Industrial pump fault intelligent diagnosis method based on GraphRAG large model

Abstract

The invention discloses an industrial pump fault intelligent diagnosis method based on a GraphRAG large model, and the method comprises the following steps: collecting fault records and vibration signals of an industrial pump, and constructing a fault and characteristic spectrum; obtaining a fault type of the feature parameters by using a convolutional neural network, and generating a fusion knowledge graph in combination with a Word2Vec model and cosine similarity; the method comprises the following steps of: inputting map data into an LLM large model, and optimizing a prompt template by using a prompttune function; extracting node information of entities, communities and the like by using a GraphRAG index construction module; extracting related entities in a user request through a convolutional neural network or LLM, and matching the related entities with the map nodes to obtain information; and mapping the information to index data, and outputting a diagnosis report by utilizing a retrieval generation module. Through the GraphRAG technology, the fault diagnosis precision and the maintenance decision-making efficiency are improved, and the industrial production stability and safety are enhanced.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing and fault diagnosis, and in particular to an intelligent fault diagnosis method for industrial pumps based on a GraphRAG large model. Background Art

[0002] In modern industrial production, pumps are key equipment for fluid transportation, and their operating status directly affects the stability and safety of the production line. However, due to long-term operation or improper maintenance, industrial pumps are prone to various faults, such as mechanical wear, seal failure, motor overheating, etc. These faults will not only lead to production interruptions, but may also cause safety accidents and economic losses. Therefore, effective fault diagnosis and prediction of industrial pumps is of great significance to ensure production safety and reduce maintenance costs.

[0003] Traditional industrial pump fault diagnosis methods mainly rely on manual experience judgment, or use a single data analysis method (such as vibration analysis, etc.). However, these methods often have problems such as low accuracy and poor real-time performance, and it is difficult to meet the high requirements of modern industry for fault diagnosis. In recent years, with the development of big data, artificial intelligence technology and natural language processing technology. These methods provide new solutions for industrial pump fault analysis and retrieval.

[0004] Large Model is an important model in the field of artificial intelligence, usually referring to a deep learning model with a very large number of parameters and a very large amount of data. Large models usually consist of millions to billions of parameters and require a large amount of data and computing resources for training and reasoning. Due to its huge size, large models have very powerful representation and generalization capabilities and can perform well in various tasks, such as speech recognition, natural language processing, computer vision, etc. Among them, Large Language Model (LLM) is a large-scale deep learning model designed specifically for natural language processing tasks. This type of model focuses on understanding and generating human language, mainly processing text data, and is used for tasks such as translation, question and answer, text summarization, and dialogue generation. In addition, GraphRAG (Graphs Retrieval Augmented Generation) technology stands for "Graph-based Retrieval Augmented Generation" and is used for question and answer on private text corpora. GraphRAG does not simply find text fragments, but builds a structured and hierarchical relationship graph of information. Combining graph-based knowledge retrieval with LLM large models captures entities, relationships, and key statements in large-scale text information. Different from traditional RAG methods, GraphRAG can enhance the ability of LLM to understand and synthesize complex data sets and their relationships, thereby producing more accurate responses.

[0005] This paper proposes an intelligent diagnosis method for industrial pump faults based on the GraphRAG large model. The method aims to capture the complex relationships between entities such as equipment, fault types, and maintenance measures by constructing an industrial pump knowledge graph, and use GraphRAG technology to effectively retrieve and utilize this knowledge. In addition, by integrating the powerful generation capabilities of the LLM large model, the system can retrieve the most relevant information fragments from existing fault analysis cases based on understanding user queries, and generate high-quality fault diagnosis reports and maintenance recommendations. Summary of the invention

[0006] The purpose of the present invention is to improve the accuracy and efficiency of industrial pump fault diagnosis, provide a scientific basis for equipment maintenance, and provide an intelligent diagnosis method for industrial pump faults based on the GraphRAG large model.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] The intelligent diagnosis method of industrial pump fault based on GraphRAG large model includes the following steps:

[0009] Step 1: Collect data and build a knowledge graph. Collect basic information and fault analysis records of industrial pumps, extract entities of industrial pumps and the relationships between them to build a fault graph, where entities include components, fault types, fault phenomena, fault causes and fault handling; collect vibration signals, extract feature parameters to build a feature graph.

