Intelligent welding robot

Abstract

The invention relates to an intelligent welding robot which comprises a robot body, a quick-change type modular welding gun assembly, an annular array sensing module, an active cooling device and an edge calculation unit. Wherein the quick-change type modular welding gun assembly is installed at the output end of the robot body and used for conducting multi-angle welding operation on a workpiece. The annular array sensing module comprises six sets of 3D line laser sensors and is used for obtaining three-dimensional point cloud data of a workpiece. The active cooling device is integrated in a joint of the robot body and used for controlling the temperature of a joint motor. And the edge calculation unit is used for rapidly processing the collected multi-modal data information, dynamically planning a welding path, adjusting welding process parameters on line by adopting a Bayesian optimization framework, and controlling the quick-change modular welding gun assembly to weld workpieces in a welding area in real time. The welding device has the effect of effectively improving the welding quality and the welding efficiency.

Description

Technical Field

[0001] The present application relates to the field of intelligent welding, and in particular to an intelligent welding robot. Background Art

[0002] Welding robots are widely used in manufacturing and other fields because of their advantages such as high welding efficiency, good welding quality and the ability to continue operating in harmful environments. With the development of science and technology, welding robots are also developing towards high intelligence and high quality.

[0003] The Chinese patent application with publication number CN116810230A discloses an artificial intelligence adaptive robot welding method, which includes the following steps: a 3D structured light camera takes a picture of the product to generate a three-dimensional point cloud; deep learning artificial intelligence software identifies the product weld and extracts the weld position coordinates and the pipe lap method; the robot path autonomous planning software generates the coordinates of the shooting point with precise positioning according to the identified weld position; generates the motion planning path required for the robot to take pictures and guides the robot to take pictures required for precise positioning; the precision positioning camera accurately repositions the product weld and feeds back the photographed weld parameters to the artificial intelligence software to identify the weld type; generates a welding program and a robot motion path program, and calls the corresponding preset welding parameters in the welding process package. This application mainly relies on deep learning artificial intelligence software for weld identification and robot path autonomous planning software to generate a path. However, its welding path is rigid, and it only calls the preset welding parameters according to the weld type, lacking the ability to optimize and adaptively adjust the parameters in real time; only uses a 3D structured light camera to obtain the product three-dimensional point cloud data for analysis, and the data source is single, resulting in poor welding accuracy and stability.

[0004] With respect to the above-mentioned related technologies, the inventors believe that the welding routes of existing welding robots are rigid and can only be welded according to preset welding parameters. There is a lack of real-time monitoring and adjustment of the welding process, resulting in poor welding accuracy and stability. Summary of the invention

[0005] In order to solve the above problems, the present application provides an intelligent welding robot.

[0006] In a first aspect, the present application provides an intelligent welding robot, which adopts the following technical solution:

[0007] An intelligent welding robot, comprising: a robot body and,

[0008] The quick-change modular welding gun assembly is installed at the output end of the robot body and is used for multi-angle welding operations on the workpiece. It includes a gas-electric hybrid quick interface for realizing synchronous and rapid connection of shielding gas, cooling water, and welding current, an embedded micro-vision unit for capturing the dynamic characteristics of the molten pool, and an electromagnetic locking mechanism for quick disassembly and locking of welding gun modules of different specifications;

[0009] The ring array sensor module includes six groups of 3D line laser sensors arranged in a 120° ring layout to obtain three-dimensional point cloud data of the workpiece;

[0010] Active cooling device, integrated inside the joints of the robot body, used to control the temperature of the joint motors;

[0011] The edge computing unit is used to quickly process the collected multimodal data information, dynamically plan the welding path, use the Bayesian optimization framework to adjust the welding process parameters online, and control the quick-change modular welding gun assembly in real time to weld the workpiece in the welding area.

[0012] Preferably, the quick-change modular welding gun assembly further comprises:

[0013] Wear detection unit, used to measure the impedance of the tungsten tip of the welding gun module and monitor the tungsten wear;

[0014] The six-dimensional force sensor unit is used to monitor the force status data of the welding gun end of the welding gun module in real time and generate force data.

[0015] Preferably, the active cooling device comprises:

[0016] The phase change material layer is arranged on the inner side of the joint motor housing, with a thickness of 1.5 mm, and is used to absorb the instantaneous heat load when the motor is running;

[0017] The microchannel structure is arranged outside the joint motor housing, and the flow channel is arranged along the axial direction, covering the entire length of the motor, and is used to remove the continuous heat load through the circulation of coolant;

[0018] Thermoelectric cooling sheets are arranged on the outside of the stator core of the joint motor and on the surface of the motor drive board, and are used to actively cool the high-temperature area to compensate for the heat dissipation capacity of the phase change material and the microchannel;

[0019] Infrared monitoring unit, used to monitor the motor surface temperature field in real time.

[0020] Preferably, the edge computing unit quickly processes the collected multimodal data information, dynamically plans the welding path based on the TD3 reinforcement learning algorithm, uses the Bayesian optimization framework to adjust the welding process parameters online, and controls the quick-change modular welding gun assembly in real time to weld the workpiece in the welding area. Specifically, the following steps are included:

[0021] The multimodal data information of the welding environment is obtained through the annular array sensor module and the six-dimensional force sensor unit, wherein the multimodal data information includes visual data information and force data information, wherein the visual data information is the three-dimensional point cloud data of the workpiece, and the force data information is the force state data of the welding gun end of the welding gun module;

[0022] The three-dimensional point cloud data of the workpiece is subjected to feature extraction and semantic segmentation through a preset PointNet++ network, and the weld trajectory features are identified and obtained. The weld trajectory features are input into a preset deep reinforcement learning model for dynamic planning to generate a dynamic path that meets the welding process requirements; the deep reinforcement learning model is obtained by iterative training based on a proximal strategy optimization algorithm;

