Distributed production management and control method based on cloud edge-end cooperation

Abstract

The application relates to the technical field of industrial automation and intelligent manufacturing, and specifically provides a distributed production management and control method based on cloud edge-end cooperation, which comprises the following steps: constructing a cloud intelligent agent, a plurality of edge-end intelligent agents and a cooperative communication module; each edge-end intelligent agent is bound with a production execution unit, collects real-time data and executes local autonomous decision, and performs millisecond-level response to low-complexity disturbance; when the disturbance exceeds the autonomous capability or involves global resource conflict, the edge-end intelligent agent reports a cooperative event to the cloud; the cloud intelligent agent evaluates the event influence through knowledge graph simulation deduction, generates global optimization scheduling instructions by using a multi-agent reinforcement learning algorithm integrated with a layered attention mechanism, and issues the instructions to the related edge-end intelligent agents; and the edge-end intelligent agents execute the instructions after verification by local digital twinning and feed back the results. The application significantly improves the response speed and flexibility of the production system to variable working conditions, and is suitable for multi-variety and small-batch discrete manufacturing scenes.

Description

Technical Field

[0001] This invention belongs to the field of industrial automation and intelligent manufacturing technology, specifically relating to a distributed production management and control method based on cloud-edge-device collaboration. Background Technology

[0002] Traditional mass production models are facing market challenges from the increasing demands for diverse product varieties, small batches, and rapid delivery. Manufacturing companies need the ability to quickly adjust production line operation modes to achieve flexible production. Simultaneously, changes in equipment status, fluctuations in material supply, and adjustments to work order priorities require manufacturing companies to respond quickly and make timely adjustments.

[0003] Currently, most mainstream flexible production management systems adopt a centralized architecture. Anomalies occurring during workshop production (such as equipment malfunctions, material shortages, and work order adjustments) are centrally reported to the enterprise dispatch center, where they are resolved manually or with the assistance of the flexible manufacturing system. Even in cloud-edge collaborative solutions that incorporate edge gateways, the essence remains a "cloud-based decision-making, edge-based execution" operating model: the edge gateway only acts as a data collector, providing on-site data support to the cloud dispatch center, lacking autonomous decision-making capabilities and unable to achieve local autonomy. This centralized dispatching method suffers from problems such as untimely response, delayed decision-making, and high costs associated with achieving flexible effects, making it difficult to meet the rapid response requirements under changing operating conditions.

[0004] Therefore, there is an urgent need to build a distributed production management and control system that has autonomous decision-making capabilities at the edge, cloud-edge collaborative optimization, and the ability to quickly respond to production changes under different operating conditions. Summary of the Invention

[0005] In view of the above-mentioned shortcomings of the prior art, the present invention provides a distributed production management and control method based on cloud-edge-device collaboration to solve the above-mentioned technical problems.

[0006] In a first aspect, the present invention provides a distributed production management and control method based on cloud-edge-device collaboration, comprising: The system constructs a cloud-based intelligent agent, multiple edge-based intelligent agents, and a collaborative communication module connecting the cloud-based intelligent agent and each edge-based intelligent agent. Each edge-based intelligent agent is uniquely bound to a production execution unit and is used to collect real-time data from that production execution unit and perform local autonomous decision-making. The cloud-based intelligent agent is used to achieve global optimization scheduling across production lines or workshops. The edge agent continuously collects real-time data from the production execution unit it is bound to, identifies the type and severity of local disturbances through a local disturbance identification model, and autonomously executes decision-making actions according to a preset autonomous strategy. When the edge agent determines that the local disturbance exceeds its autonomous capability or involves a global resource conflict, it reports the collaborative event to the cloud agent through the collaborative communication module. After receiving the collaborative event, the cloud-based agent evaluates the impact of the collaborative event and generates a global optimization scheduling instruction based on the evaluation result using a multi-agent reinforcement learning algorithm, which is then sent to the relevant edge agents through the collaborative communication module. After receiving the global optimization scheduling instruction, the edge agent confirms that the instruction is feasible, executes it, and feeds back the execution result to the cloud agent.

[0007] Secondly, the present invention provides a distributed production management and control system based on cloud-edge-device collaboration, used to execute the distributed production management and control method based on cloud-edge-device collaboration provided in the first aspect.

[0008] Thirdly, a device is provided, comprising: The storage device is used to store distributed production management and control programs based on cloud-edge-device collaboration. The processor is used to implement the steps of the distributed production management method based on cloud-edge-device collaboration as provided in the first aspect when executing the distributed production management program based on cloud-edge-device collaboration.

[0009] Fourthly, a computer-readable medium is provided, on which a distributed production management and control program based on cloud-edge-device collaboration is stored, wherein when the distributed production management and control program based on cloud-edge-device collaboration is executed by a processor, the steps of the distributed production management and control method based on cloud-edge-device collaboration provided in the first aspect are implemented.

[0010] The beneficial effects of this invention are as follows: The distributed production management and control method based on cloud-edge-device collaboration provided by this invention constructs a hierarchical architecture in which cloud-based intelligent agents and multiple edge-based intelligent agents collaborate. This enables edge-based intelligent agents to autonomously handle local disturbances, achieving millisecond or second-level responses to common disturbances such as equipment anomalies and quality fluctuations. This significantly reduces the minute-level response latency of traditional systems, greatly lowering the risk of production interruptions. The synergy between edge-based local autonomy and cloud-based global optimization ensures that even with brief interruptions in the cloud or network, each edge-based intelligent agent can maintain local production operations, greatly improving the system's flexibility and robustness. Before receiving scheduling instructions from the cloud, the edge-based agents first perform verification and simulation in their local digital twins, confirming feasibility before execution, effectively avoiding production errors that may result from direct execution. Simultaneously, driven by collaborative events reported by multiple edge-based agents, the cloud-based intelligent agents generate optimized scheduling instructions through global digital twin simulation and multi-agent reinforcement learning, achieving cross-production line resource conflict resolution and global goal optimization. This invention balances the real-time nature of local response with the high efficiency of global scheduling, providing a reliable distributed management and control solution for flexible production with multiple varieties, small batches, and rapid delivery. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic flowchart of a method according to an embodiment of the present invention.

[0013] Figure 2 This is a schematic architecture diagram of a multi-agent system according to an embodiment of the present invention.

[0014] Figure 3 This is a flowchart illustrating the application scenario of one embodiment of the present invention.

[0015] Figure 4 This is an emergency order insertion scheduling flowchart in an application scenario of one embodiment of the present invention.

[0016] Figure 5 This is a flowchart illustrating the scheduling process for a decrease in the quality pass rate under an application scenario of one embodiment of the present invention.

[0017] Figure 6 This is a flowchart of a device anomaly scheduling process in an application scenario according to an embodiment of the present invention. Detailed Implementation

[0018] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0020] The distributed production management and control method based on cloud-edge-device collaboration provided in this embodiment of the invention is executed by computer equipment, and correspondingly, the distributed production management and control system based on cloud-edge-device collaboration runs on the computer equipment.

[0021] Figure 1 This is a schematic flowchart illustrating a method according to an embodiment of the present invention. Wherein, Figure 1The executing entity can be a distributed production management and control system based on cloud-edge-device collaboration. Depending on different requirements, the order of the steps in this flowchart can be changed, and some can be omitted.

[0022] like Figure 1 As shown, the method includes: S1. Construct a cloud-based intelligent agent, multiple edge-based intelligent agents, and a collaborative communication module connecting the cloud-based intelligent agent and each edge-based intelligent agent; wherein, each edge-based intelligent agent is uniquely bound to a production execution unit and is used to collect real-time data of the production execution unit and perform local autonomous decision-making; the cloud-based intelligent agent is used to realize global optimization scheduling across production lines or workshops; S2. The edge intelligent agent continuously collects real-time data from the production execution unit it is bound to, identifies the type and severity of local disturbances through a local disturbance identification model, and autonomously executes decision-making actions according to a preset autonomous strategy; S3. When the edge agent determines that the local disturbance exceeds its autonomous capability or involves a global resource conflict, it reports the collaborative event to the cloud agent through the collaborative communication module. S4. After receiving the collaborative event, the cloud-based intelligent agent evaluates the impact of the collaborative event and generates a global optimization scheduling instruction based on the evaluation result using a multi-agent reinforcement learning algorithm, which is then sent to the relevant edge intelligent agents through the collaborative communication module. S5. After receiving the global optimization scheduling instruction, the edge agent confirms that the instruction is feasible, executes it, and feeds back the execution result to the cloud agent.

