Fault tree intelligent modeling method and device based on mbse-llm

Abstract

This application relates to the fields of artificial intelligence and reliability engineering technology, and provides a fault tree intelligent modeling method and apparatus based on MBSE-LLM. This method constructs a logical model of the analysis object based on the MBSE method, builds a knowledge query agent based on the enhanced knowledge of the analysis object, specifies a top event, and locates the top event to a node in the logical model. Starting from the node corresponding to the top event in the logical model, it iteratively unfolds the causal analysis from top to bottom using LLM with the logical model as the clue. With the help of LLM, intermediate events and bottom events are generated to construct a complete fault tree, achieving efficient and comprehensive fault tree analysis.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and reliability engineering technology, and in particular to a fault tree intelligent modeling method and apparatus based on MBSE-LLM. Background Technology

[0002] Fault tree analysis involves constructing a logical fault tree model to demonstrate how different basic events, through combinations of logic gates, lead to the occurrence of a top event or an undesirable event. Traditional fault tree construction largely relies on professionals who have a thorough understanding of the product and need to comprehensively consider various factors to conduct causal analysis. The more complex the product, the more difficult fault tree modeling becomes, and the more time-consuming manual analysis becomes, making fault tree analysis a significant challenge to implement.

[0003] Therefore, it is necessary to develop highly automated intelligent fault tree modeling methods to improve fault tree modeling efficiency and reduce labor costs. However, since the understanding of product objects varies from person to person, the comprehensiveness and coverage of the fault tree model are difficult to guarantee. Therefore, the constructed intelligent fault tree modeling method also needs to consider the comprehensiveness and coverage of the model. Summary of the Invention

[0004] In view of this, embodiments of this application provide a fault tree intelligent modeling method and apparatus based on MBSE-LLM to solve the problems of low efficiency and high cost in fault tree modeling in the prior art.

[0005] A first aspect of this application provides a fault tree intelligent modeling method based on MBSE-LLM, comprising:

[0006] A logical model of the analysis object is constructed using the Model-Based Systems Engineering (MBSE) approach, and a knowledge query agent is built based on the enhanced knowledge of the analysis object; the logical model includes at least the nodes of the hierarchical entity components corresponding to the analysis object;

[0007] Define the node attributes of each node in the fault tree; the node attributes must include at least a set of child events.

[0008] Obtain the top event and its description, determine the node to be analyzed corresponding to the top event, and determine the output set of the node to be analyzed, as well as the source node of each output in the output set;

[0009] Traverse the output set. In response to anomalies in determining the target output based on a Large Language Model (LLM), a top event occurs. Use LLM to determine the sub-event set based on the top event, top event description, node to be analyzed, target output, and source node of the target output. The target output is any output in the output set.

[0010] The union of the sub-event sets of each output in the output set is used as an intermediate event list, which is then added to the sub-event set attribute of the node to be analyzed to obtain the child nodes of the node to be analyzed.

[0011] For any child node of the node to be analyzed, use LLM to determine at least the list of environmental factor events, the list of human factor events, the list of spontaneous factor events, and the list of induced factor events based on the knowledge query agent. Add each environmental factor event, human factor event, spontaneous factor event, and induced factor event into a node as a child node of the corresponding node. Iterate through each child node until the modeling stopping condition is met to obtain the fault tree.

[0012] A second aspect of this application provides a fault tree intelligent modeling device based on MBSE-LLM, comprising:

[0013] The building module is configured to construct a logical model of the analysis object using the Model-Based Systems Engineering (MBSE) approach, and to construct a knowledge query agent based on the enhanced knowledge of the analysis object; the logical model includes at least nodes of the hierarchical entity components corresponding to the analysis object;

[0014] The definition module is configured to define the node attributes of each node in the fault tree; the node attributes include at least a set of sub-events.

[0015] The determination module is configured to obtain the top event and its description, determine the node to be analyzed corresponding to the top event, and determine the output set of the node to be analyzed, as well as the source node of each output in the output set;

[0016] The determination module is also configured to traverse the output set, and in response to an anomaly in determining the target output based on the Large Language Model (LLM) that causes a top event to occur, the LLM is used to determine the sub-event set based on the top event, the top event description, the node to be analyzed, the target output, and the source node of the target output; the target output is any output in the output set.

[0017] The modeling module is configured to take the union of the sub-event sets of each output in the output set as an intermediate event list and add it to the sub-event set attribute of the node to be analyzed, thereby obtaining the child nodes of the node to be analyzed;

[0018] The modeling module is also configured as any child node of the node to be analyzed. Using LLM, at least the list of environmental factor events, the list of human factor events, the list of spontaneous factor events, and the list of induced factor events are determined based on the knowledge query agent. Each environmental factor event, human factor event, spontaneous factor event, and induced factor event forms a node and is added as a child node of the corresponding node. Each child node is iteratively processed until the modeling stopping condition is met, and the fault tree is obtained.

[0019] 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.

[0020] The beneficial effects of this application embodiment compared with the prior art are as follows: This application embodiment constructs a logical model of the analysis object based on the MBSE method, and constructs a knowledge query agent based on the enhanced knowledge of the analysis object. Then, it specifies the top event and locates the top event to the logical model node. Starting from the node corresponding to the top event in the logical model, it iteratively unfolds the cause analysis from top to bottom using LLM with the logical model as the clue. With the help of LLM, intermediate events and bottom events are generated to construct a complete fault tree, thereby achieving efficient and comprehensive fault tree analysis. Attached Figure Description

[0021] 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.

[0022] Figure 1 This is a flowchart illustrating a fault tree intelligent modeling method based on MBSE-LLM provided in an embodiment of this application.

[0023] Figure 2 This is a schematic diagram of the logical model constructed using a ground mobile platform as the analysis object as an example.

[0024] Figure 3 This is a schematic diagram of the structure of a template for RAG knowledge query prompts provided in an embodiment of this application.

[0025] Figure 4 This is a schematic diagram of the signal flow association judgment prompt word template structure provided in the embodiments of this application.

[0026] Figure 5 This is a schematic diagram of the fault event retrieval prompt word template structure provided in the embodiments of this application.

[0027] Figure 6 This is a schematic diagram of the verification prompt word template structure provided in the embodiments of this application.

[0028] Figure 7 This is a schematic diagram of the first-level fault tree model constructed using a ground mobile platform as the analysis object.

[0029] Figure 8This is a schematic diagram of the prompt word template structure for the environmental factor analysis task provided in the embodiments of this application.

[0030] Figure 9 This is a schematic diagram of the prompt word template structure for the human factors analysis task provided in the embodiments of this application.

[0031] Figure 10 This is a schematic diagram of the prompt word template structure for the spontaneous factor analysis task provided in the embodiments of this application.

[0032] Figure 11 This is a schematic diagram of a complete fault tree model constructed using a ground mobile platform as an example.

[0033] Figure 12 This is a flowchart illustrating another method for an MBSE-LLM-based intelligent fault tree modeling device provided in this application embodiment.

