An ai-driven dynamic debt network collaborative optimization and trusted execution method and system under a supply chain scenario

Abstract

This invention discloses an AI-driven method and system for collaborative optimization and trusted execution of dynamic debt networks in supply chain scenarios, relating to the fields of supply chain finance technology, artificial intelligence, and distributed trusted computing. The method includes: S1 constructing a supply chain debt graph with three layers of dynamic weights, and storing it in fragments according to privacy levels; S2 using a privacy-graded federated graph attention enhancement model to achieve risk transmission prediction, repayment opportunity identification, and network health assessment; S3 generating a set of repayment schemes integrating financial and business objectives through a hierarchical multi-agent alternating optimization algorithm; S4 completing multi-party collaborative authorization using a two-stage enhanced authorization mechanism; S5 achieving atomic repayment based on smart contracts with dual oracles and bidirectional verification; and S6 driving the dynamic evolution of the model through business feedback. This invention achieves accurate prediction, global optimization, and trusted execution of supply chain debt networks, adapts to business characteristics, avoids existing patent risks, and improves repayment efficiency and compliance.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of supply chain finance technology, artificial intelligence, and distributed trusted computing. Specifically, it relates to an AI-driven method and system for collaborative optimization and trusted execution of dynamic debt networks in supply chain scenarios. Through federated graph learning diagnosis, hierarchical multi-agent decision-making, blockchain conditional execution, and closed-loop evolution mechanism, it realizes a method and system for debt network risk prediction, multi-objective optimized repayment, and trusted automated execution. Background Technology

[0002] In the collaboration between upstream and downstream entities in the supply chain, complex and intertwined creditor-debtor relationships are formed among core enterprises, suppliers, distributors, logistics providers, and other entities. Traditional debt settlement relies on manual reconciliation, offline negotiations, and bank transfers, which are characterized by low efficiency, high costs, and hidden risk transmission.

[0003] In existing technologies, some solutions attempt to achieve automated debt repayment through debt graphs and smart contracts (such as the applicant's prior patent CN202511715585X), but the decision-making strategy is singular and lacks predictive ability; other solutions use AI multi-objective optimization of debt restructuring (such as CN1121434799A), but do not combine the closed-loop characteristics of the supply chain debt network with the trusted execution of blockchain; there are also federated learning graph structure learning solutions (such as CN116522988A), which only focus on optimizing model training efficiency and do not address the risk prediction, repayment strategy generation, and execution closed-loop design for debt scenarios.

[0004] The existing technology has the following shortcomings: 1) It does not design a dedicated AI model for the "dynamic, closed-loop, and multi-entity interconnectedness" of supply chain debt, and general optimization algorithms are difficult to adapt to the needs of the scenario; 2) The federated learning training mechanism does not take into account the differences in the privacy level of debt data, and it is difficult to balance training efficiency and privacy protection; 3) Multi-objective optimization does not associate with the core indicators of the supply chain (such as delivery cycle and inventory turnover rate), and the repayment plan is prone to conflict with business operations; 4) AI decision-making and blockchain execution lack a two-way verification mechanism, resulting in insufficient credibility and auditability; 5) The system does not have the ability to dynamically adapt to supply chain scenarios and cannot adjust decision parameters with business fluctuations.

[0005] Therefore, there is an urgent need for a solution that fits the characteristics of supply chain business and achieves intelligent optimization and reliable execution of debt networks through differentiated technological paths, which can both solve the functional defects of existing technologies and circumvent the scope of protection of existing patents. Summary of the Invention

[0006] This invention aims to overcome the shortcomings of existing technologies and provide an AI-driven method and system for collaborative optimization and reliable execution of dynamic debt networks in supply chain scenarios. Through scenario-based technology design and differentiated paths, it achieves more accurate risk prediction, global optimization of repayment strategies, reliable and controllable execution process, and continuous evolution of system capabilities, while avoiding the protection boundaries of existing patents. Technical solution

[0007] In a first aspect, the present invention provides an AI-driven method and system for collaborative optimization and reliable execution of dynamic debt networks in a supply chain scenario, comprising the following steps: S1: Multi-dimensional Debt Data Collection and Dynamic Weight Map Construction in the Supply Chain Through multiple channels such as supply chain management system (SCM) interface, smart contract confirmation certificate, IoT device data collection, and OCR invoice parsing, we obtain structured debt data (amount, payment period, interest rate) and unstructured related data (purchase contract, delivery voucher, quality inspection report, inventory data) digitally signed by the participating parties. Define a dynamic weighting system for the "node-edge" structure of the debt network graph: Node weights are calculated using a three-layer dynamic adjustment algorithm based on real-time cash flow, credit scores, and supply chain performance ratings (such as on-time delivery rate). The weighting factors are dynamically adapted to the business cycle (such as peak purchasing seasons and collection cycles). Edge weights integrate debt amount, payment urgency coefficient, and supply chain dependence (such as core supplier dependence index), and determine the weight ratio through a combination of entropy weighting and analytic hierarchy process. A global dynamic debt graph is constructed, supporting real-time updates of node / edge weights and historical trajectory tracing. The graph data adopts a sharded storage strategy, and access permissions are divided according to the privacy level of the participants.

