Mail fusion governance device with AI identification technology

Abstract

The invention discloses a mail fusion governance device with an AI identification technology, and particularly relates to the technical field of mail governance, and the mail fusion governance device comprises a certification data receiving module, a multi-mode intelligent analysis engine, a mail account relation modeling module, a customs data collision engine and an Excel result output module. According to the method, intelligent structured conversion of multi-format attachments is realized by constructing a mail governance architecture fusing AI multi-modal identification and relation mining, and unstructured service data scattered in the attachments are automatically extracted and mapped into standard database fields by utilizing cooperative work of a large language model and a multi-modal model, so that the mail governance efficiency is improved. Establishing a dynamic mail entity relation graph, and generating an association network based on mail contact characteristics to strengthen the business traceability; a cross-source data intelligent collision engine is designed, structured mail data is compared with third-party systems such as customs declaration in real time, risk clues or business evidence closed loops are automatically output, and the time cost of manual checking is greatly reduced.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of mail governance, in particular to a mail fusion governance device with AI identification technology. BACKGROUND

[0002] The current enterprise-level mail system is under the pressure of processing massive heterogeneous data, and the traditional governance technology has significant limitations. In terms of spam filtering, the mainstream solution relies on a static rule engine (such as keyword matching and black / white list mechanism), which is difficult to dynamically adapt to the attack characteristics of new phishing emails and variant spam emails, resulting in a high misjudgment rate. For the unstructured data (including Excel, Word, PDF and image formats) commonly existing in the attachments of emails, the existing processing technology is fragmented: text attachments mainly extract fixed fields through regular expressions, and image attachments use OCR technology to recognize text content, but the collaborative analysis capability of cross-modal data is missing, making the extraction completeness of key information (such as cargo attributes and transaction terms) in business documents such as bills of lading and invoices severely insufficient. In terms of data correlation, conventional mail systems only record basic metadata (sender / recipient / timestamp), and do not build a dynamic relationship network between email entities, which restricts the traceability of complex business chains. The most prominent problem is that the mail system is completely isolated from external business data sources (such as customs and logistics systems), and cross-system data comparison completely relies on manual operation, resulting in low case handling efficiency.

[0003] The closest prior art solution adopts a two-level processing architecture: the first level uses a convolutional neural network (CNN) to analyze the header features and the content of the email to achieve initial screening of spam emails; the second level processes attachments, using a pre-trained language model to analyze text attachments (Word / PDF) and combining a computer vision model to identify the table structure in image attachments. However, this solution has fundamental flaws: multi-modal data processing is isolated from each other, text and image features are not linked for analysis, and the internal contradictions of cross-modal data cannot be identified; the structured capability is limited to pre-set field templates, and the recognition error rate for variable fields (such as mark header and container number) in non-standard business documents such as bills of lading and container manifests is high; the relationship between email entities completely relies on manual definition, and lacks an automated correlation network modeling; and there is no external system data interface designed, making it impossible to perform automatic collision analysis with business data such as customs declarations. These defects make it difficult for the existing solution to meet the core needs of enterprise-level mail intelligence governance.

[0004] Therefore, there is an urgent need for a mail fusion governance device with AI identification technology to solve the above problems. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, embodiments of the present application provide a mail fusion governance device with AI recognition technology, which constructs a mail governance architecture of fusion AI multi-modal recognition and relationship mining, realizes intelligent structured conversion of multi-format attachments, automatically extracts and maps unstructured business data scattered in attachments into standard database fields by using large language model and multi-modal model collaborative work, establishes a dynamic mail entity relationship graph, generates an association network based on mail correspondence characteristics to strengthen business traceability, designs a cross-source data intelligent collision engine, compares structured mail data with customs and other third-party systems in real time, automatically outputs risk clues or business evidence closed loop, and greatly reduces the time cost of manual checking, so as to solve the problems raised in the above background art.

