Improved rag-based substation equipment intelligent operation and maintenance method, device, equipment and medium

Abstract

The application discloses a power transformation equipment intelligent operation and maintenance method and device based on improved RAG, equipment and medium, comprising: constructing different vector similar graphs, and acquiring vertical question and answer models; optimizing a breadth-first search algorithm; generating an answer result for a business problem based on a power transformation operation and maintenance business problem, the optimized breadth-first search algorithm, a vector similar graph corresponding to a knowledge database, and a vertical question and answer model, acquiring an evaluation result for judging whether the answer result meets a standard; generating a third evaluation result for judging whether the business problem contains a fault and fault information; generating a third answer result including a fault cause, and simultaneously generating a fourth answer result for indicating a standard involved in the fault; and generating an optimal answer scheme for the business problem based on all answer results and evaluation results, the business problem, and the vertical question and answer model. The application can more comprehensively, accurately and reliably answer business problems in operation and maintenance, and improves work efficiency.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of substation equipment operation and maintenance, and particularly relates to a substation equipment intelligent operation and maintenance method, device, equipment and medium based on an improved RAG. BACKGROUND

[0002] In substation operation and maintenance work, there are many types of substation equipment, many inspection and maintenance projects, many variable fault information, complex operation procedures, dangerous work scenes, and high requirements for the professional ability of operation and maintenance personnel due to subjective consciousness of artificial, which results in that the current operation and maintenance work greatly depends on expert experience, and operation and maintenance personnel spend a long time searching for information. In order to reduce the work pressure of operation and maintenance personnel and improve the operation and maintenance efficiency, most of the existing researches are to design a question and answer system, operation and maintenance personnel provide questions or keywords through the question and answer system, the question and answer system finds matching answers based on retrieval technology to a professional database, and the found answers are fed back to the operation and maintenance personnel without missing a word to assist the operation and maintenance personnel to make further decisions. The matching answers given in this form are relatively single, cannot be associated with more comprehensive technical problems to assist decision-making, lack intelligent assistance suggestions or more reasonable solutions, and are too dependent on the input of operation and maintenance personnel, and more likely to deviate, so that the question and answer system provides invalid answers.

[0003] The emergence of large language models greatly promotes the development and application of intelligent question and answer systems. Compared with the intelligent question and answer based on database retrieval matching, the large language model technology makes the question and answer generation more consistent with the language mode of human beings, can utilize the powerful reasoning and generation ability of the model, give more human-like answers or suggestions according to the problem, generate more rich solutions, and further more intelligently assist operation and maintenance personnel to make decisions. However, due to the wide range, large quantity and complex logical relationship of the knowledge in the substation operation and maintenance field, the operation and maintenance question and answer system of the general large language model in the substation operation and maintenance work greatly depends on the experience of experts, and operation and maintenance personnel spend a long time searching for information, which cannot meet the specific application scene and demand, so that the generated solutions deviate from the actual demand, for example, in response to fault analysis and processing and other problems, it is difficult to provide accurate and reliable solutions. At the same time, due to the inevitable illusion phenomenon of the large language model, in the substation operation and maintenance business scene, the illusion phenomenon may bring huge disastrous influence to personnel safety, equipment safety and power supply reliability. SUMMARY

[0004] To solve the above technical problems, the technical scheme adopted by the present application is as follows:

[0005] An intelligent operation and maintenance method for power transformation equipment based on improved RAG, comprising the following steps:

[0006] S1, based on a pre-set database for storing different power transformation operation and maintenance knowledge, a vector similarity graph is constructed, and a vertical question and answer model is obtained based on a technical question and answer database and a large language model;

[0007] The database of different power transformation operation and maintenance knowledge includes a knowledge database, a fault database and a standard database, and the technical question and answer database belongs to the knowledge database;

[0008] S2, based on the spectral clustering method, the breadth-first search algorithm is improved to obtain a spectral clustering BFS search algorithm for similar text retrieval of the vector similarity graph in step S1 based on the retrieval vector;

