Two-stage classification equipment view maintenance decision method based on knowledge graph driving

Abstract

This invention relates to the field of equipment maintenance and repair decision-making technology, and discloses a two-stage classification method for equipment condition-based repair decision-making driven by a knowledge graph. The method includes: acquiring multi-source text data; performing a first-stage classification of equipment based on an equipment importance evaluation model, and a second-stage classification of components based on maintenance strategies; using equipment entities, fault mode entities, and maintenance measure entities as knowledge ontology, embedding state variable nodes, and constructing a knowledge graph that integrates state variables using a long short-term memory network; combining a semantic context awareness mechanism and a graph structure stability constraint mechanism to output candidate fault nodes; retrieving candidate maintenance measures corresponding to the candidate fault nodes from the knowledge graph, verifying the candidate maintenance measures using soft and hard constraints, and using an execution orchestration generator to transform the maintenance measure list into an equipment repair decision scheme. This invention, combined with a knowledge graph, achieves intelligent repair decision-making, providing decision support for maintenance personnel.

Description

Technical Field

[0001] This invention relates to the field of equipment maintenance and repair decision-making technology, specifically to a two-stage classification method for condition-based equipment repair decision-making driven by knowledge graphs. Background Technology

[0002] During operation, industrial equipment frequently experiences various faults, including mechanical, electrical, and hydraulic failures. Existing equipment maintenance processes largely rely on manual experience and paper manuals, requiring manual retrieval and comparison of unstructured fault reports, which easily leads to information omissions and diagnostic delays. Current technologies often employ single-dimensional classification methods for equipment and its components, such as classification based solely on consumption or price, or simple categorization based on component maintenance characteristics. They lack a two-stage classification mechanism that considers both equipment importance and maintenance strategy, making it difficult to achieve differentiated management and optimized allocation of maintenance resources between critical and general equipment, and between components requiring passive and preventative maintenance. As a result, fault diagnosis and maintenance decisions in condition-based maintenance scenarios remain highly dependent on manual experience, lacking automation and precision, and failing to guarantee the scientific rigor and timeliness of maintenance decisions. Therefore, how to leverage multi-source operational data, introduce a two-stage classification approach based on equipment importance and maintenance strategy, and establish an intelligent method to support condition-based maintenance decisions, achieving automated and precise equipment fault maintenance decisions, reducing manual retrieval costs, and improving the scientific rigor of equipment maintenance decisions, has become an urgent problem to be solved. Summary of the Invention

[0003] This invention provides a two-stage classification method for equipment condition-based maintenance decision-making driven by knowledge graphs, in order to solve the problem of how to make decisions on equipment fault maintenance.

[0004] In a first aspect, the present invention provides a two-stage classification-based equipment condition-based maintenance decision-making method driven by knowledge graphs, the method comprising:

[0005] Acquire multi-source text data and preprocess the multi-source text data;

[0006] The equipment is classified in the first stage to obtain critical equipment and non-critical equipment. The components are classified in the second stage to obtain components for passive maintenance and preventive maintenance. The equipment classification results and component classification results are cross-processed to obtain the two-stage classification results.

[0007] Entities and relationships between entities are extracted from preprocessed multi-source text data. Equipment entities, fault mode entities, and maintenance measure entities are used as knowledge ontology. Entities and relationships between entities are the core, and state variable nodes are embedded. The first-stage classification results are embedded in the equipment entity nodes. A knowledge graph that integrates state variables is constructed using a long short-term memory network. Relationships between entities include belonging relationships, causing relationships, and processing relationships.

[0008] Combining semantic context awareness and graph structure stability constraint mechanisms, a mechanism is established to prioritize and float candidate weights of key equipment nodes, thereby assigning priority to spare parts resource paths for preventive maintenance components and outputting candidate fault nodes.

[0009] Based on directed causal edges, candidate maintenance measures corresponding to candidate fault nodes are retrieved in the constructed knowledge graph. The candidate maintenance measures are verified by combining soft and hard constraints. The list of maintenance measures that pass the constraint verification is transformed into equipment maintenance decision schemes using an execution orchestration generator.

[0010] This invention avoids the problem of a lack of differentiated maintenance strategies caused by a one-size-fits-all approach by classifying equipment and components in two stages. Based on the two-stage classification results, entities and relationships between entities are extracted from multi-source text data, and state variables are embedded, transforming the knowledge graph from a static knowledge graph to a dynamic state-aware one, improving its adaptability to different scenarios. By combining a semantic context awareness mechanism and a graph structure stability constraint mechanism, fault nodes are accurately located. Targeted causal retrieval in the knowledge graph enables precise binding of maintenance measures and faults, achieving intelligent maintenance decision-making and providing decision support for maintenance personnel.

[0011] Secondly, the present invention provides a knowledge graph-driven two-stage classification device for condition-based maintenance decision-making, the device comprising:

[0012] The acquisition module is used to acquire multi-source text data and preprocess the multi-source text data;

[0013] The classification module is used to perform the first stage classification of equipment to obtain critical equipment and non-critical equipment, and the second stage classification of parts to obtain parts for passive maintenance and preventive maintenance. The equipment classification results and the parts classification results are cross-processed to obtain the two-stage classification results.

[0014] The knowledge graph construction module is used to extract entities and relationships between entities from preprocessed multi-source text data. It uses equipment entities, fault mode entities, and maintenance measure entities as knowledge ontology, with entities and relationships between entities as the core, embedding state variable nodes. The first-stage classification results are embedded in the equipment entity nodes. A knowledge graph that integrates state variables is constructed using a long short-term memory network. The relationships between entities include belonging relationships, causing relationships, and processing relationships.

[0015] The candidate fault node output module is used to combine the semantic context awareness mechanism and the graph structure stability constraint mechanism to filter and float the candidate weights of key equipment nodes first, and to mark the priority of spare parts resource paths for preventive maintenance parts, and output candidate fault nodes.

[0016] The maintenance decision output module is used to retrieve candidate maintenance measures corresponding to candidate fault nodes in the constructed knowledge graph based on directed causal edges, verify the candidate maintenance measures by combining soft constraints and hard constraints, and use the execution orchestration generator to transform the list of maintenance measures that have passed the constraint verification into equipment maintenance decision schemes.

[0017] Thirdly, the present invention provides an electronic device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the knowledge graph-driven two-stage classification device condition-based maintenance decision-making method described in the first aspect or any corresponding embodiment above.

[0018] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the knowledge graph-driven two-stage classification device condition-based maintenance decision-making method described in the first aspect or any corresponding embodiment thereof. Attached Figure Description

[0019] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of a two-stage classification equipment condition-based maintenance decision-making method driven by knowledge graphs according to an embodiment of the present invention;

[0021] Figure 2 This is a schematic diagram of the overall architecture of a knowledge graph-driven two-stage classification equipment condition-based maintenance decision-making system according to an embodiment of the present invention;

[0022] Figure 3 This is a schematic diagram of the maintenance decision knowledge graph structure according to an embodiment of the present invention;

[0023] Figure 4 This is a schematic diagram of the evaluation hierarchy according to an embodiment of the present invention;

[0024] Figure 5 This is a structural block diagram of a two-stage classification device for condition-based maintenance decision-making based on knowledge graphs, according to an embodiment of the present invention.

[0025] Figure 6 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

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

[0027] Currently, equipment diagnostic methods commonly use rule bases, expert systems, and machine learning-based suggestion systems. However, these methods mostly rely on structured features and are difficult to process free text.

[0028] Unstructured text generated during the operation of existing industrial equipment cannot be directly utilized by traditional rule-based systems. Furthermore, the diverse equipment types, complex fault modes, and the inability of a single model to provide comprehensive coverage necessitate a combination of expert knowledge and data-driven approaches. Existing traditional solutions lack the ability to migrate across equipment and operating conditions and lack a unified knowledge representation framework. In existing industrial enterprises, equipment fault information is mostly unstructured text, making manual summarization time-consuming and labor-intensive. Existing systems struggle to automatically and accurately match unstructured fault descriptions with suitable maintenance solutions. Therefore, this invention provides a two-stage, hierarchical equipment condition-based maintenance decision-making method driven by knowledge graphs. This method enables automated and precise equipment fault diagnosis and maintenance decision-making, reducing manual retrieval costs and accelerating maintenance decision-making.

