A Survey on Data Mining for Data-Driven Industrial Assets Maintenance

Abstract

This survey presents a comprehensive review of data-driven approaches for industrial asset maintenance, emphasizing the use of data mining and machine learning techniques, including deep learning, for condition-based and predictive maintenance. It examines 534 references from 1995 to 2023, along with three additional articles from 2024 on natural language processing and large language models in industrial maintenance. The study categorizes two main techniques, four specialized approaches, and 27 methodologies, resulting in over 100 variations of algorithms tailored to specific maintenance needs for industrial assets. It details the data types utilized in the industrial sector, with the most frequently mentioned being time series data, event timestamp data, and image data. The survey also highlights the most frequently referenced data mining algorithms, such as the proportional hazard model, expert systems, support vector machines, random forest, autoencoder, and convolutional neural networks. Additionally, the survey proposes four level classes of asset complexity and studies five asset types, including mechanical, electromechanical, electrical, electronic, and computing assets. The growing adoption of deep learning is highlighted alongside the continued relevance of traditional approaches such as shallow machine learning and rule-based and model-based techniques. Furthermore, the survey explores emerging trends in machine learning and related technologies, identifies future research directions, and underscores their critical role in advancing condition-based and predictive maintenance frameworks.

1. Introduction

The industry faces a huge transformation due to the high automation of systems and processes. This transformation started decades ago with isolated automated systems to solve large process problems with a relatively small amount of digital assets. Nowadays, the variety of digital assets and the complex networking behind them require better and automated monitoring systems, optimized maintenance policies, and advanced analysis to support engineers, technicians, and operators in their daily activities. Many studies are turning their attention to solutions with a focus on resilient architectures, expert systems for decision-making support, and advanced analytic techniques and tools for data mining (DM) and machine learning (ML).

Although DM and ML are generally studied as different areas of computer science, for didactic purposes, this work defines DM as encompassing not only the processes of extracting valuable insights and patterns from large datasets, but also the application of various ML algorithms that leverage these data for predictive and analytical purposes. In short, DM serves as a broad term that includes both the traditional techniques of knowledge extraction and the implementation of the most sophisticated ML algorithms, such as deep learning (DL), to maximize usefulness of available data.

Interesting reviews and surveys explore the field of DM with a focus on maintenance optimization, fault identification, and prognosis such as [1,2,3,4,5,6,7,8,9,10,11,12,13]. Regarding studies with a taxonomy presentation of data science and analytic applications in industrial maintenance, the most comprehensive papers to date are [14,15,16,17,18]. Nevertheless, their focus was based only on mechanic and electromechanic assets.

The revision of survey, review, and taxonomy papers led to the formulation of the following research questions, which define and convey the scope and aim of this paper:

• What techniques are commonly used for analysis in the field of industrial assets maintenance, and how have they evolved over the years?
• Which algorithms are most frequently applied, and what are their specific applications, advantages, and disadvantages?
• Are all types of assets used in the industrial sector adequately studied and analyzed? Is there a review that compiles information across all asset types?
• What categories of asset complexity are typically studied, and which techniques are best suited for analyzing these categories of assets?
• What types of data are considered for analysis? Do the techniques support processing data from assets with multiple variables?
• What are the trending topics in machine learning and the role of emerging technologies in advancing industrial maintenance?

This survey aims to advance the understanding and application of data-driven approaches for industrial asset maintenance by addressing critical gaps and broadening the scope of existing research. The primary contributions of this article are as follows:

• Introduction of data mining (DM) and machine learning (ML) techniques applied to the maintenance of industrial assets;
• Expansion of the study scope to include electric, electronic, and computing assets, in addition to the commonly studied mechanic and electromechanic assets, while proposing four categories of assets based on complexity;
• Analysis of the purposes and applications of DM and ML, with a focus on the most frequently referenced algorithms for condition-based and predictive maintenance;
• Summary of data types and working parameters used in data mining and their relevance to industrial assets maintenance analysis.

To the best of the authors’ knowledge, no existing survey comprehensively addresses all the topics covered in this research study. Therefore, the authors believe that the developed content and findings presented here are of real value for researchers, experts, and decision-makers aiming to venture into the field of industrial asset maintenance.

This work is organized as follows: Section 2 presents the materials and methods used in the research, detailing the methodologies and techniques employed for the search of research studies under interest. Section 3 delves into the application of data mining in industrial maintenance, exploring various maintenance strategies and the purposes of data mining in condition-based and predictive maintenance. Within this section, subsections cover topics such as degradation identification, failure detection, health index forecasting, data pre-processing, synthetic data generation, transfer learning, and objective optimization. Section 4 defines the data types employed in data mining, asset varieties and complexity, providing insights into the types of data utilized and the complexity of assets under consideration. Section 5 presents topics on learning models, types, and algorithms used in industrial maintenance, categorizing them based on their methodologies and techniques. This section also discusses the most cited learning algorithms by technique type. A general discussion is conducted in Section 6. Section 7 presents trending topics of machine learning and emerging technologies in industrial maintenance. Finally, the conclusions of Section 8 synthesize the key findings from the study, highlighting the importance of understanding asset characteristics, employing appropriate analytical methods, leveraging diverse learning techniques, and utilizing suitable tools for effective maintenance strategies. Lastly, as future work directions, suggestions of potential areas for further research and development in industrial maintenance and data mining are presented.

2. Materials and Methods

The primary purpose of this investigation was to identify the most relevant literature on the application of data mining (DM) in industrial maintenance contexts, including industrial asset diagnosis, fault detection, prognosis, and various other applications. To start the exploration, a strategic approach was adopted, starting with the formulation of specific keywords. These keywords were applied within the domains of article titles, keywords, or abstracts, specifically targeting the period spanning from year 2015 to May 2023. Publications and research studies that were not in English and not directly related to the industrial field were excluded from this research. The search revealed a final count of 534 works deemed relevant and included for consideration. Among the considered studies, there are 495 specific research, 22 reviews, 8 surveys, 6 taxonomies, and 3 patents. Additionally, to cover the topics of natural language processing and large language models in the field of industrial maintenance, 3 more articles were considered, being 2 of them specific research studies and 1 review, all of them of the year 2024. Totalizing 537 articles for the general repository.

Table 1. Defined keywords under consideration and the number of occurrences in industrial maintenance research within the overall repository.

Ítem	Keyword	Number of Mentions
1	Classification	211
2	Prediction	204
3	Fault diagnosis	202
4	Monitoring	192
5	Neural networks	188
6	Maintenance	180
7	Failure analysis	165
8	Predictive maintenance	117
9	Optimization	116
10	Machine learning	109
11	Artificial intelligence	100
12	Deep learning	75
13	Regression	62
14	Prognosis	63
15	Condition monitoring	60

shows part of the keywords under consideration for the exploration of technical studies, focusing on the top 15 most referenced, which are presented as a summary that details their number of mentions within the general repository. The other considered keywords are: remaining useful life (RUL) with 40 mentions, clustering with 40 mentions, preventive maintenance with 32 mentions, forecasting with 30 mentions, deterioration with 26 mentions, industry 4.0 with 26 mentions, internet of things (IoT) with 20 mentions, condition-based maintenance with 21 mentions, big data with 14 mentions, data mining with 11 mentions, industrial systems with 9 mentions, early fault detection with 7 mentions, maintenance planning with 3 mentions, reliability block diagram (RBD) with 2 mentions, failure mode effect and criticality analysis (FMECA) with 2 mentions, natural language processing with 1 mention, language models with 1 mention, and linguistic text mining with 1 mention.

The selection criteria specified the inclusion of studies since the year 2015, whereas the search for specific research studies encompassed a broader timeframe, commencing from the year 1995 up to the year 2023. Articles from 2024 are not considered due to the specific topics defined for the search criteria. Figure 1 illustrates the distribution of the 534 works identified over the entire period under review. There has been a marked increase in studies since 1995, with a rising trend continuing until 2018. Following this peak, a decline in the number of studies is observed in the subsequent years.

The research repository, consisting of 537 articles, was primarily constructed using the following academic databases: Science Direct with 205 studies, IEEE Xplore with 129 studies, Springer Link with 52 studies, MDPI with 39 studies, Sage Journals with 31 studies. An additional 81 studies were sourced from complementary platforms identified through Google Scholar, including Taylor & Francis Online, ACM Digital Library, ResearchGate, IOP Science, HINDAWI, ASME, arXiv, CORE, EXTRICA, Google Patents, WILEY Online Library, IET, IOS Press, BGU, IGI Global, OnePetro, DIVA, WIRES, CERN Document Server, SPIE Digital Library, ARC, AIP Publishing, PHIL Papers, Emerald, and BazTech.

For enhanced readability, refer to

Table 2. Used acronyms and their descriptions.

Acronym	Definition	Acronym (Cont.)	Definition (Cont.)
AB	AdaBoost	k-NN	k-nearest neighbors
AE	Autoencoder	LDA	Linear discriminant analysis
AI	Artificial Intelligence	LGR	Logistic regression
ANN	Artificial neural network	LLM	Large language models
ANOVA	Analysis of variance	LR	Linear regression
AR	Augmented reality	LSTM	Long short-term memory
ARIMA	Autoregressive integrated moving average	MB	Model-based
ARM	Association rule mining	MC	Monte Carlo
BMP	Bitmap image file	ML	Machine learning
BPTT	Back propagation through time	ML-DAE	Multi-level denoising autoencoder
CBM	Condition-based maintenance	MP	Markov process
CD-SDAE	Cross-domain stacked denoising autoencoders	NLP	Natural language processing
CHMM	Coupled hidden Markov hodel	PCA	Principal component analysis
CMYK	Color model: cyan, magenta, yellow, key	PdM	Predictive maintenance
CNN	Convolutional neural network	PHM	Proportional hazard model
DAE	Denoising autoencoder	PLC	Programmable logic controller
DBMS	Database management system	PM	Preventive maintenance
DBN	Deep beliefnetwork	PNG	Portable network graphics
DBSCAN	Density-based spatial clustering of applications with noise	PR	Polynomial regression
DDAE	Deep denoising autoencoder	RB	Rule-based
DL	Deep learning	RBM	Restricted Boltzmann machine
DM	Data mining	RF	Random forest
DNN	Deep neural network	RGB	Color model: red, green, blue
DPAE	Deep autoencoder	RIPPER	Repeated incremental pruning to produce error reduction
DRL	Deep reinforcement learning	RL	Reinforcement learning
DT	Decision tree	RNN	Recurrent neural network
ES	Expert systems	RP	Renewal process
FCM	Fuzzy C-means method	RUL	Remaining useful life
FDP	Fault diagnosis and prognosis	SAE	Sparse autoencoder
FFT	Fast Fourier transform	SCADA	Supervisory control and data acquisition
FMEA	Failure mode and effects analysis	SCAE	Stacked contractive autoencoder
GAN	Generative adversarial network	SKAE	Stacked autoencoder
GB	Gradient boosting	SKDAE	Stacked denoising autoencoder
GIFFT	General interpolated fast Fourier transform	SKSAE	Stacked sparse autoencoder
GIS	Geographic information systems	SML	Shallow machine learning
GNN	Graph neural network	SSAE-AL	Stacked sparse autoencoder with adaptation layer
GRU	Gated recurrent units	SVM	Support vector machine
HDD	Hard disk drive	SVR	Support vector regression
HHT	Hilbert-Huang transform	TbIAS	Text-based intelligent assistance system
HI	Health indicator	TIFF	Tag image file format
HIL	Health indicator learning	TM	Text mining
HMM	Hidden Markov model	TSA	Transformer for self-attention
IC	Integrated circuits	VAE-KDE	Variational autoencoder and kernel density estimation
IoT	Internet of things	VM	Virtual machine
JPEG	Joint photographic experts group	WP	Wiener process
HPO	Hyperparameter optimization	WPT	Wavelet packets transform
K-M	K-means	XAI	Explainable artificial intelligence

, which lists all the acronyms used in this work.

3. Data Mining Application in Industrial Maintenance

This section provides a technical overview of the integration of data mining (DM) techniques in industrial maintenance. It focuses on the strategic application of DM to enhance maintenance practices, beginning with an analysis of maintenance strategies and their integration with DM methodologies. The section further elucidates the role of DM in condition-based and predictive maintenance, emphasizing its significance in enabling proactive interventions and improving operational efficiency. Additionally, it examines the range of DM techniques utilized in industrial maintenance, detailing the sophisticated methods used for insight extraction, failure prediction, and decision-making optimization. A dedicated segment clarifies the foundational models and methods underlying these techniques, providing a comprehensive understanding of their application in industrial maintenance.

3.1. Exploration of Maintenance Strategies for the Application of Data Mining

The utilization of DM in modern industrial maintenance focuses on three widely recognized models: preventive, condition-based, and predictive maintenance.

These models stand as pillars frequently referenced in academic and research studies within the domain of maintenance strategies.

The summary of mentions of these maintenance strategies in the research studies is as follows: preventive maintenance with 22 mentions [4,6,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38], condition-based maintenance with 16 mentions [4,7,13,20,39,40,41,42,43,44,45,46,47,48,49,50], and predictive maintenance with 105 mentions [1,3,4,5,6,7,9,10,14,16,21,23,24,25,26,27,28,29,30,31,32,36,42,43,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87,88,89,90,91,92,93,94,95,96,97,98,99,100,101,102,103,104,105,106,107,108,109,110,111,112,113,114,115,116,117,118,119,120,121,122,123,124,125,126,127,128,129]. Corrective maintenance is infrequently mentioned in the literature concerning the application of data mining in industrial maintenance, with only 6 references [4,21,29,63,98,130]. Typically, these mentions serve to provide an introductory definition and to highlight its disadvantages in comparison to other maintenance models.

DM in industrial maintenance encompasses the application of advanced analytical techniques to extract valuable insights from maintenance-related data. It involves leveraging historical data, sensor readings, and patterns to enhance maintenance decision-making processes.

Table 3. Overview of definitions, approaches, benefits, and challenges of maintenance strategies considered for data mining application in industrial maintenance.

Maintenance Strategy	Objective	Approach	Benefits	Challenges
Preventive maintenance (PM)	PM aims to minimize the likelihood of equipment failure by performing routine maintenance or inspection tasks at scheduled intervals [19,22].	Maintenance activities are performed proactively, often based on fixed time intervals or predefined criteria [76].	PM can improve system availability and minimize losses due to breakdowns and failures, extend equipment lifespan, and ensure reliable, safe, and efficient operations [19].	Includes over-maintenance, increased costs, and difficulties in addressing issues that arise between scheduled maintenance [19].
Condition-based maintenance (CBM)	CBM focuses on periodically monitoring equipment (near real-time) to detect signs of degradation or impending failure [20]. Therefore, condition-based monitoring is a key factor of this model [6].	Continuous monitoring of key parameters, such as temperature, vibration, acoustic emissions, chemical presence, or fluid levels, enables data-driven decisions regarding maintenance [6,20].	CBM reduces maintenance costs by addressing issues only when necessary, maximizes equipment uptime, minimizes the risk of catastrophic failures, and enables safe operation [20].	Includes implementing effective monitoring systems and accurately interpreting data [2,55].
Predictive maintenance (PdM)	PdM takes CBM a step further by using advanced analytics, machine learning, or predictive models to forecast equipment failures or health conditions [21,27].	Predictive models analyze historical data, sensor readings and patterns to predict maintenance requirements [24].	PdM minimizes downtime, reduces maintenance costs, extends asset lifespan, and optimizes resource allocation [21,25,26,34].	Robust data collection, storage, analysis capabilities, as well as the need for skilled data scientists and engineers [23].

summarizes, with the corresponding references, these three key maintenance strategies, highlighting their objectives, approaches, benefits, and challenges. Each strategy provides distinct advantages while presenting specific challenges, particularly in data management and advanced analytics.

The studies considered in this research focus on industrial assets maintenance, primarily emphasizing CBM and PdM. Within these maintenance paradigms, DM plays a significant role, following the objectives detailed in the subsection.

3.2. Data Mining Application Purposes in Condition-Based and Predictive Maintenance

Considering condition-based and predictive maintenance strategies, data mining is applied in industrial maintenance mainly for the analysis purposes detailed next. Each analysis purpose and associated terminologies commonly used interchangeably to denote the same concept matter are explained.