[0008] S2: Privacy-based Federated Graph Learning Diagnosis and Risk Prediction Deploy federated learning coordination nodes and edge computing nodes, and divide training participation permissions according to the privacy level of debt data (public data, internal data, sensitive data): sensitive data nodes only participate in local model training, internal data nodes participate in encrypted interaction of model parameters, and public data nodes provide globally shared features; Design a federated graph attention enhancement model: taking a dynamic debt graph as input, local nodes extract local features based on graph convolutional networks, and coordinating nodes aggregate model parameters of heterogeneous nodes through an attention mechanism (rather than simple mean aggregation), and complete joint training of the model without leaving the local area with the original data; The trained model outputs three types of supply chain-specific diagnostic results: a) accurate prediction of risk transmission chains, marking the core nodes that trigger systemic defaults and their transmission probabilities; b) dynamic identification of repayment opportunities, combining supply chain business cycles (such as payment windows and procurement settlement periods) to predict the time for closed-loop formation; c) debt network health score, integrating three core indicators: liquidity index, performance guarantee coefficient, and business adaptability.

[0009] S3: Generation of multi-objective debt settlement schemes through hierarchical multi-agent collaboration The debt repayment problem is modeled as a hierarchical collaborative decision-making problem: the upper-level intelligent agent corresponds to the core enterprise in the supply chain and is responsible for setting the global optimization goal; the middle-level intelligent agent corresponds to the node cluster (such as the supplier cluster or the distributor cluster) and is responsible for balancing the interests within the cluster; the lower-level intelligent agent corresponds to the individual participant and is responsible for satisfying local constraints. Define a supply chain-specific composite reward function: including core objectives (maximizing debt network liquidity and minimizing systemic risk), business objectives (ensuring procurement delivery cycles and reducing inventory backlog risks), and constraint objectives (balancing cash flow pressure and optimizing tax compliance). The weight of each objective is dynamically adjusted according to the supply chain's operational status. The hierarchical agent is trained using an alternating optimization algorithm: the upper-layer agent initializes the global target weights, the middle-layer agent generates the cluster-level repayment strategy, and the lower-layer agent provides feedback on the satisfaction of local constraints. Through multiple rounds of alternating iteration, the optimal repayment scheme set is output. Each scheme includes a closed-loop path, repayment amount, execution priority, and business impact assessment report.

[0010] S4: Two-stage enhanced empowerment and explainable decision interaction Phase 1 (Pre-authorization): The settlement plan is transformed into a visual interface from the perspective of supply chain business through an interpretable AI module, showing the impact of the plan on procurement delivery, inventory turnover, and cash flow. Participants complete identity verification (biometrics + digital signature) and submit pre-authorization intentions through mobile terminals. Phase Two (Formal Authorization): The system summarizes the pre-authorization results, makes intelligent adjustments to the disputed plans (such as optimizing the repayment time point and splitting the repayment amount), and pushes the adjusted plan back to the participants to complete the formal authorization through "continuous behavioral biometrics + hardware encrypted signature". The authorization certificate adopts a hierarchical aggregation mechanism: the core enterprise's authorization certificate serves as the primary certificate, and the certificates of other participating parties serve as secondary certificates. After the primary and secondary certificates are verified, a multi-party joint authorization instruction is generated, which includes the authorization trajectory hash and timestamp.

[0011] S5: Conditional Atomic Execution of Blockchain Two-Way Verification Deploy an upgradeable smart contract for supply chain debt settlement on the consortium blockchain. The contract has two embedded oracle interfaces: a business oracle to obtain the status of supply chain business (such as delivery completion certificates and payment receipt information), and a data oracle to obtain dynamic data of the debt network. The server packages the joint authorization instruction, settlement plan, and oracle query request into a transaction and sends it to the blockchain network; the smart contract performs three verifications: verification of the validity of the authorization signature, verification of the consistency between the scheme logic and the supply chain business rules, and verification that the oracle return conditions are met. After verification, the contract updates the debt status atomically, completing the closed-loop repayment; at the same time, the AI ​​decision parameters, authorization evidence, execution results, and oracle data are packaged and uploaded to the blockchain to generate an immutable audit log, supporting supply chain finance regulatory traceability.

