An intelligent maintenance decision generation method, system, device, medium and product

Abstract

The application discloses an intelligent maintenance decision generation method, system, device, medium and product, relates to the operation and maintenance guarantee field of modern industrial equipment, and comprises the following steps: performing double-path recall retrieval and relevance reordering according to field fault condition data, obtaining a knowledge fragment set, and performing context reconstruction to obtain an enhanced prompt word; performing chain thinking reasoning through a vertical field model according to the enhanced prompt word, obtaining a fault cause chain, and combining a hard constraint set to perform semantic reorganization to obtain a preliminary maintenance decision scheme; wherein, an inference path of the chain thinking reasoning is constructed according to state transition probability calculated from a retrieval fragment in the enhanced prompt word; performing compliance detection and knowledge graph parameter reverse checking according to the preliminary maintenance decision scheme to generate a final intelligent maintenance decision, and the application provides high-quality and high-relevance context support, and improves the scientificity, rigor and safety of the maintenance decision under a complex fault scene.

Description

Technical Field

[0001] This application relates to the field of operation and maintenance support for modern industrial equipment, and in particular to an intelligent maintenance decision generation method, system, equipment, medium, and product. Background Technology

[0002] In the operation and maintenance of modern industrial equipment, with the continuous improvement of equipment integration and intelligence, the technical documents, fault cases, and expert experience generated at the maintenance site are experiencing explosive growth. However, existing maintenance decision support technologies still have the following problems in practical applications: On the one hand, existing Retrieval-Augmented Generation (RAG) technologies lack deep semantic understanding and accurate recall capabilities for multi-source heterogeneous data when facing professional maintenance scenarios. They often simply mechanically pile up search results, failing to organically integrate scattered knowledge points and providing high-quality, structured contextual support. On the other hand, existing large language models lack deep coupling and reverse verification mechanisms for the physical logic and hard safety specifications of vertical domains when dealing with professional maintenance problems. This easily leads to "knowledge illusions," generating seemingly reasonable but actually erroneous solutions that violate safety procedures or physical logic, making it difficult to guarantee the scientific nature and safety of maintenance decisions. Summary of the Invention

[0003] The purpose of this application is to provide an intelligent maintenance decision generation method, system, device, medium, and product to solve the problems of difficulty in providing high-quality, structured contextual support and difficulty in ensuring the scientific nature and security of maintenance decisions.

[0004] To achieve the above objectives, this application provides the following solution: Firstly, this application provides an intelligent maintenance decision generation method, including: Based on the on-site fault data, a dual-path recall retrieval and relevance reordering are performed to obtain a set of knowledge fragments, and then the context is reconstructed to obtain enhanced prompt words; Based on the enhanced prompts, chain-like reasoning is performed through a vertical domain model to obtain the fault cause chain, and semantic reorganization is performed in combination with a set of hard constraints to obtain a preliminary maintenance decision plan; wherein, the reasoning path of the chain-like reasoning is constructed based on the state transition probabilities calculated from the retrieval fragments in the enhanced prompts; Based on the preliminary maintenance decision plan, compliance testing and reverse verification of knowledge graph parameters are performed to generate the final intelligent maintenance decision.

[0005] Secondly, this application provides an intelligent maintenance decision generation system, including: The enhanced prompt word determination module is used to perform dual-path recall retrieval and relevance reordering based on on-site fault condition data to obtain a set of knowledge fragments, and then perform context reconstruction to obtain enhanced prompt words; The preliminary maintenance decision-making module is used to obtain a fault cause chain by performing chain-like reasoning through a vertical domain model based on the enhanced prompt words, and to perform semantic reorganization by combining a set of hard constraints to obtain a preliminary maintenance decision-making plan; wherein, the reasoning path of the chain-like reasoning is constructed based on the state transition probabilities calculated from the retrieval fragments in the enhanced prompt words; The final intelligent maintenance decision generation module is used to perform compliance checks and reverse verification of knowledge graph parameters based on the preliminary maintenance decision plan, and generate the final intelligent maintenance decision.

[0006] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described intelligent maintenance decision generation method.

[0007] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described intelligent maintenance decision generation method.

[0008] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described intelligent maintenance decision generation method.

[0009] According to the specific embodiments provided in this application, this application has the following technical effects: This application obtains a high-quality set of knowledge fragments by performing dual-path recall retrieval and relevance reordering based on on-site fault data, and then performs context reconstruction to obtain enhanced prompt words. This achieves accurate extraction and structured encapsulation of multi-source heterogeneous fault information, overcoming the shortcomings of traditional RAG technology in mechanically piling up search results, and providing high-quality, highly relevant contextual support for subsequent reasoning. At the same time, based on the enhanced prompt words, chain-like reasoning is performed through a vertical domain model to obtain the fault cause chain, and semantic reorganization is performed in combination with a set of hard constraints to obtain a preliminary maintenance decision plan. Then, compliance testing and reverse verification of knowledge graph parameters are performed based on the preliminary maintenance decision plan to obtain the final intelligent maintenance decision. This effectively eliminates the "knowledge illusion" of large language models in vertical domains, ensuring that the final output maintenance decision plan strictly complies with industry safety standards and physical logic, and improves the scientificity, rigor, and safety of maintenance decisions in complex fault scenarios. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A flowchart illustrating an intelligent maintenance decision generation method according to an embodiment of this application; Figure 2 A flowchart illustrating the overall technical solution provided in one embodiment of this application. Detailed Implementation

