Ship small assembly multi-cantilever robot collaborative welding system

Abstract

The application discloses a ship small group assembly multi-cantilever robot cooperative welding system and relates to the technical field of ship intelligent manufacturing. The application discloses a ship small group assembly multi-cantilever robot cooperative welding system and relates to the technical field of ship intelligent manufacturing. The application discloses a ship small group assembly multi-cantilever robot cooperative welding system and relates to the technical field of ship intelligent manufacturing. The application discloses a ship small group assembly multi-cantilever robot cooperative welding system and relates to the technical field of ship intelligent manufacturing. The application discloses a ship small group assembly multi-cantilever robot cooperative welding system and relates to the technical field of ship intelligent manufacturing. The application discloses a ship small group assembly multi-cantilever robot cooperative welding system and relates to the technical field of ship intelligent manufacturing. The application discloses a ship small group assembly multi-cantilever robot cooperative welding system and relates to the technical field of ship intelligent manufacturing. The application discloses a ship small group assembly multi-cantilever robot cooperative welding system and relates to the technical field of ship intelligent manufacturing. The application discloses a ship small group assembly multi-cantilever robot cooperative welding system and relates to the technical field of ship intelligent manufacturing. The application discloses a ship small group assembly multi-cantilever robot cooperative welding system and relates to the technical field of ship intelligent manufacturing. The application discloses a ship small group assembly multi-cantilever robot cooperative welding system

Description

Technical Field

[0001] This invention relates to the field of intelligent ship manufacturing technology, specifically to a collaborative welding system for multi-cantilever robots used in ship assembly. Background Technology

[0002] In the shipbuilding industry, with the increasing size and complexity of ships, the demand for automated welding of sub-assembly components is becoming more and more urgent. Sub-assembly is a typical hull structure composed of bottom plates, stiffening plates, elbow plates, etc. The welding quality and efficiency of sub-assembly directly affect the overall construction cycle. In order to cope with the characteristics of large workpiece size, complex weld distribution and diverse node forms, multi-cantilever robot collaborative welding systems are often used. This system is usually composed of multiple suspended robots, which are laid out along the conveyor roller conveyor and can cover a welding area of ​​tens of meters.

[0003] For example, a multi-robot collaborative welding system for ship section assembly, as disclosed in Chinese Patent Publication No. CN121061849A, uses high-precision visual inspection scanning and SLAM matching algorithms to construct a workpiece point cloud model in real time and accurately align it with the design drawings. This system can accurately identify the weld position and workpiece geometric deviations. Combined with robot kinematic constraints and safety boundaries, it generates a collision-free welding path, ensuring that the welding torch end pose fits the workpiece surface, reducing welding deviations, and ensuring process stability.

[0004] In existing technologies, due to the narrow and overlapping workspaces of cantilever robots, multi-robot collaborative operations face spatial deadlock problems. Although the system can visually perceive the workpiece position, the dynamic changes in the cantilever motion trajectories of each robot mean that interference areas cannot be effectively avoided through static presets. The necessary path of one robot is often unexpectedly blocked by the working posture of another robot, and the deadlock state is often only identified after it occurs, lacking the ability to predict potential interference. This easily leads to frequent emergency stops and manual intervention, further exacerbating the thermal waiting problem at the temporal level. The pre-set welding sequence in the data has rigid constraints in the small assembly line. When multiple robots work together, if one robot is forced to stop due to spatial deadlock, the subsequent weld to be executed may be the prerequisite for another robot to start the next task, forming a chain of waiting. This causes the welding torch to be in a state of idle burning or standby for a long time, which not only increases the energy consumption of a single machine, but also makes the actual efficiency of multi-machine collaboration far lower than the theoretical cycle time. The whole is caught in a collaborative dilemma of spatial conflict and time idleness. To address this problem, a multi-cantilever robot collaborative welding system for ship small assembly is proposed. Summary of the Invention

[0005] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: a collaborative welding system for ship assembly using multi-cantilever robots, comprising a collaborative control center, wherein the collaborative control center is communicatively connected to the following modules:

[0006] The twin field construction module, based on four cantilever robots deployed on a gantry-type assembly line, constructs a digital twin dynamic interference field that integrates the workpiece point cloud model and the real-time pose of the cantilever through industrial vision perception and digital modeling. This provides a high-fidelity real-time virtual mapping space for multi-machine collaborative operations and lays the foundation for accurate interference judgment.

[0007] The motion modality prediction module, based on historical trajectories and real-time motion features, uses a machine learning classifier to make short-term predictions of typical motion patterns of the cantilever robot, identify potential spatial path conflicts in advance, and update the dynamic reachability constraint graph, upgrading from passive response to active prediction, and locking potential interference areas in advance to avoid sudden deadlock.

[0008] The weld seam task decoupling and reconfiguration engine is used to decompose all weld seams across workpieces into independent task units according to spatial location, welding duration, and process requirements. This breaks the welding sequence constraints within the workpiece, establishes a dynamically reconfigurable task pool, releases the space for parallel task scheduling, and eliminates the chain waiting deadlock caused by rigid sequence.

[0009] The collaborative task scheduling module constructs a scheduling agent based on a deep Q-network (DQN). It uses the minimization of the total global path length and the minimization of the welding torch thermal waiting time as the joint reward function. It conducts millions of self-game trainings in a twin environment to output a conflict-free and low-waiting-time globally optimal collaborative welding timing scheme. It learns the optimal collaborative strategy in the virtual space to ensure conflict-free and efficient operation in actual operation.

[0010] The instruction parsing and execution module is used to convert the globally optimal collaborative welding timing scheme into low-latency, highly reliable robot control instructions. Through the 5G-based industrial private network communication channel, each cantilever controller receives the dynamic path instructions and maps them into robot motion control signals in real time, ensuring accurate, synchronous, and real-time response of multi-machine collaborative actions.

[0011] Preferably, the twin field construction module includes a multi-view vision fusion perception unit and a dynamic interference field modeling unit;

[0012] The multi-view vision fusion perception unit is used to deploy a global binocular vision system at the top of the workstation to collect the three-dimensional pose and motion trajectory of four cantilever robots in real time. Combined with the point cloud model obtained by the 3D structured light camera at the end of the cantilever, it performs unified spatial registration of the workpiece and the robot, outputs multi-source fusion perception data, realizes a high-precision unified spatiotemporal reference for robot pose and workpiece geometric features, and ensures consistent and reliable perception data.

[0013] The dynamic interference field modeling unit, based on the spatially registered data, integrates the workpiece point cloud model and the real-time pose of the cantilever to construct a digital twin dynamic interference field, including the robot's current posture, motion trend and workpiece boundary, and generates a dynamic reachability domain constraint graph, transforming abstract space occupancy into a visual constraint graph, providing real-time spatial boundaries for subsequent path planning.

[0014] Preferably, the multi-view vision fusion perception unit performs the following steps:

[0015] Using a global binocular vision system deployed at the top of the welding station, stereo image pairs of the cantilever ends and the transverse mechanism of four cantilever robots are acquired synchronously at a fixed frequency. The three-dimensional pose parameters of each robot in the global coordinate system, including spatial position coordinates and Euler angles, are calculated by stereo matching algorithm, so as to achieve millisecond-level accurate positioning of the motion state of multiple robots and provide a high-precision pose reference for the construction of dynamic interference field.

[0016] Trigger the 3D structured light camera mounted on each cantilever end to project and image the workpiece surface with structured light, obtain high-density point cloud data, and transform the local point cloud model to the global coordinate system through the point cloud registration algorithm to form a workpiece point cloud model containing the features of stiffeners, welds, and base plates, fully restore the three-dimensional geometric features of the workpiece, and ensure the authenticity and reliability of weld recognition and path planning.

[0017] The real-time pose of the cantilever calculated by the global binocular vision system is spatiotemporally aligned with the point cloud model of the workpiece acquired by the 3D structured light camera. This establishes a spatial mapping relationship between the robot's motion trajectory and the geometric features of the workpiece, outputting multi-source fused perception data under a unified spatiotemporal reference. This achieves the unification of the spatial coordinates of the robot and the workpiece, laying the foundation for data fusion for subsequent collaborative control.

[0018] Preferably, the dynamic interference field modeling unit performs the following steps:

[0019] Based on the multi-source fusion perception data output by the multi-view vision fusion perception unit, the current posture envelope of the four cantilever robots and the boundary contour of the workpiece point cloud model are extracted. The welding area is segmented into voxels using a spatial octree data structure to construct an initial static interference field and realize the structured modeling of the spatial occupancy relationship of the welding area.

