Fault analysis model construction method and system for quality characteristic analysis of airborne system

Abstract

This invention provides a method and system for constructing a fault analysis model for airborne system quality characteristic analysis, belonging to the field of general airborne system quality characteristic analysis technology. The method includes: S1, parsing multimodal airborne system task requirements based on airborne system physical data; S2, enhancing fault domain knowledge using retrieval-enhanced generation and a hybrid approach of knowledge graph and database; S3, generating an initial fault analysis model using a large language model; S4, automatically verifying and optimizing the initial fault analysis model; and S5, generating a fault analysis model for airborne system quality characteristic analysis. This invention achieves automatic construction of the fault analysis model through multimodal system physical data parsing, domain knowledge enhancement, and a large language model-driven generation mechanism. Simultaneously, it introduces syntax and semantic verification, a self-repair strategy, and a source data-driven fault text semantic alignment and difference analysis mechanism to ensure consistency between the generated model and the physical source data, thereby improving the accuracy of airborne system fault analysis results.

Description

Technical Field

[0001] This invention relates to the field of general quality characteristic analysis technology for airborne systems, and specifically to a method and system for constructing a fault analysis model for airborne system quality characteristic analysis. Background Technology

[0002] In the modern aviation industry, aviation equipment is evolving towards high integration, complexity, and intelligence. This places extremely stringent requirements on the general quality characteristics (GQC) of systems, namely reliability, maintainability, supportability, testability, safety, and environmental adaptability. Traditional model-based systems engineering (MBSE) often faces the dual challenges of "model silos" and "semantic gaps" when dealing with the safety and reliability analysis of such complex systems. Although AltaRica, as a powerful formal modeling language, can effectively support model-based safety analysis (MBSA), in practical applications, the process of converting massive amounts of system function settings, unstructured requirements documents, and multimodal design diagrams into AltaRica fault analysis models still mainly relies on manual implementation. This experience-dependent approach is not only time-consuming and labor-intensive but also highly susceptible to human error, leading to a disconnect between the analysis model and the system design state. Especially during the airworthiness verification phase, ensuring the consistency, completeness, and traceability of the safety analysis model with system design requirements has become a key bottleneck restricting the efficiency and quality of aviation equipment development.

[0003] To overcome the aforementioned limitations, the introduction of cutting-edge artificial intelligence technologies and advanced experiences, such as Large Language Models (LLM), Retrieval-Augmented Generation (RAG), Knowledge Graphs (KG), and historical model libraries, into the MBSE field is becoming a breakthrough in solving the challenges of complex system modeling. This invention proposes an intelligent generation architecture that deeply integrates Artificial Intelligence (AI) technology with system physics methodology. This architecture no longer relies solely on rule-driven transformation but leverages the powerful natural language understanding and multimodal parsing capabilities of LLM to accurately extract business logic and fault propagation paths from system documents. It uses Retrieval-Augmented Generation (RAG) technology to dynamically invoke aerospace standards and specifications such as ARP4761 and ARP4754A as constraints for generation, ensuring model compliance. It utilizes knowledge graphs to construct deep semantic relationships between entities, disambiguating and supplementing implicit system interaction logic. Furthermore, by retrieving and reusing high-quality model fragments accumulated in the past, it can effectively activate tacit knowledge from past model development. This combination of technologies aims to shift the modeling paradigm from "document-driven" to "data and knowledge-driven," providing a knowledge foundation for building AltaRica fault analysis models.

[0004] The core of this invention lies in constructing an intelligent generation method with a closed-loop capability of "understanding-generation-verification-repair-alignment". Unlike general model building methods, this invention emphasizes the controllability of the generation process and the high reliability of the results. By introducing intermediate representations of XML structures, the topological relationships and attribute features of model elements are decoupled. Combined with a formal logical verification mechanism, most logical paradoxes and circular dependencies can be eliminated before code generation. More importantly, this invention not only focuses on the syntactic correctness of the model but also strives to solve the semantic consistency problem. Through differential analysis technology, it quantifies the deviation between the generated model and the original requirements, ultimately forming a "human-machine collaborative" decision-making workflow. This method can not only significantly reduce the modeling cycle for safety analysis of key components such as aero-engines and avionics systems but also provide solid technical support for the safety assessment and airworthiness certification of airborne systems through interpretable generation paths and strict semantic alignment of fault texts, promoting the intelligent and automated advancement of general quality characteristic analysis in the aviation industry. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention aims to provide a method and system for constructing fault analysis models for airborne system quality characteristic analysis. Through multimodal system physical data parsing, domain knowledge enhancement, and a model generation mechanism driven by a large language model, it achieves automatic construction of AltaRica fault analysis models. Simultaneously, it introduces syntax and semantic verification, self-repair strategies, and a source data-driven fault text semantic alignment and difference analysis mechanism to improve the syntactic correctness, semantic consistency, and physical reliability of the generated model. This accelerates the efficiency of general quality characteristic analysis work such as safety and reliability of airborne systems during the scheme design phase, improves the physical implementation capability of complex system modeling and analysis, and ensures that the generated model remains consistent with the system's physical source data, avoiding semantic drift and structural errors. This significantly improves the model's correctness, completeness, and traceability, thereby enhancing the accuracy and physical reliability of analysis results in practical applications such as fault propagation analysis, safety assessment, and reliability prediction of airborne systems.