[0034] Figure 13 This is a schematic diagram of a fault tree intelligent modeling device based on MBSE-LLM provided in an embodiment of this application.

[0035] Figure 14 This is a schematic diagram of the electronic device provided in the embodiments of this application. Detailed Implementation

[0036] 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.

[0037] The following will describe in detail, with reference to the accompanying drawings, a fault tree intelligent modeling method and apparatus based on MBSE-LLM according to embodiments of this application.

[0038] As mentioned above, the traditional fault tree construction process relies heavily on professionals who have a thorough understanding of the product and need to comprehensively consider various factors to conduct causal analysis. The more complex the product, the more difficult the fault tree modeling becomes, and the more time-consuming the manual analysis is, making fault tree analysis a significant challenge to implement.

[0039] In view of this, this application provides a fault tree intelligent modeling method based on MBSE-LLM. Combining the concept of Model-Based Systems Engineering (MBSE), product-related knowledge is transformed into a unified model and knowledge base. By leveraging the understanding and generation capabilities of Large Language Model (LLM), fault tree automatic modeling is performed, achieving efficient and comprehensive fault tree analysis.

[0040] Figure 1 This is a flowchart illustrating a fault tree intelligent modeling method based on MBSE-LLM provided in an embodiment of this application. Figure 1 As shown, the method includes the following steps:

[0041] In step S101, the logical model of the analysis object is constructed using the MBSE method, and a knowledge query agent is constructed based on the enhanced knowledge of the analysis object.

[0042] The logical model includes at least nodes of hierarchical entity components corresponding to the analysis object.

[0043] In step S102, the node attributes of each node in the fault tree are defined.

[0044] The node attributes must include at least a set of child events.

[0045] In step S103, the top event and its description are obtained, the node to be analyzed corresponding to the top event is determined, and the output set of the node to be analyzed and the source node of each output in the output set are determined.

[0046] In step S104, the output set is traversed. In response to the fact that an anomaly in the target output determined based on LLM will cause the top event to occur, the set of sub-events is determined using LLM based on the top event, the top event description, the node to be analyzed, the target output, and the source node of the target output.

[0047] The target output is any output in the output set.

[0048] In step S105, the union of the sub-event sets of each output in the output set is used as an intermediate event list and added to the sub-event set attribute of the node to be analyzed to obtain the child nodes of the node to be analyzed.

[0049] In step S106, for any child node of the node to be analyzed, the LLM is used to determine at least the list of environmental factor events, the list of human factor events, the list of spontaneous factor events, and the list of induced factor events based on the knowledge query agent. Each environmental factor event, human factor event, spontaneous factor event, and induced factor event forms a node and is added as a child node of the corresponding node. Each child node is iteratively processed until the modeling stopping condition is met, and the fault tree is obtained.

[0050] In some embodiments of this application, the method may be executed by a server or by a terminal device with certain processing capabilities.

[0051] In some embodiments of this application, a logical model of the analysis object can first be constructed using the MBSE method, and a knowledge query agent can be constructed based on the enhanced knowledge of the analysis object. Simultaneously, node attributes of each node in the fault tree can also be defined.

[0052] In some embodiments of this application, the top event and its description can be obtained, the node to be analyzed corresponding to the top event can be determined, and the output set of the node to be analyzed and the source node of each output in the output set can be determined.

[0053] In this context, the top event of the analyzed object is the starting point and core focus of the fault tree analysis. It represents the undesirable failure or accident outcome of the analyzed object and is usually located at the top of the fault tree. The top event is the endpoint of the logical deduction, and the structure of the fault tree can be constructed by decomposing its direct causes downwards.

[0054] In some embodiments of this application, the output set can be traversed. If it is determined based on LLM that an anomaly in the target output will cause a top event to occur, then LLM can be used to determine a set of sub-events based on the top event, the top event description, the node to be analyzed, the target output, and the source node of the target output. Then, the union of the sub-event sets of each output in the output set is used as an intermediate event list and added to the sub-event set attribute of the node to be analyzed to obtain the child nodes of the node to be analyzed.

[0055] In some embodiments of this application, for any child node of the node to be analyzed, an LLM can be used to determine at least the list of environmental factor events, the list of human factor events, the list of spontaneous factor events, and the list of induced factor events based on a knowledge query agent. Each environmental factor event, human factor event, spontaneous factor event, and induced factor event forms a node and is added as a child node of the corresponding node. Each child node is iteratively processed until the modeling stopping condition is met to obtain a fault tree.

[0056] Furthermore, for any given child node, the above method can be used to iteratively determine its child nodes until the bottom event is processed. Here, the bottom event of the analysis object is the basic unit in the logical model, referring to the direct cause that leads to other events, and is usually located at the bottom level of the fault tree.

[0057] According to the technical solution provided in the embodiments of this application, a logical model of the analysis object is constructed based on the MBSE method, and a knowledge query agent is constructed based on the enhanced knowledge of the analysis object. Then, a top event is specified and located to a node of the logical model. Starting from the node corresponding to the top event in the logical model, the cause analysis is iteratively carried out from top to bottom using LLM with the logical model as the clue. Intermediate events and bottom events are generated with the help of LLM to construct a complete fault tree, thereby achieving efficient and comprehensive fault tree analysis.

[0058] In some embodiments of this application, constructing the logical model of the analysis object using the MBSE method can be done by assuming the overall analysis object system is as follows: The MBSE method is used to divide the analyzed object into multiple logical levels based on its actual hardware architecture, denoted as... , of which each layer It includes a set of components that make up this layer. .in this way, It can represent the first The first in the layer Each component, also known as a physical unit, For the first The total number of components in a layer.

[0059] Within each defined hierarchy, identify and create all entity component nodes that constitute that hierarchy, and then perform component... Define its attribute tuple ;in, A unique identifier for the component. For component name, This is for component functionality. All node components can be stored in the structured data format JSON.

[0060] After defining the component nodes at each level, you can further establish connections between components. Connections can be formed by port definitions and connection lines: Let the port set be... Each port Must be with one and only one component Associated, with a unique port identifier Port type and port direction Let the set of connection relations be... Each connection A directed edge from one output port to one input port is defined as: Among them, the connection types include three categories: matter, information, and energy, namely... Text attributes This provides a supplementary explanation of the connection relationship, highlighting that the connection lines have directionality.

[0061] in, Indicates Boolean type, Indicates integer. Indicates floating-point type, Represents a character type; Indicates the input port type. Indicates the output port type; It is a positive integer; express A unique identifier, express input port, express The output port, express Connection type; To represent information, Indicates energy. It refers to matter.

[0062] The logical model constructed through the above steps is represented as a set of nodes. and edge set The directed graph formed The data is stored in the structured data format JSON to obtain the logical model of the analyzed object. Taking a certain type of ground mobile platform as an example, a unified logical model is established. The ground mobile platform has internal units such as a control box, remote control box, power box, tracks, and lithium battery pack. The signal flow transmission relationship of each unit is established, and the resulting visual logical model is as follows: Figure 2 As shown.

