Airworthiness document review method based on dynamic active learning and knowledge graph agent

Abstract

This invention discloses an airworthiness document review method based on dynamic active learning and knowledge graph intelligent agents. It collects and processes multimodal heterogeneous data sources to obtain a standardized knowledge base. Based on this standardized knowledge base, it constructs an aviation domain knowledge system using dynamic active learning and human-machine collaboration mechanisms. It then constructs a knowledge graph and performs knowledge enhancement. Finally, it dynamically constructs a review context based on a retrieval enhancement generation method combined with the knowledge graph, and injects this context into an intelligent agent for intelligent data review and report generation. This invention achieves automated review of airworthiness documents through a dynamic active learning-intelligent agent review method, significantly improving the efficiency of manual review. It utilizes RAG technology to inject domain knowledge into the intelligent agent, improving the accuracy of general-purpose large language models in airworthiness review tasks. Through knowledge semantic driving and human-in-the-loop knowledge enhancement, it alleviates the problem of scarcity of high-quality data in the domain. Finally, it collaboratively utilizes models with different parameter scales, saving computational resources.

Description

Technical Field

[0001] This invention belongs to the field of natural language query processing, and specifically relates to a method for reviewing airworthiness documents based on dynamic active learning and knowledge graph intelligent agents. Background Technology

[0002] Airworthiness is an attribute of civil aircraft, referring to the safety and physical integrity that the aircraft as a whole, encompassing all its components and subsystems, should possess within its intended operating environment and usage restrictions. To ensure this critical attribute, civil aviation authorities worldwide have established rigorous and complex airworthiness certification systems. These systems require that the design, manufacture, testing, and maintenance of aircraft strictly adhere to airworthiness regulations, standards, circulars, and technical documents, significantly improving the overall safety and operational reliability of aircraft. This makes the system a cornerstone of ensuring modern civil aviation safety. Only after passing airworthiness certification can an aircraft be used in the civil aviation sector.

[0003] The review of airworthiness documents, such as safety analysis reports, conformity verification reports, and maintenance programs, is the core of airworthiness certification and serves as crucial evidence proving that aircraft and their components fully meet statutory airworthiness standards. Therefore, a comprehensive and accurate review of airworthiness documents is an indispensable step in ensuring the continued airworthiness of aircraft and safeguarding public safety.

[0004] However, the existing airworthiness document review process still faces significant challenges. The airworthiness review process generates a large number of documents requiring review; however, current review work is primarily manual. Faced with a vast, complex document system filled with technical jargon, traditional manual review methods are not only inefficient and time-consuming, but also struggle to systematically ensure consistency between documents and review standards. Potential risks such as content conflicts, incomplete standard coverage, and compliance oversights are sometimes difficult to effectively identify and control, posing hidden dangers to aviation safety.

[0005] In recent years, with the development of natural language processing technology, especially the rapid development of large language models such as BERT, GPT, and Deepseek, potential technical paths have been provided for achieving automated and intelligent text review. These models are trained on large-scale general corpora and can learn the complex semantics, grammatical structures, and contextual relationships of language, thus demonstrating near-human capabilities in several core tasks such as text understanding, content generation, and logical reasoning.

[0006] However, directly applying existing natural language processing (NLP) technologies to the highly specialized field of airworthiness document review still presents numerous challenges. First, the airworthiness review field has certain domain knowledge barriers; general-purpose large language models lack understanding of airworthiness terminology, complex regulations, and logical systems, making it difficult to achieve the required accuracy in document review. Second, high-quality data is scarce; constructing high-quality, domain-specific datasets requires significant domain expert resources and is costly. Meanwhile, aviation industry data is scattered across multiple sources, making it difficult to expand high-quality data, leading to data bottlenecks for supervised learning fine-tuning methods. Furthermore, computational and parameter limitations exist; fine-tuning pre-trained models with large parameter sets requires substantial computational resources and has a high implementation threshold. While lightweight models have lower computational requirements, their performance often falls short of the demands of actual review when handling complex semantic understanding and logical reasoning tasks in airworthiness documents. Summary of the Invention

[0007] To address the efficiency and accuracy bottlenecks caused by the reliance on manual review in existing airworthiness document reviews, as well as the technical challenges of domain knowledge barriers, data scarcity, and high computing costs when applying general-purpose large language models in this field, this invention proposes an airworthiness document review method based on dynamic active learning and knowledge graph intelligent agents.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] A method for reviewing airworthiness documents based on dynamic active learning and knowledge graph intelligent agents includes the following steps:

[0010] Step 1: Collect data from various multimodal heterogeneous data sources, fuse and intelligently preprocess the collected multimodal heterogeneous data to obtain a standardized knowledge base;

[0011] Step 2: Based on the standardized knowledge base obtained in Step 1, construct an aviation knowledge system using dynamic active learning and human-machine collaboration mechanisms;

[0012] Step 3: Based on the current pre-set review task objectives, construct a knowledge graph and perform knowledge enhancement based on the aviation knowledge system built in Step 2;

[0013] Step 4: Based on the retrieval enhancement generation method and the knowledge-enhanced knowledge graph obtained in Step 3, dynamically construct the review context and inject it into the intelligent agent to perform intelligent data review and report generation.

[0014] A further optimized solution, the specific process of step 1 is as follows:

[0015] Step 1.1: Collect multimodal heterogeneous data sources, including aviation compliance standards library, historical project case library, aviation software defect and verification knowledge base, domain knowledge graph, and unstructured multimodal data;

[0016] Step 1.2: Collect the multimodal heterogeneous data sources gathered in Step 1.1 using multiple channels, including batch file upload and parsing, standardized API data integration, and intelligent web crawling and parsing, to finally form a multimodal raw database;

[0017] Step 1.3: The multimodal raw database collected in Step 1.2 is fused and intelligently preprocessed, including data normalization and format standardization, data cleaning and noise reduction, to form a standardized knowledge base with a unified format.

[0018] A further preferred embodiment is that the aviation compliance standards library mentioned in step 1.1 includes international standards DO-178, DO-254, and ARP4761; the domain knowledge graph refers to existing knowledge graphs in the aviation field; and the unstructured multimodal data includes charts and images, scanned copies of design documents, and system logs.

[0019] In a further preferred embodiment, in step 2, the aviation knowledge system includes a general airborne database and a mission-specific database related to the mission.

[0020] In a further optimized approach, step 2 involves the following specific process for constructing a knowledge system in the aviation field using dynamic active learning and human-machine collaboration mechanisms:

[0021] Step 2.1: Using a lightweight language model, based on the preset review task objectives, perform semantic similarity retrieval and initial screening on the standardized knowledge base obtained in Step 1 to obtain an initial screening candidate set; using a lightweight language model, divide the initial screening candidate set into three categories by predicting probability entropy: high-confidence airborne general database samples, high-confidence task-specific database samples related to the task, and doubtful samples judged by the model to have low confidence.