3.2.1. Degradation Identification

Degradation refers to the gradual deterioration or decline in the performance, condition, or reliability of a piece of equipment or system over time. It is a progressive process where the performance of the equipment slowly declines due to wear and tear, aging, or other factors. During degradation, the equipment may still be operational, but it may no longer function at its optimal level. A more detailed explanation is presented in [16], based on the ”European Standard of Maintenance” terminology, EN 13306. Maintenance strategies often aim to detect and address degradation early to prevent it from progressing to a failure state. Related terminology and other terms used in research, in alphabetical order, include the following:

A

Anomaly detection: The process of identifying abnormal or unexpected deviations in machinery or equipment performance data. It involves using advanced monitoring and analysis techniques to detect potential faults or irregularities that could lead to breakdowns or failures in industrial systems [7,10,13,16,54,55,57,61,63,75,81,84,95,131,132,133,134,135,136,137,138];

B

Degradation: The progressive decline in the quality, efficiency, or functionality of industrial machinery or infrastructure due to different factors such as usage, environmental conditions, or lack of maintenance [1,3,7,14,16,20,22,28,29,30,31,39,40,41,43,45,49,52,53,57,61,64,67,69,73,74,83,89,90,91,92,93,99,104,137,139,140,141,142,143,144,145,146,147,148,149,150,151,152,153,154,155,156,157,158,159,160,161,162,163,164,165,166];

C

Deterioration: The gradual decline or worsening of machinery, components, or infrastructure within an industrial environment. It encompasses the wear and tear, corrosion, and other factors that lead to reduced performance or functionality over time [1,6,7,10,19,31,43,59,94,109,130,148,167];

D

Health condition: Refers to the overall operational state or status of machinery, equipment, or systems in an industrial setting. It involves assessing the well-being and functionality of these assets to ensure they operate efficiently and effectively [3,7,21,23,41,42,55,58,131,141,143,165,168,169,170,171,172,173,174,175,176,177];

E

Health monitoring: The continuous or periodic assessment of key parameters, performance indicators, or predictive maintenance data to track the condition of machinery and equipment. This procedure helps to identify potential issues before they lead to breakdowns or failures [1,2,7,64,75,120,152,159,171,178,179,180,181];

F

Health state: The specific operational condition or status of machinery, equipment, or systems at a particular time. It provides insights into the overall operational health and performance of industrial assets [1,3,12,14,23,28,41,45,141,161,181,182,183,184,185,186,187];

G

Incipient fault: The initial or early stages of a fault or issue in machinery or systems are not easily detectable in general, but they have the potential to develop into more severe problems if not addressed promptly [7,14,47,106,188,189,190,191];

H

Performance evaluation: The assessment of how effectively machinery, systems, or processes function within an industrial setting. It involves measuring key performance metrics to ensure optimal efficiency and productivity [50,84,148,192,193];

I

Reliability evaluation: The analysis and assessment of the probability that industrial systems, machines, or components will operate without failure within specified operating conditions. It involves evaluating the reliability and dependability of equipments to minimize downtime [194].

3.2.2. Failure Detection

Failure occurs when a component of the equipment or the system ceases to function as intended. It is the point at which the equipment or system cannot perform its intended function due to a critical malfunction or breakdown. A detailed definition and the formulation of a fault diagnosis are presented in [2]. Failure is usually the result of prolonged degradation that has not been addressed through maintenance activities. Related terminology and other words used in research, in alphabetical order, include the following:

A

Fault detection: The process of actively monitoring and detecting deviations, abnormalities or faults within industrial equipment or systems. It relies on sensors, monitoring tools, data analysis and algorithms to identify deviations from normal operational behavior and trigger alerts or notifications for maintenance or further investigation [6,7,30,34,39,45,51,54,57,60,61,75,80,167,169,174,178,188,193,195,196,197,198,199,200,201,202,203,204,205,206,207,208,209,210,211,212,213,214,215,216,217,218,219,220,221,222,223,224,225,226,227,228,229,230,231,232,233,234];

B

Fault diagnosis: Involves analyzing and determining the root cause or sources of faults, malfunctions, or problems within industrial equipment or systems. It often employs investigative methods, data analysis, and expertise to pinpoint specific issues causing failures [2,3,4,6,7,12,13,32,34,38,39,45,47,53,58,61,77,96,103,106,111,115,118,119,122,131,137,149,157,158,162,163,164,165,169,170,171,172,173,174,175,176,177,178,179,183,184,185,186,188,189,190,191,192,193,198,200,201,203,204,206,207,208,210,211,212,215,216,218,219,221,222,224,225,226,228,229,230,231,232,235,236,237,238,239,240,241,242,243,244,245,246,247,248,249,250,251,252,253,254,255,256,257,258,259,260,261,262,263,264,265,266,267,268,269,270,271,272,273,274,275,276,277,278,279,280,281,282,283,284,285,286,287,288,289,290,291,292,293,294,295,296,297,298,299,300,301,302,303,304,305,306,307,308,309,310,311,312,313,314,315,316,317,318,319,320,321,322,323,324,325,326,327,328,329,330,331,332,333,334,335,336,337,338,339,340,341,342,343,344,345,346,347,348,349,350,351,352,353,354,355,356,357,358,359,360,361,362,363,364,365,366,367,368,369,370,371,372,373,374,375,376,377,378,379,380,381,382];

C

Fault identification: The specific act of recognizing, categorizing, or labeling faults or abnormalities within industrial machines or systems. It involves distinguishing among different types of faults or failures to facilitate targeted troubleshooting and corrective actions [7,45,103,192,208,234,236,267,321,353,366,383,384,385,386,387,388,389].

3.2.3. Health Index Forecasting

Health index forecasting involves predicting future condition or health of equipment or machinery based on historical data and various sensor measurements. It helps maintenance teams anticipate when maintenance or repairs will be needed, enabling proactive maintenance and minimizing downtime. A comprehensive definition is presented in [2]. Related terminology and other terms used in research, in alphabetical order, include the following:

A

Early fault detection: The capability to identify and detect potential faults or anomalies in industrial machinery or systems at an early stage. It involves the use of monitoring tools, sensors, and predictive algorithms to catch abnormalities before they escalate into severe failures [7,34,80,188,203,230];

B

Fault prognosis: This focuses on predicting and forecasting potential faults or failures in industrial equipment or systems before they occur. It uses advanced analytics, machine learning, and diagnostic techniques to anticipate and mitigate impending problems [2,7,53,123,152,356];

C

Forecasting: This involves the use of historical data and statistical techniques to predict future trends or conditions related to machinery health, performance, or maintenance needs. It assists in planning maintenance schedules and resource allocation [14,16,30,41,58,67,79,85,92,95,115,137,153,390,391,392,393,394];

D

Regression: A statistical method for analyzing the relationship between variables, often employed in predictive maintenance models. Regression analysis helps in understanding how changes in one variable relate to changes in another one and it can be used to predict future outcomes based on historical data patterns [2,3,7,8,10,12,14,16,27,28,41,49,53,58,59,69,72,73,75,76,78,85,91,101,138,149,158,159,160,161,162,171,182,197,269,271,289,290,293,321,346,353,382,391,395,396,397,398,399,400,401,402,403,404,405,406,407];

E

Remaining life assessment: This involves the evaluation and estimation of the remaining operational lifespan of industrial machinery. It considers various factors such as usage patterns, environmental conditions, and maintenance history to estimate how much lifespan is left in a system [3,38,71,82,91,147,181,392,395,401,408,409,410,411,412];

F

Remaining useful life (RUL): This refers to the estimated time or remaining operational lifespan of a piece of equipment or system before it is expected to fail or become unreliable. Estimating RUL helps in scheduling maintenance activities to maximize asset utilization while minimizing unplanned downtime [2,3,7,9,10,14,16,17,26,43,45,52,53,57,61,62,69,71,74,78,82,88,89,91,93,94,97,120,139,142,144,147,153,156,161,165,166,180,198,392,395,401,402,404,408,410,411,412,413,414,415].

3.2.4. Data Pre-Processing

Data pre-processing refers to the steps taken to clean, transform, and prepare raw sensor and maintenance data for analysis. This includes noise removal, missing value handling, data normalization, and other data cleaning and formatting tasks to ensure the quality and consistency of the data used for the learning model. A brief description is presented in [12]. For more details, refer to [16], which summarizes an extensive use example. Related terminology and other terms used in research, in alphabetical order, include the following:

A

Data fusion: This involves the integration or combination of data from different sources or sensors to provide a more complete and accurate representation of the state or condition of industrial machinery. It aims to improve decision-making by using multiple data streams [3,4,7,10,16];

B

Dimension reduction: Similar to feature reduction, it involves techniques to reduce the number of dimensions or variables in a dataset. It aims to retain essential information while reducing the dimensionality of the dataset, making it more manageable for analysis [191,250,308]. In industrial maintenance, dimensionality reduction is applied for the following activities:

(a)

Dimensionality reduction of sensor data: In predictive and condition-based maintenance, the principal component analysis (PCA) technique, and other variations of it, can be applied to reduce the dimensionality of sensor data while retaining critical information. This helps to improve the modeling for the equipment failure detection and prediction [63,73,75,231,263,343,416,417];

(b)

Feature selection: Feature selection methods, such as filters, wrappers, and embedded methods, are used to choose a subset of the most relevant features to improve the modeling of maintenance prediction and condition-based tasks [12,105,130,333,418,419,420]. Some applied feature selection algorithms are chi-squared, Spearman correlation, mutual information, Fisher score, Pearson correlation, and count-based [395]. Metaheuristic optimization is employed with machine learning algorithms to obtain a new subset of the best features of data, as mentioned in [209,220,252,285,383];

C

Feature engineering: This is the process of using domain knowledge to extract and create meaningful features from raw data that improve the performance of machine learning models. It involves selecting, modifying, and creating new features (attributes or variables) to enhance the predictive power of the algorithms. This step is crucial because the quality and relevance of features can significantly impact the accuracy and efficiency of the model. Feature engineering can include techniques like normalization, encoding categorical variables, handling missing data, creating interaction features, and more [3,9,78,395,421];

D

Feature extraction: This involves the process of selecting or deriving relevant and meaningful attributes or features from raw data collected from industrial machinery or systems. These extracted features help in building predictive models or performing analysis for maintenance-related insights [2,3,7,8,11,12,14,16,33,40,51,61,75,93,131,137,142,157,158,164,170,171,174,184,192,199,200,204,215,216,222,230,231,235,239,244,246,260,261,267,268,270,275,280,284,286,289,290,293,297,303,315,316,317,320,325,328,329,330,332,333,335,336,337,339,343,347,353,354,356,359,361,362,363,364,365,366,369,370,372,373,374,378,381,389,392,416,422,423,424,425,426,427,428,429,430,431,432,433,434,435,436,437,438]. In industrial maintenance, feature extraction applications are for the following tasks:

(a)

Image processing: In visual inspections, image processing techniques such as edge detection or texture analysis can be used to extract relevant features from images of equipment components [254,265,324,424];

(b)

Signal processing: In equipment maintenance, signals from machines (such as vibration, acoustic, electric current and voltage, among others) can be processed using the fast Fourier transform (FFT), general interpolated fast Fourier transform (GIFFT), and Hilbert–Huang transform (HHT) to extract frequency domain features [6,7,14,55,142,274,319,328,439] and the wavelet packet transform (WPT) to extract time-frequency domain features [96,164,171,173,188,235,238,239,291,324,347,370,422,433,440,441] that indicate the condition of the equipment;

(c)

Statistical summarization of data: Sensor data over time can be summarized using statistical features, such as kurtosis, mean, standard deviation, maximum value, minimum value, skewness, variance, root-mean-square, central moment, and others, to provide a concise representation of the data for anomaly identification or predictive models for fault prognosis [103,250,270,418,419,441,442,443,444];

E

Feature fusion: This combines or integrates multiple features or information sources from different sensors or data streams related to industrial machinery or systems. The goal is to create more informative and comprehensive features that can enhance the understanding of equipment health or performance [7,354]. In industrial maintenance, some examples of feature fusion applications include the following:

(a)

Multi-modal data fusion: Industrial maintenance often involves diverse data sources, including not only sensor data and text reports but also images, making this different from previous presented data fusion cases. Feature fusion techniques can combine features extracted from these three different modalities to create a unified representation for more accurate fault detection or equipment health assessment [2,16,362];

(b)

Sensor data fusion: In predictive maintenance, data from multiple sensors on the same equipment can be fused by concatenating or averaging the sensor readings [3,10]. This provides a holistic view of the equipment condition and has the potential to provide better data analysis tools across all domains, including the maintenance quality and the management field [16];

(c)

Temporal and spatial data fusion: In cases where equipment data include both temporal (time series) and spatial information, feature fusion can combine features from these dimensions to model the dynamic behavior and spatial relationships of equipments [137,317,319,359];

(d)

Text and numerical data fusion: Maintenance reports and numerical sensor data can be fused to provide a more comprehensive understanding of maintenance events and their impact on equipment performance [16];

F

Feature learning: This refers to algorithms or techniques that enable systems to automatically discover or learn informative features from raw data without explicit guidance. This process helps to identify patterns or representations that can be useful for predictive maintenance or fault detection [7,9,157,166,258,260,325,351,354,363,365,372,427];

G

Feature reduction: This process consists of reducing the number of features or variables in a dataset while preserving as much relevant information as possible. It helps to simplify models, reduce computational complexity, and improve efficiency in the analysis of industrial maintenance data [2,174,175,191];

H

Pre-processing: This refers to the initial phase of data preparation where raw data collected from sensors or other sources are cleaned, normalized, transformed, and other necessary steps are taken to make them suitable for analysis. It includes handling missing values, removing noise, and standardizing data formats [2,12,14,16,28,47,78,95,142,193,197,199,200,206,343,445,446,447].

3.2.5. Data Augmentation

Data augmentation involves the creation of artificial data that simulate real-world conditions. This can be useful when real data is scarce or sensitive. In maintenance, synthetic data can be used to test and validate predictive models. Related terminology and other terms used in research, in alphabetical order, include the following:

A

Augmentation: Techniques used to expand or enhance existing datasets by introducing variations, modifications, or transformations to the original data. This process helps to increase the diversity and size of datasets for better model training or analysis [2,7,13,14,153,350];

B

Data generation: The process of creating or producing new data points or datasets relevant to industrial machinery, systems, or processes. It involves generating data to supplement existing datasets, improve model performance, or simulate different operational scenarios [2,14,60,89,226,352];

C

Synthetic data: Artificially generated data that imitate real-world scenarios and characteristics of industrial machinery or systems, also known as fake data [12]. They are used when access to real data is limited or restricted to help train models, test algorithms, or perform analyses without using sensitive or proprietary information [59,99,168,197].

In industrial maintenance, synthetic data generation applications examples include the following:

(a)

Imbalanced data augmentation: In predictive maintenance, when the majority of the data represent normal equipment behavior, synthetic data can be generated to create more balanced class distributions, ensuring that models effectively detect rare equipment faults [2,12,13,14,350]. For a production system, failure events are rare due to the unaffordable and severe consequences of running machines under fault conditions and the potentially time-consuming degradation process before a failure happens [7];

(b)

Rare event simulation: Industrial equipment may experience rare events such as extreme operating conditions or rare failures. Synthetic data can simulate these scenarios to train models for robust fault detection [14].

3.2.6. Domain Adaptation

Domain adaptation is a technique where knowledge learned from one domain or dataset is applied to improve performance in another related domain. In maintenance, this can involve using pre-trained ML models or algorithms from one type of equipment to enhance the predictive accuracy for another type, leveraging the shared knowledge. Related terminology and other terms used in research, in alphabetical order, include the following:

A

Domain adaptation: This refers to the process of modifying or adjusting models or algorithms trained on one domain to perform effectively on a different but related domain. It involves minimizing the effects of domain shift by adapting the learned knowledge or features from a source domain to improve performance on a target domain with different characteristics or distributions [7,131,149,170,183,243,399,448,449,450];

B

Domain shift: This occurs when there is a difference or discrepancy between the distribution or characteristics of data from one domain (source domain) to another domain (target domain). In industrial maintenance, this could be observed when the data collected from different machines or systems exhibit variations due to changes in operating conditions, environmental factors, or other variables [351,450];

C

Transfer learning: This refers to a machine learning technique in which knowledge or patterns learned from one domain or task are transferred and applied to another related domain or task. In industrial maintenance, it involves using knowledge or models trained on one set of machinery or data to improve learning or performance on a different but related set of machinery or data, especially when the labeled data in the target domain are limited [2,7,12,13,14,118,122,131,170,183,184,241,242,266,349,397,399,426,451].

It is most valuable when there is limited data in the target domain. In industrial maintenance, examples of transfer learning applications include the following:

(a)

Equipment fault diagnosis: Suppose that the source and target domain data are collected from an identical machine, but with different operation conditions, like different speed and load, or different working environments. In this scenario, transfer learning may be used to adapt fault detection models from a known and labeled source domain. Even when the source and target domain data are collected from different but related machines, like motors and generators, transfer learning may be a feasible technique [12,118,122,131,170,184,241,242,349];

(b)

Image data classification: Transfer learning has been established as an effective technique in computer vision for using rich labeled data in the source domain to build an accurate classifier for the target domain [426];

(c)

Knowledge transfer across sites: In a multi-site manufacturing environment, knowledge gained from maintenance and quality control at one site (e.g.,: a laboratory) can be transferred to improve operations and maintenance at another site (e.g.,: real-world) with similar processes and equipment. Usually, this is known as a digital twin approach, in which situations and conditions are developed in a laboratory (virtual space) that, in the real world (physical space), would not be easily captured due to the production cost, safety, and other factors of the industrial plant [7];

(d)

Time series forecasting: Transfer learning can be used in time series forecasting tasks, where models trained on historical data from one location can be adapted to make forecasts for a different location with similar operating conditions [399].

3.2.7. Objective Optimization

Objective optimization involves finding the best set of parameters or decisions that optimize a specific objective or goal. In the context of industrial maintenance, it refers to finding the optimal maintenance strategy that minimizes costs, downtime, or other relevant objectives while maximizing equipment reliability and performance [7]. It is used to make informed decisions and allocate resources effectively. The main characteristic of objective optimization is that it can improve system performance or decision-making. In industrial maintenance, examples of optimization applications include the following:

A

Energy consumption optimization: Optimization techniques can minimize energy consumption in manufacturing processes by adjusting equipment operating parameters, such as temperature and pressure, to meet production targets while minimizing the overall cost [100]. Another important factor that can be analyzed to optimize energy consumption is the distance between the edge equipment where data are produced and the computation center where they are processed [2];

B

Maintenance scheduling: Optimization can be used to schedule preventive maintenance tasks or inspection intervals for industrial equipment. It considers factors such as equipment criticality, maintenance costs, and operating downtime to find the most cost-effective schedule [3,6,40,76,91];

C

Spare parts inventory management: Optimization models can determine the optimal level of spare parts inventory to minimize costs while ensuring that critical components are readily available for maintenance needs [31,37,69]. This is a good example of the role of data-driven prognostics in maintenance decisions that industries usually face over the life cycle of an asset;

D

Supply chain optimization: Optimization techniques can optimize the supply chain by minimizing lead times, reducing transportation costs, and ensuring timely delivery of critical spare parts. The factories of the future will combine the efficiency of mass production with custom production, and will optimize the supply chain in real-time thanks to high Internet connectivity. These factories will handle fluctuations in demand in a fully automated and fault-tolerant manner [4], meaning this important process should have redundancy.