[0023] Please refer to Figure 2 In one embodiment of the present invention, a cloud-edge-device collaborative distributed system is constructed, comprising: constructing a cloud-based intelligent agent, multiple edge-device intelligent agents, and a collaborative communication module connecting the two.

[0024] In this embodiment, the cloud-based intelligent agent is deployed on an enterprise private cloud or hybrid cloud server, responsible for macro-level scheduling and global optimization of the entire workshop or multiple production lines. The edge-based intelligent agent is deployed in software form on computing devices (such as industrial edge gateways, embedded industrial control computers, or intelligent device controllers) at the production site and is uniquely bound to a production execution unit. The production execution unit can be a CNC machine tool, an industrial robot, an assembly line, or an automated workstation.

[0025] For example, in an automotive parts manufacturing workshop, there are 5 machining centers, 3 testing devices, and 2 assembly lines. The system deploys an independent edge agent (Agent_CNC1 to Agent_CNC5) for each machining center, and corresponding edge agents for the testing devices and assembly lines. Each edge agent interacts with its bound production execution unit in real time via protocols such as OPCUA, ModbusTCP, or MTConnect, collecting information such as equipment status (running, standby, fault), process parameters (spindle speed, feed rate, temperature), task queues, and quality inspection results.

[0026] The cloud-based intelligent agent maintains a digital twin model of the entire workshop, which includes static attributes (equipment model, processing capacity, maintenance cycle) and dynamic data (real-time status, work-in-process progress, material inventory) for all production execution units. Simultaneously, the cloud-based intelligent agent integrates a global optimization scheduler based on multi-agent reinforcement learning.

[0027] The collaborative communication module employs publish-subscribe messaging middleware such as MQTT or Kafka to establish a bidirectional data channel between the cloud and each edge agent. Each edge agent subscribes to scheduling instruction topics, model update topics, and global configuration topics issued by the cloud; cloud agents subscribe to real-time status topics, collaborative event topics, and execution feedback topics reported by each edge agent. The communication protocol supports TLS encryption to ensure secure data transmission.

[0028] Through the above construction, a three-layer distributed production management and control architecture is formed: the cloud is responsible for global optimization across production lines, the edge is responsible for local autonomous decision-making, and the collaborative communication module ensures real-time information synchronization.

[0029] In this embodiment, both the cloud-based intelligent agent and each edge-based intelligent agent have a built-in digital twin engine, supporting two working modes: production monitoring and simulation inference.

[0030] (a) Production monitoring mode Taking a machining center in an automotive parts workshop as an example, the edge agent bound to the equipment collects real-time data such as spindle load, tool life, and machining temperature at 100-millisecond intervals via the OPCUA protocol, and synchronizes the data to the local digital twin model. The cloud agent gathers the data reported by all edge agents and constructs a 3D visualization dashboard for the entire workshop.

[0031] In monitoring mode, operators can view the real-time operating status of each piece of equipment (green for normal, yellow for standby, and red for malfunction), current work order number, progress, and quality pass rate via a web interface or SCADA system. When the spindle temperature of a piece of equipment exceeds a preset threshold, the edge agent immediately updates the status attributes of the local twin model and displays an alert on the monitoring interface. Simultaneously, the cloud agent automatically highlights the location of the equipment and traces its potential impact on downstream processes. Production managers can zoom in and out at any time, penetrating from the overall plant layout to the virtual entity of a single piece of equipment, achieving transparent supervision with a "one-screen view of the entire plant."

[0032] (ii) Simulation Reasoning Mode Simulation reasoning models are used to extrapolate before decision execution, reducing the risk of production adjustments.

[0033] Cloud-based simulation scenario: When the sales department plans to insert an urgent order (with higher priority than all current work orders), the dispatcher clicks the "Insert Order Simulation" button on the cloud system interface. The cloud-based intelligent agent first replicates the current global digital twin state (i.e., a snapshot of all equipment status, work order queues, and material inventory) to form an isolated simulation copy. Then, through a context-aware event injector, the "Insert Order" event is converted into an event triple containing the work order number, process route, expected delivery time, and priority weight, and injected into the knowledge graph of the simulation copy. The cloud-based intelligent agent initiates the multi-agent simulation core: each work order agent selects equipment according to the shortest processing time priority rule, each equipment agent dynamically adjusts the processing efficiency coefficient based on its health status, and the scheduling arbitration agent coordinates resource conflicts based on a multi-agent reinforcement learning strategy. The simulation clock advances with virtual time (e.g., 10 minutes in simulation corresponds to a 2-hour actual scheduling decision cycle), and the final output is the simulation results: the estimated completion time of the new order, the delay distribution of the original order, the bottleneck equipment utilization rate change curve, and key conflict points. If the simulation shows that inserting an order will cause key orders to be delayed beyond the customer's tolerance (e.g., more than 4 hours), the system will automatically suggest rejecting the order or adjusting the process route, and the management will make the final decision accordingly.

[0034] Edge simulation scenario: When an edge agent in a machining center identifies abnormal tool wear with a severity level of 3 (high) through a vibration detection model, the edge agent cannot determine whether to immediately replace the tool or adjust cutting parameters to continue production. At this point, the edge agent switches to local simulation inference mode: it replicates its bound equipment twin state (current remaining tool life, machining parameters of the remaining 5 work orders in the task queue) to form a local simulation environment. Then, the edge agent calls a genetic algorithm to perform inference and evaluation on three candidate solutions in the simulation environment (Solution A: Immediate tool replacement, taking 10 minutes; Solution B: Reduce feed rate by 20% and continue machining, but the machining time per piece increases by 15%; Solution C: Switch to a spare tool). After 100 generations of inference for each solution, the estimated total completion time and tool scrap risk probability are output. The edge agent selects the solution with the lowest overall cost (e.g., Solution B, estimated to only delay by 30 minutes and without requiring machine downtime for tool replacement), executes this decision locally, and simultaneously encrypts and uploads the simulation process data to the cloud for subsequent model fine-tuning.

[0035] This embodiment uses an edge-end intelligent agent bound to a five-axis machining center (number M-102) in a precision manufacturing workshop as an example to illustrate in detail its autonomous processing flow for local disturbances: (I) Real-time data acquisition and feature extraction Edge agents acquire spindle vibration acceleration at 20ms intervals via the OPCUA protocol. ), spindle current ( ), cutting force ( and workpiece surface roughness ( The raw data stream is processed through a sliding window (window length...). After segmentation, the device feature extraction model (which is a pre-trained convolutional autoencoder with the structure Conv1D(8,3)-Conv1D(16,3)-MaxPool-Flatten-Dense(32)) is called in advance and updated periodically by the cloud-based intelligent agent to map the original data into a 32-dimensional feature vector. :

[0036] in, The raw time series data within the window. For feature extraction networks, These are the model parameters sent from the cloud.

[0037] (II) Parallel perturbation identification and fusion classification The edge agent maintains two parallel recognition channels: Rule matching layer: Includes a lightweight local rule base and uses the Rete inference engine. For example, rule R1: IFI > 1.5 × I 基准 AND THEN Disturbance Type = Tool Wear, Severity = 5.

[0038] Lightweight AI Model Layer: Running three TinyML-based models in parallel: Anomaly detection model: Isolation Forest, outputting anomaly scores. ; Root cause analysis model: Based on decision trees, it outputs probability vectors for each cause. ; Quality prediction model: Linear regression, predicting the pass rate of the next three products. .

[0039] The rule layer decision result (if a match is found) will output the perturbation type. and severity ) and AI model layer output (output perturbation type) Severity Confidence level Perform weighted fusion. Set the weights... (AI weights) (Rule weights), final perturbation type Take the one with the highest confidence-weighted score between the two channels:

[0040] Final severity level Priority comparison is used: if the rule layer has output, it is directly selected. Otherwise take In this example, the model output is... Confidence level Rule R1 was not triggered, so the fusion result is: Disturbance type = tool wear, confidence level = 0.85, severity level = 3 (out of 5).

