A Functional Hazard Intelligent Analysis Method and Device Based on MBSE-LLM

Abstract

This application relates to the interdisciplinary field of artificial intelligence and reliability systems engineering, and provides a functional hazard intelligent analysis method and apparatus based on MBSE-LLM. This method constructs a functional model of the analysis object using the MBSE model, obtains an enhanced vector database by injecting retrieved augmented knowledge, then determines the basic description and related documents of the target function, merges the two to obtain an enhanced description, generates a hazard factor list based on the enhanced description, and finally determines the FHA result of the target function based on the hazard factor list. This method overcomes the shortcomings of traditional FHA relying on manual completion by domain experts, comprehensively utilizes model-based systems engineering methods and large language model technology, and the resulting intelligent FHA method can significantly optimize the overall process of functional hazard analysis, effectively improve the accuracy of FHA results and the efficiency of analysis work, and is beneficial for guiding the efficient iteration of traditional FHA.

Description

Technical Field

[0001] This application relates to the interdisciplinary field of artificial intelligence and reliability systems engineering, and in particular to a functional hazard intelligent analysis method and device based on MBSE-LLM. Background Technology

[0002] Functional Hazard Analysis (FHA) is a tool for identifying hazards by rigorously evaluating the functionality of a system or subsystem (including software). In traditional analysis models, the work mainly relies on domain experts to manually advance the process. This approach is not only inefficient and heavily influenced by subjective factors, but it is also prone to overlooking potential functional hazard patterns and is difficult to meet the needs of rapid iteration for complex systems.

[0003] With the development of artificial intelligence, large language models (LLMs) have demonstrated powerful capabilities in knowledge reasoning and text generation, making intelligent functional hazard analysis (FHA) possible. However, directly using LLMs for FHA can easily lead to "illusion" phenomena, generating erroneous results that do not conform to engineering realities, and lacking a precise understanding of domain knowledge.

[0004] To alleviate this problem, Retrieval-augmented Generation (RAG) technology has been introduced, which improves the accuracy of responses by retrieving external knowledge. However, in practical applications, the performance of RAG technology heavily depends on the text slicing strategy, the quality of the embedding model, and the size of the vector database. It commonly suffers from limitations such as incomplete recall information, weak semantic relevance, and insufficient understanding of cross-document logic. Therefore, there is an urgent need for an intelligent FHA method that can deeply integrate systems engineering knowledge and simultaneously improve semantic understanding and reasoning capabilities to address these technical issues. Summary of the Invention

[0005] In view of this, the present application provides a functional hazard intelligent analysis method and apparatus based on MBSE-LLM to solve the problems of FHA relying on manual labor, low efficiency, and easy omission in the prior art.

[0006] A first aspect of this application provides a functional hazard intelligent analysis method based on MBSE-LLM, comprising:

[0007] The functional model of the analysis object is constructed using the Model-Based Systems Engineering (MBSE) approach;

[0008] By injecting retrieval enhancement knowledge into the functional model, an enhanced vector database is obtained; the vector database includes the enhanced recall results for each function in the functional model.

[0009] The basic description of the target function is determined based on the functional model, and the associated documents of the target function are retrieved from the enhanced vector database; the target function is any function in the functional model.

[0010] By using a large language model (LLM) to fuse basic descriptions and related documents, an enhanced description of the target function is obtained.

[0011] Based on the enhanced description, guiding prompts are constructed for each prompt in the preset set of hazard cues, and LLM is used to generate a list of hazard elements for the target function based on the guiding prompts;

[0012] At least the functional hazard analysis (FHA) results for the target function should be determined based on the list of hazard elements.

[0013] A second aspect of this application provides a functional hazard intelligent analysis device based on MBSE-LLM, comprising:

[0014] The building module is configured to use the MBSE method to build a functional model of the analysis object;

[0015] The enhancement module is configured to inject retrieval enhancement knowledge into the functional model, resulting in an enhanced vector database; the vector database includes the enhanced recall results for each function in the functional model.

[0016] The determination module is configured to determine the basic description of the target function based on the functional model and retrieve the associated documents of the target function from the enhanced vector database; the target function is any function in the functional model.

[0017] The fusion module is configured to use the Large Language Model (LLM) to fuse basic descriptions and related documents to obtain an enhanced description of the target function.

[0018] The generation module is configured to construct guiding prompts for each clue in a preset set of hazard clues based on enhanced descriptions, and use LLM to generate a list of hazard elements for the target function based on the guiding prompts;

[0019] The analysis module is configured to determine the functional hazard analysis (FHA) results of the target function based at least on the list of hazard elements.

[0020] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0021] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method.

[0022] The beneficial effects of this application embodiment compared with the prior art are as follows: This application embodiment constructs a functional model of the analysis object using the MBSE model, obtains an enhanced vector database by injecting retrieval augmentation knowledge, then uses the functional model to determine the basic description of the target function and retrieves related documents from the enhanced vector database, merges the two to obtain an enhanced description, generates a hazard factor list based on the enhanced description, and then determines the FHA result of the target function based on the hazard factor list. This method overcomes the shortcomings of traditional FHA relying on domain experts for manual completion, comprehensively utilizes model-based systems engineering methods and large language model technology, and the resulting intelligent FHA method can significantly optimize the overall process of functional hazard analysis, effectively improve the accuracy of FHA results and the efficiency of analysis work, and is conducive to guiding the efficient iteration of traditional FHA. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating a functional hazard intelligent analysis method based on MBSE-LLM provided in an embodiment of this application.

