Early Warning Method for Gas Turbine Bearing Vibration State Based on TCN-Transformer Prediction and PCA Statistical Monitoring

Abstract

An early-warning method combining TCN-Transformer prediction and PCA statistical monitoring is proposed for gas turbine bearings with weak early-fault features and noise interference. Multi-point vibration signals are denoised using median filtering and Savitzky-Golay smoothing. A TCN-Transformer model predicts normal vibration states, and PCA indicators are constructed from healthy prediction residuals. The 95% control limits of T2 and SPE are obtained, and a simulated degradation sample is used to calibrate rolling exceedance-proportion thresholds. Validation results show that the proposed model achieves higher prediction accuracy than comparison models. In the fault sample, T2 is more sensitive to early degradation and triggers an alarm 428 sampling points before fault confirmation, corresponding to about 7.1 h. The method provides a reference for early warning and maintenance decision-making of gas turbine bearings.