Figure 2 presents and summarizes the number of research works that mention the application of PdM or CBM considering the previously defined purposes, with fault detection, forecasting, pre-processing, and degradation being the most referenced with 517, 196, 184, 169 mentions respectively. The indicated mentions consider the referenced research studies of reviews, surveys, taxonomies, and specific studies, which is why the sum of them exceeds the quantity of the core repository.

3.3. Data Mining Methodologies and Machine Learning Types Applied to Industrial Maintenance

This subsection provides an overview of the wide range of data mining (DM) methodologies, from traditional statistical methods to advanced machine learning algorithms, and highlights their applications in failure prediction, maintenance scheduling optimization, and data-driven decision-making for efficient industrial operations and maintenance.

3.3.1. Data Mining Methodologies

DM encompasses a wide range of methodologies to make extractions, predictions, or decisions based on data.

Table 4. Summary of data mining methodologies, definitions, algorithm examples, and applications.

Methodology	Definition	Algorithm Example	Application
Probabilistic model	Represents uncertainty by assigning probabilities to different outcomes.	Naive Bayes [85], hidden Markov model (HMM) [45,163,193,233,378], coupled hidden Markov model (CHMM) [340].	Classification, sequential data analysis.
Stochastic model	Includes randomness and probability distributions.	Wiener process (WP) [9], hidden Markov model (HMM) [9], Monte Carlo (MC) simulations [26,31,40,194,197].	Simulations.
Statistical learning framework	Use statistical techniques to model relationships between input features and target variables.	Linear regression (LR) [7,75,138,398], logistic regression (LGR) [7,41,49,72,85,159,160,161,171,269,391,403].	Predicting outcomes based on relationships between variables.
Neural network	Composed of interconnected artificial neural networks.	Artificial neural network (ANN).	Image recognition [2,131,346,424], natural language processing [53,397], reinforcement learning [7,77,117,144].
Hierarchical decision support models	Provide structured decision-making frameworks, often involving hierarchical structures.	Decision tree (DT) [9,73], rule-based systems [12].	Decision-making processes.
Ensemble learning	Combines multiple models to improve predictive accuracy and reduce overfitting.	Random forest (RF) [3,398], gradient boosting (GB) [394,409].	Enhancing model performance, reducing overfitting.
Boosting learning	Focuses on improving the performance of weak learners by combining them sequentially.	Adaptive boosting (AB), gradient boosting (GB) [23,367,451,452].	Increasing the accuracy of weak models.
Density-based model	Identifies clusters based on the density of data points.	Density-based spatial clustering of applications with noise (DBSCAN) [63,453].	Anomaly detection, clustering.
Partitioning and dynamic clustering	Divides data into clusters based on similarity, with dynamic clustering adapting to new data.	K-means (k-M) [32,438,454], dynamic clustering [14].	Clustering similar data points, adapting to new data in streaming scenarios.

summarizes some well-known DM methodologies from the industrial maintenance literature, highlighting their definitions, example algorithms, and applications.

3.3.2. General Types of Machine Learning

The following is a summary of the general types of machine learning (ML) that support the previous purposes applied to CBM or PdM.

A

Classification

Classification in ML is crucial for solving various problems in industrial maintenance. The main feature of classification is to sort data into categories or classes; hence, it is a problem of automatically assigning a label to an unlabeled example [455,456]. The three most common types of classification are described below:

(a)

Binary classification, which can be used to detect a fault (yes/no) based on sensor data and maintenance history, or to predict whether an asset will fail within the next defined period such as 24 h. This helps in proactive maintenance scheduling, preventing costly breakdowns or the damage of dependant systems, as detailed in [14,38,49,83,398,457];

(b)

Multi-class classification, which can be used to categorize different types of equipment faults, such as electric, mechanic, or software-related problems, based on fault data and symptoms, as mentioned in [14,38,244,298,312,322,326]. This can enable maintenance teams to identify the root causes of problems and allocate resources efficiently;

(c)

Multiple classification (multi-label classification), which helps identify multiple defects or maintenance needs in a manufacturing process, such as categorizing products with defects in color, shape, and size. According to [108], this supports quality control and ensures comprehensive maintenance planning.

B

Regression

Regression is an ML technique used to predict the continuous evolution of numerical values based on input features [455,456]. It aims to model the relationship between independent variables (features) and a dependent variable (target) to make predictions. This prediction of signal evolution can serve as input to expert systems or users to trigger alarms related to patterns observed in the future behavior of a system [14].

The output of a regression model is a continuous numerical value, such as temperature, pressure, time to failure, or cost.

The main characteristic of regression is the forecasting of future outcomes based on historical data. In industrial maintenance, regression models are applied to various scenarios to estimate or forecast continuous outcomes. A summary with examples of applications in industrial maintenance is presented next:

(a)

Remaining useful life (RUL): Regression can be used to predict the RUL of industrial equipment based on sensor data. For instance, predicting when a pump or motor will need maintenance to prevent costly breakdowns. Research studies that have estimated the RUL include [2,3,7,9,10,14,16,17,26,43,45,52,53,57,61,62,69,71,74,78,82,88,89,91,93,94,97,120,139,142,144,147,153,156,161,165,166,180,198,392,395,401,402,404,408,410,411,412,413,414,415];

(b)

Quality control: Regression analysis can be applied to predict product quality attributes based on various process parameters; for example, predicting the quality of a product based on manufacturing conditions, as described in [390];

(c)

Cost estimation: Regression can estimate maintenance or repair costs and the system availability or performance based on factors such as equipment age, usage, and historical maintenance costs. This aids in budget planning, reducing unplanned downtime, extending equipment lifetimes, and maintenance optimization, as detailed in [7,16,25,26,28,29,41,43,50,52,56,73,74,93,104,109,140,458];

(d)

Supply chain optimization: Regression can forecast demand for industrial components, helping in supply chain planning and inventory management [4];

(e)

Process enhancement: Regression models can optimize manufacturing processes by identifying the ideal operating conditions that lead to the desired output or goal, such as product yield or efficiency [100];

(f)

Fault prognosis: Regression analysis can be used to detect abnormal patterns in sensor data that indicate equipment faults or deviations from normal operation, as mentioned in [2,7,53,123,152,356].

C

Clustering

Clustering is an ML technique used to group similar data points based on certain similarity measures or patterns; hence, it is a problem of learning to assign a label to examples by leveraging an unlabeled dataset [455,456]. It aims to discover inherent structures or patterns in data by grouping data points into clusters, where points within the same cluster are more similar to each other than to those in other clusters. The output of a clustering algorithm is a set of clusters or groups, and it is typically an unsupervised learning technique because it does not require labeled data [56].

In industrial maintenance, clustering helps identify similarities or patterns among equipments, processes, or maintenance events. It allows the determination of outliers or identifies erroneous behaviors on machinery or systems [56]. This technique is useful when there is no prior knowledge or understanding of the monitored system [16]. A summary of clustering in ML with examples from industrial maintenance is presented next:

(a)

Equipment health monitoring: Clustering can be applied to sensor data from different equipment to group similar machines based on their operating conditions or performance trends. Some examples mentioned in [7] are clustering wind turbine normal states into multiple different clusters for normal behavior identification or the feature fusion of rotary machines for clustering performance improvement;

(b)

Maintenance event categorization: Clustering can categorize maintenance events based on their characteristics, such as failure modes. This helps to understand common maintenance challenges and improve response times [14];

(c)

Quality control: Clustering can group products or components based on quality attributes. For example, clustering similar products with consistent quality or asset performance can help identify production problems or anomalies [30,390,459,460];

(d)

Energy consumption profiling: Clustering can segment energy consumption patterns in a manufacturing facility and identify unusual patterns of energy consumption of different machines. This aids in early fault identification in machines and energy cost savings [24,393,421];

(e)

Anomaly detection: In maintenance, clustering can be used for anomaly detection by identifying data points that do not belong to any cluster, which may indicate unusual equipment or system behavior, or potential faults. This helps prevent damage or failure and reduce the malfunctioning time [54,63,132,133,136,138];

(f)

Process enhancement: Clustering can group similar processes or production lines, allowing for the identification of maintenance best practices that can be applied across similar assets [100].

In specific cases, some previously explained well-known techniques of ML are combined among themselves, or with other algorithms that are used before or after the data modeling, to allow tasks such as enhanced pre-processing, synthetic data generation, transfer learning, or optimization.

4. Data Types Employed in Data Mining, Asset Variety, and Complexity

The use of data mining (DM) in industrial maintenance requires the exploration and analysis of diverse data types sourced from a large number of assets, each with varying degrees of complexity. These data types cover a broad spectrum, ranging from structured datasets comprising numerical or categorical information to unstructured data such as text, images, or sensor readings. Within industrial settings, assets exhibit multifaceted variations, including machinery, equipment, systems and infrastructure, each characterized by unique operational complexities and unique maintenance requirements. Furthermore, the complexity of these assets ranges from simple standalone machines to interconnected systems within complex industrial ecosystems. The combination of these diverse data types and asset complexities represents significant challenges in extracting valuable insights, requiring sophisticated data processing techniques and analytical methodologies designed to compact and extract valuable information from this diversity. A strategic and valuable factor for the application of data mining and machine learning in data-driven industrial maintenance strategies is their ability to effectively understand and manage the diversity of data types, asset variations, and levels of complexity of these systems.

This section provides an in-depth exploration of the data types utilized in DM, highlighting the most common types used in the field of industrial maintenance. A complimentary overview of data acquisition including the domain of analisys and categorization of data by source are given afterward. Finally, it is discussed the spectrum of asset varieties encountered within industrial environments, and the associated complexity inherent in the operational processes under consideration.

4.1. Data Used for Data Mining

In data mining (DM), various data types are used to represent and analyze information from datasets. In the first segment, an explanation of general data types and their main characteristics is presented. Afterward, the explanation focuses on the most common data types sourced from industrial assets.

4.1.1. General Data Types and Characteristics

This subsubsection details the general data types and their main characteristics.

Table 5. Summary of data types, definitions, examples, and applications.

Data Type	Definition	Example	Application
Numerical data	They are among the most prevalent in data mining. They include the following: - Continuous numerical data: real numbers that represent measurements; - Discrete numerical data: integer numbers representing counts.	- Continuous numerical data: temperature [2,23,56,57,58], vibration [2,51,57,141,142,163,192,373,414,461,462], humidity [84], or pressure [28,54,56,83,216,290]; - Discrete numerical data: number of events [16], number of well-produced products [5,42] or product identification [391].	- Continuous numerical data: Used in various scientific and engineering measurements; - Discrete numerical data: Used to count and categorize discrete events or items.
Categorical data	Represent discrete categories or labels. Common types include the following: - Nominal data: categories with no intrinsic order; - Ordinal data: categories with an inherent order or ranking.	- Nominal data: status [23,81], fault types [240,254,256,260,262,293,295,296,303,316,320,349,373,380,463,464,465] or product names [391]; - Ordinal data: degradation levels [29,39], (e.g.,: low, high, critical), maintenance policy levels of dependency [43] (from none to maximum dependency), imperfect maintenance with different levels [28], faults covering different levels of severity [219].	- Nominal data: used to classify and label of categorical data; - Ordinal data: used for ranking and ordering data based on levels or severity.
Text data	Unstructured human-generated text documents.	Information about replaced components and repair activities [16].	Used for text mining (TM) and natural language processing (NLP) to analyze and extract insights.
Time series data	Observations collected or recorded over time intervals.	Sensor readings such as acoustic data [2,3,6,7,23,93,112,161,197,244,273,274,312,317,319,338,341,438,466,467,468], current signal [2,179,221,224,322,354,380,384,387,429,434,469], electrical power [107,197,470,471].	Forecasting, trend analysis, anomaly detection.
Image data	Visual information in the form of pixel values.	Object detection, fault region extraction, fault diagnosis [265], and image classification [399,424,426,449].	Computer vision tasks.
Spatial data	Geographical or location-based information, often represented as latitude and longitude coordinates. They are commonly used in geographic information systems (GIS).	Latitude and longitude coordinates [16].	Correlation of data and equipment performance considering weather conditions (temperature, humidity, irradiation) regarding geographic information.
Graph data	Represent relationships and connections between entities.	Vertex and edges of a relationship representation.	Graph mining and network analysis for fault classification [369,371,372].
Temporal data	Include timestamps and time-related information.	Event logs [98,100,130], time-stamped or time-based data from preventive maintenance [100].	Process mining and sequence analysis.
Multi-modal data	Combine data from multiple sources or different domains [2,3,51].	Merging data of multiple types and sources, such as combining text (e.g.,: details of work orders created/completed, work order types), images, and numerical data. Studies that have focused on these topics are mentioned in [2,37,186,362].	Data analysis, recommendation systems, learning model improvement.

outlines, with the corresponding references, the various data types used in industrial maintenance with their definitions, examples, and applications.

4.1.2. Common Data Types for Data Mining

In the research studies considered for this work, the most frequently mentioned data types can be summarized as follows:

A

Time series data: In the industrial field, commonly associated with condition monitoring, time series data contains knowledge in the form of degradation patterns and other types of anomalies in the data that lead to asset degradation. Time-varying features are expected to capture these abnormal patterns, and the models fed with these features are expected to learn to distinguish between normal and abnormal pattern behavior of items and also to forecast the RUL for the monitored items [16]. Time series data have several characteristics that distinguish them from other types of data:

(a)

Temporal order: Time series data are collected and recorded in chronological order, where each data point is associated with a specific time or time interval. They are widely used in asset monitoring with periodic inspections [7,20,40,59,81]. The time intervals can be equally spaced, e.g.,: weekly, daily, hourly, minutely, secondly, etc.;

(b)

Dependence on time: Time series data exhibit a temporal dependency, meaning that the value of a data point at a given time is often related to the values of previous and/or future data points [359];

(c)

Seasonality: Many time series datasets exhibit seasonal patterns, where certain patterns or behaviors repeat at regular intervals, such as daily, weekly, monthly, or yearly cycles. Seasonality can impact data analysis and forecasting [33];

(d)

Stationarity: This refers to the characteristic where statistical properties of a time series, such as mean and variance, remain constant over time. Stationary data simplify analysis and modeling. Most signals used for fault pattern recognition are non-stationary, which adds a high complexity for the modeling [45,472];

(e)

Trends: Time series data can have underlying trends, which represent long-term changes or movements in the data over time. Trends can be upward (increasing), downward (decreasing), or flat. A very important characteristic in maintenance is the observation of degradation over time to apply early failure detection [20,61] and to evaluate forecasted values [58];

(f)

Noise: Time series data often contain noise or random variability that make it challenging to discern underlying patterns. Noise can result from measurement errors, external factors, or other sources of randomness [2,55,57,61];

(g)

Irregularities: Time series data may contain irregular or anomalous events, such as peaks [12,308,441], outliers [56,79,84,309] or sudden changes [446], which can disrupt regular patterns and need to be accounted for in analysis.

B

Event timestamp data: Often referred to as log data or event data, these data have unique characteristics that make them distinct from other types of data. These characteristics are important when working with event data for various purposes, including monitoring, analysis, troubleshooting and security.

Timestamped events include human-generated data that represent information about replaced components and repair activities performed on a certain item or parts of it. Moreover, they also consist of software-generated information, e.g.,: event data information such as alarms and faults messages, which are described in natural language. In addition, they also comprise technical metadata for devices and processes. Event data collection implies a manual or an automatic process and includes qualitative information about the monitored item, such as the description of the installation, breakdown, inspection, repair, overhaul, failure causes, etc., the severity of the failure, and the description of what was done to fix the failure [16]. The main characteristics of timestamped events are as follows:

(a)

Timestamps: Event data are associated with specific timestamps, which indicate when each event occurred. These timestamps provide a chronological order, allowing for time-based analysis and tracking of events over time [100];

(b)

Event types: Event data typically include different types or categories of events. Each event type represents a specific action, incident, or state change. Examples of event types include error events, system messages, and logs [130];

(c)

Granularity: Event data can have different levels of detail, depending on the application. Timestamps can be precise, with high sampling rates (e.g.,: milliseconds, seconds, minutes) [268] or other time units, allowing for fine-grained or coarse-grained analysis;

(d)

Volume: Event data often come in large volumes, especially in systems with high levels of activity that generate large datasets [63]. Analyzing and managing large event datasets require scalable and efficient data processing techniques [85,402];

(e)

Log levels: Many event logs include log levels that indicate the severity or importance of each event, such as different degrees of fault or system health [7,14,16,96,203,233,280]. Common log levels include INFO, WARNING, ERROR, and CRITICAL. These levels help prioritize events for action or further investigation;

(f)

Contextual information: Events often include additional contextual information or metadata, as this information is critical for determining an assertive predictive model, such as source identifiers, event descriptions, user identification, IP addresses, port, repair activities performed, the components replaced, and more [16,67];

(g)

Event correlation: Event data may include fields or attributes that allow events to be correlated or grouped, e.g. a correlation identificator can link related events, making it easier to track sequences of events to a particular process instance [100];

(h)

Irregularities and anomalies: Event logs can help identify irregular or anomalous patterns and behaviors, such as security breaches [4], forecast system failures based on historical data [92,144,407], or deviations from expected norms [391];

(i)

Aggregation and summarization: Event data can be aggregated and summarized over time to extract trends, patterns, and statistics [4,14,72]. This is useful for generating reports and identifying long-term changes;

(j)

Real-time monitoring: Event logs are frequently used for real-time monitoring and alerting. Systems can be configured to trigger alerts or notifications based on specific event conditions [30,81].

