Ultrasonic A-Scan Image Detection for 3D Braided Composites Based on Convolutional Neural Network

Abstract

Automatic ultrasonic signal classification of 3D braided composites is finding increasing use in many applications for the interpretation of signals. In this paper, we present a method for ultrasonic A-scan image classification based on deep convolutional neural network (CNN) model. The ultrasonic signal data set was obtained from Olympus Omniscan MX2 and divided into training data set and test data set. The batch normalization layer and the inception layer followed by max pool layer are added to the CNN model, which can speed up convergence of the deep learning network and improve the accuracy. With CNN model used for training the signals with debongding and the non-defect signals, experimental results showed that the accuracy rate of the method we proposed in the test data set reached 99.55%, which was greatly improved compared with AlexNet model and ResNet-34 network.