Graph neural network scheduling method and system for multi-process production scheduling of aluminum rolling

Abstract

The present invention discloses a graph neural network scheduling method and system for aluminum rolling multi-process production scheduling, which belongs to the field of production scheduling technology, including: based on the constructed digital twin model of the aluminum rolling production line, combining the logical relationship and constraint conditions of the production process, constructing a dynamic graph model of the aluminum rolling process; using the graph neural network to extract dynamic graph features, generating a state vector that integrates local and global information, and combining the multi-objective optimization algorithm to generate an initial scheduling plan; when a disturbance occurs on the production line, the digital twin model is used to perceive and quantify the impact in real time, and the incremental graph neural network is used to update the dynamic graph model, optimize the initial plan, and verify the feasibility in the digital twin environment; the optimization plan is sent to the physical production line for execution, and the digital twin model and the process dynamic graph model are dynamically calibrated through a feedback mechanism; dynamic intelligent scheduling of aluminum rolling multi-process production is realized, and production efficiency and anti-disturbance capability are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of production scheduling, and specifically relates to a graph neural network scheduling method and system for multi-process production scheduling of aluminum rolling. Background Art

[0002] In the aluminum rolling production process, production scheduling needs to comprehensively consider multiple factors such as equipment status, order demand, and process requirements. However, existing graph neural networks have a response lag problem when dealing with dynamic events such as equipment failures and order changes in the production process. It is difficult to capture changes in the dependencies between processes in real time, resulting in inflexible and inefficient production scheduling. At the same time, traditional production scheduling methods find it difficult to effectively integrate expert experience, equipment parameters, and industry standards in the field of aluminum rolling, and are unable to convert implicit constraints into explicit rules, affecting the rationality and feasibility of scheduling plans. Summary of the Invention

[0003] In response to the shortcomings of the existing technology, the present invention proposes a graph neural network scheduling method and system for aluminum rolling multi-process production scheduling. Based on the constructed digital twin model of the aluminum rolling production line, combined with the logical relationship and constraint conditions of the production process, a dynamic graph model of the aluminum rolling process is constructed; the graph neural network is used to extract dynamic graph features, generate a state vector that integrates local and global information, and combine the multi-objective optimization algorithm to generate an initial scheduling plan; when a disturbance occurs on the production line, the digital twin model is used to perceive and quantify the impact in real time, and the incremental graph neural network is used to update the dynamic graph model, optimize the initial plan, and verify the feasibility in the digital twin environment; the optimization plan is sent to the physical production line for execution, and the digital twin model and the process dynamic graph model are dynamically calibrated through the feedback mechanism; dynamic intelligent scheduling of aluminum rolling multi-process production is realized, improving production efficiency and anti-disturbance capability.

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] The graph neural network scheduling method for multi-process production scheduling of aluminum rolling includes:

[0006] S1: Collect real-time data from the physical production line and build a digital twin model of the aluminum rolling production line;

[0007] S2: Based on the digital twin model of the aluminum rolling production line, combined with the logical relationships and constraints between the aluminum rolling production processes, a dynamic graph model of the aluminum rolling process is constructed;

[0008] S3: Graph neural networks are used to extract dynamic graph features of the aluminum rolling process, generate state vectors containing local and global information, and combine them with a multi-objective optimization algorithm to generate an initial scheduling plan;

[0009] S4: When a disturbance occurs on the physical production line, the digital twin model senses and quantifies the impact of the disturbance in real time. An incremental graph neural network is used to dynamically update the dynamic graph model of the aluminum rolling process and optimize the initial scheduling plan. Simultaneously, the optimized scheduling plan is verified in the digital twin environment.

[0010] S5: The optimized scheduling plan is sent to the physical production line for execution. Through the established feedback mechanism, the digital twin model of the aluminum rolling production line and the dynamic graph model of the aluminum rolling process are dynamically adjusted.

[0011] Specifically, the specific steps of S2 include:

[0012] S2.1: Decompose the aluminum rolling process into independent steps and collect real-time data from the digital twin model of the aluminum rolling production line;

[0013] S2.2: Determine the process nodes based on the disassembled independent processes and the collected real-time data Relationship with process , where V represents the vertex set of the dynamic graph model, E represents the edge set of the dynamic graph model, i and j represent the process node index values, represents the i-th process node;

[0014] S2.3: Based on process nodes Relationship with process , standardize the attribute information of each process node and construct the process node attribute vector , and arrange all node attribute vectors in columns to form a node feature matrix At the same time, the adjacency matrix is ​​constructed based on the process relationship set E , elements in A according to Confirm and satisfy ,in, represents the attribute vector of the nth process node, T represents transposition, n represents the number of process nodes, and d represents the number of attribute features contained in each process node. Represents the process node in the process relationship and process nodes The edge attribute vector between them, f represents the Gaussian mapping function;

[0015] S2.4: Process node attribute vector and edge attribute vector Expand the time dimension to obtain the process node status and edge status , where t represents the time step;

[0016] S2.5: Constructing a time series dataset using a sliding window mechanism , where T represents the length of the time series;

[0017] S2.6: Based on process node status and edge status , combined with the actual rules of aluminum rolling production, a state transfer function is set; the state transfer function includes a node state transfer function and an edge state transfer function;

[0018] S2.7: Based on the constraints, the state transfer function results are modified, and finally a directed acyclic graph that meets all the constraints is constructed to form a dynamic graph model of the aluminum rolling process; the constraints include timing constraints, process constraints, quality constraints, and resource constraints.