[0009] S3, collect business problems, and use the vertical question and answer model obtained in step S1 to determine whether the business problem is a power transformation operation and maintenance business, if yes, execute step S4, otherwise, re-collect business problems;

[0010] S4, using the business problem obtained in step S3 as a retrieval vector, the algorithm in step S2 is used to traverse the vector similarity graph corresponding to the knowledge database, and based on the traversal result, the business problem and the vertical question and answer model, a response result for the business problem is generated, the response result is used as a retrieval vector to retrieve the vector similarity graph corresponding to the standard database, and based on the retrieval result, the response result and the vertical question and answer model, an evaluation result for judging whether the response result meets the standard in the standard database is obtained;

[0011] S5, based on the business problem, the response result obtained in step S4 and the vertical question and answer model, a third evaluation result for judging whether the business problem contains a fault and fault information is generated, and according to the third evaluation result, it is judged whether the equipment has a fault, if not, return to step S3, otherwise, execute step S6;

[0012] S6, searching the vector similarity graph corresponding to the fault class database by using the algorithm of step S2 with the third evaluation result as a search vector, generating a third answer result including the fault cause based on the search result, the third evaluation result and the vertical question and answer model, and generating a fourth answer result for indicating the standard involved in the fault based on the search result, the business question, the third evaluation result and the vertical question and answer model;

[0013] S7, generating an optimal solution for giving relevant knowledge based on the knowledge class database, determining whether the equipment is faulty and the cause of the fault, and indicating the standard involved in the fault for the business question based on all the answer results and the evaluation results obtained in steps S4, S5 and S6, the business question and the vertical question and answer model.

[0014] In step S1, the construction method of the vector similarity graph comprises the following steps:

[0015] ①, pre-processing the data in each database in step S1 to obtain a corresponding text database;

[0016] ②, segmenting and vectorizing the text database obtained in step ① to obtain corresponding vector feature data;

[0017] ③, calculating the similarity between any two vector feature data of each text database in step ② by using a similarity calculation method, constructing a corresponding vector similarity graph based on the vector feature data as nodes and a preset first similarity threshold between the nodes, and constructing an edge weight based on the similarity between the two nodes corresponding to the edge.

[0018] The step S2 comprises the following steps:

[0019] i, calculating a normalized Laplacian matrix according to the degree matrix and the similarity matrix of the vector similarity graph;

[0020] ii, performing eigenvalue decomposition on the normalized Laplacian matrix to obtain a first eigenvalue and a first eigenvector, and using a K-means algorithm to cluster and divide the first eigenvector into a plurality of clusters;

[0021] iii, calculating the similarity between the search vector and the cluster center of each cluster and each node in the cluster by using a similarity calculation method, and generating similar texts by searching the vector similarity graph based on a second similarity threshold and using a breadth-first search algorithm.

[0022] The knowledge class database comprises a technical question and answer database and a theoretical knowledge database.

[0023] An intelligent operation and maintenance system for power transformation equipment based on an improved RAG, comprising:

[0024] User interaction module: used to collect business issues;

[0025] Professional knowledge base construction module: used to collect and classify professional knowledge data in the field of substation operation and maintenance, and construct different vector similarity graphs based on the classified professional knowledge data;

[0026] Vertical large model construction module: used to build a vertical question-answering model based on the classified technical question-answering database obtained from the large language model and professional knowledge base construction module;

[0027] Search algorithm building module: used to generate a spectral clustering BFS search algorithm based on the spectral clustering method to improve the breadth-first search algorithm and obtain a spectral clustering BFS search algorithm for similar text retrieval based on the retrieval vector and the vector similarity graph obtained by the professional knowledge base building module;

[0028] Knowledge Reasoning Module: This module is used to use the business questions collected by the user interaction module as retrieval vectors to traverse the vector similarity graph corresponding to the knowledge database using the algorithm of the search algorithm building module. Based on the traversal results, business questions, and vertical question-answering models, it generates answers to the business questions. It uses the answers as retrieval vectors to search the vector similarity graph corresponding to the standard database. Based on the retrieval results, the answers, and the vertical question-answering models, it obtains an evaluation result for determining whether the answers meet the standards in the standard database.

