Graph-based flexible job shop state representation and feature extraction method

Abstract

The invention discloses a graph-based flexible job shop state representation and feature extraction method, and relates to the technical field of shop scheduling, and the method comprises the following steps: 1, constructing a heterogeneous graph; step 2, embedding key features into corresponding nodes or edges based on the heterogeneous graph constructed in the step 1; 3, dynamically updating the structure of the heterogeneous graph and the characteristics of nodes and edges according to a scheduling action at each time step; step 4, performing depth feature extraction by using a graph neural network based on the real-time heterogeneous graph updated in the step 3 at each time step; and 5, inputting the extracted features into a scheduling decision module to generate a specific scheduling action, circularly executing the steps 3-4 until all procedures are scheduled, and outputting a final scheduling scheme. According to the method, the multi-resource interaction relationship in the flexible job shop can be accurately described, and the scheduling decision efficiency and effect are improved.

Description

Technical Field

[0001] This invention relates to the field of workshop scheduling technology, and in particular to a graph-based method for flexible workshop state representation and feature extraction. Background Technology

[0002] As the manufacturing industry continues to develop towards intelligence and flexibility, modern workshop production systems are showing significant growth trends in product diversification, processing path complexity, and dynamic resource allocation. The flexible job shop scheduling problem, a typical complex optimization problem in production scheduling, is an extension of the traditional job shop scheduling problem. In a flexible job shop, each process can be performed on multiple selectable machines, and the system needs to simultaneously determine the processing sequence of processes and machine selection to achieve efficient resource utilization and improve overall production efficiency. Flexible job shop models have been widely applied in aerospace, precision manufacturing, electronic equipment, and intelligent equipment manufacturing.

[0003] Driven by smart manufacturing technologies and the Industrial Internet of Things (IIoT), an increasing number of enterprises are introducing Automated Guided Vehicles (AGVs) as a crucial resource for material transportation within workshops. The addition of AGVs necessitates the coordinated scheduling of processing and transportation resources, creating a multi-resource scheduling problem in flexible workshops that integrates three decision-making dimensions: machine, process, and transportation. In this type of problem, operations not only need to be completed on machines but also require cross-equipment transportation via AGVs. Factors such as transportation waiting time, path conflicts, and AGV reachability all affect the actual start time of processes. This necessitates that the scheduling system achieve joint optimization between processing and transportation resources to reduce waiting time, avoid resource congestion, and improve workshop throughput.

[0004] In recent years, existing research methods have generally encountered the following technical bottlenecks when facing this problem: (1) The scheduling state is usually simplified into one-dimensional or two-dimensional feature vectors such as process queues and machine loads. This "flattened" representation method is difficult to effectively depict the complex topological constraint relationship between processes, machines, and AGVs; (2) Most studies decouple machine scheduling from AGV scheduling or adopt a simple sequential strategy, ignoring the close coupling relationship between the two in time and space.

[0005] Therefore, how to construct a unified and efficient representation of the complex states of multiple resources in flexible work workshops, and extract highly generalizable features from it, is a crucial problem that urgently needs to be solved in the field of flexible work workshop scheduling and intelligent manufacturing. Graph-based state representation and feature extraction methods can effectively characterize the complex resource connections and dynamic changes within the system, providing support for constructing efficient scheduling strategies, and have significant theoretical and practical value. Summary of the Invention

[0006] To effectively describe the complex interactions among various resources in a flexible workshop integrated with AGVs, and to overcome the problems of incomplete state representation, insufficient feature extraction capabilities, and difficulty in adapting to dynamic environments in traditional scheduling methods, this invention proposes a graph-based method for state representation and feature extraction in flexible workshops. This invention constructs a dynamic graph structure containing heterogeneous nodes such as processes, machines, and AGVs, as well as various types of relationships, to uniformly characterize the resource availability, process executability, and transportation status of the workshop at any scheduling moment. Furthermore, it extracts global and local features through a graph neural network to generate high-dimensional state embeddings for scheduling decision-making.

[0007] This invention provides a graph-based method for flexible job shop state representation and feature extraction, comprising the following steps:

[0008] Step 1: Construct a heterogeneous graph, setting the processes, machines, and AGVs in the scheduling system as different types of entity nodes, and establishing edge connections between nodes;

[0009] Step 2: Based on the heterogeneous graph, embed key features into the corresponding nodes or edges to construct a multi-dimensional attribute set that reflects the real-time system state;

[0010] Step 3: Each time step The heterogeneous graph structure and the characteristics of nodes and edges described in step 2 are dynamically updated based on the scheduling actions.

