Industrial intelligent decision support method oriented to multi-agent cooperation

Abstract

The application relates to an industrial intelligent decision support method facing multi-agent cooperation, and belongs to the technical field of industrial intelligent decision making. The method comprises the following steps: acquiring multi-source operation data and agent basic attribute information of an industrial site, carrying out pretreatment and feature extraction, constructing a unified feature library of multi-agent cooperative decision making and generating cooperative constraint conditions; based on the unified feature library and the constraint conditions, a multi-agent cooperative decision making topology is configured by adopting a hierarchical decision tree, and conflict pre-elimination is performed on a decision path; a state of a decision execution process is monitored, a cooperative decision state record is maintained, a deviation state is identified, constraint limitation and cooperative adjustment are performed; based on the state record and a real-time operation state, local cooperative optimization is performed on the decision topology, multi-agent cooperative decision making reasoning is carried out, a decision result is subjected to weighted fusion, and industrial decision support information is output. The application realizes conflict elimination and dynamic optimization of multi-agent cooperative decision making, and improves the robustness and real-time performance of industrial decision making.

Description

Technical Field

[0001] This invention belongs to the field of industrial intelligent decision-making technology, specifically relating to an industrial intelligent decision support method for multi-agent collaboration. Background Technology

[0002] With the rapid development of industrial manufacturing systems towards intelligence and flexibility, multi-agent collaborative decision-making technology is widely used in complex industrial scenarios such as production scheduling, equipment coordination, and resource allocation. Multi-agent systems effectively improve the system's distributed nature and scalability by decomposing complex tasks into multiple autonomous decision-making units. Each agent makes independent decisions based on local information and completes the global task through interactive collaboration. However, in dynamic and ever-changing industrial environments, multi-agent collaborative decision-making still faces many technical challenges.

[0003] First, existing decision-making methods typically rely on manually preset rules or single-feature models, lacking the ability to uniformly represent and standardize multi-source heterogeneous data from industrial sites. Operational data collected by different agents differ in format, units, and sampling frequency; direct fusion can easily lead to feature redundancy or information loss, affecting the consistency and reliability of decisions.

[0004] Secondly, resource and execution conflicts frequently occur during multi-agent collaborative decision-making. Because each agent makes independent decisions and lacks a global resource allocation coordination mechanism, resource contention and execution deadlocks easily arise when multiple agents simultaneously request the same device, channel, or tooling. Traditional methods often use sequential locks or priority queues for conflict handling, but these are ill-suited to dynamically arriving tasks and real-time changing resource states, leading to system response delays and decreased resource utilization.

[0005] Secondly, existing systems lack sufficient monitoring and anomaly handling capabilities during the decision-making process. When some agents exhibit execution deviations, timing errors, or resource overruns, the lack of effective state awareness and constraint mechanisms allows abnormal states to easily propagate among agents, affecting the stability of the global decision-making chain. Post-event manual intervention or global replanning methods suffer from poor real-time performance, failing to meet the demands of industrial settings for highly reliable, low-latency decision-making.

[0006] Furthermore, the topology of multi-agent collaborative decision-making often adopts a static configuration approach, which cannot be locally adjusted or dynamically optimized based on real-time operating status. When the task load changes or agent nodes fail, traditional methods need to rebuild the entire decision-making topology, resulting in high computational overhead, long recovery time, and difficulty in achieving fast adaptive replanning.

[0007] In summary, how to construct a unified collaborative decision-making feature library under multi-source heterogeneous data conditions, realize conflict pre-resolution and dynamic resource coordination in decision-making paths, and establish an adaptive collaborative decision-making framework with state monitoring, constraint limitation, and local optimization capabilities are the technical problems that need to be solved in the field of industrial intelligent decision-making. Summary of the Invention

[0008] To address the aforementioned problems in existing technologies, this invention provides an industrial intelligent decision support method oriented towards multi-agent collaboration. The objective of this invention can be achieved through the following technical solutions: S1: Acquire multi-source operational data and basic attribute information of multiple agents in the industrial field, perform preprocessing and feature extraction related to collaborative decision-making on the data, construct a unified feature library for multi-agent collaborative decision-making, and generate multi-agent collaborative constraints. S2: Based on a unified feature library and multi-agent collaborative constraints, a hierarchical decision tree is used to execute multi-agent collaborative decision configuration, construct a multi-agent collaborative decision topology, and perform conflict pre-resolution on the decision path; S3: Monitor the status of the multi-agent collaborative decision-making execution process, maintain collaborative decision-making status records, identify collaborative deviation statuses based on monitoring information, and implement constraints and collaborative adjustments for deviation statuses; S4: Based on the collaborative decision-making status record and real-time industrial operation status, perform local collaborative optimization on the collaborative decision-making topology, perform multi-Agent collaborative decision-making reasoning based on the optimized collaborative decision-making topology, perform weighted fusion on the decision results, and output industrial decision support information.

[0009] Specifically, the construction process of the unified feature library for multi-agent collaborative decision-making is as follows: Timestamp alignment and missing value imputation are performed on multi-source operational data from industrial sites to obtain synchronized time-series data. Category encoding and numerical normalization are performed on basic attribute information of multiple agents to generate standardized attribute vectors. The synchronous time-series data and normalized attribute vectors are vertically concatenated using the Agent identifier to form the original feature matrix; the original feature matrix is ​​then subjected to a standardization transformation to generate a standard feature matrix. A linear dimensionality reduction transformation is performed on the standard feature matrix to extract feature components whose cumulative variance contribution rate exceeds a preset threshold, thereby forming a unified feature library for multi-agent collaborative decision-making.

[0010] Specifically, performing the linear dimensionality reduction transformation includes: Calculate the covariance matrix of the standard feature matrix, perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors, arrange the eigenvectors in descending order of eigenvalues ​​to form a projection matrix; Perform matrix multiplication on the standard feature matrix and the projection matrix to obtain the dimension-reduced feature space; calculate the variance contribution rate of each dimension after dimension reduction to form the final feature components of the unified feature library.

