A state representation modeling method and system based on a multi-domain graph attention network

Abstract

The application discloses a state representation modeling method and system based on a multi-domain graph attention network, constructs a multi-domain graph attention network, extracts global semantic features from three dimensions of operation, machine and worker respectively, and fuses to generate a unified state representation, so that complex dependency relationships among multiple entities in production scheduling can be effectively captured. A subgraph attention mechanism based on a meta path strengthens heterogeneous association modeling of machine nodes and operation nodes, and feature stacking and pooling processing reduces computational redundancy of cross-domain information fusion. Operation semantic embedding is distributed through predecessor-successor attention weight, and accurately represents timing constraint logic between processes. Dynamic aggregation of worker features and machine features further improves the explainability of resource allocation relationships. Finally, the generated global state representation outputs a scheduling action distribution through a decision network, which guarantees complex scheduling rule constraints, and significantly improves the accuracy and response efficiency of scheduling decisions in a dynamic production environment.

Description

Technical Field

[0001] This invention belongs to the field of workshop scheduling technology, and in particular relates to a state representation modeling method and system based on multi-domain graph attention network. Background Technology

[0002] In the development of intelligent manufacturing, the Flexible Job Shop Scheduling (FJSP) problem is one of the key challenges facing the manufacturing field. With the advent of Industry 5.0, production models emphasizing worker participation and worker-robot collaboration have become a trend, giving rise to Human-Network Physical Systems (HCPS). In the HCPS environment, FJSP becomes more complex, involving the allocation of appropriate machines and workers to a series of tasks under constraints such as machine availability and operational priorities, in order to achieve goals such as maximizing machine efficiency, minimizing total time, or energy consumption.

[0003] Traditional methods for solving FJSP (Functional Task Scheduling) have many limitations. Heuristic methods, such as commonly used Priority Scheduling Rules (PDRs), rely on simple rule-based strategies to generate scheduling schemes, making them difficult to adapt to complex and ever-changing production scenarios. They lack robustness and scalability, and cannot effectively capture the complexities of task dependencies, resource constraints, and dynamic production priorities. While exact algorithm-based methods can theoretically achieve optimal or near-optimal results, their computational time cost is too high and their scalability is poor when facing the vast combinatorial search space in large-scale manufacturing environments, limiting their practical applications.

[0004] In recent years, machine learning, especially deep reinforcement learning (DRL), has brought new approaches to solving FJSP. Some studies have attempted to use DRL to learn to select from given PDRs; however, this method can only learn within a limited search space of PDRs, resulting in poor scheduling performance. Other studies have focused on generating high-quality PDRs by designing different network structures to extract the relationship between operations and machines, but these methods often neglect the important impact of workers on production and fail to effectively model workers, thus failing to fully realize their potential in HCPS-FJSP scenarios.

[0005] Meanwhile, existing FJSP scheduling methods are inadequate in terms of state representation, failing to fully integrate the states and dependencies of operations, machines, and workers, resulting in low decision-making quality and making it difficult to meet the requirements of Industry 5.0 for efficient and reliable manufacturing. In summary, current technologies for solving FJSP in HCPS face challenges such as difficulty in handling complex dependencies, neglecting worker factors, and incomplete state representation, urgently requiring a new technical solution to improve scheduling performance and decision-making quality. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention proposes a state representation modeling method and system based on multi-domain graph attention networks, thereby resolving the issues present in the prior art.

[0007] Firstly, to achieve the above objectives, this invention provides a state representation modeling method based on a multi-domain graph attention network, comprising the following steps:

[0008] Based on the features of operation nodes, the features of direct predecessor and successor operation nodes, and the features of related machine nodes, the priority of operation nodes is updated through an attention mechanism to generate operation semantic embeddings.

[0009] Based on machine node features, associated operation node features, and operation-machine arc features, a machine subgraph is constructed through metapaths, and machine node priorities are updated using an attention mechanism to generate machine semantic embeddings.

[0010] Based on worker node features, associated machine node features, and machine-worker arc features, a worker subgraph is constructed, and worker node priorities are updated through an attention mechanism to generate worker semantic embeddings.

[0011] The operation semantic embedding, machine semantic embedding, and worker semantic embedding are fused to generate a global state embedding.

[0012] Based on the global state embedding, scheduling decision actions are output through the actor-critic network, including operations and machine-worker allocation pairs.

