Intelligent operation and maintenance question answering method, system and device based on large model

Abstract

This invention provides an intelligent operation and maintenance question-answering method, system, and device based on a large model. The method includes: performing a hybrid retrieval on a user question based on a preset knowledge base to determine the set of text fragments most relevant to the user question; inputting the user question and the set of text fragments into a large model for decomposition, generating a set of independent sub-questions; identifying the semantic associations and logical relationships between the sub-questions in the sub-question set, generating a causal graph; performing a secondary hybrid retrieval on each sub-question to determine the subset of text fragments most relevant to the sub-question, forming a mapping relationship set; inputting the mapping relationship set and the causal graph together into the large model for correction, determining a corrected set of sub-questions; and inputting the corrected set of sub-questions, the causal graph, and the user question together into the large model for reasoning, generating an operation and maintenance answer that conforms to the topological order of the causal graph. This invention ensures the coherence and reliability of the question decomposition and correction process, thereby improving the accuracy and completeness of complex operation and maintenance question-answering.

Description

Technical Field

[0001] This invention relates to the field of intelligent question-answering technology, and in particular to an intelligent operation and maintenance question-answering method, system and device based on a large model. Background Technology

[0002] With the rapid development of the industrial operations and maintenance (O&M) field, intelligent O&M methods are increasingly becoming important tools for enterprises to improve efficiency and reduce risks. Among these methods, problem diagnosis and troubleshooting often rely on a large amount of historical data in a pre-set O&M knowledge base, including fault descriptions, handling steps, and repair solutions. This data is not only massive but also diverse in format. How to efficiently retrieve and match solutions that match the current O&M scenario has become one of the core challenges facing intelligent O&M question-and-answer methods.

[0003] In recent years, with the rapid development of large language model technology, more and more enterprises have begun to explore intelligent operation and maintenance question-answering systems based on large models. These systems leverage the language understanding and generation capabilities of large language models to assist operation and maintenance work through automated question answering. Among related technologies, mainstream intelligent operation and maintenance question-answering systems mostly employ retrieval-enhanced generation methods, combining retrieval results with large models to generate answers, thereby improving the efficiency and accuracy of question-answering. However, when facing complex operation and maintenance problems, due to the involvement of multi-step reasoning, simple retrieval and generation methods cannot guarantee the logical consistency and reasoning completeness of the answers. This significantly affects the reasoning accuracy and reliability of intelligent operation and maintenance question-answering systems and also reduces the user experience. Summary of the Invention

[0004] This invention provides an intelligent operation and maintenance question-answering method, system, and device based on a large model to solve the technical problem that intelligent operation and maintenance question-answering systems cannot ensure the accuracy and reliability of reasoning when dealing with complex operation and maintenance issues.

[0005] This invention provides an intelligent operation and maintenance question-answering method based on a large model. The method includes: performing a hybrid retrieval on user questions based on a preset knowledge base to determine the set of text fragments most relevant to the user questions; the preset knowledge base includes semantic vectors representing historical maintenance records; the hybrid retrieval includes a semantic retrieval strategy and a keyword inverted index strategy; inputting the user questions and the set of text fragments into a large model for decomposition to generate a set of independent sub-questions; identifying the semantic associations and logical relationships between the sub-questions in the set of sub-questions to generate a causal graph; performing a second hybrid retrieval on each sub-question to determine the subset of text fragments most relevant to the sub-question to form a mapping relationship set; inputting the mapping relationship set and the causal graph into the large model for correction to determine a corrected set of sub-questions; and inputting the corrected set of sub-questions, the causal graph, and the user questions into the large model for reasoning to generate an operation and maintenance answer that conforms to the topological order of the causal graph.

[0006] In one embodiment of the present invention, before performing mixed retrieval of user questions based on a preset knowledge base, the method further includes: extracting text features from historical operation and maintenance records, wherein the text features are represented by a triplet structure of fault description, processing flow and repair plan; preprocessing the text features to obtain preprocessed text features, wherein the preprocessing includes data cleaning and normalization operations; performing semantic transformation on the preprocessed text features to generate semantic vectors, and storing them in the preset knowledge base as vector data.

[0007] In one embodiment of the present invention, a hybrid retrieval is performed on a user question based on a preset knowledge base to determine the set of text fragments most relevant to the user question. This includes: extracting a first keyword and a first query semantic vector from the user question; performing semantic retrieval on candidate text fragments of the first query semantic vector in the preset knowledge base to determine multiple first text fragments; performing keyword retrieval on candidate text fragments of the first keyword in the preset knowledge base to determine multiple second text fragments; and fusing the multiple first text fragments and the multiple second text fragments to determine the set of text fragments most relevant to the user question.

[0008] In one embodiment of the present invention, a user question and a set of text fragments are input into a large model for decomposition to generate a set of sub-questions that are independent of each other. This includes: using the large model to perform semantic parsing on the user question and constraining it based on the knowledge boundary expressed by the set of text fragments to form decomposition rules that include semantic dimensions and query targets. Each sub-question corresponds to only one query target, and the sub-questions are independent of each other, have no logical dependence, and do not contain each other; splitting the user question according to the decomposition rules to generate independent sub-questions that have no logical dependence and do not contain each other, and outputting a question set composed of sub-questions.

[0009] In one embodiment of the present invention, identifying the semantic associations and logical relationships between subproblems in a set of subproblems and generating a causal graph includes: performing semantic recognition on each subproblem in the set of subproblems to determine the query target, constraint condition, and solution target corresponding to each subproblem; determining whether there is a causal association, conditional dependency, or deductive relationship between any two subproblems based on at least one of the query target, constraint condition, and solution target, and determining the semantic associations and logical relationships; and constructing and generating a corresponding causal graph by using each subproblem as a node and the semantic and logical relationships as edges based on the semantic associations and logical relationships.

