Multi-robot path planning system for ship planar sub-assembly welding based on visual detection

Abstract

This invention discloses a vision-based multi-robot path planning system for welding planar sections of ships. The invention relates to the field of shipbuilding technology and includes a multispectral vision inspection module mounted on a gantry, comprising a laser scanning sensor and an infrared thermal imager for real-time acquisition of three-dimensional point cloud and temperature distribution data of the ship's planar sections. This vision-based multi-robot path planning system for welding planar sections of ships fundamentally solves the problem of inaccurate path planning caused by assembly errors, thermal deformation, and complex working conditions in the welding of large ship planar sections through real-time multispectral vision perception and a material-adaptive dynamic correction mechanism. By employing a composite correction vector and spatiotemporal grid collaborative prediction technology, the system effectively improves weld trajectory accuracy, ensuring that weld leg dimensions strictly meet standards in high-difficulty processes such as fillet welding and vertical welding of the largest workpieces, eliminating defects such as porosity and cracks, and significantly reducing rework rates and quality risks.

Description

Technical Field

[0001] This invention relates to the field of shipbuilding technology, specifically to a multi-robot path planning system for ship planar segment welding based on vision inspection. Background Technology

[0002] Welding of ship planar sections, such as hull sections and deck sections, requires the completion of various types of welds on large workpieces, i.e., up to 15m × 15m, including flat fillet welds, vertical fillet welds, and wrap fillet welds. Traditional robotic welding production lines use a pre-programmed offline planning method: weld information is generated through SMARTWELD software, converted into robot programs and simulated by the KCONG system, and then fixed NCDATA data is called on-site to execute welding. However, this method suffers from poor dynamic adaptability and low efficiency in multi-robot collaboration. Specifically: poor dynamic adaptability: workpiece assembly errors, assembly gaps, and on-site thermal deformation cause deviations between actual working conditions and the pre-programmed model, and existing systems lack real-time perception and path correction capabilities, easily leading to welding torch interference, such as failure to wrap fillet welds with a Δ < 35mm or substandard welding quality, with weld leg deviation allowed only 0-0.5mm. Low efficiency in multi-robot collaboration: the production line requires multiple robots to work collaboratively in a shared space, such as a gantry crane with a Y-axis travel of 11-21m, but the offline program cannot dynamically adjust the path according to real-time task allocation, resulting in robot waiting or trajectory conflicts. The pressing issue is how to achieve real-time dynamic path planning for multiple robots in large-scale, high-precision welding scenarios to address deviations in on-site working conditions and optimize collaborative efficiency. This has become a decisive challenge in improving the quality and efficiency of automated welding for ship sections. Summary of the Invention

[0003] To achieve the above objectives, the present invention provides the following technical solution: a vision-based multi-robot path planning system for ship planar segment welding, comprising a gantry crane, welding robots, and a control cabinet.

[0004] The multispectral vision inspection module, mounted on the gantry, includes a laser scanning sensor and an infrared thermal imager, used to acquire three-dimensional point cloud and temperature distribution data of the ship's planar sections in real time.

[0005] The dynamic path planning module communicates with the control cabinet, receives visual inspection data, and outputs corrected paths.

[0006] The multi-robot collaborative control module is electrically connected to the dynamic path planning module to dynamically allocate motion tasks to each robot;

[0007] The dual closed-loop verification module interacts with the dynamic path planning module and the welding robot signal respectively to perform path pre-verification and welding quality monitoring.

[0008] Preferably, the dynamic path planning module includes a deformation compensation unit and a spatiotemporal mesh generation unit; the deformation compensation unit calculates the assembly error compensation amount based on the three-dimensional point cloud data, predicts the thermal deformation compensation amount based on the temperature distribution data, and generates a composite correction vector; the spatiotemporal mesh generation unit discretizes the ship's planar segments according to a preset mesh size, marks the weld type feature points, and allocates time windows; the preset mesh size is set based on the ship segment accuracy requirements and weld feature recognition requirements.

[0009] Preferably, the deformation compensation unit performs the following operations:

[0010] Based on the deviation between the actual contour of the workpiece and the preset model identified by laser point cloud, geometric correction components are output.

[0011] Based on infrared thermal imaging data, the deformation trend is calculated through a heat conduction model, and the thermal compensation component is output.

[0012] Based on the ship steel grade, a dynamic weighting coefficient is matched, and a composite correction vector is generated by integrating geometric correction components and thermal compensation components.

[0013] Preferably, the spatiotemporal grid generation unit includes a conflict prediction subunit, which activates a dynamic avoidance mechanism when spatiotemporal grid overlap is detected within a critical distance Δ on the robot path.

[0014] If Δ is less than the minimum allowable distance for fillet welding, force the path priority to be adjusted.

[0015] If Δ is greater than the safe operating distance, maintain the original path planning.

[0016] Preferably, the multi-robot collaborative control module deploys a distributed reinforcement learning algorithm framework, wherein:

[0017] Each welding robot acts as an independent intelligent agent, generating motion strategies based on welding quality parameters and motion efficiency parameters.

[0018] The central decision-maker coordinates the action strategies of each agent based on the spatiotemporal grid model to resolve resource conflicts in the Y-axis movement of the gantry.

