Supply chain finance-oriented ai smart contract risk assessment method

Abstract

This invention relates to the field of blockchain and supply chain finance risk control technology, specifically to an AI smart contract risk assessment method for supply chain finance. The method includes: collecting multi-source circulation status data and external intervention signals from the supply chain finance network within a preset time window; fusing time-series features to obtain a node behavior feature sequence; inputting this sequence into a dynamic graph neural network model to output network state trust entropy; and calculating the probability distance between the network state trust entropy and the trust entropy distribution that triggers a cascading blocking event in the smart contract as a critical distance; based on the network state trust entropy and critical distance, calculating the system survival probability index and circulation efficiency index using a dynamic risk hedging model to determine the comprehensive risk assessment state and locate the target node; and triggering a smart contract based on the comprehensive risk assessment state to perform rigid resource recovery or flexible resource compensation on the target node, thereby achieving trust reconstruction. This invention achieves efficient early warning and accurate network-level assessment of cascading blocking events in the capital chain.

Description

Technical Field

[0001] This invention relates to the field of blockchain and supply chain finance risk control technology, specifically to an AI smart contract risk assessment method for supply chain finance. Background Technology

[0002] Existing supply chain finance risk control solutions include smart contract execution mechanisms deployed on blockchain networks; In related technologies, in order to identify node risks in the supply chain finance network, static business data such as financing, settlement, credit granting and repayment in the on-chain ledger can be used for rule judgment; or, single node credit scores, overdue records or preset threshold triggering methods can be used to freeze, liquidate or adjust the credit limit of abnormal nodes. However, in supply chain finance networks with multi-level node linkages, external public opinion shocks, and abnormal transaction interventions, the former method has a weak ability to characterize temporal changes and network topology evolution, while the latter method has a low ability to warn of cascading failure risks. Therefore, the supply chain finance smart contract risk assessment methods in related technologies are difficult to balance the foresight of risk identification, the accuracy of network-level assessment, and the adaptability of contract response. Summary of the Invention

[0003] The purpose of this invention is to provide an AI-powered smart contract risk assessment method for supply chain finance, addressing the following technical issues: This enables supply chain finance systems to complete risk warnings and adaptive responses before the cascading collapse of the funding chain actually occurs, achieving earlier detection of sudden credit contractions in the network and accurate identification of risk diffusion patterns. In turn, the assessment results are mapped into actionable on-chain actions, achieving precise contraction of abnormal nodes and targeted support for repairable nodes.

[0004] The objective of this invention can be achieved through the following technical solutions: An AI-powered smart contract risk assessment method for supply chain finance, executed on a blockchain network with deployed smart contracts, includes: Step 1: Collect multi-source circulation status data and external intervention signals, fuse them according to time series characteristics to obtain node behavior feature sequences, and combine them with the direction of fund flow to form dynamic graph data; Step 2: Input the node behavior feature sequence into the hidden state vector of the dynamic graph neural network model to construct a real-time credit evolution graph and calculate the network state trust entropy; extract the historical critical trust entropy distribution in the stress test scenario library and calculate the probability distance between it and the network state trust entropy as the critical distance; Step 3: Based on the network state trust entropy and critical distance, calculate the system survival probability index and flow efficiency index through the dynamic risk hedging model to determine the comprehensive risk assessment state; extract the absolute value of the first difference of the node behavior feature sequence at adjacent time points to calculate the variation weight, and locate the target node; Step 4: When the overall risk assessment status is lower than the preset safety threshold, a rigid resource reclamation operation is performed on the target node; when it is not lower than the safety threshold, a flexible resource compensation operation is performed.

[0005] Optionally, the specific process of step 1 is as follows: Step 1.1: Collect historical interaction records, current resource holdings, and time delay characteristic data of each node in the supply chain finance network to form multi-source circulation status data; Step 1.2: Collect external intervention signals for the supply chain finance network. The external intervention signals include transaction instructions with a transaction frequency higher than or equal to a preset frequency threshold and a difference between the number of buy and sell instructions greater than or equal to a preset ratio, as well as text data with a negative sentiment score exceeding or equal to a preset threshold. Step 1.3: Perform distortion feature identification and filtering on the multi-source flow state data, and perform tensor concatenation between the filtered multi-source flow state data and the external intervention signal to generate a node behavior feature sequence.

[0006] Optionally, in step 2, the process of calculating the network state trust entropy is as follows: The node behavior feature sequence is input into the spatiotemporal encoder contained in the dynamic graph neural network model, and the hidden state vector of each node is output. Based on the hidden state vector, the dynamic connection weight between any two nodes is calculated to construct a real-time credit evolution graph. Based on the node degree distribution in the real-time credit evolution graph and the coefficient of variation of the dynamic connection weights, the network state trust entropy of the supply chain finance network is calculated.

[0007] Optionally, in step 2, the process of calculating the critical distance is as follows: Based on a stress test scenario library containing historical extreme fluctuation data and corresponding network state distributions, the historical critical trust entropy distribution that triggers the smart contract cascading blocking event is extracted. Calculate the relative entropy between the network state trust entropy and the historical critical trust entropy distribution; Based on the relative entropy and the preset risk propagation attenuation coefficient, the probability distance between the current state and the triggering of the smart contract cascading blocking event is calculated, which is used as the critical distance.

[0008] Optionally, the specific process of step 3 is as follows: the dynamic risk hedging model includes a survival probability assessment branch and a turnover efficiency assessment branch; Step 3.1: Input the network state trust entropy into the survival probability assessment branch to calculate the system survival probability index; Step 3.2: Input the critical distance into the circulation efficiency evaluation branch to calculate the circulation efficiency index; Step 3.3: Weight and fuse the system survival probability index and the circulation efficiency index to obtain the comprehensive risk assessment status.

[0009] Optionally, in step 4, the rigid resource reclamation operation specifically includes: interrupting the resource transfer permission of the target node and triggering a preset forced liquidation logic; the flexible resource compensation operation specifically includes: suspending the preset forced liquidation logic and injecting a preset grace period parameter and an additional overdraft resource limit into the target node.

[0010] Optionally, the network state trust entropy is used to characterize the uncertainty of the underlying node's data transparency and willingness to interact with data.

[0011] Optionally, in step 1.3, the distortion feature identification and filtering is used to identify defensive data forgery features in node data; wherein, the defensive data forgery features include: the feature that the distribution of invoice amount deviates from the preset normal range, the feature that the update frequency of liability data is lower than the preset time threshold, and the feature that there are no historical interaction records in the data flow path between nodes.

[0012] Optionally, in step 1.2, the external intervention signal is intervention data with a preset selling frequency characteristic injected by the quantitative trading program.

[0013] Optionally, step 4 may include the following: Collect feedback status data of the supply chain finance network after executing the rigid resource recovery operation or the flexible resource compensation operation; The trust repair gradient is calculated based on the feedback state data; the weight parameters of the dynamic graph neural network model are iteratively updated using the trust repair gradient to achieve model evolution.

