A multi-robot cooperative control method and system for a shell manufacturing production line

Abstract

The application discloses a kind of multi-robot collaborative control method and system for shell production line, belong to control or regulating system technical field, to solve the problem of low efficiency of multi-robot collaboration in prior art, weak quality traceability ability, production scheduling rigidification.The method is applied to the three-layer hierarchical control architecture production line including cloud collaborative layer, edge control layer and field execution layer;Its core lies in: cloud collaborative layer is based on as digital twin virtual simulation model to pre-optimize production scheduling scheme;Edge control layer is based on the scheme after optimization, through resource allocation algorithm, dynamically allocates conflict-free space-time resources for multiple robots, and dynamically reschedules field anomalies;Field execution layer realizes the closed-loop control of key process variables through sensor, and associates the measured value of process parameter with time stamp with workpiece unique identification, forms parameter time sequence for quality traceability.

Description

Technical Field

[0001] This invention belongs to the field of control or regulation system technology, specifically to industrial control systems, and relates to centralized control of a central factory, i.e., centralized control of multiple machines, the structure of program control systems, and adaptive control systems based on models or simulators. In particular, it relates to a multi-robot collaborative control method and system for shell-making production lines based on a cloud-edge-field three-layer hierarchical control architecture. Background Technology

[0002] Investment casting is a key process for producing high-precision, complex-structured metal parts, with shell making being its core step. The shell-making process typically involves multiple cycles of wax model assembly, slurry application, sand application, and drying. The quality of the workpiece, such as the mold shell, directly determines the dimensional accuracy and surface quality of the final casting. Traditional shell-making processes rely heavily on manual operation, which is not only labor-intensive and creates harsh production environments, but also makes it difficult to ensure consistent process parameters, leading to large fluctuations in product quality and low production efficiency.

[0003] To address the aforementioned issues, automation technology has been introduced into shell-making production lines. For example, US6453210B1 discloses an automation system employing a remote server and a PLC controller. In this system, the remote server retrieves a preset process formula from a database based on the shell type identified by a radio frequency identification (RFID) tag, and downloads the corresponding robot motion instructions to the PLC. The PLC then controls a single robot offline to perform multiple dipping cycles. This system achieves a certain degree of automation in the shell-making process.

[0004] However, in-depth analysis revealed the following significant defects and shortcomings in the aforementioned existing technologies:

[0005] First, its two-tiered control architecture of "remote server-PLC" cannot support real-time collaborative operation among multiple robots. Each robot work unit is physically isolated and operates independently, resulting in high network communication latency. This makes it impossible to perform millisecond-level dynamic scheduling and spatial conflict detection, leading to a large amount of waiting time between robots and low overall system efficiency and space utilization.

[0006] Secondly, its quality traceability capability is very limited. RFID tags are only used to identify the type of workpiece to trigger the corresponding preset formula, and do not record the actual measured values ​​of key process parameters (such as coating material temperature, immersion time, sand flow rate, drying temperature and humidity, etc.) during the production process. Once a quality problem occurs, it is impossible to trace the specific source of parameter deviation.

[0007] Secondly, its control method is open-loop control. The system mainly relies on preset formula parameters for execution and lacks the ability to perceive and adjust for dynamic changes in the production site in real time. When the level of the coating material, the flow rate of the sand, or the ambient temperature and humidity fluctuate, the system cannot automatically compensate, which can easily lead to deviations from the standard in process execution and affect the consistency of product quality.

[0008] Finally, its production scheduling strategy is rigid and lacks a pre-verification mechanism. The system typically adopts a simple first-in, first-out (FIFO) strategy, which cannot achieve global optimization when producing multiple products on a mixed production line, and is prone to situations where long-running tasks block short-running tasks. At the same time, the introduction of new products or processes must be carried out through trial and error on the physical production line, which is not only costly but also seriously affects normal production.

[0009] Further analysis from the perspective of control system architecture reveals a fundamental deficiency in existing technologies: the lack of a hierarchical control architecture that spans multiple time scales. Centralized control of the main plant requires unified scheduling and coordination of multiple machines, while model-based adaptive control demands that the system possess the ability to perform feedforward optimization using model predictions. However, the existing remote server-PLC two-layer architecture is a general-purpose programmable control system structure, suitable only for simple joint control of single machines or a small number of devices.

[0010] Therefore, how to overcome the problems of poor real-time collaboration capability, coarse quality traceability granularity, lack of closed-loop feedback control, and rigid scheduling strategy in the existing technology, and provide a multi-robot control scheme that can achieve efficient collaboration, precise control and intelligent scheduling, is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0011] To address the aforementioned technical challenges, this invention provides a multi-robot collaborative control method and system for shell manufacturing production lines. It aims to achieve a comprehensive improvement in production efficiency, product quality, and production flexibility through an innovative three-tiered hierarchical control architecture and collaborative control strategy, consisting of a cloud-based collaborative layer, an edge control layer, and a field execution layer.

[0012] This invention provides a multi-robot collaborative control method for shell manufacturing production lines, applied to an industrial automation production line with a three-layer hierarchical control architecture including a cloud collaboration layer, an edge control layer, and a field execution layer. The method includes: at the cloud collaboration layer, simulating a production scheduling scheme based on a virtual simulation model of the production line, and optimizing the production scheduling scheme based on prediction indicators (including predicted total completion time, equipment utilization rate, and collision risk value) output by the simulation; at the edge control layer, dynamically allocating conflict-free spatiotemporal resources (including action time windows and motion paths) for multiple robots in a shared workspace to perform processing tasks based on the optimized production scheduling scheme using a resource allocation algorithm; at the field execution layer, obtaining measured values ​​of process parameters by real-time monitoring of key process variables using sensors, and automatically adjusting the actuators using a closed-loop control algorithm based on the deviation between the measured process parameter values ​​and preset setpoints; and at the field execution layer, associating and recording the timestamped measured process parameter values ​​with the unique identifier of the currently processed workpiece to form a parameter time series.

[0013] Furthermore, the method also includes: distributing the optimized production scheduling scheme from the cloud collaboration layer to the edge control layer. This ensures that the actual production scheduling scheme is virtually verified and optimized, thereby guaranteeing the feasibility and efficiency of the production plan from the outset.

[0014] Optionally, the method further includes: monitoring the actual execution status of the multiple robots at a preset monitoring frequency in the edge control layer; when the deviation between the actual execution status and the planned status determined based on the conflict-free spatiotemporal resources exceeds a preset threshold, performing local rescheduling or global rescheduling according to the deviation type, and regenerating the action time window and motion path of the tasks that have not yet been executed in the processing task. This enables the system to respond quickly to unexpected disturbances on site (such as robot motion delays, temporary equipment failures, etc.), ensuring the robustness of the production process.

[0015] In one embodiment, the method further includes: at the edge control layer, when a scheduling event is triggered, a genetic algorithm is used to solve the production sequence of multi-product mixed-line production in real time under multiple constraints, with the total completion time and equipment utilization rate as optimization objectives; wherein the chromosome of the genetic algorithm includes a first encoding segment for representing the workpiece task placement order and a second encoding segment for representing the robot allocation result, and the multiple constraints include process path sequence constraints, equipment capacity constraints, robot reachability constraints, shortest drying time constraints, and shared workspace conflict constraints.