[0023] A Bayesian optimization model between welding current, voltage, speed and weld quality was constructed, a parameter response surface was constructed through Gaussian process regression, and multi-objective parameter optimization was performed using the expected improvement acquisition function to determine the best welding parameter combination;

[0024] Control the quick-change modular welding gun assembly to start welding according to the dynamic path and the best welding parameter combination. The embedded micro-vision unit collects the molten pool image and extracts the dynamic characteristics of the molten pool. According to the results of the molten pool dynamic feature extraction, the fuzzy PID controller is used to dynamically adjust the welding parameters. The proportional coefficient, integral time, and differential time are adaptively adjusted according to the fluctuation rate of the molten width.

[0025] The physical welding process is mapped to the virtual space through a pre-set digital twin system, path pre-verification and collision detection are performed in the simulation environment, and then the optimized parameters are fed back to the robot body and quick-change modular welding gun assembly.

[0026] Preferably, the deep reinforcement learning model adopts a dual-delay deep deterministic policy gradient algorithm architecture, constructs a multidimensional state space S = (weld deviation, welding speed, molten pool morphology, ambient temperature), an action space A = (path correction, wire feeding speed, arc voltage), and sets a composite reward function:

[0027] R=0.6×(1-|ΔW / W0|)+0.3×(1-σpath / σmax)-0.1×(P / Pmax)

[0028] Wherein ΔW is the weld width deviation, W0 is the standard weld width, σpath is the path curvature standard deviation, σmax is the preset maximum value of the path curvature standard deviation; P is the real-time power consumption, and Pmax is the preset maximum allowable power consumption value.

[0029] Preferably, the Bayesian optimization framework includes a dynamic process constraint injection mechanism, which initiates constraint processing when any of the following conditions is detected:

[0030] The splash rate exceeds the preset standard threshold;

[0031] The fluctuation range of the penetration depth is greater than 30% of the preset setting value;

[0032] The temperature gradient in the heat-affected zone exceeds the critical value of material phase change.

[0033] Preferably, the constraints are integrated into the acquisition function by Lagrange multiplier method to generate an optimized parameter combination within the feasible domain.

[0034] Preferably, the multimodal data fusion process comprises the following steps:

[0035] Establish a spatiotemporal alignment model for visual-force data and use an extended Kalman filter to compensate for sensor delay;

[0036] Construct a feature fusion network based on a multi-head attention mechanism to calculate the confidence weight of each sensor data;

[0037] Compare visual data and force data in real time. When the deviation between visual and force data exceeds the threshold, start the pre-set LiDAR-assisted verification.

[0038] The fusion result is output as a seven-dimensional feature vector [position x, y, z, normal vector n_x, n_y, n_z, contact force F].

[0039] Preferably, the edge computing unit controls the quick-change modular welding gun assembly in real time to weld the workpiece in the welding area, further comprising: identifying the welding task to determine whether it is a multi-machine collaborative task, and if so, entering the multi-machine collaborative mechanism control; the multi-machine collaborative mechanism control process comprises the following steps:

[0040] Based on the task allocation mechanism of the pre-set smart contract, the overall welding task is divided into multiple subtasks. Each subtask includes the unique identification ID, location coordinates, process requirements and priority of the subtask;

[0041] Identify the intelligent welding robots that can be assigned, build the equipment capability matrix Cij = [welding accuracy, remaining life, current position] of each intelligent welding robot, dynamically match the intelligent welding robot for each subtask based on the Hungarian algorithm, and control each intelligent welding robot to weld according to the assigned subtask;

[0042] Construct a collision prediction model based on the speed obstacle method to calculate the minimum separation distance of each robot's motion vector within the Δt time window in real time; and adjust the motion trajectory or speed of the robot performing the low-priority subtask when the minimum separation distance is less than the preset safety spacing threshold;

[0043] In the multi-machine collaborative mechanism control process, a lightweight Merkle tree structure is used to store task execution logs.

[0044] Preferably, the intelligent welding robots available for allocation are clearly identified, and the equipment capability matrix C of each intelligent welding robot is constructed. ij = [welding accuracy, remaining life, current position], dynamically matching intelligent welding robots for each subtask based on the Hungarian algorithm specifically includes the following steps:

[0045] Identify the intelligent welding robots that can be allocated and build the equipment capability matrix C of each intelligent welding robot ij =[welding accuracy, remaining life, current position];

[0046] Obtaining the matching weights of each subtask, wherein the matching weights include a welding accuracy weight coefficient W1, a remaining life weight coefficient W2, and a current position weight coefficient W3;

[0047] According to the equipment capability matrix C of the intelligent welding robot ij And the matching weights of each subtask calculate the matching cost M ij , construct the matching cost matrix M; the matching cost calculation formula is: M ij =W1×f1(C ij [Welding accuracy])+W2×f2(C ij [Remaining life])+W3×f3(C ij [Current Position]); f1, f2 and f3 are all mapping functions, which are preset by the management personnel. f1 means that the greater the deviation between the robot welding accuracy and the subtask process requirement accuracy, the higher the cost score; f2 means that the shorter the remaining life of the robot, the higher the cost score; f3 means that the farther the distance between the robot's current position and the task position, the higher the cost score;

[0048] The matching cost matrix M is processed based on the Hungarian algorithm, and the intelligent welding robot is dynamically matched for each subtask through row transformation and column transformation.

[0049] During the welding process, the equipment capability matrix is ​​updated regularly, the matching cost matrix is ​​recalculated, and dynamic matching adjustments are performed based on the Hungarian algorithm.