[0014] The beneficial effects of this invention are: 1. This invention generates a node behavior feature sequence by collecting multi-source circulation status data and external intervention signals within a preset time window and performing feature fusion. This mechanism breaks through the limitations of traditional methods that rely solely on static ledger data, and can effectively capture early fluctuations caused by external public opinion shocks and abnormal transaction interventions, significantly improving the foresight of risk perception. 2. This invention utilizes a dynamic graph neural network model to perform network topology evolution analysis, outputs network state trust entropy, and calculates critical distance. This method overcomes the isolation defect of single-node rating, can accurately quantify the structural uncertainty under multi-level linkage of the supply chain, and achieves efficient early warning and accurate network-level assessment of cascading blockage events in the capital chain. 3. This invention calculates the comprehensive risk assessment status based on survival probability and circulation efficiency, and triggers smart contracts to adaptively execute rigid resource recovery or flexible resource compensation; this breaks the traditional one-size-fits-all freezing and liquidation mechanism, which can not only block the risk spread of high-risk nodes, but also provide liquidity support for repairable nodes, and achieve precise adaptive response. 4. This invention identifies and filters distortion features in multi-source circulation status data, accurately intercepting defensive data forgery features such as deviations in invoice amount distribution, delayed liability updates, and lack of historical interaction circulation paths; this design effectively prevents nodes from concealing real default risks by modifying accounting data, and significantly improves the reliability of subsequent risk assessment models; 5. This invention collects network feedback status data after smart contract execution operations and iteratively updates the dynamic graph neural network model based on the calculated trust repair gradient. This closed-loop mechanism gives the system the ability to continuously correct the evaluation logic based on the actual handling effect, effectively improving the long-term robustness of the model in dealing with complex and ever-changing market environments. Attached Figure Description

[0015] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating the AI ​​smart contract risk assessment method for supply chain finance provided in the embodiments of this application. Detailed Implementation

[0016] 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.

[0017] Please see Figure 1 An AI-powered smart contract risk assessment method for supply chain finance, executed on a blockchain network with deployed smart contracts, includes: Step 1: Collect multi-source circulation status data and external intervention signals, fuse them according to time series characteristics to obtain node behavior feature sequences, and combine them with the direction of fund flow to form dynamic graph data; Step 2: Input the node behavior feature sequence into the hidden state vector of the dynamic graph neural network model, construct the real-time credit evolution graph and solve the network state trust entropy; extract the historical critical trust entropy distribution in the stress test scenario library, and calculate the probability distance between it and the network state trust entropy as the critical distance. Step 3: Based on the network state trust entropy and critical distance, calculate the system survival probability index and circulation efficiency index through the dynamic risk hedging model to determine the comprehensive risk assessment state; extract the absolute value of the first difference of the node behavior feature sequence at adjacent time points to calculate the variation weight, and locate the target node; Step 4: When the overall risk assessment status is lower than the preset safety threshold, a rigid resource reclamation operation is performed on the target node; when it is not lower than the safety threshold, a flexible resource compensation operation is performed.

[0018] This embodiment provides an AI smart contract risk assessment mechanism for supply chain finance. Specifically, the mechanism is deployed in a consortium blockchain network jointly participated in by core manufacturing enterprises, first-tier suppliers, second-tier suppliers, factoring institutions, and settlement nodes, and is used to uniformly control on-chain factoring financing, accounts receivable transfer, credit grace period, and forced liquidation. For ease of explanation, the same main scenario will be used throughout: A new energy vehicle manufacturer, as the core enterprise, opens up on-chain financing channels to battery module manufacturers, chip packaging manufacturers, copper foil suppliers and downstream processing nodes. The system uses a 6-hour preset time window to continuously assess whether the supply chain network is approaching a state of cascading financial chain breakage. Within each time window, the system collects multi-source circulation status data and external intervention signals from the blockchain ledger, off-chain risk control platform, and public opinion interface; the multi-source circulation status data includes at least the historical interaction records of each node, the current resource holding amount, and latency characteristics; External intervention signals include at least abnormal trading instructions and negative text signals; the system aligns multiple types of data according to a unified timestamp and forms a sequence of node behavior characteristics; For ease of understanding, let's assume that within a 6-hour window, we only examine node A (a Tier 1 battery supplier), node B (a Tier 2 chip packaging plant), and node C (a Tier 3 copper foil supplier). Then, node A can form three sets of characteristics at three sampling times. For example, at the first time, it is [8 interactions, 120 resource balance, 2 average settlement delay, 0.1 external intervention score]; at the second time, it is [6 interactions, 96 resource balance, 4 average settlement delay, 0.4 external intervention score]; and at the third time, it is [3 interactions, 70 resource balance, 9 average settlement delay, 0.8 external intervention score]. After time splicing, the node behavior feature sequence of node A is represented as a short time sequence composed of three frames of features, and the same applies to nodes B and C; The system inputs the node behavior feature sequence of all nodes into a dynamic graph neural network model and outputs the network state trust entropy of the current supply chain finance network. This network state trust entropy can be understood as a comprehensive quantitative result of the stability of credit connections, the degree of data transparency, and the uncertainty of interaction willingness in the network. Taking A, B, and C as an example, assuming that after the model outputs the hidden layer representations of the three nodes, the dynamic connection weight between A and B is 0.9, the dynamic connection weight between A and C is 0.3, and the dynamic connection weight between B and C is 0.4; If the node degree distribution within the same window is such that A connects to 2 nodes, B connects to 2 nodes, and C connects to 1 node, and the connection weights fluctuate more, then the network state trust entropy at that moment is relatively higher. The system then compares the trust entropy with the critical trust entropy distribution saved when cascading blocking events were triggered in the past to calculate the critical distance between the current state and the failure boundary; here, the smart contract cascading blocking event can be predefined as: in two consecutive windows, more than 25% of the nodes experience data flow interruption, credit freeze, or settlement failure. After obtaining the network state trust entropy and critical distance, the system invokes the dynamic risk hedging model; this model does not rely solely on a single financial field, but simultaneously generates system survival probability indicators and turnover efficiency indicators. Furthermore, if the system's survival probability index is greater than or equal to the first preset entropy value, the system survival probability index can decrease from 0.82 to 0.47; if the critical distance is lower than the preset distance threshold, it indicates that the system is approaching the failure boundary, and the flow efficiency index may decrease from 0.76 to 0.35. The system weights the two factors to obtain a comprehensive risk assessment status, for example, a comprehensive value of 0.41; at the same time, the system also calculates the degree of variation of the node behavior feature sequence to identify nodes with concentrated risks. Furthermore, if node A's resource balance exceeds a preset rate of change threshold for three consecutive time points, then its mutation weight is the highest, and it is identified as the target node. The smart contract automatically executes a response based on the comprehensive risk assessment status; when the comprehensive risk assessment status is lower than the safety threshold, such as below 0.5, a rigid resource reclamation operation is executed, including freezing the target node's eligibility for new financing, suspending its resource transfer rights, and entering the forced liquidation preparation process; When the overall risk assessment status is higher than or equal to the safety threshold, flexible resource compensation operations are performed, such as suspending instant settlement, temporarily extending the payment period and injecting additional overdraft limits, in order to avoid the risk of system cascading blockage caused by rapid and abnormal outflow of resources. Furthermore, during the data collection phase, if a node does not have complete on-chain data in the current window, the system will prioritize backtracking the confirmed records from the previous window to fill the gap; if the data is still missing after backtracking, the node will be marked as a low-trust node and its contribution weight to the overall network state will be reduced in this round of evaluation. During the critical distance calculation phase, if there are no available historical extreme distribution matching samples in the current window, a preset conservative reference distribution is adopted, and the risk propagation attenuation coefficient is automatically increased, so that the system enters a defensive state earlier. During the contract execution phase, if multiple target nodes simultaneously exceed the risk threshold, they will be executed in batches in the order of core dependency priority and fracture spread priority to avoid a large-scale freeze at once that could lead to further network instability. In the aforementioned new energy vehicle supply chain scenario, one afternoon, panic-inducing news emerged in the market targeting the upstream lithium mining sector. External funds began to contract, and Node C, a third-tier copper foil supplier, initiated multiple abnormal withdrawal requests within two hours, with a significant increase in the delay in account updates. The system calculates that the network state trust entropy of the entire network has increased significantly within the current 6-hour window, and the critical distance has shortened to near the failure boundary. Meanwhile, node A's node behavior characteristic sequence also showed strong fluctuations due to its dependency relationship with node C exceeding the preset dependency threshold. After comprehensive calculation, node C was locked as the target node, with a comprehensive risk assessment status of 0.43, which is lower than the safety threshold. The smart contract then performed rigid resource recovery on node C and simultaneously restricted its continued overdraft of the upstream credit line to prevent abnormal capital outflow from spreading to more levels. The purpose of this step is to build a closed-loop chain from multi-source data collection, temporal feature fusion, network evolution assessment, critical distance determination to smart contract action execution, so that the supply chain finance system can complete risk warning and adaptive handling before the cascading failure actually occurs.

