Prediction Of Pool Penetration State In Plasma Arc Welding By Fusing Attention Mechanism

Image classification technology is the foundation of computer vision and has extensive academic research value.With the development of convolutional neural network technology,image classification technology has been widely applied in various industries.In the field of industrial welding,plasma arc welding is a new type of welding technology,which uses high temperature plasma generated by arc discharge for welding.Compared with traditional welding technology,plasma arc welding has the characteristics of low cost,strong adaptability,low requirements for joint assembly accuracy,etc.Therefore,it is widely used in large-scale industrial equipment such as ships,aviation,bridges,etc.However,with the increase of plate thickness in the welding process,the stress of molten pool and small hole becomes more complex and the state becomes more unstable,which greatly increases the difficulty of welding and limits the wide application of plasma arc welding technology.In order to solve this problem,this thesis combines the image classification technology based on deep learning with the prediction of molten pool penetration in plasma arc welding to improve the welding quality and efficiency.In deep learning,the feature learning ability of the model directly affect the final result of image classification.For the front image of plasma arc welding with small feature difference,the prediction accuracy of a simple deep learning network model is difficult to meet the needs of industrial applications.Therefore,this thesis improves the input and residual structures of the classic Res Net50 network model to increase computational complexity and prediction accuracy.Additionally,custom spatial channel attention modules and visual transformer modules are added to make the network focus more on features that have a significant impact on classification,thereby further improving the classification performance of the model.The loss function plays an important role in the deep learning network.In the training process of the network model,the loss function controls the trend of the model and directly affects the final performance of the model.Therefore,the mechanism of loss function is analyzed in detail in this thesis.Based on the characteristics of plasma arc welding front weld pool image data set,this thesis defines a joint loss function as the loss function of this model.The joint loss function takes the cross＿entropy loss function as the main loss function,and defines an auxiliary loss function to adjust the main loss function,which is used to learn more about the characteristics of differences between classes,thereby improving the performance of the model.The innovations of this study,which include improve the Res Net50 network model,introducing custom space-channel attention and visual transformer modules,and customizing the joint loss function.These innovations can help improve the performance of the model in predicting pool penetration state of plasma arc welding.In order to meet the application requirements of plasma arc welding industry,this thesis uses Grad-CAM technology to visualize the model and show the classification basis of the model.The feasibility,explicability and reliability of the network model in a complex industrial environment are validated by comparing the predicted results of the model with the empirical feature information of the actual welding process.In conclusion,this thesis successfully combines image classification technology based on deep learning with the prediction of molten pool penetration in plasma arc welding to improve welding quality and efficiency.The improved Res Net50 network model with custom spatial channel attention and visual transformer modules and the joint loss function have significantly improved the performance of the model.The use of Grad-CAM technology has also demonstrated the feasibility,explicability,and reliability of the model in a complex industrial environment.