Fault Detection for Steam Turbines Based on Non-Stationary and Autocorrelated Multivariate Time Series with Non-Gaussian Noise

Abstract

In recent years, fault detection for steam turbines based on condition-monitoring data has gained significant attention. However, the resulting data often exhibit non-stationarity, autocorrelation, and non-Gaussian noise, posing significant challenges for conventional monitoring methods. To address these issues simultaneously, this paper proposes a three-step fault detection scheme. First, a robust singular spectrum analysis (RSSA) procedure removes sparse non-Gaussian disturbances through low-rank and sparse decomposition of Hankel matrices. Next, stationary and dynamic kernel principal component analysis (SDKPCA) separates non-stationary common trends, incorporates temporal dependence through lag augmentation, and extracts nonlinear low-dimensional stationary features. Finally, a Hotelling's $T^{2}$ statistic is constructed from these features for online monitoring. Simulation studies and two real-world fault cases demonstrate that the proposed scheme provides a favorable balance between detection sensitivity and false-alarm control. In the two recorded cases, it achieves TPRs of 96.26% and 90.94%, with FPRs of 0% and 9.47%, respectively.