A fault library modeling method, device and product based on a decision chain graph

Abstract

The application discloses a fault library modeling method and device based on a decision chain graph, and a product; the method comprises the following steps: obtaining structured account data and unstructured text data; performing data format cleaning and text paragraph segmentation on the unstructured text data to obtain text segmentation segments; performing sentence-level standardization processing on natural language fields in the text segmentation segments and the structured account data to obtain target sentence-level text; identifying key semantic unit information from the target sentence-level text, the key semantic unit information being used for describing a decision unit four-tuple; performing directed chain combination, chain node merging and standardization processing on the decision unit four-tuple to obtain a plurality of decision chains, and generating a decision chain graph based on the plurality of decision chains, so as to model a subway power supply fault through the decision chain graph. The embodiment of the application can make knowledge have a decision chain structure of'state-judgment-action-result', thereby realizing flow modeling and intelligent response of the fault library.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and intelligent operation and maintenance technology for rail transit, and in particular to a fault database modeling method, equipment, and product based on decision chain graph. Background Technology

[0002] In urban rail transit systems, the subway power supply system is a critical infrastructure, and its stable operation directly affects train scheduling, passenger safety, and normal line operation. A power supply system failure can lead to serious consequences such as widespread train stoppages, passenger delays, and operational disruptions. Therefore, achieving rapid diagnosis, scientific assessment, and efficient handling of power supply failures has become a core challenge in the subway operation and maintenance system.

[0003] In the process of developing this application, the applicant discovered at least the following problems in the prior art:

[0004] Knowledge is fragmented and lacks a task-driven organizational structure: Currently, most knowledge materials are distributed in the form of "document paragraphs" or "database records," lacking an organizational method centered on "processing flow" or "judgment task chain." Operations and maintenance personnel can only retrieve knowledge points one by one, making it difficult to form a systematic "judgment-decision-execution" chain.

[0005] There is a disconnect between structured ledgers and unstructured documents: structured ledgers are accurate but lack in-depth analytical support; while regulations and expert documents contain a wealth of context and operational semantics, they cannot be linked and retrieved due to their unstructured nature. The lack of a linkage mechanism between these two types of knowledge prevents on-site personnel from effectively integrating historical experience with institutional standards.

[0006] The system cannot support state-driven intelligent recommendation and path reasoning: Current intelligent question answering systems generally rely on large language models or embedded vector semantic retrieval. While they can answer some standard questions, they lack the ability to deduce the "state judgment → conditional branch → processing suggestion" process in the context of engineering tasks. The system cannot reason out an executable processing path based on the current operating state.

[0007] Expert experience exists implicitly and is difficult to reuse digitally: Expert knowledge often exists as implicit experience in documents, oral accounts, or individual experiences, lacking explicit modeling and structural expression. Especially in critical fault handling paths, the judgment logic of experts is difficult to express and reuse through ordinary documents, making it difficult for novices to get started and resulting in non-standard processing.

[0008] In existing technologies, fault database modeling is generally carried out by establishing the following types of knowledge data systems: Power supply fault ledger system: a large number of structured fault records accumulated over a long period of time, including fault time, site, equipment type, fault phenomenon, analysis process, handling plan and recovery results, etc.; Rules and regulations and emergency response standards: formal operating procedures and rules and regulations formulated by the subway group, usually in PDF or Word format, which standardize the standard operating procedures under various typical working conditions; Expert experience documents and manuals: experiential knowledge written or summarized by senior front-line engineers, including hazard identification techniques, special fault handling experience, handling priority suggestions, etc., usually stored as unstructured text documents. Summary of the Invention

[0009] This application provides a fault database modeling method, device, and product based on decision chain graphs, which can enable knowledge to have a decision chain structure of "state-judgment-action-result", thereby realizing the process-oriented modeling of the fault database.

[0010] In a first aspect, embodiments of this application provide a fault database modeling method based on decision chain graphs, the method comprising:

[0011] Obtain structured ledger data and unstructured text data corresponding to the subway power supply system;

[0012] Data format cleaning and text segmentation are performed on unstructured text data to obtain the corresponding text segments.

[0013] Sentence-level standardization is performed on the text segments corresponding to unstructured text data and the natural language fields in structured ledger data to obtain the target sentence-level text;

[0014] Key semantic unit information in the subway power supply fault handling process is identified from the target sentence-level text. The key semantic unit information is used to describe the decision unit quadruple, which includes: state node, judgment condition node, decision action node and result node.

[0015] The decision unit quadruples are subjected to directed chain combination and chain node merging and standardization to obtain multiple decision chains. A decision chain graph is then generated based on the multiple decision chains to model subway power supply faults.

[0016] Secondly, embodiments of this application provide an electronic device, including:

[0017] One or more processors;

[0018] Memory, used to store one or more programs.

[0019] When one or more programs are executed by one or more processors, the one or more processors implement the fault library modeling method based on decision chain graphs in any embodiment of this application.

[0020] Thirdly, embodiments of this application provide a computer program product on which a computer program is stored, which, when executed by a processor, implements the fault library modeling method based on decision chain graphs of any embodiment of this application.

[0021] This application proposes a fault database modeling method, device, and product based on a decision chain graph. The method involves acquiring structured ledger data and unstructured text data corresponding to the subway power supply system; cleaning the unstructured text data and segmenting it into text segments; performing sentence-level standardization on the text segments and natural language fields in the structured ledger data to obtain target sentence-level text; identifying key semantic unit information in the subway power supply fault handling process from the target sentence-level text. This key semantic unit information describes decision unit quadruples, which include: state nodes, judgment condition nodes, decision action nodes, and result nodes; performing directed chain combination and chain node merging standardization on the decision unit quadruples to obtain multiple decision chains; and generating a decision chain graph based on these multiple decision chains to model subway power supply faults. In other words, the technical solution of this application can reconstruct subway power supply knowledge into a "decision chain graph," giving the knowledge a decision chain structure of "state-judgment-action-result," thereby enabling process-oriented modeling of the fault database. However, existing technologies are severely lacking in knowledge process modeling, state perception, task-driven approaches, and reasoning execution, failing to support an intelligent auxiliary response system that "starts from the operating condition and dynamically reasons to process the path." Therefore, compared to existing technologies, the fault database modeling method, equipment, and product based on decision chain graphs proposed in this application can give knowledge a decision chain structure of "state-judgment-action-result," thereby enabling process-oriented modeling and intelligent response of the fault database. Attached Figure Description

[0022] Figure 1 A flowchart illustrating a fault database modeling method based on decision chain graphs provided in an embodiment of this application;

[0023] Figure 2 A flowchart illustrating a fault library modeling method based on decision chain graphs, provided as another embodiment of this application;

[0024] Figure 3 A flowchart illustrating a fault database modeling method based on decision chain graphs, provided in yet another embodiment of this application;

[0025] Figure 4 A schematic diagram of the structure of a fault database modeling device based on decision chain graphs provided in another embodiment of this application;

[0026] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0027] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the application and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present application, not the entire structure.

