Semi-Supervised LSTM–Deep Autoencoder Framework for Anomaly Detection and Resilient Predictive Maintenance in Bridge Infrastructure

Early and accurate detection of structural anomalies is essential for maintaining the safety and durability of engineering assets, yet subtle damage signatures are often obscured by environmental and operational variability. This study presents a semi-supervised anomaly detection framework, termed LSTM-DeepAE, for vibration-based Structural Health Monitoring (SHM), which leverages both baseline and a limited subset of damaged data to improve sensitivity to incipient structural damage. Multichannel vibration signals are transformed into multi-domain features, including spectral modal indicators, band-limited energy, and statistical descriptors, to capture the intrinsic structural dynamics and highlight anomalies. Long Short-Term Memory (LSTM) and a deep Autoencoder (AE) are trained on predominantly baseline data, with a small proportion of damaged samples incorporated in a semi-supervised manner. The deep autoencoder learns compact latent representations of structural behaviour, while the LSTM autoencoder models temporal dynamics across sequential latent features. Anomaly scores are derived from reconstruction errors and evaluated using a robust median absolute deviation (MAD) thresholding strategy for anomaly decision-making. Validation on the Z24 bridge dataset demonstrates that the PDE framework outperforms conventional methods across AUC, precision, recall, F1-score, and accuracy, providing the effectiveness of integrating deep representation learning and strong generalisation to load-induced modal variations. These results establish a robust foundation for real-time anomaly detection, predictive maintenance, and digital twin integration in SHM applications.