Multi-constraint multi-target real-time production scheduling and scheduling method and system for continuous charging production line of combustion chamber of solid engine

Abstract

The invention discloses a multi-constraint multi-target real-time production scheduling and scheduling method and system for a continuous charging production line of a solid engine combustion chamber, and the method comprises the steps: constructing a multi-agent reinforcement learning model for distributed collaborative decision-making through the collection and fusion of multi-source data based on an edge cloud digital twin collaborative architecture; and an NSGA III multi-objective evolution algorithm is introduced to carry out Pareto optimization on the candidate scheduling scheme, so that comprehensive optimization of the production cycle, the cost, the equipment utilization rate and the process safety is realized. A lightweight reasoning model is deployed on the edge side of the system, and rescheduling and self-repairing can be rapidly completed when disturbance events such as equipment faults and material delay occur. And virtual-real synchronization and visual feedback are realized through the digital twin platform. According to the method, the problems of dynamic response lag and multi-target balance difficulty of a traditional scheduling method are effectively solved, and the operation efficiency, the safety and the intelligent level of the production line are remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of production scheduling, and particularly relates to a multi-constraint multi-objective real-time production scheduling and scheduling method and system for a solid engine combustion chamber continuous charging production line. BACKGROUND

[0002] The continuous charging process of the solid engine combustion chamber is a typical high-risk, high-coupling and high-constraint production process, and the manufacturing process comprises the following steps: powder metering, mixing, feeding, pouring, solidification and detection.

[0003] The process has the following typical problems: (1) Resource constraint complexity: equipment capacity, personnel skills, process window and safety explosion-proof distance constitute multi-dimensional constraints, and a traditional static scheduling model cannot respond in real time.

[0004] (2) Multi-objective conflict of tasks: it is necessary to minimize the cycle and cost, maximize the equipment utilization rate and safety factor, the multi-objective function is non-convex, and the solution space dynamically changes.

[0005] (3) Disturbance and uncertainty: equipment failure, material lag, environmental fluctuations and process safety alarms often lead to task reordering, and traditional GA and SA algorithms are not real-time enough.

[0006] (4) System real-time decision bottleneck: existing MES / APS systems mostly rely on manual adjustment, and cannot realize intelligent decision and cloud-edge collaborative scheduling.

[0007] Therefore, there is an urgent need for a multi-constraint multi-objective intelligent scheduling system with self-learning, self-adaptation and real-time feedback capabilities to ensure the safety and stability of the continuous charging production in a dynamic and complex environment. SUMMARY

[0008] Therefore, the application provides a multi-constraint multi-objective real-time production scheduling and scheduling method and system for a solid engine combustion chamber continuous charging production line, which aims to solve the problems of dynamic disturbance response lag, multi-objective optimization balance difficulty, lack of edge intelligent collaborative mechanism and inability to realize cloud and field data closed loop in the existing continuous charging production scheduling, so as to realize intelligent scheduling of the solid engine continuous charging production process with high safety, low energy consumption and high efficiency.

[0009] To achieve the above-mentioned purpose, the application provides the following technical scheme: The application firstly proposes a multi-constraint multi-objective real-time production scheduling and scheduling method for a solid rocket engine combustion chamber continuous charging production line, which is realized based on an edge-cloud-digital twin collaborative architecture, wherein a multi-agent reinforcement learning model and an NSGA-III multi-objective evolutionary algorithm are deployed on the cloud, the edge node is realized based on Jetson NX to realize fast response, and data communication is realized through the OPC-UA protocol and the Kafka message middleware; the method comprises the following steps: Step one: system initialization and multi-source data acquisition Initialize the scheduling task set , the equipment resource set , the human resource set ; acquire real-time environmental state data and a safety constraint set , wherein the safety constraint set at least includes an anti-explosion safety distance , a maximum safety temperature and a maximum safety pressure ; the collected data is uploaded through the OPC-UA protocol and is asynchronously fused and transmitted through the Kafka message bus; Step two: multi-agent state construction Abstract various resources in the production line into independent agents ; define the state space of the first agent as: Wherein: is the resource load rate; is the device state vector; is the process progress; is the energy consumption state function; is the environmental risk factor; define the action space , wherein the action includes task receiving, task passing, energy consumption mode switching and path adjustment; Step three: multi-agent decision and interaction Each agent selects and executes an action based on its local observation state and global information based on a strategy network ; the global reward function is defined as: Wherein: represents the total cost; is the production cycle; is the equipment utilization rate; is the safety index; , ,​ and All weights are adaptive and are dynamically updated through a policy gradient mechanism. Step 4: NSGA-III Multi-Objective Evolutionary Optimization The population of candidate scheduling schemes generated by multi-agent reinforcement learning Input the NSGA-III multi-objective evolutionary algorithm for Pareto optimization; construct the fitness function. ,in: It is a production cost function; It is a production cycle function; This is a function for equipment utilization. For safety constraint penalty function; By using non-dominated sorting, reference point association, and crowding distance calculations, the merged populations were analyzed. Select the next generation population The final output is the Pareto front solution set. ; Step 5: Real-time rescheduling and execution at the edge Edge computing nodes continuously monitor production status and detect disturbance events. At the same time, identify the set of tasks affected. Edge nodes are based on local states. Initiate a fast rescheduling process, invoke the local reinforcement learning agent update strategy, and generate a local scheduling solution. If the local solution is better than the global solution in the cloud. This will trigger uplink synchronization. Upload to the cloud; Step Six: Digital Twin Synchronization and Visual Feedback: The finalized scheduling plan, including task sequences, workstation allocations, and process parameters, is synchronized in real time to the Manufacturing Execution System and the digital twin visualization platform. The digital twin model maps the physical production line to the real world, dynamically displaying equipment status, task Gantt charts, energy consumption curves, and safety risk heatmaps, and providing operators with interpretable analysis reports on scheduling decisions.

[0010] Furthermore, in step three, the multi-agent reinforcement learning employs an Actor-Critic framework for distributed training, with each agent maintaining a policy network. and a value network By employing a centralized training and distributed execution approach, the gradient ascent method is utilized. Update the policy network parameters, where: The learning rate is the policy learning rate.

[0011] Furthermore, in step three, the method for dynamically updating the adaptive weights through the policy gradient mechanism is as follows: in: and These are the adaptive weights before and after the update; The learning rate; This is the global reward function.

[0012] Furthermore, in step four, the specific execution process of the NSGA-III multi-objective evolutionary optimization includes: 41) Initialize the population Set the maximum number of iterations. ; 42) For each generation Calculate the fitness of each individual in the population. ; 43) Perform non-dominated sorting on the population to obtain the frontier hierarchy. ; 44) Normalize and correlate individuals within each frontier level based on the reference point; 45) Calculate the crowding distance of individuals and sort them by crowding within the same frontier level to maintain diversity; 46) Generating offspring populations by simulating binary crossover and polynomial mutation. ; 47) Merging parent and offspring populations ; 48) From Choose the optimal one Individuals constitute the next generation population. ; 49) Repeat steps 42) to 48) until... Output the Pareto optimal solution set .

[0013] Furthermore, the convergence of the hybrid optimization framework composed of multi-agent reinforcement learning and NSGA-III multi-objective evolutionary algorithm is defined as follows: for scheduling policy and Pareto Front If there exist finite constants , so that: , in: This represents the number of iterations. The optimal scheduling strategy; Pareto optimal solution set; This means the hybrid optimization framework converges to " Stable Pareto equilibrium.

[0014] Further, the iteration model of the hybrid optimization framework is described by the following coupled equations: Multi-agent reinforcement learning is used to dynamically generate candidate solution sets: Wherein: and represent the current iteration state vector and the next updated state vector, respectively; is the learning rate; is the reward function of reinforcement learning; is the policy gradient; is the policy function; The NSGA-III multi-objective evolutionary algorithm performs non-dominated sorting in the Pareto space, and the population update formula is: The hybrid optimization framework is coupled to obtain a composite evolutionary equation: Wherein: represents the comprehensive evolutionary solution set at the moment, the generation; is the population of the generation (parent); is the offspring population generated by crossover and mutation; is the composite evolutionary function of the hybrid optimization framework; is the environmental disturbance term; The hybrid optimization framework constructs a joint Lyapunov function and proves that its expected value decreases along the iteration, thereby guaranteeing the convergence of the hybrid optimization framework under the condition that .

[0015] Further, in step five, the model self-evolution and online update mechanism of cloud-edge collaboration includes: Knowledge distillation: compress the large global model parameters trained on the cloud into a lightweight parameter set through the knowledge distillation technology, and periodically issue it to the edge node to realize the knowledge transfer from the complex teacher network to the lightweight student network by minimizing the distillation loss; the distillation loss is represented as: Wherein: represents the policy output of the edge student network in state ; represents the policy output (i.e., action probability distribution or value prediction) of the cloud teacher network in state ; Online incremental learning: edge node based on newly collected local data Fine-tuning the model, the parameter update formula is: Wherein: is the fine-tuning learning rate; is the fine-tuning reward function; Confidence evaluation and trigger retraining: the system calculates the confidence of the model prediction output When the confidence is lower than the preset threshold , it is determined that the model may be invalid, and the retraining of the cloud model and the replacement process of the edge model are automatically triggered; wherein: represents the action output by the model in state ; represents the current system state.

