Intelligent path planning system and optimization method for large structure robot welding

Abstract

The invention relates to the technical field of robot welding, in particular to an intelligent path planning system and optimization method for large structure robot welding, by collecting three-dimensional point cloud data of a workpiece and temperature field and stress field data in the welding process, a geometry-heat-force multi-field coupling model is established, and a machine learning algorithm is adopted for parameter correction; based on the model, a deep learning network is adopted to identify welding seam topological features and generate a path constraint parameter set; welding path nodes are optimized by using a hybrid optimization strategy of reinforcement learning and a multi-target genetic algorithm; monitoring the shape of a molten pool through a visual sensing system, and dynamically adjusting welding parameters in combination with a model predictive control algorithm; a distributed model predictive control algorithm is adopted to coordinate the motion trails of the multiple robots; and finally, constructing a welding process knowledge graph to realize continuous optimization of the system. The problems of inaccurate welding deformation prediction and poor path planning adaptability in the prior art are solved, and the welding quality and the production efficiency of the large structural part are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of robotic welding technology, and in particular to an intelligent path planning system and optimization method for robotic welding of large structures. Background Technology

[0002] Existing robotic welding path planning technologies are mostly based on offline programming using geometric models, lacking consideration for the thermo-mechanical coupling effects during welding, making it difficult to accurately predict and control welding deformation. Furthermore, traditional methods often fail to adaptively generate optimal path parameters when dealing with complex weld topologies, leading to unstable welding quality. Especially in multi-robot collaborative operation scenarios, existing systems lack effective coordination mechanisms, making it difficult to guarantee motion synchronization and trajectory accuracy among multiple robots.

[0003] In terms of welding process control, existing technologies typically employ fixed process parameters, which cannot be dynamically adjusted based on real-time changes in the molten pool morphology. Although some studies have attempted to introduce visual sensing technology, most systems lack deep integration with physical models, making it difficult to achieve precise optimization of process parameters. Furthermore, traditional welding systems lack knowledge accumulation and self-optimization capabilities, requiring readjustment of process parameters for each new workpiece, severely impacting production efficiency.

[0004] In recent years, new technologies such as digital twins, machine learning, and multi-sensor fusion have provided new solutions for welding path planning. However, existing technologies lack comprehensive consideration of multiple objectives in the path optimization process, especially in the welding of large structural components, where problems such as insufficient accuracy, poor adaptability, and low optimization efficiency still exist. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent path planning system and optimization method for welding large-scale structures by robots, which solves the problems of unstable welding quality, low production efficiency, and poor adaptability to new workpieces caused by the inability of existing large-scale structure robot welding path planning technology to accurately control welding deformation under the thermo-mechanical coupling effect.

[0006] To achieve the above objectives, this invention provides an intelligent path planning system and optimization method for robotic welding of large structures, comprising the following steps:

[0007] Collect three-dimensional point cloud data of the workpiece surface, and simultaneously collect temperature field distribution data and stress field change data during the welding process;

[0008] Using the collected point cloud data, temperature field data, and stress field data, a geometric-thermal-mechanical multi-field coupling model is established. The model parameters are corrected through machine learning algorithms, and welding deformation prediction data is output.

[0009] Based on the multi-field coupling model, a deep learning network is used to identify the topological features of the weld seam and generate a set of path constraint parameters including the curvature change rate, the penetration depth threshold and the assembly deviation tolerance.

[0010] Based on the path constraint parameter set, a hybrid optimization strategy of reinforcement learning and multi-objective genetic algorithm is adopted, with welding defect rate, energy consumption and working time as optimization objectives, and welding path nodes are optimized through deep deterministic strategy gradient algorithm.

[0011] By monitoring the morphological characteristics of the molten pool through a visual sensing system and combining the prediction results of a multi-field coupling model, the welding parameters are dynamically adjusted using a model predictive control algorithm.

[0012] The motion trajectories of multiple robots are coordinated by using a distributed model predictive control algorithm through a communication network and computing platform.

[0013] Based on various data from the welding process, a welding process knowledge graph is constructed. A neural network is used to establish a mapping relationship between process parameters and welding quality, thereby enabling continuous optimization of the path planning system.

[0014] Specifically, the method for acquiring three-dimensional point cloud data of the workpiece surface includes:

[0015] Configure the laser scanner's operating parameters, including setting the scanning accuracy and scanning speed;

[0016] Plan the scanning path and use multiple scanning angles to completely scan the workpiece surface;

[0017] Control the laser scanner to perform point cloud data acquisition along a predetermined path;

[0018] The collected point cloud data is registered to generate a complete three-dimensional geometric model.

[0019] Specifically, the method for collecting temperature field distribution data and stress field change data during the welding process includes:

[0020] An infrared thermal imager monitoring system was deployed, and the monitoring area was set up to cover the weld and its heat-affected zone.

[0021] Install a stress sensor array, arranging multiple sensors at predetermined intervals along the length of the weld seam;

[0022] Synchronously collect and record process parameters such as current, voltage, and welding speed during the welding process.

