Deep Learning-assisted Penetration Control For Thin Plate TIG Welding

In the filled wire TIG welding process with a reserved gap,experienced welders can determine whether the workpiece is completely penetrated by observing the presence of fusion holes,and obtain excellent front and back side weld forming quality by controlling the size of the fusion holes,which are closely related to the forming quality of the front and back side welds.Therefore,in the automated welding process of TIG,facing a large number of welding production activities that require single-sided welding and double-sided forming,such as pressure vessels,oil pipelines and shipbuilding,how to monitor the dynamic behavior of the fusion hole during the welding process in real time and control its behavior,and then obtain excellent welded joints,has important theoretical value and great practical use needs.This paper uses 3mm thick stainless steel with misaligned thin plate filled wire butt TIG welding process as the research object,based on the existing TIG welding experimental platform,the construction of a stepper motor control module,vision sensor module,electrical signal acquisition and control module and TIG welding module as the main components of the TIG welding experimental platform.In order to effectively determine the status and locate the fusion holes in the topside melt pool image,a lightweight and efficient network is designed based on YOLOv3 to identify the status and locate the fusion holes in the topside image.Compared with YOLOv3,this network uses a lighter ResNet34-d to replace the original backbone network in YOLOv3;the Mish function replaces the original activation function Leaky-Relu in YOLOv3;a new deformable convolution is added to the backbone network,which can ignore the deformation of the topside melt pool image and better capture the fine features of the fusion holes;The CA module is added to the backbone network,which can suppress the tendency of the backbone network to focus on irrelevant features and improve the performance of the model;the original FPN structure of YOLOv3 is modified to include a channel from below,which can make the highlevel semantic information and low-level localization better combined and improve the efficiency of detecting the fusion hole location.The network can achieve higher accuracy and better generalization performance than YOLOv3 network while the accuracy is similar to YOLOv3.The network also shows better network performance when compared with other classical target recognition networks.Based on the features of the topside fusion hole image,the area of the topside fusion hole image was obtained using the region growing method based on the inference results of the deep learning network.The visual acquisition module and the deep learning module are combined in C++to realize the real-time acquisition and processing of topside fusion hole images.The shrinkage rate parameter of the fusion hole during welding was defined using the area of the topside fusion hole,which can effectively characterize the relationship between different welding parameters and the area of the topside fusion hole.Based on orthogonal experiments designed to analyze the influence of each welding process parameters(welding current,welding speed,wire feeding speed)and the amount of misalignment on the shrinkage rate of the fusion hole.At the same time,the wire feeding speed was found to be the largest welding process parameters affecting the shrinkage rate of the fusion hole.Based on this conclusion,the paper determines the control scheme with the topside fusion hole area as the controlled parameter and the wire feeding speed as the control quantity.In order to obtain a mathematical model of the wire-fill TIG welding system,the dynamic variation between the topside fusion hole area and the wire feeding speed was obtained using the step wire feeding speed as the input signal and the topside fusion hole area as the output signal,within the welding process interval that ensures the stable existence of the fusion hole.Second-order transfer function as a system identification model,the use of least squares to fill the wire TIG welding system for system identification,to obtain a better fit system model.Then,based on the second-order transfer function model,an H∞ robust controller was designed.The controller and the system identification model were imported into Simulink,and the appropriate controller parameters were obtained through simulation.The H∞ robust controller was simulated without and with added external disturbances,and the resistance of the controller to external disturbances and its good control performance were verified.Four major types of experiments are designed:variable gap constant misalignment,variable misalignment constant gap,variable gap variable misalignment and constant gap constant misalignment.The experimental results show that the controller can control the topside fusion hole area to near the reference value and obtain good front and back side weld formation,and realize the control of penetration and weld formation during fillet TIG welding with misalignment with gap and other complex working conditions.