[0016] Further, in the fifth step, the influence range of the disturbance event is dynamically identified by analyzing the task-device dependency graph and process timing constraints.

[0017] Further, in the sixth step, the explainability analysis report is generated by SHAP analysis or feature contribution calculation, which shows the contribution weight of input features to the final scheduling decision in the form of a bar chart , and visualizes the comprehensive risk level of different process nodes in the form of a heat map ; the natural language description part of the explainability report includes at least three structured text modules: "decision reason", "key impact features and their action direction", and "current comprehensive risk level", which are used to assist safety review and quality traceability.

[0018] The application also provides a system for implementing the solid rocket engine combustion chamber continuous charging production line edge-cloud collaborative multi-constraint multi-objective intelligent scheduling and scheduling method as described above, characterized in that: a three-layer collaborative architecture of edge-cloud-digital twin is adopted, including: A data acquisition and fusion module is deployed in the production field, including a PLC controller and temperature, pressure and position sensors, for real-time acquisition of device state, process parameters and environmental monitoring data; the data acquisition and fusion module has a built-in OPC-UA client for realizing standardized data access of heterogeneous devices, and through Kafka message middleware, the fused data stream is uploaded to the cloud and simultaneously to the digital twin visualization module; ​The cloud-based optimization scheduling module, deployed on a cloud server, runs a fusion model of multi-agent reinforcement learning and the NSGA-III multi-objective evolutionary algorithm. The multi-agent reinforcement learning submodule defines the agents, constructs the state and action space, and generates high-quality candidate scheduling schemes through distributed interactive learning. The NSGA-III multi-objective evolutionary optimization submodule receives the candidate schemes and solves for the Pareto optimal front solution set through non-dominated sorting, reference point association, and crowding distance calculation. This enables comprehensive optimization of production cycle, cost, equipment utilization, and safety. The edge inference and execution module, deployed on an embedded edge computing node based on Jetson Xavier NX, receives and executes scheduling instructions issued from the cloud. This module incorporates a lightweight reinforcement learning inference model for the rapid execution of locally scheduled tasks. Furthermore, the edge inference and execution module continuously monitors the production status, and when a disturbance event is detected by the data acquisition and fusion module... When the conditions are met, it can autonomously initiate the rescheduling process and generate a local optimized solution, while comparing its performance with the global solution in the cloud. It can also trigger uplink synchronization when the conditions are met. The communication and synchronization module uses the OPC-UA protocol to achieve real-time data acquisition and control command issuance at the device layer, uses the Kafka asynchronous message bus to achieve highly reliable, low-latency transmission and decoupling of massive data between the cloud and the edge, and uses the WebSocket protocol to achieve real-time bidirectional communication with the digital twin visualization module. The digital twin visualization module runs on the monitoring terminal and constructs a virtual model consistent with the physical production line. It receives real-time data from the communication and synchronization module, dynamically maps scheduling status, equipment utilization, energy consumption curves, and safety risks, and displays these data visually in the form of Gantt charts, heat maps, etc. The digital twin visualization module also integrates an interpretable analysis interface, which reveals the internal logic and key feature contributions of scheduling decisions through visual charts and natural language reports.

[0019] The beneficial effects of this invention are as follows: The present invention provides a method for intelligent scheduling and production line scheduling of solid rocket motor combustion chamber continuous charging production line with edge-cloud collaboration, multi-constraint, and multi-objective characteristics, as described above. By constructing an edge-cloud-digital twin collaborative architecture and integrating Multi-Agent Reinforcement Learning (MARL) and the NSGA-III multi-objective evolutionary algorithm, it achieves real-time, intelligent, and multi-objective optimization of solid rocket motor continuous charging production line scheduling under dynamic perturbations, achieving significant technical effects, as follows: (1) The scheduling efficiency and production efficiency are greatly improved: through the distributed real-time decision of MARL and the global Pareto optimization of NSGA-III, the production cycle, cost and equipment utilization are cooperatively optimized, the average production cycle is effectively shortened, the equipment utilization is improved, and high-efficiency production is realized; (2) Dynamic disturbance response and system robustness are enhanced: relying on the edge side fast rescheduling mechanism, the system can quickly self-repair under disturbances such as device failure and material delay, ensuring the continuity and stability of the production process; (3) The comprehensive energy consumption and production cost are significantly reduced: the agent realizes precise energy control through the energy consumption mode switching in the action space and the global energy consumption optimization target guidance, which can effectively reduce the comprehensive energy consumption and effectively control the production cost; (4) The process safety level is essentially improved: the safety constraints are integrated into the state space and the reward function, and real-time warning and visual traceability are realized through digital twinning, which can effectively reduce the safety alarm triggering rate and provide a solid safety guarantee for high-risk charging processes.

[0020] In summary, the present application effectively solves the industry problem that the traditional scheduling method cannot consider real-time, multi-objective optimization and high safety, and realizes the cooperative optimization of safety, efficiency, cost and energy consumption. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to make the purpose, technical scheme and beneficial effects of the present application clearer, the present application provides the following drawings for illustration: Figure 1 It is a general architecture diagram of the solid rocket engine combustion chamber continuous charging production line multi-constraint multi-objective real-time production scheduling and scheduling; Figure 2 It is a scheduling flowchart of MARL+NSGA-III; Figure 3 It is a real-time feedback schematic diagram of digital twinning; Figure 4 It is an algorithm convergence and stability verification curve; Figure 5 It is a cloud-edge collaborative self-evolution learning mechanism structure diagram; Figure 6 It is an explainability analysis interface schematic diagram; Figure 7 It is a comparison diagram of convergence curves of MARL+NSGA-III fusion algorithm. DETAILED DESCRIPTION

[0022] The present application will be further described below in combination with the drawings and specific embodiments, so that those skilled in the art can better understand the present application and implement it, but the embodiments are not limiting to the present application.

[0023] The solid engine combustion chamber continuous charging production line multi-constraint multi-objective real-time production scheduling and scheduling method of the embodiment is realized based on an edge-cloud-digital twin collaborative architecture. Specifically, in the embodiment, the edge-cloud-digital twin collaborative architecture based on the edge-cloud-digital twin collaborative architecture includes a data acquisition and fusion layer, a cloud optimization scheduling layer, an edge execution and feedback layer, and a digital twin visualization layer. The core innovation lies in that the traditional manufacturing execution system (MES), the Internet of Things communication protocol (OPC-UA), the asynchronous message bus (Kafka), and the multi-agent reinforcement learning algorithm (MARL) and the multi-objective evolutionary algorithm (NSGA-III) are deeply integrated through the hierarchical architecture to realize multi-objective adaptive optimization of production tasks and real-time response to dynamic disturbance. Among them, the cloud (cloud optimization scheduling layer) deploys the multi-agent reinforcement learning (MARL) model and the NSGA-III multi-objective evolutionary algorithm, and the edge node in the edge execution and feedback layer realizes fast response based on Jetson NX and realizes data communication through the OPC-UA protocol and the Kafka message middleware.

[0024] Specifically, the data acquisition and fusion layer acquires equipment operation data (temperature, pressure, torque, current, fault code, etc.), personnel task information, raw material state and environmental monitoring parameters through PLC and OPC-UA acquisition equipment. The data is uploaded to the cloud through the Kafka message bus and is synchronized to the digital twin model. The cloud optimization scheduling layer runs the MARL+NSGA-III algorithm on the cloud server to perform global task allocation and multi-objective evolutionary optimization, and periodically updates the global optimal solution set. In the edge execution and feedback layer, the edge node (Jetson NX) receives the cloud scheduling instructions, performs local scheduling and real-time re-planning. When a disturbance event (such as equipment downtime, material lag, temperature overrun) is detected, a local reinforcement learning agent is triggered for rapid rescheduling. In the digital twin visualization layer, the MES and the twin system dynamically display scheduling status, device utilization, risk heat map and other information through the WebSocket interface to realize virtual-real synchronization.

[0025] Specifically, in the embodiment, the solid engine combustion chamber continuous charging production line multi-constraint multi-objective real-time production scheduling and scheduling method includes the following steps.

[0026] Step 1: System initialization and multi-source data acquisition.

[0027] Initialize the scheduling task set , the equipment resource set , the human resource set ; acquire real-time environmental state data and the safety constraint set , the safety constraint set at least includes the anti-explosion safety distance , the maximum safety temperature and maximum safe pressure ; the collected data is uploaded through the OPC-UA protocol and asynchronously fused and transmitted through the Kafka message bus.

[0028] Step two: multi-agent state construction.