[0029] Fault reasoning module: used to generate a third evaluation result based on the business question, all answer results obtained by the knowledge reasoning module, and the vertical question-answering model to determine whether the business question contains a fault and the corresponding fault information. Based on the third evaluation result, it is determined whether the device has a fault. If not, the process returns to step S3; otherwise, it executes step S6.

[0030] Standard output module: used to use the third evaluation result output by the fault reasoning module as a retrieval vector to search the vector similarity graph corresponding to the fault database using the algorithm of the search algorithm building module, and generate a third answer result including the cause of the fault based on the search result, the third evaluation result, and the vertical question-answering model. At the same time, based on the search result, the business questions collected by the user interaction module, the third evaluation result, and the vertical question-answering model, generate a fourth answer result indicating the standard involved in the fault;

[0031] Optimal answer generation module: used to generate the optimal answer solution for the business problem based on the knowledge database to provide corresponding knowledge, determine whether the equipment is faulty and the cause of the fault, and indicate the standards involved in the fault based on all the answer results and evaluation results obtained by the knowledge reasoning module, fault reasoning module and standard output module, the business problems collected by the user interaction module and the vertical question and answer model.

[0032] An electronic device comprises a processor and a memory, the memory stores a computer program, and the computer program is executed by the processor to realize the improved RAG-based intelligent operation and maintenance method of power transformation equipment.

[0033] A computer readable storage medium stores a computer program, and the computer program is executed by the processor to realize the improved RAG-based intelligent operation and maintenance method of power transformation equipment.

[0034] The beneficial effects of the present application are:

[0035] 1. A large amount of professional knowledge is used to fine-tune and retrieval augmented generation (RAG) of general large models, and a vertical domain large language model is constructed, which can alleviate the illusion phenomenon in the inference process of the large model, make the business scheme solution more comprehensive and accurate, and improve the work efficiency in the business field;

[0036] 2. The self-evaluation prompt optimization algorithm is adopted to provide more comprehensive, accurate and reliable answers to business problems in power transformation and operation, and to give more intelligent solutions to improve the work efficiency in the business field. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0038] Figure 1 The module structure diagram of the present application.

[0039] Figure 2 The flowchart of the present application. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0041] The operation and maintenance system in the prior art using a large language model improves intelligence, but lacks professional expert experience and is difficult to provide accurate and comprehensive answers. Moreover, in highly professional problems, such as "what should be done when a 220kV and above voltage level transformer does not trip under near-zone short-circuit impact?", the general knowledge of the large language model is superficial and cannot provide accurate values, and even hallucination occurs, giving false answers and misleading operation and maintenance personnel, causing unavoidable losses. Therefore, the current technical solution cannot provide comprehensive and effective intelligent support in professional substation operation and maintenance work.

[0042] The present application improves and upgrades the substation equipment operation and maintenance system. On the basis of traditional or large language model operation and maintenance system, a professional field knowledge base is added to construct a vertical question and answer model. The system answers through images, sounds and texts to assist decision-making. At the same time, through the powerful professional reasoning ability of the vertical question and answer model, various professional problems that occur in the operation and maintenance process, such as "what should be done when the transformer C-phase appears a light gas alarm signal, the heavy gas protection trips, and the electrical quantity protection does not act?", a more intelligent solution is provided, significantly improving work efficiency.

