Water chilling unit fault diagnosis method fusing knowledge graph and large model

Abstract

The invention discloses a water cooling unit fault diagnosis method fusing a knowledge graph and a large model, and the method comprises the steps: 1, collecting and preprocessing a real fault text, enabling the large model to extract a fault entity and relation through prompt learning, and constructing a preliminary knowledge graph; 2, generating a standardized triple through entity alignment and knowledge fusion, complementing a potential relationship in combination with map reasoning and quality evaluation, and optimizing a map structure; 3, after related knowledge nodes in the knowledge graph are retrieved through input natural language fault description, the retrieved graph knowledge is combined with the large model, and reasoning and result optimization of fault diagnosis are completed; and 4, continuously updating the graph and the model according to user feedback, and continuously optimizing the knowledge graph content and the model reasoning effect. According to the method, the accuracy and the intelligent level of fault diagnosis of the water cooling unit can be effectively improved, so that the interpretability and the adaptive capacity are enhanced, and the problems of dependence on manpower, low diagnosis efficiency and slow knowledge updating in a traditional method are solved.

Description

Technical Field

[0001] The present invention belongs to the field of water-cooling unit fault diagnosis, and specifically relates to the construction of a water-cooling unit fault knowledge graph, system construction, and enhancing the accuracy of answers of a large model through the structured knowledge of the knowledge graph, so as to realize fault diagnosis of the water-cooling unit. Background Art

[0002] The strong light magnetic water cooling unit is an important auxiliary system to ensure the operation of the strong light magnetic test device. Its main function is to quickly remove the heat generated by the device during operation, prevent the device temperature from overheating, and ensure the continuous normal operation of the entire system. The structure of the strong light magnetic water cooling unit mainly consists of an evaporator, a condenser, and a water pump: the evaporator is responsible for absorbing heat from the equipment and heating the cooling water. The water pump transports the heated cooling water from the evaporator to the condenser to ensure the circulation of the cooling water. The condenser exchanges heat with the external environment and discharges the heat in the water, thereby achieving efficient heat dissipation of the equipment.

[0003] However, under high radiation, high load, and complex operating conditions, the strong light magnetic water-cooling unit is prone to failure due to performance degradation. Once the cooling fails, it not only affects the normal operation of the strong light magnetic test device, but may even cause damage to the magnets, posing a major safety hazard. Therefore, conducting intelligent monitoring and fault diagnosis of the operating status of the strong light magnetic water-cooling unit is of great theoretical and engineering significance for extending the service life of the strong light magnetic test device and achieving efficient operation and maintenance.

[0004] Currently, traditional fault diagnosis methods for water-cooled units still rely primarily on manual inspections and empirical judgment. This is not only labor-intensive and inefficient, but also significantly affected by human factors, making it difficult to detect potential problems in a timely manner. With the development of artificial intelligence technology, deep learning intelligent diagnostic methods have been widely used in equipment status identification and fault analysis for water-cooled units, making certain progress in improving diagnostic efficiency and identification accuracy. However, these methods generally suffer from the "black box" problem, and the model reasoning process lacks transparency and explainability, making it difficult to provide users with a clear basis for decision-making, which affects their widespread application in industrial scenarios. These problems have restricted the further development of intelligent diagnostic technology for water-cooled units and have also placed higher demands on the reliability and practicality of existing diagnostic systems. Summary of the Invention

[0005] In response to the problems of low accuracy, poor efficiency, reliance on manual experience, weak interpretability and delayed knowledge updating in the above-mentioned water-cooling unit fault diagnosis, the present invention proposes a water-cooling unit fault diagnosis method that integrates knowledge graphs and large models, in order to improve the accuracy, degree of automation and system interpretability of fault diagnosis, thereby realizing intelligent monitoring of the operating status of the water-cooling unit and efficient fault handling, thereby enhancing the operating stability and safety assurance capabilities of key equipment.

