A novel unsupervised deep learning approach for vibration-based damage diagnosis using a multi-head self-attention LSTM autoencoder

•End-to-end unsupervised deep learning algorithm for structural damage diagnosis.•Proposed model excels in comparative evaluation against other unsupervised models.•Multi-head self-attention feature enhances model’s diagnostic capabilities.•Incorporation of Gaussian noise layer boosts model’s performance.•Methodology is adaptable across diverse structures without re-tuning. While vibration-based structural health monitoring (SHM) has seen advances with unsupervised deep learning, limitations remain in localizing and quantifying structural damage from raw ambient vibration signals. Moreover, attention mechanisms can help identify salient patterns in acceleration response data by capturing temporal relationships. However, integrating attention mechanisms into unsupervised deep learning for vibration-based damage diagnosis is not widely explored, with no comparative evaluations against standard unsupervised methods. To address these gaps, this study develops a multi-head self-attention long short-term memory autoencoder (MA-LSTM-AE). The proposed algorithm identifies damage by comparing reconstruction error disparities from the trained undamaged state against unknown structural conditions. Through comparative evaluations on two laboratory structures and a full-scale bridge against an LSTM autoencoder (LSTM-AE) and a basic autoencoder (AE), the MA-LSTM-AE demonstrates superior damage diagnosis performance in detecting minor damage from loosened joint bolts and identifying multiple damage locations on the bridge structure. It accurately identifies, localizes, and quantifies various damage scenarios using ambient vibration data. Results provide evidence of multi-head self-attention’s potential to enhance unsupervised structural damage diagnosis.