[0050] In summary, the present application includes at least one of the following beneficial technical effects:

[0051] 1. This application can quickly process the collected multimodal data information through the setting of the edge computing unit. First, the welding path is dynamically planned based on the TD3 reinforcement learning algorithm, and the optimal welding trajectory can be quickly generated according to the real-time welding environment and workpiece status, thereby improving the welding efficiency and quality. Then, the Bayesian optimization framework is used to adjust the welding process parameters online, which can achieve accurate optimization of welding current, voltage, speed and other parameters. Finally, during the welding process, the embedded micro-vision unit is used to collect the molten pool image and extract the dynamic characteristics of the molten pool. The welding parameters are dynamically adjusted according to the extraction results to ensure the stability of the welding process and the consistency of the welding quality, thereby effectively improving the welding quality and welding efficiency.

[0052] 2. The deep reinforcement learning model first uses the PPO algorithm for iterative training, which allows the model to quickly learn basic strategies and knowledge in the initial stage and establish a relatively stable foundation. Then the TD3 algorithm architecture is used to further optimize the performance of the model and make up for the problem of over-estimation of the value function of the deep reinforcement learning model. The model can give full play to the advantages of the two algorithms, learn and make decisions more efficiently in a complex welding environment, and achieve more accurate welding path planning and parameter adjustment, thereby improving welding quality and production efficiency;

[0053] 3. Reasonably divide the overall welding task into multiple subtasks with unique identification ID, position coordinates, process requirements and priorities, build the equipment capability matrix of each intelligent welding robot, comprehensively consider factors such as welding accuracy, remaining life and current position, and perform dynamic matching based on the Hungarian algorithm. It can assign each subtask to the most suitable robot according to the actual capabilities and task requirements of the robot. During the welding process, the robot's status (such as changes in remaining life, position movement, etc.) and task requirements may change dynamically. The Hungarian algorithm can adjust the task allocation in real time according to these changes, so that the system has good dynamic adaptability. Even in complex and changeable welding environments, it can ensure the best match between tasks and robots to ensure the smooth progress of welding work. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 is a system block diagram of the intelligent welding robot in the embodiment of the present application;

[0055] Figure 2 It is a flow chart of a method in which the edge computing unit of the intelligent welding robot in an embodiment of the present application plans a welding path, optimizes parameters and controls the welding gun in real time to perform welding;

[0056] Figure 3 is a flow chart of a method for multimodal data fusion processing in an embodiment of the present application;

[0057] Figure 4is a flow chart of a method for collaborative processing by multiple intelligent welding robots in an embodiment of the present application;

[0058] Figure 5 It is a flow chart of a method for dynamically matching an intelligent welding robot to each subtask in an embodiment of the present application.

[0059] Explanation of the accompanying drawings: 1. Robot body; 2. Quick-change modular welding gun assembly; 21. Gas-electric hybrid quick interface; 22. Embedded micro vision unit; 23. Electromagnetic locking mechanism; 24. Welding gun module; 25. Wear detection unit; 26. Six-dimensional force sensor unit; 3. Ring array sensor module; 4. Active cooling device; 41. Phase change material layer; 42. Microchannel structure; 43. Thermoelectric cooling sheet; 44. Infrared monitoring unit; 5. Edge computing unit. DETAILED DESCRIPTION

[0060] The following is combined with Figure 1-5 This application is described in further detail.

[0061] The present application embodiment discloses an intelligent welding robot. Figure 1 , an intelligent welding robot, comprising a robot body 1, a quick-change modular welding gun assembly 2, a ring array sensor module 3, an active cooling device 4 and an edge computing unit 5. The quick-change modular welding gun assembly 2 is installed at the output end of the robot body 1, and is used to perform multi-angle welding operations on the workpiece. The ring array sensor module 3 includes 6 groups of 3D line laser sensors, which are arranged in a 120° ring layout to obtain three-dimensional point cloud data of the workpiece. The active cooling device 4 is integrated inside the joint of the robot body 1 to control the temperature of the joint motor. The edge computing unit 5 is used to quickly process the collected multimodal data information, dynamically plan the welding path, use the Bayesian optimization framework to adjust the welding process parameters online, and control the quick-change modular welding gun assembly 2 in real time to weld the workpiece in the welding area. By integrating the quick-change modular welding gun assembly 2, multi-angle welding and quick gas and electricity connection are possible, saving gun changing time; the joint motor temperature is controlled by the active cooling device 4 to avoid affecting the performance due to overheating, and improve the robot's continuous endurance. Through the setting of the edge computing unit 5, the collected multimodal data information can be quickly processed, the welding path can be dynamically planned, and the optimal welding trajectory can be quickly generated according to the real-time welding environment and workpiece status, thereby improving the welding efficiency and quality; the Bayesian optimization framework is then used to adjust the welding process parameters online, which can achieve accurate optimization of welding current, voltage, speed and other parameters; finally, during the welding process, the embedded micro vision unit 22 is used to collect the molten pool image and extract the dynamic characteristics of the molten pool, and the welding parameters are dynamically adjusted according to the extraction results to ensure the stability of the welding process and the consistency of the welding quality, thereby effectively improving the welding quality and welding efficiency.

