Aero-engine fault data processing method and system

Abstract

The present application relates to the technical field of aero-engine fault data isolation processing, and in particular to an aero-engine fault data processing method and system which can achieve a good balance between time efficiency and generation quality and has good versatility. First, the original data is preprocessed through the MAW workflow, and then the fine semantic segmentation of the text and the efficient vectorization index construction are completed through the domain adaptive large language model and the dynamic code execution mechanism. Second, the two-stage progressive retrieval enhancement generation architecture TSP-RAG is proposed. At the same time, the adaptive HNSW retrieval module dynamically adjusts the search breadth according to the obtained ef weight, performs vector retrieval, and removes the results at the content level. Finally, the nodes retrieved by the FIL-knowledge graph and the HNSW retrieval results are semantically fused and reordered, and are input into the large language model together with the reasoning tree to generate answers. Compared with the prior art, the present application has the advantages of high time efficiency, high generation quality, and good versatility.

Description

Technical Field

[0001] This invention relates to the field of aircraft engine fault data isolation and processing technology, specifically to an aircraft engine fault data processing method and system that achieves a good balance between time efficiency and generation quality and has good versatility. Background Technology

[0002] As the core power unit of an aircraft, the performance of an aero-engine directly affects flight safety and operational efficiency. When an engine malfunctions or fails, it is crucial to quickly deduce and analyze the source, cause, and corresponding handling measures based on existing information—this process is known as aero-engine fault isolation. Traditionally, maintenance personnel manually consult aircraft manuals or historical maintenance records to retrieve and integrate relevant information and formulate response plans. The efficiency of this process is highly dependent on the organizational structure of the manuals and records, as well as the professional knowledge and experience of the maintenance personnel. However, with the increasing complexity of aviation system structures and functions, related technical documents and maintenance data are exploding in size. On the one hand, the heterogeneity and scale of the data significantly increase the difficulty of structured processing; on the other hand, relying on manual retrieval and integration of relevant information from massive databases is not only inefficient but also makes it difficult to guarantee the quality of the integrated results. This means that the traditional manual approach can no longer meet the high reliability and timeliness requirements of modern aviation operations and maintenance for fault response. Therefore, building a dedicated information system for aero-engine fault isolation, capable of effectively processing information from various sources and efficiently retrieving it when needed, is crucial for current aero-engine fault isolation tasks.

[0003] In recent years, artificial intelligence technology, especially large language models, has made breakthroughs in reasoning capabilities, making it possible for large language models to drive intelligent agent information systems for aero-engine fault isolation to replace manual labor in performing tedious information processing and retrieval tasks, thereby improving fault isolation efficiency. However, the construction of aero-engine fault isolation information systems faces two fundamental challenges: First, the integration and processing of multi-source heterogeneous information is difficult. Aero-engines have extremely complex structures, and the original information such as fault isolation manuals, logs, and records for different aircraft models covers unstructured text, semi-structured fault isolation procedures, and structured tables, resulting in diverse data formats. In addition, the original files contain a large amount of repetitive and invalid information, making knowledge management extremely difficult. Second, there is the challenge of deep semantic understanding and reasoning in highly coupled systems. Aero-engine systems are highly coupled, with related knowledge widely distributed across heterogeneous and dispersed data sources; coupled with numerous technical terms and complex combinations of terms, professional information retrieval faces severe challenges. Traditional single-retrieval augmented generation (RAG) or knowledge graph (knowledge graph) technologies typically rely solely on vector similarity or specific graph patterns for reasoning, making it difficult to achieve comprehensive and accurate information retrieval. Furthermore, existing knowledge graph-RAG frameworks are mostly limited to a single graph, lacking flexibility and struggling to capture subtle semantics hidden within text. This limitation weakens the fusion effect of vector retrieval and graph retrieval, causing the system to be unable to effectively balance the needs of technical terminology entity matching and logical multi-hop reasoning. Summary of the Invention

[0004] This invention addresses the challenges of processing multi-source heterogeneous data in existing aero-engine fault isolation methods by proposing a versatile aero-engine fault data processing method and system that achieves a good balance between time efficiency and generation quality.

[0005] This invention achieves its purpose through the following measures:

[0006] A method for processing aero-engine fault data is characterized by the following steps: First, the raw data is preprocessed using a MAW workflow. Then, a domain-adaptive large language model and dynamic code execution mechanism are used to complete fine semantic segmentation and efficient vectorized index construction of the text. Second, a two-stage progressive retrieval enhancement generation architecture TSP-RAG is proposed. The first stage is performed by FER-knowledge graph: based on the user question and its extracted keywords, keyword matching is performed first, followed by vector retrieval in the knowledge graph to determine the relevant task scope, and the ef weight parameters for HNSW index tuning are calculated based on the similarity density. The second stage is jointly completed by FIL-knowledge graph and adaptive HNSW retrieval module: FIL-knowledge graph performs precise retrieval using the task scope defined in the first stage. Part of the results are directly output as candidate answers, and the other part is used to construct a hierarchical reasoning path through multi-hop relationship expansion and organized into a large language model-friendly trie structure, i.e., a reasoning tree. At the same time, the adaptive HNSW retrieval module uses the obtained ef weights to perform the second stage. The search breadth is dynamically adjusted by weights, vector retrieval is performed, and the results are deduplicated at the content level. Finally, the nodes retrieved by the FIL-knowledge graph and the HNSW retrieval results are semantically fused and reordered, and then fed into the large language model along with the reasoning tree to generate a response.

[0007] Before preprocessing the raw data, this invention requires multi-source data conversion. First, through format parsing and recognition technology, discrete and inconsistent raw data are mapped into a unified standardized intermediate representation. Then, multi-dimensional automated data cleaning is performed. After data cleaning, the MD file is transferred to the data processing agent, which adopts an adaptive segmentation strategy: for short texts, semantic segmentation is performed directly using prompting engineering; for long texts, a structured processing script is generated by the model and executed in a controlled environment to achieve efficient segmentation. Finally, the model performs recursive verification to ensure reasonable segmentation granularity. The processed data and metadata are persistently stored in a high-performance relational database to support subsequent vectorized index construction and efficient retrieval. The segmented data vectorization is completed through a word embedding model.

[0008] The word embedding model described in this invention uses the Qwen3-word embedding-4B model as the base model. Considering hardware conditions and the number of samples, a LoRA-based fine-tuning strategy is adopted for domain adaptation. A multi-negative example contrastive learning format is used to fine-tune the embedding model. The training data adopts a multi-negative example format [V2][-V2], where each training sample consists of a query, a positive sample pos (i.e., a semantically highly relevant fault isolation description), and three hard negative samples neg (i.e., interference content from other irrelevant tasks). A total of 5,500 high-quality samples are constructed, divided into a 90% training set and a 10% test set. The training process is optimized using the InfoNCE loss function, the formula of which is as follows:

[0009] (1),

[0010] in, For the question, As a positive sample, For the i-th negative sample, For similarity function, The number of negative samples. This is the temperature coefficient.

[0011] In this invention, during the knowledge graph construction process, a large language model-assisted entity and relation extraction technique is employed. The Qwen3-8B model is selected as the basic extraction model, and the LoRA fine-tuning strategy is used to adapt the model to the domain. The training data adopts the Alpaca format, meaning that each sample contains a clear instruction, input context, and expected output to effectively guide the model to learn complex semantic relation extraction logic: {"instruction":"...", "input":"...", "output":"..."}. The dataset is constructed using the knowledge distillation method: a large language model with a large number of parameters and strong reasoning ability is used as the teacher model. Finally, a distillation dataset containing 5,500 samples is constructed, divided into a 90% training set and a 10% test set. The fine-tuning task adopts the Causal Language Modeling Loss method.