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] To make the objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0014] Maintenance data in vertical domains is typically characterized by scarce samples, dense technical terminology, and a large amount of proprietary knowledge. Existing technical solutions often struggle to balance the relationship between "real-time knowledge updates" and "deep logical reasoning." When complex situations arise at the maintenance site, such as sudden equipment failures or multiple concurrent failures, ensuring that the system can both obtain the most accurate reference information based on its local proprietary knowledge base and, through a deep understanding of maintenance domain logic, intelligently refine, analyze, and logically verify the search results to produce solutions with priority ranking and compliance with industry standards has become a core bottleneck in improving intelligent maintenance decision-making. Therefore, building a decision-making system that can achieve deep coupling between RAG (Reliable Equipment Architecture) and large-scale models in vertical domains, and possess efficient technical collaboration capabilities, is of paramount practical significance for reducing equipment downtime and lowering maintenance costs.

[0015] like Figures 1-2 As shown in the figure, this application provides an intelligent maintenance decision generation method, including: S1. Based on the on-site fault condition data, perform dual-path recall retrieval and relevance reordering to obtain a set of knowledge fragments, and then reconstruct the context to obtain enhanced prompt words.

[0016] S2, based on the enhanced prompt words, chain-like reasoning is performed through the vertical domain model to obtain the fault cause chain, and semantic reorganization is performed in combination with the set of hard constraints to obtain a preliminary maintenance decision scheme; wherein, the reasoning path of the chain-like reasoning is constructed based on the state transition probability calculated from the retrieval fragments in the enhanced prompt words.

[0017] S3. Based on the preliminary maintenance decision plan, perform compliance checks and reverse verification of knowledge graph parameters to generate the final intelligent maintenance decision.

[0018] In an exemplary embodiment, S1 specifically includes: S11. Extract core entities from the on-site fault data and perform sparse retrieval to obtain a sparse retrieval result set. Then, map the metadata tags to a high-dimensional vector space for dense retrieval to obtain a dense retrieval result set. The core entities include equipment components, fault phenomena, and fault codes. The metadata tags include keyword vectors, semantic feature values, and fusion features of the geometric topology features and spatial location attributes of the components. The dimension of the high-dimensional vector space is higher than a set dimension threshold.

[0019] S12, perform weighted fusion based on the sparse retrieval result set and the dense retrieval result set to obtain an initial knowledge fragment set.

[0020] S13, Based on the initial set of knowledge fragments, the RRF base score and risk weight are calculated using the high-risk weighted word adjustment mechanism to obtain the set of knowledge fragments; wherein, the risk probability of high-risk words is higher than a set probability threshold.

[0021] In an exemplary embodiment, S13 specifically includes: S131, Rank the sparse search result set and the dense search result set, and calculate the RRF base score.

[0022] S132, If a fault description is found to contain high-risk words, calculate the risk weight and increase the attenuation coefficient of the safety procedure document.

[0023] S133, calculate the final score based on the RRF base score and the attenuation coefficient.

[0024] S134. Sort the knowledge fragments from highest to lowest according to the final score to obtain the set of knowledge fragments.

[0025] In practical applications, S1 achieves accurate extraction and structured encapsulation of fault information through a progressive process of multi-source data preprocessing, dual-path recall retrieval, relevance reordering, and context reconstruction. Specifically, this includes: 1.1 Preprocessing and operatorization of multi-source heterogeneous data.

[0026] When maintenance personnel input the on-site fault condition, the system first initiates a multi-dimensional search mechanism. For different types of business data, the system employs differentiated preprocessing operators for feature parsing, outputting standardized metadata tags to provide a unified data foundation for subsequent searches. For text-based data: A recursive character text splitting algorithm (RecursiveCharacterTextSplitter) is used, combined with the hierarchical structure of the maintenance manual (such as chapters, titles, and steps), to segment it into semantically complete knowledge chunks. The segmentation window size is set to 200 characters, and the overlap window size is set to 50 characters (using the sliding window technique), preserving the context overlap area to prevent semantic breaks; simultaneously, semantic feature values ​​are calculated for each knowledge chunk. The formula is ,in This represents the number of words in the knowledge block. Assign weights to words (based on a dictionary of maintenance terminology). This represents the word vector of a word.

[0027] Multimodal data: For 3D models and images, an improved ResNet-50 visual feature extraction network is used to extract the geometric topological features of the parts. Spatial location attributes merging into ( (These are weighting coefficients). For audio data, Automatic Speech Recognition (ASR) technology combined with the Viterbi Algorithm is used for text transcription and keyword correction. The state transition probability of the Viterbi Algorithm is... The keyword vectors were extracted after training on a speech corpus in the maintenance field and then transcribed. The aforementioned multimodal features are uniformly mapped to a multimodal alignment space, transforming them into semantically searchable metadata tags, i.e. .

[0028] 1.2 Based on the metadata tags obtained in 1.1 above, perform a "dual-path recall" retrieval to obtain an initial set of knowledge fragments related to the fault.

[0029] During the retrieval execution phase, the system employs a hybrid dual-path recall strategy of "keywords + vectors" to balance retrieval accuracy and generalization. The results from the two paths are weighted and fused to obtain an initial set of knowledge fragments. The specific implementation process is as follows: Sparse retrieval (keywords): A vertical domain-specific Named Entity Recognition (NER) model, fine-tuned based on BERT, is used to accurately extract core entities from fault descriptions. The specific process is as follows: 1. Perform word segmentation, part-of-speech tagging, and stop word removal on the user-input fault description text to obtain preprocessed text. .

