Fault tree construction methods, troubleshooting question and answer methods, and their systems and media

By constructing a directed acyclic graph fault tree and adaptively extracting local subgraphs based on local topological integrity, the problem of insufficient fault tree topology verification in existing technologies is solved, thereby improving the accuracy and interpretability of fault diagnosis.

CN122797684BActive Publication Date: 2026-11-10WUXI UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611283227.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-08-24
Publication Date
2026-11-10
Estimated Expiration
2046-08-24

AI Technical Summary

Technical Problem

Existing technologies struggle to effectively verify the true physical topology during fault tree construction, leading to erroneous cross-system reasoning and reversed causal directions, thus affecting the accuracy of fault diagnosis.

Method used

By extracting explicit causal directed edges and verifying implicit causal candidate edges by combining semantic similarity and device space constraint matrix, a directed acyclic graph fault tree is generated. The minimum cut set and fault propagation path are calculated using logic gate properties, and local subgraphs are adaptively extracted by combining local topological integrity to perform troubleshooting question answering.

Benefits of technology

It improves the verifiability and quantifiability of fault trees, reduces cross-system erroneous reasoning and causal inversion, and enhances the accuracy and interpretability of fault diagnosis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122797684B_ABST
    Figure CN122797684B_ABST
Patent Text Reader

Abstract

This invention discloses a fault tree construction method, a troubleshooting question-and-answer method, and their systems and media, belonging to the field of fault diagnosis technology. The fault tree construction method extracts fault events and explicit causal edges from complex equipment documentation, and forms implicit causal candidate edges through semantic, equipment space, physical relationship, and causal direction verification. Then, a directed acyclic graph fault tree is generated through loop detection. Subsequently, the probability of minimum cut set occurrence and the likelihood of fault propagation path occurrence are calculated, and counterfactual elimination is performed based on the reachable baseline before deletion to identify single-point faults. The troubleshooting question-and-answer method of this invention adaptively extracts second- or third-order local subgraphs based on local topological integrity, inputs the topological sequence and component data into a large language model, and generates troubleshooting suggestions constrained by the real topology.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and in particular to a fault tree construction method, a troubleshooting question and answer method, and their systems and media. Background Technology

[0002] Fault tree analysis is a commonly used method in industrial reliability engineering. In recent years, large language models and retrieval-enhanced generation techniques have been gradually introduced into the field of fault diagnosis for complex industrial equipment such as smart grids, aerospace, and rail transportation to assist maintenance personnel in knowledge retrieval and troubleshooting Q&A.

[0003] Existing graph retrieval enhancement generation schemes typically input large language models after serializing subgraphs or knowledge triples. If the sequence lacks information on node hierarchy, edge direction, logic gates, and physical subsystems, the topological constraints of the fault tree will be weakened. The model may incorrectly associate physically unrelated components or reverse causal directions, resulting in a fault illusion. While methods such as Bayesian networks and minimal cut sets can perform probability or fault combination calculations, they cannot independently constrain the generation range of large language models.

[0004] Therefore, it is necessary to construct and validate the real physical topology and extract local subgraphs that retain directed relationships during question answering. Summary of the Invention

[0005] Purpose of the invention: To address the problem of how to construct a verifiable directed acyclic fault tree and use local real topology constraints for troubleshooting and question answering to reduce cross-system erroneous reasoning, this invention provides a fault tree construction method, a troubleshooting question answering method, and their system and medium.

[0006] Technical solution: In a first aspect, the present invention provides a fault tree construction method, comprising the following steps:

[0007] Explicit causal directed edges are extracted, and candidate node pairs that are verified by semantic similarity, physical relationship represented by device space constraint matrix and causal direction are determined as implicit causal candidate edges to be added to the graph.

[0008] The implicit causal candidate edges are temporarily added to the dynamic adjacency list and loop detection is performed. When a loop is detected, the implicit causal candidate edges are rolled back. The implicit causal candidate edges that have not formed a loop are retained as implicit causal directed edges, and a directed acyclic graph fault tree is generated.

[0009] The directed acyclic graph fault tree is mapped to a weighted probabilistic network. Minimum cut sets are generated based on the logic gate properties. Then, the probability ranking of minimum cut set occurrence and the probability ranking of fault propagation path occurrence are output respectively.

[0010] In the directed acyclic graph fault tree before node deletion, determine the set of valid bottom events that can reach the specified top event; construct a counterfactual temporary graph for the intermediate event nodes in the directed acyclic graph fault tree, and identify single-point fault nodes based on the changes in path connectivity of the valid bottom event set before and after node deletion.

[0011] This process involves extracting explicit causal directed edges from the acquired complex equipment documents, as well as extracting bottom events, intermediate events, and top events. Bottom events are the basic fault causes that are no longer decomposed downwards; intermediate events are fault states triggered by at least one lower-level event through a combination of logic gates and propagating upstream; and top events are the system-level target faults analyzed by the fault tree.

[0012] In this scheme, considering that explicit documentation often cannot cover implicit transmission relationships across documents and components, and that semantic approximation alone may incorrectly connect physically unrelated fault phenomena, this step calculates the semantic similarity between nodes not connected by the explicit causal directed edges. It then uses the device space constraint matrix to verify the physical relationships of candidate node pairs that reach the similarity threshold. Candidate node pairs that pass the verification are identified as implicit causal candidate edges. The output consists of structured node data and structured edge data containing the candidate edges and their verification evidence, for use in the next step (i.e., ...). Figure 1 Step S2) involves temporary map entry and loop detection.

[0013] Optionally, the structured node data includes at least the node's internal identifier, node name, event type, physical subsystem to which it belongs, and logic gate attributes, and the structured edge data includes at least the source node, target node, edge type, and evidence source.

[0014] In this scheme, the equipment space constraint matrix is ​​established based on the equipment ledger, wiring diagram, piping diagram, installation location table and / or expert confirmation data; the equipment space constraint matrix is ​​used to indicate that there is no physical relationship, a direct adjacent relationship or an indirect related relationship between the components corresponding to two nodes. The direct adjacent relationship includes direct connection through cables, pipes, mechanical interfaces or common housings. The indirect related relationship includes being in a common hot zone, a common electromagnetic influence zone, a shared cooling circuit or a shared power supply circuit.