[0019] Preferably, the dual closed-loop verification module includes:

[0020] The dynamic error envelope analysis unit compares the actual value of the solder leg size with the preset threshold in real time and triggers correction or interruption commands.

[0021] The virtual twin collision simulation unit constructs a virtual scene based on real-time 3D point clouds and injects random assembly errors to simulate collision risks.

[0022] Preferably, the virtual twin collision pre-simulation unit performs Monte Carlo simulation, the process of which includes:

[0023] Generate a random assembly error dataset that follows a normal distribution;

[0024] Perform multiple rounds of collision detection on each planned path;

[0025] When the collision probability exceeds the set threshold, a replanning instruction is sent to the dynamic path planning module.

[0026] Preferably, the weld type feature points include the end point of a flat fillet weld, the turning point of a vertical fillet weld, and the positioning point of a wrap fillet weld; the time window is calculated based on the robot's maximum motion acceleration and joint rotation angle.

[0027] Preferably, it also includes a process parameter adaptive unit that synchronizes data with the dynamic path planning module and matches welding current, wire feeding speed and gas flow parameters according to the weld type.

[0028] Preferably, the matching logic of the dynamic weighting coefficient is as follows: for AH32 / DH36 high-strength steel, increase the weight of the thermal compensation component; for ordinary marine grade A steel, increase the weight of the geometric correction component.

[0029] This invention provides a vision-based multi-robot path planning system for ship planar segment welding. It offers the following advantages:

[0030] This vision-based multi-robot path planning system for ship planar segment welding fundamentally solves the problem of inaccurate path planning caused by assembly errors, thermal deformation, and complex working conditions in large ship planar segment welding through real-time multispectral vision perception and material adaptive dynamic correction mechanisms. Employing composite correction vector and spatiotemporal grid collaborative prediction technology, it effectively improves weld trajectory accuracy, ensuring that weld leg dimensions strictly meet standards in challenging processes such as fillet welding and vertical welding of the largest workpieces, eliminating defects such as porosity and cracks, and significantly reducing rework rates and quality risks.

[0031] This vision-based multi-robot path planning system for ship section welding, with its distributed collaborative decision-making framework and a collision probability verification mechanism of thousands of levels, enables efficient collaborative operation of multiple robots within limited space resources. Through dynamic priority arbitration and real-time synchronization of path parameters, it effectively avoids mechanical interference accidents, optimizes the allocation of gantry movement resources, shortens the welding cycle, and improves the overall throughput efficiency of the production line. The dual safety closed-loop design provides reliable assurance for automated welding of ship sections, driving the upgrading of large structural component manufacturing towards intelligent manufacturing. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the module interaction of the vision-based multi-robot path planning system for ship planar segment welding of the present invention;

[0033] Figure 2This is a flowchart illustrating the multi-robot path planning method for ship planar segment welding based on vision detection, as described in this invention. Detailed Implementation

[0034] 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.

[0035] Please see Figure 1 and Figure 2 This invention provides a technical solution: a multi-robot path planning system for ship planar segment welding based on vision detection, comprising a gantry, welding robots, and a control cabinet.

[0036] The multispectral vision inspection module, mounted on the gantry, includes a laser scanning sensor and an infrared thermal imager, used to acquire three-dimensional point cloud and temperature distribution data of the ship's planar sections in real time.

[0037] The dynamic path planning module communicates with the control cabinet, receives visual inspection data, and outputs corrected paths.

[0038] The multi-robot collaborative control module is electrically connected to the dynamic path planning module to dynamically allocate motion tasks to each robot;

[0039] The dual closed-loop verification module interacts with the dynamic path planning module and the welding robot signal respectively to perform path pre-verification and welding quality monitoring.

[0040] It should be further explained that, in the specific implementation process, the implementation process of the vision-based multi-robot path planning system for ship planar segment welding is as follows:

[0041] A laser scanning sensor mounted on a gantry acquires a three-dimensional point cloud of the workpiece surface, detecting the geometric deviation between the actual assembly contour and the preset model, including assembly errors of ≤±3mm and assembly gaps of ≤2mm; an infrared thermal imager monitors the temperature distribution in the welding area, identifying local deformation trends caused by thermal deformation.

[0042] The actual weld position offset is calculated based on laser point cloud data, and a geometric correction component is generated. The deformation direction from the high-temperature region to the low-temperature region is predicted based on thermal imaging data, and a thermal compensation component is output. If the workpiece material is AH32 / DH36 high-strength steel, the weight of the thermal compensation component is increased first. If it is ordinary A-grade steel, the geometric correction component is emphasized. Finally, a composite correction vector for driving robot pose adjustment is output.

[0043] The workpiece is divided into a 50mm×50mm grid, and the characteristic positions such as the end point of the fillet weld and the turning point of the vertical weld are marked. When the planned paths of two or more robots overlap within the critical distance Δ, if Δ<35mm, the path of the robot with lower priority is forced to detour or be delayed; if Δ≥65mm, the original path is maintained. The minimum allowable distance for the fillet weld is 35mm.

[0044] Each robot acts as an independent intelligent agent, generating motion strategies based on welding quality requirements and motion efficiency. The central decision-maker dynamically allocates the gantry Y-axis resources to avoid travel conflicts. Among these, the welding quality requirements include a permissible deviation of ±0.5mm for weld leg dimensions.