[0012] (2),

[0013] in, For the first indivual , For the first One goal , Represents the original model parameters Plus Introduced low-rank update , This indicates that only the answer / output part is considered. Loss was calculated. An independent validation set consisting of 200 long text paragraphs and their corresponding triples was also constructed for fine-grained performance analysis and generalization ability verification. The specific verification method is as follows: The F1 score was calculated by combining the precision (number of correctly predicted triples / total number of predicted triples) and recall (number of correctly predicted triples / total number of labeled triples). The F1 score formula is as follows:

[0014] (3),

[0015] Table 4: Final fine-tuning results and test results. The fine-tuned model achieved a performance improvement of 4.45%.

[0016] In the FER-knowledge graph retrieval stage of the two-stage progressive retrieval enhancement generation architecture TSP-RAG described in this invention, a frequency-aware retrieval strategy is executed. This strategy dynamically switches between exact matching and semantic approximate matching searches based on entity popularity. That is, based on the number of nodes that are exactly matched, if the number of nodes is low, it is a low-frequency entity, and the exact matching results are retained. At the same time, keyword vector retrieval is performed, and entity-level and sentence-level vector retrieval are fused. If it is a high-frequency entity, exact matching and keyword vector retrieval are discarded to avoid noise, and retrieval is performed only based on relational semantics and sentence semantics. Diversity-driven ef weights are introduced to achieve semantically adaptive HNSW tuning. Finally, the function outputs the average ef weight and the deduplicated task list. The former is used to optimize the breadth of vector retrieval, and the latter is passed to the second stage FIL-knowledge graph retrieval as a task constraint.

[0017] In this invention, the FER-knowledge graph structure adopts standard triples. Nodes also have Task and Number and Type attributes, as well as 2560-dimensional vector attributes. Relationship edges also have 2560-dimensional vector attributes. The FER-knowledge graph extracts entity relationships in the format [Entity 1, Entity 1 type, Relationship, Entity 2, Entity 2 type]. The original text includes the following entity types: fault / phenomenon, system / component, task / program, manual / document, status / indication, location, action, and maintenance information. The type is stored as the entity's type attribute. After extraction, each entity and relationship is vectorized and stored as a vector attribute. The FER-knowledge graph performs retrieval based on the user-input question and keywords in the question. Based on the density of the cosine similarity values ​​of the retrieval results, it dynamically and adaptively adjusts the ef-search value of the HNSW algorithm retrieval stage. Specifically, it returns a weight value between 1 and 2 for adjustment. The calculation formula is as follows:

[0018] (4),

[0019] in w d , w m , w v These are the weighting coefficients. m The average value. v The population variance is the numerical value. d for diversity The formula for the diversity index is:

[0020] (5),

[0021] Where k is the number of distinct elements and n is the number of search results. When the density of results is high, that is, the variance and diversity indicators and the mean are small, the formula will make the returned weight value smaller, thereby reducing the ef-search value to accelerate the HNSW search speed. Conversely, the ef-search value will be increased to expand the search breadth and improve the recall rate.

[0022] In this invention, the FIL-knowledge graph employs multi-hop logical connections. Nodes also possess Task and Number attributes, as well as 2560-dimensional vector attributes. Relationship edges also have 2560-dimensional vector attributes. As the core component of the second-stage retrieval, it works in conjunction with adaptive HNSW. Its knowledge content originates from the structured logic of each task in the original aero-engine fault isolation document. A fine-tuned extraction model is used to extract knowledge from the following sections: Description, Possible Causes, Circuit Breakers, Related Data, Fault Isolation Procedure, Initial Evaluation, and Repair Confirmation. Based on semantic characteristics, the sections are divided into two categories: descriptive sections (Related Data) are extracted as triples [Task Number, Task Name, Related Description]; procedural sections (the remaining sections) are parsed into multi-level coherent steps and constructed as [Main Step, Condition, Sub-Step]. The structured path supports multi-hop reasoning; all extracted units (tasks, steps, descriptions) are vectorized through a fine-tuned word embedding model, and the embedding vectors are stored in the graph as attributes of nodes or relations, providing high-fidelity semantic representation for subsequent semantic retrieval and path reasoning.

[0023] In the two-stage progressive retrieval enhancement generation architecture TSP-RAG described in this invention, after obtaining the task scope retrieved by the FER-knowledge graph, the FIL-knowledge graph structures the results into a hierarchical ternary tree for downstream inference, including:

[0024] Task pattern extraction: Standardized task identifiers are extracted using regular expressions to achieve alignment across knowledge graphs;

[0025] Bimodal retrieval includes exclusion mode and inclusion mode. Exclusion mode retrieves knowledge irrelevant to the given task to ensure the breadth of semantic coverage. Inclusion mode focuses on task-related subgraphs to improve retrieval accuracy. To balance novelty and relevance, the retrieval score for inclusion mode is reduced.

[0026] Path expansion: Starting from the top-ranked entity, perform a bounded depth-first search, terminating when the path reaches its maximum depth or when neither end can be expanded further;

[0027] Trie structuring: The retrieved paths are organized into a prefix tree Trie to support hierarchical representation, which facilitates structured reasoning in large language models;

[0028] Post-processing: By removing null values ​​and replacing empty objects with their keys, the output is ensured to conform to the JSON format specification, while improving semantic clarity and downstream compatibility.

[0029] In the two-stage progressive retrieval enhancement generation architecture TSP-RAG described in this invention, after obtaining the ef weight parameters of the FER-knowledge graph output, the designed adaptive HNSW retrieval mechanism improves efficiency while optimizing retrieval breadth through semantic diversity-driven dynamic ef parameter tuning. It also enhances result diversity and achieves cross-knowledge graph task alignment by combining content-level deduplication and automatic task ID extraction. After retrieval, to alleviate information redundancy between heterogeneous sources, a high threshold similarity τ=0.95 is used to identify and remove entities in the FIL-knowledge graph that are highly repetitive with the HNSW retrieval results. Subsequently, a query-aware re-ranking mechanism is used to uniformly rank the fusion results using the original query semantics as the anchor point, ensuring that the output aligns with the user's intentions. Figure 1 Finally, the hierarchical reasoning structure and the reordered retrieval results are jointly input into the large language model, which then generates accurate, standardized, and interpretable answers based on the specified output format.

[0030] This invention also proposes an aero-engine fault data processing system, including a construction function block and a question-and-answer function block. The construction function block prepares the data by uploading files, using a custom segmentation strategy, and employing a knowledge graph. The question-and-answer function block relies on Langchain to implement dialogue management, long-memory storage, and dynamic loading of MCP. Its key feature is that it executes the aero-engine fault processing method described above.