[0013] Optionally, the process of updating the priority of the operation node includes:

[0014] Calculate the attention coefficients of the operation node and its neighboring operation nodes, normalize them, and then aggregate the features of the neighboring nodes.

[0015] Calculate the attention coefficients between the operation node and the associated machine node, and then normalize and aggregate the machine node features.

[0016] By capturing the complex dependencies between operation nodes through a multi-head attention mechanism, operation semantic embeddings are generated.

[0017] Optionally, the process of constructing the machine subgraph includes:

[0018] The initial machine subgraph is generated by associating machine nodes and operation nodes through metapaths;

[0019] The operation node features, machine node features, and operation-machine arc features are linearly transformed and stacked to form machine subgraph node features;

[0020] Based on the features of stacked nodes, the attention coefficients of adjacent machine nodes are calculated, and the node features are aggregated after normalization to generate machine semantic embeddings.

[0021] Optionally, the process of constructing the worker subgraph includes:

[0022] An initial worker subgraph is generated by associating worker nodes with machine nodes through meta-paths;

[0023] The machine node features, worker node features, and machine-worker arc features are linearly transformed and stacked to form worker subgraph node features.

[0024] Based on the stacked node features, the attention coefficients of adjacent worker nodes are calculated, and the node features are aggregated after normalization to generate worker semantic embeddings.

[0025] Optionally, the feature fusion process includes:

[0026] The operational semantic embedding, machine semantic embedding, and worker semantic embedding are concatenated according to dimensions to form a global state embedding.

[0027] The global state is embedded into the input critic network, and the value estimate of the current environment state is output.

[0028] Optionally, the action output process of the actor-critic network includes:

[0029] Based on operator node embedding, machine node embedding, worker node embedding and global state embedding, action scores are generated through a multilayer perceptron.

[0030] The action scores are normalized, and the probability distribution of the operation-machine-worker assignment pairs is output.

[0031] Secondly, the present invention also provides a state representation modeling system based on a multi-domain graph attention network, used to implement a state representation modeling method based on a multi-domain graph attention network, the system comprising:

[0032] The operation node modeling module is used to update the operation node priority and generate operation semantic embedding based on the operation node features, the features of the direct predecessor and successor operation nodes and the features of the associated machine nodes through an attention mechanism.

[0033] The machine node modeling module is used to construct a machine subgraph based on machine node features, associated operation node features, and operation-machine arc features, and to update machine node priorities using an attention mechanism to generate machine semantic embeddings.

[0034] The worker node modeling module is used to construct a worker subgraph based on worker node features, associated machine node features, and machine-worker arc features, update worker node priorities through an attention mechanism, and generate worker semantic embeddings.

[0035] The feature fusion module is used to concatenate the operation semantic embedding, machine semantic embedding, and worker semantic embedding to generate a global state embedding;

[0036] The scheduling decision module is used to output scheduling decision actions through the actor-critic network based on the global state embedding. The actions include operations and machine-worker allocation pairs.

[0037] Optionally, the operation node modeling module includes:

[0038] The operation attention calculation unit is used to calculate the attention coefficient between the operation node and its neighboring operation nodes and associated machine nodes;

[0039] The operation feature aggregation unit is used to aggregate the features of adjacent operation nodes and machine nodes based on the normalized attention coefficients.

[0040] The operation semantic embedding generation unit is used to capture the complex dependencies between operation nodes through a multi-head attention mechanism and generate operation semantic embeddings.

[0041] Thirdly, the present invention also provides a computer terminal device, comprising:

[0042] One or more processors;

[0043] A memory, coupled to the processor, for storing one or more programs;

[0044] When the one or more programs are executed by the one or more processors, the one or more processors implement a state representation modeling method based on a multi-domain graph attention network.

[0045] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements a state representation modeling method based on a multi-domain graph attention network.