[0011] Step 4: Each time step Based on the real-time heterogeneous graph updated in step 3, deep feature extraction is performed using graph neural networks to fuse scattered multi-source information into dense, low-dimensional feature vectors that can be used for downstream decision-making tasks.

[0012] Step 5: Input the extracted features into the scheduling decision module to generate specific scheduling actions, and repeat steps 3-4 until all processes are scheduled and the final scheduling plan is output.

[0013] Furthermore, in step 1, the method for constructing the heterogeneous graph is as follows:

[0014] The flexible job shop scheduling instance is formalized as a heterogeneous graph. ,in , , These are the sets of process nodes, machine nodes, and AGV nodes, respectively. Two virtual nodes are added to the process node set to represent the start and end points of scheduling. A set of priority constraint edges between processes is constructed. Each directed edge represents a pre- and post-order dependency between adjacent processes. Construct a process-machine edge set. either one Connecting process nodes With its compatibility with machine nodes This is used to represent feasible processing relationships. Constructing the process-AGV edge set. either one Connecting process nodes With AGV nodes capable of performing the transportation tasks of this process , used to represent candidate transportation relationships. Wherein, the edge set and In the initial state, all edges are candidate edges, used to represent all processing and transportation relationships; during the scheduling process, the corresponding candidate edges are determined as selected edges based on the scheduling actions, used to represent the actual processing and transportation allocation relationships.

[0015] Furthermore, the feature vectors of nodes and edges in the heterogeneous graph include:

[0016] (1) Process nodes The original feature vector This includes: scheduling status, number of adjacent machines, number of adjacent AGVs, number of unscheduled processes in the workpiece, processing time, transportation time, start time, and workpiece completion time; wherein, the transportation time and start time are based on the process. and its preceding processes The scheduling state is estimated recursively;

[0017] (2) Machine node The original feature vector This includes: the number of adjacent processes, available time, and utilization rate;

[0018] (3) AGV node The original feature vector This includes: the number of adjacent processes, available time, and utilization rate;

[0019] (4) OM edge original features Process To the machine Transportation time, processing time ;

[0020] (5) OA side original features Transportation time based on the current position of the AGV and the status of the preceding process.

[0021] Furthermore, in step 3, the dynamic update of the heterogeneous graph structure based on the scheduling action is as follows:

[0022] At time step Scheduling actions Triadic decision This indicates that the unscheduled processes will be... Assigned to its workable machines And assign idle AGVs Responsible for process transportation, wherein the actions described satisfy the following feasibility conditions: process It has not yet been scheduled, and it is a workpiece. The first process or its preceding process Scheduling completed; selected machine It belongs to the process Compatible machine set The selected AGV At time step It is in an idle state.

[0023] Execute action Then, the system enters the next time step, and its state update includes the structural evolution of the heterogeneous graph and the temporal progression of the environment:

[0024] (1) Structural evolution: Heterogeneous diagrams from Updated to ,in:

[0025] At the process node Above, only with the selected machine node OM edge between and with the selected AGV node OA edge between Remove Candidate OM edges between all other machine nodes and candidate OA edges between all other AGV nodes are calculated to reflect the final determination of the processing and transportation allocation for this process; the feature vectors of all nodes and edges are updated and recalculated according to the feature vector calculation method described in step 2.

[0026] (2) Timing Advancement: Update the ambient clock Until the next decision moment The rules for implementation are as follows:

[0027] If an action is performed If there are still idle AGVs, then This allows for continuous decision-making at the same time; if all AGVs are busy, the environmental clock advances to the earliest completed AGV task; in special cases, when scheduling processes... Its preceding processes When processing on the same machine, no transportation task is required, and the assigned AGV... Once allocated, it is immediately released and considered to be in an idle state.

[0028] The above process is repeated until all processes are scheduled, thus forming a complete scheduling plan.

[0029] Furthermore, in step 4, the method for extracting deep features from this heterogeneous graph using a graph neural network is as follows:

[0030] (1) Machine node embedding; For each machine node, the features of its adjacent process nodes are weighted and aggregated by a graph attention embedding module based on dynamic attention mechanism. The dynamic attention mechanism can adaptively adjust the importance of adjacent nodes according to the different query nodes. At the same time, the machine node's own features are nonlinearly transformed by a multilayer perceptron network, and finally the neighborhood aggregated features and its own updated features are fused by a gating fusion mechanism to obtain the final embedded representation of the machine node.