[0030] 2. Input the finely tuned BERT model, and the model outputs the entity label probability distribution for each token. Filter by threshold ( Extract core entities such as equipment components, fault symptoms, and fault codes. .

[0031] 3. Utilize the BM25 algorithm to perform exact matching in the inverted index. The BM25 scoring formula is as follows: ,in, For knowledge base documents, For entity query sets, , These are empirical parameters. For entities In the document word frequency in This represents the total number of documents in the knowledge base. For containing entities The number of documents, This represents the average length of knowledge base documents.

[0032] 4. Filter documents with a BM25 score ≥ 1.5, extract knowledge fragments from them, and form a sparse search result set. .

[0033] Dense retrieval (vector): Utilizing the Sentence-BERT embedding model, the fault description text and the metadata tags obtained in 1.1 are combined. The fault vector is obtained by mapping the two vectors together to a high-dimensional vector space (with a dimension of 768). The retrieval process uses the HNSW (Hierarchical Navigable Small World) algorithm to find similar knowledge fragments in the vector database. The specific process is as follows: 1. Build the HNSW index and set the layer parameters. , .

[0034] 2. With To query the vector, we traverse from the top layer of HNSW, filtering candidate nodes layer by layer, and finally obtain the bottom-level candidate vector set.

[0035] 3. Calculate the candidate vector and The cosine similarity is calculated using the formula: ,in , This is the vector representation of the candidate knowledge fragment.

[0036] 4. Select knowledge fragments with a cosine similarity ≥ 0.75 to form a dense search result set. .

[0037] 5. Merge the results from the two searches to obtain an initial set of knowledge fragments. ( (For fusion weights).

[0038] 1.3 Based on the initial knowledge fragments retrieved in 1.2 above, re-ranking is performed based on the relevance of the cross encoder.

[0039] The retrieved raw fragments will enter the "relevance re-ranking" stage based on the cross-encoder architecture. Combined with a dynamic weight-aware factor, the relevance of the knowledge fragments will be improved. The specific calculation process is as follows: The system employs the RRF (Reciprocal Rank Fusion) scoring fusion algorithm to process the initial set of knowledge fragments. The search results are then re-ranked, and a high-risk weighted keyword adjustment mechanism is introduced. First, the sparse search results are... and dense search results Assign rankings respectively and (Ranking starts from 1, with lower values ​​indicating higher relevance); then, calculate the relevance for each knowledge fragment. Baseline RRF score: ,in To adjust the parameters.

[0040] Handling of high-risk weighted words: The system pre-builds a high-risk word library. The system performs word matching on the user's fault description and detects high-risk words. Then calculate the risk weight. ( To determine the frequency of high-risk keywords (maximum 3), the decay coefficient for security procedure documents will be automatically increased. Adjust the formula to (initial ); Final score Sort the data by score from highest to lowest to obtain a set of high-quality knowledge fragments. .

[0041] 1.4 Based on the first N high-quality knowledge fragments obtained in step 1.3 above (N is 10, which can be adaptively adjusted according to the complexity of the equipment), the context is reconstructed according to a specific maintenance domain logic template to obtain enhanced prompt words.

[0042] The reconstruction process employs an attention-based long context compression technique to remove redundant noise and inject industry-standard constraints. The specific compression process and calculations are as follows: 1. Collections of high-quality knowledge fragments Each segment Calculate its relationship with the fault vector attention weights ,in For fragments The vector representation of .

[0043] 2. Calculate redundancy ,like If the attention weight is high, then the segments with higher attention weights are retained, and redundant segments are removed.

[0044] 3. Inject industry standard constraints, extract core information such as safety operating procedures and technical parameters from the fragments, and integrate them with information such as fault descriptions and equipment models to form compressed contextual text. Compression rate is controlled at 40%-60% (to ensure semantic integrity).

[0045] 4. Encapsulate it as an enhanced prompt word conforming to JSON schema.

[0046] ,in For the fault description entered by the user, For equipment model, This is a summary of the corresponding industry's safety operating procedures.

[0047] In an exemplary embodiment, based on the enhanced prompt obtained in S1, logic injection and reasoning are performed through a vertical domain model to obtain a preliminary maintenance decision scheme. S2 specifically includes: S21, Based on the enhanced prompt words, perform domain semantic alignment using a low-rank decomposition matrix to obtain aligned domain semantics.

[0048] S22, based on the aligned domain semantics, map the fault phenomenon and the retrieved fragment into a state transition vector, calculate the state transition probability, and obtain the fault cause chain.

[0049] S23. Based on the fault cause chain and the set of hard constraints, perform semantic recombination and constraint verification to obtain a preliminary maintenance decision scheme.

[0050] In one exemplary embodiment, after S23, the method further includes: S24. Based on the preliminary maintenance decision plan, calculate the semantic overlap with the local knowledge base, the compatibility with industry standards, and the historical repair success rate.

[0051] S25, calculate the comprehensive confidence score based on the semantic overlap, the fit, and the historical repair success rate, prioritize the solutions based on the comprehensive confidence score, and output the solution matrix as a preliminary maintenance decision solution.