[0010] Step 2: Fusion of knowledge graph. Input the feature parameters into the convolutional neural network for fault classification to obtain the fault type, then use the Word2Vec model to generate word vectors, and perform cosine similarity calculation to obtain similar node pairs between the fault graph and the feature graph, and then obtain the fused knowledge graph.

[0011] Step 3: Optimize the prompt template. Input the data of the fused knowledge graph into the LLM large model module, and use the prompt_tune function in GraphRAG to automatically adapt and optimize the prompt template;

[0012] Step 4: Extract node information using the GraphRAG index building module. Use LLM to segment the data text into text blocks; use the optimized prompt template to identify and extract instances of graph nodes and edges in text blocks; use LLM to provide a brief summary description for each entity and relationship, so that each element has a unified and descriptive text; divide the entities into communities using the hierarchical Leiden algorithm, and generate entity, relationship and community summary information based on this; use the Node2Vec algorithm to generate a vector representation of the graph, and finally summarize all the node information, including text blocks, entities, communities, summaries and covariates.

[0013] Step 5: Input and analysis of user requests. When the user selects vibration signal input, the characteristic parameters of the signal are extracted, and the convolutional neural network is used for fault classification. The characteristic parameters are matched with the nodes in the characteristic map part by cosine similarity, and classified as a data engine. The classification results are matched with the nodes in the fault map part, and located as a knowledge engine to obtain relevant node information. In addition, when the user selects problem consultation, LLM is used to extract relevant entities in the problem, and the nodes in the characteristic map part are matched by cosine similarity to obtain relevant node information.

[0014] Step 6: Use the GraphRAG retrieval generation module to search the index and output the diagnostic report. Map the relevant node information to the index data, and then use the local search and global search in the retrieval generation module to integrate and obtain the final fault diagnosis report, which is then output to the user.

[0015] Among them, the specific steps of step 1 include: obtaining the basic information file of the industrial pump by collecting literature materials, mainly describing the main components and fault types of the industrial pump; the relevant information on the specific industrial pump fault maintenance is some common faults and treatment methods of clean water pumps and slurry pumps, mainly involving three aspects: equipment phenomena, fault causes and treatment methods; the characteristic parameters selected are frequency domain characteristic parameters, including center of gravity frequency, average frequency, root mean square frequency, and frequency standard deviation.

[0016] The specific steps of step 2 include: inputting the feature parameters into the convolutional neural network for fault classification to obtain the fault type, and then using the Word2Vec model to generate word vectors, where the important model parameters include vector dimension, window size, minimum word frequency, number of threads, initial learning rate, training model and optimization objective function; performing cosine similarity calculation to obtain similar node pairs between the fault map and the feature map, and then obtaining the fused knowledge map.

[0017] The specific steps of step 3 include: the LLM model selected is the Tongyi Qianwen model, the chat model selected is the qwen-turbo model, and the embedding model selected is the text-embedding-v1 model. For the prompt_tune function, the parameters that need to be set mainly include the method for selecting documents, the limit for loading text units, the input language, the maximum number of tokens for generating prompts, and the chunk-size used when generating text units.

[0018] The specific steps of step 4 include: The process of the hierarchical Leiden algorithm includes: applying the standard Leiden algorithm to perform community detection on the initial network, and finding the best community division by optimizing modularity or other quality functions; then, agglomerating each detected community into a super node to construct a new agglomeration graph; recursively applying the above community detection and agglomeration process to the new graph to form a higher-level community structure; this process continues until the preset stopping condition is reached; outputting a multi-level community structure to reveal the multi-scale community characteristics in the network.

[0019] In addition, the process of the Node2Vec algorithm includes: setting two hyperparameters, the return parameter and the input and output parameter, in a given graph to control the behavior of the random walk, and generating multiple random walk sequences with biases starting from each node; these sequences are used as training data and input into the Word2Vec model to learn the low-dimensional dense vector representation of each node; after multiple rounds of iterative optimization, the model converges and outputs a node embedding vector that can reflect the structural characteristics of the graph.