[0046] Compared with the prior art, the present invention has the following advantages and technical effects:

[0047] This invention provides a state representation modeling method and system based on a multi-domain graph attention network. By constructing a multi-domain graph attention network, this invention extracts global semantic features from three dimensions: operation, machine, and worker, and fuses them to generate a unified state representation, effectively capturing the complex dependencies between multiple entities in production scheduling. Specifically, the subgraph attention mechanism based on meta-paths strengthens the heterogeneous association modeling between machine nodes and operation nodes; feature stacking and pooling reduce computational redundancy in cross-domain information fusion; operation semantic embedding accurately represents the temporal constraint logic between processes through predecessor-successor attention weight allocation; and the dynamic aggregation of worker and machine features further enhances the interpretability of resource allocation relationships. The final generated global state representation outputs the scheduling action distribution through a decision network, significantly improving the accuracy and response efficiency of scheduling decisions in dynamic production environments while ensuring the constraints of complex scheduling rules. Attached Figure Description

[0048] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0049] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0050] Figure 2 This is a schematic diagram of the overall framework of MDGAT according to an embodiment of the present invention;

[0051] Figure 3 This is a schematic diagram of a system according to an embodiment of the present invention. Detailed Implementation

[0052] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0053] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0054] This application, belonging to the field of industrial manufacturing technology, specifically applies to production scheduling systems. It addresses industrial technical problems such as resource conflicts, timing mismatches, and response lags that arise during the collaborative scheduling of multiple entities (operational processes, machinery, and skilled workers) in dynamic production environments. It constructs a decision model deeply coupled with the physical constraints of the production scenario using a multi-domain graph attention network. The solution is based on objective production laws in real manufacturing units, such as machine processing capacity, physical dependencies of processes, and worker skill attributes. It transforms industrial entity operational data, including machine status, process sequence chains, and human resource allocation, into computable features. This solves the scheduling conflicts and efficiency bottlenecks caused by traditional scheduling algorithms neglecting equipment heterogeneity, dynamic coupling of processes, and human-machine adaptation. Its technical means closely revolve around the physical attributes and operational logic of the production system, conforming to the natural laws of data correlation and the objectivity of production constraints in the industrial field.

[0055] Example 1

[0056] like Figure 1 As shown, this embodiment provides a state representation modeling method based on a multi-domain graph attention network, including:

[0057] Based on the features of operation nodes, the features of direct predecessor and successor operation nodes, and the features of related machine nodes, the priority of operation nodes is updated through an attention mechanism to generate operation semantic embeddings.

[0058] Based on machine node features, associated operation node features, and operation-machine arc features, a machine subgraph is constructed through metapaths, and machine node priorities are updated using an attention mechanism to generate machine semantic embeddings.

[0059] Based on worker node features, associated machine node features, and machine-worker arc features, a worker subgraph is constructed, and worker node priorities are updated through an attention mechanism to generate worker semantic embeddings.

[0060] The operation semantic embedding, machine semantic embedding, and worker semantic embedding are fused to generate a global state embedding.

[0061] Based on the global state embedding, scheduling decision actions are output through the actor-critic network, including operations and machine-worker allocation pairs.

[0062] Specifically, a state representation modeling method based on MDGAT consists of an operation representation modeling module, a machine representation modeling module, a work representation modeling module, and a feature fusion module. A schematic diagram of the overall framework of the Multi-Domain Graph Attention Network (MDGAT) is shown below. Figure 2 As shown.

[0063] As one implementation method in this embodiment, the process of updating the priority of the operation node includes:

[0064] Calculate the attention coefficients of the operation node and its neighboring operation nodes, normalize them, and then aggregate the features of the neighboring nodes.

[0065] Calculate the attention coefficients between the operation node and the associated machine node, and then normalize and aggregate the machine node features.

[0066] By capturing the complex dependencies between operation nodes through a multi-head attention mechanism, operation semantic embeddings are generated.

[0067] Operation node attention representation modeling. First, operation node attention representation modeling is performed, utilizing the features of operation nodes in the same job, their direct parent operation nodes, and related machine node features to update the priority nodes for different operations. Specifically, given an operation node O... ij Calculate the attention coefficients of its direct predecessor and successor nodes, as well as its own:

[0068]

[0069] Among them, W O Let 'a' represent the linear transformation of the operation node, where 'a' is a single-layer feedforward neural network. Similarly, the relationship between the relevant machine node and the operation node O... ij The attention coefficient is calculated as follows:

[0070]

[0071] Among them, W M It is a linear transformation for machine nodes.

[0072] Then, the softmax function is used to normalize all attention coefficients of the target operation node, resulting in the normalized attention coefficient α. ijq and α ijk Finally, by fusing adjacent operation nodes, adjacent machine nodes, and the node itself, the aggregate function representing the operation node embedding is obtained as follows:

[0073]

[0074] Operational semantic attention representation modeling is employed. Each operation node contains specific semantic information, and the embeddings of all operation nodes are fused using the mean method, ignoring the differences between nodes. This invention designs an operational semantic attention mechanism to learn the importance of different nodes. The output representation of the attention result of the l-th layer operation nodes is used for symbolic modeling, denoted as... Semantic importance coefficient node for each operation The calculation is as follows:

[0075]

[0076] In the formula, WO' and b O' These are the weight matrix and bias vector of the operation node, respectively, q O' Let be the semantic attention vector of the operation node, and tanh be the activation function.

