Ship planar sub-section welding multi-robot task allocation system based on digital twinning

Abstract

The application discloses a ship planar section welding multi-robot task allocation system based on digital twinning, and relates to the technical field of ship intelligent manufacturing. The four-dimensional digital twinning body module acquires workpiece assembly error, weld position and robot state data in real time, and constructs a twinning model comprising a physical engine layer, a process knowledge base and a dynamic topology graph. The ship planar section welding multi-robot task allocation system based on digital twinning dynamically constructs a weld task topology graph by fusing workpiece assembly error, robot motion state and welding process rules in real time through the digital twinning body, and realizes accurate multi-robot task allocation in combination with an anti-conflict auction algorithm. The use of a space-time gridding cooperation mechanism effectively resolves motion conflicts between large gantry and small gantry mechanisms, ensuring zero interference of the welding path; and a local path optimization module adaptively adjusts the welding gun attitude and the dry extension length in narrow spaces, eliminating the risk of physical collision.

Description

Technical Field

[0001] This invention relates to the field of intelligent ship manufacturing technology, specifically to a multi-robot task allocation system for ship planar segment welding based on digital twins. Background Technology

[0002] Welding of ship planar sections often employs robotic assembly line operations, such as sectional construction and deck section welding lines. Multiple Kawasaki RA005L robots are mounted on gantry cranes (large and small), and welding programs are generated using the SMARTWELD and KCONG systems to perform various welding operations, including fillet welding and vertical fillet welding. Such systems require multiple robots to collaborate on large-sized (15000mm long x 15000mm wide) welds of various types. However, current task allocation relies heavily on preset programs or simple scheduling, making it difficult to respond in real-time to workpiece assembly errors, changes in robot operating status, and differences in weld distribution. This leads to task overlap, path conflicts, or some robots becoming idle due to uneven load distribution. This not only reduces welding efficiency but can also affect weld quality due to interference, such as weld leg deviation and appearance defects. It fails to meet the high-precision, high-rate-of-production requirements of ship welding, necessitating a system that can dynamically optimize task allocation to solve these problems. Summary of the Invention

[0003] To achieve the above objectives, the present invention provides the following technical solution: a multi-robot task allocation system for ship planar section welding based on digital twins, comprising:

[0004] The 4D digital twin module collects workpiece assembly errors, weld positions, and robot status data in real time to build a twin model that includes a physics engine layer, a process knowledge base, and a dynamic topology graph.

[0005] Dual-loop optimization engine module: dynamically allocates weld seam tasks through a global task allocation loop and generates conflict-free paths through a local path fine-tuning loop;

[0006] Multiple verification module: Performs digital twin pre-simulation verification and physical-virtual data comparison on the optimization results to ensure that there are no path conflicts and that the welding quality meets the standards;

[0007] Load balancing module: Calculates the task load index of each robot in real time and dynamically adjusts task allocation.

[0008] Preferably, the four-dimensional digital twin module includes:

[0009] Real-time data layer: Captures workpiece assembly errors and weld position deviations by fusing laser scanning sensors and vision sensors;

[0010] Physics Engine Layer: Integrates multibody dynamics simulation model to pre-simulate robot motion path interference and welding thermal deformation effects;

[0011] Process knowledge base: Embedded welding rules generated by SMARTWELD, transforming weld leg range and extension length constraints into optimization objective functions;

[0012] Dynamic topology graph: with weld seams as nodes and robot reachable paths as edges, the weights include welding time and quality risk coefficients.

[0013] Preferably, the global task allocation ring employs a conflict-resistant auction algorithm:

[0014] Each robot competes for welding tasks based on its efficiency value, which is calculated based on welding speed, weld matching degree, and path conflict risk.

[0015] The travel of the gantry walking mechanism is discretized into a spatiotemporal grid, and path conflicts are dynamically avoided through virtual time windows.

[0016] Preferably, the local path fine-tuning loop includes:

[0017] Based on the adaptive fuzzy RRT algorithm, combined with the weld type and the robot's rotational degrees of freedom θ axis ±185°, a collision-free path is generated in a narrow space.

[0018] Adjust the welding torch extension length and angle using the welding torch posture optimizer to ensure weld fullness.

[0019] Preferably, the multi-verification module performs:

[0020] Digital twin pre-simulation verification: Real-time error data is injected into the KCONG simulation environment to detect path deviation and joint torque exceeding limits;

[0021] Physical-virtual data comparison: The state of the molten pool is inverted by the arc data of the welding power source and compared with the predicted values ​​of the twin model to trigger further optimization of process parameters.

[0022] Preferably, the load balancing module:

[0023] Calculate the robot load balancing index (LBI). When the index falls below a set threshold, the tasks of the high-loaded robots are redistributed to idle units.

[0024] The index is dynamically updated based on the ratio of the variance of the duration of each robot task to the average duration.

[0025] Preferably, in the conflict-resistant auction algorithm:

[0026] When the distance between the two robots is below the safety threshold, the robot is forced to slow down and a waiting time window is inserted.

[0027] The weighting coefficient of the efficiency value is dynamically adjusted according to the weld type: vertical welding tasks are preferentially assigned to robots with large Z-axis travel.

[0028] Preferably, it also includes a quality defect prediction unit:

[0029] Based on real-time arc data from welding power sources, a probability prediction model for weld porosity and cracks is constructed.

[0030] When the defect probability exceeds the limit, the task is paused and a local path fine-tuning loop is triggered for replanning.

[0031] Preferably, the system transmits laser scanning data and control commands through a 5G communication module, and the task allocation response cycle is lower than a set threshold.

[0032] A method for multi-robot task allocation in ship planar section welding based on a digital twin system includes the following steps:

[0033] Step S1. Capture workpiece errors and robot status in real time using a digital twin;

[0034] Step S2. Generate an initial task allocation scheme through the global task allocation ring;

[0035] Step S3. Optimize the welding torch path and attitude through the local path fine-tuning loop;

[0036] Step S4. Welding is performed after dual verification through digital twin pre-simulation and physical data comparison;

[0037] Step S5. Dynamically adjust task allocation based on the load balancing index.

[0038] This invention provides a multi-robot task allocation system for ship planar section welding based on digital twins. It has the following advantages:

[0039] This digital twin-based multi-robot task allocation system for ship planar section welding dynamically constructs a weld seam task topology map by real-time fusion of workpiece assembly errors, robot motion states, and welding process rules using a digital twin. Combined with an anti-collision auction algorithm, it achieves precise multi-robot task allocation. Utilizing a spatiotemporal gridded collaborative mechanism, it effectively resolves motion conflicts between the large and small gantry mechanisms, ensuring zero interference in the welding path. A local path optimization module adaptively adjusts the welding torch posture and extension length in narrow spaces, eliminating the risk of physical collisions. A dual verification system corrects trajectory deviations and process parameters in real time, ensuring that the weld leg dimensions of complex welds such as vertical welds and fillet welds meet standards, resulting in a smooth and full weld appearance and eliminating the problem of recurring defects such as porosity and cracks.

[0040] This digital twin-based multi-robot task allocation system for ship planar section welding utilizes a load balancing mechanism constrained by hardware capabilities. It quantifies key parameters such as lifting mechanism stroke and rotational accuracy into task allocation weights, enabling intelligent matching of vertical welding tasks with robot Z-axis stroke and directional binding of narrow weld seams with high-precision θ-axis units, thus eliminating quality risks caused by capability mismatch. A 5G-enhanced communication framework ensures closed-loop task allocation through hierarchical transmission and redundant anti-interference strategies; offline collaboration and reconnection mechanisms maintain continuous production line operation in network outage emergency modes. Ultimately, this system further improves equipment utilization and stabilizes production cycle time, providing a fully autonomous optimization solution for large-section ship welding. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the module interaction of the multi-robot task allocation system for ship planar section welding based on digital twin of the present invention.