[0062] Reference Figure 1 The quick-change modular welding gun assembly 2 includes a gas-electric hybrid quick interface 21, an embedded micro-vision unit 22, an electromagnetic locking mechanism 23, a welding gun module 24, a wear detection unit 25 and a six-dimensional force sensor unit 26. Among them, the gas-electric hybrid quick interface 21, the embedded micro-vision unit 22, the electromagnetic locking mechanism 23 and the welding gun module 24 are all prior art modules, and their specific structures are not repeated. The gas-electric hybrid quick interface 21 is used to achieve synchronous and rapid connection of protective gas, cooling water, and welding current. In this example, the embedded micro-vision unit 22 adopts a micro-image acquisition device with an integrated 5-megapixel CMOS sensor to capture the dynamic characteristics of the molten pool during welding. The electromagnetic locking mechanism 23 is used to quickly disassemble and lock welding gun modules 24 of different specifications. The wear detection unit 25 is used to measure the impedance of the tungsten tip of the welding gun module 24 and monitor the tungsten loss. During the welding process, the tungsten electrode will gradually wear due to factors such as high temperature and arc erosion, and its loss degree directly affects the welding quality. By real-time monitoring of the change in tungsten electrode impedance, the wear state of the tungsten electrode can be accurately determined, and timely reminders can be given to replace the tungsten electrode to avoid welding defects caused by excessive wear of the tungsten electrode. The six-dimensional force sensor unit 26 is used to monitor the force state data of the welding gun end of the welding gun module 24 in real time and generate force data. During welding operations, the force applied to the end of the welding gun contains a variety of information, such as contact force and arc force during welding. The changes in these forces reflect the stability of the welding state. By analyzing the force data, the welding parameters can be adjusted in real time to ensure a smooth welding process.

[0063] Reference Figure 1, the active cooling device 4 includes a phase change material layer 41, a microchannel structure 42, a thermoelectric cooling sheet 43 and an infrared monitoring unit 44. The phase change material layer 41 is arranged on the inner side of the joint motor housing, with a thickness of 1.5mm, and is used to absorb the instantaneous heat load when the motor is running. The phase change material has the characteristics of changing state at a specific temperature and absorbing or releasing a large amount of heat. In this embodiment, the phase change material layer 41 is composed of a graphene-enhanced octadecane composite material with a phase change temperature of 58±2°C. The microchannel structure 42 is arranged on the outer side of the joint motor housing, and the flow channel is arranged axially, covering the entire length of the motor, and is used to remove the continuous heat load through the circulation of the coolant. In this embodiment, the microchannel structure 42 is a 0.5mm×0.5mm rectangular flow channel, and the controllable range of the coolant flow is 5-50ml / s. The thermoelectric cooling sheet 43 is arranged on the outer side of the stator core of the joint motor and on the surface of the motor drive board, and is used to actively cool the high-temperature area to compensate for the heat dissipation capacity of the phase change material and the microchannel, so as to ensure that the temperature of each part of the shutdown motor is uniform and stable. In this embodiment, the cooling power of the thermoelectric cooling sheet 43 can be adjusted in the range of 10-100W. The infrared monitoring unit 44 is used to monitor the surface temperature field of the motor in real time. In this embodiment, the infrared monitoring unit 44 uses a 640×480 resolution thermal imager with a temperature detection accuracy of ±1°C. Through infrared thermal imaging technology, the temperature distribution of various parts of the motor surface can be quickly and accurately obtained. According to the temperature field information, the working state of the thermoelectric cooling sheet 43 and the circulation speed of the coolant can be adjusted in time to achieve precise control of the motor temperature.

[0064] Reference Figure 2 The above-mentioned edge computing unit quickly processes the collected multimodal data information, dynamically plans the welding path based on the TD3 reinforcement learning algorithm, uses the Bayesian optimization framework to adjust the welding process parameters online, and controls the quick-change modular welding gun assembly in real time to weld the workpiece in the welding area. The specific steps include:

[0065] S1. Multimodal data acquisition: acquiring multimodal data information of the welding environment through the annular array sensor module and the six-dimensional force sensor unit, wherein the multimodal data information includes visual data information and force data information, wherein the visual data information is the three-dimensional point cloud data of the workpiece, and the force data information is the force state data of the welding gun end of the welding gun module;

[0066] S2. Dynamic path generation: feature extraction and semantic segmentation of the three-dimensional point cloud data of the workpiece are performed through a preset PointNet++ network to identify and obtain weld trajectory features, and the weld trajectory features are input into a preset deep reinforcement learning model for dynamic planning to generate a dynamic path that meets the welding process requirements; the deep reinforcement learning model is obtained by iterative training based on a proximal strategy optimization algorithm; the specific training steps of the deep reinforcement learning model that need to be explained are prior art and will not be repeated here;

[0067] PointNet++ is developed on the basis of PointNet; PointNet is a pioneering network for directly processing 3D point cloud data, but it does not take into account the local structural information of point cloud data. PointNet++ can more accurately capture the local and global features of point clouds required for welding operations by introducing a hierarchical feature learning architecture; based on the hierarchical sampling strategy of metric space, it extracts features from local areas at different scales, thereby better processing the irregularity and local changes of point cloud data;

[0068] S3, parameter optimization: construct a Bayesian optimization model between welding current, voltage, speed and weld quality, construct a parameter response surface through Gaussian process regression, and use the expected improvement acquisition function to perform multi-objective parameter optimization to determine the best welding parameter combination;

