A robust strategy for sensor fault detection in nuclear power plants based on principal component analysis

Abstract

The principal component analysis (PCA) method has been widely used in sensor fault detection. However, outliers of training data may affect the projection directions of both principal component (PC) and residual space, thereby reducing the fault detection rate (FDR). The high sensitivity of PCA to random noise in the test data can also lead to an increase in the false alarm rate (FAR). To improve the performance of the PCA, this paper proposes a robust PCA approach for sensor fault detection in nuclear power plants (NPPs). A statistical method based on Euclidean distance is used to clean outliers in the training data pre-processing phase. Subsequently in the fault detection phase, the moving average (MA) filtering method is adopted to process Q-statistic to reduce false alarms caused by random noise in the test data. Simulation and plant signals are used to verify the effectiveness of the proposed method. Finally, comparisons with the conventional PCA, auto-associative kernel regression (AAKR) and multivariate state estimation technique (MSET) highlight the advantages of the proposed method.