An intelligent agricultural trade information management system based on multi-agent cooperation
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANXI ZHONGYI SHARING TECH CO LTD
- Filing Date
- 2026-02-07
- Publication Date
- 2026-05-29
Smart Images
Figure CN122114477A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of smart agricultural trade information management technology, and in particular to a smart agricultural trade information management system based on multi-agent collaboration. Background Technology
[0002] As agricultural markets continue to expand, the agricultural product transaction chain exhibits characteristics such as diverse product types, complex distribution entities, and cross-regional business processes. This has led to traditional agricultural market management methods gradually becoming inadequate in terms of data collection capabilities, responsiveness, and management precision to meet the demands of digital transformation. Existing agricultural market management systems largely rely on manual inspections, static rule matching, or linkage between single business systems. The correlation between business data is weak, and there is a lack of a unified structured processing method for multi-source data, making it difficult for the system to continuously monitor and respond to market operations in real time.
[0003] Currently, common agricultural trade information systems are generally based on modular, customized business systems, such as commodity entry and exit management systems, price monitoring systems, stall management systems, and warehouse scheduling systems. While these systems can achieve a certain level of data recording and process processing within their respective domains, their data collection methods are fixed, and they lack real-time interaction between systems, failing to form an effective overall situational awareness structure. Furthermore, most systems rely on pre-set processes, exhibiting poor adaptability to environmental changes. When transaction volume fluctuates, supply and demand structures become abnormal, or market order is disrupted, the systems lack the ability to autonomously adjust their control strategies.
[0004] On the other hand, traditional agricultural trade management systems have limited capabilities for structured processing of business data. A large amount of heterogeneous data from IoT terminals, trading platforms, monitoring equipment, and regulatory tools lacks a unified organizational form, making it difficult for the system to model the temporal changes and entity relationships inherent in multi-source data. Data processing often relies on static field alignment and simple aggregation methods, failing to achieve deep extraction of temporal dependencies, relationships, and behavioral connections, making subsequent management strategies difficult to adapt to complex scenarios.
[0005] Furthermore, existing agricultural trade systems typically employ fixed rule-driven decision-making, where a predefined set of rules is written into the system, and the system executes corresponding actions when triggered. This rule framework struggles to encompass the constantly changing transaction logic, sudden disturbances, and complex behavioral characteristics of agricultural trade scenarios, including supply and demand responses, and cannot automatically adjust behavioral strategies based on feedback data. Therefore, when market conditions change, the system's regulatory capabilities exhibit a significant lag.
[0006] While some existing research has attempted to apply multi-agent mechanisms to agricultural distribution or supply chain management scenarios, these efforts are largely limited to simulation and prediction or local optimization, failing to construct cross-business domain state, action, and interaction space structures. Existing solutions also fail to encode the dynamic structural relationships within agricultural trade scenarios, lacking multi-agent collaborative mechanisms for complex business behaviors. Furthermore, feedback-based strategy update processes remain rudimentary, often employing single-step backtracking or static weight adjustments, thus failing to achieve continuous optimization of management behaviors.
[0007] In summary, existing agricultural trade information management technologies generally suffer from the following defects: 1. The multi-source business data structure is complex and lacks a unified time-series processing and correlation modeling method; 2. Inter-system interaction is weak, making it difficult to form a dynamically adjustable collaborative management structure; 3. It relies on fixed rule-driven mechanisms and cannot form real-time response and decision adjustment in complex scenarios; 4. It lacks a sustainable iterative behavior generation and strategy update mechanism, and the system cannot automatically adjust its control strategies based on feedback.
[0008] Therefore, how to provide a smart agricultural trade information management system based on multi-agent collaboration is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0009] One objective of this invention is to propose a smart agricultural trade information management system based on multi-agent collaboration. This invention fully utilizes multi-agent modeling technology, structured feature encoding method, graph topology collaboration mechanism, and parameter iterative update method based on perturbation analysis. It describes in detail the entire process of dynamic state construction, action generation, and collaborative control in complex agricultural trade business scenarios. It has the advantages of fine structural expression, stable decision-making links, and adaptive optimization of management behavior.
[0010] According to an embodiment of the present invention, a smart agricultural trade information management system based on multi-agent collaboration includes:
[0011] The data acquisition and processing module is used to collect business data from agricultural trade scenarios, perform preprocessing, and generate a multi-dimensional feature set.
[0012] The multi-agent construction module is used to construct multi-agents and set the state space, action space, and interaction space of the multi-agents;
[0013] The local state generation module is used to structurally encode the multidimensional feature set, expand the temporal dependencies in the state space, and express the association in a matrix form. The structural encoding results are divided according to the topological structure of the interaction space, and dimensional constraints and order constraints are executed to generate a sequence of state vectors of multiple agents.
[0014] The collaborative decision-making module is used to generate action intent vectors of multiple agents in the action space, perform message propagation and multi-round merging processing in the interaction space, and use action generation parameters to perform dependency resolution and matrix assembly on the merging results to generate a joint action matrix.
[0015] The instruction generation module is used to perform hierarchical parsing, behavior binding, and structural rearrangement of the fields in the joint action matrix using a rule-based encoding instruction construction method, generate management execution instructions, and send them to the corresponding terminals.
[0016] The strategy update module is used to collect feedback data after the command is executed, perform perturbation analysis on the feedback data, and iteratively update the action generation parameters based on the analysis results.
[0017] Optionally, modules can be integrated using the following methods:
[0018] S1. Collect business data from agricultural trade scenarios, preprocess it, and generate a multi-dimensional feature set;
[0019] S2. Construct a multi-agent system based on the multi-dimensional feature set, set the state space, action space and interaction space of the multi-agent system, determine the association between the stall index and the state dimension in the state space, register the action type and action generation parameters in the action space, and establish the connection relationship and topology between the multi-agent system in the interaction space.
