Maintenance plan matching and recommendation method based on causality and equipment structure enhancement

Abstract

The present invention proposes a maintenance solution matching and recommendation method based on causality and equipment structure enhancement, which belongs to the field of equipment maintenance, including: S1: constructing a synonym library in the maintenance field; training an unsupervised SimCSE model to obtain a standardized knowledge graph IDS-EMKG-QA; optimizing the question sentence to obtain a standardized question sentence; S2: performing template-based intent recognition on the standardized question sentence, if the question is in the intent recognition template, then generating and matching the answer based on the query statement, if not, go to step S3; S3: using a graph search method, according to the standardized question sentence, inferring and screening candidate answers based on the fault causal link and the equipment composition architecture in IDS-EMKG-QA; using the RotatE graph embedding model to match and sort the candidate answers and the standardized question sentence to obtain the recommendation result. The method of the present invention solves the diversity and complexity problems in the maintenance process of complex equipment.

Description

Technical Field

[0001] The present invention belongs to the technical field of service-oriented nuclear power key equipment maintenance, and in particular relates to a maintenance plan matching and recommendation method based on causality and equipment structure enhancement. Background Art

[0002] Nuclear power is a clean and efficient new energy source. my country's vigorous development of nuclear power is in line with the world's energy utilization trend and is of great significance to improving my country's energy supply structure and improving the atmospheric environment. Over the past decade, my country's nuclear power generation has continued to grow, making important contributions to ensuring the security of power supply and promoting energy conservation and emission reduction. The circulating water pump is an important prerequisite for ensuring the normal operation of the circulating water system. Once it stops working, it will cause the entire nuclear power plant system to shut down, causing major economic losses, and will cause nuclear leakage, affecting people's lives. It is a typical key equipment in the nuclear power system.

[0003] The circulating water pumps in nuclear power plants are large in size and weight, with numerous core equipment and instruments, making maintenance difficult. Currently, the operation and maintenance of nuclear power equipment, especially the maintenance of nuclear power equipment, has problems such as low matching accuracy of maintenance plans. When formulating maintenance plans, due to the limitations of the technical capabilities of maintenance personnel, it is sometimes necessary to retrieve a large number of operation and maintenance manuals, maintenance procedures, etc. Complex faults require detailed records after experts provide technical support. Due to the problems of non-standard noun naming and difficulty in deriving fault links in the accumulated records, historical experience is difficult to reuse, resulting in difficulty in obtaining equipment maintenance plans. Operation and maintenance personnel spend a lot of time processing information but only obtain a small part of the information to guide maintenance work.

[0004] Based on the equipment maintenance knowledge graph, equipment maintenance solutions are further recommended. Analysis shows that there are two main problems in recommending key equipment maintenance solutions based on the knowledge graph:

[0005] (1) There is a contradiction between the low degree of record standardization and the high maintenance accuracy requirements in complex equipment maintenance. The equipment maintenance knowledge graph effectively integrates various maintenance bases, but some terms in the maintenance knowledge graph are not expressed in a standardized way, and there are a lot of colloquial expressions when operators enter maintenance requests.

[0006] (2) The unclear path for finding maintenance solutions makes it difficult to reason about maintenance solutions in the equipment maintenance knowledge graph. Based on the equipment maintenance knowledge graph, although simple problems related to equipment maintenance knowledge can be matched, the faults and maintenance solutions in the knowledge graph are not one-to-one corresponding. It is often necessary to reason about the cause-effect relationship of the fault and the location. This process is currently carried out by the subjective judgment of the operator, which reduces the efficiency and accuracy of maintenance. Summary of the invention

[0007] In order to solve the above technical problems, the present invention provides a maintenance plan matching and recommendation method based on causality and equipment structure enhancement, comprising the following steps:

[0008] Step S1: Through the equipment maintenance domain knowledge extraction data set and the named entity recognition model, a maintenance domain standard vocabulary and maintenance domain entity library are constructed, and a maintenance domain synonym library is generated; an incremental training method is adopted to train an unsupervised SimCSE model using a public data set and the maintenance domain standard vocabulary, maintenance domain entity library, and maintenance domain synonym library to normalize the equipment maintenance knowledge graph IDS-EMKG, and obtain the normalized equipment maintenance knowledge graph IDS-EMKG-QA; combined with IDS-EMKG-QA and the maintenance domain synonym library, the query optimization algorithm based on SimCSE is used to optimize the question sentence to obtain a normalized question sentence;

