Business risk detection method and device, equipment, medium and program product

Abstract

The invention provides a business risk detection method and device, equipment, a medium and a program product, and relates to the field of financial science and technology or the field of big data. The method comprises the steps of obtaining multi-modal business data of a target industry chain, processing the multi-modal business data in each industry to generate a joint feature vector corresponding to each industry, determining business association information among the industries according to the joint feature vectors corresponding to all the industries, and sending the business association information to the target industry chain. Generating a risk detection result of the target industrial chain based on the business association information; according to the method, the multi-modal business data of the multi-industry associated enterprises in the target industry chain is acquired and subjected to targeted processing, the joint feature vector of each industry is generated so as to accurately mine the business associated information between the industries, efficient detection and accurate identification of the target industry chain risk are finally realized, and the efficiency and accuracy of industry chain risk detection are improved.

Description

Technical Field

[0001] This application relates to the fields of financial technology or big data, and in particular to a business risk detection method, apparatus, equipment, medium, and program product. Background Technology

[0002] In the field of bank customer risk management, financial institutions face increasingly complex credit risks, supply chain transmission risks, and compliance risks, requiring real-time fusion and analysis of multi-source heterogeneous data. For example, when the real estate industry experiences a cash flow crisis, the risk can rapidly propagate along the chain of "real estate → construction → building materials → steel production," manifesting as high-frequency credit card overdrafts by senior executives, disclosures of financial difficulties in wealth management communications, and abnormal financial statements of related companies—a variety of cross-business and cross-modal signals. However, current risk control systems struggle to effectively integrate data from different sources, such as credit card systems (transaction behavior), wealth management systems (voice recordings), and loan systems (financial images), leading to delayed risk identification and untimely warnings.

[0003] In existing technologies, when a customer submits a loan application, the system automatically extracts the structured data fields from the bank and credit reporting agencies, substitutes them into a preset scoring formula, and calculates a comprehensive risk score. This score corresponds to a risk level (such as low, medium, or high) and triggers the corresponding approval strategy (such as automatic approval, manual review, or rejection).

[0004] However, data between various business systems (such as credit cards, loans, and wealth management) cannot be securely integrated, heterogeneous data (text, voice, and images) are processed separately, and the dynamic propagation path of risk in the industry chain cannot be quantified, resulting in low accuracy and efficiency of risk identification. Summary of the Invention

[0005] This application provides a business risk detection method, apparatus, equipment, medium, and program product to address the technical problem of low accuracy and efficiency in risk identification.

[0006] Firstly, this application provides a business risk detection method, the method comprising:

[0007] Acquire multimodal business data of the target industry chain, which includes multiple industries with business transmission relationships. The multimodal business data is data generated by multiple business type systems of related enterprises in each industry of the target industry chain.

[0008] The multimodal business data in each industry are processed separately to generate joint feature vectors corresponding to each industry;

[0009] Based on the joint feature vectors corresponding to all industries, business relationship information between industries is determined, and risk detection results of the target industrial chain are generated based on the business relationship information.

[0010] Secondly, this application provides a business risk detection device, comprising:

[0011] The acquisition module is used to acquire multimodal business data of the target industry chain, which includes multiple industries with business transmission relationships. The multimodal business data is data generated by multiple business type systems of related enterprises in each industry of the target industry chain.

[0012] The processing module is used to process the multimodal business data in each industry separately and generate joint feature vectors corresponding to each industry.

[0013] The determination module is used to determine the business association information between industries based on the joint feature vectors corresponding to all industries, and generate the risk detection result of the target industrial chain based on the business association information.

[0014] Thirdly, embodiments of this application provide a business risk detection device, including: a memory and a processor;

[0015] The memory stores computer-executed instructions;

[0016] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0017] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0018] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0019] The business risk detection method provided in this application overcomes the limitations of single industry and single data type by comprehensively acquiring multimodal business data generated by multi-industry related enterprises with business transmission relationships in multiple business types within the target industrial chain. This achieves the completeness and comprehensiveness of industrial chain data collection, avoiding misjudgment or omission of risks due to missing data dimensions. By specifically processing multimodal business data from various industries and generating corresponding joint feature vectors, the method effectively integrates the value of heterogeneous data under different business scenarios, improving the accuracy of feature representation and industry adaptability. Based on the mining of inter-industry business relationship information using industry-wide joint feature vectors, it can deeply capture the business transmission logic and potential relationships between upstream and downstream of the industrial chain, breaking through the bottleneck of isolated industry analysis in traditional risk detection. Finally, risk detection results are generated through business relationship information, achieving efficient positioning of industrial chain risks and improving the accuracy of risk identification through multi-dimensional data support and correlation logic analysis. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0021] Figure 1 Flowchart of the business risk detection method provided for this application Figure 1 ;

[0022] Figure 2 Flowchart of the business risk detection method provided for this application Figure 2 ;

[0023] Figure 3 Flowchart of the business risk detection method provided for this application Figure 3 ;

[0024] Figure 4 A schematic diagram of the business risk detection device provided in this application;

[0025] Figure 5 A schematic diagram of the business risk detection equipment provided in this application.

[0026] The accompanying drawings have illustrated specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to specific embodiments. Detailed Implementation

[0027] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0028] It should be noted that the business risk detection methods, devices, equipment, media, and program products provided in this application can be used in the fintech or big data fields, or in any field other than the fintech or big data fields. This application does not limit the application fields of the business risk detection methods, devices, equipment, media, and program products.

[0029] In the field of bank customer risk management, financial institutions face increasingly complex risk scenarios, encompassing not only traditional credit risk but also diverse risk types such as risks transmitted through upstream and downstream supply chains and compliance and regulatory risks. To achieve accurate risk identification and early warning, real-time fusion and analysis of multi-source heterogeneous data from multiple internal business systems and external partners is necessary. This data includes structured financial indicators and transaction records, as well as unstructured voice communications, text feedback, and image-based financial documents.