[0045] The system monitors weld leg dimensions in real time. If the actual value exceeds the preset threshold of ±0.5mm three times consecutively, welding is interrupted and an alarm is triggered. A virtual work scene is constructed based on real-time point clouds, and random assembly errors are injected to perform thousands of collision simulations. If the collision probability exceeds the safety threshold, the current path is immediately terminated and replanned. Based on weld types such as flat fillet welds and vertical fillet welds, the system automatically matches welding current and wire feed speed parameters and updates them to the execution terminal synchronously with the corrected path.

[0046] The dynamic path planning module includes a deformation compensation unit and a spatiotemporal mesh generation unit. The deformation compensation unit calculates the assembly error compensation amount based on the three-dimensional point cloud data, predicts the thermal deformation compensation amount based on the temperature distribution data, and generates a composite correction vector. The spatiotemporal mesh generation unit discretizes the ship's planar segments into 50mm×50mm grids, marks the weld type feature points, and allocates time windows.

[0047] It should be further explained that, in the specific implementation process, the three-dimensional point cloud data of the laser scanning sensor is received, the geometric deviation between the actual contour of the workpiece and the preset model is identified, including assembly errors and assembly gaps, and geometric correction components are generated; the temperature distribution data of the infrared thermal imager is analyzed simultaneously, the deformation trend of the welding area is predicted through the heat conduction model, and the heat compensation components are output.

[0048] If the workpiece material is AH32 / DH36 high-strength steel, the weight of the thermal compensation component is elevated to the dominant position; if it is ordinary A-grade steel, the geometric correction component is used first; finally, a composite correction vector is generated to drive the robot arm's pose adjustment and sent to the welding robot in real time.

[0049] The surface of the ship's planar segment is discretized into a 50mm×50mm grid, and the weld type feature points are marked on the grid nodes, such as the positioning point of the fillet weld and the turning point of the vertical fillet weld; the time window occupied by each grid is calculated based on the robot's maximum motion acceleration and joint rotation angle.

[0050] When multiple robot paths are detected to overlap in time and space within a critical distance Δ: if Δ < 35 mm, the robot path with lower priority is immediately marked as "forced avoidance", triggering a detour or delayed execution command; if Δ ≥ safe operating distance: 65 mm, the original path planning remains unchanged; the gantry Y-axis movement resources are dynamically allocated through the central decision-maker to ensure continuous welding of high-priority welds; whereby the critical distance Δ is set as the minimum allowable distance for fillet welds of 35 mm.

[0051] The deformation compensation unit performs the following operations:

[0052] Based on the deviation between the actual contour of the workpiece and the preset model identified by laser point cloud, geometric correction components are output.

[0053] Based on infrared thermal imaging data, the deformation trend is calculated through a heat conduction model, and the thermal compensation component is output.

[0054] Based on the ship steel grade, a dynamic weighting coefficient is matched, and a composite correction vector is generated by integrating geometric correction components and thermal compensation components.

[0055] It should be further explained that, in the specific implementation process, the laser scanning sensor captures the three-dimensional point cloud of the workpiece surface, registers the actual contour with the preset model, and identifies the assembly error: ±3mm deviation and the assembly gap: ≤2mm gap; for the detected deviation, the position offset of the key points of the weld is calculated, and the geometric correction component for driving the welding torch posture adjustment is generated. The key points of the weld include the end point of the fillet weld and the turning point of the vertical weld.

[0056] Infrared thermal imagers monitor the temperature gradient of the welding area in real time and construct a heat conduction model based on the thermal expansion coefficient of the steel. If a local high-temperature zone greater than 300℃ is detected, the direction and magnitude of deformation towards the low-temperature zone are predicted. For AH32 / DH36 high-strength steel, due to its high thermal sensitivity, the predicted deformation magnitude is amplified by 1.2 times. For ordinary Grade A steel, a standard deformation prediction model is used. The welding torch path compensation vector caused by thermal deformation is output as a thermal compensation component.

[0057] The fusion strategy is automatically selected based on the steel grade, including: in the case of high-strength steel: the weight of the thermal compensation component is increased to more than 70%, and the weight of the geometric correction component is ≤30%; in the case of ordinary steel: the weight of the geometric correction component is ≥80%, and the thermal compensation component is used as an auxiliary adjustment; the fused composite correction vector is sent to the welding robot controller in real time to drive the welding torch to adjust its spatial pose synchronously.

[0058] During the welding process, the weld leg size is checked three times consecutively. If the actual value still deviates from the preset threshold of ±0.5mm, the compensation is deemed to have failed, and a manual intervention protocol is triggered. If the deviation is stable within the allowable range, dynamic correction is continuously executed.

[0059] The spatiotemporal mesh generation unit has a conflict prediction sub-unit. When spatiotemporal mesh overlap is detected within a critical distance Δ on the robot path, a dynamic avoidance mechanism is activated.

[0060] If Δ is less than the minimum allowable distance for fillet welding, force the path priority to be adjusted.

[0061] If Δ is greater than the safe operating distance, maintain the original path planning.