[0012] S6: Business Feedback-Driven Dynamic Evolution Loop of the Model After the debt repayment is executed, two types of feedback data are collected: data on changes in the debt network status (liquidity, risk index) and data on supply chain business operations (delivery efficiency, inventory level). These data are then encrypted, desensitized, and enhanced for privacy to serve as training samples. The federated learning coordination node triggers incremental or full training based on the quality of feedback data and the magnitude of business changes: when business fluctuations are small (e.g., fluctuations < 10%), incremental training is performed, updating only local model parameters; when fluctuations are large, full training is initiated to optimize the model weight system and decision logic. Establish a model version management mechanism to record training data, parameter configurations, and decision-making effects for each version, supporting backtracking and rollback to ensure the evolution process is controllable.

[0013] Secondly, the present invention provides a system for implementing the above method, comprising the following modules connected via a communication network: Supply chain data perception and graph construction module: Deployed on the server side, it includes a multi-source data adaptation unit (interfacing with SCM, IoT, and OCR devices), a dynamic weight calculation engine (executing a three-layer dynamic adjustment algorithm), and a privacy-graded graph storage unit, used to execute step S1; Federated Graph Learning Diagnostic Module: Includes a privacy-graded training scheduling unit, a federated graph attention-enhanced model training unit, and a supply chain risk prediction unit, deployed on coordination nodes and edge computing nodes, used to execute step S2; Hierarchical multi-agent decision-making module: including target weight dynamic adjustment unit, hierarchical agent training engine, and multi-scheme evaluation and screening unit, deployed on high-performance computing cluster, used to execute step S3; The two-stage authorization and interaction module includes an explainable AI visualization unit, an enhanced biometric unit (integrating continuous behavior monitoring), and a hierarchical authorization aggregation unit. It consists of a mobile terminal and server coordination sub-module and is used to execute step S4. The blockchain trusted execution module includes a dual oracle gateway, a smart contract deployment and management unit, a two-way verification execution unit, and an audit log generation unit. It interacts with the consortium blockchain network to execute step S5. Business Feedback and Model Evolution Module: Includes a feedback data preprocessing unit (encryption and desensitization + privacy enhancement), a training trigger decision unit, and a model version management unit, used to execute step S6. Beneficial effects

[0014] Scenario-specific adaptation: Focusing on the characteristics of supply chain debt networks, the debt graph weight system and AI model objective function are aligned with supply chain business indicators, solving the problem of insufficient adaptability of general technologies; Significantly effective power avoidance design: Federated learning adopts a "privacy-level training + attention aggregation" mechanism, which differs from the "random sampling + mean aggregation" of CN116522988A; multi-agents adopt hierarchical collaborative decision-making, which differs from the flat agent architecture of existing technologies; blockchain execution introduces dual oracles and two-way verification, which differs from single execution logic; Decision optimization and upgrade: Integrating supply chain business objectives with debt repayment objectives, the repayment plan ensures financial optimization while avoiding conflicts with procurement, delivery and other business operations, thereby improving the feasibility of the plan; Enhanced privacy and security: Differentiated federated training based on data privacy levels balances training efficiency and privacy protection, meeting supply chain data compliance requirements; Trustworthy and auditable: AI decision-making and blockchain execution are verified in two ways, and the entire process logs are recorded on the blockchain, solving the trust problem of "black box decision-making" and adapting to the needs of financial supervision; Dynamic adaptability: The model evolves through business feedback, adjusting decision parameters as supply chain business fluctuates to improve the long-term applicability of the system. Attached Figure Description

[0015] Figure 1 is a schematic diagram of the system architecture of the present invention; Figure 2 is a flowchart of the method of the present invention; Figure 3 is a schematic diagram of hierarchical multi-agent decision-making logic; Figure 4 is a schematic diagram of the interaction between two-stage authorization and blockchain execution. Detailed Implementation

[0016] The following, in conjunction with the accompanying diagrams and a real-world supply chain scenario (core enterprise E, supplier F, G, distributor H, and logistics provider I), details the implementation process of this invention: Data perception and graph construction: E, F, G, H, and I upload accounts receivable and payable data, purchase contracts, and logistics delivery vouchers through the SCM system interface. The OCR module parses key information on the invoices. The dynamic weight calculation engine assigns higher performance rating weights to F (core suppliers) and lower cash flow weights to H (those with longer payment cycles). A dynamic debt graph containing 5 nodes and 8 edges is constructed, divided into core data shards and shared data shards according to privacy levels.