[0006] To achieve the above object, the present application provides the following technical scheme: a mail fusion governance device with AI recognition technology, comprising a verification data receiving module, a multi-modal intelligent analysis engine, a mail account relationship modeling module, a customs data collision engine and an Excel result output module; The verification data receiving module is used for receiving mail data obtained from Internet enterprises, and performing format standardization and integrity check on the verification data, and the verification data receiving module comprises a data receiving interface, a format converter and a data verifier; The multi-modal intelligent analysis engine is used for unified analysis and structured processing of Excel, Word, PDF and picture multi-format data in mail attachments, and the multi-modal intelligent analysis engine integrates a large language model processing unit and a computer vision processing unit, and realizes collaborative analysis of cross-modal data through a feature fusion layer; The mail account relationship modeling module is used for constructing a dynamic relationship network based on mail sending and receiving accounts, and the mail account relationship modeling module comprises an account identifier, a relationship extractor and a graph database; The customs data collision engine is used for intelligent comparison and analysis of structured processed mail data and customs declaration data, and the customs data collision engine comprises a customs data interface, a similarity calculation module and an anomaly detection module; The Excel result output module is used for generating a standardized Excel format analysis report, and the Excel result output module comprises a data organizer, an Excel template engine and a file exporter.

[0007] A use method of a mail fusion governance device with AI recognition technology, comprising a mail fusion governance device with AI recognition technology and the following steps: S1, verification data receiving and preprocessing stage; A, the certification data receiving module receives the certification mail data provided by the Internet enterprise through a standardized interface. The data formats include but are not limited to mail original files, database export files and structured data packets; B, the format converter uniformly converts the heterogeneous data formats provided by different enterprises, extracts the mail header information, the text content and the attachment data respectively and stores them into the standardized data structure; C, the data verifier checks the integrity and validity of the certification data, checks the legality of the key fields such as the mail timestamp, the sender address and the recipient address, and marks the abnormal or missing data records; D, the preprocessing module cleans the mail text by removing HTML tags, special characters and redundant spaces, and simultaneously performs virus scanning and format identification on the attachments, so as to ensure the safety and accuracy of subsequent processing; S2, the multi-modal attachment intelligent analysis stage; A, the attachment format identifier automatically detects the file types of the mail attachments, including Excel tables, Word documents, PDF files, JPG / PNG images, and selects the corresponding analysis strategy according to the file types; B, the large language model processing unit is responsible for analyzing text attachments. For Excel files, it extracts cell data and formula information through table structure analysis. For Word documents, it extracts paragraph, table and chart content through document object model. For PDF files, it obtains editable text content through text extraction algorithm; C, for image attachments, first perform image preprocessing including denoising, enhancement and geometric correction, then use multi-modal model to identify the text content in the image, and finally use deep learning model to understand the layout and semantic structure of the image; D, the feature fusion layer fuses the text feature vector and the image feature vector through multi-layer perception, generates a unified multi-modal feature representation, and dynamically adjusts the weights of different modal features through attention mechanism; E, the business entity recognition module extracts key business fields such as bill of lading number, packing list number, invoice number, cargo name, quantity, amount, sender and receiver from the fused features, and establishes semantic association relationships between the fields; S3, the mail account relationship network construction stage; A, the account identifier extracts the sender's mailbox address and the recipient's mailbox address from the mail header information, and standardizes the mailbox address, including uniform case, domain name standardization and alias resolution; B, the account entity analysis module identifies the enterprise or organization to which the mailbox address belongs through the domain name part of the mailbox address, and infers the personal identity information through the username part, and establishes the mapping relationship between the mailbox account and the real entity; C. Relationship extractor analyzes the mail correspondence records, calculates the interaction frequency, time distribution and content correlation between accounts, identifies direct communication relationships and indirect association relationships; D. Relationship strength calculation module quantifies the closeness of the relationship between accounts based on indicators such as the number of emails, time span, and sets a threshold to filter weakly associated relationships; E. Graph database uses a node-edge model to store account relationship networks, nodes represent mailbox accounts and their attribute information, edges represent communication relationships between accounts and their strength weights, supporting complex graph queries and path analysis operations; S4, Customs declaration data collision analysis stage; A. Customs data obtains customs declaration data through uploading customs declaration excel, including customs declaration number, import and export enterprise information, cargo list, declared value, customs clearance time and other key fields; B. Data preprocessing module cleans and standardizes customs declaration data, unifies the expression of cargo names, standardizes enterprise names and address information, and processes unit conversion of numerical fields; C. Field mapping module establishes the field correspondence relationship between email data and customs data, matches the cargo information in email attachments with the commodity information in the declaration form, and associates the enterprise information in the email with the sender and receiver information in the declaration form; D. Similarity calculation module uses multiple algorithms for data comparison, including string edit distance algorithm to compare text similarity, semantic vector algorithm to calculate semantic similarity, and numerical range algorithm to judge the matching degree of amount and quantity; E. Abnormality detection module identifies inconsistencies between email data and customs data, including discrepancies in cargo names, quantity differences, amount abnormalities, time conflicts, etc., and calculates the risk score of the abnormality degree; F. Association analysis module analyzes the cross-association between email account relationship networks and customs declaration data through graph algorithms, discovers potential business chains and capital flow directions, and identifies suspicious trade behavior patterns; S5, Excel format result output stage; A. Data organizer organizes all analysis results according to the predefined data structure, including account relationship network table and data collision table; B. Excel template engine selects the corresponding Excel template according to different analysis requirements, the template contains pre-set table headers, format styles and formula calculations, ensuring the standardization and readability of the output results; C. Data filling module fills the organized analysis data into the corresponding cells of the Excel worksheet according to the template structure, and automatically adjusts the column width, row height and font format; D. The chart generation module creates data visualization charts in Excel, including account relationship network diagrams and data collision statistics charts, enhancing the intuitiveness of analysis results.