[0062] It should be further explained that, in the specific implementation process, the real-time distance between the endpoints of any two robot welding guns is continuously calculated in the spatiotemporal grid model; when the distance Δ is detected to enter the warning interval, the dynamic avoidance strategy is activated: if high-precision welds such as fillet welds or vertical welds are being performed, the robot is marked as high priority; if ordinary flat fillet welds are being performed, it is marked as low priority; the warning interval is: 35mm≤Δ<65mm.

[0063] For high-risk interference zones with Δ < 35mm: force low-priority robots to perform avoidance maneuvers. The detour option is to shift 50mm in the positive Y-axis direction and replan the path. The delay option is to pause movement until the high-priority robot leaves the critical area. If both robots are high-priority, request arbitration from the central decision-maker to adjust the path order. For safe operating zones with Δ ≥ 65mm: maintain the original path planning and only record the time window occupied by this grid.

[0064] The central decision-maker generates a sequence of Y-axis movement instructions based on the spatiotemporal grid occupancy status: when multiple robots apply for Y-axis movement at the same time, priority is given to the robot corresponding to the high-precision weld; for robots that delay avoidance, their waiting time is compensated to subsequent empty grid windows.

[0065] After the avoidance action is executed, the Δ value is rechecked: if Δ is still <35mm, the avoidance is determined to be a failure and the secondary avoidance protocol (such as Z-axis lifting) is triggered; if Δ is ≥65mm, the avoidance lock state is released.

[0066] The multi-robot collaborative control module deploys a distributed reinforcement learning algorithm framework, in which:

[0067] Each welding robot acts as an independent intelligent agent, generating motion strategies based on welding quality parameters and motion efficiency parameters.

[0068] The central decision-maker coordinates the action strategies of each agent based on the spatiotemporal grid model to resolve resource conflicts in the Y-axis movement of the gantry.

[0069] It should be further explained that, in the specific implementation process, each welding robot acts as an independent intelligent agent, collecting its own welding quality data and motion status data in real time. The welding quality data includes weld leg size deviation, and the motion status data includes acceleration and joint angle. If high-precision welds such as fillet welds or vertical welds are being performed, weld leg size accuracy is the primary optimization target. If ordinary flat fillet welds are being performed, motion efficiency is prioritized. Subsequently, a motion strategy is generated based on the optimization target: when a path conflict risk is detected, the impact of the detour path on the welding quality is evaluated; when there is no conflict, the shortest time path is selected.

[0070] Receive spatiotemporal grid occupancy requests submitted by all agents and identify the resource contention area of ​​the gantry Y-axis; when multiple agents request the same Y-axis travel segment: the high-precision weld seam task agent automatically obtains priority use rights; the ordinary weld seam task agent accepts delay instructions or path segmentation schemes; generate resource allocation instructions and send them to each robot controller.

[0071] For agents that wait due to avoidance, the subsequent grid time window is automatically compensated; if the detour path affects the welding quality, such as vertical welding being interrupted, the central decision-maker reallocates the task boundaries of adjacent robots.

[0072] For every 10 meters of weld completed, the number of task interruptions and the total welding time are recorded: if the number of interruptions for high-precision welds exceeds the limit, their subsequent priority weight is increased; if ordinary welds time out, the path segmentation algorithm is optimized.

[0073] The dual closed-loop verification module includes:

[0074] The dynamic error envelope analysis unit compares the actual value of the solder leg size with the preset threshold in real time and triggers correction or interruption commands.

[0075] The virtual twin collision simulation unit constructs a virtual scene based on real-time 3D point clouds and injects random assembly errors to simulate collision risks.

[0076] It should be further explained that, in the specific implementation process, the actual weld leg size data is continuously collected during the welding process and compared with the preset threshold in real time; if a single over-limit is detected, the welding current and robot speed are automatically fine-tuned for compensation; if the limit is still exceeded after three consecutive compensations, it is determined to be a system-level failure, welding is immediately interrupted and an audible and visual alarm is triggered, and the robot posture is frozen for manual intervention.

[0077] A virtual working environment consistent with the physical scene is constructed based on real-time 3D point cloud, generating a random assembly error dataset that follows a normal distribution to simulate typical working condition fluctuations of ship sections; thousands of independent collision simulations are performed on each planned path: if the collision probability is lower than the safety threshold, the path is approved for execution; if the collision probability exceeds the standard, an emergency replanning instruction is sent to the dynamic path planning module.

[0078] The quality monitoring unit and the collision simulation unit share a data interface: when the collision simulation requires path replanning, the quality monitoring unit is simultaneously notified to suspend detection; after the new path is executed, the quality monitoring unit starts the enhanced detection mode.

[0079] The virtual twin collision pre-simulation unit performs Monte Carlo simulations, the process of which includes:

[0080] Generate a random assembly error dataset that follows a normal distribution;

[0081] Perform multiple rounds of collision detection on each planned path;

[0082] When the collision probability exceeds the set threshold, a replanning instruction is sent to the dynamic path planning module.

[0083] It should be further explained that, in the specific implementation process, the three-dimensional point cloud data of the laser scanning sensor is received in real time to generate a virtual ship section model that is completely consistent with the physical workpiece; the spatial envelope dimensions of the robot body, welding torch and gantry motion mechanism are accurately marked in the model to ensure that the simulation geometric accuracy reaches ±0.1mm.