[0011] Specifically, the configuration of multi-agent collaborative decision-making includes: Based on the feature components in the unified feature library and the multi-agent collaborative constraints, a starting node of the decision tree is created, and the global decision task is written into the node attributes. Based on the preset task decomposition rules, the global decision task is recursively divided into subtasks. A corresponding sub-node is created according to the division result of each sub-task, and the task parameters of the sub-task are written into the created sub-node. By comparing the Agent capability with the task matching degree, the execution Agent identifier of each sub-node is obtained, and the execution Agent identifier is stored in the attribute field of the sub-node. Extract the call relationships and data flow directions between each node, map the call relationships and data flow directions into directed edges, and construct a multi-Agent collaborative decision-making topology with nodes as vertices and directed edges as connections.

[0012] Specifically, the recursive division of the global decision-making task into sub-tasks includes: Obtain the current task node to be divided, and extract the task type parameter and resource requirement parameter from the task node; Based on the task type parameters, a preset task decomposition template is matched, and a subtask division scheme is obtained from the template; and based on the resource requirement parameters, the resource feasibility of each subtask is verified. Perform a partitioning operation on each subtask that passes the verification, and output the decision tree node sequence after task partitioning.

[0013] Specifically, the process of pre-resolving the execution conflict is as follows: Traverse each decision path in the multi-Agent collaborative decision topology and obtain the resource usage information corresponding to the nodes on each path; The resource occupancy information is grouped and sorted according to the resource identifier. By detecting the overlap of resource occupancy periods, conflict marking is performed on nodes with overlapping periods. Based on Agent priority and task urgency, the marked conflict nodes are reallocated to the idle time period of the corresponding resources. Based on the reallocated node sequence, an adjusted decision path set is formed.

[0014] Specifically, the process for detecting the overlap of resource occupancy periods is as follows: A resource-time occupancy matrix is ​​constructed using resource identifiers as row indexes and time units divided according to preset time granularity as column indexes. Traverse each decision path, obtain the resource occupation time period of each node in the decision path, find the row and continuous column interval corresponding to each occupation time period in the resource-time occupation matrix, and mark the resource occupation at the time unit in the continuous column interval; Obtain the number of resource occupancy markers at each time unit, and combine the resource identifiers and time unit positions corresponding to units where the number of resource occupancy markers exceeds a preset conflict threshold as overlapping time periods.

[0015] Specifically, the process of constraining and coordinating the deviation state is as follows: Deviation state parameters are extracted from the collaborative decision-making state record, and the deviation state parameters are matched with preset constraint limit rules to obtain the corresponding constraint execution strategy. According to the constraint execution strategy, constraint limitation processing is performed on the node constraint parameter set of the deviation-related nodes to generate constraint-limited node state data; Based on the constrained node state data, the execution logic and data interaction methods of related nodes within the decision topology are collaboratively modified to generate a decision execution sequence adapted to the current running state.

[0016] Specifically, the constraint limitation processing of the node constraint parameter set is performed as follows: Boundary conditions for various constraints are extracted from the constraint enforcement strategy, and the boundary conditions are grouped according to the constraint type to form a multi-dimensional constraint decision condition set. The deviation state parameters are compared item by item with each constraint boundary in the multi-dimensional constraint judgment condition set to obtain the parameter types and deviations that exceed the constraint boundaries. Based on the parameter type and deviation amount, the node constraint parameter set of the deviation-related node is corrected, and the node running status data is updated synchronously.

[0017] Specifically, the process of performing local collaborative optimization is as follows: Extract the node running status information corresponding to the collaborative decision-making topology from the collaborative decision-making status record, and perform comparison processing on the running status information based on a preset threshold to obtain the set of nodes to be optimized. The nodes in the set to be optimized are processed to adjust their execution order, nodes whose dependencies meet the parallel conditions are processed to be arranged in parallel, and nodes whose load information meets the migration conditions are processed to be transferred. The structural information of the collaborative decision-making topology is updated to form a collaborative decision-making topology after local collaborative optimization.

[0018] Specifically, the process of performing multi-agent collaborative decision-making reasoning is as follows: The initial nodes are extracted from the locally optimized collaborative decision-making topology, and the Agent decision-making modules corresponding to each node are started sequentially based on the order of the topology nodes. Based on the input data requirements of the current node, the output data of the upstream node is obtained, the upstream output data is input into the Agent decision module and the calculation is performed to generate the decision output data of the current node. The decision output data of the current node is transmitted to the downstream node. The nodes are started, data is acquired and processed in sequence according to the topological node order to form a complete sequence of decision reasoning results.

[0019] Specifically, the weighted fusion process is as follows: Obtain the decision results and corresponding confidence weights output by each Agent decision module. The confidence weights are generated based on the Agent's historical decision accuracy and real-time load information. Each decision result is aligned according to the decision variable dimension, and the aligned decision results are weighted based on the confidence weight to obtain the fused decision value. The fused decision values ​​are processed as industrial decision support information and simultaneously written into the collaborative decision status record.

[0020] The beneficial effects of this invention are as follows: (1) This invention achieves pre-resolution of conflicts and refined coordination of resource usage in decision-making paths by constructing a unified feature library and hierarchical decision tree for multi-agent collaborative decision-making. Specifically, this invention performs standardization processing and linear dimensionality reduction transformation on multi-source operational data and basic attribute information of agents in industrial sites, extracts key feature components to form a unified feature library, eliminates the dimensional differences and redundant information of heterogeneous data, and provides a unified data foundation for subsequent decision reasoning. On this basis, a hierarchical decision tree is used to configure the multi-agent collaborative decision-making topology, and the overlap of resource usage periods is detected by the resource-time usage matrix. Conflicting nodes are reallocated to idle periods or spare resources, effectively avoiding resource competition and timing conflicts among multiple agents during execution. This technical solution significantly improves the conflict resolution capability and resource utilization rate of multi-agent collaborative decision-making, and is suitable for real-time collaborative decision-making scenarios in complex industrial environments.

