Industrial internet flexible discrete manufacturing system dynamic optimization and reconfiguration method

Abstract

The application discloses an industrial internet flexible discrete manufacturing system dynamic optimization reconstruction method, and belongs to the field of industrial internet intelligent manufacturing systems. First, a full-factor collaborative representation model of a flexible discrete intelligent manufacturing system is established, a model basis for full-factor collaboration of the flexible discrete intelligent manufacturing system is laid, then a full-process dynamic optimization reconstruction model of the flexible discrete intelligent manufacturing system is established, dynamic optimization adjustment of the flexible discrete intelligent manufacturing system is realized, then a large-scale production task-factor deployment adaptation strategy of the flexible discrete intelligent manufacturing system is designed, a task is deployed to the most suitable production factor for completion, finally, a large-scale production task-resource supply adaptation strategy of the flexible discrete intelligent manufacturing system is designed, appropriate production resources are allocated for each production task, the smooth implementation of the industrial internet flexible discrete manufacturing task is ensured, and the large-scale application and popularization of the industrial internet flexible discrete intelligent manufacturing system are promoted.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing systems in the industrial internet, and in particular to a dynamic optimization and reconfiguration method for flexible discrete manufacturing systems in the industrial internet. Background Technology

[0002] The Industrial Internet (IIoT) is a deep integration of next-generation information and communication technologies with industrial manufacturing systems. Based on networks, centered on platforms, driven by data, and guaranteed by security, it achieves intelligent connectivity of all production elements (people, machines, and materials) to form emerging business models and applications for intelligent development. It is a crucial foundation for building a manufacturing powerhouse. Flexible discrete manufacturing within the IIoT is a new production model developed to address the demands of mass customization. Characterized by flexibility, agility, and intelligence, it has become a key focus of advanced manufacturing. A key feature of flexible manufacturing is the dynamic optimization and reconfiguration of production lines, enabling rapid response to different needs and flexible adjustments to the entire production process. This requires comprehensive intelligent connectivity of all production elements to improve control over the production process. However, the intelligent connectivity of all production elements makes the full-process optimization and reconfiguration of the IIoT flexible discrete intelligent manufacturing system a dynamic and complex global challenge. If all production elements can be dynamically coordinated to meet production task requirements, the capacity for large-scale personalized customization can be enhanced. Therefore, the dynamic optimization and reconstruction of the entire process of the flexible discrete intelligent manufacturing system in the industrial internet requires the use of theories and technologies such as artificial intelligence, big data, learning optimization, and synergetics to build a data-knowledge-driven paradigm for the dynamic reconstruction of all production factors, thereby improving the lean management and control capabilities of the flexible discrete intelligent manufacturing system and enhancing the quality, economic benefits, and core competitiveness of manufactured products.

[0003] In flexible discrete intelligent manufacturing systems of the Industrial Internet, the complex structure and connections of the intelligent interconnection of all production elements, including people, machines, and materials, lead to difficulties in the convergence and stability of dynamic optimization and reconstruction of all production elements, making it difficult to guarantee the production efficiency of flexible discrete manufacturing. Therefore, it is urgent to construct efficient and stable theories and methods for dynamic optimization and reconstruction of all production elements, quantitatively evaluate the convergence and stability of the entire process reconstruction, and guide task resource adaptation and optimization algorithm design. Thus, the dynamic optimization and reconstruction of all production elements in flexible discrete intelligent manufacturing systems of the Industrial Internet involves numerous links and elements throughout the manufacturing process, requiring comprehensive consideration of two aspects: 1. How to accurately and efficiently coordinate all production elements to construct a dynamic optimization and reconstruction model for the entire process of production, transportation, and resource scheduling, laying a model foundation for the reconstruction of flexible discrete intelligent manufacturing systems of the Industrial Internet; 2. How to explore a collaborative optimization decision-making method for the entire process of flexible discrete intelligent manufacturing systems based on the dynamic optimization and reconstruction model of all production elements, solving the problem of accurate adaptation of large-scale "task-element-resource" and ensuring the smooth implementation of flexible discrete manufacturing tasks. Summary of the Invention

[0004] To address the challenges of complex structural relationships among all production elements (humans, machines, and materials) in flexible discrete intelligent manufacturing systems of the Industrial Internet, which lead to difficulties in convergence and poor stability in dynamic optimization and reconstruction throughout the entire process, this invention provides a dynamic optimization and reconstruction method for flexible discrete manufacturing systems of the Industrial Internet.

[0005] The technical solution adopted by this invention to solve the above-mentioned technical problems is: a dynamic optimization and reconfiguration method for a flexible discrete manufacturing system in the industrial internet, comprising the following steps:

[0006] Step 1: Establish a collaborative representation model of all production elements in a flexible discrete intelligent manufacturing system.

[0007] S1.1 Establish a representation model of all elements of the flexible discrete intelligent manufacturing system, encompassing the cloud, edge, and terminal.

[0008] Step 1: Divide the factors of production into perceptual factors, cognitive factors, decision-making factors, and control factors;

[0009] Step 2: Divide the element behaviors of perception elements, cognitive elements, decision-making elements, and control elements into five categories: identification behavior, state behavior, functional behavior, performance behavior, and task behavior. Obtain element behavior descriptions through identification information, state information, functional information, performance information, and task information. :

[0010] ...(1);

[0011] In equation (1), I represents identification behavior, S represents state behavior, F represents functional behavior, P represents performance behavior, and T represents task behavior.

[0012] Identification information is used to identify and distinguish different element behaviors, including names. ,serial number ,symbol ;

[0013] Status information is used to describe the current state or conditions of feature behavior, including feature location. Production status Conveying status Resource status ;

[0014] Functional information is used to describe the capabilities and uses of an element's behavior, including communication capabilities. Computational ability Production capacity Conveying capacity ;

[0015] Performance information is used to describe the execution requirements of element behavior, including process requirements. , communication requirements Calculation requirements Quality requirements ;

[0016] Task information describes the tasks assigned to feature behaviors, the status of feature behaviors during task assignment, and task execution status, including task type. Task execution time Mission execution location Task Objectives ;

[0017] S1.2 Constructing a "cloud-edge-device" full-element collaborative representation model for a flexible discrete intelligent manufacturing system based on collaborative dynamics.

[0018] Step 1: Based on the quantitative representation of element behavior, the full-element collaboration of the flexible discrete intelligent manufacturing system "cloud-edge-device" is represented as a three-dimensional topological structure of "perception-cognition-decision-control", thus obtaining the flexible manufacturing collaborative network;

[0019] The flexible manufacturing collaborative network consists of a target layer, a perception layer, a cognition layer, a decision-making layer, and a control layer, represented as follows: ,in number of floors , This indicates the information interaction relationships within and between layers, as well as the task allocation relationships; ,express The set of production factors at each level express The first in the layer The behavior of each factor of production express The quantity of production factors at each level; This indicates the interaction between production factors within a layer. This indicates cross-layer interaction between production factors at different levels;

[0020] Step 2: Establish a multi-layered flexible manufacturing collaborative network Collaborative relationship matrix:

[0021] ...(2);

[0022] In equation (2), Let be a hyperadjacency matrix, representing A flexible manufacturing collaborative network for layers;

[0023] This indicates a multi-layer flexible manufacturing collaborative network. The adjacency matrix of each layer in the middle, , This indicates the synergistic relationship between factors of production, when Representation elements With elements If there is a synergistic relationship, then ,otherwise ;

[0024] This indicates a multi-layer flexible manufacturing collaborative network. The adjacency matrix of the inter-layer network. , Indicates the first Layer elements and the Layer elements The synergistic relationship between them, when the first Layer elements and the Layer elements If there is a collaborative relationship, then ,otherwise ;

[0025] Step 3: Construct a distributed flexible manufacturing collaborative network composed of multiple production factors. Establish a distributed collaborative dynamics model of production factor behavior:

[0026] ...(3);

[0027] In equation (3), and This represents a vector field in the state space. Represents the element behavior control vector. Indicates the control objective. Indicates the control output function;

[0028] Decoupling equation (3) yields equation (4):

[0029] ...(4);

[0030] In equation (4), , , , These represent the autonomous control inputs of the elements of perception, cognition, decision-making, and control, respectively.