[0020] The specific steps of step 5 include: when the user selects vibration signal input, the characteristic parameters of the signal are extracted, the convolutional neural network is used to classify the fault, the characteristic parameters are matched with the nodes in the characteristic spectrum part through cosine similarity, and classification is performed as a data engine, and the classification results are matched with the nodes in the fault spectrum part, and positioning is performed as a knowledge engine to obtain relevant node information; in addition, when the user selects problem consultation, the Tongyi Qianwen LLM is used to extract relevant entities in the problem, and the nodes in the characteristic spectrum part are matched through cosine similarity to also obtain relevant node information.

[0021] The specific steps of step 6 include: the basic process of local search in the query mechanism is: after the user inputs the question, the system extracts the matching entities and their descriptions and embedded information, generates candidate text units, candidate community reports, and candidate entities; these candidate units are sorted and filtered, and priority results are output; finally, all priority results are summarized, a response is generated and provided to the user.

[0022] In addition, the basic process of global search is as follows: after the user enters a question, the system processes it by assigning it to different community report batches. Each community report batch generates intermediate responses and scores them, and then sorts and filters these score results to retain the best results. Finally, the system aggregates the intermediate responses with higher scores, generates the final response, and returns it to the user.

[0023] Compared with the prior art, one or more embodiments of the present invention may have the following advantages: achieving efficient and accurate diagnosis of industrial pump failures, providing scientific and reliable decision support for equipment maintenance, and significantly improving the stability and safety of industrial production. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0025] Figure 1 is an overall flow chart of the method of the present invention;

[0026] Figure 2 Constructing a flow chart for the relevant index of the method of the present invention;

[0027] Figure 3 The diagram is a schematic diagram of the index and relationship distribution of the method of the present invention;

[0028] Figure 4 A local search flow chart of the method of the present invention;

[0029] Figure 5 This is a global search flow chart of the method of the present invention. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical solution and advantages of the implementation of the present invention clearer, the following will be further elaborated in combination with the technical solution and drawings of the present invention.

[0031] like Figure 1 As shown, the present invention is an intelligent diagnosis method for industrial pump faults based on the GraphRAG large model, comprising the following implementation steps:

[0032] Step 1: Collect data and build a knowledge graph. Collect basic information and fault analysis records of industrial pumps, extract entities of industrial pumps and the relationships between them to build a fault graph, where entities include components, fault types, fault phenomena, fault causes and fault handling; collect vibration signals, extract feature parameters to build a feature graph;

[0033] Step 2: Fusion of knowledge graph. Input feature parameters into the convolutional neural network for fault classification to obtain fault types, then use the Word2Vec model to generate word vectors, and perform cosine similarity calculation to obtain similar node pairs between the fault graph and the feature graph, and then obtain the fused knowledge graph;

[0034] Step 3: Optimize the prompt template. Input the data of the fused knowledge graph into the LLM large model module, and use the prompt_tune function in GraphRAG to automatically adapt and optimize the prompt template;

[0035] Step 4: Extract node information using the GraphRAG index building module. Use LLM to segment the data text into text blocks; use the optimized prompt template to identify and extract instances of graph nodes and edges in text blocks; use LLM to provide a brief summary description for each entity and relationship, so that each element has a unified and descriptive text; divide the entities into communities using the hierarchical Leiden algorithm, and generate entity, relationship and community summary information based on this; use the Node2Vec algorithm to generate a vector representation of the graph, and finally summarize all the node information, including text blocks, entities, communities, summaries and covariates.

[0036] Step 5: Input and analysis of user requests. When the user selects vibration signal input, the characteristic parameters of the signal are extracted, and the convolutional neural network is used to classify the fault. The characteristic parameters are matched with the nodes in the characteristic map part by cosine similarity, and classified as a data engine. The classification results are matched with the nodes in the fault map part, and located as a knowledge engine to obtain relevant node information. In addition, when the user selects problem consultation, LLM is used to extract relevant entities in the problem, and the nodes in the characteristic map part are matched by cosine similarity to obtain relevant node information.

[0037] Step 6: Use the GraphRAG retrieval generation module to search the index and output the diagnostic report. Map the relevant node information to the index data, and then use the local search and global search in the retrieval generation module to integrate and obtain the final fault diagnosis report, which is then output to the user.

