Intelligent operation and maintenance auxiliary diagnosis method and device, equipment and medium

Abstract

The application relates to an intelligent operation and maintenance auxiliary diagnosis method and device, equipment and medium, which are applied to the field of artificial intelligence technology, and the method comprises the following steps: based on a user query, performing semantic similarity retrieval on a knowledge base through a first retrieval channel; performing keyword matching on the knowledge base through a second retrieval channel; when the user query contains a predetermined entity, performing relationship retrieval in a structured knowledge base contained in the knowledge base; generating a reference context for text segments or structured information related to the first group, the second group and the third group; generating a diagnosis suggestion based on the reference context; calculating a comprehensive confidence score; performing provenance annotation on key judgments in the generated diagnosis suggestion; performing hierarchical processing on the diagnosis suggestion; identifying a knowledge gap corresponding to the user query based on a preset gap determination rule; and generating a knowledge supplement suggestion report. The application has the effect of improving the retrieval accuracy of operation and maintenance diagnosis.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an intelligent operation and maintenance auxiliary diagnosis method, device, equipment and medium. Background Technology

[0002] In modern industry, particularly in the manufacturing and maintenance of high-end equipment such as hydrogen-powered drones and wind power generation equipment, the complexity of equipment is increasing daily. When frontline maintenance personnel face sudden failures on-site, they often rely on remote headquarters experts for guidance. This not only results in low response efficiency but also heavily depends on the individual experience of the experts, making it difficult to transfer and reuse knowledge on a large scale. Although companies have accumulated a wealth of valuable knowledge, such as maintenance manuals, repair records, and fault reports, this knowledge is mostly stored in unstructured or semi-structured text formats, making it difficult to retrieve and utilize efficiently.

[0003] In recent years, large language model application technologies, such as Retrieval-Augmented Generation (RAG), have provided new solutions to the aforementioned problems. This technology enhances the answering ability of language models by retrieving relevant information from external knowledge bases, achieving significant success in areas such as general question answering and document retrieval. However, directly applying general RAG technology to serious industrial operation and maintenance diagnostic scenarios faces a series of unique challenges.

[0004] First, operational diagnostics demand extremely high accuracy in information. Common RAG systems often employ single semantic vector retrieval, which can be misleading due to semantically similar but technically completely different content (for example, "motor noise" and "motor overheating" may be very similar in vector space, but their causes and solutions are drastically different), leading to inaccurate search results and consequently incorrect diagnostic recommendations. This is especially true for uniquely identifiable technical terms such as equipment models and error codes; pure semantic retrieval often cannot guarantee precise matching.

[0005] Secondly, the reliability and interpretability of diagnostic results are crucial. General-purpose RAG systems typically output a "black box" answer, leaving users unable to ascertain its reliability or trace its source. In industrial operations and maintenance scenarios, an incorrect diagnostic recommendation can lead to equipment damage, production stoppages, or even safety incidents. Therefore, operations and maintenance personnel must be able to assess the reliability of diagnostic recommendations and trace the basis for their judgments.

[0006] Finally, existing knowledge bases inevitably contain knowledge gaps. When encountering questions not covered in the knowledge base, general RAG systems can typically only respond with "cannot answer" or offer uncertain guesses, failing to proactively identify these knowledge gaps and initiate corresponding knowledge supplementation processes. This makes the knowledge base itself static, unable to continuously evolve and improve as actual operational problems arise. Summary of the Invention

[0007] To improve the accuracy of operation and maintenance diagnosis, this application provides an intelligent operation and maintenance auxiliary diagnosis method, device, equipment, and medium.

[0008] This invention provides an intelligent operation and maintenance auxiliary diagnosis method, comprising the following steps: based on a user query, performing semantic similarity retrieval on a knowledge base through a first retrieval channel to recall a first group of related text fragments; performing keyword matching on the knowledge base through a second retrieval channel, wherein the keyword matching is used to precisely match predefined types of technical terms to recall a second group of related text fragments; when the user query contains a predetermined entity, performing relation retrieval on a structured knowledge base contained in the knowledge base through a third retrieval channel to recall a third group of related structured information; fusing and sorting the first group, the second group, and the third group of related text fragments or structured information according to a preset fusion rule to generate a reference context for subsequent processing; and based on the reference context... This method uses a language model to generate diagnostic suggestions; calculates a comprehensive confidence score, which is a combination of a retrieval quality indicator representing the relevance of the retrieval results to the user query and a generation quality indicator representing the certainty of the content generated by the language model; performs source annotation on key judgments in the generated diagnostic suggestions, with the source annotation indicating the specific records in the knowledge base on which the key judgment is based; performs hierarchical processing on the diagnostic suggestions according to the preset interval to which the comprehensive confidence score belongs; identifies the knowledge gap corresponding to the user query based on preset gap determination rules; and automatically generates a knowledge supplementation suggestion report when the knowledge gap is identified, the report including the original question, relevant but mismatched knowledge retrieved from the knowledge base, and suggested directions for knowledge supplementation. This method comprehensively improves the accuracy, reliability, and self-evolutionary capability of intelligent diagnosis through multi-channel hybrid retrieval, confidence calibration, and knowledge closed-loop management.