[0015] In this scheme, a node set is established based on structured node data, and an edge set is established based on structured edge data; the topological sorting results are used to assign levels to each node and generate node coordinates and directed connections; the successor node set and in-degree of each node are recorded in the dynamic adjacency list; when a new directed edge is added, Kahn topological sorting is performed, and if the number of vertices output is less than the current total number of vertices, it is determined that there is a cycle, and the newly added directed edge (i.e., an implicit causal directed edge) is canceled.

[0016] Since topological connections alone cannot distinguish the probability of occurrence of each fault combination, this scheme establishes the following two sets of calculations: First, for the minimum cut set, the combination of base events is expanded according to the OR gate union rule and the AND gate Cartesian product rule, and the minimum cut set is obtained through subset absorption. The product of the probability of base events within the cut set is calculated under the condition that the base events are independent of each other. Second, for the ordered fault propagation path, a non-negative probability negative logarithm weight is set according to the explicit conditional probability of the directed edge or the prior failure probability of the source node. The directed edge with the target node being an AND gate is multiplied by a penalty coefficient greater than 1, and then the cumulative weight is calculated using the shortest path algorithm.

[0017] The minimum cut set is an unordered set of events, and the fault propagation path is an ordered sequence of edges. The two are sorted and output separately, and the minimum cut set is not mapped by the cumulative weight of the path.

[0018] To avoid misjudging single-point faults due to isolated bottom events that are already unreachable from the top event in the baseline directed acyclic graph before node deletion, this solution first determines the set of bottom events based on node type. and in the baseline directed acyclic graph Perform a depth-first search to obtain the effective set of bottom events. .in, Represents any bottom event in the bottom event set. In the diagram There exists from Top Event A directed path. When When empty, output an incomplete graph prompt and stop the judgment; when If not empty, delete intermediate event nodes sequentially. and its incoming and outgoing edges, only when All the bottom events in the counterfactual diagram The middle cannot reach the top event Only then will Marked as a single point of failure node.

[0019] Optionally, key intermediate events can be selected first based on the probability of occurrence of the minimum cut set or the probability of occurrence of the fault propagation path, and then the structural irreplaceability can be independently verified by counterfactual testing.

[0020] Furthermore, considering that directly inputting the entire fault tree or only name-based triples into a large language model can easily lead to cross-subsystem associations, reversed directions, and expansion without evidence, further improvements are made to the above scheme: First, trigger nodes are located in the directed acyclic graph fault tree. Then, based on local topological integrity, second- or third-order local subgraphs are adaptively extracted, preserving edge directions, logic gates, and component boundaries. These subgraphs, along with corresponding component data, form topological and semantic dual-flow constraints, thereby limiting troubleshooting suggestions to the scope of true local causality. The probability of minimum cut set occurrence and path occurrence probability are used to rank candidate causes, and single-point fault marking is used to improve the priority of handling critical structural nodes.

[0021] Secondly, the present invention provides a troubleshooting question-and-answer method, which implements troubleshooting question-and-answer based on the fault tree construction method described in the first aspect. This troubleshooting question-and-answer method includes the following steps:

[0022] Receive troubleshooting text requests and locate the trigger node in the fault tree through entity links;

[0023] Centered on the trigger node, a second-order breadth-first search is first performed along the causal edge. When the second-order subgraph contains at least one upstream bottom event and at least one downstream top event, and there are no unincluded adjacent causal nodes at the second-order boundary node, the second-order subgraph is adopted. Otherwise, it is expanded to the third order and stops when the component boundary, the third-order depth, or the preset maximum number of nodes is reached, resulting in a locally connected subgraph.

[0024] Filter out the universally unique identifiers used only for internal graph computation from the serialized data of the locally connected subgraph, and generate a local topology sequence containing component names, event types, logic gate attributes and directed connection relationships according to the topological order;

[0025] Retrieve component semantic data corresponding to the components in the local connected subgraph;

[0026] The local topological sequence and component semantic data are input into the large language model as two-stream constraint information to generate troubleshooting suggestions limited to the local connected subgraph.

[0027] Thirdly, the present invention provides a troubleshooting question and answer system, including a troubleshooting question and answer module, wherein the troubleshooting question and answer module is configured to execute the troubleshooting question and answer method described in the second aspect.

[0028] Furthermore, it also includes an evidence fusion module, which is used to fuse the static prior basic probability assignment of the component with real-time sensor evidence and update the fault confidence of the fault tree node.

[0029] Fourthly, an electronic device according to the present invention includes at least one processor and a memory communicatively connected to the at least one processor; wherein the memory stores instructions which are executed by the at least one processor to cause the at least one processor to implement the troubleshooting and question-and-answer method described in the second aspect when executing the instructions.

[0030] Fifthly, a computer-readable storage medium storing computer-executable instructions for causing a computer to perform the troubleshooting question-and-answer method as described in the second aspect.

[0031] Compared with existing technologies, the significant advantages of this invention are: 1. At the fault tree construction level, this invention first recalls potentially missed associations from nodes without explicit edges using semantic similarity, then eliminates candidate edges lacking electrical, mechanical, thermal, or shared loop foundations using the device space constraint matrix, and cancels loop edges during incremental connections using Kahn topology sorting, thereby simultaneously reducing the risks of missing edges, incorrect edge patching, and circular causality; on this basis, it uses logic gates to expand the minimum cut set, calculates the probability of occurrence of the minimum cut set and the path weight of the probability negative logarithm, and combines the elimination of counterfactual nodes with reachable baselines before deletion to obtain the probability of occurrence results and structural vulnerability results, giving the fault tree verifiable, quantifiable, and traceable analytical capabilities.