[0038] This example constructs an intelligent diagnosis method for industrial pump faults based on the GraphRAG large model, which achieves efficient and accurate diagnosis of industrial pump faults and provides scientific and reliable decision support for equipment maintenance. The method first collects and preprocesses relevant data, then optimizes prompts, and uses LLM to build relevant indexes; finally, the problem is parsed and queried and the results are displayed.

[0039] The method is described in detail below.

[0040] In this embodiment, the step 1 is specifically as follows:

[0041] The collected industrial pump data include: basic information files of industrial pumps are compiled by collecting literature, mainly describing the main components and fault types of industrial pumps. Related data on specific industrial pump fault maintenance, including some common faults and treatment methods of clean water pumps and slurry pumps, mainly involving equipment phenomena, fault causes and treatment methods.

[0042] The characteristic parameters selected are frequency domain features that perform better in processing periodic signals and complex vibration modes, mainly including center of gravity frequency, average frequency, root mean square frequency, and frequency standard deviation.

[0043] The center frequency is a weighted average of the power spectrum amplitude as the weight, and will move to the position with a larger power spectrum amplitude. The center frequency is used to describe the frequency of the signal component with a larger component in the spectrum, which can reflect the distribution of the signal power spectrum. Its calculation formula is:

[0044]

[0045] Where S(k) is the corresponding power spectrum value; f k is the frequency amplitude of the corresponding point.

[0046] The average frequency is the average value of the power spectrum. The average frequency can provide information about the distribution of the signal frequency components and understand the main frequency components of the signal. The calculation formula is:

[0047]

[0048] Where N is the total number of sample points.

[0049] The root mean square frequency is used to describe the width of the signal frequency distribution and provides a statistical measure of the distribution of signal energy on different frequency components. The root mean square frequency can be regarded as the "root mean square frequency position" of the signal in the frequency domain, that is, the root mean square value of the signal energy distribution. The calculation formula is:

[0050]

[0051] The frequency standard deviation is used to describe the degree of dispersion of the signal frequency distribution. It provides a statistical measure of the distribution of signal energy on different frequency components and can be used to measure how the frequency components of the signal are dispersed. The larger the value of the frequency standard deviation, the more dispersed the data distribution is, and the smaller the value, the more concentrated the data distribution is. The calculation formula is:

[0052]

[0053] In this embodiment, the step 2 is specifically as follows:

[0054] The feature parameters are input into the convolutional neural network for fault classification to obtain the fault type. Then the Word2Vec model is used to generate word vectors and perform cosine similarity calculation to obtain similar node pairs between the fault graph and the feature graph, and then the fused knowledge graph is obtained.

[0055] The main parameters of the Word2Vec model are specifically set as: vector dimension is 200, window size is 20, minimum word frequency is 1, number of threads is 4, initial learning rate is 0.025, training model is Skip-gram model and optimization objective function is hierarchical softmax.

[0056] In addition, cosine similarity is a similarity calculation method based on the vector space model, which is used for text similarity calculation. Its calculation formula is:

[0057]

[0058] Among them, A and B are two vectors; ||A|| and ||B|| are their moduli respectively.

[0059] By calculating the cosine similarity between the fault graph nodes and the feature graph nodes, similar node pairs are obtained, and then the fused knowledge graph is obtained.

[0060] In this embodiment, the step 3 is specifically as follows:

[0061] Among them, the main configuration of Tongyi Qianwen LLM includes: the chat model is qwen-turbo model, the embedding model is text-embedding-v1 model. For the prompt_tune function, the method of selecting documents is set to random, the limit of loading text units is set to 15, the input language is set to Chinese, the maximum number of tokens for generating prompts is set to 2000, and the chunk-size used when generating text units is set to 20.

[0062] In this embodiment, the step 4 is specifically as follows:

[0063] like Figure 2It is a flowchart for building relevant index data. It uses Tongyi Qianwen LLM to segment the knowledge graph data text into text blocks; uses the optimized prompt template to identify and extract instances of graph nodes and edges in text blocks; uses LLM to provide a brief summary description for each entity and relationship, so that each element has a unified and descriptive text; divides the entities into communities through the hierarchical Leiden algorithm, and generates entity, relationship and community summary information based on this; uses the Node2Vec algorithm to generate a vector representation of the graph to help understand the implicit structure of the graph, and finally summarizes the index data information such as text blocks, entities, communities, summaries and covariates. Figure 3 The diagram shows the distribution of the constructed indexes.