[0019] Specifically, the process of determining process nodes and process relationships includes:

[0020] Abstracting the basic process information and real-time status information into nodes in a dynamic graph model to obtain process nodes; the basic process information includes process name, processing time, and equipment requirements; the real-time status information includes whether processing is in progress and the remaining processing time;

[0021] The dependency between processes is abstracted as edges in a dynamic graph model to obtain process relationships; the weights of the edges represent the time interval requirements between processes and the logistics delivery efficiency.

[0022] Specifically, the specific steps of S3 include:

[0023] S3.1: Input the dynamic graph model of the aluminum rolling process constructed in S2 into the graph neural network, iteratively update the hidden state of the process node through multiple layers of graph convolutional layers, and obtain the hidden state vector of the i-th process node in the c+1 layer ;

[0024] S3.2: Extract upstream and downstream constraints of a single process through neighborhood aggregation to obtain local information;

[0025] S3.3: By stacking multiple layers of graph convolution, local information is transferred to the entire graph neural network, including node features of the global process dependency network , where L represents the number of layers of the graph neural network;

[0026] S3.4: At each time step t, perform global pooling on all process node features and combine them with the gated recurrent unit to obtain the global state vector corresponding to time step t ;

[0027] S3.5: Combine the process node features with the global state vector to obtain a multi-dimensional state vector ;

[0028] S3.6: Set the optimization objective function, input the encoded multi-dimensional state vector into the optimization objective function, and generate the initial scheduling plan by solving it. ;

[0029] The optimization objective function is to achieve the minimum target after weighted summation of minimizing the total production cycle, minimizing the equipment switching cost, and maximizing the order on-time rate.

[0030] Specifically, the specific steps of S3.2 include:

[0031] S3.21: For process nodes , define its neighborhood; the neighborhood includes the in-neighborhood and out of neighborhood ;

[0032] S3.22: Obtain the feature representation of each process node based on the process node attribute vector and edge attribute vector in S2.3;

[0033] S3.23: Perform inbound and outbound neighborhood aggregation using a weighted aggregation function to obtain aggregated neighborhood information;

[0034] S3.24: Fusion of node features with aggregated neighborhood information to obtain the final local information .

[0035] Specifically, the feature representation of each process node in S3.22 includes the node's own features, the incoming edge features, and the and out-edge features ;

[0036] The node's own characteristics are process node attribute vectors .

[0037] Specifically, the specific steps of S4 include:

[0038] S4.1: Obtain real-time data from the physical production line through sensor networks ;

[0039] S4.2: Comparison of digital twin predictions for an aluminum rolling line With real-time data , identify the time and location of the disturbance, and calculate the deviation between the actual value and the predicted value of the digital twin model of the aluminum rolling production line , represents the 2-norm;

[0040] S4.3: Set the disturbance threshold;

[0041] like , then the process node An exception occurs at the time step t, and the abnormal status indicator is displayed as 1, where Indicates process node The disturbance threshold of

[0042] like , then the process node It is normal at the time step t, and the abnormal status indicator is displayed as 0;

[0043] S4.4: Determine the disturbance location based on the disturbance threshold judgment result;

[0044] S4.5: According to the disturbance location, determine the affected process nodes and their neighborhoods, and update the subgraph parameters in the dynamic graph model of the aluminum rolling process affected by the disturbance. At the same time, combine the historical state with the current disturbance; the historical state includes the hidden state of the process nodes and the graph structure information of the dynamic graph model of the aluminum rolling process.

[0045] Specifically, the digital twin model of the aluminum rolling production line in S1 includes a three-dimensional geometric model, a physical mechanism model, a process logic model and a real-time data interface, and the digital twin model of the aluminum rolling production line is synchronized with the physical production line data.

[0046] A graph neural network scheduling system for multi-process production scheduling of aluminum rolling, including: digital twin model construction module, dynamic graph model construction module, graph neural network scheduling module, dynamic disturbance response module, feedback and adjustment module;

[0047] The digital twin model construction module is used to establish a virtual mapping of the physical production line, perform real-time data synchronization and production status simulation;

[0048] The dynamic graph model building module is used to abstract the aluminum rolling production process into a dynamic graph structure to capture the process node attributes, relationships and timing characteristics;

[0049] The graph neural network scheduling module is used to extract features based on the dynamic graph model and generate and optimize the initial scheduling plan;

[0050] The dynamic disturbance response module is used to perceive production disturbances in real time, dynamically update the scheduling plan and verify its feasibility;

[0051] The feedback and adjustment module is used to establish a closed-loop feedback mechanism and optimize the digital twin model and the dynamic graph model according to the production execution results.