[0022] Step 2.2: Manually review and verify the low-confidence suspicious samples identified by the model, inject knowledge into the samples, and form a golden training sample increment.

[0023] Step 2.3: Based on the golden training samples obtained in Step 2.2, perform incremental iterative training of the lightweight language model based on LORA, update the parameters of the lightweight language model, and obtain the updated lightweight language model.

[0024] Step 2.4: If there are still low-confidence suspicious samples that have not been manually reviewed and verified, then use the updated lightweight language model to classify them again by predicting the probability entropy; then return to step 2.2 to continue execution until the loop termination condition is met.

[0025] In a further preferred embodiment, in step 2.1, the preset review task objectives specifically include: identifying knowledge related to "compliance verification", identifying knowledge related to "software review process", identifying knowledge related to "software grade", and identifying knowledge related to "verification method".

[0026] The "compliance verification" specifically includes: Declaration of Conformity (MOC0), Explanatory Documents (MOC1), Analytical Calculations (MOC2), Safety Assessment (MOC3), Laboratory Testing (MOC4), Ground Testing (MOC5), Flight Testing (MOC6), Simulator Testing (MOC8), and Equipment Conformity (MOC9).

[0027] The "software review process" specifically includes: Software Planning Phase Review (SOI1), Software Development Phase Review (SOI2), Software Verification Phase Review (SOI3), and Software Final Approval Review (SOI4).

[0028] Specifically, the "software level" includes: software levels A, B, C, D, and E;

[0029] Specifically, the "verification method" includes: review, analysis, testing (such as unit testing, integration testing, and qualification testing), and target-based verification.

[0030] In a further preferred embodiment, step 2.1 involves the following semantic similarity retrieval process:

[0031] (1) Use the preset review task objectives as the natural language query text. ,in To query the text for the first Each word is input into a lightweight language model; the lightweight language model encodes the query text and outputs a fixed-dimensional query vector. ;

[0032] (2) Calculate the query vector in the pre-constructed vector space. With each candidate text vector in the normalized knowledge base The similarity between them; the pre-constructed vector space consists of several text vectors obtained by encoding the text in the normalized knowledge base in step 1;

[0033] (3) Based on all the calculated similarity scores, the candidate text vectors in the normalized knowledge base are sorted to obtain samples with similarity scores higher than the preset threshold, which constitute the initial screening candidate set.

[0034] A further preferred approach, in step 2.1, involves classifying the initial candidate set using predicted probability entropy as follows:

[0035] (1) For each candidate sample in the initial candidate set returned by semantic similarity retrieval, a probability distribution vector is obtained by classifying it using a lightweight language model. Where n is the preset number of categories, The model indicates that the sample belongs to the first... The probabilities of each category, and satisfying ;

[0036] (2) Calculate the probability distribution entropy value ;

[0037] (3) Based on the calculated entropy value Candidate samples are divided into two categories: high-confidence samples and low-confidence samples. When the entropy value... When the entropy value is lower than or equal to a preset entropy threshold, the candidate sample is determined to have high confidence. High-confidence samples are automatically classified by a lightweight language model and automatically segmented and stored in corresponding structured databases based on their semantic content and task relevance, including a general airborne database and a task-specific database. When the entropy value... When the entropy exceeds a preset threshold, the candidate sample is determined to be of low confidence. Low-confidence samples are considered suspicious samples and will be subject to manual review.

[0038] A further optimized solution, step 3, is as follows:

[0039] Step 3.1: Based on the current pre-set review task objectives, and based on the aviation domain knowledge system constructed in Step 2, obtain standardized entities, relationships, and attributes using the general large language model and Prompt project;

[0040] Step 3.2: Transform the standardized entities, relations and attributes obtained in Step 3.1 into a data model that the graph database can recognize, and store the structured data to form an initial knowledge graph;

[0041] Step 3.3: Enhance the initial knowledge graph established in Step 3.2 with knowledge semantics-driven and human-in-the-loop knowledge; including relational reasoning and filling, entity attribute expansion, knowledge fusion, pattern-guided reasoning sample generation, and expert-in-the-loop evaluation and correction process.

[0042] A further optimized solution, step 4, involves the following steps:

[0043] Step 4.1: Perform automated preprocessing on the airworthiness documents uploaded by users for review, including text extraction and format standardization, and document type identification;

[0044] Step 4.2: Perform knowledge-driven agent collaborative reasoning review on the airworthiness documents processed in Step 4.1, including:

[0045] (1) Based on the airworthiness document text information processed in step 4.1, and according to the retrieval enhancement generation method, search in the knowledge graph finally constructed in step 3 to obtain entities and relationships related to the current review text and construct the review context;

[0046] (2) Extract multiple special instructions related to the current review task from the agent instruction library. Based on the characteristics of the instructions, the agent routes the instructions to multiple large language models integrated within the agent. Each large language model is assigned a specific review sub-task.

[0047] (3) The agent injects the dynamic review context and predefined review rules constructed in step 4.2 into each scheduled large language model, making the agent a specialized agent with multiple parallel review capabilities; the specialized agent drives all the large language models inside it to perform parallel and multi-dimensional review analysis on the same text to be reviewed, and outputs multi-dimensional review results; finally, the multi-dimensional review results are integrated to obtain a comprehensive review result.

[0048] Step 4.3: Based on the comprehensive review results obtained in Step 4.2, generate a structured intelligent agent review report for domain experts to review and obtain the final review conclusion.

[0049] Beneficial effects:

[0050] Compared with the prior art, the significant advantages of this invention are:

[0051] 1) The efficiency and automation of the review are significantly improved. By constructing a review method of "dynamic active learning-intelligent agent review", this invention can perform preliminary screening and classification of a large number of airworthiness documents and conduct automated review of the documents to be reviewed. Compared with the traditional pure manual review method, it can greatly improve the review efficiency and greatly shorten the airworthiness review cycle.

[0052] 2) The review accuracy is relatively high. By constructing a general airborne database and a mission-specific database and building a corresponding knowledge graph, domain knowledge can be injected into the intelligent agent using RAG technology during the review. This can make up for the lack of domain knowledge in the intelligent agent and improve the applicability and accuracy of the general large language model in the airworthiness review task.

[0053] 3) Improved classification accuracy: Through dynamic active learning, expert review and verification, and incremental iteration of the model based on LORA, the accuracy and efficiency of classification can be greatly improved, forming a virtuous cycle of mutual promotion between "data and model".

[0054] 4) Improved data quality: Through knowledge semantics-driven and human-in-the-loop knowledge enhancement, the problem of scarcity of high-quality data in the domain can be alleviated, and the cost of data construction can be reduced.

[0055] 5) Saves computing resources. This invention uses a lightweight language model for classification tasks and an intelligent agent composed of a general-purpose large language model for review tasks. This avoids the huge computing overhead caused by fine-tuning the general-purpose large language model, gives full play to the advantages of language models with different parameter amounts, greatly reduces computing power requirements and implementation costs, and makes the technical solution highly practical and worth promoting.

