Aircraft material traceability history generation method and system based on large language model

Abstract

The application discloses a kind of based on big language model's aviation material traceability history generation method and system, the method includes: obtaining original certificate package file and splitting into multiple single-page electronic document files according to page, generate metadata label containing page number and file identification;Calculate the visual-text joint entropy value of each single-page electronic document file, generate special prompt word according to entropy value from multi-level prompt word template library Select corresponding level template;Single-page electronic document file is input into big language model with special prompt word, obtain structured JSON string;After checking structured JSON string, store with metadata label as index, and according to the first page determination result, single-page file is merged into merged certificate file;With certificate end date as event anchor point, arrange each merged certificate file in time sequence, generate the traceability history time axis of aviation material.The application can accurately identify certificate starting page, automatically generate structured aviation material traceability history.

Description

Technical Field

[0001] This invention relates to the field of digital processing and traceability technology for aviation equipment certificates, specifically to a method and system for generating aviation equipment traceability records based on a large language model. Background Technology

[0002] In the field of aviation equipment management, the airworthiness of aviation materials relies on the completeness and traceability of various certificates and documents, including certificates of conformity, maintenance reports, disassembly tags, accident-free declarations, invoices, and waybills. Existing technologies mainly offer the following solutions:

[0003] (1) Manual entry or semi-automatic ledger method: Business personnel fill in parameters such as item number, serial number, event type and date in the ERP system field by field based on paper or scanned certificates. This method relies on manual operation, is inefficient, and each certificate file is stored in isolation, which cannot be automatically spliced ​​into a complete traceability timeline.

[0004] (2) Traditional OCR, NER and layout recognition methods: Keyword matching and region coordinate extraction are performed on fixed templates, and the recognition results are mapped to a limited number of fields. This method is suitable for documents with stable layout and single language, but it has poor scalability for documents without fixed format, multilingual variants and non-scanned certificates, and the maintenance cost is high.

[0005] (3) General large model document understanding method: Upload the entire PDF or image to the general large language model and extract key information through natural language commands. Although the model has semantic understanding capabilities, it lacks special prompts for the field of aviation material certificates, such as homepage determination, event classification, and start / end date constraints. This results in inconsistencies between the output fields and the business database table structure, and it is easy to misidentify non-homepage and auxiliary pages as independent events. It is also unable to automatically generate a structured traceability timeline based on the certificate end date. Summary of the Invention

[0006] In view of the above problems, the present invention provides a method and system for generating aircraft material traceability history based on a large language model, which aims to solve the technical problems in the prior art that the large language model cannot accurately identify the certificate start page and cannot automatically generate aircraft material traceability history according to the event timeline when extracting certificate information due to the diverse formats, multiple pages and lack of structured constraints of aircraft material certificate files.

[0007] To achieve the above objectives, in a first aspect, the present invention provides a method for generating aircraft material traceability records based on a large language model, the method comprising:

[0008] S1: Obtain the original certificate package file, which contains one or more single certificate files;

[0009] S2: Split the original certificate package file into multiple single-page electronic document files by page, and generate a metadata tag for each single-page electronic document file. The metadata tag includes at least the page number of the single-page electronic document file in the original certificate package file and the identifier of the original certificate package file to which it belongs.

[0010] S3: Calculate the visual-text joint entropy value of each of the single-page electronic document files, select the corresponding level template from the preset multi-level prompt word template library according to the visual-text joint entropy value, and generate a special prompt word; input the single-page electronic document file and the special prompt word together into the large language model, and obtain the structured JSON string output by the large language model;

[0011] The structured JSON string includes at least the homepage determination result, file type, certificate start date, and certificate end date. The homepage determination result is used to identify whether the single-page electronic document file is the starting page of a single certificate file. The special prompt word contains field extraction instructions. The large language model extracts the certificate number and certificate title from the page content of the single-page electronic document file according to the field extraction instructions. When the complete certificate number and certificate title are extracted, the homepage determination result is set to "yes"; otherwise, it is set to "no".

[0012] S4: Validate the structured JSON string, and establish a one-to-one correspondence between the validated structured JSON string and the corresponding single-page electronic document file using the metadata tag as an index, and store it in the database;

[0013] Based on the homepage determination result, the single-page electronic document file with a homepage determination result of "yes" is used as the starting page of a single certificate file, the single-page electronic document file with a homepage determination result of "no" is sequentially concatenated to its respective single certificate file according to the page number order, and the single-page electronic document file with a homepage determination result of "yes" and the single-page electronic document file with a homepage determination result of "no" concatenated thereafter are merged into a single merged certificate file;

[0014] S5: Using the certificate expiration date in the structured JSON string as the event anchor, arrange the merged certificate files in chronological order to generate a timeline for the traceability of aircraft materials.

[0015] Furthermore, in step S3, calculating the visual-text joint entropy value for each of the single-page electronic document files includes:

[0016] The single-page electronic document file is subjected to OCR recognition, and the ratio of Chinese characters to English characters in the OCR recognition result is calculated as the Chinese-English mixing degree.

[0017] The OCR recognition results are matched with a dictionary of professional terms in the field of aviation materials, and the proportion of successfully matched characters to the total number of characters is counted as the professional term density.

[0018] The layout of the single-page electronic document is analyzed to identify the table area and the plain text area in the single-page electronic document, and the ratio of the area of ​​the table area to the area of ​​the plain text area is calculated as the degree of table and text mixing.

[0019] The visual-text joint entropy value is obtained by weighting and fusing the Chinese-English mixing degree, the density of professional terms, and the mixing degree of tables and text.

[0020] Furthermore, the grade template includes a low-entropy template, a medium-entropy template, and a high-entropy template. The low-entropy template contains field extraction instructions, the medium-entropy template contains field extraction instructions and field descriptions, and the high-entropy template contains layout description instructions, field extraction instructions, and conflict detection instructions.

[0021] Based on the visual-text joint entropy value, a corresponding level template is selected from a preset multi-level prompt word template library to generate dedicated prompt words, including:

[0022] When the visual-text joint entropy value is higher than the first preset threshold, the layout description instruction, field extraction instruction and conflict detection instruction are extracted from the high-entropy template, and the layout description instruction, field extraction instruction and conflict detection instruction are spliced ​​and combined to generate the special prompt word;

[0023] When the visual-text joint entropy value is lower than or equal to the first preset threshold and higher than the second preset threshold, the field extraction instruction and field description are extracted from the medium entropy template to generate the special prompt word;

[0024] When the visual-text joint entropy value is lower than or equal to the second preset threshold, the field extraction instruction is extracted from the low-entropy template to generate the special prompt word.

[0025] Furthermore, the field extraction instruction also includes sub-instructions executed in the following order:

[0026] The homepage verification sub-instruction is used to instruct the large language model to further verify whether the current page is the certificate homepage based on the page number information after the homepage determination result is set to "yes". If the verification result is no, the homepage determination result is modified to "no".

[0027] The file type classification sub-instruction is used to instruct the large language model to match file types among a variety of preset certificate types;

[0028] The date extraction sub-instruction is used to instruct the large language model to extract the certificate start date and certificate end date;

[0029] The extended field extraction sub-instruction is used to instruct the large language model to extract the corresponding extended fields according to the file type, wherein the extended fields are additional information fields related to the certificate type.

[0030] Furthermore, in step S4, the verification rules include at least one of the following rules:

[0031] When the homepage determination result is "yes" and a valid file type, certificate start date, and certificate end date are identified, the identified valid file type, certificate start date, and certificate end date are used as the current identification result; otherwise, the main fields of file type, certificate start date, and certificate end date are set to empty.

[0032] The identified file types are matched with a variety of preset certificate type whitelists. If the identified file type is not in the whitelist or is an empty file type, the identified file type is deemed invalid.

[0033] Verify that the certificate end date is not empty, and if a certificate start date also exists, the certificate start date is not later than the end date;

[0034] The validity of extended fields is verified based on the file type, and only extended fields allowed for that file type are retained;

[0035] Perform semantic and format validation on extended fields for a specified file type.

[0036] Furthermore, the method includes:

[0037] A file storage path is established with the equipment number as the root directory. The file storage path includes at least a first sub-path and a second sub-path. The first sub-path is used to store the split single-page electronic document files and their thumbnails, and the second sub-path is used to store the merged certificate file.

[0038] Furthermore, the method includes:

[0039] Based on the file type of the merged certificate file, the airworthiness of the aircraft material is determined, and the airworthiness determination includes:

[0040] Obtain the aircraft material type identifier, which is pre-stored in the aircraft material database;

[0041] Based on the aircraft material type identifier, the corresponding certificate file type set is queried from the airworthiness rule base. The certificate file type set contains all certificate file types required for the airworthiness of this type of aircraft material.

[0042] The actual file types identified in the merged certificate file are matched item by item with the certificate file type set. When every file type in the certificate file type set exists in the actual identified file types, the aircraft material is determined to meet the airworthiness requirements.

[0043] Furthermore, in step S3, the structured JSON string also includes equipment identification information, which includes at least one of equipment serial number, equipment model, and equipment batch number;

[0044] Step S4 is followed by:

[0045] Based on the equipment identification information, a common entity is extracted from the structured JSON strings of multiple equipment. The common entity includes at least one of the following: manufacturer name, repair shop name, issuing authority name, and production batch number.

[0046] Using equipment as nodes and the common entities as associated edges, a cross-equipment association graph is constructed, wherein the association graph is stored using an adjacency matrix or an adjacency list data structure.

[0047] Based on the aforementioned association graph, a breadth-first search algorithm or a depth-first search algorithm is used to detect abnormal equipment. The abnormal equipment includes isolated equipment that has no association edges with other equipment, abnormally associated equipment that has association edges with more than a preset number of other equipment, and equipment with conflicting associations where the certificate of conformity and the repair report are from different manufacturers.

