INTELLIGENT CONDITION MONITORING AND FAULT DIAGNOSIS OF ELECTRICAL POWER AND CONTROL SYSTEMS USING MACHINE LEARNING–BASED PREDICTIVE ANALYTICS

Abstract

Intelligent condition monitoring and fault diagnosis of electrical power and control systems using machine learning–based predictive analytics was quantitatively investigated to evaluate diagnostic reliability, robustness across operating regimes, and explanatory value under realistic data conditions. The study analyzed a large multi-asset dataset consisting of 211,200 time-window observations derived from electrical, thermal, mechanical, insulation-related, and control-residual measurements, including 312 documented fault events. Descriptive analysis showed strong class imbalance, with healthy operation representing 88.2% of observations, degraded states 7.1%, and fault states 4.7%, and fault behavior occurring in temporally clustered events with a median duration of 18 minutes. Correlation and collinearity analyses indicated moderate within-domain feature dependence but low cross-domain redundancy, with mean variance-based collinearity indices remaining below conservative thresholds across predictor groups, supporting multivariate modeling. Reliability analysis demonstrated acceptable to strong internal consistency across indicator domains, with internal consistency coefficients ranging from 0.76 to 0.87 and temporal stability coefficients exceeding 0.86 under stable operating regimes. Multivariate regression and hypothesis testing results showed that multi-domain predictive models explained substantially more variance in fault outcomes than baseline context-only models, with adjusted fit indices improving from approximately 0.27 to 0.59 after integration of regime-aware normalization and multi-domain features. Hypothesis testing confirmed statistically significant and practically meaningful improvements in diagnostic performance, with effect sizes exceeding 0.50 for multi-domain versus baseline comparisons. Prognostic-oriented analyses further demonstrated statistically meaningful associations between degradation indicators and time-to-event proxies, indicating systematic risk escalation prior to documented intervention events. Overall, the findings provided quantitative evidence that regime-aware, multi-domain machine learning–based predictive analytics improved fault detection reliability, reduced confounding from operating variability, and supported interpretable diagnostic inference in electrical power and control systems.