Deep Learning Based Time-frequency Pattern Recognition Of Square Steel Defects

Used for the detection of orbit injuries work is mainly done through the corresponding auxiliary equipment,when there is damage orbit,flaw echoes are converted to electrical signals,display will appear different waveform images,detect personnel through the observation on the display echo signal to judge the defect size and location,so the experience and competence of the testing accuracy and testing personnel,Inevitably,there will be artificial misjudgments.To solve this problem,this paper proposes a time-frequency analysis method using Wigner-Ville(WV)distribution to analyze the characteristics of time-frequency graphs of different defects.With time-frequency graphs as the data set of deep learning,a improved network model based on Shufflenetv2 is proposed to classify the defects of square steel with different shapes,which effectively improves the work efficiency.It avoids the misjudgment caused by human subjective judgment.Firstly,the model of ultrasonic guided wave propagation in square steel is built in the simulation software COMSOL.Three typical geometric shapes of square,triangular and circular defects are taken as the research objects.Different defect echoes are obtained by changing the size of defects by modifying the scaling factor S.Wigner-Ville(WV)distribution was used for time-frequency processing of defect echoes to get time-fr-equency diagrams.In MATLAB,the fitting toolbox was used for data fitting of time-frequency diagram features and defect sizes of different shapes of defects.In the defect shape classification task,ShufflenetV2 is used as the basic model of network improvement,and the structural advantages and innovations of MobileNetV1 and MobileNetV2 are applied to the network model,and SM1 and SM2 network models are built respectively.Different defects are constructed by changing the area scaling factor S,the side length R of the geometric shape,the position of the defect in the square steel and the size of the excitation amplitude in the COMSOL simulation software,and the defect data set is formed after time-frequency analysis of the defect echo.The performance of Googlenet,VGG16net and Resnet,Squeezenet,MobileNet and ShuffleNet are compared on the defect data set.The feature graphs of SM1 and SM2 network models are visualized,and the feature extraction process of network models is analyzed.After comparing each data fitting method,the time length features in time-frequency diagrams were taken as the research object.For square and triangular defects,the size of defects was proportional to the time width of time-frequency diagrams.For circular defects,the regularity between the size of the defect and the time length of the time-frequency diagram is poor with the increase of the defect area.In the defect shape recognition and classification task,the accuracy of ShufflenetV2 is 0.97,and the running time of each round is 27s.The accuracy of ShufflenetV2 and SM1 are 0.993 and 0.986,respectively,and the running time of each round is 19s.That is to say,the accuracy of ShufflenetV2 is improved and the number of network model and the running time of each round are reduced.Compared with manual detection,the method proposed in this paper effectively avoids manual error detection and has higher accuracy.