[0020] By integrating the kinematic model and real-time motion parameters of the cantilever robot, the motion trend of each robot in the next control cycle is predicted. The predicted trajectory is mapped to the static interference field, and the occupancy state and occupancy probability of each voxel unit are dynamically updated to generate a digital twin dynamic interference field that evolves over time, thereby realizing the real-time dynamic update of the interference field state with the robot's motion trend.

[0021] The collision risk level of each voxel unit in the dynamic interference field is calibrated. Based on the robot's motion speed, approach direction and envelope size, the potential collision probability is calculated, and a dynamic reachability domain constraint graph containing reachable areas, prohibited areas and warning areas is generated, so as to realize the fine spatial division of the robot's motion safety boundary.

[0022] Preferably, the motion modality prediction module includes a feature extraction and classification unit and a conflict prediction and update unit;

[0023] The feature extraction and classification unit is used to combine historical trajectories and real-time motion features, and introduce a motion pattern classifier based on support vector machine (SVM) to perform real-time classification and short-term prediction of the trajectory features of four cantilever robots (Gantry 1 to Gantry 4), accurately identify typical motion patterns including cantilever crossing, dwelling extension, retreating and avoiding, and straight-line movement, accurately characterize the robot's motion intention, and provide high-confidence behavior labels for conflict prediction.

[0024] The conflict prediction and update unit is used to combine the classification results with the current pose to predict the motion path within a short time window in the future, identify potential spatial path conflicts, update the dynamic reachability constraint graph, and simultaneously mark potential conflict areas, transforming the fuzzy motion trend into a quantified conflict area, thereby achieving a refined early warning of space occupancy.

[0025] Preferably, the feature extraction and classification unit performs the following steps:

[0026] Historical motion trajectory data of four cantilever robots were collected, including joint angles, external axis positions, motion speed and acceleration. A time-series motion feature vector set was constructed, and the trajectory data was segmented by sliding windows. Statistical features and frequency domain features within the windows were extracted to provide a comprehensive and high-precision data foundation for motion pattern recognition and ensure the reliability of classification results.

[0027] Based on the Support Vector Machine (SVM) classification algorithm, the classifier is trained offline using labeled historical motion pattern samples to establish a mapping relationship between trajectory feature vectors and motion pattern categories. The motion pattern categories include cantilever crossing, dwelling extension, retreating and avoidance, and straight-line movement. A highly robust classification model is constructed to achieve accurate discrimination and generalization of complex motion behaviors.

[0028] The system receives cantilever pose data from the multi-view vision fusion perception unit in real time, extracts motion feature vectors within the current time window, inputs them into a trained SVM classifier for online classification, and outputs short-term motion pattern labels for each robot at the current moment. This enables real-time perception of the robot's motion intentions and provides crucial decision-making basis for conflict prediction.

[0029] Preferably, the conflict prediction and update unit performs the following steps:

[0030] Based on the motion pattern labels output by the feature extraction classification unit, combined with the kinematic models and dynamic reachability constraint graphs of each robot, the motion path and space occupancy sequence of each robot in the next 3-5 second time window are predicted, effectively predicting the robot motion trend and providing a spatiotemporal data basis for conflict detection.

[0031] The predicted multi-robot motion paths are overlaid and analyzed in the spatiotemporal dimensions. A spatial conflict detection algorithm is used to identify path intersections and overlapping areas of space occupation. The time, location and severity of conflict occurrence are calculated, potential spatial path conflict events are marked, and potential interference locations and times are accurately identified to avoid sudden path conflicts that cause downtime.

[0032] Based on the conflict detection results, the conflict area markers in the dynamic reachability constraint graph are updated, the occupancy probability of the conflict area is increased, and the reachability is marked as a warning state. At the same time, the updated dynamic reachability constraint graph is fed back to the subsequent process in real time, dynamically updating the spatial constraints and guiding the avoidance of risk areas in advance.

[0033] Preferably, the weld seam task decoupling and refactoring engine performs the following steps:

[0034] The workpiece production data output by the upstream CAD / CAM software is analyzed, and the spatial coordinates, welding sequence number, weld length, weld leg height and node type information of all welds are extracted. An initial weld task list is constructed and associated with the corresponding workpiece to ensure that the weld data is consistent with the physical characteristics, and to provide a complete input basis for subsequent accurate scheduling.

[0035] Based on the updated dynamic reachability constraint graph and the output motion mode labels, the long welds and complex node welds in the weld task list are spatially segmented, and a single weld is decomposed into multiple independent welding segments. This breaks the original welding sequence constraints within the workpiece, releases the parallel scheduling space, and fundamentally eliminates the waiting deadlock caused by rigid sequence.

[0036] The segmented welding sections are matched with the welding capabilities of four cantilever robots. Based on the reachability of each robot, welding process parameters, and current task load, a dynamically reconfigurable task pool is constructed to achieve optimal task and resource matching, significantly improving the overall efficiency of multi-robot collaborative operation.

[0037] Preferably, the collaborative task scheduling module performs the following steps:

[0038] The scheduling agent is constructed based on deep Q-network (DQN). The task pool state output by the decoupled reconstruction engine for welding tasks and the updated dynamic reachability constraint graph are used as the input state space, and the task allocation and path selection of each robot are used as the action space, which effectively improves the decision-making accuracy and space utilization of multi-robot collaborative operation.

[0039] A joint reward function is designed with the optimization objectives of minimizing the total global path length and minimizing the welding torch thermal waiting time. A collision penalty term and a task completion reward term are introduced. The function is trained in a digital twin environment with millions of self-games to optimize the decision-making strategy of the scheduling agent, which significantly shortens the overall welding path and reduces the welding torch dry-burning waiting time.

[0040] The trained scheduling agent is deployed to the real-time operating environment. Based on the current task pool status and dynamic interference field constraints, it outputs a conflict-free and low-wait-time globally optimal collaborative welding timing scheme online. This scheme includes the task execution order, welding path sequence, and avoidance strategy of each robot, ensuring the continuity and safety of multi-robot collaborative operation in real time and avoiding spatial conflicts and task stagnation.

[0041] Preferably, the instruction parsing and execution module performs the following steps:

[0042] The system receives the globally optimal collaborative welding timing scheme and parses the task execution order, welding path sequence, and avoidance strategy in the scheme into a sequence of motion control commands for each robot, including external axis movement commands, robot joint movement commands, and welding process parameter commands. This ensures that the commands accurately match the robot's kinematic characteristics and process requirements, guaranteeing the accuracy and consistency of the welding actions.

[0043] Through the established 5G industrial private network communication channel, the parsed motion control command sequence is sent to the controllers of four cantilever robots in a low-latency and high-reliability manner, and timestamps and synchronization signals are sent synchronously to ensure the timing coordination of the actions of multiple robots and to ensure the timing consistency and action continuity of collaborative operations.

[0044] After receiving instructions, each cantilever robot controller maps them in real time to robot servo drive signals and welding power control signals, driving the cantilever to perform predetermined movements and welding actions. At the same time, it feeds back the robot status, weld completion status and abnormal alarm information to the collaborative control center through the 5G private network, so as to perceive the execution status in real time and dynamically optimize subsequent tasks, forming a closed-loop intelligent control to improve collaborative efficiency.

[0045] This invention provides a collaborative welding system for multi-cantilever robots used in ship assembly. It offers the following advantages:

[0046] (i) This ship group erection multi-cantilever robot collaborative welding system constructs a digital twin dynamic interference field by deploying a global binocular vision unit at the top of the workstation and a 3D structured light camera mounted at the end of the cantilever. It can perceive the spatial occupancy relationship between the robot body, the traversing mechanism and the workpiece in real time, perform voxel segmentation and dynamic updates on the welding area, predict the movement trend in the short time window in the future, identify potential spatial path conflicts in advance, generate a dynamic reachability domain constraint graph, effectively avoid interference risks at the spatial level, and ensure the continuity and safety of multi-machine collaborative operation.

[0047] (II) This collaborative welding system for shipboard multi-cantilever robots introduces a motion pattern classifier based on support vector machines to collect and analyze the historical trajectories and real-time motion characteristics of four cantilever robots. It accurately identifies typical motion patterns such as cantilever crossing, dwelling and extending, retreating and avoiding, and straight-line movement. Through short-term prediction of motion patterns, combined with kinematic models and dynamic reachability constraint graphs, it can predict the motion path and space occupation sequence of each robot within a time window of a few seconds in the future. Potential conflict events are output in the form of timestamps, position coordinates and risk levels, providing high-precision input conditions for task scheduling and avoidance decisions, and significantly improving the system's adaptability to dynamic working conditions.