[0043] The embodiment provides a substation equipment intelligent operation and maintenance method based on an improved RAG, including the following steps:

[0044] S1, collecting professional knowledge data in the substation operation and maintenance field and classifying, obtaining a vertical question and answer model based on the classified professional knowledge data and the question and answer model, and constructing different types of vector similarity graphs based on the classified professional knowledge data;

[0045] The present application first classifies the professional knowledge data in the substation operation and maintenance field, including a knowledge database, a fault database and a standard database. In specific operation, each type of database can be further divided according to actual conditions. In the embodiment, the knowledge database can be further divided into a technical question and answer database, a theoretical knowledge database and a practical operation skill database. The fault database refers to a fault case database, and the standard database refers to a standard specification and strong rule database. The vertical question and answer model is obtained based on the data in the technical question and answer database and the question and answer model.

[0046] The technical question and answer database contains a plurality of question and answer pairs, each question and answer pair is a piece of data, and each includes a question and a corresponding answer in the process of substation operation and maintenance, such as: what are the basic contents of substation switching operation? The answer is: (1) line stop and send operation; (2) transformer stop and send operation; (3) busbar switching and busbar stop and send operation; (4) installation and removal of grounding line operation (closing and pulling open grounding switch); (5) parallel and decoupling operation of power grid; (6) transformer voltage regulation operation; (7) station power supply switching operation; (8) relay protection and automatic device switching operation, and operation of changing relay protection and automatic device setting value; (9) other special operation.

[0047] The theoretical knowledge database contains substation operation and maintenance theoretical knowledge such as AC and DC power supply system, anti-misoperation and management, online monitoring system of dissolved gas in oil, smart substation construction and management, etc., such as: general technical principles of anti-misoperation logic, specific requirements of circuit breaker, disconnecting switch, circuit breaker trolley, grounding switch, and net (cabinet) door, wherein (1) when the circuit breaker trolley is switched between "work", "test" and "maintenance" positions, the circuit breaker in the interval should be in the split position; (2) when the circuit breaker trolley is switched to the "work" position, the grounding switches on both sides should be in the split position, the grounding line should be in the removal state, including grounding switches and grounding lines connected through main transformers, grounding transformers, station transformers, capacitors, busbars and cables, and the rear cabinet door should be closed.

[0048] The practical operation skill database includes substation switching operation, oil-immersed transformer sampling and diagnostic analysis, GIS gas chamber diagnostic analysis, GIS ultrasonic partial discharge detection technology, GIS ultrahigh frequency partial discharge detection technology, and accurate temperature measurement of equipment, etc. Substation operation and maintenance practical operation skills, such as: during the operation ticket filling stage, the following items should be filled in the operation ticket: (1) devices that should be pulled or closed [circuit breakers (switches), disconnecting switches (switches), grounding switches (devices), etc.], voltage verification, installation and removal of grounding lines, closing (installation) or opening (removal) of air circuit breakers (switches), fuses, switching protection circuits and automatic devices, and checking whether there is no voltage, etc. (2) After pulling or closing the devices [circuit breakers (switches), disconnecting switches (switches), grounding switches (devices), etc.], check the position of the devices. (3) When performing stop and send operation, check that the circuit breaker (switch) is in the split position before pulling or closing the disconnecting switch (switch) or pulling out or pushing in the handcart type circuit breaker (switch). (4) Before and after performing load reversal or decoupling and parallel operation, check the operation of the related power supply and the load distribution. (5) After equipment maintenance, check that the grounding switches (devices) in the power supply range have been pulled open and the grounding lines have been removed before closing the circuit.

[0049] The fault case database includes typical cases of substation operation and maintenance, such as electrical misoperation accidents, substation total stop accidents, large transformer accidents, reactive compensation accidents, mutual inductor accidents, GIS and switchgear accidents, and relay protection accidents. For example, the accident of "wrong interval switching operation, misoperation of live grounding switch" involves clauses (1) The system mode, device name, number and location should be checked before operation. (Q / GDW1799.1—2013 "National Grid Company Power Safety Work Procedures Substation Part" 5.3.6.2) (2) The operator should clearly understand the operation purpose and sequence, analyze the potential risks in the operation process and take appropriate measures. Sing, repeat, monitor, operate according to the operation ticket, confirm that the device state is consistent with the operation ticket content and mark it. (National Grid Safety Supervision

