Abstract
This paper addresses how to utilize artificial neural networks (ANNs) for detecting anomalies of nuclear power plants in operation. The basic principle of this methodology is to detect the anomaly with deviation between process signals measured from the actual plant and the corresponding output signals from the plant model, which is developed using three-layered auto-associative ANN; the auto-associativity has the advantage of detecting unknown plant conditions. A new learning technique adopted here compensates for the drawback of the conventional backpropagation algorithm, and is presented to make plant dynamic models on the ANN. The test results showed that this plant monitoring system is successful in detecting the symptoms of small anomalies in real-time over the wide power range including start-up, shut-down and steady state operations.
