An intelligent high-performance collaborative welding robot system

Abstract

The application provides a kind of intelligent high-performance collaborative welding robot system, belongs to welding robot technical field, the intelligent high-performance collaborative welding robot system proposed in the application includes multi-degree-of-freedom welding robot body, intelligent control system and high-performance welding power supply and auxiliary equipment.Its welding process control module first acquires workpiece three-dimensional information based on visual inspection to construct welding path;Then according to material characteristics and weld shape, select matching process parameters;Establish welding error matrix to monitor welding process in real time;Construct parameter error adjustment matrix for dynamic adjustment;Adopt neural network analysis parameter fluctuation to identify molten pool characteristics;Use pre-trained neural network to evaluate weld quality;Finally, integrate multi-robot collaborative welding strategy to realize efficient operation, solve the technical problems that welding process parameters in the prior art welding robot are difficult to adjust in real time to adapt to complex working condition changes.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of welding robots, and particularly relates to an intelligent high-performance collaborative welding robot system. BACKGROUND

[0002] Welding robots play an important role in modern industrial manufacturing, especially in the fields of automobiles, ships, aerospace, etc. Traditional welding robot systems usually adopt a preset program control mode, which sets the welding trajectory and process parameters in advance through programming, and combines simple sensor feedback for welding operation. Such systems perform well in standardized and repetitive welding tasks, improving production efficiency and maintaining certain welding quality stability.

[0003] However, traditional welding robot systems have obvious shortcomings in dealing with complex and variable welding environments. First, preset parameters are difficult to adapt to material deformation, thermal deformation and position deviation during welding; second, simple feedback control mechanisms cannot effectively capture the dynamic characteristics of the welding pool, resulting in fluctuations in welding quality; third, the single robot operation mode limits production efficiency and flexibility. Especially in complex working conditions such as dissimilar material connection and variable thickness component welding, the difficulty of welding quality control increases significantly.

[0004] Currently, the industry attempts to improve the adaptability of welding robots by introducing more sensors and improving control algorithms, but still lacks an intelligent system that can comprehensively analyze the complex dynamic characteristics of the welding process and accurately adjust the welding parameters in real time. The traditional feedback control method lags behind and cannot predict the risk of welding defects, while the simple parameter adjustment strategy cannot balance the multi-objective optimization demand of welding quality and efficiency, resulting in difficulty in real-time and accurate adjustment of welding process parameters to adapt to complex working condition changes. That is, there is a technical problem of difficulty in real-time and accurate adjustment of welding process parameters to adapt to complex working condition changes in the welding robot in the prior art. SUMMARY

[0005] Therefore, the present application provides an intelligent high-performance collaborative welding robot system, which can solve the technical problem of difficulty in real-time and accurate adjustment of welding process parameters to adapt to complex working condition changes in the welding robot in the prior art.

[0006] The application is achieved as follows: the application provides an intelligent high-performance collaborative welding robot system, which comprises a multi-degree-of-freedom welding robot body, an intelligent control system, a high-performance welding power supply and auxiliary equipment; the intelligent control system comprises a motion control module, a welding process control module, a visual detection and feedback module, a collaborative work control module and a man-machine interaction interface; the welding process control module performs the following steps: obtaining three-dimensional information of a welding workpiece to construct a welding path; selecting matching welding process parameters; establishing a welding error matrix for real-time monitoring; constructing a welding parameter error adjustment matrix; using a neural network algorithm to analyze welding parameter fluctuations; introducing a welding molten pool fluid dynamics model; applying a welding quality multi-objective optimization function; and using a welding quality evaluation neural network model for real-time evaluation.

[0007] The multi-degree-of-freedom welding robot body has at least six degrees of freedom, comprises a jointed mechanical arm and an end effector, the jointed mechanical arm is made of high-strength lightweight material, and the end effector is equipped with a welding gun; the high-performance welding power supply and auxiliary equipment comprise a high-performance welding power supply, a wire feeding mechanism and a gas protection device.

[0008] The welding process control module performs the step of obtaining three-dimensional information of a welding workpiece to construct a welding path as follows: obtaining three-dimensional information of a welding workpiece based on the visual detection and feedback module, constructing a set of key points of a welding path, and generating a welding motion trajectory plan.

[0009] The welding process control module performs the step of selecting matching welding process parameters as follows: selecting matching welding process parameters from a welding process database according to material properties of the welding workpiece and the weld shape, and generating an initial welding parameter setting.

[0010] The welding process control module performs the step of establishing a welding error matrix for real-time monitoring as follows: establishing a welding error matrix, real-time monitoring of weld position deviation, molten pool shape change and welding depth in a welding process, and forming welding state feedback data.

[0011] The welding process control module performs the step of constructing a welding parameter error adjustment matrix as follows: constructing a welding parameter error adjustment matrix, calculating welding parameter deviation based on the welding state feedback data, and adjusting welding current, welding voltage and welding speed.

[0012] The welding process control module performs the following steps to analyze welding parameter fluctuations using a neural network algorithm: It uses a neural network algorithm to analyze welding parameter fluctuations during the welding process, identifies the dynamic characteristics of the weld pool, and establishes a dynamic compensation model for welding parameters. The dynamic compensation model for welding parameters employs a lightweight deep reinforcement learning model. Specifically, its structure is a hybrid neural network architecture composed of a multilayer perceptron network and a long short-term memory network. The input layer receives welding status feedback data and historical welding parameter data, the hidden layer contains multiple fully connected layers and long short-term memory network units, and the output layer generates welding parameter compensation values.

[0013] The step of the welding process control module to introduce the welding pool fluid dynamics model is as follows: introduce the welding pool fluid dynamics model, analyze the turbulent state of metal flow inside the pool, and predict potential welding defect risks.

[0014] The welding process control module performs the following steps to apply the welding quality multi-objective optimization function: applying the welding quality multi-objective optimization function to optimize welding parameters, and adjusting the welding parameter combination based on the dynamic characteristics of the weld pool and the welding defect risk assessment results.

[0015] The welding process control module performs the following step of real-time evaluation using a welding quality assessment neural network model: It uses a pre-trained welding quality assessment neural network model to perform real-time evaluation of the weld morphology and generate welding quality prediction indicators. The specific structure of the welding quality assessment neural network model is a hybrid network architecture combining multi-layer convolutional neural networks and Transformer structures, including a weld image feature extraction module, a welding sound signal feature extraction module, and a multi-modal information fusion module. The welding quality assessment neural network model also includes an attention mechanism, the parameters of which are determined based on the material properties of the welded workpiece, the weld shape, and the welding process type.

[0016] Compared with existing technologies, this invention provides an intelligent high-performance collaborative welding robot system. This system integrates multimodal sensing, deep learning, and fluid dynamics modeling technologies to construct a complete intelligent perception and control closed loop for the welding process. The system can acquire real-time three-dimensional information of the welded workpiece, monitor weld position deviation, molten pool shape changes, and welding depth. Based on neural network algorithms, it analyzes welding parameter fluctuations and establishes a dynamic compensation model, enabling precise adjustment of welding parameters according to changes in working conditions.

[0017] Compared with the prior art, the real-time accurate adjustment capability of the welding parameters is significantly improved. First, by establishing a welding error matrix and a parameter error adjustment matrix, the system can quickly respond to deviations in the welding process; second, by introducing a welding molten pool fluid dynamics model, the internal metal flow disorder state of the molten pool is analyzed to realize predictive control of potential welding defects; finally, combined with a lightweight deep reinforcement learning model and a multi-modal information fusion mechanism, the system can adaptively focus on key moments and parameter changes in the welding process, balancing welding quality and energy efficiency.

[0018] Through the above technical means, the technical problem that the welding process parameters in the welding robot are difficult to adjust in real time to adapt to complex working condition changes in the prior art is solved, and it is ensured that stable welding quality can still be maintained in a complex and variable welding environment, especially in complex working conditions such as dissimilar material connection and variable thickness component welding. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 The flowchart of the method of the present application.

[0020] Figure 2 The composition schematic diagram of the intelligent high-performance collaborative welding robot system in Example 2. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.

[0022] As Figure 1 shown is the flowchart of the intelligent high-performance collaborative welding robot system provided by the present application, and the method comprises the following steps:

[0023] The intelligent high-performance collaborative welding robot system comprises a multi-degree-of-freedom welding robot body, an intelligent control system, a high-performance welding power supply and auxiliary equipment.

[0024] The multi-degree-of-freedom welding robot body has at least 6 degrees of freedom, comprises a joint type mechanical arm and an end effector, the joint type mechanical arm is made of high-strength lightweight material, and the end effector is equipped with an advanced welding gun.

[0025] The intelligent control system comprises a motion control module, a welding process control module, a visual detection and feedback module, a collaborative work control module and a man-machine interaction interface.

[0026] The high-performance welding power supply and auxiliary equipment comprise a high-performance welding power supply, a wire feeding mechanism and a gas protection device.

