Credit risk data processing method and system based on artificial intelligence driving

Abstract

The application discloses an artificial intelligence driving-based credit risk data processing method and system, belongs to the technical field of data processing, and comprises the following steps: acquiring qualification risk characteristics, constructing a supplier archive library, combining an evaluation index system, generating a management and control index set, performing risk identification on external data, generating a risk index set, utilizing multi-source data fusion to generate a supply dynamic vector; setting a trigger update mechanism, monitoring the supply dynamic vector in real time, combining a dynamic trigger threshold, determining an environmental mutation event, thereby generating an event trigger signal, triggering a consensus determination mechanism, initiating an update voting request, performing consensus determination, generating a task update instruction, acquiring an incremental training data set, performing incremental training on a benchmark evaluation model, thereby generating a candidate evaluation model; utilizing the supply dynamic vector and the candidate evaluation model to perform real-time credit evaluation and grading, and guaranteeing the accuracy of supplier supervision and the stability of a supply chain under complex market environments.

Description

Technical Field

[0001] This invention relates to a credit risk data processing method and system driven by artificial intelligence, belonging to the field of data processing technology. Background Technology

[0002] In supply chain management scenarios, supplier credit assessment requires the integration of heterogeneous data from multiple sources and the output of quantitative assessment results based on assessment models. With the development of digital technology, blockchain, with its immutable and traceable characteristics, provides effective technical support for the reliable storage and sharing of supplier credit data.

[0003] A Chinese patent with authorization announcement number CN111222955B discloses a blockchain-based supplier supervision method and system. This method involves acquiring basic supplier data and procuring entity evaluation information, storing this data on the blockchain, selecting factors with high credit relevance and assigning weights, calculating supplier credit evaluation results using a factor-weighted average algorithm, and then registering the evaluation results on the blockchain. Essentially, this achieves credible data storage and on-chain static weighted evaluation. However, the evaluation logic remains at the level of linear weighting with manually preset weights, lacking the ability to adaptively adjust evaluation parameters based on data.

[0004] Regarding dynamic risk assessment, patent application CN122264921A proposes a dynamic early warning system and method for corporate credit risk based on behavioral changes. This system triggers early warning events by collecting real-time data on changes in corporate behavior and comparing it with preset thresholds. However, the assessment model itself remains static; once the early warning threshold is set, it remains fixed and cannot be adaptively adjusted based on changes in the market environment or incremental samples. When structural changes occur in the external environment, it cannot distinguish between individual supplier negligence and systemic environmental impacts, easily leading to false alarms or missed alarms.

[0005] For multi-source data fusion modeling, patent application CN121685139A proposes a joint modeling system and method for a large-scale financing credit model based on multi-source data. This involves multi-source data acquisition, standardization processing, privacy computation, and feature drift detection, and automatically triggers dynamic adjustments to model weights based on feature input distribution shifts. However, update decisions rely entirely on feature drift detection at the data level, lacking a verification mechanism at the business semantic level. The singular data-driven logic cannot distinguish whether statistical shifts originate from real risk changes. Without business rule constraints, it is difficult to guarantee the compliance and rationality of update decisions in supply chain assessment scenarios.

[0006] Furthermore, none of the aforementioned existing technologies have solved the problem of interpretability of model decision-making logic. Whether it is a static linear weighted model or a black box model based on deep learning, business personnel cannot understand the internal decision-making logic of the model. When the evaluation results conflict with business cognition, there is no way to conduct effective verification and intervention, making it difficult for the model output results to gain sufficient trust in actual business processes. Therefore, this invention proposes a credit risk data processing method and system based on artificial intelligence. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the present invention aims to provide a credit risk data processing method and system based on artificial intelligence. By integrating internal and external dynamic data to generate a supply dynamic vector and storing it on the blockchain, and generating update instructions through dynamic triggering and consensus mechanisms, the method overcomes the static and black-box problems of the model by using an incremental data iterative model and full-chain traceability, thereby achieving accurate assessment.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] AI-driven credit risk data processing methods include:

[0010] Based on internal business datasets, qualification risk characteristics are obtained, a supplier archive is built, and a set of control indicators is generated by combining the evaluation indicator system; external data is collected, risk indicator sets are generated based on risk identification, and a supply dynamic vector is generated by multi-source data fusion and stored on the blockchain.

[0011] Configure a trigger update mechanism on the blockchain to monitor the supply dynamic vector in real time, and combine it with a dynamic trigger threshold to determine environmental change events, thereby generating event trigger signals;

[0012] Upon receiving the event trigger signal, the consensus determination mechanism is immediately activated, an update voting request is initiated, consensus determination is performed, and a task update instruction is generated.

[0013] Based on the task update instruction, incremental data is extracted from the supplier archive to generate an incremental training dataset, and the benchmark evaluation model is incrementally trained to generate a candidate evaluation model.

[0014] Real-time credit assessment and classification are performed using supply dynamic vectors and candidate evaluation models.

[0015] Specifically, the steps for building a supplier database include:

[0016] Collect the identity credentials and basic information submitted by suppliers, and generate a preliminary identity information review list;

[0017] A two-factor binding mechanism is used to generate a unique digital identity for the supplier;

[0018] Set up standardized interface protocols and define data field mapping rules for different business terminals;

[0019] Configure a dual-mode synchronization strategy to collect synchronized data, generate internal business datasets, and perform quality control based on field weight integrity verification.

[0020] The qualification documents in the internal business dataset are subjected to hierarchical identification processing, valid fields are extracted, and cross-verification is performed to generate authenticity verification results.

[0021] Enter the verified valid fields into the qualification dynamic ledger, set qualification early warning thresholds and qualification validity thresholds, classify qualification status based on the current date, and generate qualification risk characteristics of suppliers;

[0022] Using the digital identity identifier as the core index, a structured supplier archive is built using a hierarchical storage architecture, and a data association engine is constructed.

[0023] Specifically, the steps for generating the supply dynamic vector include:

[0024] Set evaluation objectives and candidate indicators, and obtain the evaluation dataset from the supplier database;

[0025] Using the evaluation dataset, the similarity between candidate indicators and evaluation targets is calculated, and relevant indicators are selected.

[0026] Acquire historical business data, use the coefficient of determination to calculate the fitting contribution, and screen basic indicators;

[0027] Using the analytic hierarchy process (AHP) and entropy weighting method, the initial weights of the basic indicators are calculated to generate a set of basic indicators.

[0028] Set the monitoring update interval and extract historical signals within each monitoring update interval;

[0029] The validity of historical signals is determined by preset triggering rules in order to make basic adjustments;

[0030] Obtain the data items required for basic indicators, map them to the supplier archive, define extraction rules, and construct a path mapping table;

[0031] The raw data of each basic indicator is extracted from the supplier archive by using a multi-dimensional search index.

[0032] The interquartile range method is used to identify abnormal data, and automatic correction is performed based on similar historical data.

[0033] The basic indicator data are normalized and combined with the initial weights to generate a set of control indicators.

[0034] Specifically, the steps for generating the supply dynamic vector also include:

[0035] Collect external data from multiple sources, count the number of times the same information is verified from different sources to obtain a consistency verification coefficient, and calculate the external credibility by combining the source weights.

[0036] External data with credibility higher than the credibility rating threshold is screened, and risk identification is performed. Combined with qualification risk characteristics, a risk indicator set is generated.

[0037] Based on the control indicator set and the risk indicator set, a related dataset is generated using digital identity identifiers as the primary alignment pair and timestamps as the assignment alignment pair.

[0038] Set association weights for each indicator in the associated dataset, obtain environmental impact coefficients and business priority coefficients, calculate feature weights, and perform normalization processing to generate dynamic feature weights.

[0039] The indicators in the associated dataset are weighted, fused, and standardized to generate a supply dynamic vector;

[0040] The supply dynamic vector and its corresponding metadata are standardized and assembled, and combined with the block header information to generate a storage data block.