[0077] Then, the semantic importance coefficients are normalized using the softmax function. Each operation node O... ij The weight, denoted as Represented as:

[0078]

[0079] Using the learned operation node weights as coefficients, the global operation semantic embedding is calculated as follows:

[0080]

[0081] Furthermore, this module employs a multi-head attention mechanism, effectively capturing the complex and diverse dependencies in operational nodes and semantic attention representation modeling. Here, H represents the number of attention heads used in the representation modeling process.

[0082] As one implementation method in this embodiment, the process of constructing the machine subgraph includes:

[0083] The initial machine subgraph is generated by associating machine nodes and operation nodes through metapaths;

[0084] The operation node features, machine node features, and operation-machine arc features are linearly transformed and stacked to form machine subgraph node features;

[0085] Based on the features of stacked nodes, the attention coefficients of adjacent machine nodes are calculated, and the node features are aggregated after normalization to generate machine semantic embeddings.

[0086] In the machine representation modeling module, different operation nodes and machine nodes have more complex connections. Different machines may compete for the same operations, and the node attributes in the proposed heterogeneous scheduling graph are heterogeneous. How to explicitly model machine priorities is crucial for high-quality agent decision-making. Heterogeneous Graph Attention Networks (HGATs) (Wang et al. 2019) has become a promising method for learning heterogeneous graph representations, which have multiple node and edge types. Metapaths are used to link composite relationships between two different types of objects. In the context of HCPS-FJSP, HGATs allow us to capture intra-type and inter-type relationships, thus facilitating the representation modeling of complex dependencies. Inspired by HGATs, machine node attention representation modeling and semantic attention representation modeling were designed.

[0087] like Figure 1As shown. Here, different machine nodes are treated as different types of nodes, while operational nodes are treated as the same type of node. Therefore, metapaths are used. To describe machine node M k The relationship between nodes and operations. Utilizing metapaths. An initial subgraph was obtained. Initial subgraph It only has operator node features, ignoring machine and OM arc features. A linear transformation matrix is ​​used to convert operator node features, machine node features, and OM arc features to the same dimension, and then stacked as follows:

[0088]

[0089] in Let be the feature vectors of the stacked machine nodes, representing the forming matrix of the linear inverse OM arc. Finally, a subgraph is obtained. in This represents a stack of machine nodes in a subgraph. The connection relationships of stacked machine nodes are represented in the subgraph.

[0090] Machine node attention representation modeling. First, the stacked machine node embeddings are updated at the node level by overlaying adjacent nodes. Specifically, given a stack of machine nodes M... ijk The attention coefficient of adjacent nodes is calculated as follows:

[0091]

[0092] in Indicates stacked machine nodes M pqk To M ijk Attention coefficient.

[0093] Note that the coefficients are normalized using the Softmax function, expressed as:

[0094]

[0095] Then, by combining the embedding and attention coefficients of neighboring nodes, the stacked machine nodes M can be aggregated. ijk The embeddedness of the stacked machine nodes can be represented as follows:

[0096]

[0097] Finally, for all in the subgraph The stacked machine nodes are processed, and the result is used as the node embedding for machine k, i.e.:

[0098]

[0099] Machine semantic attention representation modeling. Given the embedding of the Lth-th layer machine node. semantic importance coefficient The method for calculating k for each machine is as follows:

[0100]

[0101] Among them W M' and b M' These are the weight matrix and bias vector of the machine node, respectively. It is the semantic attention vector of the machine node.

[0102] By normalizing the semantic importance coefficients of all machines, the weight of each machine node is determined. The calculation is as follows:

[0103]

[0104] Then, the global machine semantic embedding is computed as follows:

[0105]

[0106] As one implementation method in this embodiment, the process of constructing the worker subgraph includes:

[0107] An initial worker subgraph is generated by associating worker nodes with machine nodes through meta-paths;

[0108] The machine node features, worker node features, and machine-worker arc features are linearly transformed and stacked to form worker subgraph node features.