[0021] (2) This invention achieves adaptive adjustment and dynamic optimization of multi-Agent collaborative decision-making through collaborative decision-making status monitoring, constraint limitation, and local collaborative optimization. Specifically, this invention maintains collaborative decision-making status records during the decision execution process, identifies deviation status parameters in real time, obtains constraint execution strategies by matching preset constraint limitation rules, and performs constraint limitation processing on the operation control parameters, timing execution parameters, and resource consumption parameters of deviation-related nodes to form constrained node status data. Based on this, local collaborative optimization is performed on the collaborative decision-making topology based on the status records and real-time industrial operation status. The topology structure is updated through operations such as node reordering, parallel arrangement, and task transfer. The decision results output by each Agent decision-making module are weighted and fused to output industrial decision support information. This technical solution realizes adaptive reasoning and continuous optimization of multi-Agent collaborative decision-making, improving the robustness and real-time response capability of the industrial decision-making system. Attached Figure Description

[0022] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0023] Figure 1 This is a flowchart illustrating the industrial intelligent decision support method for multi-agent collaboration according to the present invention. Figure 2 This is an architecture diagram of the industrial intelligent decision support method for multi-agent collaboration in this invention. Detailed Implementation

[0024] To further illustrate the technical means and effects of the present invention in achieving the intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0025] Please see Figure 1-2 Industrial intelligent decision support methods oriented towards multi-agent collaboration include: S1: Acquire multi-source operational data and basic attribute information of multiple agents in the industrial field, perform preprocessing and feature extraction related to collaborative decision-making on the data, construct a unified feature library for multi-agent collaborative decision-making, and generate multi-agent collaborative constraints. S2: Based on a unified feature library and multi-agent collaborative constraints, a hierarchical decision tree is used to execute multi-agent collaborative decision configuration, construct a multi-agent collaborative decision topology, and perform conflict pre-resolution on the decision path; S3: Monitor the status of the multi-agent collaborative decision-making execution process, maintain collaborative decision-making status records, identify collaborative deviation statuses based on monitoring information, and implement constraints and collaborative adjustments for deviation statuses; S4: Based on the collaborative decision-making status record and real-time industrial operation status, perform local collaborative optimization on the collaborative decision-making topology, perform multi-Agent collaborative decision-making reasoning based on the optimized collaborative decision-making topology, perform weighted fusion on the decision results, and output industrial decision support information.

[0026] Specifically, the construction process of the unified feature library for multi-agent collaborative decision-making is as follows: Perform timestamp alignment and missing value imputation on multi-source operational data from industrial sites to generate synchronized time-series data; perform category encoding and numerical normalization on basic attribute information of multiple agents to generate normalized attribute vectors; The synchronous time-series data and the normalized attribute vector are vertically concatenated according to the Agent identifier to generate the original feature matrix; the original feature matrix is ​​then subjected to a standardization transformation to eliminate the dimensional differences of each feature dimension, generating a standard feature matrix, wherein the Agent identifier is a unique code or sequence number used to distinguish different Agents; A linear dimensionality reduction transformation is performed on the standard feature matrix to extract feature components whose cumulative variance contribution rate exceeds a preset threshold. The set of extracted feature components constitutes a unified feature library for multi-Agent collaborative decision-making.

[0027] Specifically, performing the linear dimensionality reduction transformation includes: Calculate the covariance matrix of the standard feature matrix, where each element of the covariance matrix is ​​an estimate of the covariance between any two feature dimensions; Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors. Arrange the eigenvectors in descending order of eigenvalues ​​to form a projection matrix. Multiply the standard feature matrix with the projection matrix to obtain the dimension-reduced feature space; calculate the variance contribution rate of each dimension after dimension reduction, and select the top few dimensions whose cumulative value reaches a preset threshold according to the contribution rate from high to low to form the final feature components of the unified feature library.

[0028] This embodiment uses a multi-AGV collaborative handling task in a smart factory as an application scenario. Multi-source operational data from the industrial site includes the position coordinates of each AGV, battery level, task queue length, communication signal strength, etc., with a sampling period of Δt (unit: seconds, preset to 0.1 seconds based on sensor performance). Basic attribute information for multiple agents includes the AGV's number, maximum load capacity, speed level, etc.

[0029] I. Data Preprocessing and Feature Matrix Construction Timestamp alignment is performed on multi-source operational data from industrial sites: using the system master clock as a reference, a linear interpolation method is employed to map data from different sampling times to a unified time axis, generating synchronized time-series data. For missing data points, the weighted average of data from two consecutive sampling times is used for interpolation, with the weighting coefficient inversely proportional to the time interval.

[0030] Perform category encoding and numerical normalization on the basic attribute information of multiple agents: convert the AGV number into a one-hot encoded vector, scale the numerical attributes such as maximum load and speed level to the [0,1] range, and scale them based on the minimum and maximum values ​​of each attribute in all AGVs to generate a normalized attribute vector.

[0031] Synchronous time-series data and normalized attribute vectors are vertically concatenated based on Agent identifiers (i.e., unique codes or serial numbers used to distinguish different AGVs). Assume there are N AGVs (N is the preset total number of AGVs, determined based on the actual factory deployment). Each AGV collects synchronous time-series data for T time windows (T is preset based on the task execution cycle, ranging from 10 to 100). Each time window contains F feature dimensions (F is determined by the number of sensor types). The T×F dimensional time-series data of each AGV is concatenated with the normalized attribute vector to form the original feature vector. The original feature vectors of all AGVs are stacked row-wise to construct the original feature matrix.

[0032] II. Standardization Transformation Perform a standardization transformation on the original feature matrix to eliminate the dimensional differences between the feature dimensions. Calculate the mean and standard deviation of each feature dimension, subtract the mean of that dimension from each feature value, and then divide by the standard deviation to make the mean of each feature dimension equal to 0 and the standard deviation equal to 1 after the transformation, thus generating a standard feature matrix.

[0033] III. Linear Dimensionality Reduction Transformation Principal component analysis (PCA) is performed on the standard eigenvalue matrix to achieve linear dimensionality reduction. First, the covariance matrix of the standard eigenvalue matrix is ​​calculated; its elements describe the degree of linear correlation between any two eigenvalue dimensions. Second, eigenvalue decomposition is performed on the covariance matrix to solve for all eigenvalues ​​and their corresponding eigenvectors. The eigenvalues ​​are then sorted in descending order, and their corresponding eigenvectors are arranged in the same order to form the projection matrix.