[0019] In a preferred embodiment of the present invention, the specific process of step 1 is as follows: Step 1.1: Collect historical interaction records, current resource holdings and time delay characteristic data of each node in the supply chain finance network to form multi-source circulation status data; Step 1.2: Collect external intervention signals for the supply chain finance network. External intervention signals include transaction orders with a transaction frequency higher than or equal to a preset frequency threshold and a difference between the number of buy and sell orders greater than or equal to a preset ratio, as well as text data with negative sentiment scores exceeding or equal to a preset threshold. Step 1.3: Perform distortion feature identification and filtering on the multi-source flow state data, and then perform tensor concatenation between the filtered multi-source flow state data and the external intervention signal to generate a node behavior feature sequence.

[0020] This embodiment provides a mechanism for refining the construction of input data. Specifically, in the aforementioned main scenario, relying solely on standard financial fields in the on-chain ledger will fail to capture short-term data anomalies and negative text signal aggregations in a timely manner, leading to subsequent evaluation delays. Therefore, this embodiment further performs joint collection and fusion of multi-source flow status data and external intervention signals. In step 1.1, historical interaction records may include the number of historical financing applications, the number of repayments, overdue records, and the length of the bill transfer link with upstream and downstream nodes; current resource holdings may include available credit lines, outstanding accounts receivable balances, and frozen amounts; The latency characteristics can include the latency of invoices being uploaded to the blockchain, the latency of reconciliation confirmation, and the duration of payment delay; further, node B can form the original vector in the current window [application count 5, repaid 4, overdue 1, available credit 80, accounts in transit 60, invoice upload latency 3 hours, reconciliation latency 2 hours]; In step 1.2, the external intervention signal consists of two parts; the first part is the abnormal trading instruction. The system counts the number of instructions of a certain node's associated trading program within a unit of time. If the frequency is higher than the preset frequency threshold, and the ratio of the difference between the number of buy and sell instructions to the total number of instructions is higher than the preset ratio, then it is determined that there is a targeted impact behavior. The second part is negative text signals. The system scores the sentiment of texts such as financial reports, financing community messages, and on-chain announcements. If the negative sentiment score exceeds the preset threshold, it is recorded as an effective intervention. For ease of explanation, assuming that the transaction program related to node C triggers 120 sell orders and 20 buy orders within 30 minutes, the buy-sell difference ratio is 100 / 140, which is approximately 0.71. If the system sets the threshold to 0.6, the signal is valid; for example, if the text content about upstream raw material companies that may face concentrated loan cut-offs is scored as 0.83, and the threshold is 0.75, then the text is also included in the intervention data. In step 1.3, the system first performs distortion feature identification and filtering on the multi-source flow state data, and then performs tensor splicing with the external intervention signal. The specific processing logic of tensor splicing is as follows: if a node originally has 7-dimensional internal data at a certain moment, and additionally acquires 2-dimensional external intervention data, then splice it along the feature axis to form a 9-dimensional feature. If three consecutive time points are sampled, a 3×9 node behavior feature sequence is formed. Taking node C as an example, the internal data at the first time point is [3, 3, 0, 45, 20, 1, 1], the external intervention is [0.1, 0.2], and after splicing, it becomes [3, 3, 0, 45, 20, 1, 1, 0.1, 0.2]. The internal data changes at the third time step are [6, 3, 2, 18, 50, 5, 6], and the external intervention changes to [0.71, 0.83]. Therefore, the third frame in the corresponding sequence significantly enhances the ability to express abnormalities. Furthermore, if the frequency of abnormal trading orders is high but the buy-sell spread ratio does not reach the threshold, it will only be treated as a general disturbance and retained with a low weight, rather than being directly treated as a strong intervention. If the text source is unverified or there is repeated forwarding and accumulation, its sentiment score is deduplicated and decayed to avoid the repeated amplification of a single message; if a node has no external text signal, the text dimension can be filled with zeros; if both transaction and text signals are missing, the all-zero intervention vector is retained to ensure the uniformity of subsequent tensor dimensions. In the new energy vehicle supply chain scenario, although the data of the second-level chip packaging plant node B was still acceptable, high-frequency selling orders suddenly appeared that afternoon, and multiple negative texts about downstream payment failures quickly accumulated. The system combines these external intervention information with its original financing and settlement data, so that the node behavior characteristic sequence of node B presents a composite pattern of financial stability and external shock gradient exceeding the preset gradient threshold within the same window. The purpose of this step is to enable subsequent models to see not only the accounting results, but also the external driving factors of risk propagation, thereby achieving earlier identification of sudden credit contractions in the supply chain finance network.

[0021] In a preferred embodiment of the present invention, in step 2, the process of calculating the network state trust entropy is as follows: inputting the node behavior feature sequence into the spatiotemporal encoder contained in the dynamic graph neural network model, and outputting the hidden state vector of each node. Based on the hidden state vector, the dynamic connection weight between any two nodes is calculated to construct a real-time credit evolution graph; Based on the node degree distribution and the coefficient of variation of dynamic connection weights in the real-time credit evolution graph, the network state trust entropy of the supply chain finance network is calculated.

