Work ticket device entity completion and work type determination method

Abstract

The application relates to the technical field of power safety, and discloses a work ticket equipment entity completion and operation type determination method, which comprises the following steps: step 1, based on an equipment knowledge graph, a synthetic sample library is reversely generated; step 2, a power professional large language model is constructed, and candidate equipment entity extraction is carried out to generate a candidate equipment entity set; step 3, topological verification is carried out on each candidate equipment entity extracted by the power professional large language model; step 4, for the equipment entity with missing numbers, equipment number probabilistic completion is carried out; and step 5, based on a four-level tree structure organization, a hierarchical determination rule library is established to carry out hierarchical operation type determination. According to the application, the completion model and the determination model can be efficiently cold-started and continuously optimized, so that the work ticket safety measure preparation has a stable, interpretable and generalizable equipment cognition foundation.

Description

Technical Field

[0001] This invention relates to the field of power safety technology, specifically to a method for completing work order equipment entities and determining work type. Background Technology

[0002] Work permits are the legal basis for substation maintenance personnel to enter the work site, and their preparation quality directly affects personnel safety, power grid safety, and equipment safety. With the continuous expansion of the power grid and the increasing diversity of equipment types, the traditional method of having maintenance personnel prepare work permits item by item based on experience is inefficient, prone to omissions, and reliant on individual experience. To improve the efficiency and standardization of work permit preparation, the industry has undertaken various automated and intelligent explorations.

[0003] In the process of compiling substation maintenance work orders, the physical information of the maintenance equipment (including equipment name, equipment number, voltage level, and bay to which it belongs) in the work tasks is derived from the maintenance plan or work content text described in natural language. This generally presents three engineering challenges:

[0004] (1) Missing equipment number and non-standard description: In actual production, in order to improve the efficiency of writing maintenance plans, the equipment number (such as "7531") is often omitted only by the equipment name (such as "line protection", "main transformer high voltage side circuit breaker"), or various colloquial and non-standard descriptions are used, which leads to the break of the subsequent safety measures deduction link based on the unique identity of the equipment; and there is no mechanism to complete the missing equipment number. When the existing plan encounters a work description that only gives the equipment name but not the number, it either directly abandons the subsequent reasoning or uses the most similar keyword to hard match, which has low accuracy and no confidence.

[0005] (2) Relying solely on large language models for identification carries the risk of illusion: When large language models are used to extract equipment entities from work texts, the extraction of entities from small stations, renovated and expanded stations, and new equipment that are not fully covered in the training data is prone to illusion problems such as fictitious (i.e., extracting equipment that does not exist in reality) and wrong extraction (extracting equipment of the wrong voltage level or wrong interval), and there is no explanatory or traceable basis; although existing SCD (Smart Substation Configuration File) rule matching schemes can be traced back to the rule number, they are not interpretable at the semantic level; although the simple large language model scheme is semantically close, it lacks the traceability link to specific knowledge graph nodes and historical samples. Moreover, the semantic parsing ability for non-standardized work descriptions is insufficient. Schemes that use SCD structured files or linear diagrams as input cannot understand natural language descriptions such as "line protection replacement" and "main transformer high-voltage side interval maintenance" in maintenance schemes; schemes based solely on large language models have serious entity illusions and lack reliable reverse lookup and verification mechanisms.

[0006] (3) Difficulty in determining the strong coupling between work type and electrical wiring method: Under different main wiring methods such as 3 / 2 wiring, double busbar, and single busbar segmentation, the safety measures corresponding to "side switch maintenance" and "middle switch maintenance" of the same equipment are significantly different. The work type is determined by "electrical wiring method + equipment position in wiring + work content intention". Traditional methods based on keyword matching or text classification are difficult to capture this structured semantics. Furthermore, the work type determination is separated from the electrical wiring method. Existing solutions either do not determine the work type or only perform flat text classification based on keywords, without utilizing the hierarchical structure information of "wiring method → ​​equipment topology position → work content intention". This makes it impossible to reliably distinguish whether the same "cleaning and maintenance" work description should be determined as "side switch bay operation", "middle switch bay operation" or "line bay disconnect switch operation" under 3 / 2 wiring.

[0007] (4) Difficulty in cold start of model: The historical work ticket data of power companies is limited in scale (only tens of thousands of tickets per year for a provincial company), and there are even fewer samples involving missing numbers and non-standard descriptions; existing solutions generally do not provide effective sample augmentation and cold start mechanisms.