[0042] Figure 2 This is a flowchart illustrating the multi-robot task allocation method for ship planar segment welding based on digital twins, as described in this invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] Please see Figure 1 and Figure 2 This invention provides a technical solution: a multi-robot task allocation system for ship planar section welding based on digital twins, comprising:

[0045] The 4D digital twin module collects workpiece assembly errors, weld positions, and robot status data in real time to build a twin model that includes a physics engine layer, a process knowledge base, and a dynamic topology graph.

[0046] Dual-loop optimization engine module: dynamically allocates weld seam tasks through a global task allocation loop and generates conflict-free paths through a local path fine-tuning loop;

[0047] Multiple verification module: Performs digital twin pre-simulation verification and physical-virtual data comparison on the optimization results to ensure that there are no path conflicts and that the welding quality meets the standards;

[0048] Load balancing module: Calculates the task load index of each robot in real time and dynamically adjusts task allocation.

[0049] It should be further explained that, in the specific implementation process, the construction and real-time updating of the four-dimensional digital twin is as follows:

[0050] Real-time data acquisition: Using an L2S-80 laser rangefinder and an industrial camera installed on the welding production line, the workpiece surface is scanned to capture assembly errors and weld positions. The assembly error tolerance is ±3mm. When a height difference ΔH > 3mm is detected between the spliced ​​plates, the weld is automatically divided into two independent segments.

[0051] Physics engine simulation: Load the robot's kinematic model and the geometric constraints of the gantry track into the digital twin platform to simulate the cooperative motion of the robots. If the predicted distance between the two robots is less than 2m, a collision warning is triggered.

[0052] Embedded process rules: Import welding process libraries from SMARTWELD, such as a 15mm extension length for flat fillet welds and a 5-10mm weld leg range for vertical welds, and transform quality constraints into path optimization weights. For example, vertical welding tasks are automatically assigned to robots with a Z-axis travel of ≥1000mm.

[0053] The collaborative operation process of the dual-loop optimization engine is as follows:

[0054] Global task allocation loop: Each robot bids for welding tasks based on its efficiency value. In efficiency value calculation: for corner welding tasks, when the gap between parts Δ < 35mm, it is forcibly marked as "high conflict risk" and its bidding weight is reduced; the gantry walking speed is discretized according to a spatiotemporal grid, and if a grid is occupied by two robots at the same time, a virtual waiting time window is inserted.

[0055] Real-time adjustment of local path fine-tuning loop: Adaptive fuzzy RRT algorithm is used in narrow space: If the weld is of the arc type, the θ axis rotation mechanism is called first to adjust the welding gun angle; if interference between the welding gun and the workpiece is detected, the extension length is automatically shortened to the lower limit of the process and the path is replanned.

[0056] The multiple verification and dynamic correction process is as follows:

[0057] Digital twin pre-simulation verification: Real-time error data is injected into the KCONG simulation platform, and the optimized path is run. If joint torque exceeds the limit or trajectory deviation is detected to be greater than 0.3mm, it is judged as a failed solution.

[0058] Physical-virtual data closed loop: Arc voltage and current are collected using a Panasonic YD-500GR welding machine to infer the molten pool state. If the deviation between the actual weld leg size and the twin prediction value continues to increase, the following actions are automatically triggered: pause the current welding, adjust the wire feed speed or gas flow, and re-execute the local path fine-tuning loop to generate a new path.

[0059] The load balancing and task redistribution process is as follows:

[0060] The task duration of each robot is calculated in real time. When the ratio of the maximum duration to the minimum duration exceeds the set threshold, the long weld seam task of the high-load robot is divided and transferred to the idle robot. For robots with low utilization of the Z-axis lifting mechanism, vertical welding tasks are assigned first.

[0061] Control commands are transmitted to the gantry inverter motor and robot controller via a 5G module, with a response delay lower than the system's set threshold; the gantry inverter motor is an SEW geared motor.

[0062] When the probability of porosity or cracks output by the quality defect prediction unit exceeds the limit, the torch cleaning and wire cutting mechanism is activated to clean the welding torch and then the test is repeated.

[0063] Since the preset program cannot handle welds in spliced ​​plates with ΔH > 3mm, the weld seam is segmented in real time by scanning, and the influence of thermal deformation is pre-simulated by the physics engine to ensure adaptive adjustment of the welding path and solve path conflicts caused by assembly errors. Since fixed paths are prone to welding torch collisions in the Δ < 35mm region, a fuzzy RRT algorithm is adopted to dynamically optimize the welding torch posture based on the θ-axis rotation capability and forcibly constrain the extension length to eliminate interference risks. The weld seam is dynamically migrated based on the task duration ratio and the task is allocated in conjunction with the robot's hardware capabilities to solve the problem of uneven load.

[0064] Through a progressive process of error perception, dynamic auction, path fine-tuning, closed-loop verification, and load rebalancing, digital twins are deeply integrated with multi-robot collaboration to systematically solve the problems of task allocation conflicts, quality fluctuations, and resource idleness in ship welding.

[0065] The four-dimensional digital twin module includes:

[0066] Real-time data layer: Captures workpiece assembly errors and weld position deviations by fusing laser scanning sensors and vision sensors;

[0067] Physics Engine Layer: Integrates multibody dynamics simulation model to pre-simulate robot motion path interference and welding thermal deformation effects;

[0068] Process knowledge base: Embedded welding rules generated by SMARTWELD, transforming weld leg range and extension length constraints into optimization objective functions;

[0069] Dynamic topology graph: with weld seams as nodes and robot reachable paths as edges, the weights include welding time and quality risk coefficients.

[0070] It should be further explained that, in the specific implementation process, the high-precision perception and fusion process of the real-time data layer is as follows: The workpiece surface contour is scanned by a laser rangefinder (model: L2S-80), and the actual position of the weld is captured by an industrial camera. When the height difference ΔH of the splicing plate is detected to be greater than 3mm, the cross-joint weld is automatically divided into two independent task segments. Combining the real-time positioning data of the large gantry walking mechanism and the small gantry Y-axis travel, the reachable work area of ​​the robot is dynamically calculated. If the weld is located within ±0.5m of the work area boundary, it is marked as a "low-priority task". Among them, the X-axis travel of the large gantry walking mechanism is 37m-60m, and the Y-axis travel of the small gantry is 11-21.2m.

[0071] The process of conflict pre-simulation and thermal deformation compensation in the physics engine layer is as follows: Load the kinematic model of the Kawasaki RA005L robot and the Z-axis travel of the lifting mechanism, and pre-simulate the cooperative motion path of multiple robots: When the predicted Euclidean distance between two robots is less than the safety threshold of 2m, an avoidance path point is automatically inserted; For vertical welding tasks, simulate the impact of welding thermal deformation on the Z-axis positioning accuracy. If the deformation causes the welding torch extension length to exceed the process range by 15-20mm, adjust the lifting mechanism to compensate for the height in advance; Among them, the number of axes of the robot body is 6, the Z-axis travel of the lifting mechanism is 1.4-3.8m, and the maximum height of the vertical welding task is 3300mm.

[0072] The process of constraint transformation and decision-driven operation of the process knowledge base is as follows: Analyze the welding rule base generated by SMARTWELD, such as 4-10mm for flat fillet welds and 5-10mm for vertical fillet welds without beveling, and map the quality constraints into path optimization weights: For corner welding tasks, if the gap Δ between parts is less than 35mm, assign a "high conflict risk" label and prohibit large-angle movements of the θ-axis rotation mechanism; For circular arc welds, call the KCONG preset circular arc segmentation logic, and automatically generate teaching points in segments when the center angle is greater than the maximum rotation range of the θ-axis.

[0073] The real-time reconstruction mechanism of the dynamic topology graph is as follows: Weld seam endpoints are used as topology nodes, and robot movement paths are used as directed edges. The edge weights include:

[0074] Time cost: The movement time is calculated based on the walking speed of the large gate (10-15 m / min) and the lateral movement speed of the small gate (10-15 m / min).