[0084] Generate a random error dataset that conforms to typical working conditions of ship sections: for ordinary Grade A steel workpieces, use a normal distribution error with a standard deviation of 1.5 mm; for high-strength steel (AH32 / DH36) workpieces, due to their sensitivity to thermal deformation, reduce the standard deviation to 1.0 mm; before each simulation, randomly select a set of error values ​​and load them onto the key assembly points of the virtual model.

[0085] Perform 1000 independent simulation tests on a single planned path, with each test randomly rotating the error loading direction, including the X, Y, and Z axes; dynamically simulate model deformation drift caused by welding thermal deformation; and record interference events between the welding torch and the workpiece, the robot, and the gantry structure.

[0086] The number of collisions in a thousand tests is counted. If the number of collisions is ≤1, i.e. the probability is ≤0.1%, the path is deemed safe and execution is approved. If the number of collisions is ≥2, the path is immediately frozen and a red alert is sent to the dynamic path planning module. For high-risk welds, such as fillet welds with Δ<35mm, the safety threshold is tightened to zero collisions.

[0087] When a red alert is triggered, the closest feasible solution in the historical safe path library is automatically called as a temporary path; the welding quality monitoring unit is simultaneously notified to start the highest level of detection; when the first weld section is completed after the new path is executed, the simulation verification loop is restarted.

[0088] Weld type feature points include flat fillet weld endpoints, vertical fillet weld inflection points, and wrap fillet weld positioning points; the time window is calculated based on the robot's maximum motion acceleration and joint rotation angle. It should be further noted that, in the specific implementation process, when positioning the wrap fillet weld endpoints, if the distance Δ between adjacent components is detected to be less than 65mm, it is automatically marked as a high-interference-risk point, and a warning buffer zone with a radius of 35mm is generated; when identifying vertical weld inflection points, if the workpiece height exceeds 800mm or there is a sudden change in curvature, it is marked as a high-precision control point, requiring the robot to slow down to 60% of its standard speed to pass; flat fillet weld endpoints are classified according to the component intersection angle: 75°-90° are normally marked, and <75° triggers a process verification protocol.

[0089] Based on the robot's maximum acceleration and the turntable's rotation angle limit, the shortest time window is calculated for straight weld seams based on the maximum acceleration; for circular arc weld seams, such as fillet weld R-holes, the time window is calculated in conjunction with the ±185° limit of the rotation mechanism.

[0090] When a weld crosses the joint of a spliced ​​plate: if the height difference ΔH ≤ 3mm, a continuous time window is allocated for a single weld; if ΔH > 3mm, it is divided into two independent welds and a separate window is allocated.

[0091] In the workpiece edge area, i.e., less than 200mm from the boundary: the time window is increased by an additional 20% safety margin to prevent gantry inertial vibration; for multi-layer and multi-pass welding processes: the welding time window for the first pass is compressed to 80% of the standard value, and the window duration for subsequent layers is increased according to the heat accumulation effect.

[0092] It also includes a process parameter adaptive unit that synchronizes data with the dynamic path planning module, matching welding current, wire feed speed, and gas flow parameters according to the weld type. Further explanation is needed: during implementation, the visual inspection module identifies the current welding position characteristics: if it is a fillet weld (R-hole or SNIP type), the anti-spatter mode is activated: the gas flow rate is increased to 120% of the standard value, and the welding torch angle is locked at 45°±2°; if it is a vertical fillet weld (height > 800mm), a multi-layer welding strategy is enabled: the first layer current is reduced by 10%, and the wire feed speed is simultaneously reduced; if it is a planar zigzag fillet weld, segmented pulse welding is used: the current is constant in the straight section, and the current at the inflection point increases by 15% to compensate for heat loss.

[0093] Based on the workpiece material, the parameters are dynamically adjusted as follows: For high-strength steel (AH32 / DH36): the upper limit of welding current is reduced to 90% of the standard value to prevent the deterioration of the heat-affected zone; For ordinary grade A steel: the current is allowed to be over-produced by 10% to improve the penetration efficiency; When the primer thickness is detected to be >20μm, the gas flow rate is increased by an additional 15% to ensure arc stability.

[0094] When receiving the composite correction vector from the dynamic path planning module: if the path offset is greater than 2mm, the welding current is dynamically compensated by increasing or decreasing by 2% per millimeter of offset; if the robot's avoidance action causes the welding torch angle to change by more than 5°, the gas flow rate is immediately corrected to the theoretical value; the parameter update command and the robot pose adjustment command are sent synchronously with a delay of less than 50ms.

[0095] The infrared thermal imager monitors the molten pool morphology in real time: if the molten width fluctuation exceeds the preset threshold, the wire feeding speed is automatically increased until it stabilizes; if three consecutive adjustments are ineffective, it reverts to the safety parameter template.

[0096] The matching logic for the dynamic weighting coefficients is as follows: for AH32 / DH36 high-strength steel, increase the weight of the thermal compensation component; for ordinary marine grade A steel, increase the weight of the geometric correction component. It should be further explained that, in the specific implementation process, the surface reflectance spectral characteristics of the workpiece are analyzed using a laser scanning sensor: if characteristic spectral lines of AH32 / DH36 high-strength steel are identified, the thermal deformation-dominated mode is activated; if the spectrum matches that of ordinary grade A steel, the geometric correction-dominated mode is enabled; for areas where spectral identification is not possible, the high-strength steel weighting strategy is used by default.