[0025] Figure 2 This is a schematic diagram of the hierarchical structure of a certain ground mobile platform.

[0026] Figure 3 This is a schematic diagram of the second-level logical model of a certain ground mobile platform.

[0027] Figure 4 This is a text illustration converted into Markdown format, using the first page of the ground mobile platform design document as an example.

[0028] Figure 5 This is a flowchart illustrating another functional hazard intelligent analysis method based on MBSE-LLM provided in this application embodiment.

[0029] Figure 6 This is a diagram of the MBSE-guided LLM functional hazard analysis interaction mechanism provided in the embodiments of this application.

[0030] Figure 7This is a schematic diagram of a functional hazard intelligent analysis device based on MBSE-LLM provided in an embodiment of this application.

[0031] Figure 8 This is a schematic diagram of the electronic device provided in the embodiments of this application. Detailed Implementation

[0032] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0033] The following will describe in detail, with reference to the accompanying drawings, a functional hazard intelligent analysis method and apparatus based on MBSE-LLM according to embodiments of this application.

[0034] As mentioned above, directly using LLM for functional hazard analysis can easily lead to "illusion" phenomena, generating erroneous results that do not conform to engineering reality, and lacking a precise understanding of domain knowledge. While introducing RAG can improve the accuracy of the answers, this technique generally suffers from limitations in application, such as incomplete recall information, weak semantic relevance, and insufficient understanding of cross-document logic.

[0035] In view of this, this application provides a functional hazard intelligent analysis method based on MBSE-LLM. This method integrates the structured knowledge expression capability of MBSE with the outstanding advantages of LLM in semantic deep understanding and logical reasoning. It can efficiently realize the automated identification and intelligent assessment of functional hazards of systems and subsystems (including software), and solve the pain points of traditional FHA that rely on manual labor, are inefficient, and are prone to omissions from a technical point of view.

[0036] Figure 1 This is a flowchart illustrating a functional hazard intelligent analysis method based on MBSE-LLM provided in an embodiment of this application. Figure 1 As shown, the method includes the following steps:

[0037] In step S101, the functional model of the analysis object is constructed using the MBSE method.

[0038] In step S102, retrieval enhancement knowledge is injected into the functional model to obtain the enhanced vector database.

[0039] The vector database includes enhanced recall results for each function in the functional model.

[0040] In step S103, the basic description of the target function is determined based on the functional model, and the associated documents of the target function are retrieved from the enhanced vector database.

[0041] The target function is any function in the functional model.

[0042] In step S104, the basic description and associated documents are merged using LLM to obtain an enhanced description of the target function.

[0043] In step S105, a guiding prompt is constructed for each prompt in the preset set of hazard prompts based on the enhanced description, and an LLM is used to generate a list of hazard elements for the target function based on the guiding prompts.

[0044] In step S106, the FHA results for the target function are determined at least based on the list of hazard factors.

[0045] In some embodiments of this application, the method may be executed by a server or by a terminal device with certain processing capabilities.

[0046] In some embodiments of this application, a functional model of the analysis object can first be constructed using the MBSE method. Retrieval enhancement knowledge is then injected into the functional model to obtain an enhanced vector database. By injecting retrieval enhancement knowledge into the functional model, enhanced recall results for each function in the functional model can be obtained. These enhanced recall results include knowledge slices from the injected retrieval enhancement knowledge whose relevance to each function is greater than a preset threshold. Furthermore, the enhanced vector database is constructed using the enhanced recall results of each functional model to facilitate knowledge retrieval during subsequent FHA analysis of each function.

[0047] In some embodiments of this application, a basic description of the target function can be determined based on the functional model, and the associated documents of the target function can be retrieved from the enhanced vector database. Then, the basic description and associated documents can be fused using LLM to obtain an enhanced description of the target function.

[0048] In some embodiments of this application, guiding prompts can be constructed for each prompt in a preset set of hazard prompts based on enhanced descriptions, and an LLM can be used to generate a list of hazard elements for the target function based on the guiding prompts.

[0049] In other words, a set of hazard cues for FHA can be pre-set, including various types of hazard cues that need to be analyzed. For each function, the enhanced description described above can be used to construct guiding prompts for each type of hazard cues. Then, LLM can be used to generate hazard elements based on these guiding prompts, and finally, a list of hazard elements for all hazard cues corresponding to this function can be combined.

[0050] In some embodiments of this application, the FHA results for the target function can be determined at least based on a list of hazard elements.

[0051] According to the technical solution provided in the embodiments of this application, a functional model of the analysis object is constructed using the MBSE model. An enhanced vector database is obtained by injecting retrieval augmentation knowledge. Then, the basic description of the target function is determined using the functional model, and related documents are retrieved from the enhanced vector database. The two are fused to obtain an enhanced description. A hazard factor list is generated based on the enhanced description, and then the FHA result of the target function is determined based on the hazard factor list. This method overcomes the shortcomings of traditional FHA, which relies on domain experts for manual completion. It comprehensively utilizes model-based systems engineering methods and large language model technology. The resulting intelligent FHA method can significantly optimize the overall process of functional hazard analysis, effectively improve the accuracy of FHA results and the efficiency of analysis work, and is conducive to guiding the efficient iteration of traditional FHA.

