Intelligent decision-making method, system and equipment for small enterprise credit factory based on multi-modal fusion, and medium

Abstract

The invention provides an intelligent decision-making method, system and equipment for a small enterprise credit factory based on multi-modal fusion and a medium, belongs to the technical field of risk control in the financial science and technology field, and aims to collect multi-source heterogeneous data and time sequence data of an enterprise, construct a cross-modal associated data set and generate a holographic risk portrait of the enterprise. And based on the portrait, extracting decision-making features, and constructing a hierarchical decision-making model. And optimizing decision parameters by using a multi-objective optimization module, and outputting a real-time credit line, a loan interest rate and a loan deadline. And adjusting the credit line when the risk triggering condition is met. And capturing a supervision policy file in real time, analyzing hard compliance terms, revising a decision rule threshold, and generating a compliance check report. And dynamically optimizing a risk strategy by combining the macroeconomic data, the industry risk score and the enterprise association map risk conduction analysis result. The accuracy, efficiency and compliance of credit decision making are improved, the risk identification and control capability is enhanced, and dynamic risk management and resource optimization configuration are realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of risk control in the field of financial technology, and particularly relates to a small enterprise credit factory intelligent decision-making method, system, device and medium based on multi-modal fusion. BACKGROUND

[0002] The credit factory mode has shown its advantages in small and medium-sized enterprise financing. Through professional review, process simplification, rapid approval and other ways, the efficiency and quality of credit services are improved. The small and micro loan industry serves small and micro enterprise owners with annual sales revenue less than 10 million and employee scale less than 20, which is an important part of inclusive finance.

[0003] In the related art, credit decision is mainly based on enterprise financial statements and credit reports, without considering whether there is a default problem in the contract terms, the actual scale reflected by the operating site video and the market demand decline reflected by the reduction of logistics orders. The associated risk is missed to the target enterprise, which increases the risk.

[0004] The credit model in the related art is mostly one-time offline training, which cannot adapt to the dynamic changes of enterprise operation and changes of credit external environment. The parameters cannot be optimized through the actual operation of the enterprise after the loan is refused and the repayment of the enterprise after the loan is approved, which leads to continuous deviation of the model and decline of decision-making efficiency. SUMMARY

[0005] The application provides a small enterprise credit factory intelligent decision-making method based on multi-modal fusion, which improves the accuracy, efficiency and compliance of credit decision-making, enhances the risk identification and control ability, and realizes dynamic risk management and resource optimization.

[0006] The method comprises the following steps: S101: Collecting multi-source heterogeneous data of enterprises and time series data such as power consumption, water consumption, logistics order quantity and social media public opinion, and constructing a cross-modal associated data set; S102: Multi-modal semantic fusion is performed on the constructed cross-modal associated data set, and based on the fusion result, an enterprise holographic risk portrait containing enterprise basic information features, financial risk features, operating stability features and associated risk features is generated; S103: Based on the enterprise holographic risk portrait, decision features are extracted; the decision nodes are divided according to the business type, term and guarantee mode of the small enterprise credit business, and a hierarchical decision-making model is constructed; the hierarchical decision-making model comprises a static scoring module and a dynamic risk transmission analysis module, forming a collaborative decision-making framework; S104: Call the static score result output by the static score module in the constructed hierarchical decision model and the dynamic risk analysis result output by the dynamic risk transmission analysis module, and optimize the decision parameters through the built-in multi-objective optimization module to output the real-time credit limit, loan interest rate and loan period of the small enterprise; S105: Collect the business data corresponding to the real-time credit parameters output by the multi-objective optimization algorithm, and update the parameters of the hierarchical decision model in S103 based on the online learning mechanism; a preset risk trigger condition is met, and a credit limit adjustment operation is performed; S106: Real-time capture of policy documents published by regulatory agencies, and analysis of hard compliance clauses in regulatory requirements; based on the analysis result, revise the decision rule threshold in the hierarchical decision model; at the same time, the real-time credit parameters output by the multi-objective optimization algorithm are checked for compliance, and a compliance check report containing the hard condition screening result and the flexible condition check description is generated; S107: Linking macroeconomic data, industry risk score, enterprise correlation graph risk transmission analysis result, adjusting the credit score threshold of the hierarchical decision model, the output interest rate premium ratio and the upper limit of the amount; the adjusted risk strategy is fed back to the multi-objective optimization algorithm, and the subsequent real-time credit parameter output is optimized to realize dynamic optimization of the risk strategy.

[0007] The application also provides a small enterprise credit factory intelligent decision system based on multi-modal fusion, which comprises: A data acquisition and fusion module is configured to acquire multi-source heterogeneous data and time series data of enterprises, such as electricity consumption, water consumption, logistics order volume and social media public opinion, and construct a cross-modal correlation data set; A risk portrait generation module is configured to perform multi-modal semantic fusion on the constructed cross-modal correlation data set, and generate an enterprise holographic risk portrait containing enterprise basic information features, financial risk features, operating stability features and associated risk features based on the fusion result; A decision model construction module is configured to extract decision features based on the enterprise holographic risk portrait, divide decision nodes according to the business type, period and guarantee mode of the small enterprise credit business, and construct a hierarchical decision model; the hierarchical decision model comprises a static score module and a dynamic risk transmission analysis module, forming a collaborative decision framework; An optimization and decision output module is configured to call the static score result output by the static score module in the constructed hierarchical decision model and the dynamic risk analysis result output by the dynamic risk transmission analysis module, and optimize the decision parameters through the built-in multi-objective optimization module to output the real-time credit limit, loan interest rate and loan period of the small enterprise; The model updating and risk adjustment module is configured to collect business data corresponding to real-time credit granting parameters output by the multi-objective optimization algorithm, and update parameters of the hierarchical decision-making model in real time based on an online learning mechanism; a preset risk trigger condition is provided, and when the trigger condition is met, a credit limit adjustment operation is performed; The verification and report generation module is configured to real-time capture policy documents published by a regulatory agency, and parse hard compliance clauses in regulatory requirements; based on the parsing result, decision rule thresholds in the hierarchical decision-making model are revised; and real-time credit granting parameters output by the multi-objective optimization algorithm are subjected to compliance verification, and a compliance check report containing a hard condition screening result and a flexible condition verification description is generated; The risk policy optimization module is configured to link macroeconomic data, industry risk scores, and enterprise correlation graph risk transmission analysis results, and adjust credit score thresholds of the hierarchical decision-making model, output interest rate premium ratios, and upper limits of credit limits; the adjusted risk policy is fed back to the multi-objective optimization algorithm, and subsequent real-time credit granting parameter output is optimized, thereby achieving dynamic optimization of the risk policy.

[0008] According to another embodiment of the present application, an electronic device is provided, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the small business credit factory intelligent decision-making method based on multi-modal fusion when executing the program.

[0009] According to another embodiment of the present application, an electronic device is provided, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the small business credit factory intelligent decision-making method based on multi-modal fusion when executing the program.