[0048] Different levels of early warning information are generated based on the type of anomaly, and the early warning information includes at least one of a first color indication information, a second color indication information, and a third color indication information;

[0049] The association map and the early warning information are added to the structured JSON string, wherein the association map is stored in the form of a JSON array, and each array element contains equipment identification information and a list of associated edges.

[0050] Furthermore, in step S5, the rules for generating the warning information include:

[0051] When an isolated device is detected, a first color indication is generated. The first color indication is used to indicate that the device is of unknown origin and automatically prevents the device's traceability information from being displayed to the outside world.

[0052] When an abnormally associated device is detected, a second color indicator is generated. The second color indicator is used to indicate that the device needs further verification and the device is marked as pending verification on the traceability display page.

[0053] When a conflicting device is detected, a third color indication is generated. This third color indication is used to indicate that the device requires manual review, and the device's traceability information is submitted to the review queue.

[0054] In a second aspect, the present invention provides an aircraft material traceability history generation system based on a large language model, the system comprising:

[0055] The certificate acquisition module is used to acquire the original certificate package file, which contains one or more single certificate files;

[0056] The document splitting module, connected to the certificate acquisition module, is used to split the original certificate package file into multiple single-page electronic document files by page, and generate metadata tags for each single-page electronic document file. The metadata tags include at least the page number of the single-page electronic document file in the original certificate package file and the identifier of the original certificate package file to which it belongs.

[0057] The intelligent extraction module, connected to the document splitting module, is used to calculate the visual-text joint entropy value of each single-page electronic document file, select the corresponding level template from the preset multi-level prompt word template library according to the visual-text joint entropy value, and generate special prompt words; input the single-page electronic document file and the special prompt words together into the large language model to obtain the structured JSON string output by the large language model;

[0058] The structured JSON string includes at least the homepage determination result, file type, certificate start date, and certificate end date. The homepage determination result is used to identify whether the single-page electronic document file is the starting page of a single certificate file. The special prompt word contains field extraction instructions. The large language model extracts the certificate number and certificate title from the page content of the single-page electronic document file according to the field extraction instructions. When the complete certificate number and certificate title are extracted, the homepage determination result is set to "yes"; otherwise, it is set to "no".

[0059] The verification storage module, connected to the intelligent extraction module, is used to verify the structured JSON string. The verified structured JSON string, indexed by the metadata tag, is linked to the corresponding single-page electronic document file and stored in the database. Based on the homepage determination result, the single-page electronic document file with a "yes" homepage determination result is used as the starting page of a single certificate file. The single-page electronic document files with a "no" homepage determination result are sequentially appended to their respective single certificate files according to page number order. The single-page electronic document files with a "yes" homepage determination result and the appended single-page electronic document files with a "no" homepage determination result are merged into a single merged certificate file.

[0060] The timeline generation module, connected to the verification and storage module, is used to generate a timeline of traceability history for aircraft materials by arranging the merged certificate files in chronological order, using the certificate expiration date in the structured JSON string as the event anchor.

[0061] Unlike existing technologies, the above technical solution involves a method and system for generating aircraft material traceability history based on a large language model, belonging to the field of digital processing and traceability technology for aircraft equipment certificates. The method includes: acquiring the original certificate package file and splitting it into multiple single-page electronic document files, generating metadata tags containing page numbers and file identifiers; calculating the visual-text joint entropy value of each single-page electronic document file, and selecting a corresponding level template from a multi-level prompt word template library based on the entropy value to generate a dedicated prompt word; inputting the single-page electronic document file and the dedicated prompt word into a large language model to obtain a structured JSON string containing the first page judgment result, file type, certificate start date, and certificate end date; validating the structured JSON string and storing it with the metadata tags as an index, and merging the single-page files into a merged certificate file based on the first page judgment result; finally, using the certificate end date as the event anchor point, arranging the merged certificate files in chronological order to generate a traceability history timeline for aircraft materials. This invention can accurately identify the certificate start page, automatically generate a structured aircraft material traceability history, and improve the efficiency and accuracy of certificate processing.

[0062] The above description of the invention is merely an overview of the technical solution of the present invention. In order to enable those skilled in the art to better understand the technical solution of the present invention and to implement it based on the description and drawings, and to make the above-mentioned objectives and other objectives, features and advantages of the present invention easier to understand, the following description is provided in conjunction with the specific embodiments and drawings of the present invention. Attached Figure Description

[0063] The accompanying drawings are only used to illustrate the principles, implementation methods, applications, features, and effects of specific embodiments of the present invention and other related contents, and should not be considered as limitations on the present invention.

[0064] In the accompanying drawings of the instruction manual:

[0065] Figure 1 This is a flowchart of an exemplary embodiment of the method for generating aircraft material traceability history based on a large language model according to the present invention;

[0066] Figure 2 This is a flowchart illustrating the human-machine collaborative automated processing procedure for the China Aviation Materials Certificate, as described in an exemplary embodiment of the present invention.

[0067] Figure 3This is a schematic diagram of the module of the aircraft material traceability and history generation system based on a large language model, as described in an exemplary embodiment of the present invention.

[0068] The reference numerals used in the above figures are explained as follows:

[0069] 10. Aircraft material traceability and history generation system based on large language model;

[0070] 101. Certificate Acquisition Module;

[0071] 102. Document splitting module;

[0072] 103. Intelligent extraction module;

[0073] 104. Verification storage module;

[0074] 105. Timeline generation module. Detailed Implementation

[0075] To explain in detail the possible application scenarios, technical principles, specific feasible solutions, and the objectives and effects that can be achieved by this invention, the following detailed description is provided in conjunction with the listed specific embodiments and accompanying drawings. The embodiments described herein are only used to more clearly illustrate the technical solutions of this invention, and are therefore only examples, and should not be used to limit the scope of protection of this invention.

[0076] In the first aspect, such as Figure 1 As shown, this application provides a method for generating aircraft material traceability history based on a large language model, the method comprising:

[0077] S1: Obtain the original certificate package file, which contains one or more single certificate files;

[0078] S2: Split the original certificate package file into multiple single-page electronic document files by page, and generate a metadata tag for each single-page electronic document file. The metadata tag includes at least the page number of the single-page electronic document file in the original certificate package file and the identifier of the original certificate package file to which it belongs.

[0079] S3: Calculate the visual-text joint entropy value of each of the single-page electronic document files, select the corresponding level template from the preset multi-level prompt word template library according to the visual-text joint entropy value, and generate a special prompt word; input the single-page electronic document file and the special prompt word together into the large language model, and obtain the structured JSON string output by the large language model;

[0080] The structured JSON string includes at least the homepage determination result, file type, certificate start date, and certificate end date. The homepage determination result is used to identify whether the single-page electronic document file is the starting page of a single certificate file. The special prompt word contains field extraction instructions. The large language model extracts the certificate number and certificate title from the page content of the single-page electronic document file according to the field extraction instructions. When the complete certificate number and certificate title are extracted, the homepage determination result is set to "yes"; otherwise, it is set to "no".

[0081] S4: Validate the structured JSON string, and establish a one-to-one correspondence between the validated structured JSON string and the corresponding single-page electronic document file using the metadata tag as an index, and store it in the database;

[0082] Based on the homepage determination result, the single-page electronic document file with a homepage determination result of "yes" is used as the starting page of a single certificate file, the single-page electronic document file with a homepage determination result of "no" is sequentially concatenated to its respective single certificate file according to the page number order, and the single-page electronic document file with a homepage determination result of "yes" and the single-page electronic document file with a homepage determination result of "no" concatenated thereafter are merged into a single merged certificate file;

[0083] S5: Using the certificate expiration date in the structured JSON string as the event anchor, arrange the merged certificate files in chronological order to generate a timeline for the traceability of aircraft materials.

[0084] In this embodiment, the original certificate package file refers to the PDF format file package entered by the business end. The file can contain multiple independent aviation material certificates, including various documents such as certificates of conformity, maintenance reports, and disassembly tags. It is the original data input source of this invention and supports two acquisition channels: manual upload and automatic synchronization within the enterprise's internal framework.

[0085] A single-page electronic document file refers to an independent PDF generated after the original certificate package file is split into pages. It is the smallest processing unit for large language model recognition. After pagination, a 200px PNG thumbnail is generated simultaneously for business personnel to manually review the page content.

[0086] Metadata tags are index identifiers bound to each single-page electronic document file. They store two core pieces of information: the page number of the page within the original certificate package file and the unique identifier of the original certificate package file. Metadata allows for permanent tracing of the original file to which a single-page electronic document file belongs, avoiding page ownership confusion when multiple files are mixed.

[0087] The visual-text joint entropy value is a comprehensive indicator that quantifies the complexity of a single-page electronic document file. It integrates three dimensions: the proportion of mixed language text on the page, the proportion of aviation material professional terms, and the proportion of tables and text on the page. The higher the entropy value, the more complex the page layout and the greater the difficulty of text recognition.

[0088] The multi-level prompt word template library refers to a pre-configured prompt word resource library specifically for aviation materials. It is divided into three fixed templates: low, medium, and high entropy. Different templates are equipped with differentiated recognition constraint instructions to adapt to certificate pages of varying complexity. This constrains the output logic of the large model from the source and suppresses recognition illusions.

[0089] The dedicated prompt words refer to the SYSTEM prompt words generated by matching the corresponding level template based on the page's joint entropy value. It has built-in exclusive recognition rules for aviation material certificates, which restricts the model to output content only according to a fixed process of homepage judgment, file classification, date extraction, and extended field extraction, and prohibits the model from making subjective guesses and unfounded information.

[0090] Structured JSON strings refer to the standardized output carrier of large language models. They store all recognition results in key-value pair format and are required to include three core fields: homepage judgment mark, certificate type, and certificate start and end dates. They are also compatible with equipment serial number, model, batch number, and various certificate extended business fields, and can be directly connected to the backend business database field structure.

[0091] The homepage determination result refers to the boolean tag field in the JSON, which is used to distinguish whether the current page is the first page of an independent aviation material certificate. The determination logic is based on the extraction result of the two elements "merged certificate file number + certificate title". If both elements are complete, it is marked as the homepage; otherwise, it is determined as a certificate subsidiary page.