[0008] Existing technologies are either limited to structured data in SCD files of smart substations, or to manual simulation by operators on a single diagram, or to risk optimization at the dispatching layer, or to the integration of general large language models and knowledge graphs without specific capabilities for power work ticket scenarios. To address these issues, we have invented a method for work ticket equipment entity completion and operation type determination, which solves the above technical problems. Summary of the Invention

[0009] This invention provides a method for completing equipment entities and determining job types on work orders. It solves three closely related problems: "identifying equipment entities from non-standardized job texts, probabilistically completing missing equipment numbers, and hierarchically determining job types that are strongly coupled with electrical wiring methods." Furthermore, it utilizes knowledge graphs to reverse synthesize training samples for cold-start optimization of the completion model.

[0010] The objective of this invention can be achieved through the following technical measures: a method for work order equipment entity completion and hierarchical determination of job type based on the same interval probability matrix and knowledge graph topology verification, the method comprising:

[0011] Step 1: Based on the device knowledge graph, generate a synthetic sample library in reverse engineering;

[0012] Step 2: Construct a large language model for the power industry and extract candidate equipment entities to generate a set of candidate equipment entities;

[0013] Step 3: Perform topology verification on each candidate device entity extracted from the power engineering big language model;

[0014] Step 4: For equipment entities with missing numbers, perform probabilistic number completion.

[0015] Step 5: Based on the four-level tree structure, establish a hierarchical judgment rule base to perform hierarchical judgment of job types.

[0016] The objective of this invention can also be achieved through the following technical measures:

[0017] In step 1, the equipment knowledge graph is sampled by interval and wiring method, and combined with the controlled language style perturbation template to generate a synthetic sample library. The synthetic sample library contains the text of the simulated maintenance plan and the standard answer of the corresponding simulated maintenance plan text, which is used for cold start and fine-tuning of the large language model and the numbering completion model. The equipment knowledge graph is constructed by substation and includes primary equipment entities, secondary equipment entities, electrical topology connection relationships, interval affiliation relationships, voltage level attributes, and the correspondence between protection devices and primary equipment.

[0018] Step 1 includes:

[0019] Step 11: Randomly sample an interval subgraph from the device knowledge graph;

[0020] Step 12: Select several devices within the interval sub-map as anchor devices and retain their numbers; use the remaining devices as target devices and clear their numbers.

[0021] Step 13: Select one or more templates from the preset language style perturbation template library;

[0022] Step 14: Generate simulated maintenance plan text from the large language model under the following hard constraints: (a) Only equipment within the interval subgraph can be used; (b) Target equipment can only be given by name and not by number; (c) Conforms to the selected language style perturbation template.

[0023] Step 15: Store the generated simulated maintenance plan text, along with the corresponding real number and real job type, as tags in the synthetic sample library and perform quality filtering.

[0024] Step 16: Mix the synthetic samples in the synthetic sample library with a small number of real samples at a ratio of 2:1–5:1 to obtain the training dataset for LoRA fine-tuning of the power industry large language model and number completion model.

[0025] Step 17: Periodically collect real failure samples for completion, as prompts for difficult samples in the next round of reverse generation, forming a continuous learning loop.

[0026] In step 2, based on the maintenance plan text, the work task description, and the synthetic sample library, the power industry big language model is constructed on the basis of the general big language model through power industry maintenance knowledge incremental pre-training and LoRA fine-tuning. This model is responsible for the preliminary semantic understanding of the natural language description of the work task and outputs a set of candidate equipment entities.

[0027] In step 3, for each candidate device entity extracted from the power professional big language model, topology verification is performed, and the device knowledge graph is queried to determine whether it actually exists in this substation; if it does not exist, it is rejected; if the voltage level or interval affiliation is inconsistent, the error correction and re-extraction process is entered; the anti-illusion bidirectional cross-validation of candidate device entities is completed.

[0028] In step 4, let the set of equipment entities extracted in the job description be... For some equipment entities, the ID field is empty. For equipment entities with missing IDs, the ID is probabilistically filled in. Candidates are selected and scored based on the same interval probability matrix and topological constraints. The completion result and confidence level are output.

[0029] Step 4 includes:

[0030] Step 41: Based on the statistical learning of the list of equipment with the same interval for the same job in the historical work order, obtain the conditional probability distribution and establish the same interval probability matrix. The same interval probability matrix gives the probability ranking of what value the missing equipment number should take in the context of known sibling equipment.