[0029] Various resources (devices, AGVs, personnel) in the production line are abstracted as independent agents ; the state space of the first agent is defined as: wherein: is the resource load rate; is the device state vector; is the process progress; is the energy consumption state function; is the environmental risk factor.

[0030] The state space is defined as : resource load, device state, process progress, environmental risk factor, wherein is the state of the first agent.

[0031] The action space is defined, and the action includes task receiving, task passing, energy consumption mode switching, and path adjustment. The action space is defined as : including task receiving, task passing, energy consumption mode switching, path adjustment, wherein is the action.

[0032] State transition function: wherein: is the current system state; is the action output by the model in state 𝑡 ; is the disturbance event.

[0033] Step three: multi-agent decision-making and interaction.

[0034] Each agent selects and executes an action based on its local observation state and global information based on the policy network . Specifically, the global reward function is defined as: wherein: represents the total cost, and , represents energy consumption cost, represents material cost, represents labor cost, represents equipment occupancy cost; is a production cycle, and , represents end time, represents start time; is equipment utilization, and , represents actual processing or running time of equipment; represents total time of equipment available period; is a safety index, and , represents risk sensitivity coefficient, represents risk factor in current state; , , and are adaptive weights, which are dynamically updated through a policy gradient mechanism. Specifically, the method for dynamically updating the adaptive weights through the policy gradient mechanism is as follows: wherein: and are the adaptive weights before and after updating, respectively; is a learning rate; is a global reward function.

[0035] In this embodiment, the multi-agent reinforcement learning adopts an Actor-Critic framework for distributed training, each agent maintains a policy network and a value network ; through centralized training and distributed execution, the policy network parameters are updated using the gradient ascent method , wherein: is a policy learning rate.

[0036] Step four: NSGA-III multi-objective evolutionary optimization.

[0037] The candidate scheduling scheme population generated by the multi-agent reinforcement learning is input into the NSGA-III multi-objective evolutionary algorithm for Pareto optimization; a fitness function is constructed, wherein: is a production cost function; is a production cycle function; is an equipment utilization function; is a safety constraint penalty function. Through non-dominated sorting, reference point association and crowded distance calculation, the merged population Select the next generation population The final output is the Pareto front solution set. .

[0038] Specifically, in this embodiment, the specific execution process of NSGA-III multi-objective evolutionary optimization includes: 41) Initialize the population Set the maximum number of iterations. ; 42) For each generation Calculate the fitness of each individual in the population. ; 43) Perform non-dominated sorting on the population to obtain the frontier hierarchy. ; 44) Normalize and correlate individuals within each frontier level based on the reference point; 45) Calculate the crowding distance of individuals and sort them by crowding within the same frontier level to maintain diversity; 46) Generating offspring populations by simulating binary crossover and polynomial mutation. ; 47) Merging parent and offspring populations ; 48) From Choose the optimal one Individuals constitute the next generation population. ; 49) Repeat steps 42) to 48) until... Output the Pareto optimal solution set .

[0039] Step 5: Real-time rescheduling and execution at the edge.

[0040] Edge computing nodes continuously monitor production status and detect disturbance events. At the same time, identify the set of tasks affected. Edge nodes are based on local states. Initiate a fast rescheduling process, invoke the local reinforcement learning agent update strategy, and generate a local scheduling solution. If the local solution is better than the global solution in the cloud. This will trigger uplink synchronization. Uploaded to the cloud. Specifically, in this embodiment, the disturbance event... Scope of influence Dynamic identification is achieved by analyzing the task-equipment dependency graph and process timing constraints.

[0041] Step Six: Digital Twin Synchronization and Visual Feedback: The final scheduling scheme, including task sequence, workstation allocation and process parameters, is synchronized in real time to the manufacturing execution system and the digital twin visualization platform; the digital twin model maps the physical production line virtually and physically, dynamically displays equipment status, task Gantt chart, energy consumption curve and safety risk heat map, and provides interpretable analysis report for operator scheduling decisions.

[0042] In this embodiment, the interpretable analysis report is generated by SHAP analysis or feature contribution calculation, showing the contribution weight of input features to the final scheduling decision in the form of a bar chart , and visualizing the comprehensive risk level of different process nodes in the form of a heat map ; the natural language description part of the interpretable report includes at least three structured text modules: "decision reason", "key impact features and their direction of action", and "current comprehensive risk level", which are used to assist safety review and quality traceability.

[0043] Specifically, the hybrid optimization framework composed of multi-agent reinforcement learning and NSGA-III multi-objective evolutionary algorithm has the following convergence definition: for scheduling strategy and Pareto frontier , if there exists a finite constant such that: , then the hybrid optimization framework is said to converge to "stable Pareto equilibrium".

[0044] where: is the number of iterations; is the optimal scheduling strategy; is the Pareto optimal solution set.

[0045] Further, the iterative model of the hybrid optimization framework is described by the following coupled equations: Multi-agent reinforcement learning is used to dynamically generate candidate solution set: where: and represent the current iteration state vector and the next updated state vector, respectively; is the learning rate; is the reward function of reinforcement learning; is the policy gradient; is the policy function; NSGA-III multi-objective evolutionary algorithm performs non-dominated sorting in the Pareto space, and the population update formula is: ​The hybrid optimization framework is coupled to obtain a composite evolution equation: wherein: denotes the comprehensive evolution solution set of the i-th generation at the j-th time instant; denotes the i-th generation population (parent generation); denotes the offspring population generated by crossover and mutation; is a composite evolution function of the hybrid optimization framework; is an environmental disturbance term; The hybrid optimization framework guarantees convergence under the condition that a joint Lyapunov function is constructed and its expected value is proved to decrease along the iteration. The hybrid optimization framework guarantees convergence under the condition that a joint Lyapunov function is constructed and its expected value is proved to decrease along the iteration.

[0046] The multi-constraint multi-objective real-time production scheduling and scheduling method of the solid rocket engine combustion chamber continuous charging production line of the embodiment further includes a cloud edge collaborative model self-evolution and online updating mechanism, including knowledge distillation, online incremental learning, and credibility evaluation and triggering retraining.

[0047] Knowledge distillation: the parameters of a large global model trained in the cloud are compressed into a lightweight parameter set through a knowledge distillation technique and periodically issued to the edge node, and knowledge transfer from a complex teacher network to a lightweight student network is achieved by minimizing the distillation loss; the distillation loss is represented as: wherein: denotes the policy output of the edge student network in state ; denotes the policy output (i.e., action probability distribution or value prediction) of the cloud teacher network in state ; and denotes the policy output of the edge student network in state .

[0048] Online incremental learning: the edge node fine-tunes the model based on newly collected local data , and the parameter update formula is: wherein: is a fine-tuning learning rate; is a fine-tuning reward function.

[0049] Credibility evaluation and triggering retraining: the system calculates the confidence of the model prediction output , and when the confidence is lower than a preset threshold , it is determined that the model may fail, and the retraining of the cloud model and the replacement of the edge model are automatically triggered; wherein:​​​​​ Indicates the model in state The action to be output; Indicates the current system status.

[0050] Through the cloud-edge collaborative model self-evolution and online update mechanism, the system can automatically recover optimal performance after equipment replacement, batch switching or environmental drift, achieving true "self-learning and self-evolution" intelligent scheduling.

[0051] The multi-constraint, multi-objective real-time production scheduling and scheduling method for the solid rocket motor combustor continuous charging production line in this embodiment also includes a safety fault tolerance and interpretability mechanism. Considering the high-risk characteristics of the solid rocket motor charging process, this embodiment constructs a "multi-layer safety fault tolerance-interpretable decision mechanism," including multi-layer safety fault tolerance strategies, safety index definitions, and interpretability visualization modules to ensure the safety and auditability of the algorithm output.

[0052] Specifically, the multi-layered security fault tolerance strategy achieves dynamic security control through three layers of response logic: Level 1 (Local Limitation): Automatically reduces equipment load when temperature or pressure exceeds the warning threshold; Level 2 (Global Readjustment): If the risk index This immediately triggers edge rescheduling; Level 3 (Manual Control): Automatically switches to manual supervision mode after 3 consecutive security incidents.

[0053] The safety index is defined as: in: and These are the temperature and pressure regulation coefficients, respectively. This refers to the actual temperature. For safe temperature; For actual pressure; For safety pressures.

[0054] when An alarm is triggered at any time.

[0055] In the interpretability visualization module, a new "Schedule Interpretable Analysis Graph" module has been added, such as... Figure 6 As shown, the output displays the contribution of key features to each scheduling decision: showing the proportion of input features (equipment load, process temperature, operator status, energy consumption, etc.) to the final scheduling decision weight. Through feature sensitivity and SHAP analysis, the system can provide a "scheduling decision reason report" to facilitate safety review and quality traceability.

[0056] The embodiment also provides a system for implementing the multi-constraint multi-objective intelligent scheduling and scheduling method of the edge-cloud collaborative solid engine combustion chamber continuous charging production line as above, and the system is characterized in that: an edge-cloud-digital twin three-layer collaborative architecture is adopted, and the system comprises a data acquisition and fusion module, a cloud end optimization scheduling module, an edge reasoning and execution module, a communication and synchronization module and a digital twin visualization module.

