Supply chain financial risk management system based on credit assessment

Abstract

The invention discloses a supply chain financial risk management system based on credit assessment, and relates to the field of financial science and technology. A five-dimensional data model is constructed through a dynamic data modeling unit, and transaction subject levels and legal potency dimensions are integrated; the causal analysis engine adopts a dynamic updating algorithm of time sequence feature coding and compliance score constraint, and introduces a legal potency score constraint mechanism to strengthen compliance verification. The cross-domain risk conduction model combines physical space voxel grid monitoring with a digital space fluid dynamic equation, and the risk conduction velocity is quantified through a risk diffusion coefficient and a partial differential equation; and the self-evolution defense module integrates a confrontation sample generator and an incorruption degree feedback mechanism, simulates logistics track abnormity through Levy distribution, and improves the defense capability of the system in combination with a dynamic weight adjustment strategy of approval skipping times. According to the system, the deep supply chain risk conduction path identification precision is improved, and the problems of a multi-level network risk monitoring blind area and cross-domain risk tracking lag are effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical field of fintech, and particularly relates to a supply chain finance risk management system based on credit assessment. Background Art

[0002] With the acceleration of the complexity and digital transformation of the global supply chain, supply chain finance has become a key means to solve the financing difficulties of small and medium-sized enterprises and optimize the capital efficiency of the industrial chain. The existing technologies mainly rely on two-dimensional credit assessment models and static risk index analysis, and it is difficult to effectively describe the dynamic risk conduction mechanism in multi-level supplier networks. Especially when dealing with complex scenarios such as cross-border trade and warehouse receipt pledge, the traditional methods have three core pain points: First, the limited data dimension leads to insufficient depth of risk monitoring. The industry general technology can only effectively track the anomalies of Tier 2 suppliers, and there is a lack of quantitative means for the cascading impact effect of deep supply chain nodes; Second, the coupling effect of legal compliance constraints and dynamic game of business behaviors has not been incorporated into the risk assessment system, resulting in the compliance verification of key elements such as the validity of electronic contracts and the time point of transfer of goods ownership lagging behind the occurrence of risk events; Third, there is a break in the real-time mapping between the physical warehousing environment and the digital twin system. The existing risk control models are difficult to synchronously track the cross-domain diffusion laws of the value fluctuation of goods, logistics interruption and financial risks, and the prediction response delay exceeds 48 hours when dealing with sudden supply chain crises, resulting in the accumulation of systemic financial risks. Summary of the Invention

[0003] The present invention provides a supply chain finance risk management system based on credit assessment to solve the problems of insufficient accuracy in identifying the risk conduction path of multi-level supplier networks, lagging response of dynamic legal compliance verification, and difficulty in tracking the cross-domain diffusion law of physical-digital space risks in the prior art.

[0004] The present invention provides a supply chain finance risk management system based on credit assessment, including: a dynamic data modeling unit, a causal analysis engine, a cross-domain risk conduction model, and a self-evolving defense module;

[0005] The dynamic data modeling unit constructs a five-dimensional data model including the transaction subject level L and the legal effect C of the contract, and maps the supply chain data into an N×M×t×L×C-dimensional tensor, where N is the set of participating subjects, M is the set of business indicators including the accounts receivable turnover rate and the proportion of factoring financing, t is the granularity of the dynamic time slice, the L dimension records the hierarchical association relationship between multi-level suppliers and the core enterprise. When the contract signature confidence CFCA of the L3-level supplier is less than 0.4 or At this time, the blockchain verification process of the buyer's payment ability at the core enterprise level is automatically triggered; the C dimension calculates the legal effect score of the contract through the electronic signature hash value, additionally verifies the confirmation and deposit of accounts receivable for factoring contracts, and synchronizes with the registration status of the registration agency in real time;

[0006] The causal analysis engine executes a dynamic update algorithm for the causal adjacency matrix that complies with constraints: using a dual-channel update mechanism of historical retention and incremental learning, it fuses 93% of the weight of the adjacency matrix at time t with a 7% dynamic increment, and the dynamic increment is generated by non-linearly transforming the concatenated vector of the time series features encoded by a trainable gating matrix and the hidden state; injecting a legal effect attenuation function, based on the contract signature confidence levels c i and c j to calculate the combined legal effect score with the real-time compliance index q, and generating a non-linear attenuation coefficient through an S-shaped curve with an attenuation slope parameter of 2.5; when c i +c j < 0.8, it activates the compliance acceleration attenuation mechanism, increasing the attenuation rate of the associated weight to 300% of the benchmark value.