[0034] Then, the standard feature matrix is ​​multiplied by the first few columns of the projection matrix to obtain the dimensionality-reduced feature space. The variance contribution rate of each principal component (i.e., the proportion of the eigenvalue corresponding to that principal component to the sum of all eigenvalues) is calculated, and the variance contribution rates are accumulated from front to back. A preset cumulative variance contribution rate threshold θ is set (e.g., θ=0.95, indicating that more than 95% of the original information is retained; this threshold can be adjusted according to the trade-off between information retention and computational efficiency in the actual scenario, and is generally in the range of 0.85 to 0.99). The minimum number of principal components k is selected so that the cumulative variance contribution rate reaches θ.

[0035] Finally, the projected values ​​corresponding to the extracted k principal components constitute a unified feature library for multi-agent collaborative decision-making. This feature library retains the main information of the original data while significantly reducing the feature dimensionality, providing a standardized data foundation for subsequent decision-making inference.

[0036] Specifically, the configuration of multi-agent collaborative decision-making includes: Based on the feature components in the unified feature library and the multi-agent collaborative constraints, a starting node of the decision tree is created, and the global decision task is written as the task content of the starting node into the node attribute. The global decision task is recursively divided into subtasks according to the task decomposition rules. A corresponding sub-node is created based on the division result of each sub-task, and the task parameters of the sub-task are written into the created sub-node. By comparing the Agent capability with the task matching degree, the execution Agent identifier of each sub-node is obtained, and the execution Agent identifier is stored in the attribute field of the sub-node. Extract the call relationships and data flow directions between each node, map the call relationships and data flow directions into directed edges, and construct a multi-Agent collaborative decision-making topology with nodes as vertices and directed edges as connections.

[0037] Specifically, the recursive division of the global decision-making task into subtasks according to the task decomposition rules includes: Obtain the current task node to be divided, and extract the task type parameters and resource requirement parameters contained in the task node; Match a preset task decomposition template based on the task type parameter, and obtain a subtask division scheme from the template; verify the resource feasibility of each subtask based on the resource requirement parameter, and eliminate subtask division branches that are not resource-feasible. Repeat the above partitioning operation for each subtask that passes the verification until the task granularity of all subtask nodes is lower than the preset threshold, and output the decision tree node sequence after task partitioning.

[0038] Specifically, the process of pre-resolving the execution conflict is as follows: Traverse each decision path in the multi-Agent collaborative decision topology and obtain the resource usage information corresponding to the nodes on each path; The resource occupancy information is grouped and sorted according to the resource identifier. By detecting the overlap of resource occupancy periods, conflict marking is performed on nodes with overlapping periods. Based on Agent priority and task urgency, the marked conflict nodes are reallocated to the idle time period of the corresponding resources. The relocated node sequence constitutes the adjusted decision path set.

[0039] Specifically, the process for detecting the overlap of resource occupancy periods is as follows: A resource-time occupancy matrix is ​​constructed using resource identifiers as row indexes and time units divided according to preset time granularity as column indexes. Traverse each decision path, obtain the resource occupation time period of each node in the decision path, find the row and continuous column interval corresponding to each occupation time period in the resource-time occupation matrix, and mark the resource occupation at each time unit in the continuous column interval; Traverse each time unit in the resource-time occupancy matrix, count the number of resource occupancy markers at each time unit, and combine the resource identifiers and time unit positions corresponding to units where the number of resource occupancy markers exceeds a preset conflict threshold as overlapping time periods.

[0040] This embodiment uses a multi-AGV collaborative handling task in a smart factory as an application scenario. In this scenario, there are 3 AGVs (numbered AGV1, AGV2, and AGV3) that need to complete the task of moving a batch of goods from the warehouse to the production line. The global decision task is "move all goods from warehouse A to production line B".

[0041] I. Multi-Agent Collaborative Decision Configuration (Decision Tree Construction) The system creates the starting node of the decision tree based on the feature components in the unified feature library (including the load capacity, battery power, and position coordinates of each AGV) and multi-Agent collaborative constraints (such as the AGV load not exceeding 100kg and only one AGV being allowed to pass through the same channel at the same time). The global decision task "transport all goods" is written into the node attribute as the task content of the starting node.

[0042] According to the preset task decomposition rules, the global decision-making task is recursively divided into subtasks. The task decomposition rules are as follows: first, divide by batch of goods, with each batch being a subtask; then divide by transportation route segments, with each route segment being a subtask. For example, the global task is divided into three batch subtasks: Batch 1 (5 boxes of goods), Batch 2 (3 boxes of goods), and Batch 3 (4 boxes of goods). Each batch subtask is further divided into route segment subtasks: from warehouse A to transfer station C, and from transfer station C to production line B.

[0043] Based on the division of each subtask, create corresponding sub-nodes. Write the subtask parameters (such as cargo weight, start coordinates, end coordinates, and estimated time) into the attribute fields of the created sub-nodes. Obtain the execution agent identifier for each sub-node by comparing the agent's capabilities with the task's matching degree. The matching degree is calculated by comprehensively scoring the AGV's current load, remaining battery power, and distance from the starting point; the AGV with the highest score is selected as the execution agent. Store the execution agent identifier in the sub-node's attribute fields.

[0044] Extract the call relationships and data flow between nodes: For example, after the "Warehouse A → Transfer Station C" node in batch 1 completes its execution, its output data (such as arrival time and remaining power) needs to be passed to the "Transfer Station C → Production Line B" node in the same batch as input. Convert these call relationships and data flow into directed edges, and construct a multi-Agent collaborative decision-making topology with nodes as vertices and directed edges as connections.