[0008] The technical effects and advantages of this invention are as follows: This invention constructs an email governance architecture that integrates AI multimodal recognition and relationship mining, enabling intelligent structured conversion of multi-format attachments. By leveraging the collaborative work of a large language model and a multimodal model, it automatically extracts and maps unstructured business data scattered in attachments into standardized database fields, establishes a dynamic email entity relationship graph, and generates an association network based on email communication characteristics to enhance business traceability capabilities. Furthermore, it designs a cross-source data intelligent collision engine to compare structured email data with third-party systems such as customs declarations in real time, automatically outputting risk clues or business evidence loops, significantly reducing the time and cost of manual verification. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0010] Figure 2 This is a flowchart of the email fusion management device of the present invention. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] As attached Figure 1 and attached Figure 2 As shown, the present invention provides an email fusion governance device with AI recognition technology, including an investigation data receiving module, a multimodal intelligent parsing engine, an email account relationship modeling module, a customs data collision engine, and an Excel result output module; The investigation data receiving module is used to receive email data obtained from investigations by Internet companies, and to perform format standardization and integrity verification on the investigation data. The investigation data receiving module includes a data receiving interface, a format converter, and a data validator. The multimodal intelligent parsing engine is used to uniformly parse and structure data in various formats such as Excel, Word, PDF and images in email attachments. The multimodal intelligent parsing engine integrates a large language model processing unit and a computer vision processing unit, and realizes collaborative analysis of cross-modal data through a feature fusion layer. The email account relationship modeling module is used to build a dynamic relationship network based on email sending and receiving accounts. The email account relationship modeling module includes an account recognizer, a relationship extractor, and a graph database. The customs data collision engine is used to intelligently compare and analyze structured email data with customs declaration data. The customs data collision engine includes a customs data interface, a similarity calculation module, and an anomaly detection module. The Excel output module is used to generate standardized Excel-formatted analysis reports. The Excel output module includes a data processor, an Excel template engine, and a file exporter.

