Multi-agent causal constraint scheduling system, method, program product, and medium

Abstract

A multi-agent causal constraint scheduling system, method, program product and medium relate to the technical field of intelligent manufacturing. The system obtains comprehensive data through a feature construction module and generates a standardized feature vector, models the scheduling problem as a space-time heterogeneous graph using a state diagram construction module, based on the standardized feature vector, a scheduling decision module adopts three-layer agent collaborative decision-making at the factory level, workshop level and team level to generate candidate scheduling schemes, a constraint verification module performs feasibility filtering on these candidate schemes to ensure that the schemes meet various constraint conditions, a causal effect evaluation module evaluates the intervention effect of the feasible schemes through a structural causal model to provide a scientific basis for scheme evaluation, and finally a scheduling scheme determination module selects the optimal scheduling scheme according to the effect score and supports dynamic rearrangement. This integrated scheduling decision-making process significantly shortens the scheduling decision-making time and can provide an executable scheduling scheme that meets multiple constraints and has good performance in a timely manner.

Description

Technical Field

[0001] This application relates to the field of intelligent manufacturing technology, specifically to a multi-agent causal constraint scheduling system, method, program product, and medium. Background Technology

[0002] In the field of manufacturing production management, scheduling systems, as a key link between planning and execution, are of great significance for improving resource utilization and production efficiency. This is especially true in discrete manufacturing environments, where companies need to coordinate a large number of personnel, equipment, and processes to ensure the efficient completion of production tasks.

[0003] Existing scheduling systems typically use static optimization algorithms or rule-based methods to generate scheduling plans. These methods treat constraints as post-processing checks, meaning the system first generates an initial plan, then checks if it meets all constraints. When a constraint violation is found, the plan needs to be regenerated or modified, requiring multiple iterations. This separation between constraint verification and plan generation not only prolongs scheduling decision-making time but also makes it difficult to guarantee the optimization level of the final plan. It fails to provide timely, executable scheduling plans that satisfy multiple constraints and have good performance, impacting production efficiency and resource utilization. Summary of the Invention

[0004] In view of this, this application provides a multi-agent causal constraint scheduling system, method, program product and medium.

[0005] Firstly, this application provides a multi-agent causal constraint scheduling system, the system comprising: The feature construction module is used to acquire personnel data, equipment data, production data, and constraint data, and generate standardized feature vectors. The state graph construction module is used to construct a spatiotemporal heterogeneous graph based on the standardized feature vector. The nodes of the spatiotemporal heterogeneous graph include employee nodes, process nodes, equipment nodes, and shift nodes. The scheduling decision module includes a factory-level coordination agent, a workshop-level scheduling agent, and a team-level execution agent, which are used to generate candidate scheduling schemes based on the spatiotemporal heterogeneous graph state. The constraint verification module is used to perform feasibility filtering on the candidate scheduling schemes to obtain feasible scheduling schemes. The causal effect assessment module is used to assess the intervention effect of the feasible scheduling scheme based on the structural causal model and output the effect score; The scheduling scheme determination module is used to select the optimal scheduling scheme from the feasible scheduling schemes based on the effect score, and to trigger a rescheduling after receiving a disturbance event, and output a new optimal scheduling scheme within a preset time limit.

[0006] By adopting the above technical solution, constraint verification is integrated into the scheduling decision-making process, achieving pre-processing of constraints and effectively avoiding the multiple iterations caused by the separation of constraint verification and solution generation in traditional scheduling systems. The system acquires comprehensive data and generates standardized feature vectors through a feature construction module, and models the scheduling problem as a spatiotemporal heterogeneous graph using a state graph construction module, providing a unified state representation for hierarchical decision-making. Based on this state representation, the scheduling decision-making module employs a three-layer intelligent agent collaborative decision-making process at the factory, workshop, and team levels to generate candidate scheduling solutions. The constraint verification module directly performs feasibility filtering on these candidate solutions to ensure that the solutions meet all constraints. Furthermore, the causal effect evaluation module evaluates the intervention effect of feasible solutions using a structural causal model, providing a scientific basis for solution evaluation. Finally, the scheduling solution determination module selects the optimal scheduling solution based on the effect score and supports dynamic rescheduling. This integrated scheduling decision-making process significantly shortens scheduling decision-making time, improves solution generation efficiency and quality, and can promptly provide executable scheduling solutions that satisfy multiple constraints and have good performance, thereby effectively improving production efficiency and resource utilization.

[0007] Optionally, the state graph construction module is used to construct a spatiotemporal heterogeneous graph based on the standardized feature vector, specifically including: The state diagram construction module is used to extract node feature information from the standardized feature vector to construct employee nodes, process nodes, equipment nodes, and shift nodes. Based on the resource matching information in the standardized feature vector, an executable relationship edge is established between the employee node and the process node, and an equipment availability relationship edge is established between the process node and the equipment node. Based on the dependency constraint information in the standardized feature vector, sequential dependency edges are established between process nodes; and based on the resource conflict information in the standardized feature vector, resource competition edges are established between nodes with temporal or spatial conflicts. The spatiotemporal heterogeneous graph is formed by combining various types of nodes and various types of relational edges.

[0008] Optionally, the scheduling decision module includes a factory-level coordination agent, a workshop-level scheduling agent, and a team-level execution agent, used to generate candidate scheduling schemes based on the spatiotemporal heterogeneous graph state, specifically including: The factory-level coordination agent is used to receive the spatiotemporal heterogeneous graph status, analyze the spatiotemporal heterogeneous graph according to the global production target, generate a cross-workshop resource quota allocation scheme, and distribute the resource quota allocation scheme to the workshop-level scheduling agent. The workshop-level scheduling agent is used to receive the resource quota allocation scheme, perform local analysis on the spatiotemporal heterogeneous graph under the constraints of the resource quota allocation scheme, generate a shift allocation matrix within the workshop, and send the shift allocation matrix to the shift-level execution agent. The team-level execution agent is used to receive the shift allocation matrix, refine and adjust the shift allocation matrix according to the team execution window, and generate the candidate shift scheduling scheme containing the specific employee-process-equipment-shift allocation relationship.

