Abstract Principal component analysis (PCA) is applied for fault detection, identification and reconstruction of sensors in a nuclear power plant (NPP) in this paper. Various methods are combined with PCA method to optimize the model performance. During data preparing, singular points and random fluctuations in the raw data are preprocessed with different methods. During model developing, several criteria are proposed to select the modeling parameters for a PCA model. During fault detecting and identifying, a statistics-based method is applied to reduce the false alarms of T 2 and Q statistics, and abnormal behavior is analyzed in principal and residual spaces simultaneously to locate the faulty sensor. During data reconstructing, reconstruction effects are evaluated between different methods. Finally, sensor measurements from a real NPP are acquired to evaluate the optimized PCA method. Simulations with normal measurements show that false alarms are greatly reduced, that is, the accuracy and reliability of the PCA model are greatly improved with data preprocessing and false alarm reducing methods. Meanwhile simulations with drift measurements show that the optimized PCA model is fully capable of detecting, identifying and reconstructing the faulty sensors no matter with small or major failures. Highlights Various techniques are combined with PCA method for sensor condition monitoring. Multiple methods are used to preprocess the sensor measurements from a real NPP. Various modeling parameter selection criteria are proposed to develop a PCA model. A statistics-based method is applied to reduce the false alarms of T 2 and Q statistics. The data reconstruction effects of faulty sensors are compared with various methods.
DOI 인용 스타일