[0010] In one embodiment of the present invention, a secondary hybrid retrieval is performed on each sub-problem to determine the subset of text fragments most relevant to the sub-problem and form a mapping relationship set. This includes: extracting a second keyword and a second query semantic vector for each sub-problem; performing semantic retrieval on candidate text fragments of the second query semantic vector in a preset knowledge base to determine multiple third text fragments; performing keyword retrieval on candidate text fragments of the second keyword in the preset knowledge base to determine multiple fourth text fragments; fusing the multiple third text fragments and the multiple fourth text fragments to determine the subset of text fragments most relevant to the sub-problem, and associating the subset of text fragments with the sub-problem to form a mapping relationship set.

[0011] In one embodiment of the present invention, the mapping relationship set and the causal graph are jointly input into a large model for correction to determine a sub-problem correction set. This includes: inputting the sub-problems, text fragment subsets, and the causal graph into the large model for detection; if at least one of semantic conflict, dependency error, or logical inconsistency is detected among the sub-problems, it is determined that there are defects among the sub-problems; correcting the defective sub-problems through the causal graph to generate a sub-problem correction set consisting of multiple sub-problems; wherein the correction method includes at least one of the following: adjusting the sub-problem description according to the dependency relationship of the causal graph, updating key entities or conditions based on the text fragment subset, and re-retrieving and updating the text fragment subset after detecting that the semantics have changed after correction.

[0012] In one embodiment of the present invention, the sub-problem correction set, the causal graph, and the user problem are jointly input into a large model for reasoning to generate an operation and maintenance answer that conforms to the topological order of the causal graph. This includes: inputting the sub-problem correction set, the causal graph, and the user problem into a large model, performing multi-step comprehensive reasoning based on the topological order of the causal graph, and generating an operation and maintenance answer that conforms to the topological order of the causal graph.

[0013] This invention provides an intelligent operation and maintenance question-answering system based on a large model, comprising: a hybrid retrieval module, which performs a hybrid retrieval on user questions based on a preset knowledge base to determine the set of text fragments most relevant to the user questions; the preset knowledge base includes semantic vectors representing historical maintenance records; and the hybrid retrieval includes a semantic retrieval strategy and a keyword inverted index strategy; a question decomposition module, which inputs the user questions and the set of text fragments into a large model for decomposition, generating a set of independent sub-questions; a causal graph construction module, which identifies the semantic and logical relationships between the sub-questions in the set of sub-questions and generates a causal graph; a retrieval and correction module, which performs a secondary hybrid retrieval on each sub-question, determines the subset of text fragments most relevant to the sub-question to form a mapping relationship set, and inputs the mapping relationship set and the causal graph into the large model for correction, determining a corrected set of sub-questions; and an operation and maintenance reasoning module, which inputs the corrected set of sub-questions and the user questions into the large model for reasoning, generating operation and maintenance answers that conform to the topological order of the causal graph.

[0014] The present invention provides an electronic device, the electronic device comprising: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the electronic device enables the intelligent operation and maintenance question-and-answer method based on a large model as described in any of the above embodiments.

[0015] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a computer's processor, causes the computer to perform the intelligent operation and maintenance question-and-answer method based on a large model as described in any of the above embodiments.

[0016] The beneficial effects of this invention are as follows: This invention proposes an intelligent operation and maintenance question-answering method, system, and device based on a large model. It performs a hybrid retrieval of user questions based on a pre-set knowledge base to determine the set of text fragments most relevant to the user questions. By integrating semantic retrieval and keyword retrieval, it overcomes the shortcomings of single retrieval, improving recall and precision. The user questions and text fragment sets are input into a large model for decomposition, generating a set of independent sub-questions. This breaks down complex operation and maintenance problems into independent sub-questions, reducing the complexity of single-round reasoning. The semantic associations and logical relationships between sub-questions in the sub-question set are identified, generating a causal graph. The causal graph structurally presents the problem logic, reducing the logical verification cost during reasoning. A secondary hybridization is performed on each sub-question. The process involves retrieving and identifying the subset of text fragments most relevant to the sub-questions to form a mapping set. This mapping set, along with the causal graph, is then input into the main model for refinement. A revised set of sub-questions is determined, and a secondary hybrid retrieval method is used to achieve a refined match between the sub-questions and the updated text fragments. The causal graph constrains the reasoning path, and the revised sub-questions eliminate ambiguity and conflicts. The revised set of sub-questions, the causal graph, and the user's question are then input into the main model for reasoning, generating an operational answer that conforms to the topological order of the causal graph. This approach overcomes the problems of error accumulation in chain decomposition, lack of coherence in independent decomposition, and difficulty in correcting errors. By combining the independence of sub-questions with the constraints of the causal graph and introducing consistency correction and re-retrieval mechanisms, the transparency and robustness of reasoning are improved, thereby enhancing the accuracy and completeness of complex operational question answering. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0018] In the attached diagram:

[0019] Figure 1 A schematic diagram of an exemplary system architecture provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating an intelligent operation and maintenance question-answering method based on a large model provided in one embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the implementation process of the intelligent operation and maintenance question-answering method based on a large model provided in one embodiment of the present invention; Figure 4 This is a block diagram of an intelligent operation and maintenance question-and-answer system based on a large model provided in one embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a computer system of an electronic device provided in one embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of an intelligent operation and maintenance question-and-answer system based on a large model provided in one embodiment of the present invention. Detailed Implementation

[0020] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0021] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. The drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the shape, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0022] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0023] Please see Figure 1 , Figure 1 This is a schematic diagram of an exemplary system architecture provided in an embodiment of the present invention. Figure 1As shown, the system architecture may include a data acquisition device 110 and a computer device 120. The computer device may be at least one of an embedded computer, an industrial computer, a general-purpose computer, or a neural network computer. The data acquisition device 110 includes a function to collect user input and transmit it to the computer device 120 for temperature monitoring.

