Robot self-learning intelligent welding system and method based on multi-machine cooperation

Abstract

The invention discloses a robot self-learning intelligent welding system and method based on multi-robot collaboration, and the system is partially based on a multi-robot collaboration architecture and integrates three core hardware modules of a visual large line scanning camera, a welding seam locating and tracking device and an intelligent control cabinet. The visual large line scanning camera achieves workpiece coarse positioning and initial environment modeling through large-range scanning, the tail end welding seam locating and tracking device adopts a high-precision visual sensor to correct welding path deviation in real time, and the visual large line scanning camera and the tail end welding seam locating and tracking device cooperate to complete welding seam feature extraction. The intelligent control cabinet is internally provided with multiple modules for cooperative work, and the self-learning module integrates a reinforcement learning algorithm and supports dynamic path optimization and collision avoidance decision. The method takes reinforcement learning as a core, and is divided into three stages of environment construction, dynamic optimization and task execution. Through dynamic task partitioning, a workload balancing algorithm and a multi-target reward mechanism, the efficiency and safety of multi-robot collaborative welding are effectively improved, and the method is suitable for intelligent welding production under complex working conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial robot control and intelligent welding, and specifically to a robot self-learning intelligent welding system and method based on multi-robot cooperation, which is applicable to the automated welding operations of quay crane tie rod systems, lifting lug heavy plates and other non-standard components in fields such as shipbuilding and heavy machinery. Background Art

[0002] For traditional teaching programming welding of some welding tasks with complex spatial curves or free-form surfaces, it is very difficult to accurately describe and control the motion trajectory of the welding robot by traditional teaching programming. Operators need to spend a lot of time trying and adjusting to make the welding path as close to the ideal state as possible, but it is often still difficult to achieve the best effect. And during the welding process, if there are situations such as workpiece position deviation and real-time adjustment of welding parameters, the welding robot with traditional teaching programming usually cannot adapt automatically, and it is necessary to manually pause the welding process, re-perform teaching programming or manually adjust the parameters, which will affect the continuity of welding and production efficiency.

[0003] The above methods often require specific system models, and reinforcement learning aims to let the intelligent agent explore and learn a specific unknown environment, and guide the intelligent agent to learn to make decisions and perform corresponding optimal actions for specific states through methods such as reward function setting and value function updating, so as to obtain the maximum reward. Then, the welding tasks are divided to clarify the working range of the robot, and dynamic task partitioning is performed based on distance, and real-time monitoring and communication are carried out to achieve cooperation. A target function is constructed to ensure workload balance and give consideration to high-priority welds; a composite reward function is constructed and smoothed, taking multiple objectives into comprehensive consideration. The robot calculates the coordinates of key nodes using the kinematic model and obtains the state information through sensors. Finally, the welding behavior is evaluated according to the reward function, the reward value is quantified, and the robot is guided to learn and optimize the welding strategy, improve the welding quality and safety, and achieve efficient cooperative welding operations.

[0004] The above methods often require specific system models, while reinforcement learning aims to enable the intelligent agent to explore and learn a specific unknown environment, and guide the intelligent agent to learn to make decisions and perform corresponding optimal actions for specific states through methods such as reward function setting and value function updating, so as to obtain the maximum reward. Summary of the Invention

[0005] Object of the Invention: Aiming at the above problems, the present invention provides a robot self-learning intelligent welding system and method based on multi-robot cooperation, which improve welding safety and welding efficiency. Technical Solution: To solve the above problems, the present invention adopts a robot self-learning intelligent welding equipment system based on multi-robot cooperation, and the system includes functions such as multi-robot cooperation, intelligent vision recognition, autonomous learning, adaptive obstacle avoidance, automatic tracking, and automatic control.