[0041] Specifically, the steps to trigger the update mechanism include:

[0042] Set the initial trigger threshold and trigger conditions; triggering at any threshold will determine the existence of a potential mutation event.

[0043] A dynamic threshold calibration mechanism is set up. Based on a sliding window, historical data from the past N business cycles are statistically analyzed to calculate the threshold correction coefficient, which is then used to correct the initial trigger threshold and generate a dynamic trigger threshold.

[0044] Calculate the mean of all indicator values ​​in the supply dynamic vector within the current period, combine it with the historical mean to obtain the supply deviation rate, extract the dynamic indicators in the supply dynamic vector of the current period, obtain the risk volatility coefficient, and generate the volatility dataset for the current period.

[0045] The fluctuation dataset is compared one by one with the dynamic trigger threshold to determine whether there are potential mutation events and generate a single-dimensional trigger result.

[0046] By combining qualification risk characteristics and external risk levels, the single-dimensional triggering results are verified, and environmental mutation events are generated based on the effectively triggered potential mutation events.

[0047] Specifically, the initial trigger thresholds include the initial volatility threshold, the initial mutation threshold, and the initial risk threshold; the steps for setting the initial trigger thresholds include:

[0048] Obtain the historical supply dynamic vector and calculate the volatility deviation rate using the historical standard deviation and historical mean;

[0049] Obtain the weighting coefficient and quality correction coefficient, and combine them with the fluctuation deviation rate to obtain the initial fluctuation threshold;

[0050] For a single indicator, obtain the historical average value of the indicator, and at the same time, set the mutation coefficient according to the evaluation target to which the indicator belongs, and obtain the initial mutation threshold.

[0051] Set a threshold for the number of risk events, and obtain the level weight and event impact coefficient of each individual risk event. Then, obtain an initial risk threshold by weighting the risk event threshold.

[0052] Specifically, the consensus determination mechanism includes the following steps:

[0053] Construct a three-layer verification node structure and set the responsibilities and weights of each layer of nodes;

[0054] Configure basic voting parameters, including voting options, voting validity period, and voting reasons;

[0055] Once an environmental mutation event is received, an update voting request is automatically sent to the three-layer verification nodes, and the voting results of all nodes are automatically collected after the voting validity period expires.

[0056] The voting scores of each node are assigned based on the scores of each voting option. The weighted average of the consensus is calculated by combining the layer weights of each node and a consensus determination is made.

[0057] If the weighted average is greater than the first threshold, a consensus is reached; if the weighted average is between the first and second thresholds, the opinions of the nodes that suggest the adjustment are extracted, an adjustment plan is generated, and a new vote is initiated; if the weighted average is less than the second threshold, a consensus is not reached.

[0058] Specifically, the steps for generating candidate evaluation models include:

[0059] By using a multi-dimensional retrieval index, incremental data is extracted from the supplier archive, and the incremental data is subjected to hierarchical filtering, preprocessing and enhancement to generate an incremental training dataset.

[0060] Obtain the existing benchmark evaluation model in the model version library, use the incremental learning algorithm, take the incremental training dataset as input, freeze the bottom feature extraction layer of the model, and only fine-tune the parameters of the top decision layer to generate candidate evaluation models.

[0061] An offline test set is constructed, and dynamic scenario data is simulated and generated based on the mutation type in the task update instruction. The performance differences between the candidate evaluation model and the benchmark evaluation model are evaluated, and a test report is generated.

[0062] The candidate evaluation models that pass the test are assigned version numbers and packaged with the corresponding metadata to generate version model packages.

[0063] Specifically, the steps for real-time credit assessment and rating include:

[0064] Obtain the real-time evaluation model and real-time feature vector, combine the dynamic weights of the features, perform a weighted summation, and combine the quality correction coefficient to calculate the real-time credit score;

[0065] Historical credit scores are obtained, and the K-means clustering algorithm is used to divide the initial clustering thresholds. The threshold adjustment coefficient is calculated to generate dynamic grading thresholds.

[0066] The supplier level is determined based on the dynamic grading threshold and the real-time credit score;

[0067] Build a dynamic binding system to configure differentiated automated business rules for suppliers of different tiers;

[0068] Based on the hierarchical results, the corresponding business rules are automatically triggered, execution instructions are generated, the execution status of the rules is monitored in real time, the execution results returned by the business terminal are obtained, and the execution results are synchronized to the supplier archive.

[0069] The AI-driven credit risk data processing system includes: a data acquisition module, an update triggering module, and an evaluation and optimization module.

[0070] The data acquisition module is used to obtain qualification risk characteristics, build a supplier database, generate a set of control indicators based on the evaluation indicator system, collect external data synchronously and perform risk identification, generate a set of risk indicators, and generate a supply dynamic vector by using multi-source data fusion.

[0071] The update triggering module is used to monitor the fluctuation of each indicator in the supply dynamic vector on the chain in real time using a pre-dynamic triggering threshold, determine environmental change events, and generate an event triggering signal to trigger the consensus determination mechanism, initiate an update voting request, perform consensus determination, and generate task update instructions.

[0072] The evaluation and optimization module is used to generate an incremental training dataset according to the task update instruction, to incrementally train the benchmark evaluation model, generate candidate evaluation models, and perform real-time credit evaluation and classification based on the latest supply dynamic vectors of each supplier and the candidate evaluation models.

[0073] The beneficial effects of this invention are:

[0074] By integrating internal business data with external multi-source data, a dynamic supply vector representing the real-time status of suppliers is generated and stored on the blockchain. This ensures that the data is both dynamic and trustworthy, breaking the limitations of traditional static datasets and providing high-quality data support for the model to adapt to dynamic environments. Simultaneously, the blockchain-triggered update mechanism monitors vector feature fluctuations in real time and accurately determines sudden environmental events by combining dynamic trigger thresholds, avoiding mechanical responses from the model to external changes such as force majeure. A consensus-based decision-making mechanism generates task update instructions, and the benchmark evaluation model is trained incrementally based on incremental data. This preserves the model's historical learning results while enabling rapid adaptation to new data patterns, solving the problem of lagging updates in traditional static models. Based on the iterative model, real-time credit assessment and grading are achieved. The grading results dynamically coordinate with business terminals and trigger differentiated automated rules, ensuring that assessment value is quickly implemented. This avoids decision-making misguidance caused by the disconnect between traditional assessment results and actual business operations, protects the cooperation enthusiasm of high-quality suppliers, reduces supply chain cooperation risks and efficiency losses, effectively meets the needs of enterprises for precise supplier supervision in complex and volatile market environments, and ensures the stability and sustainability of the supply chain. Attached Figure Description

[0075] Figure 1 This is a schematic diagram of an AI-driven credit risk data processing method.

[0076] Figure 2 This is a flowchart illustrating the generation of the supply dynamic vector in this invention;

[0077] Figure 3 This is a flowchart of the update triggering mechanism in this invention;

[0078] Figure 4 This is a flowchart of the consensus determination mechanism in this invention. Detailed Implementation

[0079] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0080] Example 1

[0081] refer to Figures 1 to 4 As shown in the figure, this embodiment introduces a credit risk data processing method based on artificial intelligence, including the following steps:

[0082] Based on supplier terminals, a unique digital identity is assigned to each supplier, and basic information is integrated to generate a preliminary identity information review list. Through preset interface protocols and data synchronization rules, it connects with various business terminals, such as ERP terminals, procurement terminals, and warehousing and acceptance terminals, via API interfaces to obtain internal business datasets, including order data, warehousing data, and settlement data, ensuring the real-time performance and consistency of internal data. Simultaneously, optical character recognition is used to intelligently identify supplier qualification documents and extract valid fields, such as validity period, certification level, and license scope, thereby establishing a dynamic qualification ledger for each supplier. At the same time, a qualification validity threshold is set to classify the supplier's qualification status by validity period, including normal, near expiration, and expired, thereby obtaining the supplier's qualification risk characteristics. All relevant information of the same supplier is linked to the supplier archive, so that the internal business dataset, dynamic qualification ledger, and qualification risk characteristics are saved in the corresponding locations in the supplier archive.

