Towards using convolutional neural network to locate, identify and size defects in phased array ultrasonic testing

Machine learning algorithms are widely used in image recognition. In Phased Array Ultrasonic Testing (PAUT), images are typically formed through constructive and destructive superpositions of signals backscattered from flaws or geometric features. However, all PAUT data acquisition schemes require several emissions and the duration of the acquisition may be too slow in high-speed manufacturing. In this study, the Faster R-CNN was used to identify, locate and size flat bottom holes (FBH) and side-drilled holes (SDH) in an immersed test specimen using a single plane wave insonification. The training was performed on segmented and classified data generated using GPU-accelerated finite element simulations. SDH and FBH of different diameters, depths and lateral positions were included in the training set. The thickness of the test specimen was also variable. An ultrasonic phased array probe of 64 elements was simulated. All elements of the phased array probe were fired at the same time and the time traces from each element were recorded. The individual time traces were concatenated to form a matrix, which was then used in the training. This inspection scenario enables fast acquisition of data at the expense of poor lateral resolution in the resulting image. The trained neural network was initially tested using finite element simulations. Results were assessed in terms of the intersection of the union (IoU) between the ground truth geometry and the predicted geometry. With the simulated cases, the thickness of the test specimen was detected in all cases. When using a 40% IoU threshold, the detection rate of the FBH was 87% while only 20% for the SDH. The smallest detected FBH had a 0.56 wavelength depth and a lateral extent of 1.04 wavelength. Drawing a box using the -6dB drop method around the FBH always led to an IoU under 15%. On average, the lateral extent of the FBH using the -6dB method was three times larger than the diameter predicted by the proposed method. Then, the training was continued with a small augmented dataset of experiments (equivalent to 3% of the simulated dataset). In experiments, the results show that the test specimen was always correctly identified. When using a 40% IoU threshold the experimental detection rate of the FBH was 70%. The smallest detected defect in experiments had a depth of 2 wavelengths.