A study on the applications of expert systems and neural networks for the development of operator support systems in nuclear power plants

In order to assist operators in effectively maintaining plant safety and to enhance plant availability, the need to develop operator support systems is growing to increase. The application of both expert system and neural network technologies to the operator support has the potential to increase the performance of these systems. A prototype integrated operator support system, called NSSS-DS, has been developed for multiple alarm processing, plant trip diagnosis, and the failure diagnosis of three main systems (a rod control system, reactor coolant pumps (RCPs) and a pressurizer) in the primary side of the Kori-2 nuclear power plant. This system diagnoses system malfunction quickly and offers appropriate guidance to operators. The system uses rule-based deduction with certainty factor operation. Diagnosis is performed using an establish-refine inference strategy. This strategy is to match a set of symptoms with a specific malfunction hypothesis in a predetermined structure of possible hypotheses. The diagnostic symptoms include alarms, indication lamps, parameter values and valve lineup that can be acquired at a main control room. The overall plant-wide diagnosis is performed at the main control part which can process multiple alarms and diagnose possible failure modes and failed systems in the plant. The method of alarm processing is the object-oriented approach in which each alarm can be represented as an active data element, an object. The alarm processing is performed using alarm processing meta rules and alarm processing frames. Also, the diagnosis of a plant trip can be performed at the main control part. The specific diagnosis of the three main systems can be performed followed by the diagnostic results of the main control part. The system also provides follow-up treatments to the operators. The application to these systems is described from the point of view of diagnostic strategies. For the applications of the neural network technology, two feasibility studies on pattern recognition problems in nuclear power plants have been performed. One is transient identification and the other is multiple alarm processing and diagnosis. The backpropagation network (BPN) training algorithm is applied to both of the two studies. The general mapping capability of the neural networks enables to identify a transient or an alarm's fault well. The transient identification is performed by mapping or associating patterns of symptom input vectors to patterns representing transient conditions. In the implemented network, an input layer consists of 24 sensor input nodes, a hidden layer consists of 17 nodes, and an output layer consists of 14-class transient identification nodes. The input values are various parameter trends (increasing/unchanging/decreasing), valve states (open/closed), pump states (operating/stop) and alarm data (on/off). The multiple alarm processing and diagnosis is performed by training multiple alarm patterns for the diagnosis of faults in the RCP system. In the implemented network, an input layer consists of 9-class fault identification nodes. A number of case studies are performed with emphasis on the applicability of the neural networks to the pattern recognition problems. Based on the case studies, it is revealed that the BPN training algorithm can identify the transient or the alarm's fault well, although untrained, incomplete, sensor-failed, or time-varying symptoms are given. Also, multiple transients or alarm's faults are easily identified with a given symptom input vector