[0063] In some embodiments of this application, constructing a knowledge query agent based on enhanced knowledge may include: performing dynamic semantic aggregation slicing on enhanced knowledge to obtain knowledge slices, and vectorizing the knowledge slices to obtain a knowledge slice vector set; using a preset retrieval algorithm to perform knowledge retrieval on preset user queries in the knowledge slice vector set to obtain candidate knowledge slice vectors; constructing query prompts based on preset user queries and candidate knowledge slice vectors, and guiding the LLM to determine the query answers corresponding to each preset user query based on the query prompts, thereby forming a knowledge query agent.

[0064] In other words, 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 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.

[0065] The formatted Markdown string is then divided into semantically complete text segments. The length of each text segment does not exceed a preset context window threshold, and the segmentation avoids breaking across paragraphs or tables. A pre-trained text embedding model is then used to vectorize each text segment, and the vector segments are persistently stored in a vector database.

[0066] Next, a retrieval algorithm based on Maximum Marginal Relevance (MMR) is constructed for knowledge recall. Upon receiving a user query or modeling task request, the MMR algorithm first encodes the query statement into a vector using the pre-trained text embedding model and performs a nearest neighbor search in the aforementioned vector database to obtain an initial set of candidate knowledge vectors. Subsequently, the MMR algorithm iteratively selects knowledge items from the candidate set that are highly relevant to the query vector and have low redundancy among them for recall. Its scoring function is defined as... ;in, This indicates the knowledge items to be evaluated. Indicates and Different knowledge items, for The scoring function, For the selected subset of knowledge, This indicates the semantic similarity between a knowledge item and a query. Indicates the redundancy between knowledge items. For balancing parameters.

[0067] Finally, based on the recalled knowledge items, a Retrieval-augmented Generation (RAG) function can be constructed: this function generates relevant knowledge fragments recalled by the MMR algorithm. And the questions that need to be searched By embedding RAG knowledge query prompts and inputting them into a pre-trained LLM, a knowledge query agent is formed by obtaining query responses. One template for RAG knowledge query prompts is as follows: Figure 3 As shown.

[0068] In some embodiments of this application, the attributes of each node in the fault tree also include the event type and the logical model node where the corresponding event occurred. Further, the attributes of each node in the fault tree can be represented as follows: ,in For each node in the fault tree, num is the event number; event is the event content; object is the logical model node where the corresponding event occurs; type is the event type, which is divided into top event, intermediate event and bottom event; logic is the logical relationship between child events, which is divided into OR gate, AND gate, voting gate and condition gate, initially set to OR gate; parent is the parent event; childrens is the set of child events.

[0069] The event type of the node to be analyzed corresponding to the top event is the top event type. The output set of the node to be analyzed is the output set of the logical model node in the logical model where the corresponding event occurs. The source node of each output in this output set corresponds to the internal unit of this output in the logical model.

[0070] In other words, if the top event is denoted as The top event description is denoted as The corresponding node to be analyzed in the logical model is denoted as Then it can be based on , and Construct a top-event localization thought chain to generate the first-level fault tree model.

[0071] The fault tree model can be initialized first. And initialize its various properties: ,in" " indicates an OR gate, This represents the empty set.

[0072] Then, query and analyze the logical model of the object, based on the nodes in the logical model. The input-output connection relationship yields the output set of the node. and the source nodes of each output in the output set. ,in, and All are positive integers, and .

[0073] Next, we can traverse this node. output set For each output and its source node An analysis was conducted. Among them, , , get each The set of sub-events.

[0074] In some embodiments of this application, the set of sub-events of the target output is determined in the following manner: Signal flow association judgment prompts are constructed based on the target output and its source nodes; the signal flow association judgment prompts include at least the background of the association judgment task, supplementary materials, and precautions for signal flow association judgment; the supplementary materials are determined at least based on the knowledge query agent, including the functional description of the node corresponding to the target output in the logical model and the functional description of the node corresponding to the source node of the target output in the logical model; wherein, the functional description includes the node functional description defined in the logical model and the functional description obtained through RAG knowledge query; the precautions for signal flow association judgment include at least the analysis object, signal flow association definition, analysis scope, and output requirements.

[0075] Based on the signal flow association judgment prompts, a signal flow association judgment analysis task is generated. The LLM (Local Management Model) is used for intelligent judgment analysis to determine whether the target output anomaly will lead to the occurrence of the top event. If so, fault event retrieval prompts are constructed based on the target output and its source node. These prompts include at least the retrieval task background, supplementary materials, and fault event retrieval precautions. The fault event retrieval precautions include at least the analysis object, fault event definition, fault event quantity threshold, and output requirements.

[0076] A fault event retrieval and analysis task is generated based on the fault event retrieval prompts. The LLM (Local Management Model) is used for intelligent judgment and analysis to obtain a set of intermediate events for the target output. Verification prompts are constructed for each event in the intermediate event set; these prompts include at least the verification task background, the supplementary materials, and verification considerations; the verification considerations include at least the analysis object, verification definition, and output requirements.

[0077] Based on the verification prompts, a verification task is generated, and the LLM is used for intelligent judgment and analysis. All events that pass the verification are combined to obtain the sub-event set of the target output.

[0078] In other words, signal flow association judgment prompts can be constructed first based on the target output and its source nodes. These prompts can include the background of the association judgment task, supplementary materials, and precautions for signal flow association judgment. The supplementary materials are embedded in the logical model regarding nodes. and The functional description, and the results obtained from the knowledge query agent. and Functional description. This supplementary material can be referred to as... ,in, To embed nodes in the logical model and Functional description, For what is retrieved in the knowledge query agent and Functional description. When querying within the knowledge query agent, the input can be uniformly written as "component". What is its function? The name of the cell to be queried.

[0079] Furthermore, the considerations for signal flow correlation judgment should at least include the analysis object, signal flow correlation definition, analysis scope, and output requirements, which can be uniformly written as:

[0080] Notes:

[0081] 1. Only consider this one output of the component, and do not consider any other objects;

[0082] 2. Examine whether the output anomaly can directly lead to the occurrence of the top event. If there are additional conditions or other components or processes involved, it is considered that the input will not lead to the top event.

[0083] 3. Focus on the functional descriptions in the supplementary materials to avoid over-interpreting the information;

[0084] 4. Output only the content in the given format; do not add any extra content.

[0085] 5. Finally, output the result of the judgment according to the specified format: output 1 if the judgment will lead to the result, otherwise output 0.

[0086] Furthermore, a task description for signal flow correlation judgment and analysis can be generated and embedded into the aforementioned correlation judgment task background. This task description for signal flow correlation judgment and analysis can be uniformly written as follows:

[0087] “ The top event to be analyzed is: { } The object that will experience this top event is: The top event is described in detail as { We attribute the occurrence of the top event to { The output to the outside world is considered to be caused by the abnormality of the output; now let's examine the output from { Internal components Output: .

[0088] in," " is the separator. , , , as well as These are embedded into the analysis task text. LLM will provide a judgment result by outputting 1 (representing that the output is relevant) or 0 (representing that the output is irrelevant).