[0022] This embodiment provides a specific calculation mechanism for network state trust entropy. Specifically, while anomaly scoring of a single node can detect local risks in a particular enterprise, it cannot identify whether the risk has the structural conditions to spread to upstream and downstream. Therefore, this embodiment introduces a spatiotemporal encoder and a real-time credit evolution graph to characterize network layer risks. The node behavior feature sequence of each node is fed into the spatiotemporal encoder in the dynamic graph neural network; the spatiotemporal encoder is a processing structure used to simultaneously extract the time-continuously changing patterns and the correlation patterns between nodes. Taking node A as an example, its 3×9 input sequence can be compressed into a set of hidden state vectors after encoding, such as a 4-dimensional vector [0.82, 0.36, 0.15, 0.74]; node B outputs [0.79, 0.31, 0.42, 0.65]; node C outputs [0.22, 0.88, 0.91, 0.19]; these vectors do not directly represent the financial fields themselves, but represent the overall credit status of the node in the current window; The system calculates dynamic connection weights based on the similarity, dependency strength, and historical interaction strength between the hidden state vectors of any two nodes, and constructs a real-time credit evolution graph. To improve the clarity of dynamic weight calculation, the system adopts an explicit quantification rule: the cosine similarity between the hidden state vectors of node i and node j is used as the feature similarity. Dependency strength is extracted based on the proportion of fund transfers between the two parties to their respective total turnover within a preset time window; and based on cosine similarity. Dependence strength The decay adjustment factor representing the historical interaction strength and preset allocation ratio and The formula: Calculate the dynamic connection weights between node i and node j. To facilitate sandbox simulations, the calculated weight values ​​can be normalized to a range of 0 to 1; if node A and node B have frequent recent transactions, i.e. When the value is close to 1, the flow of accounts receivable is stable, indicating a high degree of dependence. High, and the hidden layer representations are close, that is If the value is high, then the weight can be calculated as 0.88 after substituting it into the formula; If node A and node C have an upstream dependency but their recent interactions have suddenly become discontinuous, the weight can be reduced to 0.41; if the connection between node B and node C was already weak, the weight may be 0.26. This results in a simplified weight matrix: the first row represents the connection weights from node A to A, B, and C as [1, 0.88, 0.41]; the second row represents the connection weights from node B to A, B, and C as [0.88, 1, 0.26]. The third row indicates that the connection weights from node C to A, B, and C are [0.41, 0.26, 1]; where the diagonal lines represent the stable reference values ​​of the nodes themselves, and the off-diagonal lines reflect the dynamic credit relationships between nodes. Furthermore, the system statistically analyzes the node degree distribution and the coefficient of variation of connection weights in the graph; the node degree distribution can be simply understood as the number of edges above the effective connection threshold for each node. Assuming the threshold is set to 0.3, then node A is effectively connected to two nodes, node B is effectively connected to one node, and node C is effectively connected to one node. The coefficient of variation of connection weights reflects the dispersion of the weights of each edge. If the distribution of edge weights changes from a concentrated state to a state of large fluctuations, it indicates that the network trust structure is loosening. The system calculates the network state trust entropy based on this; in practical engineering, the entropy value can be generated by jointly mapping the degree of imbalance in node degree distribution with the degree of fluctuation in edge weight. To avoid relying on implicit computations in the evaluation process, the underlying solution logic of this joint mapping is implemented through structured rules: the system directly calculates the variance of the weights of all effective edges in the real-time credit evolution graph as a quantitative value of the degree of edge weight fluctuation; at the same time, it calculates the Gini coefficient of the degree of the entire network nodes to quantify the degree of imbalance in the distribution of node degrees. The calculated variance and Gini coefficient are linearly normalized and then summed using a preset weighting mechanism, such as each accounting for 50% of the weight, to obtain the network state trust entropy of the current supply chain finance network. Specifically, network state trust entropy The solution formula is: in, The variance of the effective edge connection weights. The Gini coefficient represents the degree of all nodes in the network. and Let these be the variance weight and the Gini coefficient weight, respectively, and satisfy the following conditions: , , , , These are the preset normalized extreme value parameters; Furthermore, the entropy value of a certain stable window is 0.28, while the variance and Gini coefficient surged in this round of shocks, causing the joint mapping result to rise to 0.67, indicating that the network has shifted from a low uncertainty state to a high uncertainty state; Furthermore, if some nodes do not have enough neighbor connections in the current window, resulting in a low node count, the system can look back at the stable connections in the previous window as buffer edges, but will add a decay factor to prevent old relationships from overly dominating the current evaluation. If the feature value distribution dispersion of a node exceeds the preset deviation range, causing its similarity with all nodes to be close to zero, the system can identify it as an isolated risk point and mark it separately instead of deleting it directly, so as to prevent missing critical nodes that are close to breaking the chain. If the node connectivity is lower than the preset connectivity threshold, it indicates insufficient samples or abnormally low network activity. In this case, the trust entropy calculation result will be reduced in credibility and trigger the data supplementation mechanism. In the aforementioned new energy vehicle supply chain, the connection between the core vehicle manufacturer node and Tier 1 battery supplier node A was originally stable. However, after node C, as an upstream copper foil supplier, was subjected to external shocks, its hidden state shifted significantly, resulting in a simultaneous decrease in the weights of AC and BC. Although only one upstream node deteriorated first, the incremental discreteness of the edge weights in the graph exceeded the preset discreteness increment threshold, thus raising the network state trust entropy. Based on this, the system judged that this was not an ordinary single node overdue, but a structural risk that might evolve into cross-level propagation. The purpose of this step is to convert discrete node anomalies into quantitative indicators that can describe the overall uncertainty of the network, thereby enabling more accurate identification of risk diffusion patterns.

[0023] In a preferred embodiment of the present invention, in step 2, the process of calculating the critical distance is as follows: based on a stress test scenario library containing historical extreme fluctuation data and corresponding network state distribution, extract the historical critical trust entropy distribution that triggers the smart contract cascading blocking event; Calculate the relative entropy between the network state trust entropy and the historical critical trust entropy distribution; based on the relative entropy and the preset risk propagation attenuation coefficient, calculate the probabilistic distance between the current state and the triggering of the smart contract cascading blocking event, which is used as the critical distance.