[0031] (2) AGV node embedding; For each AGV node, a dynamic attention embedding structure consistent with that of the machine node is adopted to perform attention aggregation on the features of its adjacent process nodes, and the final embedded representation of the AGV node is obtained by gating fusion with its own features.

[0032] (3) Process node embedding; For each process node, its neighborhood includes its direct predecessor process node, direct successor process node, optional processing machine node set and optional AGV node set; The process node first performs multi-layer perceptron transformation and fusion on the multi-source information from machine nodes, AGV nodes and its own features respectively; Subsequently, in order to capture the sequence dependency relationship between adjacent processes within the same workpiece, the process sequence under the workpiece is time-series embedded calculation using a bidirectional long short-term memory network, and the final embedded representation of the process node is obtained through linear projection;

[0033] (4) Graph embedding; Perform the above embedding module... Layer stacking allows each layer to independently learn high-order features of node states, resulting in a final multi-layer node embedding representation. The process node embeddings, machine node embeddings, and AGV node embeddings of the final layer are then average-pooled and concatenated to form a graph-level state vector for downstream scheduling decisions.

[0034] Furthermore, in step 5, the extracted features are input into the scheduling decision module to generate specific scheduling actions as follows:

[0035] The graph-level state vector and the embedded representation of each node obtained in step 4 are input to the scheduling decision module. The scheduling decision module evaluates the priority of feasible scheduling actions at the current time based on the input features and selects the preferred scheduling action. As the scheduling decision for the current time step; after the scheduling action is executed, it is used to update the system state and promote the decision process for the next time step; step 5 is executed in a loop with steps 3 to 4 until all processes are scheduled and the final scheduling scheme is output.

[0036] The graph-based flexible job shop state representation and feature extraction method provided by this invention has the following technical effects:

[0037] Methodological aspects:

[0038] (1) A heterogeneous graph state representation model for flexible workshops is proposed. This invention constructs a heterogeneous graph structure composed of process nodes, machine nodes and AGV nodes, and introduces multiple types of association relationships such as process-machine and process-AGV, so as to realize unified modeling of complex interactions such as processing and transportation, which significantly improves the completeness and accuracy of workshop state representation.

[0039] (2) Realize dynamic graph updates of workshop status and environmental time advancement. This invention dynamically updates the feature vectors of nodes and edges based on the execution of scheduling actions, and in conjunction with the intelligent advancement rules of the environmental clock, the entire model can respond to changes in system status in real time and accurately, providing a basis for online scheduling decisions and enhancing the adaptability of the method in dynamic environments.

[0040] (3) A graph feature extraction mechanism that integrates multi-source information was designed. This invention achieves multi-dimensional information fusion by encoding features of heterogeneous nodes separately and using graph neural networks to aggregate cross-type information, thereby enhancing the feature expression capability.

[0041] (4) Generates general and high-quality scheduling decision input features. The graph-level state vector and node embedding extracted in this invention deeply integrate the global situation and local details of the system, forming a dense, low-dimensional and information-rich feature representation. This feature representation serves as a general interface and can provide high-quality input for various downstream scheduling decision modules.

[0042] Application level:

[0043] (1) Comprehensively improve the scheduling system's ability to model complex resource constraints. By explicitly depicting the interaction dependencies between processes, machines and AGVs, this invention can handle collaborative scenarios that are difficult to cover by traditional scheduling methods, effectively reduce scheduling conflicts caused by incomplete state representation, thereby shortening the production cycle and improving production efficiency.

[0044] (2) Improve the accuracy of the scheduling system in depicting the current scheduling state. This invention encodes multi-dimensional information in a heterogeneous graph structure and updates it in real time with scheduling actions, enabling the system to accurately reflect the workshop state and express key information at any scheduling moment.

[0045] (3) Improve the overall production efficiency and resource utilization of complex workshops. This invention can solve the dual resource coupling problem of processing and transportation, accurately model the load and status of machines and AGVs, so that the optimized scheduling scheme can reduce waiting time, improve machine load balance and AGV utilization, thereby improving the overall operating efficiency of the workshop.