[0006] The specific solutions are as follows:

[0007] A robot self-learning intelligent welding system based on multi-robot collaboration, based on a multi-robot collaboration architecture, includes a vision large line scan camera, a weld seam seeking and tracking device, and an intelligent control cabinet. The multi-robot collaboration architecture includes two robots that can move along the gantry beam track, and the end of the robot is integrated with a welding torch and a weld seam seeking and tracking device. The weld seam seeking and tracking device is used to accurately locate the weld seam and correct the welding angle deviation in real time. A vision large line scan camera is installed on the gantry to be responsible for the scanning area, rough position the workpiece, quickly give the approximate position, and provide initial information for subsequent operations. The weld seam seeking and tracking device and the vision large line scan camera cooperate to determine the welding scene, the shape of the workpiece, and the position of the weld bead. An intelligent control cabinet is also installed above the gantry, which integrates: a control system module, an integrated control algorithm and self-learning module, an autonomous decision-making module, a data collection and storage module, a human-machine interaction module (used to realize information interaction between the system and the operator), a communication module, and a power management module (used to supply power to the overall system). Among them, the integrated control algorithm and self-learning module are used to realize dynamic path optimization and collision avoidance decision-making for welding. It optimizes and controls the following series of operations in the multi-robot collaborative self-learning welding operation: constructs a reinforcement learning environment, collects and processes welding data, and a multi-modal fusion deep learning model performs path planning, image annotation, data enhancement, and normalization processing on the weld seam, divides tasks, constructs a task assignment deviation minimization function, calculates the composite reward value, so as to achieve efficient and accurate welding operations.

[0008] The specific functions of the robot self-learning intelligent welding system under multi-robot collaboration are as follows: automatically collect the image information of the workpiece to be welded, and preprocess the collected workpiece images. The system divides the collected weld seams according to the welding process and production requirements. Divide the task priorities and the working ranges of the robots according to the characteristics of the welding tasks, monitor the status and progress in real time, and dynamically adjust the tasks through communication to ensure collaborative operations. The monitoring system will evaluate the workloads of the two robots and reasonably allocate the task amounts. Finally, evaluate the tasks of the robots through the set composite reward function, so that the robots can get feedback during the task process.

[0009] A robot self-learning intelligent welding method based on multi-robot collaboration includes three stages: environment construction, dynamic optimization, and task execution; specifically includes the following steps:

[0010] Step (a): Construct a reinforcement learning environment based on dual-robot collaborative welding;

[0011] Step (b): Collect welding data, make autonomous decisions on multiple weld seams, and preprocess the data in the environment built in step (a);

[0012] Step (c): Divide the weld priorities and welding difficulties according to production requirements and welding processes, and at the same time clarify the working ranges of the two robots;

[0013] Step (d): Build a dynamic task partitioning and monitoring system according to the classification of welding tasks in step (c);

[0014] Step (e): The monitoring system ensures an even distribution of the workloads of the two robots through a task allocation deviation minimization function;

[0015] Step (f): Build a variety of reward functions Y according to the completion of the robot's full scan, welding tasks, and process obstacle avoidance;

[0016] Step (h): Build a composite reward function through the construction of the full scan accuracy, welding task, and process obstacle avoidance reward values, calculate the reward values, and then guide the robot to learn better welding strategies.

[0017] Furthermore, the specific steps of step (a) are as follows:

[0018] Step (a1): Simulate and depict the space of the simulation welding environment based on the python software;

[0019] Step (a2): Simulate and add the welds, obstacles, and robots to the simulation welding environment.

[0020] Furthermore, the specific steps of step (b) are as follows:

[0021] Step (b1): In the environment constructed in step (a), through the cooperation of the visual large line scan camera (1) and the weld seam positioning and tracking device (7), the data collection and storage module collects and stores the images of the overall production line layout and the workpieces to be welded;

[0022] Step (b2): The autonomous decision-making module calculates the number of weld beads at the weld seam of the collected workpiece image and reasonably arranges the weld bead distribution;

[0023] Step (b3): Label the images collected in step (b1), and the labeling content includes the positions of single-layer welds and the positions of each weld bead in each layer of multi-layer multi-pass welds;

[0024] Step (b4): Enhance the original images by rotating, scaling, cropping, adding noise, etc., expand the data set, and improve the generalization ability of the model;

[0025] Step (b5): Normalize the pixel values of the images to the range of [0, 1] or [-1, 1].

[0026] Furthermore, the specific steps of step (c) are as follows:

[0027] Step (c1): Mark the length l for each weld in the data collection and storage module. i ;

[0028] Step (c2): The integrated control algorithm module divides the weld priority p i and the welding difficulty d i according to the welding process and production requirements.

[0029] Step (c3): The autonomous decision-making module combines the stored image data of the production line layout to divide the working ranges of the two robots R1 and R2, and at the same time counts the set of welds covered within the working ranges as W n .