[0007] Furthermore, when the contract signature confidence level CFCA of a L3-level supplier is detected to be < 0.4, it calls the Hyperledger Fabric chain to verify the L1-level core contract for on-chain evidence storage; compresses the data dimension through Tucker decomposition, retains the principal components with a variance contribution rate ≥ 85%, and generates a data quality report.

[0008] Furthermore, the cross-domain risk conduction model includes: a physical space risk mapping unit and a digital space diffusion calculation unit, specifically including:

[0009] Using the physical space risk mapping unit, discretize the warehouse into a voxel grid of 1m 3 , and monitor the value of goods, temperature and humidity, and risk concentration in real time, recording the inventory value volatility and the logistics monitoring coverage rate;

[0010] Through the digital space diffusion calculation unit, use the inventory value volatility and the logistics monitoring coverage rate, combine with the time variable t and the attenuation constant λ, and calculate the diffusion coefficient

[0011] Use the risk hydrodynamics equation to quantify the risk conduction speed of the capital flow:

[0012]

[0013] where, R i represents the risk concentration of region i, γ i is the attenuation coefficient, β is the conversion rate, S j and I j represent the capital flow intensity and investment level of region j respectively;

[0014] If the logistics monitoring coverage rate is lower than 60%, D i is magnified by 1.8 times, triggering an update of the three-dimensional risk heat map of the warehouse.

[0015] Furthermore, the physical space risk mapping unit performs the following steps:

[0016] Step 1: Obtain the three-dimensional point cloud data of the warehouse through Internet of Things sensors and generate a voxel grid;

[0017] Step 2: When the risk value of any grid > 0.8, automatically freeze the associated warehouse receipt financing quota and activate the blockchain smart contract to execute asset preservation;

[0018] Step 3: Write the disposal record into the Hyperledger Fabric ledger to ensure the immutability of the data.

[0019] Furthermore, the self-evolving defense module includes: an adversarial sample generator and an integrity feedback unit. The adversarial sample generator collects historical vulnerability data in the industry and extracts relevant data and information to generate attack samples, including contract clause semantic tampering and false logistics track injection;

[0020] The specific content of the clause semantic tampering includes: initializing the BERT-Masked model to ensure the correct model parameters and configurations; inputting the contract clause text to be tampered into the BERT-Masked model; using the masking mechanism of the BERT-Masked model to tamper with the key semantic information in the contract text to generate a text sample with a high perplexity, where the perplexity > 120, to simulate the situation where an attacker maliciously tampers with the contract clause.

[0021] The specific content of the false logistics track injection includes: generating GPS jump noise data that follows the Levy distribution according to the characteristics of the Levy distribution to simulate the abnormal jump phenomenon in the logistics track; injecting the generated GPS jump noise data into the real logistics track data to form a false logistics track sample to simulate the situation where an attacker tampers with the logistics information.

[0022] Furthermore, the integrity feedback unit: obtains the system operation log and extracts the information of the number of approval skips; uses the formula: w ij := w ij ×(0.3 + 0.7e -0.5·bypass_count ) to calculate the weight adjustment factor, where w ij represents the edge weight between nodes i and j in the causal graph, bypass_count represents the number of approval skips, and applies the calculated weight adjustment factor to the edge weights of the causal graph to achieve dynamic adjustment of the causal graph.

[0023] Furthermore, the system responds to the occurrence of risks through the following steps:

[0024] Step 1: When the hash value of the electronic creditor's right certificate deviates from the historical mean by ±2σ, call the compliance verification layer to verify the CFCA signature status;

[0025] Step 2: If the signature confidence level < 0.63, perform counterfactual causal intervention analysis:

[0026] ΔR = E[R|do(contract valid = False)] - E[R|contract valid = False]

[0027] Step 3: When ΔR > 1.4, freeze the associated account within 47 seconds and generate a risk conduction path report.

[0028] Furthermore, the counterfactual causal intervention analysis includes:

[0029] Construct a causal diagram containing 12 confounding variables;

[0030] Activate the independent audit process when the intervention effect size ΔR > 2.1.