[0020] S3. Perform structural encoding on the multidimensional feature set, expand the time dependency relationship in the state space, and express the relationship in matrix form. Divide the structural encoding result according to the topological structure of the interaction space, and perform dimensional constraints and order constraints to generate a state vector sequence of multi-agents.
[0021] S4. Generate action intention vectors of multiple agents in the action space based on the state vector sequence, perform message propagation and multi-round merging processing in the interaction space, and use action generation parameters to perform dependency resolution and matrix assembly on the merging results to generate a joint action matrix.
[0022] S5. Using a rule-based encoding instruction construction method, the fields in the joint action matrix are hierarchically parsed, behavior bound, and structure rearranged to generate management execution instructions and send them to the corresponding terminals.
[0023] S6. Collect feedback data after command execution, perform disturbance analysis on the feedback data, and iteratively update the action generation parameters based on the analysis results.
[0024] Optionally, the preprocessing specifically includes: performing time merging, missing data filling, and anomaly removal on the collected business data; performing field normalization and continuous encoding on the merged data; and performing time-series slicing and dimension mapping on the normalized fields based on a preset window.
[0025] Optionally, the multi-agent system includes: a supply and demand agent, a price agent, a traceability agent, a risk control agent, and an environmental agent. Each agent is configured within a state space, an action space, and an interaction space, specifically including:
[0026] The state space includes a set of state fields, a state index structure, and a state value structure. The set of state fields records the time field, category field, batch field, and environment field corresponding to each agent. The state index structure records the mapping relationship of each field in the multidimensional feature set. The state value structure records the value range and value dimension of each field.
[0027] The action space includes a set of action types and action generation parameters. Supply and demand agents register action types related to replenishment and reduction in the action space, price agents register action types related to price adjustment, traceability agents register action types related to batch marking, risk control agents register action types related to risk classification, and environmental agents register action types related to environmental regulation. The action field mapping table records the mapping relationship between action fields and action types. The action generation parameters are a set of numerical vectors corresponding to action types. Each action type registers a parameter vector of fixed length in the action space. The numerical positions of each parameter vector are used to control the numerical value relationship and structural arrangement of action fields during action parsing, action selection, and action assembly.
[0028] The interaction space is a connection structure between agents, including a set of agent nodes, a set of connection edges, and a set of topology indexes. Each agent serves as an interaction node. The set of connection edges records the connection relationships between agent nodes, and the set of topology indexes records the topology paths used in message propagation, merging processing, and action parsing.
[0029] Optionally, S3 specifically includes:
[0030] S31. Perform convolutional unrolling on the time field in the multidimensional feature set, and perform multi-layer convolution calculation in the temporal convolutional network according to the preset dilation rate to generate a time unrolled sequence.
[0031] S32. Based on the entity association relationships in the multidimensional feature set, calculate the degree matrix and generate the graph Laplacian matrix. Perform spectral transformation on the graph Laplacian matrix to obtain the matrix representation of the association relationships.
[0032] S33. Merge the time-expanded sequence with the matrix representation and perform singular value decomposition to generate a coding matrix;
[0033] S34. Perform sequential indexing on the encoding matrix according to the topological structure of the interaction space, perform principal component alignment on the segmented parts, adjust its dimensional structure through piecewise linear mapping, perform QR decomposition and orthogonalization operations on the adjusted matrix segments, and rearrange the matrix positions according to the topological order of the interaction space.
[0034] S35. Perform vectorization mapping on the rearranged matrix, and distribute the mapping results according to the agent index to form a multi-agent state vector sequence.
[0035] Optionally, S4 specifically includes:
[0036] S41. Perform multidimensional expansion on the state vector sequence according to the encoding structure of the action space, convert each state vector into a three-dimensional state structure unit containing channel dimension, feature dimension and structure dimension, and perform kernel function convolution calculation based on the action generation parameters in the action space to generate the response coefficient matrix of the agent in the action space.
[0037] S42. In the topology of the interaction space, with adjacent agents as boundaries, message exchange operation is performed on the response coefficient matrix. The Max-Sum-based message update mechanism is used to iteratively update each node row in the response coefficient matrix, and the updated node row is written into the channel matrix of the adjacent node.
[0038] S43. Perform hierarchical reduction on the channel matrix according to the topology of the interaction space, and perform a combination of additive and selective reduction operations on the channel matrix of nodes at the same level, reducing the dimension of the channel matrix layer by layer with the structured subtree as the reduction unit.
[0039] S44. Based on the dependency weights corresponding to the action generation parameters, perform structured index expansion on the reduced channel matrix, and perform positive semidefinite relaxation projection on the result of the index expansion.
[0040] S45. Perform column compression encoding on the channel matrix after positive semidefinite relaxation projection, arrange the compressed column vectors in the field order of the action space, and perform matrix construction operator to synthesize the structure to form a joint action matrix.
[0041] Optionally, S43 specifically includes:
[0042] S431. According to the topology of the interaction space, divide the channel matrix of each agent node into a structured subtree according to the hierarchical relationship, and generate a channel segment sequence for the nodes at the same level of each subtree.
[0043] S432. Perform a segmented convolution and merging operation on the channel segment sequence, and sum the adjacent channel segments in the sequence by performing segmented convolution according to the set convolution kernel to form an additive reduction matrix.
[0044] S433. Based on the additive reduction matrix, a Top-k selection mechanism is adopted to extract the top k matrix units with the highest numerical values in each channel row, and the extracted matrix units are combined into a selective reduction matrix block according to the channel order.
[0045] S434. According to the hierarchical relationship of the structured subtree, establish a matrix slot for each parent node with the same number of child nodes;
[0046] Write the channel fragments of the child nodes into the corresponding matrix slots according to the branching order of the structured subtree;
[0047] Perform matrix multiplication on the channel segments already written to the slots in the order of subtree branches, and perform hierarchical alignment on the product matrix;
[0048] Perform a dimension compression operation on the class-aligned matrix and write the compressed matrix into the aggregation slot of the parent node;
[0049] Repeat the above operation from bottom to top along the structured subtree to achieve the layer-by-layer reduction of the channel matrix;
[0050] S435. Perform structured folding on the matrix structure after layer-by-layer reduction, and arrange the folded matrix fragments in the hierarchical order of the structured subtrees to form the reduced matrix structure.
