Intelligent operation and maintenance disposal behavior monitoring method based on artificial intelligence

Abstract

The invention discloses an intelligent operation and maintenance disposal behavior monitoring method based on artificial intelligence, and the method comprises the steps: firstly, carrying out the judicial evaluation of the value of an initial result through preliminary retrieval and content validity judgment, and avoiding unnecessary query rewriting; only if necessary, the screened high-quality key text fragments are utilized to carry out conditional and constrained query expansion on the original alarm problem. Through the controlled expansion mode, the semantic deviation risk caused by unconstrained rewriting of a general model is avoided from the source, and it is ensured that secondary retrieval can more accurately focus on the core intention of original warning. Therefore, key information with higher signal-to-noise ratio and stronger correlation can be provided for subsequent scheme generation, and the accuracy and practicability of the finally generated disposal scheme are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of intelligent operation and maintenance, and more specifically, to a method for monitoring intelligent operation and maintenance behavior based on artificial intelligence. Background Technology

[0002] In today's rapidly evolving digital and intelligent era, the stable operation of critical infrastructure such as power information systems is crucial for ensuring business continuity. However, with the introduction of massive amounts of information devices and systems, the resulting monitoring alarm information and historical operation and maintenance data are growing explosively. Traditional operation and maintenance models struggle to provide rapid and accurate anomaly detection and handling suggestions in the face of such a large volume of data and frequent alarms. This forces operation and maintenance personnel to manually sift through various tools, remaining in a reactive work mode, which seriously affects the safe and stable operation of information systems. Therefore, there is an urgent need to build an intelligent operation and maintenance behavior monitoring solution to automatically and efficiently generate fault handling plans.

[0003] However, existing intelligent operation and maintenance solutions, especially those that utilize artificial intelligence to generate handling plans, generally face some significant technical bottlenecks. A major problem is that during query rewriting to improve retrieval recall, relying on unconstrained general models can easily lead to semantic drift. This manifests as the rewritten query deviating from the specific domain and true intent of the original alert problem; for example, a specific business module failure might be mistakenly written as a generalized software outage, resulting in decreased accuracy of secondary search results, or even contradicting the original intention of solving the problem. Therefore, an optimized AI-based intelligent operation and maintenance behavior monitoring method is needed. Summary of the Invention

[0004] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide an AI-based intelligent operation and maintenance behavior monitoring method. First, it carefully evaluates the value of initial results through preliminary retrieval and content validity judgment, avoiding unnecessary query rewriting. Only when necessary, it utilizes selected high-quality key text fragments to conditionally and constrainedly expand the original alarm issue. This controlled expansion method fundamentally avoids the semantic deviation risk caused by unconstrained rewriting of general models, ensuring that secondary retrieval can more accurately focus on the core intent of the original alarm. This provides key information with a higher signal-to-noise ratio and stronger relevance for subsequent solution generation, significantly improving the accuracy and practicality of the final generated handling solution.

[0005] According to one aspect of this application, an artificial intelligence-based intelligent operation and maintenance behavior monitoring method is provided, comprising: Obtain the original alarm issue; The original alarm issues are subjected to mixed retrieval and initial sorting to obtain a sorted list of documents; Based on the original alarm issue, the sorted document list is subjected to content validity judgment to obtain validity indicators and key text fragments; In response to the validity flag being true, the original alarm issue is expanded by conditional query based on key text fragments to obtain the expanded query; The expanded query is subjected to a secondary search and the search results are filtered to obtain the final key text fragments; Based on the final key text fragments and the original alarm issues, a handling plan is generated.

