Abstract

Sensor faults in nuclear power plants (NPPs) have the potential to propagate negative impacts on system stability, leading to false alarms and accident misdiagnosis. Existing methods seldom concurrently consider complex spatial–temporal correlations among multi-type sensors in the primary circuit. This study presents a novel sensor fault detection and isolation scheme named the knowledge-guided spatial–temporal model (KGSTM), using the knowledge-guided recurrent unit (KGRU) and the concurrent detection strategy. To organically express part and whole interdependencies from inherent sensor layout, several graphs are specifically designed with pertinent domain knowledge. KGRU consists of the multi-graph convolutional network (MGCN) for fusing various spatial information and the gate recurrent unit (GRU) for extracting dynamic temporal features, further obtaining precise reconstructed signals and residuals. The concurrent detection strategy can explicitly quantify abnormal behaviors to detect and isolate faulty sensors by characterizing spatial–temporal signal variation. Numerical results on two real-world datasets from a pressurized water reactor (PWR) with simulated faults illustrate that the KGSTM has superior performance over various state-of-the-art methods in terms of signal reconstruction and fault detection.