[0056] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0057] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0058] Figure 1 This is an overall architecture diagram of the airworthiness document review method based on dynamic active learning and knowledge graph intelligent agents provided in an embodiment of the present invention.

[0059] Figure 2 This is an architecture diagram of an airworthiness document intelligent agent review system for implementing step 4 of the method, provided in an embodiment of the present invention.

[0060] Figure 3 This is a user operation flowchart of the airworthiness document intelligent agent review system provided in an embodiment of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0062] Example 1: Combination Figure 1 This embodiment proposes an airworthiness document review method based on dynamic active learning and knowledge graph intelligent agents, including the following steps:

[0063] Step 1: Collect data from various multimodal heterogeneous data sources, fuse and intelligently preprocess the collected multimodal heterogeneous data to obtain a standardized knowledge base.

[0064] Step 1 aims to address the issues of weak knowledge foundation and difficulty in directly utilizing data in intelligent review systems due to the fragmented sources and inconsistent formats of knowledge in the aviation field. Through systematic data aggregation, collection, and cleaning, a high-quality, standardized knowledge foundation is built for subsequent processes. The specific process includes:

[0065] Step 1.1: Collect multimodal heterogeneous data sources, including aviation compliance standards library, historical project case library, aviation software defect and verification knowledge base, domain knowledge graph, and unstructured multimodal data.

[0066] As one example, the data source specifically includes:

[0067] Aviation compliance standards library: including full texts of international standards such as DO-178, DO-254, and ARP4761;

[0068] Historical project case library: including compliance verification reports, safety analysis reports, etc. of completed aviation projects;

[0069] Aviation Software Defects and Verification Knowledge Base: A database that records known software defects, test cases, and verification results;

[0070] Domain knowledge graph: The existing knowledge graph within the aviation field, serving as the initial source of knowledge;

[0071] Unstructured multimodal data: such as charts and images containing important conclusions, scanned copies of design documents, system logs, etc.

[0072] Step 1.2: Collect data from the multimodal heterogeneous data sources gathered in Step 1.1 using multiple channels, including:

[0073] Batch file upload and parsing: Supports PDF, DOCX, TXT and other formats, and uses tools such as Apache PDFBox for text parsing;

[0074] Standardized API data integration: Data is obtained by calling the management system's API;

[0075] Intelligent web crawling and parsing: For publicly available airworthiness regulations websites, use frameworks such as Scrapy to crawl and parse HTML content in a targeted manner;

[0076] Ultimately, data from different channels are aggregated to form a multimodal raw database.

[0077] Step 1.3: The multimodal raw database collected in Step 1.2 is fused and intelligently preprocessed, specifically including data normalization and format standardization, data cleaning and noise reduction, ultimately forming a standardized knowledge base with a unified format, specifically including:

[0078] Data normalization and format standardization: Convert all text data to UTF-8 encoding and convert all types of data to JSON format;

[0079] Data cleaning and noise reduction: This includes removing irrelevant special characters and text; this operation can effectively remove noise information from the original data, improve data quality, and avoid irrelevant information from interfering with subsequent operation steps, thereby ensuring the accuracy and reliability of intelligent review.

[0080] Ultimately, a high-quality, standardized knowledge base with a consistent format is output, providing a clean data foundation for subsequent steps.

[0081] The core function of Step 1 is to solve the problems of heterogeneous data sources, inconsistent formats, and varying quality faced in knowledge construction in the aviation field through systematic multimodal data fusion and intelligent preprocessing. By integrating multi-source data from standards, cases, defect and verification knowledge bases, knowledge graphs, and unstructured documents, it provides authoritative and rich knowledge sources for intelligent review and transforms them into a clean, well-organized, and high-quality standardized knowledge base. This solves the problems of model performance degradation and knowledge extraction errors caused by data quality issues, and provides a data foundation for the stable and efficient operation of the entire intelligent review process.

[0082] Step 2: Based on the standardized knowledge base obtained in Step 1, construct an aviation knowledge system using dynamic active learning and human-machine collaboration mechanisms; the aviation knowledge system includes a general airborne database and a mission-specific database related to the mission.

[0083] Step 2 aims to rapidly classify the vast amount of knowledge in the knowledge base, improving the efficiency of manual classification. Through efficient collaboration between models and experts, a high-quality domain knowledge base that can evolve automatically while ensuring authority is constructed. The specific process includes:

[0084] Step 2.1: Dynamically filter the standardized knowledge base obtained in Step 1 based on a lightweight language model; the model performs semantic similarity retrieval and initial screening on the standardized knowledge base obtained in Step 1 according to the preset review task objectives, and obtains an initial screening candidate set; using the lightweight language model, the initial screening candidate set is divided into three categories through the uncertainty measurement mechanism "predictive probability entropy": high-confidence airborne general database samples, high-confidence task-specific database samples related to the task, and doubtful samples judged by the model to have low confidence.

[0085] As an example, the lightweight language model mentioned in step 2.1 refers to a pre-trained language model with 100 million to 8 billion parameters. In this example, DeepSeek-R1-Distill-Qwen-1.5B is preferred. This parameter range can significantly reduce computational overhead and fine-tuning costs while ensuring sufficient semantic understanding capabilities, thus meeting the real-time requirements of dynamic iteration.

[0086] As an example, the preset review task objectives in step 2.1 specifically include: identifying knowledge related to "compliance verification", identifying knowledge related to "software review process", identifying knowledge related to "software grade", and identifying knowledge related to "verification method".

[0087] The "compliance verification" specifically includes: Declaration of Conformity (MOC0), Explanatory Documents (MOC1), Analytical Calculations (MOC2), Safety Assessment (MOC3), Laboratory Testing (MOC4), Ground Testing (MOC5), Flight Testing (MOC6), Simulator Testing (MOC8), and Equipment Conformity (MOC9).

[0088] The "software review process" specifically includes: Software Planning Phase Review (SOI1), Software Development Phase Review (SOI2), Software Verification Phase Review (SOI3), and Software Final Approval Review (SOI4).

[0089] Specifically, the "software level" includes: software levels A, B, C, D, and E;

[0090] Specifically, the "verification method" includes: review, analysis, testing (such as unit testing, integration testing, and qualification testing), and target-based verification.

[0091] The preset review task objective in this embodiment is to identify knowledge related to the "software review process", specifically including Software Planning Phase Review (SOI1).

[0092] As an example, the semantic similarity retrieval described in step 2.1 specifically includes matching the natural language query text with a standardized knowledge base at the deep semantic level. Its core lies in mapping textual information to a high-dimensional vector space using a lightweight language model, and quantifying semantic relevance within this space using geometric metrics. The specific implementation process includes the following steps:

[0093] (1) Query vectorization: First, the preset review task target is used as the natural language query text. ,in To query the text for the first Each word is input into the lightweight language model DeepSeek-R1-Distill-Qwen-1.5B. This model encodes the query text and outputs a fixed-dimensional, dense floating-point vector, called the query vector, denoted as . This vector is a numerical representation of the query text.