[0010] From the above technical solutions, the present application has the following advantages: The small enterprise credit factory intelligent decision-making method based on multi-modal fusion provided by the application constructs multi-source data collection and cross-modal associated data sets, integrates contracts, videos, electricity consumption, logistics, realizes enterprise data full coverage, and avoids risk misjudgment caused by one-sided data. Through text key clause extraction, image authenticity verification, time series fluctuation analysis and other fusion technologies, fragmented data is converted into enterprise holographic risk portrait, making enterprise risk characteristics more intuitive and accurate, and reducing missed judgments. Multi-objective optimization and real-time credit granting parameter output balance risk control, customer income and service experience through multi-objective optimization, which can meet the needs of small and micro enterprises to obtain loans, and can also protect the safety of financial institutions' funds and improve the efficiency of credit resource allocation. Online learning and dynamic adjustment of credit limit realize real-time iteration of model parameters, avoiding the problem of decision lag caused by traditional model one-time training. Preset risk trigger conditions, real-time adjustment of credit limit, and reduction of bad debt risk. Real-time capture of regulatory policies and automatic revision of decision rules, hard and flexible compliance verification of credit parameters, reduce the compliance risk of financial institutions, and improve the policy landing efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0011] In order to more clearly illustrate the technical solutions of the present application, the drawings needed to be used in the description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creating laborious work.

[0012] Figure 1 The flow chart of the small enterprise credit factory intelligent decision-making method based on multi-modal fusion; Figure 2 The small enterprise credit factory intelligent decision-making system based on multi-modal fusion is a schematic diagram; Figure 3 It is an electronic device schematic diagram. DETAILED DESCRIPTION

[0013] The small enterprise credit factory intelligent decision-making method based on multi-modal fusion provided by the application combines natural language processing, computer vision and structured data analysis credit approval process, through integrating multiple data modalities, including but not limited to enterprise financial data, transaction flow, tax information, credit record, operating site image, enterprise owner behavior data, etc., preprocessing and feature extraction of data in different modalities, integrating these features through multi-modal fusion model, and constructing enterprise credit portrait. Using machine learning algorithm and deep learning model, combined with pre-set credit rules and risk threshold, quickly generate credit approval decision, realize intelligent second batch of small enterprise credit. The present application can effectively improve the credit approval efficiency, and improve the accuracy and reliability of credit decision.

[0014] The intelligent decision-making method for small business credit factory based on multi-modal fusion will be described in detail below. For the purpose of illustration but not for limitation, specific details such as specific system structures, techniques, etc. are presented in order to provide a thorough understanding of the embodiments of the present application. However, it should be apparent to those skilled in the art that the present application can also be implemented in other embodiments without these specific details.

[0015] It should be understood that when used in the specification herein, the terms include, indicate, mean, etc. indicate the presence of the described features, integers, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0016] The phrase one or some embodiments described in the present application means that the specific features, structures or characteristics described in the embodiment are included in one or more embodiments of the present application. Therefore, the phrases appearing in different places in the present application in one embodiment, in some embodiments, in other embodiments, in additional embodiments, etc. do not necessarily refer to the same embodiment, but mean one or more but not all embodiments, unless otherwise specifically emphasized.

[0017] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0018] Please refer to Figure 1 The flowchart of the intelligent decision-making method for small business credit factory based on multi-modal fusion in a specific embodiment is shown, and the method comprises: S101: Collecting multi-source heterogeneous data of enterprises and time series data of power consumption, water consumption, logistics order volume, social media public opinion, and constructing a cross-modal associated data set.

[0019] Optionally, collecting multi-source heterogeneous data of enterprises includes: business registration information, tax declaration data, bank flow, credit report, judicial litigation record, contract text, invoice image, operating place video, legal person voice interview record, business license image. Industry risk parameters and macroeconomic basic indicators can also be obtained. The data is also integrated, taking the enterprise unified social credit code as the association word, establishing the data association relationship, and forming the cross-modal associated data set.

[0020] S102: Perform multi-modal semantic fusion on the constructed cross-modal correlation dataset, and generate an enterprise holographic risk portrait containing enterprise basic information features, financial risk features, operating stability features, and correlation risk features based on the fusion results.

[0021] In some embodiments, the multi-modal semantic fusion specifically includes contract breach liability, payment method, invoice seal clarity, business license validity, electricity consumption / logistics order volume periodicity, consistency of legal person voice statement and bank flow, and other related information. When generating the enterprise holographic risk portrait, the fusion results are stored by enterprise name, establishment time, industry affiliation, registered capital, asset-liability ratio, tax credit level, invoice authenticity status, flow abnormal transaction proportion, electricity consumption fluctuation range, logistics order volume growth rate, operating site video reflected scale, legal person statement consistency, judicial litigation record, and correlation enterprise risk event classification, each feature being accompanied by a risk level (no risk / low risk / medium risk / high risk) and a supporting data source.

[0022] The present embodiment extracts risk features of each modality, verifies different modal features through time, amount, subject, and other business dimensions, and eliminates contradictions between features; classifies and integrates the features according to business logic to form an enterprise risk portrait.

[0023] S103: Based on the enterprise holographic risk portrait, extract decision features; divide decision nodes according to the business type, term, and guarantee method of small enterprise credit business to construct a hierarchical decision model; the hierarchical decision model includes a static scoring module and a dynamic risk transmission analysis module to form a collaborative decision framework.

[0024] In some embodiments, when the collaborative decision framework is formed, the initial score is obtained through the static scoring module, and if the score is ≥80, the dynamic risk analysis is combined. If the score is 60-79, the dynamic risk is high risk, which is reduced to unqualified; if the score is <60, it is directly judged as high risk, forming a collaborative decision logic of static quantization + dynamic completion.

[0025] The present embodiment screens the most critical features for credit decision from the enterprise holographic risk portrait, divides the decision nodes according to the actual business scenarios of small enterprise credit, and ensures that the model fits the business needs; the risk is quantitatively evaluated through the static scoring module, and the potential impact of the correlation risk is identified through the dynamic risk transmission analysis module, and the two form a decision framework in collaboration.

[0026] S104: Call the static scoring results output by the static scoring module and the dynamic risk analysis results output by the dynamic risk transmission analysis module in the constructed hierarchical decision model, and optimize the decision parameters through the built-in multi-objective optimization module to output the real-time credit limit, loan interest rate, and loan term of the small enterprise.

[0027] In some embodiments, when the collaborative decision-making framework is input, the static score result needs to include the final score, the score details of each feature, and the score level.

[0028] The dynamic risk transmission analysis result needs to include the associated risk level, the risk transmission path, and the risk deduction suggestion. If the static score is unqualified or the dynamic risk is high, the loan rejection result is directly output, and the optimization process is terminated.