[0013] A method for using an email fusion governance device with AI recognition technology includes the aforementioned email fusion governance device with AI recognition technology and the following steps: S1, Data reception and preprocessing stage; A. The investigation data receiving module receives investigation email data from Internet companies through a standardized interface. The data formats include, but are not limited to, original email files, database exported files, and structured data packets. B. The format converter performs unified conversion of heterogeneous data formats provided by different enterprises, extracting and storing email header information, body content and attachment data into a standardized data structure. C. The data validator performs integrity and validity checks on the verification data, checks the legality of key fields such as email timestamps, sender addresses, and recipient addresses, and marks abnormal or missing data records. D. The preprocessing module cleans the email body, removing HTML tags, special characters, and redundant spaces. It also scans attachments for viruses and identifies their formats to ensure the security and accuracy of subsequent processing. S2, Multimodal Attachment Intelligent Analysis Stage; A. The attachment format recognizer automatically detects the file type of email attachments, including Excel spreadsheets, Word documents, PDF files, and JPG / PNG images, and selects the appropriate parsing strategy based on the file type; B. The large language model processing unit is responsible for parsing text attachments. For Excel files, it extracts cell data and formula information through table structure analysis. For Word documents, it extracts paragraph, table and chart content through document object model. For PDF files, it obtains editable text content through text extraction algorithm. C. For image attachments, image preprocessing is first performed, including denoising, enhancement, and geometric correction. Then, a multimodal model is used to identify the text content in the image. Finally, a deep learning model is used to understand the layout and semantic structure of the image. D. The feature fusion layer fuses text feature vectors and image feature vectors using a multi-layer perceptron to generate a unified multimodal feature representation, and dynamically adjusts the weights of different modal features through an attention mechanism. E. The business entity identification module extracts key business fields such as bill of lading number, packing list number, invoice number, goods name, quantity, amount, consignor and consignee from the fused features, and establishes semantic relationships between the fields; S3, Email Account Relationship Network Building Phase; A. The account identifier extracts the sender's and recipient's email addresses from the email header information and performs standardization processing on the email addresses, including case consistency, domain name normalization, and alias resolution. B. The account entity resolution module identifies the enterprise or organization to which the email address belongs by the domain name part and infers personal identity information by the username part, thus establishing a mapping relationship between the email account and the real entity. C. The relationship extractor analyzes email correspondence records, calculates the frequency of interactions between accounts, time distribution, and content relevance, and identifies direct communication relationships and indirect association relationships. D. The relationship strength calculation module quantifies the closeness of the relationship between accounts based on indicators such as the number of emails and the time span, and sets a threshold to filter weak relationships. E. The graph database uses a node-edge model to store the account relationship network. Nodes represent email accounts and their attribute information, and edges represent the communication relationships between accounts and their strength weights. It supports complex graph query and path analysis operations. S4, Customs declaration data collision analysis stage; A. Customs data is obtained by uploading customs declaration Excel files, including key fields such as customs declaration number, import and export company information, cargo list, declared value, and customs clearance time. B. The data preprocessing module cleans and standardizes customs declaration data, unifies the way goods are described, standardizes company names and address information, and processes unit conversions for numerical fields. C. The field mapping module establishes the field correspondence between email data and customs data, matches the goods information in the email attachment with the commodity information in the customs declaration, and associates the enterprise information in the email with the consignee and consignee information in the customs declaration. D. The similarity calculation module uses multiple algorithms for data comparison, including the string edit distance algorithm to compare text similarity, the semantic vector algorithm to calculate semantic similarity, and the numerical range algorithm to determine the degree of matching between amount and quantity. E. The anomaly detection module identifies inconsistencies between email data and customs data, including discrepancies in goods names, quantities, amounts, and time conflicts, and calculates a risk score for the degree of anomaly. F. The correlation analysis module uses graph algorithms to analyze the cross-correlation between email account relationship networks and customs declaration data, discover potential business chains and fund flows, and identify suspicious trade behavior patterns. S5, Excel format result output stage; A. The data organizer organizes all analysis results according to a predefined data structure, including an account relationship network table and a data collision table. B. The Excel template engine selects the appropriate Excel template according to different analysis needs. The template includes preset headers, format styles and formula calculations to ensure the standardization and readability of the output results. C. The data filling module fills the organized analysis data into the corresponding cells of the Excel worksheet according to the template structure, and automatically adjusts the column width, row height and font format; D. The chart generation module creates data visualization charts in Excel, including account relationship network diagrams and data collision statistics charts, enhancing the intuitiveness of analysis results.