[0069] Bayesian optimization is a global optimization method based on probability models. It can predict the weld quality under different parameter combinations based on existing welding experimental data or empirical knowledge. Gaussian process regression is a non-parametric probability model that can model the relationship between input variables and output variables. In welding parameter optimization, welding current, voltage, and speed are used as inputs, and weld quality is used as output. Gaussian process regression is used to construct a parameter response surface, which describes the change in weld quality under different welding current, voltage, and speed combinations. The expected improvement acquisition function is used to measure the expected improvement that can be obtained by conducting experiments at a certain parameter point. In Bayesian optimization, our goal is to find a parameter point so that experiments at this point can maximize the understanding of the optimal solution. The expected improvement acquisition function takes into account the current The posterior probability distribution and the known optimal solution are used to calculate the expectation of obtaining a better result than the current optimal solution when conducting experiments at each parameter point; the value of the expected improvement acquisition function is calculated in the parameter space, and the parameter point that maximizes the function value is found, which is used as the next set of parameter combinations to be tested; a new welding experiment is conducted to obtain the weld quality data under this parameter combination and add it to the existing data set; then, the Gaussian process regression model and the posterior probability distribution are updated, the expected improvement acquisition function is calculated again, and the above process is repeated until the stopping condition is met; after multiple iterations of the parameter optimization process, the posterior probability distribution and the expected improvement acquisition function are continuously updated, and finally the welding current, voltage, and speed combination that achieves the optimal weld quality is found; in practical applications, a satisfactory solution can be determined as the optimal parameter combination based on specific welding requirements and quality standards;

[0070] S4, real-time control: control the quick-change modular welding gun assembly to start welding according to the dynamic path and the best welding parameter combination, collect the molten pool image and extract the dynamic characteristics of the molten pool through the embedded micro-vision unit, and dynamically adjust the welding parameters based on the dynamic characteristics extraction results of the molten pool using a fuzzy PID controller, in which the proportional coefficient, integral time, and differential time are adaptively adjusted according to the fluctuation rate of the weld width;

[0071] During the welding process, the embedded micro-vision unit is used to collect the molten pool image in real time; the molten pool is the area formed by the melting of metal during the welding process, and its dynamic characteristics such as shape and size directly reflect the stability of the welding process and the welding quality; by processing and analyzing the molten pool image, the key dynamic characteristics of the molten pool can be extracted, such as the shape, size, temperature distribution and other information of the molten pool; PID (proportional-integral-differential) controller is a common feedback control algorithm; during the welding process, the molten pool dynamic feature extraction results are input into the fuzzy PID controller as feedback information; the controller calculates the error based on the difference between the actual state and the expected state of the molten pool, and uses fuzzy rules to adaptively adjust the proportional coefficient, integral time and differential time; for example, if the molten pool width fluctuates, the fuzzy PID controller will automatically adjust the proportional coefficient, integral time and differential time according to the size and direction of the molten width fluctuation rate, and then adjust the welding parameters (such as current, voltage, etc.) to restore the molten pool state to the expected state as soon as possible;

[0072] S5, virtual-real synchronization: The physical welding process is mapped to the virtual space through the pre-set digital twin system, and the path pre-verification and collision detection are performed in the simulation environment, and then the optimization parameters are fed back to the robot body and the quick-change modular welding gun assembly. Through the above steps, the ring array sensor module and the six-dimensional force sensor unit are used to obtain multimodal data covering vision and force perception. With the help of the PointNet++ network and the deep reinforcement learning model trained based on the proximal strategy optimization algorithm, the weld trajectory is accurately identified and a dynamic path that meets the process requirements is generated. It can adapt to the welding needs of workpieces of various material specifications, and then the Bayesian optimization model is constructed. Through Gaussian process regression and expected improvement acquisition function optimization, the optimal welding parameter combination is quickly determined to improve the welding quality. In the welding process, the embedded micro-vision unit cooperates with the fuzzy PID controller to adaptively adjust the parameters according to the dynamic characteristics of the molten pool and the fluctuation rate of the molten width, realize the dynamic control of welding, and ensure that the welding process is stable, efficient and high-quality. At the same time, based on the digital twin system, the physical and virtual mapping is realized, the path pre-verification and collision detection are performed, and the optimization parameters are fed back to further enhance the welding safety and reliability, so as to effectively improve the welding quality and welding efficiency.

[0073] In addition, the above deep reinforcement learning model adopts a double-delayed deep deterministic policy gradient algorithm architecture to construct a multi-dimensional state space S = (weld deviation, welding speed, molten pool morphology, ambient temperature), an action space A = (path correction, wire feeding speed, arc voltage), and set a composite reward function:

[0074] R=0.6×(1-|ΔW / W0|)+0.3×(1-σpath / σmax)-0.1×(P / Pmax)

[0075] Where ΔW is the weld width deviation, W0 is the standard weld width, σpath is the path curvature standard deviation, σmax is the preset maximum value of the path curvature standard deviation; P is the real-time power consumption, and Pmax is the preset maximum allowable power consumption value. In welding tasks, the environment is complex and dynamically changing. The stability and anti-overestimation ability of the TD3 algorithm can help the model learn the optimal strategy more accurately. It can efficiently optimize the strategy in the continuous action space and is suitable for controlling various continuous action parameters of the welding gun, such as path correction, wire feeding speed and arc voltage. In this application, the deep reinforcement learning model first uses the PPO algorithm for iterative training, which allows the model to quickly learn basic strategies and knowledge in the initial stage and establish a relatively stable foundation. Then, the TD3 algorithm architecture can be used to further optimize the performance of the model and make up for the problem of overestimation of the value function of the deep reinforcement learning model. Through this combination, the model can give full play to the advantages of the two algorithms, learn and make decisions more efficiently in a complex welding environment, and achieve more accurate welding path planning and parameter adjustment, thereby improving welding quality and production efficiency.