[0045] II. Specific Operations of Task Recursive Division Obtain the current task node to be divided (starting node), and extract the task type parameters (e.g., "transportation") and resource requirement parameters (e.g., "requires an AGV with a load capacity ≥ 50kg, estimated channel occupancy time 2 minutes"). Match the task type parameters to a pre-defined task decomposition template library, which stores standard decomposition schemes for different task types. For example, the decomposition template for the "transportation" task is: divide into batches based on the quantity of goods, with each batch not exceeding 5 boxes; each batch is divided into road segments based on path nodes. After obtaining the sub-task division scheme from the template, verify the resource feasibility of each sub-task based on the resource requirement parameters: check if there are AGVs that meet the load capacity requirements, and whether the channel is idle during the estimated time period. If a sub-task cannot meet the resource requirements (e.g., no idle AGVs), then remove that sub-task's division branch.

[0046] Repeat the above partitioning operation for each subtask that passes the verification until the task granularity of all subtask nodes is lower than a preset threshold (e.g., the estimated execution time of each subtask is less than 30 seconds). The final output is a sequence of decision tree nodes after task partitioning, containing all leaf nodes (i.e., indivisible atomic tasks).

[0047] III. Conflict Pre-resolution (Resource Occupation Detection and Reallocation) Traverse each decision path in the multi-agent collaborative decision-making topology to obtain resource occupancy information for each node on each path. Resource occupancy information includes: resource identifier (e.g., channel CH1, AGV1), and occupancy time period (start time t). start End time t end ), Duration Δt.

[0048] Resource occupancy information is grouped and sorted according to resource identifiers. For example, all nodes involving channel CH1 are sorted according to the start time of their occupancy period. Then, the overlap of resource occupancy periods is detected: if the occupancy periods of two nodes intersect on the timeline (i.e., the end time of one node is later than the start time of another node), it is determined to be an overlap.

[0049] Conflict marking is performed on nodes that overlap. For example, if node A (AGV1 occupies channel CH1, time period [10:00, 10:05]) and node B (AGV2 occupies channel CH1, time period [10:03, 10:08]) overlap, node A and node B are marked as conflicting nodes.

[0050] The marked conflicting nodes are reassigned based on Agent priority and task urgency. Agent priority is preset as follows: AGV1 > AGV2 > AGV3; task urgency is set according to the cargo delivery deadline. Nodes with lower priority or urgency are moved to the idle time slots of their corresponding resources (e.g., adjusting node B's occupied time slot to [10:05, 10:10]), or switched to backup resources (e.g., changing AGV2's channel CH1 occupation to channel CH2). The reassigned node sequence constitutes the adjusted decision path set, ensuring that the same resource is occupied by only one Agent at a time.

[0051] IV. Specific Implementation of Resource Occupation Period Overlap Detection A resource-time occupancy matrix is ​​constructed, using resource identifiers as row indexes and time units divided according to a preset time granularity as column indexes. The preset time granularity is set to 1 minute, and the time units cover the entire task execution cycle (e.g., from 10:00 to 11:00, a total of 60 time units). The rows of the matrix correspond to resource identifiers (such as CH1, CH2, AGV1, etc.), and the columns correspond to time unit numbers.

[0052] Traverse each decision path and obtain the time periods during which each node occupies the resource. For example, if node A occupies resource CH1 during the time period [10:00, 10:05], convert it into a continuous column interval (column 0 to 4) of the corresponding row (CH1) in the matrix. Mark the resource occupancy at each time unit within this continuous column interval: accumulate the occupancy count in the matrix cell (i.e., how many nodes occupy the resource within that time unit).

[0053] Iterate through each time unit in the matrix and count the number of resource occupancy markers at each time unit. Extract the resource identifier and time unit position corresponding to the unit whose number of markers exceeds a preset conflict threshold (the threshold is preset to 1, because the same resource can only be occupied by one Agent at a time), and use them as overlapping time period combinations. For example, if the number of markers at the time unit in the 3rd column of row CH1 is 2, it means that there is a conflict in this time unit, and the overlapping time period combination is output as (CH1, time unit 3).

[0054] Through the above operations, the system can automatically detect and resolve resource conflicts in multi-agent collaborative decision-making, ensuring that the generated set of decision paths is conflict-free and executable. In this embodiment, all preset thresholds (such as a task granularity threshold of 30 seconds, a time granularity of 1 minute, and a conflict threshold of 1) are pre-calibrated based on device performance and task requirements, and can be adjusted according to the actual scenario.

[0055] Specifically, the process of constraining and coordinating the deviation state is as follows: Deviation state parameters are extracted from the collaborative decision-making state record, and the deviation state parameters are matched with preset constraint limit rules to obtain the corresponding constraint execution strategy. According to the constraint execution strategy, constraint limitation processing is performed on the node constraint parameter set of the deviation-related nodes to generate constraint-limited node state data; the node constraint parameter set includes node operation control parameters, node timing execution parameters, and node resource usage parameters; Based on the constrained node state data, the execution logic and data interaction methods of related nodes within the decision topology are collaboratively modified to generate a decision execution sequence adapted to the current running state.

[0056] Specifically, the constraint limitation processing of the node constraint parameter set is performed as follows: Boundary conditions for various constraints are extracted from the constraint enforcement strategy, and the boundary conditions are grouped according to the constraint type to form a multi-dimensional constraint decision condition set. The deviation state parameters are compared item by item with each constraint boundary in the multi-dimensional constraint judgment condition set to obtain the parameter types and deviations that exceed the constraint boundaries. Based on the parameter type and deviation amount, the node constraint parameter set of the deviation-related node is corrected, and the node running status data is updated synchronously.

[0057] This embodiment uses a multi-AGV collaborative transport task in a smart factory as an application scenario. Assume the current collaborative decision-making topology includes three AGV nodes: AGV1 (transporting goods from warehouse A to transfer station C), AGV2 (from transfer station C to production line B), and AGV3 (from warehouse A to production line B). The collaborative decision-making status record maintained by the system stores the real-time execution status of each node, including actual start time, actual end time, actual load, and current battery level.

[0058] I. Deviation State Identification and Constraint Strategy Matching During task execution, the system extracts deviation status parameters from the collaborative decision-making status record. For example, if the actual handling time of AGV1 node exceeds the preset time by 30 seconds, and the actual power consumption exceeds the preset power consumption by 15%, these parameters are deviation status parameters. The system matches these deviation status parameters with a preset constraint rule base. The rule base pre-stores constraint execution strategies corresponding to various deviation types, such as: If the "execution time exceeds the limit", the strategy is to "compress the execution cycle of subsequent nodes". If "power consumption exceeds the limit", the strategy is to "reduce the operating speed and increase the number of charging nodes"; If "resource usage exceeds limit", the strategy is to "reduce resource quota".