[0092] A merged certificate file refers to a merged certificate file that is automatically grouped based on the homepage mark. It is formed by splicing the homepage page with subsequent consecutive pages without a homepage mark in page number order. It is the smallest complete file unit for front-end aviation material traceability display.

[0093] The event anchor point refers to the certificate end date identified in each merged certificate, which serves as the time benchmark for events in the life cycle of aviation materials. All certificates are uniformly sorted by this date and linked together to form a complete history link.

[0094] The traceability timeline refers to a visual and structured display platform that arranges all business events corresponding to certificates throughout the entire life cycle of aviation materials in chronological order. Each time node is attached with the corresponding complete merged certificate and all business parameters identified and extracted, allowing the purchaser to intuitively view the complete circulation history of the equipment.

[0095] In step S1, the original certificate package file is a PDF file package entered by the business side. This file can contain multiple independent aircraft material certificates, including certificates of conformity, maintenance reports, and disassembly / removal tags, serving as the original data input source for this solution. Certificate acquisition supports two entry channels: one is manual upload of the original certificate package file by business personnel; the other is automatic synchronization of existing certificate files by the enterprise's internal MRO and ERP management frameworks. During file upload or synchronization, a unique aircraft material identifier is bound to the file. This identifier can be a part number, serial number, or batch number, thus establishing a binding relationship between the certificate file and the corresponding aircraft material entity. The original certificate package file can contain multiple unseparated independent aircraft material certificates, eliminating the need for manual pre-sorting by business personnel.

[0096] In step S2, the system calls a PDF rendering and parsing library to slice the original certificate package file page by page, generating an independent single-page PDF document for each page, and simultaneously rendering a thumbnail of uniform size. Each page is traversed, automatically extracting the page number from the original file and the globally unique ID of the original certificate package file, encapsulating these into metadata tags and embedding them within the attributes of the single-page document. These metadata tags serve as globally unique indexes, allowing subsequent identification, verification, and storage processes to rely on the tags to reverse-locate page ownership, resolving the problem of indistinguishable page ownership after multiple certificates are mixed and paginated.

[0097] In step S3, the system reads a single-page electronic document file and performs OCR full text recognition, while simultaneously carrying out three-dimensional quantitative calculations, as follows: calculating the proportion of Chinese and English text to obtain the Chinese-English mixing degree, matching the built-in aviation material professional terminology dictionary to calculate the professional term density, and analyzing the page pixel to distinguish the area of ​​table areas and pure text areas to obtain the table text mixing degree; the three indicators are assigned fixed weights and weighted summed to finally output a unique visual-text joint entropy value, the higher the value, the more complex the page layout and text structure.

[0098] The system reads a multi-level prompt word template library, compares the joint entropy value with the preset two-level threshold, automatically matches the corresponding level template, and concatenates all the recognition instructions in the template to generate a unique prompt word for the current page; the prompt word has built-in standardized field extraction instructions, which force the model to prioritize the extraction of the two core homepage judgment elements: certificate number and certificate title.

[0099] The single-page electronic document file image and exclusive prompt words are simultaneously sent to the privately deployed large language model. The large language model completes page parsing according to the prompt word constraints. After the text is streamed out, the system automatically concatenates and formats it into a standard structured JSON string. The JSON always carries core fields such as homepage judgment mark, certificate category, certificate start date, date, and certificate end date, thus completing the structured extraction of page information.

[0100] In step S4, the system first reads the JSON string output by the LLM, executes a full set of business validation rules, and filters out invalid, contradictory, and dirty identification data. The JSON that passes the validation is stored in the business database after being bound to the corresponding single-page electronic document with the single-page metadata tag as the associated index. The data that fails the validation is uniformly transferred to the manual review queue to wait for business personnel to correct it.

[0101] The system then iterates through all stored single-page electronic document files, using the homepage marker as the dividing boundary to automatically group them: the page marked "yes" serves as the starting page of a merged certificate file, and subsequent pages marked "no" are sequentially appended to the homepage in ascending order of page number. This process merges a group of consecutive homepages and subsequent pages to generate an independent merged certificate file, achieving fully automatic splitting and restoration of PDFs with multiple certificates, without the need for manual differentiation of certificate boundaries.

[0102] The system then uses the equipment number as the root directory to construct a two-level file storage path. The first sub-path stores the split single-page electronic document files and thumbnails, and the second sub-path stores the grouped and merged certificate files, thus achieving hierarchical and orderly persistence of files.

[0103] In step S5, the system reads the structured JSON string corresponding to all merged certificate files in the database, extracts the certificate end date in each certificate as the event anchor point, and sorts all merged certificates according to the anchor point time from earliest to latest. After sorting, the system integrates the file type, start and end dates, extended business fields, and merged certificate files of each certificate to generate a visual aviation material traceability timeline. Each time node on the timeline corresponds to an aviation material lifecycle business event, completely restoring the entire process record of equipment manufacturing, maintenance, disassembly, and resale.

[0104] The above solution relies on page visual-text joint entropy adaptive matching of hierarchical prompt words to constrain the recognition logic from the model input source, significantly reducing the illusion problem caused by unfounded speculation in large models. It eliminates the need to develop static recognition rules for various non-standard aviation material certificates, adapts to complex documents with multiple languages ​​and mixed tabular text, and significantly improves scalability. Through PDF single-page splitting and automatic first-page grouping mechanisms, it completely replaces the manual differentiation of start and end boundaries for mixed certificates, eliminating the need for business personnel to pre-organize certificate files and significantly reducing manual workload. Simultaneously, using the certificate end date as a unified time anchor point to connect all lifecycle events, it automatically generates a standardized, visual timeline, filling the industry gap where traditional aviation material systems cannot intuitively display the complete circulation history of equipment. The entire process of recognition, grouping, and timeline generation is automated, with only a small amount of abnormal data requiring manual review, effectively improving the efficiency of processing large batches of aviation material certificates.

[0105] In some embodiments, calculating the visual-text joint entropy value for each of the single-page electronic document files in step S3 includes:

[0106] The single-page electronic document file is subjected to OCR recognition, and the ratio of Chinese characters to English characters in the OCR recognition result is calculated as the Chinese-English mixing degree.

[0107] The OCR recognition results are matched with a dictionary of professional terms in the field of aviation materials, and the proportion of successfully matched characters to the total number of characters is counted as the professional term density.

[0108] The layout of the single-page electronic document is analyzed to identify the table area and the plain text area in the single-page electronic document, and the ratio of the area of ​​the table area to the area of ​​the plain text area is calculated as the degree of table and text mixing.

[0109] The visual-text joint entropy value is obtained by weighting and fusing the Chinese-English mixing degree, the density of professional terms, and the mixing degree of tables and text.

[0110] In some embodiments, OCR recognition refers to optical character recognition technology, which extracts all recognizable text content from a single-page electronic document image and outputs a complete text string. It is the basic data source for calculating the three entropy dimension indicators and can simultaneously extract Chinese and English characters.

[0111] The Chinese-English language mixing degree refers to the ratio of the total number of Chinese characters to the total number of English characters in a single page of electronic document, which is used to quantify the degree of language mixing on the page. Aircraft material certificates often have bilingual Chinese and English annotations. The higher the mixing degree, the more frequently the text language is switched, and the more difficult it is for the model to recognize.

[0112] The aviation materials professional terminology dictionary is a pre-built industry-specific thesaurus that includes exclusive professional terms for the entire process of civil aviation airworthiness, aviation materials production, maintenance, dismantling, and transportation. It includes standardized terms such as part number, airworthiness approval, IRO, and dismantling and maintenance, and is used to match and count the proportion of industry professional text on the page.

[0113] Terminology density refers to the percentage of characters in the total number of OCR characters on a page that can match a terminology dictionary. Higher density means more industry-specific information on the page, requiring the model to accurately identify specialized fields and strengthening the constraint of prompt words.

[0114] Page layout analysis refers to page layout analysis technology based on image pixel segmentation. It distinguishes between table blocks and plain text paragraph blocks within a page by using image outlines, lines, and color blocks, and accurately counts the pixel area of ​​the two types of areas.

[0115] The table and text mixing degree refers to the ratio of the total pixel area of ​​the table area to the total pixel area of ​​the plain text area on the page. The higher the ratio, the more tables there are on the page, the more scattered the fields are, the more irregular the layout is, and the higher the probability of model recognition error.

[0116] Weighted fusion refers to assigning fixed weight coefficients to three indicators: the degree of mixing of Chinese and English, the density of professional terms, and the degree of mixed text in tables. The values ​​of each indicator are multiplied by their corresponding weights and then summed to obtain a unique standardized visual-text joint entropy value, thereby achieving a quantitative and unified understanding of page complexity.

[0117] In this embodiment, the system completes the index calculation and entropy value fusion in four steps:

[0118] The first step is to perform full-frame OCR text recognition on the thumbnail of a single-page electronic document and output the complete text string. The program then iterates through the text to count the total number of Chinese characters and English characters respectively, and calculates the ratio between the two as the Chinese-English mixing degree. If the page contains only a single language, the value of this indicator is close to 0, which means that the language recognition difficulty is low.

[0119] The second step is to perform a fuzzy match between each character of the complete OCR text and the locally stored aviation material terminology dictionary, count the total number of characters contained in all successfully matched terms, divide the total number of characters on the page, and obtain the terminology density. This indicator is used to determine the proportion of industry information on the page. The higher the terminology density, the more strictly the model needs to follow the aviation material business rules to extract fields.

[0120] The third step involves using a layout image segmentation algorithm to analyze the single-page electronic document file image, identifying all table borders, cell blocks, and plain text paragraphs without tables within the page, and calculating the total pixel area of ​​each type of area. The total area of ​​the tables is then divided by the total area of ​​the plain text to obtain the degree of table and text mixing. Large tables on the page can disrupt the text order, and this indicator is used to quantify the degree of layout interference.