[0023] Specifically, the method for establishing a geometric-thermal-mechanical multi-field coupling model includes:

[0024] Establish a heat conduction control equation that incorporates the latent heat effect of phase change;

[0025] Construct stress-strain equations based on material constitutive relations;

[0026] Set thermal boundary conditions and mechanical boundary conditions;

[0027] The multiphysics coupling equations are solved using a numerical iterative method.

[0028] Specifically, the method for correcting model parameters using machine learning algorithms includes:

[0029] Construct a neural network model, with inputs including material parameters and welding process parameters;

[0030] Define a loss function with prediction accuracy as the objective;

[0031] Gradient descent algorithm is used to optimize network parameters and improve model prediction accuracy.

[0032] Specifically, the method for identifying weld topology features using deep learning networks includes:

[0033] Construct a deep neural network architecture, including encoder and decoder structures;

[0034] Introducing attention mechanisms into networks enhances feature extraction capabilities;

[0035] A multi-scale feature fusion method is used to process semantic information at different levels;

[0036] By applying transfer learning techniques, pre-trained models can be used to accelerate network convergence.

[0037] In the process of identifying weld topology features using a deep learning network, the method further includes the following network training steps:

[0038] Prepare a training sample dataset containing multiple weld types;

[0039] Configure hyperparameters for network training, including learning rate and batch size;

[0040] The network weight parameters are updated using an optimized algorithm;

[0041] The training process is monitored using a validation set to prevent overfitting.

[0042] On the other hand, the present invention also includes an intelligent path planning system for large-scale robotic welding, used to implement the method, comprising the data acquisition module, multi-field coupling modeling module, weld seam recognition module, path optimization module, real-time control module, collaborative control module, and knowledge management module.

[0043] The data acquisition module is used to acquire three-dimensional point cloud data of the workpiece surface, and simultaneously acquire temperature field distribution data and stress field change data during the welding process.

[0044] The multi-field coupling modeling module uses the collected point cloud data, temperature field data, and stress field data to establish a geometric-thermal-mechanical multi-field coupling model, corrects the model parameters through machine learning algorithms, and outputs welding deformation prediction data.

[0045] The weld identification module, based on a multi-field coupling model, uses a deep learning network to identify weld topological features and generate a set of path constraint parameters including curvature change rate, penetration depth threshold, and assembly deviation tolerance.

[0046] The path optimization module is used to optimize the welding path nodes by adopting a hybrid optimization strategy of reinforcement learning and multi-objective genetic algorithm based on the path constraint parameter set, with welding defect rate, energy consumption and working time as optimization objectives, and through deep deterministic strategy gradient algorithm.

[0047] The real-time control module is used to monitor the morphological characteristics of the molten pool through a visual sensing system, and dynamically adjust the welding parameters by combining the prediction results of the multi-field coupling model with the model prediction control algorithm.

[0048] The collaborative control module is used to coordinate the motion trajectories of multiple robots through a communication network and a computing platform, using a distributed model predictive control algorithm.

[0049] The knowledge management module is used to construct a welding process knowledge graph based on various data in the welding process, and to establish a mapping relationship between process parameters and welding quality through neural networks, so as to realize the continuous optimization of the path planning system.

[0050] This invention discloses an intelligent path planning system and optimization method for robotic welding of large structures. Through the deep integration of multi-physics field coupling modeling and intelligent algorithms, it achieves intelligent optimization and control of the entire welding process for large structures. By establishing a geometric-thermal-mechanical multi-field coupling model and employing machine learning algorithms for parameter correction, it achieves accurate prediction of welding deformation, effectively controlling welding deformation and reducing the welding defect rate to below 2%. Using deep learning-based weld feature recognition technology, it achieves an accuracy rate of over 98% in recognizing various weld types such as circumferential welds and fillet welds, automatically generating optimal path parameters and reducing manual programming time. Through a hybrid optimization strategy combining reinforcement learning and multi-objective genetic algorithms, it optimizes the welding path and process parameters while ensuring welding quality, reducing energy consumption and achieving green manufacturing. This method is applicable to the welding and manufacturing of various large structural components in hydropower and thermal power plants, and can handle workpieces with thicknesses ranging from 40-250 mm, showing broad industrial application prospects. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0052] Figure 1 This is a flowchart of the intelligent path planning and optimization method for large-structure robotic welding according to the present invention.

[0053] Figure 2 This is a schematic diagram of the intelligent path planning system for large-scale robotic welding of the present invention.

[0054] In the diagram: 101-Data acquisition module, 102-Multi-field coupling modeling module, 103-Weld seam recognition module, 104-Path optimization module, 105-Real-time control module, 106-Collaborative control module, 107-Knowledge management module. Detailed Implementation

[0055] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, but should not be construed as limiting the present invention.

[0056] Please see Figure 1 ,in Figure 1 This is a flowchart of an intelligent path planning and optimization method for robotic welding of large structures.

[0057] This invention provides an intelligent path planning system and optimization method for robotic welding of large structures, comprising the following steps:

[0058] S1: Collect three-dimensional point cloud data of the workpiece surface, and simultaneously collect temperature field distribution data and stress field change data during the welding process.

