Multi-agent collaborative reasoning method and system for supply chain supply and demand collaboration

Abstract

The application discloses a multi-agent collaborative reasoning method and system for supply chain supply and demand coordination, relates to the field of supply chain supply and demand coordination intelligent reasoning, and first deploys a dedicated multi-modal agent with matched business attributes for the edge computing nodes of each participant in the supply chain. Each agent locally loads a lightweight supply and demand timing reasoning model, completes localized supply and demand feature reasoning based on multi-dimensional business data stored locally by encryption, outputs multi-period local supply and demand prediction initial values and gradient residual fragments, and original sensitive business data is stored locally throughout and is not transmitted externally. Subsequently, an encrypted interaction channel is constructed according to the upstream and downstream business topology of the supply chain, and adjacent agents only transmit the gradient residual fragments encrypted by homomorphism. The application realizes zero external transmission of original commercial information in the cross-agent collaboration process, eliminates the data security concerns of participants, avoids input data distortion problems caused by false reporting, and greatly improves the collaborative reasoning accuracy.

Description

Technical Field

[0001] This invention relates to the field of intelligent reasoning technology for supply chain supply and demand collaboration, and in particular to a multi-agent collaborative reasoning method and system for supply chain supply and demand collaboration. Background Technology

[0002] Supply chain coordination is a core element for reducing costs and increasing efficiency in the real economy. It is widely used in the entire operational process, including raw material procurement, production scheduling, inventory management, and logistics distribution. In the current digital transformation of industries, achieving dynamic coordination among multiple entities, mitigating the bullwhip effect, and ensuring data security have become core development needs in the supply chain field.

[0003] Currently, mainstream supply chain supply and demand collaboration technologies mainly fall into two categories. The first category is centralized collaborative forecasting solutions. The working principle involves a core enterprise or third-party service provider building a unified supply and demand forecasting system. All participants in the supply chain are required to upload their local raw business data, such as capacity, inventory, orders, and customers, to a central platform. The platform then completes end-to-end supply and demand calculations based on a centralized time-series forecasting model, uniformly outputting production quotas, safety stock, and logistics allocation guidelines for each node. The current advantages of this solution lie in its centralized management of model training and iteration, and its relatively low technical implementation threshold. It has already been widely applied in small, closed, vertical supply chain scenarios. However, this solution has inherent pain points that are difficult to address: First, there is a risk of leakage due to the external transmission of core business data of each participant. In order to protect their own business interests, the participants generally conceal or misreport data, which leads to the distortion of model input data and a significant decrease in prediction accuracy. Secondly, the central platform stores sensitive data across the entire supply chain, posing multiple risks such as single point of failure and data compliance, making it difficult to adapt to open supply chain scenarios that involve multiple entities and regions.

[0004] The second type is the point-to-point adjacency collaboration solution. Its working principle is that supply chain nodes only exchange structured data on orders and inventory with their direct upstream and downstream partners through standardized protocols such as EDI, completing supply and demand matching based on bilateral negotiation. The existing advantages of this solution are controllable data exchange scope and lower security concerns among participating entities, and it has been used for a long time in traditional offline trade scenarios. However, the shortcomings of this solution lie in the lack of a global collaborative perspective. It can only achieve supply and demand matching between adjacent nodes and cannot adapt to the supply and demand transmission patterns of the entire supply chain. When fluctuations in end-user demand are transmitted upwards along the supply chain, they are amplified layer by layer, forming a significant bullwhip effect. Furthermore, it cannot support global-level collaborative reasoning and rapid response to anomalies. Sudden supply and demand fluctuations often require multiple rounds of manual negotiation to complete adjustments across the entire supply chain, resulting in extremely low response efficiency.

[0005] As the complexity of supply chain networks continues to increase, open supply chain scenarios that cross regions and entities are becoming more common. The existing two types of solutions can no longer simultaneously meet the multiple requirements of privacy protection, collaborative accuracy, and response efficiency. The industry urgently needs new collaborative reasoning technologies that can adapt to complex supply chain scenarios to solve the core pain points currently faced in supply and demand collaboration. Summary of the Invention

[0006] The present invention proposes a multi-agent collaborative reasoning method and system for supply chain supply and demand coordination, in order to solve the problems of high privacy risk, low reasoning accuracy and slow anomaly response in existing supply chain supply and demand coordination schemes mentioned above.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a multi-agent collaborative reasoning method for supply chain supply and demand coordination, comprising the following steps: S1. For edge computing nodes of five types of participants in the supply chain network—raw material suppliers, manufacturers, distributors, terminal retailers, and trunk logistics service providers—deploy dedicated multimodal agents that match the business attributes of each entity. Each dedicated agent locally loads a lightweight supply and demand time-series inference model that integrates temporal convolution and causal attention mechanisms. The model parameter size is adapted to the computing power threshold of the edge node. Based on the encrypted storage of historical capacity ramp-up coefficients, safety stock turnover days, order fulfillment rates, in-transit logistics trajectories, regional supply and demand disturbance factors, etc., in the local trusted execution environment of the node, localized supply and demand feature inference is completed. The initial value of local supply and demand forecast for the next 1 to 12 statistical periods is output, as well as the gradient residual fragments obtained by splitting during the backpropagation of the local model. Sensitive business data such as original order details, capacity limits, and core customer information are stored locally on the node throughout the entire process and are not transmitted across nodes or uploaded to the cloud. S2. Construct a point-to-point cross-agent encrypted interaction channel based on the fixed topological adjacency relationship of the actual business transactions between upstream and downstream of the supply chain. Only gradient residual fragments that have been homomorphically encrypted are transmitted between adjacent node agents. Each agent performs cross-verification in the encrypted state on the received residual fragments of adjacent nodes. Combined with the historical residual fluctuation range, adaptively determine and remove abnormal residual data that exceeds the fluctuation threshold. S3. Deploy a global collaborative reasoning aggregation layer in the cloud. Based on the graph attention network with the correction of supply chain supply and demand transmission characteristics, aggregate and calculate the gradient residuals of each node that has passed the verification. Assign differentiated attention weights to neighboring agents according to the historical supply and demand correlation strength, supply and demand transmission delay, and supply and demand elasticity coefficient between nodes. Update the local reasoning model parameters of each edge-side dedicated agent through multiple rounds of iteration, and output supply and demand collaborative reasoning results covering the entire supply chain. S4. Establish a dynamic feedback link for supply and demand deviations. Combining two modes, fixed statistical cycle updates and temporary triggering of sudden anomalies, the supply and demand collaborative reasoning results of the entire link within the corresponding coverage area are pushed to the corresponding node agents with access permissions, triggering a new round of local reasoning and cross-node collaborative update process within the corresponding range, forming a closed loop for supply and demand deviation correction.

