卷积神经网络的缺陷类型识别分析

该文提出一种基于卷积神经网络直接对阵列超声检测原始信号进行缺陷类型识别的方法,该方法无需对超声回波原始信号进行特征提取.文章研究对比了不同卷积神经网络及其优化的识别性能.首先采用超声相控阵系统对不同试块上的平底孔、球底孔、通孔三种缺陷进行超声检测,然后利用LeNet5、VGG16和ResNet三种卷积神经网络对一维和二...

Recognition and analysis of defect types by convolutional neural network

In this paper, a method of the defect type recognition based on the convolutional neural network is proposed by the original signal from the array ultrasonic testing. The method does not require feature extraction of the original ultrasonic echo signals. The recognition performance and optimization characteristic of different convolutional neural networks are studied and analyzed. Firstly, the experiments are conducted by the ultrasonic phased array system to collect detection signals which comes from the flat-bottom hole, spherical-bottom hole and via hole. Then, Lenet5, VGG16 and Resnet convolutional neural networks are used to identify the defects in one and two dimensions, and the network is optimized by Leaky ReLU, Dropout and Batch Normalization respectively. Subsequently, the identification accuracy and efficiency are analyzed. It is shown that 2-d convolution has higher recognition accuracy although its training speed is slower than that of 1-d convolution. At the same time, the identification accuracy will decline if the network model structure is too complex, and the data enhancement and optimization methods can help to accelerate the convergence speed and improve the accuracy.