[0009] Optionally, the constraint verification module is used to perform feasibility filtering on the candidate scheduling schemes to obtain feasible scheduling schemes, specifically including: The constraint verification module is used to obtain a preset set of hard constraint rules and a set of soft constraint rules; Each of the candidate scheduling schemes is subjected to hard constraint verification, which includes skill matching verification, equipment capability verification, time conflict verification, and process dependency verification. Candidate scheduling schemes that violate any hard constraint rule are marked as infeasible and filtered out; Perform soft constraint evaluation on candidate scheduling schemes that pass hard constraint verification and calculate the degree of soft constraint violation; The candidate scheduling schemes that pass the hard constraint verification are output as the feasible scheduling schemes, and the soft constraint violation degree is added to each feasible scheduling scheme.

[0010] Optionally, the causal effect assessment module is used to assess the intervention effect of the feasible scheduling scheme based on a structural causal model and output an effect score, specifically including: The causal effect evaluation module is used to construct a structural causal model for the scheduling scenario. The structural causal model includes endogenous variables, exogenous variables, and causal relationship functions. The endogenous variables include output per unit working hour, delay rate, overtime duration, and personnel load dispersion. The feasible scheduling scheme is substituted into the structural causal model as an intervention variable to calculate the predicted values ​​of each endogenous variable after intervention. Based on the counterfactual reasoning framework, a control scenario is constructed in which no corresponding feasible scheduling scheme is implemented, and the baseline values ​​of each endogenous variable in the control scenario are calculated. The difference between the predicted value and the baseline value is calculated to obtain the average intervention effect of each endogenous variable; The average intervention effect of each endogenous variable is weighted and summed according to the preset weight coefficients, and the effect score of each feasible scheduling scheme is generated by combining the soft constraint violation degree of each feasible scheduling scheme.

[0011] Optionally, a scheduling scheme determination module is used to select the optimal scheduling scheme from the feasible scheduling schemes based on the effect score, and trigger a rescheduling after receiving a disturbance event, outputting a new optimal scheduling scheme within a preset time limit, specifically including: The scheduling scheme determination module is used to sort the effect scores and select the feasible scheduling scheme with the highest effect score as the optimal scheduling scheme. The optimal scheduling plan is output to the production execution system, and disturbance events during the production execution process are monitored in real time. The disturbance events include at least equipment failure, staff absence, order changes, and process delays. When the disturbance event is detected, determine the degree of impact of the disturbance event on the current scheduling plan; If the impact exceeds a preset threshold, a rescheduling process is triggered. In the rescheduling process, the personnel data, equipment data, production data, and constraint data are updated. The updated data is input into the feature construction module, and the state diagram construction, scheduling decision, constraint verification, and causal effect evaluation are re-executed. The rescheduling calculation is completed within a preset time limit, and a new optimal scheduling plan is output.

[0012] Optionally, the system further includes: The interpretation and auditing module is used to assign a unique scheme identifier to each generated optimal scheduling scheme and save the input snapshot corresponding to the optimal scheduling scheme. The input snapshot includes personnel data, equipment data, production data and constraint data at the time the current scheduling was triggered. Record the objective function value of the scheduling scheme, wherein the objective function value includes at least the effect score and the scores of each item in the calculation process; Save constraint hit details, which record the verification results of hard constraints and the violation of soft constraints; Record the manual approval actions related to the optimal scheduling plan, including the approver, approval time, approval comments, and approval results.

[0013] A second aspect of this application provides a multi-agent causal constraint scheduling method, the method comprising: Acquire personnel data, equipment data, production data, and constraint data, and generate standardized feature vectors; A spatiotemporal heterogeneous graph is constructed based on the standardized feature vectors. The nodes of the spatiotemporal heterogeneous graph include employee nodes, process nodes, equipment nodes, and shift nodes. Candidate scheduling schemes are generated based on the spatiotemporal heterogeneous graph state. Perform feasibility filtering on the candidate scheduling schemes to obtain feasible scheduling schemes; The intervention effect of the feasible scheduling scheme is evaluated based on a structural causal model, and an effect score is output. Based on the effect score, the optimal scheduling scheme is selected from the feasible scheduling schemes, and a rescheduling is triggered after receiving a disturbance event, and a new optimal scheduling scheme is output within a preset time limit.

[0014] A third aspect of this application provides a computer program product containing instructions that, when run on a multi-agent causal constraint scheduling system, cause the multi-agent causal constraint scheduling system to perform the methods described in the second aspect and any possible implementation thereof.

[0015] A fourth aspect of this application provides a computer-readable storage medium including instructions that, when executed on a multi-agent causal constraint scheduling system, cause the multi-agent causal constraint scheduling system to perform the methods described in the second aspect and any possible implementation thereof.

[0016] In summary, one or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: This system integrates constraint verification into the scheduling decision-making process, enabling pre-processing of constraints and effectively avoiding the multiple iterations caused by the separation of constraint verification and solution generation in traditional scheduling systems. The system acquires comprehensive data and generates standardized feature vectors through a feature construction module, and models the scheduling problem as a spatiotemporal heterogeneous graph using a state graph construction module, providing a unified state representation for hierarchical decision-making. Based on this state representation, the scheduling decision-making module employs a three-tiered intelligent agent collaborative decision-making process at the factory, workshop, and team levels to generate candidate scheduling solutions. The constraint verification module directly performs feasibility filtering on these candidate solutions, ensuring that the solutions meet all constraints. Furthermore, the causal effect evaluation module uses a structural causal model to evaluate the intervention effects of feasible solutions, providing a scientific basis for solution evaluation. Finally, the scheduling solution determination module selects the optimal scheduling solution based on the effect score and supports dynamic rescheduling. This integrated scheduling decision-making process significantly shortens scheduling decision-making time, improves solution generation efficiency and quality, and can promptly provide executable scheduling solutions that satisfy multiple constraints and have good performance, thereby effectively improving production efficiency and resource utilization. Attached Figure Description

[0017] Figure 1 This is an architecture diagram of a multi-agent causal constraint scheduling system provided in an embodiment of this application; Figure 2 This is a schematic diagram of a spatiotemporal heterogeneous graph state representation provided in an embodiment of this application; Figure 3 This is a hierarchical multi-agent collaborative process structure diagram provided in an embodiment of this application; Figure 4 This is an example diagram of a hard constraint verification process provided in an embodiment of this application; Figure 5 This is a flowchart of a causal effect assessment and scheme selection provided in an embodiment of this application; Figure 6 This is an event-driven rescheduling sequence diagram provided in an embodiment of this application; Figure 7 This is a schematic diagram of an explanation and audit data structure provided in an embodiment of this application; Figure 8 This is a flowchart illustrating a multi-agent causal constraint scheduling method provided in an embodiment of this application. Detailed Implementation