[0031] , , , These represent the outputs of perception, cognition, decision-making, and control elements, respectively.

[0032] , , , These respectively represent behavioral descriptions of the elements of perception, cognition, decision-making, and control;

[0033] This represents the rate of change in the behavior of perceived elements in relation to the output of perceived information and distributed cognition and decision-making effects. The impact;

[0034] This represents the rate of change in cognitive element behavior in relation to the output of cognitive information and the effects of distributed perception, decision-making, and control. The impact;

[0035] The rate of change in the behavior of decision-making elements is represented by the coupled output indicating the consistency between decision information and the effects of distributed perception, cognition, and control. The impact;

[0036] This represents the rate of change in the behavior of control elements in relation to control information and distributed cognition and decision-making effects. The impact;

[0037] Establish a finite-time collaborative consistency protocol for sensing elements:

[0038] ...(5);

[0039] Establish a finite-time consensus protocol for cognitive elements:

[0040] ...(6);

[0041] Establish a finite-time consensus protocol for decision-making elements:

[0042] ...(7);

[0043] Establish a finite-time consensus protocol for control elements:

[0044] ...(8);

[0045] In equations (5), (6), (7), and (8), , Representation elements The set of neighboring elements; It is a weighted function; , , , A positive constant coefficient gain; , For symbolic functions, ;

[0046] Step 4: Apply the finite-time collaborative consistency protocol to the distributed collaborative dynamics model of production factor behavior, and establish a full-process "cloud-edge-device" all-factor collaborative model for the flexible discrete intelligent manufacturing system. ,in Indicating a multi-layer flexible manufacturing collaborative network The optimal adjacency matrix of each layer, Indicating a multi-layer flexible manufacturing collaborative network The optimal adjacency matrix of the inter-layer network;

[0047] Step 2: Establish a dynamic optimization and reconfiguration model for the entire process and all elements of a flexible discrete intelligent manufacturing system.

[0048] S2.1 Real-time acquisition of all production elements in a flexible discrete intelligent manufacturing system

[0049] Step 1: Based on the all-factor collaborative representation model, the dynamic optimization and reconstruction of all factors is divided into four stages: collaborative perception, collaborative cognition, collaborative decision-making, and collaborative control. The prior information of the behaviors of perceived factors, cognitive factors, decision-making factors, and control factors is represented in the all-factor collaborative model. Time-based elements of behavior and historical information , , , ;

[0050] Step 2: In the collaborative perception stage, obtain intelligent elements. exist The perceptual elements of time and behavior:

[0051] ...(9);

[0052] In equation (9), Representing intelligent elements The collaborative sensing method. Indicating flexible discrete intelligent manufacturing systems in The state at any given moment, This indicates the number of intelligent elements in a flexible discrete intelligent manufacturing system.

[0053] Step 3: In the collaborative cognition stage, based on perception and behavior... and existing production knowledge To acquire intelligent elements exist Cognitive elements of time:

[0054] ...(10);

[0055] In equation (10), Representing intelligent elements Collaborative cognitive approaches;

[0056] Step 4: In the collaborative decision-making stage, based on cognitive behavior... To acquire intelligent elements exist Decision-making elements at any given moment:

[0057] ...(11);

[0058] In equation (11), Representing intelligent elements Collaborative decision-making methods;

[0059] Step 5: In the collaborative control phase, based on cognitive behavior... and decision-making behavior To acquire intelligent elements exist Time-based control elements of behavior:

[0060] ...(12);

[0061] In equation (12), Representing intelligent elements Collaborative control methods;

[0062] Step 6: Connect each smart element exist The set of production factor behaviors generated at any given time This indicates that the set of behaviors of production factors is updated in real time based on the decoupling equations of the distributed collaborative dynamics model;

[0063] In the current environmental state and control strategies Under the condition, the environmental state is controlled by the state transition function. To update, indicated as ;

[0064] S2.2 Establish a dynamic optimization and reconfiguration model for the entire process of a flexible discrete intelligent manufacturing system.

[0065] Step 1: Establish a production chain model that matches production tasks with production factors. :

[0066] ...(13);

[0067] In equation (13), Represents the set of production task objectives and constraints; It represents a space that includes perceptual elements, cognitive elements, decision-making elements, and control elements; For matching operators;

[0068] , representing a set of production chains, For the first One production chain;

[0069] , representing the edge set of dependencies between production chains;

[0070] This represents the set of production factors involved in the collaborative production chain. , , , , , They represent the first The number of synergies among the perceptual, cognitive, decision-making, and control elements;

[0071] Represents a set of edge attributes;

[0072] Step 2: In the multi-layer flexible manufacturing collaborative network model The connected subgraphs of the production chain pattern are obtained and saved as the connected subgraph search paradigm of the multi-layer flexible manufacturing collaborative network;

[0073] Step 3: Model the solution process of the production chain model as a multi-objective optimization problem:

[0074] ...(14);

[0075] In equation (14), Indicates the production chain model. Represents the production chain pattern space. Indicates the first One objective function, ;

[0076] The optimal production chain model is obtained by solving a multi-objective optimization problem. ;

[0077] Step 3: Establish a large-scale production task-factor deployment adaptation strategy for flexible discrete intelligent manufacturing systems.

[0078] S3.1, A Method for Understanding Production Tasks in Flexible Discrete Intelligent Manufacturing Systems Based on Prior Knowledge

[0079] Step 1: Define Existence Each prior knowledge of production includes production objectives, production constraints, and production costs, and is defined as follows: For a collection of production knowledge, the corresponding set of knowledge tags is: ;

[0080] Step 2: Based on existing production knowledge, calculate the similarity between two entities using cosine similarity to obtain similar entities; then, use entity alignment to eliminate the inconsistency of heterogeneous production knowledge in similar entities, and store the structured knowledge in a graph database to obtain production knowledge structure information.

[0081] Step 3: Based on the production knowledge structure information, using the Transformer-based bidirectional encoder BERT, a multi-head selection model is employed to extract production knowledge entities and relations. Furthermore, a self-attention layer is used to highlight the semantic representations of different parts of the production knowledge entities.

[0082] ...(15);

[0083] In equation (15), , , These represent the input keyword vectors for production knowledge entities. The dimension of the input;

[0084] Step 4: Perform multiple self-attention operations on the production knowledge structure information, employing a multi-head attention mechanism. We learn information from different representational subspaces of production knowledge, and then introduce the correlation between labels based on the sequence labeling layer of the Conditional Random Field (CRF). The CRF scoring function is:

[0085] ...(16);

[0086] In equation (16), Representing production knowledge In knowledge tags The score on Indicates from the label Move to label The score;

[0087] The probability of outputting a knowledge label sequence is:

[0088] ...(17);

[0089] In equation (17), Representing production knowledge Possible label sequences, For the production of knowledge The set of all possible label sequences;

[0090] According to probability Gain production knowledge With knowledge tags The relationships between them generate a production knowledge model. ;

[0091] Step 5: Utilize graph neural networks to integrate existing production knowledge models The production knowledge and relationships in the model are transformed into low-dimensional representations. Production structure information associated with the production knowledge is extracted through aggregation network operations, and semantic information of the production knowledge data is extracted using a bidirectional encoder. A multi-head attention mechanism is then used to weighted aggregate the production knowledge structure and semantic information to improve the production knowledge model. Complete the missing parts;

[0092] Step 6: Target production task Represented as ,in ...(18);

[0093] In equation (18), For the characterization function, For the prediction function:

[0094] ...(19);

[0095] ...(20);

[0096] S3.2 Establish a decomposition method for the dynamic optimization and reconstruction problem of all elements.

[0097] Step 1: Based on the production knowledge model The process involves screening and adapting the accessed production factor resources to generate a production chain model that meets multiple objectives and constraints. The solution process is a multi-objective optimization problem.