[0059] S11: Configure the operating parameters of the laser scanner, including setting the scanning accuracy and scanning speed.

[0060] S12: Plan the scanning path and use multiple scanning angles to perform a complete scan of the workpiece surface.

[0061] S13: Control the laser scanner to perform point cloud data acquisition along a predetermined path.

[0062] S14: Perform registration processing on the collected point cloud data to generate a complete three-dimensional geometric model.

[0063] S15: Deploy an infrared thermal imager monitoring system and set up a monitoring area to cover the weld and its heat-affected zone.

[0064] S16: Install a stress sensor array, with multiple sensors arranged at predetermined intervals along the length of the weld.

[0065] S17: Synchronously collect and record process parameters such as current, voltage, and welding speed during the welding process.

[0066] In this embodiment, an industrial-grade high-precision laser scanner is first configured for the 3D point cloud acquisition stage. Scanning parameters are set based on the complexity of the workpiece area: for complex curved surfaces such as the rotor winding ends and the top cover seam, a scanning accuracy of ±0.03mm and a low scanning speed of 50mm / s are used to capture detailed features; for regular areas such as the outer circle of the retaining ring and the bottom ring plane, a high-speed scanning speed of 100mm / s is used to improve efficiency. Simultaneously, dynamic noise reduction is enabled to control ambient light interference errors to ≤0.01mm, ensuring a point cloud data signal-to-noise ratio ≥30dB. The scanning path planning adopts a "layered + partitioned" strategy. Vertically, each 50mm layer covers the entire height of the workpiece. Circumferentially, for workpieces with circumferential welds, the scanning area is divided into 30-45° sections with an overlap rate of ≥30% between adjacent areas. For the weld area, an additional 2-3 15-20° tilted scanning angles are added to avoid missing key features such as bevels and root gaps. During point cloud acquisition, an industrial robot equipped with a scanner and a positioning accuracy of ≤0.02mm moves along a preset path. Data is transmitted in real-time to edge computing nodes via gigabit Ethernet with a latency of ≤10ms. A breakpoint resume scanning function is also implemented to prevent repeated scanning due to equipment failure. In the point cloud processing stage, outliers and invalid points are first removed using 3σ threshold statistical filtering and 5mm radius filtering. Then, coarse registration using SAC-IA (error ≤0.5mm) and fine registration using ICP (≥50 iterations) are performed to generate a 3D geometric model. This model is then compared with the design drawings to verify key dimensions. If the deviation exceeds 0.1mm, scanning and registration are re-executed. For temperature field acquisition, a high-resolution infrared thermal imager is used. One imager is placed every 1m along the weld length at 50-100mm positions on both sides of the weld to ensure complete coverage of the weld and a 50mm heat-affected zone. The sampling frequency is set to 10Hz to capture dynamic temperature changes during the welding process. Stress field acquisition utilizes high-temperature resistant piezoelectric stress sensors, with one sensor every 500 mm along the weld length in conventional areas and one sensor every 200 mm in stress concentration areas such as weld corners and joints. The sensors are fixed with a high-temperature adhesive and covered with a heat-insulating layer to prevent damage from high welding temperatures. Process parameters are acquired in real-time via the welding power supply data interface (current and voltage) and welding speed via a robot encoder at a frequency of 50 Hz. All sensor data carries a 1 ms timestamp and is calibrated using the workpiece design reference coordinate system as a unified spatial reference, achieving a strict spatiotemporal correlation between the 3D point cloud, temperature field, stress field, and process parameters. A real-time data verification mechanism is simultaneously activated during acquisition to filter out anomalies in the point cloud data and monitor out-of-range values ​​for temperature, stress, and process parameters, ensuring the validity of the acquired data and ultimately forming an integrated original dataset encompassing spatiotemporal, geometric, thermal, mechanical, and process parameters.

[0067] S2: Using the collected point cloud data, temperature field data, and stress field data, a geometric-thermal-mechanical multi-field coupled model is established. The model parameters are corrected through machine learning algorithms, and welding deformation prediction data is output.

[0068] S21: Establish the heat conduction control equation that includes the latent heat effect of phase change.

[0069] S22: Construct stress-strain equations based on material constitutive relations.

[0070] S23: Set thermal boundary conditions and mechanical boundary conditions.

[0071] S24: Solve the multiphysics coupling equations using a numerical iterative method.

[0072] S25: Construct a neural network model, with inputs including material parameters and welding process parameters.

[0073] S26: Define a loss function with prediction accuracy as the objective.

[0074] S27: Use gradient descent algorithm to optimize network parameters and improve model prediction accuracy.