[0006] In order to achieve the above-mentioned object, the present invention adopts the following technical solutions:

[0007] A water-cooling unit fault diagnosis method integrating a knowledge graph and a large model is characterized in that it includes the following steps:

[0008] Step 1: Collect the original fault text dataset from the operation process of the water-cooling unit, clean and remove duplicates from the original fault text dataset, and obtain the preprocessed fault text dataset. ,in, Indicates the A pre-processed fault text, Indicates the total number of pre-processed fault texts, ;

[0009] Step 2: The designed knowledge extraction prompt words are input into the large model for prompt learning extraction, and the first Triplet extraction results , thus obtaining The triple extraction result set is ,in, Indicates the Individual entities, Indicates the A tail entity, express and relationship between;

[0010] Step 3: From Extract all head entities and tail entities and form an entity set ,calculate The edit distance between any two entities in is normalized to their similarity. If the similarity is less than the set threshold , then the corresponding two entities are merged into one standard entity, otherwise, they are not merged, thus obtaining a knowledge-ablated entity set consisting of all standard entities. and its corresponding triple set ;

[0011] Step 4: The triple set after knowledge ablation , build a preliminary knowledge graph of water-cooling unit faults ,and The entity types in it include fault phenomenon, fault cause, fault location and solution;

[0012] Step 5: Use the graph embedding function to All entities and relationships in the map are mapped into a low-dimensional vector space, thereby obtaining a set of entity embedding vectors and its relation embedding vector set ,in, and Respectively Middle The embedding vector of the head entity and the The embedding vector of the tail entity, express and The embedding vector of the relationship between them; express The total number of triplets in ;

[0013] according to and ,use Inference model pair Perform potential relationship prediction and missing completion to obtain the inferred water-cooling unit fault map ; Calculated using triple score evaluation function The confidence score of each triple in , if the confidence score is less than the set confidence score threshold , then delete the corresponding triplet, otherwise, retain the corresponding triplet, and thus obtain the optimized water-cooling unit fault map ;

[0014] Step 6: Obtain a text description of the fault phenomenon and from After extracting key fault information, a keyword set is formed ,in, Indicates the Keywords, is the total number of keywords;

[0015] Calculate the Keywords Respectively After the cosine similarity of each entity embedding vector of the fault phenomenon is calculated, the entities with the same The entities with the highest similarity are embedded in the vector and form the Keywords The candidate fault phenomenon entity set ;

[0016] from Search and The related triples constitute the Subset ,and The entity types in the table cover the corresponding fault phenomena, causes, locations and solutions;

[0017] Step 7: As a priori knowledge guide, it forms a question template together with the designed prompt words, which is input into the large language model for deduction, thereby outputting a more accurate first Fault diagnosis result set ,in, Indicates the Keywords No. Fault diagnosis results, and Including fault phenomenon, fault cause, fault location and solution, Indicates the total number of fault diagnosis results;

[0018] Step 8: According to After eliminating the corresponding faults in the operation of the water-cooling unit, the elimination results are obtained, and then the Evaluate and get Evaluation results , used to update and large models.

[0019] The water-cooling unit fault diagnosis method integrating the knowledge graph and the large model described in the present invention is also characterized in that step 3 includes:

[0020] Step 3.1, from Take any two entities and denote them as and , using formula (1) to calculate the entity and another entity The minimum edit distance :

[0021] (1)

[0022] In formula (1), Representing an entity Before Characters to entities Before The minimum edit distance of characters, and Represent entities separately No. characters and entities No. characters, Representing an entity Before characters, Representing an entity Before characters, represents the replacement cost function, if , then let 0 means no replacement is performed Otherwise, let 1 means to perform replacement ; Indicates deletion. Indicates insertion;

[0023] Step 3.2: Use formula (2) to get any entity and any entity among all entities except itself Similarity index between ;

[0024] (2)

[0025] In formula (2), Representing an entity The maximum edit distance value among the edit distances of all entities except itself, Representing an entity and any entity Edit distance;

[0026] like , then the entity and entities Merge into a standard entity, otherwise, do not merge, where Is the threshold set, so as to obtain the entity set after knowledge ablation consisting of all standard entities and its corresponding triple set .