[0030] Furthermore, the step (d) specifically includes the following steps:

[0031] Step (d1): The total workload H j (j = 1, 2) of the robot R j is the sum of the allocated weld workloads h i , and the workload h i of weld i can be expressed as:

[0032] h i = αl i + βd i

[0033] where α and β are weight coefficients, and α + β = 1; then:

[0034]

[0035] Step (d2): The two robots exchange their respective status and task progress information through the communication module; the robot R1 sends its status s1(t) and the information of the completed welds to the robot R2, and through this communication, the dynamic adjustment and collaborative work of the tasks are realized;

[0036] Step (d3): The control system module monitors the status of the robots and the completion of the welds in real time; c i (t) is the completion status of weld i at time t, c i (t) = 1 indicates completed, c i (t) = 0 indicates not completed.

[0037] Furthermore, the step (e) specifically includes the following steps:

[0038] Step (e1): The purpose of the monitoring system is to balance the time for the two robots to complete their respective tasks, and at the same time give priority to processing high-priority welds; the objective function is defined as:

[0039]

[0040] H1 and H2 respectively represent the total workload assigned to robots R1 and R2, that is h i is the workload size of each small task.

[0041] Furthermore, the step (f) specifically includes the following steps:

[0042] Step (f1): Comprehensively considering full - scale scanning, welding tasks, and process obstacle avoidance, assign reasonable weights to the completion of each task to construct various types of reward functions;

[0043] Step (f2): In order to avoid drastic fluctuations in the reward function in step (f1), use the reward function smoothing algorithm to improve the stability of the algorithm.

[0044] Furthermore, the step (h) specifically includes the following steps:

[0045] Step (h1): Based on indicators such as its full - scale scanning accuracy, welding tasks, and collision situations, achieve precise measurement of welding behavior. Calculate the composite reward value, and then guide the robot to learn better welding strategies.

[0046] Furthermore, the gantry - type dual - robot collaborative system integrates visual large - line scanning and weld tracking to achieve millimeter - level positioning correction. The intelligent central control unit fuses multiple sensors to generate the optimal program, dynamically allocates tasks in real - time, collaborates across regions to balance the workload, breaks through the limitations of traditional single robots, and improves welding efficiency.

[0047] The beneficial effects of the present invention are as follows: It has the following four significant advantages:

[0048] 1. Equipment integration and collaboration advantages: The system consists of core modules such as a visual large - line scanning camera (1), a weld seam seeking and tracking device (7), and auxiliary modules such as a digital inverter welder (3). Each module works collaboratively. The visual large - line scanning camera is responsible for rough positioning of the workpiece, the weld seam seeking and tracking device accurately locates and corrects deviations, the intelligent welding monitor collects parameters for auxiliary decision - making, the intelligent robot automatically plans the path based on visual recognition, and the intelligent control cabinet integrates control and self - learning modules to optimize each link. This integrated collaborative mode effectively improves welding quality and operation safety.

[0049] 2. Advantages of Data-driven Intelligent Decision-making: The visual large-line scanning camera and the weld position searching and tracking device collect welding data, which is processed by the reinforcement learning environment to provide support for decision-making. The system divides task priorities according to data, determines the working range of the robot, and accurately allocates tasks. Combining the robot kinematic model, sensor information, and reward function, it evaluates welding behaviors and optimizes welding strategies. In path planning and collision avoidance processing, the system achieves efficient and accurate control by means of data. The integrated control algorithm and self-learning module of the control cabinet rely on data to optimize motion control, and through real-time communication and dynamic task adjustment, they adapt to complex welding scenarios, improve welding quality and safety, and highlight their key role in the intelligentization of welding equipment.

[0050] 3. Advantages of Reinforcement Learning and Self-learning: Reinforcement learning is used for path planning and collision avoidance in multi-robot collaborative self-learning welding, with obvious advantages. The environment construction adopts a modular design, which is flexible and scalable, and the exception handling mechanism ensures stability. When processing data, it collects welding images under multiple conditions, which are labeled, enhanced, and normalized to improve the generalization and training stability of the model. The task allocation is fine-grained, divided according to task characteristics, the working range of the robot is clarified, dynamic partitioning and monitoring ensure collaboration, the objective function balances the workload, and high-priority welds are processed first. The reward mechanism is scientific. The composite reward function combines multiple objectives and reasonably assigns weights. The smoothing process ensures the stability of the algorithm. The reward value updates the neural network, guiding the robot to optimize strategies, improving welding quality, safety, and operation efficiency.