[0048] (III) This ship group standing multi-cantilever robot collaborative welding system analyzes and verifies the workpiece production data output by upstream CAD / CAM software, and intelligently divides long welds and complex node welds into multiple independent welding segments. The segmentation process is based on dynamic reachability domain constraint graphs and motion mode labels, dynamically assesses the spatial position and interference field occupancy probability of each welding segment, and ensures that the segmented welding segments have independent execution conditions. All welding segments are stored in the dynamic task pool as independent task units, completely breaking the original rigid welding sequence constraints in the workpiece, providing highly flexible task input for multi-machine parallel scheduling, and significantly improving the overall welding efficiency of the production line.

[0049] (iv) This ship group multi-cantilever robot collaborative welding system is based on a deep Q-network to construct a scheduling agent. The dynamic task pool state and dynamic reachability domain constraint graph are used as the input state space, and task allocation, path selection and avoidance strategies are used as the action space. The system is trained in a digital twin environment with millions of self-games. The agent takes minimizing the total global path length and minimizing the welding torch thermal waiting time as the joint optimization objectives. It introduces collision penalty terms and task completion reward terms to output a conflict-free and low-waiting collaborative welding sequence scheme. It fully integrates robot kinematic constraints and welding process requirements to ensure that multiple robots can work collaboratively in a high-density work space according to a unified sequence, avoiding spatial deadlock and thermal waiting. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the workflow of a ship crew multi-cantilever robot collaborative welding system according to the present invention;

[0051] Figure 2 This is a schematic diagram of the multi-machine collaborative welding task allocation logic of the present invention;

[0052] Figure 3 This is a schematic diagram of the gantry layout of the framework group for this invention;

[0053] Figure 4 This is a schematic diagram of the automatic operation process of the gantry 1 of the present invention;

[0054] Figure 5 This is a schematic diagram of the automatic operation process of the gantry 2 of the present invention;

[0055] Figure 6 This is a schematic diagram of the automatic operation process of the gantry 3 of the present invention;

[0056] Figure 7 This is a schematic diagram of the automatic operation process of the gantry 4 of the present invention. Detailed Implementation

[0057] 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.

[0058] Example 1, please refer to Figure 1 , Figure 2 This invention provides a technical solution: a collaborative welding system for ship crew assembly using multi-cantilever robots, comprising a collaborative control center, which is communicatively connected to the following modules:

[0059] The twin field construction module is based on four cantilever robots deployed on a gantry-type assembly line. Through industrial vision perception and digital modeling, it constructs a digital twin dynamic interference field that integrates the workpiece point cloud model and the real-time pose of the cantilever. This provides a high-fidelity real-time virtual mapping space for multi-machine collaborative operation and lays the foundation for accurate interference judgment. The twin field construction module includes a multi-view vision fusion perception unit and a dynamic interference field modeling unit.

[0060] The multi-view vision fusion perception unit, deployed on top of the workstation, is used in a global binocular vision system to acquire the 3D pose and motion trajectory of four cantilever robots in real time. Combined with point cloud models acquired by 3D structured light cameras at the cantilever ends, it performs unified spatial registration of the workpiece and the robot, outputting multi-source fusion perception data. This achieves a high-precision unified spatiotemporal reference for robot pose and workpiece geometric features, ensuring consistent and reliable perception data. Utilizing the global binocular vision system deployed on top of the welding station, it synchronously acquires stereo image pairs of the cantilever ends and the traverse mechanism of the four cantilever robots at a fixed frequency. Through a stereo matching algorithm, it calculates the 3D pose parameters of each robot in the global coordinate system, including spatial position coordinates and Euler angles, achieving millisecond-level precise positioning of the multi-robot motion state. To provide a high-precision pose reference for the construction of dynamic interference fields, the 3D structured light cameras mounted on the ends of each cantilever are triggered to project and image the workpiece surface with structured light, acquiring high-density point cloud data. The local point cloud model is transformed into the global coordinate system through the point cloud registration algorithm, forming a workpiece point cloud model containing the features of stiffeners, welds, and base plates, fully restoring the three-dimensional geometric features of the workpiece, ensuring the authenticity and reliability of weld recognition and path planning. The real-time pose of the cantilever calculated by the global binocular vision system is spatiotemporally aligned with the workpiece point cloud model acquired by the 3D structured light camera, establishing a spatial mapping relationship between the robot's motion trajectory and the workpiece's geometric features, and outputting multi-source fused perception data under a unified spatiotemporal reference, realizing the unification of the spatial coordinates of the robot and the workpiece, laying the data fusion foundation for subsequent collaborative control;

[0061] It should be noted that the global binocular vision system deployed at the top of the welding station uses two high-resolution industrial cameras to form a stereo vision pair with a baseline length of 1200mm. It synchronously acquires stereo image pairs of the cantilever ends and traverse mechanisms of four cantilever robots at a fixed frequency of 10Hz. The resolution of the stereo image pairs is 2448×2048 pixels. After epipolar correction, a semi-global stereo matching algorithm is used to calculate the disparity map. Combined with camera calibration parameters, the three-dimensional position coordinates of each robot in the global coordinate system are calculated. Simultaneously, the spatial geometric relationships of the feature points at the cantilever ends are analyzed. The Euler angles around the X, Y, and Z axes are calculated, with position accuracy controlled within ±1.5mm and attitude angle accuracy controlled within ±0.3°. Simultaneously with the global binocular vision system completing pose calculation, the 3D structured light cameras mounted on the ends of each cantilever are triggered to scan the workpiece surface. These structured light cameras use an 850nm infrared laser as the projection light source, projecting a coded grating pattern. The working distance is 800mm to 1500mm, with a single scan range of 600mm × 500mm and a point cloud density of 5 points per square millimeter. The cantilever moves along a predetermined scanning path, allowing the structured light camera to cover the entire weld area of ​​the workpiece. After acquiring high-density point cloud data, an iterative nearest-point algorithm is used to transform the local point cloud model to the global coordinate system. This model is then registered with the global coordinate reference established by the binocular vision system, ultimately forming a complete workpiece point cloud model containing the stiffener outline, weld bevel, and base plate planar features. The point cloud registration error is controlled within ±0.5mm. The real-time pose data of the four cantilever robots calculated by the global binocular vision system is then spatiotemporally aligned with the workpiece point cloud model acquired by the 3D structured light camera. Inter-synchronous synchronization adopts the IEEE 1588 precise time protocol to ensure that the sampling time deviation between the binocular vision system and each structured light camera is less than 1ms. Spatial alignment is based on the global coordinate system established by the binocular vision system. The real-time pose parameters of each cantilever end are jointly calibrated with the feature points in the workpiece point cloud model to establish a precise spatial mapping relationship between the robot's motion trajectory and the workpiece's geometric features. The final output multi-source fusion perception data includes the real-time pose sequence of each cantilever, the workpiece point cloud model and its corresponding weld feature label, with a data update frequency of 10Hz.

[0062] The dynamic interference field modeling unit, based on spatially registered data, fuses the workpiece point cloud model with the real-time pose of the cantilever to construct a digital twin dynamic interference field, including the robot's current posture, motion trend, and workpiece boundary, and generates a dynamic reachability domain constraint graph, transforming abstract spatial occupancy into a visual constraint graph, providing real-time spatial boundaries for subsequent path planning. Based on multi-source fusion perception data output by the multi-view vision fusion perception unit, the current posture envelopes of the four cantilever robots and the boundary contours of the workpiece point cloud model are extracted. A spatial octree data structure is used to perform voxel segmentation of the welding area, constructing an initial static interference field, realizing structured modeling of the spatial occupancy relationship of the welding area. By integrating the kinematic model and real-time motion parameters of the cantilever robot, the motion trend of each robot in the next control cycle is predicted. The predicted trajectory is mapped to the static interference field, and the occupancy status and occupancy probability of each voxel unit are dynamically updated. A digital twin dynamic interference field that evolves over time is generated, and the state of the interference field is updated in real time with the robot's motion trend. The collision risk level of each voxel unit in the dynamic interference field is calibrated. Based on the robot's motion speed, approach direction and envelope size, the potential collision probability is calculated, and a dynamic reachability domain constraint graph containing reachable areas, prohibited areas and warning areas is generated, realizing a refined spatial division of the robot's motion safety boundary.