[0009] Furthermore, the first retrieval channel is a channel for retrieval based on semantic similarity, the second retrieval channel is a channel for retrieval based on keyword matching, the structured knowledge base is a database that stores knowledge in a structured form, and the relationship retrieval is a relationship retrieval performed in the structured knowledge base; moreover, the preset fusion rules include fusion sorting according to a preset weighted fusion formula. This retrieval method, which combines semantics, keywords, and structured knowledge, can take into account both the generalization ability of semantics and the precise matching ability of technical terms, significantly improving the retrieval quality.

[0010] Furthermore, the retrieval quality metric is the similarity score between the reference context and the user query, and the generation quality metric is the average probability of key tokens when the language model generates the diagnostic suggestions. By quantifying the quality of both the retrieval and generation stages, an objective basis is provided for comprehensively evaluating the credibility of the diagnostic suggestions.

[0011] Furthermore, the tiered processing includes: directly outputting the diagnostic suggestion when the overall confidence score is higher than a first threshold; outputting the diagnostic suggestion with a manual review prompt when the overall confidence score is between the first and second thresholds; and transferring the issue to a human expert when the overall confidence score is lower than the second threshold. This tiered mechanism ensures that high-risk, low-confidence suggestions can be effectively managed, guaranteeing the security of operation and maintenance.

[0012] Furthermore, the gap determination rule includes at least one of the following: the similarity between the user query and all documents in the knowledge base is lower than a preset similarity threshold; the language model uses preset uncertainty words when generating the diagnostic suggestions. This rule can proactively and intelligently identify problems that the current knowledge base cannot effectively cover.

[0013] Preferably, the traceability annotation is implemented by linking key judgments in the diagnostic recommendations to specific records in the knowledge base via hyperlinks. This approach provides users with an intuitive and convenient traceability path, ensuring that every diagnostic conclusion is verifiable, greatly enhancing the system's transparency and user trust.

[0014] Furthermore, the method also includes tracking the replenishment rate and frequency of knowledge gaps, and triggering alerts when specific types of knowledge gaps occur frequently. This approach provides a macro-level insight into the weaknesses of the knowledge base, offering data-driven guidance for the systematic construction of the knowledge system.

[0015] This invention also provides an intelligent operation and maintenance auxiliary diagnostic device, comprising: a receiving unit for receiving user queries; a retrieval unit for performing semantic similarity retrieval on a knowledge base through a first retrieval channel to retrieve a first group of related text fragments, performing keyword matching on the knowledge base through a second retrieval channel to retrieve a second group of related text fragments, and performing relation retrieval on a structured knowledge base contained in the knowledge base through a third retrieval channel when the user query contains a predetermined entity, to retrieve a third group of related structured information; and a fusion unit for processing the first group, the second group, and the third group of related text fragments. The system integrates segmented or structured information, merging and sorting it according to preset fusion rules to generate a reference context for subsequent processing; a generation unit generates diagnostic suggestions based on the reference context using a language model; an evaluation unit calculates a comprehensive confidence score; a source tracing unit annotates key judgments in the generated diagnostic suggestions; a processing unit performs hierarchical processing on the diagnostic suggestions according to a preset range to which the comprehensive confidence score belongs; and a knowledge management unit identifies the knowledge gap corresponding to the user's query based on preset gap determination rules, and automatically generates a knowledge supplementation suggestion report when the knowledge gap is identified.

[0016] The present invention also provides an electronic device including a processor coupled to a memory; the processor is configured to execute a computer program stored in the memory to cause the electronic device to perform the method described in any of the preceding claims.

[0017] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when run on a computer, causes the computer to perform the method described in any of the preceding claims.