[0032] 2. At the troubleshooting question-and-answer level, this invention determines the trigger node through entity links, adaptively extracts second- or third-order locally connected subgraphs based on local topological integrity, retains component names, event types, logic gates, and edge directions according to topological order, and forms dual-flow constraint information with the corresponding component data. This limits the retrieval and generation scope of the large language model to the real local topology, thereby reducing troubleshooting suggestions involving cross-subsystem associations, reversed causal directions, and lack of evidence, and improving the interpretability of the output. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the overall process of the fault tree construction method of the present invention;

[0034] Figure 2 This is a schematic diagram of the functional module structure of the fault tree construction system of the present invention;

[0035] Figure 3 This is a schematic diagram of the logic for fusion of live data of complex equipment based on DS evidence theory according to the present invention;

[0036] Figure 4 This is a schematic diagram of the hardware physical structure of the electronic device of the present invention;

[0037] Figure 5 This is a schematic diagram of the implicit causal mining and minimum cut set path optimization mechanism for complex equipment of the present invention;

[0038] Figure 6 This is a schematic diagram of the counterfactual stress test and single-point fault elimination deduction logic of the present invention;

[0039] Figure 7 This is a diagram of the system's actual operation interface for mining the hidden causal features of complex equipment according to the present invention;

[0040] Figure 8 This is an interface diagram of the fault topology dynamic foolproof self-calibration and interception alarm system of the present invention;

[0041] Figure 9 This is a schematic diagram of the minimum cut set occurrence probability ranking and fault propagation path occurrence probability ranking results of the present invention;

[0042] Figure 10 This is a schematic diagram of the counterfactual stress test results of the present invention;

[0043] Figure 11 This is a schematic diagram of DS evidence fusion and node state update of the present invention;

[0044] Figure 12 This is an interface diagram of the local subgraph addressing and dual-flow graph retrieval enhancement question-and-answer troubleshooting system of the present invention. Detailed Implementation

[0045] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and not intended to limit the scope of the invention. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the embodiments of the present invention, and not all structures.

[0046] In the following description, specific details such as target system architecture 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 are omitted so as not to obscure the description of this application with unnecessary detail.

[0047] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0048] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0049] Furthermore, in the description of this application and the appended claims, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0050] References to "one embodiment" or "some embodiments" in this specification mean that one or more embodiments of this application include the target features, structures, or characteristics described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized.

[0051] It should be noted that the system described in this application refers to Figure 2 The intelligent fault tree construction system for complex equipment shown includes a data extraction and mining module, a topology generation and calibration module, a probability inference module, and a counterfactual testing module, and may also include an evidence fusion and troubleshooting question-and-answer module. Furthermore, the system can convert node names, phenomena, and context into feature vectors.

[0052] The core of this application system is to transform the text corresponding to complex equipment documents into a directed acyclic fault tree with logic gates (AND / OR gates), then extract local subgraphs around the trigger nodes, and use the topology sequence and component data as topology information flow and semantic information flow, respectively.

[0053] Therefore, this application first proposes a fault tree construction method. This fault tree construction method is based on fault tree analysis, text entity relation extraction and graph data structure. It converts the fault objects, phenomena and causal statements in the maintenance procedure into bottom events, intermediate events, top events and directed edges, and uses AND gates and OR gates to represent the combination relationship.

[0054] It should also be noted that the large language model described in this application refers to a pre-trained text generation model based on a converter architecture, used for entity relation extraction and constrained troubleshooting text generation.

[0055] The node records described in this application include internal identifiers, names, event types, components, physical subsystems, logic gates, evidence locations, and prior probabilities; edge records include source nodes, target nodes, edge types, directions, evidence, and confidence levels. Components refer to independent parts such as bushings, windings, and tap changers; hidden cross-component connections refer to fault connections that are not explicitly documented but can be conducted through thermal, electromagnetic, cooling, or power supply relationships.

[0056] In one implementation, a high-voltage main transformer is taken as an example.

[0057] Figure 1 The fault tree construction method shown consists of a continuous processing relationship comprising steps S1 to S4. S1 is responsible for semantic similarity candidate recall, device spatial constraint matrix verification, and causal direction verification; S2 is only responsible for temporary graph entry of candidate edges, loop detection, and loop rollback; S3 calculates the probability of minimum cut set occurrence and the likelihood of fault propagation path occurrence; S4 performs structural counterfactual analysis based on the reachable baseline before deletion. Troubleshooting question answering is not part of the fault tree construction step, but rather involves calling the generated fault tree and the above analysis results to extract local subgraphs.

[0058] The data transfer relationships in each step are as follows: S1 outputs structured nodes, explicit edges, and implicit causal candidate edges that have undergone triple verification through semantic similarity, device space constraint matrix, and causal direction; S2 outputs a fault tree that satisfies directed acyclic constraints; this fault tree serves as the common input for S3, S4, and troubleshooting questions and answers. The two types of probability ranking output by S3 and the single-point fault label output by S4 serve as the basis for the troubleshooting order and structural criticality, respectively.

[0059] See Figure 1 The implementation methods for the above steps are as follows:

[0060] Step S1: Extract fault events and explicit causal edges; verify and complete implicit causal edges.

[0061] Considering that explicit documentation often cannot cover implicit transmission relationships across documents and components, and that semantic approximation alone may incorrectly connect physically unrelated fault phenomena, this step extracts fault events and explicit causal edges from the acquired complex equipment documents, and mines implicit causal edges verified by physical constraints. Specifically, it forms high-recall candidates based on semantic similarity, verifies physical reachability using the device space constraint matrix, and finally outputs reliable structured node data and structured edge data for use in step S2.

[0062] Complex equipment documentation includes one or more of the following: maintenance procedures, equipment manuals, and maintenance records. It should be noted that the specific complex equipment documentation varies depending on the type of equipment.

[0063] In practice, the data extraction and mining module reads maintenance procedures, equipment manuals, and maintenance records, segments them by paragraphs and sentences, and uses a large language model to extract base events, intermediate events, top events, and explicit causal directed edges. Base events are the basic fault causes that are no longer decomposed; intermediate events are fault states that are triggered by lower-level events through logic gates and continue to propagate; and top events are system-level target faults.

[0064] A causal directed edge points from the cause node to the result node, while an explicit edge refers to a connection that explicitly expresses the causal relationship in the original text or is clearly recorded in a wiring diagram or pipeline diagram.

[0065] For nodes without explicit edge connections , compute nodes , Corresponding feature vector , The cosine similarity is calculated as follows: In the formula, Represents a node With nodes Cosine similarity; and Representing nodes respectively With nodes The semantic feature vector; "·" represents the vector dot product; This represents the L2 norm of the corresponding semantic feature vector.

[0066] It should be understood that for node pairs with explicit edges, their causal relationship has been determined by the evidence in the documents or drawings, and there is no need to repeat the inference. For node pairs without explicit edge connections, the system calculates the cosine similarity of the node name, fault phenomenon and context corresponding feature vectors to discover candidate relationships that may have been missed due to scattered records or cross-component descriptions. In other words, the cosine similarity plays a role in candidate recall and cannot prove the physical causal relationship on its own.

