A power grid intrinsic safety hidden danger root cause analysis method, system, device and medium

Abstract

The application discloses a kind of power grid essential safety hidden danger root cause analysis method, system, equipment and medium, method includes: obtaining skill division dimension data, constructs essential safety skill library;Get power grid fault phenomenon description, according to the metadata of each skill package in the essential safety skill library, the intention recognition of power grid fault phenomenon description is carried out, and the matching score of each skill package is calculated, according to the matching score, skill dynamic loading is carried out;Structured evidence is extracted from the power grid fault phenomenon description, according to the structured evidence, search matching is carried out in the preset essential safety standard database, obtain candidate root cause set and corresponding matching score, according to the matching score, the preliminary root cause conclusion is obtained from the candidate root cause set;The credibility of the preliminary root cause conclusion is evaluated using probability reasoning verification method, obtain standardization power grid hidden danger root cause analysis result, with significant application value and popularization prospect.

Description

Technical Field

[0001] This invention relates to the fields of power system optimization scheduling, distributed microgrids, distribution networks containing distributed multi-energy systems, and distribution-microgrid coordinated operation control technology, and in particular to a method, system, equipment, and medium for root cause analysis of inherent safety hazards in power grids. Background Technology

[0002] Power grid fault diagnosis technologies based on rule bases or expert systems, defect cause retrieval technologies based on knowledge graphs or retrieval enhancement, and intelligent agent technologies that use general large language models for operation and maintenance question answering or fault-assisted analysis are all available. These technologies can provide risk warnings based on equipment type, fault phenomena, historical cases, or standard procedures. However, they typically treat rule bases, retrieval materials, and model prompts as a whole context, lacking a mechanism to dynamically select analysis capabilities based on equipment objects and fault tasks.

[0003] Existing rule bases and expert systems rely on fixed rules. When faced with synonyms, combinations of multiple anomalies, or missing information in natural language descriptions, manual translation into rule conditions is often required. While existing retrieval enhancement generation schemes can retrieve procedures and cases, they typically inject a large number of candidate texts into the model at once, leading to contextual redundancy and mixing of similar causes, affecting the model's attention allocation to key evidence. Existing general-purpose large models or single-agent-assisted analysis schemes lack restricted-domain output constraints, which may generate root cause names outside the standard defect database, making it difficult to establish a definite mapping between conclusions and risk points, defect codes, and remedial measures. At the same time, their output results are mostly based on natural language interpretations, lacking posterior confidence calculations based on prior probabilities, likelihood probabilities, and observational evidence, and failing to provide verifiable quantitative confidence evidence.

[0004] Therefore, a root cause analysis solution is needed for scenarios with inherent safety hazards in power grids. At the front end, user intent is identified through lightweight metadata, and only intelligent agent skills related to the current fault phenomenon are loaded. At the back end, root cause reasoning is limited to standardized risk points and root cause databases. Before output, Bayesian verification is used to perform probability verification on candidate root causes, thereby balancing analysis efficiency, standardization, and credibility. Summary of the Invention

[0005] In view of the above-mentioned existing problems, the present invention provides a method, system, equipment and medium for root cause analysis of inherent safety hazards in power grids, which solves the problems of non-compliant output, knowledge redundancy and inefficiency, and lack of credible quantitative basis for analysis results in existing technical solutions.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for root cause analysis of inherent safety hazards in power grids, including: Acquire data on skill classification dimensions to build an intrinsic safety skill library; Obtain a description of a power grid fault phenomenon, perform intent recognition on the description of the power grid fault phenomenon based on the metadata of each skill package in the intrinsic safety skill library, calculate the matching score for each skill package, and dynamically load skills based on the matching score; Structured evidence is extracted from the description of the power grid fault phenomenon. Based on the structured evidence, a search and matching process is performed in a preset intrinsic safety standard database to obtain a set of candidate root causes and corresponding matching scores. Based on the matching scores, preliminary root cause conclusions are obtained by filtering from the set of candidate root causes. The credibility of the preliminary root cause conclusions was assessed using probabilistic reasoning verification methods, resulting in standardized root cause analysis results for potential power grid hazards.

[0007] As a preferred embodiment of the root cause analysis method for inherent safety hazards in power grids described in this invention, the construction of an inherent safety skill library includes: Based on the skill classification dimension data, the power grid intrinsic safety analysis capability unit is decoupled into multiple independently packaged skill packages; In response to the fact that different power grid intrinsic safety analysis capability units have differences in at least one of the skill configuration attributes, the different power grid intrinsic safety analysis capability units are divided into different skill packages; In response to multiple scenarios sharing the same device object and having overlapping core elements, multiple scenarios are treated as sub-tasks under the same skill package. Configure at least three hierarchical storage structures for each skill pack to obtain an intrinsically safe skill library.

[0008] As a preferred embodiment of the root cause analysis method for inherent safety hazards in power grids described in this invention, the following steps are included: performing intent recognition on the description of power grid fault phenomena and calculating the matching score for each skill package: Read the metadata layer in each skill pack, perform natural language parsing on the description of power grid fault phenomena, and extract key evidence information from the description of fault phenomena. Based on key evidence information and the preset trigger conditions in the metadata of each skill pack, a multi-dimensional match is performed, and a weighted calculation is performed based on the weight of each dimension to obtain the matching score of each skill pack. If the exclusion condition is triggered, the matching score of the corresponding skill will be set to zero or reduced to the range where it is not loaded.

[0009] The beneficial effects of this preferred technical solution are that it can quickly and accurately match corresponding skills, filter out irrelevant skills, and reduce invalid loading calculations.

