Underground Fault Detection and Classification: An Explainable, SCADA-Ready Pipeline with Adaptive DWT and Compact CNN-Transformer

Abstract

Underground faults are difficult to detect and classify reliably because transient signatures are short-lived, dispersive, and often masked by operating noise. Deployment in supervisory control and data acquisition (SCADA) environments demands deterministic, sub-millisecond latency and transparent decision logic. This research presents an explainable, SCADA-ready pipeline that couples adaptive discrete wavelet transform (ADWT) feature extraction with a compact convolutional neural network-Transformer (CNN-Transformer) classifier. The ADWT adaptively selects the mother wavelet and decomposition depth using an energy-entropy criterion to obtain a concentrated representation of fault-transient energy, with db4 identified as the optimal wavelet for the evaluated dataset. Post-hoc explanations are produced using SHapley Additive exPlanations (SHAP), yielding phase- and sub-band-level attributions that align with expected fault physics (phase-to-ground impulses and inter-phase couplings). On a twelve-class test set generated from a controlled MATLAB/Simulink model of the 200-km underground-cable system, the hybrid CNN-Transformer attains the highest deep-model accuracy with a strong macro-precision and recall. The CNN-Transformer achieves a model-level CPU inference latency of 0.036 ms per sample, indicating computational feasibility for protection-adjacent and edge-oriented monitoring. The combined classification, latency, and SHAP results demonstrate an accuracy-efficiency-interpretability trade-off for underground-cable fault diagnosis.