[0029] The target setting of multi-objective optimization is specifically: risk control target, customer coverage rate target, and fund utilization rate target. The optimization decision parameters include: credit limit parameter, loan interest rate parameter, and loan term parameter. When the optimization is executed, the corresponding parameter range is called according to the business type node, for example, the flow fund, loan, short-term, and credit node, the high-quality customer revenue multiple is 0.5-0.8, the interest rate is 0.1-0.2, and the term coefficient is 0.8-1.0. Through traversing the parameter combination, the risk compliance rate, coverage rate, and fund utilization rate corresponding to each combination are calculated, and the optimal combination that meets all three indicators is selected. Alternatively, the combination of risk compliance rate 100%, coverage rate 65%, and fund utilization rate 90% is selected.

[0030] When the real-time credit parameter output of the embodiment is output, it includes credit limit, loan interest rate, loan term, parameter calculation basis, and risk prompt. It can be seen that the output result of the collaborative decision-making framework clearly indicates the multi-objective optimization direction, and the optimal solution that meets all the targets is selected by traversing the parameter combination that adapts to different business nodes. Finally, the credit parameters that meet the enterprise risk level and business demand are output.

[0031] S105: Collect the business data corresponding to the real-time credit parameter output by the multi-objective optimization algorithm, and update the parameters of the hierarchical decision-making model in real time based on the online learning mechanism; a preset risk trigger condition is provided, and when the trigger condition is met, a credit limit adjustment operation is performed.

[0032] In some embodiments, the optimization basis of the model parameters is obtained by collecting the approval results, subsequent feedback, and business updates; the model is iterated in a lightweight manner by using incremental learning, and the historical deviation is corrected by experience replay; a preset risk trigger condition is provided to monitor the enterprise operation and risk changes in real time. Once triggered, the credit limit is dynamically adjusted to ensure that the model continuously adapts to the changes of the enterprise and timely controls the risk.

[0033] S106: Real-time capture of policy documents published by regulatory agencies, and analysis of hard compliance clauses in regulatory requirements; based on the analysis result, revise the decision rule threshold in the hierarchical decision-making model; at the same time, perform compliance verification on the real-time credit parameters output by the multi-objective optimization algorithm, and generate a compliance check report containing the hard condition screening result and the flexible condition verification description.

[0034] In some embodiments, when the decision rule threshold is revised, for the hard clause: if the policy requires a small and micro enterprise interest rate cap LPR+50BP, then the original eligible customer interest rate floating coefficient 0.3-0.6 is lowered to 0.5. For the flexible clause, if the credit is encouraged for technology-based small and micro enterprises, the guarantee method weight of technology-based enterprises is lowered, and the revenue multiple coefficient of credit is increased by 0.1-0.2; after revision, the original threshold, the new threshold, the revision basis, and the revision time need to be recorded in the system background.

[0035] During compliance verification, the hard condition is screened, and it is checked whether the enterprise belongs to the two high and one remaining industry, whether the interest rate exceeds the upper limit, and whether the single household credit meets the small and micro enterprise standard. If not, it is directly determined that the compliance does not pass, the loan is rejected, and the loan rejection reason is generated. The flexible condition verification: check whether the technology-based small and micro enterprise is preferentially met, whether the local policy subsidy requirement is met, and record the compliance prompt if not met.

[0036] The present embodiment can capture regulatory policies from authoritative channels, accurately extract compliance clauses through NLP technology, and revise the rule threshold of the layered decision model based on the clause revision, to ensure that the model decision meets the latest regulatory requirements; perform hard and flexible compliance verification on real-time credit parameters, and generate a traceable compliance report.

[0037] S107: link macroeconomic data, industry risk score, enterprise correlation graph risk transmission analysis result, adjust the credit score threshold of the layered decision model, the output interest rate premium ratio and the upper limit of the amount; feedback the adjusted risk strategy to the multi-objective optimization algorithm, optimize the subsequent real-time credit parameter output, and realize the dynamic optimization of the risk strategy.

[0038] In some embodiments, when the risk strategy is fed back, the adjusted scoring threshold, interest rate premium, and amount upper limit parameters are pushed to the S103 layered decision model to update its configuration file; at the same time, the strategy rule is converted into a constraint condition of the S104 multi-objective optimization module to ensure that the subsequent approval credit parameters automatically meet the new strategy; the adjustment result is synchronized to the system management background, and a risk strategy adjustment report is generated, including the adjustment background, adjustment parameters, and impact range.

[0039] As can be seen, the present embodiment combines macroeconomic trends, industry risk levels, and enterprise correlation risk data to construct a risk transmission link. Based on the link analysis result, the parameters of the layered decision model and the constraint conditions of the multi-objective optimization are adjusted, so that the risk strategy can adapt to macroeconomic changes and cope with industry risks. It also covers enterprise correlation risks, improving the foresight and pertinence of decision-making.

[0040] In an embodiment of the present application, based on step S102, a possible embodiment will be given below to non-restrictively describe its specific implementation scheme. S102 specifically includes the following steps: S1021: Standardize and preprocess each type of data in the cross-modal correlation dataset.

[0041] In some embodiments, for text data, word segmentation, stop word removal, and entity recognition are performed; for image data, size normalization and grayscale processing are performed; for power consumption and water flow, missing value filling and noise filtering are performed.

[0042] S1022: Map data of different modalities into high-dimensional feature vectors using pre-trained models.

[0043] In some embodiments, text embedding models are used to convert text key clauses and public opinion content into text feature vectors; convolutional neural networks are used to extract visual feature vectors from images; time series encoding networks are used to convert sequence data into time series feature vectors; text converted from speech is also converted into text feature vectors.

[0044] For example, the BERT model can understand the potential legal risks of contract clauses; the ResNet can identify subtle signs of forgery of invoice seals; the LSTM can capture the operating periodicity implied in the water flow sequence. This realizes the unified representation of complex heterogeneous data into a machine-processable form, and the extracted features have high abstraction and semantic information.

[0045] S1023: Calculate the semantic correlation weight between different modal feature vectors.

[0046] For example, the text features of the interview with the legal person are attention-weighted with the time-series features of the bank water flow at the same period to verify the consistency of the statements; the invoice image features are aligned with the tax declaration text features to verify the authenticity of the information. This step aims to break down the modal barriers and achieve cross-validation of information.

[0047] In some embodiments, the features of one modality are used as queries to search for related information in the feature library of another modality, and the attention weight is calculated. The part with high weight represents strong semantic correlation between the two modalities. Contrastive learning drives the model to learn a feature space where the associated different modal data are closer by constructing positive and negative sample pairs. This realizes information fusion and cross-validation, can discover and strengthen the consistency or inconsistency between different sources of data, and enhances the insight and accuracy of the system in anti-fraud and risk assessment.

[0048] S1024: Aggregate the semantically aligned modal feature vectors. Use weighted concatenation-based or neural network-based fusion methods to integrate the feature vectors of all modalities such as text, image, and time series into a unified, high-information-density fusion feature representation.