[2018] No. 1119 "Safety Management Regulations for Preventing Electrical Misoperation" 3.1.3) (3) The switching operation should have the formal instruction of the on-duty control personnel and the operation ticket should be used. (Regulations of State Grid Corporation of China on Substation Operation and Maintenance) Article 66). The response measures for the accident are: (1) Strictly implement the switching operation management regulations, use the formal control instruction and the operation ticket, and operate according to the sequence by the monitor and the operator. Non-operation personnel and on-site inspection personnel shall not perform the operation. (2) Fully implement the "five prevention" management regulations and strictly implement the lock key management system. (3) Fully check the computer key debugging function unlock password setting and strictly prohibit the use of unlock password. Strengthen software program management, require the supplier to cancel the debugging mode and shield the unlock function, and trace the responsibility and examine the violation.

[0050] The standard specification and strong rule database includes the regulations of State Grid Corporation of China on substation operation and maintenance and the operation rules of twenty-eight types of equipment (operation rules of oil-immersed transformers, operation rules of circuit breakers, operation rules of combined electrical apparatus, etc.), eighteen major power grid anti-accident measures of State Grid Corporation, and the substation part of the power safety work procedures of State Grid Corporation, etc. National or industry standards and specifications, such as: "1.5.1 In operation, if any of the following conditions is found, the operation and maintenance personnel should immediately report to the control personnel and apply for device shutdown. a) The sleeve has serious damage and discharge phenomenon. b) The conductive loop component has serious overheating or sparking phenomenon. c) SF6 circuit breaker has serious gas leakage and sends operation blocking signal. d) The oil circuit breaker has smoke or abnormal sound in the arc chamber. e) The oil circuit breaker has serious oil leakage and the oil level is not visible. f) The multi-oil circuit breaker has explosion sound inside. g) The vacuum circuit breaker has cracks or discharge sound in the arc chamber. h) The fall tank type circuit breaker has deformation or damage of the explosion-proof membrane. i) The hydraulic or pneumatic operating mechanism loses pressure and the energy storage mechanism has damaged energy storage spring. j) Other conditions that should be considered for emergency shutdown according to the actual situation."

[0051] The question and answer model is based on a large language model such as LLaMA3, ChatGLM, etc. At present, the question and answer model in the substation operation and maintenance work is mainly based on traditional database retrieval, and a few use general large language model enhanced system intelligence. However, the general large language model lacks a large amount of professional knowledge in the power field in the training process, so that the generated solution deviates from the actual demand, and serious hallucination phenomenon occurs, making it difficult to provide accurate and reliable solutions.

[0052] In the present application, the data in the technical question and answer database is randomly divided into a training set and a test set during fine-tuning training of the vertical question and answer model. The training set is used to fine-tune the training model, and the test set is used to evaluate the performance of the fine-tuned model. In the later specific application process, new technical question and answer data can also be collected according to user feedback and business problems to further fine-tune and improve the vertical question and answer model.

[0053] The professional knowledge data based on classification is used to construct different types of vector similarity graphs, including the following steps:

[0054] ①, based on the classified professional knowledge data, a corresponding text database is constructed;

[0055] After collecting professional knowledge data, all data can be preprocessed according to business needs, such as data cleaning, word segmentation and tokenization, etc., to generate a text database. The above are all prior art, and this embodiment will not be described in detail.

[0056] ②, using the Word2Vec model to vectorize the data preprocessed in step ① to obtain a plurality of vector feature data;

[0057] The preprocessed data is vectorized using Word2Vec. Specifically, vector feature data can be generated based on a paragraph, a chapter or a sentence in the text database. In this embodiment, a vector feature data is generated based on a paragraph.