[0016] This solution can effectively address scenarios involving multi-variety, small-batch production, maximizing resource utilization by using intelligent algorithms to find the globally optimal solution.

[0017] Furthermore, the method also includes: transmitting the actual operating data collected by the edge control layer and the field execution layer back to the cloud collaboration layer; the cloud collaboration layer aligns the actual operating data with the predicted indicators output by the virtual simulation model according to the task number and timestamp of the processing task; calculates an error vector including trajectory error, process time error, and key process variable error; and calibrates the robot kinematic correction parameters, dynamic parameters, and process time estimation parameters in the virtual simulation model based on the error vector; wherein, the virtual simulation model is a digital twin model of the production line. This constitutes a closed loop between the physical world and the digital world, enabling the virtual model to continuously reflect the real state of the physical system and ensuring the accuracy of simulation and optimization.

[0018] Preferably, the resource allocation algorithm allocates resources based on the digital model of each robot's workspace. The digital model includes a robot kinematic model established based on DH parameters and a three-dimensional convex hull model of the robot links. Spatial conflicts are determined by whether the distance between the three-dimensional convex hull models of any two robots is less than a preset safety distance. By accurately modeling the robot's workspace and planning conflict-free paths and time windows, safe and efficient collaborative operation of multiple robots within a shared space is achieved.

[0019] Optionally, the method further includes: aggregating all parameter time-series data associated with the unique identifier of a single workpiece, generating a structured JSON format data packet, and uploading it to the cloud collaboration layer; wherein the JSON format data packet includes at least the workpiece unique identifier, process record array, workstation number, equipment number, measured parameter values, collection timestamp, and data validity flag. This structured data format facilitates subsequent data storage, querying, and analysis, providing a data foundation for achieving full-parameter quality traceability from the "product level" to the "single piece level".

[0020] In one implementation, the edge control layer and the field execution layer communicate via Profinet IRT real-time industrial Ethernet; and the closed-loop control algorithm is a PID control algorithm or a cascaded PID control algorithm. The use of high-speed, deterministic real-time industrial Ethernet ensures reliable transmission of control commands and status data between the edge layer and the field layer, while the mature PID control algorithm ensures precise control of process parameters.

[0021] This invention also provides a multi-robot collaborative control system for shell-making production lines, applied to an industrial automated production line for processing workpieces. The system includes a cloud collaboration layer, an edge control layer, and a field execution layer, configured as a three-layer hierarchical control architecture. Specifically: the cloud collaboration layer is configured to simulate a production scheduling scheme based on a virtual simulation model of the production line, and optimize the production scheduling scheme based on the predicted indicators output by the simulation; the edge control layer is configured to dynamically allocate conflict-free spatiotemporal resources to multiple robots within a shared workspace based on the optimized production scheduling scheme using a resource allocation algorithm; the field execution layer, including sensors and actuators, is configured to monitor key process variables in real time through the sensors to obtain measured values ​​of process parameters, and automatically adjust the actuators using a closed-loop control algorithm based on the deviation between the measured process parameter values ​​and preset values; the field execution layer is further configured to associate and record the timestamped measured process parameter values ​​with the unique identifier of the currently processed workpiece, forming a parameter time series.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] 1. An innovative three-layer hierarchical control architecture of cloud-edge-field is proposed to meet the centralized control requirements of the main plant. By separating non-real-time global optimization (cloud), millisecond-level real-time scheduling (edge), and microsecond-level physical control (field) to different computing layers, and leveraging heterogeneous network protocols (EtherNet / IP and Profinet IRT) to achieve cross-timescale collaboration, the real-time bottleneck of the traditional two-layer architecture in high-density multi-robot collaborative scenarios is solved, significantly shortening the production cycle and improving equipment utilization.

[0024] 2. A model-based adaptive control closed loop was implemented. By deploying sensors at the field execution layer and employing a closed-loop control algorithm, precise control of key process parameters was achieved, overcoming the quality instability problem inherent in existing open-loop control technologies. Combined with a parameter time-series traceability method based on unique identifiers, the granularity of quality traceability was elevated from the 'product level' to the 'single-piece, end-to-end level', providing finer-grained data support for rapid identification of quality issues and continuous process improvement.

[0025] 3. Reduce production risks and debugging costs: By introducing pre-scheduling optimization based on virtual simulation models at the cloud-based collaboration layer, "verification before execution" is achieved. The introduction of new products and processes can be simulated and iterated quickly and at low cost in virtual space, effectively avoiding the risk of physical collisions and shortening the debugging time of new products from several hours to tens of minutes, significantly reducing interference with normal production.

[0026] 4. A closed-loop intelligent manufacturing control system satisfying the planning-execution-perception-analysis-feedback mechanism has been constructed. This invention organically combines cloud-based simulation optimization, edge real-time scheduling, on-site closed-loop execution, and full-process data traceability. Actual operational data from the physical world continuously feeds back into and calibrates the virtual model in the cloud, while the optimized model guides the physical world to operate more efficiently, forming a self-learning and continuously evolving intelligent manufacturing control system. This architecture is not only applicable to precision casting shell production lines but can also be horizontally extended to various discrete manufacturing scenarios requiring multi-robot collaboration, such as automotive welding, electronic assembly, and food packaging, demonstrating significant technical versatility and industrial application value.

[0027] 5. Experiments verified the continuous self-calibration capability of the digital twin model: A 30-day production line comparison experiment showed that after activating the closed-loop calibration feedback mechanism of this invention, the prediction cycle deviation of the virtual simulation model continuously converged from an initial 8.2% to 1.3% within 30 days, the trajectory prediction error decreased from 11.3mm to 3.8mm, the process time prediction error decreased from 15.7 seconds to 4.2 seconds, and the number of rescheduling triggers caused by scheduling deviations decreased by 77.6%, demonstrating the technical effect of "the simulation model becomes more accurate with use". In contrast, under the condition of not enabling closed-loop calibration, the prediction cycle deviation continuously deteriorated from 8.2% on day 1 to 9.6% on day 15, proving that the fidelity of the digital twin model lacking feedback calibration will continuously decay over time. This comparison fully demonstrates the irreplaceable nature and continuous optimization capability of the closed-loop calibration feedback mechanism of this invention. Attached Figure Description

[0028] Figure 1 The diagram shows a three-layer hierarchical control architecture for a system provided in an embodiment of the present invention.

[0029] Figure 2 This is an overall flowchart of the multi-robot cooperative control method provided in an embodiment of the present invention.

[0030] Figure 3 The flowchart illustrates the RFID-based end-to-end quality traceability method provided in this embodiment of the invention.

[0031] Figure 4 This is a schematic diagram of the closed-loop calibration of the virtual simulation model and physical system provided in an embodiment of the present invention.

[0032] Figure 5 This provides a complete example of the structure of a JSON format data packet for an embodiment of the present invention.

