5mm
50mm - 5mm = 45 mm
Medium
(LED10)
20% o he o al leng h = 10mm
50mm - 10mm = 40 mm
Se e e
(LED15)
30% o he o al leng h = 15mm
50mm - 15mm = 35 mm
Figu e 3-16: Pho o o he impelle wi h leading edge aul . In his
case, he sligh aul (5mm om all ane leading edges we e emo ed)
is shown.
T ailing Edge Faul
The ailing edge aul (TED) consis s in emo ing pa o he ane ailing edge ( he
egion a he ip o he ane) o all anes. Th ee di e en lengh s we e used. Table 3-9
shows he o iginal leng h o he ane and he milimi e s o ane emo ed. This aul
simula es he deg ada ion o impelle s along ime due o pa icles o impu i ies ha can
be p esen in he wa e . Figu e 3-17 shows he impelle used o TED aul wi h 5mm o
ane ailing edge emo ed om all anes.

PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
98
TABLE 3-9: SEVERITY OF TRAILING EDGE FAULTS
Faul Se e i y
Leng h ane educ ion
Leng h o he emaining ane
None
0
50 mm
Sligh
10% o he o al leng h = 5mm
50mm - 5mm = 45 mm
Medium
20% o he o al leng h = 10mm
50mm - 10mm = 40 mm
Se e e
30% o he o al leng h = 15mm
50mm - 15mm = 35 mm
Figu e 3-17: Pho o o he impelle wi h ailing edge aul . In his case,
he sligh aul (5mm om all ane ailing edges we e emo ed) is
shown.
Seal ubbe aul
The seal aul was p oduced spon aneously while impelle 2 was being eco ded in
no mal condi ion. The sealing ype o he pump is RS-1. The assembly o he impelle
wi h he o o is con igu ed wi h a ce amic "O- ing" ype s a iona y sea and is also
equipped wi h a "se sc ew colla ". The RS-1 ype is an "Elas ome (o Rubbe ) bellows
seal". When we open he pump, we disco e ed ha he ce amic "O- ing" was displaced
du ing he pump ope a ion. This caused he ic ion be ween he O- ing and he impelle
and a sc eech sound was audible. In Figu e 3-18 he impelle wi h he O- ing is shown.
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 99
Figu e 3-18: Pho o o he o- ing o he impelle o he AL P800 pump.
Sys em aul s
The ALP800 pump is designed and buil o pumping wa e , clean and ee om solids.
We ha e added di e en solids o he wa e o cause damage and pa ial obs uc ion o
he pump. This causes a dec ease o wa e low.
The di e en solids added o he wa e a e enume a ed in Table 3-10. The
cha ac e is icas o he solids and hei concen a ion in wa e a e also shown in he able.
TABLE 3-10: CHARACTERISTICS OF SOLIDS ADDED TO THE PUMP
Condi ions
Cha ac e is ics
Concen a ion
Sand
Fine sand om "Las
Alca a ane as" beach
G ain diame e ange:
0.0625-2 mm
2kg/50l
Sand + Pape
S anda d pape A4 80g
2kg/50l (sand), 2kg/50l(pape )
PVC balls (no sand
anymo e)
Poly inyl chlo ide balls
used o ai so pelle s
Diame e : 6 mm
Weigh : 0.12 g
2000 balls / 50 l
(240 g / 50 l)
Figu e 3-19 shows he ese oi wi h sand and Figu e 3-20 shows he PVC balls.
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
100
Figu e 3-19: Pho o o he ese oi wi h he sand added o simula e
s ange objec s o pa icles in he sys em.
Figu e 3-20: Kind o PVC balls (6 mm o diame e ) used o simula e
s ange objec s in he sys em [29].
3.2.5.4 Expe imen al se -up: senso posi ions
The oom whe e he expe imen al se -up is placed is shown in Figu e 3-21. The oom
has a 29 dB o acous ic isola ion espec ae ial noise.
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 101
Figu e 3-21: D awing o he oom o he expe imen al se -up wi h he
measu emen s in mm. A squema ic o he se -up is also shown.
A squema ic o he expe imen se up is also shown in Figu e 3-21, whe e he
pump (in g een), he ese oi wi h wa e (in blue) and he pipes (in o ange) a e shown.
The low di ec ion is also shown. A pho og aph o he expe imen is shown in Figu e
3-22, whe e he ci cula ing pump and he senso s a e shown. The elemen s o he
expe imen al se -up a e he ollowing: a cen i ugal pump, 4 me e s o pipes o 3.81 cm
(1.5 inches) o diame e connec ed o he inle and ou le o he pump, a p essu e gauge
connec ed o he ou pu o he pump, he ese oi and a olley in which he ese oi
and he pump a e placed on. The ese oi is illed wi h 50 li e s o wa e and a closed
loop is o med wi h pipes. The pump is a he same heigh o he ese oi because is a
ci cula ing pump.
Two mic ophones we e placed in he pump inle and in he pump ou le .
Ou le Mic o poin s he pump ou le and Inle Mic o poin s he pump inle (see Figu e
3-22). The mic ophones a e expec ed o ob ain in o ma ion abou he low changes. The
dis ance be ween he mic ophone and he pump used in di e en wo ks in cen i ugal
Flow di ec ion
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
102
pumps a ies om 5cm o 100cm [16],[18],[31],[32]. These wo ks ocus on de ec ion o
ca i a ion. Inia ially, he dis ance o he mic ophones om he pump was a ied om
2cm o 100cm (in s eps o 2cm om 2cm o 20cm and in s eps o 5 cm om 20cm o
100cm) and audio signals we e aken. Thei spec a we e isually inspec ed o
sea ching peaks a he disc e e equencies ( o o equency, ane passing equency and
hei ha monics). The disc e e equencies we e obse ed in all cases bu om 50cm o
100cm he ampli ude o he disc e e equencies anished p og essi ely. Taking in o
accoun ha in indus ial scena ios, he backg ound noise is usually high, mic ophones
should be loca ed nea he sou ce [33]. Fo his eason, mic ophones we e placed a 5
cm om he pump in his Thesis.
Accele ome e s we e placed in wo o hogonal di ec ions o e he pump casing:
accele ome e CESVA AC001 wi h sensi i i y 100 ± 5 % mV/g (‘RadialInle Accel’ in
Figu e 3-22) and accele ome e CESVA AC006 wi h sensi i i y 1000 ± 10 % mV/g
(‘RadialAccel’ in Figu e 3-22). 'RadialInle Accel' is placed on he olu e in he adial
plane ( ha is, he plane pe pendicula o he o o di ec ion) and nea he inle ,
'RadialAccel' is placed o e he olu e in he adial plane and pe pendicula o
“RadialInle Accel”. A hi d accele ome e was placed in axial di ec ion ( he o o
di ec ion). Howe e , his accele ome e is no conside ed in his Thesis. The
accele ome e s a e expec ed o ob ain in o ma ion abou he lows inside he olu e.
Figu e 3-22: Senso Placemen o he Expe imen al se -up.
Ou le Mic o
Inle Mic o
RadialAccel
RadialInle Accel

Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 103
3.2.5.5 Da abase S uc u e
The iles o he da abase a e in .ma o ma . The con en o each ile in he da abase is
desc ibed in Table 3-11. The names o he a iables s o ed in each ile o he da abase
a e enume a ed in he i s column o Table 3-11. In he second column he possible
alues o he a iables a e enume a ed and desc ibed.
The name o he iles in da abase con ains all he in o ma ion o know he pump
condi ion, he impelle used in he eco ding, he se e i y o he aul and he numbe o
he ile. The name o he iles in he da abase has he ollowing o ma :
CON_ImI_Se S_XX.ma
CON is he condi ion. See a iable condi ion in Table 3-11.
I is he numbe o he impelle . See a iable impelle in Table 3-11.
S is deg ee o se e i y in a aul . See a iable se e i y in Table 3-11.
XX is he numbe o he ile.
Fo example, LED_Im3_Se 1_01.ma is a ile wi h LED condi ion in impelle 3
wi h se e i y 1. The numbe o he ile is 01.
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
104
TABLE 3-11: CONTENT OF A SAMPLE FROM THE DATABASE
condi ion
S ing wi h he name o he condi ion. Possible alues a e:
.‘no ’: no mal condi ion
.‘pla’: pla e aul condi ion.
.‘sea’: pla e aul condi ion.
.‘led’: leading edge aul condi ion.
.‘ ed’: ailing edge aul condi ion.
.‘san’: sand in wa e condi ion.
.‘sap’: sand and pape in wa e condi ion.
. ‘p c’: balls o PVC in wa e
impelle
Con ains he numbe o he impelle used in he eco ding. Possible
alues a e:
1: impelle 1
2: impelle 2
3: impelle 3
4: impelle 4
mul isigO i
ginal
Ma ix ha con ains he audio and ib a ion signals o 59 seconds (wi h a
sample equency o 22050 Hz). Each ow o he ma ix co esponds o
he signal acqui ed o a senso .
Senso posi
ion
Type cell. Each cell con ains he name o he senso s in he o de ha a e
s o ed in he ma ix mul isigO iginal.
' adialInle Accel' ' adialAccel' 'ou le Mic o' 'inle Mic o'
The da a o ' a ialInle Accel' is s o ed in he i s ow o he
mul isigO iginal ma ix. The da a o ' adialAccel' is s o ed in he second
ow o he mul isigO iginal ma ix and so on.
se e i y
The se e i y o he aul . Possible alues a e:
0: no se e i y.
1: sligh se e i y.
2: medium se e i y.
3: highe se e i y
No se e i y means ha he ile co esponds o a no mal condi ion o ha
he ile co esponds o a aul wi h no se e i y associa ed ( o example,
seal aul ).
1 can co espond o a pla e aul wi h one pa o pla e emo ed, o a LED
o TED aul o 5mm.
2 can co espond o a pla e aul wi h wo pa s o he pla e emo ed, o a
LED o TED aul o 10 mm.
3 can co espond o a pla e aul wi h h ee pa s o he pla e emo ed, o
a LED o TED aul o 15 mm.
da eacq
S ing wi h he da a o he da a acquisi ion.
Example: ‘21-Jul-2011’
File eco ded 11 July 201.
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 105
3.2.5.6 Numbe s o samples o each condi ion
Table 3-12 shows he di e en condi ions eco ded o ALP800 pump. The i s column
o he able shows he impelle used, he second column shows he pump condi ion and
he hi d column shows he numbe o samples acqui ed o each condi ion. Each
sample has ou signals o 59 seconds. Each sample was ob ained in di e en eco ding
sessions.
TABLE 3-12: RECORDED CONDITIONS FOR EACH IMPELLER
Impelle #
Condi ion
Numbe o samples pe
condi ion
1
No mal
20
1
1 Pla e
20
1
2 Pla es
20
1
3 Pla es
18
2
No mal
20
2
Seal Ring
9
2
Sand
23
2
Sand and Pape
62
2
PVC balls
18
3
No mal
20
3
LED 5 mm
20
3
LED 10 mm
20
3
LED 15 mm
20
4
No mal
20
4
TED 5 mm
20
4
TED 10 mm
20
4
TED 15 mm
20
3.2.6 Pump Da abase P ep ocessing and Baseline signals
Each o iginal ile o he da abase is 60 seconds a a sample equency o 44.1kHz (mo e
han wice he bandwid h o he mic ophones). As he ampli ude in he spec a o he
audio signals d ops a ound 15kHz and he band-wid h o he accele ome e s is up o
10kHz, each ile o he da abase was decima ed by 2. The new sample equency is
22050Hz. To decima e he iles a linea phase FIR ( ini e esponse) il e wi h o de 30
is used. Finally, we emo e he beginning o he signal and we ob ain a signal o 59
seconds.
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
106
In he ollowing igu es, examples o spec a o ib a ion and audio signals om
he cen i ugal pump es ig in no mal ( ee- aul ) condi ion a e shown.
Figu e 3-23: Spec um in ange [2.7-1000]Hz o a 8192 poin s
ib a ion ame (371.5ms) om senso RadialInle Accel. The highe
peaks a e ma ked wi h ed ci cles and he equency alue is shown.
Figu e 3-24: Spec um in ange [1000-11025]Hz o a 8192 poin s
ib a ion ame (371.5ms) om senso RadialInle Accel. The highe
peaks a e ma ked wi h ed ci cles and he equency alue is shown.
0200 400 600 800 1000
0
0.5
1
Radial Inle Accel: No mal Condi ion ([0-1000]Hz)
F equency (Hz)
No malized Ampli ude
97 Hz
339 Hz
976 Hz
876 Hz
196 Hz 683 Hz
46 Hz
2000 4000 6000 8000 10000
0
0.5
1
Radial Inle Accel: No mal Condi ion ([1000-11025]Hz)
F equency (Hz)
No malized Ampli ude
1077 Hz
2719 Hz
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 113
CHAPTER 4
4. Con ibu ions o ib a ion bea ing aul
diagnosis and aul iden i ica ion
This Chap e ocuses on ib a ion ea u es in bea ing aul diagnosis and bea ing
deg ada ion. The con ibu ions o his Thesis on bea ing aul diagnosis a e based on
nonlinea measu es. The nonlinea ene gy ope a o Teage -Kaise is p oposed as a
p ep ocessing ool o bea ing aul diagnosis ollowed by s a is ical and ene gy based
ea u e ex ac ion o diagnose be ween no mal bea ing condi ion, inne ace aul , ou e
ace aul and ball aul condi ions. A new me hodology o assess he se e i y o a aul
in bea ings (i.e.: he p og ession o he aul om an onse aul o a de eloped aul ) o
ou e ace aul and inne ace aul using wa ele package ans o m and complexi y
measu es Lempel-Zi complexi y is also p oposed.
In his Chap e , we explain ou p oposals and he expe imen a ion ca ied ou .
Finally, we show he esul s and he conclusions. The i s pa o he Chap e (sec ion
4.1) is de o ed o he expe imen o he nonlinea ene gy ope a o Teage -Kaise . The
s uc u e is based on pape s published by he au ho o his Thesis [1]-[3]. The second

PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
114
pa o he Chap e is de o ed o he deg ada ion expe imen . The s uc u e is based on a
published con e ence [4] and on an ongoing pape .
4.1 P oposal o nonlinea Teage -Kaise ene gy ope a o o bea ing aul
diagnosis and bea ing deg ada ion assessmen
As s a ed in Chap e 3, ib a ion signals gene a ed by bea ings wi h disc e e aul s
loca ed a he inne ace, a ou e ace o a he olling elemen s can be iewed as an
ampli ude modula ed signal in which he ca ie is he esonance equency o he
bea ing ( he equency exci ed by he impac s o he disc e e aul ) and he undamen al
equency o he modula ing signal ( he en elope) is he bea ing cha ac e is ic
equency o he aul y bea ing. In Figu e 4-1, epea ed he e o con enience, he
ib a ion signals o a bea ing wi h inne ace, ou e ace and ball aul a e shown along
wi h hei co esponding en elope signals.
Figu e 4-1: Time Vib a ion signals o bea ings wi h ou e ace aul ,
inne ace aul and ball aul . Sou ce: [28].
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 115
In ecen decades, much a en ion has ocused on he de elopmen o new digi al
signal analysis me hods o modeling bea ing ib a ion signals so as o disc imina e
be ween no mal ope a ion and di e en aul y condi ions.
Some con en ional me hods, a e sampling he ib a ion signal
)(
wi h sample
equency
s
, ea u e he ib a ion signal









s
n
n )(
wi h s a is ical measu es o i s ,
second o highe o de such as a iance, skewness o ku osis and wi h measu es such
as c es ac o , impulse ac o and oo mean squa e which ea u e he impulsi e na u e
o he bea ing ib a ion signal [5].
O he me hods model he ib a ion signals gene a ed by aul y bea ings as an
ampli ude modula ed (AM) signal, de ined as:
)2cos()2cos()2cos()()( n n An nAn bb b


[Eq. 4-1]
whe e
b
is he esonance equency o he bea ing and
)(nA
he AM signal o signal
en elope whose
b
is he bea ing aul cha ac e is ic equency ha ha e o be de ec ed.
The AM signal can be ex ac ed using he high equency esonance analysis o
en elope analysis [6] which implies band-pass- il e ing he
)(n
bea ing ib a ion
signal. Then, Fou ie ans o m is used o ob ain
b
. The main disad an age o his
echnique is he di icul y in he co ec selec ion o he cen al equency and he
bandwid h o he band-pass il e .
Recen me hods in bea ing aul diagnosis include mo e ad anced signal
p ocessing me hods such as spec al ku osis [7], wa ele analysis and empi ical mode
decomposi ion (EMD) [8]. Spec al ku osis o e s a way o designing op imal band-
pass il e o bea ing aul diagnosis. Wa ele analysis and EMD a e ime- equency
echniques ha decompose he aw ib a ion signal in di e en equency bands. Then,
se e al ea u es such as en opy o s a is ics a e ex ac ed om he di e en equency
bands.
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
116
Recen ly, as an al e na i e o ob ain he AM signal om he aw ib a ion signal,
[6], [10] p oposed he nonlinea Teage -Kaise ene gy ope a o (TKEO) [11]. The AM
signal and he equency modula ed (FM) signal om a mono-componen AM-FM
signal can be ex ac ed using TKEO [12]. H. Li e al. [9] p oposed TKEO o ex ac he
AM signal (AM-TK signal) om bea ing ib a ion signal, wi hou using a band-pass
il e ing p ocess, jus wi h some simple ope a ions o e TKEO. Liang e al. [10]
p oposed he applica ion o he TKEO o e he aw bea ing ib a ion signal wi hou
compu ing he AM signal. Then, he Fou ie ans o m is applied and he
b
is
iden i ied.
In his Thesis, we p opose o use TKEO o ob ain he ib a ion signal in he
Teage –Kaise domain (TK signal) and hen o ea u e i wi h s a is ical and ene gy-
based ea u es. The objec i e is o show how he s a is ical and ene gy-based ea u es
ex ac ed om he TK signal ou pe o m he diagnosis esul s when ex ac ing he same
ea u es om he aw ime ib a ion signal ( ime signal), he AM signal ob ained by
pass-band il e ing (AM signal) and he AM signal ob ained by using he TKEO (TK-
AM signal). A compa a i e analysis be ween s a is ical and ene gy-based ea u es
ex ac ed om TK signal, TK-AM signal, AM signal and ime signal (T signal) is
accomplished.
Nex , he con en ional en elope analysis and he Teage -Kaise ene gy ope a o
a e explained. F om he con en ional en elope analysis he AM signal is ex ac ed. The
TK signal is ex ac ed di ec ly om he Teage -Kaise ope a o and he TK-AM signal
is ex ac ed om he Teage -Kaise ope a o a e doing some calcula ions. The T signal
is he aw ib a ion signal.
Con en ional En elope analysis: AM signal
The con en ional analysis o ob ain he ampli ude demodula ed signal (
)(nA
in [Eq.
4-1]) is called en elope analysis o high equency esonance analysis [6]. This
p ocedu e implies he ollowing s eps: (i) band-pass il e ing he bea ing ib a ion
signal, (ii) applica ion o he Hilbe ans o m o he band-pass il e ed signal, and (iii)
low-pass il e ing he esul ing signal. The s ep (i) equi es a isual inspec ion o he
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 117
bea ing ib a ion spec um o es ima e he cen al equency and he bandwid h o he
band-pass il e .
Teage -Kaise ene gy ope a o : TK-AM signal and TK signal
The Teage -Kaise ene gy ope a o was de i ed by Kaise in 1990 [11] o measu e he
ene gy o he mechanical p ocess ha gene a ed a single ime- a ying signal. TKEO can
de ec modula ions in AM-FM signals by es ima ing he p oduc o hei ime- a ying
ampli ude and equency. I is conside ed as a high- esolu ion ene gy es ima o . The
TKEO o con inuous ime signals ( ) is:
)()()]([)]([ 2
c 