[0018] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0019] In the description of the embodiments of this application, the words "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design that is described as "for example" or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design options. Rather, the use of the words "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0020] In the description of the embodiments of this application, the term "multiple" means two or more. For example, multiple systems means two or more systems, and multiple screen terminals means two or more screen terminals. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0021] Please see Figure 1This document presents an architecture diagram of a multi-agent causal constraint scheduling system provided in an embodiment of this application. The system is based on a data aggregation and feature construction module, which receives and processes personnel, equipment, production, and constraint data to generate standardized feature vectors. Four core functional modules are configured in parallel: a state graph construction module is responsible for constructing a spatiotemporal heterogeneous graph based on feature vectors, including nodes such as employees, processes, equipment, and shifts; a hierarchical scheduling decision module adopts a three-layer architecture of factory-level coordinating agents, workshop-level scheduling agents, and shift-level execution agents to collaboratively generate candidate scheduling schemes; a hard constraint verification module performs feasibility filtering on candidate schemes to ensure that the schemes meet hard requirements such as skill matching and equipment capabilities; and a causal effect evaluation module evaluates the intervention effect of feasible schemes based on a structural causal model. The system also includes two auxiliary modules: an event-driven rescheduling module triggers a rescheduling process upon receiving a disturbance event and outputs a new optimal scheme within a preset time limit; and an interpretation and auditing module provides traceability support for scheduling schemes, recording scheme identifiers, objective function values, constraint hit details, and manual approval information to achieve complete recording and auditing of the decision-making process.

[0022] This application provides a multi-agent causal constraint scheduling system, which includes a feature construction module, a state graph construction module, a scheduling decision module, a constraint verification module, a causal effect evaluation module, and a scheduling scheme determination module.

[0023] The aforementioned feature construction module is used to acquire personnel data, equipment data, production data, and constraint data, and generate standardized feature vectors.

[0024] Specifically, the feature construction module first acquires multi-source heterogeneous data through data interfaces, including: personnel data (such as employee ID, skill level, working hours, and leave plans), equipment data (such as equipment ID, type, processing capacity, and maintenance plans), production data (such as work order number, product type, process route, and delivery date), and constraint data (such as working hour regulations, human-machine matching rules, and priority strategies). Since these data originate from different systems and have varying formats, the system uses an ETL (Extract-Transform-Load) process for processing. In the extraction phase, raw data is obtained from ERP, MES, and human resource systems through pre-configured API interfaces. In the transformation phase, the system cleans the data (removing outliers and missing values), standardizes (normalizing features of different dimensions to the [-1,1] interval), and encodes (such as converting categorical features to one-hot encoding). In the loading phase, the system organizes the processed data into structured feature vectors. In its implementation, the system employs feature engineering techniques to combine and transform original features. For example, it combines employee proficiency with equipment operating efficiency to generate the "human-machine collaboration efficiency" feature, and combines process duration with the completion status of preceding processes to generate the "process readiness" feature. Furthermore, the system captures dynamic change patterns by extracting time-series features and calculating statistics (such as mean, variance, and trend) within a rolling time window. The processed, standardized feature vectors are organized into a set of key-value pairs in the form of <entity ID, feature vector>, providing a unified data foundation for subsequent state diagram construction. This avoids the fragmented constraint processing problems caused by inconsistent data formats in traditional scheduling systems. Simultaneously, the rich feature representation improves the accuracy and adaptability of scheduling decisions, creating conditions for pre-constraint processing.

[0025] The aforementioned state graph construction module is used to construct a spatiotemporal heterogeneous graph based on standardized feature vectors. The nodes of the spatiotemporal heterogeneous graph include employee nodes, process nodes, equipment nodes, and shift nodes.

[0026] Specifically, the state graph construction module is used to extract node feature information from the standardized feature vector after receiving it, and to construct employee nodes, process nodes, equipment nodes, and shift nodes.

[0027] Please see Figure 2 This is a schematic diagram of a spatiotemporal heterogeneous graph state representation provided in an embodiment of this application. Figure 2The diagram shows employee nodes (e.g., E1, E2, E3) represented by blue circles. Each node contains attributes such as employee ID, skill level matrix, working time limit, and priority, represented as V_emp = {id, skill_matrix, time_availability, priority}. Process nodes (e.g., O1, O2) represented by pink rectangles contain attributes such as process ID, work order, standard working hours, and quality requirements, represented as V_proc = {id, order_id, std_time, quality_req}. Equipment nodes (e.g., M1, M2) represented by green ellipses contain attributes such as equipment ID, equipment type, processing capacity parameters, and reliability indicators, represented as V_equip = {id, type, capability, reliability}. Shift nodes, which are not directly shown in the diagram but exist in the system, contain attributes such as shift ID, start time, end time, and shift type.

[0028] Subsequently, relationship edges between nodes are constructed based on resource matching information. Figure 2 As shown, the system establishes executable relationship edges between employee nodes E1 and E2 and process node O1, indicating that these employees possess the skill qualifications to perform the process. Edge establishment is based on matching score calculation: match_score = Σ(skill_weight_i × skill_level_i), and a connection is formed when the matching score exceeds a preset threshold. Similarly, the system establishes equipment availability relationship edges between process node O1 and equipment node M1, and between process O2 and equipment M2, indicating that these processes can be performed on the corresponding equipment. This connection is established only when the equipment's processing capacity meets the process quality requirements.

[0029] Furthermore, the state graph construction module processes dependency constraint information between nodes. Figure 2 The connection between processes O1 and O2 indicates a sequential dependency between them. This dependency is extracted from the process route, and a directed edge e_dep(O1→O2) is established. The edge weight w_dep reflects the strength of the dependency, such as the impact of transition time or quality. For resource competition edges, an edge is established when a time or space conflict is detected between two nodes. For example, if two processes require the same equipment and their times overlap, the system establishes a resource competition edge, and the edge weight reflects the severity of the conflict.

[0030] Finally, all nodes and edges in the state graph construction module are combined to form the following: Figure 2The spatiotemporal heterogeneous graph shown is an example of a comprehensive representation method that encodes spatial relationships (such as resource matching and dependency constraints) and temporal relationships (such as temporal dependencies and resource conflicts) in the scheduling problem into a unified graph structure. This provides a complete problem representation for the scheduling decision module, realizes pre-processing of constraints, and avoids the separation between constraint verification and solution generation in traditional methods.