[0029] This embodiment provides a knowledge graph-driven two-stage classification method for equipment condition-based maintenance decision-making, which is applied to a knowledge graph-driven two-stage classification system for equipment condition-based maintenance decision-making. Figure 1 This is a flowchart of a knowledge graph-driven two-stage classification equipment condition-based maintenance decision-making method according to an embodiment of the present invention. The knowledge graph-driven two-stage classification equipment condition-based maintenance decision-making system is as follows: Figure 2As shown, the method flow includes the following steps:

[0030] Step S101: Obtain multi-source text data and preprocess the multi-source text data.

[0031] In this embodiment of the invention, data is collected from multiple sources of text, including the power generation system's operation logs, maintenance and fault work orders, equipment manuals and fault diagnosis manuals, and expert experience databases, to obtain multi-source text data. This multi-source text data undergoes preprocessing such as text cleaning, word segmentation, and part-of-speech tagging to identify key data such as equipment, components, and fault phenomena, providing a data foundation for subsequent knowledge extraction. The text cleaning aims to remove special characters, standardize terminology, and resolve issues such as the mixing of Chinese and English. Word segmentation and part-of-speech tagging are implemented using professional word segmentation tools such as jieba. Furthermore, entity annotation is performed on the multi-source text data to prepare training data. A batch of key entities from fault reports are manually annotated as a training set, providing a data foundation for subsequent expansion of the annotation scale using semi-supervised learning or remote annotation.

[0032] Step S102: Perform a first-stage classification of the equipment to obtain critical equipment and non-critical equipment; perform a second-stage classification of the components to obtain components for passive maintenance and preventive maintenance; and cross-process the equipment classification results and component classification results to obtain a two-stage classification result.

[0033] In this embodiment of the invention, to optimize the allocation of maintenance resources and address the shortcomings of traditional ABC classification and four-quadrant classification based on component maintenance characteristics, thereby improving the scientific rigor and rationality of component classification results in the production system, a two-stage classification is introduced. First, equipment is classified (first stage) to identify critical and non-critical equipment. This aims to help component managers focus their efforts and time on critical equipment. Then, components are classified (second stage), taking into account maintenance strategy factors, into components for reactive maintenance and components for preventative maintenance. Finally, the equipment classification results and component classification results are cross-referenced to form four categories of components.

[0034] It is understandable that the maintenance methods for equipment components can be divided into two types: passive maintenance and preventive maintenance. Passive maintenance means repairing after damage occurs, while preventive maintenance means repairing in advance to prevent damage.

[0035] Step S103: Extract entities and relationships between entities from the preprocessed multi-source text data. Using equipment entities, fault mode entities, and maintenance measure entities as knowledge ontology, embed state variable nodes with entities and relationships between entities as the core. Embed the first-stage classification results in the equipment entity nodes and construct a knowledge graph that integrates state variables using a long short-term memory network.

[0036] In this embodiment of the invention, entity recognition and relationship extraction are performed on the preprocessed multi-source text data. The entities include:

[0037] (1) Equipment entity: Classified according to hierarchical structure, including first-level, second-level and third-level equipment levels, such as main transformer, cooling system, circuit breaker, etc.;

[0038] (2) Failure mode entities: covering typical deterioration phenomena such as abnormal temperature, abnormal vibration, insulation aging, leakage, blockage, oil leakage, and wear;

[0039] (3) Physical maintenance measures: including technical means such as replacing parts, cleaning, adjusting parameters, verifying parameters, monitoring conditions, and preventive maintenance.

[0040] The specific maintenance measures define the specific maintenance methods and technical means, such as parts replacement, parameter adjustment, real-time status monitoring, and preventive maintenance measures.

[0041] The relationships between entities include:

[0042] (1) Belonging relationship: used to describe the hierarchical structure of a device and its sub-devices or specific components;

[0043] (2) Causation relationship: used to clarify the causal link between a device or component and a specific failure mode;

[0044] (3) Processing relationship: used to clarify the response logic between specific fault modes and corresponding maintenance measures.

[0045] By establishing three types of relationships between entities, the inherent logical connections between equipment, faults, and maintenance are clearly constructed and expressed. Using equipment entities, fault mode entities, and maintenance measure entities as knowledge ontologies, a three-hop semantic link of "equipment-fault-maintenance" is formed based on the entities and the relationships between them. At the same time, state variable nodes are embedded to realize health-driven knowledge reasoning.

[0046] The data sources for the knowledge graph include: operation logs, maintenance work orders, equipment fault manuals, expert databases, and health assessment model results. After initial graph data is constructed through preprocessing, entity recognition, and relation extraction techniques, it undergoes an expert review process and is then imported into a graph database for entity storage, querying, and maintenance. For example... Figure 3 As shown, Figure 3 This is a schematic diagram of the knowledge graph structure, which shows the structural hierarchy and semantic relationships of the three core entities: equipment, faults, and maintenance, effectively supporting dynamic decision-making on maintenance strategies under equipment conditions.

[0047] Step S104: Combining the semantic context awareness mechanism and the graph structure stability constraint mechanism, a mechanism is used to prioritize and float the candidate weights of key equipment nodes, and to mark the priority of spare parts resource paths for preventive maintenance parts, and then output candidate fault nodes.

[0048] In this embodiment of the invention, after the structured extraction of key entities, causal relationships and event chains in the fault description is completed, fault matching is combined with semantic context awareness mechanism and graph structure stability constraints. In a unified metric space, semantic similarity and graph structure correlation are comprehensively considered to achieve accurate identification of fault nodes and high-precision retrieval of interpretation subgraphs, providing a data foundation for the subsequent decision-making process.

[0049] First, the input fault text undergoes standardized preprocessing, including terminology normalization, stop word filtering, and context-enhanced modeling. Using an improved semantic representation method, operating status parameters, degradation trend information, and maintenance measures descriptions are incorporated into the semantic space. This ensures that semantic features not only cover the equipment itself but also reflect the contextual information within the maintenance scenario. The resulting semantic representation can more accurately depict the performance and evolution of actual faults.

[0050] Building upon this foundation, a multi-dimensional graph structure constraint mechanism is further introduced. Structural entropy is used to measure the importance of nodes in the graph, ensuring the stability of matched nodes within the overall structure. Node reachability metrics ensure that the retrieved explanatory subgraphs are logically traceable and deducible. Simultaneously, causal chain constraints are combined to verify the causal rationality of candidate nodes, preventing situations where semantic matching is reasonable but the logical chain is incomplete. In this way, the matching results are not only semantically reasonable but also structurally consistent and interpretable.

[0051] In addition, the system outputs the set of nodes most relevant to the target fault and generates a minimal explanatory subgraph based on this set. This subgraph visually represents the logical relationship between "fault mode - cause chain - maintenance measures," helping operations and maintenance personnel quickly understand the root cause of the problem and possible handling paths. Furthermore, the explanatory subgraph can be linked with historical case libraries and maintenance strategy libraries to integrate experiential knowledge with data-driven reasoning, improving the scientific rigor and operability of maintenance strategy recommendations.

[0052] To further enhance the priority orientation and resource targeting of fault matching, a two-stage classification result is introduced to establish a classification-driven dynamic filtering and weighting mechanism. The first-stage classification result is applied to prioritize and increase the candidate weights of key equipment nodes. The second-stage classification result is applied to prioritize the spare parts resource paths of preventive maintenance components and output the candidate fault nodes most relevant to the target fault.

[0053] Step S105: Based on the directed causal edges, retrieve the candidate maintenance measures corresponding to the candidate fault nodes in the constructed knowledge graph, verify the candidate maintenance measures by combining soft constraints and hard constraints, and use the execution orchestration generator to transform the list of maintenance measures that have passed the constraint verification into equipment maintenance decision schemes.