[0094] Specifically, the model encoding process includes:

[0095] a. Word embedding: By querying the word embedding matrix ,in It refers to the size of the vocabulary list. It is the hidden layer dimension of the model, which stores each word. Mapped to a low-dimensional word embedding vector Specifically, if The index in the vocabulary is ,but:

[0096]

[0097] This is equivalent to directly obtaining the word embedding from the word embedding matrix. A vector of columns, without the need to explicitly compute the encoding.

[0098] b. Positional encoding: To preserve the sequence order information, positional encoding is added to the word embedding vector. The input vector is obtained, where positional codes are calculated using sine and cosine functions:

[0099]

[0100] Among them, position coding The specific calculations are as follows: Indicates a dimension index. Indicates the position index:

[0101]

[0102]

[0103] c. Multi-head attention mechanism: For the first Layer Transformer ( The calculation of multi-head attention is as follows: First, the first... The input sequence obtained by the Transformer layer Generate query, key, and value matrices through linear transformations:

[0104]

[0105]

[0106]

[0107] in, These are the learnable parameter matrices. This is the dimension of each attention head; then, the scaled dot product attention is computed:

[0108]

[0109] Multi-head attention obtains its output by concatenating multiple attention heads and performing a linear transformation.

[0110]

[0111] Among them, each attention head ,in as well as Both are learnable parameter matrices, where h is the number of attention heads;

[0112] d. Feedforward Neural Network: The feature vector obtained by passing the output of the multi-head attention sublayer in step c through residual connections and layer normalization. As input, it undergoes a nonlinear transformation via a feedforward neural network; this network consists of two linear transformations and a ReLU activation function, calculated as follows:

[0113]

[0114] in, These are the weight matrices, These are the bias vectors, It is the hidden layer dimension of the feedforward network;

[0115] e. Normalization and Residual Connections: Each sublayer (multi-head attention and feedforward neural networks) is followed by residual connections and layer normalization, specifically...

[0116]

[0117] in, Presentation layer normalization operation, This represents a multi-head attention or feedforward neural network; This represents the input to a multi-head attention or feedforward neural network.

[0118] f. Pooling operation: After... After layer Transformer encoding, the output of the last layer is obtained. The entire query text is represented as a vector using mean pooling.

[0119]

[0120] Ultimately, what was obtained This is the semantic vector representation of the query text, used for subsequent semantic similarity retrieval.

[0121] (2) Vector similarity calculation: Subsequently, the query vector is calculated in the pre-constructed vector space. With each candidate text vector in the normalized knowledge base The similarity between them; the pre-constructed vector space consists of several text vectors obtained by encoding the text in the normalized knowledge base in step 1 according to the method in the above query vectorization. This embodiment uses cosine similarity as the metric, and its calculation formula is:

[0122]

[0123] in, Represents the dot product of vectors. and These are the L2 norms (i.e., magnitudes) of the vectors. Cosine similarity effectively measures the difference in direction between two vectors, and its value ranges from [-1, 1]. In this application, the closer the value is to 1, the higher the semantic similarity.

[0124] (3) Result sorting and screening: Based on all the calculated similarity scores, the candidate text vectors in the normalized knowledge base are sorted in descending order. Finally, samples with similarity scores higher than the preset threshold are selected to form the initial candidate set, which serves as the input for subsequent uncertainty measurement and classification.

[0125] As an example, step 2.1, which uses a lightweight language model to classify the initial candidate set through the uncertainty measurement mechanism "predictive probability entropy," includes the following specific steps:

[0126] (1) Model prediction: For each candidate sample in the initial candidate set returned by semantic similarity retrieval, the lightweight language model will perform a classification task and output a probability distribution vector. Where n is the preset number of categories, The model indicates that the sample belongs to the first... The probabilities of each category, and satisfying ;

[0127] (2) Entropy calculation: Calculate the probability distribution according to the definition of Shannon entropy in information theory. entropy value The calculation formula is as follows. In this embodiment, the logarithm log is base 2.

[0128]

[0129] (3) Sample partitioning: based on the calculated entropy value Candidate samples are divided into two categories: high-confidence samples and low-confidence samples. In this embodiment, the confidence threshold is preferably 0.5. When the entropy value... When the entropy value is lower than or equal to a preset entropy threshold, the candidate sample is determined to have high confidence. This type of sample is automatically classified by a lightweight language model and automatically segmented and stored in the corresponding structured database based on its semantic content and task relevance. This includes a general airborne database and a task-specific database. When the entropy value... When the entropy exceeds the preset threshold, the candidate sample is determined to be of low confidence. Such samples are considered suspicious and are pushed to the human-computer collaborative interaction interface for manual review by domain experts.

[0130] Step 2.2 involves manually reviewing and verifying the low-confidence, questionable samples identified by the model, injecting knowledge into the samples, and creating a golden training sample increment.

[0131] The manual review of low-confidence, questionable samples identified by the model specifically refers to adopting a priority scheduling strategy to review samples with the highest uncertainty entropy value first, so as to ensure that experts focus their efforts on the most perplexing samples and maximize calibration efficiency.

[0132] The knowledge injection refers to the process by which domain experts, through a human-computer interaction interface, provide accurate classification labels, compliance judgments, and necessary text annotations for low-confidence, questionable samples, transforming them into high-quality training samples.

[0133] Step 2.3: Based on the golden training samples obtained in Step 2.2, perform incremental iterative training of the lightweight language model based on LORA, update the parameters of the lightweight language model, and obtain the updated lightweight language model.

[0134] As an example, the specific process of incremental model iteration based on LORA using the golden training samples is as follows:

[0135] (1) Model freezing: Preserve the weight parameters of the original lightweight language model. No changes, no full parameter fine-tuning;

[0136] (2) Adapter Injection: In the Transformer layer of the lightweight language model, low-rank adapters are injected in parallel into the attention mechanism modules (such as Query, Key, Value, and Output projection layers). The adapter consists of two small matrices (reduced dimension matrix A and increased dimension matrix B), and its operation can be expressed as the following formula, where x is the input vector, and BA is the low-rank update of the original weights. The original weight parameters of the pre-trained model are h represents the output vector;

[0137]

[0138] (3) Efficient training: During incremental iterative training, only the injected LORA adapter parameters (matrices A and B) are optimized and updated, while the original model parameters are updated. Keep frozen, using the AdamW optimizer, with the learning rate set to... The batch size was set to 32, the LORA rank r to 8, and the training epochs to 20. This process significantly reduced training overhead and computational resource consumption while effectively avoiding catastrophic forgetting. The optimization objective of this process is determined by the cross-entropy loss function. Defined, this function measures the difference between the model's predicted distribution and the true label distribution, where, It is the number of samples in a training batch. It is the total number of task categories. It is the first Each sample in category The actual label value on the screen. The model is for the first Each sample in category The predicted probability value is:

[0139]

[0140] (4) Parameter merging: After training, the adapter parameters (BA) are merged with the original model parameters. The models are merged to form an updated model, thus introducing no additional latency during the inference phase.

