Bayesian Conditional GAN for Unsupervised Anomaly Detection in Structural Health Monitoring Time-Series Dataset

Detecting rare structural damage without labeled fault data remains a critical unsolved challenge in structural health monitoring (SHM). Prestressed concrete catenary poles are key elements of high-speed railway infrastructure, and undetected degradation can compromise safety and service reliability. This paper introduces BcDCGAN, a Bayesian conditional deep convolutional generative adversarial network designed for unsupervised anomaly detection in multivariate vibration time series from three in-service catenary poles. Trained exclusively on healthy acceleration signals with wind-speed conditioning, the model learns the normal structural dynamics and produces an uncertainty-based anomaly score that combines reconstruction quality, adversarial evaluation, and epistemic uncertainty into a single decision function. An adaptive, data-driven threshold estimate from healthy validation data enables practical deployment without damage labels. On a real 2017 catenary pole dataset (1606 signals, 70/10/20 split) with injected, physically motivated damage-like patterns, BcDCGAN achieves high anomaly recall with interpretable uncertainty signals and clear separation between normal and anomalous latent representations. The results suggest that Bayesian conditional GANs can support risk-aware monitoring of railway infrastructure under varying environmental and operational conditions.