[0027] The welding process control module performs the following steps:

[0028] S01, obtaining three-dimensional information of the welding workpiece based on the visual detection and feedback module, constructing a set of welding path key points, and generating a welding motion trajectory plan;

[0029] S02, selecting matched welding process parameters from a welding process database according to the material properties of the welding workpiece and the weld shape, and generating initial welding parameter settings;

[0030] S03, establishing a welding error matrix, real-time monitoring of weld position deviation, molten pool shape change and welding depth in the welding process, forming welding state feedback data;

[0031] S04, constructing a welding parameter error adjustment matrix, calculating welding parameter deviation based on the welding state feedback data, and adjusting welding current, welding voltage and welding speed;

[0032] S05, using a neural network algorithm to analyze welding parameter fluctuations in the welding process, identify welding pool dynamic characteristics, and establish a welding parameter dynamic compensation model;

[0033] S06, introducing a welding pool hydrodynamic model to analyze the internal metal flow disorder state of the molten pool and predict the potential welding defect risk;

[0034] S07, applying a multi-objective optimization function of welding quality to optimize welding parameters, adjusting the combination of welding parameters based on the welding pool dynamic characteristics and the welding defect risk assessment results;

[0035] S08, using a pre-trained welding quality evaluation neural network model to evaluate the weld appearance in real time, generating a welding quality prediction index, and the attention mechanism parameters in the welding quality evaluation neural network model are determined according to the material properties of the welding workpiece, the weld shape and the welding process type;

[0036] S09, integrating a multi-robot collaborative welding strategy, assigning welding areas based on a task decomposition algorithm, and realizing multi-robot synchronous welding operation;

[0037] The set of welding path key points refers to a set of discrete points obtained by sampling at a predetermined interval on the welding trajectory, each key point containing spatial coordinate information and corresponding welding pose information, used to guide the motion trajectory planning of the multi-degree-of-freedom welding robot body;

[0038] The welding error matrix refers to a multi-dimensional data structure describing the deviation between the actual weld position and the expected position in the welding process, including horizontal deviation, vertical deviation, angle deviation and other parameters, used for welding precision control and error compensation;

[0039] The welding parameter error adjustment matrix refers to a conversion matrix that maps the welding error matrix to the welding parameter adjustment amount, and is used to automatically calculate and adjust the welding current, welding voltage, welding speed and other parameters according to the detected welding error.

[0040] The welding parameter dynamic compensation model refers to a welding parameter real-time adjustment model constructed based on the neural network algorithm, which can dynamically adjust the welding parameters according to the workpiece state changes and environmental factors in the welding process, and improve the welding quality stability.

[0041] The welding pool fluid dynamics model refers to a mathematical model describing the flow behavior of the metal liquid inside the welding pool, considering the influence of surface tension, buoyancy, electromagnetic force and other factors on the pool shape, and is used to predict the weld forming quality.

[0042] The turbulent state refers to the irregular flow state of the metal liquid inside the welding pool, and excessive turbulence can cause welding defects such as pores and slag, which need to be controlled by optimizing the welding parameters.

[0043] The welding quality multi-objective optimization function is used to comprehensively consider the optimization objectives of multiple aspects of welding quality, and the input includes welding pool temperature distribution parameters, pool metal flow velocity vector, pool surface fluctuation amplitude index and weld forming geometry parameters, and the output is the optimal welding parameter combination recommendation value and the corresponding welding quality prediction score.

[0044] The specific structure of the welding quality evaluation neural network model is a hybrid network architecture combining multi-layer convolutional neural network and Transformer structure, including a weld image feature extraction module, a welding sound signal feature extraction module and a multi-modal information fusion module. The weld image feature extraction module uses a residual network structure to extract weld surface topography features, the welding sound signal feature extraction module uses a one-dimensional convolutional network to extract sound time-frequency features during the welding process, and the multi-modal information fusion module integrates different modal information through a cross-attention mechanism and generates a final welding quality evaluation result. The welding quality evaluation neural network model has self-adaptive learning ability and can dynamically adjust network parameter weights according to different welding process types and material characteristics.

[0045] The step of establishing the training data set of the welding quality evaluation neural network model specifically includes collecting a large number of welding process data under different materials and different process parameters, each group of data containing a welding process image sequence, a welding sound signal, a welding parameter record and a corresponding welding quality rating label, manually rating the welding quality by a professional welding process expert, and quantitatively analyzing the welding internal defects by combining non-destructive testing technology to form a standardized training data set, while expanding the training samples by using data enhancement technology to improve the model generalization ability.

[0046] The step of training the welding quality evaluation neural network model specifically includes first training on a large-scale general welding data set to learn the general feature representation of welding quality evaluation, then fine-tuning training for specific welding processes and materials, evaluating model performance and optimizing model hyperparameters by using cross-validation method, improving model robustness by introducing adversarial training, and finally verifying the model in the actual welding environment and continuously optimizing and updating the model according to the verification results. The specific welding process and material are determined based on the selection of the training personnel.

[0047] The motion control module is used to control the motion of the multi-degree-of-freedom welding robot body, including joint motion control, Cartesian space trajectory control and position and attitude control, to realize precise positioning and attitude adjustment of the welding gun in three-dimensional space.

[0048] The welding process control module is used to control the process parameters in the welding process, including welding current, welding voltage, welding speed, wire feeding speed and protective gas flow, to realize fine control of the welding process.

[0049] The visual detection and feedback module is used to obtain the three-dimensional information of the welding workpiece and the molten pool state in the welding process in real time, including weld position, weld shape, molten pool shape and welding depth, etc., to provide feedback data for welding process control.

[0050] The collaborative work control module is used to coordinate the collaborative work of multiple welding robots, realize multi-robot synchronous welding, and improve welding efficiency and quality.

[0051] The human-computer interaction interface is used to realize the interaction between the operator and the welding robot system, including welding parameter setting, welding path planning, welding process monitoring and welding result display, etc.

[0052] The welding parameter dynamic compensation model adopts a lightweight deep reinforcement learning model, and the specific structure is a hybrid neural network architecture composed of a multi-layer perception network and a long short-term memory network. The input layer receives welding state feedback data and welding parameter historical data. The hidden layer includes multiple fully connected layers and long short-term memory network units. The fully connected layer is used to extract the nonlinear mapping relationship between the welding parameters and the welding quality. The long short-term memory network unit is used to capture the timing characteristics and long-term dependence of the welding process. The output layer generates a welding parameter compensation value, and combines the compensation value with the basic welding parameter through a residual connection mechanism to form the final welding parameter control instruction. At the same time, an attention mechanism is introduced to weight the welding state information at different time steps, so that the model can adaptively focus on the key moments and key parameter changes in the welding process. Batch normalization and Dropout techniques are used to improve the training stability and model generalization ability. The loss function is designed as a weighted combination of the welding quality evaluation index and the energy consumption. The welding quality and energy efficiency are balanced through a multi-objective optimization method. The steps for establishing the training data set are as follows: first, collect welding process data under various welding process conditions, including welding parameter records and corresponding welding quality evaluation results of different materials, different thicknesses, and different welding positions; then, preprocess the collected raw data, including data cleaning, outlier detection, data normalization, and time sequence alignment; then, use the sliding window method to divide the continuous welding process data into fixed-length training samples, each sample containing a welding parameter sequence and a corresponding welding quality score; to solve the data imbalance problem, use stratified sampling and data augmentation techniques to expand the number of samples for rare welding conditions; at the same time, introduce expert knowledge to guide the data labeling process, and based on the experience of welding experts, construct the mapping relationship between welding parameters and welding quality; finally, divide the entire data set into training set, validation set and test set for model training, hyperparameter tuning and performance evaluation. In the training process, the cross-validation method is used to evaluate the performance stability of the model under different data divisions, and the online learning mechanism is used to update the model parameters continuously, so that the model can adapt to changes in welding processes and equipment aging, thereby improving the stability and consistency of welding quality.

[0053] The specific implementation of the above steps is described in detail below.

[0054] The system includes a multi-degree-of-freedom welding robot body, an intelligent control system, and a high-performance welding power supply and auxiliary equipment. The multi-degree-of-freedom welding robot body adopts a 6-axis articulated mechanical arm structure, with a load capacity of 10 kg and a repeat positioning accuracy of ±0.05

[0055] ±0.05mm, the maximum working radius is 1800mm; joint drive uses high-precision servo motor, equipped with 20-bit absolute value encoder; the end effector is equipped with water-cooled welding gun, with automatic adjustment function. The intelligent control system uses real-time operating system platform, the main controller uses industrial computer, the processor frequency is not less than 3.5GHz, the memory capacity is not less than 16GB; the visual detection module includes high-speed industrial camera and structured light projection device; the man-machine interface uses 19-inch touch screen, the resolution is 1920x1080

[0056] 1920x1080. The high-performance welding power supply is an inverter digital control welding power supply, the rated output current is 400A, the no-load voltage is 80V, and the output characteristic can be switched between constant current and constant voltage modes; the wire feeding mechanism adopts four-wheel drive structure, and the wire feeding speed range is 2-20m / min; the gas protection device is equipped with an electronic flow controller, and the flow range is 5-25L / min.

[0057] The welding process control module performs the following steps:

[0058] The specific implementation of step S01 is based on the process of obtaining three-dimensional information of the welding workpiece by the visual detection and feedback module. First, the structured light technology is used to scan the surface of the welding workpiece to obtain the point cloud data of the workpiece surface; then the point cloud data collected from multiple angles is fused into a complete three-dimensional model of the workpiece through a point cloud registration algorithm; then an edge detection algorithm is used to extract features from the workpiece surface to identify the weld edge contour; further, a least squares curve fitting method is applied to construct a weld path curve equation; then a set of weld path key points is generated on the curve at an interval of 5-10mm; finally, based on these key points, a smooth and continuous welding motion trajectory is generated by using a cubic spline interpolation algorithm. The purpose of this step is to provide accurate motion path planning for the welding robot to ensure the accuracy of the welding position.

[0059] The specific implementation of step S02 is to select welding process parameters according to the material properties of the welding workpiece and the weld shape. First, the physical property parameters of the workpiece material are retrieved from the database, including melting point, thermal conductivity, specific heat capacity, etc.; then the weld cross-sectional area is calculated according to the weld shape parameters; then based on the material properties and weld geometry, a fuzzy reasoning system is used to select the most matched welding process parameter template from the welding process database; further, combined with the workpiece thickness and welding position, a weight coefficient correction method is applied to optimize and adjust the template parameters, and the parameter settings of the initial welding current range of 80-320A, welding voltage range of 18-32V, and welding speed range of 5-15mm / s are obtained. The role of this step is to provide reasonable initial welding parameters to lay the foundation for subsequent welding process control.