[0051] 4. Advantages of Task Allocation and Monitoring Optimization: The system is equipped with a dynamic task partitioning and monitoring system. It divides the welding difficulty coefficient according to weld information, clarifies the working range of the robot, and sets priorities for each weld. The monitoring system achieves balanced allocation of the task time of two robots through the objective function, gives priority to processing high-priority welds, and ensures the efficient completion of tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The drawings forming a part of this application are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention. In the drawings:

[0053] Figure 1 It is a schematic diagram of a two-robot arm simulation welding environment based on Python software;

[0054] Figure 2 It is a heavy plate welding scenario;

[0055] Figure 3 It is a simulation of the robot motion range space;

[0056] Figure 4 It is a scenario where two robots detect collisions in the simulation environment;

[0057] Figure 5 Schematic diagram of two intelligent robots simultaneously welding a heavy plate;

[0058] Figure 6 Design flow chart of the robot self - learning intelligent welding equipment system based on multi - machine cooperation of the present invention;

[0059] Figure 7 Multi - robot collaborative welding task processing system and method set;

[0060] Figure 8 Schematic diagram of the three - dimensional structure of the intelligent welding equipment of the present invention;

[0061] Figure 9 Schematic diagram of the three - dimensional structure of the heavy - plate intelligent welding production line of the present invention;

[0062] In the above figures: Gray square: heavy plate; black dot: robot; black line: weld seam; where (P_site_ij, j = 1, 2, 3, 4, 5, 6, 7, 8, 9) is the position coordinate of the j - th grid, and P_site_i5 is the central position coordinate of the robot. Detailed implementation mode

[0063] The present invention will be described in detail below with reference to the accompanying drawings.

[0064] The present invention and its implementation modes are described below. This description is not restrictive, and the actual implementation modes are not limited thereto. Generally speaking, if those of ordinary skill in the art are inspired by it and, without departing from the gist of the present invention, design similar structural modes and embodiments without creative efforts, they shall fall within the protection scope of the present invention. The present invention will be further described in detail below with reference to the accompanying drawings.

[0065] As Figure 1As shown in the figure, the present invention relates to a robot self-learning intelligent welding system based on multi-robot collaboration, which is particularly applicable to path planning and collision avoidance processing in the process of multi-robot collaborative self-learning welding operations. Specifically, the steps of simulating the robot self-learning intelligent welding equipment system mainly based on the Python environment are as follows: Create a Python environment and import necessary libraries such as matplotlib (for plotting) and numpy (for mathematical calculations). Build a continuous simulation learning environment. Use matplotlib to create a 1000×1000 blank space as the basic environment, set the upper limit of the robot's speed, and when the robot moves to the boundary of the environment (the x or y coordinate exceeds the range of 0-1000), reverse its speed to keep it within the simulation environment. Add elements and set interaction rules. Add heavy plates (gray squares) and robots (black dots) to the blank space, set their initial positions with a random number generator, and use matplotlib to draw them. When the robot moves, it generates acceleration according to the external force it receives and moves in combination with the current speed. Simulate the movement process through mathematical formulas, calculate the reward values of the next moment state and action-state combination to achieve self-learning and behavior optimization.

[0066] As Figure 2 In the simulated heavy plate workpiece welding scenario shown, first, the visual large line scan camera (1) scans the production line area to quickly locate the specific position of the plate. Then, the weld seam search and tracking device (7) at the end of the robot accurately searches for the weld seam. Through the collaboration of the two, the shape of the workpiece to be welded and the weld seam position are determined. After the visual scan is completed, the two robots switch to the welding torch state and simultaneously complete the welding task of the plate within their respective working ranges.

[0067] As Figure 3 As shown in the figure, it shows the simulation of the robot's movement range space. In this space, P_site_i5 represents the exact center position of the robot. Within this specified movement range, the robot can efficiently and accurately complete various established tasks. However, when the actual welding area exceeds the pre-set range, the welding robot can move flexibly along the track of the gantry beam. Through this flexible movement method, the area covered by the robot's movement range space can be accurately adjusted according to the actual operation requirements, so as to ensure that the robot can meet the operation requirements under various complex working conditions.