[0054] In this embodiment of the invention, starting from a candidate fault node, the direct causes and maintenance measures associated with the fault are located based on directed causal edges. "Maintenance nodes" representing processes, personnel qualifications, tools, components, time windows, and safety levels are introduced, collectively forming a four-layer sub-graph of "cause—measure—execution elements—inspection standards." Applicable operating conditions, resource capabilities, and safety compliance are used as hard constraints, while cost, downtime, and risk are used as soft constraints. Achievementability checks and multi-objective satisfiability determinations are performed on candidate paths. Measures that pass constraint verification are transformed into implementable operational sequences by an execution orchestration generator, and adaptive sorting is achieved by combining historical performance and online status, outputting a solution containing priorities, resource requirements, key evidence, and verification points. On-site feedback and execution results are fed back to the graph and results system through a governance mechanism for continuous optimization.

[0055] This embodiment provides a knowledge graph-driven two-stage classification method for equipment condition-based maintenance decision-making. By classifying equipment and components into two stages, it avoids the problem of a lack of differentiated maintenance strategies due to a one-size-fits-all approach. Based on the two-stage classification results, it extracts entities and relationships between entities from multi-source text data and embeds state variables, transforming the knowledge graph from a static knowledge graph into a dynamic state-aware one. This improves the adaptability to different scenarios. Combined with a semantic context awareness mechanism and a graph structure stability constraint mechanism, it accurately locates fault nodes. By utilizing directed causal retrieval in the knowledge graph, it achieves precise binding of maintenance measures and faults, realizing intelligent maintenance decision-making schemes and providing decision support for maintenance personnel.

[0056] This embodiment provides a two-stage classification method for condition-based maintenance decision-making of equipment driven by knowledge graphs. The process includes the following steps:

[0057] Step S201: Obtain multi-source text data and preprocess the multi-source text data.

[0058] Please see details Figure 1 Step S101 of the illustrated embodiment will not be described again here.

[0059] Step S202: Perform a first-stage classification of the equipment to obtain critical equipment and non-critical equipment; perform a second-stage classification of the components to obtain components for passive maintenance and components for preventive maintenance; cross-process the equipment classification results and component classification results to obtain a two-stage classification result.

[0060] Please see details Figure 1 Step S102 of the illustrated embodiment will not be described again here.

[0061] Step S203: Extract entities and relationships between entities from the preprocessed multi-source text data. Use equipment entities, fault mode entities, and maintenance measure entities as knowledge ontology. Embed state variable nodes with entities and relationships between entities as the core. Embed the first-stage classification results in the equipment entity nodes. Use a long short-term memory network to construct a knowledge graph that integrates state variables.

[0062] Specifically, the construction of the knowledge graph of fused state variables in step S203 above includes:

[0063] Step S2031: Construct a knowledge graph structure that integrates state awareness.

[0064] Step S2032: Combine the bidirectional long short-term memory network model and the sequence labeling model of conditional random fields to extract structured entities.

[0065] Step S2033: The extracted structured entities are verified based on the dual-mode verification mechanism of boundary matching heuristic rules.

[0066] Step S2034: Import the verified structured entities into the unified entity library to construct a knowledge graph that integrates state variables.

[0067] In this embodiment of the invention, based on previously accumulated glossaries, maintenance procedures, work permits, and inspection standards, key maintenance knowledge is structurally accumulated and semantically modeled to construct a multi-level ontology system covering "equipment—components—failure modes—causes—measures—resources—safety elements," and embeds health status variables and related reasoning attributes. On this basis, a bidirectional mapping mechanism between the outcome system and the knowledge graph is constructed to ensure semantic consistency and model coupling capability.

[0068] Based on the work procedures, acceptance standards, and resource list, executable knowledge units are extracted and constructed. These units use semantic atomization to associate and bind maintenance measures with operation steps, personnel qualifications, tools, parts, working hours, inspection standards, and applicable conditions. At the same time, health status tags (such as applicable health level ranges and degradation mode correspondences) are introduced as contextual constraint attributes to record the time, working conditions, source, and version in specific execution scenarios, supporting the dynamic reasoning and execution orchestration of subsequent maintenance paths.

[0069] To achieve the linked expression of maintenance knowledge and equipment operating status, an entity extraction process based on state semantic enhancement is constructed based on preprocessed multi-source text data. In the text preprocessing stage, a domain vocabulary and standard coding unified with the output system are first introduced to unify equipment aliases, term abbreviations, and units of measurement. Subsequently, the granularity of professional terminology is preserved during word segmentation and annotation, and a part-of-speech tagging method based on joint modeling of health description and fault status is adopted to improve the accuracy of key state word identification.

[0070] Constructing knowledge graph structures, such as Figure 4 As shown, the knowledge graph structure consists of a target layer, a criterion layer, and a scheme layer. A sequence labeling model combining a bidirectional long short-term memory network and a conditional random field is employed to extract structured entities containing state variables from multi-source text data, thereby avoiding the inefficiency and errors of manual extraction. To improve the accuracy of entity recognition, a dual-mode verification mechanism based on BM (Boundary-Matching) heuristic rules is introduced to verify the extracted structured entities. Finally, the verified entities are imported into a unified entity database to form a knowledge graph incorporating state variables.

[0071] Specifically, step S2031 includes:

[0072] Step S20311: Establish an evaluation hierarchical structure model with equipment importance as the target layer, key influencing factors and equipment factors as the criterion layer, and the impact of failures on production quality, the impact of failures on environmental health and safety, the impact of failures on utilization rate, the impact of failures on production, failure frequency, maintainability, detectability, and maintenance cost as the solution layer.

[0073] Step S20312: Using the equipment importance as the evaluation set, construct a fuzzy consistency judgment matrix, and construct a fuzzy comprehensive importance evaluation model based on the product of the fuzzy consistency judgment matrix and the fuzzy matrix corresponding to the evaluation set.

[0074] Step S20313: Determine the equipment classification result based on the maximum membership principle, embed the equipment classification result into the equipment entity attributes of the knowledge graph, add a new equipment status attribute field, semantically bind the equipment status attribute field with equipment nodes and fault nodes, and construct a knowledge graph structure with integrated status awareness.

[0075] In embodiments of the present invention, such as Figure 4As shown, an evaluation hierarchical structure model is established with equipment importance as the target layer, U1 key influencing factors and U2 equipment factors as the criteria layer, and a. the impact of failure on production quality, b. the impact of failure on EHS (Environment, Health, Safety), c. the impact of failure on utilization rate, d. the impact of failure on production, e. failure frequency, f. maintainability, g. detectability, and h. maintenance cost as the solution layer.

[0076] Using equipment importance as the evaluation set, for ,in Corresponding to the main key equipment, Corresponding to non-critical equipment.

[0077] Let the weight set be ,satisfy .

[0078] Construct a pairwise comparison judgment matrix, calculate the weights using the square root method based on the scaling factor, and aggregate the multi-layer weights. Here, the scaling factor is 1 for equal importance, 3 for slightly important, 5 for significantly important, 7 for very important, 9 for extremely important, 2, 4, 6, and 8 for intermediate values, and 1 / n Indicates that the opposite is important. n =3, 5, 7, 9).

[0079] First, calculate the consistency index (CI). , To determine the largest eigenvalue of a matrix, a random consistency index (RI) is established. The correspondence between the RI index and the matrix order is as follows: 1st order -0, 2nd order -0, 3rd order -0.58, 4th order -0.90, 5th order -1.12, 6th order -1.24, 7th order -1.35, 8th order -1.41, 9th order -1.45. Consistency checks are then performed, and the consistency ratio is calculated. If the consistency ratio is less than 0.1, the consistency check passes; otherwise, the check fails.

[0080] Constructing a fuzzy consistency judgment matrix ,in, For the first The factor (from sub-item U) corresponds to the first Each evaluation level ( The degree of membership of ).

[0081] Based on fuzzy consistency judgment matrix R fuzzy matrix corresponding to the evaluation set A The multiplication results are used to construct a fuzzy comprehensive importance evaluation model. , This is the evaluation vector.