[0083] An evaluation indicator system is set up based on the company's internal needs. Feature data corresponding to the dimensions are extracted from the supplier archive. Through data cleaning and standardization, a set of control indicators is generated. Simultaneously, external data related to suppliers is collected through a pre-set external data interface. Based on risk identification, semantic parsing and risk event extraction are performed on the external data to generate external risk features. Combined with qualification risk features, a set of risk indicators is generated. Using multi-source data fusion, the control indicator set and the risk indicator set are aligned and normalized to generate a supply dynamic vector to represent the supplier's real-time operating status, risk level, and cooperation capability. The supply dynamic vector is then standardized and assembled with the corresponding timestamp, data source hash value (linked to the supplier archive), and data processing node information to generate a data storage block. This block is then uploaded to the blockchain for storage, generating a storage certificate, including the block hash, storage time, and node information, to ensure the immutability and traceability of the data.

[0084] Configure a trigger update mechanism on the blockchain, set dynamic trigger thresholds based on historical data, including overall feature fluctuation thresholds, single core indicator mutation thresholds, and risk event accumulation thresholds. Through the blockchain data monitoring interface, monitor the fluctuation of each indicator in the supply dynamic vector on the chain in real time. Determine environmental mutation events by using dynamic trigger thresholds and fluctuation data, thereby generating event trigger signals. Upon receiving an event trigger signal, immediately activate the consensus determination mechanism, initiate update voting requests to verification nodes with different permission roles, conduct consensus determination, generate task update instructions, associate the event trigger reason, voting results, and verification node signatures, and broadcast them on the chain to ensure the compliance and credibility of the model update process.

[0085] Once a task update instruction is detected, incremental data is extracted from the supplier archive based on the data range associated with the instruction. This includes newly generated data after the triggering event and long-tail data that has not been fully learned in the past. Through classification, filtering, and preprocessing, an incremental training dataset is generated to incrementally train the existing benchmark evaluation model in the supplier archive, thereby generating candidate evaluation models. The candidate evaluation models are assigned version numbers and packaged with corresponding metadata, such as supply dynamic vectors, task update instructions, and fine-tuning parameter configurations, to generate a version model package. The hash value of the version model package is obtained through the blockchain notarization interface and uploaded to the blockchain.

[0086] Based on the latest supply dynamic vectors and candidate evaluation models of each supplier, real-time credit assessment and classification are performed. The assessment results and classification information are then synchronized to the business terminal through a preset interface protocol, triggering the execution of business rules to guide order allocation and contract renewal.

[0087] Specifically, the construction of the supplier database includes:

[0088] Based on the supplier terminal, the system collects the identity credentials and basic information submitted by the supplier. The identity credentials include the unified social credit code, the legal representative's identity certificate, and the enterprise registration certificate. The basic information includes the enterprise name, registered address, contact information, and product / service category information. The system automatically summarizes these to generate a preliminary list of identity information. A two-factor binding mechanism is adopted, using the unified social credit code as the basic identifier and combining it with the SHA-256 hash algorithm to generate a unique digital identity identifier for the supplier. The digital identity identifier, identity credentials, and basic information are packaged into an identity anchor data package and uploaded to the consortium blockchain for notarization, forming an identity anchor certificate.

[0089] Based on the enterprise's existing business architecture, set up standardized interface protocols, such as WebService, and formulate unified data field mapping rules for different business terminals. Standardize fields with the same meaning but different names in different terminals. For example, unify the order number of the purchasing terminal and the associated order number of the receiving and acceptance terminal into a unique business identifier, forming a field mapping rule table.

[0090] Deploy a data synchronization middleware and configure a dual-mode synchronization strategy that includes real-time triggering and scheduled completion to collect and synchronize data. When new data is generated by the business terminal, such as the creation of a purchase order or the completion of warehousing acceptance, real-time synchronization is triggered. Scheduled completion is performed once an hour to verify and complete the data missing during the real-time synchronization process in order to generate a standardized internal business dataset.

[0091] Integrity verification based on field weights performs quality control on synchronized data in the internal business dataset. This includes: assigning interval weights to each business field and calculating the integrity score of the synchronized data. If the integrity score is less than a preset integrity threshold, a data completion notification is automatically sent to the corresponding business terminal. After the business terminal completes the data, the dual-mode synchronization strategy is re-executed to ensure the standardization and completeness of the final internal business dataset. Specifically, after obtaining all historical batches of complete field entries, the existing data of supplier credit scores and rating results are calculated normally and used as the field dataset. Each historical sample in the field dataset is traversed, and a specific business field is deleted one by one, while all other fields remain unchanged. The entire credit assessment process is re-run to generate a new version of the credit score after deleting the field. The absolute difference between the original credit score and the new version of the credit score after deleting the field is calculated as the assessment impact of the current field in a single sample. The average impact of the assessment impact of a single business field on all historical samples is calculated to obtain the average impact difference of the corresponding business field. The average impact differences are then sorted in descending order and divided according to the sorting position. The weighting intervals are as follows: the top 30% of fields are high-weighted intervals, the middle 40% are medium-weighted intervals, and the bottom 30% are low-weighted intervals. Within each interval, the interval weight of the corresponding business field is obtained by taking the average impact difference of the corresponding business field in the same interval as the sum of the average impact differences of all business fields in the interval. For the field dataset, random field missing is performed on a single historical sample one by one, i.e., multiple rounds of simulation are performed with 1 field missing and 2 fields missing respectively. In each round, the number of missing fields is fixed, and the corresponding number of fields are randomly selected and deleted. The credit score is recalculated. The number of missing fields for each category is calculated. The field missing ratio is the ratio of the number of missing fields to the total number of business fields in the corresponding historical sample. The credit score calculated with the original complete fields is used as the baseline truth. The new score generated after each round of random field missing is compared with the baseline truth. When the credit rating result of the same supplier changes in two consecutive evaluations under a certain missing ratio, the field completeness score corresponding to the sample at this time is recorded. The completeness scores corresponding to the first rating jump of all historical samples are summarized, and the maximum value among all completeness scores is taken as the completeness threshold.

[0092] The verified qualification documents are processed in layers to identify document types, such as business licenses and ISO certification certificates. Based on the document type, the corresponding field identification template is called to extract valid fields, such as validity period, certification level, scope of license, issuing authority, and date of issuance. Cross-verification is then performed through third-party authoritative institution interfaces, such as the State Administration for Market Regulation's Enterprise Credit Information Disclosure System interface, to generate authenticity verification results, including pass or fail.

[0093] The verified field information is entered into the qualification dynamic ledger. A basic early warning cycle is determined based on the qualification type. The qualification early warning threshold is calculated based on the difference between the expiration date and the basic early warning cycle. The expiration date is used as the qualification validity threshold. The qualification status is divided based on the qualification early warning threshold and the qualification validity threshold. If the current date is less than the qualification validity threshold, the qualification status is normal. If the current date is between the qualification early warning threshold and the qualification validity threshold, the qualification status is nearing expiration. If the current date is greater than the qualification validity threshold, the current qualification status is expired. Finally, the supplier's qualification risk characteristics are generated and associated with the supplier's digital identity. Among them, all historical processing records of the same qualification type from the initiation of the certificate renewal application to the completion and archiving of the review are obtained, and the arithmetic average of the processing time is calculated as the basic early warning cycle for the corresponding qualification type.

