Abstract: In view of the complexity and diversity of welding products of aerospace components, this paper improves the welding quality problem caused by assembly or welding stress deformation of the teaching and reproduction type welding robot. In this paper, under a prototype of an intelligent welding robot system that can autonomously complete the welding of complex spatial curve welds in a local environment, the real-time sensing of the welding pool is realized by the CCD camera of the sensing system. Based on the reliable communication between the penetration control computer and the central control computer, a PID controller is designed to realize the real-time dynamic process control of the welding pool of LF6 aluminum alloy variable heat dissipation workpiece. Keywords: Local autonomous intelligent welding robot, welding pool, visual sensing, process control 0 Introduction Welding is an important means of material processing and is widely used in industrial production. Due to the promotion of many factors, the automation and robotization of welding process has become a trend [1]. In particular, the research and development of intelligent welding robots and their intelligent technologies, and the realization of quality control functions such as automatic welding path planning, real-time automatic path correction and weld tracking, welding penetration, and weld formation will be the development direction of welding technology in the 21st century. The research on welding quality control is an important part of welding process automation. As the requirements for automation and intelligence of welding processes are increasing, the control of welding quality has become particularly important. In recent years, with the development of computer vision technology, the use of machine vision to directly observe the weld pool, obtain the geometric shape information of the weld pool through image processing, and perform closed-loop control of welding quality has become an important research direction [2-4]. The weld width, weld depth and other dimensional parameters are important factors affecting welding quality, and the weld pool is the most direct factor affecting the weld size parameters. Therefore, studying the changes in the weld pool during the welding process and realizing the control of certain parameters of the weld pool has practical significance for the control of welding quality. At the same time, according to actual production experience, welders adjust the welding process specifications and welding gun posture according to the changes in the weld pool to ensure welding quality. Therefore, controlling welding quality according to the changes in the weld pool is also an important part of realizing welding intelligence. 1 Introduction to the structure of the arc welding robot system for real-time control of weld penetration of local autonomous intelligent pulse GTAW The main hardware part of the experimental system is a local autonomous intelligent welding robot system, which uses the robot product RH6 of Shenyang Siasun Robot & Automation Co., Ltd. as the main body [5]. A camera follower device (joint 7) is added to the end of the robot body (joint 6), which can drive two cameras to rotate around the welding torch via belt drive [6-7]. The core of the Local Autonomous Intelligent Welding Robot system is a PIII850 general-purpose computer, which provides coordinate transformation, trajectory generation, interpolation calculation, external information integration, and control, status management, and task scheduling of the welding system. It acts as a central monitoring computer, communicating with the robot control cabinet via a CAN bus. It communicates with the guidance, weld seam tracking, and penetration control modules via Ethernet. The guidance and weld seam tracking modules use visual sensing to guide the welding robot to its initial welding position and track the weld seam. One CCD camera on the seventh axis of the LAIWR robot is connected to the image acquisition card on the guidance and weld seam tracking computer via a video cable. Another CCD camera on the seventh axis of the LAIWR robot is connected to the image acquisition card on the penetration control computer via a video cable (hardware configuration is the same as the guidance and tracking module). Meanwhile, the penetration control computer is connected to the welding power source via a self-developed control/acquisition interface circuit to achieve real-time control and acquisition of welding parameters. The welding power source is the INVERTER ELESON 500P AC/DC dual-purpose GTAW welding power source manufactured by DAIHEN Corporation of Japan. In addition, the system includes auxiliary equipment such as a water tank and protective gas cylinders. The core of the penetration control test system is a PⅡ350 desktop computer, which controls the welding power source, wire feeding mechanism, and welding motion mechanism through a data acquisition card, stepper motor control card, and interface circuit. Welding current and wire feeding speed are controlled by adjusting the input voltage at the welding power source control terminal. Welding speed is controlled by changing the number of pulses input per unit time of the stepper motor. The image acquisition card is connected to the computer, which receives the timing and shape characteristics of the welding current transmitted by the welding power source and can control the image acquisition timing. The wire feeding mechanism is located at the upper end of the robot's third axis, with the wire feeding position diagonally in front of the molten pool. This system can complete the entire welding process, from automatic arc initiation, welding process parameter setting, automatic acquisition of molten pool images, calculation of molten pool image dimensions, implementation of control algorithms, automatic adjustment of welding specifications, to automatic arc extinguishing. The structural block diagram of the entire robot pulse GTAW sensing and process control experimental system is shown in Figure 1. Figure 1: Structural block diagram of the robot pulse GTAW experimental system . 