[0018] In summary, this application includes at least one of the following beneficial technical effects: 1. Improved diagnostic accuracy. By employing a three-channel hybrid retrieval strategy combining semantics, keywords, and a structured knowledge base, and by fusing and ranking the retrieval results, this invention can simultaneously utilize the generalized understanding capability of semantics and the precise matching capability of keywords and structured knowledge. Especially when processing queries containing key technical terms such as error codes and component models, it is significantly superior to single vector retrieval, thereby providing more accurate reference information for subsequent generation stages and improving the accuracy of diagnostic suggestions from the source.

[0019] 2. Enhanced credibility and interpretability of diagnostic results. This invention transforms diagnostic recommendations from a "black box" answer into a quantifiable and assessable decision-making reference by calculating a comprehensive confidence score. Simultaneously, by tracing and annotating key judgments, every conclusion is verifiable. This "quantifiable and traceable" characteristic is crucial in industrial operation and maintenance scenarios with extremely high security and reliability requirements, greatly enhancing user trust in the system.

[0020] 3. Achieved a self-evolving closed loop for the knowledge base. This invention not only answers questions but also identifies deficiencies in the knowledge base through a proactive knowledge gap discovery mechanism, automatically generating structured knowledge supplementation suggestion reports. This transforms every unresolved problem encountered into an opportunity for knowledge base growth, enabling the system to evolve from a static knowledge query tool into a continuously learning and improving, self-evolving knowledge ecosystem with greater long-term application value. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the architecture of an intelligent operation and maintenance auxiliary diagnostic system provided in an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of the architecture of a three-channel hybrid retrieval engine provided in an embodiment of the present invention.

[0023] Figure 3 This is a schematic diagram of a diagnostic suggestion output interface provided in an embodiment of the present invention.

[0024] Figure 4 This is a schematic diagram of a knowledge gap discovery and supplementation process provided in an embodiment of the present invention. Detailed Implementation

[0025] The present application will be further described in detail below with reference to the accompanying drawings.

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0027] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0028] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0029] This invention aims to address the core problems faced by existing retrieval enhancement generation technologies when applied to the field of industrial operation and maintenance diagnosis, including insufficient retrieval accuracy, lack of credibility and interpretability in diagnostic suggestions, and inability to proactively identify and fill knowledge gaps. The core idea of ​​this invention is to significantly improve the accuracy, reliability, and self-evolutionary capability of diagnosis by constructing an intelligent operation and maintenance auxiliary diagnostic method and system that integrates hybrid retrieval, confidence calibration, and knowledge closed-loop management. For clarity, the core terms involved in this invention are explained first: the first retrieval channel refers to a channel for retrieval based on semantic similarity; the second retrieval channel refers to a channel for retrieval based on keyword matching, used for precise matching of technical terms; the third retrieval channel refers to a channel for relational retrieval in a structured knowledge base; the structured knowledge base refers to a database that stores knowledge in a structured form (such as triples); the comprehensive confidence score refers to a quantitative score that comprehensively evaluates the credibility of diagnostic suggestions; and the knowledge gap refers to user questions that the current knowledge base cannot effectively answer. The technical solution of this invention will be described in detail below.

[0030] Reference Figure 1This invention provides an overall architecture for an intelligent operation and maintenance auxiliary diagnostic system 100. The system 100 mainly includes a three-channel hybrid retrieval engine 110, a diagnostic generation and calibration module 120, and a knowledge gap management module 130. All modules interact with an underlying knowledge base 140. Users submit operation and maintenance diagnostic related queries to the system through a user interface. The three-channel hybrid retrieval engine 110 is responsible for retrieving the most relevant information from the knowledge base 140. The diagnostic generation and calibration module 120 generates diagnostic suggestions based on the retrieved information and performs confidence assessment and source tracing. The knowledge gap management module 130 is responsible for identifying and managing deficiencies in the knowledge base 140, forming a closed-loop knowledge supplementation process. This architecture, through the collaborative work between modules, achieves end-to-end optimization from problem understanding, information retrieval, intelligent generation to knowledge evolution.