[0031] The aforementioned scheduling decision module includes a factory-level coordination agent, a workshop-level scheduling agent, and a team-level execution agent, which are used to generate candidate scheduling schemes based on the spatiotemporal heterogeneous graph state.

[0032] Specifically, the Factory Coordination Agent (FCA) serves as the top-level decision-making unit of the scheduling system, responsible for global resource coordination and strategy formulation. Based on the overall state of the spatiotemporal heterogeneous graph, the FCA analyzes production demand and resource status at a macro level, formulating cross-workshop resource allocation strategies. Its main responsibilities include assessing the overall plant capacity status, predicting production bottlenecks, coordinating resource quota allocation among multiple workshops, formulating priority strategies, and monitoring the execution of the overall scheduling plan. The FCA achieves rational resource allocation and globally optimal production targets by distributing resource quota allocation plans to each workshop-level agent.

[0033] Workshop Scheduling Agent (WSA): As the mid-level decision-making unit of the scheduling system, it is responsible for optimizing scheduling decisions within the workshop under the resource quota constraints allocated by the factory-level agent. The WSA receives the resource quota allocation plan issued by the FCA (Factory Control Authority) and, combined with the specific production conditions and constraints of the workshop, generates a shift allocation matrix within the workshop. Its main responsibilities include shift allocation of personnel resources within the workshop, prioritizing work processes, formulating equipment usage plans, and handling resource conflicts within the workshop. The WSA provides the decision-making basis for the shift-level agent by transforming abstract resource quotas into concrete shift allocation plans.

[0034] Team Execution Agent (TEA): As the underlying decision-making unit of the scheduling system, the TEA is responsible for optimizing the specific execution details of the scheduling plan. The TEA receives the shift allocation matrix from the WSA and, combined with the actual execution environment and micro-constraints of the team, concretizes the abstract shift arrangement into precise employee-process-equipment-shift allocation relationships. Its main responsibilities include optimizing the matching of specific employees with specific processes and equipment, handling real-time constraints at the execution level (such as tool preparation time, material availability, etc.), and responding to small-scale disturbances during execution. The TEA generates a final, directly executable scheduling plan by refining and optimizing the shift allocation plan.

[0035] Please see Figure 3 This diagram illustrates a hierarchical multi-agent collaborative workflow structure, where a Factory Coordinating Agent (FCA) serves as the top-level decision-making unit, responsible for formulating global scheduling strategies and coordinating resource allocation. The FCA decomposes tasks and distributes them to two workshop-level scheduling agents, WSA-A and WSA-B, each responsible for scheduling decisions in different workshops. The workshop-level agents further subdivide tasks to the team level. WSA-A manages team-level execution agents TEA-1 and TEA-2, while WSA-B manages TEA-3. These team-level execution agents are responsible for optimizing the execution details of specific scheduling plans. The dashed lines in the diagram represent the horizontal collaboration mechanism between team-level agents, enabling adjacent teams to negotiate shared resources and task handover, ensuring the continuity and consistency of scheduling plans. This hierarchical agent architecture, through top-down task decomposition and bottom-up information aggregation, ensures both the rational allocation of global resources and flexible handling of local constraints, effectively solving the problem of disconnect between constraint processing and plan generation in traditional scheduling systems, thus improving the efficiency and quality of scheduling decisions.

[0036] Specifically, the Factory-Level Coordinating Agent (FCA) first receives a complete spatiotemporal heterogeneous graph, as shown in the figure below. Figure 2 The diagram shows employee nodes (E1, E2, E3), process nodes (O1, O2), equipment nodes (M1, M2), and their relationships. FCA uses graph analysis algorithms to process the spatiotemporal heterogeneous graph, identifying key resources, production bottlenecks, and process paths. For example, FCA analyzes... Figure 2 The node connection pattern in the system might identify process O1 as a critical node connecting multiple employees and equipment, requiring priority resource allocation. Subsequently, FCA, combined with the company's overall production goals (such as maximizing capacity utilization, minimizing delivery delays, or balancing the load of multiple production lines), generates a cross-workshop resource allocation plan. This plan specifically includes: the quantity and type of available human resources in each workshop (e.g., assigning complementary employees E1 and E2 to workshop 1, and E3 to workshop 2), equipment usage time slots (e.g., allocating equipment M1 to workshop 1 in the morning and workshop 2 in the afternoon), critical material quotas, and the capacity target values ​​for each workshop. FCA then distributes these resource allocation plans to the corresponding workshop-level scheduling agents through internal system interfaces, while simultaneously setting the plan's validity period and update mechanism.

[0037] The Workshop-Level Scheduling Agent (WSA) receives the resource quota allocation scheme issued by the FCA and performs local analysis based on the subgraph portion corresponding to the workshop in the spatiotemporal heterogeneous graph (e.g., nodes and relationships containing E1, E2, O1, and M1). Under the constraint of resource quotas, the WSA transforms the scheduling problem into a resource-time allocation problem, considering factors such as shift structure, personnel skill matching, and equipment availability. The WSA uses a combinatorial optimization algorithm to generate a shift allocation matrix within the workshop, which explicitly specifies each employee's work shift, the set of processes they are responsible for, and the equipment they use. For example, for... Figure 2 In this scenario, WSA might schedule employee E1 to use equipment M1 to handle process O1 during the Monday morning shift, while employee E2 continues the related work during the Monday evening shift. The system considers the dependencies between processes O1 and O2 to ensure production flow continuity. When generating the shift allocation matrix, WSA prioritizes meeting the resource needs of key processes and considers factors such as employee work continuity, skill complementarity, and fatigue management. WSA then distributes the generated shift allocation matrix to the corresponding shift-level execution agents and establishes a feedback channel to collect status updates during execution.