[0075] Quality risk coefficient: If the weld is located at the edge of the workpiece or in a bending area, the risk value increases in a gradient manner; when an assembly error is detected that causes the node position to shift, the topology graph is reconstructed and broadcast to all robot controllers.

[0076] Laser scanning data is transmitted to an industrial computer (model: ARK-2250L) via a 5G module, stored in point cloud format and associated with weld IDs; the process knowledge base is deployed in the form of an SQL database, supporting batch import and version management of SMARTWELD data packages.

[0077] By identifying height differences in real time, the system automatically segments task segments and optimizes paths independently, ensuring that the welding torch maintains a constant extension length in abrupt change areas, thus solving weld quality problems caused by assembly errors. When the gap between parts Δ < 35mm, the system forcibly locks the rotation range of the θ axis and amplifies the conflict risk through topology graph weights, guiding the auction algorithm to avoid task allocation in this area and eliminating the risk of interference between the welding torch and the workpiece.

[0078] By transforming the weld leg size constraints of SMARTWELD into quality risk coefficients of the topology graph edges, robots can prioritize welds with high process matching when bidding, thus avoiding the risk of weld leg size deviations exceeding the standard from the source. Through the four-dimensional twin construction logic of multi-source perception, conflict pre-simulation, rule mapping, and topology reconstruction, assembly errors, spatial constraints, and process requirements in ship welding are transformed into a computable optimization model in real time, solving the path conflict and quality fluctuation problems caused by static planning.

[0079] The global task allocation ring uses a conflict-resistant auction algorithm.

[0080] Each robot competes for welding tasks based on its efficiency value, which is calculated based on welding speed, weld matching degree, and path conflict risk.

[0081] The travel of the gantry walking mechanism is discretized into a spatiotemporal grid, and path conflicts are dynamically avoided through virtual time windows.

[0082] It should be further explained that, in the specific implementation process, the efficiency value-driven bidding mechanism works as follows: Each robot calculates its efficiency value Q based on its own hardware capabilities, such as the Z-axis travel of the small gantry (1.4-3.8m) and the θ-axis rotation range (±185°). For fillet welding tasks, the efficiency value focuses on welding speed, with the large gantry having a maximum X-axis speed of 15m / min. For vertical welding tasks, i.e., height ≤3300mm, the efficiency value is associated with the Z-axis travel margin, and robots with insufficient travel are automatically downgraded in the bidding. When the weld is located in a narrow area with a component gap Δ <35mm, the risk factor of path conflict increases exponentially, significantly reducing the bidding priority.

[0083] During the bidding process, the robot continuously receives updated weld topology maps from the digital twin. If a weld is marked as "high risk of conflict," such as a weld in a bending area, only the robot with the largest Z-axis travel is allowed to participate in the bidding.

[0084] The spatiotemporal gridding conflict avoidance process is as follows: The gantry walking track is discretized into spatiotemporal grids with 0.5m intervals, and each grid is associated with a time window label: After a robot wins a bid for a weld, it needs to declare the occupied grid sequence and duration, calculated based on the walking speed of 10-15m / min; if the time windows of two robots declaring to occupy the same grid overlap, the system forces the later declarer to insert a waiting delay until the time windows are staggered; for corner welding tasks, an additional safety delay is reserved at the weld end grid to ensure that the θ-axis rotation mechanism completes the attitude adjustment.

[0085] The adaptive weight adjustment process for weld type is as follows: When the weld is an arc type: KCONG's preset arc segmentation strategy is called, and welds with center angles exceeding the rotation range of the θ axis are automatically segmented; the bidding weight of the sub-welds after segmentation is increased, which incentivizes the same robot to complete the task continuously to reduce repetitive positioning.

[0086] When the weld seam spans a height difference ΔH of the splice plate greater than 3mm: it is divided into two independent tasks. The upper task is only allowed to be bid on by robots with a Z-axis travel greater than or equal to ΔH. The lower task is assigned to the robot with the lowest current load to avoid waiting due to height switching.

[0087] The winning bid result is transmitted in real time to the robot controller, namely the Kawasaki control cabinet, via the 5G module, triggering the welding program call; the large gantry walking mechanism, namely the SEW geared motor, receives the grid sequence instructions and starts and stops according to the declared time window.

[0088] By discretizing the physical trajectory into a spatiotemporal grid and dynamically inserting waiting delays through a declaration-arbitration mechanism, the risk of robot collisions is eliminated. For vertical welding tasks, the robot's Z-axis travel capability is quantified into an efficiency value weight to ensure that 3300mm high welds are handled only by robots with sufficient travel, thus solving welding defects caused by exceeding the height limit for vertical welding. When the gap between components Δ < 35mm is detected, the path conflict risk coefficient is multiplied to guide the algorithm to avoid bidding in that area and prevent physical interference between the welding torch and the workpiece.

[0089] Through a three-layer collaborative mechanism of performance bidding, spatiotemporal grid arbitration, and weld segment adaptation, the large stroke conflicts, hardware capability differences, and spatial constraints in ship welding are transformed into dynamic and solvable optimization problems, thus solving the task overlap and resource idleness caused by fixed allocation.

[0090] The local path fine-tuning loop includes:

[0091] Based on the adaptive fuzzy RRT algorithm, combined with the weld type and the robot's rotational degrees of freedom θ axis ±185°, a collision-free path is generated in a narrow space.

[0092] Adjust the welding torch extension length and angle using the welding torch posture optimizer to ensure weld fullness.

[0093] It should be further explained that, in the specific implementation process, the process of generating a conflict-free path in a narrow space is as follows: When the weld is located in the area where the stiffener gap Δ < 35mm, the adaptive fuzzy RRT algorithm is activated: based on the real-time gap data of laser scanning, the growth direction of the random tree is dynamically restricted to avoid interference between the welding torch and the workpiece; if the welding torch anti-collision sensor alarm is detected, the extension length is automatically shortened to the lower limit of the process 15mm, and the path is replanned. For vertical welding tasks, i.e., vertical welding height ≤ 3300mm, the lifting mechanism is used first to adjust the height of the welding torch, and when the space height is insufficient, the θ-axis rotation mechanism is forcibly activated to compensate for the angle; among which, the Z-axis travel of the lifting mechanism is 1.4-3.8m.

[0094] The weld type-driven attitude optimization process includes circular arc weld treatment and fillet weld treatment; where:

[0095] Circular arc weld processing: Analyze the arc center angle data output by SMARTWELD. If it exceeds the maximum rotation range of the θ axis, it is automatically divided into multiple sub-arcs that can be covered. Each sub-arc is independently optimized to ensure that the angle between the tangent direction and the normal of the workpiece surface is within the allowable range of the process.

[0096] Fillet welding: When the gap between the parts Δ≥65mm, the fillet weld is performed by directly sensing the weld end point; if 35mm≤Δ<65mm, the contour of the workpiece needs to be scanned vertically and finely adjusted within ±10° through the θ axis to avoid welding gun collision.

[0097] The dynamic obstacle avoidance and real-time replanning process is as follows: During path execution, the robot distance is continuously monitored. If the Euclidean distance between the two robots is less than 2m, the robot immediately pauses and inserts a waiting instruction. If the pause causes the welding interruption time to exceed the limit, local path replanning is triggered, bypassing the conflict area and connecting the original weld. When the welding power supply (model: Panasonic YD-500GR) detects abnormal arc fluctuations, it is determined to be a path deviation or electrode extension length misalignment. Welding is paused and the torch cleaning and wire cutting mechanism is invoked to clean the welding nozzle. The local path is regenerated based on real-time point cloud data to ensure the weld pool connection at the continuation welding position.