[0031] Combination Figure 1 and Figure 4 The complete process of an intelligent operation and maintenance auxiliary diagnosis method provided by this invention is demonstrated. First, the system receives a query input by the user. Then, the three-channel hybrid retrieval engine 110 starts three retrieval channels in parallel: the first retrieval channel performs semantic similarity retrieval on the knowledge base 140 to retrieve the first set of text fragments; the second retrieval channel performs keyword matching, especially for predefined technical terms, to retrieve the second set of text fragments; when the query contains a predetermined entity, the third retrieval channel performs relation retrieval in the structured knowledge base embedded in the knowledge base 140 to retrieve the third set of structured information. Subsequently, these three sets of information are fused and sorted according to preset fusion rules to generate a high-quality reference context. The diagnosis generation and calibration module 120 uses a language model to generate preliminary diagnostic suggestions based on this context and calculates the comprehensive confidence score, while also tracing and annotating the key judgments in the suggestions. Afterwards, the system performs hierarchical processing according to the comprehensive confidence score. At the same time, the knowledge gap management module 130 determines whether the user query constitutes a knowledge gap according to preset gap judgment rules, as shown in step S401. If the determination is yes, a knowledge supplementation suggestion report is automatically generated as shown in step S402. This report can then be submitted to human experts for review and supplementation, as shown in step S403. Finally, the new knowledge is updated to the knowledge base 140, as shown in step S404, thus completing the closed-loop knowledge supplementation.

[0032] To more clearly illustrate the technical solution of the present invention, a detailed description is provided below with reference to specific embodiments. In a basic embodiment, an intelligent operation and maintenance auxiliary diagnostic device is provided, which can be used to perform the above-described method. Figure 1As shown, the device can be a software system deployed on a server or cloud platform, comprising: a receiving unit for receiving user queries; a retrieval unit for performing multi-channel retrieval; a fusion unit for fusing retrieval results from multiple channels; a generation unit for generating diagnostic suggestions; an evaluation unit for calculating a comprehensive confidence score; a source tracing unit for tracing and labeling the diagnostic suggestions; a processing unit for performing hierarchical processing on the diagnostic suggestions; and a knowledge management unit for identifying and managing knowledge gaps.

[0033] During operation, the receiving unit of the device first receives user queries from maintenance personnel. Subsequently, the retrieval unit is activated, with its internal first, second, and third retrieval channels starting simultaneously. The first retrieval channel, based on semantic understanding, identifies document fragments from the knowledge base that are semantically similar to the query. The second retrieval channel focuses on exact matching, finding records containing the exact same technical terms as in the query. The third retrieval channel searches for known relationships related to the query entity in the structured knowledge base. This multi-channel parallel retrieval design ensures comprehensive information retrieval. After the retrieval is complete, the fusion unit receives the retrieval results from the three channels and sorts them according to preset fusion rules, selecting the most relevant and reliable information fragments to form a comprehensive reference context. This step effectively enhances the factual basis of subsequent generated content by integrating the advantages of different retrieval strategies. Upon receiving this reference context, the language model within the generation unit generates a diagnostic suggestion containing specific diagnostic steps or cause analysis.

[0034] After generating diagnostic suggestions, the evaluation unit immediately begins its work, calculating a comprehensive confidence score. This score integrates two dimensions: first, the retrieval quality indicator, which measures the relevance of the reference context to the original user query; and second, the generation quality indicator, which measures the internal certainty of the language model in generating the diagnostic suggestions. Simultaneously, the source tracing unit annotates each key assertion in the diagnostic suggestions, clearly indicating its information source from a specific record in the knowledge base. This design ensures that the diagnostic results are no longer a black box, but rather verifiable. Next, the processing unit performs tiered processing on the diagnostic suggestions based on the comprehensive confidence score calculated by the evaluation unit. For example, suggestions with high scores can be presented directly to the user; suggestions with medium scores include accompanying prompts; and suggestions with very low scores may be directly referred to human experts. Finally, the knowledge management unit analyzes the query process and determines whether the knowledge base can adequately answer the question based on preset gap assessment rules. If not, it is identified as a knowledge gap event, and a knowledge supplementation suggestion report is automatically generated, providing clear guidance for subsequent knowledge base improvement. Through the coordinated implementation of these steps, the device in this embodiment provides efficient, reliable, and self-learning support for operational and maintenance diagnostics.