[0006] Compared to existing technologies, this application provides an AI-based intelligent operation and maintenance behavior monitoring method. First, it carefully evaluates the value of initial results through preliminary retrieval and content validity judgment, avoiding unnecessary query rewriting. Only when necessary, it utilizes selected high-quality key text fragments to conditionally and constrainedly expand the original alarm issue. This controlled expansion method fundamentally avoids the semantic deviation risk caused by unconstrained rewriting of general models, ensuring that secondary retrieval can more accurately focus on the core intent of the original alarm. This provides key information with a higher signal-to-noise ratio and stronger relevance for subsequent solution generation, significantly improving the accuracy and practicality of the final generated handling solution. Attached Figure Description

[0007] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0008] Figure 1 This is a flowchart of an AI-based intelligent operation and maintenance behavior monitoring method according to an embodiment of this application; Figure 2 This is a schematic diagram of data flow in the AI-based intelligent operation and maintenance behavior monitoring method according to an embodiment of this application; Figure 3 This is a flowchart of an AI-based intelligent operation and maintenance behavior monitoring method according to an embodiment of this application. Detailed Implementation

[0009] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0010] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0011] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0012] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0013] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0014] The technical solution of this application proposes an intelligent operation and maintenance behavior monitoring method based on artificial intelligence. Figure 1 This is a flowchart of an AI-based intelligent operation and maintenance behavior monitoring method according to an embodiment of this application. Figure 2 This is a system architecture diagram of an AI-based intelligent operation and maintenance behavior monitoring method according to an embodiment of this application. Figure 1 and Figure 2 As shown, the AI-based intelligent operation and maintenance behavior monitoring method according to an embodiment of this application includes the following steps: S1, obtaining the original alarm issue; S2, performing mixed retrieval and initial sorting on the original alarm issue to obtain a sorted document list; S3, based on the original alarm issue, performing content validity judgment on the sorted document list to obtain a validity flag and key text fragments; S4, in response to the validity flag being true, performing conditional query expansion on the original alarm issue based on the key text fragments to obtain an expanded query; S5, performing secondary retrieval and retrieval result filtering on the expanded query to obtain the final key text fragments; S6, generating a handling plan based on the final key text fragments and the original alarm issue.

[0015] Specifically, S1 involves acquiring the original alarm issue. It should be understood that anomaly prediction and handling in power information systems are crucial for the safe and stable operation of power systems. Through power business anomaly analysis and prediction, alarms can be issued before faults occur, and faults can be located when they occur, giving maintenance personnel valuable time to handle the situation, reducing the probability of faults, minimizing the scope of their impact, ensuring the safe and stable operation of power information systems, and preventing level 8 or higher information system operation events. However, the existing operation and maintenance (O&M) model suffers from a lack of effective summarization of massive monitoring alarm information and historical O&M knowledge, failing to provide O&M personnel with direct handling suggestions. When a fault occurs, O&M personnel still rely on emergency meetings to locate the fault, leaving limited O&M resources bogged down in constantly making judgments among various O&M monitoring tools. O&M personnel remain in a passive work mode, which is detrimental to the safe and stable operation of the power information system. Therefore, to effectively improve the O&M level of the power information system and realize the generation of business anomaly prediction and handling solutions, the technical solution of this application first obtains the original alarm issues as the basic input for all subsequent analysis, judgment, retrieval, and solution generation. The original alarm issues refer to initial, unprocessed alarm information automatically generated or detected by various data sources such as monitoring tools, log systems, or business links within the power information system, indicating the existence of anomalies, faults, or potential problems in the system. This information is usually presented in text form and may include descriptions of equipment status, error codes, abnormal performance indicators, service interruptions, etc.

[0016] In practice, real-time alarm notifications, log files, performance monitoring data, etc., from various power information system components can be monitored or read through system integration and data acquisition interfaces.

[0017] Specifically, in step S2, a hybrid retrieval and initial sorting are performed on the original alarm issues to obtain a sorted document list. It should be understood that in new power system scenarios, it is necessary to find specific keywords or phrases, requiring the use of sparse vector retrieval during the retrieval phase. A single retrieval method cannot fully meet the needs of alarm issue handling in power operation and maintenance scenarios, resulting in problems such as insufficient semantic understanding or inadequate keyword matching. In the technical solution of this application, a hybrid retrieval method is adopted, which can fully leverage the advantages of dense retrieval in complex semantic relationships while utilizing the efficiency of sparse retrieval to effectively achieve high-precision and high-efficiency retrieval tasks. Through hybrid retrieval, the problem of massive alarm data and information overload caused by the dynamically changing operating environment and constantly updated fault modes of the power system can be overcome, thereby improving alarm response efficiency, reducing the probability of misjudgment and missed judgment, and ensuring the effectiveness and accuracy of subsequent handling solutions.