C

Image data: Image data consist of visual information in the form of images or pictures, possessing unique characteristics that distinguish them from other types of data. These characteristics are essential to consider when working with image data in various applications, including computer vision, image processing, machine learning, and multimedia analysis. Some of the challenging characteristics include image quality [30], resolution, and dimension. The basic characteristics of image data are as follows:

(a)

Visual representation: Image data are primarily represented visually as two-dimensional arrays of pixels. Each pixel contains color information and contributes to the overall appearance of the image;

(b)

Pixel depth: This is often referred to as bit depth or color depth, i.e., the number of bits used to represent each color of pixel information. Common depths include 8-bit (256 colors), 24-bit (16.7 million colors), and 32-bit (including transparency);

(c)

Image dimensions: Image data are characterized by their width and height, which define the dimensions of the image in pixels. The aspect ratio, or the ratio of width to height, affects the appearance of an image;

(d)

Image formats: Images are stored in specific file formats, such as JPEG (joint photographic experts group), PNG (portable network graphics), BMP (bitmap image file), TIFF (tag image file format), and others. Each format has its own compression, quality, and compatibility characteristics;

(e)

Metadata: Images can contain metadata, including information about the image, such as creation date, author, camera settings, and geolocation data;

(f)

Color histograms: Image data can be represented by color histograms, which summarize the distribution of colors within an image. Histograms are useful for image analysis and retrieval;

(g)

Data size: High-resolution images and image datasets can be large in terms of file size and memory requirements, requiring efficient storage and processing solutions.

Image data can be processed through a wide range of techniques, including filtering [14], resizing, rotation, cropping, enhancement, flattening [424], noise separation [427], and transformation, to extract information or improve image quality.

Regarding image features, image data often involve extracting features or patterns, such as edges, textures, shapes, objects, for tasks such as object detection, recognition, and classification [131,204,265,360,424,426], e.g., detecting typical defects such as cracks, stains, or scratches in mechanic assets [459]. Some types of commonly used image features include moment invariants, gray statistical characteristics, texture features, and the differential box-counting fractal dimension [254]. Hence, the key features for processing image data are as follows:

(a)

Resolution: Images can have different resolutions, which determine the level of detail in the image. High-resolution images contain more pixels and finer details, while low-resolution images have fewer pixels and coarser details [459]. In some research studies, images containing a large number of pixels are known as superpixel images [427];

(b)

Color spaces: Images can be represented in various color spaces, such as RGB (red, green, blue) [186,362], grayscale (single-channel) [459], CMYK (cyan, magenta, yellow, key), and others;

(c)

Image sensors: Image data originated from image sensors, such as digital cameras, scanners, thermal cameras [5,7,30,204,265,360,473], and satellite imaging devices [453].

The number of mentions for each data type in the research studies is as follows: time series data appear 572 times, event data with timestamps are mentioned 107 times, and image data are referenced 66 times. The indicated number of mentions includes the referenced research studies of reviews, surveys, taxonomies, and specific studies, which is why it exceeds 534—the number of research works of the core repository described in Section 2.

4.1.3. Data Acquisition, Analysis Domain, and Source Categorization

At the core of any analytical effort is the process of data acquisition, which provides the basis for in-depth analysis and decision support. Understanding the intricacies and methodologies surrounding data acquisition is fundamental because it sets the stage for comprehensive and meaningful analyses across diverse domains.

A

Data acquisition: There are two main methods of acquisition: embedded or connected sensor, and external sensor. The former refers to a sensor installed inside an asset, i.e., directly connected to the asset, normally included by default for the manufacturer or by request of a customer, measuring voltage, current, vibration, and speed, among other parameters. On the other hand, the latter refers to a sensor installed outside an asset, which is usually performed in the plant to meet engineering or regulatory requirements, including sensors to collect the environment temperature and humidity of a room, the acoustic signal of machinery, or a thermographic camera to verify the working temperature of assets.

In the repository under consideration, embedded or connected sensors were mentioned 580 times, making them the most commonly employed sensor type. This widespread use of embedded or connected sensors highlights their importance in monitoring and data collection, particularly in integrated systems where sensors are built into or connected to the equipment. On the other hand, external sensors were mentioned 165 times. Although less prevalent, external sensors still play a crucial role in scenarios where flexibility, retrofitting, or non-intrusive monitoring is required. The significant disparity in usage between embedded or connected and external sensors underscores the preference for integrated solutions, while also acknowledging the critical role of external sensors in specific applications. The reported number of mentions includes the sum of different sensors referenced in research studies of reviews, surveys, taxonomies, and specific studies; for that reason, it excess the number of research works of the core repository.

In the analyzed repository, individual sensors were used in 477 research studies, making them the predominant choice for sensor-based data collection. This preference for individual sensors suggests a focus on simpler, more targeted monitoring setups, where a single data stream is sufficient for analysis. However, multiple sensors were utilized in 268 research studies, indicating a significant number of cases where more complex, multi-dimensional data collection was necessary. The use of multiple sensors reflects the growing need for comprehensive monitoring in systems where interactions between different parameters are critical. The difference in the number of studies highlights the balance between simplicity and complexity in sensor deployment, with researchers choosing one or the other depending on the specific requirements of their studies.

B

Domain of analysis: For time series data, analyses can be conducted in the time, frequency, and time-frequency domains. The analysis of the repository reveals that time-frequency domain techniques were the most frequently cited, with 114 studies specifically mentioning methods such as wavelet transform and statistical feature extraction. This prominence reflects the importance of capturing both time and frequency information simultaneously, which is crucial for analyzing complex, non-stationary signals. In comparison, 41 studies focused on the frequency domain, often using techniques such as fast Fourier transform (FFT) to analyze the spectral characteristics of the data. These methods are particularly valuable for identifying periodicities and other frequency-based patterns. Finally, 24 studies were conducted in the time domain, where the focus is on analyzing data as they vary over time, without transforming it into other domains. The distribution of these references underscores the variety of analytical approaches employed in research, with a clear emphasis on techniques that can handle more complex, multi-dimensional data.

C

Categorization of data by source: The interest in categorizing the data source is to determine whether the dataset is from real-world applications or simulations. A substantial portion of the research, 354 studies, is based on simulated data, encompassing laboratory experiments, simulated public databases, and software-based simulators. This reliance on synthetic data highlights the controlled environments researchers often use to test and validate their models. However, there is a notable gap in the use of real-world data, with only 95 studies utilizing such data, indicating a potential disconnection between theoretical research and practical, real-world applications. The remaining studies in the repository do not explicitly specify their data sources, further emphasizing the need for more transparency and more works on real-world data to ensure the applicability and robustness of the findings.

D

Working parameters of assets: Regarding the working parameters data of assets included in the considered research studies for data mining (DM) applications, the most frequently mentioned are as follows: vibration (188 references), speed (65 references), temperature (62 references), mechanical power (53 references), and text data (27 references). Other working parameters with fewer mentions include: acoustic, voltage, current, pressure, sound, acceleration, torque, magnetic field, electrical power, irradiation, displacement, gas concentration, and humidity.

4.2. Types of Assets Considered for Data Mining

The revision of research studies with the application of PdM or CBM models allowed the identification of five types of assets: mechanic, electromechanic, electric, electronic, and computing.

The total number of unique components or systems mentioned in research studies categorized by asset type is as follows: mechanic (21 different assets), electronic (19 different assets), electromechanic (18 different assets), electric (13 different assets), and computing (7 different assets).

The mechanic category exhibits the highest diversity of assets, while the computing category displays the lowest diversity. Across the mechanic, electronic, and electromechanic categories, the diversity of assets appears to be relatively similar.

Regarding the cumulative mentions of components or systems within defined asset categories, the overall counts are: mechanic (1014), electromechanic (441), electronic (62), electric (42), and computing (25). The indicated cumulative mentions consider the sum of assets of the same category and they could share the same reference, i.e., for mechanic assets of the next subsubsection, the detailed assets (machine, aeroengine, bearing, rotating machinery, gears) total 835 mentions. All these assets are also mentioned in the paper [57], showing that several assets are mentioned in the same reference, which explains why the number of citations exceeds the total number of papers.

4.2.1. Mechanic Assets

Mechanic assets consist of physical components, such as gears, motors, engines, and moving parts. These components are designed to perform specific functions within a mechanical system. These kinds of assets often require regular maintenance to ensure they operate efficiently and safely. This maintenance can include tasks such as lubrication, inspection, repair, and replacement of worn or damaged parts. They are designed to perform mechanical functions, such as generating mechanical power, transmitting motion, or performing a specific task. They are commonly used in various industries, including manufacturing and transportation, to carry out tasks that involve the conversion of mechanical energy [97,360,461,468,474,475].

According to the number of mentions related to mechanic assets, the top five terms with many studies related to industrial maintenance are listed below:

A

Machine: This includes various equipment used for manufacturing, processing, or performing specific tasks. There are 328 mentions of this term in the research studies [1,2,3,4,5,6,7,8,9,10,11,12,14,16,17,21,23,24,27,31,33,34,35,36,37,38,41,42,44,47,51,53,54,55,56,57,58,59,60,61,62,66,68,71,72,73,75,78,79,80,81,82,83,84,85,86,87,88,89,90,91,93,95,96,98,99,100,102,103,104,105,106,107,108,109,112,113,114,115,116,119,120,121,123,125,130,131,137,138,148,149,151,158,159,160,162,163,165,166,167,168,170,171,175,176,177,179,181,183,184,185,186,188,189,190,191,192,193,195,197,199,200,201,202,203,205,206,207,208,210,211,217,218,219,220,221,222,223,224,225,228,232,233,235,237,239,240,241,243,244,245,246,247,248,249,250,251,252,253,254,255,257,258,259,260,262,263,266,267,270,272,273,275,276,278,281,285,286,294,296,297,298,299,300,301,302,303,304,305,306,307,308,309,310,311,312,313,314,315,316,317,318,319,320,321,322,323,324,325,326,327,328,329,330,337,338,341,344,345,347,348,349,350,357,360,361,363,365,366,367,369,371,372,373,378,379,380,381,382,383,384,385,386,389,390,391,392,394,395,398,399,400,401,402,403,405,406,407,408,411,412,414,417,418,420,421,422,423,428,429,430,431,432,436,437,438,439,442,443,445,446,447,451,452,454,457,461,463,465,466,469,472,473,474,475,476,477,478,479,480,481,482,483,484,485,486,487,488,489,490,491,492,493,494,495,496,497,498,499,500,501,502,503,504,505];

B

Aeroengine: Also known as an aircraft engine, this component generates thrust to propel an aircraft forward. There are 155 mentions of this term in the research studies [2,3,5,7,9,12,16,22,24,25,26,28,29,31,32,34,38,42,45,48,53,57,67,68,71,76,78,82,83,85,91,95,99,106,108,109,110,114,115,116,122,131,137,144,147,152,154,156,162,164,165,166,170,171,174,178,180,181,184,185,197,204,213,214,215,216,221,225,226,230,232,233,235,237,238,239,241,243,245,247,252,254,255,260,262,266,275,286,287,288,289,290,291,292,293,294,297,306,314,325,327,328,336,342,347,348,355,356,357,362,366,367,368,369,370,371,378,382,390,391,395,400,403,405,408,409,410,412,413,419,421,425,433,440,444,454,459,460,471,474,475,477,479,482,498,501,503,506,507,508,509,510,511,512];

C

Bearing: These are crucial components found in various machinery and equipment. They facilitate rotational or linear movement by reducing friction between moving parts. There are 151 mentions of this term in the research studies [2,3,6,7,12,44,45,51,53,55,57,58,75,89,92,93,99,106,119,120,141,142,151,153,155,157,160,162,163,171,172,173,174,175,177,183,184,186,188,191,199,200,202,205,207,208,218,219,220,232,233,235,237,240,241,243,244,246,249,250,252,256,257,258,259,260,261,262,263,264,266,269,274,279,280,281,282,283,284,302,306,307,311,312,313,315,316,317,318,321,326,329,330,331,332,333,334,335,336,339,340,341,342,346,348,349,351,352,357,360,363,365,367,369,373,374,378,383,384,386,388,392,393,400,404,410,418,420,430,437,439,441,443,445,446,461,462,463,464,473,481,484,485,493,508,510,512,513,514,515];

D

Rotating machinery: This term encompasses various mechanical equipment such as pumps, turbines, compressors, and motors that involve rotating components. There are 103 mentions of this term in the research studies [2,3,6,7,37,51,53,55,57,58,63,86,87,93,119,120,163,165,168,171,175,177,181,185,186,188,189,190,191,192,199,200,206,208,213,219,220,227,233,235,243,246,247,248,250,251,252,257,258,259,260,262,263,266,267,270,272,276,278,281,294,302,306,310,314,320,321,328,341,348,355,359,360,363,365,366,367,372,379,386,389,392,418,420,422,438,441,443,445,446,461,465,473,480,483,486,488,489,493,495,499,516];

E

Gears: These are mechanical components that transmit torque and motion between shafts in machinery. There are 84 mentions of this term in the research studies [2,3,6,7,12,35,41,44,45,57,61,89,90,103,106,137,141,142,150,158,174,176,177,183,186,192,197,198,200,209,210,220,221,238,239,243,245,253,258,259,261,262,263,266,268,270,272,273,280,285,301,303,304,319,320,321,337,338,342,351,354,359,361,370,375,376,377,385,416,432,441,442,467,472,492,494,496,497,499,517,518,519,520].

The following mechanic assets are also mentioned in the repository: gearbox, transmission line, gas circulation unit/system, linear motion guides, pump, planetary gearbox, valve, turbine blade, air compressor, railroad track, hydraulic brake, ball screw, exhaust fan, transmission boxes, cooling radiator, and crankshafts.

4.2.2. Electromechanic Assets

Electromechanic assets incorporate both electric and mechanic elements in their design. These assets often include motors, sensors, switches, relays, and other electrical components alongside mechanical parts such as gears, levers, or actuators [192]. This combination allows them to convert electrical energy into mechanical motion or vice versa. Electromechanic assets are typically controlled and operated by electrical systems. They can be automated, monitored, and controlled using electrical signals, often through a programmable logic controller (PLC) or a microcontroller.

According to the number of mentions related to electromechanic assets, the top five terms with many studies related to industrial maintenance are listed below:

A

Industrial machinery: This term covers a wide range of machinery, including mechanical and electrical components used in industrial processes, manufacturing, and production. There are 92 mentions of this term in the research studies [2,3,4,5,7,10,13,14,15,16,17,18,23,31,36,41,42,45,47,51,52,53,54,55,56,57,58,60,61,62,63,66,67,71,72,75,76,78,86,87,89,96,99,103,109,115,122,132,133,137,141,159,162,168,176,180,182,186,188,193,206,218,224,227,228,232,233,234,236,257,263,264,272,319,323,352,360,361,367,388,390,396,397,401,408,414,446,459,460,471,479,485];

B

Electromechanical equipment: These are devices or systems that have both electrical and mechanical components, such as motors, actuators, and relays. There are 80 mentions of this term in the research studies [1,3,4,6,7,10,14,16,23,27,35,36,37,38,41,44,50,51,53,55,56,59,61,67,70,72,73,74,75,77,79,85,87,89,90,93,97,98,100,104,107,109,114,115,116,124,130,131,132,143,147,152,163,166,181,193,197,204,229,231,236,259,260,263,341,342,349,373,387,393,394,402,414,425,452,473,499,502,505];

C

Industrial tool: These include various equipment and devices, with both electrical and mechanical components, used for manufacturing, construction, or repair purposes in industrial settings. There are 76 mentions of this term in the research studies [3,4,6,7,10,16,17,18,28,30,34,41,42,45,47,56,57,59,61,63,66,70,71,72,73,75,78,80,81,87,92,93,94,96,100,113,116,122,124,131,143,144,148,158,161,170,172,182,187,197,233,239,249,253,273,274,286,298,322,341,350,351,352,381,391,397,398,423,425,459,468,475,483,509,517];

D

Induction motor: These are commonly used in industrial applications to drive machinery and equipment, having both electrical and mechanical parts. There are 37 mentions of this term in the research studies [3,6,55,60,96,111,179,192,207,211,217,224,229,286,300,322,323,324,325,326,345,350,368,384,386,387,428,429,433,434,438,439,469,486,487,491,514];

E

Wind turbine: These are structures equipped with large blades that convert wind energy into electrical power. They have both electrical and mechanical components. There are 27 mentions of this term in the research studies [1,3,6,7,12,53,56,63,67,81,106,110,169,197,221,226,237,272,301,309,334,354,357,358,370,509,521].

The following electromechanic assets are also mentioned in the repository: rotor, hydraulic generator unit, power system, refrigerant flow system, HVAC (heating, ventilation and air conditioning), nuclear power machinery, scooter, power substation, elevator door, high-voltage circuit breaker, washing machine, and circuit breaker.

4.2.3. Electric Assets

Electric assets are equipment, devices, or systems that primarily involve the use or generation of electrical energy. These assets are designed to perform functions that rely on the manipulation or control of electrical energy. They are used for purposes such as power generation, distribution, and control [60].