[0098] ...(twenty one);

[0099] Step 2, Definition Define the set of unit vectors in the production chain pattern space. Let be the unit vector of the task. yes A subset of, if it exists ,but ;

[0100] Step 3, Definition yes a subset of yes Except A collection of other production chain models, for and Perform interaction detection:

[0101] ...(twenty two);

[0102] ...(twenty three);

[0103] if ,but and There is no interaction between them;

[0104] Step 4, when and When there is an interaction, The production chain model is divided into three mutually exclusive subsets of the same size. , and And detect each subset with Interactions between them;

[0105] if If an interaction exists with a subset, then the corresponding subset is further partitioned and decomposed until all subsets interact with it. There is no interaction between them;

[0106] S3.3, Reinforcement Learning-Based Large-Scale Production Task-Element Deployment Adaptation Strategy for Cloud-Edge-Device Collaboration

[0107] Step 1: Model the task-element deployment adaptation and optimization decision-making process of "cloud-edge-device" collaboration as a Markov decision process. ,in For actions, it represents the computing resources, communication resources, and optimization problem decomposition results of the "cloud-edge-device" architecture; The action space represents the set of actions that place the sub-optimization problem in the "cloud-edge-device" framework for solution. Represents the state transition function; To express gratitude, Discount factor;

[0108] Step 2: Optimize the reward function using maximum entropy, adding the entropy determined by the action distribution to the original reward:

[0109] ...(26);

[0110] In equation (26), To control the random parameters of the optimal policy and , Representation Strategy In state Entropy in;

[0111] The state value function based on entropy is:

[0112] ...(27);

[0113] The entropy-based policy function is:

[0114] ...(28);

[0115] In equations (27) and (28), Let the action value function be... The distribution of state-action pairs; For the policy parameters, the objective function for their update is: ;

[0116] Step 3: Introduce the ratio of the action probability of the current policy to that of the prior policy. Define the objective function using a truncation function:

[0117] ...(29);

[0118] In equation (29), This is a positive number used to limit the magnitude of policy updates and maintain the stability of policy updates;

[0119] Step 4: Use the adaptive gradient ascent algorithm to ascend along the objective function. Gradient direction, iteratively update parameters The process continues until convergence, yielding a deployment and adaptation strategy for the sub-optimization problem across all elements of the "cloud-edge-device" architecture. ;

[0120] Step 4: Establish a large-scale production task-resource supply adaptation strategy for flexible discrete intelligent manufacturing systems.

[0121] S4.1 Establish a large-scale task resource demand prediction model based on knowledge distillation

[0122] Step 1, Definition Indicates need Resource-based sub-optimization task Each sub-optimization task Characterized as a length of Feature attention encoding Each dimension represents the weight of the corresponding feature, and the sub-optimization task is calculated. The Attention weights for each feature :

[0123] ...(30);

[0124] In equation (30), Representation of features Query items With search matrix The similarity is , and its dimension is , Indicates the batch size of the mini-batch samples;

[0125] Step 2: Subtask characteristics As input to the Teacher Network (TA), the sub-optimization task is output by optimizing the network weights. resource requirements Furthermore, the teacher network (TA) is transferred to the lightweight student network (ST) through knowledge distillation. Both networks share the underlying feature embedding layer, and a sub-optimization task is established. Student resource demand prediction network;

[0126] Step 3: Establish sub-optimization tasks The loss function of the student resource demand prediction network is:

[0127] ...(31);

[0128] ...(32);

[0129] ...(33);

[0130] In equations (31), (32), and (33), Indicates a shared layer. This represents the student network's private layer. This indicates a private layer in the teacher network. and These represent the logistic mappings before the softmax output in the student network and the teacher network, respectively. Table cross-entropy loss; This indicates a loss of oversight.

[0131] By minimizing the loss function, the parameters of the student network and the teacher network are made similar, thus obtaining a large-scale task resource demand prediction model.

[0132] S4.2 Establish a large-scale task-resource relationship situational model based on graph attention networks.

[0133] Step 1: Establish a two-layer diagram of sub-optimization tasks and resources, represented as follows: ,in and These represent the networks corresponding to the sub-optimization task layer and the resource layer, respectively. This indicates the relationship between the sub-optimization layer and the resource layer;

[0134] Step 2: Construct a graph attention network encoder using multiple stacked graph attention layers. The input layer consists of node features represented as follows: ,in Indicates the first Characteristics of individual optimization tasks Indicates the first Characteristics of each resource;

[0135] By fusing features of neighboring nodes using a shared attention mechanism, the first... Layer attention coefficient :

[0136] ...(34);

[0137] In equation (34), and These represent the sub-optimization task and resources, respectively. Represents the weight matrix. Represents the attention operator;

[0138] Step 3: Filter using the adjacency matrix to obtain the graph attention network after... The output features of the layer encoder are represented as This is used as input to the decoder to obtain the graph attention network after... The output of the layer decoder is represented as ;

[0139] Step 4: Establish the feature loss function :

[0140] ...(35);

[0141] In equation (35), ;

[0142] Step 5: Establish the structural loss function :

[0143] ...(36);

[0144] In equation (36), Represents a node The neighbors;

[0145] Step 6: Minimize the feature loss function Optimize the consistency between reconstructed nodes and original nodes, and minimize the structural loss function. Optimize the similarity between adjacent nodes;

[0146] Based on the training process that minimizes the loss function, the parameters of the encoder and decoder networks are updated, and the trained network parameters are saved to obtain the task-resource association situational representation model. ;

[0147] S4.3 Establish a precise task-resource supply matching strategy based on Stackelberg game theory.

[0148] Step 1: Model the system resources and computational resources of different layers of the flexible discrete intelligent manufacturing system as follows: A heterogeneous resource pool;

[0149] Step 2: Based on the resource demand prediction model and the task-resource relationship situational characterization model, establish a Stackelberg game model between the task controller and the resource manager. ,in and These represent the resource manager and the task controller, respectively. and These represent the resource manager pricing strategy and the task controller adaptation strategy, respectively. and These represent the utility function vectors for the resource manager and the sub-optimization task, respectively.

[0150] Step 3: Establish the utility function and perform optimization calculations until the utility function satisfies the desired performance. This ensures that there is a unique Stackelberg load balancer between the task controller and the resource manager;

[0151] Step 4: Model the task-resource supply problem as a multi-agent Markov decision process. , For each intelligent agent The state space includes resource requirements, remaining resources, and task deadlines; The action space includes task selection and element adaptation selection; Indicates the state transition probability; Represents the reward function;

[0152] The value function and policy are updated using Stackelberg equilibrium, and the optimal task-resource supply decision is obtained using a multi-agent deep reinforcement learning algorithm. The formula for calculating the Stackelberg Q-value function is as follows:

[0153] ...(37);

[0154] In equation (37), The Stackelberg value function is represented. , ;

[0155] Each agent Main critic network By minimizing the loss function on the parameters Update:

[0156] ...(38);

[0157] Each agent Update the main actor network using policy gradients parameter :

[0158] ...(39);

[0159] The parameters of the main actor network and the critic network will be periodically assigned to the parameters of their target network. and Until training is complete;

[0160] Step 5: Use the predicted resource requirements of the task and the relationship between sub-optimization tasks and resources as input parameters to the trained actor network to obtain real-time data on the agent. Optimal task - resource supply decision .

[0161] Preferably, in step 1 of S1.1, the physical form of the sensing element is a sensor mounted on the power unit, the rotating unit, or the working unit;

[0162] The physical form of cognitive elements is production and manufacturing equipment loaded with machine tool process models, CNC machining control models, and machine tool dynamics models;

[0163] The physical form of decision-making elements is decision-making software, including R&D design systems, flexible scheduling decision-making systems, and AGV routing planning systems;

[0164] The physical form of the control elements is the control system, including real-time adjustment control system, interpolation control system, and production flexible control system.

[0165] Preferably, in step 4 of S3.2, and The interaction between them is detected as follows:

[0166] ...(twenty four);

[0167] ...(25);

[0168] when ,but and There is no interaction between them.