[0046] (4) Provide a general state modeling framework for flexible workshops with complex structures. The graph representation method constructed in this invention can be widely applied to scenarios such as discrete manufacturing, intelligent logistics, and highly flexible workshops. The extracted feature vectors can also serve various decision models such as rule-based schedulers and reinforcement learning, providing scalable modeling tools and feature extraction interfaces for production scheduling problems.

[0047] In summary, this invention achieves a comprehensive and detailed representation of the state of flexible workshops through heterogeneous graph modeling, dynamic graph updating, and multi-source information fusion, and generates high-quality scheduling input features. In terms of application, it significantly enhances the system's ability to model constraints and the accuracy of state characterization, improves overall production efficiency and resource utilization, and provides a general and scalable state representation framework for complex workshops. Attached Figure Description

[0048] Figure 1 This is a Gantt chart for flexible job shop scheduling provided in this invention.

[0049] Figure 2 This is the heterogeneous graph model provided in this invention.

[0050] Figures 3(a)-3(h) illustrate the heterogeneous graph update and environmental time evolution process presented in this invention.

[0051] Figure 4 This is the neural network architecture of the present invention. Detailed Implementation

[0052] The invention is described below with reference to flowchart illustrations and / or block diagrams of methods, systems, and computer program products according to embodiments of the invention. It will be understood that some blocks of the flowchart illustrations and / or block diagrams, as well as combinations of some blocks of the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be stored or implemented in microcontrollers, microprocessors, digital signal processors (DSPs), field-programmable gate arrays (FPGAs), state machines, programmable logic controllers (PLCs) or other processing circuits, general-purpose computers, or special-purpose computers. The instructions are used in computers or other programmable data processing devices (e.g., production machines) to create means or block diagram blocks for implementing the functions / actions specified in the flowcharts and / or block diagrams, which are executed by the processor of the computer or other programmable data processing device.

[0053] The specific implementation process of this invention is as follows:

[0054] Consider a small-scale example of the flexible workshop scheduling problem, which includes 3 jobs, 3 machines, and 2 AGVs. Job 1 has 3 processes, job 2 has 3 processes, and job 3 has 2 processes.

[0055] The compatible machine sets and processing times for each process are as follows: Process Available in machine , or The processing times are 4, 8, and 1 time units respectively; Process Available in machine or The processing times are 9 and 3 time units respectively. (Process) Only available on machines The processing time is 5 time units; the process is as follows: Available in machine or The processing times are 2 and 4 time units respectively; process Available in machine or The processing times are 7 and 6 time units respectively. (Process) Available in machine or The processing times are 1 and 5 time units respectively; Process Only available on machines The processing time is 3 time units.

[0056] The travel time of the AGV between different locations is set as follows: from depot to machine , , The time units are 3, 1, and 4 respectively; from the machine to depot , The time units are 2, 3, and 1 respectively; from the machine to depot , The time units are 5, 4, and 2 respectively; from the machine to depot , The time units are 1, 2, and 6 respectively.

[0057] Based on the above processing and transportation constraints, a specific and feasible scheduling scheme can be generated, and the corresponding scheduling Gantt chart is as follows: Figure 1 As shown. In this feasible schedule, the completion time (i.e., makespan) is 18 time units.

[0058] This invention provides a graph-based method for flexible job shop state representation and feature extraction, comprising the following steps:

[0059] Step 1: Construct a heterogeneous graph, setting the processes, machines, and AGVs in the scheduling system as different types of entity nodes, and establishing edge connections between nodes;

[0060] Step 2: Based on the heterogeneous graph, embed key features into the corresponding nodes or edges to construct a multi-dimensional attribute set that reflects the real-time system state;

[0061] Step 3: Each time step The heterogeneous graph structure and the characteristics of nodes and edges described in step 2 are dynamically updated based on the scheduling actions.

[0062] Step 4: Each time step Based on the real-time heterogeneous graph updated in step 3, deep feature extraction is performed using graph neural networks to fuse scattered multi-source information into dense, low-dimensional feature vectors that can be used for downstream decision-making tasks.

[0063] Step 5: Input the extracted features into the scheduling decision module to generate specific scheduling actions, and repeat steps 3-4 until all processes are scheduled and the final scheduling plan is output.

[0064] Based on the above steps, combined with Figure 1 The specific implementation process of the method of the present invention in the example scenario shown is as follows:

[0065] Step 1: Construct a heterogeneous graph.

