Sensor fault detection, isolation and reconstruction in nuclear power plants

Abstract

The widespread sensors in a nuclear power plant (NPP) provide vital support to its operation. Any fault in these sensors can be a severe issue, threatening the safety of the NPP. Thus, the detection and isolation of the problematic sensors and the reconstruction of their readings are essential. In this study, several algorithms for sensor fault detection, isolation, and signal reconstruction are proposed. Through simultaneously monitoring changes in the recordings from sensors of the same type, an anomaly can be detected while the preset thresholds are exceeded. No subjective reasoning is required in setting the thresholds. To prevent improper recovery actions, an interquartile range-based scheme is employed to avoid falsely linking the variations in sensor readings caused by an initiating event (IE) to those caused by sensor faults. Moreover, via a sequential backward selection-based approach, the faulty sensors can be effectively isolated without the need for checking all the installed ones when the variations are deemed to be caused by sensor faults. Lastly, multivariate autoregressive model-based approaches are introduced to reconstruct the readings of faulty sensors either when an NPP is in normal operation or during the occurrence of an IE. Results of experiments using data generated by a simulator of the Maanshan NPP in Taiwan are provided to illustrate the efficacy of the proposed approaches.