[0031] The present invention improves the recognition ability of the risk conduction path of the deep - layer supplier network through a five - dimensional dynamic tensor modeling and cross - domain coupling analysis mechanism, breaking through the monitoring blind area of the existing technology for secondary supplier nodes; through the synergistic effect of a dynamic constraint algorithm and a real - time causal reasoning engine, it realizes the instant verification of the entire process of e - contract compliance risks; constructs a dual - domain synchronous tracking model, effectively captures the dynamic correlation law between the abnormal changes in the warehousing environment and financial risk indicators, and significantly shortens the cross - domain risk early warning response cycle. In the technical implementation, an adversarial training defense mechanism and an adaptive weight adjustment strategy are integrated to comprehensively enhance the system's robust control ability for complex supply chain scenarios. Description of the Drawings

[0032] Figure 1 It is the system architecture diagram of a supply chain finance risk management system based on credit assessment of the present invention;

[0033] Figure 2 It is the decomposition diagram of the causal analysis process in the present invention. Detailed Embodiments

[0034] The present invention relates to a supply chain finance risk management system based on credit assessment, which realizes full - chain risk prevention and control through multi - dimensional data modeling, dynamic causal analysis, cross - domain risk conduction modeling and self - evolving defense mechanism, and is used to solve the problems of insufficient recognition accuracy of the risk conduction path of multi - level supplier networks, lag in dynamic legal compliance verification response, and difficulty in tracking the cross - domain diffusion law of physical - digital space risks.

[0035] The above technical solutions will be described in detail below in combination with the specification drawings and specific implementation manners to better understand the above technical solutions. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments of the present invention. It should be understood that the present invention is not limited to the example embodiments that only explain the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention. In addition, it should be noted that for the sake of convenience of description, only the parts related to the present invention are shown in the drawings rather than all of them.

[0036] Embodiment 1

[0037] As Figure 1 - Figure 2 shown: Dynamic Data Modeling Unit, Causal Analysis Engine, Cross-Domain Risk Propagation Model, Self-Evolving Defense Module;

[0038] The Dynamic Data Modeling Unit maps the supply chain data into an N×M×t×L×C-dimensional tensor by constructing a five-dimensional data model including the transaction subject level L and the contract legal effect C, where: N is the set of participating subjects, M is the set of business indicators including the accounts receivable turnover rate and the proportion of factoring financing, and t is the dynamic time slice granularity;

[0039] The L dimension records the hierarchical association relationship between multi-level suppliers and the core enterprise. When it is detected that the contract signature confidence CFCA of the L3-level supplier is less than 0.4 or when, the blockchain verification process of the buyer's payment ability at the core enterprise level is automatically triggered;

[0040] The C dimension calculates the contract legal effect score through the electronic signature hash value, additionally verifies the confirmation and deposit of accounts receivable for the factoring contract, and synchronizes with the registration status of the registration agency in real time;

[0041] The Causal Analysis Engine executes the dynamic update algorithm of the causal adjacency matrix with compliance constraints. The specific implementation method is:

[0042] Using a dual-channel update mechanism of historical retention and incremental learning, fusing 93% of the weight of the adjacency matrix at time t with 7% of the dynamic increment, and the increment is generated by non-linearly transforming the concatenated vector of the time series features encoded by the trainable gating matrix and the hidden state;

[0043] Injecting a legal effect attenuation function, calculating the joint legal effect score based on the contract signature confidences c i 、c j and the real-time compliance index q, and generating a non-linear attenuation coefficient through an S-shaped curve with an attenuation slope parameter of 2.5; when it is detected that c i +c jWhen <0.8, activate the compliance acceleration decay mechanism to increase the decay rate of the correlation weight to 300% of the benchmark value.

[0044] Specifically, based on the temporal feature encoding and compliance score constraint, the causal adjacency matrix is dynamically updated, and the calculation formula is: A t+1 = 0.93A t + 0.07σ(W·[X t ∥H t ) where A t : the causal adjacency matrix at the current time t, [X t ∥H t : represents the concatenation operation of the temporal feature encoding X t and the hidden state H t , and the compliance constraint function is: where c i , c j are the contract legal effect scores of the two trading parties. When c i + c j <0.8, the weight decay speed is increased by 3 times.