[0066] The flexible job shop scheduling instance is formalized as a heterogeneous graph. ,in , , These are the sets of process nodes, machine nodes, and AGV nodes, respectively. Two virtual nodes are added to the process node set to represent the start and end points of scheduling. A set of priority constraint edges between processes is constructed. Each directed edge represents a pre- and post-order dependency between adjacent processes. Construct a process-machine edge set. either one Connecting process nodes With its compatibility with machine nodes This is used to represent feasible processing relationships. Constructing the process-AGV edge set. either one Connecting process nodes With AGV nodes capable of performing the transportation tasks of this process , used to represent candidate transportation relationships. Wherein, the edge set and In the initial state, all edges are candidate edges, representing all possible processing and transportation relationships. During scheduling, the corresponding candidate edges are determined as selected edges based on the scheduling actions, representing the actual processing and transportation allocation relationships. The heterogeneous graph of the example instance and the heterogeneous graph of the feasible solutions obtained after scheduling are shown below. Figure 2 As shown.

[0067] Step 2: Embed feature vectors for nodes and edges in the heterogeneous graph.

[0068] Process nodes The original feature vector :

[0069] State: A binary value representing the state at time step Time process Has it been scheduled? (1 if scheduled, 0 if not) Number of adjacent machines : at time step The number of candidate machines that can process this step at any time; the number of adjacent AGVs. : at time step The number of candidate AGVs that can undertake the transportation task of this process; the number of workpieces. The number of processes that have not yet been scheduled; processing time: if If it has already been scheduled, then its actual processing time will be used. ( (If the machine is selected, then the set of candidate machines is used); otherwise, the set of candidate machines is used. Average processing time Transportation time Start time Workpiece completion time: workpiece The estimated or actual completion time.

[0070] Among them, transportation time With start time Based on this process and its preceding processes The scheduling state is estimated recursively. This estimation follows three unified scenarios, achieved by propagating time constraints from scheduled processes to unscheduled processes:

[0071] Scenario 1: and All have been dispatched. and In the machine and The dispatching process is complete. At this point, the transportation time... It is deterministic, and its calculation method is: AGV from its current position Move to Processing machines Time, and from Move to Processing machines The sum of the times. Start time Take it directly as the actual start time The calculation formula is as follows:

[0072]

[0073]

[0074] Scenario 2: Precedence Already scheduled, and Not scheduled. At this time, the transit time... Calculated as an average value, which is the sum of all possible AGV positions to... Processing machines Average driving time and arrive Selected machine set The sum of the average travel times of all machines in the system. Start Time Depend on The actual completion time (i.e.) ) and estimated transit time The sum is calculated. The calculation formula is as follows:

[0075]

[0076]

[0077] Scenario 3: and None were scheduled. At this time, and The estimation method is as follows:

[0078]

[0079]

[0080] machine node The original feature vector :

[0081] Number of adjacent processes : at time step Can be assigned to machine Number of candidate processes; Available time: machine Time to complete the currently assigned process; Utilization rate: machine The proportion of non-idle time to total production time.

[0082] AGV Node The original feature vector :

[0083] Number of adjacent processes : at time step can be Number of adjacent processes undertaking transportation tasks; Available time: AGV Time to complete currently assigned tasks; Utilization: AGV The proportion of non-idle time to total production time.

[0084] OM edge original features :

[0085] Process To the machine Transportation time: if the preceding process If the process has been scheduled, the transportation time will be calculated based on the processing time of that process on the machine; otherwise, it will be based on the preceding process. Average time across all candidate machines; processing time .

[0086] OA side original features :

[0087] AGV From current location information To the preceding process The transportation time of the relevant machines. If If it has been scheduled, then calculate the transportation time to the specific machine; if If no machine has been scheduled yet, then the calculation is performed across the entire set of candidate machines. The average time for each machine.

[0088] Step 3: Dynamically update the heterogeneous graph structure based on the scheduling actions.

[0089] At time step Scheduling actions Triadic decision This indicates that the unscheduled processes will be... Assigned to its workable machines And assign idle AGVs Responsible for process transportation, wherein the actions described satisfy the following feasibility conditions: process It has not yet been scheduled, and it is a workpiece. The first process or its preceding process Scheduling completed; selected machine It belongs to the process Compatible machine set The selected AGV At time step It is in an idle state.