[0169] Preferably, in step 1 of S4.3, system resources include computer storage resources, network resources, and production resources, and computing resources include device CPU resources and GPU resources.

[0170] According to the above technical solution, the beneficial effects of the present invention are:

[0171] This invention first establishes a collaborative representation model of all production elements in a flexible discrete intelligent manufacturing system, laying the foundation for the collaborative model of all production elements in such a system. Then, it establishes a dynamic optimization and reconfiguration model for the entire process of the flexible discrete intelligent manufacturing system, enabling dynamic optimization and adjustment. Next, it designs a large-scale production task-element deployment adaptation strategy for the flexible discrete intelligent manufacturing system, deploying tasks to the most suitable production elements for completion. Finally, it designs a large-scale production task-resource supply adaptation strategy for the flexible discrete intelligent manufacturing system, allocating appropriate production resources for each production task to ensure the smooth implementation of flexible discrete manufacturing tasks in the Industrial Internet, and promoting the large-scale application and promotion of flexible discrete intelligent manufacturing systems in the Industrial Internet. Attached Figure Description

[0172] Figure 1 A collaborative representation model of all elements of a flexible discrete intelligent manufacturing system;

[0173] Figure 2 A dynamic optimization and reconstruction model for the entire process and all elements of a flexible discrete intelligent manufacturing system. Detailed Implementation

[0174] This embodiment presents a dynamic optimization and reconfiguration method for a flexible discrete manufacturing system in the Industrial Internet, comprising the following steps:

[0175] Step 1: Establish a collaborative representation model of all production elements in a flexible discrete intelligent manufacturing system.

[0176] S1.1 Establish a representation model of all elements of the flexible discrete intelligent manufacturing system, encompassing the cloud, edge, and terminal.

[0177] Step 1: Divide the factors of production into perceptual factors, cognitive factors, decision-making factors, and control factors.

[0178] The physical form of the sensing element is a sensor mounted on a power unit, a rotating device, or a working device.

[0179] The physical form of cognitive elements is production and manufacturing equipment loaded with machine tool process models, CNC machining control models, and machine tool dynamics models.

[0180] The physical form of decision-making elements is decision-making software, including R&D design systems, flexible scheduling decision-making systems, and AGV routing planning systems.

[0181] The physical form of the control elements is the control system, including real-time adjustment control system, interpolation control system, and production flexible control system.

[0182] Step 2: Divide the element behaviors of perception elements, cognitive elements, decision-making elements, and control elements into five categories: identification behavior, state behavior, functional behavior, performance behavior, and task behavior. Obtain element behavior descriptions through identification information, state information, functional information, performance information, and task information. :

[0183] ...(1);

[0184] In equation (1), I represents identification behavior, S represents state behavior, F represents functional behavior, P represents performance behavior, and T represents task behavior.

[0185] Identification information is used to identify and distinguish different element behaviors, including names. ,serial number ,symbol .

[0186] Status information is used to describe the current state or conditions of feature behavior, including feature location. Production status Conveying status Resource status .

[0187] Functional information is used to describe the capabilities and uses of an element's behavior, including communication capabilities. Computational ability Production capacity Conveying capacity .

[0188] Performance information is used to describe the execution requirements of element behavior, including process requirements. , communication requirements Calculation requirements Quality requirements .

[0189] Task information describes the tasks assigned to feature behaviors, the status of feature behaviors during task assignment, and task execution status, including task type. Task execution time Mission execution location Task Objectives .

[0190] S1.2 Constructing a "cloud-edge-device" full-element collaborative representation model for a flexible discrete intelligent manufacturing system based on collaborative dynamics.

[0191] Step 1: Based on the quantitative representation of element behavior, the full-element collaboration of the flexible discrete intelligent manufacturing system "cloud-edge-device" is represented as a three-dimensional topological structure of "perception-cognition-decision-control", thus obtaining the flexible manufacturing collaborative network.

[0192] The flexible manufacturing collaborative network consists of a target layer, a perception layer, a cognition layer, a decision-making layer, and a control layer, represented as follows: ,in number of floors , This indicates the information interaction relationships within and between layers, as well as the task allocation relationships; ,express The set of production factors at each level express The first in the layer The behavior of each factor of production express The quantity of production factors at each level; This indicates the interaction between production factors within a layer. This indicates cross-layer interaction between production factors at different levels.

[0193] Step 2: Establish a multi-layered flexible manufacturing collaborative network Collaborative relationship matrix:

[0194] ...(2);

[0195] In equation (2), Let be a hyperadjacency matrix, representing A flexible manufacturing collaborative network for layers.

[0196] This indicates a multi-layer flexible manufacturing collaborative network. The adjacency matrix of each layer in the middle, , This indicates the synergistic relationship between factors of production, when Representation elements With elements If there is a synergistic relationship, then ,otherwise .

[0197] This indicates a multi-layer flexible manufacturing collaborative network. The adjacency matrix of the inter-layer network. , Indicates the first Layer elements and the Layer elements The synergistic relationship between them, when the first Layer elements and the Layer elements If there is a collaborative relationship, then ,otherwise .

[0198] Step 3: Construct a distributed flexible manufacturing collaborative network composed of multiple production factors. Establish a distributed collaborative dynamics model of production factor behavior:

[0199] ...(3);

[0200] In equation (3), and This represents a vector field in the state space. Represents the element behavior control vector. Indicates the control objective. This represents the function that controls the output.

[0201] Decoupling equation (3) yields equation (4):

[0202] ... (4).

[0203] In equation (4), , , , These represent the autonomous control inputs of the elements of perception, cognition, decision-making, and control, respectively.

[0204] , , , These represent the outputs of perception, cognition, decision-making, and control elements, respectively.

[0205] , , , These represent behavioral descriptions of the elements of perception, cognition, decision-making, and control, respectively.

[0206] This represents the rate of change in the behavior of perceived elements in relation to the output of perceived information and distributed cognition and decision-making effects. The impact.

[0207] This represents the rate of change in cognitive element behavior in relation to the output of cognitive information and the effects of distributed perception, decision-making, and control. The impact.

[0208] The rate of change in the behavior of decision-making elements is represented by the coupled output indicating the consistency between decision information and the effects of distributed perception, cognition, and control. The impact.

[0209] This represents the rate of change in the behavior of control elements in relation to control information and distributed cognition and decision-making effects. The impact.

[0210] Establish a finite-time collaborative consistency protocol for sensing elements:

[0211] ... (5).

[0212] Establish a finite-time consensus protocol for cognitive elements:

[0213] ... (6).

[0214] Establish a finite-time consensus protocol for decision-making elements:

[0215] ... (7).

[0216] Establish a finite-time consensus protocol for control elements:

[0217] ... (8).

[0218] In equations (5), (6), (7), and (8), , Representation elements The set of neighboring elements; It is a weighted function; , , , A positive constant coefficient gain; , For symbolic functions, .

[0219] Step 4: Apply the finite-time collaborative consistency protocol to the distributed collaborative dynamics model of production factor behavior, and establish a full-process "cloud-edge-device" all-factor collaborative model for the flexible discrete intelligent manufacturing system. ,in Indicating a multi-layer flexible manufacturing collaborative network The optimal adjacency matrix of each layer, Indicating a multi-layer flexible manufacturing collaborative network The optimal adjacency matrix of the inter-layer network.

[0220] Step 2: Establish a dynamic optimization and reconfiguration model for the entire process and all elements of a flexible discrete intelligent manufacturing system.

[0221] S2.1 Real-time acquisition of all production elements in a flexible discrete intelligent manufacturing system

[0222] Step 1: Based on the all-factor collaborative representation model, the dynamic optimization and reconstruction of all factors is divided into four stages: collaborative perception, collaborative cognition, collaborative decision-making, and collaborative control. The prior information of the behaviors of perceived factors, cognitive factors, decision-making factors, and control factors is represented in the all-factor collaborative model. Time-based elements of behavior and historical information , , , .

[0223] Step 2: In the collaborative perception stage, obtain intelligent elements. exist The perceptual elements of time and behavior:

[0224] ...(9);

[0225] In equation (9), Representing intelligent elements The collaborative sensing method. Indicating flexible discrete intelligent manufacturing systems in The state at any given moment, This indicates the number of intelligent elements in a flexible discrete intelligent manufacturing system.