[0082] The result is determined using the maximum membership principle: corresponding As the final level.

[0083] Embedding device classification results into device entity attributes of the knowledge graph supports dynamic query filtering, such as prioritizing extraction. The equipment's fault subgraph provides priority factors for subsequent hard constraints.

[0084] Based on the domain maintenance knowledge ontology, and taking equipment hierarchical structure, typical fault mechanisms, operating procedures, and inspection standards as core references, multi-dimensional entity expansion is performed. The entity categories, including "equipment—component—fault mode—operation and maintenance conditions—maintenance measures—resource capabilities," and their hierarchical, causal, and dependency relationships are clearly defined layer by layer. Simultaneously, to enhance the knowledge graph's responsiveness to operational status, equipment status attribute fields such as health variables, remaining lifespan, and degradation trends are added. These equipment status attribute fields are semantically bound to equipment nodes or fault nodes, forming a knowledge graph structure that integrates state awareness.

[0085] Before high-confidence entities and relation triples are entered into the database, terminology normalization, pronoun reference resolution, semantic conflict verification, and role relationship directionality verification are performed sequentially to ensure clear entity boundaries and unique semantic relationships. Simultaneously, to support contextual reasoning and conditional triggering of graph nodes, contextual attribute fields such as "source document, collection time, operating conditions, applicable scope, and data version" are set and uniformly entered into the graph. "Executable knowledge units" from the results system will be further mapped into semantic subgraphs of "fault cause—maintenance measures—execution elements," where execution elements include operating steps, personnel qualifications, tools, parts, time budgets, and inspection requirements, forming a closed-loop knowledge representation with traceability, verifiability, and executableness.

[0086] After semantic consistency filtering, the knowledge graph structure undergoes "entity integrity check + path connectivity analysis" to ensure the closure of the "equipment—fault—cause—measure—implementation element—inspection standard" path. Then, a highly consistent initial knowledge graph is constructed through a predefined quality scoring model and domain expert verification process. Once the knowledge graph is built, a version control mechanism and semantic governance module are introduced to support subsequent incremental updates, anomaly auditing, and knowledge traceability, ultimately resulting in a semantically complete, structurally stable, and state-aware equipment maintenance knowledge graph.

[0087] Specifically, step S2032 includes:

[0088] Step S20321: The word segmentation and part-of-speech tagging sequences obtained from the preprocessing of the multi-source text data of the device are input into the encoding layer of the bidirectional long short-term memory network model. The word units are mapped to word vectors, and numerical identification features are added to the numerical information of the device class. The state data is retained as a whole entity. The encoding layer of the bidirectional long short-term memory network model is used to capture contextual dependencies and output the hidden state vector sequence. It is passed to the conditional random field layer for joint label prediction and outputs the structured entity.

[0089] In this embodiment of the invention, a long short-term memory network model is introduced to model and extract the potential logical relationships between "health status - component - failure mode - maintenance measures" contained in the preprocessed text, forming a candidate set of triples that integrate state variables, laying the foundation for constructing a dynamic response knowledge graph.

[0090] To enable the automatic extraction of equipment status, fault features, and maintenance knowledge from large-scale text, the preprocessed word segmentation and part-of-speech tagging sequences are input into the encoding layer of BiLSTM (Bi-directional Long Short-Term Memory) network. Each word is mapped to a word vector, and numerical identification features are added to the equipment-related numerical information to ensure that status data such as "80℃" and "5mm" are retained as a whole entity.

[0091] The BiLSTM encoding layer performs hidden state calculations on the input word vector sequence from both forward and reverse order to obtain the hidden vector sequence.

[0092] The BiLSTM encoding layer processes the input word vector sequence in both forward and backward directions. At each position, the forward network calculates a forward semantic representation based on the word vector at the current position and the state information of the previous position; simultaneously, the backward network generates a backward semantic representation based on the word vector at the current position and the state information of the next position. Finally, the forward and backward semantic representations are concatenated to obtain the contextual feature representation of the current word, forming the hidden state sequence of the entire text. This sequence is then passed to the Conditional Random Field (CRF) layer, which performs joint annotation prediction on the entire text, considering the label transition relationships between words, and outputs the most likely label sequence for entity category recognition. The set of hidden state sequences is then passed to the CRF layer for sequence label prediction, outputting the optimal structured entity category sequence. The CRF receives the feature vector at each position output by the BiLSTM and performs joint prediction on the label sequence in conjunction with the context, obtaining the entity label for each word, such as "B-device component", "I-device component", "B-fault phenomenon", "I-fault phenomenon", "O", etc.

[0093] Specifically, step S2033 includes:

[0094] Step S20331: Input the word segmentation and part-of-speech tagging sequences into the BiLSTM-CRF model and the BERT-CRF model respectively for prediction.

[0095] Step S20332: If the entity label prediction results for the same word unit output by the BiLSTM-CRF model and the BERT-CRF model are consistent, then the structured entity verification passes.

[0096] Step S20333: If the entity label prediction results of the BiLSTM-CRF model and the BERT-CRF model for the same word conflict, and the confidence difference is significant and does not meet the preset conditions, then the entity is marked as a structured entity and needs to be verified again.

[0097] In this embodiment of the invention, a dual-mode verification mechanism based on the BM (Boundary-Matching) heuristic rule is introduced. The segmented and part-of-speech-tagged sequences are input into the BiLSTM-CRF model and the BERT-CRF model, respectively. The entity label prediction results for the same word element output by the BiLSTM-CRF model and the BERT-CRF model are compared, and the entity label assignment is determined based on confidence weighting and majority voting rules. If the entity label prediction results for the same word element output by the BiLSTM-CRF model and the BERT-CRF model are consistent, the structured entity verification passes and is directly adopted. If the entity label prediction results for the same word element output by the BiLSTM-CRF model and the BERT-CRF model conflict, and the confidence difference is significantly different and does not meet the preset conditions, the structured entity needs to be verified again, marked as "entity awaiting manual verification," and submitted to experts for review. Maintenance experts will supplement and correct the errors in subsequent verifications, and the fused entity sequence is extracted into a list of structured entities to ensure entity quality.

[0098] Step S203: Combining the semantic context awareness mechanism and the graph structure stability constraint mechanism, a mechanism is used to prioritize and float the candidate weights of key equipment nodes, and to mark the priority of spare parts resource paths for preventive maintenance parts, and then output candidate fault nodes.

[0099] Specifically, step S203 includes:

[0100] Step S2031: Calculate the similarity between the target fault text vector and the semantic vector of the fault node in the knowledge graph, filter out the knowledge graph fault nodes that are related to the semantic context of the fault text, and obtain a candidate node set.

[0101] Step S2032: Generate an aggregated graph vector based on the graph embedding vectors of key entities in the target fault text, using the weighted average result of structural entropy weights.

[0102] Step S2033: For candidate nodes in the candidate node set, extract graph embedding vectors, and combine the reachability analysis results under counterfactual path perturbation and the structural perturbation rate corresponding to the candidate nodes to correct the structural similarity scores of the candidate nodes.

[0103] Step S2034: Adjust the structural similarity scores of candidate nodes using adaptive fusion weights to obtain a comprehensive similarity score.

[0104] Step S2035: Reorder the candidate nodes in the candidate set based on the comprehensive similarity score to obtain a list of candidate fault nodes.

[0105] Step S2036: Apply the classification results of the first stage, prioritize filtering and upscaling the candidate weights of critical device nodes, and maintain the candidate weights of non-critical device nodes.

[0106] Step S2037: Apply the classification results of the second stage and inject the component classification result labels into the resource edges of the relationship aggregation.

[0107] Step S2038: Nodes that meet the preset conditions are selected as candidate fault nodes.