[0009] Step S2: Performing template-based intent recognition on the standardized question. If the query question is in the intent recognition template, then generating and matching an answer based on the query statement of the intent recognition template. If not, then going to step S3;

[0010] Step S3: Using graph search, according to the standardized question, inference is performed in IDS-EMKG-QA based on the fault causal link and the equipment composition architecture to screen out candidate answers; the RotatE graph embedding model is used to match and sort the candidate answers and the standardized question, and finally the maintenance plan recommendation result is obtained.

[0011] Beneficial effects:

[0012] 1. The present invention provides a maintenance plan matching and recommendation method based on causality and equipment structure enhancement. In the absence of a vocabulary list and a domain vocabulary library, domain vocabulary normalization based on unsupervised SimCSE is performed. Incremental training of the unsupervised SimCSE model is used to improve the model's sensitivity to domain vocabulary, which is of great significance to the construction of a domain knowledge base.

[0013] 2. The present invention proposes a normalization process of knowledge graphs and questions for knowledge reasoning, normalizes and integrates the nodes of IDS-EMKG, and finally obtains the knowledge graph IDS-EMKG-QA for equipment maintenance plan recommendation, which effectively avoids the redundancy of nodes and relationships in the knowledge base and improves the efficiency of knowledge matching and reasoning.

[0014] 3. The method proposed in the present invention simulates the maintenance plan reasoning path of the operator and combines the knowledge graph of faults and equipment with the similarity calculation method to achieve the recommendation of maintenance plans. Experiments show that the accuracy of this method is significantly better than the existing methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A schematic flow chart of a maintenance plan matching and recommendation method based on causality and equipment structure enhancement according to the present invention;

[0016] Figure 2 This is a schematic diagram of the knowledge graph normalization process;

[0017] Figure 3 Example graphs of 1-hop and multi-hop knowledge graph nodes;

[0018] Figure 4 Detailed schematic diagram of the method of the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0020] Embodiment 1

[0021] like Figure 1 As shown, a maintenance plan matching and recommendation method based on causality and equipment structure enhancement provided by an embodiment of the present invention includes the following steps:

[0022] Step S1: Through the equipment maintenance domain knowledge extraction data set and named entity recognition model, a maintenance domain standard vocabulary and maintenance domain entity library are constructed, and a maintenance domain synonym library is generated; an incremental training method is adopted to train an unsupervised SimCSE model using public data sets and maintenance domain standard vocabulary, maintenance domain entity library, and maintenance domain synonym library to normalize the equipment maintenance knowledge graph IDS-EMKG, and obtain the normalized equipment maintenance knowledge graph IDS-EMKG-QA; combined with IDS-EMKG-QA and the maintenance domain synonym library, the query optimization algorithm based on SimCSE is used to optimize the question sentence to obtain a normalized question sentence;

[0023] Step S2: Perform template-based intent recognition on the standardized question. If the query question is in the intent recognition template, generate and match the answer based on the query statement of the intent recognition template. If not, go to step S3.

[0024] Step S3: Using graph search, according to the standardized question, in IDS-EMKG-QA, the candidate answers are inferred and screened based on the fault causal link and the equipment composition architecture; the RotatE graph embedding model is used to match and sort the candidate answers and the standardized question, and finally the maintenance plan recommendation result is obtained.

[0025] The method of the present invention can combine SimCSE unsupervised learning, templated answer matching, and RotatE embedded answer reasoning with causal relationship and equipment structure enhancement in the absence of a domain annotation set, thereby realizing knowledge graph normalization, basic problem solution matching, and complex problem solution reasoning, effectively solving the diversity and complexity problems in the maintenance process of complex equipment and improving maintenance efficiency and accuracy.