[0038] The Team-Level Execution Agent (TEA) receives the shift allocation matrix from the Work Execution Advisor (WSA) and performs final refinements within the actual execution environment. Based on the team execution window (typically a specific shift or a shorter timeframe) and the actual on-site conditions, the TEA fine-tunes and optimizes the shift allocation matrix. Factors processed by the TEA include real-time employee availability, current equipment operating status, tooling and fixture readiness, and actual material arrival time. Through these refinements, the TEA generates multiple candidate scheduling schemes with different characteristics, each containing a complete employee-process-equipment-shift allocation relationship. For example, TEA might generate three candidate scheduling schemes: Candidate Scheme 1 arranges for employee E1 to use equipment M1 to complete process O1 from 08:00 to 12:00 on Monday, and employee E2 to use equipment M1 to complete process O2 from 13:00 to 17:00 on Monday; Candidate Scheme 2 arranges for employee E1 to take turns using equipment M1 to complete process O1 from 08:00 to 10:00 on Monday and employee E2 to complete process O1 from 10:00 to 12:00 on Monday, and then employee E2 to use equipment M2 to complete process O2 from 14:00 to 18:00 on Monday; Candidate Scheme 3 arranges for employee E2 to use equipment M1 all day to complete process O1 and then process O2, while employee E1 is responsible for other tasks. These candidate scheduling schemes each have their advantages and disadvantages, and optimize different objectives (e.g., Scheme 1 minimizes the number of equipment changes, Scheme 2 balances employee load, and Scheme 3 reduces waiting time between processes). TEA passes these candidate scheduling schemes to the hard constraint verification module and the causal effect evaluation module, which further filter and evaluate them to finally determine the optimal scheduling scheme.

[0039] The aforementioned constraint verification module is used to perform feasibility filtering on candidate scheduling schemes to obtain feasible scheduling schemes.

[0040] Specifically, the constraint verification module first retrieves the preset hard constraint rule set and soft constraint rule set from the system configuration library. The hard constraint rule set contains conditions that must be strictly met, such as "employees must possess the skill level required to perform the process", "equipment processing capacity must meet the process quality requirements", "the same employee or equipment can only perform one task at a time", and "processes must be executed in the order of dependencies", etc. The soft constraint rule set contains conditions that are given priority but can be violated, such as "avoiding frequent employee job changes", "reducing equipment idle time", "balancing workload", and "prioritizing high-priority work orders", etc. Each soft constraint rule is accompanied by a weight value, indicating its importance.

[0041] Please see Figure 4 This is an example diagram of a hard constraint verification process provided in an embodiment of this application. The constraint verification module adopts a strict sequential verification process for each candidate scheduling scheme (i.e., Figure 4 The "Input Candidate Action A_i" in the code performs hard constraint checks. First, "Check 1: Working Hours and Rest Rules" is performed to verify whether the scheduling scheme complies with legal working hour limits, necessary rest intervals, and maximum continuous working hours, as required by labor regulations. For example, it checks whether employee E1's schedule exceeds the limit of 8 hours per day or 40 hours per week, and whether legally mandated rest intervals are guaranteed. Next, "Check 2: Qualification and Time-Based Matching" is performed to verify whether the employee's skill level, qualification certificate, and authorization level meet the requirements of the assigned work process. For example, for... Figure 2 When employee E1 is assigned to perform process O1, the system checks whether E1's skill vector meets the skill requirement vector of O1. Then, "Verification 3: Process Dependency and Equipment Capability" ensures that processes are arranged in the correct dependency order, such as... Figure 2 In the intermediate process O1, it must be completed before process O2, and the processing capacity of the equipment must be verified to meet the technical parameter requirements of the process, such as whether equipment M1 has the processing accuracy and capability required to perform process O1. Finally, "Verification 4: Key Adjustment Quantity Minimum Limit Check" is performed to check whether the resource allocation follows the constraints of time and space location and avoids conflict allocation, such as ensuring that employees E1 and E2 are not assigned to perform work in different locations at the same time, or that equipment M1 is not repeatedly booked.

[0042] If a candidate scheduling plan fails at any verification step, the constraint verification module will mark it as infeasible and filter it out of the candidate set. For example, if employee E1's scheduling violates the maximum continuous working hours rule, or employee E2 is assigned to operate two devices simultaneously, the candidate plan will be marked as infeasible and subsequent verification will stop. Only those that pass... Figure 4Only candidate solutions from all four verification steps will be output as the set of feasible actions, A_feasible.

[0043] For candidate scheduling schemes that have passed hard constraint validation, the constraint validation module further performs soft constraint evaluation. The module calculates the degree to which each scheme violates each soft constraint rule and combines the rule weights to calculate the overall soft constraint violation rate. For example, if a scheme causes employee E1 to frequently move between different work areas throughout the day, violating the soft constraint of "reducing unnecessary movement," the module will calculate the degree of violation based on the number of moves and the distance. If a scheme causes some employees to have a significantly higher workload than others, violating the soft constraint of "balancing workload," the module will calculate the degree of imbalance based on the standard deviation. The soft constraint violation rate is weighted to obtain a total score; the lower the score, the higher the scheme quality.

[0044] Finally, the constraint verification module marks all candidate solutions that pass the hard constraint verification as feasible scheduling solutions and attaches the calculated soft constraint violation score to each solution. This information is output in structured data format, including solution identifier, feasibility indicator, total soft constraint violation score, and specific violation details of each soft constraint.

[0045] The aforementioned causal effect assessment module is used to assess the intervention effect of feasible scheduling schemes based on a structural causal model and output an effect score.

[0046] Specifically, the causal effect assessment module first constructs a structural causal model (SCM) for the scheduling scenario. This model includes three core elements: endogenous variables, exogenous variables, and causal relationship functions. Endogenous variables are key indicators in the system that are influenced by other variables, including productivity, delay rate, overtime, load variance, and switch cost. Exogenous variables are factors that affect the system but are not affected by other variables within the system, such as employee basic skill levels, inherent equipment capacity parameters, raw material quality fluctuations, and external order demand. Causal relationship functions define the quantitative dependencies between variables; for example, Delay = f1(Productivity, OrderDemand) indicates that the delay rate is jointly determined by productivity and order demand. The SCM is represented by a directed acyclic graph (DAG), capturing the causal paths and influence strengths between variables.

[0047] like Figure 5The diagram shows a flowchart of causal effect assessment and scheme selection provided in an embodiment of this application. The system first receives the set of feasible scheduling schemes A_feasible output by the constraint verification module as input. For each feasible scheme, the causal effect assessment module performs a "counterfactual effect assessment calculation c(a)", that is, the scheduling scheme a is treated as an intervention variable and substituted into the structural causal model to calculate the predicted values ​​of each endogenous variable after the intervention. For example, if a feasible scheme arranges highly skilled employee E1 to focus on the critical process O1, the model predicts that this arrangement may lead to a 10% increase in output per unit of working hours, while reducing overtime by 2 hours, but increasing the personnel workload dispersion by 0.15 (indicating a more uneven workload distribution). Simultaneously, the module constructs a control scenario without implementing this scheduling scheme based on a counterfactual reasoning framework, calculating the baseline values ​​of each endogenous variable, such as the expected output and delay rate under the condition of maintaining the current scheduling unchanged.

