| Intelligentized robotic welding system(IRWS)is one of the effective methodologies for enhancing the productivity of the traditional arc welding with thick steel plates.The process of the purely robotic autonomous welding with thick steel plates is typically hybrid dynamic,since it includes real-time weld planning,online adaptive adjustments of the welding parameter and the posture of the welding torch,and the control of the welding torch when it deviates from the determined tracking position.The problems within the above process belongs to modeling and control of the hybrid system(HS).Therefore,how to build an effective description for the process to implement the online judgments and optimization of the related problems,and to carry out the automatic control of the welding torch in this highly complicated and nonlinear system,is worth studying for control discipline.Combing the concrete project background,this thesis studied how to administer and implement the autonomous pulse MAG welding process of T-joints of thick steel plates using HS frame and artificial intelligence,which exampled the other welding methods with thick steel plates and showed a new thinking and strategy for resolving the complex control problem.The study includes as follows.Being the repeated welding with thick steel plates,the thesis analyzed the HS characteristic of the continue-two-passes welding process as a typical example,which was generalized as “Look for and adjust” stage,“Track” stage and “Return” stage.The whole stage was described with a macroscopic MLD model,namely Automaton.Through the constructed MLD,all the logical variables needed in the model were determined what they were and how they worked in the typical welding process,which can contribute to the following research and further to macroscopic administration and monitoring of the welding process in the control software.In order to realize accurate control of the welding torch for different weld seams,as well as resolving real-time weld planning,the online autonomous adjustments of the welding parameters and the posture of the welding torch,the thesis used the information carriers as follows: the entire arc region,weld seam profiles and their feature information.These information carriers are extracted from frames captured by a novel vision sensor based on structured light,which can capture laser stripes and the relatively regular arc region in the same frame.This kind of frame contributes to three aspects: one is that the real-time weld planning,autonomous adjustment of welding parameters and the main dip angle of the welding torch can be implemented using the identified feature points of the extracted laser stripe(the weld seam profile);the second aspect is that the tracking point for each pass can be appointed from the identified feature points;the last point is that the error between the welding torch and the tracking point can be calculated in real time using the geometric center of the weld pool and arc region and the extracted laser stripe,which can realize the closed-loop control.The background of the frame is complicated because of arc regions,which disturbs the extraction of the laser stripe from the background.In this thesis,to overcome the interference of the arc region,visual attention mechanism(VAM)was employed to extract the laser stripe.A visual-mutation-based visual attention model(VMVAM)belonging to the data-driven(bottom-up)manner was put forward through imitating the process and the characteristic of observing the region of interest with human eyes from the complex background.Within the model,a non-uniformity measurement(NUM)was presented to yield the orientation saliency map using the orientation feature map resulting from multi-region and multi-orientation Gabor filtering.Also,an intensity mutation measurement(IMM)was proposed to yield the intensity saliency map using the original frame.The comprehensive saliency map was acquired by linearly combing the intensity saliency map with the orientation saliency map.Eventually,the laser stripe could be extracted from the comprehensive saliency map after OTSU algorithms and nearest neighbor clustering have been applied to it.The effectiveness of the proposed model was clarified through considerable image processing experiments.Since the weld seam profile has diversity and complexity in multi-pass welding,the feature points of the weld seam profile,described as slope mutation points(SMP)were identified by the proposed method of span segmentation of slope monotone intervals(SSSMI),which was referred to the observation process where human eyes were commonly absorbed by the slope mutation positions on the laser stripe.The method need not set any slope mutation threshold.First,it determined all slope monotone intervals and then selected the intervals whose slope spans were over the average span.The selected intervals whose slope spans were segmented by OTSU were divided into two parts,while only the part were kept when their slope spans were still over the threshold from OTSU.Thus,the centers of the kept intervals were finally deemed as the slope mutation points.A large amount of feature point identification tests showed the SSSMI algorithm could satisfy different joints.Using the identified characteristic information of the weld seam profile and combing the corresponding welding knowledge,the thesis designed a simplified expert system(ES)via Visual C++ 2008,OpenCV and Matlab platforms,where Matlab GUI program was invoked by the main program based on Visual C++ 2008.The ES includes 5 inference subsystems as follows: weld planning subsystem,welding speed subsystem,wire feeding speed subsystem,wire extension and the tilt of the welding torch.In each subsystems,the welding knowledge were turned into the corresponding strategies and the correlated visual feature information was input,meanwhile,forward reasoning was adopted.The 5 subsystems respectively realized self determination of welding tracking points,namely indirectly autonomous weld planning,real-time decision-makings for welding speed,wire feeding speed,for wire extension and for the tilt of the welding torch.The time cost of these real-time decision-makings reached millimeter level.Meanwhile,the self-learning rules were also given in the front of these subsystems to avoid false SMP.Experimental results showed the effectiveness of the ES,which shows the feasibility of exploiting the visual information to implement deep intelligence welding in the traditional robotic arc welding.Modeling is difficult since the robotic welding system is highly nonlinear.Combing the characteristic of predictive control,the thesis proposed MPC to control the welding torch both in the direction of Y and Z.In the concrete implementation,two rectification methods were proposed through the size of calculated deviation in real time between the welding torch and the tracking point,and the nonlinear motion characteristics of the welding torch with saturation regions: MPC rectification for small deviation and direct rectification for large deviation.The two kinds of rectifications were approximately described as the corresponding equations.Combing the non-linear motion characteristics of the welding torch with saturation region and the above two control methods,the thesis built a time-variable switched model(SM),which was turned into a microcosmic MLD model through setting auxiliary logical variables.This model can appropriately predictive the position of the welding torch when the welding torch just runs a straight line.To solve MIQP problem,the research presented a mixed method where exterior point method(EPM)was employed to resolve inequality constraints whereas relaxation method was chosen to dispose of integer constraints like B&B method.To realize real-time optimization calculation of MPC based on MLD model,the research suggested an improved GA method which can decrease premature convergence.Simulation experiments determined the correlated setting parameters,which laid a good foundation for real seam tracking and control.There are two kinds of MPC in the thesis which are different from reference input: tracking data without arcing as reference input(TTWARI-MPC)and reference input constant in subsection in optimal time domain(RICOTD-MPC).To confirm the effectiveness of the two MPC methods,the control experiments had been carried out with backing welds and filling welding lying in different passes,and the control results had been compared with the incremental PID's.The comparison results showed the advantages and disadvantages of the above three control methods,and it concluded that RICOTD-MPC would be the best than the other methods applied to actual production. |
PDF Full Text Request