[0063] The expression for calculating the potential collision probability is as follows:

[0064] ;

[0065] In the formula: This represents the potential collision probability of the current voxel unit, with a value ranging from 0 to 1. A larger value indicates a higher collision risk. It indicates the relative speed between two robots, reflecting how quickly the robots approach each other; This represents the maximum relative motion speed allowed for the robot, and its value is the combined upper limit of the maximum speed of the external axis and the combined speed of the joints; This represents the minimum distance between the current voxel unit and the adjacent robot envelope; This indicates the set safe distance threshold; when the actual distance is lower than this value, the risk of collision increases significantly. This represents the spatially overlapping projected area of ​​the current voxel unit and the robot envelope. This represents the maximum projected area of ​​the robot's envelope in the direction of motion, used to normalize the degree of overlap. This represents the weighting coefficients of the three influencing factors, satisfying... It is used to balance the contribution of speed, distance, and overlap area to the collision probability, and the specific value is determined according to the on-site working conditions.

[0066] It should be noted that, based on the multi-source fusion perception data output by the multi-view vision fusion perception unit, the current posture envelope of the four Kawasaki RA010 cantilever robots is first extracted. The envelope size of each robot body is determined based on the DH parameter model as a hexahedral envelope with a length of 1445mm, a width of 600mm, and a height of 1450mm. The envelope size of the transverse mechanism is 3200mm in length, 400mm in width, and 500mm in height. At the same time, the boundary contour of the workpiece point cloud model is extracted. The maximum workpiece size is 20000mm×6000mm×700mm. A spatial octree data structure is used to perform voxel segmentation of the welding area. The basic voxel unit size is set to 50mm×50mm×50mm, and the maximum octree size is... With a depth of 8 layers, the robot envelope and workpiece contour are mapped to an octree space, marking the initial occupancy state of each voxel unit and constructing an initial static interference field. This initial static interference field completely describes the spatial occupancy relationship between the robot body, the traversing mechanism, and the workpiece within the welding area. The kinematic model and real-time motion parameters of the cantilever robot are fused to predict the motion trend of each robot in the next control cycle. The kinematic model is based on the maximum motion speed of each joint: the rotational speeds of joints 1 to 6 are 145° / s, 120° / s, 150° / s, 230° / s, 230° / s, and 350° / s, respectively; the external axis X-direction travel speed is 0-10 m / min; and the Y-direction travel speed is 0-10 m / min. The real-time motion parameters include the current joint angles, external axis positions, instantaneous velocity, and acceleration. A forward kinematics algorithm is used to predict the robot's spatial trajectory point sequence within the next 200ms, with a time resolution of 20ms. The envelope sequence corresponding to the predicted trajectory points is mapped to the octree space of the static interferometric field. The occupancy status and probability of each voxel are dynamically updated. The occupancy probability is calculated based on the trajectory prediction confidence and arrival time weighted, with a confidence threshold set to 0.85 and an update frequency of 10Hz. A digital twin dynamic interferometric field that evolves over time is generated. The collision risk level of each voxel in the dynamic interferometric field is calibrated, categorized into three levels: prohibited areas, warning areas, and reachable areas. The calculation is based on a comprehensive judgment of robot movement speed, approach direction, and envelope size: when the robot movement speed is higher than 0.5m / s and the distance between it and the envelope of an adjacent robot is less than 100mm, it is marked as a prohibited area; when the distance is between 100mm and 300mm and the relative movement direction angle is less than 30°, it is marked as a warning area; the remaining voxel units are marked as reachable areas. The potential collision probability is calculated using a Gaussian mixture model, and the probability threshold is set to trigger a warning if it is above 0.3. The dynamic reachable domain constraint graph is output in real time using an octree data structure, which includes the collision risk level, occupancy probability, and timestamp information of each voxel unit. This serves as the input constraint condition for the collaborative task scheduling module to guide robot task allocation and path planning.

[0067] The motion modality prediction module, based on historical trajectories and real-time motion features, uses a machine learning classifier to make short-term predictions of typical motion patterns of the cantilever robot, identify potential spatial path conflicts in advance, and update the dynamic reachability constraint graph, upgrading from passive response to active prediction, locking potential interference areas in advance to avoid sudden deadlock. The motion modality prediction module includes a feature extraction and classification unit and a conflict prediction and update unit.

[0068] The feature extraction and classification unit combines historical trajectories with real-time motion features, introducing a motion pattern classifier based on Support Vector Machine (SVM) to perform real-time classification and short-term prediction of the trajectory features of four cantilever robots (Gantry 1 to Gantry 4). It accurately identifies typical motion patterns including cantilever crossing, stationary extension, retreating and avoidance, and straight-line movement, precisely characterizing the robot's motion intentions and providing high-confidence behavioral labels for conflict prediction. The unit collects historical motion trajectory data from the four cantilever robots, including joint angles, external axis positions, motion speeds, and accelerations, constructs a temporal motion feature vector set, and performs sliding window segmentation on the trajectory data, extracting statistical and frequency domain features within the window to provide comprehensive and high-precision motion pattern recognition. Based on a solid data foundation, ensuring reliable classification results, the Support Vector Machine (SVM) classification algorithm is used to train the classifier offline with labeled historical motion pattern samples. A mapping relationship between trajectory feature vectors and motion pattern categories is established. The motion pattern categories include cantilever crossing, dwelling extension, retreating and avoidance, and straight-line movement. A highly robust classification model is constructed to accurately identify and generalize complex motion behaviors. The cantilever pose data output by the multi-view vision fusion perception unit is received in real time, and the motion feature vectors within the current time window are extracted and input into the trained SVM classifier for online classification. The short-term motion pattern labels of each robot at the current moment are output, realizing real-time perception of the robot's motion intention and providing key decision-making basis for conflict prediction.

[0069] It should be noted that historical motion trajectory data of four Kawasaki RA010 cantilever robots were collected in real time. The sampling frequency was consistent with the output frequency of the multi-view vision fusion perception unit, set to 10Hz. The collected data dimensions included six joint angles (J1 to J6), external axis X and Y axis positions, instantaneous velocities and accelerations of each joint, and external axis motion velocities. The data accuracy met the measurement requirements of joint angle ±0.01°, position ±0.1mm, and velocity ±0.01° / s. A sliding window segmentation strategy was used for the collected raw time-series data. The window length was set to 2 seconds, the window sliding step size was 0.5 seconds, and each window contained 20 sampling points. Statistical data was extracted from each window. Features were collected, including the mean, variance, peak, and root mean square values ​​of each joint angle, as well as the rate of change of the external axis position. Simultaneously, a Fast Fourier Transform was used to extract frequency domain features, including the energy spectral density, dominant frequency amplitude, and phase information, constructing a 128-dimensional temporal motion feature vector set. An offline support vector machine classifier was trained based on labeled historical motion pattern samples. The sample library contains four typical motion modes: cantilever crossing, dwelling extension, retreating avoidance, and straight-line movement. Each mode has at least 2000 labeled samples. The cantilever crossing mode is defined as the robot crossing the adjacent robot's work area in the X direction, accompanied by coordinated adjustments from joints 4 to 6. The dwelling extension mode is characterized by the robot performing multiple movements at a fixed workstation. During weld seam welding, the external axis position remains unchanged while the joints are slightly adjusted. In the retreat avoidance mode, the robot detects an interference risk and returns to a safe position along its original path, accompanied by reverse velocity and joint reversal. In the straight-line movement mode, the robot moves at a constant speed along the X or Y direction under external axis drive, with relatively stable joint angles. A radial basis function kernel is used to map the feature vectors to a high-dimensional space. The penalty parameter C and kernel function parameters are optimized through grid search to train an SVM classification model, establishing a precise mapping relationship between trajectory feature vectors and the four motion modes. During real-time operation, pose data from the four cantilever robots is received at a 10Hz frequency, including current joint angles, external axis positions, and instantaneous velocities. Based on the acceleration parameters and a time window 2 seconds prior to the current moment, 128-dimensional statistical features and frequency domain features identical to those in the offline training phase are extracted to form the current motion feature vector. This vector is then input into the trained SVM classifier for online classification. The classifier's decision function is calculated based on support vectors and kernel functions, outputting the probability distribution of each robot's current moment belonging to cantilever crossing, dwelling extension, retreating avoidance, or straight-line movement. The mode category corresponding to the highest probability is taken as the current short-term motion mode label. The classification confidence threshold is set to 0.85. When it is below the threshold, it is marked as a mode ambiguity state and the system is triggered to enter a conservative avoidance strategy. The classification results are output to the conflict prediction update unit in real time at a frequency of 10Hz.