[0064] The hierarchical Leiden algorithm is used to divide the entities into communities, and the specific steps are as follows:

[0065] Step 1: Calculate modularity Q. Modularity is a commonly used community quality indicator and is defined as:

[0066]

[0067] Among them A ij is an element of the adjacency matrix, representing the edge weight between node i and node j; k i and k j are the degrees of node i and node j respectively; m is the total weight of the edges (or the number of edges) in the network.

[0068] Step 2, community optimization, first treats each node as an independent community, and then for each node i, calculates the modularity gain ΔQ of moving it to the community to which neighbor node j belongs:

[0069]

[0070] Where N(i) is the set of neighbors of node i; C(j) is the community to which node j belongs.

[0071] Then move node i to the community with the largest modularity gain until there are no nodes to move, to increase the modularity.

[0072] Step 3: Construct a cohesive graph. First, consider each community as a supernode. Then, the edge weight ω between two supernodes in the cohesive graph is CC' Defined as:

[0073] ω CC' =∑ i∈C,j∈C' A ij

[0074] Among them: C and C' are two different communities; A ijis the edge weight between node i and node j in the original network.

[0075] Step 4, recursively apply, repeat the above process on the new cohesive graph to form a higher level of community structure until the stopping condition is met.

[0076] By recursively applying the Leiden algorithm, a multi-level community structure is gradually constructed. Each step optimizes the community division by maximizing the modularity, ensuring that each node has at least one neighbor belonging to the same community, and generating entity, relationship and community summary information based on this.

[0077] In addition, the calculation process of the Node2Vec algorithm can be divided into several main steps: graph preprocessing, random walk generation, word embedding learning, and node embedding generation.

[0078] The following is the detailed calculation process:

[0079] Step 1: Graph preprocessing: construct a graph G = (V, E) from the input data, where V is a node set and E is an edge set. The initialization parameters include:

[0080] p: Return parameter that controls the probability of the random walk returning to the previous node.

[0081] q: In and out parameters that control whether the random walk tends to explore a local subgraph or a wider graph structure.

[0082] r: length of the random walk.

[0083] n: The number of random walks generated for each node.

[0084] Step 2: Initialize and generate n random walk sequences of length r starting from each node v∈V; then generate random walks, where the random walk process is:

[0085] Starting from node v, set the current node t=v.

[0086] For each step i (from 1 to r):

[0087] Calculate the transition probability P(u|t) from the current node t to each neighbor node u.

[0088] Select the next node u according to the transition probability.

[0089] Update the current node t=u.

[0090] Record the generated random walk sequence.

[0091] Step 3: Calculate the transition probability. The transition probability P(u|t) is given by the bias term α pq (t,u) and edge weight ωtu Decide:

[0092] P(u|t)=α pq (t,u)·ω tu

[0093] where ω tu is the edge weight between node t and node u (usually 1);

[0094] α pq (t,u) is a bias term defined as follows:

[0095]

[0096] Among them, d tu represents the distance between node t and node u; d tu =0 means u is the fallback node of t (i.e. u is the previous node v); d tu =1 means u is a direct neighbor of t; d tu =2 means that u is a second-order neighbor of t (i.e., u is connected through t's neighbor nodes).

[0097] Step 4: word embedding learning. Convert the generated random walk sequence into training data and use the Word2Vec model to train the vector representation of the node.

[0098] Step 5: Node embedding generation: extract the vector representation of each node from the trained word embedding model and normalize the node embedding.

[0099] Through the above steps, the Node2Vec algorithm is used to generate vector representations of the nodes of the industrial pump knowledge graph to capture the local and global structural features of the graph.

[0100] In this embodiment, the step 5 is specifically as follows:

[0101] In order to describe in detail the processing flow when the user selects vibration signal input and problem consultation, it is divided into two main scenarios: fault classification and location based on vibration signals and knowledge retrieval based on problem consultation. Each scenario contains specific steps of feature extraction, matching and information retrieval. The following is a detailed description:

[0102] Scenario 1: Fault classification and location based on vibration signals

[0103] Step 1: The user uploads or inputs vibration signal data and extracts frequency domain feature parameters;

[0104] Step 2: Input the extracted feature parameters into the trained convolutional neural network model for classification prediction;

[0105] Step 3: Use the cosine similarity algorithm to match the feature parameters extracted by CNN with the nodes in the pre-built feature map, and match the classification results with the nodes in the fault map part. Sort the matching results according to the similarity score, select the most likely corresponding fault type node, and collect relevant node information.

