Supplier risk prevention and control system based on multi-modal AI and big data

Abstract

The invention relates to a supplier risk prevention and control system based on multi-modal AI and big data, and the system comprises a multi-modal recognizer which is used for recognizing the input information of a supplier, and carrying out the data extraction of the recognized information; the risk feature extractor is used for arranging the data extracted by the multi-modal identifier into a regular data structure and extracting effective features; the feature logic learner is used for training by using input knowledge-based data of suppliers and establishing a feature logic relationship; the feature logic matcher matches the word vectors of the risk feature extractor with supplier feature data according to the feature logic relationship, and stores the supplier feature data into a supplier feature library; the risk assessment module constructs a feature map structure of supplier data; and constructing a risk assessment coordinate system based on the feature map structure, and determining a supplier risk determination result based on a risk area represented by the supplier anti-risk capability and the supplier risk degree in the coordinate system.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent supply chain, and in particular to a supplier risk prevention and control system based on multimodal AI and big data. Background Art

[0002] Existing supplier risk prevention and control technologies are all based on risk condition restrictions or risk event determination methods. They all use rules to record the company's supplier management system and possible risk events of suppliers into the system, and write them into risk determination rules one by one in the system. When a supplier has an event that meets the rules or the weighted assessment score of multiple events reaches the system warning threshold, the supplier is judged to have potential risks or has already encountered risks. The existing technology method is a "rule condition" system, which enumerates possible risk rules and monitors and restricts them in the system.

[0003] The existing technology only lists the risk information of suppliers and gives early warnings. The final determination of risks mainly relies on people's professional judgment and / or conjecture, which is highly subjective and non-standard. It is also unable to process unstructured risk data, and is in the form of risk rule enumeration. The types of risks that can be pre-defined in the system are limited, and the risk dimensions are too single. Many unregistered and enumerated risks are often not discovered. The existing technology of risk prevention and control is generally used only as an auxiliary, and risks are often only early warnings and reminders. The effectiveness and accuracy of risk assessment are also relatively poor, and the precision is not high. Summary of the invention

[0004] The purpose of the present invention is to provide a supplier risk prevention and control system based on multimodal AI and big data, which is used to solve the problem of incomplete and low accuracy of supplier risks.

[0005] The present invention discloses a supplier risk prevention and control system based on multimodal AI and big data, which includes: a multimodal identifier, a supplier data warehouse, a risk feature extractor, a feature logic matcher, a feature logic learner, a supplier feature library and a risk assessment module; the multimodal identifier is used to identify the input information of the supplier, and extract data from the identified information, and store the extracted data in the supplier data warehouse; the risk feature extractor organizes the data extracted by the multimodal identifier into a regular data structure, extracts effective features, and converts the extracted effective features into the form of word vectors; the feature logic learner is used to use the input supplier's The knowledge-based data is trained and the vector similarity is annotated. According to the annotated vector similarity, the feature logic relationship is established; the feature logic matcher matches the word vector of the risk feature extractor with the supplier feature data according to the feature logic relationship and stores it in the supplier feature library; the risk assessment module obtains the data of the supplier data warehouse and the supplier feature library respectively, and constructs the feature graph structure of the supplier data; based on the feature graph structure, a risk assessment coordinate system is constructed, and the coordinate axes represent the supplier's risk resistance and the supplier's risk level respectively. Based on the risk areas represented by the supplier's risk resistance and the supplier's risk level in the coordinate system, the supplier risk judgment result is determined and output.

[0006] According to an embodiment of the supplier risk prevention and control system based on multimodal AI and big data of the present invention, the supplier feature library includes a supplier risk portrait feature library and a supplier word-of-mouth portrait library; the feature logic matcher matches the word vectors of the risk feature extractor into risk portrait feature data and word-of-mouth portrait feature data according to the feature logic relationship; the risk portrait feature data is stored in the supplier risk portrait feature library, and the word-of-mouth portrait feature data is stored in the supplier word-of-mouth portrait feature library; the risk assessment module reads the supplier's risk portrait feature data, word-of-mouth portrait feature data, and reads various data of the supplier from the supplier data warehouse to construct a feature graph structure.