[0058] ③, the similarity between any two vector feature data of each text database in step ② is calculated, the vector feature data is used as a node, a first similarity threshold is preset based on the nodes to construct a non-directed edge to obtain a corresponding vector similarity graph, and the edge weight is constructed based on the similarity between the two nodes corresponding to the non-directed edge;

[0059] When constructing a vector similarity graph, you can construct a separate vector similarity graph based on each text database, or you can combine two or more text databases according to actual needs, and construct a vector similarity graph based on the combined text database. This embodiment constructs a first vector similarity graph based on a technical question and answer text database for priority retrieval of business issues; constructs a second vector similarity graph based on a text database after combining theoretical knowledge and practical skills, which is used as a supplementary retrieval of the first vector similarity graph; constructs a third vector similarity graph based on a fault case text database to assist in solving fault cause analysis in operation and maintenance; constructs a fourth vector similarity graph based on a standard specification and strong article text database for answer result evaluation. This application is based on the reasoning framework of the Tree of Thoughts (ToT), and uses LLM to simulate the human decision-making process, so that the answers obtained are more in line with the human way of thinking. Compared with directly searching once to obtain retrieval results, more comprehensive answers can be obtained through multiple searches through the thinking tree.

[0060] The vector similarity graph (VSG) uses G n Indicates that G n ={V n ,E n}, V n represents the set of nodes in the nth vector similarity graph, E n represents the set of edges in the nth vector similarity graph.

[0061] In order to improve the retrieval efficiency, after calculating the similarity between vector feature data using the similarity calculation method, the edges with lower similarity are filtered based on the first similarity threshold γ. Based on this method, all vector similarity graphs are updated and optimized. The set of edges after the nth vector similarity graph is updated is E′ n The corresponding expression is:

[0062] E′ n ={(v n,m ,v n,m′ )∈E n |sim(v n,m ,v n,m′ )≥θ}; (1)

[0063] Where sim(·) represents the similarity function. In this embodiment, cosine similarity is used for measurement, θ=0.7, and m≠m′, v n,m represents the mth node in the nth vector similarity graph, v n,m′ represents the m′th node in the nth vector similarity graph, v n,m ∈V n , v n,m′ ∈V n .

[0064] S2, improving the breadth-first search (BFS) algorithm based on the spectral clustering method to obtain a spectral BFS search algorithm for similar text retrieval based on the retrieval vector on the vector similarity graph obtained in step S1;

[0065] The breadth-first search algorithm (Breadth-First Search, BFS) is used to traverse the vector similarity graph, and the vector segments with high similarity are retrieved as candidate results, i.e., similar texts, based on the second similarity threshold θ, and the search for neighbor nodes is stopped when a value lower than the second similarity threshold θ is encountered during traversal.

[0066] When the BFS algorithm is used to traverse the first vector similarity graph, multiple nodes and edges need to be traversed, and the search efficiency is low. The spectral BFS search algorithm is proposed in the present application, which classifies the first vector similarity graph by using spectral clustering in the BFS algorithm, and then guides the search process, so that the search is more efficient, and a pruning strategy is used to effectively limit the search range and reduce the search space. The specific process of the algorithm is as follows:

[0067] i. Calculate the normalized Laplacian matrix L according to the degree matrix and the similarity matrix of the vector similarity graph;

[0068] The calculation formula of the normalized Laplacian matrix L is:

[0069]

[0070] In the formula, I represents the unit matrix, the elements on the diagonal are all 1, and the other elements are all 0, D represents the degree matrix, and W represents the similarity matrix.

[0071] The calculation formula of each element D ij on the diagonal of the degree matrix D is:

[0072]

[0073] In the formula, N represents the total number of nodes in the first vector similarity graph, and W ij represents the value of the i-th row and the j-th column in the similarity matrix W. The calculation method of the degree matrix and the similarity matrix is a prior art, and will not be described in detail in the present embodiment.

[0074] ii. Perform feature decomposition on the normalized Laplacian matrix L to obtain first eigenvalues and first eigenvectors, and use the K-means algorithm to cluster and divide the first eigenvectors into a plurality of clusters;

[0075] The first eigenvector is clustered using the K-means algorithm into k clusters (k=5 in this embodiment), the cluster information of each cluster is saved, and the average value of all edge weights in the cluster is used as the cluster center.