[0008] Furthermore, the generation process of gradient residual segments in step S1 is as follows: Each dedicated agent uses local multi-period capacity, inventory, orders, logistics, and external disturbance data as model input, runs the supply and demand time series inference model, and outputs the local initial supply and demand forecast for the corresponding period. The initial forecast is compared with the actual historical supply and demand value for the corresponding period to calculate the local model loss value. First, the maximum value boundary of a single residual segment is determined based on the gradient global sensitivity constraint formula. The gradient global sensitivity calculation formula is: ;in For gradient global sensitivity, and For adjacent training datasets that differ only by a single business record, This is the gradient vector of the model's loss function with respect to the parameter θ. Represents the 2-norm of a vector; Subsequently, a perturbation mechanism satisfying the localized differential privacy budget constraint is used to split the complete parameter gradient generated by the model's backpropagation into equal-length slices, ensuring that the L2 norm of each slice does not exceed [the specified value]. One-eighth of the data is used, and each slice contains only a portion of the non-sensitive parameter updates. It is impossible to reverse the original supply and demand data, core capacity, customer information and other sensitive features of the node through a single slice. The residual fragments that correspond to the cross-node supply and demand correlation features are selected as the only transmission data for cross-node interaction. This prevents the original sensitive information from flowing out of the node from the data source, while ensuring that the residual fragments contain enough effective cross-node collaborative information.

[0009] Furthermore, the residual cross-validation process in step S2 employs a homomorphic encryption scheme that supports floating-point ciphertext operations. Each agent collects all encrypted residual fragments transmitted from directly adjacent upstream nodes and performs residual summation in the ciphertext state without decryption to obtain the predicted total supply fluctuation value of adjacent upstream nodes. This is then combined with locally stored upstream total supply fluctuation data from the same period over nearly 36 cycles to construct an adaptive anomaly detection threshold. The threshold calculation formula is as follows: ;in This is the threshold for anomaly detection. This is the arithmetic mean of the total supply fluctuations of adjacent upstream suppliers over the past 36 periods. This represents the standard deviation of the total supply fluctuation value for the corresponding period. The total supply fluctuation forecast obtained by summing the encrypted data is compared with a threshold. If the deviation exceeds the threshold range, the residual fragments transmitted by each upstream node are individually verified in encrypted state. The source nodes whose residual values ​​exceed the reasonable range are marked, and the abnormal residual data submitted by the corresponding nodes are removed. The removed abnormal residuals do not participate in the subsequent global aggregation calculation. At the same time, the abnormal submission record is stored in the node credit file. Temporary interaction restrictions are set for nodes that continuously submit abnormal residuals. After the node completes local data verification, the permissions are restored to avoid the misreporting data or malicious poisoning data of a single node from interfering with the accuracy of the whole-link collaborative reasoning results.

[0010] Furthermore, the calculation process of differentiated attention weights in step S3 introduces a node historical supply and demand correlation correction factor, a supply and demand elasticity coefficient, and a transmission delay attenuation term. The weight calculation formula is as follows: ;in, For nodes For adjacent nodes The assigned attention weights , They are nodes , The local residual eigenvectors, To share the feature transformation matrix, This represents the weight vector of a single-layer feedforward neural network. This represents a vector concatenation operation. For nodes The set of all adjacent nodes, For nodes and The historical supply and demand correlation correction factor between the two nodes is calculated by weighting three indicators: order fulfillment matching rate, supply and demand response time, and logistics delivery timeliness rate over the past 12 statistical periods. The correction factor ranges from 0 to 1, and the higher the degree of business binding between the nodes, the larger the value.

[0011] Furthermore, the end-to-end supply and demand collaborative reasoning results output in step S3 are distributed in a targeted manner using a three-level permission control mechanism. The results include quantitative indicators for each node across four dimensions: predicted production capacity, safety stock level, dynamic order quota, and logistics capacity reservation for multiple future statistical periods. Level 1 permissions correspond to the node itself, allowing it to view its own end-to-end reasoning indicators and the supply-demand deviation tracing link. Level 2 permissions correspond to directly upstream and downstream stable cooperative nodes, allowing them to view only the supply-demand matching volume and delivery cycle indicators for related businesses, but not core sensitive data such as node cost structure, customer details, and production capacity limits. Level 3 permissions correspond to the overall supply chain supervision entity, allowing them to view aggregated indicators such as the total supply-demand gap, total capacity surplus, and overall logistics load, but not specific business details for individual nodes. All reasoning results are transmitted encrypted, and only node agents with the corresponding permissions can decrypt and view the results within their authorized scope. This ensures the accuracy of end-to-end supply and demand collaborative reasoning while further reducing the risk of leakage of sensitive business data of various supply chain participants during the collaboration process, eliminating data security concerns for each node participating in supply and demand collaboration.

[0012] Furthermore, the supply-demand deviation dynamic feedback link in step S4 supports an update mode that combines fixed period and temporary triggering. The fixed period update adaptively configures the update frequency according to the industry characteristics of the node and periodically performs collaborative reasoning across the entire link. When any node agent detects significant fluctuations in local supply and demand, including scenarios such as sudden large orders, core capacity failures, trunk logistics disruptions, and regional policy controls, and the fluctuation exceeds the steady-state fluctuation threshold calculated based on the autoregressive moving average model, the node can proactively send a temporary collaborative request to the global collaborative reasoning aggregation layer. The request includes key information such as the location of the fluctuation, the initial impact range, and the deviation magnitude. Upon receiving the request, the aggregation layer can activate related neighboring node agents related to the fluctuation's impact range without waiting for a fixed statistical period, completing rapid collaborative reasoning on a small scale. This allows for dynamic correction of local supply and demand deviations within a short time, improving the response speed of the collaborative reasoning method to sudden abnormal events in the supply chain and preventing the amplification of abnormalities along the supply chain, thus avoiding a bullwhip effect.