[0051] Optionally, S5 specifically includes:
[0052] S51. Based on the rule coding system, the fields in the joint action matrix are encoded according to the state transition relationship of the finite state machine. State-driven field filtering is performed on the fields under different action categories, and the encoded field sequence is used as an instruction to construct an instruction set.
[0053] S52. Perform hierarchical parsing processing on the instruction set, specifically including:
[0054] A hierarchical index structure is built based on the hierarchical descriptors in the fields, and a corresponding hierarchical position number is assigned to each field in the instruction set.
[0055] When a field carries a formatting identifier, perform field regular expression template matching on that field to extract structured fragments from the field sequence;
[0056] When a field does not carry a format identifier, the field is divided into a set of basic segments according to the hierarchical index structure using the default segmentation strategy.
[0057] The fragments obtained by matching the field regular expression template are hierarchically aligned with the basic segment set according to the hierarchical index structure to form a multi-level field segment set;
[0058] S53. Perform behavior binding on the field segment set, map the fields to the corresponding behavior slots according to the behavior identifiers in the action space, execute the field composition operator on the mapped fields, and form a behavior field group in the behavior slot;
[0059] S54. Perform structural rearrangement on the behavior binding table. Perform matrix-style position rearrangement on the field positions in the behavior binding table according to the preset position encoding matrix. Input the rearranged field sequence into the structured combination operator to generate management execution instructions and send them to the instruction channel of the corresponding terminal.
[0060] Optionally, S6 specifically includes:
[0061] S61. Organize the feedback data after the instruction is executed into a feedback sequence according to the time sequence and the terminal identifier. Establish an index relationship between the numerical fields in the feedback sequence and the state vector sequence to generate a feedback vector.
[0062] S62. Perform perturbation analysis on the feedback vector, the perturbation analysis including:
[0063] A random perturbation vector with the same dimension as the feedback vector is generated based on the perturbation intensity;
[0064] The positive and negative perturbations are superimposed on the feedback vectors respectively to form two sets of perturbation feedback vectors;
[0065] When a certain dimension of the disturbance vector is positive, positive sampling is performed on the feedback error signal of the corresponding dimension;
[0066] When a certain dimension of the disturbance vector is negative, reverse sampling is performed on the feedback error signal of the corresponding dimension;
[0067] The cost value is calculated for the feedback error signals corresponding to the two sets of disturbance feedback vectors, and the differential scaling operation is performed based on the two sets of cost values to obtain the disturbance gradient increment.
[0068] S63. Calculate the cost function value for the feedback error signal vector under the two sets of perturbation feedback vectors respectively, perform differential scaling operation on the two sets of cost function values, and form the gradient estimation vector of the action generation parameters according to the SPSA algorithm.
[0069] S64. Based on the gradient estimation vector, the Adam optimization method is used to perform iterative updates on the action generation parameters. The update step size is calculated according to the first moment estimation and the second moment estimation. Weighted correction operations are performed on each element of the action generation parameters.
[0070] S65. Write the updated action generation parameters to the corresponding parameter storage location in the action space, and synchronously update the action generation parameters referenced within the multi-agent.
[0071] The beneficial effects of this invention are:
[0072] First, by constructing a multi-dimensional feature set of multi-source business data and completing the temporal structure expansion and matrix expression of the correlation in the state space, the present invention enables the complex data relationships in the agricultural trade scenario to be recorded and organized in a fine-grained manner, realizing a continuous and structured representation of the market operation status, and effectively avoiding the shortcomings of data isolation, structural chaos and inability to capture change trends in existing systems.
[0073] Secondly, this invention establishes a collaborative behavior generation mechanism based on the action space and interaction space of multiple agents. Through action intent modeling, message update based on Max-Sum, structured subtree reduction, and matrix-level instruction construction process, a management action link that can be adjusted in real time with changes in market structure is formed. This realizes the ability of multi-node collaborative regulation in complex agricultural trade business environments and overcomes the shortcomings of traditional systems such as rigid rules, slow response, and difficulty in cross-domain business linkage.
[0074] Finally, this invention performs perturbation analysis on the feedback data after instruction execution, and combines a simultaneous perturbation random approximation method with a parameter iterative update strategy to continuously correct the action generation parameters, enabling the management system to have adaptive behavior update capabilities. This allows for strategy enhancement and behavior optimization during long-term operation, avoiding the management capability decline caused by traditional systems relying on manual maintenance and static rule updates. As a result, the overall control accuracy and management efficiency in smart agricultural trade scenarios are significantly improved. Attached Figure Description
[0075] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0076] Figure 1 This is a module structure diagram of a smart agricultural trade information management system based on multi-agent collaboration proposed in this invention;
[0077] Figure 2 This is a flowchart of a smart agricultural trade information management system based on multi-agent collaboration proposed in this invention.
[0078] Figure 3 This is a flowchart illustrating the construction of management execution instructions for a smart agricultural trade information management system based on multi-agent collaboration proposed in this invention. Detailed Implementation
[0079] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0080] refer to Figure 1 A smart agricultural trade information management system based on multi-agent collaboration includes:
[0081] The data acquisition and processing module is used to collect business data from agricultural trade scenarios, perform preprocessing, and generate a multi-dimensional feature set.
[0082] The multi-agent construction module is used to construct multi-agents and set the state space, action space, and interaction space of the multi-agents;
[0083] The local state generation module is used to structurally encode the multidimensional feature set, expand the temporal dependencies in the state space, and express the association in a matrix form. The structural encoding results are divided according to the topological structure of the interaction space, and dimensional constraints and order constraints are executed to generate a sequence of state vectors of multiple agents.