[0010] As a preferred embodiment of the root cause analysis method for inherent safety hazards in power grids described in this invention, the dynamic skill loading based on the matching score includes: Load a single skill based on the matching score; In response to a matching score greater than or equal to the first threshold, the instruction layer content of the corresponding skill is loaded into the current dialogue context; If the matching score is less than the first threshold but greater than or equal to the second threshold, the instruction layer is not loaded, and the user is asked for supplementary information based on the missing slot. If the matching score is less than the second threshold, the instruction layer is not loaded, and the user is prompted to provide a more detailed description of the fault. Multiple skills are loaded based on their matching scores. In response to the fact that the matching scores of multiple skills are all greater than or equal to the first threshold, and the difference between the highest score and the second highest score is less than the third threshold, the highest-scoring skill is set as the primary skill, and the remaining skills, which do not exceed a preset number, are set as auxiliary skills. The corresponding instruction layers are then called for analysis. If the matching scores of multiple skills are all greater than or equal to the first threshold, and the difference between the highest score and the second highest score is greater than or equal to the third threshold, only the highest-scoring skill will be loaded. If the matching scores for multiple skills are all less than the first threshold, no skill instruction layer is loaded.

[0011] The beneficial effects of this preferred technical solution are that it loads skills on demand and in a hierarchical manner, reduces context usage, and improves analysis response speed.

[0012] As a preferred embodiment of the root cause analysis method for inherent safety hazards in power grids according to the present invention, the following steps are performed: Based on the structured evidence, a search and matching process is conducted in a preset inherent safety standard database to obtain a candidate root cause set and corresponding matching scores, including: Transform structured evidence into evidence feature vectors; Search and match in the pre-set intrinsic safety standard database. For each root cause entry in the intrinsic safety standard database, calculate the semantic similarity between the evidence feature vector and the standard feature vector of the root cause entry, and calculate the root cause matching evaluation dimension between the structured evidence and the root cause entry. Based on the pre-set weights of the root cause matching evaluation dimension, aggregate to obtain the matching score of each root cause entry. In response to root cause entries with matching scores greater than or equal to the fourth threshold, the corresponding root cause entries are taken as candidate root causes and arranged in descending order of matching scores to form a set of candidate root causes. If no root cause entry with a matching score reaching the fourth threshold is found, a prompt message is output.

[0013] The beneficial effect of this preferred technical solution is that it filters root causes through multi-dimensional weighted matching, constrains the output range, and eliminates model-generated illusions.

[0014] As a preferred embodiment of the root cause analysis method for inherent safety hazards in power grids according to the present invention, the preliminary root cause conclusions obtained by screening from the candidate root cause set based on the matching score include: If the matching score of the highest-scoring candidate root cause is greater than or equal to the fifth threshold, and the difference between the matching scores of the highest-scoring candidate root cause and the second-highest-scoring candidate root cause is greater than or equal to the sixth threshold, the highest-scoring candidate root cause is selected as the preliminary root cause conclusion. In response to the highest-scoring candidate root cause having a matching score less than the fifth threshold, or the difference between the matching scores of the highest-scoring candidate root cause and the second-highest-scoring candidate root cause being less than the sixth threshold, at least two candidate root causes are retained and output to the credibility assessment, or a follow-up question prompt indicating missing evidence is output to the user.

[0015] As a preferred embodiment of the root cause analysis method for inherent safety hazards in power grids according to the present invention, the credibility assessment of the preliminary root cause conclusions includes: Obtain the target root causes and structured evidence corresponding to the preliminary root cause conclusions; The frequency of the target root cause in historical failure data is used as the prior probability. The conditional probability of structured evidence appearing under the condition that the target root cause is true is taken as the likelihood probability. Based on the prior probability and the likelihood probability, calculate the posterior probability that the target root cause is valid given the presence of structured evidence. The confidence level of the preliminary root cause conclusion is determined based on the comparison between the posterior probability and the preset confidence threshold.

[0016] Secondly, the present invention provides a system for analyzing the root causes of inherent safety hazards in power grids, comprising: The skill library construction module is used to acquire skill classification dimension data and build an intrinsically safe skill library; The identification and calculation module is used to obtain a description of the power grid fault phenomenon, perform intent identification on the description of the power grid fault phenomenon based on the metadata of each skill package in the intrinsic safety skill library, calculate the matching score of each skill package, and dynamically load skills based on the matching score. The matching and filtering module is used to extract structured evidence from the description of the power grid fault phenomenon, perform retrieval and matching in a preset intrinsic safety standard database based on the structured evidence to obtain a set of candidate root causes and corresponding matching scores, and filter from the set of candidate root causes based on the matching scores to obtain preliminary root cause conclusions. The credibility assessment module is used to assess the credibility of the preliminary root cause conclusions using probabilistic reasoning verification methods, and obtain standardized power grid hidden danger root cause analysis results.

[0017] Thirdly, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of the method for analyzing the root causes of inherent safety hazards in power grids.

[0018] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the method for analyzing the root causes of inherent safety hazards in a power grid.