[0013] Furthermore, in the small-scale rapid collaborative reasoning process, the range of nodes participating in the collaboration is determined by the deviation amplitude carried by the temporary collaboration request, and the formula for calculating the node influence radius is: ;in This refers to the adjacency order used to trace upstream and downstream along the supply chain topology path from the fluctuating nodes. This represents the absolute magnitude of the supply-demand discrepancy detected in this monitoring. The standard deviation of supply and demand fluctuations under steady-state operation of the node. The maximum value is no more than 3, covering all associated nodes affected by deviation propagation; The collaborative process completes residual encryption interaction, cross-validation, and parameter update calculation only within a limited number of nodes. No unrelated nodes need to participate in communication and calculation. It adopts a lightweight residual interaction protocol, omitting redundant global verification steps in the full-link aggregation process. Participating nodes can directly update local model parameters after completing adjacency residual verification, without waiting for the cloud global aggregation to be completed. While ensuring that the accuracy of local supply and demand deviation correction meets business requirements, it significantly reduces cross-node communication overhead and local computing load of each node in the multi-agent collaborative process, and avoids frequent full-link collaboration consuming too much node computing resources and network bandwidth resources.

[0014] Furthermore, when a new participating entity node is added to the supply chain network, the system first matches the dedicated intelligent agent deployment package for the corresponding business category based on the main business attributes and partner registration information of the new node, and completes the parameter initialization of the local supply and demand time series inference model. The initialization parameters use the average model parameters of the nodes already connected in the same category as the benchmark, so there is no need to train the model from scratch. Once a new node is deployed, it only needs to transmit the homomorphically encrypted initial gradient residual fragment to its registered direct business neighbor nodes to officially join the collaborative inference network. The global collaborative inference aggregation layer does not need to interrupt the existing collaborative process or re-execute the full network model training when a new node is connected. It only needs to automatically include the new node in the adjacency weight calculation range in subsequent iterations, and gradually update the correlation correction factor, elasticity coefficient and propagation delay parameter according to the actual transaction performance data of the new node and its neighboring nodes to complete the progressive collaborative adaptation of the model parameters. This enables the seamless and rapid access of new nodes, reducing the system adaptation cost and business interruption risk when the supply chain network is dynamically expanded.

[0015] Furthermore, a multi-agent collaborative reasoning system for supply chain supply and demand coordination includes edge agent modules deployed locally on each participating entity, a global collaborative aggregation module deployed in the cloud, a cross-node encrypted communication module, and a full-link dynamic feedback control module. The edge agent module is equipped with a trusted execution environment, which is used to encrypt and store the original business data of the node in a hardware-isolated environment, load a lightweight supply and demand time-series inference model adapted to the node's business type, complete localized supply and demand feature inference based on local multi-dimensional business data, output local multi-period supply and demand prediction initial values, and generate non-sensitive gradient residual fragments that meet differential privacy constraints through the built-in gradient splitting unit, ensuring that the original sensitive data does not flow out of the node's local area throughout the process. The cross-node encrypted communication module is used to build a point-to-point encrypted transmission channel according to the actual business topology of the supply chain, complete the reliable transmission of encrypted residuals between adjacent nodes, support residual cross-verification in encrypted state, and automatically identify and remove abnormal residual data. The global collaborative aggregation module is used to aggregate and calculate the residual data of each node that has passed the verification based on the graph attention network with supply and demand transmission characteristics correction. It assigns differentiated attention weights to neighboring agents by combining the historical business association strength, supply and demand elasticity, and transmission delay between nodes. It completes the parameter update of the local inference model of each edge agent through multiple rounds of iteration. It has a built-in hierarchical permission management subunit to generate supply and demand collaborative inference results covering the entire link and sub-nodes according to the permission level. The dynamic feedback control module is used to determine the scope of collaboration based on temporary collaboration requests sent according to a preset fixed statistical period or node. It then pushes the inference results of the corresponding scope to the relevant nodes within the authorized scope, triggering a new round of collaborative inference process for the corresponding scope, thereby achieving dynamic closed-loop correction of supply and demand deviations.

[0016] Furthermore, the edge agent module incorporates three processing units: differential privacy processing, homomorphic encryption processing, and local anomaly detection. The differential privacy processing unit is equipped with an adaptive noise injection mechanism, which automatically adjusts the noise injection intensity based on the node's local data sensitivity and model training progress, balancing privacy protection strength and inference accuracy, and generating non-sensitive residual fragments that meet gradient sensitivity constraints. The homomorphic encryption processing unit supports batch floating-point ciphertext computation, completing encryption, decryption, and verification operations on multiple sets of residual fragments transmitted in a single transmission at once, reducing the computational overhead of the encryption and decryption process. The local anomaly monitoring unit has built-in rules for identifying various typical supply chain anomalies. It can automatically identify anomalies such as sudden large orders, capacity failures, logistics disruptions, and policy disturbances. It calculates the supply-demand deviation in real time and automatically generates a temporary collaboration request when the deviation exceeds the threshold and uploads it to the global collaboration aggregation module. This can trigger a rapid collaboration response without manual reporting, significantly reducing the operational burden on node maintenance personnel and improving the automation of anomaly response.

[0017] Compared with existing technologies, the beneficial effects of this invention are: This invention achieves zero external transmission of original business information during cross-entity collaboration by employing edge node local inference, transmitting only gradient residual fragments processed with differential privacy perturbation and homomorphic encryption, and retaining sensitive business data locally throughout the process. This effectively eliminates the data security concerns of participating entities in supply and demand collaboration, avoids the model input distortion problem caused by participants concealing or misreporting data due to concerns about data leakage in existing centralized collaboration schemes, and significantly improves the credibility of input data and the final inference accuracy of collaborative inference.

[0018] This invention introduces a graph attention network with modified supply and demand transmission characteristics for global residual aggregation, and combines a dual-trigger feedback mechanism of fixed-cycle full-link collaboration and rapid small-scale collaboration for sudden anomalies. In the global aggregation stage, differentiated weights can be allocated according to the actual business correlation strength, supply and demand elasticity, and transmission delay of nodes, which conforms to the natural transmission law of supply chain supply and demand fluctuations and effectively weakens the interference of the bullwhip effect on the inference results. At the same time, it can quickly complete targeted collaborative correction for local sudden supply and demand fluctuations without triggering collaborative calculations of all nodes in the entire link. This solves the problems of existing point-to-point adjacency collaboration schemes, such as lack of global collaborative perspective, fluctuation transmission amplification, and low efficiency of anomaly response, and balances global inference accuracy and response speed in sudden scenarios.