[0060] The specific implementation of step S03 is to establish a welding error matrix and monitor the welding process in real time. First, a four-dimensional error matrix containing lateral deviation, longitudinal deviation, angle deviation and depth deviation is defined; then the welding process images are captured in real time by a high-speed camera, with a collection frequency of 60-120 Hz; then the actual position and theoretical position of the weld are extracted using image processing technology; further, Kalman filtering algorithm is used to filter the detection results to eliminate the influence of random noise; at the same time, the temperature distribution of the molten pool is monitored by an infrared thermal imager, and the temperature detection range is 800-1600℃; finally, the dimensional deviation values are calculated according to the image analysis results and temperature distribution data, and the welding error matrix is updated. The purpose of this step is to obtain real-time state information of the welding process, which provides a basis for subsequent parameter adjustment.

[0061] The specific implementation of step S04 is to construct a welding parameter error adjustment matrix and adjust the welding parameters. First, a linear mapping relationship matrix of error vector to welding parameter adjustment amount is established; then regression analysis is performed on the historical welding data based on the least squares method to determine the numerical values of the elements in the matrix; then the welding parameter deviation is calculated according to the welding state feedback data using an adaptive control algorithm; further, the welding current adjustment step is set to 3-8 A, the welding voltage adjustment step is set to 0.5-1.5 V, and the welding speed adjustment step is set to 0.5-2 mm / s; finally, the adjusted parameters are transmitted in real time to the welding power supply control system through a serial communication interface. The function of this step is to realize closed-loop control of welding parameters and improve the stability of the welding process.

[0062] The specific implementation of step S05 is to analyze the welding parameter fluctuation using a neural network algorithm. First, a feedforward neural network with three hidden layers is constructed, with 64, 32 and 16 nodes in each layer; then the time series data of welding current, welding voltage and welding speed are used as input, and the molten pool shape feature is used as output; then the back propagation algorithm is used to train the network, with a learning rate of 0.001-0.005; further, the trained network model is used to analyze the influence of welding parameter fluctuation on molten pool dynamic characteristics in real time; finally, based on the analysis results, a welding parameter dynamic compensation model is established to realize predictive adjustment of welding parameters. The purpose of this step is to deeply analyze the parameter fluctuation law in the welding process, which provides theoretical support for fine welding control.

[0063] The specific implementation of step S06 is to introduce a welding molten pool fluid dynamics model for analysis. First, a set of molten pool fluid mechanics equations considering surface tension, buoyancy, and electromagnetic force is established; then the finite element method is used to discretize the molten pool area into grid elements with a grid size of 0.1-0.5 mm; next, the temperature field and velocity field at each element are calculated according to the welding process parameters; further, the Reynolds number is introduced as an indicator of flow turbulence, with a critical value of 2000; finally, the high-turbulence region is identified by calculating the Reynolds number distribution, and the potential risk of defects such as porosity and slag inclusion is predicted. The role of this step is to theoretically analyze the flow state inside the molten pool and predict the mechanism of welding defects.

[0064] The specific implementation of step S07 is to apply a welding quality multi-objective optimization function for parameter optimization. First, a multi-objective optimization function is constructed, including weld formation quality, penetration consistency, porosity, and energy efficiency; then the particle swarm optimization algorithm is used for parameter optimization, with a particle number of 50-100 and a maximum iteration number of 200; next, the optimization constraint conditions are set based on the dynamic characteristics of the welding molten pool and the defect risk assessment results; further, a search is performed in the three-dimensional parameter space of welding current, welding voltage, and welding speed to obtain a set of Pareto optimal solutions; finally, the parameter combination with the best comprehensive performance is selected from the optimal solution set as the recommended value. The purpose of this step is to seek the best balance among multiple welding quality indicators and provide the optimal welding parameter combination.

[0065] The specific implementation of step S08 is to evaluate the weld appearance using a pre-trained neural network model. First, the welding quality evaluation neural network model is called; then the weld image is collected in real time by a high-resolution camera with a resolution of no less than 1920x1080 pixels; next, the image is preprocessed, including grayscale, contrast enhancement, and noise filtering; further, the processed image is input into the neural network model; finally, the welding quality score is generated based on the model output results, with a score range of 0-100, where a score of 85 or above is an excellent weld, a score of 70-85 is a qualified weld, and a score below 70 is an unqualified weld. The role of this step is to evaluate the welding quality in real time and provide feedback for welding process control.

[0066] The specific implementation of step S09 is to integrate a multi-robot collaborative welding strategy. First, the overall welding task is divided into multiple sub-tasks based on the geometry of the welding workpiece; then a task allocation algorithm is used to assign welding areas to each robot to balance the workloads of the robots; next, a robot collaborative motion planning is developed to avoid collisions between robots; further, a communication protocol is established between robots to realize real-time sharing of state information; finally, the synchronization of multiple robots is coordinated through a master-slave control architecture, with the welding progress of the master robot serving as the synchronization reference, and the progress deviation of the slave robots not exceeding 10%. The purpose of this step is to improve welding efficiency and quality and shorten the welding cycle.

[0067] The detailed structure of the welding parameter dynamic compensation model is a hybrid neural network composed of a multi-layer perception network and a long short-term memory network. The input layer of the model receives 16-dimensional welding state data and historical parameter data; the hidden layer includes 3 fully connected layers and 2 long short-term memory network layers; the number of nodes of the fully connected layers is 128, 64, and 32 respectively, and the activation function is ReLU function; the number of units of the long short-term memory network layer is 64, which is used to learn the timing characteristics of the welding parameter change; the output layer generates the compensation values of the welding current, welding voltage, and welding speed; the model adopts a residual connection mechanism to combine the compensation values with the basic parameters; at the same time, an attention mechanism is introduced to give higher weight to the state information at the key moment; the network optimization adopts an Adam optimizer with a learning rate of 0.001 and a training batch size of 32. The training data set establishment process includes collecting welding process data of 5 common materials under different thicknesses and welding positions; cleaning, normalizing, and timing aligning the original data; dividing the training samples by using a 10s length sliding window; expanding the rare working condition samples by using data enhancement technology; and finally dividing the training set, validation set, and test set in a ratio of 7:2:1.

[0068] The detailed structure of the welding quality evaluation neural network model is a hybrid network combining a multi-layer convolutional neural network and a Transformer structure. The image feature extraction module adopts a ResNet-50 residual network structure, including 5 residual blocks; the sound signal feature extraction module adopts a one-dimensional convolutional network, including 4 convolutional layers with convolution kernel sizes decreasing from 16 to 4; the multi-modal information fusion module adopts a 6-layer Transformer encoder structure with 8 attention heads; and the final output layer is a fully connected layer with 5 nodes corresponding to 5 welding quality levels. The training data set establishment process includes collecting 100,000 groups of welding data under different process parameters; each group of data includes a welding process image sequence, a sound signal, and a parameter record; inviting 5 professional welding process experts to rate the welding quality; quantitatively analyzing the internal defects by combining X-ray and ultrasonic detection technologies; performing data enhancement by using image rotation, scaling, brightness adjustment, etc.; and finally forming a standardized training data set containing 150,000 groups of samples.

[0069] Further, the specific embodiments of the welding robot are described as follows: the multi-degree-of-freedom welding robot body adopts a 6-axis articulated mechanical arm structure, with a load capacity of 10 kg, a repeat positioning accuracy of ±0.05 mm, and a maximum working radius of 1800 mm; the joint drive adopts a permanent magnet synchronous servo motor equipped with a 20-bit absolute value encoder, with an angle resolution of 0.00034 degrees; the mechanical arm body is made of high-strength aluminum alloy and titanium alloy composite materials, with a weight reduction of 30% while maintaining the same rigidity; the maximum speed of each joint is 210 degrees / s, and the acceleration is 2500 degrees / s2 The end effector is equipped with a water-cooled welding gun, the cooling water flow rate is 6-8 L / min, and the cooling efficiency can control the welding gun temperature below 65°C; the welding gun adopts a modular design, the replacement time is less than 30 seconds, has an automatic angle adjustment function, and the adjustment range is ±15 degrees.

[0070] Optionally, the high-performance welding power supply is an IGBT inverter digital control welding power supply, the rated output current is 400 A, the maximum pulse current can reach 600 A, the no-load voltage is 80 V, the response time is less than 1 ms; the output characteristics can be switched between constant current, constant voltage and mixed mode; the power supply efficiency is greater than 88%, and the power factor is greater than 0.93; supports multiple welding modes, including MAG, MIG, TIG and pulse MIG welding; the pulse frequency can be adjusted in the range of 30-500 Hz, and the duty cycle can be adjusted in the range of 10%-90%. The wire feeding mechanism adopts a four-wheel drive structure, the drive wheel is made of special alloy steel, and the hardness reaches HRC58-62; the wire feeding speed range is 2-20 m / min, and the speed stability is ±1.5%; it can adapt to various welding wires with a diameter of 0.8-1.6 mm. The gas protection device is equipped with an electronic flow controller, the flow range is 5-25 L / min, and the flow accuracy is ±2%; the gas preheater can heat the gas to 10-15°C above room temperature, reducing welding spatter; the gas path system adopts a quick connector design, and the pressure loss is less than 5%.

[0071] Optionally, the whole system power supply is three-phase 380V alternating current, the total power is less than 15kW; adopts a modular base design, which can realize quick installation and disassembly; the equipment covers an area of less than 2 square meters; the protection level reaches IP54, which can adapt to dust and moisture in industrial environment; the surface of the robot arm adopts a special anti-splashing coating, and the service life can reach more than 3 years; the control cabinet adopts a forced air cooling system, and the internal temperature is controlled below 40°C; all cable connectors use military-grade waterproof connectors to ensure stable and reliable signal transmission.