[0006] Specifically, on the one hand, the present invention provides a method and system for constructing a fault analysis model for analyzing the quality characteristics of airborne systems, which includes the following steps: S1: Analyze the mission requirements of multimodal airborne systems based on the physical data of the airborne systems; classify and digitally preprocess the multi-source heterogeneous physical data of airborne system failures, and output structured basic fault analysis data; S2: Fault domain knowledge enhancement is achieved through retrieval-enhanced generation and a hybrid approach combining knowledge graphs and libraries; implicit fault mode reasoning and semantic expansion are performed using knowledge graphs, adding physical attribute relationships to each airborne system component and fault event node; physical entity linking and fault attribute reasoning are conducted using knowledge graphs to expand fault attributes, specifically: ; in, These are regional component failure parameters; The original fault attributes; For airborne system components child elements; for Adjacent component nodes in a knowledge graph; It is a set of fault reasoning rules; To trigger a fault event Influence function; Fault reasoning rules for airborne system components; To trigger a fault event; Conditional reasoning rules based on physical variables are introduced into the fault reasoning rule set, making fault mode reasoning constrained by real-time physical measurements; retrieval enhancement is used to generate fault quantification physical parameters, which are then fused with physical simulation parameters; structured fault paradigms and historical fault model libraries are matched; and a domain-enhanced fault dataset is generated through weighted fusion. ; S3: Domain-enhanced fault dataset based on the output of step S2 Build dynamic prompt word templates and use a large language model to generate an initial fault analysis model; S4: Obtain the initial fault analysis model generated in step S3, perform fault syntax and semantic verification and self-repair strategy, and realize automatic verification and optimization of the initial fault analysis model; S5: Compare the differences between the fault topology and the set of synchronization vectors based on the abstract syntax tree; use a large language model arbiter to align the fault logic semantics; generate a fault analysis model and analysis report with parameter accuracy verification.

[0007] Preferably, step S1 specifically includes: S11: Classify and digitally preprocess the unstructured fault text data, structured physical parameters, and modeled graphical data of airborne systems from multiple sources to extract physical model structural information. S12: Use natural language processing technology to extract fault text entities and business logic, and use a large language model for deep semantic parsing to extract fault text logical information; S13: Integrate the fault text logic information extracted in step S12 with the physical model structure information extracted in step S11, perform fault text semantic alignment and physical entity conflict resolution for multimodal information, and obtain the fused physical entity set. S14: Process the physical entity set merged in step S13 and output it as a structured fault text dataset. Output using a six-tuple format.

[0008] Preferably, step S13 specifically includes: ; in: The fused set of physical entities; The set of physical entities extracted from faulty text; The set of physical entities extracted from the image; The physical entity weights of the faulty text; Image primitive weights; This is the confidence weight matrix for the data source; This is a multimodal information fusion function; This is a feature fusion operation.

[0009] Preferably, step S2 specifically includes: S21: Use knowledge graphs for implicit fault mode reasoning and fault semantic expansion; for the component set output in step S14 By utilizing a pre-built knowledge graph of general aviation quality characteristics, physical entity linking and fault attribute reasoning are performed to expand fault attributes; S22: Retrieve physical parameters from unstructured standard specification libraries using retrieval enhancement generation techniques; S23: Matching structured fault paradigms with a historical fault model library; introducing a subgraph isomorphism algorithm to match fault topology with design paradigms in the historical model library; S24: Enhanced data fusion, outputting a domain-enhanced fault dataset. .

[0010] Preferably, step S23 specifically includes: ; in, The result is the optimization of the fault topology and the historical model library. This is the subgraph isomorphism determination function; A set of graph structures; This is the equivalence function for determining subgraph isomorphism; A subgraph induced for a subset of component nodes; To be a function that maximizes the value; The maturity weights for structured failure paradigms in the historical model library.

[0011] Preferably, step S3 specifically includes: S31: Domain-enhanced fault dataset output from step S2 Assemble prompt words used to drive large language models; S32: Use a context learning mechanism to analyze the type of the component to be generated and implement context-aware few-sample prompt injection; S33: Constructing chain-driven layered code; S34: The generated code is sent back to the large language model along with inspection instructions to perform initial code review.

[0012] Preferably, step S5 specifically includes: S51: Integrate fault text codes with structured source datasets The comparison is performed, and the fault text code is reverse-structured based on the abstract syntax tree; S52: Determine the set differences between the fault topology and the synchronization vector; use set theory methods to transform the source feature set in step S1. The code feature set extracted in step S51 By making comparisons, the differences in the physical structural dimensions can be determined; S53: Use a large language model determiner to align fault logic semantics; S54: For the failure rate and probability distribution required for quantitative analysis, the airborne system performs precise numerical comparisons to achieve physical parameter accuracy verification and generate difference reports.

[0013] On the other hand, the present invention provides a fault analysis model building system for a fault analysis model building method for analyzing the quality characteristics of airborne systems, which includes: a data parsing module, a model building module, a semantic verification and repair module, and a model publishing module; The data parsing module consists of a computing server, a data interface unit, and a storage unit, and is used to acquire and parse multi-source physical data of the airborne system. The model building module consists of a computing processing unit and a model generation engine. Based on the structured system physical data and combined with preset modeling rules and domain knowledge, it builds a quality characteristic analysis model of the airborne system, realizing automatic mapping from system physical data to analysis model. The semantic verification and repair module consists of a verification engine, a self-repair processing unit, and a rule base. It is used to perform syntactic correctness verification and semantic consistency analysis on the analysis model generated by the model building module, and automatically repair and optimize the model based on the verification results to ensure that the model is consistent with the original system physical data. The model publishing module consists of a model management unit and an interface service unit. It is used to manage and store the version of the verified analysis model and supports the physical application of general quality characteristic analysis of airborne systems.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention can significantly improve the efficiency of automated generation of general quality characteristic analysis models for airborne systems; through multimodal system physical data parsing, domain knowledge enhancement and large language model driven model generation mechanism, the automatic construction of AltaRica fault analysis model is realized. Compared with the traditional method of relying on manual modeling, the degree of manual participation is greatly reduced and the modeling cycle is shortened, thereby accelerating the efficiency of general quality characteristic analysis work such as safety and reliability of airborne systems in the scheme design stage and improving the physical implementation capability of complex system modeling and analysis.