[0019] The overall architecture of this invention is adapted to the general computing power threshold of edge nodes. It adopts a containerized and lightweight deployment mode, allowing new nodes to be quickly and seamlessly connected without the need to modify existing enterprise hardware. At the same time, it has a built-in hierarchical permission control mechanism, which can be adapted to supply chain scenarios with different characteristics in multiple industries such as fast-moving consumer goods, equipment manufacturing, and fresh food cold chain. It can support the daily collaboration of small closed vertical supply chains as well as adapt to open supply chain networks with multiple entities across regions. It has low barriers to entry, wide applicability, and high industry promotion value. Attached Figure Description

[0020] Figure 1 This is the overall flowchart of multi-agent collaborative reasoning for supply chain supply and demand coordination proposed in this invention; Figure 2 This is a flowchart of the local feature reasoning and residual cross-validation process of this invention; Figure 3 This is a flowchart illustrating the global collaborative aggregation and reasoning result permission distribution process of this invention. Figure 4 This is a flowchart of the dynamic feedback control and temporary rapid collaborative reasoning of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Reference Figures 1 to 4This invention discloses a multi-agent collaborative reasoning method and system for supply chain supply and demand coordination. For edge computing nodes of five types of participants in the supply chain network—raw material suppliers, manufacturers, distributors, terminal retailers, and trunk logistics service providers—dedicated multimodal agents matching the business attributes of each entity are deployed. Each dedicated agent locally loads a lightweight supply and demand temporal reasoning model that integrates temporal convolution and causal attention mechanisms. The model uses depthwise separable convolution to replace the standard convolution structure, and the number of parameters is only one-tenth of that of the traditional LSTM prediction model. It can be adapted to the general computing power threshold of edge nodes and can run stably on existing industrial gateways and local x86 / ARM architecture servers without the need for dedicated high-performance computing hardware.

[0023] Each agent relies on the hardware-level encrypted storage capabilities of its local trusted execution environment to retain multi-dimensional business data such as historical capacity ramp-up coefficients, safety stock turnover days, order fulfillment rates, in-transit logistics trajectories, and regional supply and demand disturbance factors. This enables localized supply and demand feature inference, outputting initial local supply and demand forecasts for the next 1 to 12 statistical periods, as well as gradient residual fragments obtained during the backpropagation of the local model. Throughout the entire process, sensitive business data such as original order details, capacity limits, and core customer information are stored locally on the node, without being transmitted across nodes or uploaded to the cloud. This fundamentally solves the problem of trade secret leakage caused by the external transmission of original data in the existing collaborative model, eliminating enterprises' data security concerns when participating in supply and demand collaboration.

[0024] The technical solution constructs a point-to-point encrypted interaction channel across intelligent agents based on the fixed topological adjacency relationship of actual business transactions in the upstream and downstream of the supply chain. Only gradient residual fragments that have been homomorphically encrypted are transmitted between adjacent node intelligent agents. Each intelligent agent performs cross-verification in the encrypted state on the received residual fragments of adjacent nodes. Combined with the historical residual fluctuation range, it adaptively judges and removes abnormal residual data that exceeds the fluctuation threshold, so as to avoid malicious data poisoning or node misreporting data interfering with the global inference accuracy.

[0025] The technical solution deploys a global collaborative reasoning aggregation layer in the cloud. The aggregation layer adopts a serverless architecture that can automatically scale up and down according to the number of online nodes. Based on a graph attention network that incorporates the characteristics of supply chain transmission, it aggregates and calculates the gradient residuals of each verified node. Differentiated attention weights are assigned to neighboring agents based on the historical supply and demand correlation strength, supply and demand transmission delay, and supply and demand elasticity coefficient between nodes. After multiple rounds of iteration, the local reasoning model parameters of each edge-side dedicated agent are updated, and the supply and demand collaborative reasoning results covering the entire supply chain are output. This solves the problems of traditional global average aggregation methods that do not consider actual business correlations, leading to amplified bullwhip effect and insufficient prediction accuracy.

[0026] The technical solution establishes a dynamic feedback link for supply and demand deviations. Combining two modes—fixed statistical cycle updates and temporary triggering for sudden anomalies—it pushes the full-link supply and demand collaborative reasoning results of the corresponding coverage area to the corresponding node agents with access permissions, triggering a new round of local reasoning and cross-node collaborative update processes within the corresponding range. This forms a closed loop for supply and demand deviation correction, solving the problems of slow response to fixed-cycle updates and high overhead of full-link collaboration in the traditional mode.

[0027] This invention also discloses the specific generation process of gradient residual segments: Each dedicated agent uses local multi-period capacity, inventory, orders, logistics, and external disturbance data as model input, runs a supply and demand time series inference model, and outputs the local initial supply and demand forecast for the corresponding period. The initial forecast is compared with the actual historical supply and demand value for the corresponding period to calculate the local model loss value. To accurately determine the maximum boundary of the residual slice, the technical solution introduces a gradient global sensitivity constraint to determine the maximum value boundary of a single residual segment. The method for calculating the gradient global sensitivity is as follows: ;in This is the global gradient sensitivity, used to characterize the maximum impact of a change in a single business record on the model gradient. Its dimension is consistent with the model parameter gradient, representing a dimensionless parameter change. and For adjacent training datasets that differ only by a single business record, For the model loss function with respect to parameters gradient vector, This represents the 2-norm of a vector.

[0028] When splitting gradients based on this constraint, the technical solution divides the complete parametric gradients generated by backpropagation of the local model into equal-length slices, ensuring that the L2 norm of each slice does not exceed [a certain value]. One-eighth of the slice was then used to add noise to the slice using a Laplace perturbation mechanism that satisfies the localized differential privacy budget constraint, with the Laplace noise scale set to 1 / 8. ,in For differential privacy budgeting, in everyday collaborative scenarios A value of 1 is used to provide strong privacy protection in rapid collaboration scenarios. The value is set to 2 to balance response speed and privacy strength, ensuring that the cumulative privacy loss over the entire lifecycle does not exceed 3, thus meeting the indistinguishability requirement of differential privacy.