[0094] Using digital identity identifiers as the core index, a layered storage architecture is adopted to build a structured supplier archive. This embodiment selects MySQL as the business structured data storage engine, a consortium blockchain to store supplier digital identity hash credentials, and HDFS distributed file system to store PDF and image format qualification attachments. The three-layered data table structure is as follows: The first layer is the identity foundation layer, which stores digital identity identifiers, core identity credentials, and blockchain-based evidence certificates, such as supplier ID, unified social credit code, SHA256 digital identity, identity evidence hash, registered address, and legal person information primary key fields. The second layer is the business data layer, which stores standardized internal business datasets, such as order numbers, delivery data, settlement records, and warehousing and acceptance records. The third layer is the qualification risk layer, which stores dynamic qualification ledgers, authenticity verification reports, and qualification risk characteristics, and sets data association rules, namely, using the globally unique digital identity identifier as the primary matching key and the business unique identifier as the secondary key. To assist in cross-layer entity binding via association keys, multiple duplicate data entries under the same digital identity are automatically retained with the latest verification entries to build a data association engine. This includes unique matching of digital identities and cross-layer association of unique business identifiers. Based on data association rules, automatic cross-layer data binding is achieved, accurately binding the supplier name in the identity foundation layer with the order owner in the business data layer and the qualification owner in the qualification risk layer. This ensures the logical consistency of all data from the same supplier. A data version monitoring mechanism is introduced, assigning version numbers to each updated business data and qualification information, recording changes before and after data updates, update nodes, and operators to form a data update log. Multi-dimensional indexes are built, including digital identity indexes, business type indexes, qualification type indexes, and time range indexes, supporting multi-condition combined searches, such as searches by supplier digital identity, business scenario, qualification type, and time range, improving the efficiency of subsequent data extraction.

[0095] Specifically, the steps for generating a risk indicator set include:

[0096] The evaluation objectives are set, including operational capabilities, compliance risks, and cooperation suitability. Operational capabilities are used to evaluate delivery efficiency and quality stability, compliance risks are used to evaluate qualification compliance and behavioral compliance, and cooperation suitability is used to evaluate the degree of matching with the company's development. Candidate indicators are selected based on the evaluation objectives. For example, candidate indicators for operational capabilities include on-time delivery rate, quality pass rate, order fulfillment rate, and inventory turnover rate. The quantitative data of the evaluation objectives and the raw data of the candidate indicators are extracted from the supplier database to generate an evaluation dataset.

[0097] Based on the Pearson correlation coefficient, the similarity between each candidate indicator and each evaluation target is calculated through the evaluation dataset. Weakly correlated indicators with similarity less than a preset correlation threshold are eliminated, and the remaining candidate indicators are marked as correlated indicators. Specifically, the historical evaluation dataset of suppliers within the enterprise for the past 3 years is obtained, the Pearson correlation coefficients between all candidate indicators and each evaluation target are extracted, and the calculated Pearson correlation coefficients are sorted in ascending order. The lower quartile is used as the correlation threshold.

[0098] Historical business data is acquired, and linear fitting is performed with historical evaluation results as dependent variables and relevant indicators as independent variables. The coefficient of determination is used to measure the fitting contribution. Relevant indicators with fitting contribution higher than the preset fitting threshold are retained as the basic indicators of the evaluation target. Specifically, for the historical evaluation dataset, the coefficient of determination of all relevant indicators is extracted, and the coefficients are sorted in descending order. The coefficients corresponding to the bottom 20% of low-contribution relevant indicators are removed, and the minimum value of the remaining coefficient of determination is taken as the fitting threshold.

[0099] The relative importance of basic indicators under their respective evaluation objectives is calculated using the analytic hierarchy process (AHP) and entropy weight method, which serves as the initial weights for the basic indicators, thereby generating a set of basic indicators.

[0100] Set the monitoring update interval and extract historical signals within each monitoring update interval, including environmental trend signals and business change signals (if they exist).

[0101] The validity of historical signals is determined through preset trigger rules. If the environmental trend signal meets the trend quantification conditions or a business change signal is detected, the historical signal is deemed valid, triggering basic adjustments; otherwise, the historical signal is deemed invalid, and the original set of basic indicators is maintained. Trend quantification conditions include industry standard iterations exceeding a preset standard cycle, market fluctuations exceeding a preset fluctuation threshold, or compliance standard changes exceeding a change threshold. Basic adjustments include adding indicators, removing indicators, and weight adjustments. Added indicators are used to screen candidate indicators for uncovered evaluation targets. If new evaluation targets exist, removal indicators are selected based on fit contribution. Weight adjustments are used to screen basic indicators for changes in information entropy exceeding a certain threshold. The initial weights of the basic indicators with preset amplitude thresholds are adjusted. Among them, historical statistical data on industry supervision, market prices, and compliance documents for the past 5 years are obtained. The average number of times the official regulatory documents for compliance, production, and qualification categories of the enterprise's suppliers' industries are added, revised, or abolished each year is calculated. The ratio of the total number of times to the total number of years in the statistics is used as the regulatory period. A volatility dataset on market fluctuation amplitude for the past 5 years is obtained. The upper tertiary of the volatility dataset is used as the volatility threshold. The sequence of the proportion of revisions to compliance clauses over the years is obtained. The 70th percentile of the sequence is selected as the change threshold. The entropy sequence of historical indicator time series information is obtained. The entropy change amplitude corresponding to the first change in credit rating results in the sequence is located and used as the amplitude threshold.

[0102] Obtain the data items required for calculating basic indicators, such as the data items for on-time delivery rate, which are the number of on-time delivered orders and the total number of orders. Based on business expert experience and data field correlation analysis, map each data item to a specific level and field in the supplier archive. For example, the number of on-time delivered orders is mapped to the order delivery status field in the business data layer, and the total number of orders is mapped to the order creation record field in the business data layer. Clarify the extraction conditions for each data item to define extraction rules. Combine the basic indicator name, the evaluation target, the calculation logic, the data item, the archive data layer, the corresponding field name, and the extraction rules, and associate the initial weight of each basic indicator to construct a path mapping table to index and associate the basic indicators with the supplier archive.

[0103] Based on the path mapping table and the updated set of basic indicators, the data extraction requirements for each basic indicator are determined. Through multi-dimensional retrieval index, the original data corresponding to each basic indicator is extracted in batches from the business data layer and qualification risk layer of the supplier archive, and the data version number is associated.

[0104] The interquartile range method is used to identify abnormal data and determine the anomaly judgment range. Data exceeding the anomaly judgment range is marked as abnormal, and automatic correction is performed based on similar historical data. If it is numerical data, such as the quantity of goods received, the median of similar historical data is used to replace the original abnormal data. If it is time-based data, such as the delivery time, the average value of similar historical data is used to replace the original abnormal data. If it is status-based data, such as qualification status, the standard status value is used to replace the original abnormal data according to business compliance rules. Among them, similar historical data refers to similar business data of the same supplier within the same historical period.

[0105] The basic indicator data after anomaly correction is normalized using standardization methods to eliminate the impact of dimensional differences on the fusion results. Combined with the initial weights of the basic indicators, the standardized basic indicator data is weighted and summed to generate a set of control indicators. Each basic indicator value is associated with the supplier's digital identity, data version number, and data quality score. The data quality score is based on whether the data has undergone anomaly processing, including data without anomalies and data with anomaly correction.

[0106] Through pre-defined external data interfaces, publicly available multi-source external data related to suppliers is collected, such as data from authoritative institutions, industry platforms, and media. These multi-source external data are assigned different source weights. For the same information, the number of verifications from different sources is counted, with different verification counts corresponding to different consistency verification coefficients—that is, the percentage of correct verifications out of all verification counts. This yields the consistency verification coefficient for the current information. External credibility is calculated by multiplying the source weight by the consistency verification coefficient. If the same information has multiple sources with inconsistent source weights, the consistency verification coefficient for each source with different weights is obtained separately, and then weighted and summed to obtain the external credibility. Furthermore, the accuracy rate of matching past push information with official regulatory announcements is obtained, and this accuracy rate is used as the source weight for the corresponding data source.