[0098] The optimized path instructions are sent to the Kawasaki robot controller via the 5G module, driving the θ-axis rotation mechanism, namely the Nabtesco RV-320C reducer and the Z-axis lifting mechanism; the cleaning gun and wire cutting instructions are linked to the Binzel 830.2217.1 mechanism to perform automatic cleaning.

[0099] By dynamically constraining the growth direction of the random tree and combining it with real-time adjustment of the extension length, the welding torch is ensured to pass safely through the gap between the stiffener plates, eliminating physical interference problems. During corner welding, the sensing strategy is dynamically switched according to the Δ value to avoid corner defects caused by workpiece contour errors, thus solving the problem of insufficient corner welding qualification rate.

[0100] When the arc is abnormal or the robot is too close, the system is not interrupted but the path is replanned locally. The continuity of the weld is ensured by cleaning the torch and continuing welding, thus solving the appearance defects of the weld caused by the interruption.

[0101] By employing local path closed-loop control logic that includes gap perception, attitude optimization, dynamic obstacle avoidance, and abnormal welding continuation, the spatial constraints, process limits, and sudden anomalies in ship welding are transformed into real-time solvable optimization problems. Furthermore, through adaptive exploration in narrow spaces and a process-kinematics fusion mechanism, the pain points of path conflict and quality fluctuation are resolved.

[0102] Multi-factor authentication module execution:

[0103] Digital twin pre-simulation verification: Real-time error data is injected into the KCONG simulation environment to detect path deviation and joint torque exceeding limits;

[0104] Physical-virtual data comparison: The state of the molten pool is inverted by the arc data of the welding power source and compared with the predicted values ​​of the twin model to trigger further optimization of process parameters.

[0105] It should be further explained that, in the specific implementation process, the assembly error and robot positioning data collected in real time are injected into the KCONG simulation platform, and the optimized welding path is run. If the robot joint torque is detected to exceed the rated load of Kawasaki RA005L, or the deviation of the welding torch trajectory from the theoretical weld is >0.3mm, it is judged as a failed path.

[0106] The system automatically triggers a level-three response for failed paths, as follows:

[0107] Beginner level: Fine-tune the path point coordinates and retry the simulation;

[0108] Intermediate: Release the current spacetime grid window and notify the global task allocation ring to re-bid;

[0109] Advanced: Divide the ultra-long weld seam into independent sub-tasks and assign them to different robots.

[0110] For vertical welding tasks, the Z-axis travel is 1.4-3.8m. The influence of welding thermal deformation on the lifting mechanism is simulated. If the predicted extension length exceeds the process range by 15-20mm, the Z-axis height is compensated in advance.

[0111] The arc voltage and current are collected in real time using a Panasonic YD-500GR welding machine. The molten pool state is inverted and the actual weld leg size is calculated. If the actual values ​​at three consecutive monitoring points deviate from the twin prediction values ​​and continue to increase, it is determined to be a process mismatch. The dynamic correction process is triggered: the current welding is paused, and the torch cleaning and wire cutting mechanism (model: Binzel 830.2217.1) is activated to clean the welding nozzle; a local correction path is generated based on the real-time point cloud data, and the welding position is moved back 5mm to restart the arc; the wire feed speed or shielding gas flow rate is adjusted until the arc data returns to the stable range.

[0112] For corner welding tasks, when the gap between parts is 35mm≤Δ<65mm, compare the actual corner profile with the scanned data: if the profile deviation causes the risk of welding torch interference, force the θ-axis rotation mechanism to fine-tune ±10° and re-verify.

[0113] A porosity and crack prediction model is built based on arc data: when the defect probability continues to be higher than the set threshold, the task is interrupted and the weld is marked as "high risk"; root cause analysis is initiated: if it is a path problem, such as misalignment of the weld extension length, the local path fine-tuning loop replanning is called; if it is a process parameter problem, such as insufficient gas, the parameters are automatically adjusted and the weld is re-welded.

[0114] The simulation verification results are synchronized to the industrial control computer (model: Advantech ARK-2250L) via a 5G module to update the digital twin database; process adjustment commands are sent to the welding power source (model: Panasonic YD-500GR) and wire feeding mechanism (model: YW-CNF011) to perform real-time parameter calibration.

[0115] Through a three-tiered progressive response of fine-tuning, re-bidding, and task segmentation, deviations are dynamically eliminated while ensuring welding continuity, resolving weld trajectory deviation issues caused by error accumulation. When process mismatch is detected, the system is not interrupted but a 5mm retraction and re-welding strategy is adopted, combined with torch cleaning and parameter adjustment to ensure weld consistency and resolve weld appearance defects caused by interruption. For the wrap angle area with Δ < 65mm, contour scanning errors are compensated through θ-axis fine-tuning to avoid torch collisions caused by fixed paths and improve the wrap angle welding qualification rate.

[0116] By employing a four-tiered quality protection chain of simulation pre-inspection, arc inversion, weld reconnection, and risk fusion, the risks of path deviation, process fluctuation, and defects in ship welding are transformed into a closed-loop controllable process. By utilizing a pre-screening mechanism and online correction logic, the pain points of quality fluctuation and interruption defects are resolved.

[0117] Load balancing module:

[0118] Calculate the robot load balancing index (LBI). When the index falls below a set threshold, the tasks of the high-loaded robots are redistributed to idle units.

[0119] The index is dynamically updated based on the ratio of the variance of the duration of each robot task to the average duration.

[0120] It should be further explained that, in the specific implementation process, the Load Balancing Index (LBI) is calculated based on the estimated welding time of each robot's current task queue: if a robot's task time exceeds 1.5 times the system average time, it is marked as a "high-load unit"; if a robot is idle for two consecutive task cycles, it is marked as a "low-load unit". The estimated welding time of the current task queue includes the movement time for flat fillet welding calculated based on the gantry walking speed of 15m / min, and the height adjustment time for vertical welding calculated based on the lifting mechanism speed of 10m / min.

[0121] When the time ratio of high-load units to low-load units exceeds a set threshold, task reassignment is triggered: for long weld seams longer than 10m for high-load robots, the task is automatically divided into equal-length sub-segments; the sub-segment closest to the low-load robot is handed over to it for execution, and during the handover, it must be ensured that the Z-axis travel of the target robot meets the weld seam height requirement of no more than 3300mm.

[0122] For vertical welding tasks with a height greater than 1000mm, migration is only allowed to a robot with a Z-axis travel not less than the actual weld height. If the Z-axis travel of the low-load unit is insufficient, the migration will be skipped. After migration, the spatiotemporal grid window released by the original robot will be opened for bidding immediately to avoid conflicts caused by the large gantry track.

[0123] For corner welding tasks, if the gap between the parts Δ is less than 35mm, it is forbidden to move to a robot with insufficient θ-axis rotation accuracy; if Δ is greater than or equal to 65mm, it is preferable to move to a unit with a intact gun cleaning and wire cutting mechanism to ensure continuous corner welding.

[0124] When the welding power supply (model: Panasonic YD-500GR) detects an arc anomaly causing a task interruption, all tasks of the faulty robot are suspended, and its unfinished welds are assigned to healthy units based on spatial proximity. If the Z-axis travel of a healthy unit is insufficient, the weld is split and only the lower weldable section is transferred. After the robot returns to online status, it automatically receives new tasks assigned by the system, but the initial load limit is halved until it runs stably for 3 cycles.

[0125] Robot task duration data is uploaded to the industrial control computer (model: Advantech ARK-2250L) via a 5G module, and the load dashboard is refreshed every 5 seconds; task migration commands are linked to the KCONG system to update welding program call permissions.

[0126] By forcibly verifying the target robot's capabilities during migration, vertical welding tasks are only transferred to units with Z-axis travel margin > weld height; narrow area tasks with Δ < 35mm are limited to θ-axis high-precision robots, eliminating welding defects caused by capability mismatch.