[0024] This embodiment provides a mechanism for calculating critical distance; specifically, knowing only that the current network state has high trust entropy is not enough to determine how much buffer space the system still has. For example, an increase in entropy may be just a short-term fluctuation, while another fluctuation with a similar entropy may trigger a large-scale cascading default. To address this shortcoming, this embodiment compares the current state with the critical distribution under historical extreme fluctuations to obtain a critical distance that can be used to trigger a strategy switch. The system pre-establishes a stress test scenario library, which includes multiple known high-risk scenarios, such as core enterprises delaying the confirmation of rights, upstream price changes exceeding a preset threshold, mass public opinion attacks, and downstream concentrated defaults. It also retains the network state trust entropy distribution in the corresponding window for each scenario. For example, in a certain scenario of panic over upstream raw materials, the critical trust entropy distribution may be concentrated around [0.62, 0.68, 0.74] within the three windows before the final triggering of cascading blockade; in another scenario of frozen payments at core nodes, the critical distribution may be [0.58, 0.63, 0.70]. Based on the characteristics of the current event, the system selects one or more of the most similar samples from the scenario library and extracts the historical critical trust entropy distribution. The system calculates the relative entropy between the current network state distribution corresponding to the network state trust entropy and the historical critical trust entropy distribution; Since the trust entropy result of a single solution is a scalar value, in order to solve the cross-domain abruptness problem caused by directly comparing the numerical scalar with the distribution metric, the system adopts clear rules based on parallel inference of different supply chain sub-network slices within the current preset time window, extracts a set of local network state trust entropy observations as the representative of the current state, and counts the frequency of this set of observations in the preset discrete risk interval to form a normalized current distribution expression. For ease of explanation, assume that the normalized probability distribution of the current window across the three risk intervals is represented as [0.2, 0.5, 0.3], while the extracted historical critical distribution is represented as [0.1, 0.4, 0.5]. The system directly compares the differences between the two distributions using the discrete relative entropy algorithm to obtain the relative entropy value. Specifically, the Kolb-Leibler divergence algorithm is used to extract the divergence value that approximates the current distribution features to the historical critical distribution features as the relative entropy. If the relative entropy is small, it means that the difference between the current network state and the state within the time window before the historical failure is triggered is less than the preset similarity threshold. If the relative entropy is large, it means that there is still a significant deviation. The system calculates the critical distance by combining a preset risk propagation attenuation coefficient; this attenuation coefficient is used to reflect whether the current risk is easily propagated along the supply chain hierarchy. Furthermore, if the vehicle manufacturer has a concentrated amount of accounts receivable from upstream suppliers and multiple secondary nodes rely on the same upstream raw materials, the propagation attenuation coefficient can be set to the first preset attenuation coefficient, such as 0.7, indicating that the risk is not easy to dissipate naturally; if there are many upstream and downstream substitution relationships, the propagation attenuation coefficient can be set to the second preset attenuation coefficient, such as 1.2, indicating that the risk has a greater probability of being absorbed locally. To eliminate the ambiguity of the critical distance mapping process, the system strictly follows the principle derived from relative entropy. Risk propagation attenuation coefficient The solution formula is composed of the product of: Calculate the critical distance as the probabilistic distance. In a sandbox simulation based on the above formula, if the current relative entropy... The propagation attenuation coefficient is 0.08. If the value is 0.7, then the calculated critical distance is... The direct and precise mapping is 0.056; if the system's preset distance is less than 0.1, it enters a dangerous critical region, meaning the current network is approaching the edge of triggering a cascading blocking event. Furthermore, if there are no historical samples in the scene library that are sufficiently close to the current event, the system can construct a reference distribution by mixing multiple types of scenes, or select the most conservative critical distribution to participate in the calculation. If a certain interval has a probability of zero in the relative entropy calculation, a very small smoothing amount is added to that interval to avoid distortion of the results; if the propagation attenuation coefficient cannot be reliably estimated in the current window, the coefficient of the previous stable window is temporarily used, and the final critical distance is lowered at the same time to improve the sensitivity of the system risk response. In the context of the new energy vehicle supply chain, this round of anomalies is highly consistent with the upstream raw material panic scenario database; the system found that the distribution of the current network state trust entropy is very similar to the distribution before the historical cascade break, and because multiple primary and secondary nodes rely on the same type of copper foil raw materials, the risk propagation attenuation coefficient is low. The final calculation shows that the critical distance has entered the dangerous critical zone; at this point, even if there is no overdue data exceeding the preset overdue scale threshold on the books, the system will determine in advance that the supply chain network is approaching the cascading failure boundary. The purpose of this step is to further transform the current level of risk into how far away systemic failure is, thereby enabling a more forward-looking basis for switching contract actions. In a preferred embodiment of the present invention, the specific process of step 3 is as follows: the dynamic risk hedging model includes a survival probability assessment branch and a turnover efficiency assessment branch; Step 3.1: Input the network state trust entropy into the survival probability assessment branch to calculate the system survival probability index; Step 3.2: Input the critical distance into the circulation efficiency evaluation branch and calculate the circulation efficiency index; Step 3.3: Weight and fuse the system survival probability index and circulation efficiency index to obtain the comprehensive risk assessment status.

[0025] This embodiment provides a mechanism for implementing a dynamic risk hedging model. Specifically, if decisions are made solely based on failure probability, the system tends to freeze a large number of nodes in advance. Although this may reduce the system's credit default risk in the short term, it may directly lead to the available resources of midstream and downstream nodes falling below a preset maintenance threshold. If decisions are made solely based on the efficiency of capital turnover, risk control will be overly relaxed, resulting in a lack of buffer when a real credit shock occurs. Therefore, this embodiment separates long-term survivability and short-term turnover efficiency into two branches and integrates them into a comprehensive risk assessment status. The survival probability assessment branch receives network state trust entropy as input; generally speaking, the higher the trust entropy, the more unstable the network relationship, and the lower the system survival probability. Specifically, system survival probability index The solution formula is: in, For network state trust entropy, A scaling factor based on historical default rates; turnover efficiency index The solution formula is: in, This is the critical distance. The on-chain node settlement success rate within the current preset time window. This is the sensitivity coefficient; For ease of understanding, a set of mapping rules can be used: when the trust entropy is between 0.0 and 0.3, the system survival probability index is between 0.8 and 1.0; when the trust entropy is between 0.3 and 0.6, the index is between 0.5 and 0.8; when the trust entropy exceeds 0.6, the index drops below 0.5; furthermore, if the current trust entropy is 0.67, the survival probability index can be calculated as 0.46. The flow efficiency assessment branch receives the critical distance as input; the shorter the critical distance, the closer the system is to the danger edge. In order to prevent a complete freeze caused by misjudgment, the flow efficiency index should not be simply equal to the shorter and lower it is, but should be estimated in combination with the actual situation of current resource flow. For example, if the current critical distance is 0.056, but most nodes on the chain are still making normal repayments, the flow efficiency index may still be 0.41, rather than dropping sharply to an extremely low value, indicating that the system still has some room for improvement; if the critical distance is further shortened and the transaction failure rate increases, then the index will continue to decline. During weighted fusion, the system can assign different weights to the two branches according to business strategies. Furthermore, during stable periods, the survival probability indicator weight can be set to 0.4 and the turnover efficiency indicator weight to 0.6 to balance capital utilization. During high-risk periods, the weights can be switched to 0.7 and 0.3 to prioritize survival. Taking the above values ​​as an example, if a high-risk weight is currently used, the overall risk assessment status is 0.46×0.7+0.41×0.3=0.445; the system then compares this value with the safety threshold to decide whether to enter the recovery or compensation path. To break away from the algorithm in identifying risk-concentrated nodes, the calculation of variation weights in the node behavior feature sequence directly follows clear business evolution structure rules: the system extracts the core feature dimensions of each node within the current preset time window, such as resource balance and external intervention score, and the absolute value of the first difference between adjacent moments to quantify the magnitude of the drastic fluctuations in the data over time. By summing the absolute values ​​of the differences according to a set ratio, the variation weight score of each node with high business interpretability can be obtained. Specifically, nodes Mutation weights The solution formula is: in, The total number of core feature dimensions. and Representing nodes respectively At the present moment Compared to the previous moment The Core eigenvalues, Set the scaling factor for the corresponding feature dimension; Furthermore, if the sum of the absolute values ​​of the first-order difference of resource balance and the difference of external intervention score of node A at three consecutive time points is significantly higher than that of other nodes, then its mutation weight is the highest, and it is accurately located as the target node, ensuring that subsequent resource recovery or compensation instructions are applied to the correct risk trigger source. Furthermore, if the outputs of the two branches differ greatly, for example, the survival probability index is very low while the flow efficiency index is high, it indicates that the system is in a contradictory state where the system is still in operation on paper but the structure is already fragile. At this time, the system can start conflict verification: if the critical distance has entered the dangerous critical region, the weight of the survival probability branch will be increased first. If the critical distance is still relatively sufficient, the original weights are maintained and observation is strengthened; if a branch fails to output due to missing input, the system adopts the result of another branch and adds a conservative penalty term to avoid the overall state being artificially high. In the new energy vehicle supply chain, affected by upstream panic, the network state trust entropy rose from 0.31 to 0.67 in a short period of time, corresponding to a significant decrease in the system survival probability index; at the same time, although most settlements have not completely stopped, the critical distance has approached the danger boundary, and the circulation efficiency index is also under significant pressure. After system integration, the overall risk assessment status was 0.445, which is lower than the preset safety threshold of 0.5. Therefore, the precisely located target node was transferred to a more stringent handling mode. The purpose of this step is to incorporate the survival of the network and the smooth flow of funds into a unified judgment framework, thereby achieving a more robust risk response.