[0015] (2) The present invention can improve the syntactic correctness, semantic consistency and physical credibility of the generated model; in the process of model generation, the present invention introduces syntactic and semantic verification, self-repair strategy and source data driven fault text semantic alignment and difference analysis mechanism, so that the generated model can be consistent with the physical source data of the system, avoid semantic drift and structural errors, thereby significantly improving the correctness, integrity and traceability of the model, and thus improving the accuracy and physical credibility of the analysis results of the airborne system in practical applications such as fault propagation analysis, safety assessment and reliability prediction.

[0016] (3) The present invention can form a closed-loop model generation and self-closed-loop release process for airborne system quality characteristic analysis; by designing and driving model generation, verification, repair and release, it combines the automatic generation capability and understanding and repair capability of large models to form a closed-loop process from data input, model generation, verification and repair to final release, thereby improving the level of automation while taking into account the reliability of the model, thus supporting the continuous iteration and rapid update of airborne system models in the design, verification and operation and maintenance stages, and improving the quality characteristic analysis capability and physical application efficiency of airborne systems throughout their entire life cycle. Attached Figure Description

[0017] Figure 1 This is a control block diagram of the fault analysis model construction method for airborne system quality characteristic analysis according to the present invention. Figure 2 This is a diagram illustrating the overall framework of an LLM-based hierarchical model generation method. Figure 3 This is a logical framework diagram of the atomic node generation method in the model layering generation method of this invention embodiment; Figure 4 This is a logical framework diagram of the system architecture and synchronous synthesis method in the model hierarchical generation method of this invention. Figure 5 This is a framework diagram of the model physical file generation, verification, and self-repair method in an embodiment of the present invention. Detailed Implementation

[0018] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

[0019] This invention proposes a method for constructing a fault analysis model for airborne system quality characteristic analysis, such as... Figure 1 As shown, the process involves: analyzing the mission requirements of a multimodal airborne system based on its physical data; enhancing fault domain knowledge using a hybrid approach of retrieval enhancement generation, knowledge graph, and database; generating an initial fault analysis model using a large language model; automatically validating and optimizing the initial fault analysis model; and generating a fault analysis model for analyzing the quality characteristics of the airborne system. The specific steps include: Step S1: Analyze the mission requirements of the multimodal airborne system based on the airborne system physical data.

[0020] Step S11: Classify and digitally preprocess the multi-source heterogeneous physical data of airborne system failures; first, complete the access and classification of the multi-source heterogeneous physical data of airborne system failures. The multi-source heterogeneous physical data mainly includes three categories, specifically: The first category is unstructured fault text data, which includes natural language descriptions of development task books, requirements specifications, design documents, and functional hazard assessment (FHA) reports.

[0021] The second category is structured physical parameters, which mainly refer to the qualitative and quantitative requirements for safety indicators.

[0022] The third category is model-based graphical data, which is a topological view describing the architecture of airborne systems and subsystems.

[0023] A differentiated processing strategy is adopted in the data preprocessing stage. For unstructured text, since the original documents often contain complex hierarchical descriptions, semantic-based content segmentation is first performed to extract the target airborne system content related to general quality characteristic analysis and remove irrelevant redundant information. For structured parameters, a numerical mapping table is directly established. For the third type of graphic data, the focus is on parsing the system modeling language SysML (Systems Modeling Language) and block definition diagrams (BDD) stored in Extensible Markup Language (XML), XML Metadata Interchange (XML Metadata Interchange) format. Automated parsing scripts are used to read the model files, extract primitive elements, and transform them into intermediate representations in structured lightweight data exchange formats such as JSON and JavaScript Object Notation. Specific extraction algorithms and mapping rules are detailed in Table 1.

[0024] Table 1. System Modeling Language SysML - Block Graph BDD Extraction Algorithm Rules Step S12: Use natural language processing (NLP) technology to extract fault text entities and business logic; for text and table data, use a large language model (LLM) fine-tuned for the airborne system's physical domain for deep semantic parsing, such as... Figure 2 The diagram shows the overall framework of the LLM-based hierarchical model generation method. For unstructured fault text data, named entities are extracted using Named Entity Recognition (NER). ; in, Name entities for airborne systems; Names of airborne system components; These are the state variables of the airborne system; It serves as the input / output interface for the airborne system.

[0025] Identify the control logic and data flow relationships between components through relationship extraction. Transforming the logic described in natural language into formalized implication relations Condition-action pairs. Taking threshold control as an example, the airborne system can parse the semantics and generate logical assertions in the form of system behavior constraint assert(Sensor.value>Threshold)->(Valve.status:=Closed).