[0035] like Figure 2As shown, the specific implementation of the three-channel hybrid retrieval engine 110 can be further clarified. Specifically, the first retrieval channel 111 can be implemented as dense vector retrieval. It uses a text embedding model finely tuned for a specific domain (such as the "co-hydrogen operation and maintenance corpus" in this example), such as the bge-large-zh model, to convert both the query and knowledge base documents into high-dimensional vectors. Semantic retrieval is performed by calculating the cosine similarity between vectors. The second retrieval channel 112 can be implemented as sparse keyword retrieval. It employs the classic BM25 algorithm to perform precise keyword matching on technical terms in the query, such as the equipment model "H15-ESC-V2" or the error code "E07". The third retrieval channel 113 can be implemented as retrieval on a knowledge graph. It operates on a pre-built structured knowledge base, which stores knowledge in the form of triples (such as <fault, associated component, maintenance plan>). When the query hits an entity in the graph, relevant structured information is retrieved through subgraph matching or pathfinding. To integrate the advantages of these three channels, the fusion ranking unit 114 employs a pre-defined weighted fusion rule. Specifically, the final ranking score for each search result is calculated using the following pre-defined weighted fusion formula: Final ranking score = α × dense_score + β × sparse_score + γ × kg_score. In this formula, dense_score, sparse_score, and kg_score represent the relevance scores given by the first, second, and third search channels, respectively. The weight coefficients α, β, and γ are obtained through cross-validation training on a manually annotated search quality dataset to balance the contributions of semantic relevance, keyword accuracy, and structured association. Through this hybrid retrieval and weighted fusion design, the system can overcome the limitations of a single retrieval method. For example, the second and third search channels compensate for the semantic drift problem that may occur when the first search channel handles precise technical terms, thereby significantly improving the overall accuracy and relevance of the search results.

[0036] In a preferred embodiment of the present invention, the calculation of the comprehensive confidence score is specified. The retrieval quality index is defined as the similarity score between the top-ranked (Top-1) reference context fragment and the user query after fusion ranking. This score can be the cosine similarity value calculated by the first retrieval channel. The generation quality index is defined as the average generation probability of all tokens in the sentence constituting the key diagnostic conclusion when the language model generates diagnostic suggestions. This calculation method allows the index to more accurately reflect the model's certainty regarding the core judgment. Subsequently, the comprehensive confidence score is calculated using the following formula: Comprehensive Confidence Score = 0.5 × retrieval_confidence + 0.5 × generation_confidence, where retrieval_confidence is the aforementioned retrieval quality index, and generation_confidence is the generation quality index. The 0.5 weight here is an example and can be adjusted according to the actual application scenario to emphasize the reliability of retrieval or the fluency of generation. This quantitative confidence assessment mechanism provides a solid decision-making basis for subsequent hierarchical processing. Based on the calculated comprehensive confidence score, the system performs refined hierarchical processing. Specifically, this processing logic relies on two preset thresholds: a first threshold and a second threshold. For example, the first threshold can be set to 0.85, and the second threshold to 0.6. When the overall confidence score is higher than the first threshold of 0.85, the system determines that the diagnostic suggestion is highly credible and outputs it directly to the user. When the score is between the second threshold of 0.6 and the first threshold of 0.85, for example, a calculated score of 0.75, the system outputs a diagnostic suggestion, but also adds a "manual review recommended" prompt, reminding the user that the suggestion has some uncertainty. When the score is lower than the second threshold of 0.6, the system considers the currently generated suggestion insufficient to be credible and automatically transfers the issue along with relevant context information to a human expert queue for processing. Through this confidence-based hierarchical processing strategy, this invention effectively controls the risks of erroneous diagnosis while ensuring automation efficiency, greatly enhancing the system's practicality and security in serious industrial scenarios.