[0045] Specifically, the dynamic data modeling unit maps the supply chain data into a structured tensor by constructing a five-dimensional data model N×M×t×L×C:

[0046] N dimension: the set of supply chain participants, covering the core enterprise (L0), first-tier suppliers (L1) to multi-tier suppliers (Ln);

[0047] M dimension: the set of business indicators, including dynamic indicators such as cash flow, inventory volume, and credit rating;

[0048] t dimension: the time slice granularity, automatically adjusted according to the risk density (1 minute to 1 day);

[0049] L dimension: the transaction level, recording the hierarchical association relationship between multi-tier suppliers and the core enterprise;

[0050] C dimension: the contract legal effect feature, calculating the confidence based on the CFCA electronic signature hash value and synchronizing with the registration status of the registration agency in real time.

[0051] Dynamic data modeling unit: obtains supply chain transaction data, contract texts, and logistics monitoring information in real time through the API interface; classifies and fills the original data according to the five dimensions to generate an initial tensor When the contract signature confidence of the L3-level supplier < 0.4, trigger the cross-verification process at the core enterprise level; use the Tucker decomposition algorithm to reduce the dimension of the tensor, and retain the principal components with a variance contribution rate ≥ 85%.

[0052] The blockchain verification process specifically includes:

[0053] Evidence storage initialization stage: Generate a hash digest for the key elements of the electronic contract (contract number, signing timestamp, public key of the signing party) through SHA-256, call the Fabric chain code to execute the invoke operation, and write the digest into the leaf node of the Merkle tree;

[0054] Cross-verification stage: When the L3-level supplier verification is triggered, send a query request to the sorting node, compare the contract hash stored on the chain with the locally calculated value. If there is a difference, start the view change process (View-Change Protocol) of the PBFT consensus protocol for ledger synchronization

[0055] Causal analysis engine: Perform LSTM encoding on the time series data (such as cash flow fluctuations) in the five-dimensional tensor;

[0056] According to formula A t+1 = 0.93A t + 0.07σ(W·[X t ∥H t ) to update the causal adjacency matrix; The key causal edges (such as the payment commitment of the core enterprise) are written into the Hyperledger Fabric ledger.

[0057] This solution integrates the CFCA signature confidence level and the registration status of the registration agency (such as the China Securities Depository and Clearing Corporation Limited website) into the legal effect score c through dimension C i = 0.6×CFCA_score + 0.4×China Securities Depository and Clearing Corporation Limited website status to solve the problem of false contract identification; When an L3-level supplier anomaly is detected, automatically call the blockchain evidence storage to verify the L1-level core contract, and the verification response time < 200ms to form a closed-loop verification; Through the compliance score constraint function, realize the dynamic regulation of the legal effect on the risk conduction path; The causal graph is updated every 15 minutes to support minute-level risk early warning.

[0058] Furthermore, the cross-domain risk conduction model includes: a physical space risk mapping unit and a digital space diffusion calculation unit, specifically including:

[0059] Using the physical space risk mapping unit, discretize the warehouse into a voxel grid of 1m 3 to real-time monitor the value of goods, temperature and humidity, and risk concentration, and record the inventory value volatility and logistics monitoring coverage rate;

[0060] Through the digital space diffusion calculation unit, use the inventory value volatility and logistics monitoring coverage rate, combined with the time variable t and the decay constant λ, to calculate the diffusion coefficient

[0061] Use the risk hydrodynamics equation to quantify the risk conduction speed of the capital flow:

[0062]

[0063] Among them, R i represents the risk concentration of area i, γ i is the attenuation coefficient, β is the conversion rate, S j and I j respectively represent the capital flow intensity and investment level of area j;

[0064] If the logistics monitoring coverage rate is lower than 60%, D i is magnified by 1.8 times, triggering the update of the three-dimensional risk heat map of the warehouse.

[0065] The physical space risk mapping unit performs the following steps:

[0066] Step 1: Obtain the three-dimensional point cloud data of the warehouse through Internet of Things sensors and generate a voxel grid;

[0067] Step 2: When the risk value of any grid > 0.8, automatically freeze the associated warehouse receipt financing quota and activate the blockchain smart contract to execute asset preservation;

[0068] Step 3: Write the disposal record into the Hyperledger Fabric ledger to ensure that the data cannot be tampered with.

