Visual Sensing Of Fusion Hole And Penetration State Prediction In TIG Thin Plate Welding With Reserved Gaps

In the one-side welding with back formation process with reserved gaps,in order to ensure good penetration of the workpiece,it is necessary to form a fusion hole.Fusion holes are a unique phenomenon for butt welding with reserved gaps.The existence of fusion holes can ensure that the workpiece on both sides of the gap are fully melted,and then ensure that the welding seam is fully penetrated.The dynamic behavior of the fusion hole is very important for the stability of the welding process and the quality of the welding quality in the one-side welding with back formation with reserved gaps.In manual welding,the welder can control penetration and weld formation by judging and controlling the closing and size of the fusion hole.Therefore,it is of great significance to study the dynamic behavior of fusion holes in the welding process to guide the welding production.This paper improves the welding test platform on the basis of the existing equipment,so that the welding torch is tilted at a certain angle,and a synchronous visual sensing system based on dual CMOS cameras is built.A camera is on the back of the workpiece,facing the opposite direction of the welding direction,to monitor the backside fusion hole in real time;another camera is located on the front of the workpiece and observes the topside fusion hole toward the rear of the fusion hole.In response to the different light-sensing conditions on the front and back of the workpiece,the two cameras have designed different light filtering schemes to obtain clear images of the fusion holes.According to the gray characteristics of the acquired fusion hole image,the MATLAB image processing tool is used to design the corresponding image processing algorithm,the topside and backside fusion hole edges are obtained.The parameters of the two cameras are calibrated by the small hole imaging model,the extraction algorithm of the fusion hole feature parameter is designed,and the feature parameters(geometric parameters and shape parameters)of the fusion hole are obtained.Based on the characteristic parameters of the fusion hole,the behavior of the fusion hole under the constant and dynamic parameters was quantitatively analyzed,the formation and evolution of the fusion hole were analyzed,and the main process parameters(welding current,welding speed and reserved gap)were explored effect on fusion hole behavior.The results of the study found that:According to the different state of the fusion hole,the TIG thin plate perforation welding process can be roughly divided into:blind fusion hole stage,fusion hole growth stage and stable perforation stage.As the welding process continues,the fusion hole length shows a trend of increasing first and then decreasing.Under dynamic parameters,with the step change of welding current and welding speed,the fusion hole shows some new behavior.Welding current step increase and welding speed step decrease compared to welding current step decrease and welding speed step increase,the fusion hole reached a new balance in a shorter time,and this is explained and analyzed in conjunction with the change in arc heat input.At the same time,this paper uses the BP neural network to establish the corresponding relationship model between the fusion hole characteristic parameters and welding process parameters of the TIG thin plate perforation welding process,and obtains good prediction results and verifies the relationship model.In this paper,a prediction model of penetration state of TIG thin plate perforation welding with reserved gaps based on deep learning algorithm is built.Based on the VGG-19 deep learning model,the VGG-19 model trained on ImageNet is used as a starting point,continue to train it by using the fusion hole images of TIG thin plate welding with reserved gap.The model takes the topside fusion hole image as the input and the perforation/penetration state as the output.It automatically learns the image features through the convolutional neural network,and fine-tunes the model parameters and classification methods to meet the needs of penetration state prediction.After sufficient training,98%prediction accuracy is achieved on the training data set,97%prediction accuracy is achieved on the untrained test data set.