Channel mode attention network for structural damage identification

The main idea of structural damage identification methods based on convolutional neural networks (CNNs) is to extract damage features from vibration responses and then map these features to damage identification results. The quality of these extracted damage features directly affects damage identification accuracy. Since CNNs focus mainly on local information from vibration responses and fail to fully utilize global information, the quality of the extracted damage features cannot be guaranteed, inevitably reducing damage identification accuracy. To address this problem, this paper proposes a channel mode attention network for structural damage identification. Specifically, the channel mode attention mechanism is used to recalibrate the damage features extracted from modal data, and the global and local information of the damage features are simultaneously utilized through global average pooling and convolution operations, so that the network can extract more accurate features, ultimately improving damage identification accuracy. Numerical and experimental results show that, compared to the conventional ResNet, the proposed network improves damage identification accuracy by 8.60% and 6.34%, respectively. •CNN extracts damage features and maps them to damage identification results.•A channel mode attention network for structural damage identification is proposed.•Channel mode attention network recalibrates features by global and local information.•The proposed network obtains higher accuracy when uncertainty is considered.•The superiority of the proposed network is maintained in real structures.