[0067] The rows and columns of the equipment space constraint matrix M correspond to components or subsystems. Here, the equipment space constraint matrix M is determined by the equipment ledger to determine the component affiliation, and by the wiring diagram, piping diagram and installation location table to determine the cable, pipe, mechanical interface and spatial adjacency relationship. Then, the maintenance procedures and expert confirmation data supplement the relationship of common hot zone, electromagnetic influence zone, cooling circuit or power supply circuit.

[0068] The corresponding node With nodes Device space constraint matrix The following physical relationships exist: This indicates that there is no known physical relationship; This indicates a direct connection, such as a direct connection via cable, conduit, mechanical interface, or common housing. This indicates the existence of an explainable indirect influence path, such as being located in a common thermal zone, an electromagnetic influence zone, or sharing a cooling or power supply circuit.

[0069] In simple terms, semantic similarity is used to answer "whether the texts are likely related", while the device space constraint matrix M is used to answer "whether the device has the transmission conditions". Only when both levels of judgment pass simultaneously can the candidate edge enter the direction confirmation and loop verification.

[0070] In practice, based on semantic candidates and physical relationships, cosine similarity is compared with a preset threshold. Equipment space constraint matrix Joint determination. When When the semantic association between the node pair is insufficient, the inference of the implicit edges of that node pair is stopped; when and When a node pair is considered to have textual similarity but lacks a physical transmission basis, the candidate edge is discarded in step S1 while retaining the rejection reason; when and At that time, the source node and target node are determined according to the causal verbs in the procedure, the order of time, the fault mechanism, and the connection direction of the components.

[0071] Only candidate edges that pass the above verification are temporarily added to the adjacency list in step S2 for Kahn topology verification; if no cycle is formed, implicit causal edges are written; if a cycle is formed, step S2 removes the edge and records an alarm.

[0072] The written record must include at least the source node, target node, edge type, semantic similarity, matrix value, directional evidence, and textual evidence.

[0073] Preferably, the threshold The threshold can be determined on the validation set of labeled historical samples, with the goal of balancing candidate recall and false edge completion rate; when there are insufficient samples, the initial threshold can be obtained by K-means clustering, and then adjusted by equipment experts according to the false alarm situation.

[0074] like Figure 7 As shown, taking the long-term severe overheating of the transformer top oil and irreversible breakdown of the main winding insulation as an example, this embodiment takes... The system calculates the result after encoding the names of the two nodes, the fault symptoms, and adjacent procedural statements. ,because This forms semantic candidate edges; the equipment ledger and installation location table indicate that both belong to the main transformer body; cooling and insulation data indicate that the top oil temperature can accelerate the deterioration of the main insulation through a common oil-heat environment. The causal description in the maintenance procedure that prolonged high oil temperature accelerates insulation aging and may lead to insulation breakdown is determined to be "the transformer top oil temperature is seriously exceeded for a long time", which points to "irreversible breakdown of the main insulation of the winding".

[0075] At this point, step S1 completes the verification and submits the candidate edge to step S2. After temporarily adding the edge in step S2, the number of nodes output by the Karn topology sort equals the total number of vertices, and no loop is formed, so the edge is retained.

[0076] The structured edges retained after step S2 can be represented as: {Source node name: "Transformer top layer oil temperature seriously exceeds standard for a long period", Target node name: "Irreversible breakdown of winding main insulation", Edge type: "Hidden causal edge", Semantic similarity: 0.82, Equipment space constraint matrix value: 2, Directional evidence: "Long-term high oil temperature accelerates insulation aging", Text evidence location: "Corresponding clause of maintenance procedure"}. If If the evidence for the direction is inconsistent, then step S1 rejects the edge and does not include it in the graph; if step S1 passes but a loop is formed after temporary inclusion in the graph, then step S2 rolls back the edge.

[0077] Figure 7 The candidate edges to be verified are shown in dashed lines. After the candidate edges pass the triple verification of semantic similarity, device space constraint matrix and causal direction, they are then verified by loop in step S2 before they can be converted into valid directed edges in the fault tree.

[0078] like Figure 5 As shown, an example of abstract nodes in the same processing chain is given: A, B, and C represent underlying abnormal events, X and Y represent intermediate propagation events, and T represents the top event of the system; solid lines A→X, B→X, C→Y, X→T, and Y→T represent directed edges supported by existing evidence; dashed line A→C represents candidate edges obtained from semantic recall. These candidate edges first undergo triple verification in step S1, checking semantic similarity, device space constraint matrix, and causal direction; after passing S1, they undergo loop detection in step S2. Only after passing both stages is the edge written into the graph.

[0079] What needs to be known is that Figure 5 The nodes, logic gates, and prior probabilities in the equation are used in step S3 for the calculation of minimum cut sets and path weights.

[0080] Step S2: Generate a Directed Acyclic Graph (DAG) fault tree; detect loops and roll back when adding edges. Specifically, generate a DAG fault tree based on the structured data output from Step S1, and perform dynamic error-proofing self-calibration. This step solidifies the scattered nodes, edges, and evidence into a unified graph structure with consistent orientation and topological sortability, avoiding distortion of hierarchy, probability propagation, and local subgraph traversal caused by cyclic causality, and providing a common topological foundation for Steps S3, S4, and troubleshooting question answering.

[0081] In practice, the topology generation and calibration module establishes a vertex set based on node records. Construct an edge set using explicit edges and verified implicit edges. Based on vertex set Sum of edges Constructing candidate directed graphs .

[0082] Here, a candidate directed graph is first formed, and then a directed acyclic graph fault tree is obtained through topological verification. Moreover, the nodes retain attributes such as event type, component, logic gate, probability, and evidence. Among them, the event type is used to distinguish between bottom events, intermediate events, and top events; the component attribute is used to limit the physical boundary; and the logic gate attribute is used to represent the combined triggering rules of multiple input events and directly controls the OR gate union, AND gate Cartesian product, and corresponding probability calculation in step S3.

[0083] The system Perform topological sorting, assign node levels and coordinates according to the results, and draw directed connections to generate a DAG fault tree. Internal identifiers are used for indexing; the interface only displays component and event names. Its attributes serve as common data for steps S3, S4, and troubleshooting Q&A.

[0084] The system synchronously maintains a dynamic adjacency list, using the node identifier as the key to store the successor node and in-degree, which is used for incremental updates, subgraph traversal, and loop detection.