[0019] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention innovatively restricts the scope of the agent's root cause reasoning to a pre-built standardized database, enabling the agent's analysis results to possess both natural language understanding capabilities and adherence to industry standards. The agent only maps its output from the standard library, eliminating compliance risks caused by arbitrary generation and ensuring that the results comply with power grid safety operation specifications. The dynamic loading technology based on agent skills avoids the inefficient traditional mode of loading all knowledge at once. The system only loads relevant skill instructions when a fault occurs, greatly reducing context window occupation and lexical consumption, improving reasoning speed and accuracy, while reducing the computational pressure on large models. This invention introduces an independent Bayesian credibility analysis layer, quantifying the agent's reasoning results into probabilistic indicators. By calculating the posterior probability and comparing it with a threshold, the system provides an interpretable credibility metric for each root cause judgment, helping maintenance personnel intuitively judge the reliability of the agent's recommendations. This dual insurance mechanism of agent reasoning + mathematical verification meets the stringent requirements of the power system for high reliability. This invention combines agent skills with an intrinsically safe knowledge base, dynamically loading skills through a progressive disclosure strategy to achieve controllability and efficiency in agent analysis. Simultaneously, it introduces a Bayesian verification mechanism to provide quantitative credibility checks for the output results, mathematically guaranteeing the reliability and compliance of the analysis results. This method fills a technological gap in power grid fault analysis and has significant application value and promising prospects for wider application. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the overall process logic of a root cause analysis method for inherent safety hazards in power grids, provided in one embodiment of the present invention. Figure 2A flowchart of the root cause retrieval and screening judgment process for a root cause analysis method for inherent safety hazards in power grids, provided as an embodiment of the present invention; Figure 3 A flowchart of Bayesian credibility verification for a root cause analysis method for inherent safety hazards in power grids, provided as an embodiment of the present invention. Detailed Implementation

[0022] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0023] Example 1, referring to Figures 1-3 As an embodiment of the present invention, a method for root cause analysis of inherent safety hazards in power grids is provided, comprising: S100: Acquire skill classification dimension data and build an intrinsic safety skill library; S200: Obtain a description of the power grid fault phenomenon, perform intent recognition on the description of the power grid fault phenomenon based on the metadata of each skill package in the intrinsic safety skill library, calculate the matching score of each skill package, and dynamically load skills based on the matching score; S300: Extract structured evidence from the description of power grid fault phenomena, perform retrieval and matching in the preset intrinsic safety standard database based on the structured evidence, obtain a set of candidate root causes and corresponding matching scores, and filter from the set of candidate root causes based on the matching scores to obtain preliminary root cause conclusions; S400: The credibility of the preliminary root cause conclusions is assessed using probabilistic reasoning verification methods to obtain standardized root cause analysis results for potential power grid hazards.

[0024] In one optional embodiment, the probabilistic reasoning verification method can be the Monte Carlo probabilistic simulation verification method, which extracts the prior probability of candidate root causes and the likelihood probability distribution of various types of fault evidence; randomly generates multiple sets of simulation observation evidence samples based on historical fault samples; simulates fault reasoning for each set of samples and counts the simulation frequency of successful matching of the root cause; uses the ratio of the success frequency to the total number of simulations as the credibility and judges the credibility by comparing it with the threshold. In another optional embodiment, the probabilistic reasoning verification method can also be the Hidden Markov Model Probabilistic Sequence Verification Method, which constructs an observation sequence of fault phenomena in a time sequence, with the root cause as the hidden state; imports the equipment fault transfer probability and evidence emission probability to complete the model initialization; calculates the state probability of each root cause state under the current observation sequence through a forward algorithm; takes the maximum state probability of the target root cause as the credibility, and classifies the reliability of the conclusion. In this embodiment of the invention, the probabilistic reasoning verification method includes Bayesian probabilistic reasoning verification; The preliminary root causes obtained from the screening are set as hypothetical events, and all fault phenomena reported by users are extracted as the set of observational evidence. The prior probability of the root cause is obtained from historical fault statistics, and the likelihood probability is obtained from expert experience tables or historical case statistics. The root cause is normalized and solved by substituting into Bayes' formula to obtain the posterior probability of the root cause under the existing fault evidence. The results are compared with the preset confidence threshold to distinguish between three levels: high confidence, manual review, and unreliable. If the result is unreliable, the candidate root causes are re-screened and the root cause conclusion with quantified confidence is finally output.

[0025] It should be noted that this invention reduces contextual redundancy and improves reasoning efficiency by dynamically loading a hierarchical skill library; it eliminates the illusion of large models by relying on the inherent safety standard database to search and constrain the output range, ensuring the compliance and consistency of analysis; and it completes the credible quantitative assessment by using probabilistic reasoning to output credible standardized root causes, taking into account analysis efficiency, standardization, and result verifiability.

[0026] In this embodiment of the invention, step S100 includes the following sub-steps A1-A4; In A1: Based on the skill classification dimension data, the power grid intrinsic safety analysis capability unit is decoupled into multiple independently packaged skill packages; In A2: In response to the fact that different power grid intrinsic safety analysis capability units have differences in at least one of the skill configuration attributes, the different power grid intrinsic safety analysis capability units are divided into different skill packages; In A3: In response to multiple scenarios sharing the same device object and having overlapping core elements, multiple scenarios are treated as sub-tasks under the same skill package; In A4: Configure at least three levels of hierarchical storage structure for each skill pack to obtain the intrinsically safe skill library.

[0027] In one optional embodiment, the skill segmentation dimension data can be a trigger dimension. Based on the equipment objects and potential hazards or fault scenarios in the skill segmentation dimension data, the trigger conditions of the skill package are configured. The trigger conditions include the applicable equipment category and the fault phenomenon type, which are used to match the fault phenomenon description input by the user during the intent recognition stage to determine whether to trigger the skill package. In another optional embodiment, the skill segmentation dimension data can also be a capability dimension. Based on the evidence type and disposal knowledge boundary in the skill segmentation dimension data, the capability boundary of the skill package is configured. The capability boundary includes the allowed types of evidence types and the range of standard libraries that can be called, which are used to limit the analysis methods and output specifications of the skill package during the execution phase. In this embodiment of the invention, the skill classification dimension data includes four categories of conditions: equipment object, hidden danger or fault scenario, evidence type, and handling knowledge boundary. The core elements include skill-based data beyond the equipment itself, namely, potential hazards or malfunction scenarios, evidence types, and knowledge boundaries for handling them.