[0090] Execute action Afterward, the system enters the next time step, and its state update includes the structural evolution of the heterogeneous graph and the temporal progression of the environment time. The complete dynamic scheduling process of the example instance is shown in Figures 3(a)-3(h). The specific update process is as follows:

[0091] Structural Evolution: Heterogeneous Diagrams from Updated to ,in:

[0092] At the process node Above, only with the selected machine node OM edge between and with the selected AGV node OA edge between Remove Candidate OM edges between all other machine nodes and candidate OA edges between all other AGV nodes are calculated to reflect the final determination of the processing and transportation allocation for this process; the feature vectors of all nodes and edges are updated and recalculated according to the feature vector calculation method described in step 2.

[0093] Timing Advancement: Update the ambient clock Until the next decision moment The rules for implementation are as follows:

[0094] If an action is performed If there are still idle AGVs, then This allows for continuous decision-making at the same point in time. The process is shown in Figures 3(a)-(b) and... Figures 3(e)-3(f) As shown. If all AGVs are busy, the ambient clock advances to the earliest completed AGV task time. For example, in Figure 3(c), Perform actions at all times After all AGVs are occupied, the environmental clock advances to [time value]. This moment corresponds exactly to the AGV The completion time of the transportation task. In special circumstances, when scheduling procedures... Its preceding processes When processing on the same machine, no transportation task is required, and the assigned AGV... It is released immediately after allocation and is considered to be in an idle state. This special case is shown in Figure 3(h).

[0095] The above process is repeated until all processes are scheduled, thus forming a complete scheduling plan.

[0096] Step 4: Use graph neural networks to extract deep features from this heterogeneous graph.

[0097] In one specific embodiment of the present invention, the graph neural network employs a heterogeneous graph fusion network (HGFN) architecture. This architecture integrates a specially designed gated dynamic attention embedding method (GDAEM) for machine and AGV node embedding, and a graph long short-term memory network (GraphLSTM) for process node embedding, to collaboratively complete deep feature extraction. The network architecture diagram is shown below. Figure 4 As shown. The specific implementation process is as follows:

[0098] (1) Machine node embedding. For each machine node, GDAEM is used for embedding computation, and the process is as follows:

[0099] For machine nodes Each adjacent process node Calculate its attention coefficient :

[0100]

[0101] in, , and These are the feature vectors of the process node, machine node, and OM edge, respectively. , and The weight matrix is ​​a learnable matrix; is a learnable weight vector; LeakyReLU is the activation function. This dynamic attention mechanism allows different machine nodes to assign different levels of importance to the same adjacent process nodes.

[0102] The attention coefficients of all adjacent process nodes of the same machine node are normalized using the softmax function:

[0103]

[0104] Based on the calculated attention weights, the features of adjacent process nodes are weighted and summed to obtain the neighborhood aggregation features of the machine nodes. :

[0105]

[0106] The embedding update of a machine node depends not only on the aggregated information of its neighboring process nodes but also on its own state update. The update of its own features is implemented through a multilayer perceptron containing fully connected layers and a ReLU activation function.

[0107]

[0108] Finally, a gating mechanism is used to aggregate neighborhood features. With its own update features The data is then fused to generate the final embedded representation of the machine node. :

[0109]

[0110]

[0111] in, For the sigmoid function, It is a learnable parameter vector.

[0112] (2) AGV node embedding. For each AGV node, the GDAEM method, consistent with that used for machine nodes, is employed to perform attention aggregation on the features of its adjacent process nodes, and the final embedded representation of the AGV node is obtained through gating fusion with its own features. The calculation process is as follows:

[0113]

[0114]

[0115]

[0116]

[0117]

[0118] (3) Process node embedding. For each process node, its neighborhood includes its direct predecessor process node, direct successor process node, set of optional processing machine nodes, and set of optional AGV nodes. Its embedding calculation is as follows:

[0119] Aggregates multi-source information from its neighboring machine nodes, AGV nodes, and its own characteristics:

[0120]

[0121] in, The input is , The input is , Representative process Its own feature vector, They are responsible for splicing and merging these features.

[0122] Furthermore, considering the temporal dependencies between different processes within the same workpiece (e.g. GraphLSTM is introduced to extract sequential paths from the graph and capture the dynamic evolution features of the process sequence. For each workpiece path... Embed the corresponding nodes into the sequence Input into a bidirectional LSTM network:

[0123]

[0124] in, Indicate process The hidden state is formed by concatenating the forward and backward hidden states of the BiLSTM. The hidden layer dimension for each unidirectional LSTM.

