Early detection and diagnosis of plant anomalies using parallel simulation and knowledge engineering techniques

A conventional alarm system in a nuclear power plant surveys pressures, temperatures and similar physical quantities and triggers if they get too high or too low. To avoid false alarms for a dynamic process, the alarm limits should be rather wide. This means that a disturbance may develop quite a bit before it is detected. In order to get warnings sufficiently early to avoid taking drastic countermeasures in restoring normal plant conditions, the alarm limits should be put close to the desired operating points. As a extension of several years' activity on alarm reduction methods, the OECD Halden Reactor Project started in 1985 to develop a fault detection system based on the application of reference models for process sections. The system looks at groups of variables rather than single variables. In this way each variable within the group may have a legal value, but the group as a whole may indicate that something is wrong. Therefore faults are detected earlier than by conventional alarm systems, even in dynamic situations. Reference models for the feedwater system of a PWR nuclear power plant have been successfully evaluated with real process data. A test installation is now running on the Loviisa NPP, Finland