[0018] In practice, the original alarm question is first represented by query vectorization to obtain sparse and dense query vectors. During this process, the system first performs dual encoding on the user-inputted original alarm question. The generation of sparse query vectors typically relies on statistical methods such as keywords or word frequencies. In the specific example of this application, the construction of the sparse retrieval vector library will employ a combination of the BM25 algorithm and keyword retrieval. That is, by analyzing the original alarm question, keywords are extracted, and sparse vectors are calculated using the BM25 algorithm to reflect the importance of specific words in the alarm question. Simultaneously, the generation of dense query vectors utilizes an advanced pre-trained language model. In the specific example of this application, the user question is encoded using the BGEEmbedding model to capture the semantic information of the alarm question, thereby obtaining dense query vectors. Next, the sparse query vector and the dense query vector are input into the retrieval module of the knowledge base index to obtain a list of candidate documents. During this process, the system uses the two types of query vectors generated in the previous step to perform a mixed retrieval of the pre-built knowledge base. In the embodiments of this application, firstly, for the sparse query vector, given an alarm question... Containing words ,… A text fragment The BM25 score is expressed by the formula: in yes In the text fragment Frequency of occurrence in It is the length of the text fragment 𝐷. This represents the average length of the set of text fragments. and This is called a hyperparameter. It is each word The inverse document frequency (IVF) is calculated using the following formula: in, This indicates the total number of text fragments in the set. Indicates inclusion The number of documents; Meanwhile, regarding dense query vectors, user issues The relevant score for each text segment was obtained using BM25 retrieval. Furthermore, user questions are encoded into vectors using a sentence vector model. ,Then Vectors corresponding to all text fragments Calculate the relevant fractions using the inner product. ; Then, the sparse retrieval score and the dense retrieval score are fused to calculate a mixed score. Specifically, this process is expressed by the formula: The system then sorts all document fragments based on this mixed score and selects the 50 text fragments with the highest scores as a candidate document list. Furthermore, the original alarm question and candidate document list are semantically reordered based on a cross-encoder to obtain the sorted document list. It should be understood that in the above retrieval process, the algorithm focuses more on co-occurring words and has poor retrieval ability for semantic relevance. Therefore, to further improve the accuracy of the sorting, in the embodiments of this application, a ROM semantic relevance model is used to reorder the retrieved text fragments. The ROM semantic relevance model acts as a cross-encoder, capable of deeply understanding the semantic relationship between the original alarm question and each candidate text fragment. Specifically, the question and 50 text fragments are input into the semantic relevance model respectively, relevance scores are calculated, and then the 5 text fragments with the highest scores are selected. In this way, not only can the ability of dense vectors to capture complex semantic relationships and reflect the deeper meaning of the text be utilized, but sparse vectors can also provide clear feature representations, achieving comprehensive retrieval performance, greatly improving retrieval accuracy, avoiding the problem of massive alarm data and information overload caused by the dynamically changing operating environment and constantly updated fault modes of the power system, improving alarm response efficiency, and reducing the probability of false positives and false negatives.