[0052] In practical applications, S2 takes enhanced prompts as input and, through a progressive process of Low-Rank Adaptation (LoRA) domain alignment, chain-like reasoning, semantic restructuring constraints, and confidence evaluation, achieves a leap from "information piling up" to "logical deduction." Each sub-step is connected logically through parameter passing to ensure the rationality and professionalism of the output solution. The specific process is as follows: 2.1 Based on the enhanced prompts obtained from S1, LoRA domain semantic alignment is performed through efficient parameter fine-tuning.

[0053] During the fine-tuning phase, the LoRA algorithm is employed. By adding a low-rank decomposition matrix alongside the pre-trained weight matrix, parameterized injection of specific maintenance logic is achieved, providing domain knowledge support for subsequent inference. Specifically, the weight matrix of the pre-trained large model is... ( For input dimensions, (For output dimension), LoRA adds a low-rank matrix. and ( For low-rank dimensions, take much smaller and The weight matrix of the model after fine-tuning is: .

[0054] The objective optimization function is expressed as: ,in This is a local troubleshooting dataset (containing paired data of fault descriptions, reference knowledge, and standard repair solutions). The output of the fine-tuned model. For standard repair program labels, This is the cross-entropy loss function. Through this mechanism, the model receives the enhanced prompts constructed in step one. At the same time, it can automatically identify the coupling relationship between equipment components and accurately match the semantics of the maintenance field.

[0055] 2.2 Chain Thinking (CoT) and Derivation of Fault Cause Chain.

[0056] This works closely with the LoRA domain semantic alignment in section 2.1: LoRA alignment provides domain semantic support for chain-like reasoning, ensuring that the reasoning process aligns with the logic of the maintenance industry; chain-like reasoning transforms the LoRA-aligned domain knowledge into explicit fault reasoning paths. The specific coordination and derivation process are as follows: The model received the encapsulated fault phenomenon. With search fragments Following the enhanced prompt, based on the domain semantic understanding capability after 2.1 LoRA alignment, the "chain thinking" procedure is initiated, explicitly transforming the reasoning process into a series of state transition vectors: ,in (Initial fault symptoms) The root cause of the failure, the state transition probability The output of the model after LoRA fine-tuning is calculated using the following formula: ,in To retrieve the state in the fragment Related knowledge, For the first The set of candidate states for the step.

[0057] The model transforms vague descriptions into logical chains with causal relationships by comparing and analyzing retrieved historical maintenance records with the current equipment status. For example, it can start with "abnormal pressure" as the initial node. Combined with the retrieved sensor thresholds The "sensor drift" was derived. ) or "pipe leak" probability distribution of ) , This allows us to construct a reasoning path from phenomenon to essence. (like higher).

[0058] 2.3 Semantic Reorganization and Hard Constraint Boundary Control.

[0059] The fault cause chain derived in section 2.2 provides a logical framework for semantic reorganization. Semantic reorganization then integrates the retrieved fragmented information based on this framework, and ensures that the reorganization result conforms to industry standards through hard constraint boundary control. The core role of the hard constraint set is to provide inviolable technical standards for the reasoning process, avoiding the generation of solutions that violate safety procedures or physical logic. The specific process is as follows: During the inference generation phase, the vertical domain model utilizes the Named Entity Recognition (NER) operator to identify the retrieved fragments. and key technical indicators in the fault cause chain (such as rated current) Standard torque Pressure threshold (etc.), and define it as a set of hard constraints in the reasoning process: ,in For safety operation requirements that must be performed (such as "power off operation" and "pressure relief operation").

[0060] The subsequent use of the hard constraint set is as follows: 1. During semantic reorganization, the model integrates fragmented retrieval information according to the fault cause chain logic in section 2.2, ensuring that the reorganized content matches the constraint set. Consistent.

[0061] 2. If the maintenance steps generated by the model involve numerical values ​​(such as tightening torque or test voltage), the system will perform constraint verification. The verification formula is as follows: ,in, The numerical values ​​generated for the model, To constrain the standard values ​​in the set, To allow for tolerances.

[0062] 3. If multiple retrieved data points conflict, the model will use the "source weight scoring table" (source weights) learned during the fine-tuning phase. The original manufacturer's manual (0.9), expert case studies (0.8), and general repair records (0.6) were weighted and extracted to determine the optimal solution. Remove outdated or irrelevant information.

[0063] 4. When faced with multiple concurrent faults, the model uses Bayesian network logic to analyze the coupling relationship between each fault point and, combined with the technical standards in the constraint set, identifies the primary fault (root cause).

[0064] 2.4 Confidence evaluation and priority ranking of decision-making options.

[0065] The relationship between this step and the steps above is as follows: taking the semantic foundation provided by LoRA domain alignment in 2.1, the failure cause chain reasoning results in 2.2, and the semantic reorganization and constraint verification results in 2.3 as input, the reliability of the scheme is quantified by confidence evaluation, and the optimal execution path is output by priority ranking. The specific process and calculation formula are as follows: The logic injection mechanism includes an automated "solution confidence evaluation algorithm," which the system uses to generate solutions. With local knowledge base semantic overlap Industry standards Adaptability and the success rate of historical restoration Calculate the overall confidence score: ,in , , Preset weighting coefficients (satisfying) ).

[0066] The physical meaning of each parameter: 1. Semantic overlap : Generating a scheme With local knowledge base The mean cosine similarity of related knowledge fragments, with a value range of [0,1], reflects the degree of knowledge support for the scheme.