[0048] Next, the module calculates the differences of each endogenous variable before and after the intervention, obtaining the average intervention effect (ATE) of each feasible scheduling scheme on each endogenous variable. For example, for scheme a, the system may obtain the following effect values: Delay effect is -5% (reducing the delay rate by 5%), Overtime effect is -2 hours (reducing overtime by 2 hours), LoadVar effect is +0.15 (more uneven load distribution), SwitchCost effect is +3 (increasing the number of shift changes by 3), and CausalGain effect is +8% (increasing the causal gain by 8%).

[0049] Subsequently, as Figure 5 As shown in the "Multi-objective score calculation J(a)" step, the system performs a weighted summation of these intervention effects according to preset weight coefficients to form a comprehensive score J. The score calculation formula is: minimizeJ=w1Delay+w2Overtime+w3LoadVar+w4SwitchCost-w5*CausalGain.

[0050] Among them, w1 to w5 are non-negative weight coefficients, satisfying wi>=0, reflecting the enterprise's priority for different business objectives. For example, if the enterprise focuses more on on-time delivery, w1 (delay weight) will be set to a higher value; if it focuses more on employee work experience, w3 (load balancing weight) will be higher. By adjusting these weights, the system can flexibly adapt to the demand preferences of different production scenarios. The smaller the score J value, the better the overall performance of the scheduling plan. During the calculation, the system also considers the soft constraint violation degree output by the constraint verification module, incorporating it as a penalty item into the score calculation to ensure that the final selected plan has both good causal effect and meets the soft constraint requirements as much as possible. Finally, the system ranks all feasible scheduling plans based on the comprehensive score J, such as... Figure 5As shown at the bottom, the optimal scheduling scheme and scheduling indicator pair X_i are output, that is, the scheduling scheme with the smallest J value and its corresponding predicted key performance indicator value.

[0051] The aforementioned scheduling scheme determination module is used to select the optimal scheduling scheme from the feasible scheduling schemes based on the effect score, and to trigger a rescheduling after receiving a disturbance event, and output a new optimal scheduling scheme within a preset time limit.

[0052] Specifically, the scheduling module first sorts the effect scores output by the causal effect assessment module and selects the feasible scheduling plan with the highest effect score (i.e., the smallest J value) as the optimal scheduling plan. This optimal plan includes detailed resource allocation information, such as the spatiotemporal resource arrangements like "employee E1 uses equipment M1 to complete process O1 from 8:00 to 10:00" and "employee E2 uses equipment M2 to complete process O3 from 10:30 to 12:30". The system also generates corresponding key performance indicator predictions, such as expected completion time, resource utilization rate, and production efficiency.

[0053] like Figure 6 The diagram shown is an event-driven rescheduling sequence diagram provided in an embodiment of this application. After the optimal scheduling plan is determined, it is sent to the production execution system for implementation. At the same time, the perception layer continuously monitors potential disturbance events that may occur during the production process. When the perception layer detects a disturbance event (such as equipment failure, staff absence, order changes, or process delays), the "event reporting" process will be triggered, and relevant information (including event type, occurrence time, and scope of impact) will be transmitted to the scheduling engine.

[0054] Upon receiving a disturbance event, the scheduling engine immediately executes the "Trigger Quantity Scheduling (Delta_t)" process to calculate the impact of the disturbance event on the current scheduling plan. The system assesses the severity of the disturbance based on preset rules. For example, a failure of equipment M1 may cause delays in all processes that depend on that equipment; employee E3's leave may affect the execution of the specific process they are responsible for; a new urgent order may require the reallocation of resources, etc. For instance, the system may determine that the impact of a failure of equipment M1 is 0.7 (out of 1.0), and if this value exceeds the preset threshold of 0.5, a rescheduling process is triggered.

[0055] After a rescheduling is triggered, the scheduling engine sends a "return available action score" request to the constraint / causality evaluation module to initiate the rescheduling calculation. At this time, the system updates personnel data (such as a list of available personnel), equipment data (such as equipment status), production data (such as process progress), and constraint data (such as new deadlines). The updated data is then input into the feature building module, and the complete scheduling process is re-executed: building a state diagram, generating candidate scheduling schemes, performing constraint checks, and evaluating causal effects. For example, if equipment M1 is expected to require 4 hours of maintenance, the system will mark M1 as unavailable during this period during the rescheduling process and look for alternative equipment or adjust the process sequence.

[0056] It should be noted that, as Figure 6 As shown, the rearrangement calculation needs to be completed within a preset time limit. Depending on the nature and scope of the disturbance, the time limit can be set from 30 seconds (small-scale adjustment) to 5 minutes (large-scale rearrangement). If some processes have already started, the system will try to keep the allocation of these processes unchanged, only rearranging the affected and unstarted parts to reduce production interruptions.

[0057] After the rescheduling calculation is completed, the scheduling engine sends an "adjustment scheduling instruction" to the execution system, conveying the new optimal scheduling plan. The execution system implements the updated scheduling plan and returns "execution feedback and status feedback" information to confirm the receipt and execution of the instruction. This information is then transmitted back to the perception layer, completing one full disturbance handling cycle.

[0058] like Figure 6 As shown, the perception layer will continue to monitor the production status at a later time (t+Δt). If new disturbances or anomalies are detected, the "update status issue" process will be triggered again, and a new round of rescheduling will begin. This event-driven rescheduling mechanism ensures that the scheduling system can respond promptly to various changes in the production environment, maintaining the adaptability and effectiveness of the scheduling scheme.

[0059] As an optional embodiment, the multi-agent causal constraint scheduling system may also include an interpretation and auditing module, specifically used to provide traceability support for scheduling schemes, record scheme identifiers, objective function values, constraint hit details and manual approval information, and realize the complete recording and auditing functions of the decision-making process.

[0060] Specifically, the interpretation and auditing module assigns a globally unique plan identifier (plan_id) to each generated optimal scheduling plan, such as "SCH-20260414-001". This identifier serves as the primary key for associating various audit records. When the scheduling plan determination module outputs the optimal scheduling plan, the system immediately creates a scheduling plan record (ScheduleRecord).

