Heterogeneous signal modeling and Spectral-Peephole LSTM architecture for resilient structural health monitoring in complex engineering systems

Abstract

Structural health monitoring systems play a significant role in the early detection of nonlinear damage mechanisms caused by seismic activity and structural fatigue. Damage occurring in modern structures typically leads to complex deviations in time-series data and distortions in the fundamental frequency components that represent structural stiffness. Although conventional methods can track changes in the time domain, they fall short of simultaneously capturing the subtle spectral changes and long-term dependencies within the signal as the complexity of the damage increases. Therefore, in this study, a unique Spectral-Peephole Long Short-Term Memory (LSTM) architecture is proposed that integrates time-domain and frequency-domain features within a single cellular structure, thereby enabling the detection of structural damage at the earliest stage. The datasets used during the training and testing phases have been generated with the assistance of a large language model, following physics-informed principles derived from structural dynamics. Within the scope of the study, datasets comprising critical parameters such as acceleration, velocity, natural frequency, and modal stiffness have been created for three different complexity levels: “Easy”, “Medium”, and “Hard” based on damage severity. When creating complex datasets, environmental noise and external disturbances in the easy, medium, and hard datasets have been kept at minimum, medium, and high levels, respectively. Accordingly, data with a high Signal-to-Noise Ratio (SNR) has been obtained for the easy dataset, a medium SNR for the medium dataset, and a low SNR for the hard dataset. Experimental results show that the Standard LSTM and Peephole LSTM have achieved accuracy rates of 99.75%, 93.66%, 79.65% and 99.76%, 93.81%, 79.75% for the easy, medium, and hard datasets, respectively. In the proposed Spectral-Peephole LSTM model, accuracy rates of 99.79%, 93.83%, and 79.86% could be achieved for the easy, medium, and hard datasets, respectively. Across all three complexity tiers the Spectral-Peephole LSTM attained higher accuracy than both the Standard and the Peephole LSTM, with the largest gain observed on the hard (low-SNR) dataset. The margins are numerically small (0.04–0.21% points in tier-averaged accuracy) and the individual folds overlap, but the ordering of the three architectures is preserved at every complexity level, and the pooled differences are statistically significant. The evaluation is confined to synthetic data generated according to physics-informed principles, and the reported accuracies characterize the architecture on this benchmark.