[0052] In some embodiments of this application, the functional model may include Functional levels Each functional level It includes a set of functional units that constitute this level. ; It is a positive integer greater than 1. greater than or equal to 0 and less than or equal to positive integers, For the first The first in the layer One function, It is a positive integer greater than 0.

[0053] A functional model can be a directed graph, where each node corresponds to a function and each edge corresponds to a relationship between two functions. When constructing a functional model, the modeling principle of "top-down, layer-by-layer refinement" can be followed to ensure that each level of the functional model has clear boundaries and responsibilities.

[0054] Taking a certain ground mobile platform as an example, this platform serves as the mobile vehicle for multi-purpose robots. By loading different functional modules, it can perform various tasks such as patrolling, transportation, and rescue, and is applied in fields such as industrial inspection, smart logistics, and disaster relief. Therefore, the platform is required to have strong mobility, high driving speed, and the ability to climb slopes, overcome obstacles, and wade through water; strong environmental adaptability, allowing it to move freely on terrains such as grasslands, sandy areas, mountains, roads, and indoor spaces; the ability to withstand high vibration and impact loads, a temperature range of -25 to +50℃, and harsh electromagnetic environments; and primarily remote control operation, capable of local path planning.

[0055] The physical structure of a ground mobile platform includes a control box, remote control box, power box, power supply, vehicle body, suspension, etc. Functional analysis and decomposition based on the MBSE method result in six system-level functional levels: communication, power supply, image acquisition, signal processing, obstacle crossing, and motion. The hierarchical breakdown of the "motion" function requirements can be further elaborated as follows: Figure 2 As shown. Motion requirements can be divided into four primary functions: load, structural characteristics, motion, and drive system. The load function can be further divided into two secondary functions: load capacity and traction. The structural characteristics function can be further divided into three secondary functions: shape, weight, and whether it is wheeled or tracked. The motion function can be further divided into five secondary functions: straight-line movement, turning, braking, endurance, and obstacle crossing. The straight-line movement function can be further divided into two tertiary functions: maximum speed and acceleration time. The turning function can be further divided into two tertiary functions: turning angular velocity and turning radius. The braking function can be further divided into a tertiary function: braking time. The endurance function can be further divided into two tertiary functions: endurance range and endurance time. The obstacle crossing function can be further divided into four tertiary functions: climbing angle, ditch width, wading depth, and vertical wall height. The drive system function can be further divided into three secondary functions: transmission efficiency, power, and modular structure.

[0056] Taking the "movement" functional hierarchy of a ground mobile platform as an example, the functional model is constructed as follows: Figure 3 As shown in the functional model, the push rod power supply and push rod control signals are processed by the power supply function level and the power conversion function level before being transmitted to the deformation obstacle crossing function level. The deformation obstacle crossing function level outputs monitoring signals to the signal monitoring function level. On the other hand, the motion control signals and motor drive signals are processed by the power supply function level and the power conversion function level before being transmitted to the power output function level. The power output function level outputs monitoring signals to the signal monitoring function level, so that the signal monitoring function level obtains the final monitoring signal based on the monitoring signals output by the deformation obstacle crossing function level and the monitoring signals output by the power output function level. At the same time, the power output function level also outputs support force to the support and protection function level and outputs kinetic energy to other functional levels of the ground mobile platform.

[0057] In some embodiments of this application, injecting retrieval enhancement knowledge into the functional model to obtain an enhanced vector database may include: parsing the enhanced knowledge of the analysis object to obtain a preset format string; using a dynamic semantic aggregation slicing strategy and a sliding window aggregation method to convert the preset format string into a semantically coherent set of knowledge slices; wherein, the dynamic semantic aggregation slicing is at least based on a pre-trained embedding model; each knowledge slice includes a semantically complete text segment, and each knowledge slice corresponds to an embedding vector; importing the set of knowledge slices and the embedding vectors corresponding to each knowledge slice into the vector database; extracting any The original query suggestions are used to construct the query vector for this function using a pre-trained embedding model. Then, similarity calculation and re-ranking are performed on the query vector in a vector database to obtain... The recall results; The recall results and original query suggestions are concatenated into a first suggestion keyword, which guides the LLM to filter and summarize the recall results based on the first suggestion keyword, and update the vector database based on the filtered and summarized results; traversing all And determine the traversal of all The updated vector database is the enhanced vector database.

[0058] In other words, when injecting retrieval enhancement knowledge into the functional model, the non-handwritten electronic design and operation and maintenance data of the analysis object (such as PDF, Word, WPS, scanned drawings, XML format, etc.) can first be converted into a structured text intermediate format using an optical character recognition model. Then, the document content is parsed page by page, and a Markdown format string containing semantic tags for headings, paragraphs, and lists is output.

[0059] Taking the first page of the design document for the aforementioned ground mobile platform as an example, the converted Markdown format is shown in Figure 4.

[0060] Next, a dynamic semantic aggregation slicing strategy is adopted. First, the Markdown text is initially segmented according to natural sentences to obtain a sentence sequence. Then, a pre-trained embedding model is used to map each sentence into a dense vector. Finally, sliding window aggregation is performed starting from the first sentence. The pre-trained embedding model can be, for example, Yinka.