[0029] Each slice contains only a portion of non-sensitive parameter updates. It is impossible to reverse the original supply and demand data, core capacity, customer information and other sensitive features of a node through a single slice. The technical solution selects the residual fragments that correspond to the cross-node supply and demand correlation features as the only transmitted data for cross-node interaction. This avoids the problem of excessive sensitive information due to excessively large fragments, and also avoids the loss of effective features required for cross-node collaboration due to excessively small fragments. It achieves a balance between privacy protection and collaboration accuracy from the data source end.

[0030] This invention also discloses the specific implementation process of residual cross-validation: the residual cross-validation process adopts the CKKS homomorphic encryption scheme that supports floating-point ciphertext operations. The scheme polynomial modulus is set to 8192, which can control the time consumption of single batch encryption and decryption to the millisecond level while ensuring that the ciphertext calculation accuracy meets the requirements of residual operations.

[0031] After each agent collects all encrypted residual fragments transmitted from all directly adjacent upstream nodes, it performs residual summation in the ciphertext state without decryption to obtain the total supply fluctuation prediction value of the adjacent upstream nodes. The unit is the supply and demand fluctuation value within a unit period. Then, it combines the total upstream supply fluctuation data of the same period of the past 36 periods stored locally to construct an adaptive anomaly judgment threshold. The threshold calculation method is as follows: ;in The threshold for anomaly detection is set, and its dimensions are consistent with those of supply and demand fluctuations. This is the arithmetic mean of the total supply fluctuations of adjacent upstream suppliers over the past 36 periods. This represents the standard deviation of the total supply fluctuation value for the corresponding period. The technical solution compares the total supply fluctuation forecast obtained by summing the encrypted data with a threshold. If the deviation exceeds the threshold range, it performs single-value verification on the residual segments transmitted by each upstream node in encrypted state, marks the source nodes whose residual values ​​exceed the reasonable range, and removes the abnormal residual data submitted by the corresponding nodes. The removed abnormal residuals do not participate in subsequent global aggregation calculations, and the abnormal submission record is stored in the node's credit file. A 24-hour temporary interaction restriction is set for nodes that submit abnormal residuals three times consecutively. Interaction privileges are restored after the node uploads a data verification report.

[0032] The entire verification process is completed in encrypted form, eliminating the need to decrypt the residual content transmitted by adjacent nodes. This avoids information leakage during the verification process and effectively prevents misreporting data or maliciously poisoned data from a single node from interfering with the global inference results.

[0033] This invention also discloses a specific method for calculating differentiated attention weights: When performing residual aggregation, the global collaborative reasoning aggregation layer considers the shortcomings of traditional graph attention networks that only allocate weights based on feature similarity and do not conform to the actual business transmission patterns of the supply chain. It introduces a node historical supply and demand correlation correction factor, a supply and demand elasticity coefficient, and a transmission delay attenuation term. Based on the improved graph attention network, it allocates weights to adjacent nodes. The weight calculation formula is as follows: ;in, For nodes For adjacent nodes The assigned attention weights , They are nodes , The local residual eigenvectors, To share the feature transformation matrix, This represents the weight vector of a single-layer feedforward neural network. This represents a vector concatenation operation. For nodes The set of all adjacent nodes, For nodes and The historical supply and demand correlation correction factor between the two nodes is calculated by weighting three indicators: order fulfillment matching rate, supply and demand response time, and logistics delivery timeliness rate over the past 12 statistical periods. The correction factor ranges from 0 to 1, and the higher the degree of business binding between the nodes, the larger the value.

[0034] The weights assigned based on this method can strengthen the influence of nodes with strong supply and demand correlation on the inference results, weaken the interference of residual data from nodes with no direct business dealings or occasional transactions, ensure that the global aggregation process conforms to the actual business rules of the bullwhip effect transmission in the supply chain, and avoid irrelevant node data reducing the business adaptability of the inference results.

[0035] This invention also discloses a three-level permission control and distribution mechanism for collaborative reasoning results: After the collaborative reasoning results of the entire supply and demand chain are generated, a three-level permission control mechanism is used for targeted distribution. The results include quantitative indicators in four dimensions for each node, namely, the predicted production capacity, safety stock level, dynamic order quota, and logistics capacity reservation for multiple statistical periods in the future. The first-level permission corresponds to the node itself, which can view its own full-dimensional reasoning indicators and the supply and demand deviation traceability chain, so that the node can adjust its production and inventory plans in advance. The second-level permission corresponds to the directly upstream and downstream stable cooperative nodes, which can only view the supply and demand matching volume and delivery cycle dimension indicators corresponding to the related business, and cannot view core sensitive data such as node cost structure, customer details, and production capacity limit. Level 3 access corresponds to the overall supply chain supervision entity. It can only view aggregated indicators such as the total supply and demand gap, total capacity surplus, and overall logistics load, and cannot view the specific business details of a single node.

[0036] All inference results are asymmetrically encrypted using the recipient's public key. Only authorized node agents can decrypt and view the results within their authorized scope using their locally stored private key. The transmission process uses the TLS 1.3 encryption protocol to prevent eavesdropping. While ensuring the accuracy of inference in the supply and demand collaboration across the entire chain, this further reduces the risk of leakage of sensitive business data of various participants in the supply chain during the collaboration process, and eliminates data security concerns of each node participating in supply and demand collaboration.

[0037] This invention also discloses a dual-trigger update mode for the dynamic feedback link of supply and demand deviation: the dynamic feedback link of supply and demand deviation supports an update mode that combines fixed period and temporary trigger. The fixed period update adaptively configures the update frequency according to the characteristics of the industry to which the node belongs. For the fast-moving consumer goods industry, which has large demand fluctuations, it is set to weekly full-link collaboration, and for the equipment manufacturing industry, which has long production cycles, it is set to monthly full-link collaboration, so as to avoid resource waste or response delay caused by uniform update frequency.

