ToeSiam: Self-supervised representation learning for detecting time-varying dynamic characteristics of high-rise buildings

Accurate identification of time-varying dynamic characteristics is essential for structural health monitoring of civil structures under changing environmental conditions. However, conventional modal identification techniques often rely on strong assumptions and become unreliable when structures exhibit time-varying behavior or are subjected to severe excitations, such as strong winds or typhoons. To address this limitation, this study proposes a data-driven framework, named ToeSiam, for detecting variations in structural dynamic properties without modal identification. The proposed method first constructs positive sample pairs from vibration signals measured under normal operating conditions. The signals are then transformed into Toeplitz matrices, which are commonly used in stochastic subspace identification methods, to encode the temporal correlation structure of the system responses. These matrices are subsequently fed into a self-supervised contrastive learning network based on SimSiam with a Vision Transformer encoder, enabling the extraction of compact and discriminative feature embeddings. After training, a reference feature distribution is established from data under normal operating conditions, and deviations of test signals from this distribution are quantified using the Mahalanobis distance as an anomaly score. The effectiveness of the proposed ToeSiam framework is validated through both numerical simulation and full-scale data. The results demonstrate that the proposed method can sensitively capture the changes in structural dynamic properties. The core innovation of the proposed framework lies in its ability to directly detect variations in structural dynamic properties based on dynamic responses, thereby avoiding the introduction of identification errors and uncertainties associated with the modal identification process. The proposed ToeSiam model aims to provide a practical alternative for monitoring structural dynamic property variations, offering a novel and effective solution for long-term structural health monitoring under complex environmental excitations. © 2026 Kang Xu and Qiusheng Li.