[0031] Step 42: Select the set of anchor devices with complete numbers and that pass the verification of the device knowledge graph from the extracted device entity set E. ; Query the equipment knowledge graph to obtain the interval to which these anchor devices belong. With wiring method M;

[0032] Step 43, for each device entity with a missing number From the device knowledge graph, ordered by interval ∈ I and device type = Filter the set of candidate numbers ;

[0033] Step 44, for the candidate number set For each candidate c, calculate the comprehensive score using the following formula. :

[0034] in, This represents the conditional probability of candidate c appearing given by the same interval probability matrix under the condition of known sibling devices and wiring methods; The cosine similarity between the semantic embedding of candidate c and the context embedding of the job description is denoted as c. As an indicator function, if candidate c and If the connection is achieved through the same interval relationship, the value is 1; otherwise, it is 0. α, β, and γ are weights that satisfy α + β + γ = 1. The value of α ranges from 0.4 to 0.7, the value of β ranges from 0.2 to 0.4, and the value of γ ranges from 0.05 to 0.2.

[0035] Step 45, for the candidate number set Based on overall score Sort in descending order and output the Top-K candidate numbers as the completion result and the confidence level of the completion result; when the normalized confidence level of the first candidate is lower than the threshold θ, it is marked as low confidence and manual confirmation is requested.

[0036] Step 46: Write back the completion results with confidence scores higher than the threshold to the device entity list. Each device in the device entity list is mapped to a unique node in the device knowledge graph and has a traceable link, serving as the basis for subsequent job type determination. The training of the same-interval probability matrix uses Laplace smoothing.

[0037] In step 5, a hierarchical decision rule base is established by organizing the data into a four-level tree structure: electrical wiring method → ​​equipment position in wiring → work content intent → job type. Each node corresponds to an interpretable decision rule. The parameters of the hierarchical decision rule base include: wiring method enumeration and leaf node job type granularity.

[0038] In step 5, when performing hierarchical determination of job type, for each completed piece of equipment, starting from the equipment's position in the electrical wiring, reasoning along the tree structure of the hierarchical determination rule base to the leaf node of the job type includes:

[0039] First-level wiring mode positioning: Based on the equipment knowledge graph, query the wiring mode M of the busbar corresponding to the voltage level of this substation, triggering different judgment branches;

[0040] Second-level equipment location analysis: Given the wiring method M, query the specific location of each device in the wiring method M;

[0041] The third level of work content intent classification: The power professional big language model classifies the intent of sub-tasks such as cleaning, maintenance, testing, replacement and line parameter testing, and outputs intent labels such as general maintenance, testing and replacement.

[0042] Fourth-level rule base matching: Using the device location and work content intent as the joint key, the job type is searched in the leaf nodes of the hierarchical judgment rule base; if the match is successful, the job type and the corresponding thought chain, i.e. the four-level branch that has been traversed, are output; if the match fails, the manual confirmation channel is entered and the description of the match failure is recorded in the abnormal sample pool.

[0043] The objective of this invention can also be achieved through the following technical measures: A computer-readable storage medium storing a computer program thereon, wherein the computer program, when run by a processor, executes the steps of the work order equipment entity completion and job type determination method.

[0044] The work order equipment entity completion and operation type determination method provided by this invention is supported by four major foundations: a large-scale power engineering language model, an equipment knowledge graph, a probability matrix with equal intervals, and a hierarchical rule base. After the large-scale language model completes the initial semantic parsing of the operation text, the results are hard-validated using topological constraints of the knowledge graph, soft-completed using the probability matrix with equal intervals, and structured reasoning of the operation type using the hierarchical rule base. Furthermore, a synthetic sample library generated based on the knowledge graph is used to achieve cold start and continuous fine-tuning of the two core models (completion model and operation type determination model). This invention achieves reliable identification of maintenance equipment entities, probabilistic completion of missing numbers, and dynamic determination of operation types strongly coupled with electrical wiring methods, even with only non-standardized operation description text. Generative data augmentation is used to efficiently cold start and continuously optimize the completion and determination models, providing a stable, interpretable, and scalable equipment cognitive foundation for work order safety measure preparation. Compared with existing technologies, this invention has the following technical advantages:

[0045] (1) Use the power professional big language model to complete the semantic parsing of the natural language description of the operation task and the extraction of equipment entities. And use the equipment knowledge graph topology verification to realize the anti-illusion bidirectional cross verification of the entity extraction results to ensure that each identified device exists in the real power grid and belongs to this substation.

[0046] (2) In the case of missing equipment numbers in the job description, a "same interval / same string probability matrix" obtained by statistical learning of historical work tickets is provided as a prior for completion. Combined with the confirmed sibling equipment in the job context, the candidate output recommended number is the one with the highest probability and that has passed the topology consistency check, and a confidence level is attached.