[0070] The conflict prediction and update unit combines classification results with the current pose to predict the motion path within a short time window, identify potential spatial path conflicts, update the dynamic reachability constraint map, and simultaneously mark potential conflict areas. This transforms ambiguous motion trends into quantifiable conflict areas, enabling refined early warning of space occupancy. Based on the motion pattern labels output by the feature extraction classification unit, combined with the kinematic models and dynamic reachability constraint maps of each robot, it predicts the motion path and space occupancy sequence of each robot within a 3-5 second time window, effectively predicting robot motion trends and providing a spatiotemporal data foundation for conflict detection. The predicted multi-robot motion paths are overlaid and analyzed in the spatiotemporal dimension. A spatial conflict detection algorithm is used to identify path intersections and overlapping areas of space occupancy, calculate the conflict occurrence time, conflict location, and conflict severity, mark potential spatial path conflict events, accurately identify potential interference locations and times, and avoid sudden path conflicts that lead to downtime. Based on the conflict detection results, the conflict area markings in the dynamic reachability constraint map are updated, the occupancy probability of conflict areas is increased, and the reachability is marked as a warning state. At the same time, the updated dynamic reachability constraint map is fed back to subsequent processes in real time, dynamically updating spatial constraints and guiding the early avoidance of risk areas.

[0071] It should be noted that after obtaining the motion pattern labels output by the feature extraction and classification unit, the motion path within a 3-5 second time window is predicted by combining the kinematic models and dynamic reachability constraint graphs of each Kawasaki RA010 robot. The kinematic model is based on the maximum motion speed of each joint of the robot, where the rotational speeds of joints 1 to 6 are 145° / s, 120° / s, 150° / s, 230° / s, 230° / s, and 350° / s, respectively. The walking speeds in the X and Y directions of the external axis are adjustable from 0 to 10 m / min. Real-time motion parameters include the current joint angles, external axis positions, instantaneous velocity, and acceleration, using forward motion. The algorithm generates a sequence of spatial trajectory points within the next 200ms to 5 seconds with a time resolution of 20ms. It maps the robot body envelope and the transverse mechanism envelope corresponding to the predicted trajectory points to an octree space, dynamically updating the occupancy status and probability of each voxel unit to form a complete spatial occupancy sequence. The predicted future motion paths of the four cantilever robots are overlaid and analyzed in the spatiotemporal dimensions. A spatial conflict detection algorithm identifies path intersections and overlapping areas of spatial occupancy. Conflict detection is based on the spatial occupancy status and probability of occupancy units in the octree voxel unit, calculating the minimum distance between adjacent robot envelopes. When the distance is less than 100mm and the robot's movement speed is high... At a speed of 0.5 m / s, a conflict is identified as a prohibited area conflict. When the distance between the points is between 100 mm and 300 mm and the relative motion direction angle is less than 30°, a conflict is identified as a warning area conflict. Further calculations are made of the conflict occurrence time, conflict location coordinates, and conflict severity. The severity is calculated based on a weighted average of occupancy probability and approach speed, with a probability threshold set to trigger a conflict marker if it is above 0.3. After detection, potential spatial path conflict events are output in the form of timestamps, robot IDs, conflict coordinates, and risk levels. Based on the conflict detection results, the dynamic reachability constraint map is updated in real time, increasing the occupancy probability of conflict areas and marking their reachability as a warning. The update operation is based on an octree data structure. It recalculates the collision risk level for conflicting voxel units and uses a Gaussian mixture model to update the occupancy probability value to ensure that the risk label of the warning area is consistent with the real-time working condition. The updated dynamic reachability constraint graph contains the collision risk level, occupancy probability and timestamp information of each voxel unit. It is fed back to the collaborative task scheduling module in real time at a frequency of 10Hz. This constraint graph serves as the input condition for task allocation and path planning, guiding the robot to avoid the warning area in subsequent control cycles and prioritize the selection of reachable areas to perform welding tasks. This avoids potential conflicts in advance at the spatial level and ensures the continuity and safety of multi-machine collaborative operation.

[0072] The weld seam task decoupling and reconfiguration engine is used to decompose all weld seams across workpieces into independent task units according to spatial location, welding duration, and process requirements. This breaks the welding sequence constraints within the workpiece, establishes a dynamically reconfigurable task pool, releases the space for parallel task scheduling, and eliminates the chain waiting deadlock caused by rigid sequence.

[0073] The collaborative task scheduling module constructs a scheduling agent based on a deep Q-network (DQN). It uses the minimization of the total global path length and the minimization of the welding torch thermal waiting time as the joint reward function. It conducts millions of self-game trainings in a twin environment to output a conflict-free and low-waiting-time globally optimal collaborative welding timing scheme. It learns the optimal collaborative strategy in the virtual space to ensure conflict-free and efficient operation in actual operation.

[0074] The instruction parsing and execution module is used to convert the globally optimal collaborative welding timing scheme into low-latency, highly reliable robot control instructions. Through the 5G-based industrial private network communication channel, each cantilever controller receives the dynamic path instructions and maps them into robot motion control signals in real time, ensuring accurate, synchronous, and real-time response of multi-machine collaborative actions.

[0075] Example 2, as Figure 1 , Figure 2 As shown, based on Embodiment 1, the present invention provides a technical solution: the weld seam task decoupling and reconstruction engine performs the following steps: parsing the workpiece production data output by the upstream CAD / CAM software, extracting the spatial coordinates, welding sequence number, weld length, weld leg height and node type information of all weld seams, constructing an initial weld seam task list and associating it with the corresponding workpiece, ensuring that the weld seam data is consistent with the physical characteristics, providing a complete input basis for subsequent accurate scheduling, based on the updated dynamic reachability constraint graph and the output motion mode label, spatially segmenting the long weld seams and complex node weld seams in the weld seam task list, decomposing a single weld seam into multiple independent welding segments, breaking the original welding sequence constraints within the workpiece, releasing the parallel scheduling space, fundamentally eliminating the waiting deadlock caused by rigid sequence, matching the segmented welding segments with the welding capabilities of four cantilever robots, constructing a dynamically reconfigurable task pool according to the reachability of each robot, welding process parameters and current task load, achieving optimal adaptation of tasks and resources, and significantly improving the overall efficiency of multi-machine collaborative operation;

[0076] It should be noted that the system receives workpiece production data files from upstream CATIA or Tribon software in real time via an Ethernet communication interface, supporting both .stp and .spf formats. First, it reads the welding process information from the file, extracting the spatial coordinates, welding sequence number, weld length, weld leg height (range 4-7mm), and node type (including mandatory nodes such as 30° / 45° beveling, R-holes, drainage holes, and wrap angles) for each weld. After extraction, the weld information is spatially correlated with the scanned point cloud model based on the workpiece ID. Verification ensures that the analyzed data matches the actual workpiece characteristics, ultimately generating an initial weld task list containing a unique weld identifier, spatial coordinates, process parameters, and information about the workpiece to which it belongs. Based on the dynamic reachability constraint graph (octree voxel size 50mm, update frequency 10Hz) and motion mode labels output in real time by the conflict prediction update unit, long welds and complex node welds in the initial task list are intelligently segmented. For fillet welds exceeding 10 meters in length, they are automatically segmented into multiple independent welding segments of 3-5 meters, with the segmentation points selected at the stiffener gaps or the midpoints of straight segments. For complex nodes such as R-holes and beveling, the reachability boundaries of the robots are identified based on the reachability domain constraint map. The node region is then separately divided into independent corner welding segments. During the segmentation process, the spatial position and current interference field occupancy probability of each welding segment are dynamically evaluated to ensure that the segmented welding segments have independent execution conditions. At the same time, the original rigid welding sequence constraints within the workpiece are completely broken, releasing the space for parallel task scheduling. After the weld seam segmentation is completed, the welding capabilities of each welding segment are matched in real time with those of four Kawasaki RA010 cantilever robots. The matching criteria include the dynamic status of each robot at the current moment. The system includes an reachability constraint map (containing prohibited and warning area markers), a welding process parameter database (supporting 1.2-1.6mm flux-cored wire, CO2 shielding gas, and 20mm extension length), and the current task load status (number of completed welds and remaining welding time). For each welding segment, the system calculates the robot's reachability score, process matching degree (vertical welding height not exceeding 500mm, stiffener inclination angle 75°-90°), and avoidance cost. The optimal robot is selected and the welding segment is added to its task queue. All welding segments are stored in the dynamic task pool as independent task units.

