Power field supervision question-answering method and system for splitting complex questions and enhancing retrieval based on multiple rounds of thinking

Abstract

The invention relates to an electric power field supervision question-answering method and system based on complex question splitting and retrieval enhancement of multiple rounds of thinking, and belongs to the crossing field of power grid technical supervision and artificial intelligence. According to the method, through three core technologies of multi-round thinking iterative error correction, complex problem hierarchical decomposition and retrieval enhancement accurate matching, efficient and accurate correspondence between supervision problems and violation regulations is realized under multiple electrical professional scenes such as communication, power transformation and power transmission, and the method is suitable for business scenes such as automatic judgment of field supervision of power enterprises and intelligent matching of violation regulations.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of power grid technical supervision and artificial intelligence, and particularly relates to a power site supervision question and answer method and system based on complex problem decomposition and retrieval enhancement of multi-round thinking. BACKGROUND

[0002] The current power site supervision violation matching technology mainly includes two types:

[0003] 1. Manual matching mode: the staff needs to manually compare the standard regulations such as "Electric Power Safety Work Regulations" and "Power Grid Technical Supervision Guidelines", analyze the supervision questions such as "the grounding resistance of the transmission line tower is over-standard and the communication module is offline", judge the professional fields involved and the violated clauses, and the whole process relies on manual experience without automatic decomposition or thinking mechanism;

[0004] 2. Simple automation mode: through keyword matching such as "grounding resistance" corresponding to "transmission professional grounding specification" or a single model such as basic BERT model to retrieve the regulation library, directly output 1-2 matching results without problem decomposition, multi-round verification or retrieval optimization steps.

[0005] The closest existing technology is "power supervision violation matching method based on single text retrieval", and the specific implementation process is as follows:

[0006] 1. Receive the power site supervision question text such as "abnormal SF6 gas pressure of the power transformation equipment and missing temperature and humidity monitoring data";

[0007] 2. Extract keywords from the question text by using TF-IDF or BM25 algorithm, such as "SF6 gas pressure" and "temperature and humidity monitoring";

[0008] 3. Retrieve the preset power regulations library based on the extracted keywords, and return the top 3 regulations with the highest keyword coincidence degree;

[0009] 4. Directly output the 3 regulations as the matching results without subsequent problem decomposition, multi-round verification or retrieval strategy adjustment steps.

[0010] The core limitation of this scheme is that it does not design decomposition logic for "multi-specialty cross-problems", does not introduce multi-round thinking error correction mechanism, and the retrieval only relies on keyword coincidence degree without domain knowledge enhancement or strategy optimization.

[0011] The core shortcomings of the existing technology are as follows:

[0012] 1. No multi-round thinking mechanism: only output results through single retrieval / judgment without iterative process of "result evaluation → problem review → strategy adjustment → secondary reasoning", if the initial keyword extraction is biased such as missing "temperature and humidity monitoring", it will directly lead to matching error;

[0013] 2、Complex problems cannot be decomposed: In the face of complex supervision problems such as "multi-specialty intersection" such as power transformation + communication or "multi-dimensional nesting" such as "equipment failure + operation process violation", it is impossible to be decomposed into "single-specialty sub-problems" or "single-dimensional sub-tasks", and only the whole can be matched, resulting in fuzzy regulation corresponding;

[0014] 3、No enhancement mechanism for retrieval: Retrieval only depends on the coincidence of keywords, without combining power professional attributes such as "SF6 gas" dedicated power transformation professional, regulation applicable scene such as "outdoor pole tower" vs "indoor equipment" for enhancement, and without "preliminary screening→fine screening→correlation verification" layered retrieval strategy, the matching precision is low. SUMMARY

[0015] The purpose of the present application is to solve the problems of "complex problems cannot be decomposed, single round judgment is easy to make mistakes, and regulation matching lacks knowledge support" in power site supervision, and provide a power site supervision question and answer method and system based on complex problem decomposition and retrieval enhancement of multi-round thinking, through three core technologies of multi-round thinking iteration error correction, complex problem hierarchical decomposition, and retrieval enhancement precise matching, realize efficient and accurate correspondence between supervision problems and violated regulations in communication, power transformation, power transmission and other electrical professional scenes, and applicable to power enterprise site supervision automation determination, illegal regulation intelligent matching and other business scenes.

[0016] To achieve the above purpose, the technical scheme of the present application is: a power site supervision question and answer method based on complex problem decomposition and retrieval enhancement of multi-round thinking, comprising:

[0017] S1, power field retrieval enhancement knowledge base construction;

[0018] S2, enhanced triple data set construction;

[0019] S3, retrieval enhancement type prompt word template and classifier training;

[0020] S4, dynamic multi-specialty collaborative retrieval;

[0021] S5, multi-round thinking decomposition intelligent agent construction;

[0022] S6, double confidence judgment;

[0023] S7, cycle control and result output.

[0024] Further, step S1 comprises:

[0025] S11, data preprocessing

[0026] In structured table processing, for table data containing fields "regulation id, professional category p, applicable scenario u, violation situation v", first parse the table header metadata and build records , i represents the ith table; for the text field "violation situation" , use the domain word table and NLP segmentation to get the keyword set , m is the number of segmented words, and the weight of each keyword is calculated by TF-IDF, formula:

[0027]

[0028] Among them is the word frequency of the keyword in , that is the number of occurrences of in ÷ the total number of words; is the jth keyword in the ith table ; is the "violation situation" text field of the ith regulation; is the total number of regulations, is the number of regulations containing the keyword ; The "professional-scenario-regulation" association feature is encoded as a triple , forming a structured index item;

[0029] In structured text processing, for the procedure text with obvious chapter hierarchy, the pre-trained BERT-WWM model is used to generate semantic vectors for each paragraph , and the hierarchical structure "chapter title c→clause content a→violation example e" is extracted by combining regular / title rules; Then form a structured record for each clause:

[0030]

[0031] Among them is the chapter title corresponding to the jth clause, which is the hierarchical classification identifier to which the clause belongs; is the core content text of the jth clause, that is, the compliance requirement or violation judgment standard clearly specified in the regulation; is the violation example text corresponding to the jth clause; is the sentence vector output by BERT-WWM, with a dimension of 768;

[0032] In unstructured PDF processing, for scanned or image-containing reports, use the "OCR text recognition + power field NLP segmentation" pipeline to get the original text sequence , among them This represents the i-th word / segmentation unit in the original text sequence, where L is the length of the entire text segment; subsequently, a domain-specific named entity recognition model is used to identify technical terms, forming an entity set. , This represents the i-th entity identified through named entity recognition, where m is the number of identified entities, and the "description of the inspection problem" is extracted based on syntactic dependency relations. → Structured triples involving professional P → potential violation points V ,formula:

[0033]

[0034] in This is a detailed description of the kth inspection issue; Let k be the set of electrical engineering disciplines involved in the k-th problem; This is the set of potential violation points corresponding to the k-th problem, that is, the core elements that may violate the procedure extracted from the problem description; Problem description Vector representation of;

[0035] S12, Knowledge Base Organization

[0036] The output of input step S11 includes: , , The structured records, through unified schema mapping, three-level index construction, and HNSW vector index construction, form a unified knowledge base schema. ,formula:

[0037]

[0038] in For a globally unique identifier in the knowledge base For professional labels, For scene tags, Label the nature of the problem, Indicates record type, Represented by vectors, This is a set of related rule IDs.

