Intelligent discharge state detection in micro-EDM process with cost-effective radio frequency (RF) radiation: Integrating machine learning and interpretable AI

Machining state monitoring is an enabling technology that can ensure the smoothness of production efficiency and quality by indirectly understanding various discharge patterns and anomalous events. The predominant method for identifying discharge pulses in micro electrical discharge machining (micro-EDM) relies on electrical sensing near the machining zone, which may interfere with the process. To provide a complementary approach further from the machining zone and independent from machine wiring, this paper proposes a discharge state detection scheme for micro-EDM using cost-effective and non-invasive radio frequency (RF) signals during sparking. A threshold-based method is first employed for preliminary pulse classification, followed by manual correction to ensure data reliability. Based on RF signal characteristics, domain knowledge is incorporated to extract four feature categories as inputs for machine learning (ML) models. Tree-based ensemble learning methods are applied to predict discharge pulses, with XGBoost attaining the highest accuracy. SHapley Additive exPlanations analysis reveals that distribution features of RF signals contribute most to pulse recognition, while statistical features have a smaller impact. Furthermore, ML-based classifications are compared with deep learning models. The designed hybrid architecture combining temporal convolutional networks (TCN) and gated recurrent units, can improve the identification accuracy across all pulse types, with a macro F1-score larger than 94 %. Gradient-weighted class activation mapping shows that TCN modules emphasize the significance of information on dielectric breakdown and discharge maintenance stages. Despite potential data loss from reduced sampling rates, key signal characteristics remain intact, with prediction accuracy degradation below 1 %. This study demonstrates the feasibility of RF-based discharge state detection, offering an alternative to electrical sensing and providing insights for improving servo control and machining efficiency in future EDM processes.