[0061] like Figure 7The diagram shown is a schematic representation of an interpretation and auditing data structure provided in an embodiment of this application. Figure 7 As shown in the white box on the left, this record contains the plan_id, the generation timestamp (e.g., "2026-04-14 09:23:45"), the objective function value (e.g., "J=0.18"), the selected resource allocation action (which includes a detailed allocation plan for employees, processes, equipment, and time periods), and the incremental adjustment instruction (delta_instruction, which only records the changes relative to the previous plan in the reordering scenario).

[0062] Simultaneously, the system saves a snapshot of the input corresponding to this plan, i.e., the complete data state at the time the scheduling was triggered, including personnel data (such as a list of available employees on duty and their skill matrix), equipment data (such as the current status and capacity parameters of each piece of equipment), production data (such as a list of work orders to be processed and process progress), and constraint data (such as delivery requirements and resource limitations). This data is stored in a structured format and associated with the plan_id, ensuring that the complete environmental context in which the scheduling plan was generated can be reproduced at any time.

[0063] like Figure 7 As shown in the blue box, the system creates a constraint hit record (ConstraintLog) for each scheduling plan, linked to the corresponding scheduling plan record via the foreign key plan_id (FK). The constraint hit record details the results of each check performed by the constraint verification module, including the rule identifier (rule_id, such as "HC001-Qualification Matching" or "SC003-Load Balancing"), pass or fail status (pass_or_fail), and violation details (violation_detail). For example, for the hard constraint "Employee Skill Matching," the record might show "PASS: Employee E1 meets all skill requirements of process O1"; while for the soft constraint "Reduce Equipment Idle Time," the record might show "Violation degree 0.15: Equipment M2 was idle for 1 hour from 14:00 to 15:00." These detailed records allow managers to understand how the plan meets or violates various business rules.

[0064] like Figure 7As shown in the green box, the system also generates an ExplainLog for each plan, which is linked to the scheduling plan record via the foreign key plan_id (FK). The ExplainLog contains key features (top_features) that influence scheduling decisions, such as the top 5 decision factors including "employee E1's skill proficiency" and "equipment M3's idle time"; causal score details, showing the intervention effect score of each endogenous variable, such as "delay rate improvement: 0.05, overtime reduction: 2.5 hours"; and natural language explanation, providing human-understandable reasons for the decision, such as "This scheduling plan prioritizes on-time delivery. By reasonably allocating highly skilled employees to key processes, it is expected to reduce the delay rate by 5% and reduce overtime by 2 hours, but may lead to a slightly uneven workload distribution."

[0065] Finally, as Figure 7 As shown in the yellow box on the right, when a scheduling plan requires manual review, the system maintains an approval trail record, linked to the scheduling plan via the foreign key `plan_id` (FK). The approval trail records the approver's identifier (approver_id, e.g., "MGR001"), the approval action (including approve, reject, or edit), the approval comment (e.g., "Approved, but it is recommended to distribute workload more evenly next time"), and the approval time (e.g., "2026-04-14 10:15:22"). When multiple people participate in the review or when multi-level approval is required, the system records the complete approval process in chronological order, forming an immutable decision-making trail.

[0066] This comprehensive data structure design enables the scheduling system to possess complete auditability and interpretability. For example, if production managers question the rationality of a scheduling plan, they can clearly understand the data conditions, constraints, and evaluation indicators under which the plan was generated, as well as the opinions and modification history of approval personnel at all levels, by querying all records corresponding to plan_id. This transparent mechanism not only meets compliance audit requirements but also enhances users' trust in the system's decisions, while providing valuable historical data for continuous improvement of the scheduling algorithm.

[0067] In one embodiment, please refer to Figure 8This paper presents a flowchart illustrating a multi-agent causal constraint scheduling method. This method can be implemented using a computer program or a microcontroller. The computer program can be integrated into an application or run as a standalone utility application. Specifically, in this embodiment, the method can be applied to a multi-agent causal constraint scheduling system. The method includes steps 101 to 106, as follows: Step 101: Obtain personnel data, equipment data, production data, and constraint data, and generate standardized feature vectors.

[0068] Step 102: Construct a spatiotemporal heterogeneous graph based on the standardized feature vectors. The nodes of the spatiotemporal heterogeneous graph include employee nodes, process nodes, equipment nodes, and shift nodes.

[0069] Step 103: Generate candidate scheduling schemes based on the spatiotemporal heterogeneous graph state.

[0070] Step 104: Perform feasibility filtering on the candidate scheduling schemes to obtain feasible scheduling schemes.

[0071] Step 105: Evaluate the intervention effect of the feasible scheduling scheme based on the structural causal model and output the effect score.

[0072] Step 106: Select the optimal scheduling scheme from the feasible scheduling schemes based on the effect score, and trigger a rescheduling after receiving a disturbance event, and output the new optimal scheduling scheme within the preset time limit.

[0073] The specific implementation of the above methods and steps can be found in the system section described above, and will not be elaborated further here.

[0074] This application also provides a computer storage medium that can store multiple instructions. The instructions are adapted to be loaded and executed by a processor as described in the above embodiments of a multi-agent causal constraint scheduling method. For the specific execution process, please refer to the detailed description of the above embodiments, which will not be repeated here.

[0075] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0076] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0077] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A multi-agent causal constraint scheduling system, characterized in that, The system includes: The feature construction module is used to acquire personnel data, equipment data, production data, and constraint data, and generate standardized feature vectors. The state graph construction module is used to construct a spatiotemporal heterogeneous graph based on the standardized feature vector. The nodes of the spatiotemporal heterogeneous graph include employee nodes, process nodes, equipment nodes, and shift nodes. The scheduling decision module includes a factory-level coordination agent, a workshop-level scheduling agent, and a team-level execution agent, which are used to generate candidate scheduling schemes based on the spatiotemporal heterogeneous graph state. The constraint verification module is used to perform feasibility filtering on the candidate scheduling schemes to obtain feasible scheduling schemes. The causal effect assessment module is used to assess the intervention effect of the feasible scheduling scheme based on the structural causal model and output the effect score; The scheduling scheme determination module is used to select the optimal scheduling scheme from the feasible scheduling schemes based on the effect score, and to trigger a rescheduling after receiving a disturbance event, and output a new optimal scheduling scheme within a preset time limit.