[0039] Further, step S2 includes:

[0040] S21, Generation of Illegal Scenarios

[0041] Record each basic rule in the knowledge base Starting with the DeepSeek-R1 model, complex violation scenario descriptions are generated, with a constraint that the generated descriptions are between 40 and 80 characters long. Control prompts are used to ensure that at least two professional or two problem dimensions are included. To prevent the scenario from being too monotonous, temperature sampling and diversity penalties are used to generate M=3 variant scenarios for the same regulation. ,in Let m represent the m-th variant scenario corresponding to the i-th rule, where m = 1, 2, 3, ensuring coverage of different combinations.

[0042] S22, Triple Construction

[0043] For each complex problem generated in step S21 The set of decomposition results dominated by semi-automatic rules and manually verified and supplemented with annotations. Each of them Describe a "subproblem / subtask" and include its specialty and problem nature tags; synchronize each subproblem. Mark the search path in the knowledge base Formalized as:

[0044] in For professional paths, For scene constraints, As for the nature of the problem, Additional filtering conditions are added for spatial / temporal data, ultimately forming an enhanced triplet dataset. ,formula:

[0045]

[0046] in .

[0047] Further, step S3 includes:

[0048] S31, Professional and Exclusive Prompt Template Design

[0049] For each major Build search enhancement suggestion templates The template format is as follows:

[0050]

[0051] in Used to declare one's professional identity; Used to inject professional keywords and scenario constraints; The retrieval query statement after semantic rewriting of the sub-questions;

[0052] For multidisciplinary scenarios, design combined templates The constraints include "first search by their respective specialties, then search for cross-specialty collaborative records under the 'Operations and Maintenance Processes' category" to guide the order of cross-specialty joint searches;

[0053] S32, Template Classifier Training

[0054] The input to the template classifier is "input augmentation triples" The output is "the description of the decomposed sub-problems d + the summary of the initial screening results from the knowledge base s", and the output is "the optimal template number". With search strategy parameters The classifier uses a pre-trained BERT model based on Chinese as the encoder, and calculates the probability distribution of each template through a softmax input classification layer. ,formula:

[0055]

[0056] in The template number has a range of values. ;d represents the text of the decomposed sub-problems;s represents the summary text of the initial screening results from the knowledge base; W is the classifier weight matrix with dimensions [number of templates, 768]; b is the classifier bias vector with dimensions [number of templates, 1]; softmax (・) is the activation function that maps the feature vector to a probability distribution; during supervised training, the "ideal template and retrieval path" in the augmented triples are used as labels, and cross-entropy loss is employed. The formula is:

[0057]

[0058] in This is the one-hot encoding of the real template tag; the real template corresponds to... =1, all others are 0; reinforcement learning rewards are introduced, with "professional matching degree between sub-question and template" and "semantic similarity between retrieval result and sub-question" as reward functions. The REINFORCE algorithm is used to update the classifier parameters, gradient. The calculation formula is:

[0059]

[0060] This formula, within the joint context of "subproblem d + initial screening summary s", allows the model to repeatedly... Sample different search templates Calculate the corresponding reward Then, a weighted average of these results is taken to update the classifier parameters; where The final loss is defined as the weighted sum of the proportions of correct professional and correct scenario provisions in the search results. The formula is:

[0061]

[0062] λ represents the weight of the reinforcement learning loss; The reinforcement learning loss value is obtained by integrating the reinforcement learning loss formula.

[0063] Further, step S4 includes:

[0064] S41, Attention Weight Calculation

[0065] For the decomposed subproblem vector Introducing a group of professional experts Each major Corresponding to an expert scoring function ; Calculate the attention weights for each major using softmax ,formula:

[0066]

[0067] in , This reflects the importance of the corresponding major in the current sub-problem;

[0068] S42. Hierarchical Search Strategy

[0069] Define the difference between the largest and second largest professional weight as ,like If so, it is determined to be a single-major-dominated model; if If so, it is determined to be a multi-professional collaborative model;

[0070] (1) Single major mode

[0071] In single-discipline mode, the input sub-problem text d and the output of step S41 are... ,knowledge base Select the major with the highest weight. Perform hierarchical deep search; during initial screening, based on Subproblems and their corresponding problem property labels Filtering is performed in the knowledge base, and the results are... Represented as:

[0072]

[0073] Based on cosine similarity Sort by first 1 record;

[0074] Then, fine screening is carried out. Introducing scene tags on the basis of Filtering, the result Represented as:

[0075]

[0076] like If the number is too small, the scope can be broadened to a set of similar scenes;

[0077] Finally, perform association verification. For each rule k, read its associated rule set. And add relevant regulations to the candidate set. The formula is:

[0078]

[0079] By rearranging semantic similarity again, it was confirmed that no regulations highly related to the current sub-problem were significantly omitted.

[0080] (2) Multi-disciplinary model

[0081] In a multi-disciplinary model, multiple sets of disciplines with similar attention weights are considered simultaneously. The process involves "initial screening by major + cross-major association + scenario consistency verification + result merging and deduplication".

[0082] During the initial screening by major, each Independently perform initial screening based on "professional + nature tags" and similarity ranking to obtain... Cross-professional regulation association refers to constructing a cross-professional pairing set based on the explicit relationships between various professional regulations. ,formula:

[0083]

[0084] in, Indicates professional The initial screening regulations, Indicates professional The initial screening regulations; and Each pair of regulations represents a preliminary candidate list for the corresponding major, and a joint similarity score is calculated for each pair of regulations.

[0085] Next, scenario consistency verification is performed. For each candidate rule or combination of rules, the scenario label is checked to see if it is consistent with or compatible with the sub-problem scenario. If the scenarios are inconsistent, the joint score is reduced or the rule is eliminated. Finally, the results are merged and deduplicated, combining the candidates from each major with the results of cross-major combinations into a complete set. For duplicate clauses, the highest similarity score will be used as the criterion, and the clauses will be ranked according to their overall scores.

[0086] Further, step S5 includes:

[0087] S51, Multi-round Thinking Agent Overall Architecture

[0088] The multi-turn thinking agent receives the original supervisory problem description. First, the problem parsing module identifies the "professional cross-functional features" and "dimensional nesting features" to generate an initial tag set containing professional tags and problem nature tags; then, the decomposition execution module generates a multi-round decomposition plan based on the tag set and the rule base; the inference module is responsible for calling the DeepSeek-R1 model and the retrieval module of step S4 to provide intermediate judgments; the evaluation module performs consistency and confidence checks on the results of each round.