[0141] Step 2.4: If there are still low-confidence suspicious samples that have not been manually reviewed and verified, then use the updated lightweight language model to classify them again through the uncertainty measurement mechanism "predicted probability entropy": high-confidence airborne general database samples, high-confidence task-specific database samples related to the task, and suspicious samples that the model judges to be low-confidence; then return to step 2.2 to continue execution until the loop termination condition is met.

[0142] As an example, the loop termination condition can be any of the following:

[0143] (1) All samples in the initial candidate set have been successfully classified;

[0144] (2) The number of iterations reaches the preset maximum number of iterations.

[0145] The core function of step 2 lies in achieving efficient and accurate classification of a large amount of domain knowledge through a closed-loop mechanism of dynamic active learning and human-machine collaboration. The model quickly classifies the samples it is confident in, which greatly reduces the burden on experts. At the same time, it can accurately identify the questionable samples that it cannot judge on its own, such as complex or novel ones. This allows limited human resources to be directed to where they are most needed. Experts inject knowledge into the questionable samples according to priority, forming a golden incremental training sample. Through dynamic active learning and incremental iteration of the model based on LORA, the accuracy and efficiency of classification can be greatly improved, forming a virtuous cycle of mutual promotion between "data and model". This avoids the computational overhead of full parameter fine-tuning and the risk of catastrophic forgetting, and achieves stable and efficient evolution of model performance. Finally, a general airborne database and a mission-specific database are constructed.

[0146] Step 3: Based on the current pre-set review task objectives and the aviation domain knowledge system constructed in Step 2, construct a knowledge graph and perform knowledge enhancement. The specific process includes:

[0147] Step 3.1: Based on the current pre-set review task objectives, and building upon the aviation domain knowledge system constructed in Step 2, standardized entities, relationships, and attributes are obtained using a general large language model and Prompt engineering. This step aims to dynamically and specifically construct a domain knowledge framework highly relevant to the review task objectives. Its core lies in the intelligent scheduling and fusion of knowledge sources based on the task objectives. The specific process includes:

[0148] (1) Task parsing and data scheduling: Receive the current preset review task target, retrieve and call samples related to the task target in basic concepts, principles or context from the general airborne database based on semantic similarity. These samples provide broad and supportive general knowledge. The samples retrieved from the general airborne database are fused with samples from the task-specific database to form a high-quality knowledge subset for the current task. The samples from the task-specific database provide in-depth and focused professional knowledge.

[0149] (2) Information extraction based on prompts: The above-mentioned fused knowledge subset is input into a general large language model. In this embodiment, the general large language model is DeepSeek-V3.1. The general large language model is driven to perform information extraction by setting prompts containing information extraction requirements. The "entity extraction prompt" is used to extract the entities and their attributes related to the task from the text of the knowledge subset. The "relation extraction prompt" is used to identify the semantic relationships between the above entities.

[0150] (3) Entity alignment and knowledge fusion: In order to solve the problem of diverse representations of the same entity in texts from different sources, the “Entity Alignment Prompt” is used to guide the general large language model to identify multiple entity representations pointing to the same object in the real world. Synonymous entities identified by the general large language model are merged and integrated to form a unique and standardized entity node. All related semantic relationships are merged to ensure the simplicity and consistency of the knowledge graph constructed for the current task.

[0151] Step 3.2: Graph Data Mapping and Structured Data Storage: The extracted and aligned normalized entities, relations, and attributes output in Step 3.1 are transformed into a data model recognizable by the graph database, and structured data is stored to form an initial, queryable knowledge graph; the specific process includes:

[0152] (1) Graph data model conversion: The normalized entities, relations and attributes output in step 3.1 after extraction and alignment are mapped to an attribute graph model, where entities are converted into nodes in the graph database, each node is assigned a unique ID and its attributes (e.g., entity name, type, description), and relations are converted into directed edges connecting two nodes, each edge is also assigned its relation type and attributes (e.g., confidence, source text).

[0153] (2) Knowledge graph data storage: Using the Cypher query language of Neo4j graph database (e.g., CREATE (n:Entity {id: $id, name: $name, type: $type})), the above nodes and relationships are imported into the database in batches. During this process, the database automatically maintains indexes to ensure efficient subsequent queries based on entity type, attribute or topology.

[0154] (3) Forming an initial knowledge graph: After the data import is completed, an initial domain knowledge graph is formed in Neo4j.

[0155] Step 3.3: Perform knowledge enhancement based on knowledge semantics and human-in-the-loop principles on the initial knowledge graph established in Step 3.2. This step is an iterative optimization process aimed at addressing potential knowledge gaps, incompleteness, or undiscovered implicit relationships in the initial knowledge graph. It combines the semantic reasoning capabilities of a general-purpose large language model with the judgment of domain experts to achieve continuous enhancement and correction of the knowledge graph. The specific process includes:

[0156] (1) Relational reasoning and completion: Based on the initial knowledge graph structure, relational reasoning is performed using a general large language model. By designing reasoning prompts such as "Given entity A and entity C, infer what kind of relationship might exist between them", the general large language model is driven to generate a series of candidate, unverified new relational triples, enriching the connectivity and logical depth of the knowledge graph. Relational reasoning and completion can discover and complete the missing direct or indirect relational links between entities, enriching the connectivity of the knowledge graph.

[0157] (2) Entity attribute expansion: Based on the existing entity nodes and their existing attributes and relationships in the knowledge graph, deep semantic analysis and reasoning are performed using a general large language model. By designing an attribute expansion Prompt, the model is driven to infer and generate additional attributes that the entity may have based on the entity's existing attributes and relationships (for example, inferring attributes such as "rated flow rate" and "working pressure" for "fuel pump"), generating a candidate set of new attributes as suggestions for expanding the knowledge graph;

[0158] (3) Knowledge Fusion: For a series of candidate new relation triples to be verified and candidate new attribute sets, the semantic understanding and reasoning capabilities of the general large language model are used to conduct in-depth internal review and fusion. By designing knowledge fusion prompts (e.g., merge nodes with the same function in the generated entity nodes, reason about the merged entity relations, and re-form edges), the model is driven to automatically identify and handle cases of internal conflict resolution, internal redundancy merging, and knowledge association strengthening;

[0159] (4) Pattern-guided inference sample generation: Using a general large language model, based on existing entity and relation patterns in the knowledge graph, high-quality data samples that conform to domain semantics and logic are generated. For example, based on the existing relationship of "component-with fault mode-fault mode", new and reasonable (component, with fault mode, fault mode) triples are generated to achieve data augmentation in sparse areas. Specifically, based on the existing example of "(fuel pump, with fault mode, outlet pressure too low)", by designing inference prompts such as "refer to this example, generate reasonable fault modes for 'hydraulic pump'", the model can be guided to generate reasonable new samples such as "(hydraulic pump, with fault mode, internal leakage exceeds limit)".