[0026] Table 2 Formal Specifications and Model Element Matching Table Step S13: Semantic alignment of fault text and resolution of physical entity conflicts in multimodal information; merging the logical information of fault text extracted in step S12 and the structural information of physical models extracted in step S11 to resolve potential inconsistencies between multimodal data. The airborne system establishes a mapping relationship by calculating the semantic similarity between fault text entities and image primitives. For example, the "main pump" described in the text data is identified as "Main_Pump" in the architecture diagram, and the airborne system associates the two as the same model object. When a conflict occurs, such as the text describing component A as connected to B, while the diagram shows A connected to C, the airborne system generates a conflict report based on a preset confidence weight; this fusion process is represented by the following formula: ; in: The fused set of physical entities; The set of physical entities extracted from faulty text; The set of physical entities extracted from the image; The physical entity weights of the faulty text; Image primitive weights; This is the confidence weight matrix for the data source; This is a multimodal information fusion function; The feature fusion operation specifically includes: a union operation, which takes the union of the fault text entity and the image primitive entity. For example, if the text mentions "valve A" and the drawing mentions "valve B", the operation will retain both and generate... Attribute merging: For entities where both faulty text entities and image primitives exist, the associations of the faulty text entities and the attributes of the image primitives will be merged; Conflict detection: For different descriptions with the same attributes on both sides, a conflict flag is triggered, and a conflict is detected. Conflict handling is performed on the weight matrix; The entire data consists of input data.

[0027] In the above formula As a global adjustment factor, it satisfies normalization, i.e. , here It should be determined based on the actual physical scenario. For example, in the early stages of physical design, only preliminary documents are available, and the architecture diagrams are usually not complete. In this case, it can be... =0.2, making the generation calculations more dependent on the physical document. During the physical detailed design phase, when the system modeling language SysML model has been reviewed and set... =0.7, requiring that the generated calculations be based on the SysML model as the main body, with document information as an additional supplement to the model information.

[0028] This step enables the output of a unified, structured fault text dataset that possesses characteristics of a single data source and is semantically consistent.

[0029] Step S14: Output structured basic fault analysis data for key modeling; In order to transform the fused multimodal information into a computer-readable standard input, taking advantage of the characteristics of AltaRica Dataflow, the unified structured fault text dataset from step S13 is processed and output as a structured fault text dataset. This can be represented using a six-tuple: ; ; , , ; ; ; ; in, This is a structured fault text dataset; For airborne system components and their affiliation; For the first individual airborne system components child elements; This is a unique identifier for the component, such as Main_Pump_A; type is the node template type to which the component belongs, such as PumpNode; parent is the parent component ID, and if it is a top-level airborne system, then... ; For component indexing; This represents the total number of components. Set for internal state variables; For state variables; For stream variables; The name of the state variable; For the state value range; This is the initial state value; The name of the stream variable; For the range of the stream variable; For the direction of the flow variable; For interface connection relationships, each connection It is transformed into an equality constraint: This indicates the component The output directly determines the component. enter; For the first One interface; For output interface; For input interface; For interface index; Total number of interfaces; State transition logic rules; For state transition logic triples; For conditional functions, ; To trigger a fault event; Update the mapping for the state. Rules for example: This indicates that when temp is greater than 100, the overheat event is triggered, and an update is performed. ; For the fault propagation and synchronization path, the synchronization vector to AltaRica; Synchronization vector; For the parent node, or events observed at the airborne system level; For atomic events occurring in sub-components; when all in the vector When the conditions for occurrence are met; Forced triggering is used to accurately describe common cause failures or cascading effects; To trace metadata, record the source of data items, such as physical documents and block diagram (BDD) model elements.

[0030] Step S2: Enhance fault domain knowledge using a hybrid RAG-KG-library driver.

[0031] Step S21: Use a knowledge graph (KG) for implicit fault mode reasoning and fault semantic expansion, adding physical attribute nodes and relationships to each airborne system component and fault event node; for the component set output in step S14 The system utilizes a pre-constructed knowledge graph of general aviation quality characteristics to link physical entities and infer fault attributes. The components analyzed in step S1 often only contain functional descriptions such as "electric pump," lacking the failure mode information required for safety analysis. Through a graph traversal algorithm, the component's position within the ontology hierarchy is identified, for example: It inherits common failure modes from its parent class, such as "internal leakage" and "bearing seizure." Furthermore, the graph can infer implicit propagation paths based on the physical connections of components, such as identifying the physical proximity of "hydraulic lines" and "electrical harnesses," thereby automatically marking potential area safety faults. This reasoning process is represented by attribute expansion functions, specifically: ; in, These are regional component failure parameters; The original fault attributes extracted in step S1; For airborne system components Sub-elements are extracted from step S1; for Adjacent component nodes in a knowledge graph; This is a set of fault reasoning rules, such as has_failure_mode; To trigger a fault event Influence function; Fault reasoning rules for airborne system component elements.

[0032] By introducing conditional reasoning rules based on physical variables into the fault reasoning rule set, fault mode reasoning no longer depends solely on topological relationships, but is also constrained by real-time physical measurements.

[0033] Step S22: Use retrieval enhancement to generate accurate RAG physical parameters for fault quantification and integrate them with physical simulation parameters; to meet the needs of quantitative safety assessment, use retrieval enhancement to generate RAG technology to retrieve accurate physical parameters from an unstructured standard specification library; for each component Generate specific query vectors Retrieve and calculate the most relevant vector data from the vector database. The document fragment is as follows: ; ; in, For components Most relevant A collection of document fragments; Enhanced generating functions for retrieval; It is a cosine similarity metric function; These are document fragments from a vector database; This is a vector encoding function used to convert component information into a measurable vector representation. This vector encoding function is obtained by configuring the parameters of the physical corpus and aligning them with semantics. The output query vector is a fixed-length vector. .

