Condition Assessment of Circuit Breaker Mechanical Performance During Opening and Closing Operations Using Adaptive Machine Learning

Abstract

The mechanical performance of circuit breakers during opening and closing operations directly affects the safety and stability of power systems. However, conventional threshold-based methods cannot adequately represent nonlinear relationships among multiple mechanical parameters and have limited adaptability to equipment differences and changing operating conditions. This study proposes an adaptive machine learning method for circuit breaker mechanical condition assessment. Opening and closing times, operating velocities, contact travel, overtravel, three-phase asynchronism, and the cumulative number of operations are processed through data cleaning, normalization, and deviation calculation to construct a mechanical performance feature set. Circuit breaker conditions are classified into normal, attention, abnormal, and severe levels. An adaptive XGBoost model combining data-driven and model-driven feature weights is then developed. Bayesian optimization determines key model parameters, while a recent-sample-weighted updating mechanism improves adaptability to data-distribution changes. The model generates a condition level and comprehensive health score, and SHAP analysis explains individual feature contributions. The proposed model achieves an Accuracy of 0.950, an F1-score of 0.944, and an AUC of 0.983, outperforming support vector machine, random forest, and conventional XGBoost. Under a 40% distribution shift, adaptive updating increases the F1-score from 0.796 to 0.891. Opening time, closing velocity, contact overtravel, and opening velocity are identified as the most influential indicators. The method can identify gradual deterioration and provide interpretable evidence for condition-based maintenance.