[0061] The sliding window aggregation process can be as follows: initialize the current slice vector and slice content; for subsequent statements, calculate their cosine similarity with the current slice vector; if the cosine similarity is greater than or equal to a preset similarity threshold, then merge the subsequent statements into the current slice and update the slice vector to the vector transformed from the new content; otherwise, end the current slice, treat it as a complete knowledge fragment, and start a new slice with the subsequent statement as the starting point; this process continues until all statements have been processed, and finally a semantically coherent set of knowledge slices is obtained.

[0062] Taking the first paragraph of the first page of the ground mobile platform design document as an example, after being segmented into four sentences according to the period, semantic similarity is calculated using the Yinka embedding model. The first three sentences are aggregated into one knowledge fragment because of their high semantic coherence (the similarity between sentences 1-3 and the current slice is greater than the preset similarity threshold). The fourth sentence is separated into a separate fragment because its similarity with the current slice is less than the preset similarity threshold. Finally, the original text is divided into two semantically complete knowledge fragments. The first paragraph states: "The ground mobile general-purpose platform is a mobile driving carrier for multi-purpose robots. By loading different functional modules, it can perform various tasks such as patrol, transportation, and rescue, and is applied in fields such as industrial inspection, smart logistics, and disaster relief. Therefore, the platform is required to have strong mobility, high driving speed, and the ability to climb slopes, overcome obstacles, and wade through water; strong environmental adaptability, and can move freely in terrains such as grassland, sand, mountains, highways, and indoors." The second paragraph states: "It can withstand high vibration and impact loads, a temperature range of -25~+50℃, and harsh electromagnetic environments; in terms of operation, it is mainly remote controlled and can realize local path planning."

[0063] Next, the generated knowledge slice set and its embedding vector can be imported into a vector database, and then the components can be analyzed one by one. Specific analysis steps include: first, for any function... Analyze the data and extract its attributes to generate original query suggestions. , positive integer This indicates that the query is for the component's first... Each attribute, will Input the above embedding model to obtain the query vector. Perform similarity calculations in the vector database and retrieve the M most similar slices. M is a positive integer; the recall results are evaluated using a re-ranking model. The similarity is reordered, and the K slices with the highest similarity after reordering are selected as the final recall results. K is a positive integer; the final recall result will be... Compared with the original query suggestion The complete prompt words are concatenated to guide the large model to filter and summarize the recall results, and update the vector database based on the filtering and summary results; repeat the above steps until all components in the logical model have been analyzed. At this point, the updated vector database can be recorded as the enhanced vector database.

[0064] Taking the "morphing obstacle crossing" function of the aforementioned ground mobile platform as an example, the initial query prompt can be generated as "typical hazard clues, hazard impacts, and historical cases related to the 'morphing obstacle crossing' function." After being encoded by the Yinka embedding model, this initial query retrieves knowledge slices generated from design documents, operation and maintenance data, etc., in the vector database, recalling several semantically similar fragments. After reordering, the returned Top-K results mainly include general descriptions of the platform's environmental adaptability and vibration load, such as "able to withstand high vibration and impact loads" and "can move freely in terrains such as grassland, sand, and mountains."

[0065] After concatenating these fragments with the original query to form prompts, the large model summarizes the answer: "Based on the provided data, the obstacle-crossing function needs to meet the functional requirements of a maximum climbing angle of 30 degrees, a maximum climbing height of 200mm / step and an angle of 30 degrees, and a maximum vertical wall height of 200mm. Potential functional hazards may include loss of stability during deformation, deformation failure, and non-instructional deformation. The impact may include: inability to cross obstacles, task interruption or failure; getting stuck in front of obstacles and requiring external rescue; and failure to move to the designated position, leading to subsequent dangers. The specific historical case related to the obstacle-crossing function in the provided corpus is an industrial robot case: a welding robot arm on an automobile manufacturing line failed to move to the designated position due to encoder failure, resulting in a collision with the car body and damage to the expensive gripper and robot body."

[0066] The summarized answer is used as an enhanced recall result for the "morphing obstacle crossing" function to update the vector database, thereby updating the retrieval knowledge for the "morphing obstacle crossing" function.

[0067] In some embodiments of this application, FHA can be performed on each functional node in the functional model, for any function We can first determine the basic description of the target function based on the functional model, and then retrieve the associated documents of the target function from the enhanced vector database. Then, we can use the Large Language Model (LLM) to fuse the basic description and associated documents to obtain the enhanced description of the target function.

[0068] The basic description of the target function determined based on the functional model can be the extraction of pre-filled key information from the functional model, such as the functional description of the target function, as well as the basic functional attributes, hierarchical relationships and interface definitions corresponding to the target function.

[0069] Retrieving relevant documents for the target function from the enhanced vector database can be done by using the function identifier, function keywords, and port type of the target function as the core query conditions. Related technical document fragments, such as design specifications, interface specification documents, or redundant configuration materials, can be automatically retrieved from the enhanced vector database, and the search results can be used as relevant documents for the target function.

[0070] Then, by using the Large Language Model (LLM) to fuse the basic description and related documents, an enhanced description of the target function can be obtained. Specifically, special prompt words can be designed to guide the LLM to fuse the retrieved related documents with the original basic description in the MBSE model, generating an enhanced functional description that includes key elements such as functional details, connection relationships, and redundancy mechanisms. This enhanced functional description will serve as the basic information support for all subsequent FHA work.

