Geometrical Feature Classification in Electrical Discharge Machining Using In-Process Monitoring and Machine Learning

Maintaining process stability in electrical discharge machining (EDM) during the fabrication of geometrically complex components remains challenging due to angular dependencies that modulate interelectrode gap conditions and discharge dynamics. This study demonstrates that the use of in-process discharge pulse analysis enables geometric feature classification by decoding angular geometry-mediated debris evacuation dynamics. Experimental investigations across two-dimensional connection angles (0°–180°) revealed systematic variations in pulse-type distributions: open circuits predominated at 0° (82.47%) under reciprocating motion-induced gap instability, while short-circuit incidence peaked at 90° (4.68%) due to dielectric flow stagnation. Acute angles (60°) exhibited elevated normal discharge frequencies (18.01%) linked to debris reattachment cycles, whereas obtuse angles (180°) demonstrated enhanced process efficiency through stabilized flushing regimes. These trends are attributed to angular geometry directly governing debris evacuation pathways, plasma channel stability, and gap state transitions. To isolate these effects, machine learning models were developed to capture the underlying process dynamics. ResNet15 achieved superior classification accuracy (95%, 10-fold cross-validation repeated 10 times) by hierarchically modeling multi-temporal discharge correlations, outperforming conventional convolutional neural networks (constrained by spatial invariance limitations) and long short-term memory networks (ineffective for abrupt state-driven transitions). The results establish that geometric features impose deterministic signatures on discharge pulse patterns, which can be decoded through temporal feature aggregation in machine learning architectures. This capability enables the disentanglement of geometry-dependent effects from stochastic process variability. By focusing on essential dynamics tied to geometric constraints rather than machine-specific conditions, the methodology supports extrapolation to diverse operational parameters.