[0028] This embodiment reconstructs the knowledge implicit in traditional structured ledger records, rules and regulations, and expert experience documents into a reasonable, executable, and visualized "decision chain graph" structure, thereby realizing process-oriented modeling and intelligent response to subway power supply fault knowledge.

[0029] Current knowledge systems generally suffer from technical shortcomings such as "static data, loose structure, and lack of decision-making logic," making it difficult to meet the knowledge support needs of frontline maintenance personnel throughout the entire process of "understanding operating conditions—judgment basis—handling suggestions" during fault handling. The goal of this embodiment is to achieve technological breakthroughs in the following key aspects:

[0030] This embodiment explicitly models the fault handling process in documents and ledgers as a "decision chain graph": Subway power supply knowledge data widely contains implicit chain-like decision-making processes of "state judgment—conditional basis—action execution—feedback results," but existing systems only treat these as text paragraphs or static records, lacking structured mining and logical modeling capabilities. This embodiment uses a dedicated modeling method to abstract various knowledge fragments into "state nodes," "judgment condition nodes," "decision action nodes," and "result nodes," connecting them through a graph structure to form a "decision chain graph" with deductive capabilities.

[0031] Supports graph-based reasoning and multi-step suggestion recommendations based on current operating conditions: Based on user-inputted fault phenomena or equipment status, this invention automatically identifies the most matching starting node in the decision chain graph and deduces backward along the graph, sequentially presenting judgment conditions and possible action paths, ultimately outputting professional handling suggestions and their knowledge basis. Compared to traditional question-answering systems or fuzzy matching retrieval mechanisms, the decision chain graph can reconstruct the judgment logic of the expert handling process, achieving interpretability, visualization, and traceability of the handling path.

[0032] By integrating record-keeping, regulations, and expert experience, a decision-making chain knowledge network is constructed. Through a unified modeling structure, this embodiment supports the integration of knowledge from different sources into the same decision-making chain graph, including: typical fault cases from the record-keeping (such as a short-circuit handling process); standard operating procedures from regulatory documents (such as the judgment conditions in Article X of the regulations); and implicit processing priorities, conditional judgments, and implicit rules from expert experience. The graph can be labeled with the "knowledge source type" and "authority level," providing a basis for path selection and recommendation ranking.

[0033] Enhancing the intelligence and knowledge transferability of fault handling: This embodiment not only provides frontline maintenance personnel with full-process support from "current status to recommended handling path," but also visually displays the judgment process in a graph format, significantly improving the efficiency of knowledge absorption for new employees and the efficiency of expert experience transfer. Simultaneously, the graph structure supports incremental updates, node expansion, and path fusion, adapting to the dynamic integration of new equipment, new systems, and new cases, thus constructing a sustainably growing "intelligent decision-making knowledge network for subway power supply."

[0034] This embodiment transforms the existing subway power supply knowledge from an organization of "document paragraphs + static retrieval" to an intelligent decision chain graph structure of "operating condition driven + decision logic visualization + knowledge path derivation". This enables explicit modeling of expert knowledge processes and graph reasoning support, providing efficient, intelligent and reliable auxiliary decision-making capabilities for subway power supply system fault handling.

[0035] Figure 1 This is a flowchart illustrating a fault database modeling method based on a decision chain graph, provided in one embodiment of this application. This method can be executed by an electronic device, which can be implemented in software and / or hardware, and can be integrated into any smart device with network communication capabilities. Figure 1 As shown, the fault database modeling method based on decision chain graphs may include the following steps:

[0036] S101. Obtain the structured ledger data and unstructured text data corresponding to the subway power supply system.

[0037] In this step, the structured ledger data comes from the power supply system database and can be extracted using SQL (Structured Query Language) or API (Application Program Interface) calls. The retained fields include: record number, fault time, line, station, equipment name, fault description, analysis and judgment, handling measures, and handling results. The output standard is a JSON structure, and each record is a "knowledge instance".

[0038] Unstructured text data originates from regulations, expert manuals, and training materials in the subway power supply system, including various document formats such as PDF, Word, scanned copies, PPT, and Markdown. The main text is extracted directly from editable documents, and Tesseract OCR (configured with --psm 4) is used to recognize scanned copies. The paragraph and heading structures are obtained in conjunction with Layout-Parser. The output is structured paragraph data with fields including page number, heading, main text content, and text position.

[0039] S102. Perform data format cleaning and text segmentation on the unstructured text data to obtain the corresponding text segments.

[0040] In this step, data format cleaning can remove redundant formatting elements in the document (such as headers, footers, watermarks, etc.), identify and retain structural information (such as chapter titles and serial numbers), and segment paragraph units of "title + body text", laying a structural foundation for sentence-level extraction and improving the contextual accuracy and structural restoration ability of subsequent semantic extraction.

[0041] This embodiment introduces a TextRank-based graph ranking algorithm to identify potential "heading sentences" in unstructured documents. The algorithm constructs a sentence graph network and ranks sentences based on keyword co-occurrence relationships to determine the "information centrality" score of each sentence, which assists in heading detection and paragraph segmentation. Thus, in natural paragraphs without explicitly labeled headings, sentences that may be "chapter titles" or "section titles" are identified, providing semantic anchors for subsequent paragraph structure division.

[0042] Graph node construction: Each sentence is a node; the edge weights between nodes are set according to the content similarity between sentences. Edge weight calculation method: The cosine similarity is constructed using the following formula (1) term intersection. If the similarity is greater than the set threshold (such as 0.2), then an undirected edge is established.

[0043]

[0044] In formula (1), W i W j These are the sets of stop word keywords in sentences i and j, respectively.

[0045] Construct a sentence graph G = (V, E), where V represents the set of knowledge nodes and E represents the set of directed edges. Define the decision-making logical paths between nodes. Each edge has a weight W. ij The symbol indicates the degree of semantic proximity between sentences; the sentences in the diagram are interconnected, forming a semantic network of keyword co-occurrence.