[0034] The Large Oracle Model (LLM) was used to extract key parameters from the search results, including basic failure rate, repair rate, and exposure time, along with the confidence level of the data source.

[0035] ; in, For components The basic failure rate; For components The repair rate; For components Exposure time; This is a model for the Great Prophecy.

[0036] Confidence weights of the parameter data source for: ; in, The confidence weights for the parameter data sources; This is a confidence score calculation function based on document consistency and source reliability.

[0037] For probability distribution parameters required by the AltaRica 2.0 model, such as the exponential distribution rate, the retrieval-enhanced generation (RAG) technique can automatically perform unit conversion and normalization. For example, it can convert the "probability of failure per flight hour" in the document into the external interface declaration extern law setting required by the model.

[0038] Step S23: Matching structured fault paradigms and historical fault model libraries; To improve the standardization and generation efficiency of the model, a subgraph isomorphism algorithm is introduced to match the fault topology extracted in step S1 with mature design paradigms in the historical model library. When matching structured fault paradigms with the historical fault model library, physical simulation models, such as finite element thermo-mechanical coupling models, are used to generate quantitative parameters under specific working conditions, such as thermal stress distribution and vibration modal frequencies, replacing or weighting and fusing the original purely statistical or empirical parameters.

[0039] For the fault topology, it can be formalized as a graph structure. ,in, Let be the set of nodes and the set of edges of the graph, respectively. Then the historical model library can be represented as a set of graph structures. The superscript H indicates a historical model.

[0040] Introducing a subgraph isomorphism determination function: Equivalent representation is ,in, For a subgraph induced by a subset of nodes, to improve the reliability in the case of multiple subgraphs, an optimization based on matching the fault topology with the historical model library is introduced: ; in, The result is the optimization of the fault topology and the historical model library. This is the subgraph isomorphism determination function; A set of graph structures; This is the equivalence function for determining subgraph isomorphism; A subgraph induced for a subset of component nodes; To find the maximum value of the function; The maturity weights for structured failure paradigms in the historical model library.

[0041] The focus is on identifying classic redundant structures, such as 2oo3 voting logic, dual-redundancy hot backup, and standard component interfaces. Once a match is found, the coarse logic generated in step S1 will be replaced with a verified AltaRica 2.0 Node template from the library.

[0042] For example, when a "triple-redundant flight control computer" structure is identified, the Voter_2oo3 standard node in the library is automatically invoked, instead of regenerating an unverified piece of logic code. This not only corrects potential logic vulnerabilities but also ensures that the generated code conforms to industry best practices.

[0043] Step S24: Enhanced Data Fusion and Output Settings; The failure modes of the knowledge graph (KG) obtained from the above three sub-steps, the quantitative parameters of the retrieval-enhanced RAG, the standard paradigm of the library, and the structured basic fault analysis data from S1 are weighted and fused to generate a domain-enhanced fault dataset. During the fusion process, a confidence-based conflict resolution strategy is adopted: for structural logic, historical model libraries are given priority; for quantitative parameters, the latest standards of RAG retrieval generated by retrieval enhancement are given priority; for qualitative fault descriptions, KG inference results are given priority.

[0044] The augmented dataset output in step S2: ; in, These are standardized component objects aligned with KG entities from a knowledge graph; For the set of failure modes, It can be used not only to identify failed modules, but also to correlate failure rates. For example, for the "valve stuck" pattern, the failure rate can be correlated by generating a RAG through enhanced retrieval. ; This provides a complete set of physics and probability parameters. It includes the initial state settings (init state values), flow variable type constraints, and external interface declarations (extern probability law settings) required by AltaRica 2.0. This is a pointer to the standard model library. If a standard paradigm is matched, this field points to the path of the .alt file in the library, such as lib:PowerSystem / Generator_AC_v2.alt, indicating that the code template should be reused directly in step S3.

[0045] Step S3: Generate the initial AltaRica fault analysis model using the Large Language Model (LLM).

[0046] Step S31: Dynamic prompt word template construction; first, based on the dataset output in step S2... The system dynamically assembles prompt words to drive the Large Language Model (LLM). The LLM is a language model optimized for general quality analysis. The prompt word structure adopts a three-part architecture of "role setting + task constraints + data context." Role setting: The role of "Senior General Quality Characteristics Safety Analysis Physicist" is given to the Large Language Model LLM, emphasizing that it must be proficient in AltaRica Dataflow syntax, especially the strict use of keywords such as node, sub, sync, assert.

[0047] Task constraints: The use of AltaRica 3.0's class syntax is explicitly prohibited, the generation of edon closures is mandatory, and variable declarations must include type annotations.

[0048] Data context: The JSON fragment is serialized into the input part of the Prompt.

[0049] The formal expression of this process is: ; in, Set information for the character; These are constraint rules based on AltaRica 2.0; for JSON fragment.

[0050] Step S32: Context-aware few-sample cue injection; To prevent grammatical illusions from the large language model LLM, a retrieval-based context learning mechanism is introduced. Analyze the type of the component to be generated, such as "hydraulic pump" or "voting system", and retrieve 3-5 of the most similar high-quality AltaRica2.0 code snippets from the historical codebase as "standard examples".

[0051] Strategy: If the JSON contains a complex "Sync Vector", the Prompt will inject a correct code example containing a complex sync vector clause, allowing the Large Language Model (LLM) to learn how to write synchronization logic by imitation.

[0052] Effect: By providing<Input:JSON_Example,Output:Code_Example> Paired examples significantly improve the syntactic accuracy of generated code through few-shot learning.