[0160] (5) Human Expert-In-Loop Evaluation and Correction: The results of the above automated processes (such as new relationships after knowledge fusion and newly generated triples) are pushed to the human-machine collaborative interaction interface for review and verification by domain experts. Experts confirm, correct, or reject the results recommended by the system and formally inject the approved knowledge into the knowledge graph, combining human domain authority with machine computational efficiency to form a closed loop of "human-in-the-loop" intelligent enhancement. This "human-in-the-loop" mechanism ensures the accuracy and reliability of the knowledge enhancement process.

[0161] The core function of step 3 is to provide intelligent agents with censorship contextual data sources through knowledge graph construction and knowledge enhancement. Entity recognition, relationship recognition, and entity alignment are achieved using a general-purpose large language model and the Prompt project on both general-purpose airborne databases and mission-specific databases, thus constructing a semantic knowledge graph. Through knowledge-driven semantics and human-in-the-loop knowledge enhancement mechanisms, the problem of high-quality data scarcity in the domain is alleviated, and the cost of data construction is reduced, enabling continuous optimization and expansion of the knowledge graph.

[0162] Step 4: Based on the Retrieval Enhanced Generation (RAG) method combined with the knowledge-enhanced knowledge graph obtained in Step 3, a review context is dynamically constructed and injected into an intelligent agent for intelligent data review and report generation. Specifically, this includes:

[0163] Step 4.1: Data Input and Preprocessing: Users upload airworthiness documents to be reviewed (such as Software Planning Phase Review SOI1). Automated preprocessing is performed on the uploaded airworthiness documents, specifically including:

[0164] (1) Text extraction and format standardization: Parse the document format, extract the plain text content, and convert it into a standardized JSON data structure.

[0165] (2) Intelligent document type recognition: Based on the text content, the document type (such as the Software Planning Phase Review SOI1 report) is automatically identified using a large language model, providing contextual basis for subsequent targeted reviews.

[0166] In this embodiment, the large language model in step 4 all adopts the DeepSeek-V3.1 model.

[0167] Step 4.2: Perform knowledge-driven agent-based collaborative reasoning review on the airworthiness documents processed in Step 4.1. This involves enhancing the reasoning review through RAG retrieval and agent collaboration, specifically including:

[0168] (1) Dynamic review context construction: Based on the airworthiness document text information processed in step 4.1, the RAG technology is used to search in the Neo4j knowledge graph finally constructed in step 3 to obtain entities and relationships related to the current review text and construct a rich and structured review context.

[0169] (2) Agent Task Scheduling and Routing: Multiple specific instructions related to the current review task are extracted from the agent's instruction library. Each instruction includes fields such as "role, background, introduction, skills, goal, constraints, output format, and workflow." Based on the characteristics of the instructions, the agent routes them to multiple large language models integrated within the agent. Each large language model is assigned a specific review sub-task. For example, in the SOI1 task review during the software planning phase, the specific review instructions for the software overview section (as shown in Table 1) are routed to a corresponding large language model for processing.

[0170] Table 1: Specific Review Instructions for the Software Overview Section

[0171]

[0172] (3) Collaborative review by specialized intelligent agents: The intelligent agent injects the dynamic review context and predefined review rules (as shown in Table 2) constructed in step 4.2 into each scheduled large language model. Through this process, the intelligent agent becomes a specialized intelligent agent with multiple parallel review capabilities. The specialized intelligent agent drives all the large language models inside it to perform parallel and multi-dimensional review analysis on the same text to be reviewed, and outputs multi-dimensional review results. Finally, the multi-dimensional review results are integrated to obtain a comprehensive review result.

[0173] Table 2: Predefined Review Rules

[0174]

[0175] Step 4.3: Report Generation and Final Review Conclusion. Based on the comprehensive review results obtained in Step 4.2, a structured agent review report (as shown in Table 3) is automatically generated, and a download link is provided. Finally, the report is submitted to domain experts for review. The experts make final confirmations or corrections to the agent's review conclusions, forming an authoritative final review conclusion and completing the entire automated review process.

[0176] Table 3: Agent Review Report

[0177]

[0178] As an example, the RAG technology in step 4.2 specifically includes:

[0179] (1) Query vectorization:

[0180] The text fragment G to be reviewed, processed in step 4.1, is used as the query, and the Deepseek-V3.1 model is used to convert it into a high-dimensional vector representation. Where E is the text embedding model Deepseek-V3.1, It is Dense vectors of dimension:

[0181]

[0182] (2) Semantic similarity retrieval:

[0183] In the Neo4j knowledge graph finally constructed in step 3, vector similarity is first calculated to calculate the query vector. With entity and relation vectors in a knowledge graph The cosine similarity between them, where d is the vector dimension. and These represent the query vectors respectively. sum vector In the The similarity score is calculated precisely in terms of direction between two vectors using the following formula, with a value range of [-1, 1]. In this embodiment, a positive value is typically used, and the closer the value is to 1, the higher the semantic similarity. Based on the calculated similarity score, the system sorts all candidate knowledge vectors in descending order of their similarity scores and selects those with similarity scores higher than a preset threshold. Candidate result = 0.7;

[0184]

[0185] Next, graph-structured retrieval was performed based on the Neo4j knowledge graph, including entity center queries, multi-hop relationship queries, and path analysis queries; and the results of semantic vector retrieval and graph-structured retrieval were merged.

[0186]

[0187] in =0.5 is the fusion weight coefficient. The results are based on vector similarity. Graph-structured search results;

[0188] (3) Dynamic context construction:

[0189] Based on the aforementioned fusion search results We will construct an enhanced context for review tasks. The specific construction process is as follows:

[0190]

[0191] in: Indicates splicing; =“The review was conducted based on the following airworthiness knowledge:”; The retrieved knowledge fragments include entity attributes and relationship information; =“Content to be reviewed:”; The text fragment to be reviewed; the knowledge fragment Use a uniform format to represent:

[0192]

[0193] The resulting dynamic review context P will serve as the direct input to the specialized intelligent agent, providing it with accurate and structured domain knowledge support, effectively enhancing the accuracy and reliability of review reasoning.

[0194] The core function of step 4 is to combine the knowledge system built in the preceding steps with intelligent agent technology to achieve intelligent review of airworthiness documents, thereby improving the efficiency and accuracy of manual review. By injecting domain knowledge into the intelligent agent using RAG technology during the review process, the shortcomings of the intelligent agent in lacking domain knowledge can be compensated for, improving the applicability and accuracy of the general large language model in airworthiness review tasks. At the same time, it can avoid the computational overhead caused by full parameter fine-tuning of the general large language model, greatly reducing the computing power requirements and implementation costs, making the technical solution highly practical and worthy of promotion.