[0007] According to an embodiment of the supplier risk prevention and control system based on multimodal AI and big data of the present invention, the multimodal identifier can perform OCR recognition on the supplier's image data; can perform entity recognition on the supplier's unstructured text data; and perform data extraction on the identified supplier data to extract relevant data on supplier risks.

[0008] According to an embodiment of the supplier risk prevention and control system based on multimodal AI and big data of the present invention, a risk feature extractor predefines a variety of feature templates of supplier risk data, and the risk feature extractor organizes data according to the feature templates corresponding to the types of supplier risk data; the risk feature extractor extracts valid features from the organized rule data, and converts the extracted data into the form of word vectors, and sends them to the feature logic matcher.

[0009] According to an embodiment of the supplier risk prevention and control system based on multimodal AI and big data of the present invention, the risk feature extractor also includes a dirty word filter to filter the dirty words in the rule data.

[0010] According to an embodiment of the supplier risk prevention and control system based on multimodal AI and big data of the present invention, the feature logic matcher performs similarity matching according to the distance between each word vector in the vector space, and divides the matching results into risk portrait feature data and word-of-mouth portrait feature data.

[0011] According to an embodiment of the supplier risk prevention and control system based on multimodal AI and big data of the present invention, a feature logic learner learns the feature logic relationship of supplier data, and through the process of expert labeling and knowledge-based data import, the feature values ​​and feature names of different supplier data are constructed into feature logic relationships.

[0012] According to an embodiment of the supplier risk prevention and control system based on multimodal AI and big data of the present invention, the supplier risk determination result is determined based on the risk area represented by the supplier's risk resistance ability and the supplier's risk degree in the coordinate system, specifically including: the risk assessment coordinate system includes an X-axis and a Y-axis, the X-axis represents the risk resistance ability; the value of X is calculated based on the weighted information of the supplier; the Y-axis represents the risk degree, Y = risk event type multiplied by the number of risk events, and the value of the risk event type is determined according to the severity of each risk event; according to the calculation results of X and Y, the absolute risk red line, relative risk warning line and risk level are delineated from high to low on the risk assessment coordinate system according to the risk degree. The warning line and risk red line are drawn based on the risk level value of the Y axis. If the risk level exceeds a certain threshold, it is regarded as an absolute risk. The relative risk warning line and risk warning line are drawn based on the values ​​of Y and X. The risk area is divided into absolute high-risk area, relative high-risk area, medium-risk area and low-risk area. The area above the absolute risk red line is the absolute high-risk area. The area below the absolute risk red line and above the relative risk warning line is the relative high-risk area. The area below the absolute risk red line, below the relative risk warning line and above the risk warning line is the medium-risk area. The area below the risk warning line is the low-risk area. The risk assessment result is determined based on the division of the risk areas.

[0013] According to an embodiment of the supplier risk prevention and control system based on multimodal AI and big data of the present invention, the risk assessment module reads the risk portrait feature data of the supplier from the supplier risk portrait feature library; the risk assessment module reads the reputation portrait feature data of the supplier from the supplier reputation portrait feature library; the risk assessment module reads various data of the supplier from the supplier data warehouse, processes them through the streaming computing engine, and constructs the various data into a data set on a time series; the risk assessment module constructs a feature graph structure for the acquired risk portrait feature data, reputation portrait feature data, and supplier data set of the supplier through time series, structural features, and logical relationships.

[0014] According to an embodiment of the supplier risk prevention and control system based on multimodal AI and big data of the present invention, the risk portrait feature data and the word-of-mouth portrait feature data are generated into a graph structure by using a knowledge graph method, and the relationship between nodes and edges is constructed. The nodes are composed of data, and the edges are the relationship between the data between nodes, so as to form a feature graph structure; the data in the supplier data that has no relationship with the risk portrait feature data and the word-of-mouth portrait feature data is used as an independent factor; the risk assessment module uses the feature graph structure and independent factors to construct a risk assessment coordinate system.

