Detection of internal cracks in polyethylene pipes using ultrasonic imaging and deep learning

Abstract

This study presents a two-stage ultrasonic testing (UT) approach for detecting and classifying internal cracks in polyethylene (PE) pipes. In the first stage, ultrasonic imaging is performed using a linear array of piezoelectric transducers to identify crack locations. A probabilistic triangulation imaging method is applied to successfully detect the cracks with the depths as small as 20 % of pipe wall thickness. In the second stage, a deep learning model is developed to classify crack severity using the paired transducer corresponding to the location of the imaged crack. A convolutional neural network (CNN) is trained on signal images converted from continuous wavelet transform (CWT). To generate a diverse training dataset, finite element modeling (FEM) is utilized based on preliminary UT experiments and the material damping properties are calibrated with experiments. The CNN model trained on the synthetic database achieves high classification accuracy when validated through laboratory-acquired signals. The results demonstrate that the proposed two-stage approach accurately locates cracks and enables automated severity classification, enhancing UT-based structural health monitoring of PE pipelines.