[0125] Finally, the output of GraphLSTM is reduced in dimensionality through a linear projection layer (implemented by an MLP) to obtain the final representation of each process node:

[0126]

[0127] (4) Graph embedding. The above embedding modules are then processed... Layer stacking allows each layer to independently learn high-order features of node states, resulting in a final multi-layer node embedding representation. The process nodes of the final layer are then embedded. Machine node embedding AGV node embedding Each vector is averaged separately, and the three are concatenated to form a graph-level state vector for downstream scheduling decisions:

[0128]

[0129] Step 5: Input the extracted features into the scheduling decision module to generate specific scheduling actions.

[0130] The graph-level state vector and the embedded representation of each node obtained in step 4 are input to the scheduling decision module. The scheduling decision module evaluates the priority of feasible scheduling actions at the current time based on the input features and selects the preferred scheduling action. As the scheduling decision for the current time step; the scheduling action, after execution, is used to update the system state and drive the decision-making process for the next time step. Step 5 is executed cyclically with steps 3 to 4 until all processes are scheduled and the final scheduling scheme is output.

[0131] This invention transforms complex workshop scheduling states into structured graph data through the aforementioned heterogeneous graph representation and feature extraction methods. Utilizing the powerful relational reasoning capabilities of graph neural networks, it achieves efficient and accurate perception of the system state. In terms of computational efficiency, this method avoids combinatorial explosion of the state space through structured updates of heterogeneous graphs and reduces the overhead of manual feature design and iterative computation in traditional methods by extracting features through graph neural networks. Regarding scheduling quality, the extracted deep fusion features can more comprehensively reflect the global situation and local bottlenecks of the system, thus providing richer input to downstream decision-making modules, guiding them to learn better scheduling strategies, and ultimately obtaining a scheduling scheme that shortens completion time.

[0132] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Any modifications or changes made to the present invention by those skilled in the art after reading this application and referring to the above embodiments are within the scope of protection claimed in the pending claims of this application.

Claims

1. A graph-based method for flexible job shop state representation and feature extraction, characterized in that, Includes the following steps: Step 1: Construct a heterogeneous graph, setting the processes, machines, and AGVs in the scheduling system as different types of entity nodes, and establishing edge connections between nodes; Step 2: Based on the heterogeneous graph, embed key features into the corresponding nodes or edges to construct a multi-dimensional attribute set that reflects the real-time system state; Step 3: Each time step The heterogeneous graph structure and the characteristics of nodes and edges described in step 2 are dynamically updated based on the scheduling actions. Step 4: Each time step Based on the real-time heterogeneous graph updated in step 3, deep feature extraction is performed using graph neural networks to fuse scattered multi-source information into dense, low-dimensional feature vectors that can be used for downstream decision-making tasks. Step 5: Input the extracted features into the scheduling decision module to generate specific scheduling actions, and repeat steps 3-4 until all processes are scheduled and the final scheduling plan is output.

2. The method according to claim 1, characterized in that, In step 1, the method for constructing the heterogeneous graph is as follows: The flexible job shop scheduling instance is formalized as a heterogeneous graph. ,in , , These are respectively a set of process nodes, a set of machine nodes, and a set of AGV nodes; two virtual nodes are added to the set of process nodes to represent the start and end points of scheduling; a set of priority constraint edges between processes is constructed. Each directed edge represents the pre- and post-order dependencies between adjacent processes; construct the process-machine edge set. either one Connecting process nodes With its compatibility with machine nodes Used to represent feasible processing relationships; constructing process-AGV edge sets. either one Connecting process nodes With AGV nodes capable of performing the transportation tasks of this process , used to represent candidate transportation relationships; wherein, the edge set and In the initial state, all edges are candidate edges, used to represent all processing and transportation relationships; during the scheduling process, the corresponding candidate edges are determined as selected edges based on the scheduling actions, used to represent the actual processing and transportation allocation relationships.