[0057] The data acquisition and fusion module is arranged at a production site, comprises a PLC controller and temperature, pressure and position sensors, and is used for acquiring device states, process parameters and environment monitoring data in real time; the data acquisition and fusion module is internally provided with an OPC-UA client, is used for realizing standardized data access of heterogeneous devices, and is used for uploading the fused data stream to the cloud end and the digital twin visualization module in an asynchronous mode through a Kafka message middleware.

[0058] The cloud end optimization scheduling module is arranged at a cloud end server, is used for running a fusion model of a multi-agent reinforcement learning and an NSGA-III multi-objective evolution algorithm; wherein the multi-agent reinforcement learning submodule is responsible for defining an agent, constructing a state and action space, and generating a high-quality candidate scheduling scheme through distributed interactive learning; the NSGA-III multi-objective evolution optimization submodule receives the candidate scheme, solves a Pareto optimal front solution set through non-dominated sorting, reference point association and congestion distance calculation, and realizes global comprehensive optimization of a production cycle, a cost, a device utilization rate and safety.

[0059] The edge reasoning and execution module is arranged on an embedded edge computing node based on a Jetson Xavier NX, receives and executes the scheduling instruction issued by the cloud end; the edge reasoning and execution module is internally provided with a lightweight reinforcement learning reasoning model, and is used for fast execution of local scheduling tasks; the edge reasoning and execution module continuously monitors a production state, and when a disturbance event is detected through the data acquisition and fusion module , can autonomously start a rescheduling process and generate a local optimization solution, and compares the local optimization solution with a global solution of the cloud end, and when a condition is met, triggers uplink synchronization.

[0060] The communication and synchronization module realizes real-time data acquisition and control instruction issuing of a device layer based on an OPC-UA protocol, realizes high-reliability and low-delay transmission and decoupling of massive data between the cloud end and the edge end based on a Kafka asynchronous message bus, and realizes real-time bidirectional communication with the digital twin visualization module based on a WebSocket protocol.

[0061] ​The digital twin visualization module runs on a monitoring terminal, constructs a virtual model consistent with the physical production line; the digital twin visualization module receives real-time data from the communication and synchronization module, dynamically maps the scheduling state, device utilization, energy consumption curve and safety risk, and visualizes in the form of Gantt chart, heat map, etc.; the digital twin visualization module also integrates an explainability analysis interface for revealing the internal logic of scheduling decisions and the contribution degree of key features in the form of visual charts and natural language reports.

[0062] The specific embodiments of the method and system for implementing the multi-constraint multi-objective intelligent scheduling and scheduling of the solid engine combustion chamber continuous charging production line as above are further described below.

[0063] Figure 1 The overall structure of the multi-constraint multi-objective real-time production scheduling and scheduling system of the solid engine combustion chamber continuous charging production line proposed by the present application is shown, which adopts a three-layer structure mode of "edge-cloud-twin collaboration" as a whole, which is composed of a data acquisition and fusion layer, a cloud optimization scheduling layer, an edge execution and feedback layer, and a digital twin visualization layer. The core innovation lies in the deep integration of traditional manufacturing execution system (MES), Internet of Things communication protocol (OPC-UA), asynchronous message bus (Kafka), multi-agent reinforcement learning algorithm (MARL) and multi-objective evolutionary algorithm (NSGA-III) through hierarchical architecture, realizing multi-objective adaptive optimization and dynamic disturbance real-time response of production tasks.

[0064] In Figure 1 The bottom "data acquisition and fusion layer" mainly includes various sensors, PLC acquisition modules and edge nodes. This layer collects real-time key process parameters in the mixing, pouring and curing processes, including temperature, pressure, viscosity, flow, energy consumption and environmental indicators. The collected data is transmitted to the upper layer system through the OPC-UA protocol to realize standardized communication between heterogeneous devices. Kafka, as a message middleware, asynchronously buffers and concurrently transmits different sources of data streams, thereby ensuring low latency while achieving data consistency and high throughput.

[0065] The middle layer "edge execution and feedback layer" is built based on industrial edge computing nodes (such as Jetson NX) and deploys lightweight scheduling inference models. Its main responsibility is to receive the optimal task sequence and device instructions issued by the cloud scheduling center, and make local adaptive adjustments according to actual production disturbances (such as device failure, task delay, temperature drift). When a disturbance event is detected, the edge agent will trigger a fast rescheduling mechanism to reassign task priorities and return the corrected results within 500 milliseconds, thereby realizing a localized decision-making loop.

[0066] The upper layer "cloud optimization scheduling layer" is the intelligent core of the system, which integrates multi-agent reinforcement learning (MARL) and multi-objective Pareto evolutionary solution (NSGA-III) algorithm structure. Through the analysis of the state data uploaded from the edge, the layer dynamically updates the scheduling strategy network parameters and learns the optimal strategy of resource allocation. The multi-agent structure allows the device agent, personnel agent and material agent to learn in cooperation and competition, and finally achieve multi-objective comprehensive optimization. NSGA-III is used to sort the candidate solution set generated by reinforcement learning and construct the Pareto front, ensuring that the cost, cycle, energy consumption, safety and other objectives are balanced and optimal in a global sense.

[0067] In the "digital twin visualization layer" on the right, the MES system and the virtual simulation platform are connected through the WebSocket interface to realize real-time task mapping, energy consumption visualization and safety risk warning. This layer synchronizes the actual production process with the virtual model, which can intuitively display the dynamic state of each device, workstation, personnel and task, and form a complete production situation map. When the virtual model detects deviation or abnormal state, it can send correction instructions to the edge node through Kafka in reverse, realizing the closed-loop control of "virtual-real integration".

[0068] The whole system architecture realizes multi-layer interconnection through three types of communication channels: OPC-UA, Kafka and WebSocket, which not only ensures the stability and low delay of data transmission, but also keeps the strategy and state of each layer agent consistent. Unlike the traditional MES-APS linear structure, this architecture has scalability and self-learning ability, and can quickly adapt to production line expansion, task batch changes, device addition, etc. It is especially suitable for solid rocket engine continuous charging, which is a high-risk, strongly coupled and high-security process scenario. In summary, Figure 1 Not only reflects the hierarchical communication structure and functional module logic of the system, but also embodies the cooperative mechanism of edge reasoning and cloud global optimization, which is the overall framework basis for realizing multi-constrained and multi-objective real-time scheduling of the invention.

[0069] Figure 2 The real-time production scheduling process of the multi-agent reinforcement learning (MARL) and NSGA-III multi-objective evolutionary algorithm fusion of the invention is shown. The whole process has a structure combining series and cycles from top to bottom, which embodies the intelligent scheduling logic of the whole process from initialization, state update, reinforcement learning decision to Pareto optimization.

[0070] The process first enters the system initialization phase from the "start" module. The system generates an initial task set, resource set and environment parameter in this phase, and initializes the multi-agent network structure. Each agent corresponds to a key entity in the production line, such as a device, personnel or material unit. By reading the MES task library and real-time sensor data, the system establishes the initial mapping relationship between tasks and resources.

[0071] Then, the "population initialization" and "fitness calculation" steps are entered. At this time, the scheduling problem is mapped as a multi-objective optimization problem, and each feasible scheduling scheme is regarded as an individual in the population. The fitness function is composed of multi-objective indicators, including production cycle, energy consumption, equipment utilization, and safety risk, etc. The system calculates the fitness distribution of the initial population and performs preliminary screening.

[0072] Subsequently, in the "constructing state space scheme" link, the system uses the state modeling capability of reinforcement learning to abstract the production environment into a state space , and defines an action set . Each agent selects an action according to its own state to form a scheduling decision, such as task receiving, task surrendering, and device load adjustment, etc. Multi-agent collaboration learning is carried out through reward sharing mechanism to avoid system instability caused by resource competition.

[0073] The "multi-agent reinforcement learning decision" module is the core link. The algorithm adopts the Actor-Critic structure based on policy gradient, and updates the policy network parameters through continuous interaction and iteration. After each iteration, the system calculates the global reward is the i-th objective function, represents the total cost; is the production cycle; is the equipment utilization; is the safety indicator; , , and are adaptive weights, which are dynamically updated through the policy gradient mechanism. The reinforcement learning phase aims to obtain a high-quality candidate solution set to provide initial values for subsequent Pareto optimization.

[0074] After the reinforcement learning phase is completed, the system enters the "NSGA-III optimization phase". In this phase, the candidate solutions are sorted by non-dominance and calculated by congestion distance to generate the Pareto frontier. This algorithm maintains the diversity of the objective space through the reference point guiding mechanism to ensure the uniform distribution of the optimization solution. At this time, if the termination condition (such as the maximum number of iterations or the convergence threshold) is not met, return to the "update population" step; otherwise, output the final optimal scheduling scheme.