[0077] The collaborative task scheduling module performs the following steps: A scheduling agent is constructed based on a deep Q-network (DQN). The task pool state output by the decoupling and reconstruction engine for welding tasks and the updated dynamic reachability constraint graph are used as the input state space. The task allocation and path selection of each robot are used as the action space, effectively improving the decision-making accuracy and space utilization of multi-robot collaborative operations. A joint reward function is designed with the optimization objectives of minimizing the total global path length and minimizing the welding torch thermal waiting time. Collision penalty and task completion reward terms are introduced. Millions of self-game training sessions are conducted in a digital twin environment to optimize the decision-making strategy of the scheduling agent, significantly shortening the overall welding path and reducing the welding torch dry-burning waiting time. The trained scheduling agent is deployed to the real-time operating environment. Based on the current task pool state and dynamic interference field constraints, a conflict-free, low-wait-time globally optimal collaborative welding sequence scheme is output online, including the task execution order, welding path sequence, and avoidance strategy of each robot. This ensures the continuity and safety of multi-robot collaborative operations in real time, avoiding spatial conflicts and task stagnation.

[0078] It should be noted that the scheduling agent built based on a deep Q-network has an input state space composed of a dynamic task pool output by the weld seam task decoupling and reconstruction engine and a dynamic reachability constraint graph updated in real time by the conflict prediction and update unit. The dynamic task pool contains the spatial coordinates, process parameters, and status markers of all currently unassigned welding segments. The dynamic reachability constraint graph provides the collision risk level and occupancy probability of each voxel unit within the welding area using an octree data structure. The voxel unit size is 50mm, and the update frequency is 10Hz. The action space is defined as the task allocation and path selection of four Kawasaki RA010 robots, including assigning specific welding segments to designated robots. The system selects welding path sequences and avoidance strategies, and trains the scheduling agent in a self-game environment within a digital twin environment. During training, the joint optimization objectives are minimizing the total global path length and minimizing the welding torch thermal waiting time. Collision penalties and task completion rewards are introduced, and the decision-making strategy is optimized through millions of iterations to ensure the output solution achieves global optimum in both spatial and temporal dimensions. In actual operation, the trained scheduling agent receives the current task pool state and dynamic interference field constraints at a 10Hz frequency, generating a conflict-free, low-wait-time collaborative welding timing scheme online. This scheme specifically includes the task execution order of each robot, i.e., each robot... The sequence of welding segment IDs to be completed sequentially; the welding path sequence, i.e., the complete spatial trajectory from the current position to the welding start point, along the weld path to the end point, and the transfer to the next task start point; and the avoidance strategy, i.e., reducing the movement speed to below 0.3m / s when crossing the warning area, or actively pausing and retreating to a safe position before entering the prohibited area. After the plan is generated, it is immediately sent to the instruction parsing and execution module to ensure that multiple robots work collaboratively in a unified time sequence within the next 3 to 5 seconds time window, avoiding spatial conflicts and thermal waiting; the scheduling agent is fully trained in a digital twin environment, and its decision-making strategy fully considers the Kawasaki RA010 model robot. The kinematic constraints and welding process requirements for the robot are as follows: the upper limit of joint movement speed is 145° / s, 120° / s, 150° / s, 230° / s, 230° / s, and 350° / s respectively; the external axis X / Y direction walking speed range is 0 to 10 m / min; the welding torch extension length is fixed at 20 mm; the vertical welding height must not exceed 500 mm; the rib plate inclination angle must be within the range of 75° to 90°; when the scheduling plan is generated, the intelligent agent automatically avoids the warning area with an occupation probability higher than 0.3, prioritizes the selection of reachable areas to perform welding tasks, and evenly distributes the task volume according to the current load status of each robot to ensure the stability of the overall production line cycle time.

[0079] The instruction parsing and execution module performs the following steps: It receives the output of the globally optimal collaborative welding timing scheme, and parses the task execution order, welding path sequence, and avoidance strategy in the scheme into motion control instruction sequences for each robot, including external axis movement instructions, robot joint movement instructions, and welding process parameter instructions. This ensures that the instructions accurately match the robot's kinematic characteristics and process requirements, guaranteeing the accuracy and consistency of welding actions. Through the established 5G industrial private network communication channel, the parsed motion control instruction sequences are sent to the controllers of the four cantilever robots in a low-latency and highly reliable manner, with timestamps and synchronization signals sent synchronously to ensure the timing coordination of multiple robot actions and the consistency and continuity of collaborative operations. After receiving the instructions, each cantilever robot controller maps them in real time into robot servo drive signals and welding power control signals, driving the cantilever to execute predetermined movements and welding actions. At the same time, it feeds back the robot status, weld completion status, and abnormal alarm information to the collaborative control center through the 5G private network, allowing real-time perception of the execution status and dynamic optimization of subsequent tasks, forming a closed-loop intelligent control to improve collaborative efficiency.

[0080] It should be noted that the globally optimal collaborative welding timing scheme is transmitted in real time to the instruction parsing and execution module via industrial Ethernet. First, the scheme data is unpacked and parsed, converting the task execution order of each robot into a welding segment identifier sequence arranged along the time axis. The welding path sequence is parsed into a spatial point array containing the starting coordinates, weld trajectory point set, ending coordinates, and transfer path. Simultaneously, the avoidance strategy is transformed into speed thresholds and behavior control identifiers. Based on the kinematic model and external axis control protocol of the Kawasaki RA010 robot, the path points and behavior identifiers are further compiled into specific motion control instructions, including external axis... The commands include X / Y direction movement commands (speed range 0–10 m / min, position accuracy ±0.1 mm), rotation commands for the robot's six joints (maximum rotation speeds of joints 1 to 6 are 145° / s, 120° / s, 150° / s, 230° / s, 230° / s, and 350° / s respectively, angle accuracy ±0.01°), and welding process parameter commands (1.4 mm flux-cored wire diameter, CO2 shielding gas flow rate 50 L / min, wire extension length 20 mm, weld leg height 4–7 mm). All commands are sorted by timestamp to form an independent control command sequence for each robot. (Analysis complete) The generated motion control command sequence is transmitted to the controllers of four cantilever robots via a dedicated 5G industrial network communication channel. This 5G network adopts an independent networking mode, deployed within the production line area, and supports uplink bandwidth of no less than 100Mbps and end-to-end latency of less than 10ms to ensure the real-time performance and reliability of the control commands. During the transmission process, a timestamp based on the IEEE 1588 precision time protocol is added to each command, and a synchronization signal is sent every 200ms control cycle to ensure that the four robots coordinate their actions under the same time reference. After receiving the commands, the controllers immediately perform command verification. In conjunction with caching, tasks are executed sequentially according to timestamps. For welding tasks involving multi-machine collaboration, such as adjacent robots operating simultaneously in different areas of the same workpiece, the execution progress of each robot is calibrated in real time through the two-way communication mechanism of the 5G private network to avoid loss of synchronization due to network jitter or processing delays. Each cantilever robot controller maps the received motion control commands into servo drive signals and welding power control signals in real time. The servo drive system drives the motors of each axis to move precisely according to the joint angles and external axis positions in the commands through a closed-loop control algorithm, achieving a position control accuracy of ±0.1mm and a joint angle control accuracy of ±0.At 01°, the welding power source outputs corresponding voltage and current according to process parameter instructions to achieve a stable arc welding process. During welding, the robot controller collects real-time data on the actual position, speed, current, and welding status of each joint through built-in sensors, and reports this data to the collaborative control center via a 5G private network at a frequency of 10Hz. The reported data includes the current weld completion progress, welded section identification, abnormal alarm information, and the robot's current pose and motion status. The collaborative control center dynamically adjusts subsequent scheduling strategies based on the feedback information, forming a complete perception-decision-execution closed-loop control system.

[0081] Example 3, as Figures 1 to 7 As shown, based on Embodiments 1-2, the present invention provides a technical solution: as follows Figure 2 As shown, through information recognition, area segmentation, and task scheduling, multiple welding robots can perform parallel and efficient welding of large workpieces. The specific logic is as follows: by scanning and confirming the identity information of all workpieces produced in the current batch and their specific placement positions on the roller conveyor platform, the software combines the workpiece information with the position information to accurately determine the physical placement range of each workpiece on the platform. Next, the system performs preliminary division of the task area according to the type of production equipment—in this case, a two-gantry sheet production line. The division is based on two dimensions: first, along the centerline of the equipment in the Y direction, combined with the placement position of the workpieces, the entire work area is divided into two large areas, left and right; second, along the equipment in the X direction, combined with the welding time of the workpieces, the data is further divided into smaller work units. After completing the area segmentation, the system sorts the allocated data according to the position order of the workpieces. Finally, the processed task data is sent to each gantry actuator. The robots on each gantry perform welding operations on the workpieces in their respective areas according to the pre-arranged order, realizing a highly efficient production mode of multi-machine collaboration and parallel operation.

