Research On Ultrasonic Testing Method For Internal Defects Of Castings Based On Convolutional Neural Network

Metal materials have been increasingly used in various products in recent years,which have a great impact on people's lives.Most of these metal parts are produced by casting methods,and the internal defects of the products during the casting process directly affect the quality of the products.At present,the ultrasonic testing method is a common method for detecting internal defects of castings because of its convenient detection and completely harmless to people and the environment.However,the identification of ultrasonic testing results of internal defects of castings mostly relies on manual experience,which seriously affects the efficiency in industrial production.Therefore,in order to solve this problem,based on the relevant theory of deep learning,an intelligent algorithm is designed to realize automatic identification of internal defects.The main work of this thesis is divided into the following three parts:(1)Defect detection and image acquisition: Detection of internal defects in test blocks by ultrasonic detection equipment using pulse reflection method,ultrasonic detection image of type A scan obtained on a computer using a software export device.The acquisition and export process of image requires manual operation.(2)Preprocessing of detected images: The original image captured by the detection device is more complicated,which is not conducive to the identification of internal defects.Firstly,binarize the image to reduce the amount of data.Secondly,cut the image removal device system interface.Then the edge detection algorithm is used to remove the background noise of the ultrasonic signal.Finally,the image morphology method based on connected domain realized the extraction of ultrasonic signals.(3)Internal defect recognition based on convolutional neural network: Based on the research of Convolutional Neural Network(CNN),five layers of CNN are designed to identify defects.The Softmax function is used as the activation function,and the over-fitting problem is solved by the Dropout technique.The final experimental results show that the recognition accuracy of CNN can reach over 90%.Compared with the BP neural network's defect recognition experiment results,CNN recognition algorithm performs better.In the experiment,it is found that the size of convolution kernel has an impact on the network performance of CNN.In the process of experimental research,the CNN network optimization scheme based on adaptive convolution kernel is proposed and verified by experiments.The improved learning performance is obviously improved and the test accuracy is also improved,which proves that the method has an optimized effect on the performance of CNN.Through the above three parts of research,the CNN and its adaptive improvement designed in this thesis have good defect recognition ability in ultrasonic detection of casting internal defects,and the accuracy of automatic recognition of defects is high.