[Eq. 4-2]
whe e
d
d
)(

I can be shown [11] ha he disc e e e sion o he TKEO is:
)1()1()()]([ 2 n n n n

[Eq. 4-3]
Ma agos e al. [12] de eloped a me hod o es ima e he ampli ude en elope (AM
signal) and he ins an aneous equency (FM signal) o speech o man signals using he
TKEO. The ampli ude modula ed signal can be ex ac ed om TKEO (TK-AM signal)
as ollows [12]:
)]1()1([
)]([2
)( 
n n
n
nA


[Eq. 4-4]
This echnique has been applied in bea ing aul diagnosis in [9] o ex ac he
TK-AM signal wi hou using a band-pass il e .
The di ec applica ion o TKEO o e he aw ib a ion signal
)(n
([Eq. 4-3])
[10] can pe o m mo e e ec i e bea ing aul de ec ion. In he TK domain he aul y
samples can be easily disc imina ed because he TK signal highligh s he cha ac e is ic
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
118
impulse ain which is due o he impac o he olle s wi h he de ec . This e ec can be
seen in Figu e 4-2 which shows an example o TK signals o no mal and aul
condi ions. The e o e, i is expec ed ha ea u es ob ained om he TK signal will
disc imina e e ec i ely be ween no mal and aul bea ing condi ions wi hou es ima ing
he AM en elope. The di ec compu a ion o TKEO o e he ib a ion signal has he
ollowing ad an ages: (i) i does no equi e he use o a band-pass il e . The e o e, he
app op ia e es ima ion o he cen al equency and bandwid h o he band-pass il e is
a oided.; (ii) Th ee adjacen samples o he signal a e used o compu e he TKEO. This
ac makes he TKEO implemen a ion e y simple and compu a ionally e icien .
Figu e 4-2: Teage -Kaise ans o med signals (TK signals) o
di e en bea ing condi ions: no mal condi ion (uppe -le ), inne ace
aul condi ion (uppe - igh ), ou e ace aul condi ion (bo om-le )
and ball condi ion (bo om- igh ).
F om he AM signal, he TK-AM signal and om he TK signal a se o
s a is ical, ene gy-based and en opy ea u es a e ex ac ed and a pe o mance
compa ison be ween he di e en me hods a e ca ied ou . The ex ac ed ea u es om
he AM signal, he TK-AM signal and he TK signal a e: ea u es ha cha ac e ize he
ampli ude [1]: peak alue and peak- o-peak alue; s a is ical measu es: s anda d
de ia ion, skewness and ku osis; hi d, ea u es ha quan i y he ene gy in he signal
[1]: oo mean squa e ( ms), squa ed mean oo (sm ); ou h, ea u es ha quan i y he
impulsi e na u e o he ib a ion signal [1]: c es ac o (peak alue/ ms), L ac o (peak
alue/sm ), shape ac o ( ms/peak alue), impulsi e ac o (peak alue/mean) and
inally he Shannon en opy ha measu es he diso de o complexi y o a sys em [14].
Finally, we ha e he ea u ed AM signal, he ea u ed AM-TK signal and he ea u ed
0 0.05 0.1 0.15
0
0.2
0.4
0.6
0.8
Time(s)
Ampli ude
TK signal o N condi ion
0 0.05 0.1 0.15
0
0.5
1TK signal o IR condi ion
Time(s)
Ampli ude
0 0.05 0.1 0.15
0
0.2
0.4
0.6
Time (s)
Ampli ude
TK signal o OR condi ion
0 0.05 0.1 0.15
0
0.2
0.4
0.6
0.8
Time (s)
Ampli ude
TK signal o B condi ion

Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 119
TK signal.
The p oposed me hod is i s applied o he Case Wes e n Da abase o show how
he p oposed me hod ou pe o ms he me hod ha uses s a is ical ea u es in aul
diagnosis, i.e. disc imina ing be ween di e en bea ing aul s: inne ace aul (IR),
ou e ace aul (OR) and ball aul (B). The diag am o he expe imen o aul
diagnosis is shown in Figu e 4-3. Then, he p oposed me hod is applied o he UH-60
da abase. As his da abase is a un- o- ailu e da abase ( aul deg ada ion) he me hod is
applied wi h some di e ences which a e explained in he co esponding sec ion. See
Figu e 4-4 o he diag am o he expe imen o aul e olu ion along ime. In he nex
subsec ion he singula i ies o he applica ion in each da abase a e explained and he
esul s a e shown.
Figu e 4-3: Diag am o he expe imen a ion in bea ing aul diagnosis.
The p oposal is compa ed wi h o he me hods in he s a e o he a
and an e alua ion wi h wo classi ie s is ca ied ou .
Figu e 4-4: Diag am o he expe imen a ion in bea ing aul e olu ion.
The p oposal is applied o a un- o- ailu e bea ing ib a ion da abase.
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
120
4.1.1 E alua ion and esul s o he p oposal o bea ing aul diagnosis
Following he diag am in Figu e 4-3, he p oposed me hods 1 o 4 we e e alua ed in he
Case Vib a ion Bea ing da abase o aul diagnosis. To quan i a i e e alua e he
p oposed me hod, wo classi ie s a e used o e alua e he disc imina ion abili y o he
ea u es be ween no mal condi ion, IR aul , OR aul and B aul : a neu al ne wo k
classi ie and a Leas Squa e-Suppo Vec o Machine (LS-SVM) classi ie . Mo eo e ,
he ea u es ex ac ed om he TK signal a e so ed by ele ance o de wi h he loa ing
o wa d ea u e selec ion p ocedu e. Then, ea u es om he di e en signals a e
e alua ed (in o de o ele ance) wi h he wo men ioned classi ie s. A de ailed
explana ion o he expe imen a ion is ca ied ou in he ollowing pa ag ahps and he
esul s a e shown.
F ame segmen a ion and ea u e ex ac ion
Fi s , he mean is emo ed om he ib a ion signal o each ile o he da abase and
hen he ib a ion signal is no malized be ween -1 and 1. Then, he ib a ion signal is
di ided in o ime ames o 0.17 seconds which a e o e lapped by 33%. Each ame
comp ises 5 e olu ions o he sha . Each ame is ans o med o he TK domain using
[Eq. 4-3]. Then, he men ioned ea u es a e ex ac ed om he TK signal. We call he
ea u es ex ac ed om he TK signal TK ea u es. In o de o compa e he esul s, he
same ea u es a e also ex ac ed om he ime ib a ion signal
)(n
(T ea u es), om
he AM signal by means o en elope analysis (AM ea u es) and om he AM signal by
means o he Teage -Kaise me hod (TK-AM signal).
Rele ance o he TK ea u es
In o de o selec he subse o TK ea u es wi h gi es he bes success a e, he TK
ea u es a e so ed by ele ance o de using he sequen ial loa ing o wa d selec ion
(SFFS) algo i hm [22] and hen a e e alua ed inc emen ally wi h he neu al ne wo k
classi ie and he LS-SVM classi ie . SFFS is an heu is ic algo i hm ha sea ch he
subse o ea u es o he o iginal se which has he bes success a e. The p ocedu e o
so he ea u es by ele ance o de using he SFFS is desc ibed nex .
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 121
The da abase is spli in o a aining subse and a es ing subse by choosing
andomly h ee iles om no mal condi ion and nine iles om each aul condi ion (30
iles in o al) o he aining se and lea ing he emaining iles o he es ing se (10
iles in o al). The da a in he aining and es ing se a e z-sco e no malized using he
mean and a iance o he aining se . The SFFS algo i hm is applied 60 imes wi h
di e en aining and es ing subse s andomly chosen. As a esul , a di e en subse o
bes ea u es is ob ained each ime he SFFS algo i hm is applied. The classi ie used by
he SFFS o ob ain he success a e is a 1-nea es neighbo classi ie . A he beginning
o he expe imen , a 60( imes) x 12 ( ea u es) ma ix is ini ialized a ze o alues. Each
ime he SFFS is applied a subse o bes ea u es is ob ained. Thei posi ion in he
ma ix is ma ked wi h ‘1’. A he end o he 60 epe i ions, each column is summed so
as he numbe o imes a ea u e is wi hin he subse o bes ea u es is ob ained. Finally,
he ea u es a e so ed by ele ance in descending o de : he ea u e ha mos imes
appea s in he subse o bes ea u es is he mos ele an one and he ea u e ha leas
imes appea s in he subse o bes ea u es is he leas ele an one.
Classi ica ion
Once he TK ea u es a e so ed by ele ance o de , hey a e inc emen ally e alua ed
using wo classi ie s: a neu al ne wo k classi ie and a LS-SVM classi ie . The
inc emen al e alua ion o he ea u es consis s in he ollowing: i s , he mos ele an
ea u e is e alua ed, hen he mos ele an is e alua ed along wi h he second mos
ele an and so on. As a esul , he subse o TK ea u es which gi es he bes success
a e can be selec ed. The same p ocedu e is done wi h he T, AM and TK-AM ea u es.
Neu al Ne wo k Classi ie
The ea u es a e ed in o a mul ilaye eed o wa d neu al ne wo k (NN) wi h one
hidden laye ained o disc imina e be ween no mal condi ion, IR aul condi ion, OR
aul condi ion and B aul condi ion. The neu al ne wo k is a s anda d classi ie and can
be used o ep oduce he esul s easily. Supe ised lea ning is ca ied ou using he
esilien back p opaga ion ain algo i hm [23]. The numbe o neu ons in he hidden
laye is selec ed using c oss- alida ion echnique and he ou pu laye has 4 neu ons.
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
122
The NN inpu is he ec o o ea u es belonging o one ame: p
T
R
pp ],[ 1
(T is
ansposi ion o he ec o p) and R is he numbe o ea u es in he ec o . As he
ea u es a e e alua ed inc emen ally, when one ea u e is e alua ed R = 1, when wo
ea u es a e e alua ed R = 2 and when all he ea u es a e e alua ed R = 12. The
ac i a ion unc ions o he hidden laye a e ansigmoid unc ions (hype bolic angen s).
The ou pu laye o he NN has 4 nodes co esponding o he 4 bea ing condi ions: a2
T
aaaa ],,,[ 2
4
2
3
2
2
2
1

, whe e
2
1
a
is he ou pu o no mal condi ion,
2
2
a
is he ou pu o IR
aul condi ion,
2
3
a
is he ou pu o OR aul condi ion and
2
4
a
is he ou pu o B aul
condi ion. The ac i a ion unc ions o he neu ons o he ou pu laye a e linea
ac i a ion unc ions.
In o de o ain and es he NN, he da abase is spli in o a aining subse and a
es ing subse and he da a a e no malized in he same way o he ea u e selec ion
p ocedu e. The aining subse is used o choose he size o he hidden laye (i.e. he
numbe o neu ons o he hidden laye S). The NN is ained wi h a ying S om 1 o
20. Fo each S alue, he aining subse is subdi ided using he 3- old c oss- alida ion
echnique, i.e. he aining subse is di ided in h ee di e en subse s: wo o hem a e
used o ain he neu al ne wo k and he emaining one (called alida ion subse ) is used
o compu e he success a e. The alue o S wi h he bes success a e is chosen as he
size o he hidden laye . Finally, he NN wi h he inal con igu a ion is ained and hen
e alua ed wi h he es ing subse .
In he es ing phase, each ame is classi ied acco ding o he maximum ou pu
alue o he neu al ne wo k. Fo example, i
},,,max{ 2
4
2
3
2
2
2
1
2
1aaaaa 
, he ame is
classi ied as no mal condi ion. Then each ile is classi ied be ween no mal, IR, OR o B
aul condi ions acco ding o he mo e o ed ule, i.e. i he majo i y o he ames a e
classi ied as no mal condi ion, he ile is classi ied as no mal condi ion. All he p ocess
is epea ed 60 imes, andomly choosing di e en iles o aining and es ing se and
he esul s we e a e aged. As a esul , he success a e is ob ained.
The aining p ocess is s opped when any o hese condi ions occu s: he
maximum numbe o i e a ions, se o 5000, is eached; he maximum amoun o ime,
se o in ini e, is exceeded; he pe o mance o E (e o unc ion) is minimized o he
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 129
Figu e 4-8: E olu ion o he ea u es ex ac ed om he aw ib a ion
signal (T signal).
Figu e 4-9: E olu ion o he ea u es ex ac ed om he en elope
signal o he Q band[2500-3800Hz].
4.1.3 Conclusions
In his Thesis we p opose he use o s a is ics, ene gy-based and impulsi e measu es
ex ac ed om he Teage -Kaise signal in bea ing aul diagnosis. The ib a ion signal
is ans o med in o he Teage -Kaise domain using he Teage -Kaise ope a o . A
s anda d and easily ep oducible neu al ne wo k classi ie is used o e alua e he
disc imina ion abili y o he ex ac ed ea u es be ween no mal bea ing, ou e ace aul ,
inne ace aul and ball aul . Mo eo e , a LS-SVM classi ie is also used o alida e
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
Da a Se Index
Ku osis e olu ion ( aw signal)
Skewness e olu ion ( aw signal)
RMS e olu ion ( aw signal)
C es ac o e olu ion ( aw 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
Da a Se Index
RMS e olu ion (Q band en elope )
Skewness e olu ion (Q band en elope )
Ku osis e olu ion (Q band en elope)
C es ac o e olu ion (Q band en elope)

PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
130
he esul s ob ained by he neu al ne wo k. The esul s show ha he use o s a is ical,
based-ene gy and impulsi e ea u es ex ac ed om he Teage -Kaise signal in bea ing
aul diagnosis ou pe o ms he esul s ob ained by he same ea u es ex ac ed om he
ime ib a ion signal, om he ime en elope signal (compu ed using TKEO) and om
he ime en elope compu ed using band-pass il e . In addi ion, he use o he Teage -
Kaise signal a oids isual inspec ion o he bea ing spec um o de e mine he cen e
equency and he bandwid h o he band-pass il e . The compu a ion o he ex ac ed
ea u es and o he TKEO i sel is also simple and as e han band-pass il e ing.
The p oposed me hod is also applied o a un- o- ailu e bea ing ib a ion da abase
ob ained om an endu ance es o a Black Hawk Helicop e . The esul s show ha oo
mean squa e and ku osis ea u es ex ac ed om he signal ans o med o he Teage -
Kaise domain a e good indica o o he bea ing deg ada ion.
4.2 P oposal o Lempel-Zi complexi y measu e based on wa ele package
ans o m o bea ing deg ada ion assessmen
The aim o bea ing aul diagnosis echniques is o de ec he kind o aul p esen ed in a
bea ing (i.e. ou e - ace aul , inne - ace aul o ball aul in he case o disc e e de ec s).
Once he kind o bea ing aul is diagnosed (i.e. inne ace aul , ou e ace aul o ball
aul ), i is impo an o assess he se e i y o he aul ( aul iden i ica ion), i.e. how
la ge he aul is. Ano he impo an ask in condi ion moni o ing is he ea ly de ec ion
o a aul ( o ollow he deg ada ion o a componen om no mal condi ion o aul
condi ion).
In his Thesis, we ha e de eloped a new me hodology o assess he se e i y o
bea ings wi h ou e ace aul and inne ace aul wi h di e en deg ees o se e i y
(depic ed in Figu e 4-10). The p oposed me hod combines he wa ele package
ans o m echnique and he Lempel-Zi complexi y o assess he aul se e i y o
bea ings wi h ou e ace and inne ace aul . We e alua e he p oposed me hod in he
Case Wes e n Vib a ion Bea ing da abase, he UH-60 helicop e da abase and in he
IMS bea ing da abase. Wa ele packe ans o m is a ime- equency echnique ha
basically decomposes he signal in o de ail (low equency componen s) and
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 131
app oxima ions signals (high equency componen s) in di e en le els o
decomposi ion o scales. Wa ele packe ans o m is used as a p ep ocessing ool o
selec he node wi h highe ene gy (i.e. he node whe e he aul lies). Then, he Lempel-
Zi complexi y is used o measu e he se e i y o he aul . As he Lempel-Zi
complexi y alues a e bounded be ween 0 and 1, i is a good indica o o he se e i y o
he bea ing aul .
In his sec ion, a desc ip ion o he me hods o aul se e i y assessmen in
bea ings li e a u e is ca ied ou . Nex , he p ep ocessing ool ( he wa ele packe
ans o m) and he Lempel-Zi complexi y a e explained. Then, he p oposed
me hodology is shown and inally a se o expe imen s in bea ings is ca ied ou o show
he use ulness o ou p oposal and a compa ison wi h some me hods in li e a u e is also
accomplished.
S a e o he a
The se e i y assessmen in bea ings in li e a u e is mainly add essed wi h ime-domain
ea u es including ku osis, oo mean squa e, c es ac o and ampli udes [20], [21].
Ku osis inc eases in single-poin aul y bea ings om a alue o 3 (in he case o ee-
aul bea ing) o highe alues when he bea ing has a single-poin de ec . Howe e ,
when he single-poin de ec is in ad anced s ages ku osis alues e u n o 3 [20]. Roo
mean squa e, c es ac o , ampli udes a e o he echniques o ollow he deg ada ion o
he aul . Howe e , hey ha e he same p oblems ha ku osis. Thei alues a e no
consis en wi h he e olu ion o he aul . The e a e o he echniques in equency
domain such as ene gy ea u es ex ac ed om en elope analysis. The main
disad an age o hese ea u es a e ha he y a e no bounded.
Lempel-Zi complexi y was i s ly in es iga ed in bea ing se e i y assessmen
Lempel-Zi complexi y in [15]. The au ho s p oposed LZ o ou e - ace bea ing aul
de e io a ion. They ex ac ed he LZ alues om he aw ib a ion signal. They showed
ha he LZ alue inc eases as he size o he ou e ace bea ing aul inc eases. H. Hong
e al. [16] p oposed a new e sion o he Lempel-Zi complexi y as a bea ing aul
se e i y measu e based on he con inuous wa ele ans o m (CWT). The applica ions
o he bea ing inne - and ou e - ace aul signals ha e demons a ed ha he new e sion
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
132
o Lempel–Zi complexi y can e ec i ely measu e he se e i y o bo h inne - and
ou e - ace aul s. This echnique uses he CWT o ob ain he sub-band whe e he aul
lies. Then, he signal is ampli ude demodula ed and he en elope signal and he ca ie
signal a e ob ained. Finally, he Lempel-Zi complexi y is compu ed o he en elope
signal and o he ca ie signal and a weigh ed summed is compu ed. They use di e en
weigh s depending on he aul (IR aul o OR aul ). In his Thesis, we p opose a
simple way o assess he se e i y in bea ing using wa ele package ans o m o ob ain
he sub-band whe e he aul lies and hen applying Lempel-Zi complexi y o his sub-
band.
Me hods
Wa ele Packe T ans o m
Wa ele packe ans o m (WPT) is an ex ension o he disc e e wa ele ha allows a
ine esolu ion o equencies a bo h high equencies and low equencies. WPT can
simul aneously b eak he signal up in o de ail (low equency componen s) and
app oxima ions signals (high equency componen s) in di e en le els o
decomposi ion o scales. In bea ing aul diagnosis he aul in o ma ion lies in high
equency. Fo his eason, he wa ele packe ans o m is a common analysis ool o
bea ing aul diagnosis.
Ma hema ically, wa ele packe s a e a collec ion o unc ions
},,),2(2{ 2/ ZkjNnk W j
n
j

