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

Detecting early signs of structural damage is vital for the safety of engineering assets, but natural environmental and operational changes often hide these subtle warning signs. To address this challenge, this study introduces Deep Autoencoder (DAE) and Attention-based Long Short-Term Memory Autoencoder (ALSTM-AE), a semi-supervised machine learning framework designed for Structural Health Monitoring (SHM). The proposed method operates by processing raw vibration signals from a structure and extracting key features that capture its natural movement. It then uses a two-part neural network: a DAE learns the fundamental patterns of the structure's normal behavior, while an ALSTM-AE tracks how those patterns change over time. The system flags potential damage by calculating an anomaly score, which measures how much the current data deviates from normal expectations, and applying a strict mathematical threshold, median absolute deviation (MAD), to make the final decision. When tested on the Z24 bridge dataset, the proposed framework significantly outperformed traditional methods across key metrics, including accuracy, precision, and recall. The results demonstrate that this new approach is highly effective at ignoring normal, load-induced variations while accurately catching real damage, providing a strong foundation for real-time monitoring, predictive maintenance, and digital twin applications.

Keywords: SHM, Vibration-based monitoring, Anomaly Detection, Semi-supervised Learning, Principal Component Analysis, Autoencoder, LSTM-DeepAE, Predictive Maintenance.