[0068] As Figure 4As shown, it presents a scenario of potential collision between two robots. During the simulation, once the distance between the working ranges of the two robots ≤ 50 cm, the system will prominently prompt "Collision may be detected!" above. This indicates that under the current simulated motion trajectories, the activity spaces of the two robots may overlap, predicting a possible collision. Through such simulations, potential collision risks during the operation of the robots can be detected in advance, which helps to optimize the motion planning and path control of the robots in practical applications, avoid the occurrence of collision accidents in real scenarios, and ensure the safety and efficiency of the robots' work.

[0069] As Figure 5 shown, it presents a schematic diagram of two intelligent robots simultaneously welding a heavy plate. The gray square in the center of the picture represents the heavy plate. The working ranges of the two intelligent robots are presented as black circles, located on the left and right sides of the square respectively, and the circular parts cover the gray square, indicating that the robots are welding the heavy plate. When the system commands the two robots to work together, even if their working areas overlap, the system will not issue an alarm. This not only reflects the full coverage of the working range of the robots on the heavy plate but also represents the spatial relationship between the two during collaborative operation, intuitively showing the working layout of dual robots cooperating to weld the heavy plate in an industrial scenario.

[0070] Figure 6 is the design flow chart of the present invention, which consists of steps P1 - P7, and the descriptions of each step are as follows:

[0071] 1) Step P1

[0072] The first step: Build the environmental foundation

[0073] In the Python environment, use the matplotlib library to create a 1000×1000 two-dimensional blank space as the basis of the simulation environment. This space is equivalent to the abstract plane of the actual working site of the robot, providing a carrier for subsequent element addition and robot motion simulation. Set the speed limit of the robot to prevent it from moving too fast. At the same time, set the speed to reverse when the robot moves to the environmental boundary, so that the robot always remains within the simulation environment, simulating the limitations of the site in the real working scenario.

[0074] The second step: Collect welding information

[0075] Simulate the visual large line scan camera to scan the production line area, quickly locate the specific position of the plate, and then use the weld seam search and tracking device at the end of the robot to accurately search for the weld seam. The two cooperate to determine the shape of the workpiece to be welded and the weld seam position, providing accurate information for the subsequent welding task.

[0076] The third step: Execute the welding task

[0077] After the visual scanning is completed, simulate the two robots switching to the welding torch state and simultaneously complete the welding task of the panel within their respective working ranges. During this process, the limitations of the robot's motion range should be considered. For example, when a single robot exceeds the pre-set range, it can move flexibly along a track similar to the crossbeam of a gantry to adjust the spatial coverage area of its motion range to adapt to different welding requirements.

[0078] Step 4: Collision detection mechanism

[0079] Dynamically monitor the working space during the operation of the dual robots in real time. When their working spaces approach the interference critical state, the control system module immediately triggers the collaborative regulation mechanism to dynamically plan and reposition the motion trajectories of the dual robots to avoid potential collision risks.

[0080] 2) Step P2

[0081] Step 1: Image acquisition

[0082] Divide the weld images jointly acquired by the visual large-line scan camera (1) and the weld seam searching and tracking device (7) into a set I = {I1, I2, …, I n};

[0083] Step 2: Autonomous decision-making for multi-layer and multi-pass welds

[0084] 1. Calculation formula for the number of weld layers n:

[0085] where t is the thickness of the workpiece; h is the deposited thickness of each layer of weld, generally determined according to the welding process and welding materials, k is a correction factor, generally taking 0 or 1. When the calculated is not an integer, k = 1 to round up the number of layers and ensure sufficient weld thickness.

[0086] 2. Calculation formula for the number of passes m in each layer of weld:

[0087] where W is the required width of each layer of weld; w is the width of a single-pass weld, which is related to the welding process parameters; c is a correction factor, generally taking 0 or 1. When the calculated is not an integer, c = 1 to round up the number of passes and ensure sufficient weld width.

[0088] Step 3: Image enhancement

[0089] Enhance the original image I i through operations such as rotation, scaling, cropping, and adding noise to obtain the enhanced image set I′ = {I′1, I′2, …, I′ n}, where n′ > n.