gene a ed ecu si ely om he ollowing
exp essions:

mnn m Wmh W )2()(2)(
2
[Eq. 4-5]


mnn m Wmg W )2()(2)(
12
[Eq. 4-6]
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 133
whe e h and g a e he quad a u e mi o il e s,
)(
0 W
is he scaling unc ion and
)(
1 W
is he wa ele mo he unc ion, k is a ime-localiza ion pa ame e , j is a scale pa ame e
(dep h) and n is an oscilla ion pa ame e .
The wa ele packe coe icien s o a disc e e signal can be ob ained using he
ollowing i e a i e exp essions:
120,)()2()2()()( ,,2,1  
j
lnjnjnj nlcklhkhkckc
[Eq. 4-7]
120,)()2()2()()( ,,12,1  
 j
lnjnjnj nlcklgkgkckc
[Eq. 4-8]
The o iginal signal can be econs uc ed using he ollowing i e a i e exp ession:
120,)2()2()( 12,1,2,1,    j
k
knj
k
knjnj ncklgcklhlc
[Eq. 4-9]
Wa ele packe s a e o ganized in ees, whe e he scale j de ines he dep h o le el
o he ee and n is he posi ion o node in he ee. Fo each scale j,
12,,0  j
n
.
F om each node o he ee, wo child en nodes a e ob ained. One o he child en node is
ob ained by low-pass il e ing (using he impulse esponse h) ollowed by down-
sampling and he o he child en node is ob ained by high-pass il e ing (using he
impulse esponse g) ollowed by down-sampling. I is impo an o men ion ha he
equency o de o he nodes o a le el is no necessa y he same as he node o de
because o aliasing in oduced by down-sampling. Fo example, a signal wi h 1kHz
bandwid h decomposed using a wa ele ee o decomposi ion le el (scale) j = 3 ha e
eigh nodes a le el 3 (
7,,0 n
). The equency con en o each node (in na u al
o de , i.e. ) is he ollowing: 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). The e o e, he equency o de is no he same as
node na u al o de (also called Paley o de ). To ob ain he node in equency o de
( om low equency o high equency), he nodes need o be eo de ed.
7,,0 n
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
134
Lempel-Zi complexi y
Lempel and Zi p oposed a complexi y measu e ha can cha ac e ize he deg ee o
o de o diso de and de elopmen o spa io empo al pa e ns in a ime se ies [18], [19].
The signal is ans o med in o bina y sequences and Lempel-Zi algo i hm gi es he
numbe o dis inc pa e ns con ained in he gi en ini e sequence. A e no maliza ion,
he ela i e Lempel-Zi complexi y measu e (LZC) e lec s he a e o new pa e n
occu ences in he sequence. LZC alues ange om nea 0 (de e minis ic sequence) o
1 ( andom sequence). To compu e he LZC o a gi en sequence s, he sequence has o
be coded using a ini e symbol se A. Using he s anda d coding squeme, he sequence o
be coded s is di ided in α equip obables bins, whe e α is he numbe o di e en
symbols in A. An elemen o he symbol se A is assigned o each alue o he signal
acco ding o he bin in which he alue lies. Finally, a new sequence o symbols S is
ob ained. This sequence S is subsequen ly scanned looking o o iginal subsequences o
di e en leng hs. The complexi y alue is ela ed wi h he numbe o di e en
subsequences ound. Thus, i A* deno es he se o all ini e leng h sequences o e he
ini e symbol se A, and
)(Sl
deno es he leng h o a sequence
*
AS
, hen acco ding o
[19].
0},)({ * nnSlASAn
[Eq. 4-10]
and o e e y
n
AS 
, he Lempel-Zi complexi y can be exp essed as
)(log)1(
)( n
n
nc
an



[Eq. 4-11]
whe e
0
n

i
n
and

is he numbe o di e en symbols in he symbol se A.
The uppe bound o
)(nc
is
)(log
)()(lim n
n
nbnc
a
n

[Eq. 4-12]
The e o e,
)(nc
can be no malized by using his uppe limi

Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 135
1
)(
)(
)(0  nb
nc
nC
[Eq. 4-13]
Fo
2

wo symbols a e used o code he signal (0 and 1). Fo
3

he e
exi 3 symbols (0, 1 and 2). The LZC alue depends bo h on he numbe o symbols
used in he coding squeme and on he leng h o he sequence s. Acco ding o [16] leng h
o sequence highe han 3600 is enough.
P oposal
We p opose a new me hod o aul se e i y assessmen in bea ings (depic ed in Figu e
4-10). The p oposed me hod combines he wa ele packe ans o m and he Lempel-
Zi complexi y o assess he aul se e i y o bea ings wi h ou e ace and inne ace.
As explained be o e in his Chap e , he LZC o e he aw ib a ion signal has been
used in bea ing li e a u e o ou e - ace aul se e i y assessmen [15]. Howe e , his
echnique is a ec ed by noise. We show ha ou p oposal is less a ec ed by noise.
Wa ele packe
decomposi ion
(3 d le el, 8 nodes)
Compu e Rela i e
Ene gy o each Node
Selec Node wi h
maximum ene gy
Inpu ib a ion signal
Ou pu : ex ac ed
ea u e o he inpu
signal
Compu e Lempel-Zi
complexi y
Recons uc he signal
o he node kmax
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
136
Figu e 4-10: P oposed me hodology o bea ing se e i y aul
assessmen .
Fi s , he aw ib a ion signal is decomposed using he wa ele packe ans o m
wi h a le el 3 o decomposi ion (8 nodes). Daubechies 6 (db6) (see Figu e 4-11) was
used as he mo he wa ele because has an impulsi e shape acco ding o he impulsi e
aul bea ing cha ac e is ics (see igu e 4-12). The WPT is used o ob ain he node
whe e he aul lies and emo e unwan ed signals. Ini ially, mo he wa ele s db2 o db5
and db7 o db9 and symle wa ele mo he amiliy om o de 2 o 9 we e used. The
esul s ob ained a e e y simila be ween wa ele mo he s wi h o de s highe han 3.
Figu e 4-11: Wa e o m o he Daubechies 6 wa ele mo he .
The ib a ion signals a e b oken up o 3 le el (
3j
) o decomposi ion,
ob aining 8 equency bands (o 8 nodes) in his le el o decomposi ion. Signals a e
econs uc ed om he wa ele coe icien s associa ed o each node. The econs uc ed
signal o each node is a new ime se ies. I
x
deno es he o iginal signal ( he ib a ion
signal), hen
nj
x,
is he econs uc ed signals o j h le el o decomposi ion and n h
deno es he equency-band o he signal. Fo
3j
, he e a e 8 equency-band signals
in le el 3.
Then, he ela i e ene gy o he econs uc ed signals o each node is compu ed,
i.e. he ene gy o each node di ided by he ene gy o he o iginal aw signal. Then, he
node wi h he maximum ene gy is selec ed. Rela i e ene gy ea u es a e compu ed om
he econs uc ed signals in le el 3 o decomposi ion (
3j
). E deno es he ene gy o
he o iginal aw ib a ion signal
)(nx
:
0 2 4 6 8 10 12
-1.5
-1
-0.5
0
0.5
1
1.5
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 137
1
))((
1
2




N
xix
E
N
i
[Eq. 4-14]
whe e
x
is he mean alue o
)(nx
and N is he numbe o samples o
)(nx
. Then, he
ela i e ene gy o he econs uc ed signals in le el 3 is compu ed as:
E
M
xix
e
N
i
n
n
n1
))((
1
2
,3
,3





[Eq. 4-15]
whe e
n
x,3
(n = 0,...,7) a e each o he econs uc ed signals o le el 3, M is he numbe
o samples o
n
x,3
and
n
x,3
is he mean alue o each
n
x,3
.
The node wi h he maximum ela i e ene gy co esponds o he equency band
ha con ains he mos aul ea u es. A bea ing wi h a localized aul will ha e he
maximum ela i e ene gy loca ed in high equencies because he bea ing is modelled as
an ampli ude modula ed signal wi h he equency o he ca ie being he esonance
equency o he sys em. The e o e, ea u es ex ac ed om he selec ed node will ha e
in o ma ion abou he aul .The node wi h he maximum ela i e ene gy is selec ed and
he LZC is ex ac ed om he econs uc ed signal o his node.
The LZC gi es in o ma ion abou he complexi y o he econs uc ed signal.
This is he diagnos ic ea u e ec o ex ac ed om he inpu ame. This p ocedu e is
applied o each ame and inally a ea u e ec o is ob ained.
The p oposed me hodology is applied o he Case Wes e n Bea ing Vib a ion
Da abase, o he UH-60 helicop e da abase and o he IMS da abase. In o de o
compa e wi h classical me hod in he li e a u e, we ex ac ku osis, Lempel-Zi
complexi y om aw ib a ion da a and ku osis om wa ele package ans o m.
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
138
4.2.1 E alua ion wi h he Case Bea ing Da abase
The p oposed me hodology is applied o he Case Bea ing Da abase. Each sample o he
da abase is di ided in ames o 4096 samples each and o each ame he p ocedu e
shown in igu e 4-11 is applied. As he sample equency is
12000
s
Hz, he
bandwi d h o he signal is
6000
N
Hz. The e o e,
x
has a equency in e al
(0,6000]Hz,
0,3
x
is he econs uc ed signal in le el 3 wi h a equency ange o (0,
750]Hz and
7,3
x
is he econs uc ed signal in le el 3 wi h a equency ange o (5250,
6000]Hz. The nodes a e eo de ed in equency o de .
In Figu e 4-12-Figu e 4-16 he ela i e ene gies o each node o he WPT (in
le el 3) o a ame is shown o he no mal condi ion and o aul condi ions IR, OR
and B wi h di e en se e i ies and di e en loads. In he igu es a i le is added o each
sub igu e: IRXXYY means inne ace aul wi h XX mils o diame e and load YY. XX
can be '07' o a se e i y o 7 mils in diame e , '14' o a se e i y o 14 mils in diame e
o '21' o a se e i y o 21 mills in diame e . YY can be '00' o load 0, '01' o load 1,
'02' o load 2 o '03' o load 3. The nodes o he wa ele packe we e ob ained in
na u al o de (Paley o de ) and hen we e eo de ed in equency o de .
We can obse e om he igu es ha he ene gy is concen a ed in high
equency bands, whe e he aul cha ac e is ic lies. In he case o no mal condi ion, he
ene gy is concen a ed in low equency bands.
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 145
Figu e 4-20: E olu ion o he mean alues ob ained o each me hod:
p oposed me hod ‘LZC (EmaxWPT)’ (uppe-le ), ‘LZC ( aw signal)’
(uppe - igh ), ‘Ku osis (EmaxWPT)’ (bo om-le ) and ‘Ku osis ( aw
signal)’ (bo om- igh ) o inne ace aul wi h di e en se e i ies and
di e en loads. The no mal condi ion is also conside ed.
The obus ness o he algo i hm o Gaussian noise is also s udied. Gaussian noise
was added o he o iginal ib a ion signal wi h di e en signal o noise a ios (SNR):
20dB, 15dB, 10dB, 5dB and 0dB. Then, he p oposed me hod is applied (wi h α = 4) o
each SNR and o he o iginal signal wi hou noise added. Ku osis o he aw signal,
ku osis o he maximum ene gy node and he LZC o he aw signal a e also compu ed
o each SNR and o he o iginal signal. The mo i a ion o his expe imen is o show
how he p oposed me hod is mo e obu s o noise han compu ing he LZC om he aw
ib a ion signal. The LZC measu es he complexi y in a signal. Fo andom signals, he
LZC is 1. The e o e, i he signals a e con amina ed wi h undesi ed Gaussian noise, he
LZC will inc ease i s alue. I LZC is ex ac ed om he aw ib a ion signal, i will be
di icul o de e mine i he inc easing o dec easing o he LZC is due o a aul o o
noise. Howe e , i he LZC is ex ac ed om he node o he maximal ene gy associa ed
o impulsi e aul , he Gaussian noise will no a ec in he same deg ee he signal. The
ollowing igu es (Figu e 4-21-Figu e 4-28) show he esul s o each me hod and
con i m ou hough s.
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 ( aw signal)
N IR07 IR14 IR21
0
20
35
Ku osis (Emax WPT)
N IR07 IR14 IR21
0
10
25
Ku osis ( aw signal)

PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
146
Figu e 4-21: E olu ion o he mean alues ob ained o ‘Ku osis ( aw
signal)’ me hod applied o ou e ace aul condi ion and no mal
condi ion a ying he amoun o noise added o he o iginal ib a ion
signal. The esul s a e shown o load: load 0 (uppe -le ), load
1(uppe - igh ), load 2 (bo om-le ), load 3 (bo om- igh ).
Figu e 4-22: E olu ion o he mean alues ob ained o ‘LZC ( aw
signal)’ me hod applied o ou e ace aul condi ion and no mal
condi ion a ying he amoun o noise added o he o iginal ib a ion
signal. The esul s a e shown o load: load 0 (uppe -le ), load
1(uppe - igh ), load 2 (bo om-le ) and load 3 (bo om- igh ).
N OR07 OR14 OR21
0
10
20
30
O iginal 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
Ku osis ( aw signal)
load 0 load 1
load 2 load 3
SNR
N OR07 OR14 OR21
0.4
0.6
0.8
1
O iginal 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 ( aw signal)
load 0 load 1
load 2 load 3
SNR
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 147
Figu e 4-23: E olu ion o he mean alues ob ained o ‘Ku osis
(EmaxWPT)’ me hod applied o ou e ace aul condi ion and
no mal condi ion a ying he amoun o noise added o he o iginal
ib a ion signal. The esul s a e shown o load: load 0 (uppe -le ),
load 1(uppe - igh ), load 2 (bo om-le ) and load 3 (bo om- igh ).
Figu e 4-24: E olu ion o he mean alues ob ained o he p oposed
me hod ‘LZC (EmaxWPT)’ applied o ou e ace aul condi ion and
no mal condi ion a ying he amoun o noise added o he o iginal
ib a ion signal. The esul s a e shown o load: load 0 (uppe -le ),
load 1(uppe - igh ), load 2 (bo om-le ) and load 3 (bo om- igh ).
N OR07 OR14 OR21
0
20
40
O iginal 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
Ku osis (Emax WPT)
SNR
load 0 load 1
load 2 load 3
N OR07 OR14 OR21
0,4
0,5
0,6
O iginal 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 Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
148
Figu e 4-25: E olu ion o he mean alues ob ained o ‘Ku osis ( aw
signal)’ me hod applied o inne ace aul condi ion and no mal
condi ion a ying he amoun o noise added o he o iginal ib a ion
signal. The esul s a e shown o load: load 0 (uppe -le ), load
1(uppe - igh ), load 2 (bo om-le ) and load 3 (bo om- igh ).
Figu e 4-26: E olu ion o he mean alues ob ained o ‘LZC ( aw
signal)’ me hod applied o inne ace aul condi ion and no mal
condi ion a ying he amoun o noise added o he o iginal ib a ion
signal. The esul s a e shown o load: load 0 (uppe -le ), load
1(uppe - igh ), load 2 (bo om-le ) and load 3 (bo om- igh ).
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
Ku osis ( aw signal)
O iginal 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 ( aw signal)
O iginal 20dB 15dB 10dB 5dB 0dB
SNR
load 0 load 1
load 2 load 3
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 149
Figu e 4-27: E olu ion o he mean alues ob ained o ‘Ku osis
(EmaxWPT)’ me hod applied o inne ace aul condi ion and
no mal condi ion a ying he amoun o noise added o he o iginal
ib a ion signal. The esul s a e shown o load: load 0 (uppe -le ),
load 1(uppe - igh ), load 2 (bo om-le ) and load 3 (bo om- igh ).
Figu e 4-28: E olu ion o he mean alues ob ained o he p oposed
me hod ‘LZC (EmaxWPT)’ applied o inne ace aul condi ion and
no mal condi ion a ying he amoun o noise added o he o iginal
ib a ion signal. The esul s a e shown o load: load 0 (uppe -le ),
load 1(uppe - igh ), load 2 (bo om-le ) and load 3 (bo om- igh ).
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
Ku osis (Emax WPT)
O iginal 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)
O iginal 20dB 15dB 10dB 5dB 0dB
load 3
load 2
load 1
load 0
SNR
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
150
4.2.2 E alua ion wi h he UH-60 Helicop e Da abase o bea ing deg ada ion
assessmen
The p oposed me hod o he Lempel-Zi complexi y based on wa ele packe ans o m
is also e alua ed wi h UH-60 Black Hawk Helicop e da abase whe e a olling elemen
o he bea ing was damaged du ing an endu ance es . Al hough ou p oposal does no
ollow mono onically a aul in a olling elemen , acco ding o he esul s wi h he Case
Wes e n da abase i should de ec a ball aul . Fo his eason, we apply he p oposal o
he UH-60 helicop e da abase.
As each da ase in he UH-60 da abase is 10 seconds (wi h a sample equency
o 100kHz) and he compu a ion o he LZC is e y slow wi h such amoun o da a,
each da ase is di ided in ames o 1 second each. The esul s a e a e aged pe da ase .
The esul s a e shown in Figu e 4-29 o ‘Ku osis ( aw signal)’, ‘LZC ( aw
signal)’, ‘Ku osis (EmaxWPT)’ and he p oposed me hod ‘LZC (EmaxWPT)’. As i is
obse ed, none o he ea u es can de ec he aul . The ‘LZC ( aw signal)’ and ou
p oposal ‘LZC (EmaxWPT)’ shows an inc ease in hei alues in da ase 48, long a e
he aul .
Figu e 4-29: E olu ion o he ea u es using ku osis and Lempel-Zi
complexi y om he aw signal and ku osis and Lempel-Zi
complexi y om he node wi h maximal ene gy o he wa ele pake
ans o m.
010 20 30 40 50 60
2.6
2.8
3Ku osis ( aw signal)
010 20 30 40 50 60
0.35
0.4
0.45 LZC ( aw signal)
010 20 30 40 50 60
2.6
2.8
3Ku osis (EmaxWPT)
010 20 30 40 50 60
0.25
0.3
0.35 LZC (EmaxWPT)

Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 151
4.2.3 E alua ion wi h he IMS da abase
The IMS da abase is a un- o- ailu e expe imen o 164 hou s in whe e an ou e ace
aul is de eloped in a bea ing. P e ious pape indica es ha he s a o he aul is 89
hou s a e he beginning o he expe imen [29].
In his case, he p oposed me hod is applied o each sample o he da abase (each
sample has 20000 da a poin s co esponding o 1 second o signal). As he sample
equency is
20000
s
Hz, he bandwid h o he signal is
10000
N
Hz. The e o e,
x
has a equency in e al (0,10000]Hz,
0,3
x
is he econs uc ed signal in le el 3 wi h a
equency ange o (0, 1250]Hz and
7,3
x
is he econs uc ed signal in le el 3 wi h a
equency ange o (8750, 10000]Hz. Ku osis om he aw signal, ku osis om he
node o maximal ene gy o he WPT and he LZC om he aw signal a e also ex ac ed
and he esul s compa ed.
Figu e 4-30, Figu e 4-31, Figu e 4-32 and Figu e 4-33 show he e olu ion o he
ea u es ex ac ed by he ‘Ku osis ( aw signal)’ me hod, by he ‘Ku osis (EmaxWPT)’
me hod, by he ‘LZC ( aw signal)’ me hod and by he p oposed me hod ‘LZC
(EmaxWPT)’ espec i ely. The igu es e eal ha he ku osis ex ac ed om he aw
ib a ion signal inc ease and hen dec ease again. Mo eo e , he i s indica ion o aul
is a 117 hou s om he beginning o he un- o- ailu e expe imen . The ku osis
ex ac ed wi h he me hod Ku osisEmaxWPT shows a mo e consis en endence ha
he ku osis ex ac ed om he aw ib a ion signal and he i s indica ion o aul is a
89 hou s om he beginning o he expe imen . The LZC aw shows alues nea 1 in he
no mal condi ion. The eason can be he noise and signals o he es o he gea ha
in e e e wi h he bea ing signal. 89 hou s a e he beginning o he expe imen , he
LZC alues s a dec easing. Then, he LZC alues change wi h no clea endency.
When he LZCEmaxWPT me hod (p oposed me hod) is applied, he LZC inc ease hei
alues 89 hou s a e he beginning o he expe imen (when he aul s a s de eloping)
and emains a ound he same alues o complexi y un il he end o he expe imen .
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
152
Figu e 4-30: E olu ion o he Ku osis ex ac ed om he aw
ib a ion signal o a un- o- ailu e expe imen o he bea ing ha
e en ually de eloped an ou e - ace aul .
Figu e 4-31: E olu ion o he Ku osis ex ac ed om he node o
maximal ene gy o he WPT o a un- o- ailu e expe imen o he
bea ing ha e en ually de eloped an ou e - ace aul .
0 20 40 60 80 100 120 140 160 180
0
5
10
15
20
Time (hou s)
Ku osis ( aw signal)
0 20 40 60 80 100 120 140 160 180
0
2
4
6
8
10
12
Time (hou s)
Ku osis (Emax WPT)
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 153
Figu e 4-32: E olu ion o he Lempel-Zi complexi y ex ac ed om
he aw ib a ion signal o a un- o- ailu e expe imen o he bea ing
ha e en ually de eloped an ou e - ace aul .
Figu e 4-33: E olu ion o he Lempel-Zi complexi y ex ac ed om
he node o maximal ene gy o he WPT o a un- o- ailu e expe imen
o he bea ing ha e en ually de eloped an ou e - ace aul .
4.2.4 Conclusions
We ha e p oposed a me hod o assess he aul se e i y, also called aul iden i ica ion
(how la ge he aul is), using he wa ele packe ans o m and he Lempel-Zi
complexi y o inne ace aul and o ou e ace aul . In he case o inne - ace aul ,
he alues o Lempel-Zi complexi y dec ease when he aul e ol es and in he case o
he ou e - ace aul he alues o Lempel-Zi complexi y inc ease when he aul
e ol es. The esul s ob ained wi h he p oposed me hod a e in conco dance wi h he
0 20 40 60 80 100 120 140 160 180
0.7
0.75
0.8
0.85
0.9
0.95
1
Time (hou s)
LZC ( aw signal)
0 20 40 60 80 100 120 140 160 180
0.2
0.3
0.4
0.5
Time (hou s)
LZC (Emax WPT)
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
154
esul s in li e a u e [15], [16]. In [15] he Lempel-Zi complexi y was applied o aw
bea ing ib a ion signals o no mal and ou e - ace aul condi ions. The Lempel-Zi
complexi y inc eases om no mal o ou e - ace aul wi h high le el o se e i y.
Howe e , as we ha e shown in ou expe imen s, when he aw bea ing ib a ion signal
is con amina ed wi h Gaussian noise, he Lempel-Zi complexi y is a ec ed, leading o
inco ec in e p e a ion o he Lempel-Zi complexi y alues. In [16], he au ho s
p opose he use o he Lempel-Zi complexi y and he con inuous wa ele ans o m in
ou e - ace aul and inne - ace aul se e i y assessmen . A me hod o selec he bes
scale in he con inuous wa ele ans o m based on ene gy and ku osis was p oposed.
Then, he coe icien s in he bes scale a e eco e ed and ampli ude demodula ed. The
inal alue o he Lempel-Zi complexi y is a weigh ed summed o he Lempel-Zi
complexi y ex ac ed o e he coe icien s a he bes scale and he Lempel-Zi
complexi y ex ac ed o e he ampli ude demodula ed coe icien s a he bes scale.
Using his me hod, hey also ound ha he Lempel-Zi complexi y inc ease wi h he
ou e - ace aul se e i y and dec ease wi h he inne - ace aul se e i y. In ou p oposal,
i is no necessa y o compu e he ampli ude demodula ed signal o he signal in he
node wi h maximal ene gy and he e o e we only compu e Lempel-Zi complexi y one
ime.
One o he main ad an ages o using Lempel-Zi complexi y is ha i is
bounded be ween 0 and 1. Wi h he use o he wa ele packe ans o m and he
selec ion o he node wi h maximal ene gy ( he node in which he aul lies), he e ec s
o gaussian noise con amina ion is educed.
In he case o ou e - ace aul , his me hod can ollow he aul deg ada ion om
no mal condi ion o aul condi ion mono onically. I is wo h o men ion ha he esul s
ob ained wi h he ball aul a e no mono onically inc easing o dec easing wi h he aul
se e i y. Howe e , he ball au can be de ec ed wi h he p oposed me hod.
The me hod was applied o wo un- o- ailu e expe imen s: UH-60 helicop e
da abase and IMS da abase. In he case o he UH-60 helicop e he p oposed me hod
does no de ec he aul in ea ly s ages. A possible eason is ha he wa ele packe is
no de ec ing he co ec node whe e he aul lies. Mo e esea ch is needed in his case.
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 257
o he co ela ion dimension o each emo ion a e below his quan i y (see Figu e 8-3)
whe e he da a dis ibu ions o he co ela ion dimension es ima ed o each da abase a e
shown).
The delay (τ) and he minimum embedding dimension (m) a e es ima ed o each
ame using he i s minimum o he mu ual in o ma ion unc ion echnique and he
alse neighbou s echnique espec i ely. Then, six complexi y measu es a e ex ac ed
o each ame: alue o he i s minimum o mu ual in o ma ion unc ion (MI), Taken-
Theile es ima o o he co ela ion dimension (CD), Shannon en opy (SE), co ela ion
en opy (CE), Lempel-Zi complexi y (LZC) and Hu s exponen (H). Finally, o each
measu e, ou s a is ical a e compu ed: mean (μ), s anda d de ia ion (σ), skewness (sk)
and ku osis (k). The e o e, 24 ea u es a e ex ac ed o each ame (mean o MI: μMI,
mean o CD: μCD and so on).
Fea u e selec ion and e alua ion o he selec ed ea u es
A se o ea u es a e selec ed using a ea u e selec ion echnique p oposed by he au ho
o his Thesis. In his Chap e , he p oposed ea u e selec ion p ocedu e will no be
explained (see [39]). Once he selec ed ea u es a e iden i ied, hei disc imina ion
abili y be ween he di e en emo ions is e alua ed using a neu al ne wo k. The da abase
is spli in o a aining subse and a es subse wi h 70% and 30% o each kind o
emo ional speech eco dings, espec i ely. The expe imen s a e epea ed 25 imes, each
ime using di e en aining and es se s andomly chosen and he global success a e is
compu ed as an a e age o he success a es in each i e a ion.
8.2.3 Resul s
The esul s ob ained in he expe imen s a e shown and discussed. Fi s , we show a da a
dis ibu ion analysis o he complexi y measu es (MI, SE, CD, CE, LZC and H). Then,
we show he esul s o he ea u e selec ion p ocedu e. The ea u e selec ion p ocedu e
is pe o med in he Polish emo ional speech da abase. Finally, we show he success
a es o he h ee emo ional da abases: he Polish emo ional speech da abase, he Be lin
emo ional speech da abase and he LDC emo ional speech da abase using he GOFS
p e iously iden i ied.

PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
258
Analysis o complexi y measu es
The complexi y measu es ex ac ed om he h ee da abases a e analyzed. Figu e 8-3,
Figu e 8-4 and Figu e 8-5 show he dis ibu ion o he six complexi y measu es (MI,
SE, CD, CE, LZC and H) o he Polish emo ional speech da abase, o he Be lin
emo ional speech da abase and o he LDC emo ional speech da abase espec i ely
using boxplo s. The boxes ha e lines a he lowe qua ile, median (cen e line) und
uppe qua ile alues. The whiske s a e lines ex ending om each end o he boxes o
show he ex en o he es o he da a. Boxes whose no ches do no o e lap indica e ha
he medians o he wo g oups di e a he 5% signi icance le el. In Figu e 8-3, Figu e
8-4 and Figu e 8-5 he uppe le illus a ion co esponds o he MI, he uppe igh
illus a ion co esponds o he SE, he middle le illus a ion co esponds o he CD, he
middle igh illus a ion co esponds o he CE, he bo om le illus a ion co esponds
o he LZC and inally he bo om igh illus a ion co esponds o he H.
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 259
Figu e 8-3: Da a dis ibu ion o neu al, ea and ange emo ional
speech o each complexi y measu e ex ac ed om he Polish
emo ional da abase. MI: alue o he i s minimum o he mu ual
in o ma ion unc ion (uppe le ), SE: Shannon en opy (uppe igh ),
CD: Taken's es ima o o he co ela ion dimension (middle le ), CE:
co ela ion en opy (middle igh ), LZC: Lempel–Zi complexi y
(bo om le ), H: Hu s exponen (bo om igh ) [40].
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
260
Figu e 8-4: Da a dis ibu ion o neu al, ea and ange emo ional
speech o each complexi y measu e ex ac ed om he Be lin
emo ional da abase. MI: alue o he i s minimum o he mu ual
in o ma ion unc ion (uppe le ), SE: Shannon en opy (uppe igh ),
CD: Taken's es ima o o he co ela ion dimension (middle le ), CE:
co ela ion en opy (middle igh ), LZC: Lempel–Zi complexi y
(bo om le ), H: Hu s exponen (bo om igh ) [40].
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 261
Figu e 8-5: Da a dis ibu ion o neu al, ea and ange emo ional
speech o each complexi y measu e ex ac ed om he LDC
emo ional da abase. MI: alue o he i s minimum o he mu ual
in o ma ion unc ion (uppe le ), SE: Shannon en opy (uppe igh ),
CD: Taken's es ima o o he co ela ion dimension (middle le ), CE:
co ela ion en opy (middle igh ), LZC: Lempel–Zi complexi y
(bo om le ), H: Hu s exponen (bo om igh ) [40].
Acco ding o he Figu e 8-3 he median o he alues o he i s minimum o he
mu ual in o ma ion unc ion (MI) be ween a signal and i s delayed e sion is highe in
neu al speech han in ea emo ional speech and in ange emo ional speech. This means
ha in he ime o maximum di e ence (i.e. when he i s minimum o he mu ual
in o ma ion occu s) o a signal wi h i s delayed e sion, his di e ence is lowe in
neu al speech han in ea o ange emo ional speech. This endency is he same in he
h ee da abases. In he case o he Shannon en opy dis ibu ions, he e a e no clea
di e ences be ween neu al, ea and ange emo ional speech. The median o he
dis ibu ions a e o e lapped in he case o he LDC da abase. Howe e , hey a e mo e
clea ly sepa a ed in he case o he Be lin emo ional speech da abase and in he case o
he Polish emo ional speech da abase.
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
262
The dis ibu ion o he Taken's es ima o o he co ela ion dimension (CD) also
shows a simila beha io in he h ee da abases. The median alues o he CD a e highe
in ea emo ional speech and in ange emo ional speech han in neu al speech. This is
an indica o o a mo e complex geome ical s uc u e in ange and ea emo ional
speech. Howe e , he e a e no clea di e ences in he alues o he CD be ween ange
and ea emo ional s a es in he Be lin da abase. The median o he co ela ion en opy
(CE) dis ibu ion o neu al speech shows lowe alues han he median o ange and
ea emo ional speech. This is an indica o o a mo e complex s uc u e in ange and
ea speech han in neu al speech. The CE measu e shows o be disc imina i e be ween
neu al and ea and ange emo ional speech. Howe e , CE is less disc imina i e
be ween ea and ange emo ional speech.
The Lempel-Zi complexi y (LZC) dis ibu ion shows alues mo e nea o 1 o
ea and ange emo ional speech han in he case o neu al speech. This means ha
ange and ea emo ional speech eco ds show mo e complexi y han neu al speech
eco ds. Finally, he Hu s exponen (H) shows ha he median alue o neu al speech
is highe han median o ea and ange emo ional speech. Values o H o ea and
ange emo ional speech a e close o 0.5, showing ha his signals has mo e andomness
componen s.
Acco ding o he da a dis ibu ions, he six complexi y measu es a e
disc imina i e be ween neu al speech and nega i e emo ional speech (ange emo ional
speech and ea emo ional speech). Howe e , in he case o he CE and CD he
disc imina ion be ween ea emo ional speech and ange emo ional speech a e less clea .
Mo eo e , he da a dis ibu ions a e e y simila in he h ee da abases. This means ha
he disc imina i e abili y o he complexi y measu es is independen om he language.
Resul s o ea u e selec ion
The ollowing ea u es we e selec ed in he ea u e selec ion p ocedu e: μMI, μH, sSE,
sLZC.

Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 263
Resul s o he da abases e alua ion wi h he selec ed ea u es
The selec ed ea u es (μMI, μH, sSE and sLZC) we e e alua ed wi h a neu al ne wo k
classi ie (wi h 5 neu ons in he hidden laye ) o he h ee emo ional speech da abases
(Polish, Be lin and LDC da abases). The global success a es o he di e en emo ional
da abases a e shown in Table 8-5 wi h he s anda d de ia ion (σ). Acco ding o he
esul s, he selec ed ea u es show a good disc imina ion abili y be ween h ee
emo ional speech s a es (neu al s a e, ea emo ional s a e and ange emo ional s a e) in
he h ee da abases.
TABLE 8-5: GLOBAL SUCCESS RATES OF THE SELECTED FEATURES IN
THREE EMOTIONAL DATABASES
Polish da abase
Be lin da abase
LDC da abase
Success a es (%)
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 he con usion ma ix o he selec ed
ea u es in he h ee emo ional speech da abases wi h he mean and s anda d de ia ion
(σ) alues ob ained a e aging he esul s o each indi idual expe imen . Acco ding o
he esul s, he selec ed ea u es show good disc imina ion abili y in he h ee da abases
be ween neu al, ea emo ional s a e and ange emo ional s a e. The disc imina ion
abili y be ween neu al and nega i e emo ional s a es a e highe han be ween ea and
ange emo ional speech in he case o he Polish emo ional speech da abase and he
Be lin emo ional speech da abase. Howe e , he LDC da abase shows good
disc imina ion abili y in ange emo ional s a e agains ea and neu al s a es. The
simila esul s in he h ee da abases show ha he selec ed ea u es a e independen o
he language.
TABLE 8-6: CONFUSION MATRIX OF THE SELECTED FEATURES IN
POLISH EMOTIONAL SPEECH DATABASE
Classi ie decision (%)
Ac ual emo ional s a e
Neu al
Fea
Ange
Neu al
88.00 (σ = 11.04)
8.00 (σ = 10.34 )
4.00 (σ = 4.88)
Fea
17.00 (σ = 10.06)
64.00 (σ = 14.77)
19.00 (σ = 14.54)
Ange
7.33 (σ = 8.44)
26.33 (σ = 13.32)
66.33 (σ = 15.86)
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
264
TABLE 8-7: CONFUSION MATRIX OF THE SELECTED FEATURES IN
BERLIN EMOTIONAL SPEECH DATABASE
Classi ie decision (%)
Ac ual emo ional s a e
Neu al
Fea
Ange
Neu al
78.80 (σ = 7.81)
15.00 (σ = 6.45)
6.20 (σ = 3.89)
Fea
16.00 (σ = 10.10)
63.20 (σ = 11.35)
20.80 (σ = 10.07)
Ange
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
Classi ie decision (%)
Ac ual emo ional s a e
Neu al
Fea
Ange
Neu al
77.57 (σ = 8.76)
16.70 (σ = 7.50)
5.74 (σ = 4.65)
Fea
19.30 (σ = 9.57)
72.17 (σ = 8.33)
8.52 (σ = 4.78)
Ange
3.30 (σ = 3.82)
4.17 (σ = 3.19)
92.52 (σ = 4.44)
8.2.4 Conclusions
The use ulness o complexi y ea u es in disc imina ing be ween neu al s a e, ea
emo ional s a e and ange emo ional s a e is e alua ed. Six complexi y measu es
including he alue o i s minimum o mu ual in o ma ion unc ion, he Shannon
en opy, he Takens es ima o o he co ela ion dimension, he co ela ion en opy, he
Lempel-Zi complexi y and he Hu s exponen a e ex ac ed om h ee emo ional
da abases ( he Polish emo ional speech da abase, he Be lin emo ional speech da abase
and he LCD da abase). Then, he mean, s anda d de ia ion, skewness and ku osis a e
applied o he six complexi y measu es and 24 ea u es a e ob ained. Fea u e selec ion is
accomplished o selec a educed numbe o ea u es o e he Polish emo ional da abase.
Finally, he selec ed ea u es a e e alua ed in he Be lin emo ional speech da abase and
in he LDC emo ional da abase using a neu al ne wo k classi ie .
A quali a i e analysis o he six complexi y measu es ex ac ed om he h ee
emo ional da abases is accomplished wi h he obse a ion o he da a dis ibu ion o he
complexi y measu es. F om his analysis, he ollowing conclusions a e ex ac ed.
Acco ding o he da a dis ibu ions analysis, he six complexi y measu es a e
disc imina i e be ween neu al speech, ea emo ional speech and ange emo ional
speech. In gene al, ea and ange emo ional speech eco ds show mo e complexi y han
neu al speech eco ds. The eason can be ha in ea and ange speech eco ds, people
Ad ances in p e en i e moni o ing o machine y h ough audio and ib a ion signals
Uni e sidad de Las Palmas de G an Cana ia 265
end o use mo e ica i e sounds han in neu al s a e. F ica i e sounds a e noisie han
oiced sounds. Mo eo e , he beha io o he da a dis ibu ions o he six complexi y
measu es is he same in he h ee da abases. The da abases consis o emo ional speech
eco ds in Polish, Ge man and English languages. We can conclude, he e o e, ha he
complexi y measu es a e independen om he language.
The ou selec ed ea u es (mean o he alue o he i s minimum o he mu ual
in o ma ion unc ion, he mean o he Hu s exponen , he s anda d de ia ion o he
Shannon en opy and he s anda d de ia ion o he Lempel-Zi complexi y) we e
e alua ed wi h a neu al ne wo k classi ie in he h ee da abases. Global success a es o
72.28%, 75.4% and 80.75%, we e ob ained o he Polish emo ional speech da abase,
he Be lin emo ional speech da abase and he LDC emo ional speech da abase
espec i ely in he disc imina ion be ween neu al, ea and ange emo ional s a es.
Possibly applica ions o an au oma ic ecogni ion sys em ha disc imina e be ween
neu al emo ion and nega i e emo ions such as ea and ange can be applied in call
cen e s in o de o de ec p oblems in cos ume -sys em in e ac ion and in secu i y
applica ions in o de o de ec secu i y h ea s.
8.3 Con ibu ions
The con ibu ions o his Chap e a e he s udies o nonlinea and complexi y measu es
in pa hological oice de ec ion and in emo ional oice de ec ion.
PhD Disse a ion
Uni e sidad de Las Palmas de G an Cana ia
266
8.4 Re e ences
[1] Hen íquez, P., Alonso, J. B., Fe e , M., T a ieso, C. M., Godino-Llo en e, J., &
Díaz-de-Ma ía, F. (2009). Cha ac e iza ion o heal hy and pa hological oice
h ough measu es based on nonlinea dynamics. Audio, Speech, and Language
P ocessing, IEEE T ansac ions on, 17(6), 1186-1195.
[2] Hi ano, M. Clinical Examina ion o Voice. New Yo k: Sp inge , Ve lag, 1981.
[3] Nawka, T., Ande s, L. C., & Wendle , J. (1994). Die audi i e Beu eilung heise e
S immen nach dem RBH Sys em. Sp ache S imme Gehö , 18, 130-133.
[4] Boyano , B., Hadjo odo o , S., Tes on, B., & Dosko , D. (1997). Robus hyb id
pi ch de ec o o pa hological oice analysis. In La ynx 97 (pp. 55–58).
[5] Boyano , B., I ano , T., & Cholle , G. (1993). Robus hyb id pi ch de ec o .
Elec on. Le ., 29(22), 1924–1926.
[6] H. Kasuya, Y. Endo, and S. Saliu, “No el acous ic measu emen s o ji e and
shimme cha ac e is ics om pa hologic oice,” in P oc. Eu ospeech’ 93, Be lin,
Ge many, 1993, pp. 1973–1976.
[7] C. Ludlow, C. Bassich, N. Conno , D. Coul e , and Y. Lee, “The alidi y o using
phona o y ji e and shimme o de ec la yngeal pa hology,” in La yngeal Func ion
in Phona ion and Respi a ion. Bos on, MA: B own, 1987, pp. 492–508.
[8] Yumo o, E., Gould, W. J., & Bae , T. (1982). Ha monics o noise a io as an index
o he deg ee o hoa seness. The jou nal o he Acous ical Socie y o
Ame ica, 71(6), 1544-1550.
[9] M. Yunik and B. Boyano , “Me hod o e alua ion o he noise- o-ha monic-
componen a ios in pa hological and no mal oices,” Acus ica, ol. 70, pp. 89–91,
1990.
[10] Kasuya, H., Ogawa, S., Mashima, K., & Ebiha a, S. (1986). No malized noise
ene gy as an acous ic measu e o e alua e pa hologic oice. The Jou nal o he
Acous ical Socie y o Ame ica, 80(5), 1329-1334.
[11] M. F ohlich, D. Michaelis, and H. W. S ube, “Acous ic ‘B ea hiness
Measu es’ in he desc ip ion o pa hologic oices,” in P oc. Acous ., Speech, Signal
P ocess. ICASSP’98, Sea le, WA, 1998, ol. 2, pp. 937–940.
[12] L. Gu, J. G. Ha is, R. Sh i as a , and C. Sapienza, “Diso de ed speech
assessmen using au oma ic me hods based on quan i a i e measu es,” EURASIP J.
Appl. Signal P ocess., ol. 9, pp. 1400–1409, 2005.
[13] Boyano , B., & Hadji odo o , S. (1997). Acous ic analysis o pa hological
oices. A oice analysis sys em o he sc eening o la yngeal diseases.Enginee ing
in Medicine and Biology Magazine, IEEE, 16(4), 74-82.
[14] Wallen, E. J., & Hansen, J. H. (1996, Oc obe ). A sc eening es o speech
pa hology assessmen using objec i e quali y measu es. In Spoken Language, 1996.
ICSLP 96. P oceedings., Fou h In e na ional Con e ence on (Vol. 2, pp. 776-779).
IEEE.
[15] De Oli ei a Rosa, M., Pe ei a, J. C., & Ca alho, A. C. (1998, Decembe ).
E alua ion o neu al classi ie s using s a is ic me hods o iden i ica ion o
La yngeal pa hologies. In P oc. 5 h B azilian Symp. Neu al Ne w. ( ol. 1, pp. 220–
225).
[16] Hadji odo o , S., & Mi e , P. (2002). A compu e sys em o acous ic analysis
o pa hological oices and la yngeal diseases sc eening. Medical enginee ing &
physics, 24(6), 419-429.
1El p esen e documen o ha sido ealizado a pa i de la inanciación del P og ama de Fo mación del
Pe sonal In es igado de la Agencia Cana ia de In es igación, Inno ación y Sociedad de la In o mación
del Gobie no de Cana ias con una asa de con inanciación del 85% del Fondo Social Eu opeo y de los
p oyec os TEC2009-14123-C04 y TEC2012-38630-C04-02 del Minis e io de Economía y
Compe i i idad del Gobie no 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 Doc o al
A ances en Moni o ización P e en i a de Maquina ia
a á es de Señales de Audio y Vib ación1
Au o a: Pa icia Hen íquez Rod íguez
Di ec o es: D . D. Miguel Ángel Fe e Balles e
D . D. Jesús Be na dino Alonso He nández
Las Palmas de G an Cana ia, 22 de Oc ub e de 2015

A ances en moni o ización p e en i a de maquina ia a a és de señales de audio y ib ación
Uni e sidad de Las Palmas de G an Cana ia 275
A. Resumen de la Tesis
1 In oducción
La maquina ia en gene al y los equipos indus iales en pa icula se de e io an a lo la go
del iempo po el es és que su en du an e su ida ope a i a. El allo inespe ado en una
máquina abajando en un en o no indus ial puede ene consecuencias g a es an o
desde el pun o de ida humano, po el posible acciden e que pueda causa , como desde
el pun o de is a de cos es de p oduc i idad. Es po ello que el man enimien o de
equipos indus iales se ha con e ido en un aspec o es a égico en muchas emp esas. La
palab a man enimien o se emplea pa a designa las écnicas usadas pa a asegu a el uso
co ec o y con inuo de maquina ia, equipos e ins alaciones. Moni o iza de o ma
con inua el es ado de uncionamien o (o condición) de un ac i o ísico (una máquina,
pa e de una máquina o un sis ema compues o de a ias máquinas) es de i al
impo ancia pa a la de ección emp ana de allos y iene una g an in luencia en la
con inuidad ope acional de muchos p ocesos indus iales. Ayuda a educi cos es de
Tesis Doc o al
Uni e sidad de Las Palmas de G an Cana ia
276
man enimien o e inc emen a la segu idad y con iabilidad de los equipos indus iales. En
la siguien e abla se mues an en po cen ajes las en ajas del uso de un sis ema de
moni o ización de la maquina ia [1].
TABLA 1: VENTAJAS DE UN SISTEMA DE MONITORIZACIÓN [1]
Cos es de man enimien o
Reducción del 50% al 80%
Daños en los equipos
Reducción del 50% al 60%
Gas os en ho as ex as
Reducción del 20% al 50%
Espe anza de ida de la máquina
Inc emen o del 50% al 60%
P oduc i idad o al
Inc emen o del 20% al 30%
Los cos es de man enimien o se educen debido al hecho de que se de ec an
allos incipien es, e i ando que dichos allos c ezcan y se con ie an en un p oblema
g a e y ca o. Se educe la p obabilidad de apa ición de allos des uc i os que dañen a
los equipos y a ec en a la segu idad de las pe sonan. Se educen las ac i idades de
epa ación y po lo an o las ho as ex as dedicadas a ello. La espe anza de ida de la
máquina se inc emen a así como la p oduc i idad o al pues o que se e i an pa adas
innecesa ias de la maquina ia.
El in e és c ecien e en las écnicas de moni o ización del es ado de la maquina ia
pa a la de ección, diagnós ico y deg adación de allos an o en el campo de la
in es igación como en el de la indus ia es e iden e po la g an can idad de a ículos
publicados en el campo, po los es ue zos de las o ganizaciones de es anda ización
(ISO, SAE, e c) y po la o ganización de di e en es con e encias en el campo del
diagnós ico de allos, como po ejemplo la con e encia COMADEM (Condi ion
Moni o ing and Diagnos ic Enginee ing Managemen ).
A ances en moni o ización p e en i a de maquina ia a a és de señales de audio y ib ación
Uni e sidad de Las Palmas de G an Cana ia 277
Técnicas de man enimien o
Desde iempos an iguos, la humanidad ha usado di e en es écnicas de man enimien o.
Los homb es p imi i os a ilaban sus ense es y a mas y cosían sus pieles. Du an e la
e olución indus ial, se empeza on a usa elés de p o ección de sob e-co ien e y
p o ección de allo de ie a mien as que los úl imos desa ollos incluyen écnicas de
p ocesado de la señal y econocimien o de pa ones.
Gene almen e, las écnicas de man enimien o se di iden en dos: man enimien o
co ec i o y man enimien o p e en i o. Es e úl imo se in odujo en los años 1950 y se
di ide a su ez en man enimien o p e en i o p ede e minado y en man enimien o
basado en la condición, conocido como CBM en sus siglas en inglés (condi ion-based
main enance). El man enimien o basado en la condición ambién se denomina
man enimien o p edic i o.
En el man enimien o co ec i o se oman acciones después de que el allo haya
ocu ido. Es as acciones an di igidas a a egla el allo o a pos pone su epa ación de
acue do al c i e io de pe sonal cuali icado.
En el man enimien o p e en i o p ede e minado, se ealizan ac i idades de
man enimien o plani icadas a in e alos pe iódicos pa a e i a que los componen es se
deg aden has a el pun o de que la máquina deje de unciona . La máquina o pa e de la
máquina se epa a o se cambia an es de que ocu a un allo.
Tesis Doc o al
Uni e sidad de Las Palmas de G an Cana ia
278
Las écnicas co ec i as y de man enimien o p e en i o p ede e minado han
demos ado se bas an e cos osas pues o que muchas eces se ealizan cambios de pieza
innecesa ios, al pa a la máquina pa a ealiza el cambio de pieza se pa a la p oducción
con lo cual se aumen an los cos es de p oducción. Además las ac i idades plani icadas
de man enimien o sueles se cos osas. Po es as azones, algunas indus ias empeza on a
ealiza man enimien o basado en la condición en los años 1980.
El man enimien o p e en i o basado en la condición (CBM) o man enimien o
p edic i o se e ie e a la moni o ización del es ado de la máquina en la cual se ob iene
de o ma con inua in o mación de pa áme os que indican el es ado de la máquina. La
des iación de los pa áme os de la condición no mal indica el desa ollo de un allo. La
CBM lle a a cabo acciones de man enimien o sólo cuando hay e idencia de
compo amien o ano mal. La CBM educe el núme o de ac i idades plani icadas
educiendo po an o cos es [1].
E apas de un sis ema de moni o ización basado en la condición
En la igu a 1 se mues an las e apas ípicas de un sis ema de moni o ización que
implemen a man enimien o basado en la condición. Es as e apas incluyen la adquisición
de da os, el p ocesado de los da os y un sis ema de ayuda a la decisión. La salida del
sis ema de moni o ización se á el diagnós ico del allo (de ec a el allo y sabe dónde se
ha p oducido) y posiblemen e ambién la iden i icación del allo (qué g ado de
se e idad p esen a el allo) pa a así de e mina qué iempo de ida ú il le queda al
elemen o que es á siendo supe isado.

A ances en moni o ización p e en i a de maquina ia a a és de señales de audio y ib ación
Uni e sidad de Las Palmas de G an Cana ia 279
A con inuación se de allan cada una de las e apas de un sis ema de
moni o ización de la condición. Pues o que la p esen e Tesis se cen a en di e sos
aspec os de las dis in as e apas de un sis ema de moni o ización de la condición, se
explica á en cada e apa en qué se cen a la p esen e Tesis.
Adquisición de da os: es a e apa consis e en la ob ención de in o mación ele an e sob e
el es ado de la máquina. Los da os adqui idos pueden a ia según la clase de máquina o
la na u aleza del allo. La in o mación adqui ida puede se de muchos ipos [2]: da os de
ipo alo como p esión, empe a u a, da os de análisis de acei e; da os de o ma de
onda (es deci , señales) como señales de ib ación, de audio, señales de emisión
acús ica; y da os mul idimensionales como imágenes. El conjun o de da os adqui idos
se denomina i ma de la máquina. La p esen e Tesis se cen a en señales de ib ación y
en señales de audio como uen e de in o mación.
P ocesado de da os: los da os ob enidos en la e apa an e io se analizan pa a ob ene
in o mación que pe mi a de ec a un posible allo o diagnos ica lo. En es a Tesis se usan
écnicas de p ocesado de la señal pa a analiza los da os y ex ae in o mación aliosa
pa a la moni o ización de la condición. Es e p oceso se denomina ex acción de
ca ac e ís icas.
Sis ema de ayuda a la decisión o la clasi icación de los da os p e iamen e analizados en
di e en es es ados de la condición. En es a Tesis, se usan écnicas de econocimien o de
pa ones, en conc e o dos clasi icado es que pe mi en, una ez en enados, diagnos ica
de o ma au omá ica el es ado de la máquina.
Tesis Doc o al
Uni e sidad de Las Palmas de G an Cana ia
280
Diagnós ico de allos/P edicción de allo. El obje i o de un sis ema de moni o ización
de la condición es el diagnós ico e iden i icación de allos. El diagnós ico de allo
consis e en la de ección del allo, que esponde a la p egun a “¿hay un allo?” y al
aislamien o del allo, que esponde a la p egun a “¿dónde es á el allo?” usando las
e apas mencionadas. La iden i icación del allo se e ie e a la de e minación de la
se e idad o el amaño del allo y a eces ambién a la de e minación del iempo de
comienzo del allo (o p edicción del allo). En es a Tesis, se ealiza diagnós ico de allo
e iden i icación de allo.
Figu a 1: E apas de un sis ema de moni o ización basado en la
condición
Diagnós ico de allos basado en ib ación y audio: cojine es y bombas cen í ugas
Tal y como se ha explicado en el apa ado an e io en las e apas de un sis ema de
moni o ización basado en la condición, a la ho a de adqui i da os de la máquina se
pueden usa mul i ud de da os di e en es. En es a Tesis nos cen amos en señales de
ib ación y de audio (en el espec o audible 0-20kHz). Es po eso que en es e sub-
apa ado se in oduce el diagnós ico de allos basados en es as dos señales.
Asimismo, la in es igación de la Tesis se cen a en dos aplicaciones: cojine es y
bombas cen í ugas. Los cojine es son elemen os esenciales en las máquinas o a i as
pues o que sopo an la es uc u a de la máquina pe mi iendo y acili ando su o ación.
Diagnós ico/
P edicción
Adquisición de da os
(Fi ma de la máquina)
P ocesado de
da os
Sis ema de ayuda a
la decisión
A ances en moni o ización p e en i a de maquina ia a a és de señales de audio y ib ación
Uni e sidad de Las Palmas de G an Cana ia 281
Un allo no de ec ado en un cojine e puede causa una a e ía ca as ó ica (con ac os
indeseados en e pa es ijas y mó iles de la máquina, bloqueo del mo o , e c.). Las
máquinas o a i as es án muy ex endidas en la indus ia. Ejemplos de máquinas
o a i as son los mo o es y los gene ado es. Debido a la g an u ilización de es e ipo de
máquinas, la aplicación de écnicas encaminadas a la igilancia y con ol del es ado de
los cojine es adquie e suma impo ancia. Aunque el cos e de los cojine es en
compa ación con la máquina en sí es muy bajo, el hecho de que haya que desmon a la
máquina casi en su o alidad pa a cambia un cojine e hace que las écnicas de
moni o ización basadas en la condición sean muy ú iles pa a pe mi i que es os
elemen os uncionen has a el máximo de su ida ú il.
Po o a pa e, las bombas cen í ugas son máquinas que o man pa e
impo an e de muchos sis emas, pues o que pe mi en el mo imien o de luidos en e dos
pun os del sis ema. Se usan en la indus ia eléc ica, en la química, en la indus ia de
ex acción de pe óleo, en sis emas de e ige ación, en spas y piscinas, e c. Po lo an o,
la moni o ización de es a clase de máquinas es muy impo an e pa a el co ec o
uncionamien o de muchos sis emas indus iales.
Tan o en los cojine es como en las bombas cen í ugas, así como en muchas
o as clases de máquinas su moni o ización se hace a endiendo no malmen e a señales
de ib ación. Po supues o ambién se usan o o ipo de da os como la co ien e, ol aje
o empe a u a en el caso de los cojine es y la p esión en el caso de las bombas
cen í ugas. El diagnós ico de allos basado en ib ación ( ib a ion-based aul
diagnosis) se e ie e al diagnós ico de allos usando la señal de ib ación como uen e
de in o mación. El diagnós ico de allos basado en ib ación es un á ea de es udio muy
Tesis Doc o al
Uni e sidad de Las Palmas de G an Cana ia
282
desa ollada que incluye un amplio ango de écnicas que han e olucionado de o ma
ápida du an e las úl imas décadas. En la moni o ización de la condición, las écnicas de
diagnós ico de allos basadas en ib ación han sido ampliamen e usadas debido a la
acilidad de adqui i la señal de ib ación de la máquina. Po es a azón, la mayo ía de
los a ículos de in es igación en la li e a u a del diagnós ico de allos es án en ocados al
diagnós ico de allos basado en ib ación [2],[3],[46].
El diagnós ico de allos basado en audio se e ie e al diagnós ico de allos
usando la señal de audio como uen e de in o mación (audio-based aul diagnosis o
ambién llamado ai bo ne aul diagnosis). El diagnós ico de allos basado en audio
usando mic ó onos en el ango audible (0-20kHz) es un campo eme gen e con un g an
po encial en el campo del diagnós ico de allos pues o que los mic ó onos son senso es
no in asi os (no an mon ados encima de la máquina) y ienen mayo es posibilidades
de localización que los acele óme os. Además, el uso de señales de audio pod ía
mejo a la inspección de cie os en o nos indus iales en los cuales un sis ema de
moni o ización ijo es ca o o el mon aje de los senso es de ib ación es complicado
como po ejemplo en la indus ia de ex acción de pe óleo. Po es as azones, pensamos
que el diagnós ico basado en audio necesi a más es ue zos de in es igación. De hecho,
la moni o ización de la condición con señales de audio p esen a menos es udios
cien í icos que la ib ación [2],[3],[46].
Po lo an o, es a Tesis se en oca en dos aplicaciones: cojine es y bombas
cen í ugas. En el caso de los cojine es se usan di e en es bases de da os de ib ación
públicas y nos cen amos an o en el diagnós ico de allo como en la iden i icación del
allo (se e idad del allo). La in es igación se ha en ocado en la e apa de p ocesado de
A ances en moni o ización p e en i a de maquina ia a a és de señales de audio y ib ación
Uni e sidad de Las Palmas de G an Cana ia 289
Figu a 3: Dis ibución de los a ículos usados pa a la edacción del
es ado del a e: dis ibución de a ículos basados en señales de
ib ación y basados en señales de audio (supe io izquie da); años de
publicación de los a ículos (supe io de echa); dis ibución de
elemen os donde se p oducen los allos analizados con señales de
ib ación (in e io izquie da) como con señales de audio (in e io
de echa).
De la igu a 3 se puede obse a ambién que una g an can idad de allos se
p oduce en los cojine es. De hecho, de acue do con una encues a sob e con iabilidad de
mo o es [6] el 42% de los allos de mo o es de más de 200 caballos de po encia se
deben a allos en los cojine es. Mo o es de meno caballaje p esen an aún más
po cen aje de allos en los cojine es, llegando a se del 90% [7]. Como ya se ha
comen ado en la in oducción, las écnicas de moni o ización de la condición son muy
ú iles pa a de ec a si un cojine e p esen a algún ipo de allo. También en la igu a 3 se
puede obse a que la mayo pa e de la moni o ización de cojine es se ealiza usando

