Predictive maintenance in Industrial Systems Using Machine Learning : A Review

Abstract

Predictive maintenance (PdM) has been an important strategy in modern industry, especially with the use of Machine Learning (ML) techniques to enhance equipment reliability and reduce unplanned downtime. In contrast to old and traditional maintenance strategies, that were mainly relying on reactive or scheduled interventions, PdM provides a real-time defect detection and also failure prediction through a complete environment of sensor data records. Many recent studies highlight the effectiveness of ML techniques for optimizing intervention tasks. In this study, we present a systematic mapping study (SMS) of ML classification techniques in industrial contexts. A total of 166 articles in industry and manufacturing published between the year 2000 and 2024 were identified from Scopus digital Library, after a selection process. The findings emphasize an important aspect which is that the fault diagnosis subject is frequently investigated, with Random Forest (RF) being the predominant ML classifier with 64 appearances, followed by Support Vector Machine (SVM) with 55 uses. Also, recent research highlights the increasing role of Deep Learning (DL) in PdM via the use of Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTMs) with 28 and 17 appearances respectively. We have also measured the performance of the ML and DL models across the studied papers, by calculating the average performance metric for each model, thus providing a broader explanation and a clearer view on the use of each model. Although many papers did not explicitly specify the datasets used, we found that 85.2\% of the papers that have cited their dataset have used real world datasets, thus assuring practicability. As far as the metrics are concerned, Accuracy is the most dominant metric with 100 occurrences, followed by Precision with 61 uses, Recall 57 uses and F1-score 37 uses. The most used tools are Python with 107 occurrences, R with 40 and MATLAB with 20. These findings show that there is a need for publicly available datasets, as well as the development of alternative classification techniques to advance industrial AI PdM applications.