[0038] When any node agent detects significant fluctuations in local supply and demand, including scenarios such as sudden large orders, core capacity failures, trunk logistics disruptions, and regional policy controls, and the fluctuation exceeds the steady-state fluctuation threshold calculated based on the autoregressive moving average model, the node can proactively send a temporary collaborative request to the global collaborative reasoning aggregation layer. The request includes key information such as the location of the fluctuation, the initial impact range, and the deviation magnitude. Upon receiving the request, the aggregation layer can activate related neighboring node agents related to the fluctuation's impact range without waiting for a fixed statistical period, completing rapid collaborative reasoning on a small scale. This allows for dynamic correction of local supply and demand deviations within a short time, improving the response speed of the collaborative reasoning method to sudden abnormal events in the supply chain and preventing the amplification of abnormalities along the supply chain, thus avoiding a bullwhip effect.

[0039] This invention also discloses a node range determination and lightweight interaction mechanism for small-scale rapid collaboration: During small-scale rapid collaborative reasoning, the range of nodes participating in the collaboration is determined by the deviation amplitude carried by the temporary collaboration request, and the node influence radius is calculated as follows: ;in Let Δ be the adjacency order tracing upwards and downwards from the fluctuation node along the supply chain topology path, a dimensionless positive integer. Let Δ be the absolute magnitude of the supply-demand deviation monitored in this instance, with the dimension being supply and demand per unit period. This represents the standard deviation of supply and demand fluctuations under steady-state operation of the node, with dimensions consistent with Δ. The maximum value is no more than 3, because the transmission effect of supply and demand fluctuations in the supply chain has decayed to a negligible range after exceeding the third order of adjacency, so there is no need to wake up nodes at a greater distance.

[0040] The collaborative process completes residual encryption interaction, cross-validation, and parameter update calculation only within a limited number of nodes. No unrelated nodes need to participate in communication and calculation. It adopts a lightweight residual interaction protocol, omitting the redundant global verification steps in the full-link aggregation process. After participating nodes complete the adjacency residual verification, they can directly update the local model parameters without waiting for the cloud global aggregation to be completed. Under the premise of ensuring that the accuracy of local supply and demand deviation correction meets business requirements, the cross-node communication overhead can be reduced to less than one-tenth of that of full-link collaboration, avoiding the excessive consumption of node computing resources and network bandwidth resources by frequent full-link collaboration.

[0041] This invention also discloses an adaptation process for seamless and rapid access of new nodes: When a new participating entity node is added to the supply chain network, the system first matches a dedicated intelligent agent containerized deployment package for the corresponding business category based on the new node's main business attributes and partner registration information. It supports one-click deployment without complex configuration. After deployment, it automatically initializes the parameters of the local supply and demand time-series inference model. The initialization parameters use the average model parameters of nodes already connected in the same category as the benchmark, eliminating the need for new nodes to upload historical operating data to train the model from scratch.

[0042] After a new node completes its initial deployment, it only needs to transmit the homomorphically encrypted initial gradient residual fragment to its registered direct business neighbor nodes to officially join the collaborative inference network. The global collaborative inference aggregation layer does not need to interrupt the existing collaborative process or re-execute the full network model training when a new node joins. It only needs to automatically include the new node in the neighbor weight calculation range in subsequent iterations and gradually update the correlation correction factor, elasticity coefficient and propagation delay parameter according to the actual transaction performance data of the new node and its neighboring nodes to complete the progressive collaborative adaptation of the model parameters. Generally, it can achieve inference accuracy comparable to mature nodes after 3 statistical cycles of iteration, realizing seamless and rapid access of new nodes and reducing the system adaptation cost and business interruption risk when the supply chain network is dynamically expanded.

[0043] This invention also discloses the overall architecture of a multi-agent collaborative reasoning system for supply chain supply and demand coordination. The system includes edge agent modules deployed locally on each participating entity, a global collaborative aggregation module deployed in the cloud, a cross-node encrypted communication module, and a full-link dynamic feedback control module.

[0044] The edge agent module is delivered in the form of a containerized image, supporting operation on edge devices with multiple architectures such as x86 and ARM. It is equipped with a trusted execution environment, which is used to encrypt and store the original business data of the node in a hardware-isolated environment, load a lightweight supply and demand time-series inference model adapted to the node's business type, complete localized supply and demand feature inference based on local multi-dimensional business data, output local multi-period supply and demand prediction initial values, and generate non-sensitive gradient residual fragments that meet differential privacy constraints through the built-in gradient splitting unit, ensuring that the original sensitive data does not flow out of the node's local area throughout the process.

[0045] The cross-node encrypted communication module adopts a P2P distributed networking architecture, eliminating the need for a central node to forward traffic. It is used to build point-to-point encrypted transmission channels according to the actual business topology of the supply chain, complete the reliable transmission of encrypted residuals between adjacent nodes, support residual cross-verification in ciphertext state, automatically identify and remove abnormal residual data, and control end-to-end communication latency to the level of hundreds of milliseconds.

[0046] The global collaborative aggregation module is deployed on public cloud or industry-specific cloud. It is used to aggregate and calculate the residual data of each node that has passed the verification based on the graph attention network with supply and demand transmission characteristics correction. It assigns differentiated attention weights to neighboring agents by combining the historical business correlation strength, supply and demand elasticity, and transmission delay between nodes. It completes the parameter update of the local inference model of each edge agent through multiple rounds of iteration. It has a built-in hierarchical permission management subunit to generate supply and demand collaborative inference results covering the entire link and sub-nodes according to the permission level.

[0047] The dynamic feedback control module has a built-in rule engine, which is used to determine the scope of collaboration according to the temporary collaboration requests sent by preset fixed statistical periods or nodes. The inference results of the corresponding scope are pushed to the relevant nodes within the authorized scope, triggering a new round of collaborative inference process for the corresponding scope, and realizing dynamic closed-loop correction of supply and demand deviation.