Tesis Doc o al
Uni e sidad de Las Palmas de G an Cana ia
290
señales de ib ación. Po o a pa e, a lo la go de la elabo ación del es ado del a e se
ealizó una búsqueda exhaus i a de bases de da os públicas, de lib e acceso y g a ui as.
F u o de esa búsqueda se ob u ie on es bases de da os de señales de ib ación de
cojine es [30]-[32]. Basándonos en es as obse aciones de i adas del es ado del a e, la
p ime a pa e de es a Tesis se cen a en el es udio de señales de ib ación de cojine es
pa a el diagnós ico e iden i icación de allos.
En la li e a u a de diagnós ico de allos de cojine es se encuen a una amplia
a iedad de écnicas de p ocesado de la señal usadas pa a ex ae ca ac e ís icas de la
señal de ib ación en di e en es dominios de ep esen ación: dominio del iempo [37],
de la ecuencia [9], ceps al [10], iempo- ecuencia [11]-[15], así como écnicas no
lineales basadas en medidas de dinámica no lineal y medidas de complejidad y
p edic ibilidad [28], [43]. Las écnicas no lineales pe mi en ex ae ca ac e ís icas de la
i ma de la máquina que pueden e ela un en endimien o más p eciso de la señal. De
hecho, se han encon ado indicios de no linealidad en el uncionamien o de cojine es
[28]. Po es e mo i o, en la ase de p ocesado de la señal (en donde se ex aen las
ca ac e ís icas), nos hemos cen ado en el es udio de écnicas no lineales den o de la
aplicación de los cojine es an o pa a diagnós ico como pa a iden i icación de allos.
Si bien la moni o ización usando señales de ib ación es mucho más usada en
compa ación con el audio, en los úl imos años se ha inc emen ado la in es igación en
o no a la moni o ización basada en señales de audio [5],[19]-[27]. Además, si bien la
localización de los mic ó onos debe de se lo más ce cana posible a la máquina pa a
e i a en la medida de lo posible in e e encias con o os elemen os, es os ienen más
posibilidades de localización que los senso es de ib ación [5]. Además, no ienen que
A ances en moni o ización p e en i a de maquina ia a a és de señales de audio y ib ación
Uni e sidad de Las Palmas de G an Cana ia 291
se mon ados en la máquina, po lo que son elemen os no in asi os en absolu o.
Basándonos en es as obse aciones, la segunda pa e de la p esen e Tesis se cen a en
explo a la moni o ización del es ado de la máquina usando señales de audio.
Pues o que hay una ca encia de bases de da os públicas de señales de audio en
diagnós ico de allos en maquina ia, en la p esen e Tesis se op a po g aba una base de
da os p opia. Dicha base de da os cons a de señales de audio y de señales de ib ación
adqui idas simul áneamen e de una bomba cen í uga de agua. Se adquie en ambién
señales de ib ación pa a compa a esul ados.
Como ya hemos explicado, las señales de audio son menos usadas que las de
ib ación en el campo de diagnós ico de allos, y el caso de las bombas cen í ugas no
es una excepción. Es po es o que la mayo ía de ca ac e ís icas usadas en diagnós ico de
allos en bombas de agua son ex aídas de señales de ib ación. Po es e mo i o, en es a
Tesis se ealiza un es udio de las medidas usadas en señales de ib ación pa a
diagnós ico de allos en bombas cen í ugas y es as medidas se ex aen ambién de las
señales de audio. De igual o ma se p e ende, en base al es udio de las señales de audio
y ib ación de la bomba cen í uga p opone o as medidas que sean capaces de
disc imina en e di e en es es ados de no malidad y allo de la bomba cen í uga.
O a conclusión que se puede ex ae al analiza el es ado del a e es que hay
escasos es udios en los que se abaje conjun amen e con señales de audio y ib ación
[5], [22], [23]. Es po es e mo i o que en la p esen e Tesis se ealiza un es udio de la
combinación de señales de audio y ib ación.
Tesis Doc o al
Uni e sidad de Las Palmas de G an Cana ia
292
3 Obje i os
El obje i o de la p esen e Tesis es la mejo a de los sis emas de moni o ización del
es ado de la maquina ia en diagnós ico e iden i icación de allo usando señales de
ib ación y de audio en dos aplicaciones (cojine es y bombas de agua) con especial
én asis en la e apa de ex acción de ca ac e ís icas y en la u ilización del audio como
uen e de in o mación audio.
El obje i o de la p esen e Tesis puede se di idido en obje i os pa ciales que
deben se alcanzados pa a log a el obje i o gene al. Seguidamen e se exponen los
obje i os pa ciales:
1. Realiza una e isión gene al de las écnicas de p ocesado de señal usadas
pa a la ex acción de di e en es ca ac e ís icas así como de las écnicas
econocimien o de pa ones usadas en el diagnós ico de allos de maquina ia
usando señales de ib ación y de audio.
2. Realiza una búsqueda de bases de da os de ib ación de cojine es sin allo y
con di e en es ipos de allos pa a gene a así un eposi o io de bases de
da os con las que abaja .
3. Desa olla mé odos no lineales que ayuden al diagnós ico y a la
iden i icación de allos en cojine es.
A ances en moni o ización p e en i a de maquina ia a a és de señales de audio y ib ación
Uni e sidad de Las Palmas de G an Cana ia 293
4. Gene a una base de da os de audio y ib ación de una bomba cen í uga en
di e en es es ados de uncionamien o (sin allo y con di e en es ipos de
allos) pa a pode ealiza una e aluación de la capacidad de las señales de
audio en el diagnós ico de allos y compa a las con las señales de ib ación.
5. Aplica ca ac e ís icas del es ado del a e usadas comúnmen e en señales de
ib ación de bombas cen í ugas a las señales de audio cap adas de una
bomba cen í uga y cuan i ica la capacidad de las medidas en la
disc iminación en e a ios es ados de uncionamien o de la bomba
cen í uga (es ado no mal y es ados de allo).
6. Busca nue as medidas pa a diagnós ico de allos en bombas cen í ugas y
cuan i ica su capacidad de disc iminación en e a ios es ados de
uncionamien o de la bomba cen í uga.
7. Compa a los esul ados ob enidos con señales de ib ación y de audio.
8. Realiza un es udio de u ilización conjun a de señales de audio y ib ación
en la moni o ización de una bomba cen í uga
Tesis Doc o al
Uni e sidad de Las Palmas de G an Cana ia
294
4 Me odología
En el p esen e apa ado se expone la me odología seguida en la p esen e Tesis pa a
cumpli los obje i os p opues os.
Es ado del a e
En p ime luga se ealiza un es udio exhaus i o del es ado de la écnica o es ado del
a e en el ámbi o de la moni o ización de maquina ia con señales de audio y ib ación.
Nos cen amos en el es udio de écnicas de p ocesado de la señal pa a ex ae
ca ac e ís icas y en écnicas de econocimien o de pa ones usadas pa a disc imina
en e di e en es es ados de uncionamien o de la máquina. En el capí ulo 2 de la Tesis se
encuen a el es ado del a e jun o con un análisis c í ico del mismo.
Ob ención y gene ación de bases de da os
Se ealiza una búsqueda de bases de da os que comp endan mues as de cojine es en
es ado de uncionamien o no mal (sin allo) y uncionando con di e sos allos. Las
bases de da os ob enidas son de ib ación y es án disponibles en in e ne [30]-[32]. Una
de las bases de da os [30] comp ende señales de ib ación de cojine es sin allos y de
cojine es con allos pun uales en el anillo ex e no (ou e - ace aul ), en el anillo in e no
(inne - ace aul ) y con allo en los elemen os de odamien o (ball aul ). A modo de
ilus ación, en la igu a 4 se mues an los di e en es elemen os que componen un
cojine e. Las o as dos bases de da os [31], [32] son es - o- ailu e. Es deci , se g aba la
señal de ib ación has a que el cojine e o cojine es ienen un allo.

A ances en moni o ización p e en i a de maquina ia a a és de señales de audio y ib ación
Uni e sidad de Las Palmas de G an Cana ia 295
Figu a 4: Componen es de un cojine e cuyo elemen o de odamien o
son bolas.
Po o a pa e, dado que no hay bases de da os públicas con señales de audio
p o enien e de maquina ia, se decide g aba una base de da os p opia de señales de
audio y ib ación adqui idas de o ma simul ánea de una bomba de agua cen í uga de
ci culación (modelo ALP800) en un ci cui o ce ado o mado po la bomba en sí y un
anque con 50 li os de agua. An es de p osegui con la base de da os gene ada se
explica á b e emen e los p incipales elemen os de una bomba cen í uga.
Una bomba cen í uga cons a de dos elemen os undamen ales, el ode e y la
olu a. El ode e es el elemen o o a o io y la olu a el elemen o es aciona io. El ode e
con ie e la ene gía suminis ada po el mo o en ene gía ciné ica. La o ación del
ode e ue za al líquido a ci cula a a és de la bomba desde la di ección axial has a la
di ección adial mien as se ans ie e ene gía al líquido bombeado. La olu a con ie e
la ene gía ciné ica en ene gía de p esión. En esumen, el ode e p oduce elocidad en el
líquido y la olu a con ie e dicha elocidad en p esión. A modo de ilus ación en la
igu a 5 se mues an las di e en es pa es de una bomba de agua. Se indica ambién la
di ección de luido o líquido bombeado.
Tesis Doc o al
Uni e sidad de Las Palmas de G an Cana ia
296
Figu a 5: Pa es p incipales de una bomba cen í uga. Di ección del
luido en la bomba ( igu a izquie da). Fuen e: [32]. Fluid in: en ada
del luido. Fluid ou : salida del luido. Velocidad y p esión del luido
en una bomba. Figu a modi icada de la uen e: [34].
En la igu a 6 se mues a un ode e y se indican las di e en es pa es que lo
con o man. El ode e de una bomba cen í uga p esen e una se ie de anos delimi ados
po palas cu adas. Dichas palas ambién se denominan álabes. Los álabes son los que
impa en ue za cen í uga al luido. Un ode e ce ado como el que se mues a en la
igu a 6 iene pla os en ambos lados que encie an comple amen e el ode e desde el ojo
de succión (u ojo de ode e) a sus bo des. El ojo de ode e es la pa e cen al del ode e.
La egión del álabe más ce cana al ojo del ode e se denomina leading edge y la egión
en el bo de del álabe se denomina ailing edge.
A ances en moni o ización p e en i a de maquina ia a a és de señales de audio y ib ación
Uni e sidad de Las Palmas de G an Cana ia 297
Figu a 6: Pa es de un ode e. Figu a modi icada de la uen e: [32].
En la igu a 7 se mues a un plano de la sala donde se g aba la base de da os de
la bomba cen í uga. En el plano se obse a un esquema del mon aje ealizado con la
bomba (en e de), el anque con 50 li os de agua (en celes e) y las ube ías que o man
un ci cui o ce ado (en na anja). También se indica en la igu a 7 la di ección de lujo
del agua. La sala donde se g aba la base de da os p esen a 29dB de aislamien o acús ico
espec o a uido aé eo.
En el mon aje se usan 4 me os de ube ías de 3.81 cm de diáme o. La bomba
usada es una bomba cen í uga ci culado a de agua modelo ALP800 de 0.5 caballos de
po encia y con 2925 RPM ( e oluciones po minu o), equi alen e a 48.75 Hz. El ode e
de la bomba p esen a 7 álabes.
Tesis Doc o al
Uni e sidad de Las Palmas de G an Cana ia
298
Figu a 7: Esquemá ico de la sala donde se ealiza la g abación de la
base de da os jun o con la posición y esquema del mon aje.
En las igu as 8 y 9 se obse an o os de la bomba cen í uga modelo ALP800
u ilizada pa a gene a la base de da os de la p esen e Tesis. En la igu a 7 se mues a la
bomba de agua sin desmon a y en la igu a 8 se mues a el ode e ( o o de la izquie da)
y la olu a.
Figu a 8: Fo os de la bomba de agua modelo ALP 800.
Di ección del lujo
A ances en moni o ización p e en i a de maquina ia a a és de señales de audio y ib ación
Uni e sidad de Las Palmas de G an Cana ia 305
mé odos del es ado del a e. En la igu a 14 se obse an el mé odo p opues o (mé odo 1)
y los demás mé odos del es ado del a e que se e alúan: en el mé odo 2 se ex aen de la
señal de ib ación en el dominio del iempo (señal T) las ca ac e ís icas es adís icas y de
ene gía [8], en el mé odo 3 se ex aen de la señal de ib ación demodulada en ampli ud
(señal AM) usando la écnica de análisis de en ol en e comúnmen e u ilizada en
cojine es [39] y en el mé odo 4 se ex aen de la señal demodulada en ampli ud ob enida
a pa i del ope ado Teage -Kaise (señal TK-AM) [41].
Figu a 14: Esquema de la expe imen ación en diagnós ico de allos de
cojine es.
Una ez ob enidas las ca ac e ís icas en cada uno de los mé odos, se e alúa la
habilidad de dichas ca ac e ís icas pa a disc imina en e no malidad y di e en es ipos
de allos de cojine es. Pa a ello se ealiza un es udio de ele ancia de las ca ac e ís icas
usando un mé odo de selección de ca ac e ís icas denominado sequen ial loa ing
o wa d selec ion (SFFS) [44]. El mé odo de selección de ca ac e ís icas usado
selecciona un subg upo de ca ac e ís icas que mejo es esul ados dan a la ho a de
disc imina en e las di e en es clases del p oblema (en es e caso en e no mal, allo en