[0024] For example, computer device 120 performs a hybrid retrieval on the user's question based on a preset knowledge base to determine the set of text fragments most relevant to the user's question. The preset knowledge base includes semantic vectors representing historical maintenance records, and the hybrid retrieval includes a semantic retrieval strategy and a keyword inverted index strategy. The user's question and the set of text fragments are input into a large model for decomposition, generating a set of independent sub-questions. The semantic associations and logical relationships between the sub-questions in the sub-question set are identified, and a causal graph is generated. A second hybrid retrieval is performed on each sub-question to determine the subset of text fragments most relevant to the sub-question, forming a mapping relationship set. The mapping relationship set and the causal graph are input into the large model for correction to determine the sub-question correction set. The sub-question correction set, the causal graph, and the user's question are input into the large model for reasoning to generate an operation and maintenance answer that conforms to the topological order of the causal graph.

[0025] In related technologies, existing operation and maintenance question-answering systems mostly adopt a single retrieval strategy and a direct question-and-answer mode, which have three major technical defects: First, single keyword retrieval is difficult to cover the semantic intent of user questions, and single semantic retrieval is insufficient in matching precise terms, making it difficult to balance the recall and precision of the initial retrieval results; Second, directly inputting complex operation and maintenance questions into large models can easily lead to reasoning confusion, making it impossible to decompose the problem logic and identify the causal relationships within the problem, and the reasoning process lacks structured constraints; Third, a single retrieval result is insufficient to support in-depth reasoning for complex problems, the mapping relationship between sub-questions and knowledge fragments (i.e., text fragments) is not verified, the reasoning conclusions lack causal logic support, and ultimately the operation and maintenance answers are insufficient in accuracy, logic, and interpretability, failing to meet the precise question-and-answer requirements of complex operation and maintenance scenarios.

[0026] To address the aforementioned technical problems, this invention provides an intelligent operation and maintenance question-answering method, system, and device based on a large model. The implementation details of the technical solutions in the embodiments of this invention are described in detail below.

[0027] Please see Figure 2 , Figure 2 This is a flowchart illustrating an intelligent operation and maintenance question-answering method based on a large model provided in one embodiment of the present invention. Figure 2 As shown, in an exemplary embodiment, the intelligent operation and maintenance question-answering method based on a large model includes at least steps S210 to S250, which are described in detail below: Step S210: Perform a hybrid search on the user's question based on the preset knowledge base to determine the set of text fragments most relevant to the user's question. The preset knowledge base includes semantic vectors representing historical maintenance records. The hybrid search includes semantic search strategy and keyword inverted index strategy. For example, historical maintenance records in a pre-defined knowledge base are preprocessed by performing operations such as text segmentation, stop word filtering, and stemming to construct a keyword inverted index. Simultaneously, a pre-trained semantic encoding model (e.g., bidirectional encoder representation, sentence bidirectional encoder representation) is used to convert the historical maintenance records into high-dimensional semantic vectors, forming a semantic vector library. Upon receiving a user question (e.g., a query statement within the operations and maintenance domain), semantic retrieval and keyword inverted index retrieval are initiated. Semantic retrieval encodes the user question into a semantic vector, calculates cosine similarity in the semantic vector library using an approximate nearest neighbor algorithm, and retrieves the Top-K semantically relevant text fragments. Keyword inverted index retrieval segments the user question, locates text fragments containing target keywords based on the inverted index, calculates relevance scores using the BM25 algorithm (a relevance scoring algorithm in the information retrieval domain), and sorts them to retrieve the Top-K keyword-matching text fragments.

[0028] Optionally, in some embodiments, a hybrid retrieval is performed on the user question based on a preset knowledge base to determine the set of text fragments most relevant to the user question, including: Extract the first keyword and the first query semantic vector of the user question; perform semantic retrieval on the candidate text fragments of the first query semantic vector in the preset knowledge base to determine multiple first text fragments; perform keyword retrieval on the candidate text fragments of the first keyword in the preset knowledge base to determine multiple second text fragments; merge the multiple first text fragments and the multiple second text fragments to determine the set of text fragments most relevant to the user question.

[0029] For example, keyword inverted index retrieval and vector similarity are used to match in a preset knowledge base. When matching with keyword inverted index retrieval, it is similar to coarse screening; while when performing semantic retrieval with vectors, it is similar to precise screening. By fusing multiple first text fragments and multiple second text fragments, the fusion method includes deduplication and sorting, and the top-ranked text fragments are output to determine the set of text fragments most relevant to the user's question.

[0030] Through the above methods, semantic retrieval understands the user's true intent based on the query semantic vector, and can capture synonyms, near-synonyms and implicit needs; keyword inverted index retrieval can quickly and effectively lock core words to ensure matching accuracy. The combination of the two avoids the bias of a single retrieval method and significantly improves the accuracy of hitting relevant text fragments. By fusing the dual retrieval results, low-relevance fragments are filtered and high-value text is aggregated. The resulting set of relevant text fragments is more in line with the user's question and ensures the depth of understanding of semantic retrieval intent.

[0031] Step S220: Decompose the user questions and text fragment sets into the large model to generate a set of sub-questions that are independent of each other; For example, the large model can be guided to perform sub-problem decomposition by preset prompts. For instance, the large model can be prompted to take the user's question as the core, combine the knowledge content of the text fragment set, decompose the complex question into multiple sub-problems, and constrain each sub-problem to be independent and without logical overlap, and finally output a structured set of sub-problems.

