Wavelet-Statistic-Frame with LSTM-Attention Network for Continuous Damage Identification and Localization in Bridge Vibration Signals

To address the challenges of non-stationarity and multi-scale feature extraction in bridge vibration signals for structural health monitoring, this paper proposes a deep learning framework, termed Wavelet-Statistic-Frame with LSTM-Attention Network (WSF-LANet), that integrates multi-source feature extraction with temporal modeling to extract damagesensitive features from bridge vibration signals for identification and localization of different damage states.The model architecture is designed with three parallel feature extraction pathways: Discrete Wavelet Transform (DWT) based time-frequency analysis, extract statistical descriptors for quantifying latent damage indicators, and frame-wise segmentation and extraction of spatiotemporal features.After merging the features extracted from these three paths, a multi-attention block dynamically allocates weights across feature dimensions.The Long Short-Term Memory (LSTM) network is then used to further effectively capture the temporal dependencies of the sequence.The output is the final predicted damage matrix, which contains damage identification for each channel.In order to achieve both damage identification and localization functions, we additionally use unique heat encoding to represent multi-location and multi-category labels in a unified format.Experimental results show that on a bridge dataset from Japan, the proposed method achieves an accuracy of up to 97.5% for damage classification and a macro precision of 96.82% for localization.Ablation studies further validate the effectiveness of each feature extraction path.Crossdataset evaluations also demonstrate strong generalization capability.In summary, the proposed WSF-LANet offers an efficient, accurate, and generalizable solution for intelligent damage identification and localization in bridge structural health monitoring.