Research On GTAW Weld Penetration Recognition Based On Deep Learning

Welding technology supports the development of advanced manufacturing industries such as aerospace,shipbuilding,and transportation industry.The continuous improvement of the industrial level puts forward higher requirements for welding automation and welding quality online monitoring technology.The weld penetration state is a basic indicator to evaluate the weld quality.Skilled welders can infer the real-time weld penetration state by observing the shape of the weldpool during the welding process.Inspired by this,a series of automatic weld penetration recognition technologies have been proposed successively.Traditional weld penetration recognition technology has a strong dependence on feature engineering and is difficult to promote and apply.In recent years,the end-to-end deep learning method has made great progress.The deep learning method reduces the dependence on feature engineering.Deep learning models can automatically learn feature representation from data,and have high prediction accuracy,which brings a new way for weld penetration recognition.This dissertation carries out the research of the GTAW(Gas Tungsten Arc Welding)weld penetration state recognition method based on deep learning.The dissertation first introduces the current application of deep learning in welding,and analyzes the research progress of weld penetration state recognition.Deep learning is driven by data.In order to obtain clear images of the weld pool,a platform for collecting weld pool images was built,a visual sensor system was designed,and a large number of welding experiments were carried out.By adjusting the welding process parameters,it is possible to obtain the weld pool images of different weld penetration states including lack of fusion,lack of penetration,desirable penetration and excessive penetration.The image samples of various penetration states were sampled and balanced,and the weld pool image dataset was divided for deep learning model training,verification and testing.A deep learning development platform was built to develop GTAW penetration state recognition models.Convolutional neural networks with different structures were designed and optimized.The impacts of learning rate and batch size on model training process were analyze.In order to improve the generalization ability and development efficiency of the GTAW weld penetration recognition model,the transfer learning method was adopted: the recognition accuracy of the custom convolution model was improved through instance-based transfer learning,up to 0.9254;through network-based transfer learning,the pre-trained VGG16,Inception V3,Res Net18,and Mobile Net V2 were applied to GTAW penetration state recognition,and the penetration state recognition performance of the transfer learning models under different training strategies were compared.The test results show that the transfer learning model has more advantages than the custom model,with higher recognition accuracy and development efficiency.In order to further verify the effectiveness of the developed GTAW penetration state recognition model,the output feature maps of the layers in the model and the gradient-weighted class activation map are shown,which reflect the layer feature transformation of the input weld pool image;the visualization results show that the area that the model relies on for the weld penetration recognition is similar to the area that the skilled welder pays attention to when inferring the weld penetration state.Finally,the developed transfer learning model is loaded and integrated into the control software of the welding camera through the ML.NET machine learning framework to realize the online recognition of different weld penetration states.