[0026] In a preferred embodiment of the present invention, step 4, the rigid resource reclamation operation specifically includes: interrupting the resource transfer permission of the target node and triggering a preset forced liquidation logic; the flexible resource compensation operation specifically includes: suspending the preset forced liquidation logic and injecting a preset grace period parameter and an additional overdraft resource limit into the target node.

[0027] This embodiment provides a specific action mechanism for the contract execution layer. Specifically, after the aforementioned comprehensive risk assessment status is obtained, if only an abstract risk score is output without any executable on-chain action, it is impossible to truly prevent the spread of risk or complete trust repair. Therefore, this embodiment refines the response actions into two categories: rigid resource recovery and flexible resource compensation. Rigid resource recovery operations are mainly applicable to situations where the comprehensive risk assessment status is below the safety threshold and the target node shows obvious abnormal distortion or continuous deterioration. Interrupting resource transfer permissions may include: prohibiting the node from initiating new financing requests, prohibiting the continued transfer of its accounts receivable vouchers, freezing its outstanding overdraft limits, and suspending its ability to extend credit to external parties; forced liquidation logic may include: invoking a pre-set liquidation contract, calculating its occupied resources, triggering the locking of pledged vouchers, and notifying related settlement nodes to enter the repayment queue. Furthermore, if node C currently has a disposable overdraft limit of 30 and outstanding funds in transit of 18, then after the rigid resource recovery is executed, its new limit will immediately be reduced to zero, and the outstanding funds of 18 will be locked. In the future, only repayment is allowed and expansion is not allowed. Flexible resource compensation operations are applicable when the overall risk assessment shows that the situation is under pressure but has not yet fallen below the safety threshold, or when a certain point fluctuates significantly but the system judges it to be a repairable liquidity shortage rather than malicious withdrawal. Suspending the forced liquidation logic means that the liquidation process, which was already in the countdown, will not be executed for the time being; injecting the grace period parameter can extend the debt that must be repaid within 24 hours to 48 hours or 72 hours; additional overdraft resources can be temporarily supported based on the node's historical performance and network criticality; Furthermore, if node A's original remaining overdraft limit is 10, and the system supplements the limit with 15 and extends the billing period by 48 hours, its total available buffer will increase to 25, thereby avoiding passive supply interruption due to short-term panic. Furthermore, if a target node simultaneously meets the conditions of continuous financial deterioration and a surge in external intervention, but also undertakes the function of supplying key raw materials, the system will not directly enter a single mode, but will instead implement a tiered approach: first, freeze some of its outward expansion permissions, and then retain a limited amount of internal maintenance funds to prevent the disposal action itself from causing the core production chain to break down. If multiple target nodes require flexible compensation, and the available buffer resources on the chain are insufficient, priority will be given to supporting the nodes that have the greatest impact on the production continuity of the core enterprise, while the remaining nodes will enter the queuing compensation or partial liquidation process. In the new energy vehicle supply chain, node C was identified as high-risk and exhibited abnormal capital outflow characteristics. Therefore, the system implemented rigid resource recovery on it: immediately suspending its right to continue transferring upstream accounts receivable and transferring its outstanding financing to liquidation preparation. In contrast, although node A also experienced pressure, its main problem was a temporary shortage of cash flow caused by the rise in upstream raw material prices. Moreover, it had a long-term and stable cooperation with core vehicle manufacturers. Therefore, the system suspended its original countdown for default liquidation and injected it with a 48-hour grace period and additional quota to maintain the uninterrupted supply of the first-level key components. The purpose of this step is to directly map the risk assessment results into auditable, executable, and traceable on-chain actions, thereby enabling precise shrinking of abnormal nodes and targeted support for repairable nodes.

[0028] In a preferred embodiment of the present invention, network state trust entropy is used to characterize the uncertainty of the underlying node's data transparency and willingness to interact with data.

[0029] This embodiment provides a business interpretation mechanism for network state trust entropy. Specifically, if this indicator is only understood as a general level of risk, it is easily confused with traditional indicators such as delinquency rate and bad debt rate during engineering deployment, leading to inaccurate threshold settings. Therefore, this embodiment clarifies that it is used as an uncertainty measure to characterize the data transparency and data interaction willingness of underlying nodes. Data transparency can be understood as whether nodes continuously, completely, and on time submit key data such as invoices, liabilities, payments, and voucher transfers; data interaction willingness can be understood as whether nodes are willing to continue to participate in settlement, financing, and reconciliation through on-chain rules, rather than turning to off-chain non-consensus operations or abnormal resource transfers. A high network state trust entropy does not necessarily mean that all nodes have defaulted, but rather that more and more nodes are beginning to show a tendency to be unwilling to disclose data, unwilling to continue transactions, and that relationships are no longer stable. For ease of understanding, let's assume that the transparency scores of three nodes within a window are 0.9, 0.8, and 0.3, and the interaction willingness scores are 0.85, 0.7, and 0.2, respectively. The first two nodes maintain a high degree of transparency and interaction, while the third node shows a clear retreat; if only the overdue amount is considered, the third node may not have exceeded the limit yet. However, from the perspective of network state trust entropy, its uncertainty is already rising; if we further observe that multiple nodes are sliding from the high transparency and high interaction range to the medium and low range, the entropy value will continue to rise, reflecting that the trust foundation is becoming sparse and fragile. Furthermore, if the low transparency of a node is due to objective reasons, such as a short-term failure of the above link or a temporary delay in submission due to regulatory audit, the system can classify it as a technical low transparency and record it separately, without immediately considering it as a subjective retreat. If the reduced willingness of a node to interact is caused by external entities such as upstream shutdowns, then verification should be performed in conjunction with the production planning interface to avoid misjudging the contraction of entity operations as credit evasion. In the new energy vehicle supply chain, Tier 3 copper foil supplier Node C began to delay uploading invoices, reduce reconciliation responses, and frequently attempt to transfer transactions to off-chain completion. Although its overdue payments have not yet been fully exposed, from the perspectives of data transparency and willingness to interact, this node has clearly deviated from a stable cooperative state. Therefore, the network state trust entropy has been rapidly increased, and the system has taken measures to deal with it in advance. The purpose of this step is to clarify that the network state trust entropy describes the uncertainty of the trust foundation, rather than just traditional financial results, so as to achieve early detection of hidden risks.

[0030] In a preferred embodiment of the present invention, in step 1.3, distortion feature identification and filtering is used to identify defensive data forgery features in node data; wherein, defensive data forgery features include: features where the distribution of invoice amount deviates from a preset normal range, features where the update frequency of liability data is lower than a preset time threshold, and features where there are no historical interaction records in the data flow path between nodes.