[0072] Optionally, the specific implementation of each module of the intelligent control system is described as follows: the motion control module adopts an algorithm architecture based on model predictive control (MPC), the control period is 1 ms, the position control accuracy is less than ±0.03 mm, and the speed control accuracy is less than ±1%; the module includes a feedforward compensation unit and a feedback adjustment unit, the feedforward compensation unit calculates joint torque based on the robot dynamics model, and the feedback adjustment unit adjusts the control amount in real time based on the encoder feedback signal; the module has a built-in collision detection function, which can stop moving within 10 ms when detecting a torque change exceeding the threshold value (usually set to 20% of the rated torque); at the same time, a compliant control algorithm is integrated, which allows the robot to exhibit controllable compliance characteristics in a preset direction, and the compliance can be adjusted in the range of 0-100%.

[0073] Optionally, the visual detection and feedback module is equipped with 2 high-speed industrial cameras and 1 structured light projector, the camera resolution is 2448x2048 pixels, and the maximum acquisition frequency is 120 frames / s; the structured light projector uses a blue light LED light source, the projected pattern is a pseudo-random stripe, and the projection accuracy can reach 0.05 mm; the module adopts a combination of binocular stereo vision algorithm and phase shift stripe analysis algorithm to realize three-dimensional reconstruction of the welding workpiece; the visual processing unit adopts GPU accelerated calculation, is equipped with 8 GB video memory, and the three-dimensional reconstruction speed can reach 15 frames / s; the module also integrates a molten pool monitoring subsystem, which includes 1 narrowband filtered high-speed camera and 1 infrared thermal imager, and can monitor the molten pool shape and temperature distribution in real time.

[0074] Optionally, the collaborative operation control module adopts a hierarchical control architecture, including a task planning layer, a trajectory planning layer and an execution control layer; the task planning layer decomposes the welding task into multiple subtasks based on the workpiece CAD model and the welding process requirements, and assigns the subtasks to each robot; the trajectory planning layer generates a collision-free trajectory by using an improved fast mosaic method, and the planning period is 100 ms; the execution control layer is responsible for the synchronous coordination of the motion of each robot, and the synchronization error is controlled within ±5 ms; the module adopts a distributed computing architecture, each robot is equipped with a local controller, communicates through real-time Ethernet, and the communication delay is less than 1 ms; the system supports up to 8 robots to work collaboratively at the same time.

[0075] Optionally, the human-machine interface uses a 19-inch capacitive touch screen with a resolution of 1920x1080 and a touch response time of less than 15 ms; the interface is developed based on HTML5 and WebGL technology and supports real-time rendering of three-dimensional scenes; a graphical welding path editing tool is provided, and the operator can adjust the welding trajectory by dragging; the system has a built-in welding process knowledge base and provides a parameter recommendation function; the interface supports multi-language switching, including Simplified Chinese, English and Russian; and has a remote monitoring function, which can remotely view the welding state through a mobile device.

[0076] The mathematical models or calculation processes involved in the present application are described in detail as follows.

[0077] The welding path key point set construction process in step S01 is specifically represented as follows:

[0078] P={p1,p2,...,p n};

[0079] In the formula, P is the welding path key point set; p i is the coordinate and attitude information of the i-th key point, p i =[x i ,y i ,z i ,α i ,βi , y i ] T , z i , y i , z i are the position coordinates of the key points in the space rectangular coordinate system, with the unit of mm; a i , b i , g i are the attitude angles of the welding torch at the key points, with the unit of degree; n is the total number of key points, usually 20-100.

[0080] The construction of the welding path curve equation adopts cubic spline interpolation algorithm, and is expressed as follows:

[0081] S j (t) = a j +b j (t-t j )+c j (t-t j ) 2 +d j (t-t j ) 3 , t∈[t j , t j+1 ];

[0082] In the formula, S j (t) is the jth spline curve function; t is the curve parameter, t∈[0, 1]; a j , b j , c j , d j are the spline function coefficients, which are obtained by solving the following equation group:

[0083] S j (t j ) = p j ;

[0084] S j (t j+1 ) = p j+1 ;

[0085] S j ′(t j+1 ) = S j+1 ′(t j+1 ) ;

[0086] S j ″(t j+1 ) = S j+1 ″(t j+1 ) ;

[0087] wherein S j ′(t) and S j(t) represents the first and second derivatives of S j The solution of the coefficient matrix is obtained by using the chasing method, and the calculation complexity is O(n).

[0088] The welding error matrix in step S03 is defined as follows:

[0089]

[0090] In the formula, E is the welding error matrix; e x is the transverse deviation, with the unit of mm, representing the offset of the actual position of the weld and the planned position in the direction perpendicular to the welding direction, ranging from-2.0 to 2.0 mm; e y is the longitudinal deviation, with the unit of mm, representing the offset of the actual position of the weld and the planned position in the welding direction, ranging from-3.0 to 3.0 mm; e θ is the angle deviation, with the unit of degree, representing the included angle between the actual posture of the welding torch and the planned posture, ranging from-5.0 to 5.0 degrees; e d is the depth deviation, with the unit of mm, representing the difference between the actual penetration and the target penetration, ranging from-1.0 to 1.0 mm.

[0091] The calculation of the welding error adopts the image processing technology, and is specifically represented as follows:

[0092] e x = x actual -x planned ;

[0093] e y = y actual -y planned ;

[0094] e θ = θ actual - θ planned ;

[0095] e d = d actual -d planned ;

[0096] In the formula, (x actual , y actual ) is the actual position coordinate of the weld obtained by image processing; (x planned , y planned ) is the theoretical position coordinate on the welding planning path; θ actual is the actual welding angle; θ planned is the planned welding angle; d actual is the actual penetration obtained by analyzing the molten pool shape; d planned is the target penetration required by the process.

[0097] The extraction of position coordinates adopts a method combining edge detection and Hough transform, which is expressed as:

[0098]

[0099] In the formula, G(x, y) is the image gradient amplitude; G x (x, y) and G y (x, y) are the gradients of the image in the x and y directions respectively; and θ(x, y) is the gradient direction angle. When G(x, y) is greater than a threshold T (usually T is 1.5 times the average gradient of the image), the point is considered as an edge point.

[0100] The Kalman filter algorithm is used for filtering processing of the welding error, and the state equation and the observation equation are as follows:

[0101] X k = AX k-1 + BU k + w k ;

[0102] Z k = HX k + v k ;

[0103] In the formula, X k is the state vector at the current time, containing the error and its rate of change; A is the state transition matrix; U k is the control input; B is the control matrix; w k is the process noise, which is subject to a Gaussian distribution with a mean of 0 and a covariance of Q; Z k is the observation vector, the superscript m represents the measured value; H is the observation matrix; v k is the observation noise, which is subject to a Gaussian distribution with a mean of 0 and a covariance of R.

[0104] The welding parameter error adjustment matrix in step S04 is defined as follows:

[0105] ΔP = CE

[0106] In the formula, ΔP is the welding parameter adjustment amount, ΔP = [ΔI, ΔU, Δv] T ; ΔI is the welding current adjustment amount, with a unit of A; ΔU is the welding voltage adjustment amount, with a unit of V; Δv is the welding speed adjustment amount, with a unit of mm / s; C is the error adjustment matrix, with a dimension of 3x4, representing the mapping relationship of the welding error to the parameter adjustment amount.

[0107] The solution of the error adjustment matrix C adopts the least square method based on historical welding data:

[0108] C = (E T E) -1 E T ΔP history ;

[0109] where E is the error matrix sample in the multi-group historical welding data, and ΔP history is the corresponding parameter adjustment record. Each element of matrix C has a clear physical meaning, for example, C 11 represents the influence coefficient of the transverse deviation on the welding current adjustment, and C 23 represents the influence coefficient of the angle deviation on the welding voltage adjustment.

[0110] The update formula of the welding parameters is:

[0111] I k+1 = I k + K I · ΔI;

[0112] U k+1 = U k + K U · ΔU;

[0113] v k+1 = v k + k v · Δv;

[0114] where I k , U k , and v k are the current welding current, voltage, and speed, respectively; I k+1 , U k+1 , and v k+1 are the adjusted welding current, voltage, and speed, respectively; K I , K U , and K v are the adjustment gain coefficients of the current, voltage, and speed, respectively, and K I is usually taken as 0.5-0.8, K U is taken as 0.6-0.9, and K v is taken as 0.4-0.7. These gain coefficients are obtained by welding experiment optimization, which includes parameter adjustment test under different workpiece materials and weld types, records the change of weld quality before and after adjustment, and determines the optimal gain coefficient based on the quality improvement rate.

[0115] The molten pool fluid dynamics model in step S06 adopts the Navier-Stokes equation for description:

[0116]

[0117] where ρ is the density of the metal liquid, and the unit is kg / m 3; is the flow velocity vector, with m / s; t is time, with s; p is pressure, with Pa; μ is dynamic viscosity, with Pa·s; is the volume force, including buoyancy, electromagnetic force, etc. is the buoyancy, where β is the thermal expansion coefficient, g is the acceleration of gravity, T is the local temperature, and T0 is the reference temperature; is the electromagnetic force, where is the current density, is the magnetic induction intensity; is the Marangoni force caused by the surface tension gradient, where γ is the surface tension coefficient.

[0118] The formula for calculating the Reynolds number is:

[0119]

[0120] In the formula, Re is the Reynolds number, dimensionless; ρ is the density of the metal liquid; v is the characteristic flow velocity, usually taking the maximum flow velocity in the molten pool; L is the characteristic length, usually taking the diameter of the molten pool; μ is the dynamic viscosity. When Re>2000, it is considered that the flow enters a turbulent state, and welding defects may be generated.