[0075] This process realizes the double-layer coupled optimization of MARL and NSGA-III: the reinforcement learning part is responsible for dynamic decision-making and policy generation, while the evolutionary algorithm part is responsible for global multi-objective balance. The combination of the two not only improves the search efficiency and robustness of the algorithm, but also ensures the diversity and global optimality of the solution space. Compared with traditional genetic algorithms or single-objective reinforcement learning, the hybrid algorithm of the present application can stably converge under complex constraints and dynamic disturbances, and support real-time applications. Therefore,Figure 2 The flowchart comprehensively embodies the algorithmic idea and execution logic of the present application, and is a core implementation basis of the intelligent scheduling mechanism of the system.

[0076] Figure 3 The digital twin real-time feedback mechanism in the system of the present embodiment is shown, which is used to realize virtual-real fusion, dynamic visualization and closed-loop scheduling feedback. The entire figure is composed of a "physical entity layer", a "data acquisition layer", a "virtual twin layer" and a "feedback control layer", reflecting the interaction between data flow, control flow and information flow in the actual production line and the virtual model.

[0077] The left side is the "production line" entity part, which includes the actual charging section composed of a mixing system, a pouring device, a mechanical arm, an AGV transportation unit, an operator and the like. This part is real-time collected by PLC and industrial sensors, such as stirring speed, barrel temperature, pouring flow, cavity pressure and energy consumption power, and the like. The collected data is transmitted to the data center through the OPC-UA protocol, and is buffered through the Kafka asynchronous queue to realize high reliability and low delay data communication.

[0078] The right side is the "digital twin platform", which is displayed on a computer or a monitoring screen. A virtual simulation model of the physical process is established inside, which is one-to-one mapped with the actual production line in space, time and process parameters. The virtual model is continuously updated in state through real-time data driving, and when a deviation from the real working condition is detected, an abnormality mark and an optimization request will be triggered. For example, when the pouring temperature exceeds the safe interval or the equipment utilization rate decreases, the virtual model sends correction instructions to the edge node through the feedback channel.

[0079] The middle part is the "data transmission and control feedback layer", which realizes real-time bidirectional data communication through OPC-UA. The uplink data is "working condition data", including the real-time state and performance indicators of each process; the downlink data is "control instruction", which is generated by the cloud scheduling center, transmitted through the Kafka bus and then issued to the device execution end. The digital twin system has "visualization" and "early warning" functions, and can display the device health, task progress, process deviation and energy consumption state in real time.

[0080] The "real-time feedback" arrow at the bottom indicates that when the cloud or virtual layer executes strategy update, the new scheduling result or safety correction parameter will be directly fed back to the physical layer device to realize closed-loop control. The closed loop not only includes scheduling instructions, but also parameter setting, energy consumption mode switching and priority rearrangement, and the like. The system can complete virtual verification and actual execution synchronization in milliseconds.

[0081] The digital twin feedback mechanism of the embodiment has three functions: 1) real-time mapping, realizing virtual-real synchronization of the production state; 2) prediction analysis, evaluating the effect of the scheduling scheme in the virtual space in advance; and 3) abnormal correction, correcting the actual operation parameters based on the twin model feedback. This mechanism greatly improves the observability and controllability of the system and is a key link to realize “adaptive intelligent scheduling”. Therefore, Figure 3 The digital twin real-time feedback schematic diagram comprehensively reflects the technical idea of realizing virtual-real integration at the intelligent manufacturing system level of the embodiment, and provides visual support and security for dynamic decision-making in complex manufacturing scenarios.

[0082] Figure 4 The convergence and stability verification results of the “multi-agent reinforcement learning (MARL) and multi-objective evolutionary algorithm (NSGA-III)” fusion optimization model proposed by the present application in the multi-objective scheduling solving process are shown. The vertical axis is the normalized Pareto fitness, and the horizontal axis is the iteration number . The blue solid line represents the algorithm of the present application, the gray dashed line represents the traditional GA-SA (genetic-simulated annealing) algorithm, and the shaded band represents the ±1σ standard deviation range, which describes the stability fluctuation interval of the algorithm under different random disturbances.

[0083] From Figure 4 It can be seen that the MARL+NSGA-III algorithm shows a faster growth rate in the early stage (0-50 generations), indicating that its convergence speed in the exploration period is significantly better than that of the traditional algorithm; in the middle stage (50-120 generations), it enters a stable climbing zone, and the curve fluctuation amplitude is significantly lower than that of the comparative algorithm; in the stable interval in the later stage (120-200 generations), the algorithm converges to the vicinity of the Pareto optimal frontier, and the fitness value is about 0.91, while the traditional algorithm only reaches 0.72. That is, in the embodiment, the first 50 rounds rise rapidly, and gradually stabilize in the later stage, indicating that the system still has global convergence under multi-objective. Through analysis, it is proved that the embodiment can converge to a stable Pareto frontier solution set within a limited number of iterations, meeting the timeliness and stability dual constraints of real-time scheduling requirements. Specifically, according to the statistical results, the convergence rate of the fusion algorithm is improved by about 68%, the stability variance is reduced by about 73%, and the final fitness value is improved by 24%, verifying the correctness of the theoretical convergence analysis.

[0084] Figure 5The cloud-edge collaborative self-evolution learning mechanism structure proposed in the embodiment is shown, which is the core of realizing the self-learning, self-adaptation and self-evolution ability of the scheduling system. The whole figure adopts a three-layer architecture: a cloud optimization center, an edge computing node and a device and sensor layer, and the updating of model parameters, knowledge distillation and credibility feedback are realized through bidirectional communication, thereby forming a self-circulating intelligent scheduling system. The theoretical basis is established on the basis of knowledge transfer learning, online incremental learning and confidence evaluation theory, which guarantees the convergence, robustness and evolvability of the algorithm in long-term operation. At the engineering implementation level, Figure 5 The cloud-edge collaborative architecture realizes data interaction through OPC-UA and Kafka communication buses. The cloud parameter delivery delay is less than 200 milliseconds, the edge node training interval is 0.5 seconds, and the system update period is 2 minutes. The experimental results show that the system can quickly recover the scheduling performance under the disturbance of device failure and batch replacement, with an average recovery time of 0.46 seconds and a rescheduling success rate of 98%. In the experiment of continuous operation for 72 hours, the model self-evolution mechanism makes the Pareto fitness value continuously rise from 0.83 to 0.91, reduces the energy consumption by 13%, and reduces the average scheduling delay by 27%, which fully proves the long-term effectiveness and theoretical feasibility of the mechanism. In summary, Figure 5 The cloud-edge collaborative self-evolution learning mechanism shown in the figure not only forms a complete data flow and parameter flow closed loop in structure, but also meets the convergence and stability conditions in theory. The knowledge distillation module guarantees the lightweight and consistency of the model, the online incremental learning module realizes local rapid self-adaptation, and the credibility evaluation module realizes anomaly detection and self-repair, which together constitute the mathematical closed loop of the "self-evolution intelligent scheduling system". The correctness of the mechanism has been proved by theoretical derivation and experimental verification, which shows that the self-learning intelligent scheduling of the embodiment under the edge-cloud architecture has the engineering advantages of sustainable evolution, controllable stability and verifiable, and provides a solid technical support for the high safety and high intelligence of the solid rocket engine continuous charging production line in equipment manufacturing.

[0085] Figure 6 The digital twin system scheduling decision explainability analysis interface proposed in the embodiment is shown. The interface analyzes the internal decision-making process of the intelligent scheduling system in a human-computer understandable manner, forming a full closed loop of "input features - decision contribution - risk explanation - feedback correction". The system displays the decision logic of the model through visualization means, so that the operators and system maintenance personnel can understand "why the system chooses this scheduling strategy" in real time, thereby improving the safety, transparency and credibility of the system.

[0086] Figure 6It consists of three main parts: (1) left feature contribution bar chart; (2) right risk factor heat map; (3) bottom scheduling decision explanation panel. The feature contribution bar chart (left) is used to visually display the influence strength and direction of each input feature on the current scheduling decision. The system collects real-time data from the edge nodes, such as device load, process temperature, energy consumption level, operator status, production rhythm fluctuation, task priority, etc., and calculates the feature importance weight in the decision model output . The risk factor heat map (right) reflects the risk level of different process nodes with color depth. The heat map matrix represents the comprehensive score of the process under the risk factor. The scheduling decision explanation panel (below) outputs the system's natural language explanation report, forming a structured text: (1) decision reason: the system selects to reduce the host power based on the current device load and process temperature trend to maintain the safety threshold; (2) key features: temperature rise rate, energy consumption peak, operator status; (3) influence direction: temperature rise rate (negative), energy consumption (negative), operator proficiency (positive); (4) risk level: medium-high risk (triggering warning signal). This interface not only displays the results, but also reveals the formation logic of intelligent decision-making, making the internal behavior of the algorithm black box verifiable and auditable.