[0127] By dividing welds longer than 10m into segments based on spatial location, priority is given to transferring the segments closest to the low-load robot: spatial distance is calculated based on the coordinates of the large gantry track, minimizing the Y-axis movement distance of the small gantry; after transfer, the spatiotemporal grid window occupancy status is automatically updated to resolve track conflicts caused by task transfer.

[0128] When the robot goes offline abnormally, the system is not stopped. Instead, unfinished welds are assigned according to spatial proximity. For vertical welding tasks that exceed the capacity of the healthy unit, only the lower weldable section is transferred. After recovery, the load limit is gradually increased to ensure continuous system operation and avoid production line shutdown problems.

[0129] Through a dynamic balancing chain of load monitoring, capacity verification, spatial partitioning, and fault isolation, resource idleness, capacity mismatch, and sudden interruptions in ship welding are transformed into an adaptive optimization process. Based on a task migration strategy and fault self-healing logic under hardware constraints, the problems of insufficient robot utilization and production line vulnerability are solved.

[0130] In conflict-resistant auction algorithms:

[0131] When the distance between the two robots is below the safety threshold, the robot is forced to slow down and a waiting time window is inserted.

[0132] The weighting coefficient of the efficiency value is dynamically adjusted according to the weld type: vertical welding tasks are preferentially assigned to robots with large Z-axis travel.

[0133] It should be further explained that during the specific implementation process, the Euclidean distance between the robots on the gantry track is monitored in real time. When the distance between the two robots enters the warning range of 3m to 2m, the system automatically reduces the walking speed of the following robot to a set ratio of the original speed. If the distance is further shortened to the emergency range of <2m, the following robot is forcibly paused and a waiting instruction is inserted until the preceding robot leaves the conflict grid area.

[0134] For poses with high interference risks, such as the end point of corner welding, an additional safety delay is reserved, namely: the attitude adjustment time of the θ-axis rotation mechanism + the time of the torch cleaning action, to ensure that the mesh occupancy is released after the welding torch is completely removed.

[0135] The hardware capabilities for vertical welding tasks are prioritized as follows: when the weld height is greater than 1000mm, the weighting coefficient of the Z-axis travel margin is significantly increased in the efficiency value Q calculation formula; only robots with Z-axis travel greater than or equal to the actual weld height can participate in the bidding, such as only units with a Z-axis travel of 3800mm are allowed to handle a 3300mm high weld.

[0136] In the active avoidance strategy in narrow areas, the path conflict risk coefficient in the efficiency value Q is doubled for welds with a component gap Δ < 35mm; if the contour of the bending area is detected at the same time, only robots equipped with a high-precision θ-axis rotation mechanism (model: Nabtesco RV-320C) are allowed to bid.

[0137] During the differentiated processing of height difference welds, when the weld crosses the height difference ΔH of the splicing plate > 3mm, it is divided into independent upper and lower tasks; upper task allocation: only open to robots with Z-axis travel ≥ ΔH and current load below average for bidding; lower task allocation: priority is given to nearby idle robots to avoid the large gantry repeatedly moving and consuming time.

[0138] The deceleration command is sent to the large gantry inverter motor (model: SEW geared motor) via the 5G module to adjust the walking speed, which is adjustable from 10-15 m / min. The weld segmentation results are linked to the KCONG system to update the welding program segment markers. Through three-layer collaborative optimization of distance-level response, capability weight quantification, and highly decoupled allocation, motion conflicts, hardware differences, and spatial abrupt changes in ship welding are transformed into dynamic and controllable strategies. Focusing on the balance between safety and efficiency design and capability-driven intelligent decision-making, the system solves problems such as robot collisions, resource mismatches, and cycle time fluctuations.

[0139] It also includes a quality defect prediction unit:

[0140] Based on real-time arc data from welding power sources, a probability prediction model for weld porosity and cracks is constructed.

[0141] When the defect probability exceeds the limit, the task is paused and a local path fine-tuning loop is triggered for replanning.

[0142] It should be further explained that, in the specific implementation process, real-time monitoring of arc characteristics and defect probability modeling are performed as follows: Arc voltage and current data are continuously collected using the Panasonic YD-500GR welding power supply to extract key features, such as fluctuation frequency and short-circuit peak value. A probability prediction model is constructed based on a historical welding defect sample library. The construction process is as follows: when the feature value enters the high-risk range, the probability level of porosity and cracks is output in real time; if continuous arc instability occurs in a vertical welding task, the probability level is automatically increased. For corner welding tasks with a component gap Δ≥35mm, welding torch posture data is additionally monitored, and probability level superposition is triggered when the posture is abnormal. The historical welding defect sample library includes porosity, cracks, and lack of fusion; the welding torch posture data includes the θ-axis angle and extension length.

[0143] The graded response mechanism and process self-optimization process include primary response, intermediate response and advanced response; among them, primary response corresponds to probability level 1, intermediate response corresponds to probability level 2, and advanced response corresponds to probability level 3.

[0144] Initial response: Dynamically adjust the wire feed speed and protective gas flow rate to attempt to stabilize the electric arc;

[0145] Intermediate response: Pause welding and start the torch cleaning and wire cutting mechanism (model: Binzel 830.2217.1), clean, retract 10mm and restart the arc;

[0146] Advanced Response: Interrupt the current task, mark the weld as a high-risk section, and trigger the local path fine-tuning loop to replan the path: If it is a vertical welding task, prioritize optimizing the Z-axis lifting trajectory; if it is a narrow area weld with Δ < 35mm, forcibly widen the extension length to the upper limit of 20mm to avoid interference.

[0147] In the weld continuation quality assurance mechanism, all welding tasks restarted after interruption must perform the following: laser scanning of the weld continuation position point cloud and comparison with the original path deviation; if the deviation is >0.3mm, a transition slope path is generated for smooth connection; the welding speed is reduced to a set ratio of the original rate for the first 50mm after arc initiation, and restored after the arc stabilizes. Specifically, for the arc termination stage of corner welding, a time window is reserved for adjustment of the θ-axis rotation mechanism to ensure accurate arc termination point posture.

[0148] Among them, the defect probability data is synchronized to the digital twin via the 5G module to update the process knowledge base rules; the cleaning gun command is linked to the Binzel mechanism to perform silicone oil spraying and wire cutting actions.

[0149] By integrating the θ-axis angle and extension length data in corner welding, when Δ < 35mm and the welding torch tilt angle exceeds the limit, the defect probability level is improved even if the arc is stable, thus preventing physical interference defects.

[0150] In the three-level progressive response strategy, the primary response involves online fine-tuning of parameters to ensure welding continuity; the intermediate response involves cleaning the welding torch and retracting to continue welding, eliminating molten pool contamination; and the advanced response involves path replanning to eliminate spatial conflicts at their source and systematically solve the problem of recurring defects.

[0151] By employing a proactive quality protection system that combines arc and attitude monitoring, three-level meltdown prevention, and ramp continuation assurance, the hidden defect risks in ship welding are transformed into an interventionible and recoverable closed-loop control process. Multi-source data fusion prediction and progressive response mechanisms are used to address porosity, cracks, and appearance defects.

[0152] The system transmits laser scanning data and control commands via a 5G communication module, with a task allocation response cycle lower than a set threshold. Further explanation is needed regarding the low-latency data synchronization and decision-making closed-loop process: the laser scanning sensor (model: L2S-80) captures workpiece assembly errors and weld positions in real time, transmitting point cloud data to the digital twin via a 5G private network. If a height difference ΔH > 3mm is detected between the splicing plates, a weld segmentation command is immediately triggered, ensuring the topology update is completed before the next task allocation cycle. For the moving gantry mechanism, positioning data is refreshed every 0.5 seconds to avoid spatiotemporal grid calculation deviations due to communication delays. The workpiece assembly error is ±3mm, and the moving gantry mechanism's speed is 10-15m / min.