[0041] (III) Threshold-based complexity assessment and decision-making The edge agent maintains a severity threshold that is dynamically adjusted by the cloud. (Empirical value). The judgment logic is as follows: If the current complexity is low, the system will switch to a fast response channel; otherwise, it will switch to a deep deduction channel with high complexity.

[0042] In this example Therefore, it was determined to be a highly complex disturbance, and the in-depth simulation channel was activated.

[0043] (iv) Low complexity and fast response If another device shows "low coolant pressure" ( The edge agent directly retrieves the preset action from the local fast rule base - "increase the PWM duty cycle of the cooling pump from 60% to 80%" and immediately executes it through PLC instructions. The whole process takes 8ms and requires no simulation.

[0044] (v) In-depth simulation of high complexity For tool wear severity level 3, the edge agent initiates a local digital twin simulation. The simulation environment is built on the Unity3D engine and includes the device's kinematic model, the tool wear evolution model, and the current task queue (remaining work orders JOB-101 to JOB-105). The edge agent generates three candidate decision options: Option A: Stop the machine immediately and change the tool. (Time required) minute; Option B: Reduce the spindle speed by 20% (from 8000 rpm to 6400 rpm) and continue machining, but the machining time per piece will increase. ; Option C: Switch to the same type of tool in the spare tool magazine (requires no-load tool retrieval, increases...) minute).

[0045] The system invokes a genetic algorithm to perform deduction and evaluation of each option in the simulation environment. The decision variables of the genetic algorithm are binary encoded. (Indicates whether to select a certain option), the fitness function is the estimated total cost function. :

[0046] In the formula, This represents the total estimated completion time (in minutes) for all work orders in progress. This represents the total downtime of the equipment. The probability of workpiece scrap due to tool breakage (given by the quality prediction model) is given by weighting. .

[0047] Genetic algorithm parameters: population size Crossover probability Probability of mutation iteration After considering each option individually, the fitness value is obtained: (High downtime costs for tool changes) (Extend processing time) (Time required to retrieve a spare knife; slight risk) The minimum fitness is option A, which is immediate tool change. However, the system then calls the local reinforcement learning policy network (with one parameter...). A DQN network is used for rapid evaluation, which additionally considers the impact of long-term wear on subsequent processes, and outputs a corrected Q value:

[0048] through Value comparison ( Finally, option A (immediate tool change) is selected. The edge agent sends a tool change command to the device PLC and encrypts and uploads the simulation process data (candidate solutions, fitness, Q value) to the cloud for subsequent strategy updates.

[0049] (vi) Uploading and Continuous Learning After each autonomous decision, the edge agent will transfer the feature vector. Fusion results Selected action Actual cost The data is packaged into JSON format and uploaded to the cloud-based intelligent agent via the MQTT protocol. After the cloud-based intelligent agent collects all edge data, it periodically fine-tunes the feature extraction model parameters. and reduce the complexity threshold This forms a closed-loop evolution of "data-learning-distribution-upgrade".

[0050] This embodiment uses the strategy update cycle after one week of operation in an automotive parts workshop as an example to illustrate in detail the complete process of cloud-based intelligent agents executing strategy updates: (I) Data aggregation and preprocessing The workshop is equipped with 10 edge agents (each bound to one of the eight machining centers, one testing device, and one assembly line). At the end of each production cycle (e.g., every 24 hours), each edge agent submits its local autonomous decision-making process data (including: original feature vectors). Disturbance type Severity level Selected action Actual cost after implementation After encryption, the data is uploaded to the cloud-based intelligent agent's data lake via the MQTT protocol. The uploaded data also includes final output metrics collected from the production execution unit (such as work order completion time, equipment utilization rate, and quality pass rate).

[0051] The cloud-based intelligent agent first cleans and aligns the data, removing incomplete records caused by communication anomalies, and sorts them by timestamp to form a training dataset. .in, It represents the global state (a vectorized representation of all device status, work order progress, and inventory). It is the set of local observations of each edge agent. This is a joint action issued from the cloud. For the team reward value defined later, This represents the global state at the next moment after execution.

[0052] (II) Team Reward Function Design Cloud-based intelligent agents with team reward functions The strategy is updated to optimize the objective. This function is directly linked to global production metrics. In this embodiment, the team reward function is defined as:

[0053] The specific rules for assigning values ​​to each component are as follows: Each time a work order is completed, Otherwise, it is 0.

[0054] The total waiting time is the sum of the idle times of all devices: Then the penalty item .

[0055] If a resource conflict deadlock occurs Otherwise, it is 0.

[0056] Delivery metric: If all work orders are completed ahead of schedule, If any work order is severely delayed (exceeding 20% ​​of the promised time); Otherwise, it is 0.

[0057] Experience sets weight During training, the cumulative discount reward is... discount factor .

[0058] (III) Updating of Multi-Agent Reinforcement Learning Policy Networks The cloud-based agent kernel maintains a multi-agent reinforcement learning policy network. Its structure is a neural network integrating a hierarchical attention mechanism, as described above. The parameters of the policy network... The update is performed using the Multi-Agent Proximity Policy Optimization (MAPPO) algorithm, with the agent pruning objective function as follows:

[0059] in, It is the advantage function calculated using generalized advantage estimation (GAE). To trim hyperparameters (in this embodiment, we take...) The cloud-based intelligent agent uses the recently collected data. time steps ( The data is used to analyze the policy network. Mini-batch gradient descent (batch size) Learning rate The value network is updated simultaneously to minimize the mean squared error loss.

[0060] (iv) Incremental fine-tuning of lightweight AI models The cloud-based intelligent agent also manages the lightweight AI model set distributed to the edge intelligent agent, including: device feature extraction models (encoders). These models include anomaly detection models (isolated forests), root cause analysis models (decision trees), and quality prediction models (linear regression). These models are incrementally fine-tuned as data accumulates to improve their adaptability to varying operating conditions.

[0061] Taking the device feature extraction model as an example, the cloud-based intelligent agent will upload the raw data. (From each edge) and the extracted feature vector As positive samples for self-supervised learning, a contrastive learning strategy is used to fine-tune the parameters. The loss function is:

[0062] in For cosine similarity, For temperature coefficient, positive sample pairs These are time slices from the same device under similar operating conditions. Fine-tuning is performed using the Adam optimizer with a learning rate of [missing information]. It can iterate for a maximum of 5 epochs.

[0063] The anomaly detection model and root cause analysis model adopt an incremental training approach: retain the original model structure, update the model parameters in batches (e.g., every 500 records) using new data, and discard overly old data samples through a sliding window mechanism to control memory usage.

[0064] (v) Model distribution and version management After the policy update is complete, the cloud-based agent packages all updated model and policy parameters: Global policy network parameters (For subsequent inference in the cloud only); Parameters of various types of lightweight AI models (Feature extraction), decision tree rule set update, and isolated forest threshold update; Optional local reinforcement learning policy network parameters (Used for decision support in high-complexity scenarios at the edge).

[0065] The cloud-based agent generates a personalized model package for each edge agent and publishes it to the corresponding topic (e.g., / model_update / agent_CNC01) via a publish-subscribe mechanism through the collaborative communication module. After subscribing to the topic, the edge agent downloads the model package and performs an atomic update: first, it loads the new model into a spare memory area, verifies its integrity, and then activates it, replacing the old model. If the download fails or verification fails, the edge agent automatically rolls back to the previous stable version and reports the error log.

[0066] After the update is completed, the cloud-based intelligent agent records the version number (e.g., v2.3.1) and the corresponding dataset fingerprint for rollback or auditing purposes.

[0067] This embodiment takes two adjacent machining centers (numbered M-101 and M-102) in an automotive parts workshop as an example to illustrate the process of resolving resource conflicts between edge agents through negotiation, and the complete process of reporting collaborative events to the cloud when negotiation fails or an emergency occurs.

[0068] (a) Edge agent determines the conditions for triggering reporting. Each edge agent maintains a dynamic event threshold matrix (see Table 1, which is periodically distributed and dynamically adjusted by the cloud). This matrix defines the triggering conditions for switching from "local autonomy" to "cloud-edge collaboration". The triggering conditions include: disturbance type, autonomous processing time limit, resource impact range threshold, and estimated global impact threshold.