[0027] Furthermore, in step 5 The inference model is to use formula (3) and formula (4) to Perform potential relationship prediction and missing completion to obtain the inferred water-cooling unit fault map :

[0028] (3)

[0029] (4)

[0030] In formula (4), express The scoring function of the authenticity of the embedding vector triple corresponding to the triple in , then The corresponding triplet is Otherwise, discard the triples in The corresponding triples, where The threshold parameter to be set.

[0031] Furthermore, in step 5, formula (5) is used to calculate Any x-th triple in Confidence score ,like , then delete the triple , otherwise, keep, where, The confidence threshold parameters are set as follows: (5)

[0032] In formula (5), represents the embedding vector of the x-th head entity, represents the embedding vector of the x-th tail entity, express and The relationship embedding vector between Represents a knowledge graph embedding model Pair triples The score, Indicates the measurement of the xth head entity and the xth tail entity The similarity score between Represents a triple Credibility score of 、 、 Parameters representing the three weights to be adjusted.

[0033] The electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the water cooling unit fault diagnosis method, and the processor is configured to execute the program stored in the memory.

[0034] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium. The computer program is characterized in that when the computer program is run by a processor, the program of the water-cooling unit fault diagnosis method is executed.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. The present invention extracts fault information by introducing a large language model driven by prompt learning. It can automatically identify the key entities of fault phenomena, causes, affected parts and solutions from unstructured texts, replacing the traditional manual analysis method that relies on expert experience, and effectively improving the intelligence and standardization of diagnostic data processing.

[0037] 2. This invention innovatively introduces knowledge fusion and graph enhancement, solves the inconsistent representation of fault entities through entity alignment and semantic normalization, and realizes entity standardization by combining edit distance and similarity, thus ensuring the consistency and quality of the knowledge graph. It also constructs a graph reasoning module that supports embedded models, and uses potential triple mining and confidence assessment to complete the knowledge graph and eliminate low-credibility information, thus realizing dynamic learning and continuous optimization, and significantly improving the robustness and adaptability of the fault diagnosis method.

[0038] 3. This invention deeply integrates the knowledge graph with the big model to construct the Graph RAG question-answering diagnostic system, which not only leverages the advantages of the knowledge graph's clear knowledge structure and semantic traceability, but also enhances the reasoning ability of the big model, achieving high-accuracy, multi-angle diagnostic reasoning for complex fault scenarios, and enhancing the interpretability and user trust of the fault diagnosis method.

[0039] 4. The present invention has a continuous learning and dynamic update mechanism, which can continuously optimize the knowledge graph content and model reasoning effect based on user feedback, adapt to the ever-evolving equipment operating environment and fault types, break through the bottleneck of slow knowledge update in traditional diagnostic systems, and significantly improve the adaptability and service life of fault diagnosis methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a flow chart of the intelligent question-answering method for water-cooling unit fault diagnosis that integrates knowledge graph and large model of the present invention;

[0041] Figure 2 It is a schematic diagram of the knowledge graph triple of the water cooling unit fault part of the present invention;

[0042] Figure 3 This is a system webpage diagram corresponding to the water-cooling unit fault diagnosis question-answering method based on the knowledge graph and large model of the present invention;

[0043] Figure 4 This is a system data display diagram corresponding to the water-cooling unit fault diagnosis question-answering method based on the knowledge graph database information of the present invention;

[0044] Figure 5 This is a schematic diagram of knowledge retrieval based on Graph RAG technology of the present invention. DETAILED DESCRIPTION

[0045] In this embodiment, a water-cooling unit fault diagnosis method that integrates knowledge graphs and large models is to collect and process real fault text data, use prompt learning to guide the large model to extract key entities and relationship information, construct normalized triples to generate a preliminary knowledge graph, and optimize the graph structure through graph reasoning and quality assessment mechanisms. After the user enters the fault description, the key information in the graph is matched and combined with the large model for deep reasoning to optimize the accuracy and interpretability of the diagnosis results. It is continuously iterated and updated based on user feedback to achieve dynamic learning and adaptive optimization of the fault diagnosis model. Figure 1 Specifically, the method includes the following steps:

[0046] Step 1: Collect the original fault text dataset from the operation process of the water-cooling unit, clean and remove duplicates from the original fault text dataset, and obtain the preprocessed fault text dataset. ,in, Indicates the A pre-processed fault text, Indicates the total number of pre-processed fault texts, .