[0028] Specifically, the equipment object is used to limit the scope of skill application, such as main transformers, transmission lines, switch cabinets, busbars, relay protection devices, etc. Hazardous or fault scenarios are used to limit the analysis task, such as abnormal sound, abnormal temperature rise, abnormal gas analysis, abnormal protection action, etc. Evidence type is used to limit the input data, such as natural language descriptions, operational measurements, defect records, test data, or historical cases; The knowledge boundary is used to limit the standard library, risk point library, and root cause coding system that the skill can call.

[0029] Two capability units are classified into different skill packages when they meet any of the following conditions: they are applicable to different equipment categories or voltage levels; their sets of observable evidence are significantly different; they use different standard libraries, case libraries, or computational models; they output different risk points, defect categories, or root cause coding systems; or their triggering conditions are clearly mutually exclusive.

[0030] If multiple scenarios share the same device object and the evidence, standard library, and output encoding system highly overlap, they are treated as sub-tasks under the same skill package.

[0031] Each skill pack is configured with at least a three-tiered hierarchical storage structure, including a metadata layer that records skill identifier, skill name, equipment category, applicable voltage level, risk or failure scenario, trigger keywords, thesaurus, trigger condition description, trigger condition semantic vector, required parameter slots, exclusion conditions, and priority. This layer resides in memory, uses minimal context resources, and is used for initial screening and scoring of user intent; it does not contain complete analysis prompts or large amounts of domain information.

[0032] The Instructions layer stores detailed analysis process templates, field extraction rules, candidate root cause screening rules, Bayesian validation call rules, output format constraints, and exception handling strategies for the corresponding skill. Examples include "extract device type from user description," "extract fault phenomenon features," "call standardized defect library interface for matching," and "submit candidate root causes to credibility verification module." This layer only enters the current dialogue context when the skill is triggered and the loading threshold is reached.

[0033] The resource layer includes a standardized database of power grid equipment defects, a database of typical violations, a database of risk points, a retrieval module, a Bayesian verification module, and a historical case index. Resource layer content is accessed on demand and does not directly enter the dialogue context.

[0034] For example, when constructing a skill package for transformer fault analysis, the metadata layer sets the skill name and trigger keywords such as transformer fault and transformer anomaly; the instruction layer contains analysis step prompts, such as extracting transformer equipment and the symptom keyword "humming sound" and "oil temperature rise"; and the resource layer links to the power grid equipment defect database and accident case database. Through this layered storage, the system decouples skills from the dialogue context, reduces token overhead, and avoids interference from irrelevant information.

[0035] Skill packages can be implemented using files, database records, or object storage, and are not limited to specific directories or code. Taking transformer fault analysis skills as an example, a skill package should include at least four types of content: metadata records, instruction layer records, resource index records, and verification parameter records.

[0036] Metadata records include skill ID, skill name, device object, applicable scenarios, trigger keywords, synonyms, trigger condition semantic vectors, required slots, exclusion conditions, and version number. For example, the device object for the main transformer fault analysis skill is the main transformer, applicable scenarios include abnormal sound, abnormal oil temperature, abnormal gas analysis, abnormal current or voltage, etc., and exclusion conditions can include descriptions that explicitly point to transmission lines or distribution equipment.

[0037] The instruction layer records include the evidence extraction order, candidate root cause retrieval order, restricted domain output constraints, Bayesian verification call conditions, and output format. Restricted domain output constraints stipulate that all risk points and root causes must be mapped to risk point numbers and root cause numbers already registered in the standard library, and the skill-executing agent must not generate new root cause names outside the standard library.

[0038] The resource index records include a standardized database of power grid equipment defects, a database of typical violations, a database of risk points, a historical case index, a feature thesaurus, a root cause prior probability table, and a likelihood probability table. The resource layer is invoked as needed by the retrieval and validation modules; the resource content is not integrated into the language model context as a whole.

[0039] It should be noted that by splitting and constructing a layered, three-structure skill package according to four dimensions, loading the instruction layer on demand, and isolating the resource layer, the context overhead is significantly reduced, irrelevant information interference is avoided, and the efficiency of fault matching and analysis is improved. In this embodiment of the invention, step S200 includes the following sub-steps B1-B3; In B1: Read the metadata layer in each skill pack, perform natural language parsing on the description of power grid fault phenomena, and extract key evidence information from the description of fault phenomena; In B2: Multi-dimensional matching is performed based on key evidence information and preset trigger conditions in the metadata of each skill pack. The matching score of each skill pack is obtained by weighted calculation based on the weight of each dimension. In B3: If the exclusion condition is triggered, the matching score of the corresponding skill will be set to zero or reduced to the range where it is not loaded.

[0040] In one optional embodiment, natural language parsing can be structured information extraction based on a preset template. A structured trigger condition template is pre-configured for each skill pack. The template clearly defines required and optional fields. The user inputs a fault description according to the preset template format, and the values ​​of each field are extracted from the structured input. Standardized key evidence information can be obtained without natural language parsing. In another optional embodiment, natural language parsing can also be semantic parsing based on knowledge graphs, constructing a knowledge graph of power grid equipment-fault-phenomenon, mapping the triggering conditions of each skill package to nodes and edge relationships in the knowledge graph, with each skill package corresponding to a set of graph query templates, performing named entity recognition and relationship extraction on the fault phenomenon description input by the user, identifying equipment entities, fault phenomenon entities and their semantic relationships, querying and linking the extracted entities and relationships in the knowledge graph, and completing the missing information through graph reasoning to obtain structured key evidence information containing equipment type, fault phenomenon and its associated relationships; In this embodiment of the invention, natural language parsing includes a combination of rule filtering, semantic vector matching, and slot coverage scoring to identify relevant skills. Specifically, the main agent only reads the metadata layer of each skill package, performs word segmentation, synonym normalization, and entity recognition on the user description to obtain evidence slots such as device type, abnormal phenomena, operating parameters, time and location, and supplementary descriptions; converts the normalized user description and the description of each skill trigger condition into semantic vectors, and calculates the cosine similarity between the two; calculates the keyword or synonym hit rate, device category consistency, mandatory slot coverage, and whether the exclusion conditions are triggered; and calculates the skill matching score according to the weighted formula.

