Defect detection in metallic structures through AMR C-scan images using deep learning method

摘要

Failure analysis in metal structures prior to the arrival of the specimen to the boundary of instability is one of the most important issues in the industry. Hence, the importance of the Non-Destructive Evaluation (NDE) techniques which work with high accuracy and efficiency ability in detecting the defects is rapidly increasing. Given the fact that most of the existing equipment are made of conductive and magnetic materials, therefore, the electromagnetic fields can be used for defect detection. In this study, we present utilization of a Deep Learning method for characterization of defect areas in a specimen through x, y and z planes C-scan images. Due to its low noise sensitivity and ability to operate at low frequencies, an Anisotropic Magnetoresistive (AMR) sensor is employed for the sampling of a magnetic field over specimen surface to provide C-scan images. Next, for sizing the defect existed in the specimen, a Deep Convolutional Neural Network (DCNN) is designed and implemented for the obtained C-scan images. Finally, the validation of the obtained DCNN is carried out and its efficiency in identification of defect characteristics is compared to traditional neural networks (NNs). The most significant advantage of DCNN is related to its ability in automatic extraction of learning features from the data applied to the network depending on the problem, which here it leads to an improvement in the accuracy of the defect analysis relative to the other traditional NNs.