[0071] Taking the ground mobile platform control box as an example again, after calling the vector database and summarizing by the large model, the generated enhanced description can be: "The control box is a key subsystem of the ground mobile general platform, mainly responsible for the platform's command parsing, power management, and system start-up and shutdown control functions. Its core functions include: connecting or disconnecting the power supply of the whole machine, receiving and processing drive and operation commands from the remote control terminal, outputting control signals to actuators such as the drive motor, and monitoring the power supply status of the lithium battery pack. In terms of connectivity, the control box is connected to the lithium battery pack through power supply lines and interacts with components such as drive motors, sensors, and communication modules through signal interfaces; its internal components include key components such as a main switch, fuses, electrical connectors, backup power supplies, and interface modules."

[0072] In some embodiments of this application, based on the enhanced description of the target function, guiding prompts can be constructed for each clue in the preset set of hazard clues and embedded in historical hazard cases of similar functions retrieved from the enhanced vector database. Then, the LLM generates a list of potential functional hazard elements that conforms to the actual engineering situation.

[0073] The preset set of danger clues includes at least the following danger clues: inability to execute, incorrect or erroneous execution, accidental execution, execution at the wrong time, inability to stop execution, receiving erroneous data, sending erroneous data, and information or data conflict.

[0074] The hazard list can be a list of three hazard elements, which includes at least the following hazard elements: hazard element, triggering mechanism, and object and threat.

[0075] In some embodiments of this application, the FHA result of the target function can be determined based on a defined list of hazard elements. In some embodiments, a preset task phase list can be obtained first, and LLM can be used to determine the estimated task phase for each functional hazard of the target function. Each functional hazard, its corresponding estimated task node, and its corresponding hazard clue can be bound into a structured record. Specifically, a functional hazard of the target function corresponds to an element in the hazard element list of the target function, and different functional hazards correspond to different elements.

[0076] Taking the "morphing obstacle crossing" sub-function of the ground mobile platform as an example, the determined structured record can be shown in Table 1:

[0077] Table 1. Structured record of the sub-function "morphing and overcoming obstacles"

[0078]

[0079] In some embodiments of this application, after determining the structured record of the target function, the affected functional objects can be determined for each functional hazard of the target function based on its structured record and functional model, thereby determining the propagation path of the functional hazard; wherein, the affected functional objects include at least the directly related functions corresponding to the local impact, the superior functions corresponding to the higher-level impact, and the top-level objects corresponding to the final impact.

[0080] Next, second cue words can be constructed using the functional hazard descriptions, propagation paths, and retrieved supplementary information for each functional hazard. These cue words will guide the LLM to obtain the impact analysis results for each functional hazard based on the second cue words. The retrieved supplementary information can be related documents for the target function, supplementary information retrieved again from the enhanced vector database using the functional hazard descriptions, or other supplementary information related to the target function retrieved from the enhanced vector database; there are no restrictions here.

[0081] In other words, for each functional hazard, the affected functional objects or system objects can first be identified based on the constructed functional model: including directly related functions (local impact), superior functions (higher-level impact), and top-level objects (ultimate impact), thereby obtaining the propagation path of this functional hazard; subsequently, the propagation path, functional hazard description, and retrieved supplementary information are organized into prompt words to guide LLM to generate natural language descriptions of the functional hazard's "impact on other analyzed object systems or functions" and "involved analyzed object systems"; similarly, the propagation path, functional hazard description, and retrieved supplementary information can also be organized into prompt words to guide LLM to generate natural language descriptions of the functional hazard's "impact on products, ground stations, and personnel".

[0082] Next, LLM can be used to match the impact analysis results to obtain the impact level corresponding to each functional hazard. Based on the impact analysis results and their corresponding impact levels, the FHA results for each functional hazard can be obtained. A hazard impact level evaluation table can be predefined. By semantically matching the impact analysis results with this predefined hazard impact level evaluation table, the impact level rating is automatically completed. Then, the impact level rating result, together with the impact analysis results described in natural language, constitutes a structured output for subsequent cause analysis and response formulation.

[0083] In some embodiments of this application, cause analysis and mitigation measures can also be implemented based on the impact analysis results and impact levels of the structured output.

[0084] In some implementations, a third warning word can be designed based on the impact analysis results and their corresponding impact levels; LLM is then used to analyze each functional hazard based on the third warning word, providing subsequent analysis methods and suggested avoidance measures. These subsequent analysis methods include recommended analysis methods or recommended analysis method types.

[0085] In other implementations, the recommended avoidance measures can be determined as follows: using LLM to retrieve engineering practice cases corresponding to each functional hazard based on third cue words to obtain an initial set of recommended avoidance measures; constructing a fourth cue word based on the functional model and the functional hazard, and using LLM to verify each avoidance measure in the initial set of recommended avoidance measures based on the fourth cue word to obtain a set of candidate avoidance measures; and using LLM to fuse each avoidance measure in the set of candidate avoidance measures to obtain recommended avoidance measures.

[0086] The subsequent analysis method, also known as the compliance verification method, refers to the specific methods or work to be carried out in the subsequent analysis, such as Fault Tree Analysis (FTA) and Fault Mode Effects and Criticality Analysis (FMECA). If it is not yet certain, it can be initially defined as qualitative analysis, quantitative analysis, or qualitative-quantitative analysis.