[0226] Step 3: In the collaborative cognition stage, based on perception and behavior... and existing production knowledge To acquire intelligent elements exist Cognitive elements of time:

[0227] ...(10);

[0228] In equation (10), Representing intelligent elements A collaborative cognitive approach.

[0229] Step 4: In the collaborative decision-making stage, based on cognitive behavior... To acquire intelligent elements exist Decision-making elements at any given moment:

[0230] ...(11);

[0231] In equation (11), Representing intelligent elements A collaborative decision-making approach.

[0232] Step 5: In the collaborative control phase, based on cognitive behavior... and decision-making behavior To acquire intelligent elements exist Time-based control elements of behavior:

[0233] ...(12);

[0234] In equation (12), Representing intelligent elements The collaborative control method.

[0235] Step 6: Connect each smart element exist The set of production factor behaviors generated at any given time This indicates that the set of behaviors of production factors is updated in real time based on the decoupling equations of the distributed collaborative dynamics model.

[0236] In the current environmental state and control strategies Under the condition, the environmental state is controlled by the state transition function. To update, indicated as .

[0237] S2.2 Establish a dynamic optimization and reconfiguration model for the entire process of a flexible discrete intelligent manufacturing system.

[0238] Step 1: Establish a production chain model that matches production tasks with production factors. :

[0239] ...(13);

[0240] In equation (13), Represents the set of production task objectives and constraints; It represents a space that includes perceptual elements, cognitive elements, decision-making elements, and control elements; For matching operators.

[0241] , representing a set of production chains, For the first A production chain.

[0242] , representing the edge set of dependencies between production chains.

[0243] This represents the set of production factors involved in the collaborative production chain. , , , , , They represent the first The number of synergies among the perceptual, cognitive, decision-making, and control elements.

[0244] Represents a set of edge attributes.

[0245] Step 2: In the multi-layer flexible manufacturing collaborative network model The connected subgraphs of the production chain pattern are obtained and saved as the search paradigm of the connected subgraphs of the multi-layer flexible manufacturing collaborative network.

[0246] Step 3: Model the solution process of the production chain model as a multi-objective optimization problem:

[0247] ...(14);

[0248] In equation (14), Indicates the production chain model. Represents the production chain pattern space. Indicates the first One objective function, .

[0249] The optimal production chain model is obtained by solving a multi-objective optimization problem. .

[0250] Step 3: Establish a large-scale production task-factor deployment adaptation strategy for flexible discrete intelligent manufacturing systems.

[0251] S3.1, A Method for Understanding Production Tasks in Flexible Discrete Intelligent Manufacturing Systems Based on Prior Knowledge

[0252] Step 1: Define Existence Each prior knowledge of production includes production objectives, production constraints, and production costs, and is defined as follows: For a collection of production knowledge, the corresponding set of knowledge tags is: .

[0253] Step 2: Based on existing production knowledge, calculate the similarity between two entities using cosine similarity to obtain similar entities; then, use entity alignment to eliminate the inconsistency of heterogeneous production knowledge in similar entities, and store the structured knowledge in a graph database to obtain production knowledge structure information.

[0254] Step 3: Based on the production knowledge structure information, using the Transformer-based bidirectional encoder BERT, a multi-head selection model is employed to extract production knowledge entities and relations. Furthermore, a self-attention layer is used to highlight the semantic representations of different parts of the production knowledge entities.

[0255] ...(15);

[0256] In equation (15), , , These represent the input keyword vectors for production knowledge entities. The dimension is the input dimension.

[0257] Step 4: Perform multiple self-attention operations on the production knowledge structure information, employing a multi-head attention mechanism. We learn information from different representational subspaces of production knowledge, and then introduce the correlation between labels based on the sequence labeling layer of the Conditional Random Field (CRF). The CRF scoring function is:

[0258] ...(16);

[0259] In equation (16), Representing production knowledge In knowledge tags The score on Indicates from the label Move to label The score.

[0260] The probability of outputting a knowledge label sequence is:

[0261] ...(17);

[0262] In equation (17), Representing production knowledge Possible label sequences, For the production of knowledge The set of all possible label sequences.

[0263] According to probability Gain production knowledge With knowledge tags The relationships between them generate a production knowledge model. .

[0264] Step 5: Utilize graph neural networks to integrate existing production knowledge models The production knowledge and relationships in the model are transformed into low-dimensional representations. Production structure information associated with the production knowledge is extracted through aggregation network operations, and semantic information of the production knowledge data is extracted using a bidirectional encoder. A multi-head attention mechanism is then used to weighted aggregate the production knowledge structure and semantic information to improve the production knowledge model. Complete the missing parts.

[0265] Step 6: Target production task Represented as ,in ... (18).

[0266] In equation (18), For the characterization function, For the prediction function:

[0267] ...(19);

[0268] ... (20).

[0269] S3.2 Establish a decomposition method for the dynamic optimization and reconstruction problem of all elements.

[0270] Step 1: Based on the production knowledge model The process involves screening and adapting the accessed production factor resources to generate a production chain model that meets multiple objectives and constraints. The solution process is a multi-objective optimization problem.

[0271] ...(twenty one).

[0272] Step 2, Definition Define the set of unit vectors in the production chain pattern space. Let be the unit vector of the task. yes A subset of, if it exists ,but .

[0273] Step 3, Definition yes a subset of yes Except A collection of other production chain models, for and Perform interaction detection:

[0274] ...(twenty two);

[0275] ...(twenty three);

[0276] if ,but and There is no interaction between them.

[0277] Step 4, when and When there is an interaction, The production chain model is divided into three mutually exclusive subsets of the same size. , and And detect each subset with Interactions between them;

[0278] and The interaction between them is detected as follows:

[0279] ...(twenty four);

[0280] ...(25);

[0281] when ,but and There is no interaction between them.

[0282] The detection methods for the other two subsets are the same as similar.

[0283] if If an interaction exists with a subset, then the corresponding subset is further partitioned and decomposed until all subsets interact with it. There is no interaction between them.

[0284] S3.3, Reinforcement Learning-Based Large-Scale Production Task-Element Deployment Adaptation Strategy for Cloud-Edge-Device Collaboration

[0285] Step 1: Model the task-element deployment adaptation and optimization decision-making process of "cloud-edge-device" collaboration as a Markov decision process. ,in For actions, it represents the computing resources, communication resources, and optimization problem decomposition results of the "cloud-edge-device" architecture; The action space represents the set of actions that place the sub-optimization problem in the "cloud-edge-device" framework for solution. Represents the state transition function; To express gratitude, This is the discount factor.

[0286] Step 2: Optimize the reward function using maximum entropy, adding the entropy determined by the action distribution to the original reward:

[0287] ...(26);

[0288] In equation (26), To control the random parameters of the optimal policy and , Representation Strategy In state The entropy in.

[0289] The state value function based on entropy is:

[0290] ... (27).

[0291] The entropy-based policy function is:

[0292] ... (28).

[0293] In equations (27) and (28), Let the action value function be... The distribution of state-action pairs; For the policy parameters, the objective function for their update is: .

[0294] Step 3: Introduce the ratio of the action probability of the current policy to that of the prior policy. Define the objective function using a truncation function:

[0295] ...(29);

[0296] In equation (29), This is a positive number used to limit the magnitude of policy updates and maintain the stability of policy updates.

[0297] Step 4: Use the adaptive gradient ascent algorithm to ascend along the objective function. Gradient direction, iteratively update parameters The process continues until convergence, yielding a deployment and adaptation strategy for the sub-optimization problem across all elements of the "cloud-edge-device" architecture. .

[0298] Step 4: Establish a large-scale production task-resource supply adaptation strategy for flexible discrete intelligent manufacturing systems.