[0108] In this embodiment of the invention, to achieve semantic alignment between textual fault descriptions and fault nodes in a knowledge graph, a combination of text similarity and knowledge graph embedding techniques is employed. When new equipment fault information is received, the core entities in the fault description are first extracted through text preprocessing and entity recognition. Then, these entities are matched and similarity calculated with existing entity nodes in the knowledge graph. The fault text is vectorized using the TF-IDF method, and then combined with the graph embedding vectors of the knowledge graph nodes to calculate the comprehensive similarity between the two. This determines the entity node most likely corresponding to the fault and obtains relevant historical maintenance information and expert suggestions.

[0109] The target text is preprocessed using standardization, and a context-enhanced word frequency-inverse document frequency semantic representation model is constructed.

[0110] First, the similarity between the target fault text vector and the semantic vector of the fault node in the knowledge graph is calculated to obtain the context-enhanced semantic similarity score. The calculation formula is as follows:

[0111]

[0112] in, For the target fault text vector, For the first in the knowledge graph The semantic vector of each faulty node. For terms The semantic importance weight in the context can be determined by the local attention distribution or the path density factor.

[0113] Based on the context-enhanced semantic similarity score, all faulty nodes in the graph are sorted from high to low, and the Top-K candidate nodes are selected, denoted as the candidate node set. This serves as the object for subsequent graph embedding similarity comparison.

[0114] To match the ability to distinguish structural semantics, a structure entropy-aware graph embedding representation strategy is designed. A "semantic-structural joint path set" is constructed through a path context enhancement mechanism. Node type and role distribution attributes are introduced during the training phase, and roaming path constraints are set to ensure a balance between node semantic consistency and graph structure coverage.

[0115] Suppose the target fault text contains There are 1 key entity whose graph embedding vector is 1. And by using structural entropy weighted averaging, aggregated graph vectors are generated. The calculation formula is as follows:

[0116]

[0117] in, For entity nodes The structural entropy (which can be obtained by measuring its PageRank or node hierarchical complexity in the graph). For smoothing terms, This is the normalization factor.

[0118] For candidate node set Each candidate node in Extract its graph embedding vector Furthermore, by combining the reachability analysis results under counterfactual path perturbation, the structural similarity score is corrected, and the calculation formula is as follows:

[0119]

[0120] in, Graph embedding vectors The structural perturbation rate of the subgraph after the replacement of the causal or mitigation node is used to evaluate the stability of the subgraph structure.

[0121] To comprehensively measure the matching degree between text semantics and structural graph information, a dynamic weighting mechanism based on semantic-structural fusion is designed to calculate the final comprehensive similarity score. The calculation formula is as follows:

[0122]

[0123] in, This indicates the fusion weight, which automatically adjusts the semantic and structural contributions.

[0124] The mechanism adaptively adjusts the weights based on the degree of fusion between semantic features and structural information.

[0125] Based on the overall similarity score Reorder the candidate set and set a threshold. Filter to meet The Top-M nodes form a list of candidate faulty nodes.

[0126] To improve the real-time performance and accuracy of fault diagnosis, equipment operation monitoring data (such as temperature, vibration, and load) are further incorporated, and the preliminary matching results are validated a second time based on a multi-dimensional risk assessment model. By constructing a correlation graph between operating status and historical status, the degree of degradation of the current equipment status is quantified, and combined with the fault severity labels of matching nodes, real-time priority ranking and risk level assessment of fault modes are achieved, thus providing a basis for subsequent maintenance resource scheduling.

[0127] To further enhance the priority-oriented and resource-targeted nature of fault matching, a two-stage classification result is introduced as a classification-driven dynamic filtering and weighting mechanism: First, the first-stage classification result is applied to prioritize and promote critical equipment. The candidate weights of nodes are maintained, and the candidate weights of non-critical device nodes are preserved to avoid favoring non-critical devices. Excessive traversal of nodes. Secondly, by applying the second-stage classification results and lightly touching upon the second-stage component classification, the component classification results are integrated into the "relationship aggregation" stage of fault matching in a minimally invasive and highly reusable manner. Resource edges are pre-mapped in the relationship aggregation, reserving paths for the four types of cross-combinations. This mechanism significantly addresses the "indiscriminate commonality" defect of traditional fault matching. Filtering can reduce the size of the candidate set and shorten the response time; dynamic weighting can also improve performance. Coverage of latent faults ensures early detection and high-precision location of critical equipment faults, supporting the scientific nature of maintenance resources.

[0128] In this way, the selection of faulty nodes is completed, the final candidate faulty nodes are obtained, and they are passed on to subsequent decision-making.

[0129] In some alternative implementations, the candidate screening process is expanded to the following four sub-steps before reordering:

[0130] (1) Classification attribute filtering: based on graph node labels, pre-scanning Device subset, exclude Low priority path;

[0131] (2) Semantic-structural pre-matching: Calculate the original text S score for the filtered candidates and inject classification factors. ;

[0132] (3) Component pre-mapping aggregation: When collecting causal / dependency edges, mark the priority of Class II resource paths;

[0133] (4) Preliminary risk assessment: Based on the equipment deterioration trend, The risk level of nodes is raised, which facilitates system configuration and real-time optimization, ensuring that the matching results are not only semantically and structurally robust, but also priority-sensitive.

[0134] To quantify the fusion effect driven by classification, the comprehensive similarity formula is extended by introducing a classification factor. (Based on FAHP membership) , Calculate classification to enhance similarity :

[0135]

[0136] in, Mapping weights to components For the stability of pre-aggregated resources.

[0137] Similarity is enhanced by calculating classification. The Top-M nodes are reordered, and a matching list with category labels is output. These labels are directly injected into the interpretive subgraph to support maintenance action orchestration. This extension not only maintains the semantic-structural balance of the original text but also achieves three-dimensional synergy of "classification-semantics-structure," improving the overall robustness and interpretability of fault identification.

[0138] Step S205: Based on the directed causal edges, retrieve the candidate maintenance measures corresponding to the candidate fault nodes in the constructed knowledge graph, verify the candidate maintenance measures by combining soft constraints and hard constraints, and use the execution orchestration generator to transform the list of maintenance measures that have passed the constraint verification into an equipment maintenance decision scheme.

[0139] Specifically, the verification of candidate maintenance measures in step S205 above, combining soft and hard constraints, includes:

[0140] Step S2051: Calculate the historical success rate, average maintenance time, component availability, and recurrence rate of candidate maintenance measures, and construct a historical performance index matrix based on the calculation results.

[0141] Step S2052: Based on the extended historical performance construction mechanism, extend the historical performance index matrix into a classification enhancement matrix.

[0142] Step S2053: Using the classification enhancement matrix as input for soft and hard constraints, verify the candidate maintenance measures.

[0143] Step S2054: Using fault matching degree, historical success rate, average maintenance time risk, component availability, and safety risk score as scoring elements, the candidate maintenance measures are quantitatively scored and ranked to obtain a list of maintenance measures.

[0144] In this embodiment of the invention, for a candidate fault node, a set of fault causes directly associated with it is retrieved from the constructed knowledge graph. And sort them according to causal confidence, time freshness and source weight, for any causal node Continue searching for the set of maintenance measures directly associated with it. Simultaneously, a set of maintenance nodes is introduced. This represents the specific execution elements (processes, personnel qualifications, tools, parts, time windows, etc.). Reachability and loop closure checks are performed on the "Cause-Action-Maintenance" three-layer subgraph to ensure that the current fault state can be reached through the selected actions and maintenance nodes without violating mutual exclusion and capacity constraints. If multiple feasible paths exist, the path with the shortest path length and highest relationship weight is prioritized, while retaining several parallel paths to improve coverage and fault tolerance. After deduplicating all action nodes involved in the above paths, a preliminary set of maintenance actions is formed. .

[0145] Retrieve each candidate fault node from the pre-built database. or its set of causes The corresponding historical work orders include the following fields: work order number, fault node identifier, fault cause, repair measures, repair results, repair duration, parts used, personnel feedback, and recurrence status.

[0146] For each candidate maintenance measure, the following indicators are calculated (all statistics are based on measure dimensions):

[0147] Historical success rate:

[0148]

[0149] in, measures Number of work orders successfully repaired This represents the total number of work orders.

[0150] Average repair time:

[0151]

[0152] in, For the first Repair time for each work order.