[0031] This invention addresses the challenges of processing multi-source heterogeneous data in aero-engine fault isolation and the difficulty of traditional retrieval methods in simultaneously handling professional terminology entity matching and logical multi-hop reasoning. It proposes FIKAS, a knowledge construction and retrieval system based on a multi-agent data processing workflow (MAW) and a two-stage progressive retrieval enhancement generative architecture (TSP-RAG). First, a MAW workflow is designed, utilizing multimodal document parsing and format standardization techniques combined with a rule-based hybrid cleaning strategy to achieve standardized preprocessing of raw data. Based on this, a domain-adaptive large language model and a secure and controllable dynamic code execution mechanism collaboratively complete fine semantic segmentation of text and efficient vectorized index construction. Second, the TSP-RAG retrieval framework is proposed, which constructs an entity-relationship-based retrieval system... Two complementary knowledge graphs, FER-knowledge graph and FIL-knowledge graph for fault isolation logic, were used to design a two-stage progressive retrieval mechanism of "scope limitation-parameter adaptation". In the first stage, the FER-knowledge graph is used to determine the task scope and adaptively adjust the HNSW retrieval parameters. In the second stage, the retrieval results of HNSW are organically combined with the results of multi-hop logical reasoning and structured path expansion of FIL-knowledge graph, realizing deep collaboration between structured knowledge and unstructured text. Experimental results show that the design of TSP-RAG is effective. At the same time, its performance on the constructed dataset is better than mainstream methods such as LightRAG and LlamaIndex, significantly improving retrieval accuracy and recall, and providing reliable and comprehensive knowledge support for aero-engine fault isolation. Attached Figure Description

[0032] Appendix Figure 1 This is a flowchart of the multi-source data conversion and data preprocessing process in this invention.

[0033] Appendix Figure 2 This is a flowchart of the MAW data cleaning process in this invention.

[0034] Appendix Figure 3 This is a flowchart of the data processing intelligent agent in this invention.

[0035] Appendix Figure 4 This is a flowchart of the workflow of the knowledge graph intelligent agent in this invention.

[0036] Appendix Figure 5 This is the FER-knowledge graph structure diagram in this invention.

[0037] Appendix Figure 6 This is a flowchart of the FER-knowledge graph workflow in this invention.

[0038] Appendix Figure 7 This is the FIL-knowledge graph structure diagram in this invention.

[0039] Appendix Figure 8This is a flowchart of the FIL-knowledge graph workflow in this invention.

[0040] Appendix Figure 9 This is a schematic diagram of the TSP-RAG retrieval framework in this invention.

[0041] Appendix Figure 10 This is a system framework diagram of the present invention.

[0042] Appendix Figure 11 This is a graph showing the MRR scores of each object in the comparative experiment of this invention.

[0043] Appendix Figure 12 These are the MRR scores of each object in the ablation experiment in this embodiment of the invention.

[0044] Appendix Figure 13 This is a schematic diagram comparing recall rates in an embodiment of the present invention.

[0045] Appendix Figure 14 This is a comparison curve of recall time in an embodiment of the present invention.

[0046] Appendix Figure 15 This is the overall flowchart of the present invention. Detailed Implementation

[0047] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0048] Aero-engine systems are highly coupled, with related knowledge widely distributed across heterogeneous and dispersed data sources. Furthermore, the abundance of specialized terminology and the complex relationships between these terms present significant challenges to specialized information retrieval. Traditional single-retrieval augmented generation (RAG) or knowledge graph (KEPG) techniques typically rely solely on vector similarity or specific graph patterns for reasoning, failing to achieve comprehensive and accurate information retrieval. Moreover, existing KEPG-RAG frameworks are often limited to single graphs, lacking flexibility and struggling to capture subtle semantic implications hidden within text. This limitation weakens the fusion of vector and graph retrieval, preventing the system from effectively balancing the needs of specialized terminology entity matching and multi-hop logical reasoning. Therefore, by dividing the functions of the knowledge graph, constructing multiple complementary knowledge graphs, and combining them with vector similarity retrieval, deep collaboration between structured knowledge and unstructured text can be achieved. This provides more complete information coverage and more accurate semantic associations, thus offering reliable and comprehensive knowledge support for aero-engine fault isolation.

[0049] To address the above challenges, this invention proposes an intelligent knowledge system for aircraft engine fault isolation, which effectively promotes relevant knowledge processing and retrieval to support engineering applications, including:

[0050] A multi-agent workflow for processing and organizing multi-source heterogeneous data, named MAW, is proposed. The entire process automatically constructs a vector index and knowledge graph of RAG through various preprocessing and knowledge extraction methods, achieving efficient knowledge processing and storage in the retrieval system. A novel knowledge graph + RAG retrieval architecture (two-stage progressive RAG architecture TSP-RAG) is proposed. Its core lies in the collaborative construction of two types of complementary knowledge graphs (the knowledge graph FER-knowledge graph based on fault entity relationships and the knowledge graph FIL-knowledge graph oriented towards fault isolation logic) and the fusion with the hierarchical navigable small world (HNSW) algorithm, and the adoption of a two-stage progressive retrieval mechanism. This method achieves a good balance between time efficiency and generation quality. Meanwhile, its paradigm of "dual knowledge graph + adaptive retrieval enhancement + two-stage progressive hybrid retrieval" also has good versatility, especially suitable for fields with many technical terms and strong task-driven characteristics, such as aero-engine fault isolation. It can be extended to decision support fields of other similar complex industrial systems. In addition, a knowledge system for aero-engine fault isolation (FIKAS) was proposed. FIKAS organically integrates knowledge processing and organization processes, TSP-RAG and system orchestration interaction modules using intelligent agent technology. The effectiveness of the system has been demonstrated through multiple experiments and specific case implementations.

[0051] First, this invention proposes a multi-agent workflow (MAW) for processing and organizing multi-source heterogeneous data. The MAW specifically has the following functions: multi-source data transformation, data preprocessing, data segmentation, data storage, and knowledge graph construction. To ensure that the above functions can be executed smoothly, a multi-agent + workflow approach is adopted. The agent responsible for data segmentation and data storage is called the data processing agent, and the agent responsible for knowledge graph construction is called the knowledge graph agent.

[0052] The multi-source data transformation and preprocessing pipeline first uses format parsing and recognition technology to map discrete, inconsistently formatted raw data into a unified, standardized intermediate representation, eliminating format barriers and reducing the computational overhead of subsequent processing. Based on this, the system implements a multi-dimensional automated data cleaning strategy to effectively remove unstructured noise, eliminate redundant information, and perform normalization. To address the heterogeneity of different data sources, a hybrid mechanism combining rule-based deterministic cleaning and large-model-based intelligent cleaning is employed to ensure high-quality and consistent input data. A simplified flowchart of the multi-source data transformation and preprocessing process is shown below. Figure 1 As shown.

[0053] The data processing agent integrates a large language model and word embedding model with strong reasoning and code generation capabilities, and is configured with a secure and controllable dynamic code execution mechanism. This agent employs an adaptive segmentation strategy: for short texts, semantic segmentation is performed directly using prompting engineering; for long texts, a structured processing script is generated by the model and executed in a controlled environment to achieve efficient segmentation. Finally, the model performs recursive verification to ensure reasonable segmentation granularity. The processed data and metadata are persistently stored in a high-performance relational database to support subsequent vectorized index construction and efficient retrieval.

[0054] The foundation of the RAG system lies in text vectorization, and the choice of vectorization method directly affects the accuracy and efficiency of retrieval. This system uses a word embedding model to vectorize segmented data, a crucial prerequisite for building a vector index and achieving efficient vector retrieval. General word embedding models are trained on large amounts of general text corpora, often performing poorly for highly specialized fields like aero-engines, which are rich in technical terminology. Therefore, a specialized word embedding model is needed. After evaluating performance and considering hardware conditions, we adopted the Qwen3-word embedding-4B model, which performs well on multiple standard evaluation sets and supports multiple vector dimensions, as our base model. Considering hardware limitations and sample size, this study employs a LoRA-based fine-tuning strategy for domain adaptation. We use a multi-negative-example contrastive learning format to fine-tune the embedding model. This training method significantly improves the semantic discriminativeness and structural quality of the embedding space. Specifically, the training data adopts a multi-negative format [V2][-V2], where each training sample consists of a query, a positive sample pos (a semantically highly relevant fault isolation description), and three hard negative samples neg (interference content from other irrelevant tasks). This example constructs a total of 5,500 high-quality samples, divided into a 90% training set and a 10% test set. The training process is optimized using the InfoNCE loss function, which effectively improves semantic alignment and discriminative ability in the embedding space by maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs. This optimization strategy enables the model to construct highly discriminative semantic representations in the vector space.