[0082] It should be noted that the workpiece data produced upstream includes the welding sequence. To ensure the overall welding effect of the workpiece, the welding sequence within the workpiece will not be changed during on-site production. The equipment will adjust the welding sequence of the workpiece according to the requirements. However, different processing will be carried out depending on the mechanical structure of the sheet body and the frame equipment. The sheet body will not be additionally split into workpiece areas in the Y direction, but the frame will be split.

[0083] like Figure 3 and Figure 4 As shown, under the premise of ensuring safety (no interference), tasks are assigned according to priority, and idle or conflicting gantry are dynamically scheduled for operation. The workflow is described as follows:

[0084] After the system starts, it first enters the waiting state for upper-level data. Once a task is received, gantry 1 will enter the "automatic" state first, complete the data transmission, and enter the waiting state for welding. At this time, the system will start to judge the intervention conditions of other gantry in turn.

[0085] For gantry 2, gantry 3, and gantry 4, the system first checks whether it is in "automatic" mode and whether the data has been completely transmitted. If the conditions are met, it proceeds to the interference judgment stage.

[0086] 1. Gantry 2 mainly performs interference judgment in the X direction. If there is a risk of interference with gantry 1 which is currently working, gantry 2 cannot start welding directly and must execute the avoidance procedure (enter "avoidance" and stay in the safe position) until gantry 1 completes its work and returns to the machine position, and then check whether the interference has been resolved.

[0087] 2. In addition to judging the X direction, Longmen 3 and Longmen 4 also need to judge the interference in the Y direction. If an overlap or conflict with the working area of ​​Longmen 1 is detected, they also need to perform avoidance and wait for Longmen 1 to release the working area.

[0088] While gantry 1 is performing welding tasks, the system continuously monitors the requests and interference status of other gantry. Only when a gantry confirms that it meets the welding conditions of gantry 1 (i.e., no interference) will it be allowed to enter the welding station to perform tasks.

[0089] Once a gantry completes welding and returns to the machine position, if the system detects interference in the position direction, the operation interface will prompt the gantry to be moved manually to remove the interference and ensure equipment safety. Finally, after the robot is called to complete the welding at gantry 1, the main process of this batch of tasks is marked as the end, and the system returns to standby or waiting for the next round of data.

[0090] like Figure 3 and Figure 5 As shown, during the automatic operation of gantry 2, the system first waits for data from the host computer. After the data is sent, gantry 2 reads the weld information and enters the waiting-to-weld state. At this time, the system enters the core interference judgment and coordination stage, where gantry 2 needs to confirm the conditions with gantry 1, gantry 3, and gantry 4 in sequence.

[0091] 1. Coordination with Gantry 1: The system first determines whether there is interference in the X direction. If there is no interference, the process proceeds directly and Gantry 2 can start performing the welding task. If there is interference, Gantry 1 needs to determine the welding conditions of Gantry 2. The system needs to wait for Gantry 1 to complete the current welding and return the data to the machine position before checking whether the interference in the X direction has been resolved. If it has not been resolved, the system prompts the user to manually move Gantry 1 to eliminate the conflict.

[0092] 2. Coordination with Gantry 3 / 4: After handling the interference of Gantry 1, the system then judges the interference between Gantry 2 and Gantry 3 / 4. Gantry 3 / 4 confirms the welding conditions of Gantry 2. If there is a conflict, the process enters a waiting state, waiting for Gantry 3 / 4 to complete the welding and return the data to the machine position. Then the corresponding gantry (Gantry 3 or Gantry 4) performs tracking (which may refer to following or avoidance actions). Only after all actions are completed and the interference is resolved is Gantry 2 allowed to enter the work position.

[0093] Only when all interference judgments (with gantry 1, 3, and 4) meet the conditions will gantry 2 finally begin welding and send back the data to the machine position after completion, releasing resources for subsequent requests from other gantry units. The entire process ensures that multiple gantry units can work together safely and orderly in a limited space through a polling confirmation + dynamic waiting / avoidance mechanism.

[0094] like Figure 3 and Figure 6 As shown, in the automatic operation of gantry 3, the system first waits for data from the host computer. After the data is sent, gantry 3 reads the weld information and enters the waiting-to-weld state. At this time, the system enters the core interference judgment and coordination stage. Gantry 3 needs to confirm the conditions with gantry 1, gantry 2, and gantry 4 in sequence. The judgment is mainly based on whether there is a conflict in the X region and the Y direction.

[0095] 1. Coordination with gantry 1: The system checks whether gantry 1 is in automatic mode and whether the data has been sent. If there is interference, it needs to confirm whether gantry 1 meets the welding conditions of gantry 3. If not, it waits for gantry 1 to complete welding and return to the standby position. Or, when there is interference in the Y direction, it prompts to manually move gantry 1 to eliminate the conflict.

[0096] 2. Coordination with gantry 2: Next, determine the status of gantry 2. If there is interference, gantry 2 needs to enter the avoidance process (enter "avoidance" and stay in the safe position) until gantry 2 is welded and returns to the standby position, or the interference can be removed by manually moving gantry 2.

[0097] 3. Coordination with gantry 4: Finally, determine the state of gantry 4. The logic is similar to the previous two. If there is interference, gantry 4 will perform avoidance, wait for the welding to be completed and the data to be transmitted back to the machine position, or manually move gantry 4 if necessary.

[0098] Only when all interference judgments (with gantry 1, 2, and 4) meet the conditions, i.e., after confirming no interference, will the system allow gantry 3 to call the robot for welding and officially begin executing the task. Finally, after gantry 3 completes all welding operations, the process ends with gantry 3 robot welding completed, and the system returns to the standby or waiting state for the next round of tasks.

[0099] like Figure 3 and Figure 7 As shown, in the automatic operation of gantry 4, the system first waits for data from the host computer. After the data is sent, gantry 4 reads the weld information and enters the waiting-to-weld state. At this time, the system enters the core interference judgment and coordination stage. Gantry 4 needs to confirm the conditions with gantry 1, gantry 2, and gantry 3 in sequence. The judgment is mainly based on whether there is a conflict in the X and Y directions.

[0100] 1. Coordination with gantry 1: The system first makes a judgment in the X direction. If there is interference and there is also a conflict in the Y direction, it is necessary to confirm whether gantry 1 meets the welding conditions of gantry 4. If not, wait for gantry 1 to complete welding and return to the standby position, or when there is interference in the Y direction, prompt to manually move gantry 1 to eliminate the conflict.

[0101] 2. Coordination with gantry 2: Next, determine the status of gantry 2. If there is interference, it is necessary to confirm whether gantry 2 meets the welding conditions of gantry 4. If the conditions are not met, wait for gantry 2 to complete welding and return to the standby position, or remove the interference by manually moving gantry 2.

[0102] 3. Coordination with gantry 3: Finally, determine the status of gantry 3. The logic is similar to the previous two. If there is interference, it is necessary to confirm whether gantry 3 meets the welding conditions of gantry 4, wait for it to complete the welding and return the data to the machine position, or manually move gantry 3 if necessary.

[0103] Only when all interference judgments (with gantry 1, 2, and 3) meet the conditions, i.e., no interference is confirmed, will the system allow gantry 4 to call the robot for welding and officially begin the task. Finally, after gantry 4 completes all welding operations, the process ends with gantry 4 robot welding completed, and the system returns to standby or waiting for the next round of tasks.

[0104] 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.