[0076] The above Bayesian optimization framework includes a dynamic process constraint injection mechanism, which starts constraint processing when any of the following conditions are detected:

[0077] The splash rate exceeds the preset standard threshold;

[0078] The fluctuation range of the penetration depth is greater than 30% of the preset setting value;

[0079] The temperature gradient of the heat affected zone exceeds the critical value of the material phase change. In addition, the specific starting constraint processing includes: integrating the constraint conditions into the acquisition function through the Lagrange multiplier method to generate the optimal parameter combination within the feasible domain. By setting up a dynamic injection mechanism for process constraints, the constraint processing is started when the spatter rate, the penetration depth fluctuates, and the temperature gradient of the heat affected zone is abnormal, so as to avoid welding defects and improve the stability of welding quality. And by integrating the constraint conditions into the acquisition function through the Lagrange multiplier method, the optimal parameter combination within the feasible domain can be quickly generated, which improves the optimization efficiency and accuracy. At the same time, the Lagrange multiplier method cleverly integrates the process constraints into the acquisition function, so that the optimization process is carried out within the feasible domain that meets the process constraints, avoiding invalid searches in the infeasible parameter space, and significantly improving the efficiency of parameter optimization. For example, when looking for the best combination of welding current, voltage and speed, it is ensured that the parameter combination not only meets the process requirements, but also optimizes the weld quality, reducing the number of unnecessary attempts, saving time and cost.

[0080] Reference Figure 3 , the multimodal data fusion processing of the collected visual data and force data includes the following steps:

[0081] A1. Spatiotemporal alignment and delay compensation: Establish a spatiotemporal alignment model for visual-force data and use an extended Kalman filter to compensate for sensor delay;

[0082] In actual welding scenarios, due to different working principles and data transmission paths, visual sensors (such as ring array sensor modules) and force sensors (six-dimensional force sensor units) have differences in time and space in data collection and transmission. Establishing a spatiotemporal alignment model for visual-force data can accurately match the two types of data in time and space dimensions. The extended Kalman filter is a powerful tool that can accurately estimate and compensate for sensor delays.

[0083] A2. Feature fusion and weight calculation: Construct a feature fusion network based on a multi-head attention mechanism to calculate the confidence weight of each sensor data;

[0084] The feature fusion network based on the multi-head attention mechanism can deeply extract and analyze the features of visual and force data from multiple angles. In welding tasks, visual data contains information such as the shape of the workpiece and the position of the weld, and force data reflects the force of the welding gun. The multi-head attention mechanism can capture the complex associations and dependencies between these data.

[0085] A3. Abnormal data detection and auxiliary verification: Compare visual data and force data in real time. When the deviation between visual and force data exceeds the threshold, start the pre-set LiDAR auxiliary verification.

[0086] When an anomaly is detected, the pre-set LiDAR assisted verification is activated. LiDAR has high-precision three-dimensional measurement capabilities and can provide additional spatial information. By comparing and analyzing LiDAR data with visual and force data, the source and nature of the anomaly can be further determined, improving the accuracy and reliability of anomaly detection.

[0087] A4. Fusion result output: The fusion result is output as a seven-dimensional feature vector [position x, y, z, normal vector n_x, n_y, n_z, contact force F]. Through spatiotemporal alignment, multi-head attention mechanism fusion and lidar-assisted verification, the seven-dimensional feature vector is output, making the fusion of visual and force data more accurate and reliable, providing more precise information for subsequent control.

[0088] Reference Figure 4 The edge computing unit controls the quick-change modular welding gun assembly in real time to weld the workpiece in the welding area, and further includes: identifying and determining whether the welding task is a multi-machine collaborative task, and if so, entering the multi-machine collaborative mechanism control; the multi-machine collaborative mechanism control process includes the following steps:

[0089] B1. Task allocation based on smart contracts: Based on the task allocation mechanism of the pre-set smart contract, the overall welding task is divided into multiple subtasks. Each subtask includes the unique identification ID, location coordinates, process requirements and priority of the subtask;

[0090] B2. Robot matching based on the Hungarian algorithm: Identify the intelligent welding robots that can be assigned and build the equipment capability matrix C of each intelligent welding robot ij =[welding accuracy, remaining life, current position], dynamically matching intelligent welding robots for each subtask based on the Hungarian algorithm, and controlling each intelligent welding robot to weld according to the assigned subtask;

[0091] B3. Robot collision prediction and adjustment: Construct a collision prediction model based on the speed obstacle method to calculate the minimum separation distance of each robot's motion vector within the Δt time window in real time; and adjust the motion trajectory or speed of the robot performing the low-priority subtask when the minimum separation distance is less than the preset safety spacing threshold;

[0092] B4. Task execution log storage: In the multi-machine collaborative mechanism control process, a lightweight Merkel tree structure is used to store task execution logs. Through the above steps, the overall welding task is reasonably divided into multiple subtasks with unique identification ID, position coordinates, process requirements and priorities, and the equipment capability matrix of each intelligent welding robot is constructed. Taking into account factors such as welding accuracy, remaining life and current position, dynamic matching is performed based on the Hungarian algorithm. Each subtask can be assigned to the most suitable robot according to the actual capabilities and task requirements of the robot. In the welding process, the robot's state (such as remaining life change, position movement, etc.) and task requirements may change dynamically. The Hungarian algorithm can adjust the task allocation in real time according to these changes, so that the system has good dynamic adaptability. Even in a complex and changeable welding environment, it can ensure the best match between tasks and robots and ensure the smooth progress of welding work. In addition, a collision prediction model based on the speed barrier method is constructed during the welding process to monitor the minimum interval distance of each robot's motion vector in the Δt time window in real time. When the minimum interval distance is less than the preset safety spacing threshold, the motion trajectory or speed of the robot executing the low-priority subtask is adjusted in time to effectively avoid collision accidents between robots.