[0015] The supplier risk prevention and control system based on multimodal AI and big data of the present invention realizes fully automated and intelligent supplier risk discovery and early warning prevention and control capabilities, directly outputs risk levels, and accurately gives risk judgment results. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 The figure shows the principle diagram of the supplier risk prevention and control system based on multimodal AI and big data technology of the present invention;

[0017] Figure 2 Shown is a schematic diagram of a multi-dimensional word vector for supplier data;

[0018] Figure 3 The figure shows an example of matching the word vectors of supplier data using a feature logic learner;

[0019] Figure 4 Shown is a schematic diagram of a risk assessment coordinate system. DETAILED DESCRIPTION

[0020] In order to make the purpose, content, and advantages of the present invention more clear, the specific implementation methods of the present invention are further described in detail below in conjunction with the accompanying drawings and examples.

[0021] The technical terms of the present invention are explained as follows:

[0022] Word vector: A general term for a set of language modeling and feature learning techniques in word-embedded natural language processing (NLP). Words or phrases from the vocabulary are mapped to vectors of real numbers, converting words into dense vectors, which are mathematical representations of text language.

[0023] Knowledge graph: A knowledge domain visualization or knowledge domain mapping map. It is a series of various graphics that show the knowledge development process and structural relationship, using visualization technology to describe knowledge resources and their carriers.

[0024] Figure 1 The figure shows the principle diagram of the supplier risk prevention and control system based on multimodal AI and big data technology of the present invention. Figure 1 As shown, the supplier risk prevention and control system based on multimodal AI and big data technology includes: multimodal identifier 1, supplier data warehouse 7, streaming computing engine 8, risk feature extractor 2, feature logic matcher 4, feature logic learner 3, supplier risk portrait feature library 5, supplier reputation portrait library 6, risk assessment module 9, and streaming computing engine 8.

[0025] like Figure 1 As shown, the multimodal identifier 1 is used to identify the input information of the supplier and perform data extraction.

[0026] Among them, the multimodal identifier 1 can realize multimodal recognition of various types of data such as credit information, transaction data, qualifications, capabilities, and risk information of multiple suppliers. Among them, if the business license, qualification certificate, etc. are in the form of pictures or PDF, they can be recognized by the OCR of the multimodal identifier 1. If the bidding and tender documents are unstructured large texts, they can be recognized by entity recognition. Data such as behavior, performance, and transactions are structured data in the database, and public opinion data comes from rich text public opinion on the Internet. The multimodal identifier 1 extracts effective data content through data extraction methods. Data extraction can use the BERT pre-trained model of deep learning to realize the recognition and extraction of entity nouns. The large text recognized by OCR can be input into the trained BERT model to output valid entity nouns: such as projects, amounts, locations, commodity nouns, technical nouns, etc.

[0027] Specifically, the entity recognition method can be text entity recognition. The text is generally a long and short text content, which is a paragraph composed of sentences. By identifying text entities, that is, meaningful text words and phrases, for example, key nouns, words, phrases, such as: place, time, product name, etc., they are identified from the text and then extracted.

[0028] like Figure 1As shown, the risk feature extractor 2 processes the various types of data identified by the multimodal identifier 1. The risk feature extractor 2 has a variety of defined feature templates. Different types of data correspond to different feature templates. There are many types of feature templates. The risk feature extractor 2 matches the corresponding feature templates according to the needs and matches according to the type of supplier data. The data type and feature template can be defined by yourself and are not limited to certain types. For example, the data of supplier bidding behavior, the number of bids, the number of winning bids, and the amount of winning bids.

[0029] like Figure 1 As shown, feature templates define text rules of specific structures, which can be expressed by regular expressions or dictionaries. For example, feature templates include symbol types, such as "￥" and "$" to represent RMB and USD, and complex feature templates also have various rule formulas or regular expressions.