Based on the number of mentions, the top five electric assets most frequently referenced in the repository are as follows:

A

Power transformer: These are crucial components in electrical systems used for voltage regulation and power distribution. There are six mentions of this term in the research studies [2,46,55,67,265,507];

B

Photovoltaic panel: These are devices that convert sunlight into electricity. There are seven mentions of this term in the research studies [7,30,53,67,197,203,471];

C

Power switch: These are electrical devices used to control the flow of electricity in circuits. There are five mentions of this term in the research studies [31,55,73,100,230];

D

Battery: These are used for energy storage in various applications. There are four mentions of this term in the research studies [4,7,53,147];

E

Power grid: These are extensive networks that distribute electricity from generation sources to consumers. There are four mentions of this term in the research studies [2,7,59,187].

The following electric assets are also mentioned in the repository: electrical power system, partial discharges system, reactor, ion mill etching, automotive generator, bioreactor, ion implanter tool, and electrical distribution network.

4.2.4. Electronic Assets

Electronic assets encompass a wide range of devices and systems that rely on the control and manipulation of electrical signals, typically using semiconductors [36,502]. Electronic assets are built around semiconductors, such as transistors and diodes, which allow for the precise control of electrical signals. These semiconductors are often integrated into electronic components and devices using integrated circuits (IC) or microchips.

These assets are designed to process and control electrical signals, making them highly versatile. They can perform functions like amplification, filtering, modulation, and digital signal processing. Electronic assets can operate in both digital and analog domains. Digital electronic assets use discrete values (i.e., 0 s and 1 s) for processing information, making them ideal for tasks such as computing and data storage. Analog electronic assets, on the other hand, which work with continuous electrical signals, are used for tasks such as amplification of audio signals, analog sensors, and power regulation.

Based on the number of mentions in papers on industrial maintenance, the top five electronic assets most frequently cited are listed below:

A

Control system of a vehicle: This includes electronic components such as engine control modules, sensors, actuators, and onboard computers. There are 12 mentions of this term in the research studies [1,30,45,97,105,121,267,327,342,355,467,482];

B

Cyber-physical system: These are integrated systems comprising computational elements and electronic physical components. There are 10 mentions of this term in the research studies [4,10,36,41,45,52,65,70,95,522];

C

Hard disk drive (HDD): These are electronic data storage devices used in computers and datacenters. There are six mentions of this term in the research studies [73,125,197,412,452,504];

D

Datacenter: These are composed of electronic equipment for storing, processing, and managing digital information. There are six mentions of this term in the research studies [73,125,412,452,523,524];

E

Medical equipment: This includes a wide range of electronic devices used for healthcare. There are two mentions of this term in the research studies [98,505].

The following electronic assets are also mentioned in the repository: power electronic converters, telecomunication network, analog integrated circuits, network-connected equipment, data storage, regulation system, ethernet switch, industrial personal computer, human machine interface, and bay control unit.

4.2.5. Computing Assets

Computing assets include hardware and software components used in computing systems. They are characterized by their processing power, compatibility with software, and the ability to be scaled or upgraded to meet changing computing needs. These assets are fundamental to the operation of modern computers and computing systems [407]. Computing assets rely on software components, including operating systems, applications and utilities, to perform specific functions.

Based on the number of mentions in papers on industrial maintenance, the top five computing assets most frequently cited are as follows:

A

Maintenance systems: These are software or platforms utilized to manage and organize maintenance-related tasks. There are 11 mentions of this term in the research studies [4,37,56,68,70,72,73,79,81,95,142];

B

Supervisory control and data acquisition (SCADA) systems: These are used for supervising and controlling industrial processes, such as manufacturing, power generation, or infrastructure. They collect real-time data from sensors and devices, display information on operator screens, and enable operators to control processes remotely. There are seven mentions of this term in the research studies [4,42,54,56,75,169,357];

C

Industrial software: This refers to various applications specifically designed for industrial purposes, including manufacturing, automation, design, or analysis. There are two mentions of this term in the research studies [30,501];

D

Database Management System (DBMS): This is a software used to store, organize, and manage data in databases. There is one mention of this term in the research studies [447];

E

Server environment: These, within a virtualized datacenter, comprise virtual machines (VM) running on physical servers. These VMs operate as independent instances, hosting applications, services, or operating systems. There is 1one mention of this term in the research studies [500].

Other computing assets are mentioned in the repository: virtual server and computing system.

4.3. Overview of Assets, Working Parameters, and Categorization

In the field of industrial maintenance, a comprehensive understanding of assets, working parameters, categorization, and their applications is essential to ensure efficient management. Figure 3 offers a comprehensive guide that categorizes assets according to the working parameters essential for applying data mining (DM) techniques. This structured approach helps to understand how different types of assets generate data for analysis. Computing assets typically produce text-based event logs with timestamps, providing a detailed chronological record of operations. Electronic assets, on the other hand, commonly present working parameters related to system conditions, including memory and CPU utilization, as well as available hard disk space, which are critical for monitoring and optimizing performance. Electric assets are primarily analyzed through power, current, voltage, and position parameters, which are critical for diagnosing and predicting failures. Meanwhile, mechanical assets focus on key working parameters such as vibration, temperature, speed, position, torque, and power, which are essential to assess their operational health and detect potential issues.

4.4. Asset Complexity

Various researchers have analyzed fault diagnosis or prognosis of single components considered as the most crucial due to their impact on the system or its dependence on other components. Other research studies have focused on the degradation of the system performance or product quality due to single or multiple failures on components.

Considering the contribution of research in both, component and system levels of complexity, this study proposes a classification of industrial assets into four categories:

A

Single component: An asset studied as a single component is normally conformed by one component (1C), and for this study’s approach, it is not divided into sub-components [1,29,53,147,231,245,287,328]. Some examples are bearings, gears, linear motion guides, blades, batteries, rectifiers, and photovoltaic panels.

B

Multiple components: An asset studied as multiple components is normally composed of two or more (multiple) components (1C+), and for this study’s approach, it may not be convenient to divide it into single components [1,26,29,48,197,204,293,457]. Some examples are gearboxes, combustion and induction motors, compressors, automobile hydraulic brakes, industrial tools, hard disk drives, medical equipments, and power switches.

C

Single system: An asset studied as a single system is normally composed of two or more assets with multiple components (1S), and for the purposes of this study, it may not be convenient to divide it into all its multiple components [1,26,66,347,447]. Some examples are heat recyclers, turbofan engines, power transformers, refrigerant flow systems, metal lathes, high-voltage circuit breakers, cyber-physical systems, industrial software, maintenance systems, and virtual servers.

D

Multiple systems: An asset studied as multiple systems is normally composed of two or more (individual) systems (1S+), and for this study’s approach, it may not be convenient to divide the asset into its individual systems [12,67,167,407,500]. Some examples are bioreactors, smart power grids, energy production systems, power substations, vehicle control system, server environments.

Figure 4 illustrates the number of unique assets categorized by complexity and our study reveals the following distribution: single component with 12 different assets, multiple components with 16 different assets, single system with 35 different assets, and multiple systems with 14 different assets. It is noteworthy that assets classified as a single system (1S) include all types: mechanic, electromechanic, electric, electronic, and computing. In contrast, assets classified as a single component (1C) are limited to mechanic and electric types. For assets considered as multiple systems (1S+), all types are present except mechanic, while assets considered as multiple components (1C+) include all types except computing. This distribution highlights the diverse representation of unique asset types across varying complexities in the research studies.

Regarding the cumulative mentions of assets within defined complexities, the overall counts are as follows: single component (1006), multiple components (347), single system (621), and multiple systems (58). The indicated cumulative mentions consider the sum of assets within the same category and they could share the same reference.

5. General Techniques, Learning Categories, and Algorithms Used in Industrial Maintenance

This section presents a comprehensive overview of various techniques, learning categories, and algorithms that are paramount to optimizing maintenance practices on industrial assets. It thoroughly examines learning categories, shedding light on their specific applications and implications within industrial maintenance scenarios.

This study considers two broad techniques: data mining-based techniques and machine learning techniques. Within each of these general categories, more specialized techniques emerge, tailored to address specific types of data or problems regarding industrial assets. From these specific techniques, a variety of methodologies arise. From these methodologies, algorithms are derived. An algorithm is essentially a defined set of finite instructions designed to accomplish a particular task or solve a given problem. These algorithms are flexible and can often have multiple variations or adaptations to suit different datasets, objectives, or constraints.

5.1. Data Mining Techniques Used in Industrial Maintenance

This subsection presents a comprehensive overview of both the simpler rule-based and the more elaborate model-based techniques.

5.1.1. Rule-Based Technique

In the industrial maintenance environment, the rule-based (RB) technique is related to a structured approach that relies on the extraction and utilization of rules derived from patterns and relationships within datasets. This technique is built around the construction of explicit rules or conditions inferred from the available data, allowing for precise decision-making and inference processes on industrial assets. These rules, often in the form of “if-then” statements, are generated by analyzing patterns, correlations, and associations within the data. They encapsulate actionable insights or recommendations based on predefined conditions or thresholds, enabling straightforward interpretation and implementation in maintenance procedures.

RB applications in preventive and condition-based maintenance can be described as follows:

A

Preventive maintenance (PM): RB systems use predefined rules and conditions to schedule preventive maintenance on industrial assets. For example, maintenance may be triggered after a certain number of operational hours [73].

B

Condition-based maintenance (CBM): RB systems can set thresholds for sensor readings [22] of industrial assets. When readings exceed or fall below these thresholds, predefined actions are triggered [57]. For instance, disk monitoring is currently performed mainly through threshold analysis, which is predefined by disk manufacturers. In this approach, the device firmware compares the thresholds with the measured parameters; if an attribute drops below its threshold, it indicates a potential problem with the drive [73].

5.1.2. Model-Based Technique

In the industrial maintenance environment, the model-based (MB) technique involves the creation of explicit models of the underlying industrial asset or process, which are then applied for analysis, prediction, or decision-making purposes. These techniques encapsulate the structural and operational characteristics of the systems under analysis, allowing the representation of their behavior and dynamics. Unlike some other methodologies, model-based approaches often rely on domain knowledge, physical principles, or mathematical formulations to construct these explicit models.

MB applications in predictive and condition-based maintenance can be described as follows:

A

Predictive maintenance (PdM): MB approaches involve creating mathematical models of industrial assets’ behavior. These models can simulate equipment performance and predict when maintenance is needed based on the simulation results. Finite element analysis (FEA) [308], failure mode and effect analysis (FMEA) [30], and physics-based models [89] fall into this category. The FMEA approach has been successfully used in the aerospace, nuclear, automotive, and semiconductor industries to support design fault, safety, logistic support, testability, and other related functions [30].

B

Condition-based maintenance (CBM): MB techniques can validate the condition of industrial asset by comparing real-time sensor data to expected model outcomes. Significant deviations of the actual data from the expected values generated by the model can trigger maintenance actions. A good example in this case is the physics-based model, which is suitable due to the availability of mathematical models for a system, based on its physical fundamentals. Physics-based modeling provides a means to handle the bias in the measured data and can explain the behavior of a system in a wide range of operating conditions. However, it requires a deep understanding of the failure mechanisms and associated dynamics, which can be challenging to achieve with complex systems that have multiple competing and interrelated damage mechanisms, varying operational conditions, and multifarious data characteristics [45].

5.2. Machine Learning Techniques Used in Industrial Maintenance

This subsection presents an overview of traditional or shallow machine learning and the more sophisticated deep learning techniques.

5.2.1. Shallow Machine Learning

Shallow machine learning (SML), also known as traditional or classical machine learning, includes a set of algorithms and techniques that are highly suited to handle structured data and to perform specific tasks like classification, regression, and clustering. The most relevant implemented methodology include: statistical models [253,301,436,492,494], neural networks [6,107,192,276,431], probabilistic models [14,400], decision support hierarchical models [3,23,53,57,121,395,493], density-based models [453], ensemble learning [24,33,57,73,85,114,115,197,235,272], and boosting learning [12,451]. Regarding artificial neural networks (ANNs), in SML, these techniques typically involve the use of a limited number of layers in the model, characterized by a relatively simple architecture with limited parameters.

In the context of industrial maintenance, SML techniques encompass a variety of algorithms. These algorithms rely on feature engineering, where relevant features are manually extracted or selected from the data and the model then learns to make a mapping from input data to an output label or prediction based on these features.

SML applications in predictive and condition-based maintenance can be described as follows:

A

Predictive maintenance (PdM): SML techniques, such as classification and regression, can be used to predict the optimal timing for preventive maintenance of industrial assets. Models can be trained on historical data to identify patterns that indicate when equipment is likely to fail or require maintenance [2,41,57].

B

Condition-based maintenance (CBM): SML is valuable for analyzing sensor data in real-time. Anomalies in sensor readings of industrial assets can trigger maintenance alerts or alarms, helping to identify deteriorating equipment conditions [30].

5.2.2. Deep Learning

Deep learning (DL) is a technique designed to mimic the neural network structure of the human brain [455]. It is based on artificial neural networks (ANNs). These ANNs are constructed by multiple layers (deep neural networks) to model complex patterns and representations in data [197,409]. DL is especially suited for tasks involving high-dimensional data, sequential data, and unstructured data. Hence, unlike SML, DL models are characterized by their ability to automatically learn hierarchical representations from raw data through multiple layers of interconnected nodes named neurons. These architectures can process and extract complex patterns and features directly from the data, eliminating the need for manual feature engineering [51].

In the field of maintenance of industrial assets, DL techniques outperform SML in handling unstructured data types like images, sensor data, text, and time series information. They offer powerful feature extraction capabilities for classification and regression purposes [7,53,525]. They show remarkable performance in predictive maintenance, anomaly detection, and fault diagnostics by discerning intricate patterns and relationships within the data, offering high accuracy and predictive capabilities.

There are two main areas where DL applications can be used in industrial maintenance, predictive and condition-based maintenance, and they are summarized as follows:

A

Predictive maintenance (PdM): DL models, particularly recurrent neural networks (RNNs) [88,93,203,397] and long short-term memory (LSTM) networks [79,109,412], can capture complex temporal or natural language patterns in historical maintenance data of industrial assets. These models are suitable for predicting equipment failures based on sequences of events.

B

Condition-based maintenance (CBM): DL is highly proficient in processing high-dimensional sensor data, images, and audio signals [16,51,229,360]. Convolutional neural networks (CNNs) and recurrent neural networks (RNNs) can detect anomalies in real-time sensor readings (stream of data) [53], enabling proactive maintenance of industrial assets.

The summary of DM approaches, considering the uniqueness of the techniques applied by category, reveals a diverse landscape (see Figure 5). While SML is used in most diverse cases, featuring 60 unique algorithms, DL follows with 42 unique algorithms, showcasing its significant but with slightly less diverse range of algorithms. MB approaches comprise 16 unique algorithms, indicating a more specialized application of techniques. Lastly, RB approaches are the least diverse, represented by only two algorithms. This distribution highlights the extensive variety and application of SML and DL techniques in DM for industrial maintenance, contrasted by the more focused use of MB and RB approaches.

Considering the cumulative mentions of all algorithms referenced in research studies for each technique category, SML stands out as the most frequently cited, with a total of 638 mentions. This indicates its widespread application and importance in the field. DL algorithms follow with 405 total mentions, reflecting their growing prominence and effectiveness in handling complex data. MB techniques are mentioned 147 times, suggesting a more specialized but still significant role. In contrast, RB techniques are the least mentioned, with only five references, highlighting their limited application compared to other techniques. The cumulative mentions of algorithms also consider the multiple algorithms mentioned in the 534 research studies of our repository, i.e., in the survey [7], the following algorithms of the SML category are mentioned: support vector machines (SVRs), artificial neural networks (ANNs), decision tree (DT), and k-nearest neighbors (k-NN). For the DL category, the survey mentions autoencoders (AEs), convolutional neural networks (CNNs), long short-term memory (LSTM) networks, deep belief networks (DBNs), recurrent neural networks (RNNs), generative adversarial networks (GANs), and restricted Boltzmann machines (RBMs).

DM and ML techniques can manage either a single parameter (mono-variable) or multiple parameters (multi-variable) for learning tasks. In the category of SML, 35 algorithms were found to be capable of handling a single parameter, while 34 algorithms can manage multiple parameters, demonstrating a balanced capability across different modeling complexities. In DL, 17 algorithms are designed for single-parameter tasks, and an equal number of algorithms (17) handle multi-parameter modeling, indicating versatility in addressing various data complexities for the problem in study. For MB approaches, there are four algorithms that can handle a single parameter and another four algorithms that can handle multiple parameters, reflecting a versatility depicting the limited variety of algorithms. Lastly, in RB techniques, a single algorithm is found to handle both single and multiple parameters, showcasing its flexibility despite its limited application.

The different DM techniques are used for purposes related to the PdM or CBM model strategies. A summary of the research study mentions is presented in Figure 6. The distribution of research studies mentioning the use of specific analytical approaches across different purposes—such as forecasting, fault detection, degradation, pre-processing, synthetic data generation, transfer learning, and objective optimization—reveals distinct patterns in the application of SML, DL, MB, and RB techniques. SML dominates in fault detection (47 references) and forecasting (21 references), indicating its strength in identifying and predicting operational issues. DL also shows a strong presence in fault detection (34 references) and excels in pre-processing (17 references), suggesting its usefulness in handling complex data preparation tasks. MB approaches are primarily utilized for forecasting (eleven references) and degradation analysis (eight references), while RB techniques are minimally applied, with only a single reference in fault detection.