[0107] External data with credibility exceeding the credibility rating threshold is screened, and risk identification is performed. Risk identification includes: acquiring historical external text labeled positive, neutral, and negative by manual annotation; constructing a risk keyword library; for each word, counting the negative frequency in negative samples and the positive frequency in positive samples; using the negative frequency divided by the sum of the positive and negative frequencies as the baseline score for the corresponding word; extracting risk keywords and related entities (such as supplier names) from the data; assigning values ​​based on keyword sentiment intensity, i.e., collecting the frequency of risk keywords in the data; and summing the product of the baseline score and frequency for each risk keyword to obtain the keyword sentiment score; filtering negative events through a negative screening threshold; and generating public opinion risk characteristics by combining the scope of impact, the event, and the keyword sentiment score; for market data, market risk is extracted based on industry benchmarks through deviation analysis. Features; combining qualification risk features to generate a risk indicator set; each risk feature is labeled with a credibility score, data source, and external risk level. The external risk level is determined based on a comprehensive judgment of external credibility and keyword sentiment score, that is, the product of external credibility and keyword sentiment score is used as the comprehensive risk value. At the same time, the comprehensive risk values ​​of all historical external data are obtained to construct a risk distribution sequence. The data in the risk distribution sequence are arranged in ascending order and equally divided into 3 levels, corresponding to three levels of external risk: low external risk, medium external risk, and high external risk. Among them, historical external data samples and results after manual authenticity labeling are obtained. All external credibility calculation results are traversed, and the external credibility with the best comprehensive accuracy and recall is selected as the credibility classification threshold. Historical public opinion data with manual positive and negative labels are obtained, and the keyword sentiment score at the boundary between positive and negative samples is used as the negative screening threshold.

[0108] Specifically, the steps for generating the supply dynamic vector include:

[0109] Based on the supplier's digital identity and timestamp, cross-dimensional data alignment is performed on the control indicator set and risk indicator set, with digital identity as the main alignment pair and timestamp as the assignment alignment pair, to generate a related dataset.

[0110] Assign association weights to each indicator in the associated dataset. For basic indicators, the corresponding initial weight is used as the association weight. For risk features, the corresponding source weight is used as the association weight. If a risk feature includes multiple source weights, the average of all source weights is taken as the association weight. Determine the environmental impact coefficient based on the external risk level. For each external risk level, obtain the median of the corresponding score interval in the risk distribution sequence as the impact base. Use the impact base of low external risk as the denominator and perform a ratio calculation on all impact bases to obtain the environmental impact coefficient for the corresponding external risk level. The environmental impact coefficient for low external risk is 1, and the environmental impact coefficient increases with the increase of the external risk level. Determine the business priority coefficient based on the adjustment of corporate development goals. Obtain the proportion of procurement amount for each type of material in the company's annual procurement documents and use the procurement proportion of a single type of material as the business priority coefficient for that type of supplier. Multiply the sum of the environmental impact coefficient and the business priority coefficient by the association weight, and add the association weight to obtain the feature weight of each indicator. Perform sum normalization on the feature weights of all indicators so that the sum of all feature weights is 1, thereby generating dynamic feature weights.

[0111] By using weighted fusion, the indicators in the associated dataset are fused, and the fused feature data is standardized using the max-min normalization method to generate a supply dynamic vector.

[0112] The supply dynamic vector and its corresponding metadata are standardized and assembled, including block body data and metadata. The block body data includes the supply dynamic vector, and the metadata includes the supplier's digital identity, timestamp, feature source hash (associated with the original data hash of the supplier archive), data processing node information (processing server IP, operator ID), and feature dynamic weight. Block header information is added to the block body data, including the previous block hash, the current block height (incrementing in the order of evidence storage), and the block generation timestamp, forming a standard evidence storage data block to ensure the chain association and traceability of block data.

[0113] Specifically, the steps to trigger the update mechanism include:

[0114] Based on the supply dynamic vector, supply dynamic vector metadata, and risk indicator set, initial trigger thresholds are set, including initial volatility threshold, initial mutation threshold, and initial risk threshold. The initial volatility threshold is used to evaluate the volatility of the overall indicators of the supply dynamic vector. Historical supply dynamic vectors are obtained, and the volatility deviation rate is calculated by the ratio of historical standard deviation to historical mean. At the same time, the number of all indicators in the feature dynamic weights is obtained, and the reciprocal of the number of indicators is used as the weighting coefficient.

[0115] Based on the supply dynamic vector, historical vector samples from the same batch are obtained, and the data quality score of each historical vector is calculated. The average of the data quality scores of all historical vectors is calculated to obtain the quality correction coefficient. The initial volatility threshold is obtained by multiplying the volatility deviation rate, weighting coefficient, and quality correction coefficient. The initial mutation threshold is used to assess the absolute volatility of each indicator in the supply dynamic vector. For a single indicator, historical indicators from all past business cycles of the supplier are obtained to construct an indicator time series set. The ratio of the standard deviation to the mean of the indicator time series set is used as the mutation coefficient. The initial mutation threshold is obtained by multiplying the historical indicator mean and mutation coefficient. The initial risk threshold is used to assess the risk. The risk event superimposed and fluctuated in the risk indicator set. The supplier’s historical full-cycle risk event records were retrieved, and the upper quartile value of the frequency of risk events in a single cycle was used as the risk event quantity threshold. At the same time, the environmental impact coefficient of a single risk event was obtained, and the total amount of purchase contracts and pending orders directly affected by this risk was obtained as the related amount. The total amount of all orders placed by the enterprise to the supplier in the same statistical period was obtained as the total purchase amount. The ratio of the related amount to the total purchase amount was used as the event impact coefficient of the corresponding risk event on the supplier. The initial risk threshold was obtained by weighting the level weight and the event impact coefficient and then multiplying it by the risk event quantity threshold.

[0116] For the initial trigger threshold, trigger conditions are set. Triggering any threshold is considered a potential mutation event. For a single metric, triggering any metric triggers the corresponding threshold. Simultaneously, to adapt to dynamic data changes and iterative business needs, a dynamic threshold calibration mechanism is implemented, based on a sliding window, statistically analyzing recent... Historical data from each business cycle are used to calculate threshold correction coefficients for the initial volatility threshold, initial mutation threshold, and initial risk threshold. These coefficients are then used to correct the initial trigger threshold, generating dynamic trigger thresholds, including dynamic volatility threshold, dynamic mutation threshold, and dynamic risk threshold. The dynamic trigger threshold is the product of the initial trigger threshold and the threshold correction coefficient, plus the initial trigger threshold. The threshold correction coefficient for the initial volatility threshold is obtained by comparing the standard deviation of the deviation rate within the window with the standard deviation of the historical deviation rate. The threshold correction coefficient for the initial mutation threshold is obtained by comparing the coefficient of variation of the indicator volatility value within the window with the coefficient of variation of the indicator volatility value. The threshold correction coefficient for the initial risk threshold is obtained by comparing the mean of the risk volatility coefficient within the window with the mean of the historical risk volatility coefficient.

[0117] Based on triggering conditions, the system acquires standard evidence-stored data blocks on the blockchain in real time, extracts the supply dynamic vector, metadata, and associated risk indicator set from the block body, calculates the mean value of all indicator values ​​in the supply dynamic vector within the current period, and obtains the supply deviation rate by combining the historical mean value. It also extracts dynamic indicators in the supply dynamic vector of the current period whose weights exceed a preset core threshold by combining feature dynamic weights. Simultaneously, it acquires the feature dynamic weights and corresponding environmental impact coefficients of individual indicators in the risk indicator set, obtains the double weighting factor for the corresponding indicator through multiplication, and performs a weighted summation based on the double weighting factor and the comprehensive risk value of the corresponding indicator to obtain the risk volatility coefficient, thereby generating the volatility dataset for the current period. The historical mean value for calculating the supply deviation rate is consistent with the historical mean value for calculating the volatility deviation rate.

