Explainable AI-Driven Fault Diagnosis and Predictive Maintenance for Smart Electrical Power Systems

Abstract

Continuous and intelligent monitoring of a smart electrical power system is essential to detect faults, monitor equipment degradation, predict failure and aid timely decisions for maintenance. Most previous solutions, however, concentrate on the diagnosis of a single fault or the prediction of a condition, with the result that it cannot provide an integrated, risk-aware and explanatory maintenance support. To overcome these limitations, Recurrent Interpretable Gradient Neural Network (RIGNNet) is designed to carry out fault diagnosis, degradation forecasting, failure-risk estimation, maintenance prioritization and explainable decision making, all in one. It combines the following methodologies: Inception Time for multi-scale electrical fault feature extraction, Gated Recurrent Unit (GRU) for temporal dependency learning, Neural Basis Expansion Analysis for Interpretable Time Series Forecasting (N-BEATS) for future condition and degradation forecasting, Light Gradient Boosting Machine (LightGBM) for fault classification, severity assessment and failure-risk estimation, Risk-Based Maintenance Decision (RBMD) for maintenance priority and action selection and Local Interpretable Model-Agnostic Explanations (LIME) for feature-level interpretations. The system is able to detect real-time electrical measurements and equipment-condition indicators and detect abnormal patterns, predict deterioration, estimate failure risk potential and propose maintenance activities based on asset condition and criticality. The experimental results validate the effectiveness of the integrated approach for accurate fault diagnosis and predictive maintenance, which has an accuracy rate of 96.23%. In general, the methodology provides reliable, proactive, risk-informed and explainable assistance in the reliability and operational resilience of smart electrical power systems.