[0049] S1025: Decode the fusion feature representation into risk labels and quantitative indicators. A series of concrete, interpretable risk dimension features are parsed and output from the fusion feature through a fully connected layer, including: age, industry, financial risk features, operating stability features, correlation risk features, etc. Finally, a structured enterprise holographic risk portrait is formed.

[0050] In some embodiments, the fusion feature representation is input into a multi-task learning model. The model shares the underlying fusion features, but has multiple different output layers at the top, each responsible for predicting a financial risk score, operating stability grade, etc. This approach allows the model to consider multiple related tasks simultaneously during training, sharing statistical strength, thereby improving the generalization performance of each task.

[0051] In an embodiment of the present application, based on step S103, a possible embodiment will be given below to specifically illustrate the non-limiting embodiment. S103 specifically includes the following steps: S1031: According to the combination of business type, loan period, and guarantee mode in small business credit business, a decision node network is defined in advance. Each node corresponds to a specific business scenario, and a decision logic unit is configured for each node.

[0052] In some embodiments, the credit product system is decomposed into discrete, independently managed decision points. For example, short-term working capital credit loans are one node, and medium and long-term supply chain mortgage loans are another node. Each node is a container that will load specific decision rules and thresholds for the sub-scenario, improving the accuracy of risk pricing and business flexibility.

[0053] S1032: Input the basic information features and financial risk features in the enterprise holographic risk portrait into a credit score generator. The credit score generator calculates a comprehensive credit score for each enterprise through a nonlinear function trained on historical credit data.

[0054] In some embodiments, based on the patterns learned from a large amount of historical data, the enterprise's multiple static features are mapped to a single quantitative score that can comprehensively reflect its historical credit status and current financial health. This model is essentially a prediction function, with the input being a feature vector and the output being a score. Static features can include registered capital, asset-liability ratio, and historical overdue times.

[0055] S1033: The associated risk features in the enterprise holographic risk portrait, including the association relationship data between the enterprise, the legal person, and the shareholders, are constructed into a graph structure. Using graph structure analysis technology, risk events involving other entities directly or indirectly connected to the enterprise are identified, and the impact degree of the risk events on the enterprise through the association edge is calculated, and a risk transmission intensity index is output.

[0056] In some embodiments, the enterprise and its associated parties are regarded as nodes in a network, and the equity, transaction, and other relationships are regarded as edges. By analyzing the topological structure of this network, known risk events are located, and the path and possibility of risk transmission along the connecting edge to the target enterprise are simulated, and finally quantified into an index. The identification ability of hidden risks and systemic risks is enhanced.

[0057] S1034: According to the business type, term, and guarantee method to which the current credit application belongs, it is routed to the corresponding specific node in the decision node network. Trigger the node's exclusive decision logic, which combines the credit score output by the static scoring module with the risk transmission intensity index output by the dynamic risk transmission analysis module, and refers to the pass threshold and weight preset in the node to generate a preliminary decision.

[0058] In some embodiments, static and dynamic risk information is customized according to the risk and return characteristics of different business scenarios. For example, for a high-risk credit loan node, a very high static score pass threshold is set, and the dynamic risk index is extremely sensitive; while for a node with full mortgage, the static score threshold may be lower, and the tolerance to the dynamic risk index is higher. This is reflected in different calculation formulas and threshold parameters in each node. Realize the precise matching of risk and business. Ensure that high-risk businesses apply more stringent approval standards, while low-risk businesses can enjoy more relaxed and convenient processes, thereby controlling overall risk while optimizing customer experience and capital utilization.

[0059] S1035: Establish a dispatch center, which is used to control the decision node network and integrates the output interfaces of the static scoring module and the dynamic risk transmission analysis module. For each credit application, the dispatch center first calls the static scoring module, synchronously triggers the dynamic risk transmission analysis module, and guides it to the correct decision node according to the business attributes to perform the final calculation, thereby integrating the analysis capabilities of the two modules into a coherent decision-making process.

[0060] In some embodiments, the execution order and data flow of the decision-making task are defined: first, the enterprise holographic risk portrait generated in S102 is obtained, then the static scoring module and the dynamic risk transmission analysis module are called in parallel or in sequence, then the node routing is performed according to the business keywords, and the decision synthesis is completed at the specified node. It is ensured that each independent module can cooperate in order and efficiently.

[0061] In an embodiment of the present application, based on step S104, a possible embodiment will be given below to specifically and non-limitingly illustrate the specific implementation thereof. S104 specifically includes the following steps: S1041: Configure the data calling channel between the static scoring module, the dynamic risk transmission analysis module and the built-in multi-objective optimization module in the hierarchical decision-making model, configure the RESTAPI output interface for the static scoring module, configure the RESTAPI output interface for the dynamic risk transmission analysis module, and clearly define that the static scoring module output needs to include the enterprise ID, the static scoring value, the feature score details, and the scoring generation time field, and the dynamic risk transmission analysis module output needs to include the enterprise ID, the associated risk level, the risk event type, the risk deduction value, and the analysis generation time field.

[0062] In some embodiments, by standardizing the RESTAPI interface and the JSON format, the directional data transmission link of the static scoring, the dynamic risk transmission analysis module to the multi-objective optimization module is established, the Token authentication ensures that only authorized modules can call data, the HTTPS encryption prevents the leakage of enterprise financial and risk information in the transmission process, and the unified field and the timeout retry ensure the stability of data calling.

[0063] S1042: The multi-objective optimization module calls the interfaces of the static scoring module and the dynamic risk transmission analysis module to obtain the output data of the current small enterprise credit application; performs integrity verification on the static scoring data to check whether the static scoring value is an integer of 0-100 and whether the feature score details include the scores of the eight preset features such as the establishment period and the registered capital; and performs validity verification on the dynamic risk analysis data to check whether the associated risk level is a preset enumeration value and whether the risk deduction value is a non-negative integer.

[0064] In some embodiments, by using preset numerical ranges, field integrity, enumeration values and other verification rules, data with format errors or abnormal contents are filtered out, invalid data is retransmitted, and valid input is maximized; after retransmission fails, an alarm is triggered and a log is recorded, so that the operation and maintenance personnel can quickly locate the module failure.

[0065] S1043: In the parameter configuration library of the multi-objective optimization module, configure the to-be-optimized decision-making parameters and ranges according to the industry and the business type: the credit limit parameter, the loan interest rate parameter, and the loan period parameter; and configure the constraint conditions.

[0066] In some embodiments, the optimization parameter range and constraint conditions are refined according to the operating characteristics of different industries of small enterprises, the differences in collateral values, and the actual needs of credit business, the parameter configurability is realized through database storage, and the multi-objective optimization module is clear about the optimization object, the optimization boundary, and the requirement for reaching the standard.

