Welding Vision Detection Method Based On Deep Learning Convolutional Neural Network

In recent years,with the continuous expanded market demand for battery electric vehicle,lithium battery,which is the important component of electric vehicle,have also got vigorous development.Laser welding is the key technology for battery manufacturing.In order to control the production quality,the industry has a great desire for quality inspection system for the surface welding of lithium batteries after laser welding.As one of the main methods of welding quality inspection,machine vision technology can inspect the welded products quickly and accurately.Based on the previous work of our group,this thesis proposes a welding vision inspection method based on deep learning convolutional neural network,with the aim to design a low-power,high-efficiency,high-accuracy algorithm to detect safety vent’s laser welding defect.The thesis first introduces the research background of welding defects in lithium batteries,and compares various welding defect detection methods,then decides to adopt the high accuracy deep learning algorithm as the main research algorithm.Then we designed an image acquisition system(including the selection of cameras,lenses and light sources)of the welding defect area,and collected totally over 8000 images.Firstly,a defect detection algorithm based on HALCON software was designed,and the experiment results show that the detection efficiency and the classification accuracy are low.Additionally,most deep-level convolutional neural networks based on deep learning theory cannot be run directly on industrial computers because of their large models.Therefore,after considering the accuracy and requirements for hardware,a solution to optimize the convolutional network model is proposed.Finally,regarding simplifying the structural model,improving efficiency and accuracy,the scheme of optimizing VGG-16 model based on transfer learning theory and pre-training methods has been proposed.After pre-training VGG-16 in Image Net,the saved convolution base remains unchanged when training the model,and the fully connected layers are replaced for optimization.Moreover,the thesis further analyzes the reasons and effects of using pretraining methods,that is,reducing the number of training rounds,reducing the model size and overfitting.Subsequently,several networks that have achieved excellent results in the Image Net competition are selected for comparison experiments,mainly including the comparison of accuracy,fault positive rate,recall,and accuracy.Finally,the convolutional layer and the maximum pooling layer were visualized to facilitate viewing and optimizing the model.The research results show that the VGG-16 model based on pre-trained network has the advantages of small model,low fault positive rate,short training and prediction time,etc.On the test set,the accuracy rate of the model reached 99.87%,the fault positive rate was only 0.16%,and the predicted time for single image time was about 40 ms.The proposed VGG-16 is superior to convolutional neural networks like VGG in performance and is more suitable for laser welding quality inspection in industrial environments.