[0026] In one embodiment, the above step S1: constructs a maintenance field standard vocabulary and maintenance field entity library through equipment maintenance field knowledge extraction data set and named entity recognition model, and generates a maintenance field synonym library; adopts an incremental training method, uses a public data set and a maintenance field standard vocabulary, maintenance field entity library, and maintenance field synonym library to train an unsupervised SimCSE model for normalizing the equipment maintenance knowledge graph IDS-EMKG to obtain a normalized equipment maintenance knowledge graph IDS-EMKG-QA; combines IDS-EMKG-QA and the maintenance field synonym library, uses a query optimization algorithm based on SimCSE to optimize the question sentence, and obtains a normalized question sentence, specifically including:

[0027] Step S11: construct a maintenance field standard vocabulary based on the standard words corresponding to the equipment standard codes appearing in the knowledge base or knowledge graph, perform named entity recognition on the corpus based on the trained nuclear power equipment maintenance field named entity recognition model to construct a maintenance field entity library, and then finally obtain a maintenance field synonym library through a word2vec combined with manual recognition method;

[0028] First, build a standard vocabulary in the maintenance field. In the actual production process, there is often a set of coding or abbreviation systems to more conveniently refer to and name various equipment. If the standard codes in the table appear in the knowledge base or knowledge graph, they are replaced with the corresponding standard vocabulary, and the standard vocabulary in the table is used to build a standard vocabulary in the maintenance field to prepare for the subsequent standardization of vocabulary. For scenarios where it is difficult to obtain such coding or abbreviation systems, crawlers can be used to obtain standard vocabulary from encyclopedias.

[0029] Secondly, we build a maintenance domain entity library. Maintenance experience information is another important basis for equipment maintenance. Noun expressions are highly colloquial. These colloquial nouns are called domain entities. Through the trained named entity recognition model in the field of nuclear power equipment maintenance, we perform entity recognition on "Pump Equipment Quality Defect Report" and "Maintenance Experience Feedback" to build a maintenance domain entity library.

[0030] Finally, the two are combined with manual recognition based on word2vec to obtain a synonym library in the maintenance field.

[0031] Step S12: training the SimCSE model based on the public data set to obtain an initial SimCSE model with good performance; incrementally training and fine-tuning the initial SimCSE model based on the synonym database in the maintenance field to obtain a trained SimCSE model;

[0032] Step S13: Use the trained SimCSE model and the maintenance domain synonym library to replace and remove redundancy from the existing equipment maintenance knowledge graph IDS-EMKG to obtain the standardized equipment maintenance knowledge graph IDS-EMKG-QA;

[0033] Step S14: According to the maintenance domain synonym library and IDS-EMKG-QA, the query optimization algorithm based on SimCSE is used to optimize the question sentence to obtain a standardized question sentence.

[0034] In one embodiment, the above step S13: using the trained SimCSE model and the maintenance domain synonym library to replace and remove redundancy from the existing equipment maintenance knowledge graph IDS-EMKG to obtain a standardized equipment maintenance knowledge graph IDS-EMKG-QA, specifically includes:

[0035] Step S131: traverse the node list NodeList in the IDS-EMKG, and for each node Node, replace the Node with the corresponding standard word in the maintenance field synonym library SynList;

[0036] Step S132: extracting nodes of the same type as Node from NodeList to form a node subset NodeSubList;

[0037] Step S133: Calculate the name similarity simScore between Node and each node NodeSub in NodeSubList based on cosine similarity and the trained SimCSE model. The known cosine similarity calculation function is as follows:

[0038] (1)

[0039] in, Represents the cosine similarity of vectors a and b;

[0040] The calculation formula of name similarity simScore is as follows:

[0041] (2)

[0042] If simScore is greater than the preset threshold k, the node pair [Node, NodeSub] is added to the set SynNode;

[0043] Step S134: Manually verify all node pairs in SynNode to obtain the actual node combination SynNode' that needs to be merged; for each pair of nodes [Node, Node'] in SynNode', copy all relationships of Node' in IDS-EMKG to Node, and delete Node' and all relationships related to it; finally, obtain the standardized equipment maintenance knowledge graph IDS-EMKG-QA.