[0084] The collaborative decision-making module is used to generate action intent vectors of multiple agents in the action space, perform message propagation and multi-round merging processing in the interaction space, and use action generation parameters to perform dependency resolution and matrix assembly on the merging results to generate a joint action matrix.
[0085] The instruction generation module is used to perform hierarchical parsing, behavior binding, and structural rearrangement of the fields in the joint action matrix using a rule-based encoding instruction construction method, generate management execution instructions, and send them to the corresponding terminals.
[0086] The strategy update module is used to collect feedback data after the command is executed, perform perturbation analysis on the feedback data, and iteratively update the action generation parameters based on the analysis results.
[0087] refer to Figure 2-3 In this embodiment, the modules are interconnected using the following method:
[0088] S1. Collect business data from agricultural trade scenarios, preprocess it, and generate a multi-dimensional feature set;
[0089] S2. Construct a multi-agent system based on the multi-dimensional feature set, set the state space, action space and interaction space of the multi-agent system, determine the association between the stall index and the state dimension in the state space, register the action type and action generation parameters in the action space, and establish the connection relationship and topology between the multi-agent system in the interaction space.
[0090] S3. Perform structural encoding on the multidimensional feature set, expand the time dependency relationship in the state space, and express the relationship in matrix form. Divide the structural encoding result according to the topological structure of the interaction space, and perform dimensional constraints and order constraints to generate a state vector sequence of multi-agents.
[0091] S4. Generate action intention vectors of multiple agents in the action space based on the state vector sequence, perform message propagation and multi-round merging processing in the interaction space, and use action generation parameters to perform dependency resolution and matrix assembly on the merging results to generate a joint action matrix.
[0092] S5. Using a rule-based encoding instruction construction method, the fields in the joint action matrix are hierarchically parsed, behavior bound, and structure rearranged to generate management execution instructions and send them to the corresponding terminals.
[0093] S6. Collect feedback data after command execution, perform disturbance analysis on the feedback data, and iteratively update the action generation parameters based on the analysis results.
[0094] In this embodiment, the preprocessing specifically includes: performing time merging, missing data filling, and anomaly removal on the collected business data; performing field normalization and continuous encoding on the merged data; and performing time-series slicing and dimension mapping on the normalized fields based on a preset window.
[0095] In this embodiment, the multi-agent system includes: a supply and demand agent, a price agent, a traceability agent, a risk control agent, and an environmental agent. Each agent is configured within a state space, an action space, and an interaction space, specifically including:
[0096] The state space includes a set of state fields, a state index structure, and a state value structure. The set of state fields records the time field, category field, batch field, and environment field corresponding to each agent. The state index structure records the mapping relationship of each field in the multidimensional feature set. The state value structure records the value range and value dimension of each field.
[0097] The action space includes a set of action types and action generation parameters. Supply and demand agents register action types related to replenishment and reduction in the action space, price agents register action types related to price adjustment, traceability agents register action types related to batch marking, risk control agents register action types related to risk classification, and environmental agents register action types related to environmental regulation. The action field mapping table records the mapping relationship between action fields and action types. The action generation parameters are a set of numerical vectors corresponding to action types. Each action type registers a parameter vector of fixed length in the action space. The numerical positions of each parameter vector are used to control the numerical value relationship and structural arrangement of action fields during action parsing, action selection, and action assembly.
[0098] The interaction space is a connection structure between agents, including a set of agent nodes, a set of connection edges, and a set of topology indexes. Each agent serves as an interaction node. The set of connection edges records the connection relationships between agent nodes, and the set of topology indexes records the topology paths used in message propagation, merging processing, and action parsing.
[0099] In this embodiment, S3 specifically includes:
[0100] S31. Perform convolutional unrolling on the time field in the multidimensional feature set, and perform multi-layer convolution calculation in the temporal convolutional network according to the preset dilation rate to generate a time unrolled sequence.
[0101] S32. Based on the entity association relationships in the multidimensional feature set, calculate the degree matrix and generate a graph Laplacian matrix. Perform a spectral transformation on the graph Laplacian matrix to obtain a matrix representation of the association relationships. The generation and spectral transformation of the graph Laplacian matrix specifically include:
[0102] Based on the entity association relationships recorded in the multidimensional feature set, any two entities with a connection relationship are indexed as corresponding matrix positions, and non-zero connection markers are assigned in the association matrix. For entity pairs without a connection relationship, null positions are maintained in the association matrix.
[0103] The correlation matrix is numerically scanned row by row. Non-zero connection markers appearing in each row are accumulated, and the accumulated results are written into the degree value position corresponding to the same row to generate a degree matrix that is consistent with the rows and columns of the correlation matrix.
[0104] Perform a row-by-row difference write operation on the degree matrix and the incidence matrix, record the difference results at the same position, and form a graph Laplacian matrix;
[0105] Expand the Laplacian matrix into rows and columns, treat the matrix as a linear transformation structure, perform spectral decomposition, sort the eigenvectors obtained in the decomposition process according to their corresponding eigenvalues, and combine the sorted eigenvectors into a spectral vector set.
[0106] Based on the spectral vector set, the spectral vectors of the same source nodes are written into the matrix rows in a column-wise structure to form an association matrix with multi-channel representation. Each column in the association matrix corresponds to the projection of a certain spectral vector dimension, and each row corresponds to the position index of the entity node in the spectral space.
[0107] S33. Merge the time-expanded sequence with the matrix representation and perform singular value decomposition to generate a coding matrix;
[0108] S34. Perform sequential indexing on the encoding matrix according to the topological structure of the interaction space, perform principal component alignment on the segmented parts, adjust its dimensional structure through piecewise linear mapping, perform QR decomposition and orthogonalization operations on the adjusted matrix segments, and rearrange the matrix positions according to the topological order of the interaction space.
[0109] S35. Perform vectorization mapping on the rearranged matrix, and distribute the mapping results according to the agent index to form a multi-agent state vector sequence.