[0019] Specifically, in step S3, based on the original alarm problem, the sorted document list is evaluated for content validity to obtain validity markers and key text fragments. It should be understood that the core technical reason for the defects in the validity judgment mechanism in the above embodiments lies in the singularity of its judgment dimensions and the flattening of relationships. This mechanism essentially simplifies a complex validity identification task and delegates it to a large language model (LLM) to perform a macroscopic, general semantic-based black-box judgment. This method fundamentally ignores the inherent, profound, and asymmetric special relationship between alarm queries and effective handling documents in professional fields such as Intelligent Operations and Maintenance (AIOps), namely the "causal-handling" logical chain relationship. Specifically, this mechanism fails to examine whether the document reveals the root cause of the alarm, nor does it consider the procedural and operational steps necessary for an effective handling solution. Therefore, a document that merely lists and explains error codes, or a purely theoretical analysis article, may be incorrectly judged as valid due to superficial keyword overlap, but its application value is negligible for operation and maintenance scenarios that urgently require recovery operations. Furthermore, this qualitative yes / no judgment lacks a quantifiable structural metric, making it difficult to optimize the threshold for subsequent decisions and weakening the interpretability and stability of the entire mechanism. It reduces a deep structural relationship matching problem to a simple surface semantic similarity judgment, which is its fundamental technical weakness. To address these shortcomings, a causal-disposition structure fitting measurement mechanism based on information divergence is proposed. This involves transforming the macroscopic, qualitative validity judgment into a precise, quantitative, and highly interpretable structured fitting measurement process to ensure the system can accurately distinguish between solution documents containing only surface-related information and those with genuine operational guidance value.

[0020] In practice, the system first performs a structured conceptual decomposition of the original alarm issue and the sorted document list to obtain the problem representation text and the decomposed document set. That is, to overcome the flattened relationship judgment defect in the original mechanism, the unstructured document text is explicitly decomposed into its inherent logical components—namely, cause analysis and handling steps—laying the foundation for subsequent in-depth structured analysis. Specifically, the system first directly defines the received original alarm issue as the problem representation text; then, the system traverses all the initially sorted candidate documents. For each document, a language model specifically designed to identify semantic fragments such as causal arguments and operational procedures in technical documents is used to scan and segment the entire document. During this process, all statements or paragraphs describing the root cause of the problem and analyzing its causes are extracted and integrated into an independent cause analysis text; while all statements or paragraphs containing operational instructions, execution order, script commands, etc., are extracted and integrated into a handling step text. In this way, each original, unstructured candidate document can be transformed into a structured data unit containing the original alarm, cause analysis text, and handling steps text, thereby enabling a precise quantitative evaluation of the document's internal logical structure. Next, the structural fit between the problem representation text and each document in the decomposed document set is calculated to obtain the structural fit score set for each document. This step aims to overcome the lack of quantitative metrics in the original mechanism and introduce a rigorous mathematical method to quantitatively evaluate the structural rationality of each document. Specifically, first, an information distribution that an ideal solution should possess is defined, that is, an ideal disposal document should have a balanced and highly focused content on both cause analysis and disposal steps. Based on this, an ideal two-dimensional discrete probability distribution P is established; subsequently, for each structured decomposed document, the system calculates the semantic similarity between its cause analysis text and disposal step text and the original alarm, and obtains the actual information distribution of the document through normalization; finally, the Kullback-Leibler (KL) divergence is introduced to calculate the difference between the actual information distribution and the ideal information distribution. This process is expressed by the formula: in, Represents the KL divergence value. Represents the ideal information distribution. This represents the actual information distribution of the document. Representing random variables in the information distribution, here we have two categories: cause and steps. P(x) represents the probability of variable x under this distribution. This step quantifies, at the information theory level, the degree of deviation of the causal-disposition structure of a real document from a pre-set perfect solution template in terms of information distribution. The KL divergence value here is not a simple statistical distance, but rather measures the degree of information loss or unexpected surprise in the document structure. The smaller the divergence value, the closer the document's organizational structure is to the ideal, that is, the more it conforms to a standard solution paradigm; To transform the divergence values, which represent differences, into a more intuitive goodness-of-fit score, a negative exponential function is used for mapping: in, The final structure fit score represents document D_i.