[0121] The fourth step involves the system retrieving preset weight configuration parameters, assigning weighting coefficients to the three dimensional indicators, multiplying the values ​​of the three indicators by their corresponding weights, summing the results, and finally outputting a unique visual-text joint entropy value. The entropy value can accurately quantify the recognition difficulty after the superposition of three complex factors: page language, professional content, and layout, providing a unified and quantifiable basis for subsequent adaptive matching prompt word templates.

[0122] The above solution quantifies page complexity through three dimensions: language, professional text, and layout. This replaces the previous approach of roughly judging page difficulty using a single dimension. The entropy calculation result accurately matches the business characteristics of aviation material certificates, which are multilingual, multi-table, and densely packed with professional terminology, ensuring the accuracy of subsequent prompt word template matching. The three indicators are weighted and integrated to form a unified quantitative value, avoiding problems such as insufficient or redundant prompt word configuration caused by single-dimensional judgment. Low-complexity pages are matched with lightweight prompt words to reduce model computing power consumption, while high-complexity pages are matched with enhanced prompt words to reduce the recognition error rate. This optimizes the overall pipeline processing efficiency while ensuring recognition accuracy, effectively balancing recognition accuracy and system computing resource consumption.

[0123] In some embodiments, the grade template includes a low-entropy template, a medium-entropy template, and a high-entropy template. The low-entropy template includes field extraction instructions, the medium-entropy template includes field extraction instructions and field descriptions, and the high-entropy template includes layout description instructions, field extraction instructions, and conflict detection instructions.

[0124] Based on the visual-text joint entropy value, a corresponding level template is selected from a preset multi-level prompt word template library to generate dedicated prompt words, including:

[0125] When the visual-text joint entropy value is higher than the first preset threshold, the layout description instruction, field extraction instruction and conflict detection instruction are extracted from the high-entropy template, and the layout description instruction, field extraction instruction and conflict detection instruction are spliced ​​and combined to generate the special prompt word;

[0126] When the visual-text joint entropy value is lower than or equal to the first preset threshold and higher than the second preset threshold, the field extraction instruction and field description are extracted from the medium entropy template to generate the special prompt word;

[0127] When the visual-text joint entropy value is lower than or equal to the second preset threshold, the field extraction instruction is extracted from the low-entropy template to generate the special prompt word.

[0128] In this embodiment, the low-entropy template refers to a basic lightweight template in the multi-level prompt word template library, which only carries standardized field extraction instructions and has no additional auxiliary constraint content; it is suitable for simple certificate pages with neat layout, plain text, single language, and few professional terms, with concise instructions and short model calculation time.

[0129] Medium-entropy templates refer to page adaptation templates with medium complexity. In addition to the basic field extraction instructions, they add textual explanations and descriptions for each business field. They are used for certificate pages with a slight mix of Chinese and English, a small number of tables, and a medium density of professional terminology. By using field descriptions to assist the model in accurately distinguishing similar business fields, they reduce field classification errors.

[0130] High-entropy templates refer to enhanced templates specifically designed for highly complex pages. They also incorporate three types of constraint instructions: layout description instructions, field extraction instructions, and conflict detection instructions. These instructions are adaptable to complex certificates with multiple languages, numerous tables, non-standard and chaotic layouts, and dense technical terms. They comprehensively constrain the model parsing logic and suppress model illusions and field conflicts.

[0131] Page layout description instructions are exclusive constraint instructions for high-entropy templates. They inform the large model of the page's table and text / image layout structure, guide the model to parse text in blocks, and avoid misaligned extraction of table row and column text and field confusion.

[0132] Field descriptions refer to the supplementary explanations accompanying the entropy template, which provide textual definitions for easily confused fields such as part number, serial number, repair date, and factory serial number, helping the model distinguish semantically similar business parameters.

[0133] Conflict detection instructions are exclusive verification instructions for high-entropy templates. They require the model to identify contradictory information on the page, such as logical date conflicts, inconsistent vendor names, and duplicate certificate numbers. After identification, the conflict fields are marked in the JSON to intercept dirty data in advance.

[0134] The first preset threshold and the second preset threshold refer to two fixed numerical boundaries that distinguish the complexity level of a page. The value of the first preset threshold is greater than the value of the second preset threshold. Pages with an entropy value higher than the first preset threshold are considered to be highly complex, those between the two thresholds are considered to be of medium complexity, and those below the second preset threshold are considered to be simple pages.

[0135] The working principle of this embodiment is as follows: The system pre-configures the instructions for three types of templates in a multi-level prompt word template library, and stores two fixed entropy threshold values. After completing the visual-text joint entropy calculation of the page, it logically matches the templates in three segments and generates dedicated prompt words:

[0136] The first type of judgment logic: If the joint entropy value of the page is greater than the first preset threshold, it is determined that the page layout and text complexity is high. The system retrieves all three types of instructions in the high-entropy template, and concatenates the layout description instructions, field extraction instructions, and conflict detection instructions in sequence to form a complete special prompt word and send it to the large model. The layout instructions guide the model to parse the table text by block, and the conflict detection instructions identify the internal information contradictions of the page in real time, reducing erroneous data from the model output stage.

[0137] The second type of judgment logic: If the second preset threshold is less than the page joint entropy value and the first preset threshold is not less than the page joint entropy value, the page is judged to be of medium complexity. The system retrieves the medium entropy template, splices the basic field extraction instructions and the corresponding field description text to generate prompt words; it relies on field interpretation to distinguish semantically similar aviation material business parameters, reducing the probability of incorrect field extraction, and at the same time, it does not need to load redundant layout and conflict detection instructions, saving model computing power.

[0138] The third type of judgment logic: If the joint entropy value of the page is not greater than the second preset threshold, it is determined that the page layout is simple and there is little text interference. Only the field extraction instructions in the low-entropy template are retrieved as dedicated prompt words. There are no additional auxiliary constraints, which maximizes the simplification of the model input text length and speeds up the single-page recognition processing speed.

[0139] The instructions for all three types of templates are based on industry-specific identification rules for aviation materials. By constraining the output logic of the large model from the bottom layer of prompt words, the core defect of the general large model, which produces illusions due to the lack of industry constraints, is solved.

[0140] The above solution uses a three-tiered differentiated prompt word template and relies on a joint entropy threshold to automatically match the corresponding instruction set. Simple pages use only basic extraction instructions, significantly shortening the model's input text length and improving pipeline concurrent processing speed; medium-complexity pages add field definitions to reduce confusion in recognizing similar business fields; high-complexity pages overlay layout parsing and conflict detection instructions, significantly reducing the error rate in recognizing multi-table and multilingual non-standard certificates. This tiered matching mechanism avoids wasting computational power by uniformly using high-entropy templates, while also preventing recognition distortion of complex pages due to insufficient constraints from simple templates. It balances system processing performance and recognition accuracy, suppressing the illusion of large models from the source and reducing the workload of subsequent manual review.

[0141] In some embodiments, the field extraction instruction further includes sub-instructions executed in the following order:

[0142] The homepage verification sub-instruction is used to instruct the large language model to further verify whether the current page is the certificate homepage based on the page number information after the homepage determination result is set to "yes". If the verification result is no, the homepage determination result is modified to "no".

[0143] The file type classification sub-instruction is used to instruct the large language model to match file types among a variety of preset certificate types;

[0144] The date extraction sub-instruction is used to instruct the large language model to extract the certificate start date and certificate end date;

[0145] The extended field extraction sub-instruction is used to instruct the large language model to extract the corresponding extended fields according to the file type, wherein the extended fields are additional information fields related to the certificate type.

[0146] Field extraction instructions: The core execution instructions within each level of prompt word template are broken down into four sub-instructions executed in a fixed order, forcing the large model to complete page parsing in a fixed sequence and preventing disruption of the recognition process.

[0147] In this embodiment, the homepage verification sub-instruction refers to the first-order sub-instruction with the highest priority. After the model initially determines that the page is the certificate homepage, it performs a second verification of the page number identifier. If the page number is not 1 or there is no homepage identifier, the initial determination result is overturned and the homepage determination result is corrected by changing it from "yes" to "no", thus avoiding misjudging the continuation page as the homepage based solely on the certificate title.

[0148] The document type classification sub-instruction refers to the second-level sub-instruction, which has a built-in whitelist of 8 types of aviation material certificates (such as certificates of conformity, approval release certificates, maintenance reports, removal tags, disassembled parts tags, accident-free declarations, traceability documents, and non-traceability documents). It requires the model to match the page keywords and lock the unique certificate type from the whitelist, and does not allow the output of custom types outside the whitelist.

[0149] The date extraction sub-instruction refers to the third-order sub-instruction, which forces the model to extract two types of core time fields: the certificate start date and the certificate end date. These two types of dates are the core anchor points for the generation of the resume timeline and airworthiness verification. The instruction requires distinguishing between the effective start time and the end time of the event and prohibits the fuzzy extraction of a single date.

[0150] The extended field extraction sub-instruction refers to the fourth-order differentiated sub-instruction, which matches the corresponding exclusive additional field extraction rules based on the certificate type obtained in the previous step. Different certificates are configured with independent extended field sets, such as extracting the production quantity from the certificate of conformity, extracting the contract / waybill number from the approved release certificate, extracting the repair reason from the repair report, and extracting the removal date from the disassembly tag.

[0151] Extended fields refer to additional business parameters that are bound to the certificate type. They are not core fields applicable to all certificates, but are only valid for specific certificates. They are used to enrich the content displayed on the timeline nodes of the resume and improve the completeness of the traceability information of aviation materials.