[0085] In other words, when a new directed edge is added, the system temporarily updates the adjacency list and performs a Kahn topological sort: starting from the node with an in-degree of zero, it outputs nodes one by one and updates the in-degree of its successor nodes. If the number of output nodes equals the total number of vertices, the newly added edge is retained and the node level is synchronized; if the number of output nodes is less than the total number of vertices, it means that there are still nodes whose in-degree cannot be reduced to zero, and it is determined that the newly added edge forms a cycle. The system immediately restores the adjacency list before the update and issues an alarm. This rollback mechanism ensures that the graph after each incremental edit always satisfies the DAG constraint.

[0086] like Figure 8 As shown, step S1 performs a triple check on the candidate edge pointing from "long-term severe over-temperature of transformer top oil" to "irreversible breakdown of winding main insulation," sequentially verifying semantic similarity, equipment space constraint matrix, and causal direction of the procedure. After passing the check, step S1 outputs the candidate edge and its evidence; only in step S2 is it temporarily added to the dynamic adjacency list and subjected to Kahn topology sorting. When the number of output nodes equals the total number of vertices, step S2 retains the edge and synchronously updates the node level.

[0087] If the reverse edge "Irreversible breakdown of main insulation of winding" → "Long-term serious over-limit of transformer top oil temperature" is temporarily added to the graph, this edge forms a loop with the original valid edge. Kahn sort cannot output all vertices. Step S2 immediately rolls back the newly added edge and restores the adjacency list before the update.

[0088] In the case where “irreversible breakdown of winding main insulation” is incorrectly connected as “extreme sandstorm exacerbates motor bearing wear”, step S1 determines that it has failed the physical or directional verification based on the component boundary, specification direction and physical constraint evidence, and the edge does not proceed to step S2.

[0089] Figure 8 The system displays "S1 Triple Check Rejection" and "S2 Loop Check Rollback" respectively, thus distinguishing between candidate edges that do not meet the requirements in terms of semantics, device space, or causal direction and candidate edges that form a loop.

[0090] Step S3: Generate minimal cut sets and calculate the probability ranking for both types of occurrences. Specifically, map the DAG fault tree obtained in Step S2 into a weighted probabilistic network: expand it by logic gates and calculate the probability of minimal cut set occurrences; calculate the probability of fault propagation paths by weighting the directed edge probabilities with negative logarithms. The two types of objects and ranking criteria are different, and are output separately and used together by operations and maintenance personnel to determine the troubleshooting order; Step S4 outputs structural critical results separately. See also Figure 9 The left side is sorted by the probability of minimum cut set occurrence, and the right side is sorted by the probability of fault propagation path occurrence.

[0091] The probability inference module reads the DAG fault tree and loads the logic gate attributes of each intermediate event, the prior failure probability of each base event, the conditional probability of each directed edge, and the source of evidence. Only edges that pass the loop check in step S2 participate in the unfolding and pathfinding. Probability fields without explicit numerical evidence are marked as insufficient evidence and are not replaced with fixed default values.

[0092] The system expands the combination of base events by OR gate union and AND gate Cartesian product, and obtains the minimal cut set (MCS) through subset absorption. Under the condition that the base events are independent, the probability of the minimal cut set is equal to the product of the prior failure probabilities of each base event in the cut set.

[0093] For the minimum cut set Under the condition that the basic events are independent, calculate Its equivalent negative logarithmic fraction is: ;in, Indicates the first Minimal cut set Denotes any bottom event in the cut set. Indicates bottom event The prior failure probability, This represents the joint probability that all bottom events within the cut set occur simultaneously. This represents the negative logarithmic fraction equivalent to the joint probability. The above multiplication sign indicates the relationship between the two probabilities. All Find the product. The larger, the corresponding The smaller the value, the higher the sorting priority.

[0094] For directed edges ,set up The probabilities used for path weight calculation are: first, the explicit conditional probability of the directed edge is taken; if there is no explicit conditional probability, the source node is taken. The prior failure probability is recorded. The source type. The basic edge weights are: ;in, Represents a directed edge The probability is weighted by a negative logarithm. When the target node... When using AND gate, adopt ,in , This represents the AND gate penalty coefficient. This penalty is only used for ordered path scoring and is not included in the product calculation of the minimum cut set occurrence probability.

[0095] It should be noted that the AND gate penalty coefficient can be determined based on historical fault path verification data or equipment expert annotations, and is recorded according to the equipment category or logic gate configuration.

[0096] For path Its cumulative weight is: ;in, This represents the ordered fault propagation path from the bottom event to the top event. Representing a path The cumulative sum of the weights of all edges. The system uses Dijkstra's shortest path algorithm according to... Sort the paths in ascending order, and by This represents the path occurrence probability score calculated from the cumulative weights. It is not equivalent to the probability of the minimum cut set occurring. Represented by natural constant an exponential function with base 0. The smaller, The larger.

[0097] The system outputs two types of results: sorted by the probability of minimum cut set occurrence. Sort in descending order of the probability of fault propagation path occurrence. The order is ascending. The former corresponds to an unordered combination of bottom events, while the latter corresponds to an ordered sequence of edges, and the minimum cut set is not mapped by the cumulative weight of the path. In step S4, intermediate events in the above sorting can be selected for counterfactual testing, but the structural criticality determination in S4 is independent of the probability sorting.

[0098] Combination Figure 5 As shown, let For the door, For OR gates, the confirmed solid line edge is... , , , and Then, after expansion by OR gate union, AND gate Cartesian product, and subset absorption, The minimum cut set is and .like , , ,but , Therefore, when ordered in descending order of probability of occurrence Arranged in Before.

[0099] Further set , AND gate penalty coefficient Then the path , and The cumulative weights are approximately 6.456, 6.943, and 6.215, respectively. The smaller the cumulative weight, the higher the probability score of the converted outcome. The higher, therefore The path nodes, edges, cumulative weights, probability scores, and evidence are listed first. After updating the explicit probability of live evidence, the path ranking is recalculated.

[0100] Step S4: Establish a reachable baseline before deletion, remove intermediate events and test path connectivity, and identify single-point fault nodes. Specifically, on the DAG fault tree obtained in step S2, first determine the set of valid baseline events, and then perform counterfactual node removal and connectivity testing.

[0101] In simple terms, step S3 answers "which failure combinations and paths are more likely to occur", while this step further answers "which intermediate events, if they fail, will cause the transmission link from all bottom events to top events to be interrupted". Both use the same topology but have different evaluation dimensions. You can first select key intermediate events based on the sorting in step S3, and then use counterfactual testing to verify their structural indispensability, thereby avoiding the omission of node parameters with low probability but critical to the structure.