[0087] In the solid engine continuous charging production line experimental environment, the system is tested for 48 hours of continuous operation. The explanation results and actual working condition data are recorded every 2 hours, as shown in Table 1.

[0088] It can be seen that with the improvement of system self-evolution and model stability, the accuracy and understanding rate of explainability results increase synchronously. Especially in the high-risk working condition, the heat map automatically labels the red area, enabling operators to take intervention measures before the risk expands. This verifies that the system is not only theoretically feasible, but also significantly improves safety and operational transparency in engineering applications. In summary, Figure 6 The "explainability analysis interface" not only visually realizes the transparency of model decision logic, but also mathematically proves its theoretical correctness through Shapley decomposition, sensitivity analysis and confidence measurement. The system can dynamically explain the causes of scheduling decisions, quantify feature contributions, and reveal potential risks, and achieve "verifiable-understandable-traceable" of safety decisions at the user level. The results verify the theoretical framework and algorithm correctness of "safety fault tolerance and explainable mechanism", providing an intelligent scheduling explanation system for equipment manufacturing that meets the concept of human-machine co-intelligence, realizing the qualitative change from "black box decision" to "transparent decision".

[0089] Figure 7The convergence curves of the MARL + NSGA-III fusion optimization algorithm proposed in this embodiment and the traditional NSGA-III algorithm in the multi-objective optimization process are shown. The horizontal axis is the number of iterations , and the vertical axis is the normalized Pareto fitness. The blue solid line represents the algorithm of the present application, and the orange dashed line represents the traditional algorithm. It can be seen that the blue curve rapidly stabilizes after about 160 iterations, with a small fluctuation range, and finally converges to a fitness value of 0.91; while the orange curve gradually stabilizes after about 420 iterations, and the final fitness is only 0.72. This result fully proves the theoretical correctness and performance superiority of the algorithm of this embodiment, which mainly reflects the following four aspects: improved convergence speed, enhanced stability, improved solution set diversity, and enhanced robustness.

[0090] The convergence of the fusion algorithm comes from the complementary mechanism between the reinforcement learning strategy optimization and the evolutionary algorithm population evolution. In the traditional NSGA-III algorithm, the search process relies on random selection and annealing criteria, which is easy to fall into local optimum. In the MARL + NSGA-III framework of this embodiment, the policy gradient of reinforcement learning guides the directional search, enabling the algorithm to have the ability of "targeted random exploration".

[0091] The statistical significance test is shown in Table 2. The fitness improvement of the fusion algorithm and the traditional algorithm is significant (p < 0.01) at the 95% confidence interval. In addition, the running time is reduced by about 60%, and the average computing cost is reduced from 45.6s to 18.2s, verifying its high efficiency. This is completely consistent with the "rapid rise-early stability" feature of the blue curve in Figure 7 .

[0092] In the solid rocket engine continuous charge production line scheduling problem, the optimization objectives usually include minimizing energy consumption, minimizing production cycle, and balancing equipment load. These objectives have strong coupling and conflict relationship, and traditional algorithms cannot be considered. This embodiment captures the timing characteristics (temperature change, filling beat, energy consumption dynamics) of the production system through multi-agent reinforcement learning, and uses the Pareto evolution mechanism of NSGA-III to find a balance point in the objective space. Therefore, the algorithm embodies the unity of "local optimal control-global optimal scheduling" at the physical level. When the system is disturbed by the environment (such as high temperature, device delay), the MARL part adjusts the immediate action to maintain local safety; while the NSGA-III part ensures that the overall scheduling layout is stable. The synergistic effect of the two makes the system maintain high efficiency and safety under dynamic conditions, which is the fundamental reason for the rapid stabilization of the Figure 7 blue curve.

[0093] Figure 7 The convergence curve of this embodiment verifies the theoretical correctness and actual effectiveness of the MARL + NSGA-III fusion algorithm. The fusion algorithm is not only faster than the traditional NSGA-III by about 68% in convergence speed, but also significantly superior in stability and robustness. Its convergence trajectory conforms to the random approximation theory and Lyapunov stability condition, and the improvement of solution set diversity and hyper-volume index verifies the applicability of the Pareto optimization theory. From an engineering point of view, this algorithm can quickly obtain stable, balanced and interpretable optimal solutions in real-time scheduling tasks, providing reliable mathematical and practical support for the intelligent optimization of solid rocket engine continuous charging production lines.

[0094] Table 3 lists all the input parameters and constraints involved in the scheduling model of this embodiment, which is the mathematical basis for building a multi-constrained multi-objective optimization model. The table divides variables into six categories: task parameters, device parameters, manpower parameters, safety constraints, process parameters, and energy consumption cost parameters, comprehensively describing the physical, process, and resource constraints of the solid rocket engine continuous charging production line.

[0095] At the task level, represents the th process or production task, including raw material mixing, pre-solidification, pouring, cooling, etc. Each task has start and end time constraints, satisfying . The device parameter part is represented by , the th key device, whose capacity and energy consumption constitute the constraint boundary. The manpower parameter is represented by , the operator, whose skill level directly affects the scheduling feasible region.

[0096] Safety constraints are an important feature of this table. During the continuous charging process, there are risks of high-energy fuel, combustible dust, and vacuum pouring, so the explosion-proof safety distance , the maximum safe temperature , and the pressure are all included in the model constraint set . For example, when the distance between adjacent devices of a certain process is less than , the scheduling algorithm automatically excludes this scheme, ensuring safety compliance from the source.

[0097] Process parameters such as mixing time , pouring speed reflect process window constraints, controlling scheduling rhythm and process stability. Energy consumption and cost parameters , , , and The economic indicators in the objective function are calculated to make the model achieve the optimal production efficiency while balancing energy consumption and cost.

[0098] Table 3 also specifies the constraint form, unit and data source of each parameter, ensuring the engineering feasibility of the model. For example, data such as production capacity and energy consumption are collected through OPC-UA; skill levels are obtained from MES personnel files; safety parameters are obtained from industry standards and sensor detection. All variables together constitute an optimization constraint set . Therefore, Table 3 not only provides mathematical input for algorithm solving, but also embodies the correspondence between the engineering system and the algorithm model, and is the basic data structure for the scheduling optimization of the embodiment.

[0099] Table 4 shows the experimental results of the "solid engine combustion chamber continuous charging production line edge-cloud collaborative multi-constraint multi-objective intelligent scheduling and digital twin system" of the present application on multiple performance indicators, and is compared with the traditional NSGA-III method. The test is carried out under the same hardware and process conditions, and a combination of 10 groups of random disturbance experiments and steady-state simulation is used to verify the stability and efficiency of the system under complex constraints, dynamic disturbance and high safety manufacturing scenarios. The experimental results show that the system of the embodiment has achieved significant improvement in production efficiency, real-time response, safety guarantee, energy efficiency optimization and communication performance, etc.

[0100] As can be seen from Table 4, the average production cycle is shortened from 120 minutes of the traditional system to 72 minutes, with a reduction of 40%. This significant improvement is mainly due to the following two mechanisms: (1) Edge-cloud collaborative decision-making mechanism: while the global optimization is performed in the cloud, the edge node is responsible for local disturbance repair. When the system detects task delays or device abnormalities, the edge agent completes the rescheduling within 500 milliseconds, while the traditional system needs to wait for the central server to solve again, with an average delay of 2.8 seconds. This "double parallel" scheduling mechanism effectively avoids production stagnation and waiting time, and significantly compresses the non-processing cycle.

[0101] (2) Intelligent task priority adaptive mechanism: the system adjusts the weight in real time according to the device state and task urgency through the dynamic reward adjustment strategy of reinforcement learning, so as to automatically balance the conflict between production capacity bottleneck and energy minimization. In theory, this dynamic weighting makes the scheduling decision satisfy the multi-objective optimization condition: . Among them is the cycle cost, is the energy consumption, is the device utilization. In the experiment, the device utilization rate increased from 63% to 81%, with an increase of 18%, verifying the adaptive optimization capability of the mechanism.

[0102] The system scheduling response time is reduced from 2.8 seconds to 0.46 seconds, with an improvement of 83.6%. This improvement reflects the synergistic advantage of the system between edge intelligent inference and cloud optimization. By deploying lightweight models on the Jetson NX node, the system utilizes local computing resources to achieve "no waiting" fast inference decision. In theory, the system response delay can be expressed as: where, is the communication delay; is the inference calculation delay; is the synchronization delay.

[0103] This embodiment uses Kafka asynchronous message mechanism and multi-thread parallel communication to make reduce to 190 ms (traditional system is 680 ms), and through model compression and parameter distillation, to control within 200 ms, thereby achieving a significant reduction in total response time. In addition, in 10 random disturbance experiments, the rescheduling success rate is improved from 74% to 98%. The results show that the system still has high robustness and fault tolerance under complex disturbances (including device downtime, raw material delay, environmental temperature deviation, etc.). This is due to the "edge fast repair + cloud model update" closed-loop mechanism of the invention. The experiment shows that the system can recover to the stable interval within 0.5 seconds after disturbance, verifying the correctness of the stable control theory.