2. Determination of broadband composite filter system based on continuous spectrum and definition of aluminum alloy molten pool parameters. To control weld quality using image sensing, it is essential to obtain a molten pool image. To obtain a clear molten pool image, a filter system is necessary. The basic law of arc light spectral distribution under aluminum alloy GTAW welding conditions is that it consists of many other spectral lines of varying intensities superimposed on a relatively low-intensity continuous spectrum. Under different process parameters (welding current, welding voltage), the spectral distribution will vary slightly, but the basic law remains unchanged. The spectrum near the molten pool surface mainly consists of the aluminum atom spectrum, aluminum ion spectrum, and the continuous spectrum generated by the blackbody radiation of the molten pool metal. The spectrum in the arc column region mainly consists of argon atom and ion spectral lines, and also contains vapor lines of other metals. Therefore, the method of obtaining images of the GTAW weld pool of aluminum alloy using narrowband filtering is not feasible. By analyzing the spectral distribution of actual GTAW welding of aluminum alloy, this paper uses a broadband filter with a transmission range of 590nm-710nm and a peak transmittance of 25%, and an attenuator with a transmittance of 20%. The penetration control sensor is placed directly behind the weld pool, with the sensor's centerline at a 45º angle to the welding direction, allowing for clear acquisition of the weld pool image. Controlling the formation of the aluminum alloy pulsed GTAW weld using visual sensing is crucial for obtaining information reflecting the weld formation process. Image processing is used to obtain characteristic parameters describing the weld pool state, and controlling the stable formation of the weld is essentially controlling the stability of these parameters. Therefore, this paper defines the characteristic parameters of the aluminum alloy GTAW weld pool image as: the width and length of the frontal weld pool. Furthermore, in the frontal image of the aluminum alloy GTAW weld pool obtained in this paper, the entire weld pool cannot be seen; therefore, half the length of the frontal weld pool is utilized. (See Figure 2). [align=center]Figure 2. Schematic diagram of molten pool image and feature parameters (a) Frontal image of molten pool (b) Schematic diagram of molten pool feature parameters[/align] 3. Establishment of a neural network model for the frontal parameters of the pulsed GTAW molten pool A model is a tool or means for understanding and studying a system. In recent years, with the rise of artificial neural network research, a new path has been opened for the establishment of dynamic models of the welding process. Due to the large thermal inertia of the welding process, when establishing an ANN model, we should not only consider the current welding specification input, but also the influence of the historical values ​​of the welding specification on the current molten pool morphology. Considering the fast heat dissipation of aluminum alloy, this paper considers the current value of the welding specification and the previous three historical values ​​of the specification in the model input. The welding specifications used here include the pulse peak current and the pulse duty cycle (the welding speed and pulse base current are kept constant during the welding process, so they are not considered here). At the same time, the size and morphology of the molten pool also affect the molten pool at the current moment. Therefore, this paper also uses the feature parameters of the molten pool (maximum width of the molten pool, half length of the molten pool) at the previous three moments as the input of the network. In summary, the model in this paper has a total of 14 input parameters. The output of the model is the current value of the molten pool characteristic parameters (including the maximum width and half length of the molten pool), with a total of 2 outputs. The neural network model for the frontal parameter of the pulsed GTAW molten pool is shown in Figure 3. Regarding the selection of the number of nodes in the intermediate layer of the BP neural network, different literature has different selection principles. In this paper, the hidden layer is set to one layer because a three-layer BP network can complete any N-dimensional to M-dimensional mapping. Based on experience, 19 hidden layer units are determined to be the most suitable. Figure 3 Neural network model for the frontal parameter of the aluminum alloy pulsed GTAW molten pool 4 Design of PID controller PID control is one of the most widely used control laws in process control. PID control is a combination of proportional (P), integral (I), and derivative (D) control actions. It is a transformation of the output control quantity to apply to the error. This paper selects the incremental algorithm. In order to improve the accuracy of calculation, this paper uses trapezoidal integral instead of rectangular integral and four-point difference instead of single-point difference to reduce data error and noise. The improved incremental PID algorithm is as shown in equation (1) [8]: where Kp is the proportional coefficient, Ti is the integral time constant, Td is the derivative time constant, and T is the sampling period. It can be seen that to realize PID control, the three parameters Kp, Ti, and Td must be calculated, that is, the tuning of the control parameters. Using this PID controller, a single variable control experiment was carried out on two types of workpieces under variable heat dissipation conditions. The material used was LF6, the workpiece thickness was 2.5mm, the joint type was butt joint, and the pulse non-wire feeding method was adopted [9]. The peak welding current was used as the control quantity, the weld width on the front side was used as the controlled quantity, and the ideal weld width was given as 8mm. For trapezoidal workpieces, the control parameters were simulated parameters Kp=17.45, Ti=0.75, Td=0.605; for dumbbell-shaped workpieces, the control parameters were Kp=22.45, Ti=0.585, Td=0.795. In the control experiment, to ensure good arc formation during the arc initiation stage, the welding speed was applied again after a 3-second pause following arc initiation and penetration. The first ten pulses used constant welding parameters. Furthermore, to limit overshoot, the peak current adjustment range was limited to 160–180 A. The minimum current adjustment unit was set to 1 A. Figures 4 and 5 show the front and back photographs of the welded workpieces obtained by PID control on trapezoidal and rectangular workpieces with varying heat dissipation conditions. Figure 6 shows the PID control process curve. The maximum absolute error of the actual welded front width of the trapezoidal workpiece was 0.781 mm, the average error was 0.052 mm, and the root mean square error was 0.01822 mm. The maximum absolute error of the actual welded front width of the dumbbell-shaped workpiece was 0.935 mm, the average error was 0.0135 mm, and the root mean square error was 0.0226 mm. Figure 4. Front and back photos of trapezoidal workpiece welded under PID control. Figure 5. Photo of dumbbell-shaped workpiece welded under PID control. [align=center](a) Trapezoid workpiece (b) Dumbbell-shaped workpiece Figure 6. Control curve of PID control welding process for workpiece with variable heat dissipation[/align] 5. Conclusion The experimental results show that the single-variable PID control method has the characteristic of fast adjustment speed, but the overshoot of the system output is large, which is not conducive to system stability. As can be seen from Figure 6(a), the overall trend of the control quantity of the trapezoidal workpiece decreases as the heat dissipation conditions worsen. The back weld width basically fluctuates around 8mm, but as the heat dissipation conditions gradually worsen, the fluctuation of the back weld width also gradually intensifies, and the steady-state error also gradually increases. For the dumbbell-shaped workpiece, the front weld width value fluctuates around the given ideal weld width of 8mm. It basically remains at around 8mm. There is a slight bulge in the middle of the weld width, which is caused by the sudden deterioration of heat dissipation conditions. At the same time, the control quantity is also rapidly reduced to maintain the balance of heat input and output and keep the front weld width stable and consistent. In summary, the PID control of a single variable still has the characteristics of large overshoot and large steady-state error. References: [1] Chen Shanben, Wu Lin. Overview of Research and Application of Robot Welding Technology in my country. Proceedings of the 8th National Welding Conference (Volume 1), Beijing, 1997 [2] Kovacevic R, Zhang YM Vision Sensing of 3D Weld Pool Surface. Proceedings of the 4th International Conference on Trends in Welding Research, Gatlinburg, Tennessee, USA, 5-8, June, 1995 [3] Zhang Y.M., Li L. and Kovacevic R. Monitoring of Weld Pool Appearance for Penetration Control. Proceedings of the 4th International Conference on Trends in Welding Research, Gatlinburg, Tennessee, USA, 5-8, June, 1995 [4] Kovacevic R., Zhang YM and Beardsley H. On-line Sensing of Metal Transfer for Adaptive Control of GMA Welding. Proceedings of the 4th International Conference on Trends in Welding Research, Gatlinburg, Tennessee, USA, 5-8, June, 1995 [5] Chen Wenjie. Research on the development of a locally autonomous intelligent welding robot system and its remote control method. Doctoral dissertation of Shanghai Jiaotong University. 2004.11 [6] Zhou Lv, Lin Tao, Chen Shanben et al., Intelligent arc welding robot system based on servo binocular vision, Materials Science and Technology, 2004, Vol.12 Supplement p59-61 [7] Zhou Lv, Lin Tao, Chen Shanben et al., Servo binocular vision sensor for welding robots, China, Invention Patent: Application No.: 200410067328.1, 2004 [8] Zhang Yuming, Wu Lin. Modern control analysis and design of welding process. Harbin Institute of Technology Press. 1991. [9] Yu Shangzhi, Handbook for Welding Process Personnel, Shanghai, Shanghai Science and Technology Press, 1991 (end)