[0087] Mitigation measures primarily consider measures to reduce the impact of high-level functional hazards, including preventative measures, functional improvements, and post-hazard mitigation, serving as a reference for subsequent iterations. When developing mitigation measures, for each identified functional hazard, cue keywords are used to retrieve engineering practice cases. Validation measures are then used to further validate and guide the LLM (Leadership Management Model) in providing subsequent analysis methods, ultimately leading to feasible mitigation measures. All analysis methods and mitigation measures can be linked to their corresponding functional hazards in structured data format, ultimately forming a complete FHA (Functional Health Assessment) table.

[0088] Taking the "morphing obstacle crossing" sub-function of the ground mobile platform as an example again, the FHA results are shown in Table 2:

[0089] Table 2. FHA Results for the Sub-function "Modalization Obstacle Crossing"

[0090]

[0091]

[0092] Figure 5 This is a flowchart illustrating another functional hazard intelligent analysis method based on MBSE-LLM provided in an embodiment of this application. Figure 5 As shown, the method includes the following steps:

[0093] Step 1: Construct a functional model of the object being analyzed.

[0094] Step 2: Run the retrieval-enhanced knowledge injection method.

[0095] Step 3: Based on the modeling information obtained in Step 1 and the design data obtained in Step 2, perform functional hazard analysis based on LLM.

[0096] Figure 6 This is a diagram of the MBSE-guided LLM functional hazard analysis interaction mechanism provided in an embodiment of this application. For example... Figure 6 As shown, a logical model can be built based on MBSE, and the function names, top-level objects and parent objects of each function in the model can be input into the LLM. At the same time, search enhancement methods can be injected into the LLM to obtain component names and analysis targets from the LLM and feed back RAG answers to the LLM.

[0097] LLM can be used to enhance the descriptions of each feature to generate enhanced descriptions for each feature. In some examples, enhanced descriptions can be generated by LLM based on an initial description and retrieved enhancements.

[0098] LLM receives the analysis objectives, analysis requirements, and output requirements. It then automatically generates cue words based on the enhanced descriptions of each function. Based on these cue words, it automatically performs functional hazard analysis (FHA) and outputs the FHA results. These FHA results may include information such as hazard cues, hazards, hazard elements, triggering mechanisms, objects and threats, mission phases, impact on other systems or functions, systems involved, impact on products, ground stations, and personnel, hazard impact level, follow-up analysis methods, and mitigation considerations.

[0099] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0100] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0101] Figure 7 This is a schematic diagram of a functional hazard intelligent analysis device based on MBSE-LLM provided in an embodiment of this application. Figure 7 As shown, the device includes:

[0102] Module 701 is configured to build a functional model of the analysis object using the Model-Based Systems Engineering (MBSE) approach.

[0103] Enhancement module 702 is configured to inject retrieval enhancement knowledge into the functional model to obtain an enhanced vector database; the vector database includes enhanced recall results for each function in the functional model.

[0104] The determination module 703 is configured to determine the basic description of the target function based on the functional model and retrieve the associated documents of the target function from the enhanced vector database; the target function is any function in the functional model.

[0105] Fusion module 704 is configured to use the Large Language Model (LLM) to fuse basic descriptions and associated documents to obtain an enhanced description of the target function.

[0106] The generation module 705 is configured to construct guiding prompts for each clue in a preset set of hazard clues based on enhanced descriptions, and to generate a list of hazard elements for the target function based on the guiding prompts using LLM.

[0107] Analysis module 706 is configured to determine the functional hazard analysis (FHA) results of the target function based at least on the list of hazard elements.

[0108] According to the technical solution provided in the embodiments of this application, a functional model of the analysis object is constructed using the MBSE model. An enhanced vector database is obtained by injecting retrieval augmentation knowledge. Then, the basic description of the target function is determined using the functional model, and related documents are retrieved from the enhanced vector database. The two are fused to obtain an enhanced description. A hazard factor list is generated based on the enhanced description, and then the FHA result of the target function is determined based on the hazard factor list. This method overcomes the shortcomings of traditional FHA, which relies on domain experts for manual completion. It comprehensively utilizes model-based systems engineering methods and large language model technology. The resulting intelligent FHA method can significantly optimize the overall process of functional hazard analysis, effectively improve the accuracy of FHA results and the efficiency of analysis work, and is conducive to guiding the efficient iteration of traditional FHA.

[0109] In some implementations, the functional model includes Functional levels Each functional level It includes a set of functional units that constitute this level. ;in, It is a positive integer greater than 1. greater than or equal to 0 and less than or equal to positive integers, For the first The first in the layer One function, It is a positive integer greater than 0.

[0110] In some implementations, retrieval enhancement knowledge is injected into the functional model to obtain an enhanced vector database, including: parsing the enhanced knowledge of the analysis object to obtain a preset format string; using a dynamic semantic aggregation slicing strategy and a sliding window aggregation method to convert the preset format string into a semantically coherent set of knowledge slices; wherein, the dynamic semantic aggregation slicing is implemented based at least on a pre-trained embedding model; each knowledge slice includes a semantically complete text segment, and each knowledge slice corresponds to an embedding vector; the set of knowledge slices and the embedding vectors corresponding to each knowledge slice are imported into the vector database; and any... The original query suggestions are used to construct the query vector for this function using a pre-trained embedding model. Then, similarity calculation and re-ranking are performed on the query vector in a vector database to obtain... The recall results; The recall results and original query suggestions are concatenated into a first suggestion keyword, which guides the LLM to filter and summarize the recall results based on the first suggestion keyword, and update the vector database based on the filtered and summarized results; traversing all And determine the traversal of all The updated vector database is the enhanced vector database.

