PINN-Based Online Monitoring Method for Sensor Drift

Safe and stable operation of nuclear power plants (NPPs) relies on accurate and reliable instrumentation. Conventional sensor calibration in NPPs is performed manually at fixed intervals, which can overlook abnormal behavior between outages and increase maintenance costs due to unnecessary recalibration of healthy sensors. Online monitoring (OLM) addresses these limitations by assessing sensor health in real time during plant operation and enabling condition-based maintenance. Despite extensive researches on artificial intelligence-based OLM, many approaches require large training datasets and operate as black boxes, offering limited physical interpretability. This work proposes a sensor drift detection method based on a physics-informed machine learning that combines a long short-term memory (LSTM)-based autoencoder with a physics-consistency constraint derived from a governing physical equation. By embedding the physics constraint into the loss function, the proposed model promotes physically plausible reconstructions even with limited data, thereby improving drift-detection performance and enhancing interpretability. The suggested method can provide a stable, physically consistent framework for drift monitoring and offer a practical route to reliable and explainable sensor health monitoring in NPP instrumentation.

Keywords: On-Line Monitoring, Physics-Informed Autoencoder, Signal Validation, Sensor Drift.