[0052] The graph neural network scheduling module includes: a feature extraction unit, a global state modeling unit, a state fusion unit, and a multi-objective optimization unit;

[0053] The feature extraction unit is used to iteratively update the node hidden state through multiple layers of graph convolutional layers, aggregate neighborhood information, and extract local constraints and global dependencies of a single process;

[0054] The global state modeling unit is used to globally pool the node features of each time step, and generate a global state vector in combination with the gated recurrent unit to capture the overall timing characteristics of the production process;

[0055] The state fusion unit is used to combine the local features of the nodes with the global state vector to form a multi-dimensional state representation that includes process details and the overall process;

[0056] The multi-objective optimization unit defines an optimization objective function, inputs a multi-dimensional state vector into the function to solve it, and generates an initial scheduling plan.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] 1. The present invention proposes a graph neural network scheduling system for multi-process production scheduling of aluminum rolling, and optimizes and improves the architecture, operation steps and processes. The system has the advantages of simple process, low investment and operation costs, and low production work costs.

[0059] 2. The present invention proposes a graph neural network scheduling method for multi-process production scheduling of aluminum rolling. At the production scheduling optimization level, by constructing a digital twin model and a dynamic graph model, combining a graph neural network and a multi-objective optimization algorithm, it can effectively extract the local and global characteristics of the process, and generate an initial scheduling plan that takes into account the minimization of the total production cycle, the minimization of equipment switching costs, and the maximization of order punctuality, thereby realizing intelligent planning of production processes under multi-objective optimization and improving production efficiency and the rationality of resource allocation.

[0060] 3. The present invention proposes a graph neural network scheduling method for multi-process production scheduling of aluminum rolling. At the level of dynamic response and system optimization, when disturbances occur in the physical production line, the digital twin model is used to perceive and quantify the impact in real time, and the incremental graph neural network is used to dynamically update the model and scheduling plan. At the same time, the feasibility of the plan is verified through the digital twin environment, thereby enhancing the system's anti-interference ability and flexible production capacity. In addition, the establishment of a feedback mechanism enables the digital twin model and the dynamic graph model to be continuously optimized according to the production execution results, promotes the self-learning and adaptive evolution of the production scheduling system, and ensures the efficient, stable and high-quality operation of aluminum rolling production. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 Schematic diagram of the graph neural network scheduling method for multi-process production scheduling of aluminum rolling according to the present invention;

[0062] Figure 2This is a flow chart showing the principle of the graph neural network scheduling method for multi-process production scheduling of aluminum rolling according to the present invention;

[0063] Figure 3 This is the architecture diagram of the graph neural network scheduling system for multi-process production scheduling of aluminum rolling in the present invention. DETAILED DESCRIPTION

[0064] Example 1

[0065] See also Figure 1 and Figure 2 The present invention provides an embodiment of a graph neural network scheduling method for multi-process production scheduling of aluminum rolling, comprising the following steps:

[0066] S1: Collect real-time data from the physical production line and build a digital twin model of the aluminum rolling production line;

[0067] The digital twin model of the aluminum rolling production line in S1 includes a physical mechanism model and a real-time data interface, and the digital twin model of the aluminum rolling production line is synchronized with the physical production line data.

[0068] Furthermore, the specific steps of S1 include:

[0069] (1) Classify production line elements, including hardware equipment, material flow, process parameters and production orders;

[0070] (2) Collect real-time data from the physical production line;

[0071] (3) Based on the aluminum rolling process principle, a process physical model is established to describe the input-output relationship and dynamic characteristics; the process includes a rolling process physical model and an annealing process physical model; the rolling process physical model is: ; The physical model of the annealing process is: , where F is the rolling force, B is the width of the rolled piece, and l is the length of the contact arc. represents the mean flow stress, represents the roller radius, Indicates the difference between the thickness before rolling and the thickness after rolling. represents the annealing time, A represents a constant, represents the activation energy, represents the gas constant, Indicates absolute temperature;

[0072] Among them, activation energy is the minimum energy required for a system to change from an initial state to a transition state in a chemical reaction or physical process. Only when the reactant molecules obtain enough energy to reach the activated state can the reaction or process occur.

[0073] (4) Develop a data interface for the digital twin model to achieve two-way communication with the real-time data platform, including:

[0074] Input: Receive pre-processed real-time data and update model status, such as equipment temperature and material location;

[0075] Output: Output model simulation results, such as predicted rolling force and grain size after annealing, to the upper scheduling system;

[0076] (5) Deploy the digital twin model on the edge server or cloud, and realize remote monitoring and simulation through the industrial Internet platform;

[0077] (6) Establish a timed trigger, such as once per second, or event-triggered model update mechanism to ensure that the digital twin is synchronized with the physical production line status.