[0195] Example 2:

[0196] like Figure 2 As shown, this embodiment provides an airworthiness document review system based on dynamic active learning and knowledge graph intelligent agents. The system is the entity through which users perform document review operations, and its function directly corresponds to step 4 of the method. The system includes:

[0197] (1) User Interaction Layer: This layer serves as the interface for interaction between the system and the user, primarily including front-end user interaction services, specifically:

[0198] Web-based web page services, client application services, and API interface services are used to receive airworthiness documents uploaded by users for review, provide diverse system access points, and display processing results; this layer is the direct interface for implementing data input in step 4.1 and report generation and download in step 4.3.

[0199] The expert review module corresponds to the review conclusion in step 4.3. It uses domain experts to make final confirmation or correction to the review conclusion of the intelligent agent based on their own experience, forming the final review conclusion and completing the entire automated review process.

[0200] (2) Document review engine: As the core processing engine of the system, it includes:

[0201] The document preprocessing module corresponds to the data input and preprocessing in step 4.1. The system performs automated preprocessing on user-uploaded documents, specifically including:

[0202] a. Text Extraction: Parse the document format and extract the plain text content.

[0203] b. Format standardization: The extracted plain text content is uniformly converted into a standardized JSON data structure.

[0204] c. Intelligent document type recognition: Based on text content, the Deepseek-V3.1 model is used to automatically identify the document type (such as the Software Planning Phase Review SOI1 report), providing contextual basis for subsequent targeted reviews.

[0205] The document review module, corresponding to the knowledge-driven agent collaborative reasoning review in step 4.2 and the report generation in step 4.3, achieves reasoning review and generates a review report through RAG retrieval enhancement and agent collaboration. Specifically, it includes:

[0206] a. Dynamic review context construction: Based on the text information processed in step 4.1, the RAG technology is used to retrieve entities and relationships related to the current review text from the Neo4j knowledge graph constructed in step 3, and to construct a rich and structured review context.

[0207] b. Agent task scheduling and routing: Extract multiple specific instructions related to the current review task from the agent instruction library. Based on the characteristics of the instructions, the agent routes the instructions to multiple large language models integrated within the agent. Each large language model is assigned a specific review sub-task.

[0208] c. Collaborative Review by Specialized Intelligent Agents: The intelligent agent injects the dynamic review context and predefined review rules constructed in step 4.2 into each scheduled large language model. Through this process, the intelligent agent becomes a specialized intelligent agent with multiple parallel review capabilities. The specialized intelligent agent drives all the large language models within it to perform parallel and multi-dimensional review analysis on the same text to be reviewed, and outputs multi-dimensional review results. Finally, the multi-dimensional review results are fused together.

[0209] d. Based on the comprehensive review results, the system automatically generates a structured intelligent agent review report and provides a download link for the report.

[0210] (3) Technical Support Layer: Provides the necessary underlying capabilities to the user interaction layer and document review engine when executing the review process in step 4, specifically including:

[0211] The large model technology stack provides core model reasoning capabilities for intelligent document type recognition in step 4.1 and collaborative review of specialized intelligent agents in step 4.2;

[0212] The rule base and knowledge graph, as the results of the previous process, provide real-time and accurate knowledge retrieval sources for the construction of dynamic review context in step 4.2;

[0213] GPU computing servers provide computing power support for the large model inference involved in the intelligent document type recognition in step 4.1 and the collaborative review of specialized intelligent agents in step 4.2;

[0214] Storage services are used to persistently store the documents to be reviewed uploaded by the user in step 4.1, the review report generated in step 4.3, and related process data.

[0215] The system's various levels and modules work together, receiving input through the user interaction layer, executing an intelligent review process via the document review engine, and relying on the technical support layer to provide capability support, ultimately completing the intelligent review of airworthiness documents.

[0216] Example 3:

[0217] like Figure 3 As shown in the figure, this embodiment provides an operation flowchart for an airworthiness document review system based on dynamic active learning and knowledge graph intelligent agents, which specifically includes the following steps:

[0218] Step S1: The user uploads the airworthiness documents to be reviewed (such as Software Planning Phase Review SOI1) through the system interface.

[0219] Step S2: After receiving the document, the system initiates automated preprocessing, which includes extracting text content, standardizing the format, and intelligently identifying the document type (e.g., identifying it as a Software Planning Phase Review SOI1 report), laying the data foundation for subsequent targeted reviews.

[0220] Step S3: Based on the text information processed in step S2, the system uses Retrieval Enhanced Generation (RAG) technology to retrieve entities and relationships highly relevant to the current censorship text from a pre-built domain knowledge graph, thus constructing a rich and structured censorship context.

[0221] Step S4: The system extracts multiple specific instructions related to the current review task from the agent's instruction library. Based on the characteristics of the instructions, the agent routes the instructions to multiple large language models integrated within the agent. The review context and predefined review rules constructed in step S3 are also injected into the agent.

[0222] Step S5: After receiving the review context, instructions and rules, the agent becomes a specialized agent with multiple parallel review capabilities; the specialized agent drives all its internal large language models to perform parallel, multi-dimensional review analysis on the same text to be reviewed, and outputs multi-dimensional review results.

[0223] Step S6: The system integrates the multi-dimensional review results output by the special intelligent agent in step S5, automatically generates a structured intelligent agent review report, and provides the user with a download link for the report.

[0224] Step S7: Domain experts manually review the agent review report generated by the system, and combine their own experience to make final confirmation or correction to the agent review conclusion, forming the final review conclusion and completing the entire automated review process.

[0225] Finally, it should be noted that the above embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the technical scope disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for reviewing airworthiness documents based on dynamic active learning and knowledge graph intelligent agents, characterized in that: Includes the following steps: Step 1: Collect data from various multimodal heterogeneous data sources, fuse and intelligently preprocess the collected multimodal heterogeneous data to obtain a standardized knowledge base; Step 2: Based on the standardized knowledge base obtained in Step 1, construct an aviation knowledge system using dynamic active learning and human-machine collaboration mechanisms. The aviation knowledge system includes a general airborne database and a mission-specific database related to the mission. The construction process specifically includes: Step 2.1: Using a lightweight language model, based on the preset review task objectives, perform semantic similarity retrieval and initial screening on the standardized knowledge base obtained in Step 1 to obtain an initial screening candidate set; using a lightweight language model, divide the initial screening candidate set into three categories by predicting probability entropy: high-confidence airborne general database samples, high-confidence task-specific database samples related to the task, and doubtful samples judged by the model to have low confidence. Step 2.2: Manually review and verify the low-confidence suspicious samples identified by the model, inject knowledge into the samples, and form a golden training sample increment. Step 2.3: Based on the golden training samples obtained in Step 2.2, perform incremental iterative training of the lightweight language model based on LORA, update the parameters of the lightweight language model, and obtain the updated lightweight language model. Step 2.4: If there are still low-confidence suspicious samples that have not been manually reviewed and verified, then use the updated lightweight language model to classify them again by predicting the probability entropy; then return to step 2.2 to continue execution until the loop termination condition is met. Step 3: Based on the current pre-set review task objectives, construct a knowledge graph and perform knowledge enhancement based on the aviation knowledge system built in Step 2; Step 4: Based on the retrieval enhancement generation method and the knowledge-enhanced knowledge graph obtained in Step 3, dynamically construct the review context and inject it into the intelligent agent to perform intelligent data review and report generation.