[0033] Explanation of reference numerals in the attached figures:

[0034] 10-Cloud Collaboration Layer, 11-Virtual Simulation Model, 20-Edge Control Layer, 21-Multi-Robot Collaborative Scheduling Module, 30-Field Execution Layer, 31-Industrial Robot, 32-RFID Reader / Writer, 51-Distribute Scheduling Scheme, 52-Automatic Model Calibration, 60-Physical Production Line, 61-Actual Operating Data, S100-Cloud Simulation Optimized Scheduling Scheme, S200-Edge Layer Dynamic Allocation of Spatiotemporal Resources, S300-Field Layer Closed-Loop Control and Execution, S400-Field Layer Recording Parameter Time Series, S410-Initialization and Filing, S420-Recording Dipping Process Parameters, S430-Recording Sand Spraying Process Parameters, S440-Data Encapsulation and Upload. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0036] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0037] like Figure 1As shown, this embodiment of the invention provides a multi-robot collaborative control system for an automated shell-making production line in precision casting. The system employs a three-tiered hierarchical control architecture: a cloud collaboration layer, an edge control layer, and a field execution layer. This architecture includes a cloud collaboration layer 10, an edge control layer 20, and a field execution layer 30. The cloud collaboration layer 10 handles non-real-time, long-cycle, and global tasks, and may include a virtual simulation model 11, a data analysis and optimization module, and a remote visualization monitoring module. The edge control layer 20 is a high-performance controller deployed near the production site, such as an industrial control computer running a real-time Linux operating system, responsible for millisecond-level real-time scheduling and management. It may include a multi-robot collaborative scheduling module 21, a quality traceability management module, and a dynamic production scheduling module. The field execution layer 30 consists of physical devices responsible for executing specific process actions and collecting real-time data. It mainly includes multiple industrial robots 31, various process equipment (such as slurry tanks, sandblasting machines, and drying lines), and sensing and detection devices (such as RFID readers 32, laser rangefinders, and temperature and humidity sensors).

[0038] The information and control flows between the three layers demonstrate a clear hierarchical relationship. The cloud-based collaboration layer 10 performs non-real-time simulation and optimization based on production orders and global data, generating a macro-level production scheduling plan. This plan is then sent to the edge control layer 20, which parses it into millisecond-level real-time collaboration instructions and sends them to various devices in the field execution layer 30. While executing tasks, the field execution layer 30 collects real-time operating data and process parameters through various sensors. This data is used for local closed-loop control and fed back to the edge control layer 20 for status monitoring and dynamic adjustment, ultimately converging to the cloud-based collaboration layer 10 for data analysis and model optimization. This layered and decoupled design separates long-cycle optimization from real-time control, effectively resolving the contradiction between high remote server latency and high real-time requirements in traditional two-layer architectures.

[0039] It is necessary to clearly distinguish the essential difference between the three-layer hierarchical architecture of the present invention ("cloud collaboration layer - edge control layer - on-site execution layer") and the three-layer functional layering of "perception-decision-execution" in existing technologies. The three-layer architecture disclosed in existing technologies (such as CN120952485A, etc.) is essentially a division of software functional modules within the same computing platform (or adjacent nodes within the same local area network)—the perception layer collects data, the decision layer performs logical judgments, and the execution layer outputs instructions. The layers transmit task messages in a homogeneous network environment and do not involve distributed collaboration issues across network segments or computing paradigms.

[0040] The proposed "cloud-edge-field" three-layer architecture is an architectural paradigm based on the geographical distribution of computing resources and the level of real-time performance: the cloud collaboration layer is deployed on a remote data center / cloud server, undertaking non-real-time global long-cycle simulation and optimization tasks (seconds to minutes); the edge control layer is deployed on an industrial control computer near the production site, undertaking millisecond-level real-time scheduling and multi-robot collaboration; the field execution layer consists of embedded controllers, sensors, and actuators, undertaking microsecond to millisecond-level physical signal acquisition and control. These three layers span two distinct network environments: wide area networks (cloud to edge, EtherNet / IP, latency approximately 50ms) and real-time industrial Ethernet (edge ​​to field, Profinet IRT, latency less than 1ms). Data transmission across these layers requires protocol conversion, timing alignment, and data dimensionality reduction.

[0041] This means that the present invention must address a series of technical challenges that existing functional layered architectures do not face, such as "how computing nodes with different real-time performance levels should collaborate," "how to connect high-latency networks with low-latency networks," and "how to losslessly convert global optimization results into real-time execution instructions." It is this essential difference that makes the three-layer architecture of the present invention not merely a difference in naming, but a system-level architectural innovation involving distributed computing, heterogeneous networks, and cross-timescale collaboration.

[0042] To support efficient data interaction under this hierarchical architecture, this invention preferably employs a hybrid communication network. The cloud collaboration layer 10 and the edge control layer 20 communicate using the EtherNet / IP protocol, with a communication latency controllable to within 50ms, sufficient to meet the transmission requirements of non-real-time, high-volume data such as production scheduling and optimization schemes. For the edge control layer 20, which has the highest real-time requirements, and the field execution layer 30, the Profinet IRT (Isochronous Real-Time) real-time industrial Ethernet protocol is used. Profinet IRT's communication cycle can be as low as 1ms or even lower, with a bandwidth of up to 100Mbps. Compared to the commonly used DH+ network (240kbps, latency in the hundreds of milliseconds), this represents a bandwidth increase of approximately 400 times and a latency reduction of two orders of magnitude. This design ensures that collaborative action commands and status feedback between multiple robots can be transmitted with extremely low latency and jitter, providing a fundamental communication guarantee for achieving precise spatiotemporal collaborative control.

[0043] like Figure 2As shown, one aspect of the present invention provides a multi-robot collaborative control method for shell manufacturing production lines. The method as a whole embodies the collaborative workflow of the aforementioned three-layer hierarchical architecture. First, in step S100, in the cloud-based collaborative layer 10, based on a virtual simulation model of the production line, the production scheduling scheme is simulated, and the production scheduling scheme is optimized according to the predicted indicators output by the simulation (such as predicted cycle time, equipment utilization rate, collision risk, etc.).

[0044] In one specific embodiment, step S100 includes the following sub-steps:

[0045] S101, the cloud collaboration layer receives production order data, workpiece type data, process path data, equipment availability status data, and initial pose data of each industrial robot;

[0046] S102, Based on the geometric layout of the production line, the robot kinematics model, the workstation capacity model, the process time model, and the shared workspace model, establish the virtual simulation model;

[0047] S103, generate multiple candidate production scheduling schemes based on the production order data. Each candidate production scheduling scheme includes a workpiece task number, process sequence, candidate robot number, candidate workstation number, candidate action time window, and candidate motion path number.

[0048] S104, input each candidate production scheduling scheme into the virtual simulation model and calculate the corresponding prediction index. The prediction index includes at least the predicted total completion time, equipment utilization rate, robot waiting time and collision risk value.

[0049] S105, Evaluate each candidate production scheduling scheme according to the objective function, wherein the objective function is:

[0050]

[0051] Where T represents the predicted total completion time, W represents the robot waiting time, R represents the collision risk value, and U represents the equipment utilization rate. , , , Preset weighting coefficients;

[0052] S106 Select the candidate production scheduling scheme that meets the conditions of zero or lower than the preset risk threshold and the objective function value meets the preset optimization conditions as the optimized production scheduling scheme, and output it as a scheduling table including task number, robot number, workstation number, action start time, action end time, path number and process parameter setting value.

