Predictive Maintenance of Wire Electrical Discharge Machining Using Long Short-Term Memory Networks for Improved Process Control

Abstract

Time series forecasting and anomaly detection are becoming essential in smart manufacturing for prognostics and health management of a machine, especially where traditional methods struggle with the analysis of high frequency data. This study uses Long Short-Term Memory (LSTM) networks for detecting and predicting anomalies in wire electrical discharge machining (WEDM), specifically focusing on events like no-sparking and wire breaks. In closed-loop forecasting continuous numerical predictions are challenging in high-frequency data analysis, so a centroid-based approach was chosen. This method simplifies forecasting by using representative feature values that highlight important class differences in the data. With this closed-loop LSTM and centroid approach, the model effectively forecasts machine states up to five seconds ahead; a useful time frame for detecting critical issues such as wire breakage and no sparking before they impact operations. The results show that this method, combined with LSTM ability to capture time patterns, can handle complex, shifting conditions in WEDM. This approach could improve productivity and reduce unexpected downtime in smart manufacturing, offering a practical and efficient way to monitor and predict machine conditions.