[0090] Step 4: Pixel value normalization

[0091] The pixel value matrix of image I′ i is P i , and the normalized pixel value matrix is P′ i . The normalization formula is:

[0092]

[0093] 3) Step P3

[0094] Step 1: Mark the weld length

[0095] Mark the length l of the weld processed in step (p2) i ;

[0096] Step 2: Divide the priority and welding difficulty

[0097] The integrated control algorithm module divides the weld priority p i and the welding difficulty d i according to the welding process and production requirements;

[0098] Step 3: Define the working range and weld set

[0099] The two robots are R1 and R2, and their working ranges are S1 and S2 respectively. The weld set covered by the working range is W = {w1, w2, …, w n}, where w j ∈W and w j satisfies being within S1 or S2.

[0100] 4) Step P4

[0101] Step 1: Define the workload quantification

[0102] Define the workload H j of robot R j at time t as the sum of the weld workloads assigned to it. The workload h i of weld i can be expressed as:

[0103] h i = αl i + βd i

[0104] where α and β are weight coefficients, and α + β = 1. Then:

[0105]

[0106] Step 2: Robot information interaction and task coordination

[0107] Two robots exchange their respective status and task progress information through a communication module. Robot R1 sends its own status s1(t) and the completed weld information to Robot R2. Through this communication, dynamic adjustment of tasks and collaborative work are achieved.

[0108] Step 3: Real-time monitoring of status and progress

[0109] Real-time monitor the status of the robot and the completion of the welds. c i (t) is the completion status of weld i at time t, c i (t) = 1 indicates completion, c i (t) = 0 indicates incompletion.

[0110] 5) Step P5

[0111] Step 1: Define the objective function

[0112]

[0113] Constraints:

[0114] 1. Uniqueness of task assignment:

[0115] 2. Non-negativity constraint:

[0116] 3. Avoidance of collisions and interferences:

[0117] 4. Decision-making assistance rules:

[0118] When then y = 1, at this time Robot R1 assists R2 to complete the remaining tasks, the set of remaining tasks. When then y = 2, at this time Robot R2 assists R1 to complete the remaining tasks, the set of remaining tasks. When then y = 0, Robot R1 and R2 independently complete their respective task sets.

[0119] x i1 and x i2 are decision variables, H1 and H2 respectively represent the total workload assigned to Robot R1 and R2, that is h i is the workload size of each small task.

[0120] 6) Step P6

[0121] Step 1: Divide multiple types of reward functions

[0122] The setting of the reward value for the robot in the learning environment can correctly guide its learning process and can be divided into:

[0123] 1. Full - scan reward Y1

[0124] The purpose of the full - scan is to ensure that the robot accurately obtains the information of the welded part, providing a basis for subsequent welding tasks. When the robot completes the full - scan of the welded part and the scanning accuracy reaches the set requirement, a positive reward is given. The set target accuracy is ∈0, and the scanning accuracy is ∈; when ∈≥∈0, if ∈≤∈0

[0125] 2. Welding reward Y2

[0126] The collaborative welding reward is related to the quality and efficiency of the weld seam. The quality of the weld seam can be measured by the appearance quality and internal quality of the weld seam. The weld - seam quality index is Q, and the welding - efficiency index is E. The collaborative welding reward can be expressed as: Y2 = α1Q+α2E, where α1 and α2 are weight coefficients, satisfying α1 + α2 = 1 and α1, α2≥0

[0127] 3. Process obstacle - avoidance reward Y3

[0128] The process obstacle - avoidance reward is used to train the robot to avoid collisions with obstacles or other robots during the welding process. When the robot successfully avoids an obstacle within time step t, if a collision occurs, then

[0129] Step 2: Construct the composite reward function

[0130] Taking into comprehensive consideration multiple objectives such as full - scan, welding task, and process obstacle - avoidance, reasonable weights λ1, λ2, λ3 are assigned to each objective to construct the reward function Y:

[0131]

[0132] where Y 1k represents the full - scan reward value corresponding to the k - th objective, Y 2k represents the welding reward value corresponding to the k - th objective, and where Y 3k represents the process obstacle - avoidance reward value corresponding to the k - th objective, and λ1 + λ2 + λ3 = 1.

[0133] Step 3: Smoothing the reward function

[0134] The smoothing function S(Y) is used to smooth the reward function Y to avoid sharp fluctuations and improve the stability of the algorithm.