[0153] Component availability:

[0154]

[0155] in, This represents the number of work orders that could not be executed due to a shortage of parts.

[0156] Recurrence rate:

[0157]

[0158] in, For similar faults, measures The number of work orders that recurred afterward.

[0159] By mapping the above statistical indicators to each measure, a historical performance indicator matrix is ​​constructed:

[0160]

[0161] To further enhance the resource targeting and priority stratification of maintenance decisions, an extended historical performance matrix construction mechanism is introduced, along with a two-stage equipment object classification method for maintenance decisions, serving as a classification-driven dynamic performance evaluation and constraint-sensitive framework.

[0162] First, the results of the first-stage equipment classification are applied to the matrix. measures to inject Priority filtering and weight adjustment mechanisms – in high-impact scenarios, The statistical indicators for related measures are adaptively weighted upwards to ensure prominent evaluation of critical equipment paths, while The path maintains standard traversal to avoid low-priority redundancy;

[0163] Secondly, the second-stage component classification is integrated, expanding the component availability index into a classification-sensitive expression. Finally, four component combinations are generated through pairwise cross-combinations of the two categories, as shown in Table 1, and a classification label is injected into each measure. The classification factors under this framework depend on the specific application scenario, thereby solving the inherent defects of "unhierarchical resource allocation and ambiguous priorities" in traditional maintenance decision-making, ensuring that high-impact equipment receives scientific priority, while maintaining overall economy and operability, and supporting the transformation of condition-based maintenance from experience-dependent to hierarchical intelligence.

[0164] The historical performance data construction is expanded into the following five sub-steps, forming a closed classification-performance iteration loop:

[0165] (1) Categorical attribute injection: Preload the fault subgraph labels from step three. Equipment metadata and component premapping, tagging The matrix's action row;

[0166] (2) Adjustment of performance indicators by category: Apply dynamic factor α to success rate, duration, availability, and recurrence rate;

[0167] (3) Component pre-mapping aggregation: In the action path retrieval, the relationship edges are marked, and the inventory and uncertainty are quantified;

[0168] (4) Four-category cross-evaluation: Calculate the categorical utility of each measure. ,in Assigned based on equipment importance For component strategy, this utility depends on the production environment;

[0169] (5) Matrix Reconstruction: Extension Enhancement matrix for classification New The column serves as input for subsequent constraints.

[0170] By adjusting scenario-dependent factors, performance statistics are shifted from static aggregation to dynamic and highly interpretable classification, facilitating reproduction and optimization across different equipment systems.

[0171] Table 1

[0172]

[0173] To demonstrate the theoretical depth of classification in performance integration, the formula for component availability is expanded by introducing a cross factor. (Based on four types of utility and scene threshold adaptation,) v Class 1+II Higher reward for prevention paths), calculation of classification to enhance usability :

[0174]

[0175] in, Category I: Shortage Penalties At lower Increase, highlighting the uncertainty of passive strategies; specifically (Dynamic calibration based on application scenarios, such as inventory volatility).

[0176] Similarly, reconstruction Subsequently, it was expanded into a classification compliance mechanism in hard constraint verification; and incorporated into the soft constraint description. Weighting. Based on a three-layer collaborative logic of "classification-performance-constraint", it not only inherits multi-objective trade-offs, but also injects hierarchical adaptation to ensure that decisions remain robust and interpretable in complex operation and maintenance environments.

[0177] By embedding expert experience and operational standards into the decision-making system through a "condition → action" rule structure, the system enables automated reasoning and intelligent recommendation of maintenance solutions. This strategy engine introduces multi-dimensional constraints to ensure that decision outcomes are both safe and compliant, while also considering economic efficiency and execution effectiveness.

[0178] Hard constraints are rigid requirements that directly determine whether a maintenance plan can be adopted. They cover aspects such as safety regulations, maintenance operation prohibitions, matching personnel qualifications with job levels, and mandatory inspection items. For example, in the maintenance of high-voltage equipment, if operators do not meet the qualification level requirements or do not comply with safety isolation measures, the maintenance strategy will be automatically rejected. The introduction of hard constraints ensures that maintenance activities do not violate safety and compliance standards under any circumstances.

[0179] In contrast, soft constraints embody optimization and balance in the decision-making process. Soft constraints cover factors such as cost consumption, downtime, spare parts inventory consumption, risk reduction benefits, and equipment life extension effects. The strategy engine can dynamically evaluate the merits of maintenance measures by quantifying and weighting the soft constraint indicators of different candidate solutions. For example, among two feasible maintenance solutions, if solution A has a short downtime but high cost, and solution B has a lower cost but high spare parts consumption, the system will comprehensively rank the two based on the actual operating status and user needs and preferences, and provide a recommendation result.

[0180] Furthermore, this strategy engine does not rely solely on model training or single-condition constraints. Instead, it transforms abstract knowledge items into executable "condition-action" chain rules by incorporating the experiential rules and operational practices of domain experts. For example, if "continuous temperature rise + abnormal vibration frequency" is detected, the action options of "stop and check the lubrication system" or "replace the bearing" are triggered. This rule-based design ensures that expert experience is accumulated and passed on within the system, while enabling a fast, transparent, and interpretable reasoning process when encountering complex situations.

[0181] By leveraging the synergistic effect of hard and soft constraints, the proposed strategy engine not only avoids misoperations caused by neglecting safety regulations but also balances and optimizes under multi-objective conditions, thereby generating maintenance decisions that better suit the actual situation on site. Ultimately, it achieves comprehensive improvements in safety, economy, and feasibility, significantly enhancing the intelligence and scientific level of condition-based maintenance.

[0182] Perform counterfactual simulations and reachability checks on candidate measures: virtually intervene in key processes, resources, or time windows to assess their impact on failure evolution, health improvement, and operational conflicts; solutions that do not meet constraints or are unreachable are automatically eliminated. Rule actions include "enforce / exclude / promote or depromote priority" and are recorded in a priority change mapping table for sorting and recall.

[0183] After completing historical performance statistics and rule filtering, the remaining candidate measures... Perform quantitative scoring and ranking. Define the following scoring elements: fault matching degree. Historical success rate Average repair time risk Component availability Risk of recurrence Safety risk score .

[0184] The formula for calculating the average repair time risk is as follows:

[0185]

[0186] The risk of average repair time is mapped from the mean repair time to a risk score; the longer the repair time, the higher the risk.

[0187] The formula for calculating the risk of recurrence is as follows:

[0188]

[0189] The risk of relapse is given directly by the relapse rate or after normalization.

[0190] The formula for calculating the safety risk score is as follows:

[0191]

[0192] The safety risk score is a safety risk coefficient given by experts based on the type of failure and the complexity of the measures. The mapping is obtained.

[0193] Overall rating:

[0194]

[0195] in, As weight, .

[0196] Min-Max normalization is used uniformly:

[0197]

[0198] Calculate for all candidate measures Sort the results in descending order to obtain a sorted list. Set an elimination threshold. ,when The remaining solutions are eliminated based on their scores and enter the final decision pool. The Top-L solutions are then output as the system's recommended list.

[0199] For measures entering the decision pool, the execution orchestration generator transforms the plan into an executable sequence based on the constraint diagram of "measure → process → inspection" and provides: measure details, operation steps, required parts and quantities, estimated working hours, estimated cost, safety precautions and key verification points. All Top-L plans are output in the form of a comparison table to facilitate maintenance personnel to quickly select the appropriate plan.

[0200] After on-site execution, the engineer fills out the execution result form (whether it was successful, the actual downtime, the actual cost, whether additional measures were taken, the recurrence situation, etc.).

[0201] The system periodically and incrementally absorbs new work orders and feedback:

[0202] New fault descriptions are further preprocessed and mapped to the graph. If new entities or new "fault-cause-measure" relationships are found, they are added to the graph.

[0203] Rolling updates Equivalent statistical values;

[0204] Dynamically adjust weights With risk threshold;

[0205] Make convergent modifications to rules that perform poorly or cause frequent conflicts;

[0206] The execution results are written back to the "Priority Change Mapping Table" and the version governance module to keep the graph and strategy evolving in sync.