[0153] The robot bidding decision-making closed loop is compressed into a single communication cycle, and the efficiency value Q calculation result is broadcast to all units via 5G; the winning bid result is simultaneously sent to the robot controller and the large gantry inverter drive, namely: SEW geared motor, eliminating the multi-level forwarding delay of traditional bus communication.

[0154] In the dynamic bandwidth allocation and anti-interference mechanism, 5G bandwidth is dynamically allocated based on the urgency of the task:

[0155] For high-priority data: collision warning commands, arc abnormal signals, and defect probability exceeding limits alarms, a dedicated channel is used exclusively.

[0156] For medium-priority data: weld seam scan point cloud, load balancing index (LBI) refresh value, shared balanced channel;

[0157] For low-priority data: process knowledge base history records and equipment logs, transmit during off-peak hours.

[0158] When the system detects that electromagnetic interference in the workshop is causing an increase in packet loss rate, it automatically switches to a multi-path redundant transmission mode and relays key instructions through base stations and edge computing nodes. If the redundant transmission still fails, a local caching mechanism is triggered: the robot continues to execute according to the last valid path until communication is restored.

[0159] During the offline emergency collaboration mode, when the 5G network interruption exceeds the set threshold, each robot switches to offline collaboration based on the pre-stored data of the digital twin: it maintains movement according to the last known spatiotemporal grid window, avoids obstacles in real time through local laser sensors, and stops suddenly when the detection distance is <1m; the welding parameters call the local cached values ​​of the process knowledge base.

[0160] After the network is restored, incremental data synchronization is automatically performed. Conflict operations are rolled back and corrected based on online instructions. Interrupted welds are reconnected according to the continuation mechanism. The continuation mechanism is a 5mm retraction followed by low-speed arc initiation.

[0161] The 5G module is integrated into the industrial control computer (model: Advantech ARK-2250L) and supports QoS policy configuration; the edge computing node is deployed in the welding gantry control cabinet to perform data pre-filtering.

[0162] WiFi and bus communication cannot meet the real-time requirements of multi-robot collaboration for large workpieces (15m x 15m). To address this, laser scanning data is used to directly reach the robotic twin, bypassing intermediate processing layers; a single-cycle closed loop for bidding and execution instructions resolves robot idleness caused by delayed task allocation response.

[0163] By dynamically allocating bandwidth based on task urgency, collision commands are given absolute priority. In electromagnetic interference scenarios, dual relays between base stations and edge nodes are used, combined with local caching to achieve seamless degradation and eliminate the risk of data interruption caused by strong interference in the workshop. After a network outage, the last valid spatiotemporal grid window is used to avoid robot collisions. After recovery, seamless switching is achieved by synchronously adding weld seams in the backend, ensuring continuous production line operation during network fluctuations.

[0164] By using a 5G-enhanced collaborative framework with closed-loop, dynamic anti-interference, and seamless offline continuation, the large-scale communication bottleneck in ship welding is transformed into a reliable and adaptive neural network. Based on real-time assurance and robust design, the problem of uncontrolled cycle time caused by communication delays and interruptions is solved.

[0165] A method for multi-robot task allocation in ship planar section welding based on a digital twin system includes the following steps:

[0166] Step S1. Capture workpiece errors and robot status in real time using a digital twin;

[0167] Step S2. Generate an initial task allocation scheme through the global task allocation ring;

[0168] Step S3. Optimize the welding torch path and attitude through the local path fine-tuning loop;

[0169] Step S4. Welding is performed after dual verification through digital twin pre-simulation and physical data comparison;

[0170] Step S5. Dynamically adjust task allocation based on the load balancing index.

[0171] It should be further explained that, in the specific implementation process, during the dynamic perception and real-time construction of the digital twin, the assembly error of the workpiece and the actual position of the weld are captured by the fusion of laser scanning and visual sensing. When the assembly deviation is detected, the cross-seam weld is automatically segmented into an independent task segment. The robot motion state, welding process parameters and environmental data are loaded in real time to construct a four-dimensional digital twin and update the dynamic topology map. Among them, the robot motion state includes the X-axis position of the gantry and the θ-axis rotation angle, the welding process parameters include the extension length of 15-20mm, and the environmental data includes the workshop temperature.

[0172] During global task allocation and conflict resolution, an anti-conflict auction algorithm is initiated: each robot bids for welding tasks based on its efficiency value Q. For vertical welding tasks with a height ≤ 3300mm, the Z-axis travel weight is significantly increased, and for narrow weld areas with Δ < 35mm, the conflict risk coefficient is doubled. The gantry track is discretized into a spatiotemporal grid, and the winning robot declares the grid window it occupies. If the windows overlap, a waiting delay or deceleration is inserted to avoid them. For corner welding tasks with Δ ≥ 35mm, an additional θ-axis attitude adjustment time window is reserved to avoid arc termination interference.

[0173] During local path adaptive optimization, a local path fine-tuning loop is triggered, including: in the region where the stiffener gap Δ < 35mm, the fuzzy RRT algorithm is used to constrain the welding torch movement direction; if the anti-collision sensor alarms, the extension length is shortened to the lower limit of the process; for circular arc welds, when the center angle exceeds the θ axis rotation range of ±185°, automatic segmentation is performed, and the tangent angle of each segment of the welding torch is optimized; if the real-time arc data fluctuates abnormally, welding is paused and the weld is moved back 5mm to generate a continuation ramp path; wherein, the real-time arc data includes voltage and current.

[0174] During the dual verification and online correction, the optimized path is run on the KCONG platform. If the joint torque exceeds the limit or the trajectory deviation is greater than 0.3mm, a level 3 response is triggered. The level 3 response includes fine-tuning, re-bidding, and segmentation. The molten pool state is inverted through welding machine data. If the actual weld leg size continues to deviate from the predicted value, the wire feed speed or gas flow rate is adjusted until the arc stabilizes.

[0175] During the dynamic migration of load balancing, the load balancing index (LBI) is calculated in real time. When the ratio of the duration of high and low load units exceeds the threshold, welds longer than 10m are divided into sub-segments and transferred to low-load robots according to the proximity of the gantry coordinates. Vertical welding tasks are only migrated to units with Z-axis travel greater than or equal to the actual height. Tasks in narrow areas are limited to high-precision robots with θ-axis. If a robot fails and goes offline, unfinished welds are assigned to healthy units according to spatial proximity. If the Z-axis is insufficient, only the lower weldable section is transferred.

[0176] The winning bid instruction was sent to the Kawasaki controller via 5G, driving the θ-axis rotation mechanism (model: Nabtesco RV-320C) to execute; the process adjustment instruction was synchronized to the Panasonic YD-500GR welding machine and wire feeding mechanism.

[0177] Real-time segmentation and independent optimization of each path segment ensures that the welding torch maintains a constant extension length in areas of height abrupt change, eliminating weld trajectory deviation; the large gantry track is discretized into a grid window, and waiting delays are dynamically inserted through a declaration-arbitration mechanism; an θ-axis adjustment window is reserved at the end of the corner welding to resolve the contradiction between robot collision and cycle time; when the arc is abnormal, the robot retreats 5mm to generate a ramp continuation path, reduces welding speed at the initial stage of arc initiation, and resumes after the molten pool stabilizes, resolving appearance defects in the continuation welding; based on the forced verification of Z-axis travel capability during vertical welding migration, the task in narrow areas is bound to θ-axis accuracy to avoid the risk of capability mismatch.

[0178] Through dynamic optimization of the entire process, including error segmentation, spatiotemporal obstacle avoidance, ramp continuation, and capability verification, assembly variations, spatial conflicts, and resource imbalances in ship welding are transformed into a closed-loop controllable process; the system systematically solves the problems of task overlap, path interference, and uneven load.