[0135] 7) Step P7

[0136] Step 1: Calculate the reward value

[0137] The design of the composite reward function for multi-robot collaborative welding needs to revolve around three core objectives: First, the comprehensive scanning reward to ensure that the robot accurately obtains the information of the welded part and meets the scanning accuracy requirements; second, the welding reward to motivate based on the weld quality and welding efficiency; third, the obstacle avoidance reward to avoid collisions through positive feedback within the safe distance and punishment at dangerous distances. Combine these three parts according to weights and use smoothing to reduce reward fluctuations to stabilize the training process. In practical applications, the weights need to be dynamically adjusted according to task requirements to ensure that the robot efficiently completes the welding task on the premise of safe cooperation.

[0138] As Figure 7 shown, the multi-robot collaborative welding task processing system and method set include:

[0139] a. Deep learning model for multi-modal fusion: Use the random forest combined with the path planning algorithm for the collected images of the workpiece to be welded to calculate the number of weld beads and plan the weld distribution of each layer of weld beads. Then use the MaskR-CNN algorithm to accurately label the welds on the workpiece to be welded. With the help of the Albumentations library, it is suitable for welding defect data augmentation, expanding the dataset, and improving the model's ability to identify and process welding defects. Use standardization and scaling methods to adapt the data format.

[0140] b. Task assignment deviation minimization function: Based on workload quantification and assignment constraints, considering multiple factors of the weld and the working range of the robot, balance the task assignment. The real-time information interaction of the robot provides data for real-time adjustment to ensure collaborative operation.

[0141] c. Composite smoothing reward function: Integrate various reward functions in step P6 of the smoothing process, integrate the multi-objective reward values, provide stable feedback for the robot's learning, guide the optimization behavior, echo the reward function in step P7, and comprehensively evaluate the welding task.

[0142] As Figure 8 shown, two welding and scanning robots work together and are installed on the gantry to jointly form a working unit. Among them, the welding robot can move flexibly along the track of the gantry beam and can adjust its position according to the actual operation requirements. The visual large line scan camera (1) quickly scans the working area to achieve rough positioning of the workpiece; the weld seam seeking and tracking device (7) is installed at the end of the robot, which can accurately seek the position and correct the weld angle deviation in real time. The two cooperate to determine the product model and welding parameters, and automatically generate the bead arrangement sequence and welding program. The intelligent welding monitor collects the key welding parameters and deeply analyzes them through the data processing and analysis module. The intelligent control cabinet (2) is the center to optimize the welding link. The digital inverter welder (3) accurately controls the welding energy, the circulating strong cold water tank (4) cools at a constant temperature, the intelligent wire barrel (5) automatically feeds wire and monitors, the clean air source air compressor provides high-purity gas, and the negative pressure dust removal host (6) captures the dust. Each module cooperates to improve the welding quality and safety.

[0143] As Figure 9 shown, in the intelligent equipment welding system, robots R1 and R2 are respectively responsible for the operations in areas S1 and S2. Usually, relying on precise motion control and sensing technologies, they operate independently in their respective areas, using vision systems, weld tracking algorithms, etc. to ensure welding quality and accuracy. When a robot finishes its task ahead of schedule, the system coordination mechanism is activated. The robot that has completed the task interacts with the robot in operation through a real-time communication network, judges the assistance method according to intelligent algorithms, avoids duplicate work, and at the same time, the task allocation deviation minimization function dynamically adjusts the remaining tasks to balance the workload.

Claims

1. A robot self-learning intelligent welding system based on multi-machine collaboration, characterized in that, Based on a multi-robot collaborative framework, including a vision large line scan camera (1), a weld seam positioning and tracking device (7), and an intelligent control cabinet. The multi-robot collaborative architecture includes two robots that can move along the crossbeam track of the gantry, and the end of the robot is integrated with a welding torch and a weld seam positioning and tracking device (7). The weld seam positioning and tracking device (7) is used to accurately locate the weld seam and correct the welding angle deviation in real time. A vision large line scan camera (1) is installed on the gantry, which is responsible for scanning the area, roughly positioning the workpiece, quickly giving the approximate position, and providing initial information for subsequent operations. The weld seam positioning and tracking device (7) and the vision large line scan camera (1) cooperate to determine the welding scenario, the shape of the workpiece, and the position of the weld bead. An intelligent control cabinet (2) is also installed above the gantry, which integrates: a control system module, an integrated control algorithm and self-learning module, an autonomous decision-making module, a data collection and storage module, a human-machine interaction module, a communication module, and a power management module. Among them, the integrated control algorithm and self-learning module are used to realize the dynamic path optimization and collision avoidance decision-making of welding.