[0055] The loss function formula is as follows:

[0056] ,

[0057] The training parameters are set as shown in Table 1:

[0058] Table 1. Hyperparameters for word embedding model training

[0059]

[0060] After training, the test set was used to compare and validate the fine-tuned word embedding model with the baseline model. The specific method is as follows: the validation samples were vectorized using both models, and then the cosine similarity between each question and its corresponding positive and negative examples was calculated. If the similarity of the positive example was higher than that of the negative example, the sample was considered correctly predicted; otherwise, it was considered incorrect. Based on this, the proportion of correctly predicted samples was statistically analyzed to obtain the model's accuracy in distinguishing between positive and negative examples. Simultaneously, for all correctly predicted samples, the difference between the similarity between positive and negative examples was calculated to measure the model's ability to distinguish between them. The training results are shown in Table 2. The accuracy after fine-tuning was 28.43% higher than the un-fine-tuned model, and the difference in similarity between positive and negative examples was significantly increased.

[0061] Table 2 shows the final fine-tuning results and experimental results.

[0062]

[0063] The main function of the knowledge graph agent is to extract the required entities, relations, and related attributes from the segmented data, and to generate Cypher statements to store the extracted information in the graph database. In the knowledge graph construction process, this invention employs a large language model-assisted entity and relation extraction technique. This method places high demands on the reasoning ability of the large language model; after comprehensive evaluation, we selected the Qwen3-8B model as the basic extraction model. To ensure that the nodes and relations in the constructed knowledge graph meet the requirements in terms of semantic accuracy, structural reliability, and domain adaptability, we further employ a LoRA fine-tuning strategy to adapt the model to the domain.

[0064] Specifically, the model is responsible for accurately identifying tasks, steps, conditions, and related descriptive content from the original document, providing high-quality, structured input for subsequent knowledge graph construction. The core goal of fine-tuning is to replace the cumbersome ontology modeling process, thereby improving the model's performance on domain-specific entity relation extraction tasks. Therefore, the training data adopts the Alpaca format, where each sample contains explicit instructions, input context, and expected output to effectively guide the model in learning complex semantic relation extraction logic: {"instruction": "...", "input": "...", "output": "..."}.

[0065] To overcome the bottleneck in reasoning ability faced by basic extraction models in domain knowledge extraction tasks due to their relatively limited parameter size, this invention employs a knowledge distillation method to construct the dataset: a large language model with a large number of parameters and strong reasoning ability is used as the teacher model to generate high-quality labeled data containing complete reasoning chains, thereby effectively improving the reasoning ability of the student model. A distilled dataset containing 5,500 samples was ultimately constructed, divided into a 90% training set and a 10% test set. The fine-tuning task utilizes Causal Language Modeling Loss.

[0066] (2),

[0067]

[0068] The training parameter settings are shown in Table 3:

[0069] Table 3 Extracting model training hyperparameters

[0070]

[0071] Furthermore, to evaluate the model's capabilities in long text understanding and structured knowledge extraction, we constructed an independent validation set consisting of 200 long text paragraphs and their corresponding triples for fine-grained performance analysis and generalization ability verification. The specific verification method is as follows: The F1 score was calculated by combining the precision (number of correctly predicted triples / total number of predicted triples) and recall (number of correctly predicted triples / total number of labeled triples). The results are shown in Table 4. The fine-tuned model achieved a performance improvement of 4.45%.

[0072] The formula for F1 score is as follows: (3).

[0073] Table 4: Final Fine-tuning Results and Test Results

[0074]

[0075] Subsequently, in order to fully utilize the context learning capability of the extraction model, different types of system prompt words were designed for different knowledge graph structures and tasks, which further improved the extraction capability of the knowledge graph agent.

[0076] MAW Process Introduction: The following uses a PDF file from the FIM manual (a manual consisting of various fault isolation tasks) as a processing example to introduce each process.

[0077] First, the format is converted using OCR technology to convert the PDF into a Markdown format file. Then, a cleaning process is performed. The cleaning procedure for FIM manuals is as follows: Figure 2 As shown.

[0078] After data cleaning, the MD file is transferred to the data processing agent. Given the length of the FIM manual, the agent generates segmented code using a large language model and executes it in an isolated sandbox. First, the content is divided into several coarse-grained segments according to the task number (this strategy is consistent with the requirements of the preset prompt words). Then, the large language model performs a fine-grained secondary segmentation on each segment. After the segmentation task is completed, the agent automatically calls a word embedding model to perform vectorization, and finally stores the vector data along with the original text and metadata. Figure 3 The workflow of the data processing intelligent agent was demonstrated.

[0079] Next, the knowledge graph agent extracts entities from the segmented text paragraphs. Depending on the prompt template, it can extract ordinary entities, logical node entities, and other content to adapt to different construction needs. Finally, it generates corresponding Cypher statements based on the extraction results and stores them. Figure 4 The workflow of the knowledge graph intelligent agent is demonstrated. All the aforementioned intelligent agents operate collaboratively within a unified workflow. When an anomaly occurs at any stage, the system will automatically trace the context for self-diagnosis and attempt to repair itself.

[0080] The core of TSP-RAG consists of FER-knowledge graph, which describes the structural information in knowledge text, and FIL-knowledge graph, which encodes the dynamic logical rules of fault propagation and diagnosis. The vector retrieval module adopts the current mainstream ANN algorithm HNSW, and its retrieval efficiency is highly dependent on the index quality

[30] . In order to make efficient use of the above core modules, the framework designs a two-stage retrieval mechanism (the first stage is retrieval by FER-knowledge graph, and the second stage is retrieval by FIL-knowledge graph), and innovatively introduces a feedback-driven adaptive strategy to dynamically adjust the ef-search parameter in the HNSW approximate nearest neighbor search algorithm (ef-search is the core parameter that controls the search breadth in the HNSW retrieval stage, and directly determines the trade-off between recall and latency.) and the task-focused retrieval of FIL-knowledge graph. The system can adaptively balance the retrieval accuracy and efficiency of the second stage according to the retrieval results of the first stage. Finally, the fusion and sorting of vector retrieval results and knowledge graph nodes are introduced to gradually optimize the retrieval step by step, thereby providing highly relevant, structured and logically verified domain knowledge for large language models.

[0081] The FER-Knowledge Graph completes the retrieval task in the first stage of the entire retrieval process, consisting of the relationships between entities in the original text. A fine-tuned extraction model is used to extract entity relationships, with the extraction format required to be [entity 1, entity 1 type, relationship, entity 2, entity 2 type]. The original text includes entity types such as: fault / phenomenon, system / component, task / program, manual / document, status / indication, location, action, maintenance information, etc. These types are stored as entity type attributes. After extraction, a fine-tuned word embedding model is used to vectorize each entity and relationship, and the vectorized data is stored as vector attributes. The FER-Knowledge Graph structure is as follows: Figure 5 As shown:

[0082] The FER-knowledge graph structure uses standard triples, and nodes also have Task, Number, and Type attributes, as well as 2560-dimensional vector attributes. Edges also have 2560-dimensional vector attributes. The FER-knowledge graph's function is to perform searches based on user-input questions and keywords within those questions. It dynamically and adaptively adjusts the ef-search value during the HNSW algorithm's retrieval phase based on the density of cosine similarity values ​​in the search results. Specifically, it returns a weight value between 1 and 2 for adjustment, calculated as follows:

[0083] (4),

[0084] in w d , w m , w v These are the weighting coefficients. m The average value. v The population variance is the numerical value. d for diversity The formula for the diversity index is:

[0085] (5),

[0086] Where k is the number of distinct elements and n is the number of search results.

