On-line reactor surveillance based on multivariate analysis of noise signals

A minicomputer-based system for monitoring signals from nuclear power plants to provide advanced warning of anomalous conditions has been developed and implemented for on-line applications. A characterization of the reactor state is developed from the analysis of noise signals. During an observation period when normal conditions prevail, the surveillance system utilizes these noise characterizations to form a statistical description of normal reactor behavior. At the conclusion of this observation period, the system examines each new incoming characterization to determine if it differs significantly from the normal description. The overall design of the surveillance system is discussed with the perspective of applications in nuclear power plants. The theoretical basis for the anomaly detection procedure is presented. A detailed description of the implementation of these techniques on a minicomputer is also provided. The performance of the surveillance system has been evaluated extensively. Its effectiveness was demonstrated in three on-line tests. The system proved to be sensitive to plant conditions and experienced negligible difficulty with false alarms