[0053] Step S33: Chain-driven layered code construction; The Large Language Model (LLM) does not output all the code at once, but is guided to generate it step by step following the logic of the thought chain. The generation task is broken down into two sub-tasks: "atomic node generation" and "system architecture assembly".

[0054] like Figure 3 The diagram shown is a logical framework diagram of the atomic node generation method in the hierarchical generation method of the model according to an embodiment of the present invention. Specifically, subtask A: atomic node generation, Prompt guides the large language model LLM to execute according to the following thought chain: Reasoning: Analyze the variables in the JSON to distinguish between state variables and flow variables.

[0055] Reasoning: Analyze the logic in transitions to make the Guard conditions mutually exclusive.

[0056] Generation: Write node settings and insert state, flow, event, and trans code blocks.

[0057] Generate: Based on the probability parameters in the JSON, append the external interface declaration and the description of the law.

[0058] The logic of this task process can be found in the atomic node generation diagram in the attached diagram of the instruction manual.

[0059] like Figure 4 The diagram shown illustrates the system architecture and synchronous synthesis method logic framework in the model hierarchical generation method of this invention, specifically subtask B: airborne system architecture and synchronous synthesis; the prompt requires the large language model LLM to explicitly handle the common cause failure path output by S2: Reasoning: Identify sub_instances in the JSON and generate sub-component instantiation sub clauses.

[0060] Reasoning: Analyze the synchronized list to understand the broadcast relationship between parent and child events.

[0061] Generation: Write complex synchronization vectors (sync).

[0062] For example, a large language model (LLM) needs to infer the transformation of ["power_loss","pump1.fail","pump2.fail"] in JSON into a code synchronization vector (sync).<power_loss,pump1.fail,pump2.fail> .

[0063] Generation: Write the system behavior constraint assert connection logic, establish component connections and generate architecture code.

[0064] Step S34: Initial Code Review Based on Introspection; After the Large Language Model (LLM) generates preliminary code, the airborne system triggers a lightweight "introspection" round. The generated code is sent back to the LLM with a check instruction: "Please check the above code for unclosed parentheses, undefined variables, or syntax not supported by AltaRica 2.0, such as extends." The LLM outputs a corrected version based on this instruction. This step effectively filters out 90% of low-level syntax errors, reducing the pressure on subsequent S4 steps. An example of an AltaRica fault analysis model output is shown below: node Valve Event stuck? state status: {Working, Failed}; init status := Working; trans status = Working |- stuck ->status := Failed; law stuck ~ exponential(1e-5); end Step S4: Model Syntax and Semantic Validation and Self-Repair Strategy; Addressing the common syntax and semantic issues in the initial AltaRica fault analysis model generated from the large language model in Step S3, a static validation and self-repair method for the AltaRicaDataFlow version is proposed, such as... Figure 5The diagram shown is a framework diagram of the model physical file generation, verification and self-repair method of an embodiment of the present invention. Specifically, a set of rules is set to formalize the language constraints of AltaRica DataFlow into a set of decisionable rules.

[0065] ; And a deterministic repair mapping is set for each type of violation pattern as follows: ; This enables automatic verification and optimization of the initial AltaRica fault analysis model, as shown in Table 3.

[0066] Table 3. Statistics on Automatic Verification and Optimization of the AltaRica Fault Analysis Model Step S5: Source data-driven model semantic alignment and difference analysis.

[0067] Step S51: Reverse structuring of the code based on the Abstract Syntax Tree (AST); in order to connect the faulty text code with the structured source dataset. For comparison, the airborne system first uses the AltaRica parser to generate an Abstract Syntax Tree (AST). The airborne system then traverses the AST, reverse-engineering the "actual topological structure" described by the code to construct a code-side feature set. for: ; in: This refers to the actual set of nodes implemented in the code. For the topology connections built in the code by instantiating sub components and system behavior constraints assert; The synchronization vector set by the `sync` clause in the code; Set the parameter values ​​for init and extern declarations of external interfaces to define the initial state in the code.

[0068] Step S52: Calculate the set difference between the topology and the synchronization vector. Using set theory, the source feature set from step S1 is... The code feature set extracted in step S51 A comparison is made to determine the differences in physical structure dimensions. Two types of errors are specifically detected: Omission: It exists in the source data but not in the code.

[0069] ; For example, the Functional Hazard Assessment (FHA) report requires "dual-pump common cause failure," but this vector is not included in the code's synchronization vector list.

[0070] Redundancy: It exists in the code but not in the source data.

[0071] ; For example, the Large Language Model (LLM) generates a non-existent node.

[0072] For the calculated difference values, the severity of the difference is divided into high, medium, and low by setting an interval range.

[0073] Step S53: Use the Large Language Model (LLM) discriminator to perform semantic alignment of fault logic; for state transition logic, due to the gap in expression between the natural language description in step S1 and the formal code in step S4, simple character comparison is not possible. The airborne system introduces the Large Language Model (LLM) as a semantic discriminator.

[0074] Airborne system construction verification This suggests that the Large Language Model (LLM) evaluates the logical equivalence of the two.

[0075] enter: Source description: "The valve closes when the temperature sensor reading exceeds the threshold and the main power supply is normal." Code: trans(temp>threshold)&(power_ok)|-close_event->valve:=closed.

[0076] The decision logic is as follows: ; like If it is, it is marked as a semantic inconsistency fault, where The threshold value is set according to the requirements.

[0077] Referring to step S52, the semantic judgment result value is divided into high, medium, and low severity by setting an interval range.