[0069] Specifically, the implementation steps of the cross-domain risk conduction model: Generate the warehouse point cloud data through 3D laser scanning and divide it into 1m 3 grids; Collect the value, temperature and humidity data of each grid in real time and calculate the risk concentration; When the logistics monitoring coverage rate < 60%, D i is magnified by 1.8 times, triggering the update of the three-dimensional risk heat map; When the risk value of a certain grid > 0.8, automatically freeze the associated warehouse receipt financing quota.

[0070] This solution synchronously quantifies the risk diffusion speeds of the physical warehouse and the capital flow through the fluid equation, and the RFID tags and 5G cameras are used to realize the full-process monitoring of the warehouse, and the data is uploaded to the blockchain in real time.

[0071] Furthermore, the self-evolving defense module includes: an adversarial sample generator and an integrity feedback unit. The adversarial sample generator collects the historical vulnerability data of the industry and extracts relevant data and information to generate attack samples, including semantic tampering of contract terms and injection of false logistics trajectories;

[0072] The specific semantic tampering of clauses includes: initializing the BERT-Masked model to ensure the correct model parameters and configurations; inputting the contract clause text to be tampered with into the BERT-Masked model; using the masking mechanism of the BERT-Masked model to tamper with the key semantic information in the contract text and generating text samples with high perplexity, where the perplexity PPL > 120, to simulate the situation of an attacker maliciously tampering with contract clauses.

[0073] False logistics trajectory injection (P(l) ~ l -1-α , α = 1.5) of GPS noise specifically includes: generating GPS jump noise data that follows the Lévy distribution according to the characteristics of the Lévy distribution to simulate abnormal jump phenomena in the logistics trajectory; injecting the generated GPS jump noise data into the real logistics trajectory data to form false logistics trajectory samples to simulate the situation of an attacker tampering with logistics information.

[0074] Integrity feedback unit: Obtain the system operation logs and extract the information on the number of approval skips; use the formula: w ij := w ij ×(0.3 + 0.7e -0.5·bypass_count ) to calculate the weight adjustment factor, where w ij represents the edge weight between nodes i and j in the causal graph, bypass_count represents the number of approval skips, and apply the calculated weight adjustment factor to the edge weights of the causal graph to achieve dynamic adjustment of the causal graph.

[0075] Specifically, the implementation steps of the self-evolving defense module: Input the tampered contract text into the risk detection model to evaluate the model robustness; optimize the model using the Actor-Critic framework, and the reward function is designed as:

[0076] Statistically count the number of operation violations monthly and adjust the causal graph weights according to the formula.

[0077] Furthermore, the system's response to risks includes the following steps:

[0078] Step 1: When the hash value of the electronic claim voucher deviates from the historical mean by ±2σ, call the compliance verification layer to verify the CFCA signature status;

[0079] Step 2: If the signature confidence < 0.63, perform counterfactual causal intervention analysis:

[0080] ΔR = E[R|do(contract valid = False)] - E[R|contract valid = False]

[0081] Step 3: When ΔR > 1.4, freeze the associated account within 47 seconds and generate a risk transmission path report.

[0082] Counterfactual causal intervention analysis includes:

[0083] Construct a causal diagram containing 12 confounding variables;

[0084] Activate the independent audit process when the intervention effect size ΔR > 2.1.

[0085] During the counterfactual causal intervention analysis, the system presets and processes the following twelve key confounding variables:

[0086] I. Market environment variables

[0087] 1. Interbank lending rate volatility: Calculate the annualized volatility using the GARCH(1,1) model and update the data window quarterly on a rolling basis;

[0088] 2. Industry prosperity index: Perform Z-score standardization on the original PMI value and map it to the [-1, 1] interval.

[0089] II. Policy compliance variables

[0090] 3. Cross-border payment compliance index: Generate a 0-1 standardized score based on the fuzzy comprehensive evaluation method and synchronize the update every 24 hours;

[0091] 4. Data sovereignty compliance score: Binary identification conversion, 0 indicates the existence of compliance loopholes, and 1 indicates full compliance.