[0030] The risk feature extractor 2 processes the supplier data through the feature template and forms a regular supplier data structure according to the feature template. For example, each feature data is stored as an entry. Preferably, the risk feature extractor 2 also includes a dirty word filter, and the dirty words (useless data) in the data are filtered through the dirty word filter. Each dirty filter is composed of a dirty word dictionary and regular rules, and dirty words that meet the regular rules and dictionary will be filtered out.

[0031] For a specific embodiment, the process of the risk feature extractor 2 processing supplier data through the feature template includes:

[0032] The source input data is: Beijing Telide Company is a limited liability company, which applied for registration and establishment with the Industry and Commerce Bureau on September 7, 2021.

[0033] The structured data formed by the data processed by the feature template is:

[0034] Company name: Beijing Telide Co., Ltd.

[0035] Registration date: September 7, 2021.

[0036] like Figure 1 As shown, the risk feature extractor 2 also includes an effective feature word extraction module, which extracts effective features from structured data. The effective feature extraction can be based on the dictionary, rules and acquaintance relationship of effective features, and determine the words that meet the dictionary features, rule features and acquaintance as effective feature words.

[0037] For a specific embodiment, the extracted supplier effective features may include:

[0038] National characteristics: China, People's Republic of China, United States, United States of America…;

[0039] Provincial and municipal regional characteristics: Beijing, Shanghai, Tianjin, Daxing District, Beijing, Chaoyang District, Beijing…;

[0040] Location format features: Zhongguancun, Haidian District, Beijing…;

[0041] Model specification features: length 19 meters, area 113 square meters, length, width and height 55cm*76cm*175cm…;

[0042] Time characteristics: 1998, 98, October 8, 2022, 2023-10-8…;

[0043] Catalog features: background, company introduction, performance certification, operation and maintenance plan, security configuration...

[0044] The risk feature extractor 2 has a feature vector conversion module, which is used to convert valid feature words into multi-dimensional word vectors. Specifically, the feature conversion of word vectors can be performed through the Bert model, stored in the form of word vectors, and transmitted to the feature logic matcher 4.

[0045] Figure 2 The following is a schematic diagram of a multi-dimensional word vector of supplier data. Figure 2 As shown, for a preferred embodiment, the multi-dimensional word vector can be a 300-dimensional word vector, including 300 data feature dimensions. Figure 2 Each number in is a dimension.

[0046] like Figure 1 As shown, the feature logic matcher 4 processes the word vector provided by the risk feature extractor 2 into portrait data, and stores it in the supplier risk portrait feature library 5 or the supplier reputation portrait library 6.

[0047] The feature logic matcher 4 uses the feature logic relationship to match the word vectors of the supplier data. The feature logic matcher 4 uses the feature logic relationship to match the word vectors of the supplier data, specifically including: the word vectors of similar supplier data are close in the vector space, based on which the similarity between the word vectors is calculated, and matching is performed according to the distance between the word vectors.

[0048] like Figure 1 As shown, the feature logic matcher 4 calls the learning result of the feature logic learner 3 to perform feature logic matching. The feature logic learner 3 is used to learn the feature logic relationship. The learning process of the feature logic learner 3 is a process of expert annotation and knowledge-based data import, which constructs the logical relationship between different feature values ​​and feature names of the supplier data. For example, the specific address "Beijing" of the supplier data is a "city", and the type "Forbidden City" is a "scenic spot", and "March 5" is a date.

[0049] The feature logic learner 3 performs labeling and training of word vectors, and then searches for the same or similar vectors according to the vector similarity of the labeled feature values, and performs mapping of corresponding feature names.

[0050] Figure 3 The following is an example of matching the word vectors of the supplier data of the feature logic learner. Figure 3 As shown, Figure 3 The left column is the feature value and the right column is the feature name. The feature value is matched with the feature name. Valid feature data that conforms to the feature logic will be stored in the corresponding feature structure. Valid feature data that does not conform to the feature logic will be added with an unidentified mark and stored for secondary learning. Through system learning and expert annotation, the unidentified feature logic can be constructed into a new feature relationship and stored in the warehouse.