[0152] In this embodiment, the field extraction instructions within the special prompt words strictly limit the execution order of the four sub-instructions. The large model completes the parsing step by step, and the complete execution logic is as follows:

[0153] The first step is to execute the homepage verification sub-instruction, which includes: the model first obtains the homepage judgment result based on the certificate number and title. If the homepage judgment result is initially determined to be "yes", the page number and header homepage identifier are immediately read for secondary verification. If the page number is found to be greater than 1 or there is no certificate start identifier during the secondary verification, the homepage judgment result is directly overwritten and modified to "no" to correct the homepage misjudgment problem caused by relying solely on the text title and ensure the accuracy of subsequent certificate grouping boundaries.

[0154] The second step is to execute the file type classification sub-instruction, which includes: after the homepage judgment result is verified, the model traverses all the text on the page, matches the exclusive high-frequency keywords of the 8 preset types of aviation material certificates, matches the unique corresponding file type from the whitelist, and only outputs the certificate classification within the whitelist, preventing the model from generating undefined custom certificate types and unifying the business classification standard.

[0155] The third step is to execute the date extraction sub-instruction, which includes: after locking the certificate type, identifying the relevant time text on the page for effective, invalid, repair, disassembly, and factory release, distinguishing and extracting two independent sets of time parameters: the certificate start date and the certificate end date, clarifying the business meaning of the two types of dates, and providing standardized time data for subsequent history sorting and airworthiness verification.

[0156] The fourth step is to execute the extended field extraction sub-instruction, which includes: based on the certificate type obtained in the second step, calling the exclusive extended field extraction rules for that type, and selectively extracting the corresponding additional business information; different extraction logic is executed for different certificates, irrelevant fields are not extracted, redundant dirty fields in JSON are avoided, and the structured output is fully matched with the backend business database table structure.

[0157] The above four-level sub-instructions are executed sequentially to form a standardized page parsing logic, which forcibly constrains the recognition order and output range of the large model, eliminating problems such as field omissions, classification confusion, and homepage judgment distortion caused by the free parsing of the general large model.

[0158] The above solution breaks down field extraction instructions into four levels of ordered sub-instructions, fixing the execution order of homepage verification, certificate classification, date extraction, and extended field capture. This layered constraint on the large model's parsing logic significantly reduces the probability of page grouping errors by doubly verifying the homepage judgment results. It also limits the standardization of certificates to eight categories, unifying the platform's certificate classification standards, and differentiates the extraction of extended fields specific to each certificate. The output fields are fully compatible with the aviation materials business database, eliminating the need for manual secondary field format conversion. The ordered sub-instructions restrict the model's flexibility at the execution flow level, further reducing recognition illusions, significantly improving the regularity of structured output data, greatly reducing the filtering pressure on the backend verification gate, and enhancing the overall stability of the automated processing flow.

[0159] In some embodiments, step S4 includes at least one of the following rules:

[0160] When the homepage determination result is "yes" and a valid file type, certificate start date, and certificate end date are identified, the identified valid file type, certificate start date, and certificate end date are used as the current identification result; otherwise, the main fields of file type, certificate start date, and certificate end date are set to empty.

[0161] The identified file types are matched with a variety of preset certificate type whitelists. If the identified file type is not in the whitelist or is an empty file type, the identified file type is deemed invalid.

[0162] Verify that the certificate end date is not empty, and if a certificate start date also exists, the certificate start date is not later than the end date;

[0163] The validity of extended fields is verified based on the file type, and only extended fields allowed for that file type are retained;

[0164] Perform semantic and format validation on extended fields for a specified file type.

[0165] In this embodiment, the quality gating of the recognition results is used to intercept contradictory, missing, or formatted data output by the large model, filtering out dirty data before it is entered into the database.

[0166] The main field refers to the core required field within the structured JSON string, which includes the homepage judgment result, file type, certificate start date, and certificate end date. It is the basic data for certificate grouping, resume generation, and airworthiness judgment. If the main field is missing, the recognition result of the page has no business value.

[0167] The preferred whitelist of certificate types is a fixed set of 8 types of aviation material certificates. Only the model is allowed to output the categories in the list. All other custom categories are judged as invalid identification results, thus unifying the certificate classification standards of the platform.

[0168] Date logic validation refers to the validation of business rules for start and end dates. The certificate end date represents the event termination time and cannot be earlier than the certificate start date (event effective time). If the time is reversed, the date field is deemed invalid.

[0169] Extended field validity verification refers to matching the exclusive field permission list based on the current certificate type, retaining only the extended fields allowed for this type of certificate, and eliminating mismatched, irrelevant, and redundant fields to simplify structured data. These extended fields include standardized business fields such as waybill numbers, contract numbers, serial numbers, and repair shop numbers. Regular expressions are used to validate the text format, and semantic validation is performed using an aviation materials dictionary to eliminate invalid parameters with incorrect formatting or semantic contradictions.

[0170] In this embodiment, after the large model outputs a structured JSON string, the system executes multi-layered validation rules sequentially. If any rule fails validation, the corresponding field is marked as invalid. If a main field is severely missing, all core parameters are cleared. The complete validation logic is executed layer by layer.

[0171] The first layer is the main field integrity verification, which specifically includes: the system reads the homepage judgment result from the structured JSON string. If the homepage judgment result is "yes", it synchronously checks whether all three main fields, namely file type, certificate start date and certificate end date, have valid content. If any of the three main fields is missing, all three main fields, namely file type, certificate start date and certificate end date, are cleared directly, and the page recognition result is marked as invalid to prevent invalid pages with missing core information from participating in certificate marshalling.

[0172] The second layer is certificate classification whitelist verification, which includes: extracting the file type field from the structured JSON string and comparing it with the preset 8 types of certificate whitelists one by one; if the field is empty or the text is not in the whitelist, the file type is directly determined to be invalid, and the classification field is cleared at the same time to prevent illegal classification from interfering with subsequent airworthiness determination and resume display.

[0173] The third layer is date logic validation, which includes: firstly verifying that the certificate end date is not empty; if the certificate start date also exists, comparing the timestamps of the two periods; if the timestamp of the certificate start date is later than the certificate end date, determining that the date field has a logical contradiction, clearing the two sets of date parameters, and preventing the generation of a disordered timeline by reversing the time.

[0174] The fourth layer is extended field permission verification, which specifically includes: based on the valid file type after verification, retrieving the list of allowed extended fields for that type, traversing all extended fields in the structured JSON string, deleting irrelevant fields outside the list, retaining only the additional parameters that match the business requirements, and simplifying the stored data.

[0175] The fifth layer is the semantic format verification of extended fields, which specifically includes: for the extended fields retained in the list, calling the corresponding regular expression template to verify the character format, and matching the aviation material terminology dictionary to verify the semantic rationality; for example, the waybill number is limited to an alphanumeric combination format, and the repair shop name is matched with the enterprise name thesaurus. If the format / semantics do not match, the extended field is cleared.

[0176] Multi-layered validation rules filter out dirty data at each level. Only JSON that passes all validations is bound to the single-page electronic document file and stored in the database. Pages that fail validation are uniformly transferred to the manual review process, thus intercepting erroneous data generated by the illusion of the large model from the backend business level.

[0177] The above solution constructs a backend identification quality gate through multi-layered progressive verification rules. It comprehensively filters dirty data output by the model from five dimensions: main field integrity, classification compliance, date logic, extended field permissions, and field format semantics, significantly reducing invalid data flowing into the business database. The verification rules are executed in layers, intercepting defective data in core main fields upfront, preventing invalid pages from participating in automatic certificate grouping, and ensuring the accuracy and reliability of the data foundation for merged certificate files and resume timelines. At the same time, it automatically removes irrelevant extended fields, corrects incorrectly formatted parameters, and standardizes the structured data structure. There is no need for manual cleaning of redundant and contradictory information, further reducing the workload of business personnel in reviewing and correcting, and improving the reliability of the entire process automation.

[0178] In some embodiments, the method includes:

[0179] A file storage path is established with the equipment number as the root directory. The file storage path includes at least a first sub-path and a second sub-path. The first sub-path is used to store the split single-page electronic document files and their thumbnails, and the second sub-path is used to store the merged certificate file.

[0180] In this embodiment, the equipment number is a unique identifier for aviation materials. It can be formed by combining the equipment serial number, equipment model and production batch number. Each independent piece of aviation material corresponds to a unique set of equipment numbers, which serves as the top-level classification basis of the file storage system.

[0181] The root directory is a top-level folder named after a single set of equipment numbers. All certificate files and identification metadata related to this aircraft material are collected in this root directory, achieving physical isolation storage of aircraft material files.

[0182] The file storage path follows a standardized hierarchical folder structure within the system disk, divided into two independent sub-paths based on the root directory. The content stored in these two sub-paths is completely separate, with no files mixed together. The first sub-path is dedicated to storing the single-page electronic document files generated in step S2, as well as the synchronously generated preset image format (e.g., PNG) page thumbnails. Preferably, the maximum side length of the page thumbnails is uniformly set to 200 pixels, allowing business personnel to quickly preview and verify the page content. The second sub-path is dedicated to storing the merged certificate file generated after page grouping in step S4. The merged certificate file is a complete aviation material document composed of the homepage and consecutive pages, serving as the smallest complete file unit displayed on the front-end traceability page.

[0183] The working principle of this embodiment is as follows:

[0184] After completing the single-page electronic document file verification and automatic page grouping operations in step S4, the system initiates the hierarchical file storage logic. The system first reads the equipment number corresponding to the currently processed aviation material and creates a top-level root directory named after that equipment number in the root directory of the file storage server. After the root directory is created, the system generates two independent subfolders in parallel as two-level subpaths, named the single-page source file directory and the complete merged certificate directory, respectively.

[0185] The system iterates through all verified single-page electronic document files and their corresponding thumbnails, batches them into the first sub-path, and simultaneously writes the metadata tags bound to the single-page electronic document files into the file's attribute, permanently binding the single-page file to the identification information. The system then iterates through all grouped merged certificate files, uniformly migrating them into the second sub-path. Each merged certificate file is named with the original certificate package file identifier and the first page number, facilitating quick location of the original pagination source.