[0110] In this embodiment, S4 specifically includes:
[0111] S41. Perform multidimensional expansion on the state vector sequence according to the encoding structure of the action space, convert each state vector into a three-dimensional state structure unit containing channel dimension, feature dimension and structure dimension, and perform kernel function convolution calculation based on the action generation parameters in the action space to generate the response coefficient matrix of the agent in the action space.
[0112] S42. In the topology of the interaction space, with adjacent agents as boundaries, message exchange operation is performed on the response coefficient matrix. The Max-Sum-based message update mechanism is used to iteratively update each node row in the response coefficient matrix, and the updated node row is written into the channel matrix of the adjacent node.
[0113] In the topology of the interaction space, a message channel is established for each adjacent edge according to the adjacency relationship between agents. The index positions of the sending end node and the receiving end node are recorded for each message channel. The message unit is initialized in the message channel, and the response coefficient is written to the initial position corresponding to the message unit according to the node index.
[0114] For each agent node, read the corresponding response coefficient vector, traverse all neighboring nodes, and perform the following operations for each neighboring node:
[0115] Remove the corresponding dimension from the response coefficient vector of the current node according to the index of the adjacent node to form a candidate response set;
[0116] Perform an element-wise summation operation on the candidate response set, and add the summation value to the value in the internal channel matrix of the node to form the intermediate message volume from the node to the neighboring nodes;
[0117] Perform a maximum value scan on the intermediate message volume, record the position index of the maximum value and its corresponding value, and form the Max-Sum message value;
[0118] Write the Max-Sum message value to the update slot of the current message channel;
[0119] Perform synchronous read operations on all message channels, write the Max-Sum message value in each channel to the corresponding response coefficient position of the node according to the receiving node index, perform an overwrite write for each dimension of the response coefficient, and generate an updated response coefficient vector.
[0120] The updated response coefficient vector is written into the channel matrix of the adjacent nodes according to the node index, and a row-by-row overwrite operation is performed on the channel matrix.
[0121] S43. Perform hierarchical reduction on the channel matrix according to the topology of the interaction space, and perform a combination of additive and selective reduction operations on the channel matrix of nodes at the same level, reducing the dimension of the channel matrix layer by layer with the structured subtree as the reduction unit.
[0122] S44. Based on the dependency weights corresponding to the action generation parameters, perform structured index expansion on the reduced channel matrix, and perform positive semidefinite relaxation projection on the result of the index expansion.
[0123] S45. Perform column compression encoding on the channel matrix after positive semidefinite relaxation projection, arrange the compressed column vectors in the field order of the action space, and perform matrix construction operator to synthesize the structure to form a joint action matrix.
[0124] In this embodiment, S43 specifically includes:
[0125] S431. According to the topology of the interaction space, divide the channel matrix of each agent node into a structured subtree according to the hierarchical relationship, and generate a channel segment sequence for the nodes at the same level of each subtree.
[0126] S432. Perform a segmented convolution and merging operation on the channel segment sequence, and sum the adjacent channel segments in the sequence by performing segmented convolution according to the set convolution kernel to form an additive reduction matrix.
[0127] S433. Based on the additive reduction matrix, a Top-k selection mechanism is adopted to extract the top k matrix units with the highest numerical values in each channel row, and the extracted matrix units are combined into a selective reduction matrix block according to the channel order.
[0128] S434. According to the hierarchical relationship of the structured subtree, establish a matrix slot for each parent node with the same number of child nodes;
[0129] Write the channel fragments of the child nodes into the corresponding matrix slots according to the branching order of the structured subtree;
[0130] Perform matrix multiplication on the channel segments already written to the slots in the order of subtree branches, and perform hierarchical alignment on the product matrix;
[0131] Perform a dimension compression operation on the class-aligned matrix and write the compressed matrix into the aggregation slot of the parent node;
[0132] Repeat the above operation from bottom to top along the structured subtree to achieve the layer-by-layer reduction of the channel matrix;
[0133] S435. Perform structured folding on the matrix structure after layer-by-layer reduction, and arrange the folded matrix fragments in the hierarchical order of the structured subtrees to form the reduced matrix structure.
[0134] In this embodiment, S5 specifically includes:
[0135] S51. Based on the rule coding system, the fields in the joint action matrix are encoded according to the state transition relationship of the finite state machine. State-driven field filtering is performed on the fields under different action categories, and the encoded field sequence is used as the instruction set.
[0136] S52. Perform hierarchical parsing processing on the instruction set, specifically including:
[0137] A hierarchical index structure is built based on the hierarchical descriptors in the fields, and a corresponding hierarchical position number is assigned to each field in the instruction set.
[0138] When a field carries a formatting identifier, perform field regular expression template matching on that field to extract structured fragments from the field sequence;
[0139] When a field does not carry a format identifier, the field is divided into a set of basic segments according to the hierarchical index structure using the default segmentation strategy.
[0140] The fragments obtained by matching the field regular expression template are hierarchically aligned with the basic segment set according to the hierarchical index structure to form a multi-level field segment set;
[0141] S53. Perform behavior binding on the field segment set, map the fields to the corresponding behavior slots according to the behavior identifiers in the action space, execute the field composition operator on the mapped fields, and form a behavior field group in the behavior slot;
[0142] S54. Perform structural rearrangement on the behavior binding table. Perform matrix-style position rearrangement on the field positions in the behavior binding table according to the preset position encoding matrix. Input the rearranged field sequence into the structured combination operator to generate management execution instructions and send them to the instruction channel of the corresponding terminal.
[0143] In this embodiment, S6 specifically includes:
[0144] S61. Organize the feedback data after the instruction is executed into a feedback sequence according to the time sequence and the terminal identifier. Establish an index relationship between the numerical fields in the feedback sequence and the state vector sequence to generate a feedback vector.