[0087] When the density of results is high, i.e. the variance and diversity indices and the mean are small, the formula will result in a smaller returned weight value, which will reduce the ef-search value to speed up HNSW retrieval. Conversely, the ef-search value will be increased to expand the search breadth and improve the recall rate.

[0088] The graph also features a dynamic task scope constraint mechanism: based on the task numbers corresponding to the retrieved TOP-K nodes, the retrieval scope of the FIL-knowledge graph is limited and enhanced, thereby significantly improving the recall rate while focusing on relevant task subgraphs. Figure 6 The above process is demonstrated.

[0089] The FIL-Knowledge Graph, as a core component of the second-stage retrieval, works in conjunction with the adaptive HNSW. Its knowledge content originates from the structured logic of each task in the original aero-engine fault isolation document. We employ a fine-tuned extraction model to extract knowledge from the following sections: Description, Possible Causes, Circuit Breakers, Related Data, Fault Isolation Procedure, Initial Evaluation, and Repair Confirmation. Based on semantic characteristics, these sections are divided into two categories:

[0090] Descriptive sections (Description, Possible Causes, Circuit Breakers, RelatedData): extracted as triples [Task ID, Task Name, Related Description];

[0091] Procedural sections (the rest of the sections): These are parsed into multi-level coherent steps and constructed as structured paths of [main step, condition, sub-step] to support multi-hop reasoning.

[0092] All extracted units (tasks, steps, descriptions) are vectorized using a fine-tuned word embedding model, and the embedding vectors are stored as attributes of nodes or relations in the graph, providing high-fidelity semantic representation for subsequent semantic retrieval and path reasoning. The FIL-knowledge graph structure is as follows: Figure 7 As shown.

[0093] The FIL-knowledge graph structure employs multi-hop logical connections, and nodes also have Task and Number attributes as well as 2560-dimensional vector attributes. Relationship edges also have 2560-dimensional vector attributes.

[0094] The role of the FIL-Knowledge Graph is to perform focused retrieval based on the task scope defined by the FER-Knowledge Graph: nodes belonging to the defined task are given high priority, while nodes outside the scope are penalized with weights during the Top-K ranking stage, and are only retained if they are highly similar to the query. This mechanism avoids missed detections due to task scope limitations while effectively suppressing the introduction of irrelevant or redundant content. Building on this, the FIL-Knowledge Graph further expands multi-hop relationships starting from the retrieved nodes, constructing a complete fault isolation logic chain and outputting it in a hierarchical structure, significantly improving the logical completeness and interpretability of the returned knowledge. Figure 8 The process was demonstrated.

[0095] TSP-RAG employs a two-stage collaborative architecture. The first stage is executed by the FER-knowledge graph: based on the user's question and its extracted keywords, keyword matching is performed first, followed by vector retrieval in the knowledge graph to determine the relevant task scope (Task and Number attributes), and ef weight parameters for HNSW index tuning are calculated based on similarity density. The second stage is jointly completed by the FIL-knowledge graph and the adaptive HNSW retrieval module: the FIL-knowledge graph performs precise retrieval using the task scope defined in the first stage. Some of the results are directly output as candidate answers, while others are expanded through multi-hop relationships to construct hierarchical reasoning paths and organized into a trie structure (i.e., a reasoning tree) that is friendly to large language models. Simultaneously, the adaptive HNSW retrieval module dynamically adjusts the search breadth based on the obtained ef weights, performs vector retrieval, and deduplicates the results at the content level. Finally, the system semantically fuses and re-ranks the nodes retrieved from the FIL-knowledge graph with the HNSW retrieval results, and inputs them along with the reasoning tree into the large language model to generate accurate and interpretable answers.

[0096] In the retrieval phase of the FER-knowledge graph, we propose a frequency-aware retrieval strategy. This strategy dynamically switches between exact matching and semantic approximate matching searches based on entity popularity. Specifically, if the number of nodes matched precisely is low, it indicates a low-frequency entity, and the exact matching results are retained while keyword vector retrieval is performed, integrating entity-level and sentence-level vector retrieval. For high-frequency entities, exact matching and keyword vector retrieval are discarded to avoid noise, relying solely on relational and sentence semantics for retrieval. Furthermore, a diversity-driven ef weight is introduced to achieve semantically adaptive HNSW tuning. Finally, the function outputs the average ef weight and a deduplicated task list. The former optimizes the breadth of vector retrieval, while the latter serves as a task constraint passed to the second-stage FIL-knowledge graph retrieval, achieving efficient knowledge acquisition through cross-graph collaboration and semantic adaptation.

[0097] After obtaining the task scope retrieved from the FER-knowledge graph, our FIL-knowledge graph achieves a dynamic balance between general semantic context and task-specific knowledge through task-aware filtering technology. The results are then structured into a hierarchical ternary tree for downstream inference—this design bridges the gap between intensive retrieval and symbolic knowledge representation. It mainly consists of the following key components: Task pattern extraction: Standardized task identifiers are extracted using regular expressions to achieve alignment across the knowledge graph.

[0098] Bimodal retrieval: 1. Exclusion mode: Retrieves knowledge irrelevant to the given task to ensure broad semantic coverage. 2. Inclusion mode: Focuses on task-related subgraphs to improve retrieval accuracy. To balance novelty and relevance, the score for retrieval results in the inclusion mode is moderately lowered.

[0099] Path expansion: Starting from the top-ranked entity, perform a bounded depth-first search (1 to 5 hops), terminating when the path reaches its maximum depth or when neither end can be expanded further.

[0100] Trie structuring: The retrieved paths are organized into a trie to support hierarchical representation, which facilitates structured reasoning in large language models.

[0101] Post-processing: By removing null values ​​and replacing empty objects with their keys, the output is ensured to conform to the JSON format specification, while improving semantic clarity and downstream compatibility.

[0102] After obtaining the ef weight parameters output by the FER-knowledge graph, the designed adaptive HNSW retrieval mechanism improves efficiency while optimizing retrieval breadth through dynamic ef parameter tuning driven by semantic diversity; and further enhances result diversity by combining content-level deduplication and automatic task ID extraction, and achieves task alignment across knowledge graphs.

[0103] After the retrieval is completed, to alleviate information redundancy among heterogeneous sources, we use a high similarity threshold (τ=0.95) to identify and remove entities in the FIL-knowledge graph that are highly repetitive with the HNSW retrieval results. Subsequently, through a query-aware re-ranking mechanism, the fusion results are uniformly ranked using the original query semantics as anchor points to ensure that the output aligns with the user's intended meaning. Figure 1 This strategy, while preserving complementary knowledge from multiple sources, effectively suppresses redundancy and generates a semantically coherent and query-aligned final result list. Finally, the hierarchical reasoning structure and the reordered retrieval results are jointly input into a large language model, which, combined with a prescribed output format, generates accurate, standardized, and interpretable answers. The retrieval framework consists of… Figure 9 As shown.