[0089] For the target output Out and its source node If the LLM output result is 1, then we can continue to construct fault event retrieval suggestions based on the target output and its source node, and further analyze them. Possible failure events.

[0090] Fault event search prompts should at least include the search task background, the supplementary materials, and precautions for fault event search. The supplementary materials are the same as those obtained above. Same. The notes for fault event retrieval can be uniformly written as:

[0091] Notes:

[0092] 1. The identified failure mode must be able to cause the target event when it occurs alone, without involving other conditions, components, or systems;

[0093] 2. Consider the probability of failure modes occurring. Failure modes with low probability of occurrence are less important, and vice versa. Make a comprehensive judgment to identify the most important failure modes, and the number should not exceed {N}.

[0094] 3. The following perspectives can help your analysis:

[0095] a. Function not executed as expected; b. Function executed at an out-of-time; c. Function executed incompletely or incorrectly, outputting error messages; d. Performance not up to standard, outputting unstable results.

[0096] 4. When outputting fault modes in segments, be sure to output only the fault mode itself, and do not include phrases or words such as "cause" or "make".

[0097] 5. A fault mode can only have one state. The word "or" cannot appear. If there are multiple possibilities, they should be described separately. Two situations cannot be described in one fault mode.

[0098] 6. Finally, only answer the content specified in the output format. Do not add any other content or explain your answer.

[0099] Output format (listing the fault events you identified in bullet points):

[0100] 1. Event One;

[0101] 2. Event Two;

[0102] 3. Event Three;

[0103] ..."

[0104] Where N is the maximum expected number of fault events to be analyzed.

[0105] Furthermore, a description of the fault event retrieval and analysis task can be generated and embedded into the aforementioned fault event retrieval and analysis task background. This fault event retrieval and analysis task description can be uniformly written as follows:

[0106] “ There is one Components Its output to the outside world was abnormal and caused the target event. Top event Occurred, and the output is: The top event is described in detail as follows: Please identify the abnormal output condition that caused the top event. A failure event occurred.

[0107] in, , , , as well as These are embedded into the analysis task text. Finally, a list of analyzed intermediate events can be extracted from the LLM's response. .

[0108] You can also view the intermediate event list. Each intermediate event in the process is reflected upon and verified. In some implementations, verification prompts can be constructed for each intermediate event, which include at least the background of the verification task, the supplementary materials, and verification notes.

[0109] Among them, the background of the verification task can be embedded The text content obtained can be uniformly written as:

[0110] "Analysis has shown that a fault event {EA} occurred in component A {UA}, which caused a fault event {EB} in component B {UB}. Please determine whether the fault event {EA} was analyzed correctly, i.e., whether it caused the abnormality of component B."

[0111] This supplementary material is related to the information obtained above. The same. For example, the verification notes can be uniformly written as:

[0112] Notes:

[0113] 1. Consider only this one analysis point, and do not consider any other factors;

[0114] 2. Whether this fault event directly leads to the occurrence of an anomaly. If there are additional conditions or other components, systems, or processes involved, then the analysis is incorrect.

[0115] 3. Only answer this judgment result; do not output any other content.

[0116] 4. Output 1 if the analysis is correct, otherwise output 0.

[0117] Then, input the verification prompt into the LLM and execute the verification task. When the LLM output is 1, indicating that the intermediate events were correctly analyzed, integrate them into a set of sub-events. .

[0118] You can add a collection of child events. In the fault tree model, the children of the top event TE are obtained. ,in, num is an automatic numbering system. Logic is initialized as an OR gate. , .

[0119] At the same time, it can also be Add it to the list of intermediate events to be analyzed.

[0120] In some embodiments of this application, all outputs can be traversed. Complete the localization analysis of the top event to obtain the first-level fault tree model. And a list of intermediate events to be analyzed. The newly added intermediate events can be added to... In the children property; if a new event list is added. If the result is empty, the modeling process ends directly; if only one event node is added, the node is removed after the next round of analysis is completed.

[0121] Let's continue using "Ground Mobile Platform" as the analysis object as an example. Specify the root node "Ground Mobile Platform" as the node to be analyzed. The top event is "Unable to advance," and its description is "An advance control signal was input, but the ground mobile platform did not advance as expected." The signal flow association judgment prompt word template structure is as follows: Figure 4 As shown, the structure of the fault event search suggestion term template is as follows: Figure 5 As shown, the structure of the verification prompt word template is as follows: Figure 6 As shown.

[0122] Using the pre-trained open-source large language model Qwen2.5-72 as the LLM, the number of fault event analyses was set to [value missing]. The top event is located to a node in the logical model, and the first-level model of the resulting fault tree is as follows: Figure 7 As shown. For the top event "Unable to move forward", with "Unable to move forward" as the root node G1 of the fault tree, the first-level model nodes are obtained: "G2 Remote control box power management failure", "G3 Remote control box power supply interruption", "G4 Power box cannot provide sufficient power", "G5 Track jamming", and "G6 Drive wheel failure". Each logic gate in the first-level model node is defaulted to an OR gate or a bottom event. If the source node outputting from the top level in the logic model is a logic gate type, special processing is performed for the logic gate, adding AND gates, voting gates, and condition gates as needed. Specific processing logic is explained later. If, after the first-level analysis, the top event defaults to an OR gate with only one other type of logic gate, the default OR gate is deleted, and the top event is directly connected to that type of logic gate and expanded downwards.

[0123] The above steps have enabled the analysis and location of the top event in the fault tree. Next, the intermediate events of the fault tree can be iteratively expanded.

[0124] For each first-layer model node, LLM can be used to determine at least the list of environmental factor events, the list of human factor events, the list of spontaneous factor events, and the list of induced factor events based on the knowledge query agent. Each environmental factor event, human factor event, spontaneous factor event, and induced factor event forms a node and is added as a child node of the corresponding node. Each child node is iteratively processed until the modeling stopping condition is met, and the fault tree is obtained.

[0125] In some embodiments of this application, determining the environmental factor event list, the human factor event list, and the spontaneous factor event list may include: constructing analysis prompts for environmental factor events, human factor events, and spontaneous factor events respectively; each analysis prompt includes the analysis task background, supplementary event materials, and analysis precautions; the supplementary event materials are determined at least based on the knowledge query agent, including the functional description and factor analysis information of the corresponding node in the logical model of this sub-node; the environmental factor analysis prompts include environmental load factor analysis information, the human factor analysis prompts include human factor analysis information, and the spontaneous factor analysis prompts include spontaneous factor analysis information; the analysis precautions include at least the analysis object, event definition, and output requirements; generating analysis tasks based on the analysis prompts, using LLM for intelligent judgment analysis, and determining the environmental factor event list, the human factor event list, and the spontaneous factor event list; adding each environmental factor event, human factor event, and spontaneous factor event forming node as a child node of the corresponding node is the step of adding each environmental factor event, human factor event, and spontaneous factor event forming node as a leaf node to this sub-node.