[0047] (3) Provides a hierarchical operation type determination framework of "electrical wiring method → ​​equipment location → work content intent → rule base matching", and gives differentiated determination paths for different wiring methods such as 3 / 2 wiring, double busbar, and single busbar segmentation, so that the operation type becomes an interpretable multi-level reasoning result rather than a single-level classification label;

[0048] (4) A method for generating synthetic training samples based on equipment knowledge graph is provided. The equipment topology subgraph is sampled in a controlled manner. The large language model generates job description text that conforms to engineering language habits but has missing numbers and expression variations under the constraints of structured equipment. The known correct numbers and job types are used as labels to perform cold start and continuous optimization of the completion model and job type determination model, thereby alleviating the problem of insufficient real historical samples. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0050] Figure 1 This is a flowchart of a method for completing work order equipment entities and determining work type in a specific embodiment of the present invention;

[0051] Figure 2 This is a flowchart illustrating the hierarchical determination of job types in a specific embodiment of the present invention;

[0052] Figure 3 This is a general block diagram of the work ticket equipment entity completion and operation type determination system in a specific embodiment of the present invention. Detailed Implementation

[0053] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, and / or combinations thereof.

[0054] like Figure 1 As shown, Figure 1 This is a flowchart of a method for completing equipment entities and determining the job type in a specific embodiment of the present invention. The method includes:

[0055] Step 1 addresses the issues of missing numbers and insufficient variation samples in historical work orders by generating a synthetic sample library based on the equipment knowledge graph. Sub-graph sampling is performed on the equipment knowledge graph according to bay and wiring method. Combined with controlled language style perturbation templates, "simulated maintenance plan text" and corresponding "standard answer" tags are generated for cold start and fine-tuning of the large language model and number completion model. The equipment knowledge graph is constructed on a substation basis, including primary equipment entities, secondary equipment entities, electrical topology connections, bay affiliation, voltage level attributes, and the correspondence between protection devices and primary equipment. It serves as the factual basis for subsequent topology verification and work type determination. Specifically, it includes:

[0056] Step 11: Randomly sample an interval subgraph from the device knowledge graph (containing all primary and secondary devices and their topological relationships within that interval);

[0057] Step 12: Select several devices within the sub-map as "anchor devices" (retain their numbers), and the remaining devices as "target devices" (clear their numbers).

[0058] Step 13: Select one or more templates from the pre-set language style perturbation template library (including several styles such as "standard handwriting", "abbreviated colloquial style", and "style containing typos and non-standard terminology");

[0059] Step 14: Generate simulated maintenance plan text from the large language model under the following hard constraints: (a) only equipment within the subgraph can be used; (b) target equipment can only be given by name and not by number; (c) conforms to the selected language style template;

[0060] Step 15: Store the generated text along with the corresponding real number and real job type as labels in the synthetic sample library, and then perform quality filtering (including removing contradictory samples, samples that are too close to historical samples, and excessively long samples).

[0061] Step 16: Mix synthetic samples with a small number of real samples at a ratio of 2:1–5:1, default 3:1; perform LoRA fine-tuning on the large language model and the numbering completion model; where LoRA rank r: 8–32, default 16; LoRA alpha: 16–64, default 32; learning rate: 1e-5 to 1e-4, default 5e-5; training epochs: 3–10 epochs, default 5. It should be noted that this step only provides the LoRA fine-tuning training configuration for the above two models based on the synthetic sample library; the construction and fine-tuning of the power engineering large language model are detailed in Step 2, and the use and fine-tuning of the numbering completion model are detailed in Step 4.

[0062] Step 17: Periodically collect real failed completion samples and feed them back into the synthetic sample library generator (i.e., the knowledge graph reverse generation module, which is different from the synthetic sample library itself) as "hard sample hints" for the next round of reverse generation, forming a continuous learning loop.

[0063] Step 2: Construct a large language model for the power industry and extract candidate equipment entities to generate a set of candidate equipment entities. Based on the maintenance plan text, work task description, and synthetic sample library, the large language model for the power industry is obtained through incremental pre-training of power industry maintenance knowledge and LoRA fine-tuning on the basis of the general large language model. This model is responsible for the preliminary semantic understanding of the natural language description of the work task and outputs a set of candidate equipment entities (including fields such as name, number, interval, voltage level, etc., with missing fields retained as null values ​​and accompanied by uncertainty markers).

[0064] Step 3: For each candidate device entity extracted from the power industry's large language model, perform topology verification to complete the anti-illusion bidirectional cross-validation of the candidate device entities. This involves querying the device knowledge graph to determine if it actually exists in the substation; if not, it is rejected (i.e., the candidate entity is judged as an illusion and removed, preventing it from entering the subsequent completion and judgment process, and recorded in the abnormal sample pool, rather than terminating the entire process); if the voltage level or interval affiliation is inconsistent, the process enters the error correction and re-extraction process. The error correction and re-extraction process returns to Step 2, where the power industry's large language model re-extracts the device entity under additional constraints (correct voltage level or interval hints).