[0075] In this embodiment, when constructing the geometry-thermal-mechanical multi-field coupling model, the Fourier heat conduction equation is first used as the basic framework, and the latent heat of phase change internal heat source term is introduced to form a complete heat conduction control equation; for mainstream materials such as Q235 steel, the thermophysical parameters are dynamically set: the density is taken as 7850 kg / m³. 3 The specific heat capacity varies with temperature, being 465 J / (kg·K) at 20℃ and increasing to 650 J / (kg·K) at 600℃. The thermal conductivity is 48 W / (m·K) at 20℃. The latent heat of phase change is 270 kJ / kg in the Ac1~Ac3 range (723~910℃), incorporated into the equation as a volumetric heat source. The arc heat input density is calculated using current, voltage, and welding speed collected by S17. The MIG welding thermal efficiency is set to 0.75 to ensure the heat conduction process matches the actual welding heat input. The stress-strain equation is constructed based on the von Mises yield criterion, employing an elastoplastic thermal constitutive model to reflect the influence of temperature on mechanical properties. Boundary conditions are strictly matched to the actual welding scenario: for the thermal boundary, the weld area is loaded with arc heat flux density, the outer surface uses a convection-radiation composite boundary, and the air environment convection heat transfer coefficient is 15 W / (m·K). 2 ·K), with a forced ventilation rate of 25W / (m²). 2The emissivity was set to 0.85 (K), and the tooling contact area was set to a workshop ambient temperature of 20–25℃. The mechanical boundaries were set according to the workpiece constraint characteristics. Elastic constraints were used at the bottom of the hydroelectric bottom ring, while only the Z-axis displacement was constrained at the top of the top cover. The constraint positions corresponded one-to-one with the tooling positioning points in the S14 3D model. Numerical solutions employed the finite element method combined with a sequential coupling strategy. The weld zone and heat-affected zone were divided into 2mm hexahedral structured meshes, while other areas used 5–10mm unstructured meshes, with the total number controlled between 1 million and 2 million. The time step was set to 0.1s, synchronized with the S1 temperature acquisition frequency. Each step first solved the heat conduction equation, with a convergence criterion of temperature error ≤0.5℃. Then, the temperature field was substituted as a load into the stress equation for solution, with a convergence criterion of stress error ≤5MPa. During iteration, S1 measured data was used for verification; if the deviation exceeded 10%, the thermal boundary parameters were dynamically adjusted. In the model parameter calibration stage, a three-layer perceptron neural network was constructed. The input layer contained 12 features, including 5 material parameters and 7 process parameters, all derived from the S1 data. The output layer was the predicted welding deformation value. The training samples covered 1200 historical cases, including 5 types of large structural components, and the dataset was divided in a 7:2:1 ratio. The loss function adopted was weighted mean square error, with a weight of 1.5 for critical areas such as the weld center and 1.0 for non-critical areas. L2 regularization was also introduced with λ = 0.001 to prevent overfitting. The optimization adopted the Adam gradient descent algorithm with an initial learning rate of 0.001, using cosine annealing decay, decreasing by 10% every 100 rounds, a batch size of 32, and a total of 500 rounds. An early stopping strategy was combined, stopping the algorithm if the validation set loss did not decrease for 3 consecutive rounds.

[0076] S3: Based on a multi-field coupling model, a deep learning network is used to identify weld topology features and generate a set of path constraint parameters including curvature change rate, penetration depth threshold and assembly deviation tolerance.

[0077] S31: Construct a deep neural network architecture, including encoder and decoder structures.

[0078] S32: Introduce an attention mechanism into the network to improve feature extraction capabilities.

[0079] S33: Employ a multi-scale feature fusion method to process semantic information at different levels.

[0080] S34: Apply transfer learning techniques to accelerate network convergence using pre-trained models.

[0081] S35: Prepare a training sample dataset containing multiple weld types.

[0082] S36: Set the hyperparameters for network training, including the learning rate and batch size.

[0083] S37: Update network weight parameters using an optimization algorithm.

[0084] S38: Monitor the training process using a validation set to prevent overfitting.

[0085] In this embodiment, based on a multi-field coupling model, a deep learning network is used to identify weld topological features and generate a path constraint parameter set including curvature change rate, weld penetration threshold, and assembly deviation tolerance. Specifically, this is achieved through the following process: An improved U-Net deep neural network architecture is constructed. The encoder has a 5-layer downsampling module to extract global weld topological features, and the decoder has a 5-layer upsampling module to supplement detailed information such as bevel angles. A 3-channel feature map is output, corresponding to the weld mask, centerline, and key points. A spatial attention module is embedded in the encoder-decoder splicing layer. Feature vector fusion generates a weight map to strengthen features in key areas such as bevels, reducing interference and improving weld edge recognition accuracy by more than 20%. A multi-scale feature fusion method of "lateral connection + cross-layer fusion" is adopted. Laterally splicing features at different levels preserves small weld details, while cross-layer fusion strengthens large-scale features. Weld seam recognition was performed. Transfer learning was applied, and the encoder was initialized with ResNet50 pre-trained weights. 8000 samples covering 5 typical weld seam types were prepared and divided into a 7:2:1 dataset after data augmentation. The input size was set to 512×512, with 300 iterations, an initial learning rate of 0.001, and a cosine annealing strategy. The batch size was 16 (reduced to 8 when overflowing), and the loss function was a weighted sum of cross-entropy and Dice loss (1:1). The AdamW optimization algorithm was used to update the network weights (β1 = 0.9, β2 = 0.999, weight decay 0.001), combined with a gradient clipping threshold of 5.0 to ensure training stability. Early stopping, Dropout layer (probability 0.2), and sample supplementation were validated through 10 rounds to prevent overfitting. The final test set weld seam segmentation IoU ≥ 0.92 and feature point localization error ≤ 2 pixels.