[0030] In the current risk control system, most banks adopt the traditional model of "structured data extraction + fixed scoring model" when assessing risk: when a customer submits a loan or other business application, the system only extracts the preset structured data fields from the bank's internal business system and external credit reporting agencies, substitutes them into a fixed scoring formula to calculate a comprehensive risk score, and then maps the corresponding risk level (such as low, medium, high) according to the score, and triggers approval strategies such as automatic approval, manual review or rejection.

[0031] However, this model has limitations: on the one hand, data between various business systems (such as credit card systems, loan systems, and wealth management systems) is difficult to integrate and share across systems due to security and compliance requirements and technical barriers; on the other hand, there is a lack of an effective unified processing mechanism for unstructured heterogeneous data such as text, voice, and images, which leads to the fragmentation of multi-dimensional risk information and results in low accuracy and efficiency of risk identification.

[0032] To address the aforementioned issues, the business risk detection method provided in this application is based on the business transmission relationships between various industries in the industrial chain. It comprehensively acquires multimodal business data generated by multiple business types from related enterprises across multiple industries within the target industrial chain. First, it processes the multimodal business data for each industry to form a joint feature vector corresponding to that industry. Then, it uses the joint feature vectors of all industries to mine and determine the business relationship information between industries. Finally, it uses this business relationship information as the core basis to generate the risk detection results for the target industrial chain. This solution achieves efficient and accurate detection of industrial chain risks through the integrated processing of multimodal data and the precise mining of industry relationships.

[0033] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0034] Figure 1 Flowchart of the business risk detection method provided for this application Figure 1 In this embodiment, the executing entity is, for example, a business risk detection system. Figure 1 As shown, the method includes:

[0035] S101: Obtain multimodal business data of the target industry chain. The target industry chain includes multiple industries with business transmission relationships. The multimodal business data is the data generated by multiple business type systems of related enterprises in each industry of the target industry chain.

[0036] Among them, the target industrial chain refers to an industry cluster with a clear business transmission relationship and interdependence between upstream and downstream.

[0037] Multimodal business data is a collection of heterogeneous data generated by related enterprises in the target industry chain in multiple business types such as credit cards, wealth management, loans, and supply chain transactions. It covers different forms of data, including structured data (such as transaction records and credit records) and unstructured data (such as customer voice recordings, scanned financial statements, and contract texts).

[0038] Business type systems are information platforms within banks or enterprises that support specific business operations (such as credit card transaction systems, loan approval systems, and supply chain management systems).

[0039] Specifically, firstly, the core industries and related enterprises of the target industrial chain are clearly defined, and the list of upstream and downstream enterprises is identified through enterprise registration information, supply chain cooperation agreements, and bank customer relationship maps; secondly, a multimodal data collection channel is established, which connects to business systems such as credit cards, wealth management, and loans through an internal data platform to achieve cross-system data collection that is "usable but not visible"; finally, data collection specifications are formulated, which clearly define the field collection scope of structured data (such as transaction amount, transaction time, and credit balance) and the format requirements of unstructured data (such as voice sampling rate and image resolution), while data is obtained through data anonymization techniques (such as name anonymization and address obfuscation).

[0040] For example, the target industry chain includes the real estate development industry, the construction industry, and the building materials production industry. Related companies include real estate company A, construction company B, and building materials factory C (A is B's client, and B is C's purchaser). Multimodal business data collection specifically includes: obtaining transaction texts of company A's senior executives' personal credit cards from the credit card system of a partner branch of real estate company A, recordings of conversations between company A's finance personnel and financial managers, and a scanned image of company A's financial statement for the second quarter of 2024 from the loan system; obtaining transaction texts of company B from the credit card system of a partner branch of construction company B, recordings of conversations between company B's manager and customer service, and an image of company B's financial statement for the second quarter of 2024 from the credit card system of a partner branch of building materials supplier C; and obtaining transaction texts of company C from the credit card system of a partner branch of building materials supplier C, recordings of conversations between company C's manager and customer service, and an image of company C's financial statement for the second quarter of 2024 from the credit card system of a partner branch of building materials supplier C.

[0041] S102: Process the multimodal business data in each industry separately to generate joint feature vectors corresponding to each industry.

[0042] Among them, the joint feature vector refers to the standardized numerical vector formed by integrating, transforming and fusing multimodal business data of a single industry in the target industrial chain.

[0043] Specifically, firstly, data preprocessing is performed. For structured data, missing and outlier values ​​are removed, and field formats and units of measurement are standardized. For unstructured data, cleaning (such as speech denoising, image deblurring, and text deduplication) and format standardization (such as speech-to-text, image recognition, and text segmentation) are carried out to ensure data quality. Secondly, features are extracted separately. For structured data, statistical analysis and correlation analysis are used to extract key features such as transaction frequency, credit limit, and repayment period. For unstructured data, natural language processing (such as text keyword extraction and semantic analysis) and image processing (such as feature point extraction) techniques are used to uncover hidden features. Thirdly, feature fusion is performed. Through weight allocation and dimension unification, structured and unstructured features from the same industry are integrated into a feature set with a unified dimension. Finally, a joint feature vector is generated. The fused feature set is standardized and encoded, transforming it into a numerical vector that can be used for algorithm model analysis, ensuring the consistency and comparability of the vectors.

[0044] For example, in the real estate development industry, the following steps are taken: First, preprocessing is performed on the credit card transaction texts, call recordings, and scanned financial statements of Company A (e.g., deduplication of transaction texts, transcription of call recordings, and OCR extraction of text from financial statement images). Then, features are extracted: from the structured transaction data, features such as quarterly transaction volume and average transaction amount are extracted; from the call texts, semantic features such as cooperation intentions and funding needs are extracted; and from the financial statement texts, financial features such as revenue and liabilities are extracted. Subsequently, the three types of features are weighted according to their importance and fused, ultimately generating a joint feature vector for the real estate development industry. Similarly, for the multimodal business data of Company B in the construction industry and Company C in the building materials manufacturing industry, preprocessing, feature extraction, feature fusion, and vector encoding are performed sequentially to generate corresponding joint feature vectors, forming a feature vector set covering all sectors of the industry chain.

[0045] S103: Based on the joint feature vectors of all industries, determine the business relationship information between industries, and generate risk detection results for the target industrial chain based on the business relationship information.