[0093] Reference Figure 5 , the above clearly defines the intelligent welding robots available for allocation, and constructs the equipment capability matrix C of each intelligent welding robot ij = [welding accuracy, remaining life, current position], dynamically matching intelligent welding robots for each subtask based on the Hungarian algorithm specifically includes the following steps:

[0094] C1. Build equipment capability matrix: clarify the intelligent welding robots that can be allocated, and build the equipment capability matrix C of each intelligent welding robot ij =[welding accuracy, remaining life, current position];

[0095] C2. Determine matching weights: obtain matching weights for each subtask, wherein the matching weights include a welding accuracy weight coefficient W1, a remaining life weight coefficient W2, and a current position weight coefficient W3;

[0096] C3. Calculate the matching cost: According to the equipment capability matrix C of the intelligent welding robot ij And the matching weights of each subtask calculate the matching cost M ij , construct the matching cost matrix M; the matching cost calculation formula is: M ij =W1×f1(C ij [Welding accuracy])+W2×f2(C ij [Remaining life])+W3×f3(C ij[Current position]); where f1, f2, and f3 are all mapping functions preset by the management personnel. The greater the deviation between the robot welding accuracy and the process requirement accuracy of the subtask, the higher the cost score; the shorter the remaining life of the robot, the higher the cost score; the farther the current position of the robot is from the task position, the higher the cost score;

[0097] C4. Hungarian algorithm matching: Process the matching cost matrix M based on the Hungarian algorithm, and dynamically match intelligent welding robots for each subtask through row transformation and column transformation;

[0098] C5. Dynamic update and adjustment: Regularly update the equipment capacity matrix during the welding process, recalculate the matching cost matrix, and perform dynamic matching adjustment based on the Hungarian algorithm. By clarifying the three key dimensions of welding accuracy, remaining life, and current position, comprehensively quantify the equipment capabilities of each intelligent welding robot, then determine the weight coefficients (W1, W2, W3) of welding accuracy, remaining life, and current position respectively according to the characteristics of each subtask, calculate the matching cost between each robot and each subtask, and finally, based on the optimal solution under the current cost matrix of the Hungarian algorithm, the overall matching cost can be minimized to make the resources most reasonably configured, thereby improving the performance and efficiency of the entire multi-robot collaborative welding system.

[0099] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the protection scope of the invention. Obviously, the described embodiments are only partial embodiments of the present invention, rather than all embodiments. Based on these embodiments, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art can still, without conflict, make combinations, additions, deletions, or other adjustments to the features in each embodiment of the present invention according to the situation without creative efforts, so as to obtain different technical solutions that essentially do not deviate from the concept of the present invention, and these technical solutions also belong to the scope of protection of the present invention.

Claims

1. An intelligent welding robot, characterized in that: include: A robot body (1) and A quick-change modular welding gun assembly (2) is installed at the output end of the robot body (1) and is used to perform multi-angle welding operations on a workpiece, comprising a gas-electric hybrid quick interface (21) for realizing synchronous and rapid connection of shielding gas, cooling water, and welding current, an embedded micro-vision unit (22) for capturing dynamic characteristics of a molten pool, and an electromagnetic locking mechanism (23) for quickly disassembling and locking welding gun modules (24) of different specifications; The annular array sensor module (3) includes six groups of 3D line laser sensors arranged in a 120° annular layout for acquiring three-dimensional point cloud data of the workpiece; An active cooling device (4) is integrated inside the joint of the robot body (1) and is used to control the temperature of the joint motor; The edge computing unit (5) is used to quickly process the collected multimodal data information, dynamically plan the welding path, use the Bayesian optimization framework to adjust the welding process parameters online, and control the quick-change modular welding gun assembly (2) in real time to weld the workpiece in the welding area.

2. The intelligent welding robot according to claim 1, characterized in that: The quick-change modular welding gun assembly (2) further comprises: A wear detection unit (25) is used to measure the impedance of the tungsten electrode tip of the welding gun module (24) and monitor the tungsten electrode loss; The six-dimensional force sensor unit (26) is used to monitor the force state data of the welding gun end of the welding gun module (24) in real time and generate force sense data.

3. The intelligent welding robot according to claim 1, characterized in that: The active cooling device (4) comprises: The phase change material layer (41) is arranged on the inner side of the joint motor housing and has a thickness of 1.5 mm, and is used to absorb the instantaneous heat load when the motor is running; The microchannel structure (42) is arranged outside the joint motor housing, and the flow channel is arranged along the axial direction, covering the entire length of the motor, and is used to remove the continuous heat load through the circulation of the coolant; Thermoelectric cooling sheets (43) are arranged on the outside of the stator core of the joint motor and on the surface of the motor drive board, and are used to actively cool the high-temperature area to compensate for the heat dissipation capacity of the phase change material and the microchannel; The infrared monitoring unit (44) is used to monitor the motor surface temperature field in real time.