[0051] like Figure 1 As shown, the feature logic matcher 4 can also classify the matched feature data into corresponding data types according to the type to which it belongs. Data types include risk type and reputation type. If the matched feature data is of risk type, it is stored in the supplier risk profile feature library 5, and the reputation type is stored in the supplier reputation profile feature library 6. For example, customer evaluation belongs to the reputation type, while professional qualifications can belong to the risk type.

[0052] like Figure 1 As shown, the risk assessment module 9 is respectively connected to the supplier data warehouse 7, the supplier input port, the supplier risk profile feature library 5, and the supplier reputation profile library 6. The supplier input port is used to input the supplier's information, which may be a name or a unified social credit code.

[0053] like Figure 1 As shown, the supplier input port may specifically be an API interface.

[0054] like Figure 1 As shown, the risk assessment module 9 can call the data in the supplier data warehouse 7 to query and determine whether there is corresponding data for the information of the supplier input into the supplier data warehouse 7. If the data of the supplier is not found, it means that the supplier input is incorrect or does not belong to the supplier in the supplier data warehouse 7, then the relevant query information is returned and the query ends.

[0055] If the supplier's data is found, the risk assessment module 9 will perform relevant risk assessment on the supplier's information input through the API.

[0056] The risk assessment module 9 includes a factor matching and cross-combination module 91. The specific process of the factor matching and cross-combination module 91 performing relevant risk assessment on the supplier's information may include:

[0057] Reading relevant information from the supplier risk profile feature library 5 and the supplier reputation profile feature library 6 may specifically include:

[0058] Read the characteristic data of the risk profile of the supplier from the supplier risk profile characteristic library 5 as the data factor of the risk profile;

[0059] Read the characteristic data of the supplier's reputation portrait from the supplier's reputation portrait feature library as the data factor of the reputation portrait;

[0060] The supplier's business data is read from the supplier data warehouse 7, processed by the stream computing engine 8, and the data is constructed into a data set factor on a time series. The data set factor includes the items of the read data, which may include registration date, personnel, registration location, registered capital and other information.

[0061] The above factors are indicators or labels after data processing and calculation. Raw data is the original data field after structured. For example, the number of employees in an enterprise is 186, which is the raw data read. The medium-sized enterprise (100-500 employees) is a statistical indicator and belongs to a factor.

[0062] The risk assessment module 9 constructs a graph structure of factor matching and intersection for the acquired data factors of the risk portrait, the data factors of the word-of-mouth portrait, and the data set factors of the supplier data warehouse 7 through time series, structural characteristics, and logical relationships. Preferably, the supplier data knowledge graph method can be used to construct the structure of the graph generated by each of the above factors into a relationship between nodes and edges. The node is the data entity, and the edge refers to the relationship. For example, Company A-Holdings-Company B. Company A and Company B are both nodes, and holdings are relationships. Company A, the parent of Company B (which holds a controlling stake in the company), went bankrupt, which can be used as a correlation factor (the company and the bankrupt company are in a holding relationship) to determine that Company B also has a great risk. Individual factors that have no relationship with other supplier factors are treated as independent factors.

[0063] Figure 4 The diagram shows a risk assessment coordinate system. Figure 4 As shown, the risk assessment module 9 includes a risk grading model 92. The risk grading model 92 obtains a graph structure of matching and crossing of various factors (including independent factors) to construct a risk assessment coordinate system. The coordinate system includes two coordinate axes:

[0064] The X-axis represents the risk resistance value. The basic formula for calculating risk resistance can be: risk resistance value = a*n1+b*n2+c*n3+d*n4+e*n5+f*n6… The various factors of supplier strength and scale are weighted and calculated. The score is calculated after weighting according to the factors reflecting the scale and strength of the enterprise, such as the supplier's registration time a, capital b, personnel c, finance d, intellectual property e, etc., where n1, n2, n3, n4, n5, and n6 are all weighted values.

[0065] The Y axis represents the risk level value, which is the risk level value = risk event type * number of risk events. Risk events are determined based on severity, and then the number of risks is combined to calculate the risk level value of the supplier.