[0046] Among them, business association information refers to key information such as the strength, type and stability of associations between industries in terms of business dealings, capital flow and cooperation dependence, which are extracted through algorithm analysis based on the joint feature vectors of various industries. It is a quantitative representation of the business transmission relationship between upstream and downstream industries in the industrial chain.

[0047] Risk detection results refer to a conclusive report formed by identifying and quantifying credit risk, liquidity risk, and transmission risk in the target industry chain by combining business-related information. It includes risk point location, risk level assessment, and risk transmission path prediction.

[0048] Specifically, firstly, an industry correlation analysis model is constructed. The joint feature vectors corresponding to all industries are input into the model. Through algorithms such as cosine similarity calculation, association rule mining, and graph neural network analysis, the matching degree and correlation of feature vectors between industries are quantified, determining the strength (e.g., strong, medium, weak) and type (e.g., capital dependence, supply chain support) of business correlations between industries. Secondly, a risk assessment indicator system is established, covering indicators such as correlation stability, risk transmission sensitivity, and the proportion of abnormal features. Risk thresholds for each indicator are set based on industry regulatory standards and historical risk data. Finally, based on business correlation information, it is determined whether abnormal correlations exist. The impact of abnormal correlations on the entire industry chain is simulated through a risk transmission model, ultimately generating risk detection results.

[0049] For example, by using an algorithm to calculate the business relationships among the real estate development, construction, and building materials manufacturing industries using their joint feature vectors, the algorithm identifies a strong correlation between real estate development and construction (capital dependence), a strong correlation between construction and building materials manufacturing (supply chain support), and a moderate correlation between the two (indirect transmission). Subsequently, combined with a risk assessment indicator system, anomalies were found in features such as "transaction frequency" and "timeliness of loan repayment" within the joint feature vector of the building materials manufacturing industry. These anomalies can be transmitted through the construction industry to the real estate development industry, leading to increased cash flow risk in the industrial chain. Ultimately, this results in a risk detection result: the target industrial chain as a whole is classified as medium risk.

[0050] The business risk detection method provided in this embodiment acquires multimodal business data of a target industry chain. The target industry chain includes multiple industries with business transmission relationships. The multimodal business data consists of data generated from multiple business type systems of related enterprises in each industry within the target industry chain. The multimodal business data in each industry is processed separately to generate joint feature vectors corresponding to each industry. Based on the joint feature vectors corresponding to all industries, the business relationship information between each industry is determined. Based on the business relationship information, the risk detection result of the target industry chain is generated. This method acquires multimodal business data of related enterprises in multiple industries within the target industry chain and performs targeted processing to generate joint feature vectors for each industry to accurately mine business relationship information between industries. Ultimately, it achieves efficient detection and accurate identification of risks in the target industry chain, improving the efficiency and accuracy of industry chain risk detection.

[0051] Figure 2 Flowchart of the business risk detection method provided for this application Figure 2 ,like Figure 2 As shown, in this embodiment... Figure 1 Based on the examples, the business risk detection method is described in detail, which includes:

[0052] S201: Acquire multimodal business data of the target industry chain.

[0053] Step S201 is similar to step S101, and will not be described again here.

[0054] S202: Extract features from multimodal business data in various industries to generate multimodal business feature data corresponding to each industry.

[0055] Multimodal business feature data is a collection of structured and unstructured data features. Structured data features include numerical features (such as transaction amount statistics and credit balance change rate) and categorical features (such as transaction scenario type and credit product type). Unstructured data features include textual semantic features (such as contract key clause vectors and call sentiment) and image visual features (such as financial statement key data area features and seal texture features).

[0056] Specifically, for structured data from various industries, statistical analysis (such as mean and variance calculation) and feature encoding (such as one-hot encoding and label encoding) are used to extract numerical and categorical features. For text data in unstructured data (such as contract text and transcripts of phone recordings), natural language processing techniques (such as TF-IDF and BERT models) are used to extract semantic and keyword features. For image data in unstructured data (such as scanned financial statements and scanned contracts), computer vision techniques (such as CNN models and OCR recognition) are used to extract visual features and key information extraction features. Finally, different types of features from the same industry are integrated to generate multimodal business feature data corresponding to that industry.

[0057] For example, for multimodal business data in the real estate development industry: Numerical features such as "average monthly transaction amount in the second quarter of 2024" and "month-on-month growth rate of credit balance" are extracted from Company A's transaction flow (structured data), as well as categorical features such as "main transaction scenario is building material procurement"; semantic features such as "capital turnover pressure" and "delayed project payment" and sentiment features of "worry" are extracted from the transcripts of recorded conversations between Company A's financial personnel and financial managers; key data features such as "operating revenue" and "asset-liability ratio" are extracted from Company A's second-quarter 2024 financial report scanned images using OCR recognition, and visual features such as the standardization of financial report format and the integrity of seals are extracted using a CNN model. These are then integrated to obtain multimodal business feature data for the real estate development industry.

[0058] S203: Encrypt the multimodal business feature data corresponding to each industry to obtain the encrypted feature data corresponding to each industry.

[0059] Among them, encrypted feature data is a feature form that cannot be directly interpreted after multimodal business feature data has been processed by encryption algorithms.

[0060] Specifically, firstly, an appropriate encryption algorithm is selected based on the type of multimodal business feature data (such as numerical features, semantic vector features, and visual features). For numerical features, a homomorphic encryption algorithm (such as the Paillier algorithm) is used to ensure that numerical calculations can be performed directly after encryption. For semantic vector features and visual features, privacy-preserving encryption techniques from federated learning (such as secure aggregation protocols and differential privacy encryption) are employed. Secondly, the multimodal business feature data for each industry is segmented, and corresponding encryption algorithms are applied to different types of features, with encryption keys and access permissions set. Finally, an encryption verification mechanism is used to verify the integrity and security of the encrypted data, ensuring no data loss or tampering, and generating encrypted feature data corresponding to each industry.