In terms of asset complexity, as defined in Section 4, Figure 7 presents the number of research studies that have proposed a solution ranging from simple to complex components or systems. The classification of unique assets by complexity—single component (1C), multiple components (1C+), single system (1S), and multiple systems (1S+)—reveals distinct patterns in the application of different analysis techniques. SML shows a broad application across different complexity levels, with notable mentions: ten references for single components, four for multiple components, twelve for single systems, and three for multiple systems. DL exhibits a more balanced distribution, with six, five, five, and three references, respectively, indicating its adaptability to both simpler and more complex components and systems. MB approaches have a narrower scope, primarily focused on higher complexity of multiple components and simpler single systems, with an overall of one, two, four, and one references, respectively. RB methods appear minimally, with a single mention for the most complex category (1S+), reflecting their limited application in more intricate asset systems.

5.3. Categorization of Machine Learning

In machine learning (ML), the choice of the learning paradigm depends on the nature of the data, the specific problem to be solved, and the availability of labeled data. The four categories of ML are described as follows:

A

Supervised learning is a category of ML where the algorithm is trained on labeled data [456], meaning that each input data point is associated with a corresponding target or label [2]. The goal is to learn a mapping from inputs to outputs, or predictions, based on the labeled training data. For predictive maintenance, supervised learning is used when the information about the occurrence of failures is present in the modeling dataset [16]. Supervised learning is used for tasks such as classification [95,112,484] (assigning data points to predefined categories or classes) and regression [38] (predicting numerical values). Common algorithms include decision tree, support vector machine, neural networks, boosting algorithms, and statistical methods [72].

B

Unsupervised learning involves training ML algorithms on unlabeled data [456], where there are no predefined target labels [2,3]. The goal is to discover hidden patterns, structures, or relationships within the data. For predictive maintenance, unsupervised learning is used when only the process information is available and no historical maintenance data exist [16]. Unsupervised learning is used for tasks such as clustering [54,95] (grouping similar data points), learning effective feature representation from raw signals [3,14] (reducing the number of features while preserving important information), and anomaly detection (identifying unusual data points). Common algorithms include K-means clustering, principal component analysis (PCA), fuzzy C-means method (FCM), and neural network autoencoder [14,32,507].

C

Semi-supervised learning combines elements of both supervised and unsupervised learning. In semi-supervised learning, a small portion of the data is labeled, while the majority of the data remain unlabeled [455]. The algorithm leverages the labeled data to improve its understanding of the entire dataset [14]. Semi-supervised learning is useful when obtaining labeled data is expensive, time-consuming, or difficult to achieve in an industrial environment [61]. It can be applied in scenarios where only a subset of data points is labeled, such as in text classification or image recognition with limited labeled samples [459]. Techniques include self-training and co-training. Common algorithms include an extension of support vector machine (transductive support vector machine), neural networks based on graphs (graph neural networks) or encoders/decoders (cross-domain stacked denoising autoencoder), and booting learning (TrAdaboost).

D

Reinforcement learning in this type of learning, an agent interacts with an environment and learns to make decisions or take actions to maximize a cumulative reward [144]. The agent learns through a process of trial and error, receiving feedback in the form of rewards or penalties based on its actions [455,456]. Reinforcement learning is used for tasks such as classification, regression, optimization, and domain adaptation [7,117,144,187]. Reinforcement learning is well-suited for problems involving sequential decision-making, such as game playing, robotics control, autonomous driving, and recommendation systems. Key components include the agent, environment, actions, rewards, and a policy that dictates the behavior of the agent. Common algorithms are deep learning algorithm variants, e.g.,: the deep Q-Network [117].

The number of research studies with mentions of unique algorithms of the detailed four categories are: supervised learning (59), unsupervised learning (33), semi-supervised learning (8), and reinforcement learning (3).

In research studies examining shallow machine learning (SML) and deep learning (DL) techniques, the number of mentions of unique algorithms across different categories of ML is as follows: SML has forty references for supervised learning, thirteen for unsupervised learning, five for semi-supervised learning, and none for reinforcement learning. In contrast, DL has nineteen references for supervised learning, twenty for unsupervised learning, three for semi-supervised learning, and three for reinforcement learning.

5.4. Overview of DM Algorithms Used in Industrial Maintenance

The diversity of algorithms employed in data mining (DM) within the field of industrial maintenance of industrial assets is extensive.

Figure 8 illustrates the evolution of algorithmic diversity over time. Before the year 2000, only a limited number of algorithms were explored, reflecting the incipient nature of DM applications in industrial maintenance. However, a noticeable upward trend began around 2000, with a gradual increment in the number of algorithms studied. This trend continued to gain impulse, reaching a significant number of works around 2013. Since 2013 and up to the year 2023, there has been a marked surge in the variety and application of different algorithms, indicating a growing and sustained interest in leveraging learning techniques within the industrial maintenance domain, reaching a peak in 2019. In the following figures, articles from the year 2024 are not considered due to the specific topics defined for the search criteria.

Figure 9 remarks the presence and degree of interest of each DM technique over the years by showing the number of cumulative mentions of algorithms. The indicated cumulative mentions consider the repetition of an algorithm among research studies within a year, and the sum of all mentions is more than in the core repository due to the mention of multiple algorithms in the 534 articles considered.

The revision of studies has put more than 110 different algorithms on the table for DM in industrial maintenance, as Figure 10, Figure 11 and Figure 12 present.

For the rule-based techniques, the RIPPER (repeated incremental pruning to produce error reduction) [50] algorithm was mentioned once in 2018, and expert systems (ES) [12] was mentioned three times; first in 2005, then in 2020, and last in 2022.

As Figure 10 presents, among the model-based approaches in industrial maintenance, the wiener process [1], renewal process [1], proportional hazard model [526], and linear regression [75] algorithms have been frequently employed to enhance maintenance strategies on industrial assets.

Considering the shallow machine learning algorithms, the most frequently applied are support vector machines (SVMs) [53], artificial neural networks (ANNs) [53], random forest (RF) [73], decision tree (DT) [53], and k-nearest neighbors (k-NN) [53], as Figure 11 shows. These algorithms are valued for their ability to handle complex, high-dimensional data and provide robust predictive capabilities.

Deep learning algorithms, such as convolutional neural networks (CNNs) [2], long short-term memory (LSTM) [7] networks, deep belief networks (DBNs) [53], autoencoders (AEs) [7], and recurrent neural networks (RNNs) [7], have emerged as the most frequently applied technique in industrial maintenance, as Figure 12 presents. These algorithms have proven to be particularly effective in handling complex tasks such as predictive maintenance, fault detection, and anomaly detection.

Figure 13 presents the 32 most frequently mentioned algorithms with at least three mentions that were identified from the reviewed 534 research studies. Additionally, the figure provides an overview of the specific use purposes these algorithms serve in predictive maintenance (PdM) and condition-based maintenance (CBM) strategies. This comprehensive visualization highlights not only the prevalence of certain algorithms but also their diverse applications across different maintenance approaches, offering valuable insights into the current trends and focal points within the field of industrial maintenance. The support vector machine (SVM) algorithm stands out from the other algorithms with 136 mentions. Note that the number of mentions is associated with corresponding specific use purposes. Hence, a mere mention of the algorithm is not included in the count of this representation; therefore, the total count, including mentions without a specific use purpose, may be larger.

As a general overview, Figure 14 provides a comprehensive summary of all algorithms, systematically organized in a hierarchical structure based on the learning technique category and the specific method or strategy employed in the data mining (DM) analysis. This hierarchical presentation not only categorizes the algorithms according to their learning paradigms but also emphasizes the methodological diversity within the field, showcasing how different strategies are applied across various DM approaches in industrial maintenance.

5.5. Summary of the Most Frequently Referenced Algorithms by Technique

This section briefly describes the most popular algorithms categorized by technique. Each category is explored to showcase the algorithms that have gained prominence, reflecting their applicability and impact on the field of maintenance of industrial assets.

5.5.1. Shallow Machine Learning Representative Algorithms

The following algorithms from the shallow machine learning (SML) technique are the most relevant in the field of machine learning (ML) for predictive maintenance (PdM) and condition-based maintenance (CBM) models in industrial assets. These algorithms, ordered by the number of references, represent the forefront of SML techniques applied to industrial maintenance.

A

Support vector machine (SVM): This is used for classification tasks, typically by finding the optimal hyperplane that best separates data points and maximizing the margin between them [53].

The traditional SVM, and its many variations, were implemented in more than 140 studies [3,7,8,9,10,11,12,14,16,23,24,27,35,50,53,62,63,73,76,85,93,102,103,107,108,110,121,131,151,158,162,167,179,197,199,200,201,202,205,217,218,219,220,221,222,223,224,225,232,237,238,240,244,245,246,247,248,249,250,251,252,253,254,269,275,296,297,298,299,300,301,302,303,304,305,306,307,308,309,310,311,312,313,314,315,316,317,318,319,320,321,322,323,324,325,326,327,328,373,380,381,383,384,390,397,401,402,403,405,407,409,418,420,421,428,429,430,436,437,439,440,442,443,446,448,451,464,469,472,477,478,480,481,482,490,491,493,494,495,496,497,498].

B

Artificial neural network (ANN): This class of ML models is inspired by the neural structure of the human brain. They consist of multiple interconnected nodes (neurons) arranged in layers. Each neuron receives inputs, processes them using activation functions, and passes the result to the next layer. The layers typically consist of input neurons (receiving data), hidden neurons (processing data), and output neurons (producing the output of the model) [53]. Depending on the architecture of the output layer, an ANN can perform classification or regression tasks [14].

The traditional ANN was implemented in more than 60 studies [3,6,7,8,10,12,14,16,45,47,53,73,80,106,107,119,164,167,187,194,197,200,210,211,213,214,269,278,288,328,385,391,403,422,424,432,434,441,498,516,517,521].

C

Random forest (RF): This ensemble learning method combines multiple DTs to improve predictive accuracy [73] and reduce overfitting [3].

RF RF has demonstrated superior performance compared to other classification algorithms and has been widely applied for ensemble learning. The ensemble learning model builds a set of models which are then combined to improve overall performance, particularly when dealing with complex systems. Generally, the ensemble learning model will perform better than any base learner if an adequate ensemble strategy is adopted [73].

More than 30 studies have successfully applied RF [3,7,8,10,14,16,24,33,50,57,73,85,105,110,111,115,116,121,167,197,235,257,271,272,273,345,382,394,409,447,475,478,479,506].

D

Principal component analysis (PCA): This is used to reduce the dimensionality of high-dimensional data while retaining as much variance as possible. It transforms data into a new set of uncorrelated variables called principal components.

The implementation of PCA PCA is detailed in more than 30 work studies [7,14,23,32,53,63,73,75,85,111,133,136,169,231,263,297,337,343,387,415,416,417,428].

E

k-nearest neighbors (k-NN): This algorithm classifies data points based on the majority class among their k-nearest neighbors in the feature space [53]. It is also used for regression tasks by finding the k-nearest data points to a query point and making predictions based on their values [7].

k-NN was implemented in more than 20 research studies [7,42,53,56,85,108,121,174,175,191,197,237,256,274,329,330,331,332,333,334,335,336,337,338,339,409,438,513,519].

F

Decision tree (DT): A decision tree is used for classification [16] and regression [57,395] tasks by recursively splitting data into subsets based on the most informative features. DTs are non-parametric methods that learn decision rules from input data. Starting from a root node, the model makes decisions that allow it to go through different paths until it reaches a leaf node representing a prediction. The speed of training of these models plus its rapid prediction have helped us to use them in real-time problems [53].

There are more than 20 works that have implemented DT [3,7,8,10,14,16,53,57,115,118,121,165,241,268,270,271,341,342,343,344,395,398,409,418,483,493].

G

Logistic regression (LGR): This is a binary classification tool that models the probability of a data point belonging to one of two classes. LGR models can easily provide industrial experts with interpretability [403]. LGR models have also outperformed other well-known classifiers in terms of performance [3].

This algorithm was applied in more than 15 research studies [3,49,85,159,160,161,269,394,403,527].

H

Support vector regression (SVR): This is a regression technique that uses a variation of support vector machines (SVMs) to predict continuous numerical values, based on the searching of the hyperplane containing the most points [53,197,401]. It aims to find a regression function that best fits the data while controlling for errors. In some cases, the SVR model performs better, with considerably less estimation error compared with other ML models [7].

SVR was referenced in more than 10 research studies [3,7,14,53,58,91,102,162,197,321,401,462].

I

Linear discriminant analysis (LDA): This is used for dimensionality reduction [85], but it focuses on maximizing class separability, making it suitable for classification tasks [42,397]. The algorithm is sufficiently fast to meet real-time requirements in industry [42].

LDA was implemented in seven research studies [14,42,85,129,397,407,478].

J

Gradient boosting (GB): This is effective for binary classification. It can handle complex relationships in the data. GB is an ensemble learning technique that combines multiple weak learners to create a strong predictive model.

GB was referenced in six work studies [8,10,23,54,271,398].

K

Density-based spatial clustering of applications with noise (DBSCAN): This density-based clustering algorithm is used to discover clusters in spatial data based on the density of data points [63]. It has been proven to have a good performance with many datasets.

DBSCAN was implemented in four work studies [46,63,129,453].

Table 6. Comparative analysis of the advantages and disadvantages, supported data types, and applications of shallow machine learning algorithms in industrial maintenance.

Algorithm	Advantages	Disadvantages	Supported Data Types	Application in Industrial Maintenance	Mentions Number
Support vector machine (SVM)	- Effective in high-dimensional spaces; - Versatile due to different kernel functions.	- Less effective with large datasets; - Choice of kernel and parameters can be challenging.	Numeric.	- Fault classification [62,481,520]; - Anomaly detection [308].	147
Artificial neural network (ANN)	- Ability to model complex relationships in data; - Flexibility in handling different data types and structures; - Capability to learn and adapt to nonlinear relationships.	- Prone to overfitting with small datasets; - Requires a large amount of data for training; - Complex network architecture may be challenging to optimize.	Numeric, categorical.	- Fault prognosis and remaining useful life estimation [3,7,10,53,80,106,119,328,391]; - Fault detection and classification [3,6,10,53,107,164,167,200,210,211,213,214,269,278,288,385,422,432,434,441,498,516,517,521]; - Decision support systems [12,47].	68
Random forest (RF)	- Reduces overfitting by aggregating multiple decision trees; - Provides feature importance ranking.	- Can be computationally expensive for large datasets; - Less interpretable compared to individual trees.	Numeric, categorical.	- Fault detection [16,33,111,115,273]; - Fault prognosis [110,447,479]; - Feature selection [121].	32
Principal component analysis (PCA)	- Reduces dimensionality while preserving most important information; - Helps in visualizing high-dimensional data.	- May lose interpretability of original features; - Assumes linear relationships among variables.	Numeric.	- Fault diagnosis [111,133,136]; - Sensor data compression [63,75,85,231,263]; - Feature selection [14,231,337].	31
k-nearest neighbors (k-NN)	- Simple and intuitive; - No training involved, lazy learning approach.	- Computationally expensive for large datasets; - Sensitive to irrelevant features and distance metrics.	Numeric, categorical.	- Fault detection [53,197]; - Anomaly detection [42]; - Degradation prognosis [7,53].	24
Decision tree (DT)	- Interpretability, easy to understand and visualize.	- Prone to overfitting, especially with deep trees; - Can be sensitive to small variations in data.	Numeric, categorical.	- Fault classification [3,16,53,121,493]; - Decision support [14].	21
Logistic regression (LGR)	- Simple and efficient for linearly separable data; - Provides probabilistic interpretations.	- Assumes linear relationship between features and target; - Not suitable for complex relationships in data.	Numeric.	- Failure diagnosis [159,269]; - Degradation [49,161,527]; - Failure prediction [160,394].	19
Support vector regression (SVR)	- Effective in capturing complex relationships in data; - Handles outliers and nonlinear relationships.	- Sensitive to choice of kernel and regularization parameters; - Computationally intensive for large datasets.	Numeric.	- Remaining useful life estimation and fault prognosis [3,7,14,58,91,162]; - Anomaly detection [3,321].	12
Linear discriminant analysis (LDA)	- Feature extraction technique for classification; - Maximizes separability between classes.	- Requires the assumption of normality and equal covariances; - Can overfit with small datasets.	Numeric.	- Fault forecasting [407]; - Anomaly or fault detection [14,42].	7
Gradient boosting (GB)	- Builds strong predictive models; - Handles different types of data well.	- Prone to overfitting if not properly tuned; - More complex compared to individual decision tree.	Numeric, categorical.	- Fault detection [54]; - Failure prognosis or remaining useful life estimation [8,23].	6
Density-based spatial clustering of applications with noise (DBSCAN)	- Robust to outliers and noise in data; - Does not require the number of clusters as an input.	- Sensitivity to the choice of distance threshold (epsilon); - Struggles with varying density clusters.	Numeric.	- Data clustering [453]; - Anomaly or fault detection [46].	4

provides a comprehensive comparison, with the corresponding references, of SML algorithms, highlighting their advantages, disadvantages, supported data types, applications in industrial maintenance, and the number of mentions.

5.5.2. Deep Learning Representative Algorithms

The following algorithms from the deep learning (DL) technique type are the most relevant in the field of machine learning (ML) for predictive maintenance (PdM) and condition-based maintenance (CBM) models of industrial assets. These algorithms, ordered by the number of references, represent the forefront of DL techniques applied to industrial maintenance.

A

Autoencoders (AE) are a type of neural network that can be used for dimensionality reduction and anomaly detection. They consist of an encoder and a decoder, which are connected in a determinate architecture.

AEs were referenced in more than 30 research studies [2,3,7,12,30,53,93,142,153,160,165,169,241,243,244,262,267,273,274,317,347,356,371,419,427,444,468,528].