[0069] For example, during the autonomous decision-making process, the edge agent Agent_101 of machining center M-101 detects an action about to be executed—"Occupy shared material buffer B-3 to complete emergency work order JOB-999" (estimated duration). (minutes). Agent_101 uses an influence assessment model (a lightweight LSTM network) to evaluate in real time the expected impact of the action on its neighboring edge agent Agent_102 (bound to downstream device M-102, dependent on the same buffer). Influence score vector The calculation formula is:

[0070] in, This is the current state vector of Agent_101 (task queue length, device health). Encode the candidate actions. This is a local observation of Agent_102 (limited neighbor information that can be obtained from the knowledge graph). and These are the model parameters sent from the cloud. Using the Sigmoid activation function, the output is the influence score for each dimension (range 0-1). In this example, the calculated result is... (This indicates that the delay in Agent_102's work order may exceed 20%), and this value exceeds the preset impact threshold. At the same time, Agent_101 detected the urgency level of the event. (Output from the urgency assessment model, taking into account remaining delivery time and production bottleneck pressure), but it did not exceed the urgency threshold. Therefore, Agent_101 determines that it should first attempt edge-to-edge local negotiation, rather than immediately reporting to the cloud.

[0071] The impact assessment model, deployed locally on each edge agent, is a lightweight neural network (or ensemble model) used to quickly predict the expected impact of an action (such as inserting an emergency work order, occupying shared resources, or adjusting process parameters) on other relevant edge agents (upstream feeding equipment, downstream receiving equipment, and equipment competing for shared resources) before the edge agent is about to perform an action. Its output directly influences whether the edge agent triggers local negotiation or directly reports to the cloud.

[0072] Let the current edge agent be (Referred to as the "initiator"), its candidate actions are: (For example, "Immediately occupy shared cache B") 5. (Continue for 15 minutes). For any target edge agent ( The input feature vector of the influence assessment model It consists of the following four parts: Initiator State Characteristics : Including the length of its own task queue Total processing load (minutes), equipment health (1 represents complete health) The urgency level of the current process (Normalized remaining delivery time pressure).

[0073] Candidate action features Encode the action type using one-hot encoding (e.g., resource consumption, speed change, order insertion, etc.), and simultaneously encode the expected duration of the action. Key resource number occupied .

[0074] Target observation features Edge intelligent agents Information about knowledge graphs or collaborative communication modules can be obtained through knowledge graphs or collaborative communication modules. Partial information, including Task queue length Current bottleneck status (Is it a bottleneck device?) and Material flow dependence intensity (Values ​​range from 0 to 1, with 1 indicating a strong dependency).

[0075] Contextual features Current system global clock tick and overall order delivery pressure (Normalized values ​​from 0 to 1), overall cache free rate, etc.

[0076] By concatenating all features Dimensional design is .

[0077] The influence assessment model employs a multilayer perceptron (MLP) with an attention mechanism, outputting a scalar influence score. The structure is as follows: First layer (fully connected + ReLU): The output dimension is 64.

[0078] Second layer (fully connected + ReLU): The output dimension is 32.

[0079] The third layer (fully connected + Sigmoid): ,in Output a score between 0 and 1.

[0080] If it is necessary to evaluate the impact on multiple target agents simultaneously, parallel computation can be performed (in which case the input features) Unchanged, only (Successive replacements) to obtain the influence score vector .

[0081] The impact assessment model is trained offline by a cloud-based agent and then distributed to each edge agent. The training data comes from historical records of the actual impact of actions performed by edge agents on other devices during production processes.

[0082] (ii) Structured message sending and replying between edge devices Agent_101 sends a structured intent-impact statement message (JSON format) to the affected Agent_102. An example of the message content is as follows: {"sender_id":"Agent_101", "action":"occupy_buffer_B3", "start_time":"T+5min", "duration":"15min", "impact":{"delay_to_agent_102":"+8min"}, "suggestions":["delay_agent_101_occupy_by_5min","use_alternative_buffer_B4"]}.

[0083] After receiving the message, Agent_102 will respond within the preset time limit. Within seconds, Agent_102 performs a rapid assessment based on its local status (the critical materials for its pending work order are currently stored in cache B-3; if these materials are occupied, subsequent processes will be blocked). Agent_102 simulates two options locally using a digital twin: accept the occupation (which will delay its own completion by 10 minutes) or refuse the occupation (Agent_101 will look for other options). After the simulation, Agent_102 generates a "counter-suggestion message": {"sender_id":"Agent_102", "response_type":"counter_proposal", "counter_action":"agent_101_occupy_B3_start_at_T+10min", "expected_impact":{"delay_to_agent_101":"+2min","delay_to_agent_102":"+3min"}}.

[0084] If Agent_102 is unable to respond within the time limit due to a malfunction or an urgent task (such as being in a critical process), it will be considered as a rejection of the negotiation by default.

[0085] (III) Simulation Comparison and Scheme Selection by the Initiator After receiving the counter-proposal, Agent_101 simulates and compares the original solution in its local digital twin. Immediately occupy cache area B-315 minutes) and the counter-proposal ( The simulation evaluates the execution effect of a 5-minute delay followed by a 10-minute execution. A comprehensive benefit function is used for the assessment. :

[0086] in, The action improves the completion time of the equipment itself (a negative value indicates a deterioration). The estimated impact on the system's global objectives (such as total delivery cycle) is obtained through impact chain reasoning in the knowledge graph. This represents the additional delay caused by negotiation. Weighting is applied. It emphasizes that global optimization is superior to local interests.

[0087] Simulation results show that: ; The counter-proposal offers a higher overall benefit (smaller negative value), so Agent_101 chooses to adopt the counter-proposal, sends a confirmation message to Agent_102, and adjusts its own action to "delay until T+10min to start occupying B-3".

[0088] (iv) Reporting of unsuccessful negotiations or emergency events If this negotiation exceeds the preset negotiation timeout period If no agreement is reached within seconds (e.g., both parties repeatedly propose mutually exclusive solutions), or Agent_101 detects the urgency of the event. Exceeding the preset emergency threshold (For example, if an emergency order causes the equipment to run idle, Agent_101 will immediately terminate the negotiation, encapsulate the conflict as a collaborative event, and report it to the cloud-based intelligent agent.)

[0089] The encapsulation format for collaborative events (standardized JSON) includes the following fields: {"event_id":"evt_20250427_001", "source_agent":"Agent_101", "type":"resource_conflict", "conflict_resource":"buffer_B3", "involved_agents":["Agent_101","Agent_102"], "local_attempts":[ {"proposal":"occupy_B3_at_T+5min","response":"counter_delay"}, {"proposal":"use_B4","response":"reject_due_to_space"} ], "urgency_score": 0.96, "estimated_global_impact":"delay_critical_order#P99_by_25min"}.

[0090] Agent_101 publishes the aforementioned events to the cloud agent via a publish-subscribe topic (e.g., / coordination / urgent_events) using the collaborative communication module. Upon subscribing to this topic, the cloud agent immediately triggers the global digital twin simulation and multi-agent reinforcement learning scheduling process (see the previous implementation details), arbitrates and generates a global optimization scheduling instruction (e.g., forcing Agent_101 to use the backup buffer B-4, while adjusting the work order order for Agent_102 to compensate for the 3-minute waiting time), and then distributes the instruction to the relevant edge agents for execution.

[0091] (v) Summary of successful negotiation and reporting conditions The above process demonstrates that decentralized negotiation between edge agents takes precedence over cloud reporting. Edge agents only report collaborative events when negotiation times out and fails to reach a consensus, when the urgency of the event exceeds a threshold, or when the conflict involves global resources (such as a shared inventory center). This mechanism significantly reduces cloud computing load and communication overhead while ensuring the system's rapid response capabilities.

[0092] This embodiment continues the workshop scenario described above: the cloud-based intelligent agent receives data from the edge intelligent agent. The reported collaborative event encapsulates an urgent job insertion request (new work order JOB). (The process route, expected completion time, and priority weights for step 999). The cloud-based intelligent agent then executes the following two main steps.