[0078] S2: Based on the digital twin model of the aluminum rolling production line, combined with the logical relationships and constraints between the aluminum rolling production processes, a dynamic graph model of the aluminum rolling process is constructed;

[0079] S3: Graph neural networks are used to extract dynamic graph features of the aluminum rolling process, generate state vectors containing local and global information, and combine them with a multi-objective optimization algorithm to generate an initial scheduling plan;

[0080] S4: When a disturbance occurs on the physical production line, the digital twin model senses and quantifies the impact of the disturbance in real time. An incremental graph neural network is used to dynamically update the dynamic graph model of the aluminum rolling process and optimize the initial scheduling plan. Simultaneously, the optimized scheduling plan is verified in the digital twin environment.

[0081] S5: The optimized scheduling plan is sent to the physical production line for execution. Through the established feedback mechanism, the digital twin model of the aluminum rolling production line and the dynamic graph model of the aluminum rolling process are dynamically adjusted.

[0082] Furthermore, the specific steps of S5 include:

[0083] (1) The optimized scheduling plan, such as process sequence, equipment allocation, and time arrangement, is sent to each device on the physical production line through the control system for execution;

[0084] (2) The production line starts production according to the scheduling plan, and the sensors collect the execution data of the physical production line in real time, including:

[0085] The actual start or end time of each process;

[0086] Equipment operating parameters, such as rolling speed and temperature;

[0087] Order fulfillment status, such as whether it was delivered on time;

[0088] Quality inspection results, such as plate thickness deviation and surface defects;

[0089] (3) Feedback the collected data to the digital twin model of the aluminum rolling production line through a real-time data interface to achieve data synchronization between the physical system and the virtual system;

[0090] (4) Compare the predicted values ​​of the digital twin model with the actual execution data of the physical production line and calculate the deviation value, including: adjusting the material deformation parameters and equipment operation parameters;

[0091] (5) Update the process node attribute vector and edge attribute vector based on the actual execution data;

[0092] (6) Adjust the process node state transfer function or edge state transfer function based on actual production rules;

[0093] (7) Simulate the updated aluminum rolling process dynamic graph model in the digital twin environment and use the aluminum rolling process dynamic graph model to predict the production cycle of the next order to complete the verification;

[0094] (8) If the verification passes, save the updated dynamic graph model of the aluminum rolling process; if the deviation still exceeds the preset threshold, repeat (4)-(5) until convergence.

[0095] The specific steps of S2 include:

[0096] S2.1: Decompose the aluminum rolling process into independent steps and collect real-time data from the digital twin model of the aluminum rolling production line;

[0097] Among them, the processes include billet preparation, rough rolling, intermediate rolling, finishing rolling, annealing, shearing, and quality inspection. The vertex set V of the dynamic graph model is defined according to the disassembled independent processes, and each vertex corresponds to a process node; the process relationship is determined based on the production process logic, and the edge set E is constructed.

[0098] S2.2: Determine the process nodes based on the disassembled independent processes and the collected real-time data Relationship with process , where V represents the vertex set of the dynamic graph model, E represents the edge set of the dynamic graph model, i and j represent the process node index values, represents the i-th process node;

[0099] It should be noted that the process relationship is also called the edge, so the process relationship , also called edge .

[0100] S2.3: Based on process nodes Relationship with process , standardize the attribute information of each process node and construct the process node attribute vector , and arrange all node attribute vectors in columns to form a node feature matrix At the same time, the adjacency matrix is ​​constructed based on the process relationship set E , elements in A according to Confirm and satisfy ,in, represents the attribute vector of the nth process node, T represents transposition, n represents the number of process nodes, and d represents the number of attribute features contained in each process node. Represents the process node in the process relationship and process nodes The edge attribute vector between them, f represents the Gaussian mapping function;

[0101] in, Used to describe process nodes and process nodes If the relationship between process nodes and process nodes There is a direct connection relationship between them, that is, edge (i, j) exists in the edge set, then To include process nodes and process nodes The vector of the edge attribute information between processes, such as the material transfer time and waiting time between processes. If there is no connection relationship, then Set as zero vector, it can be seen that the mathematical expression of aluminum rolling process information and logical relationship is realized through the node feature matrix and adjacency matrix.

[0102] S2.4: Process node attribute vector and edge attribute vector Expand the time dimension to obtain the process node status and edge status , where t represents the time step;

[0103] It should be noted that, considering that the aluminum rolling production process is a dynamically changing process, in order to accurately reflect the status of the process at different time stages, the process node attribute vector and edge attribute vector are expanded in the time dimension. Through the expansion of the time dimension, the static node and edge attributes are converted into process node states and edge states that change dynamically over time, so that the model can better adapt to the dynamic changes in the production process.