[0059] Through matching, the system obtains the corresponding constraint execution strategy. In this embodiment, AGV1 experiences "time exceeding limit" and "power consumption exceeding limit", and the obtained strategies are "compress the execution cycle of subsequent nodes" and "reduce the running speed".

[0060] II. Constraint Limitation Processing of Node Constraint Parameter Sets The system performs constraint limiting processing on the node constraint parameter set of the deviation-related nodes (i.e., AGV1 and its downstream nodes) according to the constraint execution strategy. The node constraint parameter set includes node operation control parameters (such as speed limit, acceleration), node timing execution parameters (such as planned start time, planned end time), and node resource occupancy parameters (such as occupied channel time period, occupied charging pile quota).

[0061] The specific constraints and limitations on the operation are as follows: Regarding the strategy of "compressing the execution cycle of subsequent nodes": advance the planned start time of AGV2 (downstream node) and shorten the planned end time of AGV2 by the compression ratio (the compression ratio is calculated based on the deviation range, the deviation range is 30 seconds, and the compression coefficient is preset to 0.8).

[0062] Regarding the "reduce operating speed" strategy: the speed limit of AGV1 is increased from the default value V. max Adjusted to V max ×0.7 (The speed adjustment coefficient is determined based on the magnitude of the power deviation; a deviation of 15% corresponds to a coefficient of 0.7).

[0063] After performing the above operations, constrained node status data is generated, including the deceleration operation parameters of AGV1 and the compression time parameters of AGV2.

[0064] III. Multi-dimensional constraint determination and item-by-item comparison The specific process of constraint handling is further refined: the system extracts the boundary conditions of various constraints from the constraint enforcement strategy. For example, for the "execution cycle compression" constraint, the boundary condition is "the node execution time must not exceed the preset maximum duration T". max (T) max =120 seconds); for the "speed reduction" constraint, the boundary condition is "the speed must not be lower than the safe speed limit V". min (V) min =0.5m / s). Boundary conditions are grouped according to constraint type (time series, resource, control) to form a multi-dimensional constraint decision condition set.

[0065] The deviation state parameters (actual time 130 seconds, actual speed 1.2 m / s) were compared item by item with each constraint boundary in the multi-dimensional constraint judgment condition set: The actual time taken was 130 seconds, exceeding the boundary T. max =120 seconds, exceeding the limit by 10 seconds; The actual speed is 1.2 m / s, which is higher than V. min =0.5m / s, not exceeding the limit.

[0066] Get the parameter type ("execution duration") and deviation (10 seconds) that exceeds the constraint boundary.

[0067] IV. Modification and State Update of Node Constraint Parameter Set Based on the parameter type and deviation amount, perform correction processing on the node constraint parameter set of the deviation-related nodes: In the subsequent tasks of AGV1, reduce the planned execution time of each subtask by 10 seconds (achieved by increasing speed or reducing pauses). Advance the planned start time of AGV2 by 10 seconds and the end time of AGV2 by 8 seconds (considering a buffer factor of 0.8).

[0068] Synchronously update node running status data: update the current speed parameter of AGV1 to 1.2m / s, and update the planned time window of AGV2 to [original start time - 10 seconds, original end time - 8 seconds]. Simultaneously, write the updated status data back to the collaborative decision-making status record.

[0069] V. Collaborative Correction Processing Generates Decision Execution Sequence Based on the constrained node state data, a collaborative correction process is performed on the execution logic and data interaction methods of associated nodes within the decision topology: Execution logic correction: After AGV1 slows down, its arrival time at transfer station C will be delayed. Therefore, AGV2 needs to be notified to delay its departure to avoid waiting idly. The system changes the trigger condition for AGV2 from "AGV1 arrives" to "AGV1 arrives and starts after a 5-second delay" to match the new time window.

[0070] Data interaction method correction: After AGV1 completes the transport, it needs to send data such as remaining power and actual time consumed to AGV2 so that AGV2 can adjust its operating strategy. The system adds "power alarm flag" and "time deviation value" fields to the output data of AGV1.

[0071] After the above corrections, the system generates a decision execution sequence adapted to the current operating state: AGV1 slows down → sends status data after arriving at transfer station C → AGV2 starts after a 5-second delay → AGV2 executes according to the compressed time window → transport completed. This decision execution sequence can automatically adapt to deviations, ensuring the reliability and timeliness of the overall task completion.

[0072] In this embodiment, all preset thresholds (such as T) max =120 seconds, V min The parameters (such as speed coefficient 0.5 m / s, compression coefficient 0.8, and speed adjustment coefficient 0.7) are all pre-calibrated based on equipment technical parameters and process requirements, and can be adjusted according to the actual scenario. Through the above steps, the system realizes automatic detection, constraint limitation, and collaborative adjustment of deviation states, effectively improving the robustness of multi-agent collaborative decision-making.

[0073] Specifically, the process of performing local collaborative optimization is as follows: Extract node running status information corresponding to the collaborative decision-making topology from the collaborative decision-making status record, perform comparison processing on the node running status information based on a preset threshold, and obtain a set of nodes to be optimized; the node running status information includes node execution efficiency information and node resource usage information. The nodes in the set to be optimized are processed to adjust their execution order, nodes whose dependencies meet the parallel conditions are processed to be arranged in parallel, and nodes whose load information meets the migration conditions are processed to be transferred. The structural information of the collaborative decision-making topology is updated to form a collaborative decision-making topology after local collaborative optimization.

[0074] Specifically, the process of performing multi-agent collaborative decision-making reasoning is as follows: The initial nodes are extracted from the locally optimized collaborative decision-making topology, and the Agent decision-making modules corresponding to each node are started sequentially based on the order of the topology nodes. Based on the input data requirements of the current node, the output data of the upstream node is obtained, the upstream output data is input into the Agent decision module and the calculation is performed to generate the decision output data of the current node. The decision output data of the current node is transmitted to the downstream node. The startup, data acquisition and calculation processing of each node are completed in sequence according to the topological node order, forming a complete sequence of decision reasoning results.