[0067] 2. Industry standard compatibility : Generating a scheme In accordance with industry safety operating procedures The degree of matching is calculated by comparing the number of steps in the solution with the requirements of the specification. The value range is [0,1], which reflects the compliance of the solution.

[0068] 3. Historical restoration success rate : with generation scheme The success rate of repairs in similar historical repair cases, with a value range of [0,1], reflects the practical feasibility of the solution.

[0069] The final output is not a single path, but a matrix of solutions ranked based on confidence scores. ( The sorting rule is as follows: Each solution includes fault phenomenon confirmation, preparation of necessary tools, detailed troubleshooting steps, and verification of expected results, ensuring that the output results have both the accuracy of local private knowledge and the generalization reasoning ability of large models in vertical domains.

[0070] In an exemplary embodiment, based on the preliminary maintenance decision plan obtained in S2, adjustments are made through compliance verification and priority optimization to obtain the final maintenance decision. S3 specifically includes: S31, the operation steps in the preliminary maintenance decision scheme are converted into standard action vectors, and the similarity is calculated with the action vectors in the standard library. If the similarity result is lower than the preset threshold, the safety actions are supplemented to obtain the aligned scheme.

[0071] S32, based on the numerical mapping of the aligned scheme to the edge attributes of the knowledge graph, backtrack to obtain the standard parameters and tolerance range, establish a verification function for verification, and obtain a verification compliant scheme.

[0072] S33. Based on the aforementioned compliance verification scheme, the priority scoring function is calculated using the risk weight matrix, and the ranking is optimized to obtain the final intelligent maintenance decision.

[0073] S3 takes the initial maintenance decision plan as input and ensures the compliance, accuracy, and feasibility of the final plan through a closed-loop process of compliance testing, knowledge graph parameter verification, dynamic priority adjustment, and structured output. Each sub-step achieves gradual optimization of the plan through logical connection and parameter passing. The specific process is as follows: 3.1 Preliminary maintenance decision based on S2 It is processed by a compliance detection operator and semantic alignment is performed.

[0074] The system outputs preliminary solutions based on a large-scale model in a vertical domain. Then, it is input into a "compliance check operator" built based on industry standards. This operator is pre-loaded with a library of mandatory Standard Operating Procedures (SOPs) for the vertical domain. The system employs Named Entity Recognition (NER) and Dependency Parsing techniques to transform the operation steps in the generated scheme into standard action vectors. ( For the first (each operation step), and the action vectors in the specification library. ( For the first Calculate the cosine similarity of each standardized action: Set a similarity preset threshold If the system finds that the solution lacks prerequisite safety actions such as "power cut-off" and "pressure relief" (i.e., there are no corresponding...), satisfy The system will automatically trigger the correction operator to forcibly complete the missing action to the beginning of the scheme sequence, and at the same time perform semantic alignment on the completed steps to ensure logical coherence between steps (such as adjusting the operation prerequisite description of subsequent steps in sync after completing "cut off the power").

[0075] 3.2 For the aligned solution obtained in the above steps, compliance verification is performed using the knowledge graph parameter reverse verification algorithm.

[0076] For specific numerical values ​​appearing in the solution (such as tightening torque, test voltage, etc.), the system employs a parameter verification model based on a knowledge graph. The system will then use the numerical values ​​generated by the model... The mapping is performed as edge attributes in the graph (e.g., "tightening torque -120 N·m"), and then backtracks to the original technical manual node in the local knowledge base. Obtain the standard parameters and tolerance ranges specified in the manual.

[0077] The system establishes a verification function: , in The standard value specified in the manual. The tolerance range (as specified in the technical manual, such as tightening torque tolerance) ).if (Right now If the error message is below the safety threshold or exceeds the tolerance range, the verification mechanism will immediately intercept it and apply the hard specifications from the original manual. The solution undergoes mandatory coverage correction. This dual verification mode of "inference + mapping verification" ensures the rigor and accuracy of the final solution in terms of technical parameters.

[0078] 3.3 Based on the above compliance verification scheme, the sorting is optimized by using a priority dynamic adjustment algorithm.

[0079] After passing the compliance verification, the system introduces a risk weight matrix. ( The number of candidate solutions, For risk assessment dimensions (including operational risk, spare parts cost, time consumption, etc.), a final optimization and ranking is performed on all generated candidate solutions. The priority scoring function is defined as: ,in, The confidence level of the scheme calculated in step 2.4, The weights for each risk dimension are: operational risk 0.5, spare parts cost 0.3, and time consumption 0.2. For the first The first scheme is in the Risk values ​​for each risk dimension (range [0,1]).

[0080] Among them, risk level It is a probability factor calculated based on a historical accident database, and the calculation formula is as follows: , For the first The first scheme is in the The number of historical accidents under each risk dimension For the first The first scheme is in the The system calculates the total number of historical executions under each risk dimension. For example, when multiple troubleshooting paths exist, the system calculates the expected risk value for each path and prioritizes paths that have been historically validated and have low operational risk. A robust solution with a value ≤0.1 is provided, and "Key Quality Control Points (KCPs)" are highlighted in the interactive interface.

[0081] 3.4 Structured sequence output of decision instructions.

[0082] Finally, the verification mechanism transforms the validated logical chain into a structured, standardized decision template. This template is automatically populated by a template engine, ensuring that the output meets the safety asset management requirements of industrial production, including fault location, protection level, operating procedures, and acceptance criteria. This closed-loop verification and sorting mechanism truly transforms massive amounts of local data into technical instructions that are both practically applicable to the field and possess industrial-grade reliability, thereby reducing equipment downtime.

