Deep learning-based identification of pipeline weld defects using automated ultrasonic testing

For the safety evaluation of pipe welds, this paper proposes an automated ultrasonic testing method focused on defect signal feature extraction and identification. First, defect features are extracted based on the ultrasonic scattering coefficient distribution. The defect scattering coefficient matrix is then compressed using principal component analysis (PCA) to obtain the feature vector that best represents the defect. A Depthwise Separable Residual Network (DS-ResNet) model is constructed to identify pipe weld defects automatically. The sparrow search algorithm (SSA) is integrated with DS-ResNet (SSA-DS-ResNet) to optimise the model and enhance its performance. This method is applied to a case study, yielding a prediction accuracy of 97.51%, which is acceptable for industrial applications. The performance of SSA-DS-ResNet was compared with two networks prior to optimisation (ResNet and DS-ResNet), and the results indicate that SSA-DS-ResNet achieves higher accuracy.