[0186] The entire storage path structure uses the equipment number as the top-level isolation standard. Certificate files for different aviation materials belong to separate root directories, avoiding file search chaos caused by the mixed storage of a large number of aviation material certificates. The two-level sub-paths physically distinguish between fragmented paginated files and complete business certificates. Subsequent traceability queries, manual review, and file export operations can all retrieve the corresponding files according to the path without having to traverse all certificate data.

[0187] The above solution achieves physical isolation of different aviation material certificate files by building a two-tiered storage architecture driven by equipment number, completely solving the problem of difficult retrieval due to mixed storage of multiple batches of aviation material certificates. Two sub-paths are used to distinguish between paginated fragmented files and complete merged certificates. Front-end preview thumbnails only retrieve lightweight image resources from the first sub-path, and merged certificate file downloads only read files from the second sub-path, reducing system file read bandwidth consumption. The standardized path structure adapts to the platform's long-term, massive certificate storage expansion needs, clearly defines file ownership, and simplifies subsequent data archiving, cleaning, and backup maintenance processes, improving the efficiency of aviation material certificate file storage management throughout its entire lifecycle.

[0188] In some embodiments, the method includes:

[0189] Based on the file type of the merged certificate file, the airworthiness of the aircraft material is determined, and the airworthiness determination includes:

[0190] Obtain the aircraft material type identifier, which is pre-stored in the aircraft material database;

[0191] Based on the aircraft material type identifier, the corresponding certificate file type set is queried from the airworthiness rule base. The certificate file type set contains all certificate file types required for the airworthiness of this type of aircraft material.

[0192] The actual file types identified in the merged certificate file are matched item by item with the certificate file type set. When every file type in the certificate file type set exists in the actual identified file types, the aircraft material is determined to meet the airworthiness requirements.

[0193] In this embodiment, the airworthiness of aircraft materials is a compliance judgment standard in the aviation field. It means that the aircraft materials have complete certification documents required by the civil aviation quality management specifications and can be used normally in aircraft assembly, maintenance and circulation. If the required certificates are missing, they are judged as non-compliant aircraft materials.

[0194] Aircraft material type identifiers are pre-entered classification tags in the aircraft material database, divided into four categories: standard parts, new parts, used parts, and disassembled parts. The airworthiness compliance standards for different categories of aircraft materials differ significantly. The airworthiness rule base is a pre-configured industry compliance database, built entirely in accordance with the Civil Aviation Administration of China's aircraft material quality management regulations. Within the database, each type of aircraft material identifier is bound to a unique set of certificate document types.

[0195] The certificate file type set is a list of all the certificate types required for compliance of a certain type of aviation material. The elements in the list are taken from the standardized certificate whitelist.

[0196] Item-by-item matching is a verification logic that compares all automatically identified merged certificate file types with the compliance list one by one. Only when all certificate types in the list have corresponding identification files can the airworthiness determination be passed.

[0197] The working principle of this embodiment is as follows:

[0198] Before generating the traceability timeline in step S5, the system automatically executes the fully automated airworthiness determination process for aircraft materials, which specifically includes:

[0199] The first step is for the system to read the equipment number of the current aviation material, retrieve the pre-entered aviation material type identifier from the aviation material database, and confirm whether the aviation material belongs to a standard part, a brand new part, a used part, or a disassembled part.

[0200] The second step involves the system using the aircraft material type identifier as the search condition to access the local airworthiness rule base and retrieve the set of certificate file types corresponding to that category of aircraft material. The set records all the certificate types that the aircraft material must have for compliant circulation.

[0201] The third step involves the system traversing all merged certificate files stored in the second sub-path, extracting the file type field from the structured JSON string corresponding to each file, and summarizing the results to obtain a list of all certificate types actually held by the current aviation material.

[0202] The fourth step involves the system comparing each certificate type in the compliant certificate set with the actual certificate list item by item to complete a full-coverage verification.

[0203] If all certificate types within the compliance set can be found in the actual identification results, the system determines that the aircraft material meets the airworthiness requirements and displays a compliance mark on the traceability history page; if any certificate type within the compliance set is missing, the aircraft material is determined to be non-airworthiness, the front-end page will indicate the name of the missing certificate type, and the purchaser will be prompted to supplement the corresponding documents.

[0204] The aforementioned solution builds an airworthiness rule base based on official civil aviation quality management standards, setting differentiated compliance verification standards for four types of aviation materials. This enables fully automated assessment of aviation material airworthiness qualifications, replacing manual verification of certificate completeness against regulations. Airworthiness assessment results are directly integrated with the front-end traceability display page, allowing purchasers to intuitively view the compliance status of aviation materials and any missing documents, reducing manual costs for compliance verification in the aviation material trading process. The assessment logic is fully integrated with the aforementioned automatic certificate classification system, eliminating the need for manual input of certificate classification information and enabling fully automated operation, thereby enhancing the compliance and risk control capabilities of the aviation material trading platform.

[0205] In some embodiments, in step S3, the structured JSON string further includes equipment identification information, which includes at least one of equipment serial number, equipment model, and equipment batch number;

[0206] Step S4 is followed by:

[0207] Based on the equipment identification information, a common entity is extracted from the structured JSON strings of multiple equipment. The common entity includes at least one of the following: manufacturer name, repair shop name, issuing authority name, and production batch number.

[0208] Using equipment as nodes and the common entities as associated edges, a cross-equipment association graph is constructed, wherein the association graph is stored using an adjacency matrix or an adjacency list data structure.

[0209] Based on the aforementioned association graph, a breadth-first search algorithm or a depth-first search algorithm is used to detect abnormal equipment. The abnormal equipment includes isolated equipment that has no association edges with other equipment, abnormally associated equipment that has association edges with more than a preset number of other equipment, and equipment with conflicting associations where the certificate of conformity and the repair report are from different manufacturers.

[0210] Different levels of early warning information are generated based on the type of anomaly, and the early warning information includes at least one of a first color indication information, a second color indication information, and a third color indication information;

[0211] The association map and the early warning information are added to the structured JSON string, wherein the association map is stored in the form of a JSON array, and each array element contains equipment identification information and a list of associated edges.

[0212] In this embodiment, the equipment identification information is an extended storage field in a structured JSON string, used to uniquely distinguish a single piece of aircraft material. The field can be selected from three types of parameters: equipment serial number, equipment model, and equipment batch number. At least one type of data must be stored as a unique identifier for the node in the association graph.

[0213] Common entities are industry organizations and production identification information that appear repeatedly in multiple sets of aviation material identification data. They include four types of content: manufacturer name, maintenance plant name, issuing authority name, and production batch number, representing the overlap in business flow between different aviation materials.

[0214] The association graph is a global data network spanning aircraft materials. The basic unit of the graph consists of nodes and associated edges. Nodes correspond to individual aircraft materials, while associated edges represent business relationships where two aircraft materials share the same common entity. Adjacency matrices and adjacency lists are two optional storage structures for the graph. Adjacency matrices record the relationships between nodes in a two-dimensional table, while adjacency lists store all associated objects of a single node in a list format. The platform can flexibly choose the appropriate structure based on the amount of equipment data.

[0215] Both breadth-first search and depth-first search algorithms are graph traversal retrieval algorithms used to globally scan all nodes and associated edges to batch identify three types of abnormal equipment. Isolated equipment refers to aircraft parts nodes without any associated edges in the graph, indicating that the equipment has no shared manufacturer, production batch, or other circulation information. Abnormally associated equipment refers to aircraft parts whose number of associated edges exceeds a second preset threshold, indicating that the equipment's circulation network is abnormally large. Conflictingly associated equipment refers to aircraft parts where the manufacturer recorded on the certificate of conformity is inconsistent with the manufacturer recorded on the maintenance report, indicating contradictory manufacturer information.

[0216] The warning information is divided into three levels of color indication information, corresponding to three types of abnormal equipment. All of them are extended and written into a structured JSON string for synchronous storage. After the front-end traceability page reads the information, it highlights the risk warning.

[0217] The JSON array is the output format of the association graph. Each element in the array records the identifier of a single piece of aviation material and a list of all associated edges of that material, ensuring that the graph data can be directly connected to the front-end visualization rendering.

[0218] In this embodiment, after completing all verifications and certificate grouping and storage in step S4, the system initiates the cross-device association map construction and anomaly detection process, the specific process of which is as follows:

[0219] The system reads all the stored structured JSON strings and extracts the equipment identification information carried in each JSON, which is used as the basic data for the graph nodes.

[0220] The system then iterates through all the text fields in the JSON, extracts four common entities in batches: manufacturer name, repair shop name, issuing authority name, and production batch number, and records all the equipment nodes corresponding to each entity.

[0221] The system then builds an association graph, with each piece of aviation material generating an independent node. When two pieces of aviation material share the same common entity, an association edge is generated between the two nodes. The graph can be stored in either an adjacency matrix or an adjacency list. For scenarios with massive aviation material data, the adjacency list is preferred to reduce storage usage.

[0222] The system then uses either a breadth-first search algorithm or a depth-first search algorithm to traverse the complete map and simultaneously detect three types of abnormal equipment.

[0223] The system then matches the corresponding color warning label to the identified abnormal equipment type: isolated equipment is matched with the first color indicator, abnormally associated equipment is matched with the second color indicator, and associated conflicting equipment is matched with the third color indicator.

[0224] The system then converts the complete association map into a JSON array format. Each element in the array is bound to the corresponding equipment identifier and the list of associated edges. The map array and the level 3 early warning information are then written into the structured JSON string of the corresponding aviation material and stored in the database for persistent storage, so that the front-end traceability page can read and render it.

