NEURAL-NETWORK-BASED FAULT DETECTIONS IN A PRESSURIZED-WATER REACTOR PRESSURIZER

The early detection of incipient failures is of paramount importance f or the safety and reliability of nuclear power plants. The feasibility of using artificial neural networks as process simulators in a fault detection device is explored. Two neural networks are trained to follo w the dynamic evolution of the system pressure in a nonfaulty pressuri zer of a pressurized water reactor. During an accident, the discrepanc y between the plant's signals and the neural networks' predictions can be used to rapidly detect the faulty condition. In reality, the signa ls will be unavoidably affected by a certain level of noise. The robus tness of neural networks to noisy patterns assures a satisfactory degr ee of accuracy in the process predictions and, therefore, a high effic iency in the detection as well.