[0041] Skill matching score is expressed as follows: in, Indicates consistency of device objects. This indicates the semantic similarity between the user description and the skill trigger description. This indicates the required parameter, slot coverage. This indicates the hit rate of keywords and synonyms. This indicates skill priority or historical hit effect modification.

[0042] If the exclusion condition is triggered, the matching score for that skill will be set to 0 or reduced to the range where it is not loaded.

[0043] The system identifies the skill with the highest matching score that meets the loading threshold as the most relevant skill. For example, when the input is "the main transformer is making a humming sound and the oil temperature is rising," the equipment entity "main transformer" matches the main transformer fault analysis skill. The phenomenological features "humming sound" and "rising oil temperature" have a high semantic similarity to the skill's trigger description, and the exclusion conditions of skills such as line and switchgear are not triggered. Therefore, the main transformer fault analysis skill receives the highest matching score.

[0044] In this embodiment of the invention, after completing steps B1-B3, step S200 also includes steps B4-B11. In B4: Single skill loading is performed based on the match score; In B5: In response to a matching score greater than or equal to the first threshold, the instruction layer content of the corresponding skill is loaded into the current dialogue context; In B6: In response to a matching score that is less than the first threshold and greater than or equal to the second threshold, the instruction layer is not loaded, and the user is asked for supplementary information based on the missing slot; In B7: In response to a matching score being less than the second threshold, the instruction layer is not loaded, and the user is prompted to provide a more detailed description of the fault. In B8: Multiple skills are loaded based on their matching scores. In B9: In response to the fact that the matching scores of multiple skills are all greater than or equal to the first threshold, and the difference between the highest score and the second highest score is less than the third threshold, the highest score skill is set as the main skill, and the remaining skills, which do not exceed the preset number, are set as auxiliary skills, and the corresponding instruction layers are called for analysis respectively. In B10: In response to multiple skills having matching scores greater than or equal to the first threshold, and the difference between the highest and second-highest scores being greater than or equal to the third threshold, only the highest-scoring skill is loaded; In B11: In response to multiple skills having matching scores all below the first threshold, no skill instruction layer is loaded.

[0045] In this embodiment of the invention, the first threshold is 0.72, the second threshold is 0.60, and the third threshold is 0.08; Specifically, when the highest skill matches the score When the value is ≥0.72, the system loads the instruction layer content of the skill into the current dialogue context; when 0.60 ≤ When the value is less than 0.72, the system does not immediately load the complete instruction layer, but instead prompts the user for information about the device type, abnormal phenomena, or operating parameters based on the missing slots; when When the value is less than 0.60, the system does not load the skill instruction layer and prompts the user to provide a description of the fault.

[0046] The system supports loading multiple skills simultaneously. When multiple skills... When all scores are not lower than 0.72 and the difference between the highest and second-highest scores is less than 0.08, the system will set the highest-scoring skill as the primary skill and the remaining no more than two skills as auxiliary skills, and will call their respective instruction layers for candidate analysis; when the difference between the highest and second-highest scores is not less than 0.08, only the highest-scoring skill will be loaded.

[0047] It should be noted that the above threshold and difference control not only prevents irrelevant skills from entering the context, but also handles complex fault descriptions across devices or combinations of multiple phenomena.

[0048] The skill execution agent extracts or queries key parameters from user input according to the prompts of the instruction layer. For example, it automatically identifies the equipment type "main transformer", extracts the phenomenon characteristics "humming sound" and "temperature rise", and confirms or supplements information such as substation name and voltage level to the user.

[0049] For example, suppose a user inputs "the main transformer is making a humming sound and the oil temperature is rising." The main agent traverses the skill metadata, selects the "main transformer fault analysis skill," and dynamically loads the skill's instruction layer into the context. The system then extracts "main transformer" as the equipment type, "humming sound" and "rising oil temperature" as symptom characteristics, and queries the user for the transformer's model and operating years based on the instruction prompts. In this way, the system gradually introduces key information into the context, avoiding loading a large amount of irrelevant data at once.

[0050] In this embodiment of the invention, step S300 includes the following sub-steps C1-C4; In C1: Structured evidence is converted into evidence feature vectors; In C2: a search and matching is performed in the preset intrinsic safety standard database. For each root cause entry in the intrinsic safety standard database, the semantic similarity between the evidence feature vector and the standard feature vector of the root cause entry is calculated. The root cause matching evaluation dimension between the structured evidence and the root cause entry is calculated. The matching score of each root cause entry is obtained by weighted aggregation according to the preset weight of the root cause matching evaluation dimension. In C3: In response to root cause entries with matching scores greater than or equal to the fourth threshold, the corresponding root cause entries are taken as candidate root causes and arranged in descending order of matching scores to form a set of candidate root causes. In C4: In response to the absence of a root cause entry with a matching score reaching the fourth threshold, a prompt message is output.