[0126] In other words, for Each unanalyzed node in It can obtain its various attribute information, including event content. and the logical model nodes where the corresponding events occur. .against Sub-event analysis can be conducted, analyzing from four perspectives: environmental factors, human factors, spontaneous factors, and triggering factors.

[0127] When conducting environmental factor analysis, LLM analysis can be used to generate the environmental factor analysis task background, supplementary event materials, and environmental factor analysis precautions. The environmental factor analysis task background section can be embedded... and The text information, specifically, is as follows:

[0128] "Component {U} experienced a failure event {E};"

[0129] Please analyze which environmental factors led to the failure event and list them in detail.

[0130] Supplementary materials for the incident include Functional description and environmental load factor analysis information . It can be obtained by querying the knowledge query agent. For example, the query question can be written as: "component". What is the possible environmental load profile? Please describe it in bullet points, and may include numerical ranges.

[0131] For example, the precautions for environmental factor analysis can be uniformly written as:

[0132] Notes:

[0133] 1. Identify the environmental conditions and abnormal events that are most likely to cause the failure. The resulting process should be direct and clear, avoiding any far-fetched or forced connections.

[0134] 2. Consider the probability of environmental anomalies occurring. Environmental anomalies with low probability of occurrence are less important, and vice versa. Based on comprehensive judgment, select only the environmental anomalies that you think are the most important, and the number should not exceed {N}.

[0135] 3. If it is believed that no environmental anomalies could cause the intermediate event to occur, this is acceptable, and no output should be provided.

[0136] 4. The content of abnormal environmental events should only consider causes from environmental conditions, and should not involve other components or internal components, human factors, etc.

[0137] 5. Abnormal events in the output environment. Do not output unnecessary information such as "leads to...", and absolutely do not use phrases or words such as "leads to", "cause", or "because of...". ;

[0138] 6. An environmental anomaly event should have only one possible state and should not contain the word "or". If there are multiple possibilities, they should be described separately. Two situations should not be described in one environmental anomaly event.

[0139] 7. Please only answer the content specified in the output format; do not add any other content.

[0140] Output format (list the environmental anomalies you identified in bullet points):

[0141] 1. Event One;

[0142] 2. Event Two;

[0143] 3. Event Three;

[0144] ..."

[0145] Then, the list of environmental factor events analyzed from the LLM is extracted, and each event is formed into a node and added to the fault tree model; for environmental factor events... Add a node: num is automatically numbered sequentially. , All environmental factor events are bottom events and do not need to be added to the list of events to be analyzed. .

[0146] When conducting human factor analysis, LLM analysis can be used to generate the human factor analysis task background, event supplementary materials, and human factor analysis precautions. The construction methods for the human factor analysis task background, event supplementary materials, and human factor analysis precautions are similar to those for environmental factor analysis, and will not be elaborated further here.

[0147] The list of human factor events analyzed from the LLM can be extracted, and each event can be formed into a node and added to the fault tree model. This process is basically the same as adding environmental factor events, so it will not be described in detail here.

[0148] When conducting spontaneous factor analysis, LLM analysis can be used to generate the task background, supplementary event materials, and precautions for spontaneous factor analysis. The construction methods for the task background, supplementary event materials, and precautions for spontaneous factor analysis are similar to those for environmental factor analysis, and will not be elaborated further here.

[0149] The list of spontaneous factor events analyzed from the LLM can be extracted, and each event can be formed into a node and added to the fault tree model. This process is basically the same as adding environmental factor events, so it will not be described in detail here.

[0150] The prompt template structure for the environmental factor analysis task is as follows: Figure 8 As shown, the prompt word template structure for the human factors analysis task is as follows: Figure 9 As shown, the prompt word template structure for the spontaneous factor analysis task is as follows: Figure 10 As shown.

[0151] When conducting inducing factor analysis, you can first query the logical model to determine the input set of this sub-node, as well as the source node of each input in the input set; the input set of this sub-node is the input set of the logical model node that caused the corresponding event in the logical model, and the source node of each input in the input set corresponds to the internal unit of this input in the logical model.

[0152] Then, the input set is traversed. In response to the determination that the anomaly of the target input will lead to the occurrence of the top event based on LLM, the source node of the target input may be used to determine the possible failure events, and a set of candidate triggering factor events is obtained. The target input is any input in the input set.

[0153] Finally, LLM is used to verify the set of candidate triggering factor events, and the list of triggering factor events is determined based on the verified triggering factor events.

[0154] In other words, the nodes can be determined by first querying the logical model. input set and the source nodes of each input. Iterate through the input set. For each of them and its source The analysis determines whether the output anomaly will lead to the top event, obtaining the judgment result through querying LLM. This process is similar to the top event localization analysis described above, which determines each output. Whether an anomaly leads to the occurrence of the top event is largely the same process, and will not be elaborated here. In this case, the analysis task when asking questions about LLM can be written as follows:

[0155] “ The event to be analyzed is: The object that will experience this event is: We attribute the occurrence of this event to... The input signal is considered to be the cause of the event to be analyzed, and it is assumed that the event is caused by an anomaly in the input signal stream. Now, let's examine the signal from... Input: .

[0156] in, , , as well as They are embedded into the analysis task text respectively.

[0157] If LLM determines If the event is related to the top event, then its corresponding node is further analyzed using LLM. Possible failure events. This process is determined during the top event localization analysis mentioned above. The process for the set of sub-events is basically the same, and will not be repeated here. At this point, the analysis task when asking about LLM can be written as follows:

[0158] The target event is: The component that will trigger this target event is: From component The input signal was abnormal, which led to the target event. This occurs when the input signal stream is: .

[0159] in, , , as well as These are embedded into the analysis task text; this task allows for the extraction of a list of intermediate events from the LLM's responses. .

[0160] Then the analyzed intermediate event set Each event in the process was reviewed and verified, and the analyzed event set was selected. Add event collection In the fault tree model, we obtain For any given node, we have:

[0161] ;

[0162] Where num is automatically numbered sequentially. Logic is initialized as an OR gate. , Newly added nodes are also recorded in the list of intermediate events to be analyzed. Meanwhile, all newly added intermediate events are added to In the children property.

[0163] The above analysis does not consider nodes. The handling method in the case of logic gate attributes. However, in actual operation, each node... The logic gates have different properties, including AND gates, OR gates, voting gates, and condition gates. Voting gates and condition gates can both be implemented using combinations of AND and OR gates; therefore, they will be discussed below using nodes as examples. Taking AND gates and OR gates as examples, let's explain in detail the logic gates, considering nodes. Fault tree construction method under the condition of logic gate attributes.

[0164] In some embodiments of this application, if the source node of the target input is an AND gate, determining the possible fault events that the source node of the target input may cause using LLM may include: determining the input set of the source node of the target input in the logic model; using all nodes in the input set as logic gate analysis objects, determining the possible fault events that the logic gate analysis objects may cause using LLM, and obtaining a set of candidate inducing factor events; the step of adding the inducing factor event forming node as a child node of the corresponding node is to add the inducing factor event forming node as a child node of this child node.