[0032] Optionally, in some embodiments, the user question and the set of text fragments input into the large model are decomposed to generate a set of sub-questions that are independent of each other, including: The system uses a large model to perform semantic parsing on user questions and constrains them based on the knowledge boundaries expressed by the text fragment set, forming decomposition rules that include semantic dimensions and query targets. Each sub-question corresponds to only one query target, and the sub-questions are independent of each other, have no logical dependence, and do not contain each other. The user questions are split according to the decomposition rules to generate independent sub-questions that have no logical dependence and do not contain each other, and the output is a question set composed of sub-questions.

[0033] For example, the large model is a deep learning model trained on a natural language processing task, possessing semantic understanding, intent recognition, and information extraction capabilities. The large model performs semantic parsing on user questions, extracting core semantic information, latent semantic dimensions, and query targets. This involves obtaining a set of text fragments, which is a collection of structured or unstructured texts storing knowledge content related to the user question. The core knowledge elements of the text fragments are extracted, determining the knowledge scope covered by the text fragment set, i.e., the knowledge boundary. Within this knowledge boundary, the types of query targets and semantic dimensions are defined. Using the knowledge boundary of the text fragment set as constraints, and combining the semantic dimensions of the user question parsed by the large model with the initial query target, decomposition rules are formulated. These rules ensure that each sub-question corresponds to only one unique query target after being split according to regulations, and that each sub-question satisfies the constraints of mutual independence, no logical dependence, and non-inclusion. Simultaneously, the semantic dimension division criteria for the sub-questions and the scope of the query target are determined.

[0034] Based on decomposition rules, user questions are broken down sentence by sentence and semantic unit by semantic unit. According to the semantic dimension division criteria in the decomposition rules, the core semantics of the user question are divided into multiple independent semantic units, each corresponding to a clear semantic dimension. In accordance with the requirement of "each sub-question corresponds to a query target" in the decomposition rules, a unique query target is matched for each semantic unit to ensure that the query target of each sub-question is clear and unique, and does not contain multiple parallel or subordinate query targets. Logical verification is performed on the decomposed sub-questions to determine whether there are logical dependencies (such as the solution of one sub-question depends on the result of another sub-question) or mutual inclusion (such as the semantic scope of one sub-question completely covers another sub-question). If so, adjustments are made according to the decomposition rules until all sub-questions meet the requirements of mutual independence, no logical dependencies, and no mutual inclusion. All adjusted sub-questions are integrated to form a question set.

[0035] By using the above methods, decomposition rules can be scientifically and rationally formulated to provide reliable constraints for splitting user issues and avoid disorderly splitting.

[0036] Step S230: Identify the semantic relationships and logical connections between the sub-problems in the sub-problem set, and generate a cause-effect graph; For example, a set of sub-problems is input into a large model, prompting the model to analyze the semantic content of each sub-problem, identify semantic dependencies, conditional associations, and other relationships between sub-problems, and output semantic association tags for sub-problem pairs. These semantic association tags include, but are not limited to, dependencies, conditions, and derivations. Based on these semantic association tags, the large model infers the causal logic between sub-problems, distinguishes causal relationships, and obtains the causal transmission direction. Using sub-problems as nodes and causal relationships as directed edges, a directed acyclic causal graph is constructed, presenting the causal topology between sub-problems.

[0037] Optionally, in some embodiments, identifying the semantic associations and logical relationships between the sub-problems in the set of sub-problems and generating a causal graph may further include: Semantic recognition is performed on each subproblem in the set of subproblems to determine the query target, constraints and solution target corresponding to each subproblem; Based on at least one of the query target, constraints and solution target, determine whether there is a causal relationship, conditional dependency or deductive relationship between any two sub-problems, and determine the semantic relationship and logical relationship. Based on semantic associations and logical relationships, each sub-problem is treated as a node, and the semantic and logical relationships are treated as edges to construct and generate the corresponding causal graph.

[0038] For example, for each sub-problem, deep semantic parsing is performed using natural language processing techniques; three key semantic elements are extracted from the parsing results: query target, which is the entity, concept, or target object of the information to be obtained that the question intends to point to; constraints, which are various conditions that impose restrictions on the query target or the solution process, including attribute conditions, time conditions, space conditions, comparison conditions, etc.; and solution target, which is the expected form of the answer or the operation to be performed, such as requesting a specific numerical value, a judgment (yes / no), a list, or an explanatory description.

[0039] After obtaining the structured semantic descriptions of all subproblems, pairwise relationship determination is performed on the set of subproblems. The determination process is based on matching according to one or more of the following logical rules: 1) If the solution objective of subproblem A is part of the query objective or constraint of subproblem B, then A and B are determined to have a causal relationship, that is, the result of A is the cause or input of B. 2) If satisfying the constraint of subproblem B requires the answer to subproblem A to be true, then A and B have a conditional dependency relationship, that is, A is a condition of B. 3) If the answers to subproblems A and B can be logically derived through axioms, theorems, domain rules, or common sense, then A and B have a deductive relationship.

[0040] The results of the relation determination are marked as directed edges between subproblem pairs. The type of the edge represents the logical relation. Based on the output of the previous steps, an empty directed graph is initialized. Each subproblem is treated as an independent node in the graph, and the node can carry its corresponding semantic elements (query target, constraint condition, solution target) as attributes. All subproblem pairs determined in step two are traversed. For each relation from subproblem A to subproblem B, a directed edge from node A to node B is added to the graph, and the edge is labeled with the specific type of the relation. Finally, a directed graph is generated that can visually represent the internal logical topology of all problems in the entire set of subproblems, i.e., a causal graph.