[0102] The counterfactual testing module traverses the intermediate event node set K and records the baseline directed acyclic graph G, the bottom event set B, and the top event T before the node is deleted.

[0103] The system starts from the bottom event set The various bottom events in Departure, on the baseline map Perform a depth-first search to obtain the effective set of bottom events. .in, Represents any bottom event, In the diagram There exists from arrive A directed path.

[0104] when When empty, the system outputs an incomplete spectrum warning and stops the SPOF determination; when When not empty, for each candidate intermediate event Constructing a counterfactual temporary graph , i.e., delete And its incoming and outgoing edges.

[0105] See Figure 6 and Figure 10 The system from The various bottom events in Departure, in the counterfactual temporary map A depth-first search is performed. Representation diagram Does it exist from to A directed path; if for all If all are false, then the path from the valid bottom event set to the top event is completely blocked. The node was already unreachable in the baseline graph before deletion. Isolated events do not participate in this judgment.

[0106] The system enters the single-point fault marking step only when the set of valid bottom events is not empty and all paths from valid bottom events to top events are blocked after deleting candidate intermediate events.

[0107] Only when the target intermediate event is deleted Afterwards, for all All meet Only then will Marked as a single point of failure node; if any valid basic event can still be reached via a redundant path. If not, then no SPOF determination is made.

[0108] by Figure 5 For example, in the baseline graph before node deletion , , All are accessible ,therefore Delete intermediate events back, , through arrive The path disappeared, but It is still connected, therefore Not a global single point of failure; delete back, , Still can be arrive , It is also not a global single point of failure. If there are other isolated bottom events... Unreachable on the baseline map ,but It is recorded as an original unreachable node, but is not used to... or It was misclassified as SpOF.

[0109] Figure 6 Further illustrating the case of a common intermediate node k: if the effective set of events... All paths in the middle are through arrive ,delete After adding and removing incoming and outgoing edges, the number of effective bottom events increases from before deletion. If it is reduced to 0, then... Marked as SPOF. This determination satisfies both the conditions of being reachable before deletion and being globally interrupted after deletion.

[0110] exist Figure 10 In the counterfactual stress test results shown, the system sequentially simulates the deletion of intermediate events, re-executes the depth-first search, and marks the nodes that meet the above judgment conditions as single-point fault nodes. The results list the set of effective bottom events, the original unreachable bottom events, the test nodes, the number of reachable nodes before and after deletion, the remaining paths, and the judgment conclusions, which are used by maintenance personnel to evaluate whether to add redundant paths or logical protection measures.

[0111] In one embodiment, see Figure 3 and Figure 11 Considering that static prior probabilities are difficult to reflect equipment operating conditions in a timely manner, and that a single sensor may be affected by noise, missing measurements or conflicting evidence, this embodiment performs DS fusion of static priors with real-time evidence such as oil temperature, partial discharge, pressure and vibration without changing the topology of step S2, dynamically updates the node fault confidence, and triggers step S3 to recalculate the minimum cut set occurrence probability and the probability of fault propagation path occurrence.

[0112] It should be understood that the application is based on the nodes and topology of step S2, and is not based on step S4 as a prerequisite for implementation; the SPOF structure markers generated in step S4 can be displayed side by side with the fused high-confidence results to improve the priority of handling, but the two are not merged into the same probability value.

[0113] Sensor data is mapped into fault, normal, and uncertain evidence by interval.

[0114] The basic probabilities of each evidence source are assigned and combined using the Demster-Schaffer evidence theory (DS).

[0115] Let the static prior base probability of a certain component be assigned the value. The dynamic basic probability assignment of real-time sensor mapping is ; Identification framework ,in Indicates a fault. This indicates that it is normal. This indicates uncertainty that it is not yet possible to distinguish between a fault and normal operation. First, calculate the coefficient of conflict of evidence K:

[0116] ;

[0117] in, This represents the sum of the conflicting qualities of two sources of evidence. At that time, the fault confidence after fusion is: ;in, This represents the basic probability assignment of a component being in a faulty state after fusion. Obtained according to the same combination rule. and These represent the basic probability assignments for the normal state and the uncertain state, respectively.

[0118] When K ≥ 0.95, to avoid the Zadeh paradox causing the denominator to approach zero, the system is based on the reliability of static prior data. and sensor data reliability Determine normalized weights , And calculate:

[0119] ;

[0120] in, For status indication, in a single calculation, the status indication It can only be a fault state. Normal state and uncertain states one of the; Indicates in At that time, the alternative basic probability is assigned based on the weighted average of the reliability of the evidence source; and represent the normalized reliability weights of static prior evidence and real-time sensor evidence, respectively, and ,in, and It can be given based on the completeness, validity, and verification status of the evidence source.

[0121] Take when there is no reliable data The system also prompts for manual review. Therefore, low-conflict evidence is combined using Dempster's method, while high-conflict evidence is weighted and downgraded based on reliability, ensuring numerical stability during the fusion process.

[0122] For example, static priors , , Real-time evidence , , When the conflict coefficient K = 0.46, which is less than 0.95, and calculated according to Dempster's rule, the following results use the low-conflict fusion notation: , , .

[0123] After fusion Write back to the corresponding node, and update the path weight accordingly in step S3; if the node is also marked as SPOF in step S4, the system will use "high confidence and structurally critical" as the highest priority alarm.

[0124] The fusion result is pushed to the front end via a network socket; the front end updates the Scalable Vector Graphics (SVG) and Cascading Style Sheets (CSS) according to the node identifier, adjusting the node colors and connection widths. Figure 11 Show the update record.

[0125] The live fusion is performed in the following order: "sensor interval mapping - conflict coefficient calculation - low-conflict Dempster combination or high-conflict weighted degradation - node confidence write-back - recalculation in step S3", without changing the topology of step S2; the structure tag in step S4 remains valid and together with the updated occurrence probability result forms the disposal priority.

[0126] In one implementation, considering that directly inputting the entire fault tree or only name triples into a large language model can easily lead to cross-subsystem associations, reversed directions, and evidenceless expansion, see [reference needed]. Figure 12 In this embodiment, the triggering node is first located in the DAG fault tree formed in step S2, and then a second- or third-order local subgraph is adaptively extracted based on local topological integrity, forming topological and semantic dual-flow constraints with the corresponding component data. The occurrence probability ranking in step S3 is used to arrange candidate causes, and the SPOF marking in step S4 is used to improve the handling priority of structural critical nodes.