[0017] Federated Graph Learning Diagnosis: Based on the data privacy level, the federated learning coordination node schedules sensitive data nodes of F and G to participate in local training, internal data nodes of H and I to participate in parameter interaction, and public data nodes of E to provide shared features; After training, the federated graph attention enhancement model predicts that if H defaults on its 1.2 million debt to E, there is a 55% probability that this will lead to F's inability to supply goods on time; at the same time, it identifies that the closed loop of E→F→I→E can form the optimal repayment conditions after 15 days (E's repayment window).

[0018] Layered multi-agent decision-making: The upper-layer agent (corresponding to E) sets the core objectives of "maximizing liquidity + ensuring delivery"; the middle-layer agent (supplier cluster) proposes a "split repayment" scheme, and the lower-layer agents (F, I) provide feedback on cash flow constraints; after three rounds of alternating optimization, the following scheme is generated: after 15 days, a closed-loop repayment of E→F→I→E (amounting to 900,000) is executed, and the remaining 300,000 debt is tied to H's subsequent purchase orders for repayment. This scheme improves network liquidity by 35% and does not affect F's supply cycle.

[0019] Two-stage authorization: Each participant views an interpretable interface (showing the impact of repayment on supply and cash flow) via a mobile app and completes pre-authorization; based on F's cash flow pressure feedback, the system fine-tunes the repayment split ratio, and each participant completes formal authorization through "facial recognition + hardware encrypted signature" to generate a joint authorization instruction.

[0020] Blockchain condition execution: 15 days later, the business oracle reports that the E payment has been received, and the data oracle confirms that the closed-loop debt status is valid; after the smart contract verifies the authorized signature and oracle data, the 900,000 debt is atomically settled, and the AI ​​decision parameters, authorization evidence and execution logs are stored on the blockchain simultaneously.

[0021] Model evolution: Data such as a 12% improvement in supply efficiency of F and an 8% optimization in inventory turnover of H after collection and settlement are collected, encrypted, and then used to trigger incremental training. The weight of "business adaptability" in the multi-agent reward function is adjusted to improve the business compatibility of subsequent solutions.

[0022] The scope of protection of this invention is not limited to the above embodiments. All technical solutions that fall within the scope of protection of this invention are included.

Claims

1. A method and system for AI-driven dynamic debt network collaborative optimization and reliable execution in a supply chain scenario, characterized in that, Includes the following steps: S1: Multi-dimensional debt data collection and dynamic weight map construction in the supply chain Obtain structured debt data and unstructured related data digitally signed by participating parties through multiple channels; The node weights adopt a three-layer dynamic adjustment algorithm, which is calculated based on real-time cash flow, credit score, and supply chain performance rating. The weight factors are dynamically adapted with the business cycle. The edge weights integrate debt amount, payment urgency coefficient, and supply chain dependence, and determine the weight ratio by combining entropy weight method and analytic hierarchy process. Construct a global dynamic debt graph, supporting real-time weight updates and historical trajectory tracing; graph data is stored in fragments according to privacy levels. S2: Privacy-based Federated Graph Learning Diagnosis and Risk Prediction Training participation permissions are divided according to the privacy level of debt data. Sensitive data nodes only participate in local training, internal data nodes participate in encrypted parameter interaction, and public data nodes provide globally shared features. Design a federated graph attention enhancement model, in which local nodes extract local features based on graph convolutional networks, and coordinating nodes aggregate model parameters of heterogeneous nodes through an attention mechanism to complete joint training; The model outputs three types of results: accurate prediction of risk transmission chains, dynamic identification of repayment opportunities, and a debt network health score. S3: Generation of multi-objective debt settlement schemes through hierarchical multi-agent collaboration The debt repayment problem is modeled as a hierarchical collaborative decision-making problem, with the upper-level intelligent agent corresponding to the core enterprise, the middle-level intelligent agent corresponding to the node cluster, and the lower-level intelligent agent corresponding to a single participant. Define a composite reward function that includes core objectives, business objectives, and constraint objectives, with the weight of each objective dynamically adjusted according to the operational status of the supply chain; The hierarchical agent is trained using an alternating optimization algorithm, and the output is a set of optimal repayment schemes including closed-loop paths, repayment amounts, execution priorities, and business impact assessment reports. S4: Two-stage enhanced empowerment and explainable decision interaction Phase 1 Pre-authorization: The business impact of the solution is demonstrated through an interpretable AI module, and participating parties submit their pre-authorization intentions via biometrics and digital signatures; The second stage of formal authorization: After the system adjusts the objection scheme, the participating parties complete the authorization through continuous behavioral biometrics + hardware encrypted signature; A hierarchical aggregation mechanism is used to generate a joint authorization instruction from multiple parties, which includes an authorization trajectory hash and a timestamp. S5: Conditional Atomic Execution of Blockchain Two-Way Verification The consortium blockchain deploys scalable smart contracts for supply chain debt repayment, with embedded dual oracle interfaces; The server-side packaged joint authorization instructions, settlement plan, and oracle query request are sent to the blockchain. The smart contract execution authorization signature validity, the consistency between the plan and business rules, and the oracle conditions are verified in three ways. After the verification is passed, the debt status is updated atomically, and the AI ​​decision parameters, authorization evidence, execution results, and oracle data are packaged and uploaded to the blockchain to generate an audit log. S6: Business Feedback-Driven Dynamic Evolution Loop of the Model Data on changes in the status of the debt network and operational data of the supply chain were collected, and then encrypted, desensitized, and enhanced for privacy as training samples. Based on the quality of the feedback data and the magnitude of business changes, trigger incremental training or full training. Establish a model version management mechanism that supports rollback.