[0065] Step 4: For equipment entities with missing numbers, perform probabilistic number completion. Candidates are selected and scored based on the same-interval probability matrix and topological constraints. The completion result and confidence level are then output. Let the set of equipment entities extracted from the job description be... The numbering field for some entities is empty. The numbering completion model includes a semantic embedding model for calculating the Sim embedding, which is obtained by cold-starting and LoRA fine-tuning the synthetic sample library described in step 1 based on the general embedding model; the same-interval probability matrix is ​​obtained by statistical learning from historical work tickets, and the two together constitute the basis for the probabilistic completion in step 4. Specifically, it includes:

[0066] Step 41: Establish the same-interval probability matrix, which is a unique statistical prior structure of this invention. It is obtained through statistical learning based on the "list of devices in the same interval for the same job" in historical work orders, and is a conditional probability distribution in the form of P(device B|device A known, wiring method = M). When a device number is missing in the job description, this matrix provides a probability ranking of what value the missing device number should take, given the known sibling devices.

[0067] Step 42, Known Equipment Anchoring: Select the set of "anchor devices" from E that have complete numbers and pass the equipment knowledge graph verification. ; Query the knowledge graph to obtain the interval to which these anchor devices belong. With wiring method M.

[0068] Step 43, Candidate Generation: For each entity with a missing number (The equipment name is known, such as "line protection"), from the equipment knowledge graph, sorted by "interval ∈ I and equipment type = "Filter the set of candidate numbers" ;

[0069] Step 44, Probability Score: (Correct) For each candidate c, calculate its comprehensive score using the following formula:

[0070] in, This represents the conditional probability of candidate c appearing, given the sibling devices and wiring methods, as provided by the same interval probability matrix. The cosine similarity between the semantic embedding of candidate c and the context embedding of the job description is denoted as c. As an indicator function, if candidate c and If the connection is through the same interval, it is 1; otherwise, it is 0. α, β, and γ are weights that satisfy α + β + γ = 1. The value of α ranges from 0.4 to 0.7 (default 0.55), the value of β ranges from 0.2 to 0.4 (default 0.30), and the value of γ ranges from 0.05 to 0.2 (default 0.15).

[0071] Step 45, Output and Confidence: For Sort by S(·) in descending order, output Top-K candidates (K ranges from 1 to 5, default value is 3) and their confidence scores; when the normalized confidence score of the first candidate is lower than the threshold θ (θ ranges from 0.5 to 0.8, default value is 0.65), mark it as "low confidence" and request manual confirmation.

[0072] Step 46, Result Write-back: Completed results with confidence levels above the threshold are written back to the device entity list as the basis for subsequent job type determination. The training of the same-interval probability matrix uses Laplace smoothing.

[0073] Where λ is the smoothing coefficient, ranging from 0.01 to 0.1, with a default value of 0.05; it is used to alleviate the zero-probability problem of low-frequency combinations. In the formula, Count(c,E) known (M) indicates a historical work order with wiring method M and a known set of sibling devices E. knownUnder the given conditions, the number of times candidate number c co-occurs with it; ΣxCount(x,E known M) represents the sum of the co-occurrence times of all candidate numbers x in the candidate space under the same conditions; |candidate space| represents the size of the candidate number set Cj, that is, the total number of candidate numbers.

[0074] Step 5: Based on the four-level tree structure, establish a hierarchical judgment rule base to perform hierarchical judgment of job types.

[0075] A hierarchical decision rule base is established using a four-level tree structure: "Electrical wiring method → ​​Equipment position in wiring → Work content intent → Job type." Each node corresponds to an interpretable decision rule. For example, under a 3 / 2 wiring configuration, "Side switch position + cleaning and maintenance" corresponds to "Side switch interval operation"; under a 3 / 2 wiring configuration, "Middle switch position + cleaning and maintenance" corresponds to "Middle switch interval operation." The parameters of the hierarchical decision rule base include:

[0076] Wiring methods include: 3 / 2 connection, double busbar, single busbar segmented connection, bridge connection, etc., which can be tailored according to the actual configuration of the substation.

[0077] Leaf node operation type granularity: defined by the combination of "interval type + work intention", with a typical scale of 30-150 leaf nodes in a single substation.

[0078] like Figure 2 As shown, when performing hierarchical determination of job types, for each completed piece of equipment, the process starts from its position in the electrical wiring and reasons along the rule base tree structure to the leaf node of the job type. Specifically, this includes:

[0079] Level 1 (Wiring Method Positioning): Based on the equipment knowledge graph, query the wiring method M of the busbar corresponding to the voltage level of this substation, triggering different judgment branches.