[0104] S2.5: Constructing a time series dataset using a sliding window mechanism , where T represents the length of the time series;

[0105] S2.6: Based on process node status and edge status , combined with the actual rules of aluminum rolling production, set the node state transfer function and edge state transfer function ,in, and represents the nonlinear transformation function, Indicates process node The neighborhood set of ;

[0106] S2.7: Based on the constraints, modify the state transfer function results and finally construct a directed acyclic graph that satisfies all the constraints. , forming a dynamic graph model of aluminum rolling process, in which, Represents the attribute feature matrix of all process nodes corresponding to time step t, The adjacency matrix representing all edges at the time step t, that is, the graph structure at the time step t; the constraints include timing constraints, process constraints, quality constraints, and resource constraints.

[0107] Among them, the timing constraint ensures that the processes are carried out in the prescribed order to avoid logical errors. For example, rough rolling must be carried out before finishing rolling. For example, if process node i is the immediate predecessor of process node j, then there is an edge ; Process constraints ensure that the production process meets the requirements of aluminum rolling process, such as restrictions on parameters such as rolling temperature and reduction; quality constraints ensure that product quality meets standards and standardize the processing accuracy and surface quality of each process; resource constraints rationally allocate equipment, manpower and other resources to avoid resource conflicts.

[0108] The process of determining process nodes and process relationships includes:

[0109] Abstracting the basic process information and real-time status information into nodes in a dynamic graph model to obtain process nodes; the basic process information includes process name, processing time, and equipment requirements; the real-time status information includes whether processing is in progress and the remaining processing time;

[0110] The dependency between processes is abstracted as edges in a dynamic graph model to obtain process relationships; the weights of the edges represent the time interval requirements between processes and the logistics delivery efficiency.

[0111] It should be noted that the basic information of the process belongs to the static attribute of the node, which is defined in the graph structure initialization stage. The real-time status information belongs to the dynamic attribute of the node, which is defined in the dynamic attribute modeling stage and updated with time steps. The dependency relationship between processes is represented by the non-zero elements of the adjacency matrix A, which is part of the graph structure initialization. The weight of the edge belongs to the dynamic attribute of the edge, which is defined in the dynamic attribute modeling stage and can change with the production status.

[0112] The specific steps of S3 include:

[0113] S3.1: Input the dynamic graph model of the aluminum rolling process constructed in S2 into the graph neural network, iteratively update the hidden state of the process node through multiple layers of graph convolutional layers, and obtain the hidden state vector of the i-th process node in the c+1 layer ,in, Represents an edge The attention weight, represents the learnable parameters, represents the hidden state vector of the j-th process node at the c-th layer, represents the learnable parameters, represents a nonlinear activation function;

[0114] S3.2: Extract upstream and downstream constraints of a single process through neighborhood aggregation to obtain local information;

[0115] S3.3: By stacking multiple layers of graph convolution, local information is transferred to the entire graph neural network, including node features of the global process dependency network , where L represents the number of layers of the graph neural network;

[0116] S3.4: At each time step t, by formula Perform global pooling on all process node features and combine them with the gated recurrent unit to obtain the global state vector corresponding to time step t ,in, represents the attention pooling function, represents a gated recurrent unit, Represents the global state vector at the time step t-1;

[0117] S3.5: Combine the process node features with the global state vector to obtain a multi-dimensional state vector ,in, represents the vector concatenation function, Represents the hidden state of the nodes in the last layer of the graph neural network;

[0118] S3.6: Set the optimization objective function and transform the multi-dimensional state vector After encoding, input into the optimization objective function, by solving Generate an initial scheduling plan ,in, It means finding the minimum function. represents the strategy space, Represents the scheduling policy, represents the mathematical expectation function, represents the discount factor, and , Represents the reward at the time step t;

[0119] The optimization objective function is to achieve the minimum target after weighted summation of minimizing the total production cycle, minimizing the equipment switching cost, and maximizing the order on-time rate.

[0120] Furthermore, the specific steps of S3.6 include:

[0121] (1) Determine the optimization objectives based on production requirements, assign weights to each objective, and construct a weighted sum optimization objective function;

[0122] (2) The process node features extracted by the graph neural network are spliced ​​with the global state vector to form a multi-dimensional state vector containing process attributes, temporal relationships, and global dependencies , and compressed into a low-dimensional feature vector through a fully connected layer or encoder as the input of the optimization objective function;

[0123] (3) Based on actual production conditions, define constraints, such as process timing constraints, equipment resource constraints, and quality standards, to ensure that the generated scheduling plan is feasible;

[0124] (4) Select a genetic algorithm to solve the optimization objective function and generate an initial scheduling plan that meets the constraints;

[0125] (5) Simulate the initial scheduling plan in the digital twin environment to verify whether it meets the process logic, resource constraints and quality requirements. If not, return to adjust the optimization parameters or algorithm.

[0126] The specific steps of S3.2 include:

[0127] S3.21: For process nodes , define its neighborhood; the neighborhood includes the in-neighborhood and out of neighborhood ,in, Indicates process node To process node The edge of , Indicates process node To process node The edge of ;

[0128] The incoming neighborhood refers to all upstream processes of the current node; the outgoing neighborhood refers to all downstream processes of the current node.