[0207] A strategy engine for condition-based maintenance is constructed, extracting expert experience into rule subgraphs and executable constraints. Rules consist of triggering conditions, applicable boundaries, priority factors, and handling suggestions, mapped to observations such as equipment status, operating conditions, risk level, and resource availability. A collaborative mechanism of hard and soft constraints is established, using a multi-objective utility function to jointly score and adaptively rank candidate maintenance measures. Counterfactual simulation and accessibility testing are introduced to virtually intervene in candidate measures to assess their impact on fault evolution, health improvement, and operational conflicts, automatically eliminating non-compliant or unreachable solutions and prioritizing solutions with stable returns and controllable risks. Combining online preference learning and closed-loop feedback, rule weights and triggering boundaries are dynamically updated based on historical handling results, expert review, and key on-site indicators, resolving rule conflicts and maintaining strategy convergence and interpretability. Finally, a list of measures with priorities, applicable ranges, key evidence chains, and risk commitments is output to support rapid decision-making. The source and weight of rule items are consistent with the results system of step one, and the handling results are fed back after execution to update rule weights and triggering boundaries.

[0208] Specifically, by using path retrieval in the knowledge graph, corresponding maintenance measures are determined for each fault mode. Once fault matching is complete, a set of candidate maintenance measures associated with the fault mode is automatically extracted along the path defined in the knowledge graph. Historical maintenance data is then used to conduct a preliminary evaluation and ranking of the effectiveness of the maintenance measures, thereby determining the optimal maintenance strategy for on-site implementation.

[0209] First, the real-time assessment of equipment operating status. During maintenance decision-making, this invention collects equipment operating status data in real time, including key parameters such as temperature, vibration frequency, and load level. Through real-time data monitoring and trend analysis, it conducts immediate assessments of equipment health status and promptly detects potential fault signs or operational anomalies.

[0210] Secondly, the process of prioritizing maintenance tasks. Based on real-time status assessment results, this invention further integrates multi-dimensional factors such as equipment importance level and fault risk severity from a knowledge graph. Through risk correlation analysis and a priority assessment model, it automatically determines the specific maintenance priority ranking of the equipment currently requiring repair and its fault modes. This process ensures that, when maintenance resources are limited, equipment faults with higher urgency and risk levels are prioritized for repair.

[0211] Secondly, in the maintenance strategy selection stage, after clarifying the specific priority order of equipment maintenance, this invention automatically matches the most suitable maintenance strategy through knowledge graph path retrieval. Specifically, the knowledge graph directly associates fault modes with specific maintenance measures, and combines historical maintenance experience data for comprehensive evaluation and ranking, ultimately automatically selecting the optimal maintenance measure for on-site decision-making and implementation.

[0212] Finally, there's the closed-loop feedback stage for maintenance decisions. After implementing maintenance strategies on-site, engineers can provide real-time feedback on the maintenance results to the knowledge graph system, including the actual maintenance outcomes. The system continuously optimizes the knowledge graph structure and knowledge base content based on the on-site feedback data, constantly improving the accuracy and reliability of future maintenance decisions.

[0213] The knowledge graph-driven two-stage classification-based equipment condition-based maintenance decision-making method provided in this embodiment has the following advantages:

[0214] (1) Automatically mapping unstructured fault text to structured knowledge graphs can significantly reduce manual retrieval and summarization time and improve maintenance decision-making efficiency.

[0215] (2) By combining BM LSTM with graph embedding and other strategies, the accuracy of fault matching and solution decision-making is improved, the misjudgment rate is reduced, and the accuracy is enhanced;

[0216] (3) The constructed knowledge graph supports continuous expansion and updates, can cover new equipment types and new fault scenarios, has good migration and expansion capabilities, and is highly scalable;

[0217] (4) By combining historical data and expert rules, intelligent decision-making on maintenance plans can be achieved, helping maintenance personnel to make the best decisions quickly.

[0218] This embodiment also provides a two-stage classification device for condition-based maintenance decision-making based on knowledge graphs. This device is used to implement the above embodiments and preferred embodiments, and will not be repeated hereafter.

[0219] This embodiment provides a knowledge graph-driven two-stage classification device for condition-based equipment maintenance decision-making, such as... Figure 5 As shown, it includes:

[0220] The acquisition module 501 is used to acquire multi-source text data and preprocess the multi-source text data.

[0221] The classification module 502 is used to perform a first-stage classification of equipment to obtain critical equipment and non-critical equipment, and a second-stage classification of parts to obtain parts for passive maintenance and preventive maintenance. The equipment classification results and the parts classification results are cross-processed to obtain the two-stage classification results.

[0222] The knowledge graph construction module 503 is used to extract entities and relationships between entities from preprocessed multi-source text data. It uses equipment entities, fault mode entities, and maintenance measure entities as knowledge ontology, with entities and relationships between entities as the core, embedding state variable nodes, embedding the first-stage classification results in the equipment entity nodes, and using a long short-term memory network to construct a knowledge graph that integrates state variables. The relationships between entities include belonging relationships, causing relationships, and processing relationships.

[0223] The candidate fault node output module 504 is used to combine the semantic context awareness mechanism and the graph structure stability constraint mechanism to output candidate fault nodes by prioritizing the filtering and floating of candidate weights of key equipment nodes and marking the priority of spare parts resource paths for preventive maintenance parts.

[0224] The maintenance decision output module 505 is used to retrieve candidate maintenance measures corresponding to candidate fault nodes in the constructed knowledge graph based on directed causal edges, verify the candidate maintenance measures by combining soft constraints and hard constraints, and use the execution orchestration generator to transform the list of maintenance measures that have passed the constraint verification into equipment maintenance decision schemes.

[0225] The knowledge graph-driven two-stage classification equipment condition-based maintenance decision-making device provided in this embodiment of the invention can execute the knowledge graph-driven two-stage classification equipment condition-based maintenance decision-making method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method execution.

[0226] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. (Reference) Figure 6 This diagram illustrates a structural schematic suitable for implementing an electronic device according to embodiments of the present invention. The electronic device may include a processor 601, which can perform various appropriate actions and processes based on a program stored in a read-only memory (ROM) 602 or a program loaded from a memory 608 into a random access memory (RAM) 603. The RAM 603 also stores various programs and data required for the operation of the electronic device. The processor 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0227] Typically, the following devices can be connected to I / O interface 605: input devices 606, including, for example, touchscreens, touchpads, etc.; output devices 607, including, for example, liquid crystal displays (LCDs); memory 608, including, for example, hard disks; and communication devices 609. Communication device 609 allows electronic devices to exchange data via wireless or wired communication with other devices. Although Figure 6 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0228] Specifically, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. Embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. This computer program can be downloaded and installed from a network via a communication device 609, or installed from a memory 608, or installed from a ROM 602. When the computer program is executed by the processor 601, it performs the functions defined in the knowledge graph-driven two-stage classification device condition-based maintenance decision-making method of the present invention. Figure 6 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0229] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and all such modifications and variations fall within the scope defined by the appended invention.