[0037] As a preferred embodiment, the present invention further refines the determination and management of knowledge gaps. The knowledge gap management module 130 employs gap determination rules including at least the following two: First, based on the similarity of search results. When the highest similarity score calculated between a user query and all document fragments in the knowledge base 140 is lower than a preset similarity threshold (for example, this threshold can be set to 0.6), the system determines that existing knowledge cannot effectively cover the problem. For example, a query about a malfunction of a new model of drone has a highest similarity of only 0.5 with documents in the database, which is lower than the threshold of 0.6, and is therefore determined to be a knowledge gap. Second, based on the semantics of the generated content. When the language model frequently uses preset uncertainty words such as "maybe," "perhaps," and "uncertain" when generating diagnostic suggestions, this also indicates that the model lacks sufficient knowledge to make a definitive judgment, which also triggers the determination of a knowledge gap. Once a knowledge gap is determined, the system automatically generates a structured knowledge supplementation suggestion report. The report is detailed and typically includes: the original user problem description, relevant but not entirely matching knowledge fragments retrieved from the knowledge base (for expert reference), and suggested supplementary knowledge directions based on the system's problem analysis (e.g., "suggestion to supplement the troubleshooting process for the hydrogen supply system of the H200 model"). This report provides valuable input for knowledge base maintainers. Furthermore, to achieve continuous evolution of the knowledge base, the method of this invention also includes tracking the supplementation rate and frequency of knowledge gaps. The system backend counts the number of occurrences of each type of knowledge gap. When a specific type of knowledge gap (e.g., a knowledge gap related to a specific model or fault code) occurs frequently within a certain time window (e.g., within one month) and the number exceeds a preset warning threshold (e.g., 5 times), the system automatically triggers a warning mechanism, such as sending a warning email to the operations manager or domain expert. This design transforms knowledge base maintenance from a passive response to proactive planning, enabling the systematic identification and filling of weaknesses in the knowledge system, thereby building a self-improving and continuously growing knowledge ecosystem.

[0038] To enhance the traceability of diagnostic recommendations, this invention designs a specific implementation method for traceability labeling. For example... Figure 3The diagram shows a schematic of a diagnostic suggestion output interface. The interface not only displays the diagnostic suggestion text 301 and the quantified comprehensive confidence score display area 302, but more importantly, it provides source annotations for the key judgments in the suggestion. In this embodiment, this source annotation is implemented by making the key judgment parts in the diagnostic suggestion text 301 into hyperlinks, namely, source annotation hyperlinks 303. Specifically, when a user sees a diagnostic suggestion on the interface, such as "It is recommended to first check the calibration status of the hydrogen cylinder pressure sensor," this sentence or the key phrase "calibration status of the hydrogen cylinder pressure sensor" will be rendered as a clickable hyperlink. When the user clicks the source annotation hyperlink 303, the system will pop up a window or display the original knowledge fragment on which the judgment is based in the sidebar of the page. For example, the pop-up window will clearly display: "This troubleshooting plan comes from the Foshan Photovoltaic Station H15-E07 Fault Repair Record #1423 in March 2024," along with the relevant original text paragraphs from that record. This design, which directly links conclusions with evidence, achieves complete transparency and traceability in the diagnostic process. Maintenance personnel not only know what happened, but also why, and can independently verify the reliability of information sources, thereby greatly enhancing their trust in the system's output. This is a crucial capability in industrial operations and maintenance, where high reliability and security are paramount.

[0039] To more comprehensively demonstrate the preferred embodiments of the present invention, a complete embodiment incorporating all the above-mentioned technical features will be described below through a specific hydrogen-powered drone diagnostic scenario. Assume the system receives a user query from a frontline maintenance worker: "H15 drone automatically landed after reporting error E07, hydrogen cylinder pressure is normal." First, the system's receiving unit receives the query and passes it to the three-channel hybrid retrieval engine 110. The first retrieval channel 111 retrieves multiple historical maintenance records through semantic similarity, including cases where E07 faults were handled due to loose hydrogen cylinder pipe connections, as well as the E07 fault code section and standard troubleshooting flowchart in the H15 drone operation manual. The second retrieval channel 112 uses the BM25 algorithm to precisely match the keywords "E07" and "hydrogen cylinder pressure," accurately hitting the fault code definition table, which clearly indicates that E07 is defined as "low hydrogen cylinder pressure triggering protective landing." The third retrieval channel 113 retrieves associated triples from the structured knowledge base based on the entity "E07" in the query, such as <Error E07, associated components, hydrogen supply pipeline> and <Error E07, associated components, hydrogen cylinder pressure sensor>, thereby quickly identifying a series of related core components.