[0105] 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 collaborative welding system for ship assembly using multi-cantilever robots, comprising a collaborative control center, characterized in that, The collaborative control center has the following communication connection modules: The twin field construction module is based on four cantilever robots deployed on a gantry-type assembly line. Through industrial vision perception and digital modeling, it constructs a digital twin dynamic interference field that integrates the workpiece point cloud model and the real-time pose of the cantilever. The motion modality prediction module, based on historical trajectories and real-time motion features, uses a machine learning classifier to make short-term predictions of typical motion patterns of the cantilever robot, identify potential spatial path conflicts in advance, and then update the dynamic reachability constraint graph. The weld seam task decoupling and reconfiguration engine is used to decompose all weld seams across workpieces into independent task units according to spatial location, welding time, and process requirements, breaking the welding sequence constraints within the workpiece and establishing a dynamically reconfigurable task pool. The collaborative task scheduling module constructs a scheduling agent based on a deep Q-network. It uses the minimization of the total global path length and the minimization of the welding torch thermal waiting time as the joint reward function. It conducts millions of self-game trainings in a twin environment to output the globally optimal collaborative welding timing scheme. The instruction parsing and execution module is used to convert the globally optimal collaborative welding timing scheme into robot control instructions. Through the established industrial private network communication channel, each cantilever controller receives the dynamic path instructions and maps them into robot motion control signals in real time.

2. The ship crew assembly multi-cantilever robot collaborative welding system according to claim 1, characterized in that: The twin field construction module includes a multi-view vision fusion perception unit and a dynamic interference field modeling unit; The multi-view vision fusion perception unit is used to deploy a global binocular vision system at the top of the workstation to collect the three-dimensional pose and motion trajectory of four cantilever robots in real time. Combined with the point cloud model obtained by the 3D structured light camera at the end of the cantilever, it performs unified spatial registration of the workpiece and the robot and outputs multi-source fusion perception data. The dynamic interference field modeling unit, based on the spatially registered data, integrates the workpiece point cloud model and the real-time pose of the cantilever to construct a digital twin dynamic interference field, which includes the robot's current posture, motion trend and workpiece boundary, and generates a dynamic reachability domain constraint graph.

3. The ship crew multi-cantilever robot collaborative welding system according to claim 2, characterized in that: The multi-view vision fusion perception unit performs the following steps: Using a global binocular vision system deployed at the top of the welding station, stereo image pairs of the cantilever end and the traverse mechanism of four cantilever robots are acquired synchronously at a fixed frequency. The three-dimensional pose parameters of each robot in the global coordinate system, including spatial position coordinates and Euler angles, are calculated by a stereo matching algorithm. Trigger the 3D structured light camera mounted on each cantilever end to project and image the workpiece surface with structured light, obtain high-density point cloud data, and transform the local point cloud model to the global coordinate system through the point cloud registration algorithm to form a workpiece point cloud model containing the features of stiffeners, welds, and base plates. The real-time pose of the cantilever calculated by the global binocular vision system is spatiotemporally aligned with the workpiece point cloud model acquired by the 3D structured light camera to establish a spatial mapping relationship between the robot's motion trajectory and the workpiece's geometric features, and output multi-source fusion perception data under a unified spatiotemporal reference.

4. The ship crew multi-cantilever robot collaborative welding system according to claim 2, characterized in that: The dynamic interferometric field modeling unit performs the following steps: Based on the multi-source fusion perception data output by the multi-view vision fusion perception unit, the current posture envelope of the four cantilever robots and the boundary contour of the workpiece point cloud model are extracted. The welding area is segmented into voxels using a spatial octree data structure to construct the initial static interference field. By integrating the kinematic model and real-time motion parameters of the cantilever robot, the motion trend of each robot in the next control cycle is predicted. The predicted trajectory is mapped to the static interference field, and the occupancy status and occupancy probability of each voxel unit are dynamically updated to generate a digital twin dynamic interference field that evolves over time. The collision risk level of each voxel unit in the dynamic interference field is calibrated. Based on the robot's motion speed, approach direction and envelope size, the potential collision probability is calculated, and a dynamic reachability constraint graph containing reachable regions, prohibited regions and warning regions is generated.

5. A ship crew multi-cantilever robot collaborative welding system according to claim 2, characterized in that: The motion modality prediction module includes a feature extraction and classification unit and a conflict prediction and update unit; The feature extraction and classification unit is used to combine historical trajectories and real-time motion features, and introduce a motion pattern classifier based on support vector machines to classify and predict the trajectory features of the four cantilever robots in real time and accurately identify typical motion patterns including cantilever crossing, dwelling extension, retreating and avoiding, and straight-line movement. The conflict prediction and update unit is used to combine the classification results with the current pose to predict the motion path within a short time window in the future, identify potential spatial path conflicts, update the dynamic reachability constraint graph, and simultaneously mark potential conflict areas.

6. A ship crew multi-cantilever robot collaborative welding system according to claim 5, characterized in that: The feature extraction and classification unit performs the following steps: Historical motion trajectory data of four cantilever robots were collected, including joint angles, external axis positions, motion speed and acceleration. A time-series motion feature vector set was constructed, and the trajectory data was segmented by sliding window to extract statistical features and frequency domain features within the window. Based on the support vector machine classification algorithm, the classifier is trained offline using labeled historical motion pattern samples to establish a mapping relationship between trajectory feature vectors and motion pattern categories. The motion pattern categories include cantilever crossing, dwelling extension, retreating and avoiding, and straight-line movement. The system receives cantilever pose data from the multi-view vision fusion perception unit in real time, extracts motion feature vectors within the current time window, inputs them into the trained SVM classifier for online classification, and outputs short-term motion pattern labels for each robot at the current moment.

7. A collaborative welding system for ship assembly using multi-cantilever robots according to claim 5, characterized in that: The conflict prediction and update unit performs the following steps: Based on the motion pattern labels output by the feature extraction classification unit, combined with the kinematic model and dynamic reachability constraint graph of each robot, the motion path and space occupancy sequence of each robot in the next 3-5 second time window are predicted. The predicted multi-robot motion paths are overlaid and analyzed in the spatiotemporal dimensions. A spatial conflict detection algorithm is used to identify path intersections and overlapping areas of space occupation. The time of conflict occurrence, conflict location and conflict severity are calculated, and potential spatial path conflict events are marked. Based on the conflict detection results, update the conflict area markers in the dynamic reachability constraint graph, increase the occupancy probability of the conflict area and mark the reachability as a warning state, and simultaneously feed the updated dynamic reachability constraint graph back to subsequent processes in real time.

8. A ship crew multi-cantilever robot collaborative welding system according to claim 5, characterized in that: The weld seam task decoupling and refactoring engine performs the following steps: The workpiece production data output by the upstream CAD / CAM software is analyzed, and the spatial coordinates, welding sequence number, weld length, weld leg height and node type information of all welds are extracted. An initial weld task list is constructed and associated with the corresponding workpiece. Based on the updated dynamic reachability constraint graph and the output motion mode labels, the long welds and complex node welds in the weld task list are spatially segmented, and a single weld is decomposed into multiple independent welding segments, breaking the original welding sequence constraints within the workpiece. The segmented welding sections are matched with the welding capabilities of four cantilever robots. Based on the reachability of each robot, welding process parameters, and current task load, a dynamically reconfigurable task pool is constructed.

9. A ship crew multi-cantilever robot collaborative welding system according to claim 8, characterized in that: The collaborative task scheduling module performs the following steps: The scheduling agent is constructed based on a deep Q-network. The task pool state output by the decoupling and reconstruction engine for welding tasks and the updated dynamic reachability constraint graph are used as the input state space, and the task allocation and path selection of each robot are used as the action space. Design a joint reward function with the optimization objectives of minimizing the total global path length and minimizing the welding torch thermal waiting time. Introduce a collision penalty term and a task completion reward term. Conduct millions of self-game training sessions in a digital twin environment to optimize the decision-making strategy of the scheduling agent. The trained scheduling agent is deployed to the real-time operating environment. Based on the current task pool status and dynamic interference field constraints, it outputs a conflict-free and low-wait-time globally optimal collaborative welding timing scheme online, including the task execution order of each robot, welding path sequence, and avoidance strategy.

10. A ship crew multi-cantilever robot collaborative welding system according to claim 9, characterized in that: The instruction parsing and execution module performs the following steps: The system receives the output of the globally optimal collaborative welding timing scheme and parses the task execution order, welding path sequence, and avoidance strategy in the scheme into a sequence of motion control commands for each robot, including external axis movement commands, robot joint movement commands, and welding process parameter commands. Through the established 5G industrial private network communication channel, the parsed motion control command sequence is sent to the controllers of four cantilever robots in a low-latency and high-reliability manner, while simultaneously sending timestamps and synchronization signals. After receiving the instructions, each cantilever robot controller maps them in real time to robot servo drive signals and welding power control signals, driving the cantilever to perform predetermined movements and welding actions. At the same time, the robot status, weld completion status and abnormal alarm information are fed back to the collaborative control center through the 5G private network.

Images