[0121] The welding quality multi-objective optimization function in step S07 is defined as follows:

[0122] F(I,U,v)=w1Q f (I,U,v)+w2Q p (I,U,v)+w3[1-P d (I,U,v)]+w4E e (I,U,v);

[0123] In the formula, F(I,U,v) is the comprehensive optimization objective function; I, U, and v are the welding current, voltage, and speed, respectively; Q f is the weld forming quality score, ranging from 0 to 1; Q p is the penetration consistency score, ranging from 0 to 1; P d is the defect rate, ranging from 0 to 1; E eThe energy efficiency score ranges from 0 to 1; w1, w2, w3, w4 are weight coefficients, and satisfy w1+w2+w3+w4=1, usually w1=0.35, w2=0.25, w3=0.3, w4=0.1. These weight coefficients can be adjusted according to specific application requirements, increase w1 when the weld appearance requirement is high, increase w2 when the welding strength requirement is high, increase w3 when the reliability requirement is high, and increase w4 when the energy cost is sensitive.

[0124] The calculation of each sub-objective function is as follows:

[0125]

[0126] P d (I, U, v)=α1P pore (I, U, v)+α2P crack (I, U, v)+α3P slag (I, U, v);

[0127]

[0128] In the formula, w a is the weld width, w a，target is the target width, w a，max is the maximum allowable width; h r is the excess height, h r，target is the target excess height, h r，max is the maximum allowable excess height; σ d is the standard deviation of penetration, d mean is the average penetration; P pore , P crack and P slag are the generation probabilities of pores, cracks and slag, respectively, which are predicted by the modified turbulence degree function, and α1, α2, α3 are weight coefficients, satisfying α1+α2+α3=1; η melt is the actual melting efficiency, and η max is the theoretical maximum melting efficiency.

[0129] Optionally, the feedforward neural network model for analyzing the welding parameter fluctuation in step S05 is represented as follows:

[0130] H1=σ(W1X+b1);

[0131] H2=σ(W2H1+b2);

[0132] H3=σ(W3H2+b3);

[0133] Y=W4H3+b4;

[0134] In the formula, X is the input vector, containing time-series data of welding current, welding voltage, and welding speed, X = [I1, I2, ..., I...]. n U1, U2, ..., U n v1, v2, ..., v n ] T H1, H2, and H3 are the output vectors of the three hidden layers, with dimensions of 64, 32, and 16, respectively; Y is the output vector representing the molten pool morphology; W1, W2, W3, and W4 are weight matrices; b1, b2, b3, and b4 are bias terms; σ is the activation function, using the ReLU function, σ(x) = max(0, x).

[0135] Optionally, the neural network training uses the backpropagation algorithm, and the loss function is defined as:

[0136]

[0137] In the formula, L is the loss function value; m is the number of training samples; Y i This is the actual output of the i-th sample; Let be the predicted output for the i-th sample; λ is the regularization coefficient, with a value of 0.001; ||W j ||2 is the L2 norm of the weight matrix, used to prevent overfitting.

[0138] Optionally, the task allocation algorithm for multi-robot collaborative welding in step S09 adopts the following mathematical model:

[0139] minZ = max i＝1，2，...，k T i ;

[0140]

[0141] x ij ∈{0, 1}, i=1, 2,...,k; j=1, 2,...,n;

[0142] In the formula, Z is the optimization objective, i.e., the maximum completion time; k is the number of robots; n is the number of welding tasks; T i t represents the total working time of the i-th robot. j x is the execution time of the j-th welding task; ij Let x be the decision variable, and when the j-th task is assigned to the i-th robot... ij =1, otherwise x ij =0. This optimization problem is NP-hard, and a greedy algorithm is used to find an approximate optimal solution.

[0143] Optionally, the robot cooperative motion planning adopts an artificial potential field-based method, and the potential field function for collision avoidance is defined as:

[0144]

[0145] In the formula, U total The total potential energy; U rep (d ij ) represents the repulsive potential field between the i-th robot and the j-th robot; d ij d0 is the minimum distance between the two robots; d0 is the safety distance threshold, usually taken as 1000mm; η is the proportional coefficient, with a value of 10. 6 The robot's motion planning needs to minimize the working time while ensuring U. total Less than the safety threshold.

[0146] Specifically, the principle of this invention is:

[0147] The technical principle of this invention for solving the problem of real-time and precise adjustment of welding process parameters is based on a closed-loop intelligent control framework of "perception-analysis-prediction-optimization-execution". Within this framework, the system first acquires the three-dimensional information of the workpiece and the state of the molten pool during the welding process through a visual detection and feedback module, constructing a welding error matrix to achieve precise perception of the welding process. Then, based on a neural network algorithm, it analyzes the fluctuations in welding parameters, identifies the dynamic characteristics of the molten pool, and combines this with a molten pool fluid dynamics model to analyze the turbulent state of metal flow within the molten pool, achieving in-depth analysis of the welding process.

[0148] The prediction stage is a key breakthrough of this invention. By establishing a dynamic compensation model for welding parameters, the system can predict the welding quality performance under different combinations of welding parameters. This model employs a lightweight deep reinforcement learning architecture, integrating a multilayer perceptron and a long short-term memory network, enabling it to capture the nonlinear mapping relationship and temporal characteristics between welding parameters and welding quality. The introduction of an attention mechanism allows the model to adaptively focus on changes in key parameters, overcoming the lag inherent in traditional feedback control.

[0149] In the optimization phase, the system applies a multi-objective optimization function for welding quality, comprehensively considering the weld pool temperature distribution, metal flow velocity, surface fluctuation amplitude, and weld formation geometry parameters to generate the optimal combination of welding parameters. This optimization process balances welding quality and energy efficiency, meeting the multi-dimensional needs of practical industrial applications. Finally, the optimized parameters are executed by the welding process control module to achieve precise control of parameters such as welding current, voltage, and speed.

[0150] The scientific and logical nature of this invention's technical solution lies in the following: First, it establishes a welding pool behavior model based on the scientific understanding of the welding physical process and combined with the principles of fluid dynamics and materials science; second, it uses modern artificial intelligence technology to handle high-dimensional nonlinear problems in the welding process, improving system robustness through a combination of data-driven and model-driven approaches; finally, it constructs a complete multi-level control architecture, forming a coordinated control strategy from key point path planning to parameter fine-tuning, ensuring that the system can effectively cope with various welding challenges.

[0151] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0152] The specific implementation of step S01 is a process of acquiring three-dimensional information of the welding workpiece based on the visual detection and feedback module. First, structured light technology is used to perform a three-dimensional scan of the welding workpiece surface to acquire point cloud data. Then, a point cloud registration algorithm is used to fuse the point cloud data acquired from multiple perspectives into a complete three-dimensional model of the workpiece. Next, an edge detection algorithm is used to extract features from the workpiece surface and identify the weld edge contour. Furthermore, a least-squares curve fitting method is applied to construct the welding path curve equation. Then, a set of key points for the welding path is generated by sampling at 5-10mm intervals on the curve. Finally, a smooth and continuous welding motion trajectory is generated based on these key points using a cubic spline interpolation algorithm. The specific representation of the set of key points for the welding path is as follows: P = {p1, p2, ..., p...} n}; where P is the set of key points in the welding path; p i p represents the coordinates and orientation information of the i-th keypoint. i =[x i y i , z i α i ,β i γ i ] T ;x i y i , z i These are the position coordinates of the key points in a Cartesian coordinate system, in mm; α i ,β i γ i S represents the welding torch attitude angle at each critical point, in degrees; n is the total number of critical points, typically 20–100. The welding path curve equation is constructed using a cubic spline interpolation algorithm, as follows: S j (t)=a j +b j (tt j )+c j (tt j ) 2 +dj (y-t j ) 3 , t e [t j , t j+1 ]; in the formula, S j (t) is the jth piecewise curve function; t is the curve parameter, t e [0, 1]; a j , b j , c j , d j are the coefficients of the spline function, which are obtained by solving the following equation group: S j (t j ) = p j ; S j (t j+1 ) = p j+1 ; S j '(t j+1 ) = S j+1 '(t j+1 ); S j ''(t j+1 ) = S j+1 ''(t j+1 ); wherein S j '(t) and S j ''(t) represent the first and second derivatives of S j (t), respectively. The solution of the coefficient matrix uses the chasing method, and the calculation complexity is O(n). The purpose of this step is to provide accurate motion path planning for the welding robot and ensure the accuracy of the welding position.

[0153] The specific implementation of step S02 is to select the welding process parameters according to the material properties of the workpiece and the shape of the weld. First, the physical property parameters of the workpiece material are retrieved from the database, including melting point, thermal conductivity, specific heat capacity, etc.; then the cross-sectional area of the weld is calculated according to the weld shape parameters; then based on the material properties and weld geometry, the fuzzy reasoning system is used to select the most matching welding process parameter template from the welding process database; further combined with the workpiece thickness and welding position, the weight coefficient correction method is applied to optimize and adjust the template parameters, and the parameter settings of the initial welding current range of 80-320 A, welding voltage range of 18-32 V, and welding speed range of 5-15 mm / s are obtained. The role of this step is to provide reasonable initial welding parameters and lay the foundation for subsequent welding process control.

