Deep Learning Method And Application Research For Virtual Welding Quality Assessment

The level of welding technology is an important indicator to measure the strength of a manufacturing country.China's demand for welding workers has always maintained a large trend,especially high-tech welders.Welding skill training is an important way to improve welding level.With the rapid development of virtual reality technology,welding training has gradually shifted from traditional physical training mode to virtual simulation and physical training.The virtual welding simulator relies on its advantages of safety,efficiency,and pollution-free,etc,is welcomed in the welding training work,and gradually promoted the application.However,how to evaluate the virtual welding quality in the computer virtual environment and point out the problem of the student's operation has always been a technical difficulty.Therefore,this dissertation deeply studies the deep learning method for virtual weld defect identification,and intends to improve the level of virtual welding training.The main research results are as follows:(1)Aiming at the problems of computational complexity and poor realism in the existing virtual weld simulation technology,a simulation method of carbon dioxide welding based on smooth particle fluid dynamics is studied,which can simulate the flow of solder liquid quickly and realistically.Based on GPU accelerated rendering technology,the weld material,welding spatter and smoke effect are detailed,and the weld formation process of welding liquid formation and solidification is realized.(2)Virtual weld defect image is taken as the research object,and the collected defect image is preprocessed.firstly transforming the defect image into gray scale,and reducing the data processing amount while preserving the original image shape feature.The Otsu optimization segmentation algorithm is used to segment the defect region.Compared with the traditional Otsu segmentation method,the segmentation of the defect region is more complete,the boundary contour is more clear,and the segmentation speed is improved,which lays a foundation for accurate extraction and recognition of subsequent defect features.(3)In order to accurately extract the characteristics of virtual weld defects,the deep learning method was used to study the convolutional neural network recognition algorithm.Combined with the characteristics and categories of weld defects,a 9-layer convolutional neural network defect recognition network model was built.The neurons were used in Relu.The activation function is used to speed up the network convergence speed,which can effectively avoid the problem of “gradient disappearance” in the training process,and adopts Dropout in the all-connection layer to increase the generalization ability of the network,and at the same time,data enhancement is performed on the basis of the acquired original image.The sample size is greatly expanded to optimize the training effect of the model.(4)For the training of deep defect learning network for weld defect identification,three common frameworks for deep learning were compared and analyzed.The TensorFlow framework was used to train and test the defect recognition network.The crossover test of multiple training batches,learning rate and optimizer parameters was carried out.In order to improve the accuracy and robustness of network training,the accuracy of recognition is as high as 96.7%,and the accuracy of 200 new sample data sets is 96%,which reflects the better generalization of the model.Finally,compared with the training results of the classical convolutional neural network model LeNet-5,the validity of the network model for the identification of weld defects is verified.