[0080] Level 2 (Equipment Location Analysis): Under a known wiring method M, query the specific location of each piece of equipment in that wiring (e.g., under a 3 / 2 wiring, distinguish between side switches, middle switches, line side, and disconnect switches; under a double busbar wiring, distinguish between I busbar, II busbar, and bus tie equipment; under a single busbar segmented wiring, query the segment to which the equipment belongs).

[0081] Level 3 (Work Content Intent Classification): The Large Language Model (LLM) classifies the intents of sub-tasks such as "cleaning", "maintenance", "testing", "replacement", and "line parameter testing", and outputs intent labels such as "general maintenance", "testing", and "replacement".

[0082] Level 4 (Rule Base Matching): Using "Equipment Location + Work Content Intent" as the joint key, the job type is searched in the leaf nodes of the hierarchical judgment rule base; if the match is successful, the job type and the corresponding thought chain (i.e. the four-level branches traversed) are output; if the match fails, the manual confirmation channel is entered and the description of the failed match is recorded in the abnormal sample pool.

[0083] like Figure 3 The diagram shown is the overall block diagram of the work order equipment entity completion and operation type determination system. The system includes an input layer, a knowledge and statistical prior layer, a generative data augmentation and fine-tuning layer, a semantic parsing and entity extraction layer, a core reasoning layer, and an output layer.

[0084] The input layer includes maintenance plan text, work task description, and substation identifier. The substation identifier is used to locate and select among multiple priors built on a substation-by-substation basis. Specifically, it loads the equipment knowledge graph, same-interval probability matrix, and hierarchical decision rule base corresponding to the substation, thereby ensuring that topology verification, probabilistic completion, and work type determination are all performed in the correct substation context.

[0085] The knowledge and statistics prior layer includes a device knowledge graph, a common-interval probability matrix, and a hierarchical decision rule base.

[0086] Among them, the equipment knowledge graph is constructed on a substation basis and includes primary equipment entities, secondary equipment entities, electrical topology connection relationships, bay affiliation relationships, voltage level attributes, and the correspondence between protection devices and primary equipment. It serves as the factual basis for subsequent topology verification and operation type determination.

[0087] Same-interval probability matrix: This is a unique statistical prior structure of this invention. It is obtained through statistical learning based on the "list of equipment in the same interval for the same job" in historical work orders, and is a conditional probability distribution in the form of P(Equipment B | Equipment A known, wiring method = M). When a device number is missing in the job description, this matrix gives the probability ranking of what value the missing device number should take, given the known sibling devices.

[0088] Hierarchical Judgment Rule Base: Organized in a four-level tree structure of "Electrical Wiring Method → ​​Equipment Position in Wiring → Work Content Intent → Work Type", with each node corresponding to an interpretable judgment rule. For example, under 3 / 2 wiring, "Side Switch Position + Cleaning and Maintenance" corresponds to "Side Switch Interval Work"; under 3 / 2 wiring, "Middle Switch Position + Cleaning and Maintenance" corresponds to "Middle Switch Interval Work".

[0089] The generative data augmentation and fine-tuning layer includes a knowledge graph reverse generation module, a synthetic sample library, and efficient LoRA parameter fine-tuning.

[0090] The knowledge graph reverse generation module samples the equipment knowledge graph by interval / connection method, and combines it with the controlled language style perturbation template to generate "simulated maintenance plan text" and corresponding "standard answer" tags, which are used for the cold start and fine-tuning of the above-mentioned large language model and number completion model.

[0091] The semantic parsing and entity extraction layer includes a large language model for the power industry and an entity extraction module, as well as entity extraction and candidate generation.

[0092] Among them, the power industry big language model and entity extraction module is obtained by pre-training with power industry maintenance knowledge increment and LoRA fine-tuning on the basis of the general big language model. It undertakes the preliminary semantic understanding of the natural language description of the operation task and outputs a set of candidate equipment entities (including fields such as name, number, interval, voltage level, etc., and retains null values ​​for missing fields and attaches uncertainty markers).

[0093] The core reasoning layer includes a topology verification and anti-illusion module, a device number probabilistic completion module, and a job type hierarchical determination module.

[0094] Topology verification and anti-illusion module: For each candidate device entity extracted from the large language model, query the device knowledge graph to determine whether it actually exists in this substation; if it does not exist, it is rejected; if the voltage level or interval affiliation is inconsistent, it enters the error correction and re-extraction process.

[0095] Equipment Number Probabilistic Completion Module: For equipment entities with missing numbers, the module follows the probabilistic completion process of this invention to screen candidates based on the same interval probability matrix and topological constraints, scores them, and outputs the completion result and confidence level.