[0051] In one optional embodiment, the retrieval matching can be based on deterministic rule matching, traversing each root cause entry in a preset intrinsically safe standard database. For each field in the structured evidence, an exact match or wildcard match is performed with the corresponding field of the root cause entry, and the number of matched fields is counted. A fixed weight is preset for each field, and the weights of the matched fields are accumulated to obtain a total matching score. Root cause entries with a total matching score reaching a preset threshold are selected as candidate root causes. In another optional embodiment, the retrieval matching can also be based on machine learning ranking. In a pre-defined intrinsically safe standard database, traditional information retrieval algorithms are used to initially screen each root cause entry to obtain an initial candidate set. For each root cause entry in the initial candidate set, multi-dimensional features are extracted, including text similarity features between evidence and root cause descriptions, device type matching features, co-occurrence features of phenomenon keywords, voltage level matching features, and historical statistical features. The extracted multi-dimensional features are input into a pre-trained machine learning ranking model, and the model outputs a matching score for each root cause entry. Root cause entries with matching scores reaching a preset threshold are selected as candidate root causes and ranked in descending order of score. In this embodiment of the invention, retrieval matching includes a method that fuses vector similarity and rule matching; The pre-defined intrinsic safety standard database includes data items such as root cause number, root cause name, associated risk point number, applicable equipment category, typical phenomenon characteristics, exclusion conditions, historical prior probability, likelihood probability configuration, and handling prompts. The retrieval module uses these data items to screen candidate root causes, and the Bayesian verification module uses prior probability and likelihood probability to calculate credibility.

[0052] Specifically, such as Figure 2 As shown, user descriptions are transformed into structured evidence vectors. This includes equipment type, operating parameters, and keywords related to abnormal phenomena. For example, It may contain {equipment type="main transformer", phenomenon=["humming sound", "oil temperature rise"]}; The resource layer retrieval is invoked to search for candidate risk points and root causes in a pre-built intrinsically safe standard database. Each risk point and root cause in the database has a unique number and is configured with applicable equipment, typical phenomenon characteristics, exclusion conditions, historical prior probabilities, and likelihood probability parameters.

[0053] First, structured evidence The equipment type, abnormal phenomena, operating parameters, and context information are converted into evidence feature vectors, and then semantic similarity is calculated with the standard feature vector of each root cause entry in the database. At the same time, the overlap of phenomenon features, the hit rate of equipment and voltage level rules, and the consistency of context parameters are calculated, and exclusion conditions are checked.

[0054] The root cause matching score can be expressed as: in, This indicates the semantic similarity between the evidence description and the root cause criterion description. Indicates the degree of overlap in the set of phenomenal features. This indicates the rule hit rate for equipment category, voltage level, risk category, etc. This indicates consistency in context, such as operating status, historical maintenance, and load level.

[0055] The fourth threshold is 0.65; when If the score is ≥0.65, the root cause is added to the candidate set, and the system returns no more than 5 candidate root causes in descending order of score. If no candidate root cause with a score of ≥0.65 exists, the user is prompted to provide supplementary evidence or the case is transferred to a human review.

[0056] In this embodiment of the invention, after completing steps C1-C4, step S300 also includes steps C5-C6; In C5: if the matching score of the highest-scoring candidate root cause is greater than or equal to the fifth threshold, and the difference between the matching scores of the highest-scoring candidate root cause and the second-highest-scoring candidate root cause is greater than or equal to the sixth threshold, the highest-scoring candidate root cause is selected as the preliminary root cause conclusion. In C6: In response to the highest-scoring candidate root cause having a matching score less than the fifth threshold, or the difference between the matching scores of the highest-scoring candidate root cause and the second-highest-scoring candidate root cause being less than the sixth threshold, at least two candidate root causes are retained and output to the credibility assessment, or a follow-up question prompt indicating missing evidence is output to the user.

[0057] In this embodiment of the invention, the fifth threshold is 0.75 and the sixth threshold is 0.08; The skill execution agent first eliminates candidates that conflict with device type, risk category, or exclusion criteria, and then processes the remaining candidates according to... Sort in descending order.

[0058] If the highest-scoring candidate is If the highest score is ≥0.75 and the difference between it and the second-highest score candidate is not less than 0.08, then the highest score candidate is taken as the preliminary root cause conclusion; if the highest score is lower than 0.75 or the difference between it and the second-highest score is less than 0.08, then multiple candidates are retained for Bayesian verification, or the user is asked to provide further key evidence.

[0059] The final best-matching root cause is the standard library entry that satisfies the hard constraints and has the highest overall matching score. The output only includes the root cause number, root cause name, associated risk point number, and matching basis for that entry; the agent is not allowed to create root cause text outside the standard library.

[0060] For example, after converting the description into a vector, the system searches a standard risk database, assuming the root causes are: multi-point grounding of the iron core, short circuit between coil turns, and cooler blockage. Based on clues such as "humming sound" and "temperature rise" in the user description, the skill-based intelligent agent evaluates the logical rationality of these candidate causes and initially selects "multi-point grounding of the iron core" as the output of this round of analysis. This process relies entirely on a pre-set database and historical experience, eliminating the possibility of unknown, illusory causes.

[0061] It should be noted that this invention combines hard constraint filtering with score ranking to strictly limit the root cause conclusions to the scope of the standard library, eliminate the risk of illusion caused by the free generation of intelligent agents, and ensure that the output results have compliance, verifiability and a unique mapping relationship under the standardized coding system, thus providing a reliable decision-making basis for the governance of power grid safety hazards.

