Predictive Maintenance Using Internet of Things Sensors and Machine Learning: An Applied Study of Small Manufacturing Enterprises

Abstract

Predictive maintenance offers small manufacturing enterprises a data-driven alternative to corrective and time-based maintenance, but adoption is constrained by limited technical, financial, and data resources. In this study, the authors assessed if machine-condition variables that could be accessed using IoT could be used to support practical equipment-failure prediction. The AI4I 2020 Predictive Maintenance Dataset with 10,000 observations was analysed for the product type, air temperature, process temperature, rotational speed, torque, and tool wear as predictors. The 80:20 split of the data was made and class-weight was adjusted for each logistic regression, decision tree and random forest classifiers, and three-fold stratified cross validation was also performed to examine the stability of the random forest model. The precision, recall, F1-scores, ROC-AUC, PR-AUC, balanced accuracy, Matthews correlation coefficient and confusion matrix were used to evaluate performance. Random forest had the highest values for precision (0.780), F1-score (0.661), ROC-AUC (0.962), and PR-AUC (0.753). The rotational speed, torque, and tool wear were most influential predictors, while failure recall was moderate and decreased when validated under order. The results show that with human supervision, threshold tuning and periodic retraining, the machine-condition variables and machine learning enabled by IoT could be used to assist the maintenance prioritization in SMEs. Before it is used in operations, more validation using real-time SME data is needed.