[0044] Figure 2 Schematic diagram of the knowledge graph normalization process.

[0045] In one embodiment, the above step S14: optimizing the question sentence using the query optimization algorithm based on SimCSE according to the maintenance field synonym library and IDS-EMKG-QA to obtain a standardized question sentence specifically includes:

[0046] Step S141: Initialize an empty entity node set NodeSet;

[0047] Step S142: by traversing the maintenance field synonym library SynList, the words in the question sentence Text are replaced with standard words in SynList;

[0048] Step S143: Input Text into the BERT-Bi-LSTM model to obtain an entity name list NameList;

[0049] Step S144: For each entity name Name in NameList, if it already exists in IDS-EMKG-QA, add it as a node to NodeSet; if it does not exist, take Name as input, find the most similar node name in IDS-EMKG through the SimCSE model, and add the node corresponding to the name to NodeSet; finally, return a set NodeSet containing all matching nodes as a normalized question.

[0050] In one embodiment, the above step S2: performing template-based intent recognition on the standardized question, if the query question is in the intent recognition template, then generating and matching the answer based on the query statement of the intent recognition template specifically includes:

[0051] Step S21: construct 6 types of question intention recognition templates based on IDS-EMKG-QA to cover basic questions and their variants: including: equipment-fault, cause-fault, fault-impact, fault-repair method, measurement point-fault, equipment-equipment;

[0052] The embodiment of the present invention designs 6 types of intention recognition templates based on rule templates by analyzing the knowledge scope of the nuclear power equipment maintenance field, as shown in Table 1:

[0053] Table 1 Intent identification template for basic questions

[0054]

[0055] The above templates cover common basic questions and their variations, and can accurately identify the types of questions asked by users, ensuring accuracy and applicability when dealing with basic questions.

[0056] Step S22: Based on the question intent recognition template and the standardized question, a query statement and answer generation template based on the neo4j knowledge graph is constructed. Based on the query statement, the nodes within one hop around the question entity node are matched and the corresponding answer is returned.

[0057] Table 2 shows examples of query statements and answer generation templates based on the neo4j knowledge graph.

[0058] Table 2 Query statement template

[0059]

[0060] like Figure 3 As shown, taking the node "gearbox" as an example, examples of directly related nodes (1 hop) and indirectly related nodes (multiple hops) of knowledge graph nodes are shown.

[0061] In one embodiment, the above step S3: using a graph search method, according to the standardized question, in IDS-EMKG-QA, based on the fault causal link and the equipment composition architecture, infers and screens candidate answers; using the RotatE graph embedding model to match and sort the candidate answers and the standardized question, and finally obtains the maintenance plan recommendation result, which specifically includes:

[0062] Step S31: Initialize boundary parameter m, device parameter n, node distance D and answer range k;

[0063] Step S32: using a query optimization algorithm based on SimCSE to obtain the fault description FaultNode and the equipment description EquipName in the standardized question sentence;

[0064] Step S33: Search for solution nodes within m hops around FaultNode in IDS-EMKG-QA to obtain the solution node set OperaNodeSet. If no result is found, increase the boundary parameter m and continue searching. Similarly, search for fault nodes within n hops around EquipName in IDS-EMKG-QA to obtain the fault node set FaultNodeSet. If no result is found, increase the equipment range parameter n and continue searching. For each node FaultNode in FaultNodeSet, if it has a triple {FaultNode, hasOperation, OperaNode} in IDS-EMKG-QA, add OperaNode to the set OperaNodeSet.