[0046] Each node v∈V contains the following attributes: node_id (unique identifier); node_type (node ​​type), with values ​​including: S: System State, C: Condition, A: Action, R: Result; text (node ​​content, such as "Block I tripped"); source (knowledge source, such as "XX maintenance document"); core (authority rating, such as system = 1.0, experience document = 0.7); tags (additional attribute tags, applicable equipment, site, line, time, etc.). Each edge e∈E represents a decision jump relationship, such as: S→C: indicating that the system state triggers a certain judgment condition; C→A: indicating the action to be performed when the judgment condition is met; A→R: indicating the target state to be achieved after a certain operation; C→R: also allows direct jump to the result state after the judgment (such as "no processing required"). Graph multi-path support: One state node can correspond to multiple judgment conditions, forming multiple reasoning branches; Condition parallel / combination: Supports the combination of multiple condition nodes into composite judgments; Graph visualization: The graph structure can be rendered in the front end as a flowchart; Citation chain integrity: Each node and edge retains the original literature source for interpretable and auditable decision-making.

[0047] TextRank is based on the PageRank principle, iteratively updating the weight value of each node to reflect its importance in the whole graph. Initially, each sentence is assigned the same weight PR0 = 1.0; the iterative calculation is shown in the following formula (2).

[0048]

[0049] In formula (2), d is the damping factor, with a default value of 0.85; In(V i ) indicates pointing to sentence V i All other sentences; w ij For sentence V j To sentence V i The edge weights. The algorithm iterates until convergence (the difference is less than ε = 1e-5 or the maximum number of iterations reaches 100).

[0050] After the iteration is completed, all sentences are sorted by PR value; sentences that are in the top 10% or have a value higher than the mean + 1σ are considered "potential title sentences"; if the sentence also meets the format features (such as numbering, keywords, bolding, etc.), it is marked as a formal paragraph title.

[0051] Thus, even in historical regulations and scanned documents lacking manual labeling, automatic structural segmentation can still be achieved; the semantic accuracy of paragraph segmentation can be improved; and the ability to restore the strong structure of "weakly structured documents" can be formed by combining rule features (numbering, blank lines, fonts).

[0052] Optionally, the unstructured text data is cleaned and segmented into text segments to obtain the corresponding text segments, including:

[0053] The algorithm identifies target noise format items in unstructured text data and uses regular expressions to clean the format of these items. It then uses a title detection algorithm to perform title detection and format feature fusion on the cleaned unstructured text data. Finally, it segments the unstructured text data into paragraphs based on preset paragraph boundary division rules, resulting in the corresponding text segments.

[0054] In this step, the target noise format item can be such as page number format (e.g., page d+ of d+), header / footer (e.g., "Internal Document", "Confidential"), or scan residue mark (e.g., "Scan Date: 20xx").

[0055] The regular expression can be exemplified as follows.

[0056]

[0057]

[0058] Title detection algorithms, such as TextRank, treat each sentence as a node and construct a graph based on the co-occurrence of keywords between sentences. A graph ranking algorithm (see 5.1) is then used to assign weights to sentences; sentences with a weight greater than 0.25 are considered title candidates. When fusing format features, the title is checked for one of the following characteristics: it begins with a number such as "1.", "1.1", "(I)", or "3.2.1"; it contains keywords such as "General Principles", "Definitions", "Operating Specifications", or "Processing Methods"; the paragraph is in a large font, bold, or centered (this can be determined when font attributes are preserved in the PDF structure); and it is judged as a formal title using a dual-condition approach (content weight + formatting rules).

[0059] The preset paragraph boundary rules are as follows: if the content between two adjacent headings is the same paragraph, if there are two consecutive blank lines, or if the paragraph begins with a heading format after a page break, then the paragraph is considered to have ended. If the content after a heading is less than 3 lines, it will be merged with the next paragraph into a "short paragraph merging". Each paragraph maintains the structure: <paragraph number, paragraph heading, paragraph body>, while retaining the original page numbers and document location information (such as position offset, page numbering, etc.) for traceability and visualization.

[0060] The output structure of the text segments corresponding to unstructured text data is a list of structured paragraph objects, and each paragraph contains the following fields.

[0061]

[0062] This approach removes invalid formatting information and segments the document into standardized semantic paragraphs for easier subsequent information extraction. It establishes a natural language processing-friendly structured document input; ensures context-traceable structure (title + body text) for subsequent sentence-level information extraction; achieves knowledge source traceability through pagination location information; and provides precise segment location capabilities for expert visual review and graph annotation interfaces.

[0063] S103. Perform sentence-level standardization on the text segments corresponding to the unstructured text data and the natural language fields in the structured ledger data to obtain the target sentence-level text.

[0064] In this step, sentence-level standardization can divide the natural language text in unstructured documents and structured ledger fields into semantically complete and uniformly formatted "atomic sentence" units, and perform standardized transformation on the sentence content, providing accurate and model-processable input for subsequent semantic information extraction.

[0065] Unstructured paragraph text (i.e., text segments corresponding to unstructured text data) 1. Segmented main text paragraphs from regulations, expert manuals, and internal documents. Natural language fields in structured ledgers: such as "fault phenomenon", "handling method", "handling result", "analysis and judgment", etc.

[0066] Optionally, sentence-level standardization is performed on the text segments corresponding to the unstructured text data and the natural language fields in the structured ledger data to obtain the target sentence-level text, including:

[0067] Chinese language processing tools were used to segment unstructured text data and natural language fields in structured ledger data to obtain sentence-level text.

[0068] In this step, the rules for identifying sentence breaks include: standard Chinese sentence-ending symbols: “.”, “!”, “?”, “;”; and logical connectors such as “Next” and “After” as auxiliary segmentation boundaries. Special handling: list sentences following phrases such as “For example:” and “As follows:” are ignored without being broken; when merging semicolon sentences into one, the continuity between the preceding and following sentences is maintained (to avoid fragmented sentences).

[0069] The following is an example of Chinese sentence segmentation.

[0070]

[0071]

[0072] The redundancy and repetition sentences of the first sentence-level text are filtered using the similarity judgment rule to obtain the second sentence-level text.

[0073] In this step, the Jaccard similarity is used to judge redundancy for adjacent sentences using formula (3).

[0074]

[0075] In formula (3), A and B are adjacent sentences to each other.

[0076] If the similarity J(A,B) ≥ 0.8 and the keyword coincidence rate is high, then retain the first sentence and delete the redundant sentence. Thus, the problem of model misrecall caused by "editing template copying" or "repeated descriptions from multiple perspectives" in the document can be effectively avoided.

[0077] The text format and encoding of the second sentence-level text are normalized to obtain the third sentence-level text;

[0078] In this step, by unifying characters, punctuation, numerical units, etc., the corpus style is ensured to be consistent and adapted to model parsing.