[0067] S1044: In the multi-objective optimization module, the optimization target weight is configured, the flow fund loan is set to risk control 45%, income 30%, and experience 25%, and the supply chain financing is set to risk control 35%, income 40%, and experience 25%; the iteration control rule is configured, the target function value is calculated according to risk reaching rate x weight + income reaching rate x weight + experience reaching rate x weight; the module result passed by the verification of S1042 is associated with the parameters, constraints, and weights of S1043, an optimization task sheet containing task ID, enterprise ID, and parameter range ID is generated, and is stored in the cache.

[0068] S1045: The multi-objective optimization module reads the optimization task sheet in the cache, calculates the credit parameters according to the optimal parameter combination: credit limit = average revenue of the enterprise in the past 12 months x revenue multiplier coefficient x collateral discount rate x (1-risk deduction value / 100); loan interest rate = (central bank benchmark interest rate x (1+upward floating coefficient)) + (risk premium coefficient x static score / 100); loan period = basic period x adjustment coefficient; after the calculation, a credit report containing parameter calculation process and target reaching situation is generated, and is pushed to the credit approval front end and synchronized to the enterprise credit archive library.

[0069] In some embodiments, based on the optimized parameters, the credit limit, interest rate, and period are accurately calculated according to a preset formula by combining enterprise operating data, external data, and risk adjustment rules; the calculation process and result are arranged into a compliance report, which is synchronized to the front end for approval and stored in the archive library for inspection, thereby improving the approval efficiency and compliance.

[0070] In an embodiment of the present application, based on step S105, a possible embodiment will be given below to non-restrictively describe the specific implementation scheme. S105 specifically includes the following steps: S1051: Configure a business data acquisition channel to connect the output data source of the multi-objective optimization module of S104, the enterprise repayment behavior data source, and the risk event data source.

[0071] In some embodiments, by connecting the interfaces of different data sources, combining the established multi-source data acquisition link, and unifying the data format, it is ensured that the whole-link business data from S104 to subsequent repayment and risk events can be accurately and safely collected, thereby providing complete data input for online learning.

[0072] S1052: Collect business data corresponding to real-time credit granting parameters, and perform preprocessing on the collected data; repeat data deduplication, missing value completion, and outlier filtering; and store the preprocessed data by enterprise ID and data type.

[0073] S1053: In the parameter mapping library, establish a correspondence between the business data and the parameters of the S103 hierarchical decision-making model. Among them, the number of overdue payments in the last three months in the repayment data corresponds to the historical overdue feature weight of the static scoring module, and the average repayment period corresponds to the debt repayment ability feature weight of the static scoring module; the number of associated enterprise risk events in the risk event data corresponds to the associated risk weight of the dynamic risk transmission analysis module; at the same time, the parameter update trigger condition is configured: single-industry business data accumulates to 100 or cross-industry accumulates to 500, or single-enterprise subsequent repayment / risk data updates to 5; and the parameter update range threshold is set.

[0074] In some embodiments, the parameter mapping library uses a MySQL database table for storage, and the fields include business data field name, model module name, model parameter name, mapping rule, and adjustment coefficient. For example, the number of overdue payments in the last three months corresponds to the historical overdue feature weight of the static scoring module, and the mapping rule is that for every 5% increase in the number of overdue payments, the weight is adjusted by 0.02; the update trigger condition is configured in the system background parameter management interface, and the trigger data volume can be set by industry.

[0075] The parameter mapping table specifies how business data affects the corresponding relationship of model parameters, the trigger condition controls the timing of parameter update, and the range threshold limits the parameter adjustment range, providing clear and controllable rules for online learning parameter update.

[0076] S1054: Start parameter update according to the trigger condition, read the preprocessed business data in the MySQL sub-table, and locate the model parameters that need to be updated according to the parameter mapping table; After the update is completed, perform parameter validity verification to check whether the parameter value is within the preset range threshold, and if it exceeds, trigger the manual confirmation process; extract nearly 1000 historical credit data, recalculate the scoring and risk analysis results using the updated model parameters, compare the overdue prediction accuracy before and after the update, and if the accuracy decreases by more than 5%, automatically roll back to the parameters before the update; after verification, record the parameter update in the parameter change log library.

[0077] In some embodiments, the parameter adjustment value is calculated based on the preprocessed business data and the preset mapping rule, and the range verification and effect verification ensure the rationality and effectiveness of the updated parameters, ensuring the stability of the hierarchical decision-making model.

[0078] S1055: push the updated parameters that pass the verification to the S103 hierarchical decision-making model, push the static scoring module parameters to the built-in feature weight library, and push the dynamic risk transmission analysis module parameters to the associated risk weight library.

[0079] In some embodiments, the parameter pushing adopts a RabbitMQ message queue, the exchange type is set to direct, the routing key is distinguished according to the model module, and the message body contains the parameter version number, the parameter name, the updated value, and the effective time. The cache setting parameter version number Key of the static scoring module is compared with the version number when loading new parameters. If the version number is higher than the current version, the cache is updated. When the dynamic risk transmission module reads the MySQLdynamic_weight table, the WHERE parameter version number = latest version number condition is added. The parameter update report contains the updated parameter list, the involved industry, the effective time, and the historical version comparison.

[0080] The embodiment realizes efficient pushing of the updated parameters to the hierarchical decision-making model through the message queue, controls the parameter loading time through the version number, records the parameter application through the report and the log, ensures that the updated parameters are applied to the hierarchical decision-making model, and realizes the whole-process traceability.

[0081] In an embodiment of the present application, based on step S106, a possible embodiment will be given below to non-restrictively describe the specific implementation scheme. S106 specifically includes the following steps: S1061: obtain the original policy text from the regulatory agency information source, analyze the text, extract the key regulatory clauses from the text, convert them into compliance rules, and store them in the compliance knowledge base; In some embodiments, a conversion process from regulatory information publishing to machine-readable rules is established, and the policy documents and prohibitive clauses issued by the regulatory agency information source are identified through text analysis.

[0082] S1062: input the real-time credit granting parameters output by the multi-objective optimization module and the related features in the enterprise holographic risk portrait into the compliance rule module. The compliance rule module compares the content of the credit granting mode with the rules in the compliance knowledge base one by one.

[0083] In some embodiments, the credit granting scheme data to be verified is loaded into the working memory of the rule module as a fact object, and the fact object is matched with the rules in the compliance knowledge base to trigger all rules that meet the conditions. The automatic matching of the regulatory rules and the massive credit applications is realized.

[0084] S1063: According to the result of rule matching, the checking result of each compliance rule is determined. The checking result is divided into completely satisfied rules, hard prohibition type rules violated, and soft suggestion type rules violated, and the specific clause content of each checking result is attached.

[0085] In some embodiments, based on the preset rule type label, the output original trigger result is classified and importance ranked. The checking result is no longer simply pass / fail, but presents the violation level and specific reason, facilitating manual review or policy adjustment, and improving the transparency and explainability of the decision.

[0086] S1064: For each type of checking result, the compliance risk weight is allocated according to the effectiveness level of the corresponding regulatory clause. The weighted results of all rules are calculated to generate a compliance risk assessment score.