4. The intelligent welding robot according to claim 2, characterized in that: The edge computing unit (5) quickly processes the collected multimodal data information, dynamically plans the welding path based on the TD3 reinforcement learning algorithm, uses the Bayesian optimization framework to adjust the welding process parameters online, and controls the quick-change modular welding gun assembly in real time to weld the workpiece in the welding area. Specifically, the following steps are included: The multimodal data information of the welding environment is obtained through the annular array sensor module and the six-dimensional force sensor unit, wherein the multimodal data information includes visual data information and force data information, wherein the visual data information is the three-dimensional point cloud data of the workpiece, and the force data information is the force state data of the welding gun end of the welding gun module; The three-dimensional point cloud data of the workpiece is subjected to feature extraction and semantic segmentation through a preset PointNet++ network, and the weld trajectory features are identified and obtained. The weld trajectory features are input into a preset deep reinforcement learning model for dynamic planning to generate a dynamic path that meets the welding process requirements; the deep reinforcement learning model is obtained by iterative training based on a proximal strategy optimization algorithm; A Bayesian optimization model between welding current, voltage, speed and weld quality was constructed, a parameter response surface was constructed through Gaussian process regression, and multi-objective parameter optimization was performed using the expected improvement acquisition function to determine the best welding parameter combination; Control the quick-change modular welding gun assembly to start welding according to the dynamic path and the best welding parameter combination. The embedded micro-vision unit collects the molten pool image and extracts the dynamic characteristics of the molten pool. According to the results of the molten pool dynamic feature extraction, the fuzzy PID controller is used to dynamically adjust the welding parameters. The proportional coefficient, integral time, and differential time are adaptively adjusted according to the fluctuation rate of the molten width. The physical welding process is mapped to the virtual space through a pre-set digital twin system, path pre-verification and collision detection are performed in the simulation environment, and then the optimized parameters are fed back to the robot body and quick-change modular welding gun assembly.

5. The intelligent welding robot according to claim 4, characterized in that: The deep reinforcement learning model adopts a dual-delay deep deterministic policy gradient algorithm architecture to construct a multi-dimensional state space S = (weld deviation, welding speed, molten pool morphology, ambient temperature), an action space A = (path correction, wire feeding speed, arc voltage), and set a composite reward function: R=0.6×(1-|ΔW / W0|)+0.3×(1-σpath / σmax)-0.1×(P / Pmax) Wherein ΔW is the weld width deviation, W0 is the standard weld width, σpath is the path curvature standard deviation, σmax is the preset maximum value of the path curvature standard deviation; P is the real-time power consumption, and Pmax is the preset maximum allowable power consumption value.

6. The intelligent welding robot according to claim 4, characterized in that: The Bayesian optimization framework includes a dynamic process constraint injection mechanism that initiates constraint processing when any of the following conditions are detected: The splash rate exceeds the preset standard threshold; The fluctuation range of the penetration depth is greater than 30% of the preset setting value; The temperature gradient in the heat-affected zone exceeds the critical value of material phase change.

7. The intelligent welding robot according to claim 6, characterized in that: The startup constraint processing specifically includes: integrating the constraint conditions into the acquisition function through the Lagrange multiplier method to generate an optimization parameter combination in the feasible domain.

8. The intelligent welding robot according to claim 4, characterized in that: The multimodal data fusion process comprises the following steps: Establish a spatiotemporal alignment model for visual-force data and use an extended Kalman filter to compensate for sensor delay; Construct a feature fusion network based on a multi-head attention mechanism to calculate the confidence weight of each sensor data; Compare visual data and force data in real time. When the deviation between visual and force data exceeds the threshold, start the pre-set LiDAR-assisted verification. The fusion result is output as a seven-dimensional feature vector [position x, y, z, normal vector n_x, n_y, n_z, contact force F].

9. The intelligent welding robot according to claim 1, characterized in that: The edge computing unit controls the quick-change modular welding gun assembly in real time to weld the workpiece in the welding area, and further includes: identifying and determining whether the welding task is a multi-machine collaborative task, and if so, entering the multi-machine collaborative mechanism control; the multi-machine collaborative mechanism control process includes the following steps: Based on the task allocation mechanism of the pre-set smart contract, the overall welding task is divided into multiple subtasks. Each subtask includes the unique identification ID, location coordinates, process requirements and priority of the subtask; Identify the intelligent welding robots that can be allocated and build the equipment capability matrix C of each intelligent welding robot ij =[welding accuracy, remaining life, current position], dynamically matching intelligent welding robots for each subtask based on the Hungarian algorithm, and controlling each intelligent welding robot to weld according to the assigned subtask; Construct a collision prediction model based on the speed obstacle method to calculate the minimum separation distance of each robot's motion vector within the Δt time window in real time; and adjust the motion trajectory or speed of the robot performing the low-priority subtask when the minimum separation distance is less than the preset safety spacing threshold; In the multi-machine collaborative mechanism control process, a lightweight Merkle tree structure is used to store task execution logs.

10. The intelligent welding robot according to claim 9, characterized in that: The intelligent welding robots that can be allocated are clearly identified, and the equipment capability matrix C of each intelligent welding robot is constructed. ij = [welding accuracy, remaining life, current position], dynamically matching intelligent welding robots for each subtask based on the Hungarian algorithm specifically includes the following steps: Identify the intelligent welding robots that can be allocated and build the equipment capability matrix C of each intelligent welding robot ij =[welding accuracy, remaining life, current position]; Obtaining the matching weights of each subtask, wherein the matching weights include a welding accuracy weight coefficient W1, a remaining life weight coefficient W2, and a current position weight coefficient W3; According to the equipment capability matrix C of the intelligent welding robot ij And the matching weights of each subtask calculate the matching cost M ij , construct the matching cost matrix M; the matching cost calculation formula is: M ij =W1×f1(C ij [Welding accuracy])+W2×f2(C ij [Remaining life])+W3×f3(C ij [Current Position]); f1, f2 and f3 are all mapping functions, which are preset by the management personnel. f1 means that the greater the deviation between the robot welding accuracy and the subtask process requirement accuracy, the higher the cost score; f2 means that the shorter the remaining life of the robot, the higher the cost score; f3 means that the farther the distance between the robot's current position and the task position, the higher the cost score; The matching cost matrix M is processed based on the Hungarian algorithm, and the intelligent welding robot is dynamically matched for each subtask through row transformation and column transformation. During the welding process, the equipment capability matrix is ​​updated regularly, the matching cost matrix is ​​recalculated, and dynamic matching adjustments are performed based on the Hungarian algorithm.

Images