Robust Fault Detection with One-Class Training

Anomaly detection is a critical capability in modern industrial systems, particularly in the energy sector where early fault identification can prevent catastrophic failures, minimize downtime, and reduce maintenance costs. However, the scarcity of labeled fault data in real-world applications makes traditional supervised learning approaches infeasible. This motivates the need for methods trained using only healthy data, a paradigm known as one-class training. One-class approaches are especially relevant for deployment in safety-critical domains such as nuclear power generation, grid monitoring, and process control, where failure data is rare, diverse, and expensive to collect. This study evaluates the performance and generalization capabilities of four data-driven methods trained exclusively on healthy data. The first method uses Principal Component Analysis to reduce data dimensionality and leverages reconstruction error for anomaly scoring. The second approach applies sequence modeling via a Long Short-Term Memory forecasting model, predicting future time steps based on past behavior and flagging sequences that deviate significantly from predicted values. The third is a one-dimensional convolutional autoencoder designed to reconstruct multivariate time-series inputs, with deviations in reconstruction used to identify potential anomalies. The fourth method, termed Deep Center Encoding, employs a neural network encoder trained to map healthy data to a compact region in latent space centered around a learned centroid, with outliers identified based on distance from this center. All methods are evaluated on sensor data from a real, operating nuclear power plant and tested for their ability to detect previously unseen fault distributions. Our results highlight trade-offs in sensitivity and generalization across the approaches, with Deep Center Encoding showing promising robustness to distribution shifts. These findings reinforce the feasibility and importance of one-class training frameworks for generalizable, fault-agnostic condition monitoring in industrial environments, supporting broader efforts in reliable artificial intelligence and predictive maintenance.