[0165] In other words, for an AND gate in a logic model, it can be considered that each input signal of the AND gate affects the node. Therefore, the analysis process is necessary, and the nodes not considered above are also included. The steps are the same when dealing with the logic gate attributes.

[0166] For an OR gate in a logic model, we can consider each input signal of the OR gate as... They are redundant. Therefore, time nodes can be set. And add that time point to the fault tree node:

[0167] Then Add to the input list to be analyzed, and perform the above steps (not considered). Steps for handling logic gate attributes. The parent event attribute of the newly analyzed event node is changed to... and add accordingly The `childs` property.

[0168] Iteratively execute the above steps until the list of intermediate events to be analyzed is reached. Empty; finally, a complete fault tree model is constructed. .

[0169] Taking the ground mobile platform as the analysis object again as an example, the complete fault tree model constructed is as follows: Figure 11 As shown.

[0170] Figure 12 This is a flowchart illustrating another method for an MBSE-LLM-based intelligent fault tree modeling device provided in an embodiment of this application. Figure 12 As shown, the method includes the following steps:

[0171] Step 1: Establish a unified logical model. Based on the product information of the object being analyzed, a product tree can be constructed, signal flows established, and node information added sequentially to obtain the logical model of the object being analyzed. Data in the product logical model can be stored in JSON format and injected into the large language model.

[0172] Step Two: Building the Product Knowledge Vector Base. Based on product information, embedding model vectorization, MMR retrieval, and RAG knowledge query can be performed sequentially to obtain the product knowledge vector base. A product knowledge query agent can then be built based on this product knowledge vector base and injected into the large language model.

[0173] Steps 3 & 4: Users can input the top event, its description, and a specified node into the large language model, which then performs top event localization analysis and intermediate event iterative expansion. During top event localization analysis, the large language model can be used to obtain an intermediate event list through output signal correlation judgment, fault event generation, and reflection verification. In the intermediate event iterative expansion, the large language model can be used to perform environmental factor analysis, spontaneous factor analysis, and human factor analysis to generate bottom events; simultaneously, it can also be used to perform precipitating factor analysis to generate intermediate events and update the intermediate event list. For the generated intermediate events, iterative expansion analysis can be performed according to the logical model signal flow until the intermediate event list is traversed.

[0174] One approach to conducting various analyses is to ask thought chain questions to the large language model to obtain the results generated by the large language model.

[0175] The technical solution provided in this application can solve the problems of large workload and low efficiency when constructing fault trees mainly by manual means, and provide technical support for automatic fault tree modeling of complex systems.

[0176] 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.

[0177] 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.

[0178] Figure 13 This is a schematic diagram of a fault tree intelligent modeling device based on MBSE-LLM provided in an embodiment of this application. Figure 13 As shown, the device includes:

[0179] Module 1301 is configured to construct a logical model of the analysis object using the Model-Based Systems Engineering (MBSE) method, and to construct a knowledge query agent based on the enhanced knowledge of the analysis object; the logical model includes at least nodes of the hierarchical entity components corresponding to the analysis object.

[0180] Module 1302 is defined to define the node attributes of each node in the fault tree; the node attributes include at least a set of sub-events.

[0181] The determination module 1303 is configured to obtain the top event and the top event description, determine the node to be analyzed corresponding to the top event, and determine the output set of the node to be analyzed, as well as the source node of each output in the output set.

[0182] The determination module 1303 is also configured to traverse the output set, and in response to an anomaly in the determination of the target output based on the large language model LLM, which would cause a top event to occur, the LLM is used to determine the sub-event set based on the top event, the top event description, the node to be analyzed, the target output, and the source node of the target output; the target output is any output in the output set.

[0183] Modeling module 1304 is configured to take the union of the sub-event sets of each output in the output set as an intermediate event list and add it to the sub-event set attribute of the node to be analyzed, thereby obtaining the child nodes of the node to be analyzed.

[0184] The modeling module 1304 is also configured as any child node of the node to be analyzed. Using LLM, it determines at least the list of environmental factor events, the list of human factor events, the list of spontaneous factor events, and the list of induced factor events based on the knowledge query agent. Each environmental factor event, human factor event, spontaneous factor event, and induced factor event forms a node and is added as a child node of the corresponding node. The process iteratively processes each child node until the modeling stopping condition is met, and a fault tree is obtained.

[0185] According to the technical solution provided in the embodiments of this application, a logical model of the analysis object is constructed based on the MBSE method, and a knowledge query agent is constructed based on the enhanced knowledge of the analysis object. Then, a top event is specified and located to a node of the logical model. Starting from the node corresponding to the top event in the logical model, the cause analysis is iteratively carried out from top to bottom using LLM with the logical model as the clue. Intermediate events and bottom events are generated with the help of LLM to construct a complete fault tree, thereby achieving efficient and comprehensive fault tree analysis.

[0186] In some implementations, constructing a knowledge query agent based on enhanced knowledge includes: dynamically semantically aggregating and slicing the enhanced knowledge to obtain knowledge slices, and vectorizing the knowledge slices to obtain a knowledge slice vector set; using a preset retrieval algorithm to perform knowledge retrieval on preset user queries in the knowledge slice vector set to obtain candidate knowledge slice vectors; constructing query hints based on preset user queries and candidate knowledge slice vectors, and guiding the LLM to determine the query answers corresponding to each preset user query based on the query hints, thus forming a knowledge query agent.

[0187] In some implementations, the attributes of each node in the fault tree also include the event type and the logical model node where the corresponding event occurs; the event type of the node to be analyzed is the top event type; the output set of the node to be analyzed is the output set of the logical model node where the corresponding event occurs in the logical model; the source node of each output in the output set corresponds to the internal unit of this output in the logical model.

[0188] In some implementations, the set of sub-events of the target output is determined as follows: Signal flow association judgment prompts are constructed based on the target output and its source nodes; these prompts include at least the background of the association judgment task, supplementary materials, and precautions for signal flow association judgment; the supplementary materials are determined at least based on a knowledge query agent, including the functional descriptions of the corresponding nodes of the target output in the logical model and the functional descriptions of the corresponding nodes of the source nodes of the target output in the logical model; the precautions for signal flow association judgment include at least the analysis object, the definition of signal flow association, the analysis scope, and the output requirements; a signal flow association judgment analysis task is generated based on the prompts, and intelligent judgment analysis is performed using LLM to determine whether an abnormal target output will lead to the occurrence of the top event; if so, then based on the target output and its source nodes... The process involves: constructing fault event retrieval prompts; these prompts should include at least the retrieval task background, supplementary materials, and precautions for fault event retrieval; the precautions should include at least the analysis object, fault event definition, fault event quantity threshold, and output requirements; generating a fault event retrieval analysis task based on the prompts, using LLM for intelligent judgment and analysis to obtain the intermediate event set for the target output; constructing verification prompts for each event in the intermediate event set; these prompts should include at least the verification task background, supplementary materials, and verification precautions; the verification precautions should include at least the analysis object, verification definition, and output requirements; generating a verification task based on the verification prompts, using LLM for intelligent judgment and analysis, and combining all verified events to obtain the sub-event set for the target output.