[0093] Step 1: Evaluate the impact of the event through knowledge graph simulation. 1.1 Constructing a knowledge graph for the production system The cloud-based intelligent agent first replicates the current global digital twin state, constructing an isolated simulation environment. The core of this environment is a production system knowledge graph, denoted as... Knowledge graphs consist of two levels: Static Ontology Layer: Defines the basic concepts, attributes, and invariant relationships of the production line. For example, entity types include equipment, work orders, processes, and materials; relationship types include equipment-availability-capacity, work orders-demand-process, and process-consumption-material. This layer forms the system's basic architecture and remains unchanged during production.

[0094] Dynamic Instance Layer: Instantiates the current production snapshot within the framework of the static ontology. For example, device M 101 is currently in a "busy" state, and the work order (JOB) is active. 888's current progress is 45%. The dynamic instance layer features dynamic relationship edges, whose weights or attributes are dynamically updated as the simulation clock advances or events are injected. For example, a work order (JOB). 888 and Equipment M There is a "pending processing" relationship edge between 101, which has an attribute "estimated waiting time". Its value will be continuously recalculated as new events are injected and the simulation progresses.

[0095] 1.2 Event Injection and Disturbance Potential Energy Calculation The cloud-based intelligent agent uses a context-aware inferential event injector to transform received collaborative events (urgent order insertion requests) into event triples with spatiotemporal constraints and influencing factors. The triple takes the following form:

[0096] Write the triple into the dynamic instance layer of the knowledge graph: that is, instantiate a new work order entity JOB. 999, and establish dynamic relationship edges between it and the required equipment and materials. After writing, the system automatically calculates the perturbation potential energy of this event on each existing relationship edge in the graph. Define relation edges (Connecting devices) With work orders The perturbation potential energy of ) is:

[0097] in: The urgency of the event (0.95); It is the number of critical resources that the relation edge depends on (for example, if a process requires both machine tools and special fixtures to be used simultaneously, then...). ); It is the buffer time of the process corresponding to the current relation edge (i.e., the difference between the latest start time and the current simulation clock, in minutes). It is a small constant (take) To avoid division by zero.

[0098] According to calculations, JOB 999 and Equipment M The relationship between 103 and the perturbation potential energy , with device M Between 104 The perturbation potential energy of M is less than 0.5, while that of other devices is less than 0.5. This indicates that M... 103 and M 104 will become the bottleneck most severely affected by the incident.

[0099] 1.3 Multi-agent simulation and deduction The cloud-based intelligent agent creates agent intelligent agents for the device entities and work order entities in the knowledge graph respectively: Device agent intelligent agent (e.g.) Its strategic objective is to maximize its own utilization while meeting quality constraints. Each device agent can dynamically report the processing efficiency decay coefficient based on its health status obtained from its corresponding edge twin. (1 indicates complete health, 0 indicates complete failure). The equipment agent's actions include selecting the next work order to be processed, reporting estimated idle time, etc.

[0100] Work order agent intelligent agent (e.g.) Its strategic goal is to minimize its own completion time. The work order agent dynamically selects processing equipment based on the current global load (by querying the weights of all device relationship edges in the knowledge graph). Its decision can be rule-based (such as selecting the earliest idle equipment) or based on reinforcement learning training.

[0101] All agent agents run concurrently, driven by a simulated clock. The simulated clock advances in a discrete event manner, with each event (such as a device completing processing or a new work order arriving) triggering a decision by the relevant agent. Each agent agent autonomously observes the current state of the knowledge graph (e.g., obtaining estimated waiting times by querying the weights of dynamic relation edges), communicates with other agents (by writing to or reading shared relation edges in the graph), and executes actions (such as occupying equipment, starting processing, or releasing resources). Each action feeds back into the knowledge graph, updating the weights of dynamic relation edges in real time.

[0102] Assuming the simulation has gone through Wheel interaction (e.g.) (Discrete events) eventually emerge to predict the future system state. This state includes the task queue change curve of each device over time, the expected start / end time of each work order, the evolution of material inventory levels, and the occupancy records of key resources (such as buffers and fixtures).

[0103] 1.4 Multi-dimensional deconstruction and evaluation report generation After the simulation, the evaluator deconstructs the dynamic knowledge graph from three dimensions after deep traversal inference: Time dimension: Extract the estimated start / end time of all work orders and calculate the total delivery cycle variation. Average delay For example, insert jobs 999 requires 35 minutes (processing) + 12 minutes (inspection) + 20 minutes (assembly), which can be completed on time, but causes the original job order to be affected. 756 and JOB 782 delayed respectively minute, minute.

[0104] Resource dimension: Analyze the utilization rate change curves of each piece of equipment to identify bottleneck shifts. Assume the equipment... utilization rate The deduction shows that M The utilization rate of 103 increased from 78% to 96%, while M The utilization rate of 104 decreased from 72% to 68%, indicating that the bottleneck shifted from M. 101 moved to M 103. Meanwhile, the testing equipment T... The peak length of the inspection queue for item 202 reached 12 items, exceeding its normal capacity by 3 times.

[0105] Impact Dimension: Utilizing the impact chain tracing function in the knowledge graph. The evaluator automatically generates an impact chain report by traversing the dependencies of dynamic relationship edges. For example, an impact chain can be represented as:

[0106] The evaluator compiles all the above analysis results into a structured impact assessment report, the core contents of which include: List of devices with the highest disturbance potential energy: ; List of affected work orders: JOB 756 (delay 240min), JOB 782 (delay 150min), JOB 801 (delayed 30 minutes); Bottleneck transfer path: from M 101→M 103; Recommended intervention point: M Task queue reordering for 103, or enable standby device M. 106.

[0107] The report is stored in vector form and serves as one of the input features for the next reinforcement learning algorithm.

[0108] Step 2: Generate globally optimized scheduling instructions using multi-agent reinforcement learning with integrated hierarchical attention mechanisms. 2.1 Problem Formalization (Partially Observable Markov Decision Process) The cloud-based intelligent agent formalizes the collaborative scheduling problem between cross-edge intelligent agents as a partially observable Markov decision process, defined as a six-tuple. ,in: for This is a collection of decision-making agents, each corresponding to an edge agent (i.e., an equipment-level controller in the actual production site). In this embodiment... (5 machining centers + 2 testing equipment + 1 assembly line).

[0109] This is the global state space. A global state... It is an integrated vector of all production unit statuses, all work-in-process statuses, resource inventory statuses, etc. in the cloud-based digital twin, with hundreds of dimensions.

[0110] Each agent Only local observations can be obtained. . This includes the device's own bound state (such as task queue length, health, and current processing parameters) and limited "neighbor" information that can be perceived through the knowledge graph (such as the output buffer status of upstream devices and the idle signal of downstream devices). Formally, ,in It is an observation projection function.

[0111] Action space Each intelligent agent action This could be executing the next task, requesting materials, reporting estimated idle time to neighbors, or providing components of joint actions to the cloud scheduler during training.

[0112] State transition probability Joint actions are simulated and executed by a cloud-based digital twin simulation environment. Subsequently, the environment enters a new state according to physical laws and scheduling rules. The transfer process is nondeterministic because it includes random disturbances (such as random equipment degradation).

[0113] Team reward function This is directly linked to the global objective. The hierarchical differential reward used in this embodiment will be reflected in the specific calculation (see the policy network training section below for details).

[0114] 2.2 Policy Network Structure Integrating Hierarchical Attention Mechanism The cloud-based intelligent agent has a pre-trained policy network built in. ,in The innovation of this network lies in its three-layer structure of in-layer attention + output layers, as detailed below: Bottom layer: Individual encoders. Each intelligent agent Local observation First, individual feature vectors are encoded using a multilayer perceptron (MLP). :

[0115] Different agents can share encoder parameters (parameter sharing) to cope with scenarios with varying numbers of agents.

[0116] Middle layer: Shop floor level attention layer. This layer simulates the process of a "scheduler observing the shop floor globally." A learnable shop floor level context vector is introduced. Then calculate the characteristics of each individual. and Attention weights:

[0117] in It is a trainable weight matrix. These weights allow the network to dynamically focus on the critical components that have the greatest impact on the overall progress (e.g., M with the highest perturbation potential). 103). The weighted summation yields the workshop-level features. This feature will be concatenated into the subsequent representation of each agent.