[0083] The technical solution of this application will be further illustrated below with specific examples.

[0084] Example 1: Implementation process of logical priority determination when multiple faults occur concurrently.

[0085] This embodiment details how, in complex operating conditions where multiple equipment failures occur simultaneously, the system leverages the synergistic effect of RAG and a large vertical domain model to perform logical analysis and determine maintenance priorities.

[0086] 1. Initial input and feature extraction of complex fault phenomena.

[0087] When maintenance personnel encounter multiple faults in the hydraulic system, such as "abnormal fluctuations in output pressure" accompanied by "excessive main pump housing temperature," they input the above description through the system's interactive interface. The enhanced context building module in the system's backend first performs multi-dimensional parsing of the input. At this point, the system not only extracts the two core feature values ​​of "pressure" and "temperature," but also automatically calls the real-time data monitoring interface of the sensors (if available) to obtain the environmental parameters at the time of the fault. These dynamic features are then initially aligned with the static "Hydraulic System Maintenance Manual" to form an initial fault vector.

[0088] 2. RAG-based retrieval of associated fault chains.

[0089] The system initiates a multi-path retrieval mechanism. The RAG module searches the local knowledge base not only for troubleshooting steps related to "low pressure" but also for related cases of "high temperature." More importantly, the system performs a "cross-index retrieval," searching historical records for compound fault cases where "both pressure and temperature are abnormal simultaneously." The search results show that in the historical maintenance records of a certain type of pump, there is a correlation between "internal leakage" causing a pressure drop and simultaneously generating frictional heat. The RAG module prioritizes and weights these logically related knowledge chunks, providing background support for subsequent large-scale model inference.

[0090] 3. Root Cause Reasoning for Vertical Domain Models.

[0091] After acquiring the enhanced context, the large model for the vertical domain begins to execute logic-injected reasoning. The model does not simply treat the two failures as isolated events, but rather simulates the "causal chain deduction" of an expert.

[0092] Logical deduction: The model, through reasoning and analysis, concludes that repairing the "high temperature" issue first (such as replacing the radiator) may not resolve the "low pressure" problem. Furthermore, if the original fault is an internal leak, continued operation will cause the newly replaced radiator to overload again. Conversely, if the original fault is an internal leak, repairing the seal will restore normal pressure and eliminate frictional heat, thus resolving both faults simultaneously.

[0093] Priority determination: Based on the above logic, the vertical domain model determines "internal leakage of the system" as the primary fault, and sets its corresponding maintenance priority as "extremely high (P0)", while "abnormal oil temperature" is marked as a "derivative symptom" and its priority is set as "medium (P2)".

[0094] 4. Solution generation under the constraints of industry standards.

[0095] Before outputting recommendations, the system enters a verification process. The system reviews the "Safety Operation Specifications" in the local knowledge base and identifies that when multiple faults occur concurrently and involve high temperature and high pressure, "pressure relief shutdown" and "oil sampling and analysis" must be performed as mandatory prerequisite steps. Based on this, the vertical domain large model modifies the output plan, placing "safe shutdown" at the top of all maintenance steps.

[0096] 5. Interactive presentation of priority sorting scheme.

[0097] Ultimately, the maintenance personnel's terminal interface presented a logically ordered, comprehensive solution: Preferred solution (priority 1): Investigate and replace the seals to address the internal leak. System prompt: "Pressure and temperature coupling detected, internal leak probability determined to be 85%, it is recommended to prioritize this option for comprehensive repair."

[0098] Secondary option (priority 2): If the internal leakage is found to be correct, then perform a cooling system circulation efficiency test.

[0099] Safety warning: The torque parameter constraints and high-voltage operation protection requirements (120 N·m, etc.) are clearly marked in a prominent position.

[0100] 6. Decision-making loop and knowledge accumulation.

[0101] When the maintenance personnel completed the repair according to priority 1 and reported "pressure rebound and temperature begins to drop," the system verified the accuracy of the logical judgment. This successful "multi-fault collaborative handling case" will be automatically tagged with "correct logical priority" and fed back to the reinforcement learning module of the local knowledge base as a "golden case" for handling similar concurrent faults in the future, realizing a closed loop from "single-time assistance" to "system self-evolution."

[0102] Example 2: The continuous iteration process of dynamic local knowledge base.

[0103] This embodiment details how the system uses the daily interaction behavior of maintenance personnel to achieve incremental updates of the local knowledge base and continuous performance optimization of vertical domain models, thereby forming a closed loop of "knowledge assets" with specific industry attributes.

[0104] 1. Capture and fragmented collection of original maintenance logs.

[0105] After the maintenance work is completed, the system automatically enters the knowledge accumulation phase. The maintenance personnel input the final result of the maintenance in the dialog interface, for example: "Inspection revealed that the problem was a loose pin inside the cable connector, not a sensor malfunction. After tightening the pin, the system pressure returned to normal." At this point, the dynamic data acquisition module in the system's background captures this interactive text and associates it with the initial fault phenomenon of the task, the reference suggestions retrieved by the system, and the final practical result. Using natural language processing technology, the system extracts key features from this unstructured log, such as "core component: cable connector," "root cause of fault: loose pin," and "repair action: tightening."

[0106] 2. Knowledge conflict detection and consistency verification.