[0086] S4: Based on the path constraint parameter set, a hybrid optimization strategy of reinforcement learning and multi-objective genetic algorithm is adopted, with welding defect rate, energy consumption and working time as optimization objectives, and welding path nodes are optimized through deep deterministic strategy gradient algorithm.

[0087] S41: Establish a multi-objective optimization function and define the weight coefficients of each optimization objective.

[0088] S42: Use a multi-objective optimization algorithm to search for the Pareto optimal solution set.

[0089] S43: Apply deep reinforcement learning algorithms to further optimize path nodes.

[0090] S44: Set up an experience replay buffer to store state-action data during the optimization process.

[0091] In this implementation, based on the path constraint parameter set generated by S3, the optimization of welding path nodes is achieved through a hybrid strategy of "global optimization using a multi-objective genetic algorithm + local fine-tuning using deep reinforcement learning." The core objective is to simultaneously balance the three objectives of welding defect rate, energy consumption, and time. Specifically, a multi-objective optimization function is first constructed, using welding defect rate, energy consumption per unit length, and total time as the core optimization dimensions. The weight coefficients of each objective are dynamically set according to the workpiece type and production requirements. For critical structures such as thermal power rotors, which have extremely high welding quality requirements, the defect rate weight is set to the highest (50%), while energy consumption and time account for 20% and 30%, respectively. For batch-produced components such as hydropower bottom rings, the time weight is increased to 40%, with defect rate and energy consumption each accounting for 30%. The weight coefficients are calibrated using process expert experience and historical data to ensure they align with actual production priorities. Subsequently, a multi-objective genetic algorithm is used for global search, encoding the three-dimensional coordinates of the welding path nodes into chromosomes of real numbers. The initial population size is set to 100, and all initial individuals must meet the S3 constraint parameter requirements. The population is iteratively updated using operators such as tournament selection, simulated binary crossover, and polynomial mutation. Each generation uses non-dominated sorting to select path schemes not dominated by other solutions, and combines crowding distance to ensure solution diversity. After 50 iterations, a Pareto optimal solution set containing 20 to 30 effective path node sequences is obtained, with each solution corresponding to a path scheme that balances the three objectives. Then, a deep deterministic policy gradient algorithm is applied to locally fine-tune the Pareto solution, transforming path optimization into a decision problem in a continuous action space: information such as current path node coordinates, curvature change rate deviation, real-time welding defect rate, and remaining working time are integrated into a "state," and the fine-tuning amount of the node coordinates is taken as an "action." A reward function is designed: if the current fine-tuning reduces the defect rate, energy consumption, or working time, a positive reward is given; if the S3 constraint parameters are violated, an additional penalty is imposed. Simultaneously, an Actor-Critic dual-network architecture is constructed: the Actor network outputs node fine-tuning actions through a 3-layer fully connected structure, and the Critic network outputs value evaluations based on the matching degree between state and action. The two networks learn collaboratively to optimize path nodes, improving path accuracy in key areas. To ensure the stability of reinforcement learning training, an experience replay buffer with a capacity of 100,000 data points was set up. The data set of "current state, action executed, reward obtained, and next state" for each decision during the optimization process was stored in the buffer. During each network training, 256 data sets were randomly selected from the buffer for batch updates to avoid training oscillations caused by data correlation. Finally, after 2,000 rounds of training, the set with the best overall performance was selected from the optimized path schemes as the final path node sequence for robot welding.

[0092] S5: By monitoring the morphological characteristics of the molten pool through a visual sensing system and combining the prediction results of a multi-field coupling model, the welding parameters are dynamically adjusted using a model predictive control algorithm.

[0093] S51: Configure the image acquisition system, including selecting appropriate camera equipment and lighting devices.

[0094] S52: Design the image processing workflow, including image preprocessing, feature extraction, and parameter calculation.

[0095] S53: Establish a model predictive controller and set the prediction time domain and control time domain parameters.

[0096] S54: Design the weight matrix of the control system to balance the system response speed and stability.