[0041] By using the above methods, natural language problems are transformed into machine-processable semantic element triples through semantic recognition. This changes the previous shallow understanding method that relied solely on keyword matching, and enables a refined analysis of the problem's intent, conditions, and expectations, thereby improving the depth and structure of problem understanding. By systematically determining the causal, dependency, and deductive relationships between sub-problems and constructing a causal graph, the previously isolated sub-problems are integrated into a unified logical framework, establishing a logical network between sub-problems and breaking down information silos.

[0042] Step S240: Perform a secondary hybrid retrieval for each sub-problem, determine the subset of text fragments most relevant to the sub-problem to form a mapping relationship set, input the mapping relationship set and the causal graph together into the large model for correction, and determine the sub-problem correction set; For example, for each sub-problem in the sub-problem set, the above hybrid retrieval process is repeated, performing semantic retrieval and keyword inverted index retrieval respectively, and merging them to obtain the subset of text fragments most relevant to each sub-problem, forming a mapping relationship set of "sub-problem - text fragment subset". The mapping relationship set and the causal graph are input into the large model together, prompting the model to combine the topological constraints of the causal graph to verify the rationality of the mapping relationship; if the text fragment subset of the sub-problem conflicts with the causal logic, or if the sub-problem statement is ambiguous or redundant, the model is guided to correct the sub-problem, and finally a corrected sub-problem set is generated.

[0043] Optionally, in some embodiments, a secondary hybrid retrieval is performed on each sub-question to determine the subset of text fragments most relevant to the sub-question, forming a mapping relationship set, including: Extract the second keyword and the second query semantic vector for each sub-question; Semantic retrieval is performed on the candidate text fragments of the second query semantic vector in the preset knowledge base to determine multiple third text fragments; keyword retrieval is performed on the candidate text fragments of the second keyword in the preset knowledge base to determine multiple fourth text fragments; Multiple third and fourth text fragments are merged to determine the subset of text fragments most relevant to the subproblem, and the subset of text fragments is associated with the subproblem to form a mapping relationship set.

[0044] Following the above approach, a mixed search is performed on the sub-questions in the preset knowledge base to obtain the subset of text fragments most relevant to the sub-questions from the perspective of the sub-questions. The subset of text fragments is then associated with the sub-questions to form a mapping relationship set.

[0045] Optionally, in some embodiments, the mapping set and the causal graph are jointly input into the large model for correction to determine the sub-problem correction set, including: Sub-problems, subsets of text fragments, and causal graphs are input into a large model for detection. If at least one of semantic conflict, dependency error, or logical inconsistency is detected among the sub-problems, then a defect is determined to exist among the sub-problems. The defective sub-problems are corrected by using a causal graph to generate a sub-problem correction set consisting of multiple sub-problems. The correction methods include at least one of the following: a. adjusting the sub-problem description according to the dependencies of the causal graph; b. updating key entities or conditions based on a subset of text fragments; c. re-retrieving and updating the subset of text fragments after detecting semantic changes after correction.

[0046] It should be noted that if no defects are detected between subproblems, no correction is required for the subproblems. Conversely, if any of the aforementioned defects are detected between subproblems, if a dependency error occurs between any two subproblems, correction step a is executed; if a semantic conflict occurs between any two subproblems, at least one of correction steps a, b, and c is executed; if a logical inconsistency occurs between any two subproblems, at least one of correction steps b and c is executed.

[0047] By using the above method, not only can defects be detected based on the logical relationships of the cause-effect graph, but sub-problems with obvious logical defects can also be optimized. In this way, by gradually correcting from the local to the global, the accuracy of the sub-problem correction set is improved.

[0048] Step S250: Input the sub-problem correction set, cause-effect graph and user problem into the large model for reasoning to generate an operation and maintenance answer that conforms to the topological order of the cause-effect graph.

[0049] For example, the modified set of sub-problems, the causal graph, and the original user question are input into the large model. The model is prompted to follow the topological order of the causal graph (from cause to effect), and reason about each modified sub-problem in turn. Combining the knowledge content of the corresponding text fragment subset, the answer is deduced step by step. The reasoning results of each sub-problem are integrated to generate an operation and maintenance answer that conforms to causal logic and has a complete structure.

[0050] Optionally, in one embodiment, the sub-problem correction set, the causal graph, and the user problem are jointly input into the large model for reasoning to generate an operational answer that conforms to the topological order of the causal graph, including: The sub-problem correction set, causal graph, and user problem are input into the large model. Based on the topological order of the causal graph, multi-step comprehensive reasoning is performed to generate an operation and maintenance answer that conforms to the topological order of the causal graph.

[0051] For example, by inputting the sub-problem correction set, the causal graph, and the user problem into the large model, during reasoning, not only can multi-dimensional mutual verification be performed from the overall, local, and topological order of the causal graph, but also multi-step individual reasoning can be used to decompose complex problems into multiple independent sub-problems, which greatly improves the accuracy of operation and maintenance reasoning.

[0052] Optionally, based on the above embodiments, that is, before step S210 performs a mixed retrieval of user questions based on a preset knowledge base, the intelligent operation and maintenance question-answering method based on a large model further includes: Step S200 ( Figure 2(Not shown) Extract text features from historical operation and maintenance records. The text features are represented by a triplet structure of fault description, processing flow and repair plan. The text features are preprocessed to obtain preprocessed text features. The preprocessed data includes cleaning and normalization operations. The preprocessed text features are semantically transformed to generate semantic vectors, which are stored in a preset knowledge base as vector data.