[0047] The dataset used in this example is based on the operating mechanism of a strong optical magnetic water-cooling unit and was generated through multi-dimensional simulation of water-cooling unit maintenance scenarios under different operating conditions. The dataset includes textual information on fault descriptions, repair procedures, maintenance measures, and troubleshooting records during the operation of the water-cooling unit. It also systematically integrates key information from the maintenance process of the strong optical magnetic water-cooling unit, providing reliable data support for fault diagnosis research in strong optical magnetic water-cooling units.

[0048] Step 2: The designed knowledge extraction prompt words are input into the large model for prompt learning extraction, and the first Triplet extraction results , where the entity Including fault symptoms ( )、Cause of failure( )、Fault location( ) and solution ( ),relation Derived from the relationship set {cause, location, solution}, these entities Type and relationship set It is manually defined based on the original fault text data, thus obtaining The triple extraction result set is ,in, Indicates the Individual entities, Indicates the A tail entity, express and The relationship between them.

[0049] Step 3: From Extract all head entities and tail entities and form an entity set ,from Take any two entities and denote them as and , using formula (1) to calculate the entity and another entity The minimum edit distance between :

[0050] (1)

[0051] Formula (1) means: If and The prefix length of an entity in is 0, then let and The minimum edit distance between is the length of a non-empty entity, otherwise the calculation will and The minimum number of operations required to transform one entity into another entity, and as ; The operations include: deleting ,insert or replace ;

[0052] In formula (1), Representing an entity Before Characters to entities Before The minimum edit distance of characters, and Represent entities separately No. characters and entities No. characters, Representing an entity Before characters, Representing an entity Before characters, represents the replacement cost function, if , then let 0 means no replacement is performed ,otherwise, 1 means to perform replacement , the entity With entity collection All entities except itself Calculate the edit distance.

[0053] Using formula (2) to get any entity With entity collection All entities except itself Similarity index between ;

[0054] (2)

[0055] In formula (2), Representing an entity With entity collection The maximum edit distance value among the edit distances of all entities except itself, Representing an entity and entities The edit distance is calculated. , then the entity and entities Merge, otherwise, do not merge, where It is the set threshold. For example, after calculating the edit distance between "high voltage alarm" and "high voltage protection alarm", it is found that , so they can be unified into the standard entity "high voltage alarm", thus obtaining the entity set after knowledge ablation composed of all standard entities and its corresponding triple set .

[0056] Step 4: The triple set after knowledge ablation Constructing a preliminary knowledge graph of water-cooling unit faults ,and The entity types in it include fault phenomenon, fault cause, fault location and solution, and the relationships include cause, location and solution.

[0057] Step 5: Use the graph embedding function to All entities and relationships in the map are mapped into a low-dimensional vector space, thereby obtaining a set of entity embedding vectors and its relation embedding vector set ,in, and Respectively Middle The embedding vector of the head entity and the The embedding vector of the tail entity, express and The embedding vector of the relationship between them; express The total number of triplets in ;

[0058] according to and ,use Inference model pair Perform potential relationship prediction and missing completion to obtain the inferred water-cooling unit fault map ;in The inference model is to use formula (2) and formula (3) to Perform potential relationship prediction and missing completion ;

[0059] (3)

[0060] (4)

[0061] In formula (4), express The scoring function of the authenticity of the embedding vector triple corresponding to the triple in , then The corresponding triplet is Otherwise, discard the triples in The corresponding triples, where The threshold parameter to be set.