[0299] S4.1 Establish a large-scale task resource demand prediction model based on knowledge distillation

[0300] Step 1, Definition Indicates need Resource-based sub-optimization task Each sub-optimization task Characterized as a length of Feature attention encoding Each dimension represents the weight of the corresponding feature, and the sub-optimization task is calculated. The Attention weights for each feature :

[0301] ...(30);

[0302] In equation (30), Representation of features Query items With search matrix The similarity is , and its dimension is , This indicates the batch size of the mini-batch samples.

[0303] Step 2: Subtask characteristics As input to the Teacher Network (TA), the sub-optimization task is output by optimizing the network weights. resource requirements Furthermore, the teacher network (TA) is transferred to the lightweight student network (ST) through knowledge distillation. Both networks share the underlying feature embedding layer, and a sub-optimization task is established. A student resource demand prediction network.

[0304] Step 3: Establish sub-optimization tasks The loss function of the student resource demand prediction network is:

[0305] ...(31);

[0306] ...(32);

[0307] ...(33);

[0308] In equations (31), (32), and (33), Indicates a shared layer. This represents the student network's private layer. This indicates a private layer in the teacher network. and These represent the logistic mappings before the softmax output in the student network and the teacher network, respectively. Table cross-entropy loss; This indicates a loss of oversight.

[0309] By minimizing the loss function, the parameters of the student network and the teacher network are made similar, thus obtaining a large-scale task resource demand prediction model.

[0310] S4.2 Establish a large-scale task-resource relationship situational model based on graph attention networks.

[0311] Step 1: Establish a two-layer diagram of sub-optimization tasks and resources, represented as follows: ,in and These represent the networks corresponding to the sub-optimization task layer and the resource layer, respectively. This indicates the relationship between the sub-optimization layer and the resource layer.

[0312] Step 2: Construct a graph attention network encoder using multiple stacked graph attention layers. The input layer consists of node features represented as follows: ,in Indicates the first Characteristics of individual optimization tasks Indicates the first The characteristics of each resource.

[0313] By fusing features of neighboring nodes using a shared attention mechanism, the first... Layer attention coefficient :

[0314] ...(34);

[0315] In equation (34), and These represent the sub-optimization task and resources, respectively. Represents the weight matrix. This represents the attention operator.

[0316] Step 3: Filter using the adjacency matrix to obtain the graph attention network after... The output features of the layer encoder are represented as This is used as input to the decoder to obtain the graph attention network after... The output of the layer decoder is represented as .

[0317] Step 4: Establish the feature loss function :

[0318] ...(35);

[0319] In equation (35), .

[0320] Step 5: Establish the structural loss function :

[0321] ...(36);

[0322] In equation (36), Represents a node The neighbors.

[0323] Step 6: Minimize the feature loss function Optimize the consistency between reconstructed nodes and original nodes, and minimize the structural loss function. Optimize the similarity between adjacent nodes.

[0324] Based on the training process that minimizes the loss function, the parameters of the encoder and decoder networks are updated, and the trained network parameters are saved to obtain the task-resource association situational representation model. .

[0325] S4.3 Establish a precise task-resource supply matching strategy based on Stackelberg game theory.

[0326] Step 1: Model the system resources and computational resources of different layers of the flexible discrete intelligent manufacturing system as follows: The system resources include computer storage resources, network resources, and production resources, while the computing resources include device CPU resources and GPU resources.

[0327] Step 2: Based on the resource demand prediction model and the task-resource relationship situational characterization model, establish a Stackelberg game model between the task controller and the resource manager. ,in and These represent the resource manager and the task controller, respectively. and These represent the resource manager pricing strategy and the task controller adaptation strategy, respectively. and These represent the utility function vectors for the resource manager and the sub-optimization task, respectively.

[0328] Step 3: Each sub-optimization task has four stages: task upload, task transfer, task execution, and result download. However, there is no task transfer stage for isolated task nodes, and the result download stage only exists for the last-hop node. In the task-resource adaptation problem, adapting related sub-optimization tasks to the same resource pool can reduce task transfer costs. A utility function is established and optimized until the utility function is satisfied. This ensures that there is a unique Stackelberg load balancer between the task controller and the resource manager.

[0329] Step 4: Model the task-resource supply problem as a multi-agent Markov decision process. , For each intelligent agent The state space includes resource requirements, remaining resources, and task deadlines; The action space includes task selection and element adaptation selection; Indicates the state transition probability; This represents the reward function.

[0330] The value function and policy are updated using Stackelberg equilibrium, and the optimal task-resource supply decision is obtained using a multi-agent deep reinforcement learning algorithm. The formula for calculating the Stackelberg Q-value function is as follows:

[0331] ...(37);

[0332] In equation (37), The Stackelberg value function is represented. , .

[0333] Each agent Main critic network By minimizing the loss function on the parameters Update:

[0334] ... (38).

[0335] Each agent Update the main actor network using policy gradients parameter :

[0336] ... (39).

[0337] The parameters of the main actor network and the critic network will be periodically assigned to the parameters of their target network. and Until training is complete.

[0338] Step 5: Use the predicted resource requirements of the task and the relationship between sub-optimization tasks and resources as input parameters to the trained actor network to obtain real-time data on the agent. Optimal task - resource supply decision .

[0339] This embodiment addresses the agility and stability requirements of flexible manufacturing processes in the Industrial Internet. First, it establishes a collaborative representation model of all production elements in a flexible discrete intelligent manufacturing system, laying the foundation for the collaborative model of all production elements in a flexible discrete intelligent manufacturing system.

[0340] Secondly, a dynamic optimization and reconfiguration model for the entire process of a flexible discrete intelligent manufacturing system is established, which theoretically enables dynamic optimization and adjustment of the flexible discrete intelligent manufacturing system.

[0341] Then, a large-scale production task-factor deployment adaptation strategy is designed for flexible discrete intelligent manufacturing systems to deploy tasks to the most suitable production factors for completion.

[0342] Finally, a resource supply adaptation strategy for large-scale production tasks in the flexible discrete intelligent manufacturing system is designed to allocate appropriate production resources for each production task, ensuring the smooth implementation of industrial internet flexible discrete manufacturing tasks and promoting the large-scale application and promotion of industrial internet flexible discrete intelligent manufacturing systems.