[0154] The specific implementation of step S03 is to establish a welding error matrix and monitor the welding process in real time. First, a four-dimensional error matrix containing lateral deviation, longitudinal deviation, angle deviation and depth deviation is defined; then the welding process image is captured in real time by a high-speed camera, and the acquisition frequency is 60-120 Hz; then the actual position and the theoretical position of the weld are extracted by using image processing technology; further, the detection results are filtered by using Kalman filtering algorithm to eliminate the influence of random noise; at the same time, the temperature distribution of the molten pool is monitored by an infrared thermal imager, and the temperature detection range is 800-1600 ℃; finally, the deviation values in each dimension are calculated according to the image analysis results and the temperature distribution data, and the welding error matrix is updated. The welding error matrix is defined as follows: In the formula, E is the welding error matrix; e x is the lateral deviation, the unit is mm, which represents the offset of the actual position of the weld and the planned position in the direction perpendicular to the welding direction, and the range is-2.0-2.0 mm; e y is the longitudinal deviation, the unit is mm, which represents the offset of the actual position of the weld and the planned position in the welding direction, and the range is-3.0-3.0 mm; e θ is the angle deviation, the unit is degree, which represents the included angle between the actual attitude of the welding torch and the planned attitude, and the range is-5.0-5.0 degrees; e d is the depth deviation, the unit is mm, which represents the difference between the actual penetration and the target penetration, and the range is-1.0-1.0 mm. The calculation of the welding error adopts the image processing technology, and the specific representation is as follows: e x =x actual -x planned ; e y =y actual -y planned ; e θ =θ actual -θ planned ; e d =d actual -d planned ; In the formula, (x actual , y actual ) is the actual position coordinate of the weld obtained by image processing; (x planned , y planned ) is the theoretical position coordinate on the welding planning path; θ actual is the actual welding angle; θ planned is the planned welding angle; d actual is the actual penetration obtained by analyzing the molten pool shape; d planned is the target penetration required by the process. The extraction of the position coordinate adopts the method combining edge detection and Hough transformation, which is represented as: In the formula, G(x, y) is the image gradient amplitude; G x (x, y) and Gy (x, y) are the gradients of the image in x and y directions respectively; θ(x, y) is the gradient direction angle. When G(x, y) is greater than a threshold T (usually T is 1.5 times of the average gradient of the image), the point is considered as an edge point. Kalman filtering algorithm is used for filtering of welding errors, and the state equation and observation equation are as follows: k = AX k-1 + BU k + w k ; Z k = HX k + v k ; in the formula, X k is a state vector at the current time, containing error and its rate of change; A is a state transition matrix; U k is a control input; B is a control matrix; w k is a process noise, obeying a Gaussian distribution with a mean of 0 and a covariance of Q; Z k is an observation vector, the superscript m represents a measurement value; H is an observation matrix; v k is an observation noise, obeying a Gaussian distribution with a mean of 0 and a covariance of R. The purpose of this step is to obtain real-time state information of the welding process, providing a basis for subsequent parameter adjustment.

[0155] The specific implementation of step S04 is to construct a welding parameter error adjustment matrix and adjust the welding parameters. First, a linear mapping relationship matrix of the error vector to the welding parameter adjustment amount is established; then, based on the least square method, regression analysis is performed on the historical welding data to determine the numerical values of the elements in the matrix; then, the welding parameter deviation is calculated according to the welding state feedback data by using an adaptive control algorithm; further, the welding current adjustment step is set to 3-8 A, the welding voltage adjustment step is set to 0.5-1.5 V, and the welding speed adjustment step is set to 0.5-2 mm / s; finally, the adjusted parameters are transmitted to the welding power supply control system in real time through a serial communication interface. The welding parameter error adjustment matrix is defined as follows: ΔP = CE; in the formula, ΔP is the welding parameter adjustment amount, ΔP = [ΔI, ΔU, Δv] T ; ΔI is the welding current adjustment amount, with the unit of A; ΔU is the welding voltage adjustment amount, with the unit of V; Δv is the welding speed adjustment amount, with the unit of mm / s; C is the error adjustment matrix, with the dimension of 3x4, representing the mapping relationship of the welding error to the parameter adjustment amount. The solution of the error adjustment matrix C adopts the least square method, based on the historical welding data: C = (E T E) -1 E T ΔP history ; in the formula, E is the error matrix sample in the historical welding data; ΔP historycorresponding parameter adjustment record. Each element of matrix C has a clear physical meaning, for example, C 11 represents the influence coefficient of transverse deviation on welding current adjustment, C 23 represents the influence coefficient of angle deviation on welding voltage adjustment. The update formula of welding parameters is: I k+1 = I k + K I · ΔI; U k+1 = U k + K U · ΔU; v k+1 = v k + K v · Δv; in the formula, I k , U k , v k are the current welding current, voltage and speed respectively; I k+1 , U k+1 , v k+1 are the adjusted welding current, voltage and speed respectively; K I , K U , K v are the adjustment gain coefficients of current, voltage and speed, and K I usually takes a value of 0.5-0.8, K U takes a value of 0.6-0.9, and K v takes a value of 0.4-0.7. These gain coefficients are obtained by optimizing welding experiments, which include parameter adjustment tests under different workpiece materials and weld types, record the change of weld quality before and after adjustment, and determine the optimal gain coefficient based on the quality improvement rate. The role of this step is to realize the closed-loop control of welding parameters and improve the stability of the welding process.

[0156] The specific implementation of step S05 is to analyze the welding parameter fluctuation by using a neural network algorithm. First, a feedforward neural network with three hidden layers is constructed, and the number of nodes in each layer is 64, 32 and 16 respectively; then the time series data of welding current, welding voltage and welding speed are taken as input, and the molten pool shape feature is taken as output; then the back propagation algorithm is used to train the network, and the learning rate is set to 0.001-0.005; further, the trained network model is used to analyze the influence of welding parameter fluctuation on the dynamic characteristics of the molten pool in real time; finally, based on the analysis result, a welding parameter dynamic compensation model is established to realize predictive adjustment of welding parameters. The feedforward neural network model for analyzing welding parameter fluctuation by using a neural network is represented as follows: H1=σ(W1X+b1); H2=σ(W2H1+b2); H3=σ(W3H2+b3); Y=W4H3+b4; in the formula, X is an input vector, containing time series data of welding current, welding voltage and welding speed, X=[I1, I2,..., I n , U1, U2,..., Un v1, v2,..., v n ] T ; H1, H2, H3 are the output vectors of three hidden layers, with dimensions of 64, 32, 16 respectively; Y is the output vector, representing the molten pool morphology features; W1, W2, W3, W4 are weight matrices; b1, b2, b3, b4 are bias terms; σ is the activation function, using ReLU function, σ(x) = max(0, x). The neural network training uses back propagation algorithm, and the loss function is defined as: where L is the loss function value; m is the number of training samples; Y i is the actual output of the i-th sample; is the predicted output of the i-th sample; λ is the regularization coefficient, taking a value of 0.001; ||W j ||2 is the L2 norm of the weight matrix, used to prevent overfitting. The purpose of this step is to analyze the parameter fluctuation law in the welding process in depth, providing theoretical support for fine welding control.

[0157] The specific implementation of step S06 is to introduce a welding molten pool fluid dynamics model for analysis. First, the molten pool fluid mechanics equation set considering surface tension, buoyancy, and electromagnetic force is established; then the finite element method is used to discretize the molten pool area into grid elements, with a grid size of 0.1-0.5 mm; then the temperature field and velocity field at each element are calculated according to the welding process parameters; further, the Reynolds number is introduced as a judgment index of flow turbulence state, with a critical value of 2000; finally, the high turbulence area is identified by calculating the Reynolds number distribution, and the potential risk of defects such as porosity and slag inclusion is predicted. The molten pool fluid dynamics model uses Navier-Stokes equation to describe: where ρ is the metal liquid density, with a unit of kg / m 3 ; is the flow velocity vector, with a unit of m / s; t is the time, with a unit of s; p is the pressure, with a unit of Pa; μ is the dynamic viscosity, with a unit of Pa·s; is the body force, including buoyancy, electromagnetic force, etc., is the buoyancy, where β is the thermal expansion coefficient, g is the acceleration of gravity, T is the local temperature, and T0 is the reference temperature; is the electromagnetic force, where is the current density, is the magnetic induction intensity; is the Marangoni force caused by the surface tension gradient, where γ is the surface tension coefficient. The calculation formula of the Reynolds number is: where Re is the Reynolds number, dimensionless; p is the liquid metal density; v is the characteristic flow velocity, usually the maximum flow velocity in the pool; L is the characteristic length, usually the pool diameter; and m is the dynamic viscosity. When Re > 2000, the flow is considered to be turbulent, and welding defects can occur. The purpose of this step is to theoretically analyze the flow state in the pool and predict the mechanism of welding defects.

[0158] The specific implementation of step S07 is to apply a welding quality multi-objective optimization function to optimize the parameters. First, a multi-objective optimization function including weld forming quality, penetration consistency, porosity rate, and energy efficiency is constructed; then a particle swarm algorithm is used for parameter optimization, with the number of particles set to 50-100 and the maximum number of iterations set to 200; next, optimization constraints are set based on the dynamic characteristics of the welding pool and the defect risk assessment results; further, a search is performed in the three-dimensional parameter space of welding current, welding voltage, and welding speed to obtain a set of Pareto optimal solutions; finally, the parameter combination with the best comprehensive performance is selected from the set of optimal solutions as the recommended value. The welding quality multi-objective optimization function is defined as follows: F(I, U, v) = w1Q f (I, U, v) + w2Q p (I, U, v) + w3[1-P d (I, U, v)] + w4E e (I, U, v); where F(I, U, v) is the comprehensive optimization objective function; I, U, and v are the welding current, voltage, and speed, respectively; Q f is the weld forming quality score, ranging from 0 to 1; Q p is the penetration consistency score, ranging from 0 to 1; P d is the defect rate, ranging from 0 to 1; E e is the energy efficiency score, ranging from 0 to 1; w1, w2, w3, and w4 are weight coefficients, satisfying w1 + w2 + w3 + w4 = 1, and typically w1 = 0.35, w2 = 0.25, w3 = 0.3, and w4 = 0.1. These weight coefficients can be adjusted according to specific application requirements: increasing w1 for high weld appearance requirements, increasing w2 for high weld strength requirements, increasing w3 for high reliability requirements, and increasing w4 for energy cost sensitivity. The calculation of each sub-objective function is as follows: P d (I, U, v) = a1P pore (I, U, v) + a2P crack (I, U, v) + a3P slag (I, U, v); where w a is the weld width, w a，target is the target width, w a，max is the maximum allowed width; h r is the excess height, hr，target For the target height, h r，max σ is the maximum allowable residual height. d d represents the standard deviation of the melting depth. mean P represents the average melting depth. pore P crack and P slag η represents the probability of the formation of pores, cracks, and inclusions, respectively, which is predicted by a modified turbulence function. α1, α2, and α3 are weighting coefficients, satisfying α1 + α2 + α3 = 1. melt For the actual melting efficiency, η max This represents the theoretical maximum melting efficiency. The purpose of this step is to find the optimal balance among multiple welding quality indicators and provide the best combination of welding parameters.