[0062] In this embodiment of the invention, step S400 includes the following sub-steps D1-D5; In D1: Obtain the target root cause and structured evidence corresponding to the preliminary root cause conclusions; In D2: the frequency of the target root cause in historical failure data is used as the prior probability; In D3: the conditional probability of structured evidence appearing under the condition that the target root cause is true is taken as the likelihood probability; In D4: Based on the prior probability and the likelihood probability, calculate the posterior probability that the target root cause is valid given the presence of structured evidence; In D5: The confidence level of the preliminary root cause conclusion is determined based on the comparison between the posterior probability and the preset confidence threshold.

[0063] In embodiments of the present invention, such as Figure 3 As shown, let This is a hypothetical event, i.e., the root cause chosen by the skill agent. , The set of observed fault phenomena.

[0064] Prior probabilities include statistically analyzing the frequency of occurrence of the root cause in similar equipment failures from the power system fault management system (PMS) or historical data, using this as a priori information. For example, statistics might show that "multi-point grounding of the iron core" accounts for 30% of past main transformer failure records.

[0065] Likelihood probability includes, in the assumption Phenomena that occurred upon establishment The probability of this can be obtained from expert experience tables or historical case statistics. For example, if it is indeed a case of multi-point grounding of the iron core, the combined probability of the two phenomena of "humming sound" and "high oil temperature" is 80%.

[0066] The posterior probability calculated using Bayesian methods is expressed as follows: in, This can be estimated by normalizing all candidate root causes or experience. For prior probability, Let be the likelihood probability.

[0067] like Marked as high confidence, the system directly outputs the result and indicates high confidence level. For high thresholds, such as 0.8.

[0068] like Marked as requiring manual review, the posterior probability value and calculation basis are displayed to the user for the dispatcher to judge, where is a low threshold, such as 0.4.

[0069] like If the system determines that the agent's analysis result may be a hallucination or unreliable, it will automatically revert to step S300, select the next possible root cause from the candidates, or prompt the user to provide more information for reassessment.

[0070] For example, the selected root cause "multi-point grounding of the iron core" is verified, assuming prior... (Historical Fault Records), Likelihood After calculation, we obtained If the result exceeds 0.8, the system marks the conclusion as highly reliable and provides direct feedback to the user; if the result is between 0.4 and 0.8, the system prompts the dispatcher that "manual verification is required" and displays the probability value; if the result is below 0.4, the system considers the conclusion unreliable and automatically returns to search for other candidate reasons.

[0071] It should be noted that this step adds a mathematical credibility check to the agent analysis, effectively avoiding potential analytical biases or erroneous conclusions.

[0072] Example 2: The above is an illustrative scheme of a method for analyzing the root causes of inherent safety hazards in power grids according to this embodiment. It should be noted that the technical solution of this method for analyzing the root causes of inherent safety hazards in power grids is based on the same concept as the technical solution of the method described above. Details not described in detail in the technical solution of this method for analyzing the root causes of inherent safety hazards in power grids can be found in the description of the technical solution of the method described above.

[0073] This embodiment of a system for analyzing the root causes of inherent safety hazards in power grids includes: The skill library construction module is used to acquire skill classification dimension data and build an intrinsically safe skill library; The identification and calculation module is used to obtain a description of the power grid fault phenomenon, perform intent identification on the description of the power grid fault phenomenon based on the metadata of each skill package in the intrinsic safety skill library, calculate the matching score of each skill package, and dynamically load skills based on the matching score. The matching and filtering module is used to extract structured evidence from the description of the power grid fault phenomenon, perform retrieval and matching in a preset intrinsic safety standard database based on the structured evidence to obtain a set of candidate root causes and corresponding matching scores, and filter from the set of candidate root causes based on the matching scores to obtain preliminary root cause conclusions. The credibility assessment module is used to assess the credibility of the preliminary root cause conclusions using probabilistic reasoning verification methods, and obtain standardized power grid hidden danger root cause analysis results.

[0074] This embodiment also provides a computer device applicable to a method for analyzing the root causes of inherent safety hazards in power grids, including: The system includes a memory and a processor. The memory stores computer-executable instructions, and the processor executes these instructions to implement a root cause analysis method for inherent safety hazards in power grids, as proposed in the above embodiments.

[0075] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements a root cause analysis method for inherent safety hazards in power grids as proposed in the above embodiments.

[0076] The storage medium proposed in this embodiment and the method for analyzing the root causes of inherent safety hazards in power grids proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0077] Based on the above description of the implementation methods, those skilled in the art will clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computing device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0078] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for root cause analysis of inherent safety hazards in power grids, characterized in that, include: Acquire data on skill classification dimensions to build an intrinsic safety skill library; Obtain a description of a power grid fault phenomenon, perform intent recognition on the description of the power grid fault phenomenon based on the metadata of each skill package in the intrinsic safety skill library, calculate the matching score for each skill package, and dynamically load skills based on the matching score; Structured evidence is extracted from the description of the power grid fault phenomenon. Based on the structured evidence, a search and matching process is performed in a preset intrinsic safety standard database to obtain a set of candidate root causes and corresponding matching scores. Based on the matching scores, preliminary root cause conclusions are obtained by filtering from the set of candidate root causes. The credibility of the preliminary root cause conclusions was assessed using probabilistic reasoning verification methods, resulting in standardized root cause analysis results for potential power grid hazards.

2. The method for root cause analysis of inherent safety hazards in power grids as described in claim 1, characterized in that, Building an intrinsically safe skillset includes: Based on the skill classification dimension data, the power grid intrinsic safety analysis capability unit is decoupled into multiple independently packaged skill packages; In response to the fact that different power grid intrinsic safety analysis capability units have differences in at least one of the skill configuration attributes, the different power grid intrinsic safety analysis capability units are divided into different skill packages; In response to multiple scenarios sharing the same device object and having overlapping core elements, multiple scenarios are treated as sub-tasks under the same skill package. Configure at least three hierarchical storage structures for each skill pack to obtain an intrinsically safe skill library.