[0096] Hierarchical determination module for job type: According to the hierarchical determination process of this invention, for each completed device, starting from its position in the electrical wiring, reasoning along the rule base tree structure to the leaf node of the job type.

[0097] The output layer includes a structured list of equipment entities, job types and confidence levels, and a visualization of the thought chain.

[0098] In summary, this invention proposes a bidirectional cross-validation mechanism combining a "power industry large language model + equipment knowledge graph topology verification." This mechanism organically combines semantic extraction from the large language model with factual constraints from the knowledge graph, effectively suppressing the illusion problem of domain-specific named entities in the large language model. Each identified device can be found in the real power grid and has a traceable link. This invention uses a "same interval / same string probability matrix" obtained from historical work orders as a statistical prior in scenarios where device numbers are missing. Furthermore, it designs a three-factor weighted scoring model S(c | e) based on "probability matrix × semantic embedding similarity × topology indicator function". jThis invention, combined with Laplace smoothing to handle low-frequency combinations of zero-probability problems, achieves reliable, interpretable, and quantifiable confidence completion in scenarios with missing numbers. It proposes a four-level hierarchical work type determination framework: "wiring method → ​​equipment location → work content intent → rule base matching," decomposing the originally flat text classification problem into multi-level reasoning with interpretable thought chains. Differentiated determination paths are provided for different wiring methods such as 3 / 2 wiring, double busbars, and single busbar segmentation, ensuring that the same work content description is accurately distinguished into different work types under different wiring methods. This invention employs a synthetic training sample enhancement method based on reverse generation of equipment knowledge graphs. Through controlled sampling of equipment topology subgraphs, target equipment number removal, and reverse generation of a large language model under language style perturbation template control, a large synthetic sample library with "missing numbers and expression variations" features is constructed. This library is mixed with a small number of real samples for LoRA fine-tuning, effectively alleviating the cold start difficulty caused by insufficient real samples in power work order scenarios. This invention proposes a three-module collaborative pipeline architecture of "topology verification anti-illusion → probabilistic completion → hierarchical judgment". The intermediate conclusions of each stage serve as interpretable inputs for the next stage. The final output includes not only the equipment list and job type, but also a visualization of the thought chain consistent with the thinking of human engineers, which facilitates quick verification by work order reviewers.

[0099] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0100] The present invention also provides a storage medium, which is a computer-readable storage medium storing a computer program thereon, the computer program being executable when it is run. Figure 1 The method shown may include some or all of the steps. The storage medium may include read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, etc. The storage medium may also include non-volatile memory or non-transitory memory, etc.

[0101] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A work ticket device entity completion and work type determination method, characterized by, The method includes: Step 1: Based on the device knowledge graph, generate a synthetic sample library in reverse engineering; Step 2: Construct a large language model for the power industry and extract candidate equipment entities to generate a set of candidate equipment entities; Step 3: Perform topology verification on each candidate device entity extracted from the power engineering big language model; Step 4: For equipment entities with missing numbers, perform probabilistic number completion. Step 5: Based on the four-level tree structure, establish a hierarchical judgment rule base to perform hierarchical judgment of job types.

2. The method for completing work order equipment entities and determining work type according to claim 1, characterized in that, In step 1, the equipment knowledge graph is sampled by interval and wiring method, and combined with the controlled language style perturbation template to generate a synthetic sample library. The synthetic sample library contains the text of the simulated maintenance plan and the standard answer of the corresponding simulated maintenance plan text, which is used for cold start and fine-tuning of the large language model and the numbering completion model. The equipment knowledge graph is constructed by substation and includes primary equipment entities, secondary equipment entities, electrical topology connection relationships, interval affiliation relationships, voltage level attributes, and the correspondence between protection devices and primary equipment.

3. The method for completing work order equipment entities and determining work type according to claim 2, characterized in that, Step 1 includes: Step 11: Randomly sample an interval subgraph from the device knowledge graph; Step 12: Select several devices within the interval sub-map as anchor devices and retain their numbers; use the remaining devices as target devices and clear their numbers. Step 13: Select one or more templates from the preset language style perturbation template library; Step 14: Generate simulated maintenance plan text from the large language model under the following hard constraints: (a) Only equipment within the interval subgraph can be used; (b) Target equipment can only be given by name and not by number; (c) Conforms to the selected language style perturbation template. Step 15: Store the generated simulated maintenance plan text, along with the corresponding real number and real job type, as tags in the synthetic sample library and perform quality filtering. Step 16: Mix the synthetic samples in the synthetic sample library with a small number of real samples at a ratio of 2:1–5:1 to obtain the training dataset for LoRA fine-tuning of the power industry large language model and number completion model. Step 17: Periodically collect real failure samples for completion, as prompts for difficult samples in the next round of reverse generation, forming a continuous learning loop.