[0145] S62. Perform perturbation analysis on the feedback vector, the perturbation analysis including:
[0146] A random perturbation vector with the same dimension as the feedback vector is generated based on the perturbation intensity;
[0147] The positive and negative perturbations are superimposed on the feedback vectors respectively to form two sets of perturbation feedback vectors;
[0148] When a certain dimension of the disturbance vector is positive, positive sampling is performed on the feedback error signal of the corresponding dimension;
[0149] When a certain dimension of the disturbance vector is negative, reverse sampling is performed on the feedback error signal of the corresponding dimension;
[0150] The cost value is calculated for the feedback error signals corresponding to the two sets of disturbance feedback vectors, and the differential scaling operation is performed based on the two sets of cost values to obtain the disturbance gradient increment.
[0151] S63. Calculate the cost function value for the feedback error signal vector under the two sets of perturbation feedback vectors respectively, perform differential scaling operation on the two sets of cost function values, and form the gradient estimation vector of the action generation parameters according to the SPSA algorithm.
[0152] S64. Based on the gradient estimation vector, the Adam optimization method is used to perform iterative updates on the action generation parameters. The update step size is calculated according to the first moment estimation and the second moment estimation. Weighted correction operations are performed on each element of the action generation parameters.
[0153] S65. Write the updated action generation parameters to the corresponding parameter storage location in the action space, and synchronously update the action generation parameters referenced within the multi-agent.
[0154] Example 1:
[0155] To verify the feasibility of this invention in practice, it was applied to a large-scale agricultural trade complex. This complex includes numerous stall operators, different types of agricultural product trading channels, real-time monitoring equipment, price recording terminals, and supply scheduling. The overall business structure is characterized by a wide variety of commodities, large transaction fluctuations, and rapid changes in supply and demand. Traditional technologies in this scenario generally suffer from problems such as difficulty in unifying the organization of multi-source data, inability of management decisions to respond in real time to complex and changing structures, and inability of rule engines to cover dynamic transaction behavior. This leads to frequent occurrences of supply and demand imbalances, chaotic stall management, and untimely scheduling responses.
[0156] During the trial operation in this scenario, the invention first collected transaction volume data, inventory change data, stall sales behavior data, price change data, and environmental monitoring data from the transaction flow recording terminal, weighing and measuring equipment, price monitoring equipment, inventory registration module, and monitoring and analysis system. The preprocessing module then formatted the raw data, extracted key fields, and removed noise to form a structured multidimensional feature set. This feature set includes time, transaction volume, inventory, association marker, and behavioral parameter fields, enabling fine-grained depiction of the changing trends of each business entity over different periods.
[0157] After the data enters the processing chain, the feature structure encoding process of this invention expands the time field into a continuous time series matrix and constructs a graph structure for the association markers between business entities. It uses the degree matrix and the graph Laplacian matrix to generate a spectral space expression reflecting the association structure, enabling the system to obtain the mutual influence relationships and directions of change between each stall. The state vector sequence generated through structure encoding can accurately reflect the current market state in the subsequent action generation process, including transaction growth rate, inventory pressure, short-term abnormal sales behavior, price fluctuation trajectory, etc.
[0158] In the multi-agent collaborative behavior generation stage, this invention constructs state spaces and action spaces for the supply scheduling agent, price monitoring agent, order management agent, and supply coordination agent, respectively. These agents form a collaborative framework in the interaction space through response coefficients. Each agent generates an action intent vector based on the state vector sequence, exchanges response information through a message update mechanism based on the Max-Sum algorithm, and iteratively corrects its respective response coefficients, enabling multiple agents to gradually form coordinated behaviors that conform to the current market structure. For example, when the transaction volume in a certain area experiences an abnormal increase in a short period of time, the response coefficient of the supply scheduling agent will rise due to message propagation, and the associated supply coordination agent will also form a synchronized action trend after iterative updates.
[0159] In the instruction construction phase, this invention employs a rule-based encoding-based instruction construction method. This method performs hierarchical parsing, behavior binding, and structural rearrangement of the joint action matrix, ultimately generating management execution instructions for each terminal. These instructions include replenishment instructions, price warnings, order guidance instructions, and regional dispatch control instructions. These instructions are then distributed to dispatch terminals, monitoring terminals, and on-site execution terminals, enabling the system to respond consistently and systematically to environmental changes.
[0160] During system operation, this invention performs perturbation analysis on the feedback data after executing instructions and iteratively updates the action generation parameters, enabling the agent's behavior pattern to continuously adjust according to changes in business conditions. For example, when some regions experience prolonged periods of sluggish sales, changes in the values in the feedback sequence will alter the gradient direction obtained from the perturbation analysis, thereby correcting the action generation parameters of the price monitoring agent and the supply coordination agent, making subsequent generated instructions more aligned with the actual situation. Through continuous feedback loops, the agent's regulatory behavior gradually stabilizes in a direction suitable for the current market structure.
[0161] During a period of continuous trial operation, this invention conducted statistical analysis on market operation data, selecting indicators such as transaction volume stability, inventory turnover efficiency, replenishment delay, and response time to abnormal behavior as evaluation criteria. Statistical results show that after using this invention, transaction volume stability improved by an average of approximately 18.7%, inventory turnover efficiency improved by approximately 22.4%, replenishment delay decreased by approximately 31.5%, and response time to abnormal behavior shortened by approximately 27.8%. Furthermore, the system's accuracy in identifying abnormal price fluctuations improved by approximately 15.3%, and the supply scheduling agent was able to accurately predict inventory pressure in most volatile scenarios after continuous disturbance updates. During peak trading hours, the scheduling and order instructions received by each terminal significantly reduced backlogs, making the overall market operation smoother.