[0097] In this embodiment, a high-precision image acquisition system is used to capture millisecond-level dynamic changes in the molten pool. The illumination device is equipped with a coaxial laser-assisted light source, which enhances the contrast of the molten pool edge through diffuse reflection. The camera and light source are integrated at a 30° tilt to the side of the robotic welding torch, ensuring complete field of view coverage of the molten pool and its surrounding 5mm area. Data transmission uses a fiber optic link to control latency to within 10ms. The image processing workflow adopts a "layered processing + feature locking" strategy: in the preprocessing stage, Gaussian filtering removes image noise, adaptive threshold segmentation separates the molten pool area from the background, and morphological operations are used to eliminate arc artifacts; the feature extraction stage focuses on identifying the geometric and dynamic features of the molten pool, fitting the molten pool edge curve using the least squares method, and calculating its instantaneous morphological parameters; the parameter calculation module quantifies the extracted features into controllable indicators, such as the rate of change of molten pool area and edge irregularity, with the processing cycle synchronized with the camera frame rate. The model predictive controller is built upon a multi-field coupled model: the prediction time domain is set to 50ms. Based on the temperature field prediction data output by S2 (such as the trend of the molten pool center temperature over time), combined with the current molten pool morphology characteristics, the evolution state of the molten pool within the next 50ms is predicted. The controller integrates a dynamic error correction mechanism, using the deviation between the real-time molten pool characteristics and the model prediction values ​​as feedback to correct the prediction model parameters for the next time domain. The design of the control system's weight matrix balances response speed and adjustment stability: high weights (60%) are assigned to key molten pool feature deviations (such as width and area deviations) to ensure rapid correction of deviations from the target morphology (such as rapidly reducing welding current when the molten pool is too wide); 30% weight is assigned to the rate of change of welding parameters (such as current adjustment amplitude and voltage fluctuation) to avoid drastic fluctuations in the molten pool caused by sudden parameter changes; 10% weight is reserved for the stress field prediction results of the multi-field coupled model (such as when local stress exceeds the material yield strength), prioritizing the structural mechanical properties. Ultimately, this achieves stable and controllable molten pool during welding, improving the consistency of welding quality for large structural components.

[0098] S6: Through communication networks and computing platforms, a distributed model predictive control algorithm is used to coordinate the motion trajectories of multiple robots.

[0099] S61: Deploy a high-speed communication network to ensure the real-time performance and reliability of data transmission.

[0100] S62: Configure edge computing nodes and allocate computing resources and storage space.

[0101] S63: Design a distributed optimization algorithm to achieve collaborative control of multiple robots.

[0102] S64: Establish a time synchronization mechanism to ensure the coordinated and consistent movement of multiple robots.

[0103] In this implementation, an integrated architecture of "high-speed communication - edge computing - distributed optimization - time synchronization" is used, relying on distributed model predictive control algorithms to achieve precise coordination of multi-robot motion trajectories, ensuring weld seam connection accuracy and motion safety. Specifically, a high-speed communication network integrating "5G + Industrial Ethernet" is first deployed: 5G adopts an independent networking mode, with the core network deployed locally in the workshop, and the base station coverage radius controlled within 50m to ensure signal stability during robot welding; the Industrial Ethernet uses a gigabit ring network architecture to connect each robot controller to the edge computing nodes, with the two network links being redundant. The edge computing nodes adopt a "master-slave" distributed deployment: one master edge node is set up in the center of the workshop, responsible for global task allocation and conflict arbitration; each robot corresponds to one slave edge node, which handles the robot's trajectory planning and sensor data preprocessing tasks locally. The computing resource allocation adopts a dynamic scheduling strategy: when the robot performs complex circumferential weld welding, the master node temporarily allocates additional computing power to the corresponding slave node to ensure the real-time performance of trajectory optimization calculation; the storage space is designed with "local caching + global sharing", the slave node caches the real-time data of the local robot, and the master node stores the historical collaborative data of all robots, supporting fast data retrieval and backtracking analysis. The distributed optimization algorithm is designed based on the Distributed Model Predictive Control (DMPC) framework: each robot is regarded as an independent control subsystem, and each subsystem exchanges its own motion state (position, velocity, acceleration) and trajectory planning results in real time through a communication network; the master edge node decomposes the local path tasks of each robot according to the global path optimized by S4 (e.g., when welding the top cover of water and electricity, two robots are responsible for the symmetrical side weld seam, and a 50mm overlap area is reserved at the path connection); each slave node calculates its own motion command for the next moment through rolling time domain optimization based on local sensor data (e.g., weld seam position deviation monitored by vision) and the state information of adjacent robots. If a collision risk with other robots is predicted (distance ≤ 50mm), a conflict arbitration mechanism is triggered, and the trajectory is adjusted according to task priority (the priority of the critical weld seam welding robot is higher than that of the auxiliary robot), with the adjustment range controlled within ±0.2mm to ensure that the position accuracy of the weld seam connection is ≤0.1mm.The time synchronization mechanism adopts the IEEE 1588 Precision Time Protocol (PTP): the GPS clock module on the main edge node is used as the global reference clock, and each robot controller, slave edge node, and sensor device acts as a slave clock, periodically receiving the time synchronization signal of the master clock through industrial Ethernet; a time compensation module is set in each robot controller to correct the deviation between the local clock and the master clock in real time, ensuring that the execution time of the motion command and the acquisition time of the sensor data of all robots are strictly aligned. For example, when two robots start welding simultaneously, the motion start time difference is ≤3ms, and the welding speed synchronization error at the weld joint is ≤0.1mm / s, avoiding problems such as weld misalignment and incomplete fusion caused by time asynchrony. Ultimately, the efficiency of multi-robot collaborative welding is improved by more than 30%, and the collaborative trajectory accuracy fully meets the requirements of overall welding of large structural components.

