The Study Of Defect Map Processing And Intelligent Classification Technology In Ultrasonic Phased Array Exploration

The traditional method of detecting the welds of workpieces is usually to use ultrasonic flaw detector to generate a defect map,and then to observe it by some experienced workers,so as to realize the analysis of the defect types,The result is not only time-consuming,but also has a high error rate and can not provide a suitable warning signal.In order to solve the problem of low accuracy and slow classification speed of traditional weld defects classification methods,an intelligent classification of weld defects map based on transfer convolutional neural network is proposed in this paper.The main contents are as follows:(1)A bench-type ultrasonic phased array flaw detector was used to scan the weld defects and obtain the ultrasonic spectrum of the weld defects.Geometric transformation and image enhancement methods are used to pre-process the ultrasonic phased array detection pattern of weld defects in order to filter out the high frequency clutter noise while keeping the texture edge characteristics of the defect pattern,the original image is cropped and normalized.From the original image data,the clearer image containing the whole defect region is selected.(2)The experimental platform was built,and the data set was divided into training set and test set according to the ratio of 9:1.The classified network model was built by using ResNet-34,Mobile Net-v2 and AlexNet.The data were trained before and after the migration learning pre-training model was added to the three network models.The classification accuracy of the three networks is 94.5%,42.7% and 94.5%respectively before the migration learning pre-training model is added.The classification accuracy can reach 98.6%,84.5% and 96.5% when the migration learning pre-training model is added to the three network models respectively.The experiments show that the classification accuracy of weld defect map can be improved effectively by using convolutional neural network,and the classification speed is much faster than that of manual classification,it effectively speeds up the classification speed of the defect map.(3)The database of defect atlas is constructed,which is used to store the standard defect Atlas(that is,the atlas collected from the standard defect specimen).Using the trained convolutional neural network model,a defect map classification system is constructed,which is used to realize the intelligent recognition of the natural defect map and the fast matching with the standard defect map,the defect report is generated according to the type of defect and damage level.Give the proper warning.