[0092] III. Operational risk variables

[0093] 5. Probability of signature device tampering: The probability value output by the Bayesian anomaly detection algorithm, accurate to four decimal places;

[0094] 6. Risk of operator identity impersonation: Generate a 0-100 risk score using the multimodal fusion scoring algorithm, and trigger an alarm when the score is above 80.

[0095] IV. Legal validity variables

[0096] 7. Transferability identifier of electronic claim vouchers: Boolean logic mapping, 0 indicates prohibited transfer, and 1 indicates permitted transfer;

[0097] 8. Completeness of contract dispute resolution clauses: Semantic matching score based on the BERT model.

[0098] V. Supply chain topology variables

[0099] 9. Depth of the core enterprise guarantee chain: Network layer counting method, L1 is directly associated, and the count increases by one for each additional level of extension;

[0100] 10.Supplier Substitutability Index: calculated using the Herfindahl-Hirschman Index (HHI), with the formula Σ(market share2)×10000.

[0101] 6. Technical Credibility Variables

[0102] 11. Blockchain consensus delay: take the median of the confirmation time of the last 100 transactions;

[0103] 12. Invoice verification hash conflict rate: the percentage of conflicting invoices to the total number of verifications.

[0104] All confounding variables need to complete the following preprocessing before entering the counterfactual model:

[0105] 1. Missing value filling: Use multivariate interpolation with chain equations (MICE), with ≥ 5 iterations

[0106] 2. Dimensional unification: Scale each variable to the interval [0,1] through min-max normalization

[0107] 3. Collinearity test: calculate the variance inflation factor (VIF) and remove redundant variables with VIF>10

[0108] In one example: When the system detects that a factoring financing request has a risk of duplicate financing, the latest data of the above 12 confounding variables are automatically loaded. For example:

[0109] If the volatility of interbank lending rates suddenly increases to 18.7% (threshold 15%)

[0110] The invoice verification hash collision rate is 2.3% (the industry average is 0.7%)

[0111] The system will activate a dual machine learning model, first calculating the propensity score through XGBoost, then predicting the potential results with random forest, and finally outputting the intervention effect value. When Δ>1.2, the fund transfer will be frozen immediately and a risk disposal report will be generated.

[0112] It should be understood that the embodiments disclosed in the present invention and the above description can enable those skilled in the art to use the present invention to implement the present invention. At the same time, the present invention is not limited to the above-mentioned embodiments. It should be understood that those skilled in the art can still modify the technical solutions recorded in the above-mentioned embodiments, or replace some of the technical features therein by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. A supply chain financial risk management system based on credit assessment, characterized in that: include: Dynamic data modeling unit, causal analysis engine, cross-domain risk transmission model, self-evolving defense module; The dynamic data modeling unit constructs a five-dimensional data model including the transaction subject level L and the contract legal effect C, and maps the supply chain data into a N×M×t×L×C dimensional tensor, where N is the set of participating entities, M is the set of business indicators including accounts receivable turnover rate and factoring financing ratio, t is the dynamic time slice granularity, and the L dimension records the hierarchical relationship between multi-level suppliers and core enterprises. When the L3 supplier contract signature confidence CFCA is less than 0.4 or the factoring financing ratio is greater than When the blockchain verification process of the buyer's payment ability is triggered automatically at the core enterprise level; the C dimension calculates the contract legal effect score through the hash value of the electronic signature, and additionally verifies the receivables for the factoring contract, confirms the ownership and keeps the evidence, and synchronizes the registration status with the registration agency in real time; The causal analysis engine implements a dynamic update algorithm for the causal adjacency matrix with compliance constraints: using a dual-channel update mechanism of history retention and incremental learning, 93% of the weight of the adjacency matrix at time t is merged with 7% of the dynamic increment. The dynamic increment is generated by performing a nonlinear transformation on the splicing vector of the temporal feature encoding and the hidden state through a trainable gating matrix; a legal effect attenuation function is injected, based on the contract signature confidence c of the transaction parties. i 、c j The joint legal effectiveness score is calculated with the real-time compliance index q, and a nonlinear attenuation coefficient is generated through an S-shaped curve with an attenuation slope parameter of 2.5; when c i +c j When <0.8, the compliance accelerated decay mechanism is activated, increasing the associated weight decay rate to 300% of the baseline value.