[0162] To visually demonstrate the actual effects of this invention in application scenarios, the core data statistics during the trial operation are presented in Table 1:
[0163] Table 1. Statistical Table of Operational Effectiveness of the Smart Agricultural Trade Multi-Agent Collaborative System
[0164] Indicator Categories Average value before trial operation Average value after trial run Improvement rate (%) Trading volume stability volatility (%) 12.40 10.10 18.70 Inventory turnover efficiency (%) 78.30 95.80 22.40 Replenishment response time (minutes) 14.6 10.0 31.50 Abnormal response time (minutes) 11.2 8.1 27.80 Price anomaly detection accuracy (%) 82.60 95.20 15.30 Scheduling strategy accuracy (%) 73.50 88.40 20.30 Equilibrium supply-demand ratio deviation 0.27 0.19 29.60 Number of people stranded during peak hours 112 78 30.40
[0165] As shown in Table 1, the multi-agent collaborative behavior generation mechanism of this invention can continuously converge and regulate strategies in complex agricultural trade scenarios, dynamically adapting to changes in the transaction structure, and achieving more efficient supply and demand regulation, clearer order guidance, and more stable price monitoring. This invention, through the structured expression of multi-source features, the agent collaborative mechanism, and feedback-based strategy updates, achieves collaborative management capabilities that traditional systems cannot reach, significantly improving the operational efficiency and data response capabilities of the entire agricultural trade scenario.
[0166] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A smart agricultural trade information management system based on multi-agent collaboration, characterized in that, include: The data acquisition and processing module is used to collect business data from agricultural trade scenarios, perform preprocessing, and generate a multi-dimensional feature set. The multi-agent construction module is used to construct multi-agents and set the state space, action space, and interaction space of the multi-agents; The local state generation module is used to structurally encode the multidimensional feature set, expand the temporal dependencies in the state space, and express the association in a matrix form. The structural encoding results are divided according to the topological structure of the interaction space, and dimensional constraints and order constraints are executed to generate a sequence of state vectors of multiple agents. The collaborative decision-making module is used to generate action intent vectors of multiple agents in the action space, perform message propagation and multi-round merging processing in the interaction space, and use action generation parameters to perform dependency resolution and matrix assembly on the merging results to generate a joint action matrix. The instruction generation module is used to perform hierarchical parsing, behavior binding, and structural rearrangement of the fields in the joint action matrix using a rule-based encoding instruction construction method, generate management execution instructions, and send them to the corresponding terminals. The strategy update module is used to collect feedback data after the command is executed, perform perturbation analysis on the feedback data, and iteratively update the action generation parameters based on the analysis results.
2. The intelligent agricultural trade information management system based on multi-agent collaboration according to claim 1, characterized in that, The modules are connected in the following way: S1. Collect business data from agricultural trade scenarios, preprocess it, and generate a multi-dimensional feature set; S2. Construct a multi-agent system based on the multi-dimensional feature set, set the state space, action space and interaction space of the multi-agent system, determine the association between the stall index and the state dimension in the state space, register the action type and action generation parameters in the action space, and establish the connection relationship and topology between the multi-agent system in the interaction space. S3. Perform structural encoding on the multidimensional feature set, expand the time dependency relationship in the state space, and express the relationship in matrix form. Divide the structural encoding result according to the topological structure of the interaction space, and perform dimensional constraints and order constraints to generate a state vector sequence of multi-agents. S4. Generate action intention vectors of multiple agents in the action space based on the state vector sequence, perform message propagation and multi-round merging processing in the interaction space, and use action generation parameters to perform dependency resolution and matrix assembly on the merging results to generate a joint action matrix. S5. Using a rule-based encoding instruction construction method, the fields in the joint action matrix are hierarchically parsed, behavior bound, and structure rearranged to generate management execution instructions and send them to the corresponding terminals. S6. Collect feedback data after command execution, perform disturbance analysis on the feedback data, and iteratively update the action generation parameters based on the analysis results.
3. The intelligent agricultural trade information management system based on multi-agent collaboration according to claim 2, characterized in that, The preprocessing specifically includes: performing time merging, missing data filling, and anomaly removal on the collected business data; performing field normalization and continuous encoding on the merged data; and performing time-series slicing and dimension mapping on the normalized fields based on a preset window.
4. The intelligent agricultural trade information management system based on multi-agent collaboration according to claim 2, characterized in that, The multi-agent system includes: a supply and demand agent, a price agent, a traceability agent, a risk control agent, and an environmental agent. Each agent is configured within a state space, an action space, and an interaction space, specifically including: The state space includes a set of state fields, a state index structure, and a state value structure. The set of state fields records the time field, category field, batch field, and environment field corresponding to each agent. The state index structure records the mapping relationship of each field in the multidimensional feature set. The state value structure records the value range and value dimension of each field. The action space includes a set of action types and action generation parameters. Supply and demand agents register action types related to replenishment and reduction in the action space, price agents register action types related to price adjustment, traceability agents register action types related to batch marking, risk control agents register action types related to risk classification, and environmental agents register action types related to environmental regulation. The action field mapping table records the mapping relationship between action fields and action types. The action generation parameters are a set of numerical vectors corresponding to action types. Each action type registers a parameter vector of fixed length in the action space. The numerical positions of each parameter vector are used to control the numerical value relationship and structural arrangement of action fields during action parsing, action selection, and action assembly. The interaction space is a connection structure between agents, including a set of agent nodes, a set of connection edges, and a set of topology indexes. Each agent serves as an interaction node. The set of connection edges records the connection relationships between agent nodes, and the set of topology indexes records the topology paths used in message propagation, merging processing, and action parsing.
5. A smart agricultural trade information management system based on multi-agent collaboration according to claim 2, characterized in that, S3 specifically includes: S31. Perform convolutional unrolling on the time field in the multidimensional feature set, and perform multi-layer convolution calculation in the temporal convolutional network according to the preset dilation rate to generate a time unrolled sequence. S32. Based on the entity association relationships in the multidimensional feature set, calculate the degree matrix and generate the graph Laplacian matrix. Perform spectral transformation on the graph Laplacian matrix to obtain the matrix representation of the association relationships. S33. Merge the time-expanded sequence with the matrix representation and perform singular value decomposition to generate a coding matrix; S34. Perform sequential indexing on the encoding matrix according to the topological structure of the interaction space, perform principal component alignment on the segmented parts, adjust its dimensional structure through piecewise linear mapping, perform QR decomposition and orthogonalization operations on the adjusted matrix segments, and rearrange the matrix positions according to the topological order of the interaction space. S35. Perform vectorization mapping on the rearranged matrix, and distribute the mapping results according to the agent index to form a multi-agent state vector sequence.