3. The method for root cause analysis of inherent safety hazards in power grids as described in claim 2, characterized in that, Intent recognition is performed on descriptions of power grid fault phenomena, and a matching score is calculated for each skill pack, including: Read the metadata layer in each skill pack, perform natural language parsing on the description of power grid fault phenomena, and extract key evidence information from the description of fault phenomena. Based on key evidence information and the preset trigger conditions in the metadata of each skill pack, a multi-dimensional match is performed, and a weighted calculation is performed based on the weight of each dimension to obtain the matching score of each skill pack. If the exclusion condition is triggered, the matching score of the corresponding skill will be set to zero or reduced to the range where it is not loaded.

4. The method for root cause analysis of inherent safety hazards in power grids as described in claim 3, characterized in that, Dynamic skill loading based on the matching score includes: Load a single skill based on the matching score; In response to a matching score greater than or equal to the first threshold, the instruction layer content of the corresponding skill is loaded into the current dialogue context; If the matching score is less than the first threshold but greater than or equal to the second threshold, the instruction layer is not loaded, and the user is asked for supplementary information based on the missing slot. If the matching score is less than the second threshold, the instruction layer is not loaded, and the user is prompted to provide a more detailed description of the fault. Multiple skills are loaded based on their matching scores. In response to the fact that the matching scores of multiple skills are all greater than or equal to the first threshold, and the difference between the highest score and the second highest score is less than the third threshold, the highest-scoring skill is set as the primary skill, and the remaining skills, which do not exceed a preset number, are set as auxiliary skills. The corresponding instruction layers are then called for analysis. If the matching scores of multiple skills are all greater than or equal to the first threshold, and the difference between the highest score and the second highest score is greater than or equal to the third threshold, only the highest-scoring skill will be loaded. If the matching scores for multiple skills are all less than the first threshold, no skill instruction layer is loaded.

5. The method for root cause analysis of inherent safety hazards in power grids as described in claim 4, characterized in that, Based on the structured evidence, a search and matching process is performed in a pre-defined intrinsic safety standard database to obtain a set of candidate root causes and corresponding matching scores, including: Transform structured evidence into evidence feature vectors; Search and match in the pre-set intrinsic safety standard database. For each root cause entry in the intrinsic safety standard database, calculate the semantic similarity between the evidence feature vector and the standard feature vector of the root cause entry, and calculate the root cause matching evaluation dimension between the structured evidence and the root cause entry. Based on the pre-set weights of the root cause matching evaluation dimension, aggregate to obtain the matching score of each root cause entry. In response to root cause entries with matching scores greater than or equal to the fourth threshold, the corresponding root cause entries are taken as candidate root causes and arranged in descending order of matching scores to form a set of candidate root causes. If no root cause entry with a matching score reaching the fourth threshold is found, a prompt message is output.

6. The method for root cause analysis of inherent safety hazards in power grids as described in claim 5, characterized in that, Preliminary root cause conclusions obtained from filtering the candidate root cause set based on the matching scores include: If the matching score of the highest-scoring candidate root cause is greater than or equal to the fifth threshold, and the difference between the matching scores of the highest-scoring candidate root cause and the second-highest-scoring candidate root cause is greater than or equal to the sixth threshold, the highest-scoring candidate root cause is selected as the preliminary root cause conclusion. In response to the highest-scoring candidate root cause having a matching score less than the fifth threshold, or the difference between the matching scores of the highest-scoring candidate root cause and the second-highest-scoring candidate root cause being less than the sixth threshold, at least two candidate root causes are retained and output to the credibility assessment, or a follow-up question prompt indicating missing evidence is output to the user.

7. A method for root cause analysis of inherent safety hazards in power grids as described in claim 1 or 6, characterized in that, The credibility assessment of the preliminary root cause conclusions includes: Obtain the target root causes and structured evidence corresponding to the preliminary root cause conclusions; The frequency of the target root cause in historical failure data is used as the prior probability. The conditional probability of structured evidence appearing under the condition that the target root cause is true is taken as the likelihood probability. Based on the prior probability and the likelihood probability, calculate the posterior probability that the target root cause is valid given the presence of structured evidence. The confidence level of the preliminary root cause conclusion is determined based on the comparison between the posterior probability and the preset confidence threshold.

8. A system for analyzing the root causes of inherent safety hazards in power grids, comprising applying the root cause analysis method for inherent safety hazards in power grids as described in any one of claims 1-7, characterized in that, include: The skill library construction module is used to acquire skill classification dimension data and build an intrinsically safe skill library; The identification and calculation module is used to obtain a description of the power grid fault phenomenon, perform intent identification on the description of the power grid fault phenomenon based on the metadata of each skill package in the intrinsic safety skill library, calculate the matching score of each skill package, and dynamically load skills based on the matching score. The matching and filtering module is used to extract structured evidence from the description of the power grid fault phenomenon, perform retrieval and matching in a preset intrinsic safety standard database based on the structured evidence to obtain a set of candidate root causes and corresponding matching scores, and filter from the set of candidate root causes based on the matching scores to obtain preliminary root cause conclusions. The credibility assessment module is used to assess the credibility of the preliminary root cause conclusions using probabilistic reasoning verification methods, and obtain standardized power grid hidden danger root cause analysis results.

9. A computer device, characterized in that, include: A memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the root cause analysis method for inherent safety hazards in power grids as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed by a processor, implements the steps of the root cause analysis method for inherent safety hazards in power grids as described in any one of claims 1 to 7.

Images