[0075] Specifically, the weighted fusion process is as follows: Collect the decision results and corresponding confidence weights output by each Agent decision module. The confidence weights are generated based on the Agent's historical decision accuracy and real-time load information. Each decision result is aligned according to the decision variable dimension, and the aligned decision results are weighted based on the confidence weight to obtain the fused decision value. The fused decision data is processed as industrial decision support information and simultaneously stored in the collaborative decision status record.

[0076] This embodiment uses a multi-AGV collaborative transport task in a smart factory as an application scenario. The initial collaborative decision-making topology includes multiple AGV nodes, and the task is to transport a batch of goods from warehouse A to production line B. During execution, the collaborative decision-making status record stores the operating status information of each node.

[0077] I. Local Cooperative Optimization The system extracts node operation status information corresponding to the collaborative decision-making topology from the collaborative decision-making status record, including the execution efficiency information of each node (such as actual handling time and waiting time) and resource consumption information (such as channel occupancy rate and AGV load rate). The actual execution efficiency of each node is compared with a preset threshold, and nodes exceeding the threshold are added to the set of nodes to be optimized. Simultaneously, resource consumption is checked, and nodes with resource consumption exceeding the threshold, along with their associated nodes, are also added to the set of nodes to be optimized.

[0078] The process for adjusting the execution order of nodes in the set to be optimized involves: analyzing the dependencies between nodes, identifying nodes with no dependencies and non-overlapping paths, adjusting their original sequential execution order to parallel execution, and aligning their start times. Simultaneously, load information is monitored, and some tasks from nodes with loads exceeding a preset migration threshold are transferred to idle nodes with loads below the threshold to balance the load.

[0079] The structural information of the collaborative decision-making topology is updated: the start time of relevant nodes is modified, new child nodes are added, and migrated task nodes are deleted. In the updated topology, parallel nodes run simultaneously, the load is balanced, and a collaborative decision-making topology optimized locally is formed.

[0080] II. Multi-Agent Collaborative Decision-Making Reasoning Initial nodes are extracted from the locally optimized collaborative decision-making topology. Based on the order of the topology nodes (the sequence sorted by dependencies), the Agent decision-making modules corresponding to each node are started sequentially.

[0081] Taking any Agent node as an example: the current node's input data requirement is the output data of the upstream node. The system obtains the output data from the upstream node and inputs it into the current Agent's decision module. The Agent's decision module performs calculations: based on the current state, task objectives, and constraints, it plans an action plan and generates decision output data (such as path sequences and estimated completion times). The current node's decision output data is then transmitted to downstream nodes as their input.

[0082] The startup, data acquisition, and computation processing of each node are completed sequentially according to the topological node order, forming a complete sequence of decision reasoning results.

[0083] III. Execution of Weighted Fusion Collect the decision results and corresponding confidence weights output by each Agent's decision-making module. Assume there are n Agents participating in the decision-making process, where the decision result of the i-th Agent (i=1,2,...,n) is r. i (For example, the estimated time to reach the target location), the corresponding confidence weight is w. i .

[0084] weight w i The calculation formula is generated based on the agent's historical decision accuracy and real-time load information: w i =α·a i +β·(1-l i ), in: a iLet be the historical decision accuracy of the i-th Agent, with a value ranging from 0 to 1, which is obtained from the statistics of historical execution records; l i The current load rate of the i-th Agent is 0 to 1, and is obtained from real-time monitoring. α and β are preset coefficients that satisfy α+β=1. In this embodiment, α=0.6 and β=0.4, indicating that the weight is composed of the weighted sum of historical accuracy and idle rate (i.e., 1 minus load rate).

[0085] Align the decision results according to the dimensions of the decision variables, so that all r i Corresponding to the same physical quantity (such as time, distance, etc.).

[0086] The fusion decision value R is calculated using the weighted average formula. fused The formula is as follows: , in: This represents the summation of the products of the decision results of all agents and their corresponding weights; This represents the summation of the weights over all agents; The formula calculates a weighted arithmetic mean proportional to the confidence level, so that decisions with high confidence levels have a higher weight in the fusion result.

[0087] The calculated fusion decision value R fused The output is processed as industrial decision support information. Simultaneously, the fused decision values ​​are stored in the collaborative decision status record for subsequent periodic optimization or historical analysis.

[0088] Through the above steps, the system achieves local optimization, distributed inference, and weighted fusion of results in multi-agent collaborative decision-making. All preset thresholds (such as execution efficiency threshold, resource consumption threshold, load migration threshold, and coefficients α and β) are pre-calibrated based on historical data and process requirements, and can be dynamically adjusted according to the actual scenario.

[0089] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An industrial intelligent decision support method for multi-agent collaboration, characterized in that, include: S1: Acquire multi-source operational data and basic attribute information of multiple agents in the industrial field, perform preprocessing and feature extraction related to collaborative decision-making on the data, construct a unified feature library for multi-agent collaborative decision-making, and generate multi-agent collaborative constraints. S2: Based on a unified feature library and multi-agent collaborative constraints, a hierarchical decision tree is used to execute multi-agent collaborative decision configuration, construct a multi-agent collaborative decision topology, and perform conflict pre-resolution on the decision path; S3: Monitor the status of the multi-agent collaborative decision-making execution process, maintain collaborative decision-making status records, identify collaborative deviation statuses based on monitoring information, and implement constraints and collaborative adjustments for deviation statuses; S4: Based on the collaborative decision-making status record and real-time industrial operation status, perform local collaborative optimization on the collaborative decision-making topology, perform multi-Agent collaborative decision-making reasoning based on the optimized collaborative decision-making topology, perform weighted fusion on the decision results, and output industrial decision support information.

