Research On Time Domain Ultrasonic Signal Recognition Technology Based On Deep Learning

One-dimensional time-domain ultrasonic signal contains the state information of the measured object.among all kinds of defect identification methods,ultrasonic echo signal analysis is one of the most important and useful means.The health state of the measured object can be identified by analyzing and processing it.In this study,the aluminum plate is taken as the research object,and the one-dimensional convolution neural network(1-D CNN)is used as the theoretical analysis tool to automatically extract features from the original time-domain ultrasonic signals and carry out the research work on the recognition method of one-dimensional ultrasonic time-domain signals.The main research contents are as follows:(1)A defect recognition method of time-domain ultrasonic signal based on 1-D CNN is constructed.It solves the problem that the traditional method is highly dependent on other factors,and simplifies the identification operation steps.JPR-600 C air-coupled ultrasonic testing system is used to collect data to prepare the data set for model construction.The method consists of four main feature extraction layers and two classification layers.Part of the parameter selection is determined with the help of visualization operation.On the defect data set,LSTM,CNN-LSTM,BP neural network and SVM model are compared respectively.The experimental results show that the performance of the proposed 1-D CNN model is satisfactory,and the accuracy of defect recognition is about 97%.(2)The ultrasonic defect recognition method based on one-dimensional residual network is improved.The structural advantages and innovation points of the residual neural network are applied to the network model to further improve the recognition accuracy and slow down the gradient dispersion problem.Taking 1-D CNN model as the basic model of network improvement,the training process of network model is visualized,and the feature extraction process of network model is analyzed.By using the same experimental data,the proposed method is compared with the above basic model,and the experimental results show that this method can better identify defect categories and achieve a recognition accuracy of more than 98%.