[0129] S3.22: Obtain the feature representation of each process node based on the process node attribute vector and edge attribute vector in S2.3;

[0130] The feature representation includes node features, edge features and out-edge features ;

[0131] The node's own characteristics are process node attribute vectors ;

[0132] Among them, the input edge feature represents the process node To process node The edge feature of the outgoing edge represents the process node To process node edge features.

[0133] S3.23: Perform inbound and outbound neighborhood aggregation using a weighted aggregation function to obtain aggregated neighborhood information;

[0134] Furthermore, we perform neighborhood aggregation: ,in, Indicates process node The in-neighborhood aggregation features, Indicates the upstream process node For the current process node The attention weight, Represents the multi-layer perceptron for upstream process nodes The features and edge features are transformed nonlinearly. Represents the multilayer perceptron function entering the neighborhood, Represents vector concatenation;

[0135] Perform neighborhood aggregation: ,in, Indicates process node The out-of-neighborhood aggregation features, Indicates the current process node For downstream process nodes The attention weight, Represents the multilayer perceptron for downstream process nodes The features and edge features are transformed nonlinearly. Multilayer perceptron function representing the neighborhood;

[0136] S3.24: Fusion of node features with aggregated neighborhood information to obtain the final local information ,in, Represents the multilayer perceptron aggregation function.

[0137] The specific steps of S4 include:

[0138] S4.1: Obtain real-time data from the physical production line through sensor networks ;

[0139] S4.2: Comparison of digital twin predictions for an aluminum rolling line With real-time data , identify the time and location of the disturbance, and calculate the deviation between the actual value and the predicted value of the digital twin model of the aluminum rolling production line , represents the 2-norm;

[0140] S4.3: Set the disturbance threshold;

[0141] like , then the process node An exception occurs at the time step t, and the abnormal status indicator is displayed as 1, where Indicates process node The disturbance threshold of

[0142] like , then the process node It is normal at the time step t, and the abnormal status indicator is displayed as 0;

[0143] S4.4: Determine the disturbance location based on the disturbance threshold judgment result;

[0144] S4.5: According to the disturbance location, determine the affected process nodes and their neighborhoods, and update the subgraph parameters in the dynamic graph model of the aluminum rolling process affected by the disturbance. At the same time, combine the historical state with the current disturbance; the historical state includes the hidden state of the process nodes and the graph structure information of the dynamic graph model of the aluminum rolling process.

[0145] Example 2

[0146] See also Figure 3 Another embodiment provided by the present invention is a graph neural network scheduling system for multi-process production scheduling of aluminum rolling, comprising:

[0147] Digital twin model construction module, dynamic graph model construction module, graph neural network scheduling module, dynamic disturbance response module, feedback and adjustment module;

[0148] A digital twin model building module is used to establish a virtual mapping of the physical production line, enabling real-time data synchronization and production status simulation;

[0149] A dynamic graph model building module is used to abstract the aluminum rolling production process into a dynamic graph structure to capture the process node attributes, relationships and timing characteristics;

[0150] The graph neural network scheduling module is used to extract features based on dynamic graph models, generate and optimize initial scheduling plans, and achieve multi-objective production optimization;

[0151] Dynamic disturbance response module, used to perceive production disturbances in real time, dynamically update scheduling plans and verify feasibility;

[0152] The feedback and adjustment module is used to establish a closed-loop feedback mechanism, optimize the digital twin model and dynamic graph model based on the production execution results, and improve the long-term scheduling accuracy of the system.

[0153] The dynamic graph model construction module includes: process decomposition unit, data acquisition unit, node definition unit, relationship definition unit, matrix construction unit, time series expansion unit, and constraint correction unit;

[0154] Process disassembly unit, used to disassemble the entire process into independent processes;

[0155] Data acquisition unit, used to extract real-time data from the digital twin model;

[0156] Node definition unit, used to abstract basic process information and real-time status into graph vertices;

[0157] The relationship definition unit is used to abstract the dependencies between processes into graph edges, and the edge weights represent the time interval and logistics efficiency;

[0158] The matrix construction unit is used to standardize node attributes to form a node feature matrix, and to construct an adjacency matrix based on process relationships to describe the connection strength between nodes.

[0159] The time series expansion unit is used to expand the time dimension of node and edge attributes to generate a state sequence with multiple time steps;

[0160] The constraint correction unit is used to correct the state transition results by combining timing, process, quality, and resource constraints to ensure that the generated directed acyclic graph conforms to actual production rules.

[0161] The graph neural network scheduling module includes: feature extraction unit, global state modeling unit, state fusion unit, and multi-objective optimization unit;

[0162] Feature extraction unit, which is used to iteratively update the node hidden state through multiple layers of graph convolutional layers, aggregate neighborhood information, and extract local constraints and global dependencies of a single process;

[0163] The global state modeling unit is used to globally pool the node features of each time step and generate a global state vector in combination with the gated recurrent unit to capture the overall timing characteristics of the production process;

[0164] The state fusion unit is used to combine the local features of the nodes with the global state vector to form a multi-dimensional state representation that includes process details and the overall process;

[0165] The multi-objective optimization unit defines the optimization objective function, inputs the multi-dimensional state vector into the function to solve it, generates an initial scheduling plan, and balances production efficiency and cost.

