Intelligent defect recognition for pipeline girth welds based on automated ultrasonic testing

Abstract

Automated ultrasonic testing (AUT) is an effective technique for non-destructive evaluation (NDE) of pipeline girth welds. In this paper, a deep learning method is proposed for automatic identification of defects in such welds. Criteria for defect classification in the girth weld are extracted from AUT mapping. A deep learning model, namely the depthwise separable residual network (DSResNet), is proposed by optimising the depthwise separable convolutional layer and weighted loss function in ResNet-18. The hyperparameters in the network are optimised using an improved sparrow search algorithm (ISSA). The AUT data are preprocessed and augmented to form a database for pipeline girth weld identification. The proposed deep learning network is used to analyse experimental AUT data obtained from pipeline girth welds and it is found that the proposed method can be effectively used for automatic identification of defects in these welds, with the identification accuracies for five typical defect types (including non-defects (NDs)) all above 90%.