[0079] The following Python standard library can be used for text format and encoding normalization processing. The replacement content includes:

[0080] Full-width punctuation → Half-width punctuation; English punctuation → Chinese equivalent characters; Unify Chinese numerical units: such as "megaohm", "MΩ" are unified as "MΩ"; Convert mixed unit formats: "2mΩ" → "2MΩ"; "2 megaohm" → "2MΩ"; Remove non-semantic symbols: such as "……", "~~", "---", etc.

[0081]

[0082] If there are no suspected sentence incomplete sentences in the third sentence-level text, then determine the third sentence-level text as the target sentence-level text; if there are suspected sentence incomplete sentences in the third sentence-level text, then perform short sentence completion and repair on the third sentence-level text to obtain the target sentence-level text;

[0083] Among them, if there is a sentence in the third sentence-level text with a sentence length less than the preset quantity word, then it is determined that there is a suspected sentence incomplete sentence in the third sentence-level text.

[0084] For example, if the sentence length is less than 4 words, judge the sentence as a suspected sentence incomplete: if it is an imperative sentence (such as "Observe insulation"), supplement the context of the previous sentence; if it is a symbol incomplete sentence (such as "As shown in the figure"), eliminate it. Normalize the numbers of list sentences ("(1) Replace the circuit breaker", "(2) Detect the grounding resistance") and merge the number prefixes. Convert incomplete dialogue sentences into declarative sentences (such as "Is there an alarm?" → "The system is detecting alarms.").

[0085] Thus, the text is divided into single sentences with complete semantic boundaries; redundant, repetitive, and noisy sentences are removed; text encoding and format are standardized, and punctuation, units, and numerical symbols are unified; and semantic atomic units that meet the requirements of the extraction model are output.

[0086] S104. Identify key semantic unit information from the target sentence-level text during the subway power supply fault handling process. The key semantic unit information is used to describe the decision unit quadruple.

[0087] In this step, the decision unit quadruple includes: a state node, a judgment condition node, a decision action node, and a result node. The decision unit quadruple is a set of four semantic units that are key to handling subway power supply faults, accurately identified from sentence-level text: State (S, Status): representing the abnormal behavior of equipment or system; Condition (C, Condition): judgment basis, environmental state, fault characteristics, etc.; Action (A, Action): handling measures, recommended operation or maintenance behavior; Result (R, Result): the system state after handling, the target recovery conditions or effect description.

[0088] S105. Perform directed chain combination and chain node merging and standardization on the decision unit quadruple to obtain multiple decision chains, and generate a decision chain graph based on the multiple decision chains, so as to model the subway power supply fault through the decision chain graph.

[0089] In this step, the extracted decision-making unit quadruples (S: state, C: condition, A: action, R: result) are combined into a directed process chain structure to express the causal, judgment, and handling logic in the subway power supply fault handling process. The constructed chain not only has machine-understandability but can also be used in application scenarios such as graph traversal, path recommendation, and process replay in inference engines.

[0090] Next, each general chain is attached with its data instance from the ledger, forming a "knowledge support set". For example, a decision_instance table structure is created, with fields including: record_id, chain_id, city, line, station, happen_time, recover_time, attachments[]. Attachments are uniformly uploaded to OSS, and the URI is written into the graph node metadata. Each chain can be bound to multiple instances to achieve a many-to-one support structure.

[0091] The credibility assessment and version archiving of the map content are performed, and the scoring function is shown in the following formula (4).

[0092]

[0093] In formula (4), N is the number of chain nodes, and score i The node score is given, and M is the number of bound instances. If σ < 0.7, it enters the "manual review queue"; the finished graph is exported in GraphML / JSON-L format; the graph version naming format is vYYYY.MM.DD-r sequence number; the graph content is finally deployed to the graph database (Neo4j / ArangoDB).

[0094] To enhance the practical applicability and data support of each decision chain in the graph, this embodiment introduces a decision instance binding mechanism. Each historical ledger record is treated as a "decision instance" and bound to a general decision chain to form an "instance set." The many-to-one binding model is as follows: Each decision chain has a unique chain_id; one or more ledger records are bound to this chain_id after being matched with the decision chain; each instance retains its original metadata fields, including the following example:

[0095]

[0096] The example in this embodiment supports filtering decision chains by dimensions such as geography, time, and device, and can be used to provide "supporting cases" during knowledge response, as well as perform various enhanced operations such as frequency statistics, processing time comparison, and linkage of attachment documents.

[0097] The fault database modeling method based on decision chain graph proposed in this application involves acquiring structured ledger data and unstructured text data corresponding to the subway power supply system; cleaning the unstructured text data and segmenting it into text segments; performing sentence-level standardization on the text segments and natural language fields in the structured ledger data to obtain target sentence-level text; identifying key semantic unit information in the subway power supply fault handling process from the target sentence-level text, where the key semantic unit information describes decision unit quadruples, which include: state nodes, judgment condition nodes, decision action nodes, and result nodes; performing directed chain combination and chain node merging standardization on the decision unit quadruples to obtain multiple decision chains; and generating a decision chain graph based on these multiple decision chains to model subway power supply faults. In other words, the technical solution of this application can reconstruct subway power supply knowledge into a "decision chain graph," giving the knowledge a "state-judgment-action-result" decision chain structure, thereby enabling process-oriented modeling of the fault database. In existing technologies, there are serious deficiencies in knowledge flow modeling, state perception, task-driven approaches, and reasoning execution, making it impossible to support an intelligent auxiliary response system that "starts from the working condition and dynamically reasons to process the path." Therefore, compared with existing technologies, the fault database modeling method, device, and product based on decision chain graphs proposed in this application can enable knowledge to have a decision chain structure of "state-judgment-action-result," thereby realizing process-oriented modeling and intelligent response of the fault database.

[0098] Figure 2 This is a flowchart illustrating a fault database modeling method based on a decision chain graph, provided as another embodiment of this application. Further optimizations and extensions can be made based on the above technical solution, and it can be combined with the various optional implementation methods described above. For example... Figure 2 As shown, the fault database modeling method based on decision chain graphs may include the following steps:

[0099] S201. Execute each preset regular expression rule sequentially on the target sentence-level text to obtain multiple entity fragments, and annotate the multiple entity fragments with nodes to obtain the rule-extracted sentences.

[0100] In this step, a pre-designed rule base covering 25 typical industry sentence structures is used to adapt to content with highly regular semantic expressions (especially regulations and standard expressions). The rules utilize regular expressions (RE2 engine) for high-performance matching. Typical templates are shown in Table 1.