[0118] Upper layer: Collaborative attention layer between agents. For each agent... Calculate its individual characteristics With all other intelligent agents Mutual attention between individual characteristics:

[0119] in and For learnable parameters, This represents vector concatenation. Then, the agent... The fusion features are represented as follows:

[0120] This representation enables the agent The decision-making process implicitly incorporates information about "neighbors" that have a significant impact on it (such as upstream feeding equipment and downstream receiving equipment).

[0121] Output layer: based on For each agent, a policy head network (another MLP) is used. Probability distribution of output actions:

[0122] For discrete action spaces, the output is a multinomial distribution; for continuous action spaces (such as process parameter adjustment), the output is a Gaussian distribution with mean and variance.

[0123] 2.3 Feedforward reasoning generates joint actions During the execution phase (online inference), the cloud-based agent loads the pre-trained policy network parameters. (The training process uses the MAPPO algorithm; see other implementation methods described above for details.) When generating scheduling instructions, the input consists of two parts: Real-time local observation of each edge agent (Obtained via the collaborative communication module).

[0124] Key features in the impact assessment report include: the device identifier with the highest disturbance potential energy and key conflict points marked in the impact chain. These features are encoded into vectors. And spliced ​​with the local observations of each agent, i.e. .

[0125] Will Input the policy network and perform one forward propagation: The underlying encoder will each Mapped to .

[0126] Mid-level workshop-level attention layer calculation and .

[0127] Upper-layer collaborative attention layer computation and .

[0128] The output layer is for each agent. Output action probability distribution .

[0129] The cloud-based agent selects the action with the highest probability from each distribution (or samples according to a pre-defined exploration strategy) to obtain the joint action. In this embodiment, the generated joint action includes: right Action "JOB" 756 postponed to JOB "Processing after 999", execution probability 0.85; right The action "increase the frequency of quality inspections and enable overtime mode" has an execution probability of 0.92. right Action "Prioritize JOB Assembly" 999, JOB The assembly start time of 756 is delayed by 4 hours, with an execution probability of 0.78. For other intelligent agents: Maintain the original schedule.

[0130] The cloud-based intelligent agent encapsulates this sequence of joint actions into globally optimized scheduling instructions, which are then published through the collaborative communication module. The subscribed topic is distributed to all relevant edge agents.

[0131] This embodiment continues the previous scenario: the cloud-based intelligent agent generates a global optimized scheduling command for an emergency order insertion event and distributes it to relevant edge intelligent agents through the collaborative communication module. Taking processing center M as an example... 103 bound edge intelligent agents Taking this as an example, we will explain in detail the processing procedure after receiving the instruction.

[0132] (a) Receiving and parsing instructions You have subscribed to a cloud-based scheduling instruction topic (e.g., / cloud / cmd / global_schedule). When a globally optimized scheduling instruction is released in the cloud... The instruction was received within 50ms. The instruction is a JSON message containing the following key fields: target_agent: "Agent_103"; action_type: "reorder_queue"; action_detail: This will generate the work order JOB. Insert 999 at the head of the task queue, and move the original header work order JOB to the top. 756 postponed to JOB Processing after 999; expected_start_time: "current time + 1 minute"; expected_duration: "35min"; global_constraints: Other dependencies (such as M) After processing 103 is completed, the workpiece must be sent to T within 2 minutes. 202 test).

[0133] First, perform syntax validation and integrity checks on the command to confirm that the command format is correct and the target is the local machine.

[0134] (ii) Local digital twin verification (confirmation of command feasibility) To prevent cloud commands from becoming invalid due to sudden changes in local conditions (such as sudden equipment failure or material shortage during command transmission), Before implementation, rapid verification and simulation must be performed in its local digital twin.

[0135] The local digital twin is a high-fidelity simulation environment that includes the kinematic model of its bound device, the tool wear model, a snapshot of the current task queue, and a real-time inventory of the local buffer. Copy the current device's real state (recorded as time). A separate simulation replica is constructed. Within this replica, the sequence of actions specified by the cloud-based instructions is applied to the task queue, and a verification time window is simulated forward. (This embodiment takes) (Minutes, covering the entire process of the instruction's expected execution).

[0136] Let the new task queue after the instruction is executed be... Each work order In device M The standard processing time on 103 is Considering the actual processing efficiency coefficient of the equipment. (Assessed in real-time by edge intelligence agents, reflecting effects such as tool wear and spindle temperature rise), then the work order... The estimated actual processing time is ,in ( (Indicates complete health).

[0137] The edge agent calculates the total completion time after executing instructions in the simulation. Risk of resource conflict Define a feasibility cost function. :

[0138] in: This is the latest allowed completion time for the instruction (given by the cloud based on the global objective). This refers to the overdue period; This indicates the availability of various resources. The demand for (such as coolant, tool life, and buffer space) exceeds the supply. It is an indicator function; if the task queue length exceeds the preset maximum value (e.g., 20), it takes the value 1; otherwise, it takes the value 0. For the weighting coefficients, take... .

[0139] The simulation was run in the local digital twin, and the calculations were obtained. .because Preset acceptable threshold ,and Therefore, the instruction is deemed feasible. If If the edge agent refuses to execute the command, it will immediately report a "command not feasible" collaborative event to the cloud, along with simulation results for the cloud to reschedule.

[0140] (iii) Execute instructions After confirming feasibility, The instructions are translated into specific device control commands. Based on the `action_detail` in the instruction, Send two commands to its associated PLC (Programmable Logic Controller): Reorder the task queue, JOB The 999 machining program is loaded into the spindle and machining begins immediately (execution time 35 minutes). The original work order JOB Processing item 756 is temporarily paused and marked as "Waiting - Priority Downgraded," pending JOB processing. Automatic feeding will occur after 999 is completed.

[0141] During the execution, The system continuously monitors the equipment status, collecting data such as spindle current, vibration, and temperature at 10ms intervals, and comparing the actual progress after command execution with the simulation prediction. If the deviation exceeds a set threshold (e.g., the actual processing time exceeds the predicted time by 20%), it triggers local autonomous adjustments (such as fine-tuning the feed rate) or reports to the cloud again.

[0142] (iv) Feedback on Implementation Results After the instruction is executed (i.e., JOB) 999 completed processing and from M 103 Uninstall) An execution feedback report is generated and uploaded to the cloud-based intelligent agent via the collaborative communication module in an encrypted manner. The report contains the following information: Command identifier (corresponding to the ID of the command issued from the cloud); Actual start time Compared with the actual end time ; Actual total processing time ; Abnormal events that occur during execution (such as whether there is a tool wear alarm or whether there is buffer blockage). Equipment health status updates (such as remaining tool life after machining, cumulative spindle load); Whether the instructions are fully followed (if deviations are caused by local autonomous adjustments, the reasons and results of the adjustments must be explained).

[0143] Set cloud command requirements for JOB 999 Completed within minutes (at standard efficiency). Actual efficiency may vary due to slight tool wear. Actual time consumed minutes, delay minute. The delay was documented in the feedback report and attributed to "tool wear causing a 3% decrease in efficiency".

[0144] In addition, the edge agent also uploads snapshots of key data from the execution process, including vibration spectrum feature vectors during processing. Spindle power curve, and final workpiece quality inspection results (such as surface roughness). m, within acceptable limits). This data will be used for subsequent cloud-based policy updates (such as adjusting M). An important sample of the 103 processing efficiency prediction model.

[0145] (v) Closed loop completed After receiving feedback, the cloud-based intelligent agent incorporates the execution result into the global digital twin model, updates the device status and work order progress, and may trigger a new round of global optimization (for example, if the feedback shows M). The efficiency of device 103 was lower than expected, and the cloud can allocate a longer processing time buffer to this device in subsequent scheduling. Simultaneously, the cloud will record the complete cycle of relevant data from the issuance of the command to the feedback, for use in analyzing the actual performance of the policy network.

[0146] Please refer to Figure 3 In one application scenario, the distributed production management and control method based on cloud-edge-device collaboration includes an initialization phase and a production phase.

[0147] The initialization phase includes: S1: Deploy a flexible production distributed control system on-site, build a production line-level digital twin model in the cloud, and build an equipment-level digital twin model at the edge to ensure that the system modules work normally.