[0109] Based on the stacked node features, the attention coefficients of adjacent worker nodes are calculated, and the node features are aggregated after normalization to generate worker semantic embeddings.

[0110] In the worker representation modeling module, the superimposed worker node feature vector is obtained by superimposing machine node features, worker node features, and MW arc features. And expressed as:

[0111]

[0112] Among them W D and W W It is a linear transformation matrix of the MW arc and the working node. Finally, a subgraph is obtained. in Subgraph Stacked worker nodes Subgraph representing the connection relationships between stacks The working nodes in the process.

[0113] The working nodes should be represented in the model. (In the subgraph) In the middle, stacked worker nodes W kl The attention coefficient for each adjacent node is calculated as follows:

[0114]

[0115] in, For stacked worker nodes W pl For W kl Attention coefficient.

[0116] The formula for calculating the normalized worker attention coefficient is:

[0117] Then, the embedding of each stacked worker node can be updated as follows:

[0118]

[0119] Finally, average pooling is applied to the pair graph. All stacked worker nodes are processed. Therefore, the node embedding representation of worker l is:

[0120]

[0121] Worker semantic attention representation modeling. Similarly,

[0122] Embedded based on the l-th layer working node Semantic importance coefficient for each machine Represented as:

[0123]

[0124] Among them W W' and b W' These are the weight matrix and bias vector of the working node, respectively.

[0125] and This is the semantic attention vector of the worker node. Meanwhile, the weight of each worker node is represented as:

[0126]

[0127] Therefore, the computation method for the semantic embedding of the global worker is as follows:

[0128]

[0129] As one implementation method in this embodiment, the feature fusion process includes:

[0130] The operational semantic embedding, machine semantic embedding, and worker semantic embedding are concatenated according to dimensions to form a global state embedding.

[0131] The global state is embedded into the input critic network, and the value estimate of the current environment state is output.

[0132] As one implementation method in this embodiment, the action output process of the actor-critic network includes:

[0133] Based on operator node embedding, machine node embedding, worker node embedding and global state embedding, action scores are generated through a multilayer perceptron.

[0134] The action scores are normalized, and the probability distribution of the operation-machine-worker assignment pairs is output.

[0135] In the representation fusion module, global semantic embeddings from different domains are integrated into a rich global embedding to help the agent fully understand the environment state. Therefore, the resulting global operational semantic embedding h... O Global machine semantic embedding h M semantic embedding of global workers h W The connection is globally embedded as follows:

[0136] h G =[h O ||h M ||h W ]

[0137] This fusion module captures combined information from all domains, enabling decision components to make high-quality scheduling decisions.

[0138] decision making:

[0139] The decision network is the final component of the decision-making process and is based on an actor-critic network design. The actor and critic networks consist of MLP networks with parameters θ and w, respectively. Note that an action is a compatible action on an OMW pair. t =(O ij M k W l ). Utilizing the embedded environment state, the policy network outputs a scale μ(a) for each feasible action. t |s t ),Right now:

[0140]

[0141] Then, the probability of choosing an action is calculated as follows:

[0142]

[0143] Furthermore, the global embedding is used as input to the critique network to estimate the current environment state s. t The value v w (s t ).

[0144] Based on this, this invention provides a state representation modeling method based on multi-domain graph attention networks (MDGAT). This invention significantly improves the decision-making quality and efficiency of flexible workshop scheduling problems in intelligent manufacturing through a state representation modeling method based on MDGAT. This method utilizes a multi-domain attention mechanism of operation, machine, and worker nodes to accurately capture task dependencies, resource constraints, and dynamic features of human-machine collaboration, solving the shortcomings of traditional heuristic rules and single graph networks in modeling complex interactions. Through meta-path-driven subgraph construction and feature stacking, it explicitly integrates heterogeneous information such as worker skills and machine load, achieving for the first time a systematic deep fusion of worker state and scheduling decisions, avoiding resource mismatch problems in the Human-Network Physical System (HCPS) scenario in Industry 5.0. The feature fusion module embeds and concatenates multi-dimensional semantics into a global state representation, providing a unified high-order decision basis for the actor-critic network. At a task scale of thousands, the decision response time is reduced by more than 40% compared to existing deep reinforcement learning methods. It also supports real-time adjustments in dynamic environments, significantly improving the robustness and scalability of the scheduling scheme, meeting the core requirements of intelligent manufacturing for efficient collaboration and real-time optimization.