[0104] The safety alarm triggering rate is reduced from 6.2% to 1.1%, with a decrease of 82.3%. The result comes from the multi-layer safety fault-tolerant mechanism and digital twin risk prediction module introduced in the system. In 72 hours of continuous charging simulation test, all over-temperature and over-pressure risk events are predicted and corrected in advance by the system, and there is no major alarm. This shows that the invention realizes a "early warning - intervention - recovery" closed-loop control in the safety protection layer, which is equivalent to a dynamic risk self-regulating system in theory. The energy consumption level is also significantly reduced. The experimental results show that the average single batch energy consumption is reduced from 320 kWh to 240 kWh, with a decrease of 25%. The effect comes from two aspects: (1) The reinforcement learning agent adjusts the energy consumption mode dynamically to realize load adaptive allocation; (2) The cloud optimization module minimizes the objective function on the global level to achieve energy consumption - production balance. Through the above two mechanisms, the system of the embodiment realizes the dual benefits of energy consumption optimization and safety guarantee, which has obvious economic and strategic value in manufacturing.

[0105] The cloud-edge communication delay is reduced from 680 ms to 190 ms, a decrease of 72.1%. This verifies the high real-time and reliability of the OPC-UA combined with Kafka mechanism. Kafka provides an asynchronous message queue and batch confirmation mechanism, making the data packet loss rate between the edge node and the cloud less than 0.01%, and ensuring that the task scheduling and MES system state synchronization delay is not more than 200 ms. The improvement in communication performance also indirectly improves the virtual-real synchronization efficiency of the system, ensuring that the error between the digital twin simulation state and the actual production state is less than 2%. The high communication stability of the system also provides a technical guarantee for multi-agent strategy synchronization.

[0106] Table 4 verifies the comprehensive performance improvement of the system in this embodiment in multiple dimensions, mainly including: (1) Efficiency improvement: production cycle shortened by 40%, task delay significantly reduced; (2) Intelligence enhancement: the system can respond to external disturbances within milliseconds; (3) Safety improvement: early warning rate and trigger rate significantly decreased; (4) Energy efficiency optimization: overall energy consumption reduced by 25%, unit output energy efficiency improved by 18%; (5) Stability: delay reduced by 72%, synchronization reliability up to 99%.

[0107] In summary, the system in this embodiment not only shows high reliability and verifiability at the experimental level, but also theoretically meets the convergence conditions of dynamic robust control and multi-objective optimal scheduling model. The system can achieve the comprehensive goal of "high safety, high efficiency, low energy consumption, and strong intelligence" in high-risk scenarios such as solid rocket engine continuous charging, fully proving the scientificity and engineering correctness of the performance comparison results in Table 4.

[0108] Table 5 shows the comparison results of the MARL + NSGA-III fusion algorithm proposed in this embodiment and the traditional NSGA-III algorithm in multi-objective optimization tasks. The test is conducted under the same task set, constraint set, and hardware conditions. Through multiple rounds of simulation calculation of average convergence time, computational complexity, Pareto front quality, diversity, and error bias, the convergence, stability, and computational efficiency of the algorithm are systematically verified.

[0109] As can be seen from Table 5, the number of iterations required for the fusion algorithm to reach the optimal solution is reduced from 500 times of the traditional algorithm to 160 times, with an increase of 68% in convergence speed. This is because the MARL part introduces a directional policy gradient update mechanism for the search process, which can quickly locate high-yield areas. The traditional NSGA-III relies on random crossover and mutation, with low exploration efficiency; while MARL converts the search into an expectation maximization problem by learning the state-action-reward relationship: When the local optimal solution obtained by the reinforcement learning agent is transmitted into the NSGA-III, the algorithm starts from a higher quality initial population, and accelerates the convergence process. In theory, this double-layer mechanism is equivalent to embedding an "intelligent explorer" outside the "global search" of the genetic algorithm, thereby reducing the algorithm complexity from to . In the experiment, the calculation time is reduced from 45.6 seconds to 18.2 seconds, with a performance improvement of about 60%.

[0110] The Pareto front diversity index Δ is improved from 0.32 to 0.67, with an increase of 109%. This result shows that the fusion algorithm can maintain the balance of the solution space while maintaining fast convergence. In the algorithm design, NSGA-III adopts the "reference point guidance + crowding distance" strategy: wherein: is the distance between adjacent non-dominated solutions, is the average distance. The MARL introduces a "strategy disturbance term" in the solution set generation stage: ,

[0111] which enhances the coverage of the solution space, enabling the system to avoid falling into local extrema too early. Therefore, the embodiment is obviously superior to the traditional NSGA-III in terms of solution set diversity and global balance. The hyper volume (HV) index is increased from 0.72 to 0.91, with an increase of 26%. The HV index is used to measure the overall quality and coverage range of the Pareto front, and the increase indicates that the fusion algorithm is closer to the theoretical optimal Pareto boundary. Figure 7 The blue curve in the convergence curve is smoother and tends to be stable earlier, which also verifies this result.

[0112] The target deviation rate is reduced from 8.4% to 2.6%, with a decrease of about 69%. This result shows that the algorithm has stronger anti-interference and stability. When the system input is disturbed , the reinforcement learning agent can dynamically adjust the output through strategy updating, thereby maintaining stable performance. The system error will converge exponentially to zero. The low target deviation rate and stable rounds in the experimental results verify that this theoretical condition is established. In addition, the recovery response time of the algorithm in the face of disturbance events (device failure, task delay) is 0.46 seconds, while the traditional algorithm takes 2.8 seconds. This is due to the local retraining mechanism of the edge node, which can update the local strategy parameters immediately after receiving new data, and synchronize to the cloud model through knowledge distillation, achieving fast repair and consistency maintenance.

[0113] ​The "learning stable round" of the fusion algorithm reaches 1200 rounds, indicating that the system can still maintain the stability and effectiveness of the strategy parameters in long-term operation. In the 72-hour continuous operation experiment, the Pareto fitness value does not show obvious degradation, and the balance between multiple objectives remains consistent. This shows that the algorithm has the ability to continuously learn in a dynamic environment, and its long-term stability is guaranteed by two parts: (1) the strategy gradient smooth updating mechanism ensures small step optimization; (2) the cloud-edge synchronization correction mechanism prevents parameter drift. The system stability variance σ decreases from 0.11 to 0.03, a decrease of 73%, indicating that the solution set fluctuation is significantly reduced, and the convergence trend is more smooth. This also proves that the convergence trajectory of the algorithm conforms to the random approximation theory and asymptotic stability conditions.

[0114] From a theoretical point of view, the fusion algorithm of the embodiment combines the dynamic adaptability of reinforcement learning with the global searchability of evolutionary algorithms, forming a hybrid nonlinear convergence system. From an engineering point of view, the algorithm can achieve fast, stable, and low-energy optimal scheduling solutions in complex manufacturing systems, providing a verifiable intelligent optimization tool for continuous charging production.

[0115] In summary, Table 5 verifies the efficiency, robustness, and stability of the MARL + NSGA-III algorithm, which outperforms the traditional NSGA-III algorithm in performance indicators, theoretically meets the stable Pareto convergence condition, and realizes real-time scheduling and continuous self-learning closed loop in engineering, providing a solid data and theoretical support for algorithm innovation.

[0116] The above-described embodiments are only preferred embodiments of the present application, and the protection scope of the present application is not limited thereto. Any equivalent replacement or transformation made by those skilled in the art based on the present application is within the protection scope of the present application. The protection scope of the present application is subject to the claims.

Claims

1. A multi-constraint, multi-objective real-time production scheduling and scheduling method for a solid rocket motor combustion chamber continuous charging production line, characterized in that: The method is realized based on an edge-cloud-digital twin collaborative architecture, in which a multi-agent reinforcement learning model and an NSGA-III multi-objective evolutionary algorithm are deployed on the cloud, the edge node is realized based on Jetson NX to realize fast response, and data communication is realized through an OPC-UA protocol and a Kafka message middleware; the method comprises the following steps: Step one: system initialization and multi-source data acquisition Initialization of the scheduling task set , device resource set , human resource set ; real-time acquisition of environmental state data and safety constraint set , the safety constraint set at least includes the anti-explosion safety distance , the maximum safety temperature and the maximum safety pressure ; the acquired data is uploaded through the OPC-UA protocol and is asynchronously fused and transmitted through the Kafka message bus; Step two: multi-agent state construction Various resources in the production line are abstracted as independent agents ; the state space of the first agent is defined as: wherein: is a resource load rate; is a device state vector; is a process progress; is an energy consumption state function; is an environmental risk factor; Defining action space , the action includes task receiving, task passing, energy consumption mode switching and path adjustment; Step three: multi-agent decision and interaction Each agent selects and executes an action based on its local observation state and global information, based on a policy network The global reward function is defined as: wherein: represents the total cost; is the production cycle; is the equipment utilization; is the safety index; , , and are adaptive weights, dynamically updated by a policy gradient mechanism; Step four: NSGA-III multi-objective evolutionary optimization A population of candidate scheduling schemes generated by multi-agent reinforcement learning Inputting the NSGA-III multi-objective evolutionary algorithm for Pareto optimization; constructing a fitness function Wherein: is a production cost function; is a production cycle function; is a device utilization function; is a safety constraint penalty function; By non-dominated sorting, reference point correlation and crowding distance calculation, the next generation population is selected from the merged population The final output is a Pareto front solution set The final output is a Pareto front solution set ; Step five: edge-side real-time rescheduling and execution The edge computing node continuously monitors the production status, when detecting a disturbance event , identifies the affected task set ; the edge node initiates a fast rescheduling process based on the local state , invokes a local reinforcement learning agent to update the policy, and generates a local scheduling solution ; if the local solution is better than the cloud global solution , triggers an uplink synchronization, and uploads it to the cloud ; Step six: digital twin synchronization and visual feedback: The final determined scheduling scheme, including a task sequence, a station allocation and a process parameter, is synchronized in real time to a manufacturing execution system and a digital twin visualization platform; the digital twin model performs virtual-real mapping on the physical production line, dynamically displays equipment state, task Gantt chart, energy consumption curve and safety risk heat map, and provides an explainable analysis report for the operator to make scheduling decisions.

