Abstract
Fault diagnosis system usually includes a prediction module and a diagnosis module. The prediction module is supposed to predict sensor values based on sensor observations, and the diagnosis module is used to determine whether a sensor has degradation or failure by comparing the difference between sensor observations and predictions. Auto-Associative Neural Network (AANN) is a widely applied data-driven prediction method consisting of input layer, mapping layer, bottleneck layer, de-mapping layer and output layer. AANN compresses the information contained in the input data (i.e. sensor observations) into bottleneck layer through mapping layer, and then reconstructs the sensor values at output layer based on the extracted nonlinear features stored in bottleneck layer through de-mapping layer. This paper focuses on the feasibility study of AANN model in fault diagnosis for nuclear power plants. The performance of AANN model was studied by using simulation data, and the AANN model was optimized by combining prejudgment and secondary prediction. The results show that the optimized AANN model can improve the performance comparing to original AANN model, and can be applied to the fault diagnosis for nuclear power plant sensors.
