Sensors are essential detection components in nuclear power instrumentation and control systems. Their proper functioning plays a crucial role in the operation of nuclear power systems and equipment. Sensor failures may lead to inaccurate detection data and delayed identification of nuclear system or equipment malfunctions. Moreover, both sensor and system failures can result in unstable or inaccurate sensor readings, thereby increasing the risk of nuclear accidents and threatening the safe operation of nuclear power plants. To accurately distinguish between sensor failures and nuclear system failures during abnormal nuclear plant operation, it is necessary to determine whether the sensors themselves are experiencing anomalies or failures. Based on this premise, this paper conducts research on the failure detection method of nuclear power sensors.
Firstly, using the simulation platform of the Fu Qing Nuclear Power Plant, historical operational data of sensors under both steady-state and transient conditions are collected to establish a dataset for sensor failure detection. Then, based on this dataset, the T2 and Q statistics of KPCA, along with corresponding control limits, are utilized to determine whether anomalies occur during nuclear power plant operation. If anomalies are detected, the Random Forest algorithm is employed to distinguish between sensor failures and system failures. In the case of sensor failure determination, the Q statistic contribution rates of KPCA (SPE) and the autoencoder are used separately for sensor fault localization. Finally, combining the D-S evidence theory, the decision fusion of the Q statistic contribution rates of KPCA (SPE) and the autoencoder yields the final localization result. Experimental analysis demonstrates that the proposed method effectively detects nuclear power sensor failures with high accuracy.