6. The intelligent agricultural trade information management system based on multi-agent collaboration according to claim 2, characterized in that, S4 specifically includes: S41. Perform multidimensional expansion on the state vector sequence according to the encoding structure of the action space, convert each state vector into a three-dimensional state structure unit containing channel dimension, feature dimension and structure dimension, and perform kernel function convolution calculation based on the action generation parameters in the action space to generate the response coefficient matrix of the agent in the action space. S42. In the topology of the interaction space, with adjacent agents as boundaries, message exchange operation is performed on the response coefficient matrix. The Max-Sum-based message update mechanism is used to iteratively update each node row in the response coefficient matrix, and the updated node row is written into the channel matrix of the adjacent node. S43. Perform hierarchical reduction on the channel matrix according to the topology of the interaction space, and perform a combination of additive and selective reduction operations on the channel matrix of nodes at the same level, reducing the dimension of the channel matrix layer by layer with the structured subtree as the reduction unit. S44. Based on the dependency weights corresponding to the action generation parameters, perform structured index expansion on the reduced channel matrix, and perform positive semidefinite relaxation projection on the result of the index expansion. S45. Perform column compression encoding on the channel matrix after positive semidefinite relaxation projection, arrange the compressed column vectors in the field order of the action space, and perform matrix construction operator to synthesize the structure to form a joint action matrix.
7. A smart agricultural trade information management system based on multi-agent collaboration as described in claim 6, characterized in that, Specifically, S43 includes: S431. According to the topology of the interaction space, divide the channel matrix of each agent node into a structured subtree according to the hierarchical relationship, and generate a channel segment sequence for the nodes at the same level of each subtree. S432. Perform a segmented convolution and merging operation on the channel segment sequence, and sum the adjacent channel segments in the sequence by performing segmented convolution according to the set convolution kernel to form an additive reduction matrix. S433. Based on the additive reduction matrix, a Top-k selection mechanism is adopted to extract the top k matrix units with the highest numerical values in each channel row, and the extracted matrix units are combined into a selective reduction matrix block according to the channel order. S434. According to the hierarchical relationship of the structured subtree, establish a matrix slot for each parent node with the same number of child nodes; Write the channel fragments of the child nodes into the corresponding matrix slots according to the branching order of the structured subtree; Perform matrix multiplication on the channel segments already written to the slots in the order of subtree branches, and perform hierarchical alignment on the product matrix; Perform a dimension compression operation on the class-aligned matrix and write the compressed matrix into the aggregation slot of the parent node; Repeat the above operation from bottom to top along the structured subtree to achieve the layer-by-layer reduction of the channel matrix; S435. Perform structured folding on the matrix structure after layer-by-layer reduction, and arrange the folded matrix fragments in the hierarchical order of the structured subtrees to form the reduced matrix structure.
8. A smart agricultural trade information management system based on multi-agent collaboration according to claim 2, characterized in that, S5 specifically includes: S51. Based on the rule coding system, the fields in the joint action matrix are encoded according to the state transition relationship of the finite state machine. State-driven field filtering is performed on the fields under different action categories, and the encoded field sequence is used as the instruction set. S52. Perform hierarchical parsing processing on the instruction set, specifically including: A hierarchical index structure is built based on the hierarchical descriptors in the fields, and a corresponding hierarchical position number is assigned to each field in the instruction set. When a field carries a formatting identifier, perform field regular expression template matching on that field to extract structured fragments from the field sequence; When a field does not carry a format identifier, the field is divided into a set of basic segments according to the hierarchical index structure using the default segmentation strategy. The fragments obtained by matching the field regular expression template are hierarchically aligned with the basic segment set according to the hierarchical index structure to form a multi-level field segment set; S53. Perform behavior binding on the field segment set, map the fields to the corresponding behavior slots according to the behavior identifiers in the action space, execute the field composition operator on the mapped fields, and form a behavior field group in the behavior slot; S54. Perform structural rearrangement on the behavior binding table. Perform matrix-style position rearrangement on the field positions in the behavior binding table according to the preset position encoding matrix. Input the rearranged field sequence into the structured combination operator to generate management execution instructions and send them to the instruction channel of the corresponding terminal.
9. A smart agricultural trade information management system based on multi-agent collaboration according to claim 2, characterized in that, S6 specifically includes: S61. Organize the feedback data after the instruction is executed into a feedback sequence according to the time sequence and the terminal identifier. Establish an index relationship between the numerical fields in the feedback sequence and the state vector sequence to generate a feedback vector. S62. Perform perturbation analysis on the feedback vector, the perturbation analysis including: A random perturbation vector with the same dimension as the feedback vector is generated based on the perturbation intensity; The positive and negative perturbations are superimposed on the feedback vectors respectively to form two sets of perturbation feedback vectors; When a certain dimension of the disturbance vector is positive, positive sampling is performed on the feedback error signal of the corresponding dimension; When a certain dimension of the disturbance vector is negative, reverse sampling is performed on the feedback error signal of the corresponding dimension; The cost value is calculated for the feedback error signals corresponding to the two sets of disturbance feedback vectors, and the differential scaling operation is performed based on the two sets of cost values to obtain the disturbance gradient increment. S63. Calculate the cost function value for the feedback error signal vector under the two sets of perturbation feedback vectors respectively, perform differential scaling operation on the two sets of cost function values, and form the gradient estimation vector of the action generation parameters according to the SPSA algorithm. S64. Based on the gradient estimation vector, the Adam optimization method is used to perform iterative updates on the action generation parameters. The update step size is calculated according to the first moment estimation and the second moment estimation. Weighted correction operations are performed on each element of the action generation parameters. S65. Write the updated action generation parameters to the corresponding parameter storage location in the action space, and synchronously update the action generation parameters referenced within the multi-agent.