As minor damage accumulates in civil engineering structures during long-term service, the importance of structural health monitoring and refined damage identification is increasingly prominent. However, existing intelligent structural sensing and control technologies suffer from problems such as reliance on human experience for feature extraction, complex vibration signal processing, and unstable damage identification accuracy. To address these issues, this paper proposes a deep learning-based structural damage detection method. This method fuses and analyzes features in the time, frequency, and wavelet domains using CNNs and RNNs, and combines this with GNNs to model structural topology information, achieving automatic classification and quantitative prediction of structural damage. In the experiments, vibration signal data generated from physical experiments and simulation models are used to construct training, validation, and test sets, and features are standardized and data augmented. The findings show that in classification tasks, the accuracy of the CNN combined with the LSTM (Long Short-Term Memory) model reaches 97.2%, and the F1 score is 96.9%, which is about 1% to 2% higher than that of a single CNN (Convolutional Neural Network) or LSTM model. In regression tasks, the model’s mean squared error in predicting stiffness loss rate is 0.07, the mean absolute error is 0.06, and the coefficient of determination reaches 0.99, demonstrating a high-precision ability to identify minor damage.
Xu et al. (Thu,) studied this question.