[0048] This invention also discloses the internal unit configuration of the edge agent module: the edge agent module has three types of processing units built in: differential privacy processing, homomorphic encryption processing and local anomaly detection. The differential privacy processing unit is configured with an adaptive noise injection mechanism, which automatically adjusts the noise injection intensity according to the node's local data sensitivity and model training progress. When the parameters fluctuate greatly in the early stage of model training, the noise intensity is appropriately increased to strengthen privacy protection. After the model converges, the noise intensity is appropriately reduced to improve inference accuracy, thus balancing the privacy protection intensity and inference accuracy and generating non-sensitive residual fragments that meet the gradient sensitivity constraints. The homomorphic encryption processing unit integrates a CPU hardware acceleration instruction set, performs instruction-level optimization of AES encryption operations, supports batch floating-point ciphertext calculation, and completes encryption, decryption and verification operations on multiple sets of residual fragments transmitted in a single transmission at one time. The encryption and decryption speed is more than 5 times faster than the native implementation, reducing the computing power overhead of the encryption and decryption process. The local anomaly monitoring unit has a built-in rule library for identifying typical supply chain anomalies across multiple industries. It can automatically identify anomalies such as sudden large orders, production capacity failures, logistics disruptions, and policy disturbances. It calculates the supply-demand deviation in real time and automatically generates a temporary collaboration request when the deviation exceeds the threshold and uploads it to the global collaboration aggregation module. This triggers a rapid collaboration response without manual reporting, significantly reducing the operational burden on node maintenance personnel and improving the automation of anomaly response.

[0049] Reference Figure 1 This flowchart illustrates the overall architecture of a multi-agent collaborative reasoning method for supply chain supply and demand coordination. The system first deploys dedicated agents at the edge nodes of each participant in the supply chain network. These agents utilize locally stored historical data to perform supply and demand time-series feature inference, outputting gradient residual fragments while ensuring the original data remains within its domain. Subsequently, the residual fragments are transmitted through an encrypted interaction channel constructed according to the supply chain topology, and cross-validation between adjacent nodes is performed to eliminate abnormal data. Next, the global collaborative reasoning aggregation layer in the cloud uses a graph attention network to aggregate and calculate the validated residuals, iteratively updating the model parameters of each edge node, and finally generating the end-to-end supply and demand collaborative reasoning result. The system establishes a dynamic feedback loop, periodically or on-demand pushing the results to the corresponding nodes, driving a new round of collaborative iteration.

[0050] Reference Figure 2 This flowchart details the data protection mechanism for local inference at the edge and the residual verification process between nodes. A dedicated agent inputs local multi-period data on production capacity, inventory, etc., to run the inference model and calculate the loss between the initial local prediction and the actual historical values. To prevent the leakage of sensitive information, the system introduces a differential privacy perturbation mechanism to slice the model parameter gradients into equal-length slices, selecting only non-sensitive gradient residual fragments for homomorphic encryption transmission. At the receiving end of adjacent nodes, the system directly performs residual summation in encrypted form and compares the total supply fluctuation prediction value with a confidence interval constructed based on historical data from the same period. If a deviation exceeds the limit, a single-value verification procedure is immediately initiated to accurately locate and remove residual data from the abnormal source node, ensuring the quality and security of data participating in subsequent global calculations.

[0051] Reference Figure 3This flowchart illustrates the working logic of the cloud-based global collaborative reasoning aggregation layer and the secure distribution mechanism of the final prediction results. After acquiring the residual data verified by each node, the aggregation layer introduces a historical supply-demand correlation correction factor calculated based on indicators such as order fulfillment and response time, accurately assigning differentiated attention weights to neighboring agents. After completing residual aggregation based on the graph attention network, the system iteratively updates the local inference model parameters of each node, thereby generating quantitative indicator results covering four dimensions: predicted production capacity, safety stock level, dynamic order quota, and logistics capacity reservation. In the distribution stage, the system strictly implements a hierarchical permission control mechanism, pushing the indicator prediction data of a single node only to itself and directly upstream and downstream agents with stable cooperative relationships.

[0052] Reference Figure 4 This flowchart illustrates the system's dual-track dynamic feedback control mode and small-scale anomaly emergency response mechanism. During normal operation, the system periodically performs collaborative inference updates covering the entire supply chain according to preset weekly or monthly statistical cycles. When the anomaly monitoring unit within the edge intelligence body detects a drastic fluctuation in local supply and demand exceeding the steady-state threshold, the mechanism switches. The node proactively sends a temporary collaboration request to the cloud, and the system then precisely traces two adjacent nodes upwards and downwards along the supply chain topology, quickly defining a limited scope of participation. Within this small scope, residual encryption interaction, cross-validation, and parameter updates are urgently performed, enabling the supply chain to respond quickly to localized sudden disturbances and avoiding resource waste and latency caused by recalculating the entire network.

[0053] The above are merely preferred embodiments 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 multi-agent collaborative reasoning method for supply chain supply and demand coordination, characterized in that, Includes the following steps: S1. Deploy corresponding dedicated intelligent agents for the edge nodes of each participating entity in the supply chain network. Each dedicated intelligent agent loads the supply and demand time series reasoning model locally, completes local supply and demand feature reasoning based on the historical capacity, inventory and order data stored locally on the node, and outputs the local initial value of supply and demand forecast and the gradient residual fragment generated during the local model training process. The original supply and demand data are kept locally on the node and not output to the outside. S2. Construct a cross-agent encrypted interaction channel according to the fixed topological adjacency relationship between upstream and downstream of the supply chain. Only gradient residual fragments that have been homomorphically encrypted are transmitted between adjacent node agents. Cross-validation is performed on the received residual fragments of adjacent nodes. Abnormal residual data that exceed the fluctuation threshold are removed by combining the historical residual fluctuation range. S3. Deploy a global collaborative reasoning aggregation layer in the cloud, aggregate and calculate the gradient residuals of each verified node based on the graph attention network, assign differentiated attention weights to neighboring agents according to the strength of the supply and demand correlation between nodes, iteratively update the local reasoning model parameters of each edge-side dedicated agent, and output supply and demand collaborative reasoning results covering the entire supply chain. S4. Establish a dynamic feedback link for supply and demand deviations, and push the results of the supply and demand collaborative reasoning of the entire link to the corresponding node agents according to the preset statistical cycle, triggering a new round of local reasoning and cross-node collaborative update process.

2. The multi-agent collaborative reasoning method for supply chain supply and demand coordination according to claim 1, characterized in that, The specific process of generating the gradient residual segment in step S1 is as follows: Each dedicated intelligent agent uses local continuous multi-period capacity, inventory, order, and logistics data as input to run a supply and demand time series inference model and output local initial values ​​for supply and demand forecasts for multiple future periods. The initial values ​​are then compared with the actual historical supply and demand values ​​for the corresponding periods to calculate the local model loss value. A perturbation mechanism that satisfies the localized differential privacy budget constraint is used to split the complete parameter gradient generated during the model backpropagation into equal-length slices. Each slice contains only a portion of non-sensitive parameter updates. The residual fragments corresponding to the cross-node supply and demand correlation features are selected as the only transmission data for cross-node interaction, thus preventing the original sensitive information from flowing out of the node's local area from the data source.