[0061] For example, in the multimodal business feature data of the real estate development industry, numerical features such as "average monthly transaction amount of 5 million yuan in the second quarter of 2024" and "month-on-month growth rate of 15% in credit balance" are encrypted using the Paillier homomorphic encryption algorithm to obtain encrypted numerical features; semantic feature vectors such as "capital turnover pressure" and "project payment delay" are generated by adding small noise using differential privacy encryption technology; key data features and visual features such as "asset-liability ratio of 60%" in the financial statement scan image are encrypted using a secure aggregation protocol; multimodal business feature data of the construction industry and building materials production industry are processed with the same encryption logic, and finally, encrypted feature data corresponding to the three industries are obtained respectively, and the encrypted data can only be processed by authorized nodes.

[0062] Optionally, the multimodal business feature data corresponding to each industry can be encrypted to obtain encrypted feature data corresponding to each industry. Specific implementation methods include:

[0063] The multimodal business feature data corresponding to each industry are encrypted sequentially according to modality type to obtain the encrypted feature data corresponding to each industry. The encrypted feature data is a set of multiple homomorphically encrypted modal feature data.

[0064] Modality type refers to the morphological classification of multimodal business feature data, including feature categories of different dimensions such as structured data features (e.g., numerical and categorical features), text semantic features, and image visual features.

[0065] Homomorphic encryption is a special encryption technique that allows specific operations to be performed directly on encrypted data. The decrypted result is consistent with the result of the operation on the original data, without exposing the original feature information.

[0066] Encrypted feature data is a set of features that cannot be directly interpreted. It is formed by homomorphically encrypting multimodal business feature data from various industries according to their modal types. This preserves the computational usability of the features while ensuring data privacy and security.

[0067] Specifically, firstly, the multimodal business feature data of various industries are classified into modalities, clarifying the boundaries and data formats of different modalities such as structured data features, text semantic features, and image visual features. Secondly, suitable homomorphic encryption algorithms are selected for different modal types: numerical structured features use additive homomorphic encryption algorithms (such as the Paillier algorithm), categorical structured features use encoding conversion algorithms based on homomorphic encryption, text semantic feature vectors use a scheme combining homomorphic encryption and semantic preservation technology, and image visual features use feature compression encryption algorithms adapted to homomorphic encryption. Then, in the order of "structured data features → text semantic features → image visual features", different modal features of each industry are encrypted sequentially. After each modality is encrypted, integrity verification is performed to ensure that the data is not lost or tampered with. Finally, all homomorphically encrypted modal feature data of the same industry are integrated to form the corresponding encrypted feature data for that industry.

[0068] For example, numerical structured features such as "average monthly transaction amount of 3 million yuan in the second quarter of 2024" and "month-on-month growth rate of credit balance of 8%" are first categorized and organized; textual semantic features such as "tight supply of raw materials" and "product inventory backlog" are categorized; and image visual features such as "regional features of key financial data" and "visual features of seals on purchase contracts" are categorized. The numerical structured features are encrypted using the Paillier additive homomorphic encryption algorithm to obtain encrypted transaction amount and credit growth rate features. The textual semantic feature vectors are encrypted using a semantic preservation encryption scheme based on BFV homomorphic encryption to generate encrypted semantic features. The image visual features are encrypted using a CNN feature compression encryption algorithm adapted to homomorphic encryption to obtain encrypted visual features. After verifying the integrity of each modality of encrypted data, the three types of encrypted features are integrated to form encrypted feature data for the building materials production industry. Similarly, numerical features such as "asset-liability ratio of 60%" and semantic features such as "delayed project payment" in the real estate development industry, and numerical features such as "contract amount" and semantic features such as "delayed construction progress" in the construction industry, are all processed according to the above modal order and corresponding homomorphic encryption algorithms to obtain their respective encrypted feature data.

[0069] S204: Perform multimodal fusion analysis on the encrypted feature data corresponding to each industry to generate joint feature vectors corresponding to each industry.

[0070] Among them, multimodal fusion parsing refers to integrating different types of feature data (such as encrypted numerical features, encrypted semantic features, and encrypted visual features) under the encrypted state of the same industry through an adapted fusion model, mining the inherent relationship between different modal features, and forming a unified dimension feature representation.

[0071] The joint feature vector is a high-dimensional vector data obtained after multimodal fusion analysis, which centrally carries the core information of multimodal business characteristics in the same industry.

[0072] Specifically, firstly, a multimodal fusion model under the federated learning framework is constructed, which supports the processing of different types of features under encrypted conditions. Secondly, encrypted numerical features, encrypted semantic features, and encrypted visual features from various industries are input into the fusion model. Weights of different modal features are assigned through attention mechanisms (such as Multi-Head Attention) to highlight key features related to business and risk assessment. Then, fusion strategies such as feature concatenation and weighted summation are used to integrate the encrypted features of different modalities into a vector of a unified dimension. Finally, the validity and consistency of the joint feature vector are ensured through a model output verification mechanism to generate joint feature vectors corresponding to each industry.

[0073] For example, taking the construction industry as an example, its encrypted numerical features such as "average monthly transaction amount in the second quarter of 2024" and "credit balance quarter-on-quarter growth rate" are input into a federated multimodal fusion model, along with encrypted semantic features such as "project progress delay" and "rising raw material procurement costs" and image features such as "encrypted visual features of key data areas in financial statements". The model assigns higher weights to semantic features such as "project progress delay" and "rising raw material procurement costs" through an attention mechanism, and lower weights to visual features such as the standardization of financial statement formats. A weighted summation strategy is used to integrate the various encrypted features into a vector of dimension 128. After validity verification, a joint feature vector for the construction industry is obtained. Similarly, joint feature vectors for the real estate development industry and the building materials production industry are generated respectively.

[0074] Optionally, the encrypted feature data includes: text data, voice data, and image data. Multimodal fusion parsing is performed on the encrypted feature data corresponding to each industry to generate joint feature vectors corresponding to each industry. Specific implementation methods include:

[0075] Semantic features are extracted from the text data to obtain text feature vectors;

[0076] Voiceprint features and emotion index are extracted from the speech data to obtain speech feature vectors;

[0077] Target business indicators are identified from image data to obtain image feature vectors;

[0078] The text feature vector, speech feature vector, and image feature vector are weighted and fused to generate a joint feature vector.

