Terahertz time domain spectral classification and identification of defects in multilayer composite materials based on ResNet101 network

With the rapid development of aerospace industry, multilayer structural composites are more widely used. Debonding, bubbles, inclusions and other defects in multilayer structural composites are in urgent need of new non-destructive testing methods for detection and identification. In this paper, the terahertz time domain spectroscopy system is used to detect the polymer materials with different embedded defects. In order to improve the accuracy of defect classification and recognition, the terahertz signal is transformed into the form of wavelet time-frequency graph through Morlet wavelet basis, which enrichesthe local feature and structure information. Then the ResNet101 network with residual block structure is used and compared with VGG16 and DenseNet201 networks. The experimental results show that Compared with the VGG16 network model, the ResNet101 network model has more stable training process, higher accuracy, and faster convergence speed. In training, the accuracy is raised to 100%, and the simultaneous loss rate is reduced to 0. Compared with the DenseNet201 network model, it has higher accuracy and more reasonable training process. It is proved that the ResNet101 network model is more suitable for the classification and identification of terahertz signals of defects in multilayer composite materials. This method can provide a new idea for the defect classification of multi-layer structure composites.