[0101] Table 1. Examples of Typical Templates

[0102]

[0103] The matching process is as follows: each regular expression is executed sequentially on the sentence text; after a successful match, the corresponding entity fragment is extracted; and it is automatically labeled as an S, C, A, or R type node.

[0104] S202. Perform instruction-based model extraction on the text that meets the preset conditions in the target sentence-level text to obtain the model-extracted sentences.

[0105] In this step, the preset conditions include: rule not matched, high degree of freedom in sentence structure, and existence of nested structures. When a rule is not matched, or when there is high degree of freedom in sentence structure or nested structures, an imperative model is used for fallback extraction.

[0106] Optionally, imperative model extraction is performed on text that meets preset conditions in the target sentence-level text to obtain model-extracted sentences, including:

[0107] The sequence labeling model is used to extract the text that meets the preset conditions in the target sentence-level text using an instructional model to obtain the extracted sentences; and / or, the large language model is used to extract the text that meets the preset conditions in the target sentence-level text using an instructional model to obtain the extracted sentences.

[0108] In this step, the sequence labeling model is BERT-CRF, with a model structure based on chinese-bert-wwm-ext. Output labels are in BIO format: S, C, A, R, O (others). Sequence length is limited to 128. The training data consists of 3000 real-world labeled sentences from ledgers and regulatory documents. The labeling format is [BS]I segment [IS]bus trip [O], [BA]check [IA]grounding wire. Training parameters: optimizer: Adam, learning rate: 3e-5, batch size: 16, training epochs: 2; final evaluation metric: F1≈0.92. The model is exported in ONNX format; ONNXRuntime is used for accelerated deployment, with an average inference rate >30000 sentences / s (GPU).

[0109] Large language models such as PromptIE compress and express information across sentences, but these expressions contain ellipses, repetitions, ambiguities, and other elements that are difficult to template.

[0110] Example as follows:

[0111]

[0112] S203. Perform four-tuple structure transformation on the rule extraction statement and the model extraction statement respectively to obtain the key semantic unit information in the subway power supply fault handling process.

[0113] In this step, after extraction, each sentence will be converted into a four-tuple structure, which will then be used as input for the graph nodes. An example is shown below:

[0114]

[0115] The fault database modeling method based on decision chain graphs proposed in this application transforms natural language sentences into knowledge units in the form of quadruples through structural transformation.<S,C,A,R> It supports high recall rates covering various industry expressions; it outputs node components with a unified structure and semantic indexing capabilities for subsequent chain building and graph organization.

[0116] Figure 3 This is a flowchart illustrating a fault database modeling method based on decision chain graphs, provided as another embodiment of this application. Further optimizations and extensions of the above technical solution are possible, and it can be combined with the various optional implementation methods described above. For example... Figure 3 As shown, the fault database modeling method based on decision chain graphs may include the following steps:

[0117] S301, Custom chain structure.

[0118] In this step, the chain structure is constructed sequentially according to the state nodes, judgment condition nodes, decision action nodes, and result nodes. The decision chain is modeled using a directed graph (DAG) structure, with the chain structure being S→C→A→R, where each arrow represents the fault handling flow process of "trigger-judgment-action-feedback".

[0119] S302. Based on the chain structure, determine the connection direction of each semantic node in the decision unit quadruple according to the statement order and semantic modifiers in the corresponding statements.

[0120] In this step, the priority order (i.e., the order of statements) is as follows: if all four types of nodes exist, they are strictly arranged in the order S→C→A→R; if some nodes are missing, the "shortest connected path" is constructed, such as C→A→R or S→A→R. The causal / temporal order between sentences is determined by: using keyword priority, such as "subsequently," "then," and "finally" to indicate the order; words like "due to," "because," and "leading to" are used to infer the direction of causal edges; if the context lacks explicit guiding words, edges are constructed by default according to the extraction order.

[0121] It should be noted that before constructing the chain, the system will perform a semantic consistency check on the content of each quadruple. If the nodes extracted from multiple sentences in the same paragraph have semantic repetition or equivalent expression, they will be merged into the same node (by calling the merging module). Each node in the chain is identified by a unique node_id to avoid redundancy.

[0122] S303. Perform node text vectorization on each semantic node in the decision unit quadruple, and determine multiple semantic nodes with semantic repetition relationship through the node text vector corresponding to each semantic node, and merge the multiple semantic nodes with semantic repetition relationship.

[0123] In this step, each chain consists of multiple semantic nodes (S: state, C: condition, A: action, R: result). However, in different documents or instances, there may be nodes with different expressions but consistent semantics (e.g., "cable aging" and "deterioration of cable insulation performance" essentially refer to the same state). To improve the consistency of the graph's expression, node compression, and query efficiency, this embodiment performs merging and standardization processing on all extracted nodes.

[0124] This embodiment adopts a joint judgment strategy of "semantic vector similarity + text edit distance" to ensure the accuracy and stability of semantic merging. Node text vectorization uses a Chinese semantic vector model (such as text2vec-base-chinese or sentence-transformers / paraphrase-mpnet-base-v2 Chinese adaptation); each node text is encoded as a 768-dimensional vector to represent its semantic embedding. Similarity calculation and threshold setting: for any two nodes of the same type (i.e., S and S, C and C, etc.), the cosine similarity is calculated using formula (5). If the distance Sim(i,j) between the two node text strings is <18, it is judged that the text descriptions are similar; only when both indicators are satisfied simultaneously is the node semantics considered to be repeated and can be merged.

[0125]

[0126] In formula (5), i and j represent two different semantic nodes.

[0127] The node merging logic is as follows: Node content from authoritative sources with higher priority is retained, in the order of regulations > expert experience documents > structured ledgers; if the sources are the same, nodes with higher scores (extraction confidence or usage frequency) are retained; the list of nodes to be merged is recorded in the alias_list field of the master node, preserving the original representation for fuzzy search and visualization. Merged nodes are redirected from the chain structure to the master node's node_id, updating references in all chains.

[0128] The merged semantic node structure is shown in the following example:

[0129]

[0130] At the same time, the edge relationships in all chain structures will automatically map the old node_id to the new node_id, maintaining the consistency of the chain structure.

[0131] S304. Based on preset standardization rules, semantic standardization processing is performed on the semantic nodes after node merging to obtain multiple decision chains.