[0127] The system performs aliasing and entity linking on troubleshooting requests, and locates the triggering node by name, subsystem, and context; user confirmation is required when the confidence level is low.

[0128] The system performs a breadth-first search starting from the trigger node: expanding along incoming edges when tracing the cause and along outgoing edges when determining the impact. The system first generates a second-order subgraph; if the second-order subgraph already covers at least one upstream bottom event and at least one downstream top event, and the second-order boundary node has no unincluded directly adjacent causal nodes, then the second-order subgraph is used; where directly adjacent causal nodes include nodes directly connected to the boundary node via incoming or outgoing edges. Otherwise, it expands to a third-order subgraph and stops when the component boundary, the third-order depth, or the preset maximum number of nodes is reached.

[0129] During serialization, universally unique identifiers (UUIDs) used only for indexing are filtered out, while nodes, logical gates, and directed connections are preserved, and a topological information flow is formed in topological order.

[0130] The system retrieves relevant component information from regulations and manuals, forming a semantic information flow. These two types of information, along with constraints such as "only using given nodes and edges, no cross-subsystem reasoning, and prompting for review when evidence is insufficient," are input into the large language model.

[0131] Figure 12 In the process, after the user inputs "What to do about severe electrical and mechanical failure of OLTC", the system first normalizes "OLTC" to "on-load tap changer". Then, it locates "severe electrical and mechanical failure of on-load tap changer (OLTC)" as the trigger node by using the node name, the subsystem to which it belongs and the context. If multiple candidate nodes have similar scores, the system returns the candidate name for the user to confirm, instead of directly generating troubleshooting suggestions.

[0132] After localization, the system first extracts a second-order local subgraph along the incoming and outgoing edges. If the second-order boundary nodes still have unincluded incoming or outgoing neighbors, or if the upstream bottom event and downstream top event are not simultaneously covered, the system expands to the third order. This yields upstream events such as drive motor overcurrent and mechanism dead zone thermal lock-up, severe leakage due to deterioration of the switch oil chamber seal, overheating due to poor contact between the tap changer's moving and stationary contacts, and fatigue fracture of the energy storage spring in the fast mechanism, and retains the impact path up to "substation main transformer tripping". Expansion stops after reaching the third order, component boundary, or maximum number of nodes; if a complete context still cannot be formed, manual review is prompted.

[0133] The system deletes UUIDs used only for indexing and writes the sequence "upstream event → OLTC severe electrical and mechanical fault → main transformer tripping" along with the event types and logic gate attributes of each node into the topology information stream according to topological order. Simultaneously, it retrieves clauses related to OLTC isolation, contact resistance measurement, drive mechanism inspection, and reset verification from maintenance procedures and equipment manuals, forming a semantic information stream. This dual-stream information, along with constraints such as "only using given nodes and edges, no cross-subsystem reasoning, and prompting for manual verification when evidence is insufficient," is input into the large language model.

[0134] The large language model generates a "isolation-detection-verification" process: first, it performs safety isolation; then, it checks high-risk upstream causes according to step S3; it prioritizes the verification of key structural nodes marked in step S4; and finally, it verifies the downstream tripping status and procedural evidence. The output retains component names and evidence sources but does not display internal UUIDs. If a local subgraph or component data cannot support the conclusion, the expansion stops and manual confirmation is prompted.

[0135] In one embodiment, the present invention also provides, as Figure 2 The fault tree construction and troubleshooting application system shown includes a data extraction and mining module, a topology generation and calibration module, a probability inference module, and a counterfactual testing module. These four modules correspond to S1 to S4. It also includes an evidence fusion and troubleshooting question and answer module, which are used to apply the constructed fault tree.

[0136] The data extraction and mining module is used to obtain complex equipment documents, extract fault events and explicit causal edges, and perform physical relationship and causal direction verification by combining semantic similarity and equipment spatial constraint matrix, and output structured nodes, explicit edges and implicit causal candidate edges; candidate edges that fail physical or directional verification are directly rejected in this module.

[0137] The topology generation and calibration module is used to temporarily add candidate edges that have passed the aforementioned verification to the dynamic adjacency list, generate a directed acyclic graph fault tree, and only perform loop detection, loop rollback, and alarm.

[0138] The probability inference module is used to generate minimum cut sets and output the probability ranking of minimum cut set occurrence and the probability ranking of fault propagation path occurrence, respectively.

[0139] The counterfactual testing module is used to establish a set of valid base events before deletion, construct a counterfactual temporary graph, and identify single-point fault nodes based on changes in path connectivity before and after deletion.

[0140] The evidence fusion module is used to assign static prior basic probability values ​​and real-time sensor evidence based on the Demster-Schaffer evidence theory (DS) fusion component, obtain posterior fault confidence, and trigger probability network updates.

[0141] The troubleshooting question-and-answer module is used to locate the trigger node in the fault tree, adaptively extract second- or third-order locally connected subgraphs based on local topological integrity, and input the local topological sequence and component semantic data as two-stream constraint information into the large language model; when the third-order range still cannot form a complete context, it outputs an insufficient evidence prompt. The above modules can be implemented by software code, or by field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), or a combination thereof.

[0142] In one embodiment, based on the method described in the above embodiments, Figure 4An example of the physical structure of an electronic device is provided, including: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus. The processor can invoke logical instructions in the memory to execute the methods described in the above embodiments.

[0143] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application.

[0144] Based on the methods in the above embodiments, in one embodiment a computer-readable storage medium is also provided, which stores a computer program that, when run on a processor, causes the processor to perform the methods in the above embodiments.

[0145] Based on the methods in the above embodiments, this application provides a computer program product that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0146] It is understood that the processor in the embodiments of this application 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, transistor logic devices, hardware components, or any combination thereof. The general-purpose processor can be a microprocessor or any conventional processor.

[0147] The method steps in the embodiments of this application can be implemented in hardware or by a processor executing software instructions.

[0148] Software instructions may consist of corresponding software modules, which may be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art.

[0149] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. Both the processor and the storage medium can reside within an ASIC.

[0150] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0151] A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted through a computer-readable storage medium.