[0021] This allows for the generation of a precise, quantifiable structure fit score for each candidate document, ranging from (0,1). This score not only reflects the content's matching degree but, more importantly, its structural integrity as a solution, thus providing a reliable quantitative basis for subsequent decision-making. Furthermore, based on the structure fit scores of each document, the decomposed document set is subjected to validity judgment and fragment extraction based on aggregate fit to obtain validity markers and key text fragments. It should be understood that simply selecting the best document may overlook valid information in other suboptimal documents; therefore, a more robust mechanism is needed to make the final judgment and extract the most critical information. Specifically, in this process, the first step is to determine the validity marker. This step first extracts the maximum structure set score from the structure fit scores of each document; if the maximum structure set score is greater than a preset validity threshold, the validity marker is true; otherwise, it is false. This Boolean validity marker will determine whether to proceed to the subsequent query rewriting process. The second step is to extract key text fragments: If the validity flag is true, the system does not simply select the document with the highest score, but first normalizes the set of structure fit scores for each document based on the Softmax function to obtain the weight distribution; this process is expressed by the formula: in, This represents the weight of document i. This represents the set of structure fit scores for all documents. and Let i and j represent the scores of document i and document j, respectively, and N be the total number of candidate documents. This step borrows the attention mechanism from neural networks, widening the score gap through nonlinear transformation. This ensures that the document with the most reasonable structure receives the majority of attention weights, while not completely ignoring other suboptimal documents that may contain supplementary information. This makes the final information extraction a weighted and focused insight into the entire retrieval result. Then, based on this weight distribution, the system extracts at least one document with the highest score from the decomposed document set. Furthermore, it extracts the cause analysis text and the action step text from at least one document and merges them to obtain the key text fragment. In this way, it ensures that the key information in the final output not only comes from the document with the most reasonable structure, but also represents a weighted synthesis of the information value of all candidate documents, thus providing the highest quality and clearest contextual information for the subsequent large model generation stage.

[0022] This optimization mechanism transforms the macroscopic, qualitative effectiveness judgments of the original mechanism into a precise, quantitative, and highly interpretable structured fitting measurement process. Simultaneously, by explicitly modeling and measuring the core logical chain of causal-disposal in the intelligent operation and maintenance scenario, it ensures the system can accurately distinguish between solution documents containing only surface-level information and those with genuine operational guidance value. Furthermore, by introducing mathematical tools such as information divergence and the Softmax function, this mechanism not only provides theoretical support for the reliability of the judgments but also makes the decision-making process more robust and refined. Ultimately, by providing the generative model with high-quality, high signal-to-noise ratio key text fragments that have undergone structured verification and weighted fusion, this mechanism significantly reduces the risk of content illusion generated by large models, thereby ensuring that the generated alarm handling solutions possess stronger practicality, accuracy, and security, achieving a crucial leap from semantic relevance to structural usability.

[0023] Specifically, in step S4, in response to the validity flag being true, the original alarm question is conditionally expanded based on key text fragments to obtain an expanded query. It should be understood that query rewriting is performed by default before retrieval, and the rewritten query is prone to semantic drift, leading to poor retrieval results. In other words, relying solely on the original alarm question or query rewriting may cause the retrieval results to deviate from the user's true intent. In the technical solution of this application, content validity judgment is performed through text filters to initiate query rewriting based on synonym expansion, which helps alleviate the problem of user intent deviation caused by unnecessary query rewriting. This means that query expansion is only performed when the system confirms that the content validity judgment has been passed. At this time, a true validity flag indicates that the system has found a structurally sound potential solution containing a "cause-effect" logical chain. These key text fragments provide valuable contextual information and more professional domain terminology for the alarm question. Expanding the original alarm question based on this high-quality contextual information can avoid semantic drift caused by blind expansion, enabling the expanded query to more accurately capture the deeper meaning of the alarm question and its potential solutions, thereby further improving the accuracy and recall of secondary retrieval.