[0089] S52, Multi-round decomposition process

[0090] The multi-round decomposition follows a three-round progressive strategy of "professional dimension → nature dimension → scenario dimension":

[0091] The first round of professional dimension decomposition breaks down the complex problem Q into several sets of "single-professional sub-problems". The formula is:

[0092]

[0093] The second round of property dimension decomposition addresses each single-discipline sub-problem. Further break it down into a set of problem-related subtasks. Let the universal set of all problem properties and subtasks be:

[0094]

[0095] Any specific subtask is denoted as ;

[0096] Each subtask carries metadata including "whether quantitative judgment is required" and "whether cross-disciplinary collaboration is required";

[0097] The third round of scenario-dimensional decomposition adds scenario constraints to the sub-tasks, for each sub-task... Generate subtasks with context. ,formula:

[0098]

[0099] S53, Multi-round Reasoning Linkage

[0100] After each round of decomposition, the newly generated set of sub-problems / subtasks is input into the hybrid expert model in step S41 to calculate the professional attention weights. The sub-question d and the summary of the initial screening results s of the knowledge base are input into the template classifier in step S3 to select the optimal prompt word template and retrieval strategy. Then, dynamic retrieval is performed in step S4 to obtain the candidate clauses and intermediate reasoning conclusions of the current round.

[0101] The evaluation module performs "professional consistency" and "scenario adaptability" checks on the results of each round, including checking whether the professional tags of high-weight professions are consistent with the professional tags of the search results, and whether the applicable scenarios of the search rules cover or include sub-task scenarios.

[0102] If the verification passes, the corresponding round result is written into the decomposition path and reasoning path records, and the next round of decomposition begins; if it fails, the process backtracks to the decomposition execution module, adjusts the decomposition logic, and re-triggers the retrieval and reasoning.

[0103] Further, step S6 includes:

[0104] S61, Two-dimensional confidence calculation

[0105] (1) Inference confidence

[0106] The answer sequence output by the DeepSeek-R1 model Calculate the average log probability of the sequence using the token probability. The formula is:

[0107]

[0108] in, This represents the complete sequence of answers generated by the model, each... This is the t-th output token; This represents the prefix sequence generated before the generation of the t-th token. T represents the total number of tokens in the answer sequence; x represents the text input to the DeepSeek-R1 model, i.e., the sub-problem / scenario description; then, it is mapped to the [0,1] interval using a Sigmoid or normalization function, resulting in... The formula is:

[0109]

[0110] (2) Retrieval confidence

[0111] Search confidence score comprehensively considers factors including "professional matching", "scenario matching", and "semantic matching", and is defined as follows: ,formula:

[0112]

[0113] in For professional matching, For scene matching degree, This represents the average semantic similarity. , , Each has its own weight; and + + =1;

[0114] (3) Global overall confidence level

[0115] The global overall confidence score is used to adjust the degree of "more trusted inference" or "more trusted retrieval" according to business requirements, and is expressed as follows: ,formula:

[0116]

[0117] in For inference weights;

[0118] S62, Low Confidence Handling and Backtracking Mechanism

[0119] The system presets an inference confidence threshold. With retrieval confidence threshold The output of input step S61 , When any confidence level is below the threshold, i.e. This triggers the low-confidence processing flow:

[0120] Record the current "low confidence results + decomposition records + retrieval logs", and execute them according to the problem type. If the problem is mainly due to insufficient retrieval recall, return to step S4, adjust the retrieval strategy, switch from single-discipline mode to multi-discipline mode, relax the scenario filtering, or increase the number of candidates. If the problem is mainly due to unstable reasoning, return to step S5, readjust the problem decomposition logic: split the sub-problem granularity, add scenario or time dimension constraints, and re-execute multiple rounds of reasoning on the new decomposition results.

[0121] Furthermore, in step S61(3), the reasoning priority scenario... =0.7, search priority scenario =0.3.

[0122] Further, step S7 includes:

[0123] S71, Iteration Count Control

[0124] Introducing a loop counter Record the number of iterations for "dynamic retrieval → multi-round decomposition → confidence verification", with the maximum number of iterations set to [number to be filled in]. The loop control logic is formalized as follows: ;when When the confidence level is reached, the loop is exited and the result processing branch is entered, regardless of whether the confidence level is met.

[0125] S72, Result Processing Logic

[0126] If in Within the round, the inference confidence level is ≥ "And "retrieval confidence ≥" Given the dual conditions of "", the system outputs the following:

[0127] Final set of matching rules This includes the corresponding regulation number and a summary of the original text for each major.

[0128] Decompose path record This includes the order of decomposition of the professional, nature, and scenario dimensions in each round of decomposition;

[0129] Search path records This includes the retrieval mode used, the filtering conditions in each round, and the changes in the candidate set;

[0130] like If the double reset reliability threshold is still not reached, the system will decompose the current record. Search history The system packages and outputs "intermediate results and confidence curves" and transfers them to the manual processing channel; it also outputs "current decomposition / retrieval records" to help manual personnel quickly locate problems.

[0131] This invention also provides a power field supervision question-and-answer system based on multi-round thinking-based complex problem decomposition and retrieval enhancement, including:

[0132] The knowledge base and dataset construction module builds an enhanced knowledge base for retrieval in the power industry and constructs a triplet dataset based on the knowledge base.

[0133] The model training module, based on the constructed dataset, trains classifiers with customizable prompt word templates for different electrical engineering disciplines.

[0134] The collaborative retrieval module, based on a trained template classifier, dynamically selects a single-discipline or multi-discipline hierarchical retrieval strategy according to weight differences;

[0135] The agent construction module constructs a multi-round thinking agent based on the original inspection problem description;

[0136] The output module outputs the final result through multiple rounds of iterative thinking and confidence level closed-loop control.

[0137] Compared with the prior art, the present invention has the following beneficial effects:

[0138] 1. Multiple rounds of thinking lead to a "low misjudgment rate": Compared with the single judgment of existing technologies, this application reduces the misjudgment rate of complex problem matching by more than 60% through iterative error correction within 5 rounds. For example, the problem of missing "communication module" can be completed through secondary decomposition, and keyword deviation can be corrected by adjusting the search strategy.

[0139] 2. Complex problem decomposition brings "high adaptability": It can efficiently handle complex problems involving "multi-disciplinary intersection" such as substation + communication, and "multi-dimensional nesting" such as equipment + process. Compared with existing technologies that can only match the overall problem, this application improves the matching accuracy of sub-problems to over 80%.

[0140] 3. Enhanced matching brings "high accuracy": Combining professional classification, scenario constraints, and hierarchical search strategies, compared with traditional keyword search, the "professional accuracy" of regulation matching, for example, the proportion of substation regulations matched for substation issues has increased to over 80%. Detailed Implementation

[0141] The technical solution of the present invention will now be described in detail.