[0087] In some embodiments, a weight allocation system is first established, different risk weights are given to each rule according to the authority and severity of different regulatory rules, and a compliance risk score is obtained by weighted calculation. The scattered compliance check results are aggregated into a quantifiable risk indicator, so that the system can sort and compare the compliance risks of different credit schemes and support risk management.

[0088] S1065: According to the predefined report template, the classification information of the checking result, the specific clause content violated, and the overall compliance risk assessment score are integrated to assemble a structured compliance check report document containing summary, detailed description and conclusion.

[0089] In some embodiments, the template is used to fill the classified result, the weighted summary score and other data into the corresponding position of the preset report template, to generate a standard report with uniform format and complete content.

[0090] In an embodiment of the present application, based on step S107, a possible embodiment will be given below to non-limitingly illustrate the specific implementation scheme. S107 specifically includes the following steps: S1071: Collect and pre-process macroeconomic and industry risk data, and convert the data into year-on-year and month-on-month change rates using time series.

[0091] In some embodiments, macroeconomic and industry data can be obtained, and macroeconomic environment changes can be converted into quantifiable risk signals. Standardization processing can also be performed to make different dimensions of data comparable for correlation analysis.

[0092] S1072: Extract the risk conduction quantitative index of the correlation graph. The overdue rate of the associated enterprise in the preset time period, the default rate of the industry, and the order interruption probability of the supply chain core enterprise are calculated. The three indicators are weighted according to the associated enterprise overdue rate 40%, the industry default rate 30%, and the supply chain interruption probability 30%, and the enterprise correlation risk conduction coefficient is generated.

[0093] In some embodiments, the correlation graph based on S102 obtains the correlation risk characteristics. The associated enterprise overdue rate is calculated from the overdue state field of the enterprise-association enterprise edge. The industry default rate is extracted from the industry risk score field of the enterprise-industry edge. The supply chain interruption probability is calculated from the order performance rate field of the enterprise-supply chain edge. The weighting coefficient is verified based on historical data. The influence weight of the associated enterprise overdue on the enterprise is the highest 40%, followed by the industry overall default 30%, and finally the supply chain fluctuation 30%.

[0094] The correlation risk of the present embodiment is scattered in the associated enterprise, industry, and supply chain. These scattered risks are integrated into the total correlation risk faced by the enterprise. The risk conduction coefficient calculated by weighting reflects the risk degree with a single numerical value, which is convenient for model processing.

[0095] S1073: Establish a mapping rule of risk parameters and model strategy, and define the corresponding relationship between macroeconomic indicators, industry risk score, correlation risk coefficient and hierarchical decision model parameters.

[0096] In some embodiments, the mapping rule is defined based on historical matching data of the risk. Optionally, when the correlation risk coefficient exceeds 0.6, the enterprise overdue rate is 20% higher than the average, and therefore the upper limit of the quota is reduced by 10% to control the risk exposure.

[0097] It should be noted that the mapping rule is a bridge from external risk to model strategy, which converts abstract risk changes into specific model parameter adjustment; the rule is based on historical data verification to ensure the rationality and effectiveness of the adjustment. Let the strategy adjustment have a clear quantitative basis to avoid the randomness of subjective judgment; ensure that the linkage between risk changes and model strategy is a precise match.

[0098] S1074: Adjust the risk strategy parameters of the hierarchical decision model. According to the mapping rule of S1073, modify the basic risk score threshold of the static scoring module, the interest rate premium calculation coefficient of the dynamic risk conduction analysis module, and the industry limit proportion of the upper limit of the quota, and write the adjusted parameters into the configuration file of the hierarchical model.

[0099] In some embodiments, the parameter adjustment directly modifies the configuration file of the hierarchical model. Based on the credit score threshold of the static scoring module, the score is changed from 70 to 65, which means that a lower basic risk score can also enter the subsequent process. The interest rate premium calculation coefficient of the dynamic risk transmission analysis module is changed from 0.5% to 0.55%, which means that the interest rate is higher under the same risk, covering additional risks. The industry limit ratio of the credit limit is changed from 25% to 22%, which means that the overall industry credit is shrinking.

[0100] It should be noted that the decision logic of the hierarchical model is based on the parameters in the configuration file, and modifying the parameters directly changes the decision results of the model; the configuration file is the adjustable part of the module, and strategy updating can be realized without reconstructing the model.

[0101] S1075: feedback the adjusted strategy to the target optimization algorithm, and use the adjusted credit score threshold, interest rate premium ratio, and credit limit as new constraint conditions of the multi-objective optimization algorithm to update the input parameter set of the optimization module, so that the subsequent generated credit scheme adapts to the new risk strategy requirements.

[0102] In some embodiments, the input of the multi-objective optimization algorithm originally includes enterprise holographic risk portrait, business type, term, etc., and the three constraint conditions of the adjusted credit score threshold, interest rate premium ratio, and credit limit are added.

[0103] The goal of the multi-objective optimization algorithm is to generate a credit scheme that meets the requirements, and the added constraint conditions make the optimization results directly adapt to the new risk strategy; ensure that the subsequent generated credit scheme always meets the latest risk strategy, so that risk control is more sustainable and can dynamically adapt to external environmental changes.

[0104] It should be understood that the size of the serial number of each step in the above embodiments does not mean the order of execution, and the execution order of each process should be determined according to its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0105] The following is an embodiment of the small business credit factory intelligent decision system based on multi-modal fusion provided by the embodiments of the present disclosure. The system and the small business credit factory intelligent decision method based on multi-modal fusion of each embodiment described above belong to the same inventive concept. Details not described in the embodiment of the small business credit factory intelligent decision system based on multi-modal fusion can be referred to the embodiment of the small business credit factory intelligent decision method based on multi-modal fusion.

[0106] As shown in Figure 2 The system includes: A data acquisition and fusion module 201 is configured to acquire multi-source heterogeneous data of an enterprise and time series data such as electricity consumption, water consumption, logistics order volume, and social media public opinion, and construct a cross-modal correlation data set. The risk portrait generation module 202 is configured to perform multi-modal semantic fusion on the constructed cross-modal correlation data set, and generate an enterprise holographic risk portrait containing enterprise basic information features, financial risk features, operating stability features and correlation risk features based on the fusion result; The decision model construction module 203 is configured to extract decision features based on the enterprise holographic risk portrait, divide decision nodes according to the business types, terms and guarantee methods of small enterprise credit business, and construct a hierarchical decision model. The hierarchical decision model includes a static scoring module and a dynamic risk transmission analysis module, forming a collaborative decision framework. The optimization and decision output module 204 is configured to call the static scoring result output by the static scoring module and the dynamic risk analysis result output by the dynamic risk transmission analysis module in the constructed hierarchical decision model, and optimize the decision parameters through the built-in multi-objective optimization module to output the real-time credit limit, loan interest rate and loan term of the small enterprise. The model update and risk adjustment module 205 is configured to collect business data corresponding to the real-time credit parameters output by the multi-objective optimization algorithm, and update the parameters of the hierarchical decision model in real time based on an online learning mechanism. When the preset risk trigger condition is met, a credit limit adjustment operation is performed. The verification and report generation module 206 is configured to real-time capture policy documents published by regulatory agencies, and parse hard compliance clauses in the regulatory requirements. Based on the parsing result, the decision rule threshold in the hierarchical decision model is revised. At the same time, the real-time credit parameters output by the multi-objective optimization algorithm are checked for compliance, and a compliance check report containing the hard condition screening result and the flexible condition check description is generated. The risk strategy optimization module 207 is configured to link macroeconomic data, industry risk scores and enterprise correlation graph risk transmission analysis results, adjust the credit score threshold of the hierarchical decision model, the output interest rate premium ratio and the upper limit of the credit limit, and feed back the adjusted risk strategy to the multi-objective optimization algorithm to optimize the subsequent real-time credit parameter output, thereby realizing dynamic optimization of the risk strategy.