4. The method for completing work order equipment entities and determining work type according to claim 1, characterized in that, In step 2, based on the maintenance plan text, the work task description, and the synthetic sample library, the power industry big language model is constructed on the basis of the general big language model through power industry maintenance knowledge incremental pre-training and LoRA fine-tuning. This model is responsible for the preliminary semantic understanding of the natural language description of the work task and outputs a set of candidate equipment entities.

5. The method for completing work order equipment entities and determining work type according to claim 1, characterized in that, In step 3, for each candidate device entity extracted from the power professional big language model, topology verification is performed, and the device knowledge graph is queried to determine whether it actually exists in this substation; if it does not exist, it is rejected; if the voltage level or interval affiliation is inconsistent, the error correction and re-extraction process is entered; the anti-illusion bidirectional cross-validation of candidate device entities is completed.

6. The method for completing work order equipment entities and determining work type according to claim 1, characterized in that, In step 4, let the set of equipment entities extracted in the job description be... For some equipment entities, the ID field is empty. For equipment entities with missing IDs, the ID is probabilistically filled in. Candidates are selected and scored based on the same interval probability matrix and topological constraints. The completion result and confidence level are output.

7. The method for completing work order equipment entities and determining work type according to claim 6, characterized in that, Step 4 includes: Step 41: Based on the statistical learning of the list of equipment with the same interval for the same job in the historical work order, obtain the conditional probability distribution and establish the same interval probability matrix. The same interval probability matrix gives the probability ranking of what value the missing equipment number should take in the context of known sibling equipment. Step 42: Select the set of anchor devices with complete numbers and that pass the verification of the device knowledge graph from the extracted device entity set E. ; Query the equipment knowledge graph to obtain the interval to which these anchor devices belong. With wiring method M; Step 43, for each device entity with a missing number From the device knowledge graph, ordered by interval ∈ I and device type = Filter the set of candidate numbers ; Step 44, for the candidate number set For each candidate c, calculate the comprehensive score using the following formula. : in, This represents the conditional probability of candidate c appearing given by the same interval probability matrix under the condition of known sibling devices and wiring methods; The cosine similarity between the semantic embedding of candidate c and the context embedding of the job description is denoted as c. As an indicator function, if candidate c and If the connection is achieved through the same interval relationship, the value is 1; otherwise, it is 0. α, β, and γ are weights that satisfy α + β + γ = 1. The value of α ranges from 0.4 to 0.7, the value of β ranges from 0.2 to 0.4, and the value of γ ranges from 0.05 to 0.

2. Step 45, for the candidate number set Based on overall score Sort in descending order and output the Top-K candidate numbers as the completion result and the confidence level of the completion result; when the normalized confidence level of the first candidate is lower than the threshold θ, it is marked as low confidence and manual confirmation is requested. Step 46: Write back the completion results with confidence scores higher than the threshold to the device entity list. Each device in the device entity list is mapped to a unique node in the device knowledge graph and has a traceable link, serving as the basis for subsequent job type determination. The training of the same-interval probability matrix uses Laplace smoothing.

8. The method for completing work order equipment entities and determining work type according to claim 1, characterized in that, In step 5, the hierarchical judgment rule base is established by organizing the four-level tree structure of electrical wiring method → ​​equipment position in wiring → work content intention → job type, with each node corresponding to an interpretable judgment rule; The parameters of the hierarchical decision rule base include: wiring method enumeration and leaf node job type granularity.

9. The method for completing work order equipment entities and determining work type according to claim 8, characterized in that, In step 5, when performing hierarchical determination of job type, for each completed piece of equipment, starting from the equipment's position in the electrical wiring, reasoning along the tree structure of the hierarchical determination rule base to the leaf node of the job type includes: First-level wiring mode positioning: Based on the equipment knowledge graph, query the wiring mode M of the busbar corresponding to the voltage level of this substation, triggering different judgment branches; Second-level equipment location analysis: Given the wiring method M, query the specific location of each device in the wiring method M; The third level of work content intent classification: The power professional big language model classifies the intent of sub-tasks such as cleaning, maintenance, testing, replacement and line parameter testing, and outputs intent labels such as general maintenance, testing and replacement. Fourth-level rule base matching: Using the device location and work content intent as the joint key, the job type is searched in the leaf nodes of the hierarchical judgment rule base; if the match is successful, the job type and the corresponding thought chain, i.e. the four-level branch that has been traversed, are output; if the match fails, the manual confirmation channel is entered and the description of the match failure is recorded in the abnormal sample pool.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it executes the steps of the work ticket equipment entity completion and job type determination method as described in any one of claims 1 to 9.

Images