Claims

1. A two-stage classification-based equipment condition-based maintenance decision-making method driven by knowledge graphs, characterized in that, The method includes: Acquire multi-source text data and preprocess the multi-source text data; The equipment is classified in the first stage to obtain critical equipment and non-critical equipment. The components are classified in the second stage to obtain components for passive maintenance and preventive maintenance. The equipment classification results and component classification results are cross-processed to obtain the two-stage classification results. Entities and relationships between entities are extracted from preprocessed multi-source text data. Equipment entities, fault mode entities, and maintenance measure entities are used as knowledge ontology. With entities and relationships between entities as the core, state variable nodes are embedded. The first-stage classification results are embedded in the equipment entity nodes. A knowledge graph that integrates state variables is constructed using a long short-term memory network. The relationships between entities include belonging relationships, causing relationships, and processing relationships. Combining semantic context awareness and graph structure stability constraint mechanisms, a mechanism is established to prioritize and float candidate weights of key equipment nodes, thereby assigning priority to spare parts resource paths for preventive maintenance components and outputting candidate fault nodes. Based on directed causal edges, candidate maintenance measures corresponding to the candidate fault nodes are retrieved in the constructed knowledge graph. The candidate maintenance measures are verified by combining soft and hard constraints. The list of maintenance measures that pass the constraint verification is transformed into an equipment maintenance decision scheme by using an execution orchestration generator. The mechanism, which combines semantic context awareness and graph structure stability constraint, prioritizes and floats candidate weights for key equipment nodes to prioritize spare parts resource paths for preventative maintenance components, outputting candidate fault nodes, including: Calculate the similarity between the target fault text vector and the semantic vector of the fault node in the knowledge graph, filter the knowledge graph fault nodes that are related to the contextual semantics of the fault text, and obtain a candidate node set; Based on the graph embedding vectors of key entities in the target fault text, an aggregated graph vector is generated using a weighted average of structural entropy weights. For the candidate nodes in the candidate node set, the graph embedding vector is extracted. Combined with the reachability analysis results under counterfactual path perturbation and the structural perturbation rate corresponding to the candidate node, the structural similarity score of the candidate node is corrected. The structural perturbation rate is used to evaluate the stability of the subgraph structure to which the candidate node belongs. The structural similarity scores of candidate nodes are adjusted using adaptive fusion weights to obtain a comprehensive similarity score; Based on the comprehensive similarity score, the candidate nodes in the candidate set are reordered to obtain a list of candidate fault nodes; Applying the classification results from the first stage, priority is given to filtering and increasing the candidate weights of critical device nodes, while maintaining the candidate weights of non-critical device nodes. Apply the classification results from the second stage and inject the component classification result labels into the resource edges of the relationship aggregation; Nodes that meet the preset conditions are selected as candidate fault nodes.

2. The method according to claim 1, characterized in that, The construction of the knowledge graph that integrates state variables includes: Construct a knowledge graph structure that integrates state awareness; By combining a bidirectional long short-term memory network model and a conditional random field sequence labeling model, structured entities are extracted. A dual-mode verification mechanism based on boundary matching heuristic rules is used to verify the extracted structured entities. The validated structured entities are imported into a unified entity library to construct a knowledge graph that integrates state variables.

3. The method according to claim 2, characterized in that, The construction of the knowledge graph structure with fusion state awareness includes: An evaluation hierarchical structure model is established with equipment importance as the target layer, key influencing factors and equipment factors as the criteria layer, and the impact of failures on production quality, environmental health and safety, utilization rate, production, failure frequency, maintainability, detectability and maintenance cost as the solution layer. Using equipment importance as the evaluation set, a fuzzy consistency judgment matrix is ​​constructed, and a fuzzy comprehensive importance evaluation model is built based on the product of the fuzzy consistency judgment matrix and the fuzzy matrix corresponding to the evaluation set. The equipment classification results are determined based on the principle of maximum membership. The equipment classification results are then embedded into the equipment entity attributes of the knowledge graph. A new equipment status attribute field is added, and the equipment status attribute field is semantically bound to equipment nodes and fault nodes to construct a knowledge graph structure that integrates status awareness.

4. The method according to claim 2, characterized in that, The sequence labeling model combining a bidirectional long short-term memory network model and a conditional random field extracts structured entities, including: The word segmentation and part-of-speech tagging sequences obtained from the preprocessing of multi-source text data of the device are input into the encoding layer of the bidirectional long short-term memory network model. The word units are mapped to word vectors, and numerical identification features are added to the numerical information of the device class. The state data is retained as a whole entity. The encoding layer of the bidirectional long short-term memory network model is used to capture contextual dependencies and output a hidden state vector sequence. This sequence is passed to the conditional random field layer for joint label prediction and outputs a structured entity.

5. The method according to claim 4, characterized in that, The dual-mode verification mechanism based on boundary matching heuristic rules verifies the extracted structured entities, including: The word segmentation and part-of-speech tagging sequences are respectively input into the BiLSTM-CRF model and the BERT-CRF model for prediction; If the BiLSTM-CRF model and the BERT-CRF model produce the same entity label prediction results for the same word, then the structured entity verification passes. If the entity label prediction results of the BiLSTM-CRF model and the BERT-CRF model conflict for the same word, and the confidence difference is significant and does not meet the preset conditions, then the entity labeled as a structured entity needs to be verified again.

6. The method according to claim 1, characterized in that, The verification of candidate maintenance measures by combining soft and hard constraints includes: Calculate the historical success rate, average maintenance time, parts availability, and recurrence rate of candidate maintenance measures, and construct a historical performance index matrix based on the calculation results; The historical performance index matrix is ​​extended into a classification enhancement matrix based on the mechanism of constructing extended historical performance. The candidate maintenance measures are validated using the classification enhancement matrix as input for both soft and hard constraints. Using fault matching degree, historical success rate, average maintenance time risk, component availability, and safety risk score as scoring factors, the candidate maintenance measures are quantitatively scored and ranked to obtain a list of maintenance measures.

7. A two-stage classification equipment condition-based maintenance decision-making device driven by knowledge graphs, characterized in that, The device includes: The acquisition module is used to acquire multi-source text data and preprocess the multi-source text data; The classification module is used to perform the first stage classification of equipment to obtain critical equipment and non-critical equipment, and the second stage classification of parts to obtain parts for passive maintenance and preventive maintenance. The equipment classification results and the parts classification results are cross-processed to obtain the two-stage classification results. The knowledge graph construction module is used to extract entities and relationships between entities from preprocessed multi-source text data. It uses equipment entities, fault mode entities, and maintenance measure entities as knowledge ontology, with entities and relationships between entities as the core, embedding state variable nodes, embedding the first-stage classification results in equipment entity nodes, and using a long short-term memory network to construct a knowledge graph that integrates state variables. The relationships between entities include belonging relationships, causing relationships, and processing relationships. The candidate fault node output module is used to combine the semantic context awareness mechanism and the graph structure stability constraint mechanism to filter and float the candidate weights of key equipment nodes first, and to mark the priority of spare parts resource paths for preventive maintenance parts, and output candidate fault nodes. The maintenance decision output module is used to retrieve candidate maintenance measures corresponding to the candidate fault nodes in the constructed knowledge graph based on directed causal edges, verify the candidate maintenance measures by combining soft constraints and hard constraints, and use the execution orchestration generator to transform the list of maintenance measures that have passed the constraint verification into an equipment maintenance decision scheme. The candidate fault node output module is specifically used for: calculating the similarity between the target fault text vector and the semantic vector of the fault node in the knowledge graph; filtering knowledge graph fault nodes related to the semantic context of the fault text to obtain a candidate node set; generating an aggregated graph vector based on the graph embedding vectors of key entities in the target fault text, using a weighted average of structural entropy weights; extracting graph embedding vectors from the candidate nodes in the candidate node set, and correcting the structural similarity score of the candidate nodes by combining the reachability analysis results under counterfactual path perturbation and the structural perturbation rate corresponding to the candidate nodes, wherein the structural perturbation rate is used to evaluate the stability of the subgraph structure to which the candidate node belongs; adjusting the structural similarity score of the candidate nodes using adaptive fusion weights to obtain a comprehensive similarity score; reordering the candidate nodes in the candidate set based on the comprehensive similarity score to obtain a candidate fault node list; applying the first-stage classification results to prioritize filtering and upsizing the candidate weights of key equipment nodes, while maintaining the candidate weights of non-key equipment nodes; applying the second-stage classification results to inject the component classification result labels into the resource edges in the relationship aggregation; and selecting nodes that meet preset conditions as candidate fault nodes.

8. An electronic device, characterized in that, include: A memory and a processor are interconnected, the memory storing computer instructions, and the processor executing the computer instructions to perform the knowledge graph-driven two-stage classification device condition-based maintenance decision-making method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the knowledge graph-driven two-stage classification device condition-based maintenance decision-making method as described in any one of claims 1 to 6.

Images