[0118] The volatility dataset is compared one by one with the dynamic trigger threshold to determine whether there are potential sudden events, including volatility trigger, indicator sudden event trigger, and risk accumulation trigger, and a single-dimensional trigger result is generated. If the supply deviation rate is greater than the dynamic volatility threshold, it is determined that there is a volatility trigger in the current cycle. If any dynamic indicator value is greater than the dynamic sudden event threshold, it is determined that there is an indicator sudden event trigger in the current cycle. If the risk volatility coefficient is greater than the dynamic risk threshold, it is determined that there is a risk accumulation trigger in the current cycle.

[0119] By combining qualification risk characteristics with external risk levels, the results of single-dimensional triggering are validated. Based on the potential mutation events that are effectively triggered, environmental mutation events are generated. If there is a related risk, such as the existence of corresponding qualification risk characteristics or external high-risk events, it is determined to be a valid trigger. If there is no related risk, the trigger record is temporarily stored and continuously monitored. If a single-dimensional trigger still exists within a cycle, it is considered a valid trigger; otherwise, it is considered an invalid trigger. Among these, high-risk external events are obtained through the risk level of the risk indicator set.

[0120] Specifically, the consensus determination mechanism includes the following steps:

[0121] A three-tiered verification node system is constructed, comprising business layer nodes, risk control layer nodes, and technical layer nodes. The responsibilities and weights of each layer are defined. Business layer nodes correspond to procurement specialists and department managers, used to verify the correlation between sudden events and business scenarios, such as whether they affect order delivery. Risk control layer nodes correspond to risk control specialists, used to verify the risk level and compliance impact of sudden events, such as whether compliance risks and supply chain stability are involved. Technical layer nodes correspond to technical maintenance personnel, used to verify data integrity and the compliance of triggering logic, such as whether supply dynamic vector data is complete and threshold comparisons are accurate. The system retrieves the final review results of all historical sudden event voting events, calculates the total number of times each layer's voting opinion was ultimately adopted and used as a decision-making basis, and obtains the layer weight of each layer by summing and normalizing the total number of votes from each layer.

[0122] Configure basic voting parameters, including voting options, voting validity period, and voting reasons. Voting options include agree, disagree, and suggest adjustments. Voting reasons include specific suggested adjustment directions.

[0123] Once an environmental mutation event is received, an update voting request is automatically sent to the Layer 3 verification nodes. The request includes the environmental mutation event, the currently effective version information, the candidate update direction suggestions, and the voting options. At the same time, the distribution time and the node's reception status are recorded to build a voting traceability log.

[0124] After the voting period expires, the voting results of all nodes are automatically collected. The scores of each voting option are assigned to obtain the voting scores of each node. Combined with the layer weights of each node, a weighted calculation is performed to obtain the consensus weighted ratio. Since the number of nodes in each layer is limited by the number of senders, there are multiple nodes in one layer. At this time, the layer weight of each node is the ratio of the total layer weight of the business layer nodes to the number of nodes.

[0125] A voting threshold system is set, including a first threshold and a second threshold. Consensus is determined based on the weighted average of the votes, generating a consensus determination report. If the weighted average of the votes is greater than the first threshold, the current vote is directly considered to have reached a consensus, generating a consensus determination report including a list of valid votes, the weighted average calculation process, and the consensus result. If the weighted average of the votes is within the first and second thresholds, the updated node opinions after the proposed adjustment are extracted, an adjustment plan is generated, and a new vote is initiated. If the weighted average of the votes is less than the second threshold, the current vote is considered to have failed to reach a consensus, the update process is terminated, and a report of failure to reach a consensus is generated, including a list of valid votes, the weighted average calculation process, and a summary of reasons for opposition. Specifically, all historical voting records that have been closed-loop and marked with the final decision type are retrieved, and the records are divided into three categories: direct consensus, proposed adjustment and re-vote, and rejection without consensus. The weighted average of the votes corresponding to all samples that directly reached a consensus is extracted, and the minimum value is taken as the first threshold. The weighted average of the votes corresponding to all samples that proposed adjustment and re-vote is extracted, and the minimum value is taken as the second threshold.

[0126] Based on the consensus determination report, a standardized task update instruction is generated, including basic information, data scope, and iteration requirements. The basic information includes the unique task update ID, trigger event type, consensus achievement time, and verification node signature list. The data scope includes the time range of incremental data to be extracted, the list of associated suppliers, and dynamic indicators / risk events. The iteration requirements include the update direction.

[0127] The task update instruction is hash-associated with environmental mutation events and consensus judgment reports to generate an associated hash value, which is then packaged into an update decision data package. This data package is uploaded to the blockchain via the consortium blockchain's notarization interface to achieve immutable notarization and generate a task update instruction notarization certificate. The task update instruction and notarization certificate are then synchronized to the business data layer and model server of the supplier archive, and update notifications are sent to relevant nodes, specifying the instruction execution time limit and feedback requirements. A task update instruction traceability index is generated, which is associated with the supplier's dynamic vector hash, trigger threshold, and consensus results. This allows for quick querying of the entire chain of decision information through the unique task update ID.

[0128] Specifically, the steps for generating candidate evaluation models include:

[0129] The system monitors task update instructions on the blockchain in real time. By associating the unique task update ID with the traceability index, it obtains the data range specified in the instruction, such as the incremental data time range, the list of associated suppliers, and the core data dimensions. Through multi-dimensional search indexes, it extracts incremental data from the supplier archive, including indicator data, risk feature data, and metadata supplements. The indicator data is the original data of dynamic core indicators of associated suppliers extracted from the business data layer, such as the order and warehousing data corresponding to the on-time delivery rate and quality pass rate. The risk feature data is the qualification risk features and external risk features extracted from the qualification risk layer and risk indicator set, including risk level and credibility score. The metadata supplement is the synchronous extraction of data quality score and feature source hash to ensure data traceability. The incremental data is layered and filtered, and the data priority is divided through correlation analysis. Only the data related to the iteration requirements of the task update instruction is retained, and preprocessing and enhancement are performed, such as deduplication and noise reduction, format unification, and data augmentation, to finally generate an incremental training dataset.

[0130] Obtain existing benchmark evaluation models from the model repository, including model structure, historical training configuration, and performance benchmark data. Utilize incremental learning algorithms, such as gradient accumulation-based fine-tuning algorithms, where gradients from multiple batches of samples are accumulated within each iteration batch before unified parameter updates are performed to avoid gradient oscillations in small samples within a single batch. During the training phase, freeze the weights of the bottom feature extraction layer of the benchmark evaluation model and do not participate in parameter updates throughout the process. Only the top-level decision output layer is opened for gradient backpropagation. Combined with L2 regularization to constrain the magnitude of the decision layer parameters, use the incremental training dataset as input, freeze the bottom feature extraction layer of the model, and only fine-tune the parameters of the top-level decision layer to generate candidate evaluation models. At the same time, introduce a regularization mechanism to prevent model overfitting and ensure generalization ability.

[0131] Historical data not used in benchmark and incremental training is extracted from the supplier database to generate an offline test set. Simultaneously, dynamic scenario data is simulated based on the mutation types in the task update instructions. The performance differences between the candidate evaluation model and the benchmark evaluation model are evaluated using multiple metrics, including accuracy, recall, evaluation bias, and variance over 10 consecutive prediction rounds. Both the candidate and benchmark evaluation models are deployed to a simulation environment, inputting an online simulation set and a dynamic scenario set to simulate actual business processes and evaluate the model's response speed and decision rationality in dynamic scenarios. If the candidate evaluation model's multiple metrics improve by more than a preset threshold compared to the benchmark evaluation model, and online simulation verification shows no anomalies, the test is considered passed. If the threshold is not reached, incremental data is added, with a maximum of three iterations. Finally, a complete test report is generated, clearly defining the performance differences between the candidate and benchmark evaluation models and the dynamic scenario adaptation effect.

