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). This paper introduces BcDCGAN, a Bayesian conditional deep convolutional generative adversarial network designed for unsupervised anomaly detection in multivariate vibration time series from prestressed concrete catenary poles. The architecture integrates variational Bayesian inference over generator and critic weights with temporal convolutional networks, enabling epistemic uncertainty alongside reconstruction and critic objectives. Trained exclusively on healthy acceleration signals with wind speed conditioning, the model produces a log-space Bayesian anomaly score that jointly combines normalized reconstruction error, critic evaluation, and epistemic uncertainty estimates into a single weighted decision function. An adaptive threshold is calibrated from the validation data for deployment-ready performance. Evaluation on a real 2017 catenary pole dataset (1606 signals, 70/10/20 split) with injected anomalies achieves 99.2% recall while revealing clear latent space separation and appropriate uncertainty signaling for out-of-distribution samples. Progressive posterior uncertainty reduction during training confirms robust learning of healthy structural dynamics, supporting interpretable, risk-aware decisions in safety-critical railway infrastructure.