[0024] In practice, the original alarm problem is first input into a domain-specific named entity recognition (NER) model fine-tuned to obtain a set of domain entities. During this process, the system first performs named entity recognition (NER) on the original alarm problem. Unlike general NER models, this uses a domain-specific NER model, meaning it has been trained and optimized for the specific domain of power operation and maintenance, enabling it to more accurately identify specialized entities within that domain, such as equipment models (e.g., "main transformer"), system components (e.g., "database connection pool"), and fault phenomena (e.g., "abnormal temperature rise," "slow service response"). The model's input is text containing the original alarm information, and its output is a set of categorized domain entities. For example, if the original alarm problem is "abnormal temperature rise of the main transformer in a substation," the model might identify entities such as "substation" (location), "main transformer" (equipment), and "abnormal temperature rise" (fault phenomenon). Next, based on the domain entity set and the power operation and maintenance domain knowledge dictionary, a mapping table from entities to their extended terms is determined. That is, after identifying the domain entity set, the system will use a pre-built power operation and maintenance domain knowledge dictionary to find extended terms for these entities. This power operation and maintenance domain knowledge dictionary is a specialized knowledge base that may contain: synonyms / near-synonyms: for example, "abnormal temperature rise" may be associated with "overheating" or "high temperature warning"; abbreviations / full names: for example, "transformer" may be associated with abbreviations or full names such as "TRAFO"; hierarchical relationships or related equipment: for example, "main transformer" may be associated with "cooling system" or "oil tank," or "substation" may include "circuit breaker" or "disconnect switch"; causal association words: if a "key text fragment" indicates a cause, then words related to that cause can be used as extended terms. Based on this knowledge, the system will build one or more related extended terms for each domain entity and store them in the entity-to-extended-term mapping table. The construction of this mapping table is based on the process of querying and associating domain entities from the professional knowledge dictionary. Furthermore, based on the mapping table from entities to their extended terms, the original alert question is reconstructed using a controlled Boolean query to obtain the expanded query. That is, the original alert question is reconstructed using the mapping table from entities to their extended terms, forming an expanded query. Controlled Boolean query reconstruction means that expansion is not simply appending words, but rather using Boolean logical operators (AND, OR, etc.) to construct a structured query to ensure the accuracy and effectiveness of the expansion. For example, for each identified entity in the original query, the pattern "(entity OR extended term 1 OR extended term 2)" can be used, and these expanded entity phrases can be connected using "AND". If key text fragments provide specific cause analysis text, the technical terms or core concepts contained in this text can also be incorporated into the Boolean query to further refine the query scope. In this way, the expanded query not only retains the information of the original alert question but also incorporates rich professional domain knowledge and related concepts, enabling it to match more comprehensive and in-depth solution documents in subsequent secondary searches.

[0025] Specifically, in step S5, a secondary retrieval and retrieval result filtering are performed on the expanded query to obtain the final key text fragments. It should be understood that using all documents obtained during the retrieval phase as input to the large model may exceed the length limit of the model's context window. In this case, the large model will truncate the excess portion, resulting in the large model being unable to obtain comprehensive contextual information. Furthermore, the relevant documents obtained during the retrieval phase may contain content irrelevant to the current question. The presence of this irrelevant information will interfere with the large model, making it difficult for it to accurately capture the key information for answering the question, thus leading to a decrease in the quality of the large model's output. Therefore, in the technical solution of this application, a secondary retrieval is used to obtain more comprehensive and detailed information, and irrelevant information is effectively removed through retrieval result filtering to extract key text blocks that can directly answer the alarm question, while mitigating the information loss problem caused by the input exceeding the context window length limit. Ultimately, by reducing the interference of a large amount of irrelevant information on the model, the model's understanding ability is enhanced, thereby improving the quality of the large model's answer generation to generate a more practical and accurate solution.

[0026] In practice, the first step is a secondary retrieval. During this process, the system vectorizes the expanded query again, generating new sparse and dense query vectors. Then, using these new query vectors, the system accesses the retrieval module in the knowledge base index again, performing a process similar to or more optimized than the initial retrieval (e.g., targeting a larger search space or more stringent semantic matching) to obtain a series of documents or text fragments that better match the expanded query as preliminary secondary retrieval results. The next step is filtering the search results. First, the expanded query is merged with the documents or text fragments obtained from the secondary search, serving as input to the text filter. Then, the merged input is sent to the pre-trained Qwen1.5-14b-chat model. This model performs specific filtering tasks through prompting engineering. Specifically, it identifies and removes text content that is not directly related to the alert issue or has little value in answering the question by performing deep semantic analysis and understanding of the input content. By removing redundant information, it ensures that the most valuable information is carried as much as possible within a limited context window. After processing and filtering by the Qwen1.5-14b-chat model, highly refined text content directly addressing the alert issue and possessing processing value is finally obtained—the final key text fragments.