2. The multi-agent causal constraint scheduling system according to claim 1, characterized in that, The state graph construction module is used to construct a spatiotemporal heterogeneous graph based on the standardized feature vector, specifically including: The state diagram construction module is used to extract node feature information from the standardized feature vector to construct employee nodes, process nodes, equipment nodes, and shift nodes. Based on the resource matching information in the standardized feature vector, an executable relationship edge is established between the employee node and the process node, and an equipment availability relationship edge is established between the process node and the equipment node. Based on the dependency constraint information in the standardized feature vector, sequential dependency edges are established between process nodes; and based on the resource conflict information in the standardized feature vector, resource competition edges are established between nodes with temporal or spatial conflicts. The spatiotemporal heterogeneous graph is formed by combining various types of nodes and various types of relational edges.

3. The multi-agent causal constraint scheduling system according to claim 1, characterized in that, The scheduling decision module, comprising a factory-level coordination agent, a workshop-level scheduling agent, and a team-level execution agent, is used to generate candidate scheduling schemes based on the spatiotemporal heterogeneous graph state, specifically including: The factory-level coordination agent is used to receive the spatiotemporal heterogeneous graph status, analyze the spatiotemporal heterogeneous graph according to the global production target, generate a cross-workshop resource quota allocation scheme, and distribute the resource quota allocation scheme to the workshop-level scheduling agent. The workshop-level scheduling agent is used to receive the resource quota allocation scheme, perform local analysis on the spatiotemporal heterogeneous graph under the constraints of the resource quota allocation scheme, generate a shift allocation matrix within the workshop, and send the shift allocation matrix to the shift-level execution agent. The team-level execution agent is used to receive the shift allocation matrix, refine and adjust the shift allocation matrix according to the team execution window, and generate the candidate shift scheduling scheme containing the specific employee-process-equipment-shift allocation relationship.

4. The multi-agent causal constraint scheduling system according to claim 1, characterized in that, The constraint verification module is used to perform feasibility filtering on the candidate scheduling schemes to obtain feasible scheduling schemes, specifically including: The constraint verification module is used to obtain a preset set of hard constraint rules and a set of soft constraint rules; Each of the candidate scheduling schemes is subjected to hard constraint verification, which includes skill matching verification, equipment capability verification, time conflict verification, and process dependency verification. Candidate scheduling schemes that violate any hard constraint rule are marked as infeasible and filtered out; Perform soft constraint evaluation on candidate scheduling schemes that pass hard constraint verification and calculate the degree of soft constraint violation; The candidate scheduling schemes that pass the hard constraint verification are output as the feasible scheduling schemes, and the soft constraint violation degree is added to each feasible scheduling scheme.

5. The multi-agent causal constraint scheduling system according to claim 4, characterized in that, The causal effect assessment module is used to assess the intervention effect of the feasible scheduling scheme based on a structural causal model and output an effect score, specifically including: The causal effect evaluation module is used to construct a structural causal model for the scheduling scenario. The structural causal model includes endogenous variables, exogenous variables, and causal relationship functions. The endogenous variables include output per unit working hour, delay rate, overtime duration, and personnel workload dispersion. The feasible scheduling scheme is substituted into the structural causal model as an intervention variable to calculate the predicted values ​​of each endogenous variable after intervention. Based on the counterfactual reasoning framework, a control scenario is constructed in which no corresponding feasible scheduling scheme is implemented, and the baseline values ​​of each endogenous variable in the control scenario are calculated. The difference between the predicted value and the baseline value is calculated to obtain the average intervention effect of each endogenous variable; The average intervention effect of each endogenous variable is weighted and summed according to the preset weight coefficients, and the effect score of each feasible scheduling scheme is generated by combining the soft constraint violation degree of each feasible scheduling scheme.

6. The multi-agent causal constraint scheduling system according to claim 1, characterized in that, The scheduling scheme determination module is used to select the optimal scheduling scheme from the feasible scheduling schemes based on the effect score, and trigger a rescheduling after receiving a disturbance event, outputting a new optimal scheduling scheme within a preset time limit, specifically including: The scheduling scheme determination module is used to sort the effect scores and select the feasible scheduling scheme with the highest effect score as the optimal scheduling scheme. The optimal scheduling plan is output to the production execution system, and disturbance events during the production execution process are monitored in real time. The disturbance events include at least equipment failure, staff absence, order changes, and process delays. When the disturbance event is detected, determine the degree of impact of the disturbance event on the current scheduling plan; If the impact exceeds a preset threshold, a rescheduling process is triggered. In the rescheduling process, the personnel data, equipment data, production data, and constraint data are updated. The updated data is input into the feature construction module, and the state diagram construction, scheduling decision, constraint verification, and causal effect evaluation are re-executed. The rescheduling calculation is completed within a preset time limit, and a new optimal scheduling plan is output.

7. The multi-agent causal constraint scheduling system according to claim 1, characterized in that, The system also includes: The interpretation and auditing module is used to assign a unique scheme identifier to each generated optimal scheduling scheme and save the input snapshot corresponding to the optimal scheduling scheme. The input snapshot includes personnel data, equipment data, production data and constraint data at the time the current scheduling was triggered. Record the objective function value of the scheduling scheme, wherein the objective function value includes at least the effect score and the scores of each item in the calculation process; Save constraint hit details, which record the verification results of hard constraints and the violation of soft constraints; Record the manual approval actions related to the optimal scheduling plan, including the approver, approval time, approval comments, and approval results.

8. A multi-agent causal constraint scheduling method, characterized in that, The method includes: Acquire personnel data, equipment data, production data, and constraint data, and generate standardized feature vectors; A spatiotemporal heterogeneous graph is constructed based on the standardized feature vectors. The nodes of the spatiotemporal heterogeneous graph include employee nodes, process nodes, equipment nodes, and shift nodes. Candidate scheduling schemes are generated based on the spatiotemporal heterogeneous graph state. Perform feasibility filtering on the candidate scheduling schemes to obtain feasible scheduling schemes; The intervention effect of the feasible scheduling scheme is evaluated based on a structural causal model, and an effect score is output. Based on the effect score, the optimal scheduling scheme is selected from the feasible scheduling schemes, and a rescheduling is triggered after receiving a disturbance event, and a new optimal scheduling scheme is output within a preset time limit.

9. A computer program product containing instructions, characterized in that, When the computer program product is run on a multi-agent causal constraint scheduling system, the multi-agent causal constraint scheduling system performs the method as described in claim 8.

10. A computer-readable storage medium comprising instructions, characterized in that, When the instruction is executed on a multi-agent causal constraint scheduling system, the multi-agent causal constraint scheduling system performs the method as described in claim 8.

Images