[0053] For example, a fault description is a precise depiction of the fault phenomenon, the scenario in which it occurs, the scope of its impact, and its abnormal characteristics. A handling process is a standardized set of steps for handling a fault that has occurred, such as "what to do first, what to do next, who is responsible, and the goal of each step." A repair plan is a specific solution developed based on the handling process to address the root cause of the fault, such as "how to completely resolve the fault, the repair methods, the required resources, and precautions."

[0054] The fault description, processing flow and repair plan are represented by a triple structure to form text features (vectors). Then, the text features are processed by data cleaning (including but not limited to anomalies, missing, redundancy and inconsistencies) and normalization operations to obtain preprocessed text features. Finally, semantic vectors are generated by semantic transformation and stored in a preset knowledge base as vector data.

[0055] Through the above methods, the pre-set knowledge base defines the scope of knowledge for reasoning, provides factual basis for reasoning, and can effectively avoid invalid reasoning that exceeds the scope of knowledge. At the same time, it can supplement the lack of training data for large models, especially scarce and professional domain knowledge, making the reasoning results more in line with the facts. Especially when reasoning about complex problems, which requires a lot of computing power for semantic operations and knowledge retrieval, the pre-set knowledge base can store core knowledge and reasoning relationships in advance, and large models can directly retrieve and call them without repeating the calculations.

[0056] Please see Figure 3 The following is a schematic diagram illustrating the implementation process of the intelligent operation and maintenance question-answering method based on a large model provided in one embodiment of the present invention, which is described in detail below: 1) Construct a pre-defined knowledge base and vectorized representation; extract text information such as fault descriptions, processing steps and repair solutions from historical operation and maintenance records, clean and normalize them, and then use an embedding model to convert the text into semantic vector representations and store them in a vector database to form a pre-defined knowledge base. The pre-defined knowledge base is used to support semantic retrieval and problem reasoning.

[0057] 2) Initial retrieval of the original question; when the user enters an operation and maintenance question... After identifying the user's question, a hybrid retrieval strategy is employed, combining vector similarity retrieval with keyword inverted indexing to obtain several text fragments most relevant to the question. .

[0058] 3) Decompose the problem; break down the operation and maintenance problem. and several related text fragments retrieved. Input a large model (i.e., a large language model) and generate multiple sets of independent sub-problems. .

[0059] 4) Construct a causal graph; After obtaining the sub-problems, use the large model to determine the semantic and logical relationships between the sub-problems, identify their causal relationships, and construct a causal graph accordingly; In the causal graph, nodes represent sub-problems, and edges represent causal relationships, thereby explicitly depicting the logical paths between sub-problems and providing a structured basis for consistency checks and reasoning corrections.

[0060] 5) Knowledge retrieval at the sub-problem level; for each sub-problem It independently performs hybrid retrieval, extracting the most relevant text fragments from a pre-defined knowledge base. Establish mapping relationships ( , This forms a set of mapping relationships, providing targeted knowledge support for each sub-problem.

[0061] 6) Sub-question correction and re-retrieval based on cause-effect graphs; This involves refining and re-retrieval all sub-questions, their corresponding search results, and cause-effect graphs. Figure 1 The data is then input into a large language model to determine whether there are semantic conflicts, dependency errors, or logical inconsistencies among the sub-problems. If defects are detected (e.g., conflicts), the sub-problems of the conflicting nodes are corrected based on the causal graph relationships and retrieval evidence, generating a corrected set of sub-problems. The correction process includes at least one of the following: adjusting the sub-question description based on causal dependencies; updating key entities or conditions based on search results; if the semantics change significantly after correction, triggering a hybrid search again to update the corresponding text fragments. .

[0062] 7) Generate global comprehensive reasoning and answers; input the corrected sub-problems, updated knowledge paragraphs, cause-effect graph structure and operation and maintenance questions into the large language model, perform multi-step comprehensive reasoning according to the topological order of the cause-effect graph, and generate the final operation and maintenance answer that conforms to the logical chain; the operation and maintenance answer not only includes the direct answer, but can also output the reasoning path description, improving the interpretability of operation and maintenance Q&A.

[0063] Through the above methods, the intelligent operation and maintenance question-answering system based on a large model performs a hybrid retrieval on user questions based on a pre-set knowledge base to determine the set of text fragments most relevant to the user questions. By integrating semantic retrieval and keyword retrieval, it compensates for the shortcomings of single retrieval and improves recall and precision. The user questions and text fragment sets are input into the large model for decomposition, generating a set of independent sub-questions. This breaks down complex operation and maintenance problems into independent sub-questions, reducing the complexity of single-round reasoning. The semantic associations and logical relationships between the sub-questions in the sub-question set are identified, generating a causal graph. The causal graph presents the problem logic in a structured way, reducing the logical verification cost in the reasoning process. A second hybrid retrieval is performed on each sub-question to determine the most relevant text fragments. A subset of relevant text fragments forms a mapping relationship set. This mapping relationship set, along with the causal graph, is input into a large model for correction, determining a sub-problem correction set. A secondary hybrid retrieval achieves refined matching between sub-problems and updated text fragments. The causal graph constrains the reasoning path, correcting sub-problems to eliminate ambiguity and conflict. The sub-problem correction set, the causal graph, and the user's question are input into the large model for reasoning, generating an operation and maintenance answer that conforms to the topological order of the causal graph. This invention overcomes the problems of error accumulation in chain decomposition, lack of coherence in independent decomposition, and difficulty in correcting errors. By combining the independence of sub-problems with causal graph constraints and introducing consistency correction and re-retrieval mechanisms, the transparency and robustness of reasoning are improved, thereby enhancing the accuracy and completeness of complex operation and maintenance question answering.