2. The method according to claim 1, characterized in that, In step S1, the multiple channels include supply chain management system (SCM) interface, smart contract confirmation certificate, IoT device collection, and OCR invoice parsing; the structured debt data includes amount, payment period, and interest rate; and the unstructured related data includes purchase contracts, delivery vouchers, quality inspection reports, and inventory data.

3. The method according to claim 1, characterized in that, In step S1, the supply chain performance rating includes on-time delivery rate; the supply chain dependence includes core supplier dependence index; and the business cycle includes peak procurement season and payment collection cycle.

4. The method according to claim 1, characterized in that, In step S2, the privacy level of the debt data is divided into public data, internal data, and sensitive data; the debt network health score integrates three core indicators: liquidity index, performance guarantee coefficient, and business adaptability.

5. The method according to claim 1, characterized in that, In step S3, the core objective is to maximize the liquidity of the debt network and minimize systemic risk; the business objective is to ensure procurement delivery cycle and reduce inventory backlog risk; and the constraint objective is to balance cash flow pressure and optimize compliance with taxation.

6. The method according to claim 1, characterized in that, In step S5, the dual-oracle interface includes a business oracle and a data oracle; the business oracle obtains the supply chain business status, including delivery completion certificates and payment receipt information; the data oracle obtains dynamic data of the debt network.

7. The method according to claim 1, characterized in that, In step S6, the threshold for business fluctuation is set to 10%. When the fluctuation is less than 10%, incremental training is performed, and only local parameters of the model are updated. When the fluctuation is greater than or equal to 10%, full training is started to optimize the model weight system and decision logic.

8. A system for implementing the method of any one of claims 1-7, characterized in that, This includes the following modules connected via a communication network: Supply chain data perception and graph construction module: Deployed on the server side, including multi-source data adaptation unit, dynamic weight calculation engine, and privacy-graded graph storage unit, used to execute step S1; Federated Graph Learning Diagnostic Module: Includes a privacy-graded training scheduling unit, a federated graph attention-enhanced model training unit, and a supply chain risk prediction unit, deployed on coordination nodes and edge computing nodes, used to execute step S2; Hierarchical multi-agent decision-making module: including target weight dynamic adjustment unit, hierarchical agent training engine, and multi-scheme evaluation and screening unit, deployed on high-performance computing cluster, used to execute step S3; Two-stage authorization and interaction module: including an explainable AI visualization unit, an enhanced biometric unit, and a hierarchical authorization aggregation unit, composed of a mobile terminal and server coordination sub-module, used to execute step S4; The blockchain trusted execution module includes a dual oracle gateway, a smart contract deployment and management unit, a two-way verification execution unit, and an audit log generation unit. It interacts with the consortium blockchain network to execute step S5. Business Feedback and Model Evolution Module: Includes a feedback data preprocessing unit, a training trigger decision unit, and a model version management unit, used to execute step S6.

9. The system according to claim 8, characterized in that, The enhanced biometric unit integrates continuous behavior monitoring; the feedback data preprocessing unit performs encryption, desensitization, and privacy enhancement processing.

10. The system according to claim 8, characterized in that, The multi-source data adaptation unit interfaces with the supply chain management system (SCM), IoT devices, and OCR devices; the dynamic weight calculation engine executes a three-layer dynamic adjustment algorithm.

Images