Sensor fault detection and signal restoration using scenario-guided supervised autoencoders for nuclear power plant accidents

Abstract

Accurate detection and timely restoration of faulty sensor signals are critical for ensuring safety and operational integrity in nuclear power plants, especially under accident conditions characterized by highly dynamic and nonlinear parameter behaviors. In this study, we propose a scenario-guided supervised autoencoder (SAE) framework that performs three integrated functions: (1) detecting faulty sensor signals during accident transients, (2) restoring corrupted signals to their unfaulted trajectories, and (3) preserving the quality of downstream accident diagnosis. Our framework employs a supervised variational autoencoder (VAE) with a long shortterm memory encoder, trained on accident scenario simulations from a compact nuclear simulator replicating key thermal-hydraulic behaviors of a pressurized water reactor. A classification decoder provides scenario-guided supervision, enabling the model to distinguish between fault-induced signal deviations and legitimate accident transient changes. Evaluation across nine accident scenarios and seven different sensor fault types shows that the proposed SAE achieves a fault detection true positive rate of 99.09% with 95.52% precision, compared to 95.50% and 91.99% for the conventional VAE. For signal restoration, 97.70% of restored signals fall within 15% mean absolute error of their unfaulted trajectories, compared to 69.50% for the VAE. The restoration-based approach recovers accident diagnosis accuracy to levels comparable to unfaulted conditions, outperforming fault isolation strategies for most fault types. Additional robustness analyses show that the proposed framework retains detection performance under additive Gaussian measurement noise up to sigma = 0.05 (5% of the normalized signal range) and achieves a 100% fault detection rate for up to three simultaneous sensor faults. These results suggest that scenario-guided supervised autoencoders can improve sensor signal integrity in safety-critical nuclear applications.