[0066] like Figure 4 As shown, the risk distribution area is determined based on the values ​​of the X-axis and Y-axis represented by the supplier information. Specifically, the absolute risk red line, relative risk warning line, and risk reminder line can be drawn according to the calculated values ​​of the X-axis and Y-axis. Among them, the delineation of the risk red line only considers the risk level value of the Y-axis, and exceeding a certain risk level threshold is regarded as an absolute risk. The relative risk warning line and risk reminder line are comprehensively delineated based on the supplier's risk level and risk resistance ability. Among them, high-risk (absolute) area: the area above the R absolute risk red line; high-risk (relative) area: the area below the R absolute risk red line and above the O relative risk warning line; medium-risk area: the area below the R absolute risk red line, below the O relative risk warning line, and above the B risk reminder line; low-risk area: the area below the B risk reminder line.

[0067] like Figure 1 As shown, the risk assessment module 9 outputs the risk determination result. The risk determination result includes: the risk level determined by the risk assessment module 9 and the specific risk event factor, risk resistance ability score and other data.

[0068] The present invention uses multimodal information recognition technology to identify effective features from information in various modes such as pictures, texts, and unstructured large texts, and constructs a set of supplier risk portraits and word-of-mouth portraits through means such as feature extraction, denoising, and association. By cross-combining factors from the image features stored in the supplier risk portrait, the supplier word-of-mouth portrait, and the data sequence set obtained from the supplier data, the risk level of the supplier is determined through calculation formulas and risk grading models.

[0069] The present invention uses a supplier risk assessment module to achieve more quantitative, scientific and objective risk indicator calculations. Through a multi-factor cross-supplier risk assessment module, various risk factors are uniformly measured and combined for calculation, solving the existing technical defects of "subjective assumptions" and "single solidification". Through multimodal technology, various types of supplier risk information are obtained to build supplier risk and reputation portraits, analyze and evaluate from various dimensions that reflect supplier risk behavior, and achieve a more comprehensive supplier risk assessment. The present invention establishes a risk assessment module, a risk grading model and a risk feature matcher. Through the above model and matcher, a fully automated and intelligent supplier risk discovery and early warning prevention and control capability is realized, and risk levels and metrics are directly output to accurately give risk determination results.

[0070] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A supplier risk prevention and control system based on multimodal AI and big data, characterized in that: include: Multimodal identifier, supplier data warehouse, risk feature extractor, feature logic matcher, feature logic learner, supplier feature library and risk assessment module; The multimodal identifier is used to identify the supplier's input information and extract data from the identified information, and the extracted data is stored in the supplier data warehouse; The risk feature extractor organizes the data extracted by the multimodal identifier into a regular data structure, extracts effective features, and converts the extracted effective features into word vectors; The feature logic learner is used to use the input supplier's knowledge data for training and to annotate vector similarities, and to establish feature logic relationships based on the annotated vector similarities; The feature logic matcher matches the word vector of the risk feature extractor with the supplier feature data according to the feature logic relationship and stores it in the supplier feature library; The risk assessment module obtains data from the supplier data warehouse and the supplier feature library respectively, and constructs a feature graph structure of the supplier data; A risk assessment coordinate system is constructed based on the characteristic graph structure, and the coordinate axes represent the supplier's risk resistance and the supplier's risk level, respectively. Based on the risk areas represented by the supplier's risk resistance and the supplier's risk level in the coordinate system, the supplier risk assessment result is determined and output.

2. The supplier risk prevention and control system based on multimodal AI and big data as claimed in claim 1, characterized in that: The supplier feature database includes the supplier risk profile feature database and the supplier reputation profile database; The feature logic matcher matches the word vectors of the risk feature extractor into risk profile feature data and word-of-mouth profile feature data according to the feature logic relationship; The risk profile feature data is stored in the supplier risk profile feature database, and the reputation profile feature data is stored in the supplier reputation profile feature database; The risk assessment module reads the supplier's risk profile feature data, reputation profile feature data, and various supplier data from the supplier data warehouse to build a feature graph structure.

