A Threshold-Free and Label-Free Pipeline for Adaptive Pulse Classification in Electrical Discharge Machining

The surface quality and process accuracy of Electrical Discharge Machining (EDM) are critical performance metrics, which rely on accurate estimations of tool wear rate (TWR) and material removal rate (MRR). These estimations, in turn, necessitate effective classification of electrical discharge pulses. However, the hybrid nature of modern EDM power supplies—comprising resistor-capacitor (RC), inductor-capacitor (LC), and transistor modules—renders traditional threshold-based pulse classification methods highly machine-specific, limiting their applicability. Moreover, the manual labeling of pulses is both time-consuming and resource-intensive. This study proposes a threshold-free and label-free data-driven pipeline for pulse classification in EDM. Dynamic Time Warping (DTW) is utilized to estimate the similarity among pulses, independent of their duration, peak voltage, and predefined thresholds for time or voltage. Based on the constructed similarity matrix, an unsupervised learning model is developed to autonomously cluster pulses. This framework provides an adaptive method for pulse classification, eliminating the need for predefined criteria or extensive labeling, thereby enhancing the generalizability and efficiency of EDM process monitoring and control.