[0145] Example 2

[0146] In this embodiment, a computer terminal device is provided, including:

[0147] One or more processors;

[0148] A memory, coupled to the processor, for storing one or more programs;

[0149] When the one or more programs are executed by the one or more processors, the one or more processors implement the methods in the above embodiments.

[0150] In this embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the methods described in the above embodiments.

[0151] In this embodiment, an electronic device is also provided, including a memory and a processor. The memory stores a computer program, and the processor is configured to run the computer program to perform the methods described in the above embodiments.

[0152] The aforementioned program can run on a processor or be stored in memory (or a computer-readable medium). Computer-readable media include both permanent and non-permanent, removable and non-removable media, and information storage can be achieved by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0153] These computer programs may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes can be implemented using different modules, and different steps can be implemented using different modules.

[0154] like Figure 3 As shown, this embodiment provides such an apparatus or system. The system, referred to as a state representation modeling system based on multi-domain graph attention networks, includes:

[0155] The operation node modeling module is used to update the operation node priority and generate operation semantic embedding based on the operation node features, the features of the direct predecessor and successor operation nodes and the features of the associated machine nodes through an attention mechanism.

[0156] The machine node modeling module is used to construct a machine subgraph based on machine node features, associated operation node features, and operation-machine arc features, and to update machine node priorities using an attention mechanism to generate machine semantic embeddings.

[0157] The worker node modeling module is used to construct a worker subgraph based on worker node features, associated machine node features, and machine-worker arc features, update worker node priorities through an attention mechanism, and generate worker semantic embeddings.

[0158] The feature fusion module is used to concatenate the operation semantic embedding, machine semantic embedding, and worker semantic embedding to generate a global state embedding;

[0159] The scheduling decision module is used to output scheduling decision actions through the actor-critic network based on the global state embedding. The actions include operations and machine-worker allocation pairs.

[0160] As one implementation method in this embodiment, the operation node modeling module includes:

[0161] The operation attention calculation unit is used to calculate the attention coefficient between the operation node and its neighboring operation nodes and associated machine nodes;

[0162] The operation feature aggregation unit is used to aggregate the features of adjacent operation nodes and machine nodes based on the normalized attention coefficients.

[0163] The operation semantic embedding generation unit is used to capture the complex dependencies between operation nodes through a multi-head attention mechanism and generate operation semantic embeddings.

[0164] As one implementation method in this embodiment, the machine node modeling module includes:

[0165] The machine subgraph construction unit is used to associate machine nodes and operation nodes through metapaths to generate the initial machine subgraph.

[0166] The machine feature stacking unit is used to perform linear transformations on the operation node features, machine node features, and operation-machine arc features and stack them to form machine subgraph node features.

[0167] The machine semantic embedding generation unit is used to calculate the attention coefficients of adjacent machine nodes based on the features of stacked nodes, and then normalize and aggregate the node features to generate machine semantic embeddings.

[0168] As one implementation method in this embodiment, the worker node modeling module includes:

[0169] The worker subgraph construction unit is used to associate worker nodes and machine nodes through meta-paths to generate the initial worker subgraph;

[0170] The worker feature stacking unit is used to perform linear transformation and stack the machine node features, worker node features and machine-worker arc features to form worker subgraph node features.

[0171] The worker semantic embedding generation unit is used to calculate the attention coefficients of adjacent worker nodes based on the stacked node features, and then normalize and aggregate the node features to generate worker semantic embeddings.

[0172] As one implementation method in this embodiment, the feature fusion module includes:

[0173] The global embedding splicing unit is used to splice the operation semantic embedding, machine semantic embedding and worker semantic embedding by dimension to generate a global state embedding.

[0174] The state value estimation unit is used to embed the global state into the input critic network and output a value estimate of the current environment state.

[0175] As one implementation method in this embodiment, the scheduling decision module includes:

[0176] The action scoring generation unit is used to generate action scores through a multilayer perceptron based on operation node embedding, machine node embedding, worker node embedding and global state embedding.

[0177] The action probability distribution unit is used to normalize action scores and output the probability distribution of operation-machine-worker assignment pairs.

[0178] The system or apparatus is used to implement the functions of the methods in the above embodiments. Each module in the system or apparatus corresponds to each step in the method, as has been described in the method and will not be repeated here.