3. The method according to claim 1, characterized in that, In step 2, the feature vectors of the nodes and edges in the heterogeneous graph include: (1) Process nodes The original feature vector This includes: scheduling status, number of adjacent machines, number of adjacent AGVs, number of unscheduled processes in the workpiece, processing time, transportation time, start time, and workpiece completion time; wherein, the transportation time and start time are based on the process. and its preceding processes The scheduling state is estimated recursively; (2) Machine Node The original feature vector This includes: the number of adjacent processes, available time, and utilization rate; (3) AGV node The original feature vector This includes: the number of adjacent processes, available time, and utilization rate; (4) OM edge original features Process To the machine Transportation time, processing time ; (5) OA side original features Transportation time based on the current position of the AGV and the status of the preceding process.

4. The method according to claim 1, characterized in that, In step 3, the dynamic update of the heterogeneous graph structure based on the scheduling action is as follows: At time step Scheduling actions Triadic decision This indicates that the unscheduled processes will be... Assigned to its workable machines And assign idle AGVs Responsible for process transportation, wherein the actions described satisfy the following feasibility conditions: process It has not yet been scheduled, and it is a workpiece. The first process or its preceding process Scheduling completed; selected machine It belongs to the process Compatible machine set The selected AGV At time step It is in an idle state; Execute action Then, the system enters the next time step, and its state update includes the structural evolution of the heterogeneous graph and the temporal progression of the environment: (1) Structural evolution: Heterogeneous diagrams from Updated to ,in: At the process node Above, only with the selected machine node OM edge between and with the selected AGV node OA edge between Remove Candidate OM edges between all other machine nodes and candidate OA edges between all other AGV nodes are calculated to reflect the final determination of the processing and transportation allocation for this process; the feature vectors of all nodes and edges are updated and recalculated according to the feature vector calculation method described in step 2. (2) Timing Advancement: Update the ambient clock Until the next decision moment The rules for implementation are as follows: If an action is performed If there are still idle AGVs, then This allows for continuous decision-making at the same time; if all AGVs are busy, the environmental clock advances to the earliest completed AGV task; in special cases, when scheduling processes... Its preceding processes When processing on the same machine, no transportation task is required, and the assigned AGV... Once allocated, it is immediately released and considered to be in an idle state; The above process is repeated until all processes are scheduled, thus forming a complete scheduling plan.

5. The method according to claim 1, characterized in that, In step 4, the method for extracting deep features from this heterogeneous graph using a graph neural network is as follows: (1) Machine node embedding; For each machine node, the features of its adjacent process nodes are weighted and aggregated by a graph attention embedding module based on dynamic attention mechanism. The dynamic attention mechanism can adaptively adjust the importance of adjacent nodes according to the different query nodes. At the same time, the machine node's own features are nonlinearly transformed by a multilayer perceptron network, and finally the neighborhood aggregated features and its own updated features are fused by a gating fusion mechanism to obtain the final embedded representation of the machine node. (2) AGV node embedding; For each AGV node, a dynamic attention embedding structure consistent with that of the machine node is adopted to perform attention aggregation on the features of its adjacent process nodes, and the final embedded representation of the AGV node is obtained by gating fusion with its own features. (3) Process node embedding; For each process node, its neighborhood includes its direct predecessor process node, direct successor process node, optional processing machine node set and optional AGV node set; The process node first performs multi-layer perceptron transformation and fusion on the multi-source information from machine nodes, AGV nodes and its own features respectively; Subsequently, in order to capture the sequence dependency relationship between adjacent processes within the same workpiece, the process sequence under the workpiece is time-series embedded calculation using a bidirectional long short-term memory network, and the final embedded representation of the process node is obtained through linear projection; (4) Graph embedding; Perform the above embedding module... Layer stacking allows each layer to independently learn high-order features of node states, thus obtaining the final multi-layer node embedding representation. The process node embedding, machine node embedding, and AGV node embedding of the final layer are averaged and pooled, and then concatenated to form a graph-level state vector for downstream scheduling decisions.

6. The method according to claim 1, characterized in that, In step 5, the extracted features are input into the scheduling decision module to generate specific scheduling actions as follows: The graph-level state vector and the embedded representation of each node obtained in step 4 are input to the scheduling decision module. The scheduling decision module evaluates the priority of feasible scheduling actions at the current time based on the input features and selects the preferred scheduling action. As the scheduling decision for the current time step; after the scheduling action is executed, it is used to update the system state and promote the decision process for the next time step; step 5 is executed in a loop with steps 3 to 4 until all processes are scheduled and the final scheduling scheme is output.

Images