[0104] This invention constructs a knowledge system FIKAS for fault isolation in aero-engines. Table 5 shows the key configurations of the system: the main model uses Qwen3-MAX and is deployed through the API of Alibaba Cloud Bailian platform; the isolation environment uses the API of Modal platform to create a sandbox; and the components communicate with each other through FastAPI. Figure 10 The components of the FIKAS system are shown.

[0105] Table 5 FIKAS System Configuration

[0106]

[0107] The entire system is designed based on the open-source framework LangChain, which features workflow orchestration, intelligent agents, and middleware. The MAW module uses a workflow engine to encapsulate a dual-agent collaborative mechanism. The retrieval section follows the principle of intelligent triage: first, a domain classifier determines whether the user's question falls within a specific professional category; for non-domain questions, the general capabilities of the large language model are directly invoked to provide an answer, thereby avoiding invalid retrieval and significantly reducing computational resource consumption. TSP (Knowledge Graph) is the core cognitive engine of the retrieval agent. By parsing the entity and task semantics in the user's question, it locates key nodes in the graph, triggering a hybrid recall strategy of vector retrieval and graph structure retrieval to generate high-quality context for subsequent reasoning. In addition, the system has a built-in scalable MCP service layer, integrating basic functions such as database query, web page query, and regional time acquisition. All MCP services follow a hot-swappable architecture design, supporting dynamic registration and uninstallation, achieving plug-and-play functionality and seamless online upgrades, ensuring flexible evolution of system functions.

[0108] Finally, an interactive platform was created based on this system, which is divided into two basic functional blocks: construction and question answering. The construction functional block can complete the data preparation by uploading files, customizing the segmentation strategy, and using the knowledge graph function. The question answering module relies on LangChain to realize AI native functions such as dialogue management, long-term memory storage, and dynamic loading of MCP. The simple interface reduces the operation threshold for knowledge system construction and querying, so as to support the completion of knowledge system construction and querying through simple operations.

[0109] Before the experiment began, we constructed a knowledge graph and HNSW vector index based on the MAW module of the FIKAS system, using some content from the fault isolation manual of a certain engine model, to support TSP-knowledge graph retrieval. The final graph contains 8,257 nodes and 10,218 relationships, and the vector index covers 1,562 manual fragments.

[0110] TSP-Knowledge Graph Retrieval Experiment

[0111] (1) Experimental setup

[0112] To systematically evaluate the performance of the TSP-knowledge graph, we constructed two aero-engine fault isolation knowledge datasets: FI-QA and FI-QA-mini. Both datasets are derived from the manual document described earlier for constructing the knowledge graph and vector index, and will be used for comparative and ablation experiments to evaluate the performance of the retrieval module under different data scales. Each data point follows the format {query, answer, source}, which consists of a question and answer based on a fragment from the manual document, along with the fragment ID. Table 6 shows the information for these two datasets.

[0113] Table 6 Experimental Dataset

[0114]

[0115] In this experiment, the key hyperparameters of the HNSW index were ultimately set as follows: M=128, ef-construction=300, dim=2560. The number of results returned by both the Cypher query and the HNSW retrieval was set to 20.

[0116] Both the comparative and ablation experiments used the same evaluation metrics, MRR and HIT@n, which accurately reflect the agent's core ability to present the correct answer to the user in practical applications from two complementary dimensions: first-order ranking accuracy and Top-n recall stability. The formula is as follows:

[0117] (6), (7), of which ,

[0118] Q: Query set, |Q| represents the total number of queries, rank i Hit: The ranking position of the first relevant document in the search results list in the i-th query; n: the truncation position, indicating that only the first n results are viewed (set to 1, 5, and 10 in this experiment); hit i : A binary function. If the most relevant document is in the nth position or earlier, the value is 1 (hit), otherwise it is 0 (miss).

[0119] (2) Comparative experiment

[0120] To comprehensively evaluate the performance advantages of TSP-knowledge graph, this example designed a comparative experiment, selecting current mature retrieval models as comparison objects, namely the mainstream pure vector retrieval scheme HNSW, the high-performance knowledge graph and RAG fusion framework LightRAG, and the recently popular multi-index driven retrieval framework LlamaIndex.

[0121] Here, the hyperparameter settings for HNSW are consistent with those above, and the recall counts for LightRAG and LlamaIndex are initially set to 20. Both the indexing and retrieval processes for LightRAG and LlamaIndex require a large language model to drive them. Here, the model service is accessed via the API-Key from Alibaba Cloud's Bailian platform. The largest language model selected is Qwen3-MAX, the most powerful model in the Qwen series, and the word embedding model is a finely tuned version of Qwen3-WordEmbedding-4B.

[0122] This example uses two constructed datasets for experiments, and the results are shown in Table 7 and 8. Figure 11 As shown, the best score is highlighted in bold, and the second-best score is underlined. On the FI-QA-mini dataset, although HNSW performs better on the Hit@1 and Hit@10 metrics, on the larger FI-QA dataset, the model of this invention achieves the best performance on all metrics. This fully demonstrates that the performance advantage of the model of this invention becomes increasingly significant with the increase in data scale. Furthermore, experiments show that LightRAG performs relatively poorly, and during the experiments, we found that the retrieval recall of LightRAG must be limited to around 10; otherwise, it will cause errors due to exceeding the context window limit of the large language model, thus limiting the performance of LightRAG. Finally, the index construction time of LightRAG and LlamaIndex is also significantly longer than that of the method proposed in this invention. Tables 7, 8, and... Figure 12 The results of the comparative experiment are presented.

[0123] Table 7 Comparative Experiment Results under FI-QA-mini

[0124] Table 8 Comparative Experiment Results under FI-QA

[0125]

[0126] Comparing the MRR scores of each object in the experiment

[0127] (3) Ablation Experiment: To verify the effectiveness of the core idea of ​​TSP-Knowledge Graph, an ablation experiment was conducted in this case. By removing the FIL-Knowledge Graph module and retaining HNSW and FER-Knowledge Graph, we specifically evaluated the impact of the adaptive Ef-search strategy on the retrieval performance of HNSW. The experimental results are shown in the table below. The results show that on the small-scale dataset FI-QA-mini, the performance of adaptive Ef-search is comparable to that of the fixed parameter setting; however, on the larger FI-QA dataset, HNSW using adaptive Ef-search significantly outperforms the fixed parameter version on all evaluation metrics. This phenomenon indicates that in a large-scale data environment, the performance of HNSW is more sensitive to the Ef-search parameters, and adaptive Ef-search can effectively improve the retrieval effect of HNSW. Tables 9 and 10 show the results of this ablation experiment.

[0128] Table 9. Impact of Adaptive Ef-search on Performance under FI-QA-mini

[0129]

[0130] Table 10. Impact of Adaptive Ef-search on Performance under FI-QA

[0131] To evaluate the contribution of the scope constraint mechanism to the overall performance of the TSP-knowledge graph, we conducted the following ablation experiment: We retained the HNSW and FIL-knowledge graphs, but removed the scope constraint of the FER-knowledge graph on the FIL-knowledge graph. The experimental results are shown in the table below. The results demonstrate that this scope constraint mechanism has a significant impact on the retrieval performance of the TSP-knowledge graph framework on both datasets. Tables 11 and 12 present the results of this ablation experiment.

[0132] Table 11 Impact of Range Constraints on Performance under FI-QA-mini

[0133]

[0134] Table 12 Impact of Range Constraints on Performance under FI-QA

[0135]

[0136] The MRR scores of the two ablation experiments are as follows: Figure 13 As shown.

