A Guided-Wave Based Damage Diagnosing Method with Energy Spectrum and Multi-scale Network

As a promising approach of structural health monitoring (SHM), Guided-wave is being widely used for damage diagnosis of carbon fiber reinforced composite (CFRP) in aeronautics. The traditional methods usually employ threshold for damage detection. However, as Guided-wave is easily influenced by structural and environmental factors, the man-made thresholds may vary significantly from case to case, which deteriorates its universal application. To overcome this weakness, a deep-learning-based method is proposed. Firstly, the sample is produced in the form of grey scale, so that its label is determined objectively rather than by human judgement. Secondly, it employs a multi-scale convolutional neural network with dedicated settings of kernel size and number of convolutional layers, which can effectively extract damage features from multi-dimensional signals while keeping the negative influences of increased network width low. Additionally, Gradient-class activation mapping (GRAD-CAM) technique is used to reveal the sample areas where the deep learning model concerns, which helps to interpret the physical significance of damage features. The experimental results demonstrate that the accuracy of damage identification and localization reaches 92.6%, which is superior to the traditional methods.