3. The supplier risk prevention and control system based on multimodal AI and big data as claimed in claim 1, characterized in that: The multimodal identifier can perform OCR recognition on the supplier's image data; it can perform entity recognition on the supplier's unstructured text data; and it can extract data from the identified supplier data to extract relevant data on supplier risks.

4. The supplier risk prevention and control system based on multimodal AI and big data as claimed in claim 1, characterized in that: The risk feature extractor pre-defines a variety of feature templates for supplier risk data, and the risk feature extractor organizes data according to the feature templates corresponding to the types of supplier risk data; The risk feature extractor extracts effective features from the sorted rule data, converts the extracted data into word vectors, and sends them to the feature logic matcher.

5. The supplier risk prevention and control system based on multimodal AI and big data as claimed in claim 1, characterized in that: The risk feature extractor also includes a dirty word filter to filter the dirty words in the rule data.

6. The supplier risk prevention and control system based on multimodal AI and big data as claimed in claim 2, characterized in that: The feature logic matcher performs similarity matching based on the distance between each word vector in the vector space, and divides the matching results into risk portrait feature data and word-of-mouth portrait feature data.

7. The supplier risk prevention and control system based on multimodal AI and big data as claimed in claim 1, characterized in that: The feature logic learner learns the feature logic relationship of supplier data. Through the process of expert labeling and knowledge-based data import, the feature values ​​and feature names of different supplier data are constructed into feature logic relationships.

8. The supplier risk prevention and control system based on multimodal AI and big data as claimed in claim 1, characterized in that: Based on the risk area represented by the supplier's risk resistance and the supplier's risk level in the coordinate system, the supplier risk assessment results are determined, including: The risk assessment coordinate system includes the X-axis and the Y-axis, where the X-axis represents the ability to resist risks; The value of X is calculated based on the weighted information from suppliers; The Y axis represents the risk level, Y = risk event type multiplied by the number of risk events. The value of the risk event type is determined according to the severity of each risk event. Based on the calculation results of X and Y, the absolute risk red line, relative risk warning line and risk reminder line are drawn from high to low according to the risk level on the risk assessment coordinate system. The risk red line is drawn only considering the risk level value on the Y axis. If it exceeds a certain risk level threshold, it is considered as an absolute risk; the relative risk warning line and risk reminder line are drawn based on the values ​​of Y and X; The risk areas are divided into absolute high-risk areas, relatively high-risk areas, medium-risk areas and low-risk areas; the areas above the absolute risk red line are absolute high-risk areas; the areas below the absolute risk red line and above the relative risk warning line are relatively high-risk areas; the areas below the absolute risk red line, below the relative risk warning line and above the risk warning line are medium-risk areas; the areas below the risk warning line are low-risk areas; The risk assessment results are determined based on the risk area division results.

9. The supplier risk prevention and control system based on multimodal AI and big data as claimed in claim 2, characterized in that: The risk assessment module reads the risk profile feature data of the supplier from the supplier risk profile feature library; The risk assessment module reads the reputation profile feature data of the supplier from the supplier reputation profile feature library; The risk assessment module reads various supplier data from the supplier data warehouse, processes it through the streaming computing engine, and constructs the data into a time series data set; The risk assessment module constructs a feature graph structure for the acquired risk profile feature data, word-of-mouth profile feature data, and supplier data set through time series, structural features, and logical relationships.

10. The supplier risk prevention and control system based on multimodal AI and big data as claimed in claim 9, characterized in that: Using the knowledge graph method, the risk profile feature data and the word-of-mouth profile feature data are generated into a graph structure, and the relationship between nodes and edges is constructed. The nodes are composed of data, and the edges are the relationship between the data between nodes, so as to form a feature graph structure; The data in the supplier data that has no relationship with the risk profile feature data and the reputation profile feature data are taken as independent factors; The risk assessment module uses the characteristic graph structure and independent factors to construct a risk assessment coordinate system.

Images