[0132] The candidate evaluation models that pass the test are assigned version numbers and packaged with the corresponding metadata to generate version model packages. The version model packages are then subjected to SHA-256 hash calculation through the blockchain notarization interface to obtain the hash value of the version model package. After being associated with the version number and task update instructions, the hash value is stored on the blockchain and the model version registry on the blockchain is updated.

[0133] Specifically, the steps for real-time credit assessment and rating include:

[0134] The latest version of the currently active model is obtained as the real-time evaluation model. At the same time, the latest supply dynamic vector of each supplier is obtained from the supplier archive through multi-dimensional search index, which is used as the real-time feature vector. Combined with the feature dynamic weight, the value of each indicator in the real-time feature vector is weighted and summed. Combined with the quality correction coefficient, the real-time credit score is calculated.

[0135] Obtain the credit score distribution data of all suppliers over the past three business cycles, use the K-means clustering algorithm to determine the initial clustering threshold, and combine the current overall supply chain status (such as industry prosperity and risk event incidence) with the company's business objectives (such as cost control and compliance priority) to calculate the threshold adjustment coefficient, thereby generating a dynamic grading threshold, as shown in the following expression:

[0136]

[0137] In the formula, For dynamic grading thresholds, The initial clustering threshold is... The supply status coefficient represents the difference between the risk event occurrence rate and the industry prosperity level. This is the business target coefficient, which is the ratio of the supplier constraint and control procurement budget amount for this product category to the total procurement budget amount for all products in this category;

[0138] Based on dynamic grading thresholds and real-time credit scores, the supplier level is determined, including special level, priority level, observation level, and elimination level, and a real-time evaluation report is generated for the supplier.

[0139] Based on the supplier classification results, cross-terminal synchronization priorities are set to ensure that core business terminals obtain key information first, including first-level, second-level, and third-level priorities. First-level priority includes procurement terminals and contract terminals, second-level priority includes ERP terminals and financial settlement terminals, and third-level priority includes supplier terminals. At the same time, based on the preset interface protocol, the evaluation results are synchronized to each business terminal, and the synchronization success confirmation signal returned by each business terminal is received simultaneously.

[0140] A dynamic binding system is built, which configures differentiated automated business rules for suppliers of different levels based on enterprise business objectives and supply chain status. The rules can be adjusted in real time. For special-level suppliers, priority order allocation is granted, more flexible payment terms are provided, and annual renewal assessments are automatically triggered. For priority suppliers, normal order allocation is provided, standard cooperation policies are enjoyed, and regular (e.g., quarterly) risk reviews are conducted. For observation-level suppliers, the right to accept large orders is restricted, risk warnings are sent (including specific improvement directions, such as qualification renewal and on-time delivery rate improvement), and the frequency of warehousing inspection is increased. For elimination-level suppliers, bidding qualifications are frozen, unfulfilled non-core contracts are terminated, and they are removed from the qualified supplier pool.

[0141] Based on the hierarchical results, the corresponding business rules are automatically triggered to generate execution instructions, such as order allocation instructions and early warning notification templates. The execution status of the rules is monitored in real time, the execution results returned by the business terminal are obtained, and the execution results are synchronized to the supplier archive.

[0142] Example 2

[0143] Another embodiment of the present invention provides: an artificial intelligence-driven credit risk data processing system, comprising: a data acquisition module, an update triggering module, and an evaluation and optimization module;

[0144] The data acquisition module is used to configure digital identity identifiers for each supplier and build a preliminary identity information review list. Through preset interface protocols and data synchronization rules, it acquires internal business datasets, intelligently identifies qualification documents to establish a dynamic qualification ledger, determines the qualification status of suppliers, obtains the qualification risk characteristics of suppliers, builds a supplier archive, sets an evaluation indicator system based on internal enterprise needs, extracts feature data of corresponding dimensions from the supplier archive, and generates a set of control indicators through data cleaning and standardization. Simultaneously, it collects external data related to suppliers through preset external data interfaces, performs risk identification, generates external risk characteristics, and generates a risk indicator set by combining qualification risk characteristics. Using multi-source data fusion, it aligns and normalizes the control indicator set and the risk indicator set to generate a dynamic supply vector.

[0145] The update trigger module is used to monitor the fluctuation of various indicators in the supply dynamic vector on the chain in real time by using a preset dynamic trigger threshold. It determines environmental change events by using the dynamic trigger threshold and fluctuation, thereby generating an event trigger signal. Upon receiving the event trigger signal, the consensus determination mechanism is immediately activated to send update voting requests to verification nodes with different permission roles to perform consensus determination, generate task update instructions, associate the event trigger reason, voting results, and verification node signatures, and broadcast them on the chain to ensure the compliance and credibility of the model update process.

[0146] The evaluation and optimization module is used to acquire incremental data according to task update instructions, and generate incremental training datasets through classification, filtering and preprocessing. This dataset is used to incrementally train the existing benchmark evaluation models in the supplier database, thereby generating candidate evaluation models. Based on the latest supply dynamic vectors of each supplier and the candidate evaluation models, real-time credit evaluation and classification are performed. The evaluation results and classification information are synchronized to the business terminal through a preset interface protocol, and business rules are triggered to guide order allocation and contract renewal.

[0147] Working principle and effects:

[0148] By integrating internal business data with external multi-source data, a dynamic supply vector representing the real-time status of suppliers is generated and stored on the blockchain to ensure dynamic and reliable data, breaking the limitations of traditional static data. A blockchain-triggered update mechanism monitors fluctuations in vector characteristics, accurately identifies environmental changes by combining dynamic thresholds, and generates update instructions through a consensus mechanism, avoiding mechanical responses of the model to external changes. Incremental data is extracted based on these instructions, and the benchmark evaluation model is trained incrementally in a targeted manner, preserving historical learning results while quickly adapting to new data patterns, solving the problem of lagging model updates. Simultaneously, end-to-end traceability breaks down the model's black box, allowing business personnel to understand and intervene. The iterative model enables real-time evaluation and grading, with results synchronized to business terminals and triggering differentiated rules, ensuring rapid implementation of evaluation value, avoiding misleading decisions and supply chain risks, and guaranteeing the accuracy of supplier supervision and supply chain stability in complex market environments.

[0149] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A credit risk data processing method based on artificial intelligence, characterized in that, include: Based on internal business datasets, qualification risk characteristics are obtained, a supplier database is built, and a set of control indicators is generated by combining the evaluation indicator system. Collect external data, and based on risk identification, perform semantic parsing and risk event extraction on the external data to generate a risk indicator set; By using multi-source data fusion, the control indicator set and the risk indicator set are aligned and normalized to generate a supply dynamic vector, which is then stored on the blockchain. Configure a trigger update mechanism on the blockchain to monitor the supply dynamic vector in real time, and combine it with a dynamic trigger threshold to determine environmental change events, thereby generating event trigger signals; Upon receiving the event trigger signal, the consensus determination mechanism is immediately activated, an update voting request is initiated, consensus determination is performed, and a task update instruction is generated. Based on the task update instruction, incremental data is extracted from the supplier archive to generate an incremental training dataset, and the benchmark evaluation model is incrementally trained to generate a candidate evaluation model. Real-time credit assessment and classification are performed using supply dynamic vectors and candidate evaluation models.