[0031] This embodiment provides a refined mechanism for distortion feature identification and filtering. Specifically, in the aforementioned data fusion process, if defensive data forgery features are not identified first, the system may treat modified accounting data as real and stable signals, thereby concealing risks. Therefore, this embodiment further identifies spoofing behavior in node data. The first type of characteristic is that the distribution of invoice amounts deviates from the preset normal range; this does not require that the invoice amounts must strictly follow a certain theoretical distribution, but rather uses the distribution of amounts during the historical stable operating period of the node as the reference range; Furthermore, in the past 90 days, the invoice amount of Node B has mainly been concentrated between 80,000 and 120,000, with an average of about 100,000, and the fluctuations are relatively stable; If more than 20 invoices with amounts of 49,800, 49,900, and 50,100 suddenly appear in the current window, and they are clearly clustered near a certain boundary to circumvent the audit, the system will mark them as having a distribution deviation feature. This type of feature often reflects that the node is concealing its true financial difficulties by splitting invoices or lowering the amount of each transaction. The second type of characteristic is that the frequency of liability data updates is lower than the preset time threshold. If a node usually updates its liability snapshot every 12 hours during a stable period, but has not updated for 36 consecutive hours and its transaction activity has not decreased accordingly, it indicates that the node may be intentionally delaying the exposure of its true liability level. The system can use the ratio of the update interval to the historical average update interval as a criterion; further, the historical average is 12 hours, the current average is 36 hours, the ratio is 3, and if the threshold is set to 2, it is judged as abnormal. The third type of feature is that there is no historical interaction record in the data flow path between nodes; it is assumed that node C usually only has stable transactions with node A, but in the current window, a key certificate is transferred through nodes D and E that have never appeared before, forming a new ADE-clearing path; If the system checks the historical graph and finds that the path did not exist before, and nodes D and E both lack sufficient credit accumulation, it is considered an abnormal path; this feature can be used to identify behaviors such as transferring assets through temporary paths, falsifying transaction backgrounds, or circumventing existing risk control rules. During the filtering process, the system does not delete all data with distorted features. Instead, it stratifies the data according to the level of anomaly. Mild anomalies are simply reduced in weight, moderate anomalies are entered into a manual review pool or on-chain secondary verification, and high anomalies are removed from the high-confidence training input of this round and formed into a separate suspected forgery feature channel for subsequent models. For example, if node B experiences only one liability update delay, it is retained but its weight is reduced; if both invoice splitting and aggregation and unfamiliar path flow occur simultaneously, its master data can be removed, while the anomaly marker is retained separately. Furthermore, if the deviation in invoice amount is caused by a change in the actual business structure, such as the addition of small-amount repair parts procurement business, the system will cross-validate the procurement catalog and contract type; if the delay in liability update is due to on-chain interface downtime maintenance, it will be exempted through system logs; if the new path flow comes from an approved newly added cooperative node, it will not be considered abnormal after completing whitelist registration. In the new energy vehicle supply chain, after encountering financial pressure, node C began to split a raw material purchase invoice of 150,000 yuan into three smaller invoices of nearly 50,000 yuan each. It also failed to update the real debt snapshot for two consecutive days and introduced a transfer voucher from an intermediate node that had never participated in the settlement of the industry chain. The system identifies these three features as a high-level defensive data forgery signal, thereby preventing subsequent models from being misled by manipulated accounting data. The purpose of this step is to strip or label inputs that may have spoofing intentions before the data enters the risk model, thereby improving the credibility of subsequent network evaluation results.

[0032] In a preferred embodiment of the present invention, in step 1.2, the external intervention signal is intervention data with a preset selling frequency characteristic injected by a quantitative trading program.

[0033] This embodiment provides a specific implementation mechanism for external intervention signal sources; specifically, in some extreme scenarios, the supply chain finance network does not experience actual node defaults, but is first impacted by programmed selling in the capital market and triggers a chain of risk transmission. If the system only focuses on on-chain operational data, it will miss the earliest sources of impact; therefore, this embodiment uses intervention data injected by a quantitative trading program, which has preset selling frequency characteristics, as a dedicated external intervention signal. The system can extract programmatic trading traces from market interfaces related to the supply chain, such as bulk raw materials, core stocks, bond prices, and industry exchange-traded funds. The preset selling frequency characteristics here are not limited to large single selling volume, but focus more on multiple consecutive selling behaviors with regular rhythm and consistent direction within a unit of time. Furthermore, within a 10-minute observation window, if a certain related asset experiences a fixed-size sell order every 30 seconds, triggering 18 times consecutively, and the total number of sell orders is far greater than the total number of buy orders, then it can be determined that there are characteristics of programmed selling frequency. The system maps the intervention data to relevant supply chain nodes, for example, mapping the impact of the sell-off of copper foil raw material-related assets to node C and its upstream and downstream dependent nodes; Furthermore, the system does not directly use market price fluctuations themselves as the intensity of intervention, but instead combines the frequency of selling, directional consistency, and coverage of related assets into an intervention score; For example, if the selling frequency is standardized to 0.7, the directional consistency to 0.9, and the related asset coverage to 0.8, then the external intervention score can be synthesized into a higher value; this can avoid treating occasional large-scale trading orders and continuous programmatic interventions the same way. Furthermore, if high-frequency selling is detected but market depth is sufficient and prices recover quickly, the signal will only be retained as a medium intervention and will not be immediately upgraded to a high-risk level; if the market interface is temporarily interrupted, the system can switch to using cached transaction segments from the most recent period with added time decay. If a sell-off occurs on an asset unrelated to this supply chain, even if it is frequent, it will not be mapped to the current network to avoid false alarms. In the new energy vehicle supply chain, within a certain preset time window, multiple trading targets related to upstream lithium battery materials simultaneously issued regular and continuous sell orders, with the frequency of sell orders within 10 minutes far exceeding the usual average. After the system identifies the intervention data injected by this type of quantitative trading program, it writes it as an external intervention signal into the current window features of node A and node C, so that the model can perceive the pressure of credit contraction in advance before the on-chain account has completely deteriorated. The purpose of this step is to incorporate the quantifiable characteristics of panic transmission in the capital market into the supply chain finance risk assessment process, thereby enabling a more timely response to external shocks.

[0034] In a preferred embodiment of the present invention, after step 4, the method further includes: collecting feedback status data of the supply chain finance network after performing rigid resource recovery operations or flexible resource compensation operations; calculating the trust repair gradient based on the feedback status data; and using the trust repair gradient to iteratively update the weight parameters of the dynamic graph neural network model to achieve model evolution.