[0189] In some implementations, determining the lists of environmental factor events, anthropogenic factor events, and spontaneous factor events includes: constructing analysis prompts for environmental factor events, anthropogenic factor events, and spontaneous factor events respectively; each analysis prompt includes the analysis task background, supplementary event materials, and analysis precautions; the supplementary event materials are determined at least based on the knowledge query agent, including the functional description and factor analysis information of the corresponding node in the logical model for this sub-node; environmental factor analysis prompts include environmental load factor analysis information, anthropogenic factor analysis prompts include anthropogenic factor analysis information, and spontaneous factor analysis prompts include spontaneous factor analysis information; the analysis precautions at least include the analysis object, event definition, and output requirements; generating analysis tasks based on the analysis prompts, using LLM for intelligent judgment analysis, and determining the lists of environmental factor events, anthropogenic factor events, and spontaneous factor events; adding each environmental factor event, anthropogenic event, and spontaneous factor event forming node as a child node of the corresponding node involves adding each environmental factor event, anthropogenic event, and spontaneous factor event forming node as a leaf node to this sub-node.

[0190] In some implementations, determining the triggering event list includes: querying the logic model to determine the input set of the current child node and the source node of each input in the input set; the input set of the current child node is the input set of the logic model node where the corresponding event occurs in the logic model, and the source node of each input in the input set corresponds to the internal unit of the current input in the logic model; traversing the input set, in response to determining that an anomaly of the target input based on LLM will cause the top event to occur, using LLM to determine the possible fault events that the source node of the target input may cause, and obtaining a candidate triggering event set; wherein, the target input is any input in the input set; using LLM to verify the candidate triggering event set, and determining the triggering event list based on the verified triggering events.

[0191] In some implementations, the source node of the target input is a logic gate in the logic model; the logic gate includes an AND gate or an OR gate; in response to determining that the source node of the target input is an AND gate, the LLM is used to determine the possible fault events that the source node of the target input may cause, including: determining the input set of the source node of the target input in the logic model; using all nodes in the input set as the logic gate analysis object, using LLM to determine the possible fault events that the logic gate analysis object may cause, and obtaining a set of candidate inducing factor events; the step of adding the inducing factor event forming node as a child node of the corresponding node is to add the inducing factor event forming node as a child node of this child node.

[0192] In some implementations, in response to determining that the source node of the target input is an OR gate, LLM is used to determine the possible fault events that the source node of the target input may cause, including: determining the input set of the source node of the target input in the logic model; adding time nodes in the fault tree; the parent node of the time node is the node corresponding to the top event, and the event content attribute of the time node includes that all nodes in the input set are abnormal; using all nodes in the input set as logic gate analysis objects, LLM is used to determine the possible fault events that the logic gate analysis objects may cause, and a set of candidate triggering factor events for the time node is obtained; the step of adding the triggering factor event forming node as a child node of the corresponding node is to add the triggering factor event forming node as a child node of the time node.

[0193] 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.

[0194] Figure 14 This is a schematic diagram of the electronic device provided in an embodiment of this application. For example... Figure 14As shown, the electronic device 14 of this embodiment includes: a processor 1401, a memory 1402, and a computer program 1403 stored in the memory 1402 and executable on the processor 1401. When the processor 1401 executes the computer program 1403, it implements the steps in the various method embodiments described above. Alternatively, when the processor 1401 executes the computer program 1403, it implements the functions of each module / unit in the various device embodiments described above.

[0195] Electronic device 14 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 14 may include, but is not limited to, processor 1401 and memory 1402. Those skilled in the art will understand that... Figure 14 This is merely an example of electronic device 14 and does not constitute a limitation on electronic device 14. It may include more or fewer components than shown, or different components.

[0196] The processor 1401 may 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.

[0197] The memory 1402 can be an internal storage unit of the electronic device 14, such as a hard disk or RAM of the electronic device 14. The memory 1402 can also be an external storage device of the electronic device 14, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, FlashCard, etc., equipped on the electronic device 14. The memory 1402 can also include both internal and external storage units of the electronic device 14. The memory 1402 is used to store computer programs and other programs and data required by the electronic device.

[0198] 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.

[0199] 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.

[0200] 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 method for intelligent modeling of fault tree based on MBSE-LLM, characterized in that, include: A logical model of the analysis object is constructed using the Model-Based Systems Engineering (MBSE) method, and a knowledge query agent is constructed based on the enhanced knowledge of the analysis object. The logical model includes at least nodes corresponding to the hierarchical entity components of the analysis object; Define the node attributes of each node in the fault tree; the node attributes shall include at least a set of sub-events. Obtain the top event and its description, determine the node to be analyzed corresponding to the top event, and determine the output set of the node to be analyzed, as well as the source node of each output in the output set; Traverse the output set, and in response to the anomaly that the target output is determined based on the Large Language Model (LLM), the top event will occur. Use the LLM to determine the set of sub-events based on the top event, the top event description, the node to be analyzed, the target output, and the source node of the target output. The target output is any output in the output set; The union of the sub-event sets of each output in the output set is used as an intermediate event list and added to the sub-event set attribute of the node to be analyzed to obtain the sub-nodes of the node to be analyzed. For any child node of the node to be analyzed, the LLM is used to determine the list of environmental factor events, the list of human factor events, the list of spontaneous factor events, and the list of induced factor events based at least on the knowledge query agent. Each environmental factor event, human factor event, spontaneous factor event, and induced factor event forms a node and is added as a child node of the corresponding node. Each child node is iteratively processed until the modeling stopping condition is met, and the fault tree is obtained. The set of sub-events for the target output is determined in the following way: Signal flow association judgment prompts are constructed based on the target output and its source nodes; the signal flow association judgment prompts include at least the background of the association judgment task, supplementary materials, and precautions for signal flow association judgment; the supplementary materials are determined based at least on the knowledge query agent, including the functional description of the node corresponding to the target output in the logical model and the functional description of the node corresponding to the source node of the target output in the logical model; the precautions for signal flow association judgment include at least the analysis object, the definition of signal flow association, the analysis scope, and the output requirements; Based on the signal flow association judgment prompt words, a signal flow association judgment analysis task is generated, and the LLM is used for intelligent judgment analysis to determine whether the target output anomaly will cause the top event to occur. If so, then fault event retrieval prompts are constructed based on the target output and its source nodes; the fault event retrieval prompts include at least the retrieval task background, the supplementary materials, and fault event retrieval precautions; the fault event retrieval precautions include at least the analysis object, fault event definition, fault event quantity threshold, and output requirements; Based on the fault event retrieval prompts, a fault event retrieval and analysis task is generated, and the LLM is used for intelligent judgment and analysis to obtain the intermediate event set of the target output. For each event in the intermediate event set, construct a verification prompt word; the verification prompt word shall at least include the verification task background, the supplementary materials, and verification notes; the verification notes shall at least include the analysis object, the verification definition, and the output requirements. Based on the verification prompts, a verification task is generated, and the LLM is used for intelligent judgment and analysis. All events that pass the verification are combined to obtain the sub-event set of the target output.

