Remaining useful lifetime prediction for predictive maintenance in manufacturing

Traditional maintenance approaches often result in either premature replacement of machine parts or downtime in production lines due to malfunctions. Consequently, these lead to significant amount waste in material, time and, ultimately, money. In this study, a machine learning-based predictive maintenance approach is proposed to predict the Remaining Useful Life of production lines in manufacturing. Using data collected from integrated IoT sensors in a real-world factory, we attempted to address the problem of predicting potential equipment failures on assembly-lines before they occur through machine learning models in real-time. To evaluate the effectiveness of the approach, we developed several predictive models using ML algorithms, including Random Forests (RF), XGBoost (XGB), Multilayer Perceptron (MLP) and Support Vector Regression (SVR) and compared the results for all possible variations. Furthermore, the impact of noise filtering, smoothing and clustering techniques on the performance of ML models were investigated. Among all the methods evaluated, RF, an ensemble bagging method, showed the best performance, followed by XGB and implemented in production systems. The implemented prediction model achieved successful results and was able to prevent about 42 percent of actual production line failures.

Automated summary

This study proposes a machine learning-based predictive maintenance approach to predict the Remaining Useful Life of production lines in manufacturing. By using data from integrated IoT sensors in a real-world factory, the approach aims to predict potential equipment failures on assembly lines in real-time and prevent downtime, resulting in resource savings.