[0064] Please see Figure 4 , Figure 4 This is a block diagram of an intelligent operation and maintenance question-and-answer system based on a large model provided in one embodiment of the present invention. This device can be applied to... Figure 1 The implementation environment shown is specifically configured in computer device 120. This device can also be applied to other exemplary implementation environments and specifically configured in other devices. This embodiment does not limit the implementation environment to which the device is applicable.

[0065] like Figure 4 As shown, a large-model-based intelligent operation and maintenance question-and-answer system 400 according to an embodiment of the present invention includes: The hybrid retrieval module 410 performs a hybrid retrieval on the user's question based on a preset knowledge base to determine the set of text fragments most relevant to the user's question. The preset knowledge base includes semantic vectors representing historical maintenance records, and the hybrid retrieval includes semantic retrieval strategies and keyword inverted index strategies. The problem decomposition module 420 is used to decompose the user's problem and the set of text fragments into the large model, generating a set of sub-problems that are independent of each other. The causal graph construction module 430 is used to identify the semantic and logical relationships between sub-problems in the set of sub-problems and generate a causal graph. The retrieval and correction module 440 is used to perform secondary mixed retrieval on each sub-problem, determine the subset of text fragments most relevant to the sub-problem to form a mapping relationship set, and input the mapping relationship set and the causal graph into the large model for correction to determine the sub-problem correction set. The operation and maintenance reasoning module 450 is used to input the sub-problem correction set, cause-effect graph and user problem into the large model for reasoning, and generate operation and maintenance answers that conform to the topological order of the cause-effect graph.

[0066] Based on the above embodiments, before the hybrid retrieval module 410, the intelligent operation and maintenance question-and-answer system further includes: a preset knowledge base ( Figure 4 (Not shown) is used to extract text features from historical operation and maintenance records. The text features are represented according to the triple structure of fault description, processing flow and repair plan. The text features are preprocessed to obtain preprocessed text features. The preprocessed data includes cleaning and normalization operations. The preprocessed text features are semantically transformed to generate semantic vectors, and stored in vector data to form a preset knowledge base.

[0067] For example, such as Figure 6 As shown, historical operation and maintenance data is processed through a preset knowledge base construction and vectorization module to obtain a preset knowledge base. The original operation and maintenance problem is then combined with the preset knowledge base and retrieved using a hybrid retrieval module. The retrieval results, along with the original operation and maintenance problem, are input into a problem decomposition module to break down the problem into sub-problems. A causal graph is then constructed using a causal graph construction module, providing a basis for the retrieval and correction modules to correct the sub-problems. Finally, the corrected sub-problems and the causal graph are input into a larger model for comprehensive reasoning, outputting an operation and maintenance answer that conforms to the topological order of the causal graph.

[0068] It should be noted that the intelligent operation and maintenance question-and-answer system based on a large model provided in the above embodiments and the intelligent operation and maintenance question-and-answer method based on a large model provided in the above embodiments belong to the same concept. The specific ways in which each module and unit performs operations have been described in detail in the method embodiments, and will not be repeated here. In practical applications, the intelligent operation and maintenance question-and-answer system based on a large model provided in the above embodiments can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above, and this is not a limitation here.

[0069] Embodiments of the present invention also provide an electronic device, including: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the electronic device enables the intelligent operation and maintenance question-and-answer method based on a large model provided in the above embodiments.

[0070] Please see Figure 5 , Figure 5 This is a schematic diagram of the structure of a computer system for an electronic device provided in one embodiment of the present invention. Figure 5 The computer system 500 of the electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of the present invention.

[0071] like Figure 5 As shown, the computer system 500 includes a central processing unit 501, which can perform various appropriate actions and processes based on a program stored in the read-only memory 502 or a program loaded from the storage section 508 into the random access memory 503, such as performing the methods described in the above embodiments. The random access memory 503 also stores various programs and data required for system operation. The central processing unit 501, the read-only memory 502, and the random access memory 503 are interconnected via a bus 504. An input / output interface 505 is also connected to the bus 504.

[0072] The following components are connected to the input / output interface 505: an input section 506 including a keyboard, mouse, etc.; an output section 507 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 508 including a hard disk, etc.; and a communication section 509 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 509 performs communication processing via a network such as the Internet. A drive 510 is also connected to the input / output interface 505 as needed. A removable medium 511, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 510 as needed so that computer programs read from it can be installed into the storage section 508 as needed.

[0073] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing computer programs for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 509, and / or installed from removable medium 511. When the computer program is executed by central processing unit 501, it performs various functions defined in the system of the present invention.

[0074] In the above embodiments, unless otherwise specified, the use of ordinal numbers such as "first" and "second" to describe common objects only indicates that they refer to different instances of the same object, rather than indicating that the objects being described must be in a given order, whether temporally, spatially, sequentially, or in any other way.

[0075] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A smart operation and maintenance question-answering method based on a large model, characterized in that, The method includes: A hybrid retrieval is performed on the user's question based on a preset knowledge base to determine the set of text fragments most relevant to the user's question. The preset knowledge base includes semantic vectors representing historical maintenance records, and the hybrid retrieval includes a semantic retrieval strategy and a keyword inverted index strategy. The user question and the set of text fragments are input into a large model and decomposed to generate a set of sub-questions that are independent of each other. Identify the semantic associations and logical relationships between the sub-problems in the set of sub-problems, and generate a cause-effect graph; For each sub-problem, a secondary hybrid retrieval is performed to determine the subset of text fragments most relevant to the sub-problem, forming a mapping relationship set. The mapping relationship set and the causal graph are then input into the large model for correction to determine the sub-problem correction set. The modified set of sub-problems, the causal graph, and the user problem are input into the large model for reasoning to generate an operation and maintenance answer that conforms to the topological order of the causal graph.