2. The method of claim 1, wherein, Based on the enhanced knowledge, a knowledge query agent is constructed, including: The enhanced knowledge is dynamically semantically aggregated and sliced ​​to obtain knowledge slices, and the knowledge slices are vectorized to obtain a knowledge slice vector set; Using a preset retrieval algorithm, a preset user query is retrieved from the knowledge slice vector set to obtain candidate knowledge slice vectors. Based on the preset user query and the candidate knowledge slice vector, query prompt words are constructed. Based on the query prompt words, the LLM is guided to determine the query answer corresponding to each preset user query, thus forming the knowledge query agent.

3. The method of claim 1, wherein, The attributes of each node in the fault tree also include the event type and the logical model node where the corresponding event occurred; The event type of the node to be analyzed is the top event type; The output set of the node to be analyzed is the output set of the logical model node in the logical model where the corresponding event occurs; The source node of each output in the output set corresponds to the internal unit of that output in the logical model.

4. The method of claim 1, wherein, Determine the lists of environmental factor events, human factor events, and spontaneous factor events, including: Analysis prompts are constructed for environmental factor events, human factor events, and spontaneous factor events respectively. Each analysis prompt includes the analysis task background, supplementary event materials, and analysis precautions. The supplementary event materials are determined at least based on the knowledge query agent and include the functional description of the corresponding node in the logical model and factor analysis information. Environmental factor analysis prompts include environmental load factor analysis information, human factor analysis prompts include human factor analysis information, and spontaneous factor analysis prompts include spontaneous factor analysis information. The analysis precautions include at least the analysis object, event definition, and output requirements. Based on the analysis prompts, an analysis task is generated, and the LLM is used for intelligent judgment and analysis to determine the list of environmental factor events, the list of human factor events, and the list of spontaneous factor events. The steps to add the nodes that form each environmental factor event, human factor event, and spontaneous factor event as child nodes of the corresponding node are as follows: add each environmental factor event, human factor event, and spontaneous factor event as a leaf node to this child node.

5. The method according to claim 1, characterized in that, Identify a list of triggering events, including: Query the logic model to determine the input set of this child node, and the source node of each input in the input set; the input set of this child node is the input set of the logic model node where the corresponding event occurs in the logic model, and the source node of each input in the input set corresponds to the internal unit of this input in the logic model; Traverse the input set, and in response to the determination based on the LLM that an anomaly of the target input will cause the top event to occur, use the LLM to determine the possible fault events that the source node of the target input may cause, to obtain a candidate triggering factor event set; wherein, the target input is any input in the input set; The LLM is used to verify the candidate triggering factor event set, and the triggering factor event list is determined based on the verified triggering factor events.

6. The method according to claim 5, characterized in that, The source node of the target input is a logic gate in the logic model; the logic gate includes an AND gate or an OR gate; In response to determining that the source node of the target input is an AND gate, the LLM is used to determine possible fault events that the source node of the target input may cause, including: Determine the set of inputs of the source node of the target input in the logical model; Using all nodes in the input set as logic gate analysis objects, the LLM is used to determine the possible fault events that the logic gate analysis objects may cause, and a candidate inducing factor event set is obtained. The steps to add the triggering event formation node as a child node of the corresponding node are as follows: add the triggering event formation node as a child node of this child node.

7. The method according to claim 6, characterized in that, In response to determining that the source node of the target input is an OR gate, the LLM is used to determine possible fault events that the source node of the target input may cause, including: Determine the set of inputs of the source node of the target input in the logical model; Add time nodes to the fault tree; the parent node of the time node is the node corresponding to the top event, and the event content attribute of the time node includes that all nodes in the input set are abnormal; Using all nodes in the input set as logic gate analysis objects, the LLM is used to determine the possible fault events that the logic gate analysis objects may cause, and a set of candidate triggering factor events for the time node is obtained. The step of adding the triggering event formation node as a child node of the corresponding node is to add the triggering event formation node as a child node of the time node.

8. A fault tree intelligent modeling device based on MBSE-LLM, characterized in that, include: The building module is configured to construct a logical model of the analysis object using the Model-Based Systems Engineering (MBSE) approach, and to construct a knowledge query agent based on the enhanced knowledge of the analysis object. The logical model includes at least nodes corresponding to the hierarchical entity components of the analysis object; The definition module is configured to define the node attributes of each node in the fault tree; the node attributes include at least a set of sub-events. The determination module is configured to obtain the top event and the top event description, determine the node to be analyzed corresponding to the top event, and determine the output set of the node to be analyzed, as well as the source node of each output in the output set; The determining module is also configured to traverse the output set, and in response to an anomaly in the determination of the target output based on the Large Language Model (LLM) causing the top event to occur, the LLM is used to determine a set of sub-events based on the top event, the top event description, the node to be analyzed, the target output, and the source node of the target output. The target output is any output in the output set; The modeling module is configured to take the union of the sub-event sets of each output in the output set as an intermediate event list and add it to the sub-event set attribute of the node to be analyzed, thereby obtaining the sub-nodes of the node to be analyzed. The modeling module is further configured to, for any child node of the node to be analyzed, use the LLM to determine at least based on the knowledge query agent a list of environmental factor events, a list of human factor events, a list of spontaneous factor events, and a list of induced factor events, and add each environmental factor event, human factor event, spontaneous factor event, and induced factor event into a node as a child node of the corresponding node, iteratively process each child node until the modeling stopping condition is met, and obtain the fault tree; The set of sub-events for the target output is determined in the following way: Signal flow association judgment prompts are constructed based on the target output and its source nodes; the signal flow association judgment prompts include at least the background of the association judgment task, supplementary materials, and precautions for signal flow association judgment; the supplementary materials are determined based at least on the knowledge query agent, including the functional description of the node corresponding to the target output in the logical model and the functional description of the node corresponding to the source node of the target output in the logical model; the precautions for signal flow association judgment include at least the analysis object, the definition of signal flow association, the analysis scope, and the output requirements; Based on the signal flow association judgment prompt words, a signal flow association judgment analysis task is generated, and the LLM is used for intelligent judgment analysis to determine whether the target output anomaly will cause the top event to occur. If so, then fault event retrieval prompts are constructed based on the target output and its source nodes; the fault event retrieval prompts include at least the retrieval task background, the supplementary materials, and fault event retrieval precautions; the fault event retrieval precautions include at least the analysis object, fault event definition, fault event quantity threshold, and output requirements; Based on the fault event retrieval prompts, a fault event retrieval and analysis task is generated, and the LLM is used for intelligent judgment and analysis to obtain the intermediate event set of the target output. For each event in the intermediate event set, construct a verification prompt word; the verification prompt word shall at least include the verification task background, the supplementary materials, and verification notes; the verification notes shall at least include the analysis object, the verification definition, and the output requirements. Based on the verification prompts, a verification task is generated, and the LLM is used for intelligent judgment and analysis. All events that pass the verification are combined to obtain the sub-event set of the target output.

9. 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 7.

Images