[0078] Step S54: Parameter accuracy verification and difference report generation; the airborne system performs precise numerical comparisons based on the failure rate and probability distribution required for quantitative analysis.

[0079] For example, if the source file is 1.5e-6 and the generated code is 1.5e-5, it will be considered a difference.

[0080] Airborne system setting parameter deviation function: ; Refer to step S52, By setting a range, its severity can be divided into high, medium, and low.

[0081] For example, the high severity range of the parameter deviation function value is set to (0.01, 1), that is, if If the error exceeds 1%, the airborne system will consider it a high-severity difference.

[0082] Step S5 outputs a structured difference report. The report data structure is: [Difference Type, Location, Severity, Description], where the type is one of the following: structural omission, structural redundancy, semantic conflict, or parameter anomaly; the location is the line number and the source document page number, based on the information from step S14. Source mapping is used to obtain the severity level, which is either high, medium, or low. The description is automatically generated based on the type. For example, the description for a missing type difference item in a structure is "The synchronization vector sync on line 45 of the code is missing the pump2.fail event, which is set on page 12 of the Functional Hazard Assessment (FHA) document."

[0083] Model generation and self-closed-loop deployment for performing airborne system quality characteristic analysis.

[0084] Model file generation, dynamic verification, and automatic repair: Based on the AltaRica fault analysis model code corrected in step S4, the DataFlow model physical file is generated using the AltaRica modeling tool. The compiler is used to parse and verify the model file. For syntax problems encountered during compilation and semantic differences output in step S5, fragment repair is performed using a Large Language Model (LLM). The specific steps are as follows: Let the set of problems be: ,in For grammar problem elements, These are semantic difference elements, each containing the question type, location, and description.

[0085] For E-formation repair units: ),in For the question element, For the present Context code, This refers to the syntax rules for AltaRica DataFlow.

[0086] Define the constrained repair function ,in, For the constrained language model repair function, output For code snippets The repaired fragment. Constraints are: That is, the scope of repair must not exceed the structural boundary of the original problematic fragment.

[0087] The repaired code will be recompiled and verified. If it passes the verification, the repaired code will be retained. If it fails, the repair unit will be dynamically updated. , This involves adding invalid results to repair and obtain new fragments. Repeat this step until a specified number of repairs is reached, or until the correct fragment is obtained, depending on actual needs.

[0088] Model release: After successful model compilation and verification, a release version is generated for use in airborne system analysis, simulation, or fault assessment. The final model conforms to the AltaRica DataFlow specification and can be directly used for quantitative analysis and to support the analysis of general quality characteristics of airborne systems.

[0089] The second aspect of the present invention provides a fault analysis model construction system for a fault analysis model construction method for airborne system quality characteristic analysis, comprising: a data parsing module, a model construction module, a semantic verification and repair module, and a model publishing module.

[0090] The data parsing module consists of a computing server, a data interface unit, and a storage unit. It is used to acquire and parse multi-source physical data of the airborne system, including requirement data, structural data, and functional logic data. It also performs structured processing on the parsing results to generate a unified data representation. The data parsing module is connected to the model building module to provide standardized input data to the model building module.

[0091] The model building module consists of a computing processing unit and a model generation engine. Based on the structured system physical data and combined with preset modeling rules and domain knowledge, it constructs a quality characteristic analysis model of the airborne system, realizing automatic mapping from system physical data to the analysis model. The model building module is connected to the data parsing module and the semantic verification and repair module, respectively, and outputs the generated initial AltaRica fault analysis model to the semantic verification and repair module.

[0092] The semantic verification and repair module consists of a verification engine, a self-repair processing unit, and a rule base. It is used to perform syntactic correctness verification and semantic consistency analysis on the analysis model generated by the model building module, and automatically repair and optimize the model based on the verification results to ensure that the model is consistent with the original system physical data. The semantic verification and repair module is connected to the model publishing module and outputs the verified and repaired model to the model publishing module.

[0093] The model publishing module consists of a model management unit and an interface service unit. It is used to manage, store, and publish the validated analysis models, enabling them to be called by airborne system security analysis, reliability assessment, and fault propagation analysis tools, thereby supporting the physical application of general quality characteristic analysis of airborne systems. The model publishing module connects with external analysis systems to realize the physical application of the models.

[0094] The beneficial effects of the embodiments of the present invention are as follows: Compared with the traditional method that relies on manual modeling, the present invention significantly reduces the degree of human involvement, shortens the modeling cycle, accelerates the efficiency of general quality characteristic analysis work such as safety and reliability of airborne systems in the scheme design stage, and improves the physical implementation capability of complex system modeling and analysis; further improves the accuracy and physical credibility of analysis results of airborne systems in practical applications such as fault propagation analysis, safety assessment and reliability prediction; supports the continuous iteration and rapid updating of airborne system models in the design, verification and operation and maintenance stages, and improves the quality characteristic analysis capability and physical application efficiency of the system throughout its entire life cycle.