[0053] Subsequently, in step S200, the edge control layer 20 receives and parses the production scheduling scheme from the cloud, and dynamically allocates conflict-free spatiotemporal resources to multiple robots in the shared workspace through a resource allocation algorithm.

[0054] In one specific implementation, step S200 includes the following sub-steps:

[0055] S201, the edge control layer receives the optimized production scheduling plan from the cloud collaboration layer and forms a queue of tasks to be executed according to the task start time;

[0056] S202, for each task in the task queue to be executed, generate at least one candidate action time window and at least one candidate motion path based on the robot's current pose, target workstation position and process action type;

[0057] S203, calculate the link pose of the robot at each sampling moment on the candidate motion path based on the DH kinematic model of each robot, and determine the spatial occupancy area of ​​the robot in the shared workspace based on the three-dimensional convex hull model of each link.

[0058] S204: Calculate the distance between the spatial occupancy areas of any two robots at the same sampling time. If the distance is greater than or equal to the preset safety distance, it is determined that there is no conflict between the corresponding candidate action time window and the candidate motion path.

[0059] S205, if the distance is less than the preset safe distance, the action time window of the lower priority task will be delayed first; if the delay exceeds the preset maximum waiting time, the candidate motion path of the task will be regenerated.

[0060] S206, when the same task fails to generate a conflict-free action time window and motion path after a preset number of consecutive attempts, the edge control layer triggers dynamic rescheduling.

[0061] S207, the edge control layer outputs conflict-free execution instructions, which include robot number, action type, action start time, action end time, path point sequence and target workstation number.

[0062] Next, in step S300, at the field execution layer 30, the robot and process equipment execute tasks according to the assigned instructions. Simultaneously, key process variables are monitored in real time via sensors, and based on the deviation between the monitored values ​​and set values, a closed-loop control algorithm automatically adjusts the actuators. Finally, in step S400, also at the field execution layer 30, the measured values ​​of process parameters, including timestamps, measured in real time by sensors, are associated with and recorded with the unique identifier of the currently processed workpiece, forming a parameter time series. This provides a foundation for quality traceability and data analysis. These four core steps work together to solve a series of problems in existing technologies, such as the inability to perform pre-scheduling verification, lack of real-time collaborative capabilities, open-loop control leading to quality instability, and lack of refined quality traceability.

[0063] In a preferred embodiment, the method further includes distributing the production scheduling scheme optimized by virtual simulation from the cloud collaboration layer 10 to the edge control layer 20. The virtual simulation model 11 of the cloud collaboration layer 10 can virtually run the production scheduling scheme at a speed far exceeding real-time speed (e.g., 50 times the real-time speed), thereby predicting potential production bottlenecks, robot path conflicts, and other issues before actual execution on the physical production line. Based on the simulation results, the cloud can automatically iteratively optimize the scheduling parameters until an optimized scheme that meets preset goals (e.g., cycle time ≤ 4 minutes and zero collisions) is obtained. Distributing this optimized scheme is equivalent to providing a high-quality "initial solution" for real-time scheduling in the edge layer, avoiding expensive and time-consuming trial and error debugging in the physical world, and significantly shortening the introduction time of new products and the production preparation cycle.

[0064] Furthermore, to enhance the system's adaptability and robustness to dynamic changes on-site, the method also includes monitoring the robot's actual execution state, including joint pose, speed, and torque, at a preset monitoring frequency (e.g., 10ms) in the edge control layer 20. When the deviation between the robot's actual execution state and the planned state exceeds a preset threshold (e.g., position deviation greater than 5mm or time delay greater than 500ms), the system will automatically trigger a dynamic rescheduling of the production scheduling plan. The multi-robot collaborative scheduling module 21 of the edge control layer 20 will immediately re-perform global or local timing planning and path planning based on the current actual state of all robots, generating new conflict-free scheduling instructions. This rapid response mechanism enables the system to complete adjustments and resume collaborative work within 200ms when robot actions experience minor delays or external interference, avoiding collisions or production interruptions caused by error accumulation, and improving response time by several orders of magnitude compared to manual intervention (typically greater than 5 minutes).

[0065] In one specific implementation, dynamic rescheduling includes the following steps: the edge control layer collects the robot's actual joint position, end effector pose, speed, torque, and task execution progress at a preset monitoring frequency, and compares the actual execution state with the planned state; when the deviation between the actual execution state and the planned state is less than a first threshold, the original production scheduling scheme remains unchanged; when the deviation is greater than or equal to the first threshold and does not cause a risk of conflict in the shared workspace, local rescheduling is performed, which includes adjusting the action time window of the affected robot and its subsequent associated tasks; when the deviation is greater than or equal to a second threshold, or when a collision risk is determined based on the space occupancy prediction result, global rescheduling is performed; when global rescheduling is performed, the edge control layer divides the tasks into completed tasks, currently executing and uninterruptible tasks, interruptible tasks, and unexecuted tasks; among them, completed tasks remain unchanged, the current path of currently executing and uninterruptible tasks is locked, and interruptible tasks and unexecuted tasks re-participate in the time window and motion path allocation;

[0066] After rescheduling is completed, the edge control layer generates a new scheduling version number and overwrites the tasks that have not yet been executed in the original scheduling table with the new scheduling table. At the same time, it records the rescheduling event number, triggering reason, time window before adjustment and time window after adjustment.

[0067] In handling complex scheduling scenarios such as multi-product mixed-line production, as a more optimized implementation, the method further includes, at the edge control layer 20, employing a multi-objective optimization algorithm for real-time solution when scheduling events (such as new product launch, order changes, or equipment malfunctions) are triggered. Specifically, a genetic algorithm such as NSGA-II (Non-Dominated Sorting Genetic Algorithm) can be used, with at least one production efficiency indicator as the optimization objective, such as simultaneously minimizing the total completion time (makespan) and maximizing the overall equipment utilization rate. During the solution process, the algorithm considers multiple constraints, including process paths for different products, time requirements for each process, equipment capacity limitations, robot accessibility, and the minimum time limit for the drying process. By setting a reasonable population size (e.g., 100) and iteration generations (e.g., 200), the algorithm can search for a set of Pareto optimal solutions in a short time, providing production decision-makers with optimal production sequences under different preferences. Experimental data shows that, compared with the traditional FIFO (First In First Out) strategy, the multi-objective optimization algorithm can reduce the total completion time of 600 sets of three-product mixed production lines from 48.2 hours to 34.8 hours, a reduction of 27.8%. At the same time, it reduces the congestion waiting time at bottleneck workstations from 8.6 hours to 1.1 hours, significantly improving the flexibility and overall efficiency of the production line.

[0068] In one specific implementation, the genetic algorithm employs a task sequence encoding method. Each chromosome includes a first encoding segment and a second encoding segment. The first encoding segment represents the delivery order of multiple workpiece tasks, and the second encoding segment represents the robot allocation result corresponding to each task. During population initialization, the edge control layer generates N feasible chromosomes based on the current order, completed tasks, currently executing tasks, and equipment availability. For chromosomes that do not meet constraints related to process path sequence, equipment capacity, robot reachability, or shortest drying time, a penalty function is used to correct them or reduce their fitness. The genetic algorithm aims to minimize the total completion time and maximize equipment utilization, and further introduces robot waiting time and constraint violation penalty terms.

