A Deep Learning-Based Ultrasonic Pattern Recognition Method for Inspecting Girth Weld Cracking of Gas Pipeline
Yan Yan, Dong Liu, Bin Gao, Gui Yun Tian, Zhichao Cai
Abstract
Electromagnetic Acoustic Transducer (EMAT) has become one of the fastest-growing solutions for pipeline weld inspection over the past decade due to its non-contact advantage. One primary problem of EMAT is that it has relatively lower energy transition efficiency compared to widely used piezoelectric transducers, coupled with the effect of lift-off and the non-uniformity issue of welding material, the Signal-to-Noise Ratio (SNR) can be significantly restricted. This brings great difficulty in interpreting the EMAT signal measured from pipeline girth welds. To overcome this challenge, this paper presents a deep learning-based ultrasonic pattern recognition method to identify the pipeline girth weld cracking automatically. The proposed method utilizes a deep Convolution Neural Network (CNN) integrated with a pre-trained Support Vector Machine (SVM) classifier to extract the high-level features from the time-frequency representation of A-scan signals measured by bulk-wave EMAT and classify these signals into defective or non-defective groups. To validate the proposed method, a set of experiments is carried out to classify A-scan signals measured from the girth welds of an ex-service type 813-X70 gas pipeline. A comparative investigation is also undertaken to demonstrate the superiority of the proposed method against the conventional ultrasonic pattern recognition methods for evaluation.