[0014] The specific implementation method is as follows: When using this invention, the investigation data receiving module receives investigation email data provided by Internet companies through a standardized interface. The data formats include, but are not limited to, original email files, database exported files, and structured data packets. The format converter performs unified conversion on the heterogeneous data formats provided by different companies, extracting and storing the email header information, body content, and attachment data into a standardized data structure. The data validator performs integrity and validity checks on the investigation data, checking the legality of key fields such as email timestamps, sender addresses, and recipient addresses, and marking abnormal or missing data records. The preprocessing module performs text cleaning on the email body, removing HTML tags, special characters, and redundant spaces. Simultaneously, attachments are scanned for viruses and their formats are identified to ensure the security and accuracy of subsequent processing. The attachment format identifyer automatically detects the file type of email attachments, including Excel spreadsheets, Word documents, PDF files, and JPG / PNG images, and selects the appropriate parsing strategy based on the file type. The large language model processing unit is responsible for parsing text attachments. For Excel files, it extracts cell data and formula information through table structure analysis; for Word documents, it extracts paragraph, table, and chart content through the document object model; and for PDF files, it obtains editable text content through text extraction algorithms. For image attachments, image preprocessing is first performed, including noise reduction, enhancement, and geometric correction, and then a multimodal model is used. The system identifies text content in images and then uses a deep learning model to understand the image's layout and semantic structure. The feature fusion layer fuses text and image feature vectors using a multi-layer perceptron to generate a unified multimodal feature representation. An attention mechanism dynamically adjusts the weights of different modal features. The business entity recognition module extracts key business fields such as bill of lading number, packing list number, invoice number, goods name, quantity, amount, and consignor / consignee from the fused features and establishes semantic relationships between these fields. The account recognizer extracts the sender's and recipient's email addresses from the email header and standardizes these addresses, including case sensitivity, domain name normalization, and alias resolution. The account entity resolution module then uses the email address... The domain name identifies the company or organization to which the account belongs, while the username infers personal identity information, establishing a mapping between email accounts and real entities. A relationship extractor analyzes email correspondence, calculating the frequency, time distribution, and content relevance of interactions between accounts to identify direct and indirect communication relationships. A relationship strength calculation module quantifies the closeness of relationships between accounts based on indicators such as email quantity and time span, and sets thresholds to filter weak relationships. The graph database uses a node-edge model to store the account relationship network, where nodes represent email accounts and their attribute information, and edges represent communication relationships and their strength weights between accounts. It supports complex graph queries and path analysis operations. Customs data is obtained by uploading customs declaration Excel files.The system includes key fields such as customs declaration number, import / export company information, cargo list, declared value, and customs clearance time. The data preprocessing module cleans and standardizes customs declaration data, unifying the way goods are described, standardizing company names and addresses, and handling unit conversions for numerical fields. The field mapping module establishes the correspondence between email data and customs data, matching cargo information in email attachments with commodity information in customs declarations, and associating company information in emails with consignor / consignee information in customs declarations. The similarity calculation module uses multiple algorithms for data comparison, including string edit distance algorithms to compare text similarity, semantic vector algorithms to calculate semantic similarity, and numerical range algorithms to determine the degree of matching between amount and quantity. The anomaly detection module identifies inconsistencies between email data and customs data, including discrepancies in cargo names, quantities, amounts, and time conflicts, and calculates a risk score for the degree of anomaly. The association analysis module uses graph algorithms to analyze the cross-correlation between email account relationship networks and customs declaration data, discovering potential business chains and fund flows, and identifying suspicious trade behavior patterns. The data processor organizes all analysis results according to a predefined data structure. This invention includes an account relationship network table and a data collision table. An Excel template engine selects appropriate Excel templates based on different analysis needs. The templates contain preset headers, format styles, and formula calculations to ensure standardized and readable output results. A data filling module fills the organized analysis data into the corresponding cells of the Excel worksheet according to the template structure, automatically adjusting column width, row height, and font format. A chart generation module creates data visualization charts in Excel, including account relationship network diagrams and data collision statistical charts, enhancing the intuitiveness of the analysis results. This invention constructs an email governance architecture that integrates AI multimodal recognition and relationship mining to achieve intelligent structured conversion of multi-format attachments. Utilizing a large language model and a multimodal model working together, it automatically extracts and maps unstructured business data scattered in attachments (such as bills of lading numbers, cargo details, and amounts) into standardized database fields, establishing a dynamic email entity relationship graph. Based on email communication characteristics, it generates a relational network to strengthen business traceability capabilities. A cross-source data intelligent collision engine is designed to compare structured email data with third-party systems such as customs declarations in real time, automatically outputting risk clues or business evidence loops, significantly reducing the time cost of manual verification. ,