[0104] S7: Based on various data in the welding process, construct a welding process knowledge graph, establish a mapping relationship between process parameters and welding quality through neural networks, and realize continuous optimization of the path planning system.

[0105] In this implementation, a basic library is first constructed by integrating multi-source data generated from S1 to S6, covering material properties, process parameters, process monitoring data, quality results, and path planning records. After cleaning outliers, standardizing formats, and associating annotations, a structured dataset with feature dimensions is formed. Based on this, a welding process knowledge graph is constructed, adopting a three-layer architecture of "entity-relationship-attribute": the entity layer contains core elements such as material type, process parameters, and quality indicators; the relationship layer defines the association rules between entities; and the attribute layer supplements details. Knowledge extraction combines rules and machine learning: structured data is mined using association rules to generate triples, and unstructured data uses the BERT model to extract implicit knowledge, ultimately forming a graph containing entities and relationships. Next, a graph neural network (GNN) model is constructed, using entities in the knowledge graph as nodes and relationships as edges, with process parameters as node features and welding quality indicators (defect rate, deformation) as output targets. Finally, the system is continuously optimized through this mapping relationship: the GNN prediction results are fed back to the hybrid optimization strategy of S4; under new operating conditions, the knowledge graph is automatically called to query similar cases and the optimization target weights are dynamically adjusted; the knowledge graph is updated with new data every month, and the GNN model is retrained every quarter to adapt to process changes; a human-machine collaboration interface is set up to allow experts to correct abnormal rules in the graph.

[0106] On the other hand, please see Figure 2 , Figure 2This is a schematic diagram of an intelligent path planning system for large-structure robotic welding. The invention also includes an intelligent path planning system for implementing the method, comprising a data acquisition module 101, a multi-field coupling modeling module 102, a weld identification module 103, a path optimization module 104, a real-time control module 105, a collaborative control module 106, and a knowledge management module 107. The data acquisition module 101 is connected to the multi-field coupling modeling module 102 and the knowledge management module 107, respectively. The multi-field coupling modeling module 102 is connected to the weld identification module 103, the real-time control module 105, the collaborative control module 106, and the knowledge management module 107, respectively. The knowledge management module 107 is connected, and the weld seam recognition module 103 is connected to the path optimization module 104 and the knowledge management module 107 respectively. The path optimization module 104 is connected to the real-time control module 105, the collaborative control module 106, and the knowledge management module 107 respectively. The real-time control module 105 is connected to the collaborative control module 106 and the knowledge management module 107 respectively. The collaborative control module 106 is connected to the knowledge management module 107 respectively. The knowledge management module 107 is connected to the multi-field coupling modeling module 102, the path optimization module 104, and the real-time control module 105 respectively.

[0107] The data acquisition module 101 is used to acquire three-dimensional point cloud data of the workpiece surface and simultaneously acquire temperature field distribution data and stress field change data during the welding process.

[0108] The multi-field coupling modeling module 102 uses the collected point cloud data, temperature field data and stress field data to establish a geometric-thermal-mechanical multi-field coupling model, corrects the model parameters through machine learning algorithms, and outputs welding deformation prediction data.

[0109] The weld identification module 103, based on a multi-field coupling model, uses a deep learning network to identify weld topological features and generate a set of path constraint parameters including curvature change rate, penetration depth threshold and assembly deviation tolerance.

[0110] The path optimization module 104 is used to optimize the welding path nodes by adopting a hybrid optimization strategy of reinforcement learning and multi-objective genetic algorithm according to the path constraint parameter set, with welding defect rate, energy consumption and working time as optimization objectives, and through deep deterministic strategy gradient algorithm.

[0111] The real-time control module 105 is used to monitor the morphological characteristics of the molten pool through a visual sensing system, and dynamically adjust the welding parameters by combining the prediction results of the multi-field coupling model with the model prediction control algorithm.

[0112] The collaborative control module 106 is used to coordinate the motion trajectories of multiple robots through a communication network and a computing platform using a distributed model predictive control algorithm.

[0113] The knowledge management module 107 is used to construct a welding process knowledge graph based on various data in the welding process, and to establish a mapping relationship between process parameters and welding quality through a neural network, so as to realize the continuous optimization of the path planning system.

[0114] The above-disclosed embodiments are merely one or more preferred embodiments of this application and should not be construed as limiting the scope of this application. Those skilled in the art can understand that all or part of the processes for implementing the above embodiments and equivalent changes made in accordance with the claims of this application still fall within the scope of this application.