Claims

1. A dynamic optimization and reconfiguration method for a flexible discrete manufacturing system in the Industrial Internet, characterized in that, Includes the following steps: Step 1: Establish a collaborative representation model of all production elements in a flexible discrete intelligent manufacturing system. S1.1 Establish a representation model of all elements of a flexible discrete intelligent manufacturing system, encompassing the cloud, edge, and terminal. Step 1: Divide the factors of production into perceptual factors, cognitive factors, decision-making factors, and control factors; Step 2: Divide the element behaviors of perception elements, cognitive elements, decision-making elements, and control elements into five categories: identification behavior, state behavior, functional behavior, performance behavior, and task behavior. Obtain element behavior descriptions through identification information, state information, functional information, performance information, and task information. : ...（1）； In equation (1), I represents identification behavior, S represents state behavior, F represents functional behavior, P represents performance behavior, and T represents task behavior. Identification information is used to identify and distinguish different element behaviors, including names. ,serial number ,symbol ; Status information is used to describe the current state or conditions of feature behavior, including feature location. Production status Conveying status Resource status ; Functional information is used to describe the capabilities and uses of an element's behavior, including communication capabilities. Computational ability Production capacity Conveying capacity ; Performance information is used to describe the execution requirements of element behavior, including process requirements. , communication requirements Calculation requirements Quality requirements ; Task information describes the tasks assigned to feature behaviors, the status of feature behaviors during task assignment, and task execution status, including task type. Task execution time Mission execution location Task Objectives ; S1.2 Constructing a "cloud-edge-device" full-element collaborative representation model for a flexible discrete intelligent manufacturing system based on collaborative dynamics. Step 1: Based on the quantitative representation of element behavior, the full-element collaboration of the flexible discrete intelligent manufacturing system "cloud-edge-device" is represented as a three-dimensional topological structure of "perception-cognition-decision-control", thus obtaining the flexible manufacturing collaborative network; The flexible manufacturing collaborative network consists of a target layer, a perception layer, a cognition layer, a decision-making layer, and a control layer, represented as follows: ,in number of floors , This indicates the information interaction relationships within and between layers, as well as the task allocation relationships; ,express The set of production factors at each level express The first in the layer The behavior of each factor of production express The quantity of production factors at each level; This indicates the interaction between production factors within a layer. This indicates cross-layer interaction between production factors at different levels; Step 2: Establish a multi-layered flexible manufacturing collaborative network Collaborative relationship matrix: ...（2）； In equation (2), Let be a hyperadjacency matrix, representing A flexible manufacturing collaborative network for layers; This indicates a multi-layer flexible manufacturing collaborative network. The adjacency matrix of each layer in the middle, , This indicates the synergistic relationship between factors of production, when Representation elements With elements If there is a synergistic relationship, then ,otherwise ; This indicates a multi-layer flexible manufacturing collaborative network. The adjacency matrix of the inter-layer network. , Indicates the first Layer elements and the Layer elements The synergistic relationship between them, when the first Layer elements and the Layer elements If there is a collaborative relationship, then ,otherwise ; Step 3: Construct a distributed flexible manufacturing collaborative network composed of multiple production factors. Establish a distributed collaborative dynamics model of production factor behavior: ...（3）； In equation (3), and This represents a vector field in the state space. Represents the element behavior control vector. Indicates the control objective. Indicates the control output function; Decoupling equation (3) yields equation (4): ...（4）； In equation (4), , , , These represent the autonomous control inputs of the elements of perception, cognition, decision-making, and control, respectively. , , , These represent the outputs of perception, cognition, decision-making, and control elements, respectively. , , , These respectively represent behavioral descriptions of the elements of perception, cognition, decision-making, and control; This represents the rate of change in the behavior of perceived elements in relation to the output of perceived information and distributed cognition and decision-making effects. The impact; This represents the rate of change in cognitive element behavior in relation to the output of cognitive information and the effects of distributed perception, decision-making, and control. The impact; The rate of change in the behavior of decision-making elements is represented by the coupled output indicating the consistency between decision information and the effects of distributed perception, cognition, and control. The impact; This represents the rate of change in the behavior of control elements in relation to control information and distributed cognition and decision-making effects. The impact; Establish a finite-time collaborative consistency protocol for sensing elements: ...（5）； Establish a finite-time consensus protocol for cognitive elements: ...（6）； Establish a finite-time consensus protocol for decision-making elements: ...（7）； Establish a finite-time consensus protocol for control elements: ...（8）； In equations (5), (6), (7), and (8), , Representation elements The set of neighboring elements; It is a weighted function; , , , A positive constant coefficient gain; , For symbolic functions, ; Step 4: Apply the finite-time collaborative consistency protocol to the distributed collaborative dynamics model of production factor behavior, and establish a full-process "cloud-edge-device" all-factor collaborative model for the flexible discrete intelligent manufacturing system. ,in Indicating a multi-layer flexible manufacturing collaborative network The optimal adjacency matrix of each layer, Indicating a multi-layer flexible manufacturing collaborative network The optimal adjacency matrix of the inter-layer network; Step 2: Establish a dynamic optimization and reconfiguration model for the entire process and all elements of a flexible discrete intelligent manufacturing system. S2.1 Real-time acquisition of all production elements in a flexible discrete intelligent manufacturing system Step 1: Based on the all-factor collaborative representation model, the dynamic optimization and reconstruction of all factors is divided into four stages: collaborative perception, collaborative cognition, collaborative decision-making, and collaborative control. The prior information of the behaviors of perceived factors, cognitive factors, decision-making factors, and control factors is represented in the all-factor collaborative model. Time-based elements of behavior and historical information , , , ; Step 2: In the collaborative perception stage, obtain intelligent elements. exist The perceptual elements of time and behavior: ...（9）； In equation (9), Representing intelligent elements The collaborative sensing method. Indicating flexible discrete intelligent manufacturing systems in The state at any given moment, This indicates the number of intelligent elements in a flexible discrete intelligent manufacturing system. Step 3: In the collaborative cognition stage, based on perception and behavior... and existing production knowledge To acquire intelligent elements exist Cognitive elements of time: ...（10）； In equation (10), Representing intelligent elements Collaborative cognitive approaches; Step 4: In the collaborative decision-making stage, based on cognitive behavior... To acquire intelligent elements exist Decision-making elements at any given moment: ...（11）； In equation (11), Representing intelligent elements Collaborative decision-making methods; Step 5: In the collaborative control phase, based on cognitive behavior... and decision-making behavior To acquire intelligent elements exist Time-based control elements of behavior: ...（12）； In equation (12), Representing intelligent elements Collaborative control methods; Step 6: Connect each smart element exist The set of production factor behaviors generated at any given time This indicates that the set of behaviors of production factors is updated in real time based on the decoupling equations of the distributed collaborative dynamics model; In the current environmental state and control strategies Under the condition, the environmental state is controlled by the state transition function. To update, indicated as ; S2.2 Establish a dynamic optimization and reconfiguration model for the entire process of a flexible discrete intelligent manufacturing system. Step 1: Establish a production chain model that matches production tasks with production factors. : ...（13）； In equation (13), Represents the set of production task objectives and constraints; It represents a space that includes perceptual elements, cognitive elements, decision-making elements, and control elements; For matching operators; , representing a set of production chains, For the first One production chain; , representing the edge set of dependencies between production chains; This represents the set of production factors involved in the collaborative production chain. , , , , , They represent the first The number of synergies among the perceptual, cognitive, decision-making, and control elements; Represents a set of edge attributes; Step 2: In the multi-layer flexible manufacturing collaborative network model The connected subgraphs of the production chain pattern are obtained and saved as the connected subgraph search paradigm of the multi-layer flexible manufacturing collaborative network; Step 3: Model the solution process of the production chain model as a multi-objective optimization problem: ...（14）； In equation (14), Indicates the production chain model. Represents the production chain pattern space. Indicates the first One objective function, ; The optimal production chain model is obtained by solving a multi-objective optimization problem. ; Step 3: Establish a large-scale production task-factor deployment adaptation strategy for flexible discrete intelligent manufacturing systems. S3.1, A Method for Understanding Production Tasks in Flexible Discrete Intelligent Manufacturing Systems Based on Prior Knowledge Step 1: Define Existence Each prior knowledge of production includes production objectives, production constraints, and production costs, and is defined as follows: For a collection of production knowledge, the corresponding set of knowledge tags is: ; Step 2: Based on existing production knowledge, calculate the similarity between two entities using cosine similarity to obtain similar entities; then, use entity alignment to eliminate the inconsistency of heterogeneous production knowledge in similar entities, and store the structured knowledge in a graph database to obtain production knowledge structure information. Step 3: Based on the production knowledge structure information, using the Transformer-based bidirectional encoder BERT, a multi-head selection model is employed to extract production knowledge entities and relations. Furthermore, a self-attention layer is used to highlight the semantic representations of different parts of the production knowledge entities. ...