[0106] Scenario 2: Knowledge retrieval based on question consultation

[0107] In step 1, the user asks a question or request in natural language, such as “What failures may occur in the impeller of a centrifugal pump?”.

[0108] Step 2: Use Tongyi Qianwen LLM to extract key entities from users’ questions, such as “centrifugal pump”, “impeller”, “fault”, etc.

[0109] Step 3: Also use the cosine similarity algorithm to match the extracted related entities with the nodes in the feature graph to find the part most relevant to the question.

[0110] In this embodiment, the step 6 is specifically as follows:

[0111] The GraphRAG query mechanism mainly includes two search methods: local search and global search.

[0112] The local search process is a complex and multi-stage process that aims to efficiently and accurately retrieve relevant information from the knowledge graph and provide users with valuable responses. The system can also provide links or references to original materials or more detailed documents to help users explore deeper. The specific process diagram of local search is as follows Figure 4 As shown, the following are the detailed steps of the process:

[0113] Step 1: Map the obtained relevant node information to the index data and match it to the descriptions of these entities, communities, summaries, and covariates;

[0114] Step 2: Based on the matched entities and their descriptions, the system searches for relevant nodes and edges in the knowledge graph and generates a series of candidate results.

[0115] Step 3: To ensure the relevance and quality of the final output, the system will sort and filter all candidate units:

[0116] Step 4: After sorting and filtering, the system will determine a set of results with higher priority. These results are usually arranged in a certain order so that users can quickly find the most relevant information;

[0117] In step 5, the system aggregates all priority results into a comprehensive answer and presents the generated response to the user, ensuring that the user can easily understand and utilize the information provided.

[0118] The global search process is a distributed processing method that aims to improve query efficiency and result quality by processing multiple community report batches in parallel. Figure 5 As shown, the following is a detailed description of the process:

[0119] Step 1: Map the obtained relevant node information to the index data and match the descriptions of the communities, summaries, and covariates related to these entities;

[0120] Step 2: The system selects several most relevant communities for processing based on the input information; it breaks down the input information into multiple subtasks and assigns them to the selected community report batches. Each batch is responsible for processing a part of the problem content to ensure load balancing and efficient concurrent processing.

[0121] Step 3: Each community report batch performs a local search process, including entity recognition, feature extraction, cosine similarity calculation, etc., to generate an intermediate response for the batch. Then the intermediate responses generated by each batch are scored, and the scoring criteria include relevance, confidence source, reliability, diversity, etc.

[0122] Step 4: Based on the scoring results of all batches, sort and filter the intermediate responses according to the comprehensive scores, retaining the most valuable parts to ensure the quality of the final output.

[0123] Step 5: Aggregate the intermediate responses with higher scores to form a comprehensive and coherent final answer; and further polish and optimize the final answer through natural language generation technology to make it more in line with the user's reading habits and comprehension ability.

[0124] The present invention is not limited to the above specific embodiments. Those skilled in the art of the present invention may make various modifications, supplements or substitute the described specific embodiments in similar ways without departing from the spirit of the present invention or exceeding the scope defined by the attached claims, which are all within the scope of protection required by the present invention.