[0069] Fitness functions include:

[0070]

[0071]

[0072] in, Let U represent the total completion time, U represent the equipment utilization rate, and P represent the constraint violation penalty term. In each iteration, non-dominated sorting, crowding distance calculation, selection, crossover, and mutation operations are performed on the population. The iteration stops when the number of iterations reaches a preset number of generations, or when the Pareto solution set has not changed within a consecutive preset number of generations. The edge control layer selects the solution that satisfies the current production preference from the Pareto solution set as the optimal production sequence.

[0073] To enable continuous learning and self-evolution of the production system, the method further includes transmitting the actual operating data collected by the edge control layer 20 and the field execution layer 30 back to the cloud collaboration layer 10. The field execution layer 30 and its associated equipment together constitute the physical production line 60. Figure 4As shown, this constitutes a closed-loop feedback between the physical production line 60 and the virtual simulation model 11. After receiving the actual operation data 61, including the robot's actual motion trajectory, the actual time of each process, and the measured values ​​of sensors, the cloud collaboration layer 10 compares and analyzes it with the predicted values ​​of the virtual simulation model 11. When a systematic deviation is found, the cloud data analysis and optimization module will automatically calibrate and optimize the internal parameters of the virtual simulation model 11 through algorithms such as gradient descent and particle swarm optimization. For example, it calibrates the robot's kinematic parameters (such as link length error), dynamic parameters (such as friction coefficient), and the time estimation model for each process. This automatic model calibration mechanism 52 continuously improves the fidelity of the virtual simulation model as a digital twin model, and the cycle deviation of the simulation prediction can converge from more than 8% initially to less than 3%, and the trajectory deviation is less than 5mm. A more accurate virtual simulation model can in turn generate a better dispatching scheme 51, thus forming a virtuous cycle of "planning-execution-learning-optimization".

[0074] In one specific implementation, the cloud collaboration layer automatically calibrates the virtual simulation model based on actual operational data, including the following steps:

[0075] S501, receive the actual operation data uploaded by the edge control layer and the field execution layer. The actual operation data includes the task number, robot number, actual trajectory points, actual start time of the process, actual end time of the process, sensor measured values, and rescheduling event records.

[0076] S502 aligns the actual running data with the predicted data output by the virtual simulation model according to the task number, robot number, and timestamp;

[0077] S503, Calculate the error vector, which includes trajectory error, process time error and key process variable error;

[0078] S504 uses the sum of squared errors as the calibration objective function:

[0079]

[0080] in, Indicates trajectory error. This indicates the error in process time. This represents the error of key process variables, where α, β, and γ are preset weights.

[0081] S505, Based on the calibration objective function, the robot kinematic correction parameters, dynamic friction parameters and process time estimation parameters in the virtual simulation model are iteratively updated;

[0082] S506: When the error decrease after a preset number of iterations is less than the convergence threshold, or when the maximum number of iterations is reached, calibration is stopped, and the calibrated virtual simulation model is used for simulation optimization of the next batch of production scheduling schemes.

[0083] To verify the technical effectiveness of the aforementioned digital twin closed-loop calibration feedback mechanism (feature g), the inventors conducted a 30-day comparative experiment on an actual precision casting shell production line, quantitatively evaluating the continuous improvement effect of closed-loop calibration on the prediction accuracy of the virtual simulation model and the quality of production scheduling.

[0084] The same production line was operated in two phases. Phase 1 (first 15 days, "open-loop mode"): The virtual simulation model used only initial theoretical parameters for simulation prediction; actual operating data was collected but not used for model calibration. Phase 2 (last 15 days, "closed-loop calibration mode"): The closed-loop calibration feedback mechanism of this invention was activated. After each production batch (approximately 200 workpieces), the virtual simulation model parameters were automatically calibrated using actual operating data. The experimental results are as follows:

[0085] In the "open-loop mode," the prediction accuracy of the virtual simulation model not only failed to improve itself, but also deteriorated due to time-varying factors such as equipment wear and environmental temperature changes. The prediction cycle time deviation worsened from 8.2% on day 1 to 9.6% on day 15. The average trajectory prediction error was 11.3 mm, the average process time prediction error was 15.7 seconds, and the number of rescheduling triggers due to scheduling scheme deviations reached 18.3 times per 100 workpieces. This means that without closed-loop calibration, the fidelity of the digital twin model will continuously decrease over time, and the simulation prediction will gradually lose its reference value.

[0086] After initiating closed-loop calibration, the predicted cycle time deviation decreased from an initial 8.2% to 4.7% on day 5, 2.9% on day 10, and 2.1% on day 15 (an improvement of 78.1%), further converging to 1.3% on day 30. The trajectory prediction error decreased from 11.3 mm to 3.8 mm (an improvement of 66.4%), the process time prediction error decreased from 15.7 seconds to 4.2 seconds (an improvement of 73.2%), and the number of rescheduling triggers decreased from 18.3 times to 4.1 times (an improvement of 77.6%). The calibration mechanism demonstrates continuous improvement capability—that is, the simulation model becomes more accurate with use.

[0087] The experimental data above show that feature g (digital twin closed-loop calibration feedback) is not a simple application of existing digital twin technology, but rather produces technical effects that cannot be expected by existing technologies and continue to improve over time through a closed-loop value chain of "actual operating data - error vector - automatic calibration of model parameters - improved prediction accuracy - reduced scheduling deviation - reduced rescheduling frequency".

[0088] Specifically, in the resource allocation algorithm of the edge control layer 20, the spatiotemporal resources include the action time window and motion path allocated to each robot. The action time window defines the start time, end time, and action type of a robot performing an action. The motion path is a series of pose points of the robot in three-dimensional space. To ensure no collisions, the resource allocation algorithm is based on the digital model of each robot's workspace. This model can be established using the Denavit-Hartenberg (DH) parameter method to build the robot's kinematic model, and a three-dimensional convex hull model can be built for each link of the robot. During path planning, the algorithm ensures that at any given time, a minimum safe distance (e.g., 100mm) is always maintained between the three-dimensional convex hull models of any two robots. Through this collision detection accurate to the three-dimensional geometric level, the system can plan a close cooperative path where robots "brush past" each other, thereby maximizing the utilization of the shared workspace, reducing the idle time caused by waiting and avoidance of robots from 3.2 minutes per cycle to 0.3 minutes, and increasing the average robot utilization rate from 48.2% to 87.3%.

[0089] In terms of data traceability, to facilitate data storage, transmission, and parsing, the method also includes summarizing all parameter time-series data associated with the unique identifier (UUID) of a single mold shell. For example... Figure 3 As shown, after a mold shell M1 completes all processes, all data recorded in stages such as the slurry application process S420 and the sand application process S430 will be integrated into a structured JSON (JavaScript Object Notation) format data package and uploaded to the cloud collaboration layer 10 for data archiving S440. The JSON format has good readability and cross-platform compatibility, and can clearly express complex nested data structures. For example... Figure 5 As shown, a complete JSON data packet can contain basic information about the mold shell (such as UUID and variety code), detailed process parameters for each coating layer (such as layer sequence number, slurry temperature, residence time, sand spraying flow rate, etc.), and the corresponding equipment number and timestamp. This standardized data format greatly facilitates subsequent data analysis, quality traceability, and visualization, reducing the time for locating quality problems from the traditional 85 minutes to 2.8 minutes.