Variations of the AE algorithm were also applied for PdM and CBM. These variations included the deep autoencoder (DPAE) [511], stacked autoencoder (SKAE) [142,234,347], stacked sparse autoencoder (SKSAE) [346], sparce autoencoder (SAE) [7,241], variational autoencoder and kernel density estimation (VAE-KDE) [148], stacked contractive autoencoder (SCAE) [427], deep denoising autoencoder (DDAE) [448], multi-level-denoising autoencoder (ML-DAE) [169], stacked denoising autoencoder (SKDAE) [185,267], and cross-domain stacked denoising autoencoder (CD-SDAE) [243].

B

Convolutional neural networks (CNNs) are inspired by biological visual perception mechanisms. They have unique structural characteristics, such as local connection, weight sharing, and pooling, which enables CNNs with strong feature learning and representation ability [2]. CNNs are designed for image and spatial data [14].

CNNs were referenced in more than 40 research studies [2,7,8,9,10,12,14,24,51,53,67,86,93,112,131,137,170,183,186,197,204,227,229,230,238,239,242,265,266,267,359,360,361,362,363,364,365,366,392,393,415,424,459,465,471,529].

C

Long short-term memory (LSTM) networks are a specialized variant version of recurrent neural networks (RNNs), usually studied as a different algorithm in the technical literature. LSTM was specially designed to address the vanishing gradient problem, which is enhanced by adding “forget” gates and has shown great capability in memorizing and modeling long-term dependency in data. LSTM is one of the most commonly used models when working with time-dependent data [7].

LSTM was referenced in more than 35 work studies [7,8,9,10,14,28,36,53,58,69,79,83,91,95,99,109,142,152,156,180,236,356,358,359,369,392,393,397,406,408,410,412,413,415,424,474,526].

D

Recurrent neural networks (RNNs) are designed for sequential data analysis and modeling [14,79]. They have feedback in their hidden connections that allows information to persist over time and keep a memory of previous inputs in the internal state of the network [7], making them suitable for tasks like natural language processing [397,424] and time series analysis [7,203]. However, due to being trained with back propagation through time (BPTT), RNNs always have the notorious gradient vanishing/exploding issue [7].

RNNs were referenced in more than 30 research studies [2,7,8,14,51,53,79,83,88,93,95,123,152,166,203,283,289,293,356,364,366,397,402,406,408,417,423,424,454,465].

E

Deep belief networks (DBNs) are a type of generative neural network composed of multiple layers of stochastic latent variables. They can be viewed as a composition or stacking of simple, unsupervised networks such as restricted Boltzmann machines (RBMs) [7,53] or other similar structures. DBNs, with multiple hidden layers, can remove the dependence on prior knowledge and adaptively extract features. They are also able to process nonlinear high-dimensional data, thereby effectively avoiding problems, such as dimensional disaster [2]. DBNs are trained in a layer-wise manner, enabling unsupervised feature learning and efficient representation of complex data distributions.

DBNs were referenced in more than 20 work studies [2,7,10,12,45,53,154,155,176,240,260,263,264,353,354,355,389,463,473,508,512,515].

F

Generative adversarial networks (GAN) are an unsupervised method that can generate realistic samples via a minimax game between two networks. They consist of a generator network to generate samples and a discriminator network to judge the likeness of the generated samples, which are trained simultaneously through adversarial training. The generated realistic fake data fit within the distribution of the training data, outperforming traditional over-sampling methods by a large margin.

GANs were referenced in more than 15 research studies [2,7,12,13,14,153,226,227,228,348,350,352,388,510,511,530].

G

Gated recurrent units (GRUs) are a variant of recurrent neural networks (RNNs), usually studied as a different algorithm in the technical literature. GRUs were designed with an architecture to address the vanishing gradient problem in standard RNNs and to capture long-term dependencies in sequential data efficiently [7]. GRUs have gates that regulate the flow of information, allowing them to retain important information and discard irrelevant details over sequences.

There are more than 10 research studies that have implemented GRUs in their studies [2,7,10,120,166,356,357,397,454,465,531].

H

Transformer for self-attention (TSA) models are a neural network architecture that is primarily used in natural language processing tasks, but has found applications in various domains due to its ability to handle sequential data efficiently. At its core, the transformer model utilizes a mechanism called self-attention or scaled dot-product attention. Self-attention allows the model to weigh the importance of different input tokens (words, symbols, or segments) when processing a specific token, enabling it to capture dependencies between tokens in an input sequence [366].

TSA was referenced in more than five research studies [157,266,267,268,366,367,393,531].

I

Graph neural networks (GNNs) are specialized neural networks designed to operate on graph-structured data. They can capture complex relationships and dependencies between entities in graphs, making them suitable for tasks involving relational data [2].

GNNs have been referenced in more than 5 research studies [2,206,368,369,370,371,372,484,522].

J

Deep reinforcement learning (DRL), a combination of reinforcement learning (RL) and deep neural networks (DNNs), has shown its potential to be the new generation of decision-making frameworks for complex systems. DRL agents can typically learn by themselves to establish successful optimal policies for gaining maximum long-term rewards.

There are three research studies that have implemented DRL [7,144,187].

Table 7. Comparative analysis of the advantages and disadvantages, supported data types, and applications of deep learning algorithms in industrial maintenance.

Algorithm	Advantages	Disadvantages	Supported Data Types	Application in Industrial Maintenance	Mentions Number
Autoencoder (AE)	- Unsupervised feature learning; - Data denoising and feature extraction.	- Limited interpretability; - Reconstruction loss might not capture meaningful features; - Sensitivity to hyperparameters and architecture choices.	Various data types (numeric, image).	- Feature extraction [2,7]; - Anomaly detection [7]; - Multi-sensor data fusion [7].	33 (+34 variations)
Convolutional neural network (CNN)	- Excellent for image feature extraction; - Effective in capturing spatial hierarchies in data; - Parameter sharing reduces the number of parameters.	- Limited ability to capture sequential dependencies; - Can be computationally expensive, especially with large images; - Ussually requires large amounts of labeled data for training.	Various data types (numeric, image).	- Fault detection [2,24,67,93,227,360,363]; - Feature extraction [2,7,360]; - Fault prognosis and remaining useful life estimation [7,197].	49
Long short-term memory network (LSTM)	- Captures long-term dependencies in sequential data; - Addresses vanishing gradient problem in RNNs.	- Complexity increases with larger network architectures; - May overfit with small datasets.	Sequential data.	- Remaining useful life estimation and fault prognosis [58,69,79,109,142,412]; - Anomaly detection without pre-processing [180,358].	38
Recurrent neural network (RNN)	- Handles sequential/temporal data effectively; - Variable-length inputs/outputs.	- Can suffer from vanishing/exploding gradient problems; - Long-range dependencies can be challenging to capture.	Sequential data.	- Fault prognosis and remaining useful life estimation [7,88,93]; - Fault diagnosis [7,203,293].	31
Deep belief networks (DBNs)	- Unsupervised feature learning; - Efficient representation learning of complex data structures.	- Complex training procedure; - Sensitivity to hyperparameters and architecture choices.	Various data types (numeric, image, text).	- Fault diagnosis and prognosis [2,10,53,176,240,260,264,389,463,473,508,515]; - Feature learning [2,7,12,263,353,354]; - Feature fusion [354,355]; - Health index determination or remaining useful life estimation [12,45,154,155,512].	25
Generative adversarial network (GAN)	- Generates realistic synthetic data; - Enables unsupervised learning of data distributions.	- Training instability and mode collapse can occur; - Requires careful tuning and balancing of generator and discriminator.	Various data types (numeric, image, text).	- Synthetic data generation [2,7,226,352,510]; - Data augmentation [2,14,227,348,350]; - Fault identification [7,13,228,511]; - Remaining useful life estimation [153].	21
Gated recurrent units (GRU)	- Captures long-range dependencies in sequential data; - Handles vanishing gradient problem better than standard RNNs.	- May suffer from overfitting with small datasets; - Limited memory compared to LSTM.	Sequential data.	- Health indicator prediction or remaining useful life estimation [7,120,454]; - Early fault diagnosis and prognosis [2,10,356,357,531]; - Text classification [397]; - Feature learning [166,465].	13
Transformer for self-attention (TSA)	- Handles sequential data efficiently; - Captures long-range dependencies effectively; - Computation Parallelization in order to enhance efficiency.	- Complexity in understanding and training; - Computationally intensive, especially with large input sequences; - Limited interpretability compared to other models.	Sequential data.	- Fault detection and diagnostics [157,267,531]; - Health state prognosis [393]; - Feature extraction [157]; - Transfer learning [266].	12
Graph neural network (GNN)	- Captures relationships in graph-structured data.	- Computationally intensive, especially with large graphs.	Graph data.	- Dependency analysis [2]; - Fault diagnosis and prognosis [2,368,369,370,371,372,522].	9
Deep reinforcement learning (DRL)	- Learns optimal decision-making policies through trial and error.	- Sample inefficiency; - High computational requirements.	Various data types (numeric, image, text).	- Proactive maintenance management [7,144,187]; - Remaining useful life estimation [144]; - Health indicator learning (HIL) [7]; - Fault diagnosis [7].	3

provides a comprehensive comparison, with the corresponding references, of DL algorithms used in industrial maintenance, highlighting their advantages, disadvantages, supported data types, applications in industrial maintenance, and the number of mentions.

5.5.3. Model-Based Technique Representative Algorithms

The following model-based (MB) technique algorithms, ordered by the number of references, are the most relevant in the field of data mining for predictive maintenance (PdM) and condition-based maintenance (CBM) models of industrial assets.

A

Proportional hazard model (PHM): A statistical model used to analyze time-to-failure data [526]. It assesses the impact of multiple factors on the hazard rate (probability of failure per unit time).

PHM was referenced in more than 20 work studies [1,2,3,41,45,58,61,68,69,71,73,78,83,89,141,351,392,402,411,458,475].

B

Markov process (MP): A stochastic model that describes transitions from one state to another in a sequence of discrete steps. An extension of MP is the hidden Markov model (HMM), which deals with both observable and hidden states. It is a statistical model that includes observed states (visible or measurable) and hidden states (unobserved or latent). Some CBM models that assume discrete-state deterioration are modeled by MP [1].

MP and its extension, HMM, were referenced in more than 15 research studies [1,9,30,39,40,45,109,163,193,233,340,378,379,461,462,509].

C

Linear regression (LR): A statistical method used to establish a relationship between dependent and independent variables by fitting a linear equation [75].

LR was referenced in eight research studies [7,58,75,101,138,398,409,506].

D

Autoregressive integrated moving average (ARIMA): A time series forecasting method that models the relationship between a series of observations and commonly is used for predicting future points in a time series. It incorporates autoregression, differencing, and moving average components.

ARIMA was referenced in six research studies [10,30,58,147,390,391].

E

Wiener process (WP): A continuous-time stochastic process used to model random fluctuations in various systems [1].

WP was referenced in five research studies [1,9,20,139,146].

F

Analysis of variance (ANOVA): A statistical technique used to compare the means of two or more groups to determine if there are statistically significant differences between them.

ANOVA was referenced in four technical studies [6,42,409,459].

G

Polynomial regression (PR): A method that extends linear regression by fitting a polynomial equation to the data. It is useful when the relationship between variables is nonlinear. PR can model more complex relationships of data.

PR was referenced in two studies [75,182].

H

Renewal process (RP): A stochastic process used in probability theory and statistics to model sequences of events occurring in continuous time, where the time intervals between events follow a certain probability distribution. It represents a series of events that occur at irregular intervals, with each event marking the occurrence of some specific action or state change.

RP was referenced in two research studies [1,25].

I

Failure mode and effects analysis (FMEA): A semi-qualitative method used to prevent failures and analyze the risks of a process by identifying causes and effects on the system to determine the actions that will be used to inhibit failures and prioritizing them based on severity, occurrence, and detectability. This approach has been used in the aerospace, nuclear, automotive, and semiconductor industries to support design fault, safety, logistic support, testability, and other related functions [30].

Table 8. Comparative analysis of the advantages and disadvantages, supported data types, and applications of model-based algorithms in industrial maintenance.

Algorithm	Advantages	Disadvantages	Supported Data Types	Application in Industrial Maintenance	Mentions Number
Proportional hazard model (PHM)	- Predicts time-to-failure; - Assesses impact of multiple factors.	- Assumptions on hazard proportionality may not hold; - Sensitive to outliers.	Time-to-failure data.	- Degradation [1,141]; - Optimal maintenance policy [458]; - Performance indicator calculation [526].	21
Markov process (MP)	- Models sequential dependencies; - Defines transition probabilities.	- Assumes Markov property which may not be applicable in all cases; - Limited memorization property.	Discrete-state data.	- Maintenance planning and strategy optimization [30,39,40,509]; - Fault prediction [45]; - Multi-sensor fusion [340]; - Fault diagnosis [163,193,233,378,379,461,462].	19
Linear regression (LR)	- Simple to implement and interpret; - Works well with linear relationships.	- Assumes linear relationship between variables; - Sensitive to outliers.	Continuous data.	- Fault prognosis and remaining useful life estimation [7,75,101,398,409,506]; - Anomaly detection [138].	8
Autoregressive integrated moving average (ARIMA)	- Handles time series data; - Captures trends and seasonality.	- Sensitive to parameter selection; - Assumes stationarity.	Time series data.	- Failure prognosis and remaining useful life estimation [10,58,147,390]; - Degradation [391].	6
Wiener process (WP)	- Models continuous-time stochastic processes; - Captures random fluctuations.	- Complex modeling for non-stationary processes.	Continuous data.	- Fault prediction [9]; - Degradation [1,20,139].	5
Analysis of variance (ANOVA)	- Tests differences between groups; - Identifies significant factors.	- Assumes independence and normality; - Sensitivity to outliers.	Categorical and continuous data.	- Feature identification [6,42]; - Performance assessment [459].	4
Polynomial regression (PR)	- Captures nonlinear relationships; - More flexible than linear regression.	- Susceptible to overfitting; - Interpretability decreases with higher-degree polynomials.	Continuous data.	- Health state estimation [182].	2
Renewal process (RP)	- Models recurring events; - Considers inter-arrival times.	- Requires detailed event data; - Assumptions on event occurrence may not hold.	Time series data.	- Condition-based maintenance strategy [1]; - Maintenance cost estimation [25].	2
Failure mode and effects analysis (FMEA)	- Systematic identification of potential failure modes; - Prioritization of risks.	- Subjective scoring; - Difficulty in quantifying severity, occurrence, and detectability.	Qualitative and Quantitative.	- Proactive maintenance [30].	1

provides a comprehensive comparison, with the corresponding references, of the most relevant MB algorithms used in industrial maintenance, highlighting their advantages, disadvantages, supported data types, applications in industrial maintenance, and the number of mentions.

5.5.4. Rule-Based Technique Representative Algorithms

The following algorithms, ordered by the number of references, from the rule-based (RB) technique type are the most relevant in the field of data mining for predictive maintenance (PdM) and condition-based maintenance (CBM) models of industrial assets.

A

Expert systems (ES): These systems incorporate knowledge-based rules to make decisions or provide expert-level advice in specific domains. ES are computer applications that simulate the decision-making ability of a human expert in a specific domain by employing a knowledge base and inference engine [12]. These systems use a collection of rules and knowledge to make decisions or solve problems [14,483].

B

Repeated incremental pruning to produce error reduction (RIPPER): This rule-based data mining algorithm is used for classification tasks. It operates by creating a set of rules that can predict a target variable by iteratively growing and refining rules through incremental improvements. It starts with a set of rules and then prunes or refines them to reduce prediction errors [50].

Table 9. Comparative analysis of the advantages and disadvantages, supported data types, and applications of rule-based algorithms in industrial maintenance.

Algorithm	Advantages	Disadvantages	Supported Data Types	Applications in Industrial Maintenance	Mentions Number
Expert systems (ES)	- Captures domain-specific knowledge and expertise; - Provides explanations for decision-making.	- Highly dependent on the accuracy of rules and knowledge base; - Can be complex to build and maintain.	Text, numeric, categorical.	- Fault diagnosis [12,294,483].	3
Repeated incremental pruning to produce error reduction (RIPPER)	- Offers interpretability and readability due to rule-based nature; - Incrementally refines rules for better error reduction.	- May suffer from overfitting with noisy or imbalanced data; - May require parameter tuning for optimal performance.	Numeric, categorical.	- Failure classification [50].	1

shows the advantages, disadvantages, supported data types, applications in industrial maintenance, and the number of mentions, with the corresponding references, of each RB algorithm in the context of maintenance-related tasks.

Each type of DM technique has its strengths and is suited to different types of data, tasks, and expertise availability. The choice of algorithm depends on the specific problem, dataset characteristics, and the desired outcomes.

6. Discussion

The industrial sector faces numerous challenges to maintain operational efficiency and extend lifecycle of industrial assets. Traditional maintenance approaches are mainly comprised of corrective strategies, involving costly downtime and suboptimal asset performance. In recent years, the adoption of DM techniques has emerged as a promising solution to address these challenges. By leveraging the power of data analytics, DM algorithms can find hidden patterns in operational data, predict potential failures, and optimize maintenance programs to avoid costly downtime.

As presented in Figure 8, there has been a noticeable increase in the diversity and application of different DM algorithms since 2013, indicating a growing and persistent interest in leveraging learning techniques within the industrial assets domain. This is as a result of improvements in computing power and data accessibility as well as a better understanding of the potential benefits these techniques offer in predictive maintenance, optimization, and overall asset management.

This discussion is an immersion into the various dimensions of industrial assets maintenance and highlights the role of DM in optimizing maintenance in the industrial field.