[0035] This embodiment provides a mechanism for model evolution based on execution feedback. Specifically, the aforementioned process can achieve risk assessment and contract response, but if the model remains fixed for a long time, it will be difficult to adapt to changes in the behavior of supply chain nodes, market disturbances, and settlement effects. Therefore, this embodiment continues to collect feedback status data after contract execution and uses it to update the dynamic graph neural network model. Feedback status data can include at least the following categories: whether the target node stops abnormal withdrawals after the rigid resource recovery is executed, whether its upstream and downstream nodes experience new supply disruptions, and whether the locked funds are successfully recovered. After implementing flexible resource compensation, whether the target node resumes invoice uploading, whether the reconciliation delay decreases, and whether the additional overdraft limit is used for normal procurement rather than abnormal transfer; Furthermore, if after Node A is granted a 48-hour grace period, its invoice upload delay recovers from 6 hours to 2 hours, its liability snapshot update returns to normal, and its downstream production replenishment resumes, then the compensation action has indeed restored some trust. Conversely, if Node C's associated nodes continue to exhibit abnormal transfers after being frozen, then single-point recovery is insufficient to stop the spread. Based on this feedback, the system calculates the trust repair gradient; this can be understood as a measure of whether the network state changes towards stability or deterioration after a certain response action, and how much the change is. For ease of sandbox simulation, we assume that the network state trust entropy is 0.67 before the implementation of flexible compensation, drops to 0.49 after two windows, and the weights of key connections recover significantly. In this case, the trust repair gradient can be recorded as positive. If the entropy value rises from 0.67 to 0.74 after the implementation of rigid recycling, it indicates that the disposal caused greater fluctuations. In this case, the repair gradient is negative. The system uses this gradient as a supervision signal to update the parameters related to temporal encoding and connection weight prediction in the dynamic graph neural network model, so that the model gradually learns which actions are more conducive to network stability in which states. Specifically, the process of calculating the trust repair gradient is as follows: Define the trust repair objective function: in, and These represent the network state trust entropy before and after executing the response action, respectively. This represents the change in settlement failure rate in the target node's associated subgraph. and A preset penalty coefficient is used; the backpropagation algorithm is used to calculate... The partial derivatives with respect to the weight parameters of the dynamic graph neural network model yield the trust repair gradient: The weight parameters are updated using gradient descent. During iterative updates, the system can use a window-by-window fine-tuning approach instead of complete retraining each time. For example, if the current model is too harsh on short-term panic nodes, and multiple feedbacks reveal that flexible compensation can reduce subsequent entropy values, the model will gradually increase its tolerance for such nodes. If some nodes transfer assets faster after being granted a grace period, the model will increase the risk weight of their abnormal patterns; in this way, the model evolution and the smart contract execution results form a closed loop. Furthermore, if the feedback data after a certain execution is insufficient, such as if the observation time within the window is too short or the node is temporarily offline, the system will not update the parameters for the time being, but will only cache the record for later completion. If the feedback results are affected by a major external event, such as a sudden regulatory crackdown causing a sharp drop in the total transaction volume across the network, the data in that window will be marked as a low-learnability sample to avoid mislearning exogenous shocks as the effect of the strategy. If updates of multiple consecutive windows cause the coefficient of variation of the model output features to exceed the preset stable threshold, the system can trigger a rollback mechanism to revert to the parameters of the previous stable version. In the new energy vehicle supply chain, the system first performs rigid resource recovery on node C and flexible resource compensation on node A; feedback after 12 hours showed that node A resumed normal reconciliation with upstream purchases, and the settlement failure rate of its associated nodes decreased. After node C is frozen, its abnormal outflow stops, but some secondary supply nodes still experience delays. Based on this, the system calculates the trust repair gradients corresponding to the two types of actions and updates the model parameters so that when faced with similar situations such as upstream panic, pressure on key primary supply, and outflow of secondary nodes, the model can more accurately distinguish between nodes that should be compensated and nodes that should be recovered. The purpose of this step is to ensure that the model does not remain at the static identification level, but continuously corrects the assessment logic based on the actual handling results, thereby achieving the dynamic evolution of risk assessment capabilities.

[0036] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. An AI-powered smart contract risk assessment method for supply chain finance, characterized in that, Execution in a blockchain network where smart contracts are deployed includes: Step 1: Collect multi-source circulation status data and external intervention signals, fuse them according to time series characteristics to obtain node behavior feature sequences, and combine them with the direction of fund flow to form dynamic graph data; Step 2: Input the node behavior feature sequence into the hidden state vector of the dynamic graph neural network model to construct a real-time credit evolution graph and calculate the network state trust entropy; extract the historical critical trust entropy distribution in the stress test scenario library and calculate the probability distance between it and the network state trust entropy as the critical distance; Step 3: Based on the network state trust entropy and critical distance, calculate the system survival probability index and flow efficiency index through the dynamic risk hedging model to determine the comprehensive risk assessment state; extract the absolute value of the first difference of the node behavior feature sequence at adjacent time points to calculate the variation weight, and locate the target node; Step 4: When the overall risk assessment status is lower than the preset safety threshold, a rigid resource reclamation operation is performed on the target node; when it is not lower than the safety threshold, a flexible resource compensation operation is performed.

2. The AI ​​smart contract risk assessment method for supply chain finance according to claim 1, characterized in that, The specific process of step 1 is as follows: Step 1.1: Collect historical interaction records, current resource holdings, and time delay characteristic data of each node in the supply chain finance network to form multi-source circulation status data; Step 1.2: Collect external intervention signals for the supply chain finance network. The external intervention signals include transaction instructions with a transaction frequency higher than or equal to a preset frequency threshold and a difference between the number of buy and sell instructions greater than or equal to a preset ratio, as well as text data with a negative sentiment score exceeding or equal to a preset threshold. Step 1.3: Perform distortion feature identification and filtering on the multi-source flow state data, and perform tensor concatenation between the filtered multi-source flow state data and the external intervention signal to generate a node behavior feature sequence.

3. The AI ​​smart contract risk assessment method for supply chain finance according to claim 1, characterized in that, In step 2, the process of calculating the network state trust entropy is as follows: The node behavior feature sequence is input into the spatiotemporal encoder contained in the dynamic graph neural network model, and the hidden state vector of each node is output. Based on the hidden state vector, the dynamic connection weight between any two nodes is calculated to construct a real-time credit evolution graph. Based on the node degree distribution in the real-time credit evolution graph and the coefficient of variation of the dynamic connection weights, the network state trust entropy of the supply chain finance network is calculated.

4. The AI ​​smart contract risk assessment method for supply chain finance according to claim 3, characterized in that, In step 2, the process of calculating the critical distance is as follows: Based on a stress test scenario library containing historical extreme fluctuation data and corresponding network state distributions, the historical critical trust entropy distribution that triggers cascading blocking events of smart contracts is extracted. Calculate the relative entropy between the network state trust entropy and the historical critical trust entropy distribution; Based on the relative entropy and the preset risk propagation attenuation coefficient, the probability distance between the current state and the triggering of the smart contract cascading blocking event is calculated, which is used as the critical distance.

5. The AI ​​smart contract risk assessment method for supply chain finance according to claim 1, characterized in that, The specific process of step 3 is as follows: The dynamic risk hedging model includes a survival probability assessment branch and a turnover efficiency assessment branch; Step 3.1: Input the network state trust entropy into the survival probability assessment branch to calculate the system survival probability index; Step 3.2: Input the critical distance into the circulation efficiency evaluation branch to calculate the circulation efficiency index; Step 3.3: Weight and fuse the system survival probability index and the circulation efficiency index to obtain the comprehensive risk assessment status.

6. The AI ​​smart contract risk assessment method for supply chain finance according to claim 1, characterized in that, In step 4, the rigid resource reclamation operation specifically includes: interrupting the resource transfer permission of the target node and triggering a preset forced liquidation logic; the flexible resource compensation operation specifically includes: suspending the preset forced liquidation logic and injecting a preset grace period parameter and an additional overdraft resource limit into the target node.

7. The AI ​​smart contract risk assessment method for supply chain finance according to claim 1, characterized in that, The network state trust entropy is used to characterize the uncertainty of the underlying node's data transparency and willingness to interact with data.

8. The AI ​​smart contract risk assessment method for supply chain finance according to claim 2, characterized in that, In step 1.3, the distortion feature identification and filtering is used to identify defensive data forgery features in node data; wherein, the defensive data forgery features include: the feature that the distribution of invoice amount deviates from the preset normal range, the feature that the update frequency of liability data is lower than the preset time threshold, and the feature that there are no historical interaction records in the data flow path between nodes.

9. The AI ​​smart contract risk assessment method for supply chain finance according to claim 2, characterized in that, In step 1.2, the external intervention signal is intervention data with a preset selling frequency characteristic injected by the quantitative trading program.

10. The AI ​​smart contract risk assessment method for supply chain finance according to claim 1, characterized in that, Step 4 is followed by: Collect feedback status data of the supply chain finance network after executing the rigid resource recovery operation or the flexible resource compensation operation; The trust repair gradient is calculated based on the feedback state data; the weight parameters of the dynamic graph neural network model are iteratively updated using the trust repair gradient to achieve model evolution.

Images