[0107] As shown in Figure 3 The present application also provides an electronic device comprising a display module 103, a memory 102, a processor 101, a communication module 104, and a computer program stored in the memory and executable on the processor 101, wherein the processor 101 implements the steps of the small enterprise credit factory intelligent decision method based on multi-modal fusion when executing the program.

[0108] In embodiments of the application, the electronic device includes, but is not limited to, a laptop computer, a desktop computer, a workstation, a personal digital assistant, a server, a blade server, a mainframe computer, and other suitable computers. The electronic device can also represent various forms of mobile devices such as personal digital processors, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections, and relationships, and their functions, are shown as examples only and are not meant to limit implementations of the embodiments described and / or claimed in this application.

[0109] In embodiments of the application, the processor 101 can be implemented by using at least one of application specific integrated circuits, programmable logic devices, field programmable gate arrays, processors, controllers, micro-controllers, microprocessors, electronic units designed to perform the functions described herein, and in some cases such implementation can be implemented in a controller. For software implementation, embodiments of such processes or functions can be implemented with separate software modules, which allow at least one function or operation to be performed, by a software application (or program) written in any suitable programming language to be executed by a controller. Software code can be implemented by the software application (or program) written in any suitable programming language to be executed by a controller, and can be stored in the memory.

[0110] The display module 103 is used to display information input by a user or information provided to a user. The display module 103 can include a display panel, which can be configured in the form of a liquid crystal display, an organic light emitting diode, or the like.

[0111] The memory 102 can be used to store software programs as well as various data. The memory 102 can include a high-speed random access memory, and can also include a non-volatile memory such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state memory device.

[0112] The communication module 104 transmits and / or receives radio signals to and / or from at least one of a base station, an external terminal, and a server. Such radio signals can include voice call signals, video call signals, or various types of data according to text and / or multimedia message transmission and reception.

[0113] The application further provides a storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the small enterprise credit factory intelligent decision-making method based on multi-modal fusion.

[0114] The storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example,, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0115] The storage medium stores a program product capable of implementing the methods described above in this specification. In some possible implementations, various aspects of this disclosure can also be implemented as a program product comprising program code that, when run on a terminal device, causes the terminal device to perform the steps described in the exemplary methods section of this specification according to various exemplary embodiments of this disclosure.

[0116] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A smart decision-making method for small business credit factories based on multimodal fusion, characterized in that, The methods include: S101: Collect multi-source heterogeneous data and time-series data of enterprises, including electricity consumption, water consumption, logistics order volume, and social media sentiment, to construct a cross-modal association dataset; S102: Perform multimodal semantic fusion on the constructed cross-modal association dataset, and based on the fusion results, generate a holographic risk profile of the enterprise that includes basic enterprise information features, financial risk features, operational stability features, and association risk features; S103: Based on the enterprise's holographic risk profile, extract decision-making features; segment decision-making nodes according to the business type, term, and guarantee method of small business lending, and construct a hierarchical decision-making model; the hierarchical decision-making model includes a static scoring module and a dynamic risk transmission analysis module, forming a collaborative decision-making framework; S104: Call the static scoring results output by the static scoring module and the dynamic risk analysis results output by the dynamic risk transmission analysis module in the constructed hierarchical decision model, and optimize the decision parameters through the built-in multi-objective optimization module to output the real-time credit limit, loan interest rate and loan term for small businesses. S105: Collect business data corresponding to the real-time credit parameters output by the multi-objective optimization algorithm, and update the parameters of the S103 hierarchical decision model in real time based on the online learning mechanism; preset risk trigger conditions, and execute the credit limit adjustment operation when the trigger conditions are met; S106: Capture policy documents issued by regulatory agencies in real time, analyze the hard compliance clauses in regulatory requirements; based on the analysis results, revise the decision rule thresholds in the hierarchical decision model; at the same time, perform compliance verification on the real-time credit parameters output by the multi-objective optimization algorithm, and generate a compliance inspection report that includes hard condition screening results and flexible condition verification instructions; S107: Link macroeconomic data, industry risk scores, and enterprise correlation graph risk transmission analysis results to adjust the credit scoring threshold, output interest rate premium ratio, and credit limit of the hierarchical decision-making model; feed the adjusted risk strategy back to the multi-objective optimization algorithm to optimize the subsequent real-time credit parameter output and achieve dynamic optimization of the risk strategy.

2. The intelligent decision-making method for small business credit factories based on multimodal fusion according to claim 1, characterized in that, S102 specifically includes the following steps: Standardize and preprocess various types of data in the cross-modal association dataset; Pre-trained models are used to map data from different modalities into high-dimensional feature vectors; Calculate the semantic relevance weights between feature vectors of different modalities; Aggregate the semantically aligned feature vectors of each modality; The fusion feature representation is decoded into risk labels and quantitative indicators.

3. The intelligent decision-making method for small business credit factories based on multimodal fusion according to claim 1, characterized in that, S103 specifically includes the following steps: Based on the combination of business types, loan terms, and guarantee methods in small business lending, a decision node network is predefined, with each node corresponding to a specific business scenario and each node configured with a decision logic unit. The basic information features and financial risk features in the enterprise's holographic risk profile are input into the credit score generator, and a comprehensive credit score is calculated for the enterprise through a nonlinear function trained on historical credit data. The associated risk characteristics of the enterprise, legal person and shareholder relationship data in the enterprise holographic risk profile are constructed into a graph structure. The graph structure analysis technology is used to identify risk events of entities directly or indirectly connected to the enterprise, calculate the degree of impact of risk events on the enterprise, and output the risk transmission intensity index. Based on the current credit application's business type, term, and guarantee method, it is routed to the corresponding specific node in the decision node network, triggering the node's exclusive decision logic. The credit score is combined with the risk transmission intensity index for calculation, and preliminary decision opinions are generated with reference to the node's preset pass threshold and weight. A scheduling center is established to manage the decision-making node network and integrate the output interfaces of the credit scoring generator and the risk transmission analysis module. For each credit application, the scheduling center first calls the credit scoring generator and simultaneously triggers the risk transmission analysis module. Then, based on the business attributes, it guides the application to the correct decision-making node to perform the final calculation, thus integrating the analytical capabilities of the two modules into a coherent decision-making process.