[0015] Working principle of this invention: Refer to the instruction manual appendix Figure 1 and attached Figure 2When using this invention, it constructs an email governance architecture that integrates AI multimodal recognition and relationship mining to achieve intelligent structured conversion of multi-format attachments. By utilizing large language models and multimodal models working together, it automatically extracts and maps unstructured business data scattered in attachments (such as bill of lading numbers, cargo details, amounts, etc.) into standardized database fields, establishes a dynamic email entity relationship graph, and generates an association network based on email communication characteristics to enhance business traceability capabilities. It also designs a cross-source data intelligent collision engine to compare structured email data with third-party systems such as customs declaration in real time, automatically outputting risk clues or business evidence loops, and significantly reducing the time cost of manual verification.

[0016] Finally, the following points should be noted: First, in the description of this application, it should be noted that, unless otherwise specified and limited, the terms "installation", "connection", and "linkage" should be interpreted broadly, and can be mechanical or electrical connections, or internal connections between two components, or direct connections. "Up", "down", "left", "right", etc. are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may change. Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An email fusion management device with AI recognition technology, characterized in that: It includes a data receiving module, a multimodal intelligent parsing engine, an email account relationship modeling module, a customs data collision engine, and an Excel result output module; The investigation data receiving module is used to receive email data obtained from investigations by Internet companies, and to perform format standardization and integrity verification on the investigation data. The investigation data receiving module includes a data receiving interface, a format converter, and a data validator. The multimodal intelligent parsing engine is used to uniformly parse and structure data in various formats such as Excel, Word, PDF and images in email attachments. The multimodal intelligent parsing engine integrates a large language model processing unit and a computer vision processing unit, and realizes collaborative analysis of cross-modal data through a feature fusion layer. The email account relationship modeling module is used to build a dynamic relationship network based on email sending and receiving accounts. The email account relationship modeling module includes an account recognizer, a relationship extractor, and a graph database. The customs data collision engine is used to intelligently compare and analyze structured email data with customs declaration data. The customs data collision engine includes a customs data interface, a similarity calculation module, and an anomaly detection module. The Excel output module is used to generate standardized Excel-formatted analysis reports. The Excel output module includes a data processor, an Excel template engine, and a file exporter.