[0148] S2: The edge agent detects the status of the production execution unit, encapsulates the necessary data fields, and reports the status to the remote agent in heartbeat mode.

[0149] S3: Cloud-based intelligent agent digital twin module, which synchronizes the data and status reported by distributed edge intelligent agents, and presents the current operational overview of the production workshop in a virtual and real synchronous manner.

[0150] Production stage: S4: The cloud-based intelligent agent receives the work plan issued by the production control MES system. Based on the current production task schedule and execution progress, the cloud-based intelligent agent's digital twin module simulates and deduces the feasibility of issuing new tasks by MES under the macro optimization objectives (shortest delivery cycle, shortest inventory time), decomposes them into process-level production plans, and issues the production plans to the edge intelligent agents.

[0151] S5: The edge agent receives the production task, evaluates its feasibility on its own digital twin platform based on the current state of the production execution unit, generates the work instructions for the unit, and controls the production execution unit to work.

[0152] S6: During normal production processing, the edge agent collects data from the production execution unit, including equipment status, material status, production cycle time, produced products, product quality, etc., and sends the data to the cloud agent after standardization.

[0153] S7: The cloud-based intelligent agent receives data reported from the edge, and on the one hand, it synchronizes the data to the digital twin module for virtual and real representation, and on the other hand, it reports the data to the production control MES system.

[0154] If the edge agent receives processing requests for work orders B and C while executing work order task A.

[0155] S8: Based on the processing progress of the current work order task A, the edge agent simulates and extrapolates the execution of work order tasks B and C on the digital twin model, obtains the start time and estimated completion time of the work order tasks, and reports the simulation results to the cloud agent.

[0156] S9: Edge agent, based on the order in the task queue, prepares to start the next processing task, including material allocation, process parameter confirmation, and downloading processing files.

[0157] Production Phase - Abnormal Scheduling: S10: The edge intelligent agent monitors the operating status of this unit in real time. When a local disturbance is detected (equipment malfunction, material malfunction, product qualification rate decline), the local intelligent decision-making agent is activated to make autonomous decisions and handle the situation in order to maintain production.

[0158] S11: If a local disturbance meets the preset event triggering conditions, the edge agent immediately sends an event notification to the cloud agent, which includes the disturbance type, impact assessment and measures already taken.

[0159] S12: The cloud-based intelligent agent receives event notifications, updates the global digital twin model, and triggers a re-simulation to assess the scope and extent of the impact of abnormal fluctuations. Based on the simulation results, it determines whether the global plan needs adjustment. If so, it generates new macro-level target instructions and sends them to the affected edge nodes.

[0160] S13: The cloud-based intelligent agent receives event notifications such as order insertion, order reduction, work order priority adjustment, and cross-warehouse material call. Based on the current production progress, it triggers the digital twin system to re-simulate, assesses the scope and degree of the impact of abnormal fluctuations, generates an adjusted work plan, and, under the condition of optimal global objectives, issues scheduling tasks to the corresponding edge intelligent agents.

[0161] S14: After receiving a new task, the relevant edge agent calls its own digital twin model to verify and evaluate it, outputs local decision opinions, adjusts the local execution strategy, realizes the re-coordination with the global goal, and enters the next cycle.

[0162] Production completed: S15: If there are no new processing tasks in the edge agent queue or a manual operation instruction is received, the edge agent controls the production execution unit to standby or stop, and production ends.

[0163] Scenario 1: Normal production and emergency orders, please refer to the following: Figure 4 : Task assignment (corresponding to process S4): The cloud-based intelligent agent receives production work orders from the MES system, including ordinary work orders P1, P2, P3 and emergency work orders P4.

[0164] The cloud-based intelligent agent digital twin module immediately switched to simulation inference mode, simulating the impact of inserting P4 on the current production plan within the model. The simulation revealed that, if the original plan were followed, P4 would result in severe delivery delays and lead to product backlogs and congestion in the L3 and L4 stages.

[0165] Global optimization and command issuance (corresponding process S13): The global optimization scheduling module is based on a multi-agent reinforcement learning algorithm to coordinate the goals of each edge agent and calculate a new global optimal schedule: "Let line L1 pause P2, produce the semi-finished product of P4 first, and notify lines L2 and L3 to prioritize the flow of P4."

[0166] The cloud distributes this adjusted process plan to the edge agents L1, L2, L3, and L4. The multi-agent reinforcement learning algorithm is the value function-based decomposition algorithm VDN / QMIX.

[0167] Edge verification and local decision-making (corresponding processes S5 and S8): After receiving the new instruction, edge agent L1 simulates the process in its job simulation module: pausing P2 and switching to P4 will require a 15-minute line changeover time. The simulation confirms its feasibility, and the estimated start and finish times for P4 are reported to the cloud agent.

[0168] The autonomous decision-making optimization module of edge agent L1 is then activated, using a genetic algorithm to locally rearrange the remaining tasks of P2 and the new tasks of P4 in order to minimize the impact of this line change on the overall equipment utilization.

[0169] At the same time, agent L1 automatically requests the material system to call the specific model of components required by P4.

[0170] Cloud-edge collaborative execution: Agent L1 shares the estimated completion time of P4, which is used by agents L2 and L3 to locally optimize their own production tasks. Edge agents execute according to the new plan. The cloud-based agent digital twin model is synchronized in real time, and the global view clearly displays the scheduling process path of P4.

[0171] Scenario 2: Source tracing and parameter self-adjustment of quality fluctuations (corresponding to process S10), please refer to... Figure 5 : The edge agent L3 detected a sudden increase in the solder joint defect rate of a batch of boards from line L1. Its data processing module extracted features from the defect images.

[0172] Agent L3 immediately reports the "pass rate decline" event to the cloud agent and sends a quality warning directly to agent L1 through the collaborative communication module, along with the defective feature data.

[0173] After receiving the warning, the edge agent L1 immediately calls the AI ​​model to analyze the correlation between the current process parameters and defective characteristics, and finds the main reasons affecting product quality defects.

[0174] The model deduced that the squeegee pressure during solder paste printing was slightly insufficient. Therefore, without cloud intervention, agent L1 invoked the mechanistic model simulation results, automatically fine-tuned the pressure parameters of the solder paste printer, and produced 10 test boards.

[0175] After the test board passed the L3 line test, the pass rate returned to normal. Agent A recorded the decision and effect of this "parameter self-adjustment" and reported it to the cloud for subsequent model iterations in the AI ​​model training and optimization module.

[0176] The AI ​​models used by the edge agent are delivered by the cloud agent and include a quality defect root cause analysis algorithm, a ReNet50 encoder anomaly detection algorithm, and a regression model of process parameters and quality pass rate.

[0177] Scenario 3: Self-healing and global rescheduling of equipment malfunctions. Please refer to [link / reference]. Figure 6 : Anomaly detection and local autonomy (corresponding to process S10): The edge agent L2's data processing module detected small but continuous temperature fluctuations in the fifth temperature zone in real time. Its lightweight AI model predicted an 85% risk of temperature deviation occurring within 20 minutes.

[0178] The L2 autonomous decision-making optimization engine immediately initiates local autonomy: First, the parameters of adjacent temperature zones are automatically fine-tuned in an attempt to compensate for the fluctuations in the 5th temperature zone.

[0179] Secondly, slightly reduce the conveying speed of the board currently being processed (belonging to work order P1) to extend the dwell time in the stable temperature zone and ensure welding quality.

[0180] At the same time, the "equipment performance degradation early warning" and "compensation measures already taken" are reported to the cloud through the multi-agent collaborative communication module.

[0181] Global event response and decision-making (corresponding processes S11, S12): The cloud-based fault status monitoring module received an alert, and the status of the L2 line device in the global digital twin model changed to "yellow alert".

[0182] Cloud-based global intelligent agent triggers re-simulation: Simulates the impact on the global plan if the L2 line needs to be shut down for maintenance within the next hour.

[0183] Simulation results show that without adjustment, P1 work order will be delayed and block all subsequent orders. Therefore, the global agent generates a new scheduling strategy: "Temporarily schedule some P3 work order tasks that do not have high requirements for furnace temperature to another backup reflow soldering line L2-B for production." Cross-side collaboration (corresponding to cross-side collaboration mechanism): The cloud-based agent distributes some tasks from P3 to the agents on the L2-B line. Agents L2 and L2-B share current process parameters and task status through a collaborative communication module, ensuring consistent product quality.