[0225] The above solution automatically constructs a global correlation graph based on aviation material identification data, eliminating the need for manual input of manufacturer distribution relationships and fully reconstructing the business relationship network between all aviation materials. Relying on a graph traversal algorithm, it automatically and in batches screens three types of high-risk equipment, achieving proactive risk identification in aviation material trade and overcoming the shortcomings of traditional traceability systems that only display information on individual equipment and cannot provide comprehensive risk control. The graph and early warning information are uniformly embedded in the existing structured JSON string, eliminating the need for additional independent data tables to store correlation data, ensuring compatibility with existing business database architectures. Three-level color-coded warnings intuitively distinguish risk levels, facilitating tiered handling of abnormal aviation materials by operations personnel, and comprehensively improving the risk management capabilities throughout the entire aviation material transaction process.

[0226] In some embodiments, in step S5, the rules for generating the warning information include:

[0227] When an isolated device is detected, a first color indication is generated. The first color indication is used to indicate that the device is of unknown origin and automatically prevents the device's traceability information from being displayed to the outside world.

[0228] When an abnormally associated device is detected, a second color indicator is generated. The second color indicator is used to indicate that the device needs further verification and the device is marked as pending verification on the traceability display page.

[0229] When a conflicting device is detected, a third color indication is generated. This third color indication is used to indicate that the device requires manual review, and the device's traceability information is submitted to the review queue.

[0230] In this embodiment, the first color indication information corresponds to a unique warning identifier for isolated equipment. Isolated equipment lacks any shared manufacturer or production batch association, making it impossible to trace its upstream and downstream distribution links, thus it is determined to be aviation materials of unknown origin. Automatically blocking external display is part of the platform's access control logic. For aviation materials carrying the first color warning, the front-end e-commerce site cannot load and display their traceability history and certificate files; only back-end administrators can see the complete data.

[0231] The second color indicator corresponds to an abnormal associated equipment warning sign. This indicates that the number of associated edges for this type of aircraft material exceeds the system's preset second threshold, and the scale of the circulation network exceeds the standard for regular aircraft materials, posing a risk of bulk circulation and multiple layers of intermediaries. The "Pending Verification" status is a unique marker on the front-end page. The purchaser can view the basic traceability data for this aircraft material, and the page will display a fixed risk warning text, prompting the purchaser to proactively verify the circulation background with the platform operators.

[0232] The third color indicator corresponds to a warning sign for conflicting equipment. Inconsistent manufacturer information on different certificates for the same aircraft material raises risks of document forgery and data entry errors. The review queue is a list of tasks processed manually in the background. Aircraft materials with a third-color warning are automatically pushed to the top of the queue, forcing operations personnel to manually review and correct them. The traceability page will be restricted from display until the review is completed.

[0233] The aforementioned solution establishes tiered control rules for three categories of risky equipment, differentiating the control intensity for risks related to unknown origin, abnormal circulation, and conflicting documentation. This approach strictly blocks the external circulation of aircraft materials with unknown origins while avoiding complete blocking of low-risk, abnormally circulating equipment, thus balancing platform transaction compliance with the needs of purchasers for information access. The early warning and control logic is deeply integrated with the record generation process, eliminating the need for manual marking of risky equipment and configuring page permissions; the entire process is automated, reducing the platform's risk control workload. The tiered display mechanism clearly distinguishes risk levels, allowing purchasers to intuitively identify potential compliance risks in aircraft materials, assisting in procurement decisions and reducing disputes arising from non-compliant aircraft material transactions.

[0234] like Figure 2As shown in the attached diagram, the human-machine collaborative automated processing flow for aircraft material certificates in this embodiment is as follows: First, the aircraft material certificates are accessed, with manual import or synchronization of certificate files; the information center automatically completes PDF splitting, and then uses a large language model to sequentially perform homepage determination, file type classification, key information extraction, and other information extraction; the system further conducts quality control of the recognition results; finally, PDF merging, recognition result adjustment, and saving are completed. The certificate access and recognition result adjustment stages are configured with manual operation entry points, while the PDF splitting, model parsing, and quality control stages are fully automated; when the recognition result verification is abnormal, a manual review and adjustment mechanism is initiated.

[0235] In the second aspect, such as Figure 3 As shown, the present invention provides an aircraft material traceability history generation system 10 based on a large language model, the system comprising:

[0236] The certificate acquisition module 101 is used to acquire the original certificate package file, which contains one or more single certificate files;

[0237] The document splitting module 102, connected to the certificate acquisition module, is used to split the original certificate package file into multiple single-page electronic document files by page, and generate metadata tags for each single-page electronic document file. The metadata tags include at least the page number of the single-page electronic document file in the original certificate package file and the identifier of the original certificate package file to which it belongs.

[0238] The intelligent extraction module 103, connected to the document splitting module, is used to calculate the visual-text joint entropy value of each single-page electronic document file, select the corresponding level template from the preset multi-level prompt word template library according to the visual-text joint entropy value, and generate special prompt words; input the single-page electronic document file and the special prompt words together into the large language model to obtain the structured JSON string output by the large language model;

[0239] The structured JSON string includes at least the homepage determination result, file type, certificate start date, and certificate end date. The homepage determination result is used to identify whether the single-page electronic document file is the starting page of a single certificate file. The special prompt word contains field extraction instructions. The large language model extracts the certificate number and certificate title from the page content of the single-page electronic document file according to the field extraction instructions. When the complete certificate number and certificate title are extracted, the homepage determination result is set to "yes"; otherwise, it is set to "no".

[0240] The verification storage module 104, connected to the intelligent extraction module, is used to verify the structured JSON string, establish a one-to-one correspondence between the verified structured JSON string and the corresponding single-page electronic document file using the metadata tag as an index, and store it in the database; according to the homepage determination result, the single-page electronic document file with a homepage determination result of "yes" is used as the starting page of the single certificate file, the single-page electronic document file with a homepage determination result of "no" is concatenated to its corresponding single certificate file in page number order, and the single-page electronic document file with a homepage determination result of "yes" and the concatenated single-page electronic document file with a homepage determination result of "no" are merged into a single merged certificate file;

[0241] The timeline generation module 105, connected to the verification and storage module, is used to generate a timeline of traceability history of aircraft materials by arranging the merged certificate files in chronological order, using the certificate end date in the structured JSON string as the event anchor.

[0242] In this embodiment, the certificate acquisition module 101 is the system data entry point, supporting two channels for obtaining original certificate package files: manual file upload and automatic synchronization within the enterprise framework. It also binds aviation material identification information to establish the association between the certificate and the aviation material.

[0243] The document splitting module 102 receives the original certificate package file output by the certificate acquisition module, performs all operations such as pagination splitting, thumbnail rendering, and metadata tag generation, and outputs a single-page electronic document file carrying metadata to the downstream intelligent extraction module.

[0244] The intelligent extraction module 103 integrates an entropy calculation unit, a prompt word template library calling unit, and a large model calling unit to complete the quantification of page complexity, adaptive prompt word generation, structured information extraction, and output standardized JSON recognition results to downstream.

[0245] The verification storage module 104 has a built-in complete set of multi-layer business verification rules and file storage scheduling unit, which completes dirty data filtering, single page and identification data binding, automatic page grouping, hierarchical file storage, and outputs complete merged certificate files and structured data for database storage.

[0246] The timeline generation module 105 receives all the data persisted by the verification and storage module, integrates the airworthiness determination unit, the association map construction unit, the risk warning generation unit, and the history visualization rendering unit, and finally outputs a timeline of aircraft material traceability history that can be displayed externally.

[0247] The working principle of the aircraft material traceability and history generation system 10 based on a large language model is as follows:

[0248] The certificate acquisition module receives the original certificate package file from the outside, synchronously binds the corresponding aviation material's equipment number, model and other identification information, and after completing the data entry verification, pushes the original certificate package file to the document splitting module.

[0249] The document splitting module reads the original certificate package file, calls the PDF parsing library to cut it page by page to generate independent single-page electronic documents, renders and generates thumbnails of uniform size, generates metadata tags for each page containing page numbers and the original certificate package file identifier, and transmits all single-page electronic document files and metadata together to the intelligent extraction module.

[0250] The intelligent extraction module sequentially performs visual text joint entropy calculation, matches corresponding level prompt word templates to generate dedicated prompt words, sends the single-page electronic document image and dedicated prompt words into the internal private large language model, receives the model's streaming output text and concatenates it into a structured JSON string. The JSON string carries core fields such as the homepage judgment result, file type, and certificate start and end dates. The JSON string and the corresponding single-page electronic document file are then synchronously pushed to the verification and storage module.

[0251] The verification storage module retrieves the complete set of verification rules to perform multi-layer filtering on the JSON. Data that passes verification is stored in the database by binding the corresponding single-page electronic document file with metadata tags as indexes. The system traverses all single-page homepage judgment tags, automatically concatenates the homepage and consecutive pages to generate a merged certificate file, and completes file layered persistence based on two-level storage paths. After storage is completed, the merged certificate file and the corresponding structured JSON string are pushed to the timeline generation module.

[0252] The timeline generation module reads the certificate end dates corresponding to all merged certificates as event anchors, sorts all certificates in chronological order, and simultaneously performs airworthiness determination, cross-material association map construction, and three-level risk warning generation operations. It integrates all data to render and generate a visualized aircraft material traceability timeline, which is then pushed to the front-end e-commerce site for external display.

[0253] The above solution involves a complete system that covers the entire business chain, including certificate access, pagination processing, intelligent extraction of large models, data verification and storage, and resume generation. It integrates the core technical logic of all the methods mentioned above, and is implemented in a software and hardware integrated manner. It can be directly deployed on the aviation material integrated service platform, aviation material e-commerce, and MRO operation and maintenance system, and is suitable for the digital traceability transformation needs of various aviation enterprises.

[0254] Finally, it should be noted that although the above embodiments have been described in the description and drawings of this invention, this should not limit the scope of patent protection of this invention. Any technical solutions that are based on the essential concept of this invention, utilize the content described in the description and drawings of this invention to make equivalent structural or procedural substitutions or modifications, as well as the direct or indirect application of the technical solutions of the above embodiments to other related technical fields, are all included within the scope of patent protection of this invention.