Tesis Doc o al
Uni e sidad de Las Palmas de G an Cana ia
306
cojine e ex e no, allo en cojine e in e no y allo en odamien o). La selección de
ca ac e ís icas se aplica a ias eces pa a ob ene inalmen e las ca ac e ís icas en o den
de ele ancia. Finalmen e se usan dos clasi icado es, uno de edes neu onales y o o de
máquinas de sopo e ec o ial de mínimos cuad ados (LS-SVM) pa a e alua las
medidas o denas po ele ancia. Los esul ados ob enidos con el mé odo p opues o
mejo an los ob enidos con los demás mé odos compa ados [47].
Es impo an e ecalca que a la ho a de e alua las ca ac e ís icas con los
clasi icado es se di ide la base de da os en un conjun o de en enamien o o mado po el
70% de las mues as de cada clase y en un conjun o de es o mado po el 30% de las
mues as es an es. Con el conjun o de en enamien o se ajus an los pa áme os de los
clasi icado es usando la écnica de alidación c uzada k- old con k = 3. En el caso de las
edes neu onales, donde el clasi icado usado iene la capa de en ada, la capa de salida
y una capa ocul a, se ajus a el núme o de neu onas de la capa ocul a. En el caso del
clasi icado LS-SVM que ha sido p og amado pa a que u ilice como ke nel unciones de
base adial Gaussiana, se ajus a el pa áme o de egula ización y el ancho de banda de la
unción de base adial Gaussiana.
Una ez aplicada la écnica p opues a a la base de da os de cojine es que
p esen a allos en di e en es pa es de los mismos (base de da os Case Wes e n) [30] se
aplica la misma écnica a la base de da os de deg adación del helicóp e o [31] en la que
un cojine e si uado en una pa e bas an e inaccesible de uno de los enes p incipales de
ansmisión de un helicóp e o Black-Hawk iene un allo en uno de sus elemen os
odan es al inal de un es de esis encia. En la igu a 15 se mues a el esquema de la
p opues a aplicado a la deg adación del cojine e. Los esul ados usando el mé odo
A ances en moni o ización p e en i a de maquina ia a a és de señales de audio y ib ación
Uni e sidad de Las Palmas de G an Cana ia 307
p opues o mejo an esul ados ob enidos con esa misma base de da os a la ho a de
de ec a el comienzo del allo [49].
Figu a 15: Diag ama de la aplicación del mé odo p opues o a la
deg adación de un cojine e.
En la p ime a pa e del capí ulo 4 se encuen a el desa ollo del mé odo
p opues o y de la expe imen ación ealizada jun o con los esul ados ob enidos.
Me odología pa a de ección e iden i icación de allo en cojine es
Pa a desa olla un índice que si a pa a iden i ica la se e idad del allo en cojine e se
ealiza p ime amen e un es udio del es ado de la écnica en écnicas de iden i icación de
allos (es deci , écnicas que son capaces de de e mina la se e idad del allo en
cues ión). Basándonos en ese es udio se p opone la aplicación de la ans o mada
wa ele packe a la señal de ib ación seguida de la complejidad de Lempel-Zi pa a la
iden i icación de allos en la ca a ex e na y en la ca a in e na de cojine es. El mé odo es
capaz de de ec a allo en los elemen os de odamien o, pe o no de de e mina la
se e idad del mismo. Además, el mé odo p opues o es capaz de segui de o ma
Tesis Doc o al
Uni e sidad de Las Palmas de G an Cana ia
308
monó ona la e olución de allos en ca a ex e na. Se compa a el mé odo p opues o con
o os es mé odos usados en la li e a u a: ku osis, complejidad Lempel-Zi aplicada a
la señal de ib ación sin p ep ocesa , ku osis aplicado al nodo de máxima ene gía de la
wa ele packe [37], [43]. También se ealiza un es udio de la obus ez del mé odo
p opues o en e a uido blanco Gaussiano.
La igu a 16 mues a un diag ama del mé odo p opues o pa a la iden i icación de
allos en la ca a ex e na e in e na de cojine es.
Figu a 16: Mé odo p opues o pa a la iden i icación de allos en
cojine es.
P ime o la señal de ib ación se descompone usando la ans o mada wa ele
packe con ni el de descomposición 3 (donde hay 8 nodos). La wa ele mad e usada
pa a la descomposición es la Daubechies 6 (db6) ( e igu a 17) po que iene una o ma
Descomposición
“Wa ele packe ”
(3e ni el, 8 nodos)
Cálculo de la ene gía
ela i a po nodo
Selección del nodo con
mayo ene gía
En ada: Señal de
ib ación
Salida: ca ac e ís ica
ex aída de la señal de
en ada
Cálculo de la
complejidad Lempel-
Zi
Recons ucción de la
señal del nodo kmax
A ances en moni o ización p e en i a de maquina ia a a és de señales de audio y ib ación
Uni e sidad de Las Palmas de G an Cana ia 309
de onda impulsi a, igual que la o ma de onda de las señales de ib ación de cojine es
con allos pun uales. Una ez descompues a la señal de ib ación se econs uye cada
una de las señales de los nodos. Se calcula luego la ene gía de esas 8 señales con
espec o a la ene gía de la señal de ib ación o iginal. Finalmen e se selecciona el nodo
con mayo ene gía ela i a. Pues o que se ha usado una wa ele mad e con o ma
impulsi a, el nodo con mayo ene gía se á aquel donde se encuen e el allo si lo
hubiese. Así la ans o mada wa ele elimina pa es de la señal indeseada y se cen a en
el allo. Una ez ob enido el nodo con mayo ene gía se calcula la complejidad de
Lempel-Zi sob e la señal econs uida del nodo de mayo ene gía. Con la complejidad
de Lempel-Zi se e alúa, como su p opio nomb e indica, la complejidad de la señal
ob enida.
Figu a 17: Fo ma de onda de la wa ele mad e Daubechies 6.
El mé odo p opues o se aplica a la base de da os Case Wes e n [30] pues o que
es a base de da os p esen a señales de ib ación de cojine e con allos de di e en e
se e idad (se e idad le e, media y al a) en ca a ex e na, ca a in e na y odamien o.
Asimismo ambién se aplican los o os es mé odos de la li e a u a mencionados
an e io emen e [42]. Los esul ados ob enidos demues an que el mé odo p opues o da
mejo es esul ados [52] sob e odo cuando se añade uido blanco Gaussiano a las
señales pa a comp oba lo obus o del algo i mo p opues o en e al uido.
0 2 4 6 8 10 12
-1.5
-1
-0.5
0
0.5
1
1.5
Tesis Doc o al
Uni e sidad de Las Palmas de G an Cana ia
310
El mé odo p opues o ambién se aplica a las o as dos bases de da os de
cojine es, las base de da os de deg adación de cojine es del IMS [32] y del helicóp e o
UH-60 (Black Hawk) [31]. En la base de da os del IMS el cojine e p esen a un allo de
ca a ex e na a lo la go del es un- o- ailu e. El mé odo p opues o ob iene mejo es
esul ados que los o os es mé odos con los que se ha compa ado. En el caso de la base
de da os del helicóp e o, donde el allo se p oduce en el elemen o odan e del cojine e,
los esul ados no son an sa is ac o ios y el mé odo p opues o se compo a de o ma
simila el mé odo de ex ae la complejidad de Lempel-Zi de la señal de ib ación sin
p e-p ocesa .
En la segunda pa e del capí ulo 4 se encuen a el desa ollo del mé odo
p opues o pa a la iden i icación de allos en la ca a in e na y la ca a ex e na de cojine es
así como la expe imen ación ealizada jun o con los esul ados ob enidos.
Me odología usada pa a el diagnós ico de allos en bomba de agua usando señales
de audio y ib ación
Una ez gene ada la base de da os de audio y ib ación de la bomba cen í uga, nos
cen amos en el diagnós ico de los allos gene ados en la bomba. El obje i o p incipal
en es e caso es de e mina si con el audio como uen e de in o mación se es capaz de
dis ingui en e los di e en es allos de la bomba cen í uga. De igual o ma se p e ende
compa a los esul ados con los ob enidos usando señales de ib ación.
La me odología usada pa a el diagnós ico de allos en bomba de agua usando
señales de audio y ib ación se explica con la ayuda de la siguien e igu a ( igu a 18).

A ances en moni o ización p e en i a de maquina ia a a és de señales de audio y ib ación
Uni e sidad de Las Palmas de G an Cana ia 311
Figu a 18: Esquema de la me odología lle ada a cabo pa a cuan i ica
la habilidad de las ca ac e ís icas ex aídas de la señal de audio y
ib ación en diagnós ico de allos en bombas cen í ugas.
Ex acción de ca ac e ís icas
En el bloque de Ex acción de ca ac e ís icas se ex ae un conjun o de ca ac e ís icas
an o de la señal de ib ación como de la señal de audio. En ealidad, es o se hace po
senso . Pues o que hay 4 senso es (2 de audio y 2 de ib ación) se ex aen
ca ac e ís icas de las señales p o enien es de los 4 senso es.
Las ca ac e ís icas que se ex aen son ca ac e ís icas usadas en la li e a u a del
diagnós ico de allos en bombas cen í ugas usando señales de ib ación y de audio.
Pa a iden i ica las ca ac e ís icas p ime o se debe ealiza un es udio de las écnicas de
p ocesado de la señal usadas en bombas cen í ugas pa a el diagnós ico de allos. Es
necesa io menciona que pues o que la mayo ía del diagnós ico de allo en bombas
cen í ugas se ealiza con señales de ib ación así como con señales de p esión, se
encuen an escasas ca ac e ís icas que se hayan ex aído usando el audio. Po ello las
mismas ca ac e ís icas que se ex aen de la señal de ib ación se ex aen ambién de la
señal de audio.
Basándose en la inspección isual de las señales ob enidas en la base de da os
gene ada, se p oponen una se ie de nue as medidas no u ilizadas an e io men e en el
diagnós ico de allos de bombas cen í ugas.
Ex acción de
ca ac e ís icas
Clasi icación
Selección de
ca ac e ís icas
selec ion
Señal
Diagnós ico de
clasi icación
Tesis Doc o al
Uni e sidad de Las Palmas de G an Cana ia
312
Una ez iden i icadas las ca ac e ís icas del es ado del a e y p opues as nue as
medidas, és as se deben implemen a .
Pa a ex ae las ca ac e ís icas cada obse ación (o mues a) de la base de da os
gene ada se di ide en segmen os, ambién llamados amas, de 8192 mues as cada uno,
equi alen es a 371.5 ms. Las amas se oman una cada es po lo que si cada
obse ación du a 59 segundos, enemos 53 amas po obse ación. Po cada una de las
amas se ex ae el conjun o de ca ac e ís icas implemen ado. An es de ex ae las
ca ac e ís icas a cada una de las obse aciones se le qui a la media y se no maliza en e
-1 y 1.
Una ez ex aídas las ca ac e ís icas po obse ación, se calcula el p omedio de
cada una de las ca ac e ís icas po obse ación.
Selección de ca ac e ís icas
Una ez que se ex aen las ca ac e ís icas po obse ación de la base de da os, el
obje i o es iden i ica aquellas más ap opiadas po senso que mejo disc iminen en e
los di e en es es ados de la bomba cen í uga. El p oceso de selecciona ca ac e ís icas
de un conjun o de ca ac e ís icas se denomina selección de ca ac e ís icas. Es e p oceso
es impo an e pues o que la selección de ca ac e ís icas que apo en in o mación
elacionada con el allo y el desca e de ca ac e ís icas que no apo en in o mación
mejo a la asa de éxi o en el diagnós ico. Así pues, en el segundo bloque de la igu a 18
se ealiza la selección de ca ac e ís icas y un análisis de ele ancia, el cual se explicó
an e io men e cuando se explicó la me odología pa a el diagnós ico de allos en
cojine es. Simplemen e eco da que la écnica u ilizada pa a la selección de
A ances en moni o ización p e en i a de maquina ia a a és de señales de audio y ib ación
Uni e sidad de Las Palmas de G an Cana ia 313
ca ac e ís icas se denomina sequen ial loa ing o wa d selec ion (SFFS) [44]. La azón
po la que se ha elegido es a écnica de selección de ca ac e ís icas es po que es un
mé odo de e minis a y po lo an o se ob end án los mismos esul ados cuando se epi a.
De igual o ma el cos e compu acional de la écnica es bajo.
Después de aplica la selección de ca ac e ís icas y de ealiza el análisis de
ele ancia se ob end á un conjun o de ca ac e ís icas po senso .
E aluación de las ca ac e ís icas: Clasi icación
El bloque de clasi icación de la igu a 18 se u iliza pa a e alua la capacidad de las
ca ac e ís icas a la ho a de disc imina en e los di e en es es ados de la bomba
cen í uga. Así que el clasi icado nos da á una asa de éxi o en la clasi icación de los
di e en es es ados de uncionamien o de la máquina.
El clasi icado clasi ica las obse aciones en di e en es unidades de clasi icación
ambién llamadas clases con la in o mación que ob iene las ca ac e ís icas. Las unidades
de clasi icación co esponden con los di e en es es ados o condiciones de la máquina
que se quie an clasi ica . En la p esen e Tesis, se ha decidido implemen a dos
con igu aciones di e en es, así que las unidades de clasi icación conside adas son las
siguien es (en e pa én esis es á el ac ónimo de la clase):
 8 unidades de clasi icación: es ado no mal (NOR), allo en el pla o del ode e
(PLA), allo en el leading edge (LED), allo en el ailing edge (TED), allo en
Tesis Doc o al
Uni e sidad de Las Palmas de G an Cana ia
314
la goma del sellado (SEA), a ena añadida al agua (SAN), a ena y papel añadidos
conjun amen e al agua (SAP) y bolas de PVC añadidas al agua (PVC).
 17 unidades de clasi icación. En es e caso se ienen en cuen a las se e idades en
los casos de o u a del pla o, del allo leading edge y del allo ailing edge. Las
17 unidades de clasi icación son: es ado no mal (NOR), allo en el pla o del
ode e con una po ción de pla o eliminada (PLA1), con dos po ciones de pla o
eliminadas (PLA2) y con es po ciones de pla o eliminadas (PLA3), allo en el
leading edge (LED1) de 5 mm de amaño, de 10 mm de amaño (LED2) y de 15
mm de amaño (LED3), allo en el ailing edge de 5 mm de amaño (TED1), de
10 mm de amaño (TED2) y de 15 mm de amaño (TED3), allo en la goma del
sellado (SEA), a ena añadida al agua (SAN), a ena y papel añadidos
conjun amen e al agua (SAP) y bolas de PVC añadidas al agua (PVC).
Al igual que en el caso del diagnós ico de allos en cojine e se usan dos
clasi icado es: un clasi icado de ed neu onal y o o de máquinas de sopo e ec o ial
de mínimos cuad ados (LS-SVM). El clasi icado de ed neu onal iene es uc u a
eed o wa d y un algo i mo de esilien backp opaga ion. El clasi icado LS-SVM usa
unciones Gaussianas de base adial. Al igual que en el caso de diagnós ico de allos en
cojine es, la me odología pa a e alua los di e en es casos (8 unidades de clasi icación y
17 unidades de clasi icación) consis e en di idi la base de da os en un conjun o de
en enamien o o mado po el 70% de las obse aciones de cada una de las clases de la
base de da os y en un conjun o de es o mado po el 30%. El conjun o de
en enamien o a su ez se usa pa a ajus a los pa áme os de los clasi icado es. El
p ocedimien o se epi e 20 eces y se p omedian los esul ados.
A ances en moni o ización p e en i a de maquina ia a a és de señales de audio y ib ación
Uni e sidad de Las Palmas de G an Cana ia 321
uido ecuencial (2 ca ac e ís icas), ene gía de uido ecuencial (1 ca ac e ís ica),
medidas no lineales (7 ca ac e ís icas). Pa e de es as con ibuciones han sido publicadas
en con e encias [51], [52]. En el capí ulo 5 se de allan las medidas apo adas.
Apo ación o iginal 7: e aluación de las medidas en señales de audio pa a el
diagnós ico de allos en bomba cen í uga
La capacidad pa a disc imina en e di e en es es ados de la maquina ia (sin allo y con
di e en es ipos de allos) de las ca ac e ís icas ex aídas se ha e aluado usando dos
clasi icado es, uno basado en edes neu onales y o o basado en máquinas de sopo e
ec o ial de mínimos cuad ados (LS-SVM). Los esul ados de la e aluación de las
ca ac e ís icas usando la señal de audio son muy sa is ac o ios, llegándose a asas de
acie o del 96.53% en la disc iminación en e 8 es ados (es ado no mal de
uncionamien o y 7 es ados de allo) usando sólo 7 ca ac e ís icas seleccionadas y de
88.29% en la disc iminación en e 17 es ados (es ado no mal y 10 es ados de allo en
los que se sepa an allos po se e idades) usando 16 ca ac e ís icas seleccionadas.
También en el capí ulo 5 se desc ibe la e aluación de las medidas y los
esul ados ob enidos en dicha e aluación an o pa a las señales de audio como pa a las
señales de ib ación po senso indi idual.
Apo ación o iginal 8: es udio de la usión de señales de audio y ib ación en el
diagnós ico de allos en bomba cen í uga
Finalmen e se ha ealizado un es udio de la usión de las señales de audio y ib ación
pa a de e mina si su unión mejo a el diagnós ico de allos en la aplicación e bomba