[0137] (4) Supplementary Experiment

[0138] With the continuous improvement of the performance of large language models (MLMs) and the ongoing expansion of their context windows, relying solely on the first 10 search results is no longer sufficient to comprehensively and accurately evaluate the actual performance of a retrieval framework. Therefore, this example includes supplementary experiments, comparing only the best-performing TSP-Knowledge Graph and the second-best performing HNSW from the previous comparative experiments. In this experiment, we increased the number of returned search results from 20 to 50, while keeping other parameters unchanged. The experimental dataset consists of 64 samples randomly selected from FI-QA. This experiment employs a new evaluation criterion to determine whether the returned search results are sufficient to support the answer to the question; the specific calculation formula is as follows:

[0139] (8),

[0140] Where: R: The number of datasets is 64 here, R * The number of search results is sufficient to support the sample size of the question answers.

[0141] Experimental results are as follows Figure 14 , 15 As shown, TSP-RAG fully utilizes the structural information of the knowledge graph and the logical chain constructed by multi-hop logical reasoning capabilities through a two-stage progressive retrieval mechanism, and achieves excellent retrieval performance by combining it with HNSW vector retrieval. Although the retrieval time of TSP-RAG is several times that of HNSW, the overhead of this retrieval stage is acceptable considering the time consumed in the subsequent large language model inference stage.

[0142] The following case study will illustrate the complete query process of the system.

[0143] When a user enters the question: "What should be done if the correct voltages are not found between pin 1 and pin 2 of electrical connector DP1203?", the system first uses its built-in large language model to extract the key entity "Electrical connector DP1203" from the question. Then, based on this entity, the system performs entity matching and retrieval in the FER-knowledge graph, returning relevant fragment numbers (e.g., ['74-21 _05', '79-05 _12', '77-05 _05', '73-00 _032', ……]).

[0144] Meanwhile, the system calculated that the weight parameter for adjusting Ef-search was 0.8691, which means that the size of the Ef-search parameter can be appropriately reduced to speed up the retrieval. Next, the system passed the aforementioned fragment number and Ef-search weight parameter to the second-stage FIL-knowledge graph and HNSW for retrieval.

[0145] The following is an example of a retrieval chain for a FIL-knowledge graph:

[0146] "if Do the Repair Confirmation at the end of this task.": {"If theRepair Confirmation is not MAWtisfactory, then continue. ": {"if Do not find the maintenance mesMAWges": {" Continue. ": {"if The inspection isMAWtisfactory": {…….}}}}}.

[0147] Here are some examples of search results for HNSW:

[0148] ['START LEVER POSITIONCONNECTOR DP1203RANGECUTOFFPIN 1 TO PIN2GREATER THAN 22 VDCIDLEPIN 1 TO PIN 2LESS THAN 5 VDC\n(f) If you find thecorrect voltages above, then replace the HMU.\n1) These are the tasks: HMURemoval, AMM TASK 73-21-10-000-801-F00, HMU Installation,…']

[0149] Finally, the system sends the fragments and chains retrieved from the FIL-knowledge graph, along with the fragments retrieved from HNSW, to the large language model for inference and integration. The system's final response is: "Examine and repair the wiring between pin 2 and the applicable circuit breaker, then open the circuit breakers and install MAWfety tags, and connect the connector DP1203 to the HMU."

[0150] The FIKAS system developed in this example aims to automate the extraction of effective fault isolation knowledge from massive, multi-source data, assisting engineers in information retrieval and thereby improving the efficiency of aero-engine fault isolation. Specifically, the system constructs an HNSW index and two knowledge graphs (knowledge graph storage) to support the TSP-RAG retrieval process by processing and extracting knowledge from multiple source files. Through comparative and ablation experiments, we verified the significant performance advantages of TSP-RAG and the effectiveness of each module. Finally, the system adopts a front-end and back-end separation architecture, and the simplified front-end interface greatly facilitates the practical use by engineers. The implementation path is as follows: Figure 15 .

[0151] The knowledge and information involved in aero-engine fault isolation is vast and wide-ranging, and traditional manual retrieval and integration methods are often inefficient. Therefore, the FIKAS system developed in this paper aims to solve two core challenges in this field: processing multi-source information and accurately retrieving complex information. In summary, this invention first designs a multi-source file knowledge extraction workflow driven by dual agents, realizing the entire process from raw data identification, preprocessing, segmented storage to vectorized storage. Subsequently, based on dual knowledge graphs, a two-stage progressive hybrid retrieval mechanism, and parameter adaptive mechanisms, a TSP-RAG architecture is constructed, effectively balancing efficiency and recall in the retrieval stage. Through comparative and ablation experiments, we verify the significant performance advantages of TSP-RAG and the effectiveness of each module. Case studies further confirm the application value of this system in practical fault isolation. Finally, the system adopts a front-end and back-end separation architecture and designs a simplified front-end interface, greatly facilitating engineers' actual operation and effectively improving the efficiency of aero-engine fault isolation. Furthermore, the ideas involved in the system can be applied to other aero-engine tasks.

Claims

1. A method for processing aircraft engine fault data, characterized in that, First, the raw data is preprocessed using the MAW workflow. Then, semantic segmentation and vectorized index construction of the text are completed through a domain-adaptive large language model and dynamic code execution mechanism. Second, a two-stage progressive retrieval enhancement generation architecture TSP-RAG is proposed. The first stage is executed by the knowledge graph FER-knowledge graph based on fault entity relationships: based on the user question and its extracted keywords, keyword matching is performed first, followed by vector retrieval in the knowledge graph to determine the relevant task scope, and the ef weight parameters for HNSW index tuning are calculated based on the similarity density. The second stage is completed jointly by the knowledge graph FIL-knowledge graph oriented towards fault isolation logic and the adaptive HNSW retrieval module: the FIL-knowledge graph performs precise retrieval using the task scope defined in the first stage. Part of the results are directly output as candidate answers, and the other part is used to construct a hierarchical reasoning path through multi-hop relationship expansion and organized into a trie structure, i.e., a reasoning tree. At the same time, the adaptive HNSW retrieval module uses the obtained ef... The search breadth is dynamically adjusted by weights, vector retrieval is performed and the results are deduplicated at the content level. Finally, the nodes retrieved by FIL-knowledge graph and the results retrieved by HNSW are semantically fused and reordered, and then fed into the large language model along with the reasoning tree to generate a response. In the FER-knowledge graph retrieval stage of the two-stage progressive retrieval enhancement generation architecture TSP-RAG, a frequency-aware retrieval strategy is executed. This strategy dynamically switches between exact matching and semantic approximate matching based on entity popularity. That is, based on the number of nodes that are exactly matched, if the number of nodes is low, it is a low-frequency entity, and the exact matching result is retained. At the same time, keyword vector retrieval is performed, and entity-level and sentence-level vector retrieval are integrated. If it is a high-frequency entity, exact matching and keyword vector retrieval are discarded to avoid noise, and retrieval is performed only based on relational semantics and sentence semantics. Diversity-driven ef weights are introduced to achieve semantically adaptive HNSW tuning. Finally, the function outputs the average ef weight and the deduplicated task list. The former is used to optimize the vector retrieval breadth, and the latter is passed to the second stage FIL-knowledge graph retrieval as a task constraint. The FER-knowledge graph structure uses standard triples. Nodes also have Task, Number, and Type attributes, as well as 2560-dimensional vector attributes. Edges also have 2560-dimensional vector attributes. The FER-knowledge graph extracts entity relations. It searches based on the user-input question and keywords within the question. Based on the density of cosine similarity values ​​in the search results, it dynamically and adaptively adjusts the ef-search value in the HNSW algorithm's retrieval phase. Specifically, it returns a weight value between 1 and 2 for adjustment, calculated as follows: （4）， in w d , w m , w v These are the weighting coefficients. m The average value. v The population variance is the numerical value. d for diversity The formula for the diversity index is: （5）， Where k is the number of distinct elements and n is the number of search results. When the density of results is high, that is, the variance and diversity indicators and the mean are small, according to the formula, the returned weight value will be smaller, thereby reducing the ef-search value to accelerate the HNSW search speed. Conversely, the ef-search value will be increased to expand the search breadth and improve the recall rate. FIL-Knowledge Graph uses multi-hop logical connections, and nodes also have Task and Number attributes as well as 2560-dimensional vector attributes. Relationship edges also have 2560-dimensional vector attributes.