[0159] The specific implementation of step S08 involves using a pre-trained neural network model to evaluate the weld morphology. First, the welding quality evaluation neural network model is invoked; then, weld images are acquired in real-time using a high-resolution camera with a resolution of at least 1920×1080 pixels; next, the images are pre-processed, including grayscale conversion, contrast enhancement, and noise filtering; the processed images are then input into the neural network model; finally, a welding quality score is generated based on the model's output, ranging from 0 to 100, where 85 points or above is considered excellent, 70-85 points is acceptable, and below 70 points is unacceptable. The purpose of this step is to evaluate welding quality in real-time and provide feedback for welding process control.

[0160] The specific implementation of step S09 is to integrate a multi-robot collaborative welding strategy. First, based on the geometry of the workpiece, the overall welding task is decomposed into multiple sub-tasks. Then, a task allocation algorithm is used to assign welding areas to each robot, balancing the workload of each robot. Next, a collaborative motion plan for the robots is formulated to avoid collisions between them. Furthermore, a communication protocol between the robots is established to achieve real-time sharing of status information. Finally, a master-slave control architecture coordinates the synchronous operation of multiple robots, with the welding progress of the master robot serving as the synchronization benchmark, and the progress deviation of the slave robots not exceeding 10%. The task allocation algorithm for multi-robot collaborative welding adopts the following mathematical model: minZ = max i＝1，2，...，k T i ; x ij ∈{0,1}, i=1,2,...,k; j=1,2,...,n; where Z is the optimization objective, i.e., the maximum completion time; k is the number of robots; n is the number of welding tasks; T i t represents the total working time of the i-th robot. j x is the execution time of the j-th welding task; ij Let x be the decision variable, and when the j-th task is assigned to the i-th robot...ij = 1, otherwise x ij = 0. The optimization problem belongs to NP-hard problem, and a greedy algorithm is used to solve the approximate optimal solution. The method based on artificial potential field is used for cooperative motion planning of robots, and the potential field function for avoiding collision is defined as: where U total is the total potential energy; U rep (d ij ) is the repulsive potential field between the i-th robot and the j-th robot; d ij is the minimum distance between the two robots; d0 is the safety distance threshold, usually 1000 mm; η is the proportional coefficient, taking the value of 10 6 . The motion planning of the robot needs to ensure that U total is less than the safety threshold while minimizing the working time. The purpose of this step is to improve the welding efficiency and quality and shorten the welding cycle.

[0161] The detailed structure of the dynamic compensation model of welding parameters is a hybrid neural network composed of a multi-layer perception network and a long short-term memory network. The input layer of the model receives 16-dimensional welding state data and historical parameter data; the hidden layer contains 3 fully connected layers and 2 long short-term memory network layers; the number of nodes of the fully connected layer is 128, 64, and 32 respectively, and the activation function is ReLU function; the number of units of the long short-term memory network layer is 64, which is used to learn the time sequence characteristics of the welding parameter changes; the output layer generates the compensation values of the welding current, welding voltage, and welding speed; the model uses a residual connection mechanism to combine the compensation values with the basic parameters; at the same time, an attention mechanism is introduced to give higher weight to the state information at key moments; the network optimization uses Adam optimizer with a learning rate of 0.001 and a training batch size of 32. The process of establishing the training data set includes collecting welding process data of 5 common materials under different thicknesses and welding positions; cleaning, normalizing, and time alignment of the original data; dividing the training samples using a 10s length sliding window; expanding the rare working condition samples through data enhancement technology; finally, dividing the training set, validation set, and test set according to the ratio of 7:2:1.

[0162] The detailed structure of the welding quality evaluation neural network model is a hybrid network combining a multi-layer convolutional neural network and a Transformer structure. The image feature extraction module adopts a ResNet-50 residual network structure and contains 5 residual blocks. The sound signal feature extraction module adopts a one-dimensional convolutional network and contains 4 convolutional layers with a convolution kernel size decreasing from 16 to 4. The multi-modal information fusion module adopts a 6-layer Transformer encoder structure with 8 attention heads. The final output layer is a fully connected layer with 5 nodes corresponding to 5 welding quality levels. The training data set establishment process includes collecting 100,000 sets of welding data under different process parameters; each set of data contains a sequence of welding process images, sound signals and parameter records; inviting 5 welding process experts to rate the welding quality; combining X-ray and ultrasonic detection technologies for quantitative analysis of internal defects; using image rotation, scaling, brightness adjustment and other methods for data enhancement; and finally forming a standardized training data set containing 150,000 samples.

[0163] In order to better understand and implement the present application, the following provides an embodiment 2 of a specific application scenario of the present application: researchers apply the present application to the welding processing of high-strength aluminum alloy structural parts in the production process of large aviation parts. The part is a connecting structure of a certain type of aircraft wing stringer and skin, made of 7075 aluminum alloy material, with a length of 3.6 m, a complex structure, a total weld length of more than 12 m, and containing various welding positions and weld forms. In view of the welding process difficulties of the part, the researchers built an intelligent high-performance collaborative welding robot system, realizing high-quality and high-efficiency welding production.

[0164] The welding robot system in this embodiment is composed as shown in Figure 2 The system hardware configuration adopts 3 ABB IRB4600 type 6-axis articulated robots, each robot has a load of 20 kg and a repeat positioning accuracy of ±0.05 mm, the mechanical arm is made of carbon fiber composite material, with a weight reduction of 28%, and the maximum working radius is 2.55 m. The end effector is equipped with a water-cooled TPS5000 MIG / MAG welding gun, with a cooling water flow of 7.2 L / min, and the welding gun adopts a modular design with a ±12 degree automatic angle adjustment function. The robot control cabinet adopts IRC5 multi-machine type, integrating high-precision robot control software with a minimum interpolation period of 0.4 ms.

[0165] The welding power source adopts Fronius TPS 5000 CMT digital pulse MIG welding power source, the rated output current is 500 A, the maximum pulse current can reach 680 A, the no-load voltage is 70 V, and the response time is less than 0.8 ms. The power source supports CMT (cold metal transfer) process, which can effectively reduce heat input, deformation and spatter. The wire feeding mechanism adopts a four-wheel drive structure, the wire feeding speed range is 2-25 m / min, the speed stability is ±1.2%, and the 1.2 mm ER5356 aluminum alloy welding wire is adapted. The gas protection device adopts a high-precision electronic flow controller, the flow range is 8-22 L / min, the flow accuracy is ±1.8%, and high-purity argon is used as the protective gas.

[0166] The visual detection system is equipped with 3 Basler acA2440-35um industrial cameras with a resolution of 2448x2048 pixels and a maximum acquisition frequency of 35 frames / s, and a Cognex PatMax algorithm is used to realize sub-pixel level accuracy of weld seam recognition. At the same time, a FLIR A655sc infrared thermal imager is equipped, which has a temperature measurement range of -40-650°C and a thermal sensitivity of 0.03°C, and can monitor the molten pool temperature distribution in real time. The structured light projection system uses a blue light LED light source, and the projection accuracy can reach 0.03 mm.

[0167] In terms of computing equipment, an industrial control computer equipped with an Intel Xeon E5-2697 v4 processor is used as the main control unit, which has 18 cores and 36 threads, a main frequency of 3.6 GHz, and a memory capacity of 64 GB. The graphics processing unit uses an NVIDIA Quadro P5000 with 8 GB of video memory, which is used to accelerate visual processing and neural network inference calculation. Data storage uses an NVMe SSD array with a capacity of 4 TB, and the sequential read and write speeds reach 3500 MB / s and 2700 MB / s respectively.

[0168] In this embodiment, the welding process control module performs the following steps for welding process control:

[0169] Firstly, the three-dimensional point cloud data of the wing stringer and the skin is obtained by the structured light scanning system, and the point cloud density is 0.5 points / mm 2 Based on the point cloud registration algorithm, the point cloud data collected from 8 different angles is fused. The weld seam profile is extracted by an edge detection algorithm, and the extracted weld seam accuracy reaches ±0.12 mm. Based on the extracted weld seam profile, a smooth and continuous welding trajectory is generated by using a cubic spline interpolation algorithm, and the trajectory is sampled at intervals of 8 mm to generate a set of welding path key points, a total of 1520 key points.

[0170] According to the material properties of 7075 aluminum alloy and the shape of T-shaped weld, the most matching parameter template is selected from the welding process database. The melting point of the material is 635℃, the thermal conductivity is 130W / (m·K), and the specific heat capacity is 960J / (kg·K). Welding is carried out according to the parameters shown in Table 1:

[0171] Table 1 Welding process parameters of 7075 aluminum alloy T-shaped weld

[0172]

[0173] During welding, the weld position and molten pool shape are monitored in real time by a high-speed camera with a frequency of 80Hz and an image resolution of 1280x1024 pixels. The actual position of the weld is extracted from the image by applying an edge detection algorithm, and the edge detection threshold is set to 1.6 times the average gradient. At the same time, the temperature distribution of the molten pool is monitored by an infrared thermal imager with a frequency of 60Hz. According to the detection results, the welding error matrix is calculated, and the measured data is shown in Table 2:

[0174] Table 2 Error data statistics during welding

[0175] Error type Max Min Mean Std. Dev. Lateral deviation (mm) 1.86 -1.65 0.08 0.42 Longitudinal deviation (mm) 2.21 -2.35 0.12 0.56 Angular deviation (°) 4.35 -4.12 0.28 0.98 Depth deviation (mm) 0.86 -0.92 0.05 0.24

[0176] Based on the error data, an error adjustment matrix C is constructed, and a least squares method is used to perform regression analysis on the historical welding data to obtain the error adjustment matrix as shown in Table 3:

[0177] Table 3 Error adjustment matrix of welding parameters

[0178] Coefficient Lateral deviation Longitudinal deviation Angular deviation Depth deviation Current adjustment (A / mm) 6.8 4.2 2.5 12.6 Voltage adjustment (V / mm) 1.2 0.8 0.6 2.4 Speed adjustment (mm / s mm) 0.5 0.8 0.3 1.6

[0179] A feedforward neural network is applied to analyze the influence of welding parameter fluctuations on the characteristics of the molten pool. The network contains 3 hidden layers with node numbers of 64, 32 and 16 respectively. The network is trained using the backpropagation algorithm with a learning rate of 0.003 and a training sample size of 25000 groups. The trained network model is used to analyze the parameter fluctuation law and identify the dynamic characteristics of the molten pool, and a dynamic compensation model of welding parameters is established.