2. The airworthiness document review method based on dynamic active learning and knowledge graph intelligent agents according to claim 1, characterized in that: The specific process of step 1 is as follows: Step 1.1: Collect multimodal heterogeneous data sources, including aviation compliance standards library, historical project case library, aviation software defect and verification knowledge base, domain knowledge graph, and unstructured multimodal data; Step 1.2: Collect the multimodal heterogeneous data sources gathered in Step 1.1 using multiple channels, including batch file upload and parsing, standardized API data integration, and intelligent web crawling and parsing, to finally form a multimodal raw database; Step 1.3: The multimodal raw database collected in Step 1.2 is fused and intelligently preprocessed, including data normalization and format standardization, data cleaning and noise reduction, to form a standardized knowledge base with a unified format.

3. The airworthiness document review method based on dynamic active learning and knowledge graph intelligent agents according to claim 2, characterized in that: The aviation compliance standards library mentioned in step 1.1 includes international standards DO-178, DO-254, and ARP4761; the domain knowledge graph refers to existing knowledge graphs in the aviation field; the unstructured multimodal data includes charts and images, scanned copies of design documents, and system logs.

4. The airworthiness document review method based on dynamic active learning and knowledge graph intelligent agents according to claim 1, characterized in that: In step 2.1, the preset review task objectives specifically include: identifying knowledge related to "compliance verification", identifying knowledge related to "software review process", identifying knowledge related to "software grade", and identifying knowledge related to "verification method". The "compliance verification" specifically includes: Declaration of Conformity (MOC0), Explanatory Documents (MOC1), Analytical Calculations (MOC2), Safety Assessment (MOC3), Laboratory Testing (MOC4), Ground Testing (MOC5), Flight Testing (MOC6), Simulator Testing (MOC8), and Equipment Conformity (MOC9). The "software review process" specifically includes: Software Planning Phase Review (SOI1), Software Development Phase Review (SOI2), Software Verification Phase Review (SOI3), and Software Final Approval Review (SOI4). Specifically, the "software level" includes: software levels A, B, C, D, and E; Specifically, the "verification method" includes: review, analysis, testing, and target-based verification.

5. The airworthiness document review method based on dynamic active learning and knowledge graph intelligent agents according to claim 1, characterized in that: In step 2.1, the semantic similarity retrieval process is as follows: (1) Use the preset review task objectives as the natural language query text. ,in To query the text for the first Each word is input into a lightweight language model; the lightweight language model encodes the query text and outputs a fixed-dimensional query vector. ; (2) Calculate the query vector in the pre-constructed vector space. With each candidate text vector in the normalized knowledge base The similarity between them; the pre-constructed vector space consists of several text vectors obtained by encoding the text in the normalized knowledge base in step 1; (3) Based on all the calculated similarity scores, the candidate text vectors in the normalized knowledge base are sorted to obtain samples with similarity scores higher than the preset threshold, which constitute the initial screening candidate set.

6. The airworthiness document review method based on dynamic active learning and knowledge graph intelligent agents according to claim 1, characterized in that: In step 2.1, the process of classifying the initial candidate set by predicting probability entropy is as follows: (1) For each candidate sample in the initial candidate set returned by semantic similarity retrieval, a probability distribution vector is obtained by classifying it using a lightweight language model. Where n is the preset number of categories, The model indicates that the sample belongs to the first... The probabilities of each category, and satisfying ; (2) Calculate the probability distribution entropy value ; (3) Based on the calculated entropy value Candidate samples are divided into two categories: high-confidence samples and low-confidence samples. When the entropy value... When the entropy value is lower than or equal to a preset entropy threshold, the candidate sample is determined to have high confidence. High-confidence samples are automatically classified by a lightweight language model and automatically segmented and stored in corresponding structured databases based on their semantic content and task relevance, including a general airborne database and a task-specific database. When the entropy value... When the entropy exceeds a preset threshold, the candidate sample is determined to be of low confidence. Low-confidence samples are considered suspicious samples and will be subject to manual review.

7. The airworthiness document review method based on dynamic active learning and knowledge graph intelligent agents according to claim 1, characterized in that: The specific process of step 3 is as follows: Step 3.1: Based on the current pre-set review task objectives, and based on the aviation domain knowledge system constructed in Step 2, obtain standardized entities, relationships, and attributes using the general large language model and Prompt project; Step 3.2: Transform the standardized entities, relations and attributes obtained in Step 3.1 into a data model that the graph database can recognize, and store the structured data to form an initial knowledge graph; Step 3.3: Enhance the initial knowledge graph established in Step 3.2 based on knowledge semantics and human-in-the-loop knowledge; It includes relational reasoning and filling, entity attribute expansion, knowledge fusion, pattern-guided reasoning sample generation, and expert-in-the-loop evaluation and correction processes.

8. The airworthiness document review method based on dynamic active learning and knowledge graph intelligent agents according to claim 1, characterized in that: The specific process of step 4 is as follows: Step 4.1: Perform automated preprocessing on the airworthiness documents uploaded by users for review, including text extraction and format standardization, and document type identification; Step 4.2: Perform knowledge-driven agent collaborative reasoning review on the airworthiness documents processed in Step 4.1, including: (1) Based on the airworthiness document text information processed in step 4.1, and according to the retrieval enhancement generation method, search in the knowledge graph finally constructed in step 3 to obtain entities and relationships related to the current review text and construct the review context; (2) Extract multiple special instructions related to the current review task from the agent instruction library. Based on the characteristics of the instructions, the agent routes the instructions to multiple large language models integrated within the agent. Each large language model is assigned a specific review sub-task. (3) The agent injects the dynamic review context and predefined review rules constructed in step 4.2 into each scheduled large language model, making the agent a specialized agent with multiple parallel review capabilities; the specialized agent drives all the large language models inside it to perform parallel and multi-dimensional review analysis on the same text to be reviewed, and outputs multi-dimensional review results; finally, the multi-dimensional review results are integrated to obtain a comprehensive review result. Step 4.3: Based on the comprehensive review results obtained in Step 4.2, generate a structured intelligent agent review report for domain experts to review and obtain the final review conclusion.

Images