2. The intelligent operation and maintenance question-answering method based on a large model according to claim 1, characterized in that, Before performing a mixed retrieval of user questions based on a preset knowledge base, the process also includes: Extract text features from historical operation and maintenance records. The text features are represented by a triplet structure of fault description, processing flow and repair plan. The text features are preprocessed to obtain preprocessed text features. The preprocessed data includes cleaning and normalization operations. The preprocessed text features are semantically transformed to generate semantic vectors, which are then stored in the preset knowledge base as vector data.

3. The intelligent operation and maintenance question-answering method based on a large model according to claim 1, characterized in that, The step of performing a mixed search on the user's question based on a preset knowledge base to determine the set of text fragments most relevant to the user's question includes: Extract the first keyword and the first query semantic vector of the user's question; Semantic retrieval is performed on the candidate text fragments of the first query semantic vector in the preset knowledge base to determine multiple first text fragments; keyword retrieval is performed on the candidate text fragments of the first keyword in the preset knowledge base to determine multiple second text fragments; Multiple first text fragments and multiple second text fragments are merged to determine the set of text fragments most relevant to the user's question.

4. The intelligent operation and maintenance question-answering method based on a large model according to claim 1, characterized in that, The step of decomposing the user question and the text fragment set into the large model to generate a set of independent sub-questions includes: The large model is used to perform semantic parsing on the user's question, and constraints are imposed based on the knowledge boundary expressed by the text fragment set to form decomposition rules that include semantic dimensions and query targets. Each sub-question corresponds to only one query target, and the sub-questions are independent of each other, have no logical dependence, and do not contain each other. The user problem is split according to the decomposition rules to generate independent sub-problems that are not logically dependent on each other and do not contain each other, and a problem set consisting of the sub-problems is output.

5. The intelligent operation and maintenance question-answering method based on a large model according to claim 1, characterized in that, The step of identifying the semantic associations and logical relationships between the sub-problems in the set of sub-problems and generating a causal graph includes: Semantic recognition is performed on each subproblem in the set of subproblems to determine the query target, constraints and solution target corresponding to each subproblem; Based on at least one of the query target, the constraint conditions and the solution target, determine whether there is a causal relationship, conditional dependency or inference relationship between any two sub-problems, and determine the semantic relationship and logical relationship. Based on the semantic associations and logical relationships, each sub-problem is used as a node and the semantic and logical relationships are used as edges to construct and generate the corresponding causal graph.

6. The intelligent operation and maintenance question-answering method based on a large model according to any one of claims 1-5, characterized in that, The step of performing a secondary hybrid retrieval for each sub-question to determine the subset of text fragments most relevant to the sub-question, forming a mapping relationship set, includes: Extract the second keyword and the second query semantic vector for each sub-question; Semantic retrieval is performed on the candidate text fragments of the second query semantic vector in the preset knowledge base to determine multiple third text fragments; keyword retrieval is performed on the candidate text fragments of the second keyword in the preset knowledge base to determine multiple fourth text fragments; Multiple third text fragments and multiple fourth text fragments are merged to determine the subset of text fragments most relevant to the sub-problem, and the subset of text fragments is associated with the sub-problem to form a mapping relationship set.

7. The intelligent operation and maintenance question-answering method based on a large model according to any one of claims 1-5, characterized in that, The step of inputting the mapping set and the causal graph into the large model for correction, and determining the sub-problem correction set, includes: The sub-problems, the subset of text fragments, and the causal graph are input into the large model for detection. If at least one of semantic conflict, dependency error, or logical inconsistency is detected among the sub-problems, it is determined that there are defects among the sub-problems. The defective sub-problems are corrected using the causal graph to generate a sub-problem correction set consisting of multiple sub-problems; wherein the correction method includes at least one of the following: adjusting the sub-problem description according to the dependency relationship of the causal graph, updating key entities or conditions based on the text fragment subset, and re-retrieving and updating the text fragment subset after detecting that the semantics have changed after correction.

8. The intelligent operation and maintenance question-answering method based on a large model according to any one of claims 1-5, characterized in that, The process of inputting the modified sub-problem set, the causal graph, and the user problem into a large model for reasoning to generate an operational answer that conforms to the topological order of the causal graph includes: The sub-problem correction set, the causal graph, and the user problem are input into the large model. Multi-step comprehensive reasoning is performed based on the topological order of the causal graph to generate an operation and maintenance answer that conforms to the topological order of the causal graph.

9. An intelligent operation and maintenance question-and-answer system based on a large model, characterized in that, The system includes: The hybrid retrieval module performs a hybrid retrieval on the user's question based on a preset knowledge base to determine the set of text fragments most relevant to the user's question. The preset knowledge base includes semantic vectors representing historical maintenance records, and the hybrid retrieval includes a semantic retrieval strategy and a keyword inverted index strategy. The problem decomposition module is used to decompose the user problem and the set of text fragments into the large model, generating a set of sub-problems that are independent of each other. The causality graph construction module is used to identify the semantic and logical relationships between sub-problems in the set of sub-problems and generate a causality graph. The retrieval and correction module is used to perform secondary mixed retrieval on each sub-problem, determine the subset of text fragments most relevant to the sub-problem to form a mapping relationship set, and input the mapping relationship set and the causal graph into the large model for correction to determine the sub-problem correction set. The operation and maintenance reasoning module is used to input the sub-problem correction set and the user problem into the large model for reasoning, and generate operation and maintenance answers that conform to the topological order of the causal graph.

10. An electronic device, characterized in that, The electronic device includes: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the electronic device to implement the intelligent operation and maintenance question-and-answer method based on a large model as described in any one of claims 1 to 8.

Images