[0079] Specifically, firstly, for encrypted text data, a privacy-preserving NLP model under the federated learning framework (such as federated BERT) is used to extract semantic features without decrypting the original data, filtering out invalid information and generating a fixed-dimensional text feature vector. Secondly, for encrypted speech data, a privacy-adapted speech processing algorithm is used to first extract voiceprint features (such as Mel-frequency cepstral coefficients), and then an emotion index is obtained by analyzing parameters such as speech tone and speech rate through an emotion recognition model. After integration, a speech feature vector is formed. Then, for encrypted image data, a federated computer vision model (such as federated CNN) combined with OCR technology is used to identify target business indicators in the image (such as revenue amount in financial reports and transaction amount in contracts), converting them into numerical features and generating an image feature vector. Finally, weights are set based on the importance of the business scenario (such as a weight of 0.4 for text semantic features, 0.2 for speech emotion index, and 0.4 for image business indicators). The three types of feature vectors are integrated into a joint feature vector of a unified dimension through a weighted summation formula, ensuring that the fused vector retains the core information of each modality while highlighting the influence of key features.

[0080] For example, encrypted text data includes encrypted engineering contract text between Company B and real estate company A, and encrypted procurement contract text between Company B and building materials factory C. Semantic features such as "total project price of 120 million yuan", "construction period of 18 months", and "raw material purchase unit price increase of 5%" are extracted using a federated BERT model to generate a text feature vector with a dimension of 128. Encrypted voice data is a recording of a call between the person in charge of Company B and a bank customer service (encrypted). After extracting voiceprint features, the emotion index is determined to be 0.7 (leaning towards anxiety) using an emotion recognition model, and integrated to generate a 128-dimensional voice feature vector. Encrypted image data is a scanned copy of Company B's financial report for the second quarter of 2024 (encrypted). Target business indicators such as "operating revenue of 35 million yuan" and "asset-liability ratio of 72%" are identified using a federated CNN combined with OCR, and converted into numerical features to generate a 128-dimensional image feature vector. Weighted fusion is performed according to preset weights (text 0.4, voice 0.2, image 0.4) to calculate a joint feature vector of the construction industry with a dimension of 128. Similarly, the text features of "delayed project payment" for real estate company A, the voice features of "anxious sentiments of financial personnel" and the image features of "declining revenue in financial reports" for real estate company A, and the text features of "tight supply of raw materials" for building materials factory C, the voice features of "anxious sentiments of the person in charge" and the image features of "inventory backlog" for building materials factory C are all processed according to the above process to generate joint feature vectors for their respective industries.

[0081] S205: Based on the joint feature vectors of all industries, extract the business correlation and risk assessment indicator changes for each industry.

[0082] Among them, business relevance refers to the degree of relevance between different industries in the target industrial chain based on business transmission relationship. It is obtained through similarity analysis of joint feature vectors and reflects the strength of relevance between industries in terms of transactions, capital flow, etc. Change in risk assessment indicators refers to the fluctuation range of key risk indicators of each industry (such as changes in asset-liability ratio, changes in cash flow stability, and changes in default probability) relative to the benchmark value. It is calculated through the mapping model of joint feature vectors and risk indicators.

[0083] Specifically, firstly, an industry correlation calculation model is constructed. Based on the cosine similarity algorithm under the federated learning framework, the similarity between joint feature vectors of different industries is calculated. The higher the similarity, the stronger the business correlation. Secondly, a mapping model between joint feature vectors and risk assessment indicators is established. The correspondence between feature vectors and key risk indicators (such as debt-to-equity ratio, cash flow coverage ratio, and default probability) is obtained through training with historical risk data. Then, the joint feature vectors of each industry are input into the mapping model to calculate the current value of the key risk indicators of each industry. The current value is compared with the preset benchmark value (such as the historical average value or industry standard value) to obtain the change in risk assessment indicators. Finally, the calculation results are standardized to ensure the comparability of business correlation and the change in risk assessment indicators.

[0084] For example, using an industry correlation calculation model, comparing the joint feature vectors of the real estate development industry and the construction industry yields a similarity of 0.85, classifying the business correlation as "strong correlation"; comparing the joint feature vectors of the construction industry and the building materials production industry yields a similarity of 0.78, classifying the business correlation as "relatively strong correlation"; and comparing the joint feature vectors of the real estate development industry and the building materials production industry yields a similarity of 0.62, classifying the business correlation as "moderate correlation". Using a risk assessment indicator mapping model, the current asset-liability ratio of the real estate development industry is 75%, compared to a benchmark of 65%, representing a change of +10%; the current cash flow coverage ratio of the construction industry is 1.2, compared to a benchmark of 1.5, representing a change of -0.3; and the current default probability of the building materials production industry is 3.2%, compared to a benchmark of 2.0%, representing a change of +1.2%.

[0085] S206: Generate risk detection results based on business relevance and changes in risk assessment indicators.

[0086] Specifically, when the change in a risk assessment indicator of an industry indicates that the business correlation between the industry and its upstream and downstream industries reaches "strong correlation" or above, the industry is identified as a key risk point. Secondly, a risk transmission path analysis model is constructed. Based on the business correlation ranking results, the risk transmission direction of the key risk point to the upstream and downstream industries is identified, and the overall risk level of the industrial chain is classified. Finally, the key risk points, risk transmission paths, and overall risk levels are integrated to generate risk detection results that include risk descriptions, risk levels, transmission paths, and prevention and control recommendations.

[0087] Based on the risk detection results, a five-level decision tree is further generated. The root node of the decision tree is set to the customer's comprehensive risk score of 85 points, and the branch nodes are set to the risk ratio of industrial chain transmission of 35%. Based on the logical relationship between the root node and the branch nodes, targeted risk control measures are output at the leaf nodes. The risk control measures include "reducing the credit line of construction company B by 20%" and "requiring building materials company C to provide additional collateral".