4. The intelligent decision-making method for small business credit factories based on multimodal fusion according to claim 1, characterized in that, S104 specifically includes the following steps: S1041: Configure the data call interface between the static scoring module, the dynamic risk transmission analysis module, and the multi-objective optimization module, and define the output data fields; S1042: The multi-objective optimization module calls the data call interface to obtain output data and performs integrity and validity checks on the output data; S1043: Configure the parameter range and constraints for credit limit, loan interest rate, and loan term in the multi-objective optimization module; S1044: Configure the optimization objective weights and iteration control rules of the multi-objective optimization module, and generate an optimization task list; S1045: The multi-objective optimization module reads the optimization task order, calculates the credit limit, loan interest rate and loan term, generates a credit report and pushes it.

5. The intelligent decision-making method for small business credit factories based on multimodal fusion according to claim 1, characterized in that, S105 specifically includes the following steps: S1051: Configure the business data acquisition channel to connect to the output data source of the multi-objective optimization module, the enterprise repayment behavior data source, and the risk event data source; S1052: Collect business data corresponding to real-time credit parameters, preprocess the collected data, and classify and store the preprocessed data. S1053: Establish the correspondence between business data and hierarchical decision model parameters in the parameter mapping library, and configure parameter update trigger conditions and range thresholds; S1054: Initiate parameter update according to the trigger condition, read the preprocessed business data, update the model parameters, and verify the validity of the updated parameters; S1055: Push the validated updated parameters to the corresponding modules of the hierarchical decision model.

6. The intelligent decision-making method for small business credit factories based on multimodal fusion according to claim 1, characterized in that, S106 specifically includes the following steps: The original policy texts are obtained from information sources of regulatory agencies, and the texts are parsed to extract key regulatory clauses, which are then transformed into compliance rules and stored in the compliance knowledge base. The real-time credit parameters output by the multi-objective optimization module and the relevant features in the enterprise's holographic risk profile are input into the compliance rules module. The compliance rules module compares the content of the credit method with the rules in the compliance knowledge base one by one. Based on the results of rule matching, the verification result of each compliance rule is determined; the verification results are divided into rules that are fully satisfied, hard prohibition rules that are violated, and soft recommendation rules that are violated, and the specific clauses that are violated or satisfied are attached to each verification result. For each type of verification result, a compliance risk weight is assigned based on the effectiveness level of the corresponding regulatory provisions; The weighted results of all rules are aggregated and calculated to generate a compliance risk assessment score; Based on a predefined report template, the verification results classification information, the specific clauses violated, and the overall compliance risk assessment score are integrated and assembled into a structured compliance inspection report document containing a summary, detailed description, and conclusion.

7. The intelligent decision-making method for small business credit factories based on multimodal fusion according to claim 1, characterized in that, S107 specifically includes the following steps: Collect macroeconomic data and industry risk data, preprocess the data, convert them into year-on-year change rate and month-on-month change rate, and perform standardization to make data from different dimensions comparable; The enterprise association graph is used to calculate the overdue rate of related enterprises, the default rate of their respective industries, and the order interruption probability of core enterprises in the supply chain within a preset time period. The three indicators are weighted and calculated according to the weights of 40% for the overdue rate of related enterprises, 30% for the default rate of the industry, and 30% for the probability of supply chain interruption, to generate the enterprise association risk transmission coefficient. Based on historical risk matching data, establish mapping rules between macroeconomic indicators, industry risk scores, associated risk transmission coefficients and hierarchical decision-making model parameters, and clarify the correspondence between each risk parameter and model strategy parameters; According to the established mapping rules, modify the basic risk scoring threshold of the static scoring module, the interest rate premium calculation coefficient of the dynamic risk transmission analysis module, and the industry limit ratio of the quota limit in the hierarchical decision model, and write the adjusted parameters into the configuration file of the hierarchical decision model. The adjusted credit scoring threshold, interest rate premium ratio, and credit limit are added as new constraints to the multi-objective optimization algorithm, updating the set of input parameters for the algorithm so that the credit granting schemes generated subsequently are adapted to the new risk strategy requirements.

8. A smart decision-making system for small business credit factories based on multimodal fusion, characterized in that, The system is used to implement the intelligent decision-making method for small business credit factories based on multimodal fusion as described in any one of claims 1 to 7; The system includes: The data acquisition and fusion module is used to collect multi-source heterogeneous data and time-series data from enterprises, such as electricity consumption, water consumption, logistics order volume, and social media sentiment, to build a cross-modal association dataset. The risk profile generation module is used to perform multimodal semantic fusion on the constructed cross-modal association dataset. Based on the fusion results, it generates a holographic risk profile of the enterprise that includes basic enterprise information features, financial risk features, operational stability features, and associated risk features. The decision-making model construction module extracts decision-making features based on the enterprise's holographic risk profile; it segments decision-making nodes according to the business type, term, and guarantee method of small business lending, and constructs a hierarchical decision-making model; the hierarchical decision-making model includes a static scoring module and a dynamic risk transmission analysis module, forming a collaborative decision-making framework. The optimization and decision output module is used to call the static scoring results output by the static scoring module and the dynamic risk analysis results output by the dynamic risk transmission analysis module in the constructed hierarchical decision model, and optimize the decision parameters through the built-in multi-objective optimization module to output the real-time credit limit, loan interest rate and loan term for small businesses. The model update and risk adjustment module is used to collect business data corresponding to the real-time credit parameters output by the multi-objective optimization algorithm, and update the parameters of the hierarchical decision model in real time based on the online learning mechanism; preset risk trigger conditions, and when the trigger conditions are met, execute the credit limit adjustment operation; The verification and report generation module is used to capture policy documents issued by regulatory agencies in real time, parse the hard compliance clauses in regulatory requirements, revise the decision rule thresholds in the hierarchical decision-making model based on the parsing results, and perform compliance verification on the real-time credit parameters output by the multi-objective optimization algorithm to generate a compliance inspection report that includes hard condition screening results and flexible condition verification instructions. The risk strategy optimization module links macroeconomic data, industry risk scores, and enterprise correlation graph risk transmission analysis results to adjust the credit scoring threshold, output interest rate premium ratio, and credit limit of the hierarchical decision-making model. The adjusted risk strategy is then fed back to the multi-objective optimization algorithm to optimize the subsequent real-time credit granting parameter output, thereby achieving dynamic optimization of the risk strategy.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the intelligent decision-making method for small business credit factories based on multimodal fusion as described in any one of claims 1 to 7.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent decision-making method for small business credit factories based on multimodal fusion as described in any one of claims 1 to 7.

Images