[0090] At the hardware and low-level control levels, to ensure the precise execution of control commands, this invention preferably employs specific communication protocols and control algorithms. As mentioned earlier, the edge control layer 20 and the field execution layer 30 communicate via ProfinetIRT real-time industrial Ethernet. For closed-loop control, PID (proportional-integral-derivative) control algorithms or their variants, such as cascade PID control algorithms, are widely used. For example, in the drying process, a cascade PID control strategy can be adopted: the outer loop uses the regional average temperature measured by a distributed temperature and humidity array as the control target, outputting the heater power setpoint required by the inner loop; the inner loop uses the actual heater power as the control target, rapidly adjusting the power output. This cascade control can effectively suppress the influence of internal disturbances such as grid voltage fluctuations on temperature control, achieving high-precision temperature and humidity curve tracking. Experiments show that using this control strategy and combining it with parameter-level traceability data for data-driven optimization of PID parameters can reduce the standard deviation of workpiece wall thickness from 0.21 mm to 0.06 mm, a reduction of 71.4%.

[0091] It is important to note that integrating PID closed-loop process control with real-time motion control of multiple robots into a single three-layer hierarchical architecture presents significant technical challenges and is not simply a combination of the two. Specifically: First, in the three-layer architecture of this invention, the edge control layer simultaneously undertakes robot motion planning (millisecond-level state monitoring and collision detection) and data analysis and aggregation tasks, while the PID closed-loop control of the field execution layer requires sensor sampling, deviation calculation, and output adjustment to be completed within a fixed period. Improper design of the competition for computational resources and data flow coupling between the two can lead to increased motion planning delays, thereby increasing the risk of collisions. This invention achieves spatiotemporal decoupling of motion control and process control by completely integrating the PID control loop into the local controller of the field execution layer, with the edge layer only receiving aggregated process parameter data without participating in real-time PID calculation. Second, there is an implicit mutual constraint relationship between robot motion planning and process variable control. For example, the motion path and immersion speed of the robot performing the slurry application directly affect the degree of slurry surface fluctuation, thus affecting the reading of the level sensor and the stability of the level PID control. If designed independently, the robot's pursuit of maximum speed will lead to excessive surface disturbance, PID control overshoot, and even oscillation. This invention introduces process-constrained motion planning into the resource allocation algorithm—when generating the robot's action time window, the response time constant of the process equipment is used as a boundary condition to ensure that the robot's action rate does not exceed the stability margin of the process equipment, thus enabling harmonious coexistence between robot motion and process control.

[0092] The discovery and resolution of the aforementioned technical problems demonstrate the creative contribution of this invention to multi-domain collaborative control under the three-layer architecture, far exceeding the scope of conventional techniques that use PID control alone in a single robot system.

[0093] The solution to this technical challenge, along with the synergistic gains from PID closed-loop control and RFID full-parameter traceability mentioned in the previous section on beneficial effects—which, based on the separate introduction of PID closed-loop control (yield rate of 92.0%) and parameter-level RFID traceability (yield rate of 84.3%), achieves a yield rate of 96.8% when combined, resulting in an additional synergistic improvement of 2.7 percentage points—mutually corroborates this. It is precisely because the problems of computing resource competition and motion-process mutual constraints are solved at the three-layer architecture level that PID closed-loop control can effectively collaborate with the traceability system, producing a superimposed effect that cannot be achieved by individual deployments.

[0094] This invention also provides a multi-robot collaborative control system for shell-making production lines, configured to implement the above-described method. For example... Figure 1 As shown, the system includes a cloud collaboration layer 10, an edge control layer 20, and a field execution layer 30, configured as a three-tiered hierarchical control architecture. The cloud collaboration layer 10 is configured as a virtual simulation model 11 based on the production line to simulate and optimize production scheduling schemes. The edge control layer 20 is configured to dynamically allocate conflict-free spatiotemporal resources to multiple robots 31 based on the received production scheduling scheme through its internal multi-robot collaborative scheduling module 21. The field execution layer 30 includes various sensors (such as laser rangefinders, flow sensors, temperature and humidity sensors, RFID readers 32, etc.) and actuators (such as robot servo motors, slurry pumps, frequency converters, etc.), configured to monitor key process variables in real time through sensors and automatically adjust the actuators using a closed-loop control algorithm. Simultaneously, the field execution layer 30 is also configured to associate and record the measured values ​​of process parameters, including timestamps, measured by sensors with the unique identifier of the currently processed workpiece, forming a parameter time series.

[0095] The invention will be described in more detail below through a basic embodiment. In this embodiment, the production line is used only for producing mold shells of a single product type. First, in the cloud collaboration layer 10, workers use a preset virtual simulation model to simulate the shell-making process for this product type. The cloud collaboration layer 10 performs virtual operation based on at least two candidate production scheduling schemes, and calculates the predicted cycle time, robot waiting time, and path conflict risk for each candidate production scheduling scheme; then, it selects the candidate production scheduling scheme with the shorter predicted cycle time and path conflict risk below a preset threshold as the optimized production scheduling scheme, and sends the optimized production scheduling scheme to the edge control layer 20. This scheme only includes basic process sequence and time estimation. The scheme is sent to the edge control layer 20. After receiving the scheme, the edge control layer 20 allocates spatiotemporal resources to the robots based on a simplified workspace model (e.g., setting a fixed cuboid bounding box as the activity area for each robot) through a basic resource allocation algorithm (e.g., a time-based fixed avoidance strategy, i.e., another robot can only enter after one robot has completely left the shared area). At the on-site execution layer 30, industrial robot 31 executes processes such as slurry application, sand application, and drying sequentially according to instructions issued by the edge control layer. During slurry application, a level sensor detects that the slurry level is lower than the set value. Upon receiving this deviation signal, a basic PID controller starts a slurry replenishment pump to compensate until the level returns to the set range. At each workstation, an RFID reader 32 reads the RFID tag on the mold shell to confirm its identity and associates the key parameters of the current process (such as drying start time and drying temperature) with the mold shell's unique identifier (UUID), recording them to form a parameter time series. This data is stored in a local database at the edge control layer for basic production records. The edge layer does not perform real-time dynamic rescheduling; it does not use multi-objective optimization algorithms; the data collected on-site is stored locally and not transmitted back to the cloud for calibrating the virtual simulation model; the data recording format is also a simple database table, rather than a structured JSON data packet.