2. A supply chain financial risk management system based on credit assessment as claimed in claim 1, characterized in that: When it is detected that the L3 supplier contract signature confidence CFCA is less than 0.4, the Hyperledger Fabric chain evidence is called to verify the L1 core contract; the data dimensions are compressed through Tucker decomposition, the main components with variance contribution rate ≥ 85% are retained, and a data quality report is generated.

3. A supply chain financial risk management system based on credit assessment as claimed in claim 1, characterized in that: The cross-domain risk transmission model includes: a physical space risk mapping unit and a digital space diffusion calculation unit, specifically including: Using the physical space risk mapping unit, the warehouse is discretized into 1m 3 The voxel grid monitors the value of goods, temperature and humidity, and risk concentration in real time, and records the volatility of inventory value and logistics monitoring coverage; The diffusion coefficient is calculated by using the inventory value volatility and logistics monitoring coverage, combined with the time variable t and the decay constant λ, through the digital space diffusion calculation unit. The risk fluid dynamics equation is used to quantify the risk transmission speed of capital flow: Among them, R i represents the risk concentration of region i, γ i is the attenuation coefficient, β is the conversion rate, S j and I j They represent the capital flow intensity and investment level of region j respectively; If the logistics monitoring coverage is less than 60%, D i Zoom in 1.8 times to trigger the update of the warehouse's 3D risk heat map.

4. A supply chain financial risk management system based on credit assessment as claimed in claim 3, characterized in that: The physical space risk mapping unit performs the following steps: Step 1: Obtain warehouse 3D point cloud data through IoT sensors and generate voxel grids; Step 2: When the risk value of any grid is greater than 0.8, the associated warehouse receipt financing amount will be automatically frozen, and the blockchain smart contract will be activated to execute asset preservation; Step 3: Write the disposal record into the Hyperledger Fabric ledger to ensure that the data cannot be tampered with.

5. A supply chain financial risk management system based on credit assessment as claimed in claim 1, characterized in that: The self-evolution defense module includes: an adversarial sample generator and an integrity feedback unit. The adversarial sample generator collects industry historical vulnerability data and extracts relevant data and information to generate attack samples, including semantic tampering of contract terms and injection of false logistics tracks; The semantic tampering of the clauses specifically includes: initializing the BERT-Masked model to ensure that the model parameters and configuration are correct; inputting the contract clause text to be tampered into the BERT-Masked model; using the masking mechanism of the BERT-Masked model to tamper with the key semantic information in the contract text, generating a text sample with a high perplexity, the perplexity > 120, to simulate the situation where the attacker maliciously tampered with the contract clauses; The false logistics trajectory injection specifically includes: generating GPS jump noise data that obeys the Levy distribution according to the characteristics of the Levy distribution to simulate the abnormal jump phenomenon in the logistics trajectory; injecting the generated GPS jump noise data into the real logistics trajectory data to form a false logistics trajectory sample to simulate the situation where an attacker tamperes with the logistics information.

6. A supply chain financial risk management system based on credit assessment as claimed in claim 5, characterized in that: The integrity feedback unit: obtains the system operation log, extracts the number of approval skipping information therein; uses the formula according to the number of approval skipping: w ij :=w ij ×(0.3+0.7e -0.5·bypass_count ) calculates the weight adjustment factor, where w ij It represents the edge weight between node i and node j in the causal graph, bypass_count represents the number of times the approval is skipped, and the calculated weight adjustment factor is applied to the edge weight of the causal graph to achieve dynamic adjustment of the causal graph.

7. A supply chain financial risk management system based on credit assessment as claimed in claim 1, characterized in that: The system includes the following steps to deal with the risk: Step 1: When the hash value of the electronic debt certificate deviates from the historical mean ±2σ, the compliance verification layer is called to verify the CFCA signature status; Step 2: If the signature confidence is less than 0.63, perform counterfactual causal intervention analysis: ΔR=E[R|do(contract valid=False)]-E[R|contract valid=False] Step 3: When ΔR>1.4, freeze the associated accounts within 47 seconds and generate a risk transmission path report.

8. A supply chain financial risk management system based on credit assessment as claimed in claim 7, characterized in that: Counterfactual causal intervention analysis includes: Construct a causal diagram containing 12 confounding variables; The independent audit process was activated when the intervention effect size ΔR>2.1.

Images