[0065] Figure 4 The specific process of maintenance path reasoning based on the fusion of fault causal links and equipment structure includes the following three paths, which are used to find the maintenance plan of the fault, that is, the algorithm described in step S33:

[0066] Path 1: Extract the fault directly from the question and find the corresponding maintenance solution for the fault;

[0067] Path 2: Extract the fault directly from the question, find the nearest fault cause with a maintenance plan through the fault causal chain, and return the maintenance plan;

[0068] Path 3: Extract equipment and faults from the question, find faults similar to those in the question from neighboring equipment, and find the corresponding maintenance plan for the fault;

[0069] Step S34: Input OperaNodeSet, FaultNode, and hasOperation into the RotatE model to obtain OperaEmbSet, FaultEmb, and hasOperaEmb respectively; it is known that the distance function defined by RotatE is as follows:

[0070] (3)

[0071] in, represents the Hadamard product, h (head) represents the subject entity of the knowledge graph triple (h, r, t), r (relation) represents the relationship between the subject and the object in the knowledge graph triple, and t (tail) represents the object entity of the knowledge graph triple;

[0072] For each device embedding vector OperaEmb in OperaEmbSet, the distance between it and the fault embedding vector FaultEmb is calculated by formula 3 , the result Add to set D;

[0073] (4)

[0074] Arrange the set D from small to large according to the absolute value, and select the first k with the smallest distance Corresponding A maintenance method set FinOperaSet is formed and returned as the final recommended maintenance solution.

[0075] Figure 4 A detailed flowchart of the maintenance plan matching and recommendation method based on causality and equipment structure enhancement is shown.

[0076] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A maintenance plan matching and recommendation method based on causality and equipment structure enhancement, characterized in that: include: Step S1: Through the equipment maintenance domain knowledge extraction data set and the named entity recognition model, a maintenance domain standard vocabulary and maintenance domain entity library are constructed, and a maintenance domain synonym library is generated; an incremental training method is adopted to train an unsupervised SimCSE model using a public data set and the maintenance domain standard vocabulary, maintenance domain entity library, and maintenance domain synonym library to normalize the equipment maintenance knowledge graph IDS-EMKG, and obtain the normalized equipment maintenance knowledge graph IDS-EMKG-QA; combined with IDS-EMKG-QA and the maintenance domain synonym library, the query optimization algorithm based on SimCSE is used to optimize the question sentence to obtain a normalized question sentence, which specifically includes: Step S11: construct a maintenance field standard vocabulary based on the standard words corresponding to the equipment standard codes appearing in the knowledge base or knowledge graph, perform named entity recognition on the corpus based on the trained nuclear power equipment maintenance field named entity recognition model to construct a maintenance field entity library, and then finally obtain a maintenance field synonym library through a word2vec combined with manual recognition method; Step S12: training the SimCSE model based on the public data set to obtain an initial SimCSE model with good performance; performing incremental training and fine-tuning on the initial SimCSE model based on the maintenance field synonym library to obtain a trained SimCSE model; Step S13: Use the trained SimCSE model and the maintenance field synonym library to replace and remove redundancy from the existing equipment maintenance knowledge graph IDS-EMKG to obtain a standardized equipment maintenance knowledge graph IDS-EMKG-QA; Step S14: optimizing the question sentence using the query optimization algorithm based on SimCSE according to the maintenance field synonym library and IDS-EMKG-QA to obtain a standardized question sentence; Step S2: Performing template-based intent recognition on the standardized question. If the query question is in the intent recognition template, then generating and matching an answer based on the query statement of the intent recognition template. If not, then going to step S3; Step S3: Using graph search, according to the standardized question, in IDS-EMKG-QA, inference is performed based on the fault causal link and the equipment composition architecture to screen candidate answers; the RotatE graph embedding model is used to match and sort the candidate answers and the standardized question, and finally the maintenance plan recommendation result is obtained.

2. The maintenance plan matching and recommendation method based on causality and equipment structure enhancement according to claim 1 is characterized in that: The step S13: using the trained SimCSE model and the maintenance domain synonym library to replace and remove redundancy from the existing equipment maintenance knowledge graph IDS-EMKG to obtain a standardized equipment maintenance knowledge graph IDS-EMKG-QA, specifically includes: Step S131: traverse the node list NodeList in the IDS-EMKG, and for each node Node, replace the Node with a corresponding standard word in the maintenance field synonym library SynList; Step S132: extracting nodes of the same type as Node from NodeList to form a node subset NodeSubList; Step S133: Calculate the name similarity simScore between Node and each node NodeSub in NodeSubList based on the cosine similarity and the trained SimCSE model. The known cosine similarity calculation function is as follows: (1) in, Represents the cosine similarity of vectors a and b; The calculation formula of name similarity simScore is as follows: (2) If simScore is greater than the preset threshold k, the node pair [Node, NodeSub] is added to the set SynNode; Step S134: Manually verify all node pairs in SynNode to obtain the actual node combination SynNode' that needs to be merged; for each pair of nodes [Node, Node'] in SynNode', copy all relationships of Node' in IDS-EMKG to Node, and delete Node' and all relationships related to it; finally, obtain the standardized equipment maintenance knowledge graph IDS-EMKG-QA.