6.1. Asset Characteristics

The effectiveness of DM algorithms in the maintenance of industrial assets is highly dependent on the quality and characteristics of the data used for analysis.

As presented in Section 4.1, time series data from sensors provide valuable information about the health of assets, text data provide valuable information on maintenance actions, event data give punctual information about assets, systems, or processes, and image data from assets allow fault region extraction or condition classification. Hence, multi-modal data fusion enables a comprehensive understanding of assets performance and health.

The results of the reviewed studies, presented in Section 4.2, show the current landscape of DM applications in industrial assets maintenance, particularly with a focus on mechanic and electromechanic assets. Most research efforts are directed toward these types of assets, with vibration analysis and mechanical power being the predominant parameters used for analysis. This emphasis on rotating machinery reflects the critical role such assets play in industrial operations and the potential impact of failures on overall productivity and profitability. However, the relatively limited attention given to other types of assets presents a significant opportunity for further exploration and research. Assets related to automation, encompassing electric, electronic, and computing components or systems, are becoming increasingly prevalent in modern industrial settings, offering rich potential for investigation, and could generate valuable insights into optimizing maintenance practices for these emerging technologies.

The information presented in Section 4.4 indicates that the complexity of automation systems is increasing, which requires advanced predictive maintenance strategies and sophisticated monitoring techniques to preemptively address potential failures on industrial assets. By prioritizing these assets, organizations can leverage their automation infrastructure to achieve greater operational excellence and competitive advantage. The study of complex assets, such as integrated systems comprising multiple interconnected components, presents significant challenges but is essential due to the high degree of control system integration. These systems often involve various fault modes that can impact their overall performance, making it crucial to adopt comprehensive analysis and maintenance strategies. The integration of multiple subsystems and control mechanisms increases the complexity of diagnosing and addressing faults, because issues in one component can propagate or interact with other parts of the system, making it more challenging to identify and resolve problems. Effective management of such complex systems involves developing robust models that account for the interactions and dependencies between components. This may include using digital twins, system simulations, and machine learning algorithms to predict and mitigate potential failures. By focusing on these complexities, industries can enhance their ability to maintain high operational standards and minimize the risk of disruptions in critical processes or assets.

Finally, the considerable lack of faulty data of industrial assets is another problem to face, with this scarcity being due to the high costs that it would represent for industries or the regulatory requirements that would need to be complied with. Here, the use of real-world data to feed digital twins is the path to achieving better model development. The more sensors that feed the digital twin, the more data for the learning task of models to infer key features for early detection of degradation of assets or systems.

6.2. Analytical Techniques

DM provides a wide range of analytical techniques to address a variety of maintenance challenges on industrial assets, including fault detection, degradation, failure prognosis, health index forecasting, feature extraction, transfer learning across different domains, and maintenance plan optimization.

The distribution of research studies across these various analytical approaches reveals distinct patterns in the application of shallow machine learning (SML), deep learning (DL), model-based (MB), and rule-based (RB) techniques. As presented in Figure 6, SML is prominently used for fault detection and forecasting, showcasing its effectiveness in identifying and predicting operational issues. DL also excels in fault detection and is particularly strong in pre-processing, highlighting its capability in managing complex data tasks. MB approaches are mainly employed for forecasting and degradation analysis, while RB techniques are rarely used, with minimal application in fault detection. This pattern demonstrates a heavy reliance on SML and DL for key predictive and diagnostic functions in industrial assets maintenance, with MB techniques being more specialized and RB approaches remaining largely underutilized. The presented analysis emphasizes the need to choose the appropriate technique based on the specific maintenance objective, with SML and DL offering broad versatility for various applications, while other methods serve more specific roles. SML algorithms currently dominate the industrial maintenance landscape. These algorithms, while effective in many cases, often require manual feature engineering to achieve optimal performance.

SML techniques are often more interpretable, allowing experts to understand the factors that contribute to predictions or classifications. However, SML techniques struggle with complex patterns or unstructured data types such as images, text, or high-dimensional sensor data. While they are valuable tools for many industrial maintenance tasks, their performance might decrease when dealing with intricate, nonlinear relationships or when faced with datasets that demand a higher level of abstraction and representation.

DL techniques excel in scenarios with vast amounts of data and complex, nonlinear relationships, allowing them to capture intricate dependencies and nuances that might be difficult to achieve for SML approaches. However, their remarkable performance comes with high computational demands, requiring substantial computational resources for training and fine-tuning the models. Furthermore, the inherent complexity and black-box nature of DL architectures can make their decision processes less interpretable compared to SML techniques. However, in industrial maintenance, DL is a powerful tool for extracting insights from complex datasets and pushing the boundaries of predictive accuracy and analysis.

Regarding industrial assets complexity, the analysis presented in Figure 7 suggests that while SML remains the most versatile across different asset complexities, DL is emerging as a strong contender, particularly in more complex scenarios.

The research studies emphasizes how decision-making processes have experienced a fundamental transformation as a result of the application of DM techniques in industrial domains, which have become increasingly data-driven. Decision-making is now based on thorough data analysis, rather than just intuition or experience, as more and more industries incorporate sophisticated machine learning algorithms into their operational and maintenance processes. This change makes it possible for management and stakeholders to make more informed and unbiased decisions by utilizing DM models for predictive analytics, pattern recognition, and anomaly detection capabilities.

6.3. Learning Categorization

Different machine learning (ML) algorithms and learning categories offer unique advantages in industrial maintenance applications.

Section 5.3 shows that shallow machine learning (SML) is primarily used in supervised learning, emphasizing its strength in working with structured, labeled data where clear input–output relationships exist. However, its application in unsupervised and semi-supervised learning is limited, and it does not extend to reinforcement learning, reflecting its restricted use in scenarios with unlabeled data or dynamic decision-making environments. In contrast, deep learning (DL) demonstrates a more versatile application across learning paradigms, effectively handling both labeled and unlabeled data. The balanced DL use of supervised and unsupervised learning, along with its emerging role in semi-supervised and reinforcement learning, demonstrates its ability to adapt to complex and evolving data scenarios. This shift highlights the growing prominence of DL in ML, offering advanced capabilities for handling unstructured data and dynamic environments, while SML remains focused on its strengths in supervised tasks. For example, reinforcement learning techniques can optimize maintenance decision-making by learning from past experience and feedback from the environment. Semi-supervised learning is useful for automatically labeling data. Similarly, domain adaptation methods enable DL models to adapt to new environments and generalize across different datasets, improving their performance and scalability in real-world applications.

7. Trending Topics of Machine Learning and Emerging Technologies in Industrial Maintenance

This section summarizes the trending topics of machine learning (ML) application in domains of industrial assets maintenance, focusing on how they are used and can improve the decision-making of engineers and technicians.

The content of

Table 10. Trending topics of machine learning and emerging technologies in industrial maintenance.

Topic	Description	Applications in Industrial Maintenance
Virtual sensor	Virtual sensors are computational models or algorithms that estimate variables indirectly using available data and system knowledge. They offer flexibility, cost-effectiveness, and can measure variables that are difficult or impractical to capture directly. They integrate diverse data sources such as process models, historical data, and real-time measurements to provide accurate estimates, adapt to changing conditions, and identify faults or anomalies [3,4,195,271].	- Used for damage detection, industrial robot interaction, fault detection, and digital twin applications [4]; - Employed in scenarios such as a bicomponent mixing machine for wind generator blades to predict sensor measurements and replace erroneous readings, improving accuracy and reliability in maintenance processes [195]; - Applied to a variable refrigerant flow system as fault indicators for diagnosing multiple-simultaneous faults, estimating fault levels for each fault, such as improper refrigerant charges and condenser fouling [271].
Digital twin	Digital twins are virtual models that simulate physical assets, processes, and systems in real-time. They are continuously updated to reflect the current state of their physical counterparts, enabling precise monitoring and advanced fault detection. This approach not only facilitates the collection of extensive data on failure scenarios but also supports the development of predictive maintenance (PdM) strategies [4,7,59,95].	- Enables the leverage of data from corporate assets to optimize various facets of operations, including production, maintenance, and inventory tracking [4]; - Enhances maintenance by providing real-time data and predictive analytics to detect pre-failure events [95]; - Digital twins are used for extensive run-to-failure data collection, critical for effective fault detection and prediction, thereby optimizing production and reducing downtime [7].
Explainable Artificial Intelligence (XAI)	XAI enhances deep learning models by providing explanations for their output predictions, which is essential for transparency and interpretability [7,9,11,398,459].	- Allows end-users to understand decisions, recommendations, and actions produced by AI systems, building trust, especially in predictive maintenance applications such as estimating the remaining useful life of machinery [9]; - XAI facilitates understanding of deep learning models by implementing methods that provide insights into model behavior and facilitate understanding through data visualization [7].
Augmented reality (AR)	AR overlays digital information onto the physical world, providing real-time guidance and visual instructions, enhancing efficiency and accuracy in maintenance operations, reducing downtime and minimizing human error [532]. AR facilitates collaboration and interaction between humans and digital production systems [127].	- Improves communication during product design and development, identifying design errors early and reducing the need for physical prototypes, thereby saving time and costs for enterprises [533]; - Enhances maintenance tasks by providing technicians with real-time access to manuals, reducing errors and downtime. Additionally, facilitates remote expert support for complex repairs and enables better decision-making and task optimization by integrating real-time virtual information into the work environment [534]; - The objective is to reduce production times and costs, and improve quality by addressing issues such as limited information availability, inadequate training, and poor communication. AR can simplify tasks, optimize decisions, and enhance the efficiency of workers by integrating real-time virtual information into their environment. For example, AR can display sensor data from industrial equipment directly in the technician’s field of view, enabling immediate action based on real-time data [535];
Natural language processing (NLP)	NLP enables the extraction and analysis of insights from unstructured text data like maintenance logs, reports, and technician notes, predicting failures and suggesting preventive measures [128,397,536,537,538,539].	- A text-based intelligent assistance system (TbIAS) is presented by [397], which automates the selection of appropriate data mining methods based on natural language problem descriptions; - NLP methods enhance failure diagnosis accuracy, decision-making capabilities, and reduce maintenance time by automating the analysis of textual data and integrating it with data-driven maintenance strategies [536]; - Large language models (LLMs) have emerged as powerful tools for a wide range of tasks, including NLP, machine translation, and question-answering. These models leverage deep learning techniques, to learn and understand the complex patterns and structures present in language data. LLMs can process vast amounts of data, and capture semantic relationships between words and phrases. These models can also process visual, audio, audiovisual, as well as multi-modal data [538]; - Combining association rule mining (ARM) with pre-trained LLMs in the context of maintenance requests has been explored to enhance predictive maintenance (PdM) capabilities. Maintenance requests, which are textual descriptions of damages or upcoming work, often contain relevant information for PdM in a facility management context [540].
Security and safety challenge	As technology becomes more integrated, cybersecurity issues become crucial, particularly in industrial environments where data collected from sensors must be secured, and the storage and processing by production systems must be guaranteed. Cybersecurity measures include authentication, cryptographic verification, and security protocols to ensure data integrity and protect against external attacks [4].	- Essential for protecting data integrity in industrial maintenance systems, especially in environments with internet of things (IoT) devices and interconnected networks. Proactive cybersecurity measures help prevent data breaches and ensure safe communication, maintaining the reliability and safety of industrial operations [4].
Standards for predictive maintenance	Established guidelines and standards that define the best practices for predictive maintenance technologies and methodologies implementation [7].	- Ensure consistency and reliability in the application of predictive maintenance strategies across different industries, facilitate the integration of PdM systems into existing processes, including system design, fault diagnosis, and prognostic workflow. Also, support interoperability between various technologies and platforms, leading to better maintenance management practices and facilitating the widespread adoption of innovative technologies in the field [7].

presents an overview of the application of various advanced technologies for the maintenance of industrial assets.

8. Conclusions and Future Work

The reviewed studies have shown that the integration of advanced data mining (DM) techniques within the domain of industrial environments has transformed maintenance strategies of industrial assets. This survey highlights the critical role of DM techniques in enabling condition-based and predictive maintenance frameworks by utilizing extensive datasets from a wide range of assets. In addition to traditionally studied mechanic and electromechanic assets, this survey also considers automation-related assets, such as electric, electronic, and computing components or systems, as well as assets of varying complexity, providing a detailed categorization. The data-driven approaches analyzed and detailed in the studies facilitates a paradigm shift from reactive to predictive maintenance, which is essential for minimizing unscheduled downtime, optimizing operational costs, and enhancing the reliability and performance of industrial assets. This survey has provided a comprehensive DM hierarchical organization, structured into two general techniques, four specific techniques, 27 methodologies and numerous algorithms. This hierarchical framework categorizes algorithms by their learning techniques while highlighting the methodological diversity within the field, illustrating how distinct strategies are applied across various DM approaches in industrial maintenance. Furthermore, the survey identifies the 32 most frequently mentioned algorithms from the 534 analyzed studies, providing valuable insights into their prevalence and application in different maintenance contexts. Among these, the support vector machine (SVM) algorithm stands out as the most frequently cited, with 136 mentions, underscoring its widespread adoption and effectiveness in addressing maintenance challenges.

Regarding machine learning (ML) techniques, it is remarkable that while shallow machine learning (SML) techniques remain widely used due to their computational efficiency and straightforward implementation, the growing adoption of deep learning (DL) techniques reflects a shift toward advanced architectures capable of processing high-dimensional data and automating feature extraction. In this context, some research highlights the importance of a balanced integration of SML and DL techniques, leveraging their complementary strengths to address the diverse challenges of modern industrial maintenance. This synergy is particularly crucial for managing the increasing complexity of contemporary industrial systems. Moreover, emerging technologies such as digital twins for real-time simulation of physical assets, augmented reality (AR) to enhance maintenance visualization and operator training, and advanced natural language processing (NLP) and large language models (LLMs) for intuitive interaction with maintenance systems and enhanced data-driven decision-making, present significant opportunities for innovation in industrial asset management. Together, these advancements promise to redefine the strategies and tools available for optimizing maintenance processes, enabling greater precision, efficiency, and resilience in industrial operations.

The presented survey highlights the diverse aspects of industrial assets and their management throughout their lifecycle, along with the set of techniques employed to optimize their maintenance. This consolidated data will serve as a foundation for guiding targeted studies, supporting advancements in the field of data-driven support within the industrial sector.

Future research should focus on developing robust ML models designed for the specific needs of industrial assets maintenance, to support efficiency, reliability, and sustainability in manufacturing and other industrial sectors. In this paper, we propose the following future research directions:

A

Integration of multi-modal data: One promising path for future research is the integration of multi-modal data [51] in industrial maintenance applications. By combining information from different sources such as time series data, images, maintenance, or operation text descriptions, as well as systems logs, researchers can develop more comprehensive models for fault detection, failure prediction, and anomaly detection. Taking a comprehensive approach to data analysis like this could provide valuable insights into the health and performance of assets, leading to the development of more precise maintenance strategies and enhancing operational efficiency;

B

Hyperparameter optimization (HPO) of machine learning algorithms: Hyperparameter optimization (HPO) is a critical area of study with potential for impactful contributions, particularly in improving the robustness and efficiency of algorithms through optimized hyperparameter configurations. The selection of hyperparameters significantly influences the performance and applicability of machine learning (ML) models, and this is especially pronounced in deep learning (DL) models [541];

C

Explainable deep learning models: As DL techniques continue to gain interest and space in industrial maintenance, there is a growing need for explainable artificial intelligence (XAI) [9,11,398,459] methods to enhance model interpretability. Research efforts should focus on developing explainable DL models that provide insights into the underlying factors driving predictions and decisions. By incorporating transparency and interpretability into DL frameworks, researchers can facilitate better understanding and acceptance of AI-driven maintenance solutions among industrial practitioners and stakeholders;

D

Advanced reinforcement learning algorithms: While reinforcement learning (RL) [7,27,77,117,144] has shown promise in optimizing maintenance processes, there is ample room for advancement in algorithmic development and optimization. Future research could explore novel RL techniques designed specifically for the challenges of industrial maintenance, such as dynamic environments, sparse failure data, and safety constraints. By leveraging the power of advanced RL algorithms, researchers can design more adaptive and efficient maintenance policies that maximize asset performance and longevity for each of the different asset types that shape industry 4.0 [4,7,10,14,93,182,536,542];

E

Edge computing for real-time analysis: With the proliferation of IoT devices [4,7,12,23,42,59,60,67,109,112,127,390,537,542] and edge computing capabilities, there is growing interest in deploying real-time analysis techniques for industrial maintenance applications. Research should focus on developing lightweight ML models that can be deployed directly on edge devices to enable real-time data analysis and decision-making. By leveraging edge computing resources, industrial sectors can minimize latency, reduce bandwidth requirements, and improve the scalability of their maintenance systems;

F

Collaborative maintenance frameworks: Collaboration and knowledge sharing among industrial stakeholders are essential for advancing maintenance practices and addressing complex challenges effectively. Future research could explore the development of collaborative maintenance frameworks that facilitate data sharing, model transferability, and collective learning across different organizations and industry sectors. By encouraging collaboration and synergy among stakeholders, researchers can accelerate innovation in industrial maintenance and drive continuous improvement in asset performance and reliability.

Based on the conclusions and identified directions for future work, the authors propose the development of a complementary article that presents a comprehensive taxonomy for data-driven maintenance of industrial assets. This taxonomy aims to provide the scientific community with a detailed and structured overview of key topics, including data mining techniques, types of data utilized, asset complexities, and asset categories. By doing so, it will contribute to a deeper understanding of the field and help identify gaps for further research.