Tesis Doc o al
Uni e sidad de Las Palmas de G an Cana ia
322
cen í uga. Los esul ados ob enidos son sa is ac o ios llegándose a aumen a la asa de
acie o en el diagnós ico del allo has a un 100%. En el capí ulo 6 se explican las
écnicas de usión usadas y se mues an los esul ados ob enidos.
Po úl imo, y como ya se comen ó en la in oducción, du an e el anscu so de la
p esen e Tesis la doc o anda siguió con la in es igación en señales de oz. En es a á ea
se ealizan las siguien es apo aciones:
Apo ación o iginal 9 ( oz): p opues a de ca ac e ís icas basadas en dinámica no
lineal pa a la de ección de pa ologías la íngeas del sis ema onado
Se p oponen una se ie de ca ac e ís icas basadas en dinámica no lineal ( eo ía del caos)
pa a la de ección de pa ologías del sis ema onado usando la señal de oz. F u o de es e
es udio se gene a un a ículo en e is a de índica de impac o [53].
Apo ación o iginal 10 ( oz): p opues a de ca ac e ís icas basadas en dinámica no
lineal y de medidas de complejidad pa a la de ección de emociones a a és de la
señal de oz
Se p oponen una se ie de ca ac e ís icas basadas en dinámica no lineal, así como
medidas de complejidad pa a la disc iminación de di e en es es ados de emoción usando
la señal de oz. Se gene a de es e abajo dos a ículos en e is as con índice de impac o
[54], [55].
En el anexo de la Tesis se hace un esumen de ambos abajos en oz,
explicando la me odología usada, las ca ac e ís icas ex aídas y los esul ados ob enidos.
Se debe menciona que pa e de las medidas p opues as en oz ambién han sido
p opues as en es a Tesis pa a el diagnós ico de allos en bombas cen í ugas.
A ances en moni o ización p e en i a de maquina ia a a és de señales de audio y ib ación
Uni e sidad de Las Palmas de G an Cana ia 323
6 Conclusiones
El obje i o de la p esen e Tesis es mejo a la asa de acie o en los sis emas de
moni o ización del es ado de la máquina en diagnós ico e iden i icación (se e idad del
allo) de allos usando señales de audio y ib ación y cen ándonos en dos aplicaciones
(cojine es y bombas cen í ugas) con especial én asis en la e apa de ex acción de
ca ac e ís icas y en el uso de señales de audio como uen e de in o mación. Basándonos
en la in es igación ealizada en las dos á eas de aplicación de la Tesis, cojine es y
bombas cen í ugas, y en los esul ados ob enidos podemos conclui que an o el uso de
audio como uen e de in o mación como el uso de écnicas no lineales a la ho a de
ex ae ca ac e ís icas mejo a la asa de éxi o en las dos aplicaciones en las que se ha
cen ado la Tesis: cojine es y bombas cen í ugas.
En la aplicación de cojine es se han usado señales de ib ación como uen e de
in o mación pues o que se han usado bases de da os públicas disponibles en In e ne con
mues as de cojine es uncionando en es ado no mal y con allos pun uales en las
di e en es pa es de los cojine es. Se han p opues o dos mé odos basados en écnicas no
lineales pa a el diagnós ico de allos en la ca a ex e na, en la ca a in e na y en los
elemen os odan e de cojine es y pa a la iden i icación de allos en la ca a ex e na y en
la in e na. Con el segundo mé odo p opues o se ha gene ado un índice que sigue de
o ma monó ona la deg adación de un cojine e con allo no mal a allo en la ca a
ex e na. Los mé odos p opues os se han compa ado con mé odos de la li e a u a y los
esul ados ob enidos han sido sa is ac o ios, mejo ando los esul ados. Además, en el
segundo mé odo p opues o se ha ealizado un es udio de la obus ez del algo i mo en e
a uido Gaussiana. El es udio demues a que el algo i mo p opues o p esen a más
Tesis Doc o al
Uni e sidad de Las Palmas de G an Cana ia
324
obus ez al uido blanco Gaussiano que los mé odos de la li e a u a con los que se ha
compa ado. Po lo an o se puede conclui que el uso de écnicas no lineales pa a el
diagnós ico e iden i icación de allos en cojine es es una línea de in es igación
p ome edo a.
La aplicación de la bomba cen í uga se ha usado pa a explo a el uso de la señal
de audio en el diagnós ico de allos de bombas cen í ugas. La moni o ización de
maquina ia en gene al, y en pa icula la e e ida a bombas de agua y a cojine es, u iliza
en la mayo ía de los casos la señal de ib ación como uen e de in o mación. El uso de
audio ( ango de ecuencias [0-20kHz]) ha sido menos explo ado, po es o en la
p esen e Tesis nos hemos cen ado en su es udio. Pa a ello se necesi a ob ene una base
de da os de audio ob enida de una bomba cen í uga. Pues o que no hay ninguna
disponible, se ha gene ado una base de da os p opia en la que se adquie en señales de
audio y ib ación de o ma simul ánea de una bomba cen í uga uncionando con
no malidad y a la que se gene an di e en es ipos de allos. Un conjun o de
ca ac e ís icas del es ado del a e de diagnós ico de allos en bombas cen í uga se
ex aen de las señales de ib ación ob enidas y se aplican ambién a las señales de
audio. También se p oponen un conjun o de 31 nue as medidas ex aídas en el dominio
de la ecuencia, en el dominio de los ceps um y en el dominio no lineal. Se ealiza una
selección de ca ac e ís icas y luego se e alúan po senso con dos clasi icado es. Los
esul ados mues an asas de acie o al as y bas an e simila es en senso es de audio y de
ib ación. Se ob ienen asas de éxi o del 99.82% y del 99.59% pa a los senso es de
audio y del 99.91% y 99.59% pa a los senso es de ib ación en el caso de disc imina
en e 8 es ados de la bomba cen í uga. En el caso de disc imina en e 17 es ados se
ob ienen asas de éxi o del 87.48% y 94.23% pa a los senso es de audio y del 86.71% y
A ances en moni o ización p e en i a de maquina ia a a és de señales de audio y ib ación
Uni e sidad de Las Palmas de G an Cana ia 325
87.30% pa a los senso es de ib ación. De los esul ados ob enidos, se puede conclui
que el audio puede se usado como uen e de in o mación en el diagnós ico de allos de
la bomba cen í uga u ilizada. De odas o mas, es necesa io ealiza más in es igación
al espec o con di e en es bombas cen í ugas.
En la Tesis ambién se ha ealizado un es udio de la combinación de las señales
de audio y ib ación ob enidas en la bomba cen í uga pa a de e mina si la usión de la
in o mación de ambas señales mejo a la asa de acie o. Las asas de éxi o se
inc emen a on signi ica i amen e a la ho a de usiona un senso de audio y o o de
ib ación en el caso de disc imina en e 17 es ados de la bomba cen í uga. En es e
caso se log ó una asa de acie o del 99.55%, bas an e supe io a los 94.93% ob enidos
po uno de los senso es de audio. En el caso de disc imina en e 8 condiciones de la
bomba cen í uga, se llega on a ob ene asas del 100% a la ho a de usiona un senso
de audio y o o de ib ación. De los esul ados del es udio de la usión se puede conclui
que si bien los esul ados indi iduales po senso son buenos, la usión aumen a la asa
de acie o. Es e aumen o es especialmen e e iden e en el caso del diagnós ico de 17
condiciones de la bomba cen í uga.
Tesis Doc o al
Uni e sidad de Las Palmas de G an Cana ia
326
7 Re e encias
[1] Vilela, R.M., Me ôlho, J. C., & Ca doso, J. C. (2004, May). Machine and
insdus ial moni o iza ion sys em by analysis o acous ic signa u es. In
Elec o echnical Con e ence, 2004. MELECON 2004. P oceedings o he 12 h IEEE
Medi e anean (Vol. 1, pp. 277-279). IEEE.
[2] Ja dine, A. K., Lin, D., & Banje ic, D. (2006). A e iew on machine y diagnos ics
and p ognos ics implemen ing condi ion-based main enance. Mechanical sys ems
and signal p ocessing, 20(7), 1483-1510.
[3] Vach se anos, G., Lewis, F. L., Roeme , M., Hess, A. & Wu, B. In elligen Faul
Diagnosis and P ognosis o Enginee ing Sys ems, 1s ed. Hoboken, New Je sey:
John Wiley & Sons, Inc., 2006.
[4] Pal idge, B. (2002). Thesis and disse a ion w i ing: an examina ion o published
ad ice and ac ual p ac ice. English o Scien i ic Pu poses, 21(2), 125–143.
[5] Bayda , N., & Ball, A. (2003). De ec ion o gea ailu es ia ib a ion and acous ic
signals using wa ele ans o m. Mechanical Sys ems and Signal P ocessing, 17(4),
787-804.
[6] Bell, R.N., McWilliams, D.W., O’Donnell, P., Singh, C., & Wells, S. J. (1985).
Repo o la ge mo o eliabili y su ey o indus ial and comme cial ins alla ions
Pa I and II. IEEE T ansac ions on Indus y Applica ions, 21(4), 853-872.
[7] Immo illi, F., Bellini, A., Rubini, R., & Tassoni, C. (2010). Diagnosis o bea ing
aul s in induc ion machines by ib a ion o cu en signals: A c i ical
compa ison. IEEE T ansac ions on Indus y Applica ions, 46(4), 1350-1359.
[8] Saman a, B., Al-Balushi, K. R., & Al-A aimi, S. A. (2006). A i icial neu al
ne wo ks and gene ic algo i hm o bea ing aul de ec ion. So Compu ing, 10(3),
264-271.
[9] McFadden, P. D., & Smi h, J. D. (1984). Vib a ion moni o ing o olling elemen
bea ings by he high- equency esonance echnique—a e iew. T ibology
in e na ional, 17(1), 3-10.
[10] Choi, Y. C., & Kim, Y. H. (2007). Faul de ec ion in a ball bea ing sys em using
minimum a iance ceps um. Measu emen Science and Technology, 18(5), 1433.
[11] Zhang, Y., & Randall, R. B. (2009). Rolling elemen bea ing aul diagnosis
based on he combina ion o gene ic algo i hms and as ku og am.Mechanical
Sys ems and Signal P ocessing, 23(5), 1509-1517.
[12] An oni, J., Randall, R. B. (2006). The spec al ku osis: applica ion o he
ib a o y su eillance and diagnos ics o o a ing machines. Mechanical Sys . and
Signal P ocess., 20, 308-331.
[13] Wang, D., Pe e , W. T., & Tsui, K. L. (2013). An enhanced Ku og am me hod
o aul diagnosis o olling elemen bea ings. Mechanical Sys ems and Signal
P ocessing, 35(1), 176-199.
[14] Peng, Z. K., Pe e , W. T., & Chu, F. L. (2005). A compa ison s udy o imp o ed
Hilbe –Huang ans o m and wa ele ans o m: applica ion o aul diagnosis o
olling bea ing. Mechanical sys ems and signal p ocessing, 19(5), 974-988.
[15] Ali, J. B., Fnaiech, N., Saidi, L., Chebel-Mo ello, B., & Fnaiech, F. (2015).
Applica ion o empi ical mode decomposi ion and a i icial neu al ne wo k o
au oma ic bea ing aul diagnosis based on ib a ion signals. Applied Acous ics, 89,
16-27.

A ances en moni o ización p e en i a de maquina ia a a és de señales de audio y ib ación
Uni e sidad de Las Palmas de G an Cana ia 327
[16] Yang, D. M., S onach, A. F., MacConnell, P., & Penman, J. (2002). Thi d-o de
spec al echniques o he diagnosis o mo o bea ing condi ion using a i icial
neu al ne wo ks. Mechanical sys ems and signal p ocessing, 16(2), 391-411.
[17] Wang, W. J., Wu, Z. T., & Chen, J. (2001). Faul iden i ica ion in o a ing
machine y using he co ela ion dimension and bispec a. Nonlinea Dynamics,
25(4), 383-393.
[18] Shiba a, K., Takahashi, A., & Shi ai, T. (2000). Faul diagnosis o o a ing
machine y h ough isualisa ion o sound signals. Mechanical Sys ems and Signal
P ocessing, 14(2), 229-241.
[19] Cudina, M. (2003). De ec ion o ca i a ion phenomenon in a cen i ugal pump
using audible sound. Mechanical Sys ems and Signal P ocessing 17(6), 1335–1347.
[20] Singh, V., Meena, N. (2009, Janua y). Engine Faul Diagnosis using DTW,
MFCC and FFT. In P oc. o he 1s In . Con . on In elligen Human Compu e
In e ac ion, (pp. 83-94).
[21] Mansoo , A. B., Ahmed, H., Mahmood, Z. (2009, Feb ua y). Design and
De elopmen o an Au oma ed Acous ic based Je Engine Pe o mance E alua o .
In P oc. o he IEEE In . Con . on Indus ial Technol. (pp. 1-4).
[22] Wu, J. -D., Chuang, C. -Q. (2005). Faul diagnosis o in e nal combus ion
engines using isual do pa e ns o acous ic and ib a ion signals. NDT&E In ., 38,
605-614.
[23] Al Thobiani, A., Gu, F., & Ball, A. (2010 June 22-24). The moni o ing o
ca i a ion in cen i ugal pumps based on he analysis o ib o-acous ic
measu emen s. In CM 2010 and MFPT 2010, UK.
[24] Al-Hashmi, S.A. (2012, Sep embe 23-26). Spec um Analysis o Acous ic
Signals o Ca i a ion De ec ion. In 2012 IEEE Symposium on Indus ial
Elec onics and Applica ions (ISIEA2012), Bandung, Indonesia.
[25] Saeid Fa okhzad, H.A., (2013). Acous ic Based Ca i a ion De ec ion o
Cen i ugal Pump by Neu al Ne wo k. Jou nal o Mechanical Enginee ing and
Technology, 1(1), 1-5.
[26] Al-Hashmi, S.A. (2013, Oc obe ). S a is ical analysis o acous ic signal o
ca i a ion de ec ion. In e na ional Jou nal o Eme ging Technology and Ad anced
Enginee ing, 3(4).
[27] Benko, U., Pe o cic, J., Ju icic, D., Ta ca , J., Rejec, J. (2005). An app oach o
aul diagnosis o acuum cleane mo o s based on sound analysis. Mechanical Sys .
and Signal P ocess., 19, 427-445.
[28] Janja asji , S., Ocak, H., Lopa o, K. A. (2008). Bea ing condi ion diagnosis and
p ognosis using applied nonlinea dynamical analysis o machine ib a ion signal. J.
o Sound and Vib a ion, 317, 112-126.
[29] De i, S., Si a, L., Shanke N. R., & P abaka an, K. (2010). A Compa a i e
S udy be ween Vib a ion and Acous ic Signals in HTC Cooling Pump and Chilling
Pump. In . J. o Eng. and Technol., 2.
[30] Case Wes e n Bea ing Da a Cen e ,
h p://cseg oups.case.edu/bea ingda acen e /home. Úl ima isi a: No iemb e 2014.
[31] h p://qsun.eng.ua.edu/cpw_web_new/in o.h m. Úl ima isi a: Ene o 2014
[32] Lee, J., Qiu, H., Yu, G., Lin, J., & Rexno d Technical Se ices (2007). Bea ing
Da a Se . IMS, Uni e si y o Cincinna i. NASA Ames P ognos ics Da a Reposi o y,
[h p:// i.a c.nasa.go /p ojec /p ognos ic-da a- eposi o y], NASA Ames, Mo e
Field, CA."
Tesis Doc o al
Uni e sidad de Las Palmas de G an Cana ia
328
[33] Webpage Explo e he Wo ld o Piping
h p://www.we mac.o g/equipmen /pumps_gene al.h ml. Las isi ed: 19 No embe
2014
[34] "Pump P inciples Manual: global aining educa ion and de elopmen p og am"
eds. Ches e on.
[35] Zhao, X., Zuo, M. J., & Pa el, T. H. (2012). Gene a ing an indica o o pump
impelle damage using hal and ull spec a, uzzy p e e ence-based ough se s and
PCA. Measu emen Science and Technology, 23(4), 045607.
[36] Zhao, X. M., Hu, Q. H., Lei, Y. G., & Zuo, M. J. (2010). Vib a ion-based aul
diagnosis o slu y pump impelle s using neighbou hood ough se
models.P oceedings o he Ins i u ion o Mechanical Enginee s, Pa C: Jou nal o
Mechanical Enginee ing Science, 224(4), 995-1006.
[37] Saman a, B., Al-Balushi, K. R., & Al-A aimi, S. A. (2006). A i icial neu al
ne wo ks and gene ic algo i hm o bea ing aul de ec ion. So Compu ing, 10(3),
264-271.
[38] Randall, R. B. (2004). S a e o he a in moni o ing o a ing machine y-pa 1.
Sound and ib a ion, 38(3), 14-21.
[39] McFadden, P. D., & Smi h, J. D. (1984). Vib a ion moni o ing o olling elemen
bea ings by he high- equency esonance echnique—a e iew. T ibology
in e na ional, 17(1), 3-10.
[40] Kaise , JF., On a simple algo i hm o calcula e he ene gy o a signal. In:
In e na ional con e ence on acous ics, speech, and signal p ocessing; 1990. p. 381–
384.
[41] Li, H., Fu, L., & Zhang, Y. (2009, Ap il). Bea ing aul s diagnosis based on
eage ene gy ope a o demodula ion echnique. In IEEE In e na ional Con e ence
on Measu ing Technology and Mecha onics Au oma ion (ICMTMA'09). (Vol. 1,
pp. 594-597).
[42] Al-Ghamd, A. M., & Mba, D. (2006). A compa a i e expe imen al s udy on he
use o acous ic emission and ib a ion analysis o bea ing de ec iden i ica ion and
es ima ion o de ec size. Mechanical sys ems and signal p ocessing, 20(7), 1537-
1571.
[43] Yan, R., & Gao, R. X. (2004). Complexi y as a measu e o machine heal h
e alua ion. Ins umen a ion and Measu emen , IEEE T ansac ions on, 53(4), 1327-
1334.
[44] Pudil, P., No o ičo á, J., & Ki le , J. (1994). Floa ing sea ch me hods in ea u e
selec ion. Pa e n Recogni ion Le e s, 15(11), 1119-1125.
[45] Kunche a, L. I. (2004). Combining pa e n classi ie s: me hods and algo i hms.
John Wiley & Sons.
[46] Hen íquez, P., Alonso, J. B., Fe e , M. A., & T a ieso, C. M. (2014). Re iew o
au oma ic aul diagnosis sys ems using audio and ib a ion signals. IEEE
T ansac ions on Sys ems, Man, and Cybe ne ics: Sys ems, 44(5), 642-652.
[47] Hen íquez, P., Alonso, J. B., Fe e , M. A., & T a ieso, C. M. (2013).
Applica ion o he Teage –Kaise ene gy ope a o in bea ing aul diagnosis. ISA
ansac ions, 52(2), 278-284.
[48] Hen íquez, P., Alonso, J. B., Fe e , M. A., T a ieso, C. M. (2011, May-June).
Applica ion o Highe O de S a is ics o Teage -Kaise Ene gy T ans o med
Vib a ion Signal o Bea ing Faul Diagnosis. In P oc. o he 24 h In . Cong ess on
Condi ion Moni o ing and Diagnos ics Enginee ing Managemen (pp. 265-274),
S a ange , No way.
A ances en moni o ización p e en i a de maquina ia a a és de señales de audio y ib ación
Uni e sidad de Las Palmas de G an Cana ia 329
[49] Hen íquez, P., Whi e, P., Alonso, J. B., Fe e . M. A. (2011, Oc obe ).
Applica ion o Teage -Kaise Ene gy Ope a o o he Analysis o Deg ada ion o a
Helicop e Inpu Pinion Bea ing. In P oc. o he In e na ional Con e ence
Su eillance 6 (pp. 265-274), Uni e si y o Technology o Compiègne, F ance.
[50] Hen íquez, P., Alonso, J. B., Fe e , M. A., T a ieso, C. M. (2014, Sep embe ).
Deg ada ion assessmen in bea ings wi h ou e ace aul . In 27 h In e na ional
Cong ess o Condi ion Moni o ing and Diagnos ic Enginee ing Managemen ,
B isbane, Aus alia.
[51] Hen íquez, P.; Alonso, J. B.; Fe e , M. A.; T a ieso, C. M.; Gómez, G. (2012).
Faul diagnosis using audio and ib a ion signals in a ci cula ing pump. Jou nal o
Physics: Con e ence Se ies 25 h In . Cong ess on Condi ion Moni o ing and
Diagnos ic Enginee ing, 364, 012135.
[52] Hen íquez, P.; Alonso, J. B.; Fe e , M. A.; T a ieso, C. M. (2014). Applica ion
o complexi y measu es o pump aul diagnosis. In 27 h In . Cong ess on Condi ion
Moni o ing and Diagnos ic Enginee ing, B isbane, Aus alia.
[53] Hen íquez, P., Alonso, J. B., Fe e , M., T a ieso, C. M., Godino-Llo en e, J., &
Díaz-de-Ma ía, F. (2009). Cha ac e iza ion o heal hy and pa hological oice
h ough measu es based on nonlinea dynamics. IEEE T ansac ions on Audio,
Speech, and Language P ocessing, 17(6), 1186-1195.
[54] Hen íquez, P., Alonso, J. B., Fe e , M. A., T a ieso, C. M., & O ozco-
A oya e, J. R. (2013). Global Selec ion o Fea u es o Nonlinea Dynamics
Cha ac e iza ion o Emo ional Speech. Cogni i e Compu a ion, 5(4), 517-525.
[55] Hen íquez, P., Alonso, J. B., Fe e , M. A., T a ieso, C. M., O ozo-A oya e, J.
R. (2014). Nonlinea Dynamics Cha ac e iza ion o Emo ional Speech.
Neu ocompu ing, 132, 126-135.