[0095] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for constructing a fault analysis model for airborne system quality characteristic analysis, characterized in that, It includes: S1: Analyze the mission requirements of the multimodal airborne system based on the airborne system physical data; Classify and digitally preprocess multi-source heterogeneous physical data of airborne system failures to output structured basic failure analysis data; S2: Fault domain knowledge enhancement is achieved through retrieval-enhanced generation and a hybrid approach combining knowledge graphs and libraries; implicit fault mode reasoning and semantic expansion are performed using knowledge graphs, adding physical attribute relationships to each airborne system component and fault event node; physical entity linking and fault attribute reasoning are conducted using knowledge graphs to expand fault attributes, specifically: ； in, These are regional component failure parameters; The original fault attributes; For airborne system components child elements; for Adjacent component nodes in a knowledge graph; It is a set of fault reasoning rules; To trigger a fault event Influence function; Fault reasoning rules for airborne system components; To trigger a fault event; Conditional reasoning rules based on physical variables are introduced into the fault reasoning rule set, making fault mode reasoning constrained by real-time physical measurements; retrieval enhancement is used to generate fault quantification physical parameters, which are then fused with physical simulation parameters; structured fault paradigms and historical fault model libraries are matched; and a domain-enhanced fault dataset is generated through weighted fusion. ; S3: Domain-enhanced fault dataset based on the output of step S2 Build dynamic prompt word templates and use a large language model to generate an initial fault analysis model; S4: Obtain the initial fault analysis model generated in step S3, perform fault syntax and semantic verification and self-repair strategy, and realize automatic verification and optimization of the initial fault analysis model; S5: Compare the differences between the fault topology and the set of synchronization vectors based on the abstract syntax tree; use a large language model arbiter to align the fault logic semantics; generate a fault analysis model and analysis report with parameter accuracy verification.

2. The fault analysis model construction method for airborne system quality characteristic analysis according to claim 1, characterized in that: Step S1 is as follows: S11: Classify and digitally preprocess the unstructured fault text data, structured physical parameters, and modeled graphical data of airborne systems from multiple sources to extract physical model structural information. S12: Use natural language processing technology to extract fault text entities and business logic, and use a large language model for deep semantic parsing to extract fault text logical information; S13: Integrate the fault text logic information extracted in step S12 with the physical model structure information extracted in step S11, perform fault text semantic alignment and physical entity conflict resolution for multimodal information, and obtain the fused physical entity set. S14: Process the physical entity set merged in step S13 and output it as a structured fault text dataset. Output using a six-tuple format.

3. The fault analysis model construction method for airborne system quality characteristic analysis according to claim 1, characterized in that: Step S13 is as follows: ； in: The fused set of physical entities; The set of physical entities extracted from faulty text; The set of physical entities extracted from the image; The physical entity weights of the faulty text; Image primitive weights; This is the confidence weight matrix for the data source; This is a multimodal information fusion function; This is a feature fusion operation.

4. The fault analysis model construction method for airborne system quality characteristic analysis according to claim 1, characterized in that: Step S2 is as follows: S21: Use knowledge graphs for implicit fault mode reasoning and fault semantic expansion; for the component set output in step S14 By utilizing a pre-built knowledge graph of general aviation quality characteristics, physical entity linking and fault attribute reasoning are performed to expand fault attributes; S22: Retrieve physical parameters from unstructured standard specification libraries using retrieval enhancement generation techniques; S23: Matching structured fault paradigms with a historical fault model library; introducing a subgraph isomorphism algorithm to match fault topology with design paradigms in the historical model library; S24: Enhanced data fusion, outputting a domain-enhanced fault dataset. .

5. The fault analysis model construction method for airborne system quality characteristic analysis according to claim 1, characterized in that: Step S23 is as follows: ； in, The result is the optimization of the fault topology and the historical model library. This is the subgraph isomorphism determination function; A set of graph structures; This is the equivalence function for determining subgraph isomorphism; A subgraph induced for a subset of component nodes; To find the maximum value of the function; The maturity weights for structured failure paradigms in the historical model library.

6. The fault analysis model construction method for airborne system quality characteristic analysis according to claim 1, characterized in that: Step S3 is as follows: S31: Domain-enhanced fault dataset output from step S2 Assemble prompt words used to drive large language models; S32: Use a context learning mechanism to analyze the type of the component to be generated and implement context-aware few-sample prompt injection; S33: Constructing chain-driven layered code; S34: The generated code is sent back to the large language model along with inspection instructions to perform initial code review.

7. The fault analysis model construction method for airborne system quality characteristic analysis according to claim 1, characterized in that: Step S5 is as follows: S51: Integrate fault text codes with structured source datasets The comparison is performed, and the fault text code is reverse-structured based on the abstract syntax tree; S52: Determine the set differences between the fault topology and the synchronization vector; use set theory methods to transform the source feature set in step S1. The code feature set extracted in step S51 By making comparisons, the differences in the physical structural dimensions can be determined; S53: Use a large language model determiner to align fault logic semantics; S54: For the failure rate and probability distribution required for quantitative analysis, the airborne system performs precise numerical comparisons to achieve physical parameter accuracy verification and generate difference reports.

8. A fault analysis model construction system for the fault analysis model construction method for airborne system quality characteristic analysis as described in any one of claims 1 to 7, characterized in that, It includes: The module includes a data parsing module, a model building module, a semantic verification and repair module, and a model publishing module. The data parsing module consists of a computing server, a data interface unit, and a storage unit, and is used to acquire and parse multi-source physical data of the airborne system. The model building module consists of a computing processing unit and a model generation engine. Based on the structured system physical data and combined with preset modeling rules and domain knowledge, it builds a quality characteristic analysis model of the airborne system, realizing automatic mapping from system physical data to analysis model. The semantic verification and repair module consists of a verification engine, a self-repair processing unit, and a rule base. It is used to perform syntactic correctness verification and semantic consistency analysis on the analysis model generated by the model building module, and automatically repair and optimize the model based on the verification results to ensure that the model is consistent with the original system physical data. The model publishing module consists of a model management unit and an interface service unit. It is used to manage and store the version of the verified analysis model and supports the physical application of general quality characteristic analysis of airborne systems.

Images