[0097] The heat compensation component is given a core position. If welding heat accumulation > 300℃·s is detected, the heat compensation weight is increased by an additional 10%. In areas where the vertical welding height is > 800mm, the geometric correction component is used to assist in fine-tuning. The geometric correction component is only used to compensate for sudden assembly deviations.

[0098] When the geometric correction component is dominant, i.e., when the geometric correction component exceeds 80%, the geometric correction weight increases to 95% when the assembly gap is >1.5mm; when the primer thickness is detected to be >20μm, the thermal compensation component intervenes to prevent arc drift; thermal deformation compensation is only used in high-temperature dense areas, i.e., continuous welding at >250℃ for more than 30s.

[0099] After weight allocation, weld quality is checked three times consecutively. If the weld leg size deviation is still >0.4mm, switch to reverse weight mode, i.e., high-strength steel switches to geometry-dominated mode and ordinary steel switches to heat-dominated mode. If the switch is still ineffective, a system-level process check is triggered.

[0100] It should be further explained that, in the specific implementation process, the system uses a multispectral vision inspection module mounted on the gantry to capture the working status of the ship's planar sections in real time. Laser scanning sensors acquire three-dimensional contour data of the workpiece surface, accurately identifying geometric deviations between the actual assembly position and the preset model, including assembly errors and assembly gaps. An infrared thermal imager simultaneously monitors the temperature distribution in the welding area, tracking the deformation trend caused by heat conduction. When a geometric deviation is detected, the system generates a geometric correction component; when the influence of thermal deformation is predicted, the system generates a thermal compensation component. The system automatically selects a correction strategy based on the workpiece material type: for high-strength steel, the system prioritizes the thermal compensation component as the dominant correction amount; for ordinary steel, the system uses the geometric correction component as the core correction basis. The two components are dynamically weighted and fused to form the final correction command, driving the welding robot to adjust the welding torch's spatial pose in real time.

[0101] The system divides the workpiece surface into uniform grid cells, with each grid labeled with weld feature type and assigned a time window. When the planned paths of multiple robots overlap spatiotemporally within a critical distance range, a conflict prediction mechanism is immediately activated: if a high-precision weld task such as fillet weld or vertical weld is detected, it is given path execution priority; if a normal flat fillet weld task is detected, it is required to perform an avoidance maneuver. Avoidance strategies include path detour or delay waiting, with the specific choice depending on the degree of impact of path deviation on welding quality. The central decision-maker dynamically coordinates gantry movement resources, automatically compensating waiting robots for subsequent idle time windows when high-priority tasks occupy critical areas. After completing every ten meters of weld length, the system counts the number of task interruptions and total time consumption, dynamically optimizing subsequent path allocation strategies.

[0102] During welding, the actual weld leg size is continuously monitored and compared with a preset threshold in real time. If a single exceedance occurs, the process parameters are automatically fine-tuned; if three consecutive fine-tunings fail, the system is deemed to have failed, welding is immediately interrupted, and an alarm is triggered. Before path execution, a virtual working scene is constructed based on real-time 3D point clouds, and a random assembly error dataset conforming to typical ship section working conditions is injected. Thousands of independent collision simulation tests are performed on each path, with each test randomly rotating the error loading direction and simulating thermal deformation drift. The collision frequency is statistically analyzed: if the frequency exceeds the safety threshold, the path is immediately abandoned and a historical safe alternative is invoked; if the frequency meets the standard, execution is approved. The quality monitoring unit and the collision pre-simulation unit are interconnected; quality inspection is suspended during path replanning, and enhanced monitoring mode is activated after the new path is executed.

[0103] When identifying the endpoints of fillet welds, the system automatically detects the spacing between adjacent components. When the spacing is less than the safe operating distance, a ring-shaped warning zone is generated and a high-interference warning is activated. For vertical welding turning points with a height exceeding 800 mm, the robot's passing speed is forcibly reduced to ensure accuracy. When a height difference is detected at the workpiece splice seam, the continuity of the weld is automatically determined: a height difference within 3 mm is treated as a single weld, while a difference exceeding 3 mm is divided into independent weld segments. Process parameters are dynamically matched to the weld type: for fillet welds, the shielding gas flow rate is increased and the welding torch tilt angle is locked; for vertical welds, a layered current reduction strategy is adopted; for zigzag fillet welds, current pulse compensation is triggered at the turning point. When the system receives a path correction command, the welding parameters are adjusted synchronously: for every 1 mm increase in path offset, the welding current increases or decreases proportionally; when the welding torch angle changes by more than 5 degrees, the gas flow rate is corrected immediately.

[0104] The system identifies steel grades through spectral analysis. In high-strength steel scenarios, the heat deformation compensation component dominates, its weight is further increased when welding heat accumulation exceeds a critical value, and geometric correction is only activated when a sudden large-scale assembly deviation is detected. In ordinary steel scenarios, the geometric correction component serves as the core compensation basis, and heat deformation compensation is introduced locally when assembly gaps or primer thickness exceed limits. Welding quality is verified three times consecutively after weight allocation: if the deviation continues to exceed limits, the system switches to reverse weight mode; if still ineffective, a full-system process check is initiated. This fault-tolerant design ensures stability under extreme operating conditions.