[0142] This invention provides a power field supervision question-and-answer method based on multi-round thinking-based complex problem decomposition and retrieval enhancement, including:

[0143] S1. Construction of an enhanced knowledge base for power sector retrieval;

[0144] S2, Enhanced triplet dataset construction;

[0145] S3. Retrieval Enhancement Prompt Template and Classifier Training;

[0146] S4. Dynamic multi-disciplinary collaborative search;

[0147] S5, Multi-round thinking decomposes intelligent agent construction;

[0148] S6. Double-reset reliability assessment;

[0149] S7, Loop Control and Result Output.

[0150] This invention also provides a power field supervision question-and-answer system based on multi-round thinking-based complex problem decomposition and retrieval enhancement, including:

[0151] The knowledge base and dataset construction module builds an enhanced knowledge base for retrieval in the power industry and constructs a triplet dataset based on the knowledge base.

[0152] The model training module, based on the constructed dataset, trains classifiers with customizable prompt word templates for different electrical engineering disciplines.

[0153] The collaborative retrieval module, based on a trained template classifier, dynamically selects a single-discipline or multi-discipline hierarchical retrieval strategy according to weight differences;

[0154] The agent construction module constructs a multi-round thinking agent based on the original inspection problem description;

[0155] The output module outputs the final result through multiple rounds of iterative thinking and confidence level closed-loop control.

[0156] The following is a detailed implementation process of the present invention.

[0157] The method of this invention achieves a deep integration of "multi-round thinking, complex problem decomposition, and enhanced retrieval matching" through seven steps: "knowledge base construction → dataset augmentation → template training → dynamic retrieval → multi-round decomposition → confidence verification → loop control". The specific process is as follows:

[0158] 1. Construction of an Enhanced Knowledge Base for Power Industry Retrieval

[0159] This step lays the foundation for subsequent enhanced matching in retrieval, and includes two sub-modules: data preprocessing and knowledge base organization, both using a structured intermediate representation. Storage (structured intermediate representation) The temporary storage format for records from various data sources is ultimately mapped to a unified knowledge base schema. ).

[0160] 1.1 Data Preprocessing

[0161] In processing structured tables (such as the "Regulations on Violations of Power Equipment" table), for table data containing fields such as "Regulations Number id, Professional Category p, Applicable Scenario u, Violation Circumstance v", the table header metadata is first parsed and records are constructed. Regarding the text field "Violation Situation" Keyword set is obtained by using domain thesaurus and NLP word segmentation. The weight of each keyword can be calculated using TF-IDF, with the formula:

[0162]

[0163] in Keywords Text regarding violations word frequency in exist Number of times it appears ÷ (Total word count) For the i-th regulation The j-th keyword in (belonging to) gather); For the "Violation Circumstances" text field of the i-th regulation; For the total number of regulations, The number of regulations containing this word is used for subsequent similarity retrieval and vectorization representation. Based on this, the "profession-scenario-regulation" association feature is encoded into triples. This creates structured index items, facilitating quick filtering by profession and scenario.

[0164] In processing structured texts (such as the "Guideline for Supervision of Power Transmission Engineering"), for procedural texts with clear chapter hierarchies, a pre-trained BERT-WWM model is used to generate semantic vectors for each paragraph. (Based on the pre-trained BERT-WWM model with an output dimension of 768), and using regular expressions / title rules to extract the hierarchical structure "section title c → clause content a → violation example e". Then, a structured record is generated for each clause.

[0165]

[0166] in The title is the chapter title corresponding to the j-th clause, which is the hierarchical classification identifier of the clause. The core text of the j-th clause is the compliance requirements or violation judgment criteria clearly stipulated in the regulations; Here is the example text of the violation corresponding to the j-th clause; The sentence vector is output by BERT-WWM, with a dimension of 768, and is used for subsequent cosine similarity retrieval.

[0167] In processing unstructured PDFs (such as on-site inspection reports), for scanned versions or reports containing images, a pipeline of "OCR text recognition + NLP word segmentation in the power industry" is used to obtain the original text sequence. Subsequently, a domain-specific named entity recognition model was used to identify technical terms such as "pole grounding" and "optical cable splicing," forming an entity set. And extract the "description of the inspection problem" based on syntactic dependency relations. → Structured triples involving professional P → potential violation points V ,formula:

[0168]

[0169] in This is a detailed description of the kth inspection issue; Let k be the set of electrical engineering disciplines involved in the k-th problem; This is the set of potential violation points corresponding to the k-th problem, that is, the core elements that may violate the procedure extracted from the problem description; Problem description The vector representation of .

[0170] 1.2 Knowledge Base Organization

[0171] Input step 1.1 Output , , Structured records are used to form a unified knowledge base schema through unified schema mapping, three-level index construction, and HNSW vector indexing (based on ANN). ,formula:

[0172]

[0173] in For a globally unique identifier in the knowledge base For professional tags (communication / substation / transmission, etc.) For scene tags, Label the nature of the problem (equipment malfunction / process violation / data anomaly, etc.). Indicate the record type (details / question-answer pair / triple). Represented by vectors, This is a set of related rule IDs. The knowledge base uses a three-level hierarchical index: "professional dimension → question nature dimension → specific details / question-answer pairs / triples", and the vector field... Associated with HNSW vector index, it supports vector similarity retrieval after filtering by tags, enabling multi-dimensional filtering and multi-category association.

[0174] 2. Enhanced Triple Dataset Construction

[0175] This step focuses on constructing training data for multi-round thinking and complex problem decomposition. It generates violation scenarios that cover multiple disciplines and multiple dimensions, and annotates each scenario with "decomposition result + retrieval index" to enhance the triplet.

[0176] 2.1 Generation of violation scenarios

[0177] Record each basic rule in the knowledge base Starting with the DeepSeek-R1 model, a description of a complex violation scenario is generated. Constraint generation length The description should be between 40 and 80 characters, and should include at least two professional aspects or two problem dimensions using control prompts, such as "SF6 gas pressure in substation equipment is below the threshold, and the corresponding communication module has been offline for 3 hours, and maintenance records have not been updated." To prevent the scenario from being too monotonous, temperature sampling and nucleus sampling are used to generate M=3 variant scenarios for the same regulation. This ensures coverage of different combinations. The output is... The set of variant scenarios provides data support for subsequent triple construction.

[0178] 2.2 Triple Construction

[0179] For each complex problem generated in step 2.1 The set of decomposition results dominated by semi-automatic rules and manually verified and supplemented with annotations. Each of them Describe a "subproblem / subtask" and include its specialty and problem nature tags; synchronize each subproblem. Mark the search path in the knowledge base Formalized as:

[0180] in For professional paths (such as "substation major") For scenario constraints (such as "outdoor 220kV substation") For the nature of the problem (e.g., "equipment malfunction"), Additional filtering conditions are added for spatial / temporal data. This ultimately results in an enhanced triplet dataset. ,formula:

[0181]

[0182] in This provides supervision signals for subsequent template classifiers and multi-round inference.

[0183] 3. Retrieval Enhancement Tip Templates and Classifier Training

[0184] This step achieves end-to-end optimization of "sub-problem → optimal prompt word template → retrieval strategy" by designing exclusive prompt word templates for different electrical engineering disciplines and training a template classifier.