Claims

1. A method for generating aircraft material traceability history based on a large language model, characterized in that, The method includes: S1: Obtain the original certificate package file, which contains one or more single certificate files; S2: Split the original certificate package file into multiple single-page electronic document files by page, and generate a metadata tag for each single-page electronic document file. The metadata tag includes at least the page number of the single-page electronic document file in the original certificate package file and the identifier of the original certificate package file to which it belongs. S3: Calculate the visual-text joint entropy value of each of the single-page electronic document files, select the corresponding level template from the preset multi-level prompt word template library according to the visual-text joint entropy value, and generate a special prompt word; input the single-page electronic document file and the special prompt word together into the large language model, and obtain the structured JSON string output by the large language model; The structured JSON string includes at least the homepage determination result, file type, certificate start date, and certificate end date. The homepage determination result is used to identify whether the single-page electronic document file is the starting page of a single certificate file. The special prompt word contains field extraction instructions. The large language model extracts the certificate number and certificate title from the page content of the single-page electronic document file according to the field extraction instructions. When the complete certificate number and certificate title are extracted, the homepage determination result is set to "yes"; otherwise, it is set to "no". S4: Validate the structured JSON string, and establish a one-to-one correspondence between the validated structured JSON string and the corresponding single-page electronic document file using the metadata tag as an index, and store it in the database; Based on the homepage determination result, the single-page electronic document file with a homepage determination result of "yes" is used as the starting page of a single certificate file, the single-page electronic document file with a homepage determination result of "no" is sequentially concatenated to its corresponding single certificate file according to the page number order, and the single-page electronic document file with a homepage determination result of "yes" and the single-page electronic document file with a homepage determination result of "no" concatenated thereafter are merged into a single merged certificate file; S5: Using the certificate expiration date in the structured JSON string as the event anchor, arrange the merged certificate files in chronological order to generate a timeline for the traceability of aircraft materials.

2. The method for generating aircraft material traceability history based on a large language model as described in claim 1, characterized in that, In step S3, calculating the visual-text joint entropy value for each of the single-page electronic document files includes: The single-page electronic document file is subjected to OCR recognition, and the ratio of Chinese characters to English characters in the OCR recognition result is calculated as the Chinese-English mixing degree. The OCR recognition results are matched with a dictionary of professional terms in the field of aviation materials, and the proportion of successfully matched characters to the total number of characters is counted as the professional term density. The layout of the single-page electronic document is analyzed to identify the table area and the plain text area in the single-page electronic document, and the ratio of the area of ​​the table area to the area of ​​the plain text area is calculated as the degree of table and text mixing. The visual-text joint entropy value is obtained by weighting and fusing the Chinese-English mixing degree, the density of professional terms, and the mixing degree of tables and text.

3. The method for generating aircraft material traceability history based on a large language model as described in claim 1 or 2, characterized in that, The grade templates include low-entropy templates, medium-entropy templates, and high-entropy templates. The low-entropy templates contain field extraction instructions, the medium-entropy templates contain field extraction instructions and field descriptions, and the high-entropy templates contain layout description instructions, field extraction instructions, and conflict detection instructions. Based on the visual-text joint entropy value, a corresponding level template is selected from a preset multi-level prompt word template library to generate dedicated prompt words, including: When the visual-text joint entropy value is higher than the first preset threshold, the layout description instruction, field extraction instruction and conflict detection instruction are extracted from the high-entropy template, and the layout description instruction, field extraction instruction and conflict detection instruction are spliced ​​and combined to generate the special prompt word; When the visual-text joint entropy value is lower than or equal to the first preset threshold and higher than the second preset threshold, the field extraction instruction and field description are extracted from the medium entropy template to generate the special prompt word; When the visual-text joint entropy value is lower than or equal to the second preset threshold, the field extraction instruction is extracted from the low-entropy template to generate the special prompt word.

4. The method for generating aircraft material traceability history based on a large language model as described in claim 1 or 2, characterized in that, The field extraction instruction also includes sub-instructions executed in the following order: The homepage verification sub-instruction is used to instruct the large language model to further verify whether the current page is the certificate homepage based on the page number information after the homepage determination result is set to "yes". If the verification result is "no", the homepage determination result is modified to "no". The file type classification sub-instruction is used to instruct the large language model to match file types among a variety of preset certificate types; The date extraction sub-instruction is used to instruct the large language model to extract the certificate start date and certificate end date; The extended field extraction sub-instruction is used to instruct the large language model to extract the corresponding extended fields according to the file type, wherein the extended fields are additional information fields related to the certificate type.

5. The method for generating aircraft material traceability history based on a large language model as described in claim 4, characterized in that, In step S4, the verification rules include at least one of the following rules: When the homepage determination result is "yes" and a valid file type, certificate start date, and certificate end date are identified, the identified valid file type, certificate start date, and certificate end date are used as the current identification result; otherwise, the main fields of file type, certificate start date, and certificate end date are set to empty. The identified file types are matched with a variety of preset certificate type whitelists. If the identified file type is not in the whitelist or is an empty file type, the identified file type is deemed invalid. Verify that the certificate end date is not empty, and if a certificate start date also exists, the certificate start date is not later than the end date; The validity of extended fields is verified based on the file type, and only extended fields allowed for that file type are retained; Perform semantic and format validation on extended fields for a specified file type.

6. The method for generating aircraft material traceability history based on a large language model as described in claim 1, characterized in that, The method includes: A file storage path is established with the equipment number as the root directory. The file storage path includes at least a first sub-path and a second sub-path. The first sub-path is used to store the split single-page electronic document files and their thumbnails, and the second sub-path is used to store the merged certificate file.

7. The method for generating aircraft material traceability history based on a large language model as described in claim 1, characterized in that, The method includes: Based on the file type of the merged certificate file, the airworthiness of the aircraft material is determined, and the airworthiness determination includes: Obtain the aircraft material type identifier, which is pre-stored in the aircraft material database; Based on the aircraft material type identifier, the corresponding certificate file type set is queried from the airworthiness rule base. The certificate file type set contains all certificate file types required for the airworthiness of this type of aircraft material. The actual file types identified in the merged certificate file are matched item by item with the certificate file type set. When every file type in the certificate file type set exists in the actual identified file types, the aircraft material is determined to meet the airworthiness requirements.

8. The method for generating aircraft material traceability history based on a large language model as described in claim 1, characterized in that, In step S3, the structured JSON string also includes equipment identification information, which includes at least one of equipment serial number, equipment model, and equipment batch number; Step S4 is followed by: Based on the equipment identification information, a common entity is extracted from the structured JSON strings of multiple equipment. The common entity includes at least one of the following: manufacturer name, repair shop name, issuing authority name, and production batch number. Using equipment as nodes and the common entities as associated edges, a cross-equipment association graph is constructed, wherein the association graph is stored using an adjacency matrix or an adjacency list data structure. Based on the aforementioned association graph, a breadth-first search algorithm or a depth-first search algorithm is used to detect abnormal equipment. The abnormal equipment includes isolated equipment that has no association edges with other equipment, abnormally associated equipment that has association edges with more than a preset number of other equipment, and equipment with conflicting associations where the certificate of conformity and the repair report are from different manufacturers. Different levels of early warning information are generated based on the type of anomaly, and the early warning information includes at least one of a first color indication information, a second color indication information, and a third color indication information; The association map and the early warning information are added to the structured JSON string, wherein the association map is stored in the form of a JSON array, and each array element contains equipment identification information and a list of associated edges.

9. The method for generating aircraft material traceability history based on a large language model as described in claim 8, characterized in that, In step S5, the rules for generating the early warning information include: When an isolated device is detected, a first color indication is generated. The first color indication is used to indicate that the device is of unknown origin and automatically prevents the device's traceability information from being displayed to the outside world. When an abnormally associated device is detected, a second color indicator is generated. The second color indicator is used to indicate that the device needs further verification and the device is marked as pending verification on the traceability display page. When a conflicting device is detected, a third color indication is generated. This third color indication is used to indicate that the device requires manual review, and the device's traceability information is submitted to the review queue.

10. A system for generating traceability and history of aircraft materials based on a large language model, characterized in that, The system includes: The certificate acquisition module is used to acquire the original certificate package file, which contains one or more single certificate files; The document splitting module, connected to the certificate acquisition module, is used to split the original certificate package file into multiple single-page electronic document files by page, and generate metadata tags for each single-page electronic document file. The metadata tags include at least the page number of the single-page electronic document file in the original certificate package file and the identifier of the original certificate package file to which it belongs. The intelligent extraction module, connected to the document splitting module, is used to calculate the visual-text joint entropy value of each single-page electronic document file, select the corresponding level template from the preset multi-level prompt word template library according to the visual-text joint entropy value, and generate special prompt words; input the single-page electronic document file and the special prompt words together into the large language model to obtain the structured JSON string output by the large language model; The structured JSON string includes at least the homepage determination result, file type, certificate start date, and certificate end date. The homepage determination result is used to identify whether the single-page electronic document file is the starting page of a single certificate file. The special prompt word contains field extraction instructions. The large language model extracts the certificate number and certificate title from the page content of the single-page electronic document file according to the field extraction instructions. When the complete certificate number and certificate title are extracted, the homepage determination result is set to "yes"; otherwise, it is set to "no". The verification storage module, connected to the intelligent extraction module, is used to verify the structured JSON string. The verified structured JSON string, indexed by the metadata tag, is linked to the corresponding single-page electronic document file and stored in the database. Based on the homepage determination result, the single-page electronic document file with a "yes" homepage determination result is used as the starting page of a single certificate file. The single-page electronic document files with a "no" homepage determination result are sequentially appended to their respective single certificate files according to page number order. The single-page electronic document files with a "yes" homepage determination result and the appended single-page electronic document files with a "no" homepage determination result are merged into a single merged certificate file. The timeline generation module, connected to the verification and storage module, is used to generate a timeline of traceability history for aircraft materials by arranging the merged certificate files in chronological order, using the certificate expiration date in the structured JSON string as the event anchor.

Images