[0179] It should be further explained that, in the specific implementation process, after the system is started, it first uses a high-precision laser scanning device and an industrial vision sensor to jointly scan the ship section workpiece. The scanning process captures the surface contour features of the workpiece in real time, focusing on detecting weld positions and assembly deviations. When a significant height difference is detected between adjacent spliced ​​plates, the system automatically marks the weld in that area as a special task segment and performs precise segmentation based on the location of the height abrupt change. Simultaneously, it integrates the position feedback of the gantry crane, robot joint angle data, and environmental temperature and humidity parameters to construct a dynamically updated digital twin model. This model merges the physical space equipment status, workpiece characteristics, and welding process rules into a virtual mapping, providing a basis for subsequent decision-making.

[0180] The task allocation engine operates based on an improved collaborative bidding algorithm. Each robot generates a comprehensive performance score based on its own hardware capabilities. The score calculation covers three dimensions, including the following:

[0181] Welding efficiency dimension: The movement time is calculated by combining the upper limit of the gantry's walking speed with the current load;

[0182] Process matching dimensions: For vertical welding tasks, verify the travel margin of the lifting mechanism; for weld seams in narrow areas, assess the risk level of path conflict.

[0183] Quality assurance dimension: correlate with historical data on the pass rate of similar welds.

[0184] During the bidding process, the system discretizes the gantry's walking track into equidistant spatiotemporal units. The winning robot must declare the sequence and duration of unit occupancy. When overlapping unit occupancy declarations are detected across multiple robots, time differential compensation is automatically inserted: for warning-level conflicts, the movement speed of the following robot is reduced; for emergency-level conflicts, the following robot is forced to pause and wait for the conflicting unit to be released. An additional time window for adjusting the rotation mechanism's attitude is reserved for critical finishing processes such as corner welding.

[0185] The path planning engine is activated under specific conditions, including:

[0186] Working in confined spaces: When the weld is located in an area with dense stiffeners, the system dynamically generates a safe passage boundary using laser ranging data. The planning algorithm explores feasible paths under boundary constraints. If the real-time anti-collision sensor triggers an alarm, the welding torch extension length is automatically shortened to the lower limit allowed by the process, and the path is recalculated.

[0187] Complex trajectory processing: For circular arc welds, the system analyzes the geometric center angle data. When the angle value exceeds the range of motion of the rotating mechanism, intelligent segmentation is performed and the travel angle of each segment of the welding torch is independently optimized to ensure that the end of the welding torch always maintains the optimal process tilt angle.

[0188] Abnormal response mechanism: If the arc monitoring data continues to be abnormal during the welding process, the system will immediately suspend the operation and initiate a multi-level response. The response includes: cleaning the welding torch nozzle and retracting to the nearest qualified weld point, generating a welding transition path based on real-time scanning data, and restarting the arc in low-speed mode until the molten pool is stable.

[0189] The optimized path requires dual verification: virtual environment pre-simulation and physical data closure; among which:

[0190] Virtual environment simulation: Load the current digital twin model onto the simulation platform and inject the real-time acquired assembly error data and execution path. When the robot joint load exceeds the limit or the welding torch trajectory deviation exceeds the standard, the system automatically triggers a three-level processing procedure:

[0191] Primary response: Fine-tuning the spatial coordinates of path points;

[0192] Intermediate Response: Release the current task and re-bid;

[0193] Advanced Response: Multi-robot collaboration for the handover of segmented ultra-long weld seams.

[0194] Physical data closed loop: Arc characteristic parameters are acquired in real time through the welding power source, and the actual molten pool state is calculated by inversion. When the actual weld leg size deviates from the predicted value continuously, the wire feeding mechanism speed and shielding gas flow rate are dynamically adjusted until the process parameters return to the stable range. For corner welding tasks, the deviation between the actual contour and the predicted model is checked simultaneously, and the spatial posture of the welding torch is finely adjusted by the rotation mechanism when necessary.

[0195] The system continuously monitors the task load status of each robot, including routine load balancing and handling of abnormal operating conditions; among which:

[0196] Routine load balancing: When a high-load robot's task duration is detected to significantly exceed the system average, the long weld seam it is handling is automatically divided according to spatial location. The divided sub-task segments are preferentially transferred to adjacent low-load robots. During the transfer process, the travel of the target robot's lifting mechanism and the accuracy of its rotating mechanism are strictly verified to ensure they meet the process requirements.

[0197] Abnormal operating condition handling: When a robot goes offline due to a malfunction, the system allocates its unfinished welds to healthy units based on spatial proximity. For vertical welding sections beyond the target robot's capabilities, only the operable parts are transferred and the areas to be processed are marked. After the malfunctioning robot returns to online status, the system employs a gradual task loading strategy to slowly reintegrate it into the production line.

[0198] The system constructs a data transmission channel through a dedicated 5G network, including real-time data hierarchical transmission: collision warning commands and process anomaly signals occupy a high-priority channel; weld scan data and equipment status information are allocated balanced bandwidth; and historical data are transmitted during idle periods.

[0199] It also includes anti-interference strategies: when electromagnetic interference in the workshop causes a decline in communication quality, a dual-path redundant transmission mode is automatically activated, relaying critical commands in parallel through base stations and edge computing nodes. In extreme network outage scenarios, each robot switches to offline collaborative mode: continuing operation using the last effective path, performing emergency obstacle avoidance through local sensors, and recalling pre-stored process parameters to maintain basic welding quality. After the network is restored, the system automatically performs data synchronization and path calibration to ensure consistency between the physical space and the digital twin.

[0200] Unlike static planning schemes, this method actively identifies abrupt changes in the height of the spliced ​​panels through real-time scanning. When a height difference exceeds a process threshold, the weld seam is automatically segmented and the path parameters for each segment are independently optimized. This mechanism ensures that the welding torch maintains a constant extension length in areas of height abrupt changes, eliminating weld seam trajectory deviation caused by assembly errors.

[0201] By discretizing the physical trajectory into a spatiotemporal cell grid, the motion of multiple robots is dynamically coordinated through a declaration-arbitration mechanism. A time delay reserve window is designed for special processes such as corner welding, ensuring arc termination quality while avoiding production line cycle time losses caused by emergency stops.

[0202] When the welding process is interrupted due to an anomaly, the system executes a three-step recovery process, including: laser positioning of welding coordinates to replace manual alignment; generation of a transition ramp path to bridge trajectory deviations; and initial low-speed welding to ensure the quality of the reconstructed molten pool. This solves the problem of unstable quality at the weld continuation points in ship section welding.

[0203] The task redistribution process establishes a dual verification mechanism: vertical welding tasks are strictly matched with the travel capacity of the lifting mechanism; narrow area operations are bound to high-precision rotating units; thus avoiding the risk of welding defects caused by hardware mismatch during load balancing.

[0204] By implementing a complete technology chain encompassing error perception, dynamic allocation, path optimization, closed-loop verification, and load migration, fully autonomous optimization of multi-robot collaborative operations in ship planar section welding has been achieved. By transforming pain points such as assembly variations, spatial constraints, and resource imbalances into calculable and controllable optimization parameters, the system effectively improves production line equipment utilization and production cycle stability while ensuring weld leg size accuracy and weld appearance quality.

[0205] A multi-robot task allocation method for ship planar section welding based on digital twins includes the following steps:

[0206] Step S1: Start the laser scanning equipment and vision sensor to capture the workpiece surface contour and weld position; collect the coordinates of the gantry crane, robot joint angles, and ambient temperature and humidity data in real time; and construct a digital twin model that integrates physical state and process rules.