[0166] The dynamic disturbance response module includes: disturbance identification unit, threshold judgment unit, dynamic graph update unit, and digital twin verification unit;

[0167] The disturbance identification unit is used to collect physical production line data in real time through a sensor network, compare it with the predicted value of the digital twin model, and identify the time, location and degree of deviation of the disturbance;

[0168] The threshold judgment unit sets the disturbance threshold of the process node, judges whether the node is abnormal by comparing the actual deviation with the threshold, and locates the affected process and neighborhood;

[0169] The dynamic graph update unit is used to update the subgraph parameters affected by the disturbance, combine the historical status, and use the incremental graph neural network to re-optimize the scheduling plan to reduce the impact of the disturbance on production;

[0170] The digital twin verification unit is used to simulate the optimized scheduling plan in a virtual environment to verify its feasibility and effectiveness, avoiding the risk of directly sending it to the physical production line.

[0171] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific embodiments. The above-mentioned specific embodiments are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also change, modify, replace and modify the above-mentioned embodiments without departing from the purpose and scope of protection of the present invention. These are all protected by the present invention.

Claims

1. A graph neural network scheduling method for multi-process production scheduling of aluminum rolling, characterized by: include: S1: Collect real-time data from the physical production line and build a digital twin model of the aluminum rolling production line; S2: Based on the digital twin model of the aluminum rolling production line, combined with the logical relationships and constraints between the aluminum rolling production processes, a dynamic graph model of the aluminum rolling process is constructed; S3: Graph neural networks are used to extract dynamic graph features of the aluminum rolling process, generate state vectors containing local and global information, and combine them with a multi-objective optimization algorithm to generate an initial scheduling plan; S4: When a disturbance occurs on the physical production line, the digital twin model senses and quantifies the impact of the disturbance in real time. An incremental graph neural network is used to dynamically update the dynamic graph model of the aluminum rolling process and optimize the initial scheduling plan. Simultaneously, the optimized scheduling plan is verified in the digital twin environment. S5: Send the optimized scheduling plan to the physical production line for execution. Through the established feedback mechanism, dynamically adjust the digital twin model of the aluminum rolling production line and the dynamic graph model of the aluminum rolling process. The specific steps of S3 include: S3.1: Input the dynamic graph model of the aluminum rolling process constructed in S2 into the graph neural network, iteratively update the hidden state of the process node through multiple layers of graph convolutional layers, and obtain the hidden state vector of the i-th process node in the c+1 layer ; S3.2: Extract upstream and downstream constraints of a single process through neighborhood aggregation to obtain local information; S3.3: By stacking multiple layers of graph convolution, local information is transferred to the entire graph neural network, including node features of the global process dependency network , where L represents the number of layers of the graph neural network; S3.4: At each time step t, perform global pooling on all process node features and combine them with the gated recurrent unit to obtain the global state vector corresponding to time step t ; S3.5: Combine the process node features with the global state vector to obtain a multi-dimensional state vector ; S3.6: Set the optimization objective function, input the encoded multi-dimensional state vector into the optimization objective function, and generate the initial scheduling plan by solving it. ; The optimization objective function is to achieve the minimum target after weighted summation of minimizing the total production cycle, minimizing the equipment switching cost, and maximizing the order on-time rate.

2. The graph neural network scheduling method for aluminum rolling multi-process production scheduling according to claim 1 is characterized in that: The specific steps of S2 include: S2.1: Decompose the aluminum rolling process into independent steps and collect real-time data from the digital twin model of the aluminum rolling production line; S2.2: Determine the process nodes based on the disassembled independent processes and the collected real-time data Relationship with process , where V represents the vertex set of the dynamic graph model, E represents the edge set of the dynamic graph model, i and j represent the process node index values, represents the i-th process node; S2.3: Based on process nodes Relationship with process , standardize the attribute information of each process node and construct the process node attribute vector , and arrange all node attribute vectors in columns to form a node feature matrix At the same time, the adjacency matrix is ​​constructed based on the process relationship set E , elements in A according to Confirm and satisfy ,in, represents the attribute vector of the nth process node, T represents transposition, n represents the number of process nodes, and d represents the number of attribute features contained in each process node. Represents the process node in the process relationship and process nodes The edge attribute vector between them, f represents the Gaussian mapping function; S2.4: Process node attribute vector and edge attribute vector Expand the time dimension to obtain the process node status and edge status , where t represents the time step; S2.5: Constructing a time series dataset using a sliding window mechanism , where T represents the length of the time series; S2.6: Based on process node status and edge status , combined with the actual rules of aluminum rolling production, a state transfer function is set; the state transfer function includes a node state transfer function and an edge state transfer function; S2.7: Based on the constraints, the state transfer function results are modified, and finally a directed acyclic graph that meets all the constraints is constructed to form a dynamic graph model of the aluminum rolling process; the constraints include timing constraints, process constraints, quality constraints, and resource constraints.