[0105] By combining steel grade and operating conditions, the system autonomously selects either geometric correction or thermal deformation compensation as the primary strategy, overcoming the limitations of a single model in handling complex conditions. For high-strength steel, the core objective is to suppress hot cracking, while for ordinary steel, the primary task is to overcome assembly errors. A tiered avoidance mechanism is implemented based on the importance of weld processes, with high-precision welds receiving path priority. Through time window compensation and task boundary reorganization, production line congestion caused by multiple machines waiting is eliminated.

[0106] The system accurately replicates on-site shipboard conditions through thousands of random error simulations, implementing a zero-collision standard in high-risk areas. A three-tiered quality control system, consisting of single fine-tuning, three verifications, and system interruption checks, eliminates the risk of batch rework. Synchronized issuance of parameter adjustment commands and welding torch pose change commands ensures process stability during dynamic correction. An offset current compensation algorithm keeps weld leg size fluctuations within limits.

[0107] It should be further explained that the vision-based detection-based multi-robot path planning method for ship planar segment welding includes the following steps:

[0108] Step S1: The laser scanning sensor mounted on the gantry captures the three-dimensional point cloud of the workpiece surface to accurately identify assembly errors and assembly gaps; simultaneously, the infrared thermal imager is activated to monitor the temperature gradient distribution in the welding area and record the characteristics of thermal deformation trends.

[0109] Step S2: Generate geometric correction components based on laser point cloud data; predict deformation based on thermal imaging data and output thermal compensation components; if the workpiece material is high-strength steel, prioritize the thermal compensation component as the dominant correction strategy; if it is ordinary steel, use geometric correction components as the core correction basis; fuse to generate a composite correction vector and send it to the robot controller.

[0110] Step S3: Discretize the workpiece surface into a fixed-size grid, and mark the characteristic positions such as the end point of the fillet weld and the turning point of the vertical weld; calculate the time window occupied by each grid; when multiple robot paths overlap within the critical distance, give priority to the high-precision weld task, and force the ordinary weld task to perform detour or delayed avoidance; the central decision-maker dynamically compensates the subsequent time window of the affected robot.

[0111] Step S4: Each robot generates motion strategies as an independent intelligent agent: high-precision welds prioritize quality, while ordinary welds prioritize efficiency; the central decision-maker coordinates the gantry movement resources to resolve travel conflicts; and task boundaries are redistributed after path changes.

[0112] Step S5: Continuously monitor the deviation between the actual weld leg size and the preset threshold; automatically fine-tune the process parameters if the deviation exceeds the limit once; if the fine-tuning fails three times in a row, interrupt the welding and trigger a system alarm.

[0113] Step S6: Construct a virtual work scene based on real-time point cloud; inject normally distributed random assembly errors that conform to ship working conditions; perform thousands of independent collision simulations on a single path: if the number of collisions exceeds the safety threshold, call the historical safe path to replace it; implement the zero-collision mandatory standard in high-risk weld areas.

[0114] Step S7: Identify the current weld type and match parameters: increase gas flow and lock the welding torch tilt angle for fillet welds, and enable the layered current reduction strategy for vertical welds; adjust parameters synchronously when receiving path correction instructions: compensate current for path offset according to linear relationship, and correct gas flow in real time when welding torch angle changes abruptly.

[0115] Step S8: Identify the steel grade through spectral features; for high-strength steel, thermal deformation compensation is the main focus, and geometric correction is only enabled for sudden large-size assembly deviations; for ordinary steel, geometric correction is the core, and thermal compensation is introduced locally when the primer exceeds the standard; after weight allocation, the welding quality is verified three times consecutively.

[0116] Step S9: If the quality deviation continues to exceed the standard after weight allocation, switch to reverse weight mode; if it is still ineffective, freeze the system and start full process verification; reset the compensation parameters after the verification is passed.

[0117] Step S10: Count the number of task interruptions and total time spent after each 10-meter weld is completed; dynamically optimize conflict prediction rules and resource allocation strategies based on statistical data; update the historical safe path library.

[0118] By employing a multispectral vision real-time perception and material-adaptive dynamic correction mechanism, the problem of inaccurate path planning caused by assembly errors, thermal deformation, and complex working conditions in the planar segment welding of large ships is fundamentally solved. The use of composite correction vectors and spatiotemporal grid collaborative prediction technology effectively improves weld trajectory accuracy, ensuring that weld leg dimensions strictly meet standards in challenging processes such as fillet welding and vertical welding of the largest workpieces, eliminating defects such as porosity and cracks, and significantly reducing rework rates and quality risks.

[0119] A distributed collaborative decision-making framework and a collision probability verification mechanism of thousands enable efficient collaborative operation of multiple robots in limited space. Dynamic priority arbitration and real-time synchronization of path parameters effectively avoid mechanical interference accidents, optimize gantry movement resource allocation, shorten welding cycles, and improve overall production line throughput efficiency. The dual safety closed-loop design provides reliable assurance for automated welding of ship sections, driving the upgrading of large structural component manufacturing towards intelligent manufacturing.