[0096] Building upon the basic embodiments described above, the present invention may further incorporate optimized technical features to enhance system performance and intelligence. In one progressive embodiment, the system emphasizes enhanced digital twin and dynamic scheduling capabilities. For example... Figure 4As shown, the virtual simulation model 11 of the cloud collaboration layer 10 is a high-fidelity digital twin model, which includes not only the geometric model of the equipment, but also the robot kinematic and dynamic models based on DH parameters. During resource allocation, the multi-robot collaborative scheduling module 21 of the edge control layer 20 employs a precise collision detection algorithm based on three-dimensional convex hulls, enabling the planning of more compact and efficient robot motion paths. After the virtual simulation-optimized scheduling scheme 51 is issued, the edge control layer 20 closely monitors the robot's actual execution status at 10ms intervals. To illustrate the effect of dynamic rescheduling, a scenario can be simulated: an industrial robot 31 responsible for sand application experiences a 50ms delay in its grasping action due to air pressure fluctuations. The edge control layer 20 detects this deviation within 10ms using encoder data. Since this deviation does not exceed the 500ms rescheduling threshold, the system may only perform local time window adjustments. If another delay reaches 550ms, the system immediately determines it as a significant execution deviation and triggers global rescheduling. The multi-robot collaborative scheduling module 21 will recalculate the time windows and paths for all subsequent tasks based on the current actual positions and speeds of all robots within a very short time (e.g., within 200ms), and issue new instructions. Simultaneously, all this actual operational data 61, including execution deviations and rescheduling records, will be transmitted back to the cloud-based collaborative layer 10. The cloud-based data analysis module will use this data, through machine learning methods such as gradient descent, to automatically calibrate the time estimation model parameters in the virtual simulation model 11, making its predictions for similar working conditions increasingly accurate.

[0097] In another progressive embodiment, the system focuses on achieving efficient mixed-line production and refined quality traceability. This scenario simulates the mixed production of three different process paths for mold shells (e.g., requiring 3, 6, and 8 layers of coating, respectively) on the same production line. When the RFID reader 32 at the loading station detects a new type of mold shell M1, the dynamic production scheduling module of the edge control layer 20 immediately triggers a rescheduling. At this time, a multi-objective optimization algorithm (such as the NSGA-II genetic algorithm) is activated, aiming to minimize total completion time and maximize equipment utilization as dual optimization objectives. Under the condition of satisfying the process constraints of each product, it quickly searches for the optimal production deployment sequence. For example, the algorithm might decide to "insert" a short-cycle mold shell with 3 layers of coating into the long drying waiting interval of a long-cycle mold shell with 8 layers of coating, thereby achieving a high degree of reuse of spatiotemporal resources. Experimental comparison data shows that compared to a simple FIFO strategy, this method can reduce the total completion time by 27.8%. Simultaneously, the quality traceability management module also works synchronously. Figure 3As shown, mold shell M1 is initialized and archived at the loading station (S410), with its unique identifier (UUID) written into it. Subsequently, in each process, such as the slurry dipping process (S420) and the sand pouring process (S430), the measured values ​​of relevant process parameters (such as slurry temperature, immersion depth, residence time, sand flow rate, etc.) along with timestamps are continuously appended and recorded to the data file associated with this UUID. When the mold shell completes all processes, all fragmented data is aggregated, packaged into a structured JSON data packet, and uploaded to the cloud. Figure 5 As shown, the JSON packet has a clear structure. The top layer contains metadata such as "shell_uuid", and below it is a "process_log" array. Each element in the array represents a process step, which records the "step", "timestamp" and a "parameters" object containing the measured values ​​of all key parameters.

[0098] Another progressive embodiment focuses on the details of hardware and low-level control. In this embodiment, the edge control layer (e.g., an Advantech industrial PC) is connected to multiple KUKA or FANUC robot controllers and sensor I / O modules in the field via a Profinet IRT bus to form a ring or star topology network. The Profinet IRT time synchronization mechanism ensures that the clock error of all devices in the network is within 1μs, which provides a guarantee for the edge control layer to issue coordinated commands with precise start execution times. Taking the drying process in the shell-making process as an example, the temperature control in the three-stage drying line adopts a cascade PID control strategy. The outer loop PID controller reads the average temperature measured by three distributed temperature and humidity sensors in each temperature zone at a 1-second cycle and compares it with the target temperature set in the process recipe (e.g., 22°C for the first stage and 25°C for the second stage). Its output serves as the power setpoint for the inner loop PID controller. The inner loop PID controller monitors the actual output power of the heater at a faster frequency (e.g., 100ms cycle) and quickly adjusts the conduction angle of the thyristor to accurately track the power setpoint given by the outer loop. This control strategy effectively isolates the temperature field from disturbances such as grid voltage fluctuations, achieving high-precision temperature control of ±0.5℃. PID parameter tuning can be optimized using a data-driven approach based on historical data to achieve the best control performance.

[0099] To comprehensively demonstrate the technical solution of this invention, a complete embodiment incorporating all the aforementioned technical features is described below. The system is configured as follows: the cloud collaboration layer 10 is deployed on a microservice cluster of Alibaba Cloud (or AWS); the edge control layer 20 uses an Advantech industrial control computer equipped with a real-time Linux kernel and a CODESYS soft PLC environment; the field execution layer 30 includes multiple KUKA and FANUC six-axis industrial robots 31, equipped with SICK laser rangefinders, IFM flow sensors, distributed temperature and humidity sensor arrays, and an Impinj UHF RFID system, including RFID readers 32 deployed at each key workstation. Now, taking a complex new product introduction and mixed-line production order as an example, the entire collaborative workflow of this system is demonstrated. First, the production planner inputs a mixed-line production order containing a new product and two regular products into the cloud MES interface module. The virtual simulation model 11 of the cloud collaboration layer 10 is immediately launched. Based on a high-fidelity digital twin model, it performs rapid virtual simulation of the process flow of the new product (50 times faster than real-time simulation), completing hundreds of iterations within minutes, automatically optimizing and generating an initial production scheduling scheme with zero collision risk and optimal predicted cycle time. This optimized production scheduling scheme is then sent to the edge control layer 20. Upon receiving the scheme, the edge control layer 20's dynamic production scheduling module initiates a multi-objective optimization algorithm (such as the NSGA-II algorithm), combining it with the actual tasks on the current production line to seamlessly integrate the new product into the existing production sequence, generating an optimal mixed-line production plan. The multi-robot collaborative scheduling module 21 then begins executing the plan, allocating millisecond-precise motion time windows and conflict-free motion paths to each robot 31 based on the DH+ convex hull model, and issuing instructions through the Profinet IRT network. During operation, a robot experiences a slight delay due to unforeseen circumstances, which is captured by the 10ms periodic status monitoring system and triggers dynamic rescheduling. The system recovers normal collaborative cycle time within 200ms. Meanwhile, in each process stage, such as slurry application and drying, the cascaded PID closed-loop control algorithm, based on sensor feedback, precisely maintains key process parameters such as slurry level and drying temperature and humidity within ±1% of the set value. Throughout the process, the RFID tag on each mold shell M1 serves as its unique digital identity. At each workstation, the RFID reader 32 at that workstation writes all measured values ​​of the current process parameters (temperature, time, flow rate, etc.) along with a timestamp into a data file associated with the mold shell's UUID. After production is completed, a complete, structured JSON data package is generated and uploaded to the cloud. This massive amount of actual operational data not only provides data records for quality traceability analysis but is also continuously used to provide feedback and calibration for the virtual simulation model in the cloud, continuously improving its simulation accuracy.