[0062] Calculated using triple score evaluation function The confidence score of each triple in , if the confidence score is less than the set confidence score threshold , then delete the corresponding triplet, otherwise, keep the corresponding triplet;

[0063] calculate Any x-th triple in Confidence score : (5)

[0064] In formula (5), represents the x-th head entity embedding vector, represents the x-th tail entity embedding vector, express and The relationship embedding vector between Represents a knowledge graph embedding model Pair triples The score, Indicates the measurement of the xth head entity and the xth tail entity The similarity score between Represents a triple The credibility score of 、 、 Represents the three parameters whose weights are to be adjusted, so as to obtain the optimized fault map of the water-cooling unit , the knowledge graph finally constructed is as follows Figure 2 shown.

[0065] Step 6: Build a water-cooling unit fault diagnosis question-answering system. The system inputs information on the web page such as Figure 3 Get the text description of the fault phenomenon and from Extract key fault information and form a keyword set ,in, Indicates the Keywords, is the total number of keywords;

[0066] Use formula (6) to calculate the Keywords Respectively The cosine similarity of each entity embedding vector about the fault phenomenon;

[0067] (6)

[0068] In formula (6), represents each keyword embedding vector, Represents the embedding vector of the fault phenomenon entity in the knowledge graph, and after calculation and screening, The entities with the highest similarity are embedded in the vector and form the Keywords The candidate fault phenomenon entity set ,from Search and The related triples constitute the Subset ,and The entity types in the table cover the corresponding fault phenomena, causes, locations and solutions;

[0069] Step 7: Under the Graph RAG technical framework, its specific architecture is as follows Figure 4 As shown, As a priori knowledge guide, it forms a question template together with the designed prompt words and is input into the large language model for deduction. By using the structured triples as priori knowledge guides and combining the general background knowledge and language reasoning ability mastered by the LLM itself, the large model can complement and expand the existing knowledge and further deduce a more accurate first Fault diagnosis result set , thereby improving the accuracy and explainability of fault diagnosis, among which, Indicates the Keywords No. Fault diagnosis results, and Including fault phenomenon, fault cause, fault location and solution, Indicates the total number of fault diagnosis results. Graph RAG technology embodies the collaborative diagnosis mechanism of "knowledge retrieval + reasoning generation", effectively integrating knowledge graph knowledge with large model intelligence;

[0070] Step 8: According to After eliminating the corresponding faults in the operation of the water-cooling unit, the elimination results are obtained, and then the Evaluate and get Evaluation results , used to update And large models, the system updates the information such as Figure 5 shown.

[0071] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0072] In this embodiment, a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above method are executed.

[0073] In summary, the present invention proposes a water-cooling unit fault diagnosis method that integrates knowledge graphs and large models. The method drives the large model to construct a fault knowledge graph through prompt learning, generates standardized triples through entity alignment and fusion, and then completes potential relationships and optimizes the knowledge graph structure through graph reasoning and quality assessment. The Graph RAG technology architecture is combined with the large model for fault reasoning and diagnosis, realizing system adaptive learning, significantly improving the accuracy, efficiency, interpretability and practicality of diagnosis, and overcoming the shortcomings of traditional diagnostic methods.