[0185] 3.1. Professionally Designed Prompt Templates

[0186] For each major Build search enhancement suggestion templates The template generally takes the following form:

[0187]

[0188] in Used to declare one's professional identity, such as "[Substation Professional]"; Used to inject professional keywords and scenario constraints; This is the retrieval query statement after semantic rewriting of the sub-problem.

[0189] For example, the template for substation specialization can be specifically described as: "[Substation Specialization] Please prioritize searching for regulations under the 'Equipment Failure' category, which must contain keywords such as 'SF6 gas pressure' and 'insulating oil dielectric loss', and the applicable scenario must be 'outdoor 220kV substation'."

[0190] For multidisciplinary scenarios, design combined templates The constraints include “first search by their respective specialties, then search for cross-specialty collaborative records under the ‘Operations and Maintenance Processes’ category” to guide the order of cross-specialty joint searches.

[0191] 3.2 Template Classifier Training

[0192] The input to the template classifier is "input augmentation triples" The output is "the description of the decomposed sub-problems d + the summary of the initial screening results from the knowledge base s", and the output is "the optimal template number". With search strategy parameters (e.g., professional weights, scene filtering strength, etc.)". The main body of the classifier uses a BERT-Based-Chinese model pre-trained in Chinese as the encoder, and calculates the probability distribution of each template through a softmax input classification layer. ,formula:

[0193]

[0194] in The template number has a range of values. ;d represents the text of the decomposed sub-problems;s represents the summary text of the initial screening results from the knowledge base; W is the classifier weight matrix (dimension [number of templates, 768]); b is the classifier bias vector (dimension [number of templates, 1]); softmax (・) is the activation function, mapping the feature vector to a probability distribution (ΣP=1). During supervised training, the "ideal template and retrieval path" in the augmented triples are used as labels, and cross-entropy loss is employed. The formula is:

[0195]

[0196] in One-hot encoding of the real template tag (corresponding to the real template) =1, the rest are 0). Based on this, reinforcement learning rewards are introduced, with the reward function being "the professional matching degree between the sub-question and the template" and "the semantic similarity between the retrieval result and the sub-question". The REINFORCE algorithm is used to update the classifier parameters, gradient. The calculation formula is:

[0197]

[0198] in This can be defined as the weighted sum of the proportions of correct professional and correct scenario-based clauses in the search results. Final loss. The formula is:

[0199]

[0200] λ is the weight of the reinforcement learning loss (λ=0.3, determined through optimization on the validation set); The reinforcement learning loss value is obtained by integrating the above reinforcement learning loss formula.

[0201] 4. Dynamic multi-disciplinary collaborative search

[0202] This step is the core of enhanced matching for retrieval. It combines a hybrid expert model to calculate the attention weights of each specialty and dynamically selects a single-specialty or multi-specialty hierarchical retrieval strategy based on the differences in weights.

[0203] 4.1 Attention Weight Calculation

[0204] For the decomposed subproblem vector Introducing a group of professional experts Each major Corresponding to an expert scoring function The attention weights for each specialty are calculated using softmax. ,formula:

[0205]

[0206] in , This reflects the importance of the major in the current sub-problem.

[0207] For example, for the "SF6 pressure subproblem", the substation-specific weights are usually obtained after training. The weight of telecommunications major It has a lower weight in supporting other professional fields.

[0208] 4.2 Hierarchical Search Strategy

[0209] Define the difference between the largest and second largest professional weight as ,like If so, it is determined to be a single-major-dominated model; if If so, it is determined to be a multi-professional collaborative mode.

[0210] (1) Single major mode

[0211] In single-discipline mode, the input subproblem d and the output of step 4.1 are... ,knowledge base Select the major with the highest weight. Perform a hierarchical depth search. During the initial screening, based on... Subproblems and their corresponding problem property labels Filtering is performed in the knowledge base, and the results are... Represented as:

[0212]

[0213] And based on cosine similarity Sort by first 1 record.

[0214] Then, fine screening is carried out. Introducing scene tags on the basis of Filtering, the result Represented as:

[0215]

[0216] like If the number is too small, the scope can be broadened to a set of similar scenarios (such as "outdoor 220kV substation" and "outdoor 110kV substation").

[0217] Finally, perform association verification. For each rule k, read its associated rule set. And add relevant regulations to the candidate set. The formula is:

[0218]

[0219] By rearranging semantic similarity again, it was confirmed that there were no obvious omissions of regulations that are highly related to the current sub-problem.

[0220] (2) Multi-disciplinary model

[0221] In a multi-disciplinary model, multiple sets of disciplines with similar attention weights are considered simultaneously. The process involves "initial screening by specialty + cross-specialty association + scenario consistency verification + result merging and deduplication".

[0222] During the initial screening by major, each Independently perform initial screening based on "professional + nature tags" and similarity ranking to obtain... Cross-disciplinary regulation association refers to constructing a cross-disciplinary pairing set based on the explicit relationships between various professional regulations (such as the substation "equipment failure" regulation and the communication "data transmission" regulation being associated through the same equipment ID or the same maintenance work order). ,formula:

[0223]

[0224] in, Indicates professional The initial screening regulations, Indicates professional The initial screening regulations; and Each pair of regulations represents a preliminary set of candidate regulations for the corresponding major, and a joint similarity score is calculated for each pair of regulations.

[0225] Next, scenario consistency verification is performed. For each candidate rule or combination of rules, the scenario label is checked to see if it is consistent with or compatible with the sub-problem scenario. If the scenarios are inconsistent (e.g., one rule is "indoor substation" and another is "outdoor line"), its joint score is reduced or it is eliminated. Finally, the results are merged and deduplicated, combining the candidates from each specialty with the results of cross-specialty combinations into a complete set. For duplicate clauses, the highest similarity is used as the criterion, and they are ranked according to the comprehensive score to provide input for subsequent confidence assessment.

[0226] 5. Multi-round thinking to decompose the construction of intelligent agents

[0227] This step utilizes a multi-round thinking agent comprised of a "problem analysis module → decomposition and execution module → reasoning module → evaluation module" to achieve layer-by-layer decomposition and multi-round reasoning linkage of complex problems.

[0228] 5.1 Overall Architecture of Multi-Turn Thinking Agent

[0229] The multi-turn thinking agent receives the original supervisory problem description. First, the problem analysis module identifies the problem's "professional crossover characteristics" and "dimensional nesting characteristics." For example, if Q simultaneously contains "SF6" and "communication module," it is determined that it involves crossover between substation and communication disciplines; if it simultaneously contains "fault" and "missing record," it is determined to be a nested dimension of equipment and process, and an initial tag set containing professional tags and problem nature tags is generated. Subsequently, the decomposition execution module generates a multi-round decomposition plan based on the tag set and rule base; the inference module is responsible for calling the DeepSeek-R1 model and the retrieval module in step 4 to provide intermediate judgments; and the evaluation module performs consistency and confidence checks on the results of each round.

