Comparative study of machine learning techniques for remaining useful life estimation in simulated mini factory environment

Abstract

Predictive maintenance (PdM) has become one of the most effective approaches for enhancing the longevity of industrial assets and minimizing the costs associated with their failure. This study aims to design and evaluate different machine learning algorithms to determine the RUL of industrial machines based on temperature, vibration, and RPM sensors. The assessed models include Linear Regression, Random Forest Regressor, Decision Tree Regressor, K–Nearest Neighbors Regressor, and Gradient Boosting Regressor. To establish real-life industrial environment and to put our method of predictive maintenance under realistic setting, the mini factory environment with use of Arduino was created. This setup involves a DC motor with temperature, vibration, and RPM sensors to monitor its state. The DC motor is chosen as one of the industrial machinery components that must be monitored, and the sensors record the parameters of its functioning. This mini factory environment allows for controlled experiments and data collection without disrupting actual production lines. In the preprocessing of data, there are some techniques to help our model become more predictive, which we call feature engineering and hyperparameter tuning in the model. Model performance was quantified using the Mean Squared Error, Root Mean Squared Error, R-squared score, and cross-validation score. Experiments prove that to accurately predict RUL, the best ensemble learning methods applied are gradient boosting and random forests. This research benefits industries as it assists PdM in providing an accurate estimation of the remaining useful life of industrial machinery using sensor data, thus reducing the time and cost of maintenance and increasing productivity.