Claims

1. A method for intelligent path planning and optimization in robotic welding of large structures, characterized in that, Includes the following steps: Collect three-dimensional point cloud data of the workpiece surface, and simultaneously collect temperature field distribution data and stress field change data during the welding process; Using the collected point cloud data, temperature field data, and stress field data, a geometric-thermal-mechanical multi-field coupling model is established. The model parameters are corrected through machine learning algorithms, and welding deformation prediction data is output. Based on the multi-field coupling model, a deep learning network is used to identify the topological features of the weld seam and generate a set of path constraint parameters including the curvature change rate, the penetration depth threshold and the assembly deviation tolerance. Based on the path constraint parameter set, a hybrid optimization strategy of reinforcement learning and multi-objective genetic algorithm is adopted, with welding defect rate, energy consumption and working time as optimization objectives, and welding path nodes are optimized through deep deterministic strategy gradient algorithm. By monitoring the morphological characteristics of the molten pool through a visual sensing system and combining the prediction results of a multi-field coupling model, the welding parameters are dynamically adjusted using a model predictive control algorithm. The motion trajectories of multiple robots are coordinated by using a distributed model predictive control algorithm through a communication network and computing platform. Based on various data from the welding process, a welding process knowledge graph is constructed. A neural network is used to establish a mapping relationship between process parameters and welding quality, thereby enabling continuous optimization of the path planning system.

2. The intelligent path planning and optimization method for large-structure robotic welding as described in claim 1, characterized in that, In acquiring three-dimensional point cloud data of the workpiece surface, the method specifically includes: Configure the laser scanner's operating parameters, including setting the scanning accuracy and scanning speed; Plan the scanning path and use multiple scanning angles to completely scan the workpiece surface; Control the laser scanner to perform point cloud data acquisition along a predetermined path; The collected point cloud data is registered to generate a complete three-dimensional geometric model.

3. The intelligent path planning and optimization method for large-structure robotic welding as described in claim 2, characterized in that, The method specifically includes the following for collecting temperature field distribution data and stress field change data during the welding process: An infrared thermal imager monitoring system was deployed, and the monitoring area was set up to cover the weld and its heat-affected zone. Install a stress sensor array, arranging multiple sensors at predetermined intervals along the length of the weld seam; Synchronously collect and record process parameters such as current, voltage, and welding speed during the welding process.

4. The intelligent path planning and optimization method for large-structure robotic welding as described in claim 1, characterized in that, In establishing a geometric-thermal-mechanical multi-field coupling model, the method specifically includes: Establish a heat conduction control equation that incorporates the latent heat effect of phase change; Construct stress-strain equations based on material constitutive relations; Set thermal boundary conditions and mechanical boundary conditions; The multiphysics coupling equations are solved using a numerical iterative method.

5. The intelligent path planning and optimization method for large-structure robotic welding as described in claim 4, characterized in that, In correcting model parameters using machine learning algorithms, the method specifically includes: Construct a neural network model, with inputs including material parameters and welding process parameters; Define a loss function with prediction accuracy as the objective; Gradient descent algorithm is used to optimize network parameters and improve model prediction accuracy.

6. The intelligent path planning and optimization method for large-structure robotic welding as described in claim 1, characterized in that, In identifying weld topological features using deep learning networks, the method specifically includes: Construct a deep neural network architecture, including encoder and decoder structures; Introducing attention mechanisms into networks enhances feature extraction capabilities; A multi-scale feature fusion method is used to process semantic information at different levels; By applying transfer learning techniques, pre-trained models can be used to accelerate network convergence.

7. The intelligent path planning and optimization method for large-structure robotic welding as described in claim 6, characterized in that, In identifying weld topology features using deep learning networks, the method further includes the following network training steps: Prepare a training sample dataset containing multiple weld types; Configure hyperparameters for network training, including learning rate and batch size; The network weight parameters are updated using an optimized algorithm; The training process is monitored using a validation set to prevent overfitting.

8. An intelligent path planning system for robotic welding of large structures, used to implement the method as described in any one of claims 1-7, characterized in that, It includes a data acquisition module, a multi-field coupling modeling module, a weld seam recognition module, a path optimization module, a real-time control module, a collaborative control module, and a knowledge management module; The data acquisition module is used to acquire three-dimensional point cloud data of the workpiece surface, and simultaneously acquire temperature field distribution data and stress field change data during the welding process. The multi-field coupling modeling module uses the collected point cloud data, temperature field data, and stress field data to establish a geometric-thermal-mechanical multi-field coupling model, corrects the model parameters through machine learning algorithms, and outputs welding deformation prediction data. The weld identification module, based on a multi-field coupling model, uses a deep learning network to identify weld topological features and generate a set of path constraint parameters including curvature change rate, penetration depth threshold, and assembly deviation tolerance. The path optimization module is used to optimize the welding path nodes by adopting a hybrid optimization strategy of reinforcement learning and multi-objective genetic algorithm based on the path constraint parameter set, with welding defect rate, energy consumption and working time as optimization objectives, and through deep deterministic strategy gradient algorithm. The real-time control module is used to monitor the morphological characteristics of the molten pool through a visual sensing system, and dynamically adjust the welding parameters by combining the prediction results of the multi-field coupling model with the model prediction control algorithm. The collaborative control module is used to coordinate the motion trajectories of multiple robots through a communication network and a computing platform, using a distributed model predictive control algorithm. The knowledge management module is used to construct a welding process knowledge graph based on various data in the welding process, and to establish a mapping relationship between process parameters and welding quality through neural networks, so as to realize the continuous optimization of the path planning system.

Images