Prediction of Depth of Defect from Phased Array Ultrasonic Testing Data Using Neural Network

Ultrasonic non-destructive testing (NDT) requires the involvement of an expert operator for inspection and interpretation. This makes the process outcome sensitive to multiple forms of human error, leading to inaccuracy in results. The present demands of society have increased the volume of inspection and testing costs exponentially. The potential solution to these problems is to use machine learning (ML), and recently researchers and industries have started exploring the diversity of ML techniques to use in NDT. This paper investigates the prospect of the artificial neural network (ANN) to characterize the defect in NDT through simulation and experiment. First, synthetic A-scan data was generated from an angle beam ultrasonic model using COMSOL, and using these data the depth of the defect was characterized using a feed-forward neural network. It is found that a simple topology of 10:10:2 network performs well and gives a correlation coefficient of 0.95 between the output and target. Second, an experiment was performed by preparing samples (mild steel blocks) with artificial defects at different depths. The depth characterization was performed by extracting the features from A-scan data using a phased array ultrasonic testing (PAUT) device. The result shows that the feed-forward network can predict the depth of defect with a mean squared error of 0.0701.