Claims

1. A water-cooling unit fault diagnosis method integrating knowledge graph and large model, characterized in that: The following steps are involved: Step 1: Collect the original fault text dataset from the operation process of the water-cooling unit, clean and remove duplicates from the original fault text dataset to obtain the preprocessed fault text dataset. ,in, Indicates the A pre-processed fault text, Indicates the total number of pre-processed fault texts, ; Step 2: The designed knowledge extraction prompt words are input into the large model for prompt learning extraction, and the first Triplet extraction results , thus obtaining The triple extraction result set is ,in, Indicates the Individual entities, Indicates the A tail entity, express and relationship between; Step 3: From Extract all head entities and tail entities and form an entity set ,calculate The edit distance between any two entities in is normalized to their similarity. If the similarity is less than the set threshold , then the corresponding two entities are merged into one standard entity, otherwise, they are not merged, thus obtaining a knowledge-ablated entity set consisting of all standard entities. and its corresponding triple set ; Step 4: The triple set after knowledge ablation , build a preliminary knowledge graph of water-cooling unit faults ,and The entity types in it include fault phenomenon, fault cause, fault location and solution; Step 5: Use the graph embedding function to All entities and relationships in the map are mapped into a low-dimensional vector space, thereby obtaining a set of entity embedding vectors and its relation embedding vector set ,in, and Respectively Middle The embedding vector of the head entity and the The embedding vector of the tail entity, express and The embedding vector of the relationship between them; express The total number of triplets in ; according to and ,use Inference model pair Perform potential relationship prediction and missing completion to obtain the inferred water-cooling unit fault map ; Calculated using triple score evaluation function The confidence score of each triple in , if the confidence score is less than the set confidence score threshold , then delete the corresponding triplet, otherwise, retain the corresponding triplet, and thus obtain the optimized water-cooling unit fault map ; Step 6: Obtain a text description of the fault phenomenon and from After extracting key fault information, a keyword set is formed ,in, Indicates the Keywords, is the total number of keywords; Calculate the Keywords Respectively After the cosine similarity of each entity embedding vector of the fault phenomenon is calculated, the entities with the same The entities with the highest similarity are embedded in the vector and form the Keywords The candidate fault phenomenon entity set ; from Search and The related triples constitute the Subset ,and The entity types in the table cover the corresponding fault phenomena, causes, locations and solutions; Step 7: As a priori knowledge guide, it forms a question template together with the designed prompt words, which is input into the large language model for deduction, thereby outputting a more accurate first Fault diagnosis result set ,in, Indicates the Keywords No. Fault diagnosis results, and Including fault phenomenon, fault cause, fault location and solution, Indicates the total number of fault diagnosis results; Step 8: According to After eliminating the corresponding faults in the operation of the water-cooling unit, the elimination results are obtained, and then the Evaluate and get Evaluation results , used to update and large models.

2. A water-cooling unit fault diagnosis method integrating knowledge graph and large model according to claim 1, characterized in that: Step 3 includes: Step 3.1, from Take any two entities and denote them as and , using formula (1) to calculate the entity and another entity The minimum edit distance : (1) In formula (1), Representing an entity Before Characters to entities Before The minimum edit distance of characters, and Represent entities separately No. characters and entities No. characters, Representing an entity Before characters, Representing an entity Before characters, represents the replacement cost function, if , then let 0 means no replacement is performed Otherwise, let 1 means to perform replacement ; Indicates deletion. Indicates insertion; Step 3.2: Use formula (2) to get any entity and any entity among all entities except itself Similarity index between ; (2) In formula (2), Representing an entity The maximum edit distance value among the edit distances of all entities except itself, Representing an entity and any entity Edit distance; like , then the entity and entities Merge into a standard entity, otherwise, do not merge, where Is the threshold set, so as to obtain the entity set after knowledge ablation consisting of all standard entities and its corresponding triple set .

3. The water-cooling unit fault diagnosis method integrating knowledge graph and large model according to claim 1 is characterized in that: In step 5 The inference model is to use formula (3) and formula (4) to Perform potential relationship prediction and missing completion to obtain the inferred water-cooling unit fault map : (3) (4) In formula (4), express The scoring function of the authenticity of the embedding vector triple corresponding to the triple in , then The corresponding triple is Otherwise, discard the triples in The corresponding triples, where The threshold parameter to be set.

4. The water-cooling unit fault diagnosis method integrating knowledge graph and large model according to claim 1 is characterized in that: In step 5, formula (5) is used to calculate Any x-th triple in Confidence score ,like , then delete the triple , otherwise, keep, where, The confidence threshold parameters are set as follows: (5) In formula (5), represents the embedding vector of the x-th head entity, represents the embedding vector of the x-th tail entity, express and The relationship embedding vector between Represents a knowledge graph embedding model Pair triples The score, Indicates the measurement of the xth head entity and the xth tail entity The similarity score between Represents a triple Credibility score of 、 、 Parameters representing the three weights to be adjusted.

5. An electronic device comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the water-cooling unit fault diagnosis method according to any one of claims 1 to 4, and the processor is configured to execute the program stored in the memory.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the computer program executes the water cooling unit fault diagnosis method according to any one of claims 1 to 4.

Images