Quantification of Defects with Point-Focusing Shear Horizontal Guided Wave EMAT Using Deep Residual Network

In this work, a deep residual network named GFresNet-2D is proposed for a point-focusing shear horizontal guided wave electromagnetic acoustic transducer, which can be used to quantify different types of defects, such as pinholes, cracks, and corrosion, in materials. As the traditional feature extraction and statistical machine learning methods are too complex and rely on artificial recognition, an automatic feature extraction model based on deep learning is applied for defect detection and quantification. Owing to their similarity with the ultrasonic guided wave signals, the measured 1D signals from the experiments cannot be directly applied to train neural networks. Therefore, we used the normalization, minimum suppression, and continuous wavelet transform methods to convert the initial measured 1D signals into processed 2D images, and constructed a data set containing 1,440,000,000 signal/image data. The performance of the proposed GFresNet-2D model for this new data set was also compared with those of traditional models, and sensitivity analyses were performed for some of the representative parameters. The results confirm that the proposed method can contribute to the development of deep-learning-based defect quantification using the ultrasonic guided wave focusing method.