Data-driven structural transition detection using vibration monitoring and LSTM networks

Structural health monitoring (SHM) is essential for ensuring the safety and durability of civil infrastructure. Traditional SHM approaches, based on manual inspections or threshold-based analyses, often fail to detect early or subtle structural changes. In this work, we propose a data-driven framework for detecting structural regime transitions using long short-term memory (LSTM) networks trained on power spectral density data. This method does not require prior knowledge of the excitation sources or structural dynamics, enabling robust and interpretable transition detection under real-world conditions. A key component of the framework is the empirical transition point, $T_a$, computed from prediction probabilities through persistence thresholding and entropy filtering. This allows for precise and automated detection of regime shifts, even in the absence of explicit ground truth. The model is validated on vibration data collected under ambient and train-induced excitations, achieving high accuracy in distinguishing pre- and post-retrofitting states. It demonstrates strong robustness across a range of operational and dynamic conditions. To enhance interpretability, we introduce the confidence variability index (CVI), which quantifies the temporal stability of the model's predictions and serves as an indicator of transition consistency. While the framework does not currently identify the physical causes of transitions, its sensitivity to dynamic changes makes it a valuable early-warning tool in SHM. Despite the inferential nature of transition detection due to the lack of ground truth, the approach offers a scalable, interpretable, and real-time solution for structural regime monitoring—contributing to the advancement of SHM systems through uncertainty quantification, causal inference, and intelligent infrastructure management.