[0100] The core advantage of this invention lies in the powerful synergistic effect generated between its various inventive points. For example, in terms of quality improvement, experimental data shows that introducing PID closed-loop control alone (Mode B) or introducing parameter-level RFID traceability alone (Mode C) increases the yield to 92.0% and 84.3%, respectively. When the two are combined (Mode D, the solution of this invention), the yield reaches 96.8%. By calculating (96.8% - 83.5%) > (92.0% - 83.5%) + (84.3% - 83.5%), a synergistic gain of 2.7 percentage points can be found. The mechanism is that the parameter time series traceability system stores and analyzes a large amount of instantaneous data generated by sensors during the closed-loop control process in a structured manner, transforming the tuning of PID parameters from a "trial and error" method relying on engineer experience to a "data-driven" method based on historical data statistical optimization, thereby improving the control effect. From a broader perspective, the virtual simulation model pre-scheduling provides a high-quality "warm start" initial solution for real-time collaborative scheduling at the edge layer, enabling it to converge to the optimal schedule more quickly. Meanwhile, the actual operational data accumulated at the edge layer continuously calibrates the virtual simulation model, improving its simulation accuracy. The "variety-parameter-quality" correlation revealed by traceability data can be input into the virtual simulation model, upgrading its simulation capability from simply "verifying whether production is possible" to "predicting production quality." Ultimately, these three key inventions collectively construct a continuously self-optimizing flywheel: "planning (virtual simulation model pre-scheduling) - execution (edge ​​collaborative scheduling and closed-loop control) - recording (RFID full-parameter traceability) - learning (data feedback to optimize the model and parameters)." In a 30-day comprehensive verification experiment, the effective output (number of qualified products) of the complete solution of this invention increased by 123.7% compared to existing solutions (from 5,010 groups to 11,209 groups). Among them, after quantitative decomposition, the synergistic gain generated by the interaction of the three major invention points contributed 41.9% of the total increment. This fully proves that the invention is not a simple superposition of technologies, but rather a superposition of technologies through architectural innovation and deep integration, which produces a superposition effect of multi-module synergy and realizes an unprecedented continuous self-optimization capability of the production system.

Claims

1. A multi-robot collaborative control method for shell manufacturing production lines, applied to an industrial automation production line with a three-layer hierarchical control architecture including a cloud collaboration layer, an edge control layer, and a field execution layer, wherein the field execution layer includes multiple robots, sensors, and actuators, characterized in that, The method includes: In the cloud-based collaborative layer, based on the virtual simulation model of the production line, the production scheduling scheme is simulated, and the production scheduling scheme is optimized according to the simulation output, which includes at least the predicted total completion time, equipment utilization rate, and collision risk value. In the edge control layer, based on the optimized production scheduling scheme, a resource allocation algorithm is used to dynamically allocate conflict-free spatiotemporal resources, including action time windows and motion paths, to the multiple robots in the shared workspace for performing processing tasks. In the field execution layer, the sensor monitors key process variables in real time to obtain the measured values ​​of process parameters, and based on the deviation between the measured values ​​of process parameters and the preset set values, a closed-loop control algorithm is used to automatically adjust the actuator. In the field execution layer, the measured values ​​of the process parameters containing timestamps are associated with and recorded with the unique identifier of the currently processed workpiece to form a parameter time series; In the edge control layer, the actual execution status of the multiple robots is monitored at a preset monitoring frequency. When the deviation between the actual execution status and the planned status determined based on the conflict-free spatiotemporal resources exceeds a preset threshold, local rescheduling or global rescheduling is performed according to the deviation type, and the action time window and motion path of the tasks that have not yet been executed in the processing task are regenerated. The actual operating data collected by the edge control layer and the field execution layer is transmitted back to the cloud collaboration layer. The cloud collaboration layer aligns the actual operating data with the predicted indicators output by the virtual simulation model according to the task number and timestamp of the processing task, calculates an error vector including trajectory error, process time error and key process variable error, and calibrates the robot kinematic correction parameters, dynamic parameters and process time estimation parameters in the virtual simulation model based on the error vector; wherein, the virtual simulation model is a digital twin model of the production line.

2. The method according to claim 1, characterized in that, The method further includes: sending the optimized production scheduling scheme from the cloud collaboration layer to the edge control layer.

3. The method according to claim 1, characterized in that, The method further includes: in the edge control layer, when a scheduling event is triggered, a genetic algorithm is used to solve the production sequence of multi-product mixed-line production in real time under multiple constraints, with the total completion time and equipment utilization rate as optimization objectives; wherein, the chromosome of the genetic algorithm includes a first encoding segment for representing the workpiece task placement order and a second encoding segment for representing the robot allocation result, and the multiple constraints include process path sequence constraints, equipment capacity constraints, robot reachability constraints, shortest drying time constraints, and shared workspace conflict constraints.

4. The method according to claim 1, characterized in that, The resource allocation algorithm allocates resources based on the digital workspace model of each robot. The digital workspace model includes a robot kinematic model established based on DH parameters and a three-dimensional convex hull model of the robot links. The algorithm determines whether there is a spatial conflict by judging whether the distance between the three-dimensional convex hull models of any two robots is less than a preset safety distance.

5. The method according to claim 1, characterized in that, The method further includes: aggregating all parameter time series data associated with the unique identifier of a single workpiece, generating a structured JSON format data packet, and uploading it to the cloud collaboration layer; wherein the JSON format data packet includes at least the workpiece unique identifier, process record array, workstation number, equipment number, measured parameter value, collection timestamp, and data validity flag.

6. The method according to claim 1, characterized in that, The edge control layer and the field execution layer communicate using Profinet IRT real-time industrial Ethernet; and the closed-loop control algorithm is a PID control algorithm or a cascaded PID control algorithm.

7. A multi-robot collaborative control system for shell-making production lines, applied to an industrial automated production line employing a three-layer hierarchical control architecture, characterized in that... The system includes: The cloud collaboration layer, edge control layer, and field execution layer are configured as a three-layer hierarchical control architecture. in: The cloud-based collaboration layer is configured to simulate the production scheduling scheme based on the virtual simulation model of the production line, and optimize the production scheduling scheme based on the predicted indicators output by the simulation. The cloud collaboration layer is further configured to: transmit the actual operating data collected by the edge control layer and the field execution layer back to the cloud collaboration layer, and the cloud collaboration layer aligns the actual operating data with the predicted indicators output by the virtual simulation model according to the task number and timestamp of the processing task, calculates an error vector including trajectory error, process time error and key process variable error, and calibrates the robot kinematic correction parameters, dynamic parameters and process time estimation parameters in the virtual simulation model based on the error vector, wherein the virtual simulation model is a digital twin model of the production line; The edge control layer is configured to dynamically allocate conflict-free spatiotemporal resources to multiple robots in a shared workspace based on the optimized production scheduling scheme and through a resource allocation algorithm. The edge control layer is also configured to: monitor the actual execution status of the multiple robots at a preset monitoring frequency; when the deviation between the actual execution status and the planned status determined based on the conflict-free spatiotemporal resources exceeds a preset threshold, perform local rescheduling or global rescheduling according to the deviation type, and regenerate the action time window and motion path of the tasks that have not yet been executed in the processing task. The field execution layer, including sensors and actuators, is configured to monitor key process variables in real time through the sensors to obtain measured values ​​of process parameters, and to automatically adjust the actuators based on the deviation between the measured values ​​of process parameters and preset values ​​using a closed-loop control algorithm. The field execution layer is also configured to associate and record the measured values ​​of the process parameters containing timestamps with the unique identifier of the currently processed workpiece, forming a parameter time series.

Images