2. The method according to claim 1, characterized in that, The construction process of the unified feature library for multi-agent collaborative decision-making is as follows: Timestamp alignment and missing value imputation are performed on multi-source operational data from industrial sites to obtain synchronized time-series data. Category encoding and numerical normalization are performed on basic attribute information of multiple agents to generate standardized attribute vectors. The synchronous time-series data and normalized attribute vectors are vertically concatenated using the Agent identifier to form an original feature matrix; a standardization transformation is then performed on the original feature matrix to generate a standard feature matrix. A linear dimensionality reduction transformation is performed on the standard feature matrix to extract feature components whose cumulative variance contribution rate exceeds a preset threshold, thereby forming a unified feature library for multi-agent collaborative decision-making.

3. The method according to claim 2, characterized in that, The linear dimensionality reduction transformation specifically includes: Calculate the covariance matrix of the standard feature matrix, perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors, arrange the eigenvectors in descending order of eigenvalues ​​to form a projection matrix; Perform matrix multiplication on the standard feature matrix and the projection matrix to obtain the dimension-reduced feature space; calculate the variance contribution rate of each dimension after dimension reduction to form the final feature components of the unified feature library.

4. The method according to claim 1, characterized in that, The configuration for multi-agent collaborative decision-making specifically includes: Based on the feature components in the unified feature library and the multi-agent collaborative constraints, a starting node of the decision tree is created, and the global decision task is written into the node attributes. Based on the preset task decomposition rules, the global decision task is recursively divided into subtasks. A corresponding sub-node is created according to the division result of each sub-task, and the task parameters of the sub-task are written into the created sub-node. By comparing the Agent capability with the task matching degree, the execution Agent identifier of each sub-node is obtained, and the execution Agent identifier is stored in the attribute field of the sub-node. Extract the call relationships and data flow directions between each node, map the call relationships and data flow directions into directed edges, and construct a multi-Agent collaborative decision-making topology with nodes as vertices and directed edges as connections.

5. The method according to claim 4, characterized in that, The recursive division of the global decision-making task into sub-tasks specifically includes: Obtain the current task node to be divided, and extract the task type parameter and resource requirement parameter from the task node; Based on the task type parameters, a preset task decomposition template is matched, and a subtask division scheme is obtained from the template; and based on the resource requirement parameters, the resource feasibility of each subtask is verified. Perform a partitioning operation on each subtask that passes the verification, and output the decision tree node sequence after task partitioning.

6. The method according to claim 1, characterized in that, The specific process for the pre-resolution of execution conflicts is as follows: Traverse each decision path in the multi-Agent collaborative decision topology and obtain the resource usage information corresponding to the nodes on each path; The resource occupancy information is grouped and sorted according to the resource identifier. By detecting the overlap of resource occupancy periods, conflict marking is performed on nodes with overlapping periods. Based on Agent priority and task urgency, the marked conflict nodes are reallocated to the idle time period of the corresponding resources. Based on the reallocated node sequence, an adjusted decision path set is formed.

7. The method according to claim 6, characterized in that, The process for detecting overlapping time periods of resource occupancy is as follows: A resource-time occupancy matrix is ​​constructed using resource identifiers as row indexes and time units divided according to preset time granularity as column indexes. Traverse each decision path, obtain the resource occupation time period of each node in the decision path, find the row and continuous column interval corresponding to each occupation time period in the resource-time occupation matrix, and mark the resource occupation at the time unit in the continuous column interval; Obtain the number of resource occupancy markers at each time unit, and combine the resource identifiers and time unit positions corresponding to units where the number of resource occupancy markers exceeds a preset conflict threshold as overlapping time periods.

8. The method according to claim 1, characterized in that, The specific process of execution constraint limitation and coordinated adjustment is as follows: Deviation state parameters are extracted from the collaborative decision-making state record, and the deviation state parameters are matched with preset constraint limit rules to obtain the corresponding constraint execution strategy. According to the constraint execution strategy, constraint limitation processing is performed on the node constraint parameter set of the deviation-related nodes to generate constraint-limited node state data; Based on the constrained node state data, the execution logic and data interaction methods of related nodes within the decision topology are collaboratively modified to generate a decision execution sequence adapted to the current running state.

9. The method according to claim 8, characterized in that, The execution constraint restriction processing is specifically as follows: Boundary conditions for various constraints are extracted from the constraint enforcement strategy, and the boundary conditions are grouped according to the constraint type to form a multi-dimensional constraint decision condition set. The deviation state parameters are compared item by item with each constraint boundary in the multi-dimensional constraint judgment condition set to obtain the parameter types and deviations that exceed the constraint boundaries. Based on the parameter type and deviation amount, the node constraint parameter set of the deviation-related node is corrected, and the node running status data is updated synchronously.

10. The method according to claim 1, characterized in that, The specific process of performing local collaborative optimization is as follows: Extract the node running status information corresponding to the collaborative decision-making topology from the collaborative decision-making status record, and perform comparison processing on the running status information based on a preset threshold to obtain the set of nodes to be optimized. The nodes in the set to be optimized are processed to adjust their execution order, nodes whose dependencies meet the parallel conditions are processed to be arranged in parallel, and nodes whose load information meets the migration conditions are processed to be transferred. The structural information of the collaborative decision-making topology is updated to form a collaborative decision-making topology after local collaborative optimization.

11. The method according to claim 1, characterized in that, The specific process of performing multi-agent collaborative decision-making reasoning is as follows: The initial node is extracted from the locally optimized collaborative decision-making topology, and the Agent decision-making module corresponding to each node is started sequentially based on the order of the topology nodes. Based on the input data requirements of the current node, the output data of the upstream node is obtained, the upstream output data is input into the Agent decision module and calculation processing is performed to generate the decision output data of the current node; The decision output data of the current node is transmitted to the downstream node. The nodes are started, data is acquired and processed in sequence according to the topological node order to form a complete sequence of decision reasoning results.

12. The method according to claim 1, characterized in that, The specific process of performing weighted fusion is as follows: Obtain the decision results and corresponding confidence weights output by each Agent decision module. The confidence weights are generated based on the Agent's historical decision accuracy and real-time load information. Each decision result is aligned according to the decision variable dimension, and the aligned decision results are weighted based on the confidence weight to obtain the fused decision value. The fused decision values ​​are processed as industrial decision support information and simultaneously written into the collaborative decision status record.

Images