[0184] Although the present invention has been described in detail with reference to the accompanying drawings and preferred embodiments, the present invention is not limited thereto. Various equivalent modifications or substitutions can be made to the embodiments of the present invention by those skilled in the art without departing from the spirit and essence of the invention, and such modifications or substitutions should all be within the scope of the present invention. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should also be covered within the protection scope of the present invention.

Claims

1. A distributed production management and control method based on cloud-edge-device collaboration, characterized in that, include: The system constructs a cloud-based intelligent agent, multiple edge-based intelligent agents, and a collaborative communication module connecting the cloud-based intelligent agent and each edge-based intelligent agent. Each edge-based intelligent agent is uniquely bound to a production execution unit and is used to collect real-time data from that production execution unit and perform local autonomous decision-making. The cloud-based intelligent agent is used to achieve global optimization scheduling across production lines or workshops. The edge agent continuously collects real-time data from the production execution unit it is bound to, identifies the type and severity of local disturbances through a local disturbance identification model, and autonomously executes decision-making actions according to a preset autonomous strategy. When the edge agent determines that the local disturbance exceeds its autonomous capability or involves a global resource conflict, it reports the collaborative event to the cloud agent through the collaborative communication module. After receiving the collaborative event, the cloud-based intelligent agent assesses the impact of the collaborative event and generates a global optimization scheduling instruction based on the assessment result using a multi-agent reinforcement learning algorithm, which is then sent to the relevant edge agents through the collaborative communication module. After receiving the global optimization scheduling instruction, the edge agent confirms that the instruction is feasible, executes it, and feeds back the execution result to the cloud agent.

2. The method according to claim 1, characterized in that, The collaborative communication module adopts a hybrid communication architecture, including: The cloud-edge communication submodule adopts a publish-subscribe communication mode. The cloud agent and each edge agent publish and subscribe to messages through predefined topics, which are used to issue global scheduling instructions and report collaborative events. The edge-to-edge communication submodule is configured with an event-driven sparse negotiation protocol for decentralized local negotiation between edge agents.

3. The method according to claim 2, characterized in that, The sparse negotiation protocol includes: Each edge agent maintains an influence assessment model to evaluate in real time the expected impact of its upcoming actions on other edge agents and generate an influence score vector. When the influence score of an edge agent on any other edge agent exceeds a preset influence threshold, the initiating edge agent sends a structured message to the affected receiving edge agent. The message includes the sender ID, action description, expected impact, and suggested coordination scheme. After receiving the message, the receiving edge agent will reply with a counter-suggestion message or a confirmation message within a preset response time limit; if the receiving edge agent is unable to reply due to its own emergency situation, it will be regarded as a rejection by default. After receiving a response, the initiating edge agent simulates and compares the execution effects of the original solution and the counter-suggestion solution in the counter-suggestion message in its local digital twin, and selects the solution with the best estimated overall benefit; if the receiver does not reply within the timeout or replies with a rejection, the initiating agent directly executes the original solution or adjusts it according to the local strategy. If the two parties cannot reach an agreement within the preset negotiation timeout period, or if the initiating edge agent detects that the urgency of the event exceeds the preset urgency threshold, the initiating edge agent will encapsulate the conflict as a collaborative event and report it to the cloud agent for arbitration.

4. The method according to claim 1, characterized in that, Both the cloud-based intelligent agent and the edge-based intelligent agent have a digital twin dual-mode, which includes a production monitoring mode and a simulation reasoning mode. In the production monitoring mode, the production process is visualized in real time. In the simulation reasoning mode, the impact assessment and deduction of scenarios such as order insertion, order reduction, or equipment failure are performed.

5. The method according to claim 1, characterized in that, The edge agent continuously collects real-time data from its bound production execution unit, identifies the type and severity of local disturbances through a local disturbance identification model, and autonomously executes decision-making actions based on a pre-set autonomous strategy, including: Raw data is obtained by parsing device protocols or collecting data from sensors, and the device feature extraction model issued by the cloud-based intelligent agent is called to convert the raw data into feature vectors. The feature vector is simultaneously input into the rule matching layer and the lightweight AI model layer. The outputs of the two are weighted and fused or prioritized and compared to obtain the perturbation type, confidence level, and severity level. If the severity level is lower than the preset threshold, it is determined to be a low-complexity disturbance, and the corresponding preset action is retrieved from the local fast rule base and executed. If the severity level is equal to or higher than the preset threshold, it is determined to be a high-complexity perturbation. Local digital twin simulation is initiated, and a genetic algorithm or a local reinforcement learning policy network issued by the cloud agent is called to perform inference and evaluation on multiple candidate decision options. The scheme with the lowest estimated cost is selected for execution, and the execution result is uploaded to the cloud agent for policy update.

6. The method according to claim 5, characterized in that, The method for updating the policy by the cloud-based intelligent agent includes: Receive local processing data and execution result data uploaded by each edge intelligent agent; By aggregating all received data and using the team reward function as the optimization objective, the multi-agent reinforcement learning policy network carried within the cloud-based intelligent agent and the lightweight AI model distributed to the edge intelligent agent are retrained or incrementally fine-tuned. The updated model and policy are then distributed to the corresponding edge agents.

7. The method according to claim 6, characterized in that, The team reward function adopts a hierarchical differential reward design, specifically including: Define a global team reward to measure the achievement of overall production goals; Define a localized reward for each edge agent to measure the performance of its associated production execution unit; Calculate the differential reward for each edge agent, whereby the differential reward is equal to the difference between the global reward when the agent takes an actual action and the global reward when the agent takes a baseline action and the actions of other agents remain unchanged. The global team reward, local personalized reward, and differential reward are weighted and synthesized to obtain the total reward for each edge agent, incentivizing the agent to take actions that make a positive contribution to the global outcome.

8. The method according to claim 1, characterized in that, The impact of the cooperative event is assessed, and based on the assessment results, a multi-agent reinforcement learning algorithm is used to generate a globally optimized scheduling instruction, including: The impact of events is assessed through knowledge graph simulation and deduction. A multi-agent reinforcement learning algorithm with an integrated hierarchical attention mechanism is used to generate globally optimized scheduling instructions based on the impact of the events.

9. The method according to claim 8, characterized in that, The impact of events is assessed through knowledge graph simulation and deduction, including: A knowledge graph of the production system is constructed, which includes a static ontology layer and a dynamic instance layer. The dynamic instance layer has dynamic relation edges, and the weights of the dynamic relation edges are dynamically updated as the simulation progresses or events are injected. The collaborative events are converted into event triples with spatiotemporal constraints and influence factors through an event injector. The event triples are written into the knowledge graph, and the perturbation potential energy value of the collaborative events on each relation edge in the graph is calculated. For the device entity and work order entity in the knowledge graph, respectively create device agent intelligent agent and work order agent intelligent agent. Each agent intelligent agent makes autonomous decisions and updates the graph according to its own strategy and the current state of the graph. After multiple rounds of interaction, the inferred future system state is obtained. The future system state is deconstructed from the time, resource, and related impact dimensions using an evaluator, generating an impact assessment report; wherein the related impact dimension includes tracing the impact chain.

10. The method according to claim 9, characterized in that, A multi-agent reinforcement learning algorithm with an integrated hierarchical attention mechanism is used to generate globally optimized scheduling instructions based on the impact of the events, including: The collaborative scheduling problem of cross-edge agents is formalized as a partially observable Markov decision process, in which each edge agent corresponds to a decision agent, and the cloud digital twin provides the global state. Load a pre-trained policy network integrating a hierarchical attention mechanism, the policy network including: a bottom-level individual encoder, a middle-level workshop-level attention layer, an upper-level agent-to-agent collaborative attention layer, and an output layer; Using the disturbance potential energy and impact chain information in the impact assessment report as input features of the policy network, the probability distribution of output actions for the affected edge agents is obtained through forward reasoning. Actions are selected from this distribution to form joint actions, thereby generating a global optimization scheduling instruction.

Images