[0132] In this step, semantic standardization processes execute standardization rules such as: unit and numerical standardization (e.g., "0.2 megohms" is standardized to "0.2 MΩ"); action verb standardization (e.g., "replace" and "replace" are standardized to "replace"); synonym standardization replacement (refer to the domain thesaurus); node text format cleaning, removing redundant modifiers (e.g., "needs to do......" → directly extract the subject-verb structure).

[0133] Thus, nodes with similar expressions and the same semantics are merged into unified nodes; graph redundancy is reduced, structural compactness and knowledge clarity are improved; and the merged results are given unified identifiers and metadata to support semantic tracing.

[0134] S305. If there are chains with missing nodes among multiple decision chains, then perform similarity matching between the existing nodes in the chains with missing nodes and the typical decision chain template to obtain the corresponding matching nodes.

[0135] In this step, the typical decision chain template is a set of typical decision chain templates (such as "grounding fault handling chain" and "cable overheating handling chain") pre-built for the system; each template contains standard S / C / A / R node text styles and path structures; the templates are stored in a dual manner of graph structure + text vector.

[0136] S306. Complete the chain with missing nodes based on the matching nodes.

[0137] In this step, for chains with missing nodes, existing nodes in the current chain are used as query conditions to calculate the similarity with the template chain (using BM25 or vector recall). If the matching similarity is ≥0.6, the content of the missing type of node in the template is automatically introduced as completion. The completed node will be marked with "inferred" for easy manual review later.

[0138] Each completed chain structure is stored as a standard graph, as shown in the following example:

[0139]

[0140]

[0141] Each node includes: node ID (node_id); node type (S / C / A / R); extracted text (text); source field (e.g., regulations / ledger) and extraction method (rules / model); and completion flag (inferred, if it is a node generated by template inference).

[0142] As a large amount of text data is continuously input, new decision chains will be continuously built in the system. To avoid duplicate chains and achieve unique chain identification and knowledge aggregation management, this embodiment performs a "matching judgment" operation on each completed chain. The system will determine whether it belongs to one of the existing general decision chains. If not, a new chain number will be generated and included in the graph.

[0143] Optionally, it also includes: concatenating preset fields of each semantic node in the decision chain and performing hash calculation on the concatenation result to obtain the chain fingerprint value; if the chain fingerprint value of the target chain is different from the chain fingerprint value of the decision chain, then performing graph structure similarity matching between the target chain and the existing chain template, the target chain being the decision chain constructed after generating the decision chain graph; if it is determined based on the similarity matching result that the target chain does not match the existing chain template, then updating the target chain as a new decision chain in the decision chain graph.

[0144] In this step, the chain fingerprint generation mechanism is as follows: concatenate the text field of each node in each chain: fingerprint_text = S_text + C_text + A_text + R_text; apply the standard SHA-256 hash algorithm to the concatenation result to generate the chain fingerprint value as shown in the example below.

[0145]

[0146] If an existing chain has the same hash fingerprint, it is determined to be an "identical chain" and instance binding is performed; otherwise, the structure matching judgment stage is entered.

[0147] The subgraph structure matching algorithm (Weisfeiler-Lehman isomorphism judgment) uses the Weisfeiler-Lehman Subtree Kernel to compare the graph structure similarity between the current chain and existing chain templates; it uses node type (S / C / A / R) and node text as node labels; and determines whether the structure is isomorphic or a near variant. Similarity calculation is based on: node type sequence matching (S→C→A→R); average cosine similarity of node text ≥ 0.8; and overall structure label overlap ≥ 0.75 (i.e., at least three nodes match and the edge order is consistent). If the matching result shows a complete match (identical fingerprints), it belongs to an existing chain and is instance-bound; if the matching result shows a structural match (different fingerprints but isomorphic structure), it is determined to be an instance of the general chain and bound to the original chain_id; if the matching result shows a mismatch, a new chain number is generated, written to the general chain library, and the construction source is recorded.

[0148] When the current chain is determined to be a new chain, the system generates a new chain_id, with the following naming rules: CHAIN-<4-digit year><2-digit month><2-digit day>-<serial number>, for example: CHAIN-20250423-0031; records the source information, construction time, and construction method (rule / model / hybrid); and enters the chain library as a general chain template for subsequent instance binding and query reuse.

[0149] The output chain and its matching result in this embodiment are structurally illustrated as follows:

[0150]

[0151]

[0152] The fault database modeling method based on decision chain graph proposed in this application assembles dispersed decision units into a logically coherent, directed acyclic chain structure; ensures that each chain covers a complete processing semantic path (S→C→A→R); automatically completes and standardizes chains with missing structural information; and provides edge structure and upstream and downstream relationships for graph structured modeling.

[0153] In addition, this embodiment also provides a decision chain matching and response generation mechanism based on prompt words. It supports users to input prompt words in natural language, automatically match "general decision chains" in the graph that match their semantics, and generate semantically clear and structurally interpretable response information based on the matching results. This processing link integrates the semantic understanding capabilities of large language models with the stability of structured graphs, realizing a decision support mechanism that combines "generalizability" and "traceability". The processing flow is as follows.

[0154] The system is driven by a large model to perform semantic parsing of prompt words. The input is a language prompt word, and an LLM (such as Qwen or ChatGLM) is called to generate a set of semantic keywords to construct a structured query intent as shown below. The output is a prompt word semantic intent vector or a structured semantic query.

[0155]

[0156] The decision chain is driven by graph-based graph matching, generating chain query vectors (BERTembedding or Graph-BERT) using the aforementioned intent. Similarity retrieval (top-k) is performed in the chain graph vector library. For structural queries, subgraph pattern matching can also be used. The returned results are: a set of matched chains (with similarity scores); and matching method labels (vector, structure, keyword, or combination). The large model then participates in generating the response, generating a natural language explanation based on the matched chain structure and the user's original prompts by calling an LLM (with a specified template): structural chain → multi-step decision explanation; accompanied by suggested text, a visual graph, historical example examples, etc. An example output is shown below.

[0157]

[0158] Figure 4 This is a schematic diagram of a fault database modeling device based on a decision chain graph, provided as another embodiment of this application. Figure 4 As shown, the fault database modeling device based on decision chain graphs includes:

[0159] The acquisition module 401 is used to acquire structured ledger data and unstructured text data corresponding to the subway power supply system.

[0160] The first processing module 402 is used to perform data format cleaning and text segmentation on unstructured text data to obtain the text segments corresponding to the unstructured text data.

[0161] The second processing module 403 is used to perform sentence-level standardization processing on the text segments corresponding to the unstructured text data and the natural language fields in the structured ledger data to obtain the target sentence-level text.