[0111] In some implementations, the basic description of the target function includes at least the function description, function attributes, hierarchical relationship, and interface definition of the target function; the associated documents are retrieved from the enhanced vector database by using the function identifier, function keywords, and port type of the target function as the core query conditions; the enhanced description of the target function includes at least the function details, connection relationships, and redundancy mechanisms of the target function.

[0112] In some implementations, the preset set of danger clues includes at least the following danger clues: inability to execute, incorrect or erroneous execution, accidental execution, execution at the wrong time, inability to stop execution, receiving erroneous data, sending erroneous data, and information or data conflict; the list of danger elements includes at least the following danger elements: danger element, triggering mechanism, object, and threat.

[0113] In some implementations, the FHA results for the target function are determined at least based on a list of hazard elements, including: obtaining a list of preset task stages; using LLM to determine the estimated task stage for each functional hazard of the target function; and binding each functional hazard, its corresponding estimated task node, and its corresponding hazard clue into a structured record; wherein, a functional hazard of the target function corresponds to an element in the hazard element list of the target function, and different functional hazards correspond to different elements; for each functional hazard, the affected functional objects are determined based on its structured record and functional model, thereby determining the propagation path of the functional hazard; the affected functional objects include at least the directly related functions corresponding to the local impact, the superior functions corresponding to the higher-level impact, and the top-level objects corresponding to the final impact; using the functional hazard description, propagation path, and retrieved supplementary information for each functional hazard to construct second prompt words, guiding LLM to obtain the impact analysis results for each functional hazard based on the second prompt words; using LLM to match the impact analysis results to obtain the impact level corresponding to each functional hazard, and obtaining the FHA results for each functional hazard based on the impact analysis results and their corresponding impact levels.

[0114] In some implementations, after determining the impact level, the process further includes: designing a third warning word based on the impact analysis results and their corresponding impact levels; using LLM to analyze each functional hazard based on the third warning word, and providing subsequent analysis methods and suggested avoidance measures; wherein, the subsequent analysis methods include recommended analysis methods or recommended analysis method types.

[0115] The recommended avoidance measures are determined as follows: An initial set of recommended avoidance measures is obtained by using LLM to retrieve engineering practice cases corresponding to each functional hazard based on the third cue word; a fourth cue word is constructed based on the functional model and the functional hazard, and LLM is used to verify each avoidance measure in the initial set of recommended avoidance measures based on the fourth cue word to obtain a set of candidate avoidance measures; finally, LLM is used to fuse the avoidance measures in the candidate avoidance measures set to obtain the recommended avoidance measures.

[0116] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0117] Figure 8 This is a schematic diagram of the electronic device provided in an embodiment of this application. For example... Figure 8As shown, the electronic device 8 of this embodiment includes a processor 801, a memory 802, and a computer program 803 stored in the memory 802 and executable on the processor 801. When the processor 801 executes the computer program 803, it implements the steps in the various method embodiments described above. Alternatively, when the processor 801 executes the computer program 803, it implements the functions of each module / unit in the various device embodiments described above.

[0118] Electronic device 8 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 8 may include, but is not limited to, processor 801 and memory 802. Those skilled in the art will understand that... Figure 8 This is merely an example of electronic device 8 and does not constitute a limitation on electronic device 8. It may include more or fewer components than shown, or different components.

[0119] The processor 801 can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0120] The memory 802 can be an internal storage unit of the electronic device 8, such as a hard disk or RAM of the electronic device 8. The memory 802 can also be an external storage device of the electronic device 8, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, etc., equipped on the electronic device 8. The memory 802 can also include both internal and external storage units of the electronic device 8. The memory 802 is used to store computer programs and other programs and data required by the electronic device.

[0121] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0122] If an integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program may include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. A computer-readable medium may include: any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0123] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A functional hazard intelligent analysis method based on MBSE-LLM, characterized in that, include: The functional model of the analysis object is constructed using the Model-Based Systems Engineering (MBSE) method. The functional model is injected with enhanced retrieval knowledge to obtain an enhanced vector database; the vector database includes the enhanced recall results for each function in the functional model. Based on the functional model, a basic description of the target function is determined, and the associated documents of the target function are retrieved from the enhanced vector database; the target function is any function in the functional model. By fusing the basic description and the associated documents using a large language model (LLM), an enhanced description of the target function is obtained. Based on the enhanced description, guiding prompts are constructed for each clue in the preset set of hazard clues, and the LLM is used to generate a list of hazard elements for the target function based on the guiding prompts; The functional hazard analysis (FHA) results for the target function should be determined at least based on the aforementioned list of hazard elements. The FHA result for the target function is determined based at least on the list of hazard factors, including: Obtain a list of preset task stages, use the LLM to determine the estimated task stage for the occurrence of each functional hazard of the target function, and bind each functional hazard, the estimated task node corresponding to the functional hazard, and the hazard clue corresponding to the functional hazard into a structured record; wherein, a functional hazard of the target function corresponds to an element in the hazard element list of the target function, and different functional hazards correspond to different elements; For each functional hazard, the affected functional objects are determined based on its structured record and the functional model, thereby determining the propagation path of the functional hazard; the affected functional objects include at least the directly related functions corresponding to the local impact, the superior functions corresponding to the higher-level impact, and the top-level objects corresponding to the final impact; A second cue word is constructed using the functional hazard descriptions, propagation paths, and retrieved supplementary information for each functional hazard, guiding the LLM to obtain the impact analysis results of each functional hazard based on the second cue word; The impact level of each functional hazard is obtained by matching the impact analysis results using the LLM, and the FHA result of each functional hazard is obtained based on the impact analysis results and their corresponding impact levels.