[0088] For example, a change of +10% in the asset-liability ratio of the real estate development industry (exceeding the +5% threshold) and a strong correlation with the construction industry are identified as key risk points; a change of -0.3 in the cash flow coverage ratio of the construction industry (exceeding the -0.2 threshold) and a relatively strong correlation with both upstream and downstream industries are identified as key risk points; a change of +1.2% in the default probability of the building materials production industry (not exceeding the +2% threshold) is identified as a general risk point. Through a risk transmission path analysis model, the risk transmission path is identified as "real estate development industry → construction industry → building materials production industry". Using a weighted scoring method, the overall risk score of the industrial chain is calculated to be 82 points (out of 100), classifying the overall risk level as "high risk". The final risk detection result is: the overall risk level of the target industrial chain is high risk, the key risk points are the real estate development industry (excessively high asset-liability ratio) and the construction industry (tight cash flow), and the risk transmission path is the spread of risk from the upstream real estate industry to the downstream construction and building materials production industries.

[0089] The business risk detection method provided in this embodiment first acquires multimodal business data of the target industry chain (the target industry chain includes multiple industries with business transmission relationships, and the multimodal business data comes from multiple business type systems of related enterprises in each industry). Then, feature extraction is performed on the multimodal business data of each industry to generate corresponding multimodal business feature data for each industry. Next, the multimodal business feature data of each industry is encrypted to obtain corresponding encrypted feature data for each industry. Then, multimodal fusion parsing is performed on the encrypted feature data to generate joint feature vectors for each industry. Finally, based on the joint feature vectors of all industries, the business correlation degree and risk assessment indicator changes for each industry are extracted, and the risk detection results of the target industry chain are generated based on the business correlation degree and risk assessment indicator changes. This method, through targeted processing of multimodal business data of multi-industry related enterprises in the industry chain through "feature extraction-encryption processing-fusion parsing," ensures data security while accurately mining the business correlation patterns between industries and the dynamic changes in risk indicators. Ultimately, it achieves efficient detection and accurate identification of risks in the target industry chain, improving the efficiency, accuracy, and data security capabilities of industry chain risk detection.

[0090] Figure 3 Flowchart of the business risk detection method provided for this application Figure 3 ,like Figure 2 As shown, in this embodiment... Figure 2 Based on the examples, this paper provides a detailed explanation of how to generate risk detection results based on business relevance and changes in risk assessment indicators. The method includes:

[0091] S301: Determine the correlation values ​​for each industry based on the business relevance and the changes in risk assessment indicators.

[0092] Among them, the correlation value is a quantitative indicator that comprehensively reflects the risk level of a single industry in the target industrial chain and its ability to influence upstream and downstream industries. Its core is to weight and integrate the change of the industry's own risk assessment indicators with the business correlation between the industry and other industries. The higher the value, the stronger the risk transmission potential of the industry and the greater the impact on the overall risk of the industrial chain.

[0093] Specifically, the formulas for calculating the correlation values ​​for each industry are as follows:

[0094]

[0095] The business relevance weights are divided into "strong relevance (weight 0.6), relatively strong relevance (weight 0.4), medium relevance (weight 0.2), and weak relevance (weight 0.1)". The risk weights are determined according to the type of risk assessment indicator (e.g., core risk indicators such as debt-to-asset ratio and probability of default have a weight of 0.7, while secondary indicators such as cash flow coverage ratio have a weight of 0.3). Finally, for each industry, the weighted sum of its business relevance with all related industries is first calculated, and then multiplied by the absolute value of the change in the industry's risk assessment indicator and the corresponding risk weight to obtain the industry's relevance value.

[0096] For example, the real estate development industry has a "strong correlation" with the construction industry (correlation degree 0.85) and a "medium correlation" with the building materials production industry (correlation degree 0.62). The weighted sum of business correlation is 0.85×0.6+0.62×0.2=0.51+0.124=0.634. The risk assessment indicator for this industry is the debt-to-asset ratio (core indicator, risk weight 0.7), with an absolute change of 10%. Therefore, the correlation value of the real estate development industry is 0.634×10%×0.7=0.04438. Similarly, the construction industry has a "strong correlation" (0.85) with the real estate industry and a "relatively strong correlation" (0.78) with the building materials industry. The weighted sum of the business correlation is 0.85×0.6+0.78×0.4=0.51+0.312=0.822. The risk indicator is the cash flow coverage ratio (a secondary indicator with a weight of 0.3), with an absolute value of 0.3 for the change. The correlation value is 0.822×0.3×0.3=0.07398. The building materials production industry has a "strong correlation" (0.78) with the construction industry and a "moderate correlation" (0.62) with the real estate industry. The weighted sum of business correlation is 0.78×0.4+0.62×0.2=0.312+0.124=0.436. The risk indicator is the probability of default (core indicator, weight 0.7), with an absolute change of 1.2%. The correlation value is 0.436×1.2%×0.7=0.0036624.

[0097] S302: Determine whether there is a risk in the target industry chain based on the associated values ​​and preset thresholds.

[0098] Among them, the preset threshold is a risk judgment threshold set based on the historical risk data of the target industrial chain and the risk prevention and control requirements. It is used to measure whether the overall risk of the industrial chain exceeds the acceptable range.

[0099] Specifically, firstly, based on the industry attributes of the target industrial chain (such as the real estate industrial chain), historical risk event data, and regulatory requirements, preset thresholds are set: the warning threshold for single-industry correlation values ​​is 0.05, and the risk threshold is 0.07; the warning threshold for the overall correlation values ​​of the industrial chain (the sum of correlation values ​​of all industries) is 0.1, and the risk threshold is 0.15; secondly, the correlation values ​​of each industry and the overall correlation values ​​of the industrial chain are calculated separately; finally, based on the correlation values ​​and preset thresholds, it is determined whether there are risks in the target industrial chain.

[0100] Optionally, based on the associated values ​​and preset thresholds, it can be determined whether there is a risk in the target industry chain. Specific implementation methods include:

[0101] If the correlation value is greater than the preset threshold, the target industry chain is at risk;

[0102] If the associated value is less than or equal to the preset threshold, the target industry chain is at risk.