3. The graph neural network scheduling method for aluminum rolling multi-process production scheduling according to claim 2 is characterized in that: The process of determining process nodes and process relationships includes: Abstracting the basic process information and real-time status information into nodes in a dynamic graph model to obtain process nodes; the basic process information includes process name, processing time, and equipment requirements; the real-time status information includes whether processing is in progress and the remaining processing time; The dependency between processes is abstracted as edges in a dynamic graph model to obtain process relationships; the weights of the edges represent the time interval requirements between processes and the logistics delivery efficiency.

4. The graph neural network scheduling method for aluminum rolling multi-process production scheduling according to claim 3 is characterized in that: The specific steps of S3.2 include: S3.21: For process nodes , define its neighborhood; the neighborhood includes the in-neighborhood and out of neighborhood ; S3.22: Obtain the feature representation of each process node based on the process node attribute vector and edge attribute vector in S2.3; S3.23: Perform inbound and outbound neighborhood aggregation using a weighted aggregation function to obtain aggregated neighborhood information; S3.24: Fusion of node features with aggregated neighborhood information to obtain the final local information .

5. The graph neural network scheduling method for aluminum rolling multi-process production scheduling according to claim 4 is characterized in that: The feature representation of each process node in S3.22 includes the node's own features, the incoming edge features and out-edge features ; The node's own characteristics are process node attribute vectors .

6. The graph neural network scheduling method for aluminum rolling multi-process production scheduling according to claim 5 is characterized in that: The specific steps of S4 include: S4.1: Obtain real-time data from the physical production line through sensor networks ; S4.2: Comparison of digital twin predictions for an aluminum rolling line With real-time data , identify the time and location of the disturbance, and calculate the deviation between the actual value and the predicted value of the digital twin model of the aluminum rolling production line , represents the 2-norm; S4.3: Set the disturbance threshold; like , then the process node An exception occurs at the time step t, and the abnormal status indicator is displayed as 1, where Indicates process node The disturbance threshold of like , then the process node It is normal at the time step t, and the abnormal status indicator is displayed as 0; S4.4: Determine the disturbance location based on the disturbance threshold judgment result; S4.5: According to the disturbance location, determine the affected process nodes and their neighborhoods, and update the subgraph parameters in the dynamic graph model of the aluminum rolling process affected by the disturbance. At the same time, combine the historical state with the current disturbance; the historical state includes the hidden state of the process nodes and the graph structure information of the dynamic graph model of the aluminum rolling process.

7. The graph neural network scheduling method for aluminum rolling multi-process production scheduling according to claim 6, characterized in that: The digital twin model of the aluminum rolling production line in S1 includes a three-dimensional geometric model, a physical mechanism model, a process logic model and a real-time data interface, and the digital twin model of the aluminum rolling production line is synchronized with the physical production line data.

8. A graph neural network scheduling system for aluminum rolling multi-process production scheduling, which is used to implement the graph neural network scheduling method for aluminum rolling multi-process production scheduling according to any one of claims 1 to 7, characterized in that: include: Digital twin model construction module, dynamic graph model construction module, graph neural network scheduling module, dynamic disturbance response module, feedback and adjustment module; The digital twin model construction module is used to establish a virtual mapping of the physical production line, perform real-time data synchronization and production status simulation; The dynamic graph model building module is used to abstract the aluminum rolling production process into a dynamic graph structure to capture the process node attributes, relationships and timing characteristics; The graph neural network scheduling module is used to extract features based on the dynamic graph model and generate and optimize the initial scheduling plan; The dynamic disturbance response module is used to perceive production disturbances in real time, dynamically update the scheduling plan and verify its feasibility; The feedback and adjustment module is used to establish a closed-loop feedback mechanism and optimize the digital twin model and the dynamic graph model according to the production execution results.

9. The graph neural network scheduling system for aluminum rolling multi-process production scheduling according to claim 8, characterized in that: The graph neural network scheduling module includes: a feature extraction unit, a global state modeling unit, a state fusion unit, and a multi-objective optimization unit; The feature extraction unit is used to iteratively update the node hidden state through multiple layers of graph convolutional layers, aggregate neighborhood information, and extract local constraints and global dependencies of a single process; The global state modeling unit is used to globally pool the node features of each time step, and generate a global state vector in combination with the gated recurrent unit to capture the overall timing characteristics of the production process; The state fusion unit is used to combine the local features of the nodes with the global state vector to form a multi-dimensional state representation that includes process details and the overall process; The multi-objective optimization unit defines an optimization objective function, inputs a multi-dimensional state vector into the function to solve it, and generates an initial scheduling plan.

Images