Identification of the major trends in current maintenance policies using lDA methodology for semantic analysis of the published research results

International Journal for Quality Research 18(2) 587–602 ISSN 1800-6450 Corresponding author: Goran D. Putnik Email: [email protected] 587 Goran D. Putnik1 Leonilde Varela Pedro Pinheiro Sara Maia Cátia Alves Article info: Received 20.10.2023. Accepted 03.02.2024. UDC – 004.4`414 DOI – 10.24874/IJQR18.02-17 IDENTIFICATION OF THE MAJOR TRENDS IN CURRENT MAINTENANCE POLICIES USING LDA METHODOLOGY FOR SEMANTIC ANALYSIS OF THE PUBLISHED RESEARCH RESULTS Abstract: This paper presents a semantic analysis of a collection of articles on maintenance policies through the Latent Dirichlet Allocation (LDA) methodology. The research comprises two periods. The first, from 2000 and 2018, aims to reveal the habits of organizational maintenance policies before the second period, from 2019 to 2023, and verify the differences in current trends in maintenance policies compared to the past. With technologies associated with cyber-physical systems and Industry 4.0, a change in organizational models and maintenance policies is expected. The results obtained through the LDA methodology show changes in the maintenance design, namely the intensification of the maintainability concept, schedule optimization, and the integration of technologies enablers of smart production. Keywords: Maintenance Policies, Semantic analysis, Latent Dirichlet Allocation (LDA) 1. Introduction Maintenance activities are essential for ensuring systems' reliability and smooth operation (Kamalraj, Verma, & Joshi, 2023) in industrial context, where numerous sources of uncertainty exist and is critically important to treat these sources of uncertainty properly (Kim, An, & Choi, 2017). This article, aims to apply a Machine Learning (ML) technique the Latent Dirichlet Allocation (LDA) methodology, for extraction of “hidden” information from scientific publications addressing maintenance policies. The hypothesis of the article is that there are differences of the research topics regarding maintenance policies in last years in comparison with the earlier period. The collection of articles for analyses by the LDA methodology is extracted from Scopus data base (scopus.com), by the keywords "maintenance policies" AND (manufacturing OR industry). The collection extracted is divided in 2 collections of articles for 2 periods, from year 2000 to 2018 and from 2019 to 2023. The analyses of 2 different periods allows to observe whether current trends in maintenance policies diverges from previous ones and whether new concepts emerge in maintenance policies scientific field. The LDA methodology is applied for a semantic analysis for identification of relevant topics associated to the maintenance policies. The model considers the words in the text that are most plausible for a topic. Putnik et al., Identification of the major trends in current maintenance policies using IDA methodology for semantic analysis of the published research results 588 The words considered are from the articles abstracts. Following, the article presents a traditional literature review related to maintenance policies, in chapter 2. Next, the LDA methodology is presented in chapter 3. After, the results obtained by application of LDA methodology are shown in chapter 4 with commentaries. The article finish with conclusions in chapter 5. This article, is an extended version of the conference paper presented to “The Quality Festival 2023”, which took place 24-27 May 2023, at Faculty of Engineering, University of Kragujevac, Serbia. 2. Literature review "Maintenance strategies" is often considered a synonym of "maintenance policies". In the literature, "maintenance strategies" and "maintenance policies" are used interchangeably (Santos, Cavalcante, & Wu, 2022). Many may suspect that production and scheduling problems deal with systems remaining available forever. This fallacy is far from reality, as there are times when machines break down and/or require repair to be maintained (Mobley, 2002), implying that production planning and scheduling cover an infinite number of models. In general, the production process planning/scheduling focuses on a single machine, modelling its deterioration and, simultaneously, allowing the analysis and configuration of more complex group technology (Wang, 2002), such as achieving integrated production and maintenance scheduling, optimizing maintenance time and avoiding costs associated with machine failures (Ladj, Varnier, & Tayeb, 2016). However, the joint optimization of maintenance and production plans deserves greater attention and more in-depth studies (Bajestani, Banjevic, & Beck, 2014; Rastgar, Rezaeian, Mahdavi & Fattahi, 2023; Liao, Pan, & Xi, 2010; Pan, Liao, & Xi, 2010), in addition to taking on integrated maintenance scheduling and machine degradation modelling, where they are taken into consideration separately (Mourtzis & Vlachou, 2018; Zhai, Riess, & Reinhart), aiming for perfect reliability, and availability for complex systems can be simplified by scheduling the maintenance staff schedules. During the last decades, maintenance has changed significantly taking into account the new manufacturing paradigms (Vrignat, Kratz, & Avila, 2022). Nowadays organizations try to implement smart and efficient maintenance processes to raise productivity, increase customer satisfaction, reduce machine downtime, and decrease delays (Dakkak, Irhirane, & Bounit, 2019; Hajej, Nidhal, Anis, & Bouzoubaa, 2020). Maintenance policies aim, for example, to reduce costs, increase reliability or improve the quality of services and products. Increasingly complex manufacturing systems have made maintenance more relevant to the industry with the increase in dynamics of production activities (Sezer, Romero, Guedea, Macchi, & Emmanouilidis, 2018). With the challenges that occur due to stochastic failures and multistate deterioration in production systems, and with the prospect of increased flexibility in production scheduling management (Sun & Geng, 2019), corrective maintenance (CM) and preventive maintenance (PM) policies attempt to keep different production systems functional (Cui, Lu, Li, & Han, 2018). CM was the first to be applied. CM is done only after the failure occurs and is detected. Over time, PM began to be implemented. PM has specific schedules for performing maintenance (Matyas, Nemeth, Kovacs, & Glawar, 2017; Nemeth, Ansari, Sihn, Haslhofer, & Schindler, 2018). Although several researches have focused on flexible workshop scheduling and often assume a static environment in production, with machines unoccupied and available for production, it does not reflect realistic production environments, with dynamic International Journal for Quality Research, 18(2), 587–602, 2024, doi: 10.24874/IJQR18.02-17 589 environments, with constant changes in the entry of new orders (Gao et al., 2015), variations in processing times and changes in expiration dates (Baykasoğlu, Madenoğlu, & Hamzadayı, 2020). As machines age, the likelihood of the need for downtime caused by scheduled PM tasks increases (Zandieh, Khatami, & Rahmati, 2017; Zhang, Liao, Zeng, Shi, & Zhao, 2021), as well as occasional machine failures, which trigger CM. In 1970, a Japanese concept of Total Productive Maintance (TPM) emerged, with the main objective of increasing the productivity of existing equipment, expanding PM to become more productive maintenance and complementary to other global strategies (Zlatić, 2019). The implementation of TPM allows you to maximize the use of all equipment, eliminating losses resulting from failures, maximizing the efficiency of equipment operation and building a comprehensive PM system (Rasztorf, Urbaniak, & Zimon, 2023). PM encompasses time-based preventive maintenance (TBPM) and condition-based maintenance (CBM) (Ben-Daya & Rahim, 2000). TBPM is a traditional version of PM as maintenance is also predetermined for specific schedules, CBM, on the other hand, selects maintenance activities considering the true conditions of the system, i.e., it monitors degradation and predicts failures. However, this subcategory encompasses predictive maintenance (PdM), which uses data and time understanding to signal the possibility of a failure and to prevent downtime as well. Predictive maintenance (PdM) and predictive analytics represent one of the recent innovations that provokes curiosity in researchers and industry. PdM skills are essential with the rise of complex industrial processes (Shrivastava, Singhal, & Bhuvana, 2023). There are relevant developments concerning complex ML models (Putnik, Manupati, Pabba, Varela, & Ferreira, 2021), and considering Remaining Useful Lifetime (RUL) (Manupati et al., 2019). Real-time optimization models are the most required to be applied in practice, the expression "real time" is related to the system response to stops and unforeseen events that do not interrupt the production process (Harmonosky & Robohn, 1991). Manufacturing systems are exceedingly dynamic and prone to interruptions and unforeseen events. Optimization models allow to assist the manufacturing system to withstand and adjust to these unforeseen events in real time (Cheng et al., 2018; Zheng et al., 2018). These systems include smart objects that represent a sub-system of broader concepts of advanced manufacturing systems, such as Industrial Internet of Things (IIoT), ubiquitous and cloud manufacturing systems, cyber-physical systems, digital factories, factories of the future, and Industry 4.0, which is related to the phenomena of Big Data and associated technologies and techniques (Putnik et al., 2015). It emphasizes that without an information system based on the application of modern information technology, it cannot achieve an efficient and effective management of maintenance, which must be open and adaptable to grow with the information system and included processes, facilitating completion of this information system after a few years (Arsovski, Pavlovic, & Arsovski, 2008). The data produced by machines can improve maintenance in different areas (Kim et al., 2017) based on the IIoT and Artificial Intelligence (AI) (Zonta et al., 2020), allowing to improve fault detection, quality control and decision support (Shivajee, Singh, & Rastogi, 2019). With the accelerated development of the IIoT, Industry 4.0, Big Data and Cloud Computing (Chen, Mao, & Liu, 2014; Xu, He, & Li, 2014), companies are looking for strategies to differentiate