[0103] Specifically, the warning threshold for single-industry correlation values ​​is 0.05, and the risk threshold is 0.07. The warning threshold for the overall correlation values ​​of the industrial chain (the sum of correlation values ​​of all industries) is 0.1, and the risk threshold is 0.15. Next, the correlation values ​​of each industry and the overall correlation values ​​of the industrial chain are calculated separately to ensure that the calculation results accurately correspond to the industry composition of the target industrial chain. Finally, the risk judgment logic is executed: if any industry correlation value is greater than the single-industry risk threshold (0.07), or the overall correlation value of the industrial chain is greater than the overall risk threshold (0.15), then the target industrial chain is judged to have risk. If all industry correlation values ​​are less than or equal to the single-industry warning threshold (0.05) and the overall correlation value is less than or equal to the overall warning threshold (0.1), then the target industrial chain is judged to have no risk. If the industry correlation value is between the warning threshold and the risk threshold, or the overall correlation value is between the warning threshold and the risk threshold, it is judged as a potential risk (potential risks fall under the risk warning category and are not considered explicitly existing risks).

[0104] For example, the target industry chain consists of real estate development, construction, and building materials production. The correlation values ​​for each industry are calculated as follows: real estate development: 0.04438; construction: 0.07398; building materials production: 0.0036624. The overall correlation value for the industry chain is 0.04438 + 0.07398 + 0.0036624 = 0.1220224. Compared to the preset thresholds: the construction industry correlation value (0.07398) exceeds the single-industry risk threshold (0.07), satisfying the condition that "any industry correlation value > single-industry risk threshold." Simultaneously, the overall correlation value (0.1220224) falls between the overall warning threshold (0.1) and the overall risk threshold (0.15). According to the judgment rules, because a single-industry correlation value exceeds the risk threshold, the target industry chain is ultimately determined to be at risk.

[0105] S303: If there are risks in the target industry chain, the risk transmission factors and risk levels shall be determined based on the degree of business relevance.

[0106] Among them, the risk transmission factor is a parameter that quantifies the intensity of risk transmission between upstream and downstream industries in the industrial chain. Its value is equal to the product of the business correlation between industries and the correlation value between the risk source industry, reflecting the probability and degree of impact of risk transmission from the risk source to related industries.

[0107] Risk levels are classifications of risk severity based on the distribution of risk transmission factors and the overall risk level, typically divided into three levels: "general risk," "significant risk," and "major risk."

[0108] Specifically, the process begins with identifying the risk source industries: industries with correlation values ​​≥ the risk threshold or the highest correlation values ​​are selected as core risk sources. Next, risk transmission factors are calculated: the business correlation between the risk source industry and its upstream and downstream industries is multiplied by the correlation value of the risk source industry to obtain the risk transmission factor for each transmission path. Then, risk level assessment criteria are established: if the maximum risk transmission factor is <0.03 and the sum of all transmission factors is <0.08, it is classified as a general risk; if the maximum risk transmission factor is between 0.03 and 0.06 and the sum of all transmission factors is between 0.08 and 0.12, it is classified as a significant risk; if the maximum risk transmission factor is ≥0.06 or the sum of all transmission factors is ≥0.12, it is classified as a major risk. Finally, combining the numerical distribution of risk transmission factors and the assessment criteria, the risk transmission factors and corresponding risk levels of the industrial chain are determined.

[0109] For example, given that the target industry chain is known to have risks, with the core risk source industry being the construction industry (correlation value 0.07398). Based on business correlation: the construction industry has a business correlation of 0.85 with the real estate development industry and 0.78 with the building materials production industry. Calculating the risk transmission factors: real estate development industry transmission factor = 0.85 × 0.07398 ≈ 0.06288, building materials production industry transmission factor = 0.78 × 0.07398 ≈ 0.05770. According to the judgment criteria, the maximum risk transmission factor 0.06288 ≥ 0.06, therefore, the main risk transmission factors are determined to be 0.06288 (real estate direction) and 0.05770 (building materials direction), and the risk level of this target industry chain is classified as significant risk.

[0110] Figure 4 A schematic diagram of the business risk detection device provided in this application is shown below. Figure 4 As shown, the business risk detection device 400 provided in this embodiment includes:

[0111] The acquisition module 401 is used to acquire multimodal business data of the target industry chain. The target industry chain includes multiple industries with business transmission relationships. The multimodal business data is data generated by multiple business type systems of related enterprises in each industry of the target industry chain.

[0112] The processing module 402 is used to process the multimodal business data in each industry and generate joint feature vectors corresponding to each industry.

[0113] The determination module 403 is used to determine the business relationship information between industries based on the joint feature vectors corresponding to all industries, and generate the risk detection results of the target industrial chain based on the business relationship information.

[0114] In one possible implementation, the business risk detection device 400 further includes: a generation module 404;

[0115] The generation module 404 is used to extract features from multimodal business data in various industries and generate multimodal business feature data corresponding to each industry.

[0116] The processing module 402 is also used to encrypt the multimodal business feature data corresponding to each industry to obtain the encrypted feature data corresponding to each industry.

[0117] The processing module 402 is also used to perform multimodal fusion analysis on the encrypted feature data corresponding to each industry, and generate joint feature vectors corresponding to each industry.

[0118] In one possible implementation, the processing module 402 is further configured to encrypt the multimodal business feature data corresponding to each industry in sequence according to the modality type, so as to obtain the encrypted feature data corresponding to each industry. The encrypted feature data is a set of multiple homomorphically encrypted modal feature data.

[0119] In one possible implementation, the processing module 402 is further configured to extract semantic features from the text data to obtain a text feature vector.

[0120] The processing module 402 is also used to extract voiceprint features and emotion index from the speech data to obtain a speech feature vector;

[0121] The processing module 402 is also used to identify target business indicators from image data and obtain image feature vectors;

[0122] The processing module 402 is also used to perform weighted fusion of text feature vectors, speech feature vectors and image feature vectors to generate a joint feature vector.

[0123] In one possible implementation, the processing module 402 is further configured to extract the business correlation degree and risk assessment indicator change of each industry based on the joint feature vectors corresponding to all industries.

[0124] The generation module 404 is also used to generate risk detection results based on the business relevance and the change in risk assessment indicators.

[0125] In one possible implementation, the determining module 403 is further configured to determine the correlation values ​​of each industry based on the business correlation degree and the change in risk assessment indicators.