2. The multi-constraint multi-objective real-time production scheduling and dispatching method for the solid rocket motor combustion chamber continuous charging production line according to claim 1, characterized in that: In step three, the multi-agent reinforcement learning uses the Actor-Critic framework for distributed training, with each agent maintaining a policy network. and a value network By employing a centralized training and distributed execution approach, the gradient ascent method is utilized. Update the policy network parameters, where: The learning rate is the policy learning rate.

3. The multi-constraint multi-objective real-time production scheduling and dispatching method for the solid rocket motor grain continuous production line according to claim 1, characterized in that: In the step three, the method for dynamically updating the adaptive weight through the policy gradient mechanism is as follows: wherein: and are the adaptive weights before and after the update, respectively; is the learning rate; is the global reward function.

4. The multi-constraint multi-objective real-time production scheduling and dispatching method for the solid rocket motor grain continuous production line according to claim 1, characterized in that: In the step four, the specific execution process of the NSGA-III multi-objective evolutionary optimization comprises: 41) initialize population , set maximum number of iterations ; 42) for each generation , calculate the fitness of each individual in the population ; 43) Non-dominated sorting of the population, resulting in front ranks ; 44) normalizing and correlating each individual in the same front level based on a reference point; 45) calculating the crowding distance of the individual, and sorting according to the crowding distance in the same front level to maintain diversity; 46) Generating a population of offspring by simulated binary crossover and polynomial mutation ; 47) Merge parent and offspring populations ; 48) from the optimal individuals constitute the next generation population ; 49) repeat steps 42) through 48) until , output the pareto-optimal solution set .

5. The multi-constraint multi-objective real-time production scheduling and sequencing method for a solid rocket motor combustion chamber continuous grain production line according to claim 4, characterized in that: The mixed optimization framework constituted by the multi-agent reinforcement learning and the NSGA-III multi-objective evolutionary algorithm has convergence defined as: for the scheduling strategy and the Pareto front , if there is a finite constant such that: , wherein: is the number of iterations; is the optimal scheduling policy; is the Pareto optimal solution set; then the hybrid optimization framework is said to converge to a stable Pareto equilibrium.

6. The multi-constraint multi-objective real-time production scheduling and sequencing method for a solid rocket motor combustion chamber continuous grain production line according to claim 5, characterized in that: The iterative model of the hybrid optimization framework is described by the following coupled equations: The multi-agent reinforcement learning is used to dynamically generate a candidate solution set: wherein: and denote the current iteration state vector and the next updated state vector, respectively; is the learning rate; is the reward function of reinforcement learning; is the policy gradient; is the policy function; The NSGA-III multi-objective evolutionary algorithm performs non-dominated sorting in the Pareto space, and the population update formula is as follows: The hybrid optimization framework is coupled to obtain a composite evolutionary equation: wherein: represents the moment in time, the comprehensive evolutionary solution set; is the population of the generation; offspring population resulting from crossover and mutation; is the composite evolutionary function of the hybrid optimization framework; is the environmental perturbation term; The hybrid optimization framework guarantees convergence under the conditions that the hybrid optimization framework is constructed by constructing a joint Lyapunov function and proving that the expected value of the joint Lyapunov function is decreasing along the iterations. converges.

7. The method of claim 1, wherein the method is a multi-constrained multi-objective real-time production scheduling and dispatching method for a solid rocket motor grain continuous production line, characterized in that: Further comprising a model self-evolution and online updating mechanism based on cloud-edge collaboration, comprising: Knowledge distillation: compressing the parameters of a cloud-trained large global model into a light-weight parameter set through knowledge distillation techniques and periodically pushed to the edge nodes, achieving knowledge transfer from the complex teacher network to the light-weight student network by minimizing the distillation loss; the distillation loss is expressed as: wherein: represents the policy output of the edge student network in state ; represents the policy output of the cloud teacher network in state ; Online incremental learning: edge node based on new acquisition of local data Fine-tuning the model, the parameter update formula is: wherein: is a fine-tuning learning rate; is a fine-tuning reward function; is a current edge model parameter; is a local dataset of the edge node at time step . Confidence evaluation and trigger retraining: the system calculates the confidence of the model's prediction output When the confidence is lower than a preset threshold , it is determined that the model may fail, and the retraining of the cloud model and the replacement process of the edge model are automatically triggered; wherein: represents the action output by the model in state ; represents the current system state.

8. The method of claim 1, wherein the method is a multi-constrained multi-objective real-time production scheduling and dispatching method for a solid rocket motor grain continuous production line, characterized in that: In the fifth step, the impact range of the disturbance event is determined by analyzing the task-device dependency graph and the process timing constraints.

9. The method of claim 1, wherein: In the sixth step, the explainability analysis report is generated by SHAP analysis or feature contribution calculation, showing the contribution weight of input features to the final scheduling decision in the form of a bar chart , and the comprehensive risk level of different process nodes is visualized in the form of a heat map ; the natural language description part of the explainability report includes at least three structured text modules of "decision reason", "key impact features and their action direction" and "current comprehensive risk level", which are used to assist safety review and quality traceability.

10. A system for implementing the multi-constraint multi-objective intelligent scheduling and dispatching method for the side cloud collaboration of the continuous charging production line of the combustion chamber of the solid engine according to any one of claims 1-9, characterized in that it comprises: A three-layer collaborative architecture of edge-cloud-digital twin is adopted, comprising: A data acquisition and fusion module is deployed in the production field, comprising a PLC controller and temperature, pressure and position sensors, for real-time acquisition of equipment state, process parameters and environmental monitoring data; the data acquisition and fusion module is provided with an OPC-UA client for realizing standardized data access of heterogeneous devices, and the fused data stream is uploaded to the cloud in an asynchronous mode through a Kafka message middleware, and is simultaneously uploaded to a digital twin visualization module; The cloud-based optimization scheduling module, deployed on a cloud server, runs a fusion model of multi-agent reinforcement learning and the NSGA-III multi-objective evolutionary algorithm. The multi-agent reinforcement learning submodule defines the agents, constructs the state and action space, and generates high-quality candidate scheduling schemes through distributed interactive learning. The NSGA-III multi-objective evolutionary optimization submodule receives the candidate schemes and solves for the Pareto optimal front solution set through non-dominated sorting, reference point association, and crowding distance calculation. This enables comprehensive optimization of production cycle, cost, equipment utilization, and safety. An edge reasoning and execution module is deployed on a Jetson Xavier NX-based embedded edge computing node, receives and executes scheduling instructions issued by the cloud; the edge reasoning and execution module has a built-in lightweight reinforcement learning reasoning model for rapid execution of local scheduling tasks; the edge reasoning and execution module continuously monitors the production state, and when a disturbance event is detected through the data acquisition and fusion module , can autonomously start a rescheduling process and generate a local optimization solution, and compare it with the global solution of the cloud, and trigger uplink synchronization when the conditions are met. A communication and synchronization module realizes real-time data acquisition and control instruction issuing of the device layer based on an OPC-UA protocol, realizes high-reliability and low-delay transmission and decoupling of massive data between the cloud and the edge based on a Kafka asynchronous message bus, and realizes real-time bidirectional communication with the digital twin visualization module based on a WebSocket protocol; The digital twin visualization module runs on the monitoring terminal and constructs a virtual model consistent with the physical production line; the digital twin visualization module receives real-time data from the communication and synchronization module, dynamically maps the scheduling state, device utilization, energy consumption curve and safety risk, and visually displays them in the form of Gantt chart, heat map and the like; the digital twin visualization module also integrates an explainability analysis interface for revealing the internal logic of scheduling decisions and the contribution degree of key features in the form of visual charts and natural language reports.

Images