themselves from competitors and improve quality of its processes, products and services, which requires the collection of data on machines, and through its analysis, quality flaws can be corrected (Köksal, Batmaz, & Testik, 2011). Some Putnik et al., Identification of the major trends in current maintenance policies using IDA methodology for semantic analysis of the published research results 590 studies (Vaidya et al., 2018; Pierdicca et al., 2017; Jones et al., 2019) agree on the evolution of technology in 9 pillars, and these pillars are necessary to provide the improvement in the areas pointed out. The nine pillars encompass the creation of information from horizontal and vertical systems, big data and analytics, cloud computing, the IIoT, simulation, Autonomous Robots, Cyber Security, Augmented Reality, and Additive Manufacturing. These pillars transform the factory, meaning that it becomes fully autonomous, integrated and optimizes the production process. Using previous technologies, Prognosis and Health Management (PHM) has emerged as a higher-level adaptation of CBM, intending to achieve more efficient maintenance that applies RUL for decision support (Vrignat et al., 2022). However, the PHM does not end with the RUL forecast. System health management goes beyond failure time predictions and supports optimal maintenance and logistics decisions, considering available resources, the operational context and the economic consequences of different failures. Health management is the process that enables timely and optimal maintenance actions to be taken based on diagnostic and prognostic results, available resources and operational demand (Fink et al., 2020). Uncertainty in forecasts is an inevitable part (Kim et al., 2017), requiring the use of advanced predictive tools that reduce the impacts of these uncertainties (Lee, Lapira, Bagheri, & Kao, 2013). PHM systems occupy a central place in the opportunities created by cyber-physical systems and Industry 4.0, aiming to reduce the probability of extreme failure events (Biggio & Kastanis, 2020). However, most maintenance optimization works do not consider PHM. Eventually, this gap arises because most studies apply detection, diagnostic and prognostic models to individual components of different types, while optimization systems require the analysis of integrated systems (Pinciroli, Baraldi, & Zio, 2023). PHM can be considered a holistic approach to a management system (Fink et al., 2020), with integration into complex industrial manufacturing systems being a differentiating element of the algorithms applied in PHM compared to other methodologies (Lee, Bagheri, & Kao, 2014). With Industry 4.0, production planning is a priority, as achieving efficiency and speed in maintenance requires organizations to have a profitable production system (Rødseth, Schjølberg, & Marhaug, 2017). Therefore, the maintenance activities should be oriented and adapted to maximize the availability of installed systems and equipment in line with customer needs, with efficiency, effectiveness and sustainability, considering costs, production quality, environmental protection, safety and legality (Moutinho & Oliveira, 2015). 3. Methodology In this semantic analysis, two sets of article abstracts extracted from Scopus (scopus.com) were used. The first collection contains 335 articles, published from 2000 to 2018 and the second one contains 214 articles, published from 2019 to 2023. The keywords used in the search were: “maintenance policies” AND (manufacturing OR industry). In this article, the data analysis method chosen was topic modelling. Topic modelling automatically maps hidden topics in a set of text through ML algorithms. Among the different ML techniques, LDA methodology is chosen. LDA is a set of topic modelling techniques that automatically finds latent structures in a corpus of unstructured documents using word frequency statistics (Jelodar et al., 2019). A word is seen as a basic unit of data measurement, the corpus as a set of documents that encompass the data set, and each document as a sequence of words. The International Journal for Quality Research, 18(2), 587–602, 2024, doi: 10.24874/IJQR18.02-17 591 different words in a corpus represent the vocabulary and the topics represent the probability distributions of the vocabulary words (Vayansky & Kumar, 2020). Therefore, it is expected that certain words stand out more frequently in documents, as each document is reflexively related to a certain topic. Topics are semantic sets formulated by words that stand out in documents and in the corpus (Blei, 2012). Fitting the LDA model requires an assessment of the results. Among the bestknown measures are relevance (Sievert & Shirley, 2014) and coherence (Stevens, Kegelmeyer, Andrzejewski, & Buttler, 2012). Given the reasons presented by Sievert and Shirley (2014), we adopted the value of λ=0,6 for the measure of relevance for defining the topics. Considering the size of the corpus, and because it is a coherence measure that considers the total size of the samples, the choice of the coherence CUmass measure seems to be the most appropriate (Stevens et al., 2012). The CUmass coherence metric (Mimno, Wallach, Talley, Leenders, & McCallum) is given by Equation 1: ∑ ∑ ( ) (1) Where, in a document D, D(wi, wj) is the document frequency of the set of words wi and wj and D(wj) is the document frequency of only wj. When fitting LDA models, a CUmass value close to zero is desirable. Considering the constitution and size of the corpus of the two periods analyzed, the following criteria were established in the LDA algorithm: - Reduce the search to 4 topics due to the small size of the corpus and guarantee an intertopic distance that avoids the creation of topic clusters. - Use 30 relevant terms per topic. Although there is the possibility of increasing redundancy in the terms that define the topics, this redundancy makes it possible to accentuate the validity of the name attributed to the topic. - Prepare the documents belonging to the corpus. Considering that LDA is an ML technique, the algorithm must include specific routines to eliminate inaccuracies, duplications, redundancies and inconsistencies during data processing (Sharma, 2023). In this process, the use of stopwords is restricted. All these steps were supported by the NLTK library (Phand & Chakkarwar), with the most common words, such as prepositions and numbers, being eliminated. In order not to lose the expression “4.0”, “4.0” was replaced by "fourdotzero" in the texts. There is a routine that textualizes the acronyms with the information present in the articles and eliminates any text that is in parentheses. The final parts with additional information from the editors of the articles is removed. All words became singular. Only nouns were considered. - Run a routine with 1000 simulations to create n-grams for the LDA model. The selection criterion for the most favorable simulation is the CUmass measure. The methodology is summarized in figure 1. Figure 1. Proposed methodology for the automatic identification of the major trends in current maintenance policies. Putnik et al., Identification of the major trends in current maintenance policies using IDA methodology for semantic analysis of the published research results 592 4. Results Figure 2 shows the plot of the CUmass coherence scores. The best results of 1000 simulations confirms that the selection of the number of 4 topics gives the value of coherence CUmass closest to zero as required, for each collection. a) b) Figure 2. Topic coherence index plot for data between a) 2000 and 2018 and b) 2019 and 2023. The results, in Figure 3 to 10, present intertopic distance maps that show the 4 topics and the most relevant terms, considering λ=0.6. The Principal Component (PC) analysis projection compiles the similarity, allowing to obtain the intertopic distance (Chuang, Ramage, Manning, & Heer, 2012). PC1 embodies the most relevant variation in the data, while PC2 represents the second most relevant variation in the data. Although they represent the same data set, PC2 is independent of PC1. PC analysis allows the visualization of topics in a two-dimensional orthogonal graph, simplifying subsequent interpretation. Figure 3. Intertopic Distance Map and the list of the Top-30 most relevant terms in topic 1 (2000-2018). International Journal for Quality Research, 18(2), 587–602, 2024, doi: 10.24874/IJQR18.02-17 593 Figure 4. Intertopic Distance Map and the list of the Top-30 most relevant terms in topic 2 (2000-2018). Figure 5. Intertopic Distance Map and the list of the Top-30 most relevant terms in topic 3 (2000-2018). Putnik et al., Identification of the major trends in current maintenance policies using IDA methodology for semantic analysis of the published research results 594 Figure 6. Intertopic Distance Map and the list of the Top-30 most relevant terms in topic 4 (2000-2018). Figure 7. Intertopic Distance Map and the list of the Top-30 most relevant terms in topic 1 (2019-2023). International Journal for Quality Research, 18(2), 587–602, 2024, doi: 10.24874/IJQR18.02-17 595 Figure 8. Intertopic Distance Map and the list of the Top-30 most relevant terms in topic 2 (2019-2023). Figure 9. Intertopic Distance Map and the list of the Top-30 most relevant terms in topic 3 (2019-2023). Putnik et al., Identification of the major trends in current maintenance policies using IDA methodology for semantic analysis of the published research results 602 Zonta, T., Da Costa, C. A., da Rosa Righi, R., Lima, M. J. de, da Trindade, E. S., & Li, G. P. (2020). Predictive maintenance in the Industry 4.0: A systematic literature review. Computers & Industrial Engineering, 150, 106889. Goran D. Putnik Department of Production and Systems Engineering, University of Minho, Portugal ALGORITMI Research Centre /LASI, University of Minho, Portugal [email protected] ORCID 0000-0003-3378-6866 Leonilde Varela Department of Production and Systems Engineering, University of Minho, Portugal ALGORITMI Research Centre /LASI, University of Minho, Portugal ORCID 0000-0002-2299-1859 Pedro Pinheiro ALGORITMI Research Centre /LASI, University of Minho, Portugal ORCID 0000-0001-7849-4109 Sara Maia ALGORITMI Research Centre /LASI, University of Minho, Portugal Cátia Alves ALGORITMI Research Centre /LASI, University of Minho, Portugal; 2Ai – School of Technology, IPCA, Portugal LASI – Associate Laboratory of Intelligent Systems, Guimarães, Portugal ORCID 0000-0001-6445-5475 View publication stats