[0152] Computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0153] Computer-readable storage media can be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

[0154] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A fault tree construction method, characterized in that, Includes the following steps: Extract bottom events, intermediate events, and top events from the acquired complex equipment documents; Explicit causal directed edges are extracted, and candidate node pairs verified by semantic similarity, physical relationships represented by the device space constraint matrix, and causal direction are identified as implicit causal candidate edges to be added to the graph; where the corresponding nodes With nodes Device space constraint matrix The following physical relationships exist: This indicates that there is no known physical relationship; Indicates a direct connection; This indicates the existence of an interpretable indirect influence path; where, when cosine similarity... Less than the preset threshold When the cosine similarity is high, stop inferring the implicit causal candidate edges for that node pair; when the cosine similarity is low, stop inferring the implicit causal candidate edges for that node pair. Greater than or equal to the preset threshold ,and When the cosine similarity is high, discard the implicit causal candidate edge and retain the rejection reason; when the cosine similarity is low, discard the implicit causal candidate edge and retain the rejection reason. Greater than or equal to the preset threshold ,and At that time, the source node and target node are determined according to the causal verbs in the procedure, the order of time, the fault mechanism, and the connection direction of the components; The implicit causal candidate edges are temporarily added to the dynamic adjacency list and loop detection is performed. When a loop is detected, the implicit causal candidate edges are rolled back. The implicit causal candidate edges that do not form loops are retained as implicit causal directed edges, and a directed acyclic graph fault tree is generated. The step of generating the directed acyclic graph fault tree includes: establishing a vertex set based on node records. An edge set is constructed using explicit causal directed edges and verified implicit causal directed edges. Based on vertex set Sum of edges Constructing candidate directed graphs For candidate directed graphs Perform topological sorting, assign node levels and coordinates according to the results, and draw directed connections to generate a directed acyclic graph (DAG) fault tree. Synchronously update the dynamic adjacency list, using node identifiers as keys to store successor nodes and in-degrees for incremental updates, subgraph traversal, and loop detection. The loop detection steps include: temporarily updating the dynamic adjacency list; starting from a node with an in-degree of zero, outputting nodes sequentially and updating the in-degree of successor nodes; if the number of output nodes equals the total number of vertices, retaining the newly added implicit causal directed edges and synchronizing the node levels; if the number of output nodes is less than the total number of vertices, determining that the newly added implicit causal directed edges form a loop, and rolling back to ensure that the graph after each incremental edit always satisfies the DAG constraint. The directed acyclic graph fault tree is mapped to a weighted probabilistic network. Minimum cut sets are generated based on the logic gate properties. Then, the probability ranking of minimum cut set occurrence and the probability ranking of fault propagation path occurrence are output respectively. In the directed acyclic graph fault tree before node deletion, determine the set of valid bottom events that can reach the specified top event; construct a counterfactual temporary graph for the intermediate event nodes in the directed acyclic graph fault tree, and identify single-point fault nodes based on the changes in path connectivity of the valid bottom event set before and after node deletion.

2. The fault tree construction method according to claim 1, characterized in that, The steps for generating the minimum cut set occurrence probability ranking and the fault propagation path occurrence probability ranking include: Read the directed acyclic graph fault tree and load the logical gate attributes of each intermediate event, the prior failure probability of each bottom event, the conditional probability of each directed edge and its evidence source; the probability field without clear numerical evidence is marked as insufficient evidence and is not replaced by the preset probability. The combination of base events is expanded by OR gate union and AND gate Cartesian product, and the minimum cut set is obtained by subset absorption. Under the condition that the basic events are independent of each other, calculate the product of the prior failure probabilities of the basic events in each minimal cut set, and generate a minimum cut set occurrence probability sort in descending order of the product; Prioritize setting the negative logarithmic path weights based on the explicit conditional probability of directed edges; if there is no explicit conditional probability, use the prior failure probability of the source node as a substitute and mark the probability source type; multiply directed edges with AND gates as target nodes by a penalty coefficient greater than 1; calculate the cumulative weight of each fault propagation path using the shortest path algorithm, and generate a sorted order of the probability of fault propagation paths by ascending order of cumulative weights. The probability ranking of the fault propagation path and the probability ranking of the minimum cut set are calculated and output respectively.

3. The fault tree construction method according to claim 1, characterized in that, The steps for identifying a single point of failure include: Starting from the set of bottom events identified by node type, perform a depth-first search in the original directed acyclic graph to determine the set of valid bottom events that can reach the specified top event. When the set of valid bottom events is empty, output an incomplete graph prompt and stop single-point fault determination. Candidate intermediate event nodes and their associated edges are deleted sequentially. Only when the set of valid bottom events is not empty, and all bottom events in the set of valid bottom events cannot reach the specified top event after deleting the candidate intermediate event node, the candidate intermediate event node is marked as a single-point fault node, and a structural stress test report comparing the reachability before and after deletion is output.

4. A troubleshooting question-and-answer method based on the fault tree construction method described in claim 1, characterized in that, This troubleshooting question-and-answer method includes the following steps: Receive troubleshooting text requests and locate the trigger node in the directed acyclic graph fault tree through entity links; Centered on the trigger node, perform a second-order breadth-first search along the causal edges. When the second-order subgraph has covered the upstream bottom event and the downstream top event, and there are no unincluded directly adjacent causal nodes at the boundary nodes, retain the second-order subgraph. The directly adjacent causal nodes include incoming and outgoing neighbors. Otherwise, expand to the third order. Stop when the third order, component boundary, or preset maximum number of nodes is reached, and a locally connected subgraph is obtained. Filter out the universally unique identifiers used only for internal graph computation from the serialized data of the locally connected subgraph, and generate a local topology sequence containing component names, event types, logic gate attributes and directed connection relationships in topological order; Retrieve component semantic data corresponding to the components in the local connected subgraph; The local topological sequence and component semantic data are input into the large language model as two-stream constraint information to generate troubleshooting suggestions limited to the local connected subgraph.

5. A troubleshooting question-and-answer system, characterized in that, It includes a troubleshooting Q&A module, which is configured to execute the troubleshooting Q&A method as described in claim 4.

6. An electronic device, characterized in that, It includes at least one processor and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that are executed by the at least one processor to cause the at least one processor to implement the troubleshooting question-and-answer method as described in claim 4 when executing the instructions.

7. A computer-readable storage medium, characterized in that, The storage medium stores computer-executable instructions, which are used to cause the computer to execute the troubleshooting question-and-answer method as described in claim 4.

Citation Information

Patent Citations

  • Fault detection method of flight power coupling system based on Bayesian model reasoning

    CN118133650A

  • Fault tree design method and system based on graphical interface, storage medium and program product

    CN121598223A