A Study Of The Visual Morphology Of The Weld Pool In Non-melting Electrode Welding

In the industrial welding environment,the research of molten pool shape characteristics plays a key role in the on-line monitoring of welding process quality.In this paper,the extraction algorithm of molten pool contour and the prediction method of formed weld width are studied.The main research work is as follows:(1)The image acquisition system of molten pool based on passive visual sensing method is established.Based on the characteristics of TIG welding stainless steel molten pool image,in order to obtain the weak edge information and the closed molten pool contour,an algorithm of molten pool contour extraction(Operator Template Matching based on Edge Direction Guidance,OTM-EDG)is designed based on the improvement of traditional image algorithm.The algorithm mainly includes the enhancement of weak edge area based on nonlinear grayscale transformation and gradient calculation based on edge oriented operator template matching.Through the analysis of the experimental results,it is confirmed that the algorithm in this chapter is more accurate than the traditional edge detection algorithm for weak edge detection,and can obtain the closed contour edge.(2)A scheme of extracting of molten pool contour based on convolution neural network is proposed.A semantic segmentation network Res-Seg is designed based on residual network,in order to improve the robustness of the trained network model,the data set expansion method based on DCGAN(Deep Convolutional GAN)is combined and the data augmentation based on color and morphology is carried out before network training.By comparing with the traditional edge detection algorithm and other semantic segmentation networks,it is verified that the network model proposed in this chapter has high segmentation accuracy and strong generalization ability for molten pool images under various welding process parameters in industrial environment.(3)A method of predicting weld width by using BP neural network is proposed,which takes four parameters of molten pool: molten area pixel width,welding current intensity,welding speed and wire feeding speed as the input of the network,and the actual weld width as the output of the network to train the BP neural network.After the comparison and analysis of the test results,it is verified that the average test error of this method is less than 0.25 mm,which can be used to predict the weld width in industrial welding environment.