[0107] Before newly generated maintenance experience is incorporated into the local knowledge base, it must be evaluated by a conflict detection algorithm. The system compares the vector representation of the new case with similar cases in the existing knowledge base. If a logical conflict is found between the new experience and the existing manual (e.g., the manual suggests replacing the sensor, but actual operation shows that tightening the pins solves the problem), the system will not directly overwrite the original knowledge, but will mark the case as "expert experience correction". The vertical domain model will then perform a logical check on the conflict: the model analysis considers loose pins to be a latent mechanical fault, and its abnormal electrical signal is extremely similar to that of a sensor fault. Based on this logic, the system will automatically generate a supplementary explanation, attach it below the original "sensor fault troubleshooting steps", and prioritize it as a "high-frequency error-prone troubleshooting item".

[0108] 3. Structured storage of knowledge base based on active learning.

[0109] To ensure efficient retrieval of new knowledge using RAG technology, the system performs "knowledge reconstruction" on fragmented logs using a large vertical domain model. The system fine-tunes the model through instructions, transforming the aforementioned maintenance logs into standardized "fault-cause-solution" triple structures and automatically generating corresponding summaries and index keywords. Subsequently, the system synchronously stores this structured data in a local vector database and a relational database. In this embodiment, the system also automatically updates the edge relationships in the knowledge graph, establishing a new association link between "pressure anomaly" and "cable plug," and assigning it a weight value based on actual maintenance frequency. This approach ensures that when a similar fault occurs again, the system can prioritize recommending this proven, low-cost solution.

[0110] 4. Data augmentation and fine-tuning iterations based on model generation capabilities.

[0111] When the number of new cases accumulated locally reaches a certain threshold (e.g., every 50 new records), the system will automatically initiate the performance enhancement process of the vertical domain model. Due to the limited number of real repair cases, the system will utilize the generation capabilities of domestically developed large-scale basic models, using these new cases as seeds, to generate hundreds of logically related simulated data through data augmentation techniques.

[0112] Variant generation: The model simulates different repair personnel's tone and different fault combination scenarios (such as the behavior of loose pins under different ambient temperatures) to generate an expanded dataset.

[0113] Automated training: Utilizing these enhanced, locally specific datasets, the system initiates lightweight fine-tuning (such as LoRA or QLoRA) on a local area network server. This process requires no manual intervention and its main goal is to adjust the model's semantic understanding weights on specific device components, making the model more aligned with the language habits and maintenance preferences of the specific team in subsequent decisions.

[0114] 5. Quality assessment and graded protection of intellectual assets.

[0115] The system implements a hierarchical management logic for the knowledge assets generated through continuous iteration. Cases that have undergone multiple maintenance verifications and demonstrate high accuracy are automatically upgraded to "Gold Cases" and given the highest weight during retrieval; while data-enhanced samples generated by the model are labeled as "Inference Aid Data." Simultaneously, for sensitive vertical domain materials, the system strictly implements localized anonymization processing during iteration. All data involved in fine-tuning flows only within the local area network, and model weight updates are stored only on the local server. This mechanism ensures that the company's core maintenance technical secrets and experience data are not leaked during continuous iteration, truly achieving technological autonomy and control, and preserving and increasing the value of knowledge assets.

[0116] 6. Evaluation of decision-making effectiveness improvement after closed-loop feedback.

[0117] At the end of this embodiment, the system generates a performance comparison report after the iteration. By comparing the system's response speed, solution accuracy, and maintenance personnel's evaluation indicators for the same type of complex fault (such as latent cable faults) before and after the iteration, the report quantitatively demonstrates the improved decision-making efficiency brought about by "dynamic iteration." Experimental data shows that after a quarter of daily use and dynamic iteration, the system's diagnostic accuracy for occasional faults in specific equipment can be improved by more than 25%, reducing spare parts waste caused by misdiagnosis.

[0118] This application provides an intelligent maintenance decision generation system, including: The enhanced prompt word determination module is used to perform dual-path recall retrieval and relevance reordering based on on-site fault condition data to obtain a set of knowledge fragments, and then perform context reconstruction to obtain enhanced prompt words.

[0119] The preliminary maintenance decision-making module is used to perform chain-like reasoning based on the enhanced prompts, using a vertical domain model to obtain a fault cause chain, and then perform semantic reorganization in conjunction with a set of hard constraints to obtain a preliminary maintenance decision-making plan; wherein, the reasoning path of the chain-like reasoning is constructed based on the state transition probabilities calculated from the retrieval fragments in the enhanced prompts.

[0120] The final intelligent maintenance decision generation module is used to perform compliance checks and reverse verification of knowledge graph parameters based on the preliminary maintenance decision plan, and generate the final intelligent maintenance decision.

[0121] In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments. The computer device can be a server or a terminal. The computer device includes a processor, a memory, an input / output interface (I / O), and a communication interface. The processor, memory, and I / O interface are connected via a system bus, and the communication interface is connected to the system bus via the I / O interface. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device stores data to be processed. The I / O interface of the computer device is used for exchanging information between the processor and external devices. The communication interface of the computer device is used for communicating with an external terminal via a network connection. When the computer program is executed by the processor, it implements the above-described methods.