2. A method for using an email fusion management device with AI recognition technology, characterized in that: Includes the email fusion and governance device with AI recognition technology as described in claim 1, and the following steps: S1, Data reception and preprocessing stage; A. The investigation data receiving module receives investigation email data from Internet companies through a standardized interface. The data formats include, but are not limited to, original email files, database exported files, and structured data packets. B. The format converter performs unified conversion of heterogeneous data formats provided by different enterprises, extracting and storing email header information, body content and attachment data into a standardized data structure. C. The data validator performs integrity and validity checks on the verification data, checks the legality of key fields such as email timestamps, sender addresses, and recipient addresses, and marks abnormal or missing data records. D. The preprocessing module cleans the email body, removing HTML tags, special characters, and redundant spaces. It also scans attachments for viruses and identifies their formats to ensure the security and accuracy of subsequent processing. S2, Multimodal Attachment Intelligent Analysis Stage; A. The attachment format recognizer automatically detects the file type of email attachments, including Excel spreadsheets, Word documents, PDF files, and JPG / PNG images, and selects the appropriate parsing strategy based on the file type; B. The large language model processing unit is responsible for parsing text attachments. For Excel files, it extracts cell data and formula information through table structure analysis. For Word documents, it extracts paragraph, table and chart content through document object model. For PDF files, it obtains editable text content through text extraction algorithm. C. For image attachments, image preprocessing is first performed, including denoising, enhancement, and geometric correction. Then, a multimodal model is used to identify the text content in the image. Finally, a deep learning model is used to understand the layout and semantic structure of the image. D. The feature fusion layer fuses text feature vectors and image feature vectors using a multi-layer perceptron to generate a unified multimodal feature representation, and dynamically adjusts the weights of different modal features through an attention mechanism. E. The business entity identification module extracts key business fields such as bill of lading number, packing list number, invoice number, goods name, quantity, amount, consignor and consignee from the fused features, and establishes semantic relationships between the fields; S3, Email Account Relationship Network Building Phase; A. The account identifier extracts the sender's and recipient's email addresses from the email header information and standardizes the email addresses, including case consistency, domain name normalization, and alias resolution. B. The account entity resolution module identifies the enterprise or organization to which the email address belongs by the domain name part and infers personal identity information by the username part, thus establishing a mapping relationship between the email account and the real entity. C. The relationship extractor analyzes email correspondence records, calculates the frequency of interactions between accounts, time distribution, and content relevance, and identifies direct communication relationships and indirect association relationships. D. The relationship strength calculation module quantifies the closeness of the relationship between accounts based on indicators such as the number of emails and the time span, and sets a threshold to filter weak relationships. E. The graph database uses a node-edge model to store the account relationship network. Nodes represent email accounts and their attribute information, and edges represent the communication relationships between accounts and their strength weights. It supports complex graph query and path analysis operations. S4, Customs declaration data collision analysis stage; A. Customs data is obtained by uploading customs declaration Excel files, including key fields such as customs declaration number, import and export company information, cargo list, declared value, and customs clearance time. B. The data preprocessing module cleans and standardizes customs declaration data, unifies the way goods are described, standardizes company names and address information, and processes unit conversions for numerical fields. C. The field mapping module establishes the field correspondence between email data and customs data, matches the goods information in the email attachment with the commodity information in the customs declaration, and associates the enterprise information in the email with the consignee and consignee information in the customs declaration. D. The similarity calculation module uses multiple algorithms for data comparison, including the string edit distance algorithm to compare text similarity, the semantic vector algorithm to calculate semantic similarity, and the numerical range algorithm to determine the degree of matching between amount and quantity. E. The anomaly detection module identifies inconsistencies between email data and customs data, including discrepancies in goods names, quantities, amounts, and time conflicts, and calculates a risk score for the degree of anomaly. F. The correlation analysis module uses graph algorithms to analyze the cross-correlation between email account relationship networks and customs declaration data, discover potential business chains and fund flows, and identify suspicious trade behavior patterns. S5, Excel format result output stage; A. The data organizer organizes all analysis results according to a predefined data structure, including an account relationship network table and a data collision table. B. The Excel template engine selects the appropriate Excel template according to different analysis needs. The template includes preset headers, format styles and formula calculations to ensure the standardization and readability of the output results. C. The data filling module fills the organized analysis data into the corresponding cells of the Excel worksheet according to the template structure, and automatically adjusts the column width, row height and font format; D. The chart generation module creates data visualization charts in Excel, including account relationship network diagrams and data collision statistics charts, enhancing the intuitiveness of analysis results.

Images