2. The method for processing aircraft engine fault data according to claim 1, characterized in that, Before preprocessing the raw data, multi-source data conversion is required. First, through format parsing and recognition technology, the discrete and inconsistent raw data are mapped into a unified standardized intermediate representation. Then, multi-dimensional automated data cleaning is performed. After data cleaning is completed, the MD file is transferred to the data processing agent, which adopts an adaptive segmentation strategy: for short texts, semantic segmentation is performed directly using prompting engineering. For long texts, a structured processing script is generated by the model and executed in a controlled environment to achieve efficient segmentation. Finally, the model performs recursive verification to ensure that the segmentation granularity is reasonable. The processed data and metadata are persistently stored in a high-performance relational database to support subsequent vectorized index construction and efficient retrieval. The segmented data vectorization is completed through a word embedding model.

3. The method for processing aircraft engine fault data according to claim 2, characterized in that, The word embedding model uses the Qwen3-word embedding-4B model as the base model. Considering hardware conditions and the number of samples, a LoRA-based fine-tuning strategy is adopted for domain adaptation. A multi-negative example contrastive learning format is used to fine-tune the embedding model. The training data adopts a multi-negative example format [V2][-V2], where each training sample consists of a query, a positive sample pos (i.e., a semantically highly relevant fault isolation description), and three hard negative samples neg (i.e., interference content from other irrelevant tasks). A total of 5,500 high-quality samples are constructed, divided into a 90% training set and a 10% test set. The training process is optimized using the InfoNCE loss function, the formula of which is as follows: （1）， in, For the question, As a positive sample, For the i-th negative sample, For similarity function, The number of negative samples. This is the temperature coefficient.

4. The method for processing aircraft engine fault data according to claim 3, characterized in that, In the knowledge graph construction process, a large language model-assisted entity and relation extraction technique is adopted. The Qwen3-8B model is selected as the basic extraction model, and the LoRA fine-tuning strategy is used to adapt the model to the domain. The training data adopts the Alpaca format, that is, each sample contains explicit instructions, input context, and expected output, so as to effectively guide the model to learn complex semantic relation extraction logic: {"instruction": "..." ,"input": "...", "output": "..."}. The dataset is constructed using the knowledge distillation method: a large language model with strong reasoning ability and a large number of parameters is used as the teacher model. Finally, a distillation dataset containing 5,500 samples is constructed, which is divided into a 90% training set and a 10% test set. The fine-tuning task adopts Causal Language Modeling Loss. （2）， in, For the first indivual , For the first One goal , Represents the original model parameters Plus Introduced low-rank update , This indicates that only the answer / output part is considered. Loss was calculated. An independent validation set consisting of 200 long text paragraphs and their corresponding triples was also constructed for fine-grained performance analysis and generalization ability verification. The specific verification method is as follows: The F1 score was calculated by combining the precision (number of correctly predicted triples / total number of predicted triples) and recall (number of correctly predicted triples / total number of labeled triples). The F1 score formula is as follows: （3）， Table 4: Final fine-tuning results and test results. The fine-tuned model achieved a performance improvement of 4.45%.

5. The method for processing aircraft engine fault data according to claim 4, characterized in that, The format requirement for FER-Knowledge Graph Extraction is [Entity 1, Entity 1 Type, Relationship, Entity 2, Entity 2 Type]. The original text includes the following entity types: fault / phenomenon, system / component, task / program, manual / document, status / indication, location, action, maintenance information. The type is stored as the type attribute of the entity. After extraction, each entity and relation is vectorized and stored as a vector attribute.

6. The method for processing aircraft engine fault data according to claim 5, characterized in that, FIL-Knowledge Graph, as the core component of the second-stage retrieval, works in conjunction with Adaptive HNSW. Its knowledge content originates from the structured logic of each task in the original aero-engine fault isolation document. Employing a finely tuned extraction model, it extracts knowledge from the following sections: Description, Possible Causes, Circuit Breakers, Related Data, Fault Isolation Procedure, Initial Evaluation, and Repair Confirmation. Based on semantic characteristics, the sections are divided into two categories: descriptive sections, extracted as triples [task number, task name, related description]; and procedural sections (the remaining sections), parsed into multi-level coherent steps and constructed as structured paths [main step, condition, sub-step] to support multi-hop reasoning. All extraction units are vectorized using a finely tuned word embedding model. Extraction units include tasks, steps, and descriptions, and the embedded vectors are stored in the graph as attributes of nodes or relationships, providing high-fidelity semantic representations for subsequent semantic retrieval and path reasoning.

7. The method for processing aircraft engine fault data according to claim 6, characterized in that, In the two-stage progressive retrieval enhancement generation architecture TSP-RAG, after obtaining the task scope retrieved from the FER-knowledge graph, the FIL-knowledge graph structures the results into a hierarchical ternary tree for downstream inference, including: Task pattern extraction: extracting standardized task identifiers through regular expressions to achieve alignment across knowledge graphs; Bimodal retrieval: including exclusion mode and inclusion mode, where exclusion mode: retrieves knowledge irrelevant to the given task to ensure the breadth of semantic coverage; inclusion mode: focuses on task-related subgraphs to improve retrieval accuracy. To balance novelty and relevance, the retrieval result score of inclusion mode is reduced; Path expansion: starting from the top-ranked entities, performing a bounded depth-first search, terminating when the path reaches its maximum depth or both ends cannot be expanded further; Trie structuring: organizing the retrieved paths into a prefix tree Trie to support hierarchical representation, facilitating structured inference by large language models; Post-processing: by removing null values ​​and replacing empty objects with their keys, ensuring that the output results conform to the JSON format specification, while improving semantic clarity and downstream compatibility.

8. The method for processing aircraft engine fault data according to claim 7, characterized in that, After the two-stage progressive retrieval enhancement generation architecture TSP-RAG completes the retrieval, a high threshold similarity τ=0.95 is used to identify and remove entities in the FIL-knowledge graph that are duplicated with the HNSW retrieval results. Subsequently, through a query-aware re-ranking mechanism, the fusion results are uniformly ranked with the original query semantics as the anchor. Finally, the hierarchical reasoning structure and the re-ranked retrieval results are jointly input into the large language model, which generates the answer.

9. An aero-engine fault data processing system, comprising a construction function block and a question-and-answer function block, wherein the construction function block prepares data by uploading files, using a custom segmentation strategy, and employing a knowledge graph, and the question-and-answer function block utilizes Langchain to implement dialogue management, long-memory storage, and dynamic loading of MCP, characterized in that... Perform the aircraft engine fault data processing method as described in any one of claims 1-8.

Images