[0040] Next, the fusion ranking unit 114 ranks the three recall results according to the weighted fusion formula (final ranking score = α × dense_score + β × sparse_score + γ × kg_score). Since the results from the second and third retrieval channels precisely indicate a strong correlation between E07 and pressure, and the user query explicitly states the contradiction of "hydrogen cylinder pressure is normal," the fusion algorithm assigns higher weight to clues revealing this contradiction (such as cases concerning sensor malfunction). Therefore, in the fused Top-3 reference context, diagnostic clues about "sensor malfunction" are ranked highly. Subsequently, the language model in the generation unit generates diagnostic suggestions based on this high-quality reference context. The model identified a logical conflict between "E07 is defined as low pressure" and "the on-site pressure display is normal," leading to the conclusion that the sensor itself might be faulty. It then generated a diagnostic suggestion: "E07 is triggered when the hydrogen cylinder pressure sensor reading is below a threshold. Given your report that the hydrogen cylinder pressure display is normal, there is a possibility that the sensor itself may be malfunctioning (e.g., the reading is too high or stuck at the normal value). It is recommended to check the calibration status of the hydrogen cylinder pressure sensor first, rather than directly checking the hydrogen supply line." While generating recommendations, the evaluation unit calculated a comprehensive confidence score of 0.9 for this diagnosis. This score is based on highly relevant but contradictory evidence retrieved and the average token probability of the model generating a definitive conclusion. Simultaneously, the tracing unit added a tracing hyperlink (303) to the key judgment "It is recommended to first check the calibration status of the hydrogen cylinder pressure sensor," linking it to a similar sensor calibration error repair case in the knowledge base. Finally, based on the judgment that the comprehensive confidence score of 0.9 is higher than the first threshold of 0.85, the processing unit directly outputs the diagnostic recommendation with the tracing link to the user. Since the problem was effectively solved, the knowledge gap management module 130 did not trigger a knowledge gap judgment. This embodiment fully demonstrates how the present invention, through the synergistic effect of various technical features—hybrid retrieval to discover deep contradictions, confidence assessment to ensure reliable decision-making, and evidence tracing to enhance user trust—ultimately achieves a deep diagnostic capability that surpasses that of a single RAG system, providing accurate, reliable, and explainable intelligent assistance for complex industrial operation and maintenance scenarios.

[0041] In another alternative implementation, the construction process of the structured knowledge base can be automated. Specifically, an automatic extraction technique based on a large language model can be used to replace traditional manual annotation. By designing specific prompt word engineering and combining it with a small number of annotated samples for example learning, the large language model can automatically identify and extract triple relationships in the format of <fault, associated parts, maintenance plan> from unstructured text such as operation and maintenance manuals and maintenance records. This allows for the efficient construction and updating of the structured knowledge base, reducing the labor costs of knowledge base maintenance and improving update efficiency.

[0042] In another optional implementation, the training process for the weight coefficients α, β, and γ in the weighted fusion formula can be further refined. Specifically, the weights can be trained on a manually labeled retrieval quality dataset using optimization methods such as grid search, Bayesian optimization, or genetic algorithms.

[0043] In another alternative implementation, the calculation and calibration of the overall confidence score can employ a statistical guarantee framework based on conformal prediction. Unlike the aforementioned method of retrieving and generating quality indicators through weighted combination, conformal prediction provides a statistically significant confidence interval or prediction set for the generated diagnostic recommendations. For example, the system can output a set containing multiple possible diagnostic outcomes, guaranteeing that the truly correct diagnostic solution falls within this set with a preset probability (e.g., 95%). This method provides stronger theoretical assurance for confidence assessment, and is particularly suitable for scenarios with extremely high risk control requirements.

[0044] Those skilled in the art will understand that the above embodiments are preferred embodiments for implementing the present invention. In practical applications, various changes and modifications can be made to the structure and implementation method according to specific circumstances. For example, the method of the present invention can be implemented by an electronic device including a processor and a memory coupled to the processor, the processor being used to execute a computer program stored in the memory to cause the electronic device to perform the method as described in any of the embodiments. Furthermore, the present invention can also be provided in the form of a computer-readable storage medium storing a computer program or instructions that, when run on a computer, cause the computer to perform any of the methods described in any of the embodiments. Those skilled in the art can make modifications and variations to these embodiments without departing from the spirit and scope of the present invention.

[0045] In another alternative embodiment, the method and apparatus of the present invention can be further extended to support multimodal input. Specifically, the system can introduce a large multimodal model for directly analyzing user-uploaded images of fault locations. This multimodal model can identify equipment components and abnormal conditions (such as leaks, burn marks, instrument readings, etc.) in the images and convert this visual diagnostic information into text descriptions or feature vectors. These results extracted from visual information can be used as additional contextual information, fed into the retrieval engine along with information from text queries, or used as a new retrieval dimension, enabling the system to comprehensively utilize both textual and visual information for a more comprehensive diagnosis, further improving diagnostic accuracy.

[0046] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the foregoing application concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions claimed in this application.

