An attention-enhanced residual network framework for structural damage detection based on structural health monitoring data

Abstract

Local damage in steel structures often produces only subtle differences in vibration responses, and measurement noise can further reduce identification robustness. To address these challenges, this study proposes a structural damage identification method that combines Markov transition field (MTF) encoding with a coordinate-attention-enhanced residual network. First, vibration responses are discretized into a finite number of states using approximate equal-frequency binning and then converted into MTF images. Coordinate attention (CA) is subsequently embedded into residual blocks to model channel dependencies and directional positional information. The method was validated using experimental data from the Qatar University Grandstand Simulator (QUGS) and the IASC–ASCE benchmark structure. Comparative experiments were conducted across image-encoding methods, attention mechanisms, noise levels, and baseline models. The results show that MTF generally outperforms the Gramian angular difference field (GADF), recurrence plot (RP), and short-time Fourier transform (STFT). Under noise-free conditions, ResNet-50-CA achieves an accuracy of 94.5% and also demonstrates strong robustness under noisy conditions. Through one-dimensional coordinate pooling, ResNet-50-CA jointly calibrates spatial and channel information. Feature visualization further shows that CA suppresses irrelevant redundancy in MTF images through coordinate-aware spatial modeling, producing highly compact and nearly non-overlapping feature clusters with the best discriminability.