2. The method according to claim 1, characterized in that, The functional model includes Functional levels ; Each functional level It includes a set of functional units that constitute this level. ; in, It is a positive integer greater than 1. greater than or equal to 0 and less than or equal to positive integers, For the first The first in the layer One function, It is a positive integer greater than 0.

3. The method according to claim 2, characterized in that, By injecting retrieval enhancement knowledge into the functional model, an enhanced vector database is obtained, including: The enhanced knowledge of the analyzed object is parsed to obtain a preset format string; The preset format string is converted into a set of semantically coherent knowledge slices using a dynamic semantic aggregation slicing strategy and a sliding window aggregation method; wherein, the dynamic semantic aggregation slicing is implemented based at least on a pre-trained embedding model; each knowledge slice includes a semantically complete text segment, and each knowledge slice corresponds to an embedding vector; Import the knowledge slice set and the embedding vectors corresponding to each knowledge slice into a vector database; Extract any The original query suggestion is used to construct a query vector for this function based on the pre-trained embedding model. Then, similarity calculation and re-ranking are performed on the query vector in the vector database to obtain... The recall results; Will The recall results and the original query suggestions are concatenated into a first suggestion word, which guides the LLM to filter and summarize the recall results based on the first suggestion word, and updates the vector database based on the filtering and summary results; Traverse all And determine the traversal of all The updated vector database is the enhanced vector database.

4. The method according to claim 1, characterized in that, The basic description of the target function includes at least the function description, function attributes, hierarchical relationship and interface definition of the target function; The associated documents are retrieved from the enhanced vector database by using the function identifier, function keywords and port type of the target function as the core query conditions. The enhanced description of the target function includes at least the functional details, connectivity relationships, and redundancy mechanisms of the target function.

5. The method according to claim 1, characterized in that, The preset set of danger clues includes at least the following danger clues: inability to execute, incorrect or erroneous execution, accidental execution, execution at the wrong time, inability to stop execution, receiving erroneous data, sending erroneous data, and information or data conflict. The list of hazard elements includes at least the following hazard elements: hazard elements, triggering mechanisms, and objects and threats.

6. The method according to claim 1, characterized in that, After determining the impact level, the method further includes: Based on the impact analysis results and their corresponding impact levels, a third prompt word was designed; The LLM is used to analyze the functional hazards based on the third cue word, and subsequent analysis methods and suggested avoidance measures are given. The subsequent analysis methods include recommended analysis methods or recommended analysis method types; The recommended avoidance measures are determined in the following manner: Using the LLM to retrieve engineering practice cases corresponding to each functional hazard based on the third cue word, an initial set of suggested avoidance measures is obtained; Based on the functional model and the functional hazard, a fourth cue word is constructed. The LLM is then used to verify each avoidance measure in the initial set of suggested avoidance measures based on the fourth cue word, thereby obtaining a set of candidate avoidance measures. The LLM is used to fuse the various avoidance measures in the candidate avoidance measure set to obtain the suggested avoidance measure.

7. A functional hazard intelligent analysis device based on MBSE-LLM, characterized in that, include: The building module is configured to construct a functional model of the analysis object using the Model-Based Systems Engineering (MBSE) approach. An enhancement module is configured to inject enhanced retrieval knowledge into the functional model to obtain an enhanced vector database; the vector database includes enhanced recall results for each function in the functional model. The determination module is configured to determine the basic description of the target function based on the functional model, and retrieve the associated documents of the target function from the enhanced vector database. The target function is any function in the functional model; The fusion module is configured to use a large language model (LLM) to fuse the basic description and the associated document to obtain an enhanced description of the target function. The generation module is configured to construct guiding prompts for each clue in a preset set of hazard clues based on the enhanced description, and use the LLM to generate a list of hazard elements for the target function based on the guiding prompts; The analysis module is configured to determine the functional hazard analysis (FHA) results for the target function based at least on the list of hazard elements. The FHA result for the target function is determined based at least on the list of hazard factors, including: Obtain a list of preset task stages, use the LLM to determine the estimated task stage for the occurrence of each functional hazard of the target function, and bind each functional hazard, the estimated task node corresponding to the functional hazard, and the hazard clue corresponding to the functional hazard into a structured record; wherein, a functional hazard of the target function corresponds to an element in the hazard element list of the target function, and different functional hazards correspond to different elements; For each functional hazard, the affected functional objects are determined based on its structured record and the functional model, thereby determining the propagation path of the functional hazard; the affected functional objects include at least the directly related functions corresponding to the local impact, the superior functions corresponding to the higher-level impact, and the top-level objects corresponding to the final impact; A second cue word is constructed using the functional hazard descriptions, propagation paths, and retrieved supplementary information for each functional hazard, guiding the LLM to obtain the impact analysis results of each functional hazard based on the second cue word; The impact level of each functional hazard is obtained by matching the impact analysis results using the LLM, and the FHA result of each functional hazard is obtained based on the impact analysis results and their corresponding impact levels.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 6.

Images