Claims

1. An intelligent operation and maintenance auxiliary diagnostic method, characterized in that, include: Based on user queries, semantic similarity retrieval is performed on the knowledge base through the first retrieval channel to recall the first set of relevant text fragments; The knowledge base is matched with keywords through a second retrieval channel. The keyword matching is used to accurately match predefined types of technical terms in order to recall a second set of relevant text fragments. When the user query contains a predetermined entity, a relationship retrieval is performed in the structured knowledge base contained in the knowledge base through the third retrieval channel to recall the third set of related structured information; The text fragments or structured information related to the first group, the second group and the third group are merged and sorted according to a preset fusion rule to generate a reference context for subsequent processing. Based on the aforementioned reference context, diagnostic suggestions are generated using a language model; Calculate the overall confidence score, which is a combination of a retrieval quality index that characterizes the relevance of the retrieval results to the user query and a generation quality index that characterizes the certainty of the content generated by the language model; The key judgments in the generated diagnostic suggestions are traced and annotated, and the traceability annotations indicate the specific records in the knowledge base on which the key judgments are based; The diagnostic recommendations are graded according to the preset interval to which the comprehensive confidence score belongs; Based on preset gap determination rules, the system identifies the knowledge gap corresponding to the user's query; when the knowledge gap is identified, a knowledge supplementation suggestion report is automatically generated.

2. The method according to claim 1, characterized in that, The first retrieval channel is a dense vector retrieval channel, the second retrieval channel is a sparse keyword retrieval channel, the structured knowledge base is a knowledge graph, and the relation retrieval is a triple retrieval; furthermore, the preset fusion rules include fusion sorting according to a preset weighted fusion formula.

3. The method according to claim 1, characterized in that, The retrieval quality metric is the similarity score between the reference context and the user query, and the generation quality metric is the average probability of the key token when the language model generates the diagnostic suggestion.

4. The method according to claim 1, characterized in that, The hierarchical processing includes: When the overall confidence score is higher than the first threshold, the diagnostic suggestion is output directly. When the overall confidence score is between the first threshold and the second threshold, the diagnostic suggestion is output and a manual review prompt is added. When the overall confidence score is lower than the second threshold, the issue will be referred to a human expert.

5. The method according to claim 1, characterized in that, The gap determination rule includes at least one of the following: the similarity between the user query and all documents in the knowledge base is lower than a preset similarity threshold; the language model uses preset uncertainty words when generating the diagnostic suggestions.

6. The method according to claim 1, characterized in that, The source tracing annotation is achieved by linking key judgments in the diagnostic recommendations to specific records in the knowledge base via hyperlinks.

7. The method according to claim 1, characterized in that, The method also includes: tracking the replenishment rate and frequency of knowledge gaps, and triggering an alert when specific types of knowledge gaps occur frequently.

8. An intelligent operation and maintenance auxiliary diagnostic device, characterized in that, include: The receiving unit is used to receive user queries; The retrieval unit is configured to, based on the user query, perform semantic similarity retrieval on the knowledge base through a first retrieval channel to recall a first set of related text fragments, perform keyword matching on the knowledge base through a second retrieval channel to recall a second set of related text fragments, and when the user query contains a predetermined entity, perform relation retrieval on the structured knowledge base contained in the knowledge base through a third retrieval channel to recall a third set of related structured information. The fusion unit is used to merge and sort the related text fragments or structured information of the first group, the second group and the third group according to a preset fusion rule to generate a reference context for subsequent processing. The generation unit is used to generate diagnostic suggestions based on the reference context and through a language model; An evaluation unit is used to calculate a comprehensive confidence score, which is a combination of a retrieval quality index that characterizes the relevance of the retrieval results to the user query and a generation quality index that characterizes the certainty of the content generated by the language model. The source tracing unit is used to trace the source of key judgments in the generated diagnostic suggestions. The source tracing annotation indicates the specific record in the knowledge base on which the key judgment is based. The processing unit is used to perform tiered processing on the diagnostic recommendations based on the preset interval to which the comprehensive confidence score belongs; The knowledge management unit is used to identify the knowledge gap corresponding to the user's query based on preset gap determination rules, and automatically generate a knowledge supplementation suggestion report when the knowledge gap is identified.

9. An electronic device, characterized in that, Includes a processor, which is coupled to a memory; The processor is configured to execute a computer program stored in the memory to cause the electronic device to perform the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It includes a computer program or instructions that, when run on a computer, cause the computer to perform the method as described in any one of claims 1 to 7.

Images