[0230] 5.2 Multi-round decomposition process

[0231] The multi-round decomposition follows a three-round progressive strategy of "professional dimension → nature dimension → scenario dimension":

[0232] The first round of professional dimension decomposition breaks down the complex problem Q into several sets of "single-professional sub-problems". The formula is:

[0233]

[0234] For example, it can be broken down into "electronics sub-problem: SF6 pressure anomaly" and "communication sub-problem: module offline", and each sub-problem is labeled with a professional tag.

[0235] The second round of property dimension decomposition addresses each single-discipline sub-problem. Further break it down into a set of problem-related subtasks. For example, "equipment fault determination task + violation matching task". Let the complete set of all problem-related subtasks be:

[0236]

[0237] Any specific subtask is denoted as ;

[0238] Each subtask carries metadata such as "whether quantitative judgment is needed" and "whether cross-disciplinary collaboration is needed," which is used for subsequent template selection and retrieval strategy adjustments.

[0239] The third round of scenario-dimensional decomposition adds scenario constraints to the sub-tasks, for each sub-task... Generate subtasks with context. ,formula:

[0240]

[0241] For example, the "equipment fault determination task" can be further refined into "equipment fault determination task → scenario: outdoor 220kV substation".

[0242] 5.3 Multi-turn reasoning linkage

[0243] After each round of decomposition, the newly generated set of sub-problems / sub-tasks is input into the hybrid expert model (step 4.1) to calculate the expert attention weights. Then, input the sub-question d and the summary of the initial screening results s from the knowledge base to the template classifier in step three to select the optimal prompt word template and retrieval strategy. Then, perform dynamic retrieval in step four to obtain the candidate clauses and intermediate reasoning conclusions for the current round.

[0244] The evaluation module performs "professional consistency" and "scenario adaptability" checks on the results of each round. For example, it checks whether the professional tags of high-weight professions are consistent with the professional tags of the search results, and whether the applicable scenarios of the search rules cover or include sub-task scenarios.

[0245] If the verification passes, the result of that round is written into the decomposition path and reasoning path records, and the next round of decomposition begins; if it fails, the process is backtracked to the decomposition execution module, the decomposition logic is adjusted (such as supplementing scenario constraints or splitting excessively long sub-problems), and the retrieval and reasoning are triggered again.

[0246] 6. Double-reset reliability assessment

[0247] This step, as the core error correction link of multi-round thinking, quantifies the reliability of the current result from two dimensions: "inference confidence" and "retrieval confidence", and triggers backtracking and re-decomposition in the case of low confidence.

[0248] 6.1 Two-dimensional confidence calculation

[0249] (1) Inference confidence

[0250] The answer sequence output by the DeepSeek-R1 model Calculate the average log probability of the sequence using the token probability. The formula is:

[0251]

[0252] in, This represents the complete sequence of answers generated by the model, each... This is the t-th output token; This represents the prefix sequence generated before the generation of the t-th token. T represents the total number of tokens in the answer sequence; x represents the text input to the DeepSeek-R1 model, i.e., the sub-problem / scenario description; then, it is mapped to the [0,1] interval using a Sigmoid or normalization function, resulting in... The formula is:

[0253]

[0254] For example, when the overall output probability sequence of "SF6 pressure anomaly corresponding to regulation A" is high, A value close to 1 indicates that the reasoning result is stable and reliable.

[0255] (2) Retrieval confidence

[0256] The retrieval confidence score comprehensively considers factors such as "professional matching degree," "scenario matching degree," and "semantic matching degree," and is defined as follows: ,formula:

[0257]

[0258] in For professional matching degree (such as the proportion of substation professional regulations in the search results of substation-related issues). For scene matching degree (such as the proportion of outdoor regulations in the search results of outdoor scene sub-question). This represents the average semantic similarity. , .3, (The weights sum to 1, optimized through grid search).

[0259] For example, if the proportion of substation-related regulations in the search results for substation-related issues is ≥ 0.95, the proportion of outdoor scenario regulations is ≥ 0.90, and the average semantic similarity is ≥ 0.85, then... This was determined to be a high-confidence search.

[0260] (3) Global overall confidence level

[0261] The global overall confidence score is used to adjust the degree of "more trusted inference" or "more trusted retrieval" according to business requirements, and is expressed as follows: ,formula:

[0262]

[0263] in For inference weight (default 0.5; inference-first scenario) =0.7, search priority scenario =0.3).

[0264] 6.2 Low Confidence Handling and Backtracking Mechanism

[0265] The system presets an inference confidence threshold. (0.85) and the retrieval confidence threshold (0.90), the output of step 6.1 , When any confidence level is below the threshold ( This triggers the low-confidence processing flow:

[0266] Record the current "low-confidence results + decomposition records + retrieval logs", and execute them according to the problem type. If the problem mainly manifests as insufficient retrieval recall (such as low professional matching or scenario matching), return to step 4, adjust the retrieval strategy, switch from single-professional mode to multi-professional mode, relax scenario filtering, or increase the number of candidates. If the problem mainly manifests as unstable reasoning (such as large differences in answers generated in multiple rounds), return to step 5, readjust the problem decomposition logic (split the granularity of sub-problems, add scenario or time dimension constraints), and re-execute multiple rounds of reasoning on the new decomposition results. This bidirectional backtracking mechanism ensures that the system can gradually improve the overall matching quality through multiple rounds of correction when encountering boundary scenarios and highly complex multi-professional problems.

[0267] 7. Loop Control and Result Output

[0268] To achieve a multi-round closed loop of thinking, this step uses iteration number control and result processing logic to ensure that the system converges to a high-confidence result or is handed over to human in an orderly manner within a limited number of rounds.

[0269] 7.1 Iteration Count Control

[0270] Introducing a loop counter Record the number of iterations for "dynamic retrieval → multi-round decomposition → confidence verification", with the maximum number of iterations set to [number to be filled in]. The loop control logic is formalized as follows: ;when When the confidence level is reached, the loop will exit and the result processing branch will be entered, regardless of whether the confidence level is met.

[0271] 7.2 Result Processing Logic

[0272] If in Within the round, the inference confidence level is ≥ "And "retrieval confidence ≥" Given the dual conditions of "", the system outputs the following:

[0273] Final set of matching rules This includes the corresponding regulation number and a summary of the original text for each major.

[0274] Decompose path record This includes the order in which the professional, nature, and scenario dimensions are broken down in each round, such as "professional → nature → scenario";

[0275] Search path records This includes the search mode used (single-discipline / multi-discipline), the filtering conditions in each round, and the changes in the candidate set. For example, "single-discipline in-depth search: substation discipline → outdoor equipment → equipment failure → SF6 pressure regulations".

[0276] like If the double reset reliability threshold is still not reached, the system will decompose the current record. Search history The system packages and outputs "intermediate results and confidence curves" and transfers them to the manual processing channel; it also outputs "current decomposition / retrieval records" to help manual personnel quickly locate problems.

