Damage identification of spatial steel truss structures by the combination of CNN and BiLSTM neural networks

To address the comprehensive detection of structural damage type, location, and severity, a novel deep learning-based method is proposed. Frequency response functions are employed as model inputs to characterize structural dynamics. Convolutional Neural Network (CNN) is firstly used to extract local dynamic features, which are subsequently processed by a Bidirectional Long Short-Term Memory (BiLSTM) network to capture sequential dependencies within the frequency-domain data. A dual-output prediction framework is designed, comprising a Softmax function for classifing damage type and location, and a linear activation regression branch fro quantifying damage severity. The proposed method was validated using a laboratory-scale steel truss model. The model was trained on simulated datasets corresponding to two common damage scenarios, member damage and bolt loosening. To further evaluate the generalization capacity of the method, corrosion damage data which not included in the training process, were introduced for testing. The results indicate that the proposed method can accurately localize and quantify the damages. Moreover, Moreover, the accuracy remains above 88.9% for the unseen corrosion damage, which demonstrates its robust generalization performance.