Claims

1. An intelligent diagnosis method for industrial pump faults based on the GraphRAG large model, characterized by: Step 1: Collect data and build a knowledge graph. Collect basic information and fault analysis records of industrial pumps, extract entities of industrial pumps and the relationships between them to build a fault graph, where entities include components, fault types, fault phenomena, fault causes and fault handling; collect vibration signals, extract feature parameters to build a feature graph. Step 2: Integrate knowledge graph. The feature parameters are input into the convolutional neural network for fault classification to obtain the fault type. Then the Word2Vec model is used to generate word vectors, and the cosine similarity is calculated to obtain similar node pairs between the fault graph and the feature graph, and then the fused knowledge graph is obtained. Step 3: Optimize the prompt template. Input the data of the fused knowledge graph into the LLM large model module, and use the prompt_tune function in GraphRAG to automatically adapt and optimize the prompt template. Step 4: Extract node information using the GraphRAG index building module. Use LLM to segment the data text into text blocks; use the optimized prompt template to identify and extract instances of graph nodes and edges in text blocks; use LLM to provide a brief summary description for each entity and relationship, so that each element has a unified and descriptive text; divide the entities into communities using the hierarchical Leiden algorithm, and generate entity, relationship and community summary information based on this; use the Node2Vec algorithm to generate a vector representation of the graph, and finally summarize all the node information, including text blocks, entities, communities, summaries and covariates. Step 5: Input and analysis of user requests. When the user selects the vibration signal input, the characteristic parameters of the signal are extracted, and the convolutional neural network is used to classify the faults. The characteristic parameters are matched with the nodes in the characteristic map part through cosine similarity, and the classification is performed as a data engine. The classification results are matched with the nodes in the fault map part, and the positioning is performed as a knowledge engine to obtain relevant node information; In addition, when the user chooses to consult a question, LLM is used to extract the relevant entities in the question, and the nodes in the feature graph part are matched through cosine similarity to obtain the relevant node information. Step 6: Use the GraphRAG retrieval generation module to search the index and output the diagnostic report. Map the relevant node information to the index data, and then use the local search and global search in the retrieval generation module to integrate them to obtain the final fault diagnosis report and output it to the user.

2. According to the basic information, fault analysis record and characteristic parameters of the industrial pump described in claim 1, it is characterized in that: By collecting data, we get the basic information files of industrial pumps, which mainly describe the main components and fault types of industrial pumps; the relevant data of industrial pump fault maintenance are some common faults and treatment methods of clean water pumps and slurry pumps, which mainly involve equipment phenomena, fault causes and treatment methods. The characteristic parameters mainly include four frequency domain characteristic parameters: center of gravity frequency, average frequency, root mean square frequency and frequency standard deviation.

3. The method of generating word vectors using the Word2Vec model as described in claim 1, wherein: The main parameters that need to be set for the Word2Vec model include vector dimension, window size, minimum word frequency, number of threads, initial learning rate, training model, and optimization objective function.

4. According to claim 1, the data of the fused knowledge graph is input into the LLM module, and the prompt template is automatically adapted and optimized using the prompt_tune function, characterized in that: The LLM model selects the Tongyi Qianwen model, the chat model selects the qwen-turbo model, and the embedding model selects the text-embedding-v1 model. For the prompt_tune function, the parameters that need to be set mainly include the method for selecting documents, the limit for loading text units, the input language, the maximum number of tokens for generating prompts, and the chunk-size used when generating text units.

5. The method of dividing the entities into communities by using the hierarchical Leiden algorithm according to claim 1, characterized in that: The process of the hierarchical Leiden algorithm includes: using the standard Leiden algorithm to perform community detection on the initial network, and finding the best community division by optimizing the modularity; condensing each detected community into a super node and constructing a new condensed graph; recursively applying the community detection and condensation process to the new graph to form a higher-level community structure; this process continues until the preset stopping condition is reached; and finally outputting a multi-level community structure to reveal the multi-scale community characteristics in the network.

6. The method of claim 1, wherein: The process of the Node2Vec algorithm includes: in a given graph, setting two hyperparameters, the return parameter and the input and output parameter, to control the behavior of the random walk, and generating multiple biased random walk sequences starting from each node; these sequences are used as training data and input into the Word2Vec model to learn the low-dimensional dense vector representation of each node; after multiple rounds of iterative optimization, the model converges and outputs the node embedding vector that can reflect the structural characteristics of the graph.

7. The method of claim 1 wherein the local search and the global search in the retrieval generation module are integrated to obtain a final fault diagnosis report, characterized in that: The basic process of local search includes: after the user inputs a request, the system extracts the entity and its description and embedded information, generates candidate text units, candidate community reports, candidate entities, etc.; these candidate units are sorted and filtered, and priority results are output; finally, all priority results are aggregated, a response is generated and returned to the user. In addition, the basic process of global search includes: after the user enters the request, the system processes it by assigning the question to different community report batches; each community report batch generates intermediate responses and scores them, and then sorts and filters these scoring results to retain the best results; finally, the system aggregates the intermediate responses with higher scores, generates the final response and returns it to the user.

Images