3. The maintenance plan matching and recommendation method based on causality and equipment structure enhancement according to claim 1 is characterized in that: The step S14: optimizing the question sentence using the query optimization algorithm based on SimCSE according to the maintenance domain synonym library and IDS-EMKG-QA to obtain a standardized question sentence, specifically includes: Step S141: Initialize an empty entity node set NodeSet; Step S142: by traversing the maintenance field synonym library SynList, the words in the question sentence Text are replaced with standard words in SynList; Step S143: Input Text into the BERT-Bi-LSTM model to obtain an entity name list NameList; Step S144: For each entity name Name in NameList, if it already exists in IDS-EMKG-QA, add it as a node to NodeSet; if it does not exist, take Name as input, find the most similar node name in IDS-EMKG through the SimCSE model, and add the node corresponding to the name to NodeSet; finally, return a set NodeSet containing all matching nodes as a normalized question.

4. The maintenance plan matching and recommendation method based on causality and equipment structure enhancement according to claim 1 is characterized in that: The step S2: performing template-based intent recognition on the standardized question, if the query question is in the intent recognition template, generating and matching an answer based on the query statement of the intent recognition template, specifically includes: Step S21: construct 6 types of question intention recognition templates based on IDS-EMKG-QA to cover basic questions and their variants: including: equipment-fault, cause-fault, fault-impact, fault-repair method, measurement point-fault, equipment-equipment; Step S22: Based on the question intention recognition template and the standardized question, a query statement and answer generation template based on the neo4j knowledge graph is constructed, and based on the query statement, nodes within one hop around the question entity node are matched, and the corresponding answer is returned.

5. The maintenance plan matching and recommendation method based on causality and equipment structure enhancement according to claim 1 is characterized in that: Step S3: using a graph search method, according to the standardized question, in IDS-EMKG-QA, infer and screen candidate answers based on the fault causal link and the device composition architecture; The RotatE graph embedding model is used to match and sort the candidate answers and the standardized question sentences, and finally the maintenance plan recommendation results are obtained, which specifically include: Step S31: Initialize boundary parameter m, device parameter n, node distance D and answer range k; Step S32: using the query optimization algorithm based on SimCSE to obtain the fault description FaultNode and the equipment description EquipName in the standardized question sentence; Step S33: Search for solution nodes within m hops around FaultNode in IDS-EMKG-QA to obtain the solution node set OperaNodeSet. If no result is found, increase the boundary parameter m and continue searching. Similarly, search for fault nodes within n hops around EquipName in IDS-EMKG-QA to obtain the fault node set FaultNodeSet. If no result is found, increase the equipment range parameter n and continue searching. For each node FaultNode in FaultNodeSet, if it has a triple {FaultNode, hasOperation, OperaNode} in IDS-EMKG-QA, add OperaNode to the set OperaNodeSet. Step S34: Input OperaNodeSet, FaultNode, and hasOperation into the RotatE model to obtain OperaEmbSet, FaultEmb, and hasOperaEmb respectively; it is known that the distance function defined by RotatE is as follows: (3) in, represents the Hadamard product, h represents the subject entity of the knowledge graph triple, r represents the relationship between the subject and the object in the knowledge graph triple, and t represents the object entity of the knowledge graph triple; For each device embedding vector OperaEmb in OperaEmbSet, the distance between it and the fault embedding vector FaultEmb is calculated by formula (3): , the result Add to set D; (4) Arrange the set D from small to large according to the absolute value, and select the first k with the smallest distance Corresponding A maintenance method set FinOperaSet is formed and returned as the final recommended maintenance solution.

Images