RESEARCH ON STRUCTURAL DAMAGE DETECTION METHOD BASED ON ONE-DIMENSIONAL DILATED CONVOLUTION NEURAL NETWORK

It is of great significance to use vibration response for intelligent detection and diagnosis of structural damage. However, the traditional deep convolution neural network model has some problems in processing structural vibration signal, such as large model parameters, loss of signal detail information and poor generalization performance. Therefore, a structural damage detection method based on one-dimensional dilated convolution neural network was proposed. The neural network model used in this method used dilated convolution to replace the traditional combination layer of convolution and pooling, and increased the receptive field while keeping the number of parameters unchanged; global pooling layer was used to replace the traditional full connection layer to reduce the model parameters and prevent the occurrence of over fitting, and then according to the phenomenon of class imbalance in the actual vibration signal data set, the cost sensitive classifier was trained by setting penalty weights for different types of signals, which could effectively improve the structural damage detection accuracy under unbalanced samples. The verification and application analysis of the damage experiment on the grandstand scale model of the University of Qatar showed that the method proposed in the paper could automatically extract the optimal features from the original acceleration signals and classify them without losing the detailed information of the signals, thus achieving high recognition accuracy and classification success ratio, which could be used for real-time damage detection.