[0162] The identification module 404 is used to identify key semantic unit information in the subway power supply fault handling process from the target sentence-level text. The key semantic unit information is used to describe the decision unit quadruple, which includes: state node, judgment condition node, decision action node and result node.

[0163] The third processing module 405 is used to perform directed chain combination and chain node merging and standardization processing on the decision unit quadruple to obtain multiple decision chains, and generate a decision chain graph based on the multiple decision chains, so as to model the subway power supply fault through the decision chain graph.

[0164] Optionally, the first processing module 402 is specifically used for:

[0165] The algorithm identifies target noise format items in unstructured text data and uses regular expressions to clean the format of these items. It then uses a title detection algorithm to perform title detection and format feature fusion on the cleaned unstructured text data. Finally, it segments the unstructured text data into paragraphs based on preset paragraph boundary division rules, resulting in the corresponding text segments.

[0166] Optionally, the second processing module 403 is specifically used for:

[0167] A Chinese language processing tool is used to segment unstructured text data and natural language fields in structured ledger data to obtain first-sentence-level text. A similarity-based filtering method is used to remove redundant and repetitive sentences from the first-sentence-level text, resulting in second-sentence-level text. The second-sentence-level text undergoes text format and encoding normalization to obtain third-sentence-level text. If no potentially incomplete sentences are found in the third-sentence-level text, it is identified as the target sentence-level text. If potentially incomplete sentences are found in the third-sentence-level text, short sentence completion and repair are performed to obtain the target sentence-level text. Furthermore, if a sentence in the third-sentence-level text has a sentence length shorter than a preset number of words, it is determined that a potentially incomplete sentence exists.

[0168] Optional, the recognition module 404 is specifically used for:

[0169] Each preset regular expression rule is executed sequentially on the target sentence-level text to obtain multiple entity fragments, and node annotations are performed on these entity fragments to obtain rule-extracted statements. Instructional model extraction is then performed on text in the target sentence-level text that meets preset conditions to obtain model-extracted statements. These preset conditions include: rule not hit, high sentence structure freedom, and the existence of nested structures. Four-tuple structure transformations are then performed on both the rule-extracted statements and the model-extracted statements to obtain key semantic unit information in the subway power supply fault handling process.

[0170] Optional, the recognition module 404 is specifically used for:

[0171] The sequence labeling model is used to extract the text that meets the preset conditions in the target sentence-level text using an instructional model to obtain the extracted sentences; and / or, the large language model is used to extract the text that meets the preset conditions in the target sentence-level text using an instructional model to obtain the extracted sentences.

[0172] Optionally, the third processing module 405 is specifically used for:

[0173] A custom chain structure is defined, consisting of state nodes, judgment condition nodes, decision action nodes, and result nodes in sequence. Based on this chain structure, the connection direction of each semantic node in the decision unit quadruple is determined according to the statement order and semantic modifiers in the corresponding statements. Each semantic node in the decision unit quadruple is vectorized into a node text, and multiple semantic nodes with semantic repetition relationships are identified through the node text vectors corresponding to each semantic node. These semantic nodes are then merged. Based on preset standardization rules, the merged semantic nodes are semantically standardized to obtain multiple decision chains.

[0174] Optional features also include: a matching module and a completion module.

[0175] The matching module is used to perform similarity matching between the existing nodes in the chain with missing nodes and the typical decision chain template if there are chains with missing nodes among multiple decision chains, and obtain the corresponding matching nodes.

[0176] The completion module is used to complete the nodes of chains with missing nodes based on the matching nodes.

[0177] Optional features include: a splicing module and an update module.

[0178] The splicing module is used to splice the preset fields of each semantic node in the decision chain and perform hash calculation on the splicing result to obtain the chain fingerprint value.

[0179] The matching module is also used to perform graph structure similarity matching between the target chain and the existing chain template if the chain fingerprint value of the target chain is different from that of the decision chain. The target chain is the decision chain constructed after generating the decision chain graph.

[0180] The update module is used to update the target chain as a new decision chain in the decision chain graph if the target chain is determined to be mismatched with an existing chain template based on the similarity matching result.

[0181] The above-described fault database modeling apparatus based on decision chain graphs can execute the methods provided in any embodiment of this application, and has the corresponding functional modules and beneficial effects for executing the methods. Technical details not described in detail in this embodiment can be found in the fault database modeling method based on decision chain graphs provided in any embodiment of this application.

[0182] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 5 A block diagram of an exemplary electronic device suitable for implementing embodiments of the present application is shown. Figure 5 The electronic device 12 shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0183] like Figure 5 As shown, the electronic device 12 is represented in the form of a general-purpose computing device. The components of the electronic device 12 may include, but are not limited to: one or more processors or processing units 16, system memory 28, and bus 18 connecting different system components (including system memory 28 and processing unit 16).

[0184] Bus 18 represents one or more of several bus architectures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus using any of the various bus architectures. For example, these architectures include, but are not limited to, the Industry Standard Architecture (ISA) bus, the Micro Channel Architecture (MAC) bus, the Enhanced ISA bus, the Video Electronics Standards Association (VESA) local bus, and the Peripheral Component Interconnect (PCI) bus.

[0185] Electronic device 12 typically includes a variety of computer system readable media. These media can be any available media that can be accessed by electronic device 12, including volatile and non-volatile media, removable and non-removable media.

[0186] System memory 28 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) 30 and / or cache memory 32. Electronic device 12 may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, storage system 34 may be used to read and write non-removable, non-volatile magnetic media (… Figure 5 Not shown; usually referred to as a "hard drive"). Although Figure 5 Not shown, a disk drive for reading and writing to a removable non-volatile disk (e.g., a "floppy disk") and an optical disk drive for reading and writing to a removable non-volatile optical disk (e.g., a CD-ROM, DVD-ROM, or other optical media) may be provided. In these cases, each drive may be connected to bus 18 via one or more data media interfaces. Memory 28 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of the embodiments of this application.

[0187] A program / utility 40 having a set (at least one) of program modules 42 may be stored, for example, in memory 28. Such program modules 42 include, but are not limited to, an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. Program modules 42 typically perform the functions and / or methods described in the embodiments of this application.

[0188] Electronic device 12 can also communicate with one or more external devices 14 (e.g., keyboard, pointing device, display 24, etc.), and with one or more devices that enable a user to interact with electronic device 12, and / or with any device that enables electronic device 12 to communicate with one or more other computing devices (e.g., network card, modem, etc.). This communication can be performed via input / output (I / O) interface 22. Furthermore, electronic device 12 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 20. As shown, network adapter 20 communicates with other modules of electronic device 12 via bus 18. It should be understood that, although... Figure 5 As not shown, other hardware and / or software modules may be used in conjunction with electronic device 12, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0189] The processing unit 16 executes various functional applications and data processing by running programs stored in the system memory 28, such as implementing the fault library modeling method based on decision chain graph provided in the embodiments of this application.