[0027] Specifically, S6 generates a handling plan based on the final key text fragments and the original alarm question. Through a series of rigorous hybrid retrieval, validity judgment, and query expansion processes, the system has obtained highly focused and structurally sound final key text fragments, which are verified, high-quality professional knowledge. Combining this knowledge with the original alarm question and using a large-scale generative model to generate a solution avoids the limitations of traditional models that rely entirely on ontology training data and are prone to providing fictitious or fabricated answers. This reduces the risk of content illusion generated by large models and ensures that the generated alarm handling plan has stronger practicality, accuracy, and security. This allows operations and maintenance personnel to directly obtain specific handling suggestions, thereby achieving rapid and accurate problem handling.

[0028] In practice, the final key text fragment and the original alarm question are first injected into a knowledge-enhanced generation template to obtain reasoning-based prompts. During this process, the system inputs both the final key text fragment and the original alarm question into the knowledge-enhanced generation template. This template, as a prompting tool in RAG technology, uses a preset structure and guidance. For example, the template might contain a format like, "Original Alarm: [Original Alarm Question]; Background Knowledge (Cause Analysis and Handling Steps): [Final Key Text Fragment]; Please generate a detailed handling plan based on the above information." Through this structured injection, the reasoning-based prompts clearly inform the large-scale generative model of the task objectives, available information, and expected output format, thereby guiding the model to perform more effective information integration and logical reasoning. Next, the inference-based prompts are input into a large-scale generative model to obtain the original text output by the model. The large-scale generative model is a language model fine-tuned with knowledge specific to a particular domain (such as power operation and maintenance). It leverages its powerful language understanding and generation capabilities, combined with the domain knowledge contained in the final key text fragments provided in the inference-based prompts, and its understanding of the original alarm question, to generate an original text answering the alarm question. This original text is the result of the model's semantic understanding and logical inference based on the prompt content, containing the model's initial ideas on how to handle the alarm question. Furthermore, the raw text output by the model undergoes structured post-processing to obtain a disposal plan. It should be understood that while the text directly output by large generative models may possess the flexibility of natural language, it often lacks the rigorous structure and standardized format required for operational scenarios. Therefore, further structured post-processing of the raw text output by the model is performed to transform the model's raw output into a clear, standardized disposal plan that is easy for operations personnel to understand and implement. This step specifically includes: Key extraction steps: Identify and extract the main actions, sequence, and conditions from the original text; Formatting: Convert unstructured text into lists, numbered steps, tables, or files in specific formats such as Markdown and JSON, making them easier to read and parse automatically; Supplementing metadata: Based on preset standards, add key metadata to the solution, such as alarm level, expected recovery time, responsible team, required tools, etc. Consistency check: Check whether the technical terminology, instruction format, etc. in the plan comply with the standards and specifications in the field of power operation and maintenance; Through this series of post-processing steps, the final output is a high-quality handling solution that can be directly used to guide operation and maintenance personnel in handling alarms.

[0029] In summary, as Figure 3 As shown in the embodiments of this application, an AI-based intelligent operation and maintenance behavior monitoring method is explained. First, it carefully evaluates the value of initial results through preliminary retrieval and content validity judgment, avoiding unnecessary query rewriting. Only when necessary, it utilizes selected high-quality key text fragments to conditionally and constrainedly expand the original alarm issue. This controlled expansion method fundamentally avoids the semantic deviation risk caused by unconstrained rewriting of general models, ensuring that secondary retrieval can more accurately focus on the core intent of the original alarm. This provides key information with a higher signal-to-noise ratio and stronger relevance for subsequent solution generation, significantly improving the accuracy and practicality of the final generated solution.