[0180] A fluid dynamics model is introduced to analyze the metal flow state inside the molten pool. The finite element method is used to discretize the molten pool area into 0.2mm grid elements. Based on the Navier-Stokes equation, the flow field inside the molten pool is calculated with an average calculation time of 5ms / frame. The potential turbulent area inside the molten pool is identified by analyzing the Reynolds number, and the measured Reynolds number distribution range is 850-2200. When the Reynolds number exceeds 2000, it is determined to be turbulent flow, and the defect risk of the corresponding area increases by 46%.

[0181] The welding parameters are optimized by a multi-objective optimization function, and the weight coefficients are set as w1=0.32, w2=0.28, w3=0.27 and w4=0.13. The particle swarm algorithm is used for parameter optimization, and the particle number is set to 80 and the maximum iteration number is 180. The optimized welding parameters are shown in Table 4:

[0182] Table 4 Optimized welding process parameters

[0183]

[0184] Real-time evaluation of weld quality, using a hybrid neural network model based on ResNet-50 and Transformer structure, the image feature extraction module contains 5 residual blocks, the sound signal feature extraction module contains 4 convolution layers, and the multi-modal information fusion module adopts a 6-layer Transformer encoder structure. The model inference speed reaches 25 frames / second, and the quality evaluation accuracy reaches 93.2%.

[0185] A multi-robot collaborative welding strategy is adopted, and the welding task is allocated to 3 robots based on a task decomposition algorithm, and the allocation result is shown in Table 5:

[0186] Table 5 Multi-robot welding task allocation

[0187] Robot number Welding area Weld length (m) Estimated time (min) Workload ratio (%) Robot-1 Root area 3.8 22.5 33.8 Robot-2 Mid-wing area 4.2 21.8 32.8 Robot-3 Tip area 4.0 22.2 33.4

[0188] The synchronization of the 3 robots is coordinated through the master-slave control architecture, and the welding progress deviation is controlled within ±5.8%, effectively avoiding interference and collision between robots, and the total welding cycle is shortened by 65.2% compared with traditional single robot operation.

[0189] Traditional welding process control mainly relies on manual experience or simple closed-loop control, and there are problems such as insufficient parameter optimization, large welding quality fluctuation and low production efficiency. The traditional method usually uses a preset parameter table for welding, which cannot dynamically adjust parameters according to changes in the welding process, resulting in unstable welding quality. Especially in the welding of complex aluminum alloy structural parts, it is difficult to control the molten pool, and defects such as pores and cracks are easily produced, with low pass rate and high repair rate, which seriously affects production efficiency and product quality.

[0190] The present application realizes intelligent control and optimization of the whole welding process by introducing an intelligent high-performance collaborative welding robot system. Compared with the traditional method, the system has the following advantages: the welding defect rate is reduced by 28.5%, and the welding quality stability is improved by 26.8%. Especially in the field of welding of complex aluminum alloy structural parts, the difficult molten pool control problem that the traditional method cannot overcome is solved, and the product quality and production efficiency are significantly improved.

[0191] It should be noted that the variables involved in the present application are explained in detail as shown in the following Tables 6, 7, 8.

[0192] Table 6 Variable Explanation Table (First Part)

[0193]

[0194]

[0195] Table 7 Variable Explanation Table (Second Part)

[0196]

[0197]

[0198] Table 8 Variable Explanation Table (Third Part)

[0199]

[0200] The above description is merely a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. An intelligent high performance collaborative welding robot system, characterized by, The application relates to a multi-degree-of-freedom welding robot body, an intelligent control system, a high-performance welding power supply and auxiliary equipment; the intelligent control system comprises a motion control module, a welding process control module, a visual detection and feedback module, a collaborative work control module and a man-machine interface; the welding process control module performs the following steps: Step 1: obtaining three-dimensional information of a welding workpiece to construct a welding path; Step 2: selecting matching welding process parameters; Step 3: establishing a welding error matrix to monitor a welding process in real time, specifically: establishing a welding error matrix, defining a four-dimensional error matrix containing a transverse deviation, a longitudinal deviation, an angle deviation and a depth deviation, monitoring a weld position deviation, a molten pool shape change and a welding depth in a welding process in real time to form welding state feedback data; Step 4: constructing a welding parameter error adjustment matrix to adjust welding parameters; specifically: establishing a linear mapping relationship matrix from an error vector to a welding parameter adjustment amount; based on a least square method, historical welding data are analyzed to determine the numerical values of elements in the error adjustment matrix; welding parameter deviations are calculated according to welding state feedback data by using an adaptive control algorithm; welding current, welding voltage and welding speed are adjusted, wherein the welding parameter error adjustment matrix is defined as follows: ； In the formula, is a welding parameter adjustment amount, ; is a welding current adjustment amount; is a welding voltage adjustment amount; is a welding speed adjustment amount; is an error adjustment matrix, representing a mapping relationship of the welding error to the parameter adjustment amount; Error adjustment matrix The solution of the error adjustment matrix is based on historical welding data using a least squares method: ； In the formula, is an error matrix sample in a plurality of sets of historical welding data; is a corresponding parameter adjustment record; The updating formula of the welding parameters is: ； ； ； wherein, , , are the current welding current, welding voltage and welding speed, respectively; , , are the adjusted welding current, welding voltage and welding speed, respectively; , , are the adjustment gain factors for the welding current, welding voltage and welding speed, respectively. Step 5: analyzing welding parameter fluctuations by using a neural network algorithm, specifically: analyzing welding parameter fluctuations in a welding process by using a neural network algorithm, identifying welding molten pool dynamic characteristics and establishing a welding parameter dynamic compensation model; the welding parameter dynamic compensation model adopts a lightweight deep reinforcement learning model, and the specific structure is a hybrid neural network architecture composed of a multilayer perception network and a long short-term memory network; welding state feedback data and welding parameter historical data are received by an input layer; a hidden layer comprises multiple fully connected layers and long short-term memory network units; and welding parameter compensation values are generated by an output layer; Step 6: introducing a welding molten pool fluid dynamics model to analyze internal metal flow disorder states of the molten pool, and predicting potential welding defect risks, specifically: establishing molten pool fluid dynamics equation groups considering surface tension, buoyancy and electromagnetic force; the molten pool region is discretized into grid units by using a finite element method; temperature fields and velocity fields at each unit are calculated according to welding process parameters; a Reynolds number is introduced as a judgment index of flow disorder states; high disorder areas are identified by calculating the Reynolds number distribution, and potential welding defect risks are predicted; Step 7: applying a welding quality multi-objective optimization function to optimize parameters, specifically: constructing a multi-objective optimization function comprising weld forming quality, depth consistency, porosity and energy efficiency; parameters are optimized by using a particle swarm algorithm; optimization constraint conditions are set based on welding molten pool dynamic characteristics and defect risk evaluation results; a three-dimensional parameter space of welding current, welding voltage and welding speed is searched to obtain a Pareto optimal solution set; a parameter combination with the best comprehensive performance is selected from the optimal solution set as a recommended value. Step 8: Real-time evaluation of weld appearance using a welding quality evaluation neural network model, specifically: using a pre-trained welding quality evaluation neural network model to evaluate the weld appearance in real time, generating a welding quality prediction index, the specific structure of the welding quality evaluation neural network model is a hybrid network architecture combining multi-layer convolutional neural network and Transformer structure, including a weld image feature extraction module, a welding sound signal feature extraction module and a multi-modal information fusion module; the welding quality evaluation neural network model also includes an attention mechanism, the parameters of the attention mechanism are determined according to the welding workpiece material characteristics, the weld shape and the welding process type; Step 9: Integrate multi-robot collaborative welding strategy, assign welding area based on task decomposition algorithm, realize multi-robot synchronous welding operation.

2. The intelligent high-performance collaborative welding robot system according to claim 1, wherein, The multi-degree-of-freedom welding robot body has at least 6 degrees of freedom, including a joint type mechanical arm and an end effector, the joint type mechanical arm is made of high-strength lightweight material, and the end effector is equipped with a welding gun; the high-performance welding power supply and auxiliary equipment include a high-performance welding power supply, a wire feeding mechanism and a gas protection device.

3. The intelligent high-performance collaborative welding robot system according to claim 2, wherein, The welding process control module performs the step of acquiring the three-dimensional information of the welding workpiece to construct the welding path as follows: acquiring the three-dimensional information of the welding workpiece based on the visual detection and feedback module, constructing a set of key points of the welding path, and generating a welding motion trajectory plan.

4. The intelligent high-performance collaborative welding robot system according to claim 3, wherein, The welding process control module performs the step of selecting matching welding process parameters as follows: selecting matching welding process parameters from the welding process database according to the welding workpiece material characteristics and the weld shape, and generating initial welding parameter settings.

Images