[0126] The determination module 403 is also used to determine whether there is a risk in the target industry chain based on the associated values ​​and preset thresholds;

[0127] Module 403 is also used to determine the risk transmission factor and risk level based on the degree of business relevance if there are risks in the target industrial chain.

[0128] In one possible implementation, the determining module 403 is further configured to determine that if the associated value is greater than a preset threshold, the target industrial chain is at risk.

[0129] The determination module 403 is also used to determine that if the associated value is less than or equal to a preset threshold, the target industrial chain is at risk.

[0130] The business risk detection device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0131] Figure 5 A schematic diagram of the business risk detection equipment provided in this application. Figure 5 As shown, the electronic device of this embodiment may include: at least one processor 501; and a memory 502 communicatively connected to the at least one processor; wherein the memory 502 stores instructions executable by the at least one processor 501, the instructions being executed by the at least one processor 501 to cause the electronic device to perform the method as described in any of the above embodiments.

[0132] Optionally, the memory 502 can be either standalone or integrated with the processor 501. When the memory 502 is set up independently, the device also includes a bus for connecting the memory 502 and the processor 501.

[0133] The implementation principle and technical effects of the electronic device provided in this embodiment can be found in the foregoing embodiments, and will not be repeated here.

[0134] This application also provides a computer-readable storage medium storing computer-executable instructions. When the computer-executable instructions are executed by a processor, the methods provided in any of the foregoing embodiments can be implemented.

[0135] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the method provided in any of the foregoing embodiments.

[0136] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0137] It should be further noted that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0138] It should be understood that the above-described device embodiments are merely illustrative, and the device of this application can also be implemented in other ways. For example, the division of units / modules in the above embodiments is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units, modules, or components may be combined, or integrated into another system, or some features may be ignored or not executed.

[0139] Furthermore, unless otherwise specified, the functional units / modules in the various embodiments of this application can be integrated into one unit / module, or each unit / module can exist physically separately, or two or more units / modules can be integrated together. The integrated units / modules described above can be implemented in hardware or as software program modules.

[0140] Unless otherwise specified, the processor can be any suitable hardware processor, such as a CPU, GPU, FPGA, DSP, and ASIC. Unless otherwise specified, the storage unit can be any suitable magnetic or magneto-optical storage medium, such as resistive random access memory (RRAM), dynamic random access memory (DRAM), static random access memory (SRAM), enhanced dynamic random access memory (EDRAM), high-bandwidth memory (HBM), hybrid memory cube (HMC), etc.

[0141] If the integrated unit / module is implemented as a software program module and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0142] In the above embodiments, the descriptions of each embodiment have their own emphasis. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0143] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0144] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A business risk detection method, characterized in that, The method includes: Acquire multimodal business data of the target industry chain, which includes multiple industries with business transmission relationships. The multimodal business data is data generated by multiple business type systems of related enterprises in each industry of the target industry chain. The multimodal business data in each industry are processed separately to generate joint feature vectors corresponding to each industry; Based on the joint feature vectors corresponding to all industries, business relationship information between industries is determined, and risk detection results of the target industrial chain are generated based on the business relationship information.

2. The method according to claim 1, characterized in that, The process of processing the multimodal business data in each industry to generate joint feature vectors corresponding to each industry includes: Feature extraction is performed on the multimodal business data in various industries to generate multimodal business feature data corresponding to each industry; The multimodal business feature data corresponding to each industry is encrypted to obtain the encrypted feature data corresponding to each industry. Multimodal fusion analysis is performed on the encrypted feature data corresponding to each industry to generate the joint feature vector corresponding to each industry.

3. The method according to claim 2, characterized in that, The encryption process is performed on the multimodal business feature data corresponding to each industry to obtain encrypted feature data corresponding to each industry, including: The multimodal business feature data corresponding to each industry are encrypted sequentially according to modality type to obtain the encrypted feature data corresponding to each industry. The encrypted feature data is a set of multiple homomorphically encrypted modal feature data.

4. The method according to claim 2, characterized in that, The encrypted feature data includes: text data, voice data, and image data. The step of performing multimodal fusion analysis on the encrypted feature data corresponding to each industry to generate the joint feature vector corresponding to each industry includes: Semantic features are extracted from the text data to obtain a text feature vector; The voice data is processed to extract voiceprint features and emotion index to obtain a voice feature vector; The image data is subjected to target business indicator identification to obtain image feature vectors; The text feature vector, the speech feature vector, and the image feature vector are weighted and fused to generate the joint feature vector.

5. The method according to claim 1, characterized in that, The step of determining business relationship information between industries based on the joint feature vectors corresponding to all industries, and generating risk detection results for the target industrial chain based on the business relationship information, includes: Based on the joint feature vectors corresponding to all industries, extract the business correlation degree and risk assessment indicator changes for each industry. The risk detection result is generated based on the business relevance and the change in the risk assessment indicators.

6. The method according to claim 5, characterized in that, The step of generating the risk detection result based on the business relevance and the change in the risk assessment indicators includes: Based on the business relevance and the changes in the risk assessment indicators, the correlation values ​​for each industry are determined; Based on the associated values ​​and preset thresholds, it is determined whether the target industry chain poses a risk; If the target industry chain has risks, then the risk transmission factor and risk level are determined based on the business relevance.

7. The method according to claim 6, characterized in that, The step of determining whether the target industry chain has risks based on the correlation values ​​and preset thresholds includes: If the correlation value is greater than the preset threshold, then the target industrial chain is at risk. If the correlation value is less than or equal to the preset threshold, then the target industrial chain is at risk.

8. A business risk detection device, characterized in that, include: The acquisition module is used to acquire multimodal business data of the target industry chain, which includes multiple industries with business transmission relationships. The multimodal business data is data generated by multiple business type systems of related enterprises in each industry of the target industry chain. The processing module is used to process the multimodal business data in each industry separately and generate joint feature vectors corresponding to each industry. The determination module is used to determine the business association information between industries based on the joint feature vectors corresponding to all industries, and generate the risk detection result of the target industrial chain based on the business association information.

9. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-7.

11. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method described in any one of claims 1-7.

Images