Vibration-Based OLTC Mechanical Condition Monitoring with CatBoost Algorithm

The on-load tap-changing switch (OLTC) is essential for voltage stability, and its reliability directly impacts transformer and grid safety. However, long-term operation often causes mechanical and electrical failures, while traditional no-load diagnostic methods struggle to reflect real conditions, limiting effectiveness. This study focuses on the M-type OLTC, establishing a load simulation platform to introduce five typical mechanical faults. Sound, vibration, and motor current signals were collected for multi-source fault diagnosis. For sound signal analysis, LPCC, MFCC, GFCC and their derivatives were used to build composite feature vectors, combined with XGBoost, LightGBM, and CatBoost classifiers. The LightGBM model optimized by the Grey Wolf Optimization Algorithm (GWO) achieved 97.9% accuracy. For vibration signals, waveform denoising, CMAE feature selection, multi-domain features, and texture analysis formed a 29-dimensional feature vector. The CatBoost model, optimized by the Sparrow Search Algorithm (SSA), reached 95.8% accuracy. Results demonstrate that the proposed platform and multi-feature fusion method significantly enhance accuracy and robustness of OLTC fault diagnosis, offering a new approach for transformer condition monitoring and fault warning.