3. The multi-agent collaborative reasoning method for supply chain supply and demand coordination according to claim 1, characterized in that, The residual cross-validation process in step S2 is as follows: Each agent collects all homomorphically encrypted residual fragments transmitted by directly adjacent upstream nodes. Relying on the homomorphic encryption ciphertext calculation characteristics, it performs residual summation in the ciphertext state without decryption to obtain the total supply fluctuation prediction value of the adjacent upstream nodes. Then, it constructs a confidence interval by combining the upstream total supply fluctuation data of the same period in the past three years stored locally. The total supply fluctuation prediction value obtained by ciphertext summation is compared with the confidence interval. If the deviation between the two exceeds the threshold range corresponding to the preset confidence level, then the residual fragments transmitted by each upstream node are individually verified in the ciphertext state. The source nodes whose residual values ​​exceed the reasonable range are marked, and the abnormal residual data submitted by the corresponding nodes are removed. The removed abnormal residuals do not participate in the subsequent global aggregation calculation.

4. The multi-agent collaborative reasoning method for supply chain supply and demand coordination according to claim 1, characterized in that, The calculation process of differentiated attention weights in step S3 introduces a node historical supply and demand correlation correction factor, and the weight calculation formula is as follows: ;in, For nodes For adjacent nodes The assigned attention weights , They are nodes , The local residual eigenvectors, To share the feature transformation matrix, This represents the weight vector of a single-layer feedforward neural network. This represents a vector concatenation operation. For nodes The set of all adjacent nodes, For nodes and The historical supply and demand correlation correction factor between the two nodes is calculated by weighting three indicators: order fulfillment matching rate, supply and demand response time, and logistics delivery timeliness rate over the past 12 statistical periods. The correction factor ranges from 0 to 1, and the higher the degree of business binding between the nodes, the larger the value.

5. The multi-agent collaborative reasoning method for supply chain supply and demand coordination according to claim 1, characterized in that, The end-to-end supply and demand collaborative reasoning results output in step S3 are distributed in a targeted manner using a hierarchical permission control mechanism. The results include quantitative indicators for each node in four dimensions: predicted production capacity, safety stock level, dynamic order quota, and logistics capacity reservation for multiple statistical periods in the future. The indicators for a single node are only pushed to itself and its direct upstream and downstream stable cooperative intelligent agents, and are not open to unrelated nodes or third parties without permission.

6. The multi-agent collaborative reasoning method for supply chain supply and demand coordination according to claim 1, characterized in that, The supply-demand deviation dynamic feedback link in step S4 supports an update mode that combines fixed period and temporary triggering. Fixed period updates are performed periodically according to preset weekly and monthly statistical cycles to conduct collaborative reasoning across the entire link. When any node agent detects a significant fluctuation in the local supply and demand status, exceeding the steady-state threshold, the node can proactively send a temporary collaborative request to the global collaborative reasoning aggregation layer. Upon receiving the request, the aggregation layer does not need to wait for a fixed statistical period.

7. The multi-agent collaborative reasoning method for supply chain supply and demand coordination according to claim 6, characterized in that, In the aforementioned small-scale rapid collaborative reasoning process, the range of nodes participating in the collaboration is determined by tracing two adjacent nodes upwards and downwards along the supply chain topology path from the node that sends the temporary collaboration request. The collaboration process only completes residual encryption interaction, cross-validation, and parameter update calculation within the determined limited range of nodes, without requiring irrelevant nodes to participate in communication calculations.

8. The multi-agent collaborative reasoning method for supply chain supply and demand coordination according to claim 1, characterized in that, When a new participating node joins the supply chain network, the system deploys a dedicated intelligent agent matching the business type for the new node and completes the initialization of local inference model parameters. The global collaborative inference aggregation layer does not need to re-execute the training process of the entire network model from scratch when a new node joins.

9. A multi-agent collaborative reasoning system for supply chain supply and demand coordination, used to implement the multi-agent collaborative reasoning method according to any one of claims 1-6, characterized in that, This includes edge intelligence modules deployed locally on each participating entity's premises, a global collaborative aggregation module deployed in the cloud, a cross-node encrypted communication module, and a full-link dynamic feedback control module; The edge agent module is used to load a local supply and demand time-series reasoning model adapted to the node's business type. Based on the historical capacity, inventory, order, and logistics data encrypted and stored locally on the node, it completes localized supply and demand feature reasoning, outputs the initial value of local multi-period supply and demand forecast and the gradient residual fragment generated by model training, and retains the node's original supply and demand data throughout the process without outputting it to the outside world. The cross-node encrypted communication module is used to build a point-to-point encrypted channel according to the supply chain business topology, complete the reliable transmission of encrypted residuals between adjacent nodes and the cross-verification of ciphertext, and eliminate abnormal residuals. The global collaborative aggregation module is used to aggregate and calculate the residual data of each node that has passed the verification based on the graph attention network with correlation correction. It assigns differentiated attention weights to neighboring agents by combining the historical business correlation strength between nodes. Through multiple rounds of iteration, it completes the parameter update of the local inference model of each edge agent and outputs the supply and demand collaborative inference results covering the entire link and the hierarchical permissions of each node. The dynamic feedback control module is used to push the inference results of the corresponding range to the relevant nodes within the permission range according to the temporary collaborative request sent by the preset fixed statistical period or node, thereby triggering a new round of collaborative inference process for the corresponding range.

10. The multi-agent collaborative reasoning system for supply chain supply and demand coordination according to claim 9, characterized in that, The edge agent module has three types of processing units built in: differential privacy, homomorphic encryption and local anomaly detection. The differential privacy processing unit is used to slice and decompose the complete gradient of the local model and perturb it with noise to generate non-sensitive gradient residual fragments, thus preventing the leakage of sensitive information from the source. The homomorphic encryption unit is used to encrypt and decrypt the transmission residual. It supports residual calculation in the ciphertext state without decryption, reducing the risk of leakage in the transmission process. The local anomaly monitoring unit is used to collect local supply and demand status in real time. When fluctuations exceed the threshold, it automatically generates a temporary collaborative request and triggers a rapid response.

Images