[0122] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0123] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0124] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0125] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by hardware related to computer program instructions. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0126] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0127] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0128] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. An intelligent maintenance decision generation method, characterized in that, include: Based on the on-site fault data, a dual-path recall retrieval and relevance reordering are performed to obtain a set of knowledge fragments, and then the context is reconstructed to obtain enhanced prompt words; Based on the enhanced prompts, chain-like reasoning is performed through a vertical domain model to obtain the fault cause chain, and semantic reorganization is performed in combination with a set of hard constraints to obtain a preliminary maintenance decision plan; wherein, the reasoning path of the chain-like reasoning is constructed based on the state transition probabilities calculated from the retrieval fragments in the enhanced prompts; Based on the preliminary maintenance decision plan, compliance testing and reverse verification of knowledge graph parameters are performed to generate the final intelligent maintenance decision.

2. The intelligent maintenance decision generation method according to claim 1, characterized in that, Based on the on-site fault condition data, a dual-path recall retrieval and relevance reordering are performed to obtain a set of knowledge fragments, specifically including: The core entities are extracted from the on-site fault data and sparse search is performed to obtain a sparse search result set. The metadata tags are then mapped to a high-dimensional vector space for dense search to obtain a dense search result set. The core entities include equipment components, fault phenomena, and fault codes. The metadata tags include keyword vectors, semantic feature values, and fusion features of the geometric topology features and spatial location attributes of the components. The dimension of the high-dimensional vector space is higher than a set dimension threshold. The initial set of knowledge fragments is obtained by weighting and fusing the sparse and dense search result sets. Based on the initial set of knowledge fragments, the RRF base score and risk weight are calculated using a high-risk weighted word adjustment mechanism to obtain the knowledge fragment set; wherein, the risk probability of high-risk words is higher than a set probability threshold.

3. The intelligent maintenance decision generation method according to claim 2, characterized in that, Based on the initial set of knowledge fragments, and combined with the high-risk weighted word adjustment mechanism, the RRF base score and risk weight are calculated to obtain the knowledge fragment set, which specifically includes: The sparse and dense search result sets are ranked, and the base RRF score is calculated. If a fault description is found to contain high-risk words, calculate the risk weight and increase the attenuation coefficient of the safety procedure document. The final score is calculated based on the RRF baseline score and the attenuation coefficient. Sort the knowledge fragments from highest to lowest according to the final score to obtain a set of knowledge fragments.

4. The intelligent maintenance decision generation method according to claim 1, characterized in that, Based on the enhanced prompts, chain-like reasoning is performed using a vertical domain model to obtain the fault cause chain. Then, semantic reorganization is performed using a set of hard constraints to obtain a preliminary maintenance decision plan, specifically including: Based on the enhanced prompt words, domain semantic alignment is performed using a low-rank decomposition matrix to obtain the aligned domain semantics; Based on the aligned domain semantics, the fault phenomenon and the retrieved fragment are mapped to state transition vectors, the state transition probability is calculated, and the fault cause chain is obtained. Based on the fault cause chain and the set of hard constraints, semantic recombination and constraint verification are performed to obtain a preliminary maintenance decision scheme.

5. The intelligent maintenance decision generation method according to claim 4, characterized in that, After semantic recombination and constraint verification based on the fault cause chain and hard constraint set to obtain a preliminary maintenance decision scheme, the following steps are also included: Based on the preliminary repair decision plan, calculate the semantic overlap with the local knowledge base, the compatibility with industry standards, and the historical repair success rate. A comprehensive confidence score is calculated based on the semantic overlap, the fit, and the historical repair success rate. The comprehensive confidence scores are then used to prioritize the solutions, and a solution matrix is ​​output as a preliminary maintenance decision-making scheme.

6. The intelligent maintenance decision generation method according to claim 1, characterized in that, Based on the preliminary maintenance decision plan, compliance checks and reverse verification of knowledge graph parameters are performed to generate the final intelligent maintenance decision, specifically including: The operation steps in the preliminary maintenance decision scheme are converted into standard action vectors, and the similarity is calculated with the action vectors in the specification library. If the similarity result is lower than a preset threshold, the safety actions are supplemented to obtain the aligned scheme. Based on the numerical mapping of the aligned scheme to the edge attributes of the knowledge graph, the standard parameters and tolerance range are obtained by backtracking, and a verification function is established for verification to obtain a compliant scheme. Based on the aforementioned compliance verification scheme, a priority scoring function is calculated using a risk weight matrix, and the scores are optimized and ranked to obtain the final intelligent maintenance decision.

7. An intelligent maintenance decision generation system, characterized in that, The intelligent maintenance decision generation method according to any one of claims 1-6 includes: The enhanced prompt word determination module is used to perform dual-path recall retrieval and relevance reordering based on on-site fault condition data to obtain a set of knowledge fragments, and then perform context reconstruction to obtain enhanced prompt words; The preliminary maintenance decision-making module is used to obtain a fault cause chain by performing chain-like reasoning through a vertical domain model based on the enhanced prompt words, and to perform semantic reorganization by combining a set of hard constraints to obtain a preliminary maintenance decision-making plan; wherein, the reasoning path of the chain-like reasoning is constructed based on the state transition probabilities calculated from the retrieval fragments in the enhanced prompt words; The final intelligent maintenance decision generation module is used to perform compliance checks and reverse verification of knowledge graph parameters based on the preliminary maintenance decision plan, and generate the final intelligent maintenance decision.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the intelligent maintenance decision generation method according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the intelligent maintenance decision generation method as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the intelligent maintenance decision generation method as described in any one of claims 1-6.

Images