2. The credit risk data processing method based on artificial intelligence as described in claim 1, characterized in that, The steps involved in building a supplier profile include: Collect the identity credentials and basic information submitted by suppliers, and generate a preliminary identity information review list; A two-factor binding mechanism is used to generate a unique digital identity for the supplier; Set up standardized interface protocols and define data field mapping rules for different business terminals; Configure a dual-mode synchronization strategy to collect synchronized data, generate internal business datasets, and perform quality control based on field weight integrity verification. The qualification documents in the internal business dataset are subjected to hierarchical identification processing, valid fields are extracted, and cross-verification is performed to generate authenticity verification results. Enter the verified valid fields into the qualification dynamic ledger, set qualification early warning thresholds and qualification validity thresholds, classify qualification status based on the current date, and generate qualification risk characteristics of suppliers; Using the digital identity identifier as the core index, a structured supplier archive is built using a hierarchical storage architecture, and a data association engine is constructed.

3. The credit risk data processing method based on artificial intelligence as described in claim 2, characterized in that, The steps for generating the supply dynamic vector include: Set evaluation objectives and candidate indicators, obtain evaluation datasets from the supplier database, calculate the similarity between candidate indicators and evaluation objectives, and filter relevant indicators; Acquire historical business data, calculate the fitting contribution using the coefficient of determination, screen basic indicators, calculate the initial weights of the basic indicators, and generate a set of basic indicators. Set the monitoring update interval and extract historical signals within each monitoring update interval; The validity of historical signals is determined by triggering rules, and basic adjustments are made accordingly. Obtain the data items required for basic indicators, map them to the supplier archive, define extraction rules, and construct a path mapping table; The raw data of each basic indicator is extracted from the supplier archive by using a multi-dimensional search index. The interquartile range method is used to identify abnormal data, and automatic correction is performed based on similar historical data. The basic indicator data are normalized and combined with the initial weights to generate a set of control indicators.

4. The credit risk data processing method based on artificial intelligence as described in claim 3, characterized in that, The steps for generating the supply dynamic vector also include: Collect external data from multiple sources, count the number of times the same information is verified from different sources to obtain a consistency verification coefficient, and calculate the external credibility by combining the source weights. External data with credibility higher than the credibility rating threshold is screened, and risk identification is performed. Combined with qualification risk characteristics, a risk indicator set is generated. Based on the control indicator set and the risk indicator set, a related dataset is generated using digital identity identifiers as the primary alignment pair and timestamps as the assignment alignment pair. Set association weights for each indicator in the associated dataset, obtain environmental impact coefficients and business priority coefficients, calculate feature weights, and perform normalization processing to generate dynamic feature weights. The indicators in the associated dataset are weighted, fused, and standardized to generate a supply dynamic vector; The supply dynamic vector and its corresponding metadata are standardized and assembled, and combined with the block header information to generate a storage data block.

5. The credit risk data processing method based on artificial intelligence as described in claim 4, characterized in that, The steps to trigger the update mechanism include: Set the initial trigger threshold and trigger conditions; triggering at any threshold will determine the existence of a potential mutation event. A dynamic threshold calibration mechanism is set up. Based on a sliding window, historical data from the past N business cycles are statistically analyzed to calculate the threshold correction coefficient, which is then used to correct the initial trigger threshold and generate a dynamic trigger threshold. Calculate the mean of all indicator values ​​in the supply dynamic vector within the current period, combine it with the historical mean to obtain the supply deviation rate, extract the dynamic indicators in the supply dynamic vector of the current period, obtain the risk volatility coefficient, and generate the volatility dataset for the current period. The fluctuation dataset is compared one by one with the dynamic trigger threshold to determine whether there are potential mutation events and generate a single-dimensional trigger result. By combining qualification risk characteristics and external risk levels, the single-dimensional triggering results are verified, and environmental mutation events are generated based on the effectively triggered potential mutation events.

6. The credit risk data processing method based on artificial intelligence as described in claim 5, characterized in that, Initial trigger thresholds include initial volatility threshold, initial mutation threshold, and initial risk threshold; The steps to set the initial trigger threshold include: Obtain the historical supply dynamic vector and calculate the volatility deviation rate using the historical standard deviation and historical mean; Obtain the weighting coefficient and quality correction coefficient, and combine them with the fluctuation deviation rate to obtain the initial fluctuation threshold; For a single indicator, obtain the historical average value of the indicator, and at the same time, set the mutation coefficient according to the evaluation target to which the indicator belongs, and obtain the initial mutation threshold. Set a threshold for the number of risk events, and obtain the level weight and event impact coefficient of each individual risk event. Then, obtain an initial risk threshold by weighting the risk event threshold.

7. The credit risk data processing method based on artificial intelligence as described in claim 6, characterized in that, The steps of the consensus determination mechanism include: Construct a three-layer verification node structure and set the responsibilities and weights of each layer of nodes; Configure basic voting parameters, including voting options, voting validity period, and voting reasons; Once an environmental mutation event is received, an update voting request is automatically sent to the three-layer verification nodes, and the voting results of all nodes are automatically collected after the voting validity period expires. The voting scores of each node are assigned based on the scores of each voting option. The weighted average of the consensus is calculated by combining the layer weights of each node and a consensus determination is made. If the weighted average is greater than the first threshold, a consensus is reached; if the weighted average is between the first and second thresholds, the opinions of the nodes that suggest the adjustment are extracted, an adjustment plan is generated, and a new vote is initiated; if the weighted average is less than the second threshold, a consensus is not reached.

8. The credit risk data processing method based on artificial intelligence as described in claim 7, characterized in that, The steps to generate candidate evaluation models include: By using a multi-dimensional retrieval index, incremental data is extracted from the supplier archive, and the incremental data is subjected to hierarchical filtering, preprocessing and enhancement to generate an incremental training dataset. Obtain the existing benchmark evaluation model in the model version library, use the incremental learning algorithm, take the incremental training dataset as input, freeze the bottom feature extraction layer of the model, and only fine-tune the parameters of the top decision layer to generate candidate evaluation models. An offline test set is constructed, and dynamic scenario data is simulated and generated based on the mutation type in the task update instruction. The performance differences between the candidate evaluation model and the benchmark evaluation model are evaluated, and a test report is generated. The candidate evaluation models that pass the test are assigned version numbers and packaged with the corresponding metadata to generate version model packages.

9. The credit risk data processing method based on artificial intelligence as described in claim 8, characterized in that, The steps involved in real-time credit assessment and rating include: Obtain the real-time evaluation model and real-time feature vector, combine the dynamic weights of the features, perform a weighted summation, and combine the quality correction coefficient to calculate the real-time credit score; Historical credit scores are obtained, and the K-means clustering algorithm is used to divide the initial clustering thresholds. The threshold adjustment coefficient is calculated to generate dynamic grading thresholds. The supplier level is determined based on the dynamic grading threshold and the real-time credit score; Build a dynamic binding system to configure differentiated automated business rules for suppliers of different tiers; Based on the hierarchical results, the corresponding business rules are automatically triggered, execution instructions are generated, the execution status of the rules is monitored in real time, the execution results returned by the business terminal are obtained, and the execution results are synchronized to the supplier archive.

10. An AI-driven credit risk data processing system, used to implement the AI-driven credit risk data processing method as described in any one of claims 1-9, characterized in that, include: Data acquisition module, update triggering module, evaluation and optimization module; The data acquisition module is used to obtain qualification risk characteristics, build a supplier database, generate a set of control indicators based on the evaluation indicator system, collect external data synchronously and perform risk identification, generate a set of risk indicators, and generate a supply dynamic vector by using multi-source data fusion. The update triggering module is used to monitor the fluctuation of each indicator in the supply dynamic vector on the chain in real time using a pre-dynamic triggering threshold, determine environmental change events, and generate an event triggering signal to trigger the consensus determination mechanism, initiate an update voting request, perform consensus determination, and generate task update instructions. The evaluation and optimization module is used to generate an incremental training dataset according to the task update instruction, to incrementally train the benchmark evaluation model, generate candidate evaluation models, and perform real-time credit evaluation and classification based on the latest supply dynamic vectors of each supplier and the candidate evaluation models.

Images