Prediction Of Stamping Defects For Automobile Panels Based On Deep Learning

The automobile panels is the main component of the automobile body,and the modernaesthetic concept has increasingly increased the requirements for the structure and appearance of the automobile body,therefore,the quality of stamping forming for automotive panels is becoming stricter.Whether a company can produce high-quality panels is critical to the production of a complete vehicle.In the early process design of the panels,it is necessary to predict the stamping defects of the panel design in real time.Compared with traditional CAE,finite element analysis is relatively time consuming and cannot realize real-time prediction of automobile panels.With the development of artificial neural networks and deep learning techniques,it is possible to use deep learning techniques to realize real-time formability prediction of early product design,reducing the early design flaws of products.Convolutional neural networks can process complex images,use feature maps to achieve image classification and target detection,etc.,and have great advantages in the field of target detection.According to the shape analysis target of the panels,this paper based on the Refine Det target detection network to predict the stamping defects of the automobile panels in real time.A large amount of data is required for deep learning training,and there is no open source data set for stamping and forming of the automobile panels.Therefore,the KMAS/One-step simulation analysis is used to generate the formed thickness cloud image of the cover piece and the Gaussian curvature cloud image data picture,a total of 7743 sheets,and use the label generation tool Labelimg to pre-mark the defect parts.The generated Gaussian curvature cloud map is used to train the Refine Det network and test the network accuracy.The results show that the average accuracy of the Refin Det neural network for the prediction of stamping defects is 75.48%,and the prediction accuracy of the Crack is 72.92%,and the prediction accuracy of the Wrinkling is 78.04%.At the same time,the Res Net-101 feature extraction network is used for prediction,and the accuracy is slightly improved,The average accuracy of defect prediction is 77.46%.It is proved that deep learning is feasible for real-time prediction of defective parts of automobile panels.On this basis,the paper analyzes the influence of the selection of the blanking force,friction coefficient,die clearance and drawbead resistance coefficient on the forming qualityof the blanking ring in the forming of the workpiece.The neural network is constructed,and the neural network is trained by using the process parameter data in the orthogonal table,and then the relationship between the process parameters and the forming and thinning rate of the cover is predicted.The test shows that the prediction result of the neural network on the thinning rate and the Autoform stamping simulation software is less than 5%,which proves the feasibility of the neural network for the prediction of the thinning rate of the panels forming.