[0277] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A power site supervision question and answer method based on multi-round thinking-based complex problem splitting and retrieval enhancement, characterized in that, Comprise: S1, power field search enhancement knowledge base construction; S2, enhance the construction of triple data set; S3, search enhancement prompt word template and classifier training; S4, dynamic multi professional collaborative search; S5, multi round thinking split agent construction; S6, double confidence judgment; S7, cycle control and result output.

2. The multi-round thinking based complex problem splitting and retrieval enhanced power field supervision question and answer method according to claim 1, characterized in that, Step S1, comprising: S11, data preprocessing In structured table processing, for table data containing fields "regulation id, professional category p, applicable scenario u, violation situation v", first parse the table header metadata and build records , i represents the i-th table; for the text field "violation situation" , use the domain glossary and NLP segmentation to obtain a keyword set , m is the number of segmented words, and the weight of each keyword is calculated by TF-IDF, formula: wherein is the keyword is the frequency of the keyword in the text ; and is the total number of words in the text ; and is the jth keyword in the ith regulation ; and is the number of regulations containing the keyword ; and is a structured index item​ In structured text processing, for the regulation text with obvious chapter hierarchy, the pre-trained BERT-WWM model is used to generate semantic vector for each paragraph and combined with the regular / header rule to extract the hierarchical structure "chapter title c→clause content a→violation example e"; then form a structured record for each clause: wherein is the chapter title corresponding to the jth clause, which is the hierarchical classification identifier to which the clause belongs; is the core content text of the jth clause, i.e., the compliance requirement or violation determination standard explicitly specified in the regulation; is the violation example text corresponding to the jth clause; is the sentence vector output by BERT-WWM, with a dimension of 768; In the unstructured PDF processing, for scanned version or report containing pictures, the "OCR text recognition + power field NLP segmentation" pipeline is adopted to obtain the original text sequence , wherein represents the i-th word / segmentation unit in the original text sequence, and L is the length of the entire text; then a domain named entity recognition model is used to identify professional terms to form an entity set , represents the i-th entity obtained by named entity recognition, and m is the number of identified entities, and the syntactic dependency relationship is used to extract the structured triple of "supervision question description → involves professional P→ potential violation point V" , formula: wherein is the specific description text of the kth supervision problem; is the set of power specialties involved in the kth problem; is the set of potential violation points corresponding to the kth problem, i.e., the core elements that may violate the regulations extracted from the problem description; is the problem description in the vector representation; S12, knowledge base organization The structured record outputted by the inputting step S11 includes , , The structured record outputted by the inputting step S11 includes , formula: wherein is a knowledge base global unique identifier, is a professional tag, is a scenario tag, is a question nature tag, represents a record type, is a vector representation, is a set of associated regulation IDs.

3. The multi-round thinking based complex problem splitting and retrieval enhanced power field supervision question and answer method according to claim 2, characterized in that, Step S2, comprising: S21, violation scene generation Record each basic rule in the knowledge base Starting with the DeepSeek-R1 model, complex violation scenario descriptions are generated, with a constraint that the generated descriptions are between 40 and 80 characters long. Control prompts are used to ensure that at least two professional or two problem dimensions are included. To prevent the scenario from being too monotonous, temperature sampling and diversity penalties are used to generate M=3 variant scenarios for the same regulation. ,in Let m represent the m-th variant scenario corresponding to the i-th rule, where m = 1, 2, 3, ensuring coverage of different combinations. S22, triple construction For each complex question generated in step S21 The corresponding decomposition result set is marked by semi-automatic rules and manually checked and supplemented , wherein each is a "sub-question / sub-task" description, and is accompanied by its professional and problem nature label; for each sub-question Mark the retrieval path in the knowledge base , formalized as: wherein is a professional path, is a scenario constraint, is a problem property, is a spatial / temporal, etc. supplementary filter condition; resulting in an enhanced triple dataset , formula: wherein .

4. The multi-round thinking based complex problem splitting and retrieval enhanced power on-site supervision question and answer method according to claim 3, characterized in that, Step S3, comprising: S31, professional exclusive prompt word template design For each specialty , build a search enhanced prompt template , template form is: wherein for declaring professional identity; for injecting professional keywords and scenario constraints; the search query statement after semantic rewriting for the sub-problems; For multi-specialty cross-scenarios, design a combination template Its constraints include "first search by each specialty respectively, then search cross-specialty coordination records under the 'operation and maintenance process' category" to guide the cross-specialty joint search order; S32, template classifier training The input of the template classifier is "input augmented triplets , decomposed sub-problem description d+ knowledge base preliminary screening result summary s", and the output is "optimal template number and search strategy parameters "; the classifier body uses a Chinese pre-trained BERT model as an encoder, and calculates the probability distribution of each template through a softmax output classification layer , formula: wherein is the template number, and the value range is ; d is the decomposed sub-problem text; s is the summary text of the knowledge base preliminary screening result; ; W is the classifier weight matrix, with dimensions [template number, 768]; b is the classifier bias vector, with dimensions [template number, 1]; softmax(・) is an activation function that maps the feature vector to a probability distribution; in the supervised training stage, the "ideal template and retrieval path" in the enhanced triple is taken as the label, and the cross-entropy loss is used, and the formula is: wherein one-hot encoding of the true template label, the true template corresponding to = 1, and the rest are 0; introduce reinforcement learning rewards, and take the "professional matching degree of the sub-problems and the templates" and the "semantic similarity of the retrieval results and the sub-problems" as the reward functions , and the REINFORCE algorithm is used to update the classifier parameters, and the gradient The calculation formula is: The formula lets the model make multiple passes over the joint context of "subproblem d + prescreening summary s" sampling different retrieval templates , compute the corresponding rewards , and take a weighted average of the results to update the classifier parameters; where is defined as the weighted sum of the fraction of correct professions and correct scene regulations in the retrieval results, and the final loss is given by: λ is a reinforcement learning loss weight; is a reinforcement learning loss value, obtained by integrating the reinforcement learning loss formula.

5. The multi-round thinking based complex problem splitting and retrieval enhanced power field supervision question and answer method according to claim 4, characterized in that, Step S4, comprising: S41, attention weight calculation For the decomposed sub-problem vector , introduce a set of professional experts , each professional corresponds to an expert scoring function ; use softmax to calculate the attention weight of each professional , formula: wherein , , reflects the importance of the respective specialty in the current sub-problem; S42, hierarchical search strategy Define the difference between the maximum professional weight and the second maximum weight as If , it is determined as single-professional dominant mode; if , it is determined as multi-professional collaborative mode; (1) single professional mode In the single-professional mode, input the sub-problem text d, and the output of step S41 , the knowledge base , select the professional with the largest weight to perform hierarchical depth search; during preliminary screening, according to the and the problem property label corresponding to the sub-problem , perform filtering in the knowledge base, and the obtained result is represented as: by cosine similarity ranked top records; Subsequently, fine screening is performed, and Introducing scene tags on the basis Filtering, the resulting result Is expressed as: If Too few, the scene can be relaxed as a similar scene set; Finally, the correlation verification is performed, and the relevant regulations are supplemented into the candidate set For each regulation k, read its correlation regulation set , and supplement the relevant regulations into the candidate set , the formula is: Through the rearrangement of semantic similarity again, confirm that there is no obvious omission in the regulations highly related to the current sub problem; (2) multi professional mode In the multi-professional mode, multiple professional sets with similar attention weights are considered simultaneously , a "specialty preliminary screening + cross-professional association + scene consistency verification + result merging and deduplication" process is performed; When screening by specialty, for each The "specialty + property tag" screening and similarity ranking are performed independently, and the results are obtained ; Cross-specialty regulation association, that is, according to the explicit association relationship between various professional regulations, a cross-specialty pairing set is constructed , formula: wherein, representing the initial screening regulations from the respective professional , representing the initial screening regulations from the respective professional ; and are the sets of initial screening candidate regulations for the respective professional, and the joint similarity score is computed for each pair of regulations; Subsequently, scene consistency check is performed, i.e. for each candidate regulation or regulation combination, it is checked whether its scene label is consistent or compatible with the sub-problem scene, and if the scene is inconsistent, its joint score is reduced or it is removed; finally, result merging and deduplication are performed, and the professional candidate and cross-professional combination results are merged into a full set The repeated regulations are subject to the highest similarity, and are sorted according to the comprehensive score.