[0179] The above implementation method solves the problem of state representation modeling based on multi-domain graph attention networks in related technologies, thereby ensuring that the problems existing in the prior art are resolved.

[0180] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A state representation modeling method based on multi-domain graph attention networks, characterized in that, Includes the following steps: Based on the features of operation nodes, the features of direct predecessor and successor operation nodes, and the features of related machine nodes, the priority of operation nodes is updated through an attention mechanism to generate operation semantic embeddings. Based on machine node features, associated operation node features, and operation-machine arc features, a machine subgraph is constructed through metapaths, and machine node priorities are updated using an attention mechanism to generate machine semantic embeddings. Based on worker node features, associated machine node features, and machine-worker arc features, a worker subgraph is constructed, and worker node priorities are updated through an attention mechanism to generate worker semantic embeddings. The operation semantic embedding, machine semantic embedding, and worker semantic embedding are fused to generate a global state embedding. Based on the global state embedding, scheduling decision actions are output through the actor-critic network, including operations and machine-worker allocation pairs.

2. The method according to claim 1, characterized in that, The process of updating the priority of the operation node includes: Calculate the attention coefficients of the operation node and its neighboring operation nodes, normalize them, and then aggregate the features of the neighboring nodes. Calculate the attention coefficients between the operation node and the associated machine node, and then normalize and aggregate the machine node features. By capturing the complex dependencies between operation nodes through a multi-head attention mechanism, operation semantic embeddings are generated.

3. The method according to claim 1, characterized in that, The process of constructing the machine subgraph includes: The initial machine subgraph is generated by associating machine nodes and operation nodes through metapaths; The operation node features, machine node features, and operation-machine arc features are linearly transformed and stacked to form machine subgraph node features; Based on the features of stacked nodes, the attention coefficients of adjacent machine nodes are calculated, and the node features are aggregated after normalization to generate machine semantic embeddings.

4. The method according to claim 1, characterized in that, The process of constructing the worker subgraph includes: An initial worker subgraph is generated by associating worker nodes with machine nodes through meta-paths; The machine node features, worker node features, and machine-worker arc features are linearly transformed and stacked to form worker subgraph node features. Based on the stacked node features, the attention coefficients of adjacent worker nodes are calculated, and the node features are aggregated after normalization to generate worker semantic embeddings.

5. The method according to claim 1, characterized in that, The feature fusion process includes: The operational semantic embedding, machine semantic embedding, and worker semantic embedding are concatenated according to dimensions to form a global state embedding. The global state is embedded into the input critic network, and the value estimate of the current environment state is output.

6. The method according to claim 1, characterized in that, The action output process of the actor-critic network includes: Based on operator node embedding, machine node embedding, worker node embedding and global state embedding, action scores are generated through a multilayer perceptron. Normalize the action scores and output the probability distribution of the operation-machine-worker assignment pairs.

7. A state representation modeling system based on a multi-domain graph attention network, characterized in that, The system includes: The operation node modeling module is used to update the operation node priority and generate operation semantic embedding based on the operation node features, the features of the direct predecessor and successor operation nodes and the features of the associated machine nodes through an attention mechanism. The machine node modeling module is used to construct a machine subgraph based on machine node features, associated operation node features, and operation-machine arc features, and to update machine node priorities using an attention mechanism to generate machine semantic embeddings. The worker node modeling module is used to construct a worker subgraph based on worker node features, associated machine node features, and machine-worker arc features, update worker node priorities through an attention mechanism, and generate worker semantic embeddings. The feature fusion module is used to concatenate the operation semantic embedding, machine semantic embedding, and worker semantic embedding to generate a global state embedding; The scheduling decision module is used to output scheduling decision actions through the actor-critic network based on the global state embedding. The actions include operations and machine-worker allocation pairs.

8. The system according to claim 7, characterized in that, The operation node modeling module includes: The operation attention calculation unit is used to calculate the attention coefficient between the operation node and its neighboring operation nodes and associated machine nodes; The operation feature aggregation unit is used to aggregate the features of adjacent operation nodes and machine nodes based on the normalized attention coefficients. The operation semantic embedding generation unit is used to capture the complex dependencies between operation nodes through a multi-head attention mechanism and generate operation semantic embeddings.

9. A computer terminal device, characterized in that, include: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the state representation modeling method based on multi-domain graph attention networks as described in any one of claims 1-6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the state representation modeling method based on multi-domain graph attention network as described in any one of claims 1-6.

Images