（15）； In equation (15), , , These represent the input keyword vectors for production knowledge entities. The dimension of the input; Step 4: Perform multiple self-attention operations on the production knowledge structure information, employing a multi-head attention mechanism. We learn information from different representational subspaces of production knowledge, and then introduce the correlation between labels based on the sequence labeling layer of the Conditional Random Field (CRF). The CRF scoring function is: ...（16）； In equation (16), Representing production knowledge In knowledge tags The score on Indicates from the label Transfer to label The score; The probability of outputting a knowledge label sequence is: ...（17）； In equation (17), Representing production knowledge Possible label sequences, For the production of knowledge The set of all possible label sequences; According to probability Gain production knowledge With knowledge tags The relationships between them generate a production knowledge model. ; Step 5: Utilize graph neural networks to integrate existing production knowledge models The production knowledge and relationships in the model are transformed into low-dimensional representations. Production structure information associated with the production knowledge is extracted through aggregation network operations, and semantic information of the production knowledge data is extracted using a bidirectional encoder. A multi-head attention mechanism is then used to weighted aggregate the production knowledge structure and semantic information to improve the production knowledge model. Complete the missing parts; Step 6: Target production task Represented as ,in ...(18); In equation (18), For the characterization function, For the prediction function: ...（19）； ...（20）； S3.2 Establish a decomposition method for the dynamic optimization and reconstruction problem of all elements. Step 1: Based on the production knowledge model The process involves screening and adapting the accessed production factor resources to generate a production chain model that meets multiple objectives and constraints. The solution process is a multi-objective optimization problem. ...（21）； Step 2, Definition Define the set of unit vectors in the production chain pattern space. Let be the unit vector of the task. yes A subset of, if it exists ,but ; Step 3, Definition yes a subset of yes Except A collection of other production chain models, for and Perform interaction detection: ...（22）； ...（23）； if ,but and There is no interaction between them; Step 4, when and When there is an interaction, The production chain model is divided into three mutually exclusive subsets of the same size. , and And detect each subset with Interactions between them; if If an interaction exists with a subset, then the corresponding subset is further partitioned and decomposed until all subsets interact with it. There is no interaction between them; S3.3, Reinforcement Learning-Based Large-Scale Production Task-Element Deployment Adaptation Strategy for "Cloud-Edge-Device" Collaboration Step 1: Model the task-element deployment adaptation and optimization decision-making process of "cloud-edge-device" collaboration as a Markov decision process. ,in For actions, it represents the computing resources, communication resources, and optimization problem decomposition results of the "cloud-edge-device" architecture; The action space represents the set of actions that place the sub-optimization problem in the "cloud-edge-device" framework for solution. Represents the state transition function; To express gratitude, Discount factor; Step 2: Optimize the reward function using maximum entropy, adding the entropy determined by the action distribution to the original reward: ...（26）； In equation (26), To control the random parameters of the optimal policy and , Representation Strategy In state Entropy in; The state value function based on entropy is: ...（27）； The entropy-based policy function is: ...（28）； In equations (27) and (28), Let the action value function be... The distribution of state-action pairs; For the policy parameters, the objective function for their update is: ; Step 3: Introduce the ratio of the action probability of the current policy to that of the prior policy. Define the target function using a truncation function: ...（29）； In equation (29), This is a positive number used to limit the magnitude of policy updates and maintain the stability of policy updates; Step 4: Use the adaptive gradient ascent algorithm to ascend along the objective function. Gradient direction, iteratively update parameters The process continues until convergence, yielding a deployment and adaptation strategy for the sub-optimization problem across all elements of the "cloud-edge-device" architecture. ; Step 4: Establish a large-scale production task-resource supply adaptation strategy for flexible discrete intelligent manufacturing systems. S4.1 Establish a large-scale task resource demand prediction model based on knowledge distillation Step 1, Definition Indicates need Resource-based sub-optimization task Each sub-optimization task Characterized as a length of Feature attention encoding Each dimension represents the weight of the corresponding feature, and the sub-optimization task is calculated. The Attention weights for each feature : ...（30）； In equation (30), represents the similarity between the query item of the feature and the search matrix, and its dimension is , which represents the batch size of the mini-batch samples; Step 2: Use the sub-task features as input to the teacher network (TA), optimize the network weights, output the resource requirements of the sub-optimization task, and transfer the teacher network (TA) to the lightweight student network (ST) through knowledge distillation. The two share the underlying feature embedding layer to establish a student resource requirement prediction network for the sub-optimization task. Step 3: Establish the loss function for the student resource demand prediction network of the sub-optimization task: ...（31）； ...（32）； ...（33）； In equations (31), (32), and (33), represents the shared layer, represents the private layer of the student network, represents the private layer of the teacher network, and represents the logistic mapping before the softmax output in the student network and teacher network, respectively; represents the cross-entropy loss; represents the supervised loss; By minimizing the loss function, the parameters of the student network and the teacher network are made similar, thus obtaining a large-scale task resource demand prediction model. S4.2 Establish a large-scale task-resource relationship situational model based on graph attention networks. Step 1: Establish a two-layer graph of sub-optimization task and resource, represented as , where and represent the networks corresponding to the sub-optimization task layer and the resource layer, respectively, and represents the relationship between the sub-optimization layer and the resource layer; Step 2: Construct a graph attention network encoder using multiple stacked graph attention layers. The input layer is represented by node features, where represents the node i.e., . Characteristics of individual optimization tasks Indicates the first Characteristics of each resource; By fusing features of neighboring nodes using a shared attention mechanism, the first... Layer attention coefficient : ...（34）； In equation (34), and These represent the sub-optimization task and resources, respectively. Represents the weight matrix. Represents the attention operator; Step 3: Filter using the adjacency matrix to obtain the graph attention network after... The output features of the layer encoder are represented as follows: This is used as input to the decoder to obtain the graph attention network after... The output of the layer decoder is represented as ; Step 4: Establish the feature loss function : ...（35）； In equation (35), ; Step 5: Establish the structural loss function : ...（36）； In equation (36), Represents a node The neighbors; Step 6: Minimize the feature loss function Optimize the consistency between reconstructed nodes and original nodes, and minimize the structural loss function. Optimize the similarity between adjacent nodes; Based on the training process that minimizes the loss function, the parameters of the encoder and decoder networks are updated, and the trained network parameters are saved to obtain the task-resource association situational representation model. ; S4.3 Establish a precise task-resource supply matching strategy based on Stackelberg game theory. Step 1: Model the system resources and computational resources of different layers of the flexible discrete intelligent manufacturing system as follows: A heterogeneous resource pool; Step 2: Based on the resource demand prediction model and the task-resource relationship situational characterization model, establish a Stackelberg game model between the task controller and the resource manager. ,in and These represent the resource manager and the task controller, respectively. and These represent the resource manager pricing strategy and the task controller adaptation strategy, respectively. and These represent the utility function vectors for the resource manager and the sub-optimization task, respectively. Step 3: Establish the utility function and perform optimization calculations until the utility function satisfies the desired performance. This ensures that there is a unique Stackelberg load balancer between the task controller and the resource manager; Step 4: Model the task-resource supply problem as a multi-agent Markov decision process. , For each intelligent agent The state space includes resource requirements, remaining resources, and task deadlines; The action space includes task selection and element adaptation selection; Indicates the state transition probability; Represents the reward function; The value function and policy are updated using Stackelberg equilibrium, and the optimal task-resource supply decision is obtained using a multi-agent deep reinforcement learning algorithm. The formula for calculating the Stackelberg Q-value function is as follows: ...（37）； In equation (37), The Stackelberg value function is represented. , ; Each agent Main critic network By minimizing the loss function on the parameters Update: ...（38）； Each agent Update the main actor network using policy gradients parameter : ...（39）； The parameters of the main actor network and the critic network will be periodically assigned to the parameters of their target network. and Until training is complete; Step 5: Use the predicted resource requirements of the task and the relationship between sub-optimization tasks and resources as input parameters to the trained actor network to obtain real-time data on the agent. Optimal task - resource supply decision .

2. The dynamic optimization and reconfiguration method for a flexible discrete manufacturing system in the industrial internet according to claim 1, characterized in that: In step 1 of S1.1, the physical form of the sensing element is a sensor mounted on the power unit, rotation unit, or working unit; The physical form of cognitive elements is production and manufacturing equipment loaded with machine tool process models, CNC machining control models, and machine tool dynamics models; The physical form of decision-making elements is decision-making software, including R&D design systems, flexible scheduling decision-making systems, and AGV routing planning systems; The physical form of the control elements is the control system, including real-time adjustment control system, interpolation control system, and production flexible control system.

3. The dynamic optimization and reconfiguration method for a flexible discrete manufacturing system in the industrial internet according to claim 1, characterized in that: In step 4 of S3.2, and The interaction between them is detected as follows: ...（24）； ...（25）； when ,but and There is no interaction between them.

4. The dynamic optimization and reconfiguration method for a flexible discrete manufacturing system in the industrial internet according to claim 1, characterized in that: In step 1 of S4.3, system resources include computer storage resources, network resources, and production resources, while computing resources include device CPU resources and GPU resources.

Images