6. The multi-round thinking based complex problem splitting and retrieval enhanced power field supervision question and answer method according to claim 5, characterized in that, Step S5, comprising: S51, multi round thinking agent overall architecture The multi-round thinking intelligent agent receives an original supervision question description , first, the question analysis module identifies its "professional cross characteristics" and "dimensional nested characteristics", generating an initial label set containing professional labels and question property labels; then, the decomposition execution module generates a multi-round decomposition plan according to the label set and the rule library; the reasoning module is responsible for calling the DeepSeek-R1 model and the retrieval module of step S4 to give intermediate judgments; the evaluation module checks the consistency and confidence of each round of results; S52, multi round decomposition process Multi round decomposition follows the three round progressive strategy of "professional dimension→property dimension→scene dimension": The first round of professional dimension decomposition will split the complex problem Q into a number of "single-professional sub-problems" sets The formula is: Second round of property dimension decomposition, for each single professional sub-problem , further split into a set of problem property sub-tasks , let the full set of all problem property sub-tasks be: Any of the specific subtasks is noted as ; Each sub task carries the meta information including "whether quantitative judgment is needed" and "whether cross professional collaboration is needed"; The third round of scene dimension decomposition, on the basis of the nature of sub task to supplement the scene constraints, for each sub task Generate sub task with scene , formula: S53, multi round reasoning linkage After each round of decomposition, the newly generated sub-problems / sub-tasks set is input into the hybrid expert model of step S41 to calculate professional attention weights And the sub-problems d and the knowledge base preliminary screening result summary s are input into the template classifier of step S3 to select the optimal prompt word template and retrieval strategy, and then the dynamic retrieval of step S4 is performed to obtain the candidate regulations and intermediate reasoning conclusions of the current round. The evaluation module checks the "professional consistency" and "scene adaptability" of each round result, including checking whether the professional tags of high weight professional and search result are consistent, and whether the applicable scene of search regulation covers or contains the scene of sub task; If the check is passed, the corresponding round result is written into the decomposition path and reasoning path record, and the next round decomposition is entered; If not, backtrack to the decomposition execution module, adjust the decomposition logic, and trigger the search and reasoning again.

7. The multi-round thinking based complex problem splitting and retrieval enhanced power field supervision question and answer method according to claim 1, characterized in that, Step S6, comprising: S61, double dimension confidence calculation (1) reasoning confidence sequences of answers output by the DeepSeek-R1 model , using token probability to compute sequence average log probability , where: where, represents the complete answer sequence generated by the model, each is the tth output token; represents the prefix sequence generated before the tth token is generated , T is the total number of tokens in the answer sequence; x is the text input into the DeepSeek-R1 model, i.e., the sub-problem / scenario description; then it is mapped to the interval [0, 1] through the Sigmoid or normalization function, and the result is , the formula is: (2) search confidence The search confidence comprehensively considers factors including "professional matching degree", "scene matching degree" and "semantic matching degree", and defines the search confidence as , formula: wherein is a professional match degree, is a scene match degree, is an average semantic similarity; , , are respective weights; and + + = 1. (3) global comprehensive confidence The global comprehensive confidence is used to adjust the degree of "more trust reasoning" or "more trust retrieval" according to the business side requirements, expressed as , formula: wherein is the inferred weight; S62, low confidence processing and backtracking mechanism The system presets a threshold of inference confidence With the threshold of retrieval confidence , the input step S61 outputs , When any of the confidences is lower than the threshold, i.e. , a low-confidence processing flow is triggered: Record the current "low confidence result + decomposition record + search log", and execute according to the problem type respectively; If the problem mainly shows that the search recall is insufficient, return to step S4 and adjust the search strategy: switch from single professional mode to multi professional mode, relax scene filtering or increase the number of candidates; If the problem mainly shows that the reasoning is unstable, return to step S5 and re adjust the problem decomposition logic: split the granularity of sub problem, increase the scene or time dimension constraint, and execute multi round reasoning on the new decomposition result.

8. The multi-round thinking based complex problem splitting and retrieval enhanced power field supervision question and answer method according to claim 1, characterized in that, In (3) of Step S61, the inference priority scenario = 0.7, the search priority scenario = 0.

3.

9. The multi-round thinking based complex problem splitting and retrieval enhanced power field supervision question-answer method according to claim 1, characterized in that, Step S7, comprising: S71, iteration number control Introducing a loop counter , recording the iteration number of "dynamic search → multi-round decomposition → confidence check", the maximum iteration number is set to ; the loop control logic is formalized as: ; when , whether the confidence meets the standard or not, the loop is exited and the result processing branch is entered; S72, result processing logic If in Within the round, the inference confidence level is ≥ "and "retrieval confidence ≥ Given the dual conditions of "", the system outputs the following: Final matching regulation set , including each professional corresponding regulation number and original text abstract; Decomposition path record , including professional dimension, nature dimension and scene dimension decomposition order of each round of decomposition; Retrieval path record including the retrieval mode employed, the filter conditions and candidate set changes at each round. If If the double confidence threshold is still not reached, the system will output the current "decomposition record / search record / intermediate results and confidence curves of all previous runs" packed together, and transfer to the manual processing channel; at the same time, the "current decomposition / search record" is output to assist manual fast positioning of the problem.

10. A power site supervision question and answer system based on multi-round thinking-based complex problem splitting and retrieval enhancement, characterized in that, Comprise: Knowledge base and data set construction module, construct power field search enhancement knowledge base, and construct triple data set based on knowledge base; Model training module, based on the constructed data set, train different electrical professional design exclusive prompt word template classifier; Collaborative search module, based on the trained template classifier, dynamically select single professional or multi professional hierarchical search strategy according to the weight difference; Agent construction module, based on the original supervision question description, construct multi round thinking agent; Output module, through multi round thinking iteration and confidence closed loop control, output the final result.

Images