XAI-Enabled Quantification of Insulation Aging from Reference-Signal Response

Abstract

Reliable and auditable assessment of insulation aging in high-voltage assets is essential for condition-based maintenance and failure prevention. We present a methodology that combines controlled excitation using a broadband reference signal (MHz-range chirp) with explainable AI (XAI) to enable interpretable classification of aging levels. The system records the response and constructs a multi-channel input comprising time-domain representations, spectral features, and differential-to-reference channels for a convolutional neural network (CNN). For interpretability, we employ SHAP (SHapley Additive exPlanations), Grad-CAM (Gradient-weighted Class Activation Mapping), and Integrated Gradients (IG) to provide attribution maps indicating the time-frequency regions and specific channels that drive the predicted class. In controlled aging experiments, the explanations consistently emphasize discriminative spectral bands and support error analysis by flagging likely misclassifications. The approach increases confidence in model outputs and supports deployment in condition-based maintenance workflows.