2. A robot self-learning intelligent welding method based on multi-machine cooperation, characterized in that, Based on the system described in claim 1, it includes three stages: environment construction, dynamic optimization, and task execution. Specifically, it includes the following steps: Step (a): Construct a reinforcement learning environment based on dual-robot collaborative welding. Step (b): Collect welding data in the environment built in step (a), make autonomous decisions on multiple weld seams, and preprocess the data. Step (c): Divide the weld seam priority and welding difficulty according to production requirements and welding processes, and at the same time clarify the working ranges of the two robots. Step (d): Perform dynamic task partitioning and construct a monitoring system according to the priority division of the welding tasks in step (c). Step (e): The monitoring system ensures the balanced distribution of the workloads of the two robots through a task allocation deviation minimization function. Step (f): Construct various types of reward functions Y according to the completion of the robot's comprehensive scanning, welding tasks, and process obstacle avoidance. Step (h): Construct a composite reward function through the comprehensive scanning accuracy, welding tasks, and process obstacle avoidance reward values, calculate the reward values, and then guide the robot to learn better welding strategies.

3. The method according to claim 2, characterized in that, The specific steps of step (a) include the following: Step (a1): Simulate and depict the space of the simulation welding environment based on the python software. Step (a2): Simulate and add a reinforcement learning environment with weld seams, obstacles, and robots.

4. The method according to claim 3, characterized in that The specific steps of step (b) include the following: Step (b1): In the environment constructed in step (a), through the cooperation of the vision large line scan camera (1) and the weld seam positioning and tracking device (7), the data collection and storage module collects and stores the images of the overall production line layout and the workpiece to be welded. Step (b2): The autonomous decision-making module calculates the number of weld beads at the weld seam of the collected workpiece weld area image and reasonably arranges the weld bead distribution. Step (b3): Label the images collected in step (b1), and the labeling content includes the position of the single-layer weld seam and the position of each weld bead in each layer of the multi-layer multi-pass weld seam. Step (b4): Enhance the original image by rotation, scaling, cropping, and adding noise to expand the dataset and improve the generalization ability of the model. Step (b5): Normalize the pixel values of the image to the range of [0, 1] or [-1, 1].

5. The method according to claim 4, characterized in that, The said step (c) specifically includes the following steps: Step (c1): Mark the length l for each weld seam in the data collection and storage module i ; Step (c2): The integrated control algorithm module divides the weld priority p i and the welding difficulty d i for classification; Step (c3): The autonomous decision-making module divides the working ranges of the two robots R1 and R2 in combination with the stored production line layout image data, and simultaneously counts the set of weld seams covered within the working ranges as W n .

6. The method according to claim 5, wherein The said step (d) specifically includes the following steps: Step (d1): Robot R j The total workload H j is the sum of the allocated weld workloads h i where j = 1, 2; the workload h of weld i i is expressed as: h i = αl i + βd i Where α and β are weight coefficients, and α + β = 1; then: Step (d2): The two robots exchange their respective status and task progress information through the communication module; robot R1 sends its own status s1(t) and the completed weld information to robot R2, and through this communication, dynamic adjustment and collaborative work of the task are realized. Step (d3): The monitoring system monitors the status of the robot and the completion of the weld seam in real time; c i Ci(t) is the completion status of weld seam i at time t, c i Ci(t) = 1 indicates completion, c i Ci(t) = 0 indicates non - completion.

7. The method according to claim 6, characterized in that The said step (e) specifically includes the following steps: Step (e1): The purpose of the monitoring system is to balance the time for the two robots to complete their respective tasks and give priority to processing high-priority welds; define the objective function as: H1 and H2 respectively represent the total workloads assigned to robots R1 and R2, that is h i is the workload size of each small task.

8. The method according to claim 7, wherein The said step (f) specifically includes the following steps: Step (f1): Considering comprehensively full-scan, welding tasks, and process obstacle avoidance, allocate reasonable weights for the completion of each task to construct various types of reward functions. Step (f2): To avoid drastic fluctuations in the reward function in step (f1), smooth the reward function to improve the stability of the algorithm.

9. The method according to claim 8, characterized in that The said step (h) specifically includes the following steps: Step (h1): Through the indicators of full-scan accuracy, welding tasks, and collision situations, achieve accurate measurement of the welding behavior, calculate the composite reward value, and further guide the robot to learn better welding strategies.

Images