[0120] 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 said element.

[0121] 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 path planning system for ship planar segment welding based on vision inspection, characterized in that, It includes a gantry crane, a welding robot, and a control cabinet, characterized in that: The multispectral vision inspection module, mounted on the gantry, includes a laser scanning sensor and an infrared thermal imager, used to acquire three-dimensional point cloud and temperature distribution data of the ship's planar sections in real time. The dynamic path planning module communicates with the control cabinet, receives visual inspection data, and outputs corrected paths. The multi-robot collaborative control module is electrically connected to the dynamic path planning module to dynamically allocate motion tasks to each robot; The dual closed-loop verification module interacts with the dynamic path planning module and the welding robot signal respectively to perform path pre-verification and welding quality monitoring. The dynamic path planning module includes a deformation compensation unit and a spatiotemporal grid generation unit. The deformation compensation unit calculates the assembly error compensation amount based on the three-dimensional point cloud data, predicts the thermal deformation compensation amount by combining the temperature distribution data, and generates a composite correction vector. The spatiotemporal grid generation unit discretizes the ship's planar segments according to a preset grid size, marks the weld type feature points, and allocates time windows. The deformation compensation unit performs the following operations: Based on the deviation between the actual contour of the workpiece and the preset model identified by laser point cloud, geometric correction components are output. Based on infrared thermal imaging data, the deformation trend is calculated through a heat conduction model, and the thermal compensation component is output. Based on the ship steel grade, a dynamic weighting coefficient is matched, and a composite correction vector is generated by integrating geometric correction components and thermal compensation components. The spatiotemporal grid generation unit is equipped with a conflict prediction subunit. When spatiotemporal grid overlap is detected within a critical distance Δ on the robot path, a dynamic avoidance mechanism is activated. If Δ is less than the minimum allowable distance for fillet welding, force the path priority to be adjusted. If Δ is greater than the safe operating distance, maintain the original path planning; The dual closed-loop verification module includes: The dynamic error envelope analysis unit compares the actual value of the solder leg size with the preset threshold in real time and triggers correction or interruption commands. The virtual twin collision simulation unit constructs a virtual scene based on real-time 3D point clouds and injects random assembly errors to simulate collision risks.

2. The vision-based multi-robot path planning system for ship planar segment welding according to claim 1, characterized in that: The multi-robot collaborative control module deploys a distributed reinforcement learning algorithm framework, wherein: Each welding robot acts as an independent intelligent agent, generating motion strategies based on welding quality parameters and motion efficiency parameters. The central decision-maker coordinates the action strategies of each agent based on the spatiotemporal grid model to resolve resource conflicts in the Y-axis movement of the gantry. Each welding robot acts as an independent intelligent agent, collecting its own welding quality data and motion status data in real time. The welding quality data includes weld leg size deviation, and the motion status data includes acceleration and joint angles. If high-precision welds such as fillet welds or vertical welds are being performed, weld leg size accuracy is the primary optimization goal. If ordinary fillet welds are being performed, motion efficiency is prioritized. Subsequently, a motion strategy is generated based on the optimization goals: when a path conflict risk is detected, the impact of the detour path on the welding quality is evaluated; when there is no conflict, the shortest time path is selected. Receive spatiotemporal grid occupancy requests submitted by all agents and identify the resource contention area of ​​the gantry Y-axis; when multiple agents request the same Y-axis travel segment: the high-precision weld seam task agent automatically obtains priority use rights; the ordinary weld seam task agent accepts delay instructions or path segmentation schemes; generate resource allocation instructions and send them to each robot controller; For agents that wait due to avoidance, their subsequent grid time windows are automatically compensated; if the detour path affects the welding quality, such as vertical welding being interrupted, the central decision-maker reallocates the task boundaries of adjacent robots. For every 10 meters of weld completed, the number of task interruptions and the total welding time are recorded: if the number of interruptions for high-precision welds exceeds the limit, their subsequent priority weight is increased; if ordinary welds time out, the path segmentation algorithm is optimized.

3. The vision-based multi-robot path planning system for ship planar segment welding according to claim 1, characterized in that: The virtual twin collision pre-simulation unit performs Monte Carlo simulation, the process of which includes: Generate a random assembly error dataset that follows a normal distribution; Perform multiple rounds of collision detection on each planned path; When the collision probability exceeds the set threshold, a replanning instruction is sent to the dynamic path planning module.

4. The vision-based multi-robot path planning system for ship planar segment welding according to claim 1, characterized in that: The weld type feature points include the end point of the flat fillet weld, the turning point of the vertical fillet weld, and the positioning point of the wrap fillet weld; the time window is calculated based on the robot's maximum motion acceleration and joint rotation angle.

5. The vision-based multi-robot path planning system for ship planar segment welding according to claim 1, characterized in that: It also includes a process parameter adaptive unit that synchronizes data with the dynamic path planning module and matches welding current, wire feed speed and gas flow parameters according to the weld type.

6. The vision-based multi-robot path planning system for ship planar segment welding according to claim 1, characterized in that: The matching logic of the dynamic weighting coefficient is as follows: for AH32 / DH36 high-strength steel, increase the weight of the thermal compensation component; for ordinary marine grade A steel, increase the weight of the geometric correction component.

Images