[0207] Step S2: Detect whether the height difference of the splicing plates exceeds the process threshold; if it exceeds the threshold, automatically divide the weld seam across the height difference area into independent task segments; configure welding path parameters independently for each task segment to ensure that the welding torch extension length is constant;

[0208] Step S3: Each robot generates a performance score based on its hardware capabilities, which includes welding efficiency, process matching degree, and quality history; initiates a conflict-resistant auction algorithm to bid for the weld seam task; discretizes the gantry track into a spatiotemporal cell grid, and the winning robot declares the occupied cells and time window;

[0209] Step S4: Detect whether the spatiotemporal unit declarations of multiple robots overlap; if the warning levels overlap, reduce the moving speed of the following robot; if the emergency levels overlap, pause the following robot until the unit is released; for corner welding tasks, reserve a time window for the attitude adjustment of the rotating mechanism.

[0210] Step S5: When the weld is located in a dense stiffener area, generate a safe passage boundary based on laser ranging data; plan a collision-free path under boundary constraints; if the anti-collision sensor triggers an alarm, shorten the welding torch extension length to the lower limit of the process and replan.

[0211] Step S6: Analyze the geometric center angle of the circular arc weld; if the angle exceeds the range of motion of the rotating mechanism, divide the weld into multiple segments; independently optimize the welding torch travel angle of each segment to maintain the best process tilt angle;

[0212] Step S7: If the arc monitoring data continues to be abnormal, pause welding and clean the welding torch nozzle; rewind to the nearest qualified weld point; generate a transition ramp path based on real-time scanning; restart the arc in low-speed mode until the molten pool stabilizes;

[0213] Step S8: Load the digital twin model to simulate the path on the simulation platform; if the joint load exceeds the limit or the trajectory deviation exceeds the standard, trigger a three-level response: fine-tune the path coordinates, release the task re-bidding, and split the weld seam collaboration; invert the molten pool state through arc data and dynamically adjust the wire feeding speed and gas flow rate;

[0214] Step S9: Calculate the task load index of each robot in real time; when the task duration of a high-load unit significantly exceeds the average, divide its long weld seam according to its spatial location; after strictly verifying the lifting mechanism stroke and rotation accuracy of the target robot, transfer the sub-task; if the robot fails and goes offline, allocate the unfinished weld seam to the healthy unit according to proximity.

[0215] Step S10: Transmit data in a hierarchical manner through the 5G private network: collision warning and process abnormality signals occupy the high-priority channel exclusively; dual-path redundant transmission is enabled when there is strong interference in the workshop; during network interruption, the robot follows the last effective path and avoids obstacles in an emergency using local sensors; after the network is restored, the data is synchronized and the path is calibrated.

[0216] By integrating workpiece assembly errors, robot motion states, and welding process rules in real time using a digital twin, a weld seam task topology map is dynamically constructed. Combined with an anti-collision auction algorithm, precise allocation of multi-robot tasks is achieved. A spatiotemporal gridded collaborative mechanism effectively resolves motion conflicts between the large and small gantry mechanisms, ensuring zero interference in the welding path. A local path optimization module adaptively adjusts the welding torch posture and extension length in narrow spaces, eliminating the risk of physical collisions. A dual verification system corrects trajectory deviations and process parameters in real time, ensuring that the weld leg dimensions of complex welds such as vertical welds and fillet welds meet standards, resulting in a smooth and full weld appearance and eliminating the problem of recurring defects such as porosity and cracks.

[0217] The system employs a load balancing mechanism based on hardware capability constraints, quantifying key parameters such as the lifting mechanism's stroke and rotational accuracy into task allocation weights. This enables intelligent matching of vertical welding tasks with the robot's Z-axis stroke and directional binding of narrow weld seams with high-precision θ-axis units, eliminating quality risks caused by capability mismatch. The 5G enhanced communication framework ensures closed-loop task allocation through hierarchical transmission and redundant anti-interference strategies; offline collaboration and reconnection mechanisms maintain continuous production line operation in network outage emergency modes. Ultimately, this results in further improved equipment utilization and stable production cycle time, providing a fully autonomous optimization solution for the welding of large ship sections.

[0218] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0219] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-robot task allocation system for ship planar section welding based on digital twins, characterized in that, include: The 4D digital twin module collects workpiece assembly errors, weld positions, and robot status data in real time to build a twin model that includes a physics engine layer, a process knowledge base, and a dynamic topology graph. Dual-loop optimization engine module: dynamically allocates weld seam tasks through a global task allocation loop and generates conflict-free paths through a local path fine-tuning loop; Multiple verification module: Performs digital twin pre-simulation verification and physical-virtual data comparison on the optimization results to ensure that there are no path conflicts and that the welding quality meets the standards; Load balancing module: Calculates the task load index of each robot in real time and dynamically adjusts task allocation; The global task allocation ring employs a conflict-resistant auction algorithm. Each robot competes for welding tasks based on its efficiency value, which is calculated based on welding speed, weld matching degree, and path conflict risk. The travel of the large gantry walking mechanism is discretized into a spatiotemporal grid, and path conflicts are dynamically avoided through virtual time windows; In the conflict-resistant auction algorithm: When the distance between the two robots is below the safety threshold, the robot is forced to slow down and a waiting time window is inserted. The weighting coefficient of the efficiency value is dynamically adjusted according to the weld type: vertical welding tasks are preferentially assigned to robots with large Z-axis travel.

2. The multi-robot task allocation system for ship planar section welding based on digital twin as described in claim 1, characterized in that: The four-dimensional digital twin module includes: Real-time data layer: Captures workpiece assembly errors and weld position deviations by fusing laser scanning sensors and vision sensors; Physics Engine Layer: Integrates multibody dynamics simulation model to pre-simulate robot motion path interference and welding thermal deformation effects; Process knowledge base: Embedded welding rules generated by SMARTWELD, transforming weld leg range and extension length constraints into optimization objective functions; Dynamic topology graph: with weld seams as nodes and robot reachable paths as edges, the weights include welding time and quality risk coefficients.

3. The multi-robot task allocation system for ship planar section welding based on digital twin as described in claim 1, characterized in that: The local path fine-tuning loop includes: Based on the adaptive fuzzy RRT algorithm, combined with the weld type and the robot's rotational degrees of freedom θ axis ±185°, a collision-free path is generated in a narrow space. Adjust the welding torch extension length and angle using the welding torch posture optimizer to ensure weld fullness.

4. The multi-robot task allocation system for ship planar section welding based on digital twin as described in claim 1, characterized in that: The multi-factor authentication module performs the following: Digital twin pre-simulation verification: Real-time error data is injected into the KCONG simulation environment to detect path deviation and joint torque exceeding limits; Physical-virtual data comparison: The state of the molten pool is inverted by the arc data of the welding power source and compared with the predicted values ​​of the twin model to trigger further optimization of process parameters.

5. The multi-robot task allocation system for ship planar section welding based on digital twin as described in claim 1, characterized in that: The load balancing module: Calculate the robot load balancing index (LBI). When the index falls below a set threshold, the tasks of the high-loaded robots are redistributed to idle units. The index is dynamically updated based on the ratio of the variance of the duration of each robot task to the average duration.

6. The multi-robot task allocation system for ship planar section welding based on digital twin as described in claim 1, characterized in that, It also includes a quality defect prediction unit: Based on real-time arc data from welding power sources, a probability prediction model for weld porosity and cracks is constructed. When the defect probability exceeds the limit, the task is paused and a local path fine-tuning loop is triggered for replanning.

7. The multi-robot task allocation system for ship planar section welding based on digital twin as described in claim 1, characterized in that: The system transmits laser scanning data and control commands through a 5G communication module, and the task allocation response cycle is lower than a set threshold.

8. A method for multi-robot task allocation in the system for welding planar sections of a ship as described in any one of claims 1-7, characterized in that, Includes the following steps: Step S1. Capture workpiece errors and robot status in real time using a digital twin; Step S2. Generate an initial task allocation scheme through the global task allocation ring; Step S3. Optimize the welding torch path and attitude through the local path fine-tuning loop; Step S4. Welding is performed after dual verification through digital twin pre-simulation and physical data comparison; Step S5. Dynamically adjust task allocation based on the load balancing index.

Images