[0190] This application also provides a computer program product.

[0191] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer program products, which may include one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be an application-specific or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0192] Note that the above description is merely a preferred embodiment and the technical principles employed in this application. Those skilled in the art will understand that this application is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of this application. Therefore, although this application has been described in detail through the above embodiments, this application is not limited to the above embodiments. Many other equivalent embodiments may be included without departing from the concept of this application, and the scope of this application is determined by the scope of the appended claims.

Claims

1. A fault library modeling method based on a decision chain graph, characterized in that, The method comprises: obtaining structured account data and unstructured text data corresponding to a subway power supply system; performing data format cleaning and text paragraph segmentation on the unstructured text data to obtain text segmentation segments corresponding to the unstructured text data; performing sentence-level standardization processing on the text segmentation segments corresponding to the unstructured text data and natural language fields in the structured account data to obtain target sentence-level text; identifying key semantic unit information in a subway power supply fault handling process from the target sentence-level text, the key semantic unit information being used to describe a decision unit quadruple; wherein the decision unit quadruple comprises: a state node, used to represent abnormal performance of equipment or a system; a judgment condition node, used to represent conditions, environmental states or fault features relied on for fault judgment; a decision action node, used to represent disposal measures, recommended operations or maintenance behaviors taken for the fault; and a result node, used to represent a state, a recovery condition or an effect achieved by the system after the decision action is executed; based on inherent causal, judgment and disposal logical relationships between the state node, the judgment condition node, the decision action node and the result node in the fault handling process, performing directed chain combination and chain node merging and standardization processing on the decision unit quadruple to obtain a plurality of decision chains, and generating a decision chain graph based on the plurality of decision chains to model the subway power supply fault through the decision chain graph.

2. The method of claim 1, wherein, The data format cleaning and text paragraph segmentation on the unstructured text data to obtain text segmentation segments corresponding to the unstructured text data comprises: identifying target noise format items in the unstructured text data, and performing format cleaning on the target noise format items in the unstructured text data using a regular expression; performing title detection and format feature fusion on the unstructured text data obtained after format cleaning using a title detection algorithm; performing text paragraph segmentation on the unstructured text data obtained after title detection and format feature fusion based on a preset paragraph boundary division rule to obtain text segmentation segments corresponding to the unstructured text data.

3. The method of claim 1, wherein, The sentence-level standardization processing on the text segmentation segments corresponding to the unstructured text data and the natural language fields in the structured account data to obtain target sentence-level text comprises: performing sentence breaking processing on the text segmentation segments corresponding to the unstructured text data and the natural language fields in the structured account data using a Chinese language processing tool to obtain first sentence-level text; performing redundant and repetitive sentence filtering on the first sentence-level text using a similarity judgment rule to obtain second sentence-level text; performing text format and encoding normalization processing on the second sentence-level text to obtain third sentence-level text; if there is no suspected incomplete sentence in the third sentence-level text, determining the third sentence-level text as the target sentence-level text; if there is the suspected incomplete sentence in the third sentence-level text, performing short sentence completion and repair on the third sentence-level text to obtain the target sentence-level text; If the third sentence-level text contains a sentence with a sentence length less than a preset number of words, it is determined that the third sentence-level text contains the suspected incomplete sentence.

4. The method of claim 1, wherein, The key semantic unit information in the subway power failure handling process is identified from the target sentence-level text, including: The target sentence-level text is sequentially executed for each preset regular rule to obtain a plurality of entity segments, and the plurality of entity segments are node-labeled to obtain a rule extraction sentence; The target sentence-level text is sequentially executed for each preset regular rule to obtain a plurality of entity segments, and the plurality of entity segments are node-labeled to obtain a rule extraction sentence; The rule extraction sentence and the model extraction sentence are respectively converted into a four-tuple structure to obtain the key semantic unit information in the subway power failure handling process.

5. The method of claim 4, wherein, The target sentence-level text is sequentially executed for each preset regular rule to obtain a plurality of entity segments, and the plurality of entity segments are node-labeled to obtain a rule extraction sentence; The target sentence-level text is sequentially executed for each preset regular rule to obtain a plurality of entity segments, and the plurality of entity segments are node-labeled to obtain a rule extraction sentence; And / or, The target sentence-level text is sequentially executed for each preset regular rule to obtain a plurality of entity segments, and the plurality of entity segments are node-labeled to obtain a rule extraction sentence.

6. The method of claim 1, wherein, The decision unit four-tuple is combined into a directed chain and the chain nodes are merged and standardized to obtain a plurality of decision chains, including: The chain structure is self-defined, and the chain structure is sequentially composed according to the order of the state node, the judgment condition node, the decision action node and the result node; Based on the chain structure, the connection direction of each semantic node in the decision unit four-tuple is determined according to the order of the sentences and the semantic modifiers in the corresponding sentences; The node text of each semantic node in the decision unit four-tuple is vectorized, and the plurality of semantic nodes with semantic repetition relationship are determined by the node text vector corresponding to each semantic node, and the plurality of semantic nodes with semantic repetition relationship are node-merged; The semantic nodes after node merging are subjected to semantic standardization processing based on a preset standardization rule to obtain a plurality of decision chains.

7. The method of claim 6, wherein, Further comprising: If there is a chain with missing nodes in the plurality of decision chains, similarity matching is performed between the existing nodes in the chain with missing nodes and a typical decision chain template to obtain corresponding matching nodes; The chain with missing nodes is node-completed based on the matching nodes.

8. The method of claim 6, wherein, Further comprising: The preset fields of each semantic node in the decision chain are spliced, and the spliced result is subjected to hash calculation to obtain a chain fingerprint value; If the chain fingerprint value of the target chain is different from the chain fingerprint value of the decision chain, graph structure similarity matching is performed between the target chain and an existing chain template, the target chain being a decision chain constructed after generating the decision chain graph; If it is determined based on the similarity matching result that the target chain does not match the existing chain template, the target chain is updated as a new decision chain in the decision chain graph.

9. An electronic device, comprising: Comprising: One or more processors; Memory for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement a decision chain graph-based fault library modeling method as claimed in any one of claims 1 to 8.

10. A computer program product comprising a computer program, characterized in that, The computer program, when executed by a processor, implements a decision chain graph-based fault library modeling method as claimed in any one of claims 1 to 8.

Images