Assessing Micro Wall-Thinning Damage in Aluminum Frames of Subsonic Speed Vehicles: A Comprehensive Approach Integrating Higher - Order Harmonic Nonlinear Guided Wave Analysis and Convolutional Neural Networks

Owing to advancements in propulsion technology, high-speed trains are subjected to significant fatigue stress on their frames and bodies. These effects can elevate the possibility of significant accidents on railways. In this study, a deep learning approach is used to identify an evaluation technique for verifying an initial wall-thinning crack using frequency spectrogram images. The wall-thinning effects of 6.3 mm thick Al-2024 T6 plates, primarily used for frames and bodies, are confirmed through frequency responses of guided waves. To verify the effects of wall thinning, higher-order harmonic frequencies of guided waves and ultrasonic vibration techniques are simultaneously applied. To identify the tendency of frequency response, frequency spectrogram images, illustrated by short-time Fourier transforming analysis, are chosen. To train the detection system using the region-based convolutional neural network (R-CNN), spectrogram images are generated for wall-thinning in steps of 0.12 mm (2% of the thickness). To train the R-CNN engine, the ResNet-50 architecture and the ReLU activation function are used in the convolutional layer to address the issue of gradient vanishing. From the results of this research, the depth of wall thinning in the plates can be identified. Based on the R-CNN training set, the initial wall-thinning failure can be verified with reliable accuracy.