[0030] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for monitoring intelligent operation and maintenance behavior based on artificial intelligence, characterized in that, include: Obtain the original alarm issue; The original alarm issues are subjected to mixed retrieval and initial sorting to obtain a sorted list of documents; Based on the original alarm issue, the sorted document list is subjected to content validity judgment to obtain validity indicators and key text fragments; In response to the validity flag being true, the original alarm issue is expanded by conditional query based on key text fragments to obtain the expanded query; The expanded query is subjected to a secondary search and the search results are filtered to obtain the final key text fragments; Based on the final key text fragments and the original alarm issues, a handling plan is generated.

2. The method for monitoring intelligent operation and maintenance behavior based on artificial intelligence according to claim 1, characterized in that, The original alert issues are subjected to a mixed search and initial sorting to obtain a sorted list of documents, including: The original alarm problem is represented by query vectorization to obtain sparse query vectors and dense query vectors; Input the sparse query vector and the dense query vector into the retrieval module in the knowledge base index to obtain a list of candidate documents; The original alarm issues and candidate document list are semantically reordered based on a cross-encoder to obtain the sorted document list.

3. The method for monitoring intelligent operation and maintenance behavior based on artificial intelligence according to claim 1, characterized in that, Based on the original alert issue, the sorted document list is subjected to content validity assessment to obtain validity indicators and key text fragments, including: The original alert issues and the sorted document list are subjected to structured conceptual decomposition of alert-documents to obtain the issue representation text and the decomposed document set; The problem involves calculating the structural fit between the text and each document in the decomposed document set to obtain a set of structural fit scores for each document. Based on the structure fit score set of each document, the decomposed document set is subjected to validity judgment and fragment extraction based on aggregate fit to obtain validity indicators and key text fragments.

4. The method for monitoring intelligent operation and maintenance behavior based on artificial intelligence according to claim 3, characterized in that, Based on the structure fit scores of each document, the decomposed document set is subjected to validity judgment and fragment extraction based on aggregate fit to obtain validity indicators and key text fragments, including: Extract the maximum structure set score from the structure fit score set of each document; If the maximum structure set degree score is greater than the preset validity threshold, the validity flag is true; otherwise, it is false.

5. The method for monitoring intelligent operation and maintenance behavior based on artificial intelligence according to claim 4, characterized in that, Based on the structure fit scores of each document, the decomposed document set is subjected to validity judgment and fragment extraction based on aggregate fit to obtain validity indicators and key text fragments, including: In response to the validity flag being true, the set of structure fit scores for each document is normalized using the Softmax function to obtain the weight distribution; Based on the weight distribution, extract at least one document with the highest score from the decomposed document set; and Extract the cause analysis text and the action procedure text from at least one document, and merge the cause analysis text and the action procedure text to obtain the key text fragment.

6. The method for monitoring intelligent operation and maintenance behavior based on artificial intelligence according to claim 1, characterized in that, In response to a validity flag being true, the original alert issue is expanded using conditional queries based on key text fragments to obtain an expanded query, including: The original alarm problem is input into the named entity recognition model, which has been fine-tuned with domain knowledge, to obtain the domain entity set; Based on the domain entity set and the power operation and maintenance domain knowledge dictionary, a mapping table is determined from the entities to their extended words; Based on the mapping table from entities to their extended terms, the original alarm problem is reconstructed using a controlled Boolean query to obtain the expanded query.

7. The method for monitoring intelligent operation and maintenance behavior based on artificial intelligence according to claim 1, characterized in that, Based on the final key text fragments and the original alarm issue, a handling plan is generated, including: The final key text fragments and the original alert questions are injected into a knowledge-enhanced generation template to obtain hints with reasoning; Input inference-based prompts into a large generative model to obtain the original text output by the model; The original text output by the model is post-processed in a structured manner to obtain a disposal plan.

Images