[0076] The expression of the eigendecomposition of the normalized Laplace matrix L is:

[0077] L=UΛU T ; (4)

[0078] Where U is an orthogonal matrix whose columns are the first eigenvectors of L, and Λ is a diagonal matrix whose diagonal elements are the first eigenvalues ​​of L.

[0079] iii. Calculate the similarity between the retrieval vector and the cluster center and nodes of each cluster, and use the breadth-first search (BFS) algorithm to search the vector similarity graph based on the second similarity threshold θ to generate similar texts;

[0080] During traversal, the similarity between the search vectors (such as the business question and answer) and each cluster center is first calculated. If the similarity is below the second similarity threshold, no further search is performed for that cluster. Otherwise, the cluster is searched starting from the cluster center. During retrieval, when the similarity between a node and the business question falls below the second similarity threshold, the node's subsequent nodes are pruned, removing its second-order neighboring nodes, further reducing retrieval complexity. Here, the second similarity threshold is θ = 0.7. In each cluster, the first k′ most similar vectors are selected as similar texts, where k′ is a positive integer.

[0081] S3, collect business questions from operation and maintenance personnel, and use the vertical question-answering model obtained in step S1 to determine whether the business question is related to substation operation and maintenance business. If so, execute step S4, otherwise collect business questions again;

[0082] The type of the business question can be an image, a sound or a text. If the business question is an image or a sound, it needs to be processed and converted into a text by using a multi-modal model before being input into the vertical question answering model. If the business question is a text, the text can be directly input into the vertical question answering model. When processing an image, the multi-modal model is an image-to-text model, and when processing a sound, the multi-modal model is a speech-to-text model. The image-to-text model includes CLIP-ViL, ViT-BERT, SimVLM, BLIP, Florence, L-Verse, Oscar, VinVL, MURAL, CoCa, etc., and the speech-to-text model includes a convolutional neural network (VGG, GoogLeNet, ResNet), a recurrent neural network (LSTM, GRU), a generative adversarial network (GAN), a Transformer network, an attention mechanism, etc. The above is prior art, and the process of requesting a text sequence will not be described in detail in the present embodiment.

[0083] The operation and maintenance personnel propose a business question to the vertical question answering model through a user interaction interface and obtain a business solution. The solution generated by the vertical question answering model can be a text or converted into an image or sound and returned to the user.

[0084] S4, using the business question obtained in step S3 as a search vector, traversing the vector similarity graph corresponding to the knowledge database by using the algorithm in step S2, generating an answer result for the business question based on the traversal result, the business question and the vertical question answering model, using the answer result as a search vector to search the vector similarity graph corresponding to the standard database, and obtaining an evaluation result for judging whether the answer result meets the standard in the standard database based on the search result, the answer result and the vertical question answering model.

[0085] In the present embodiment, the vector similarity graph corresponding to the vector similarity graph corresponding to the knowledge database is a first vector similarity graph and a second vector similarity graph. Specifically, when performing this step, the step S4 includes:

[0086] a, the first similar text is obtained by traversing and screening the first vector similarity graph based on the service question obtained in step S3 using the method in step S2, the first similar text and the service question are filled into the first preset prompt, the first preset prompt after filling is input into the vertical question and answer model to generate a first answer result, the first evaluation text is obtained by traversing and screening the fourth vector similarity graph based on the first answer result using the method in step S2, the first answer result and the first evaluation text are filled into the second preset prompt, the second preset prompt after filling is input into the vertical question and answer model to obtain a first evaluation result for judging whether the first answer result meets the standard in the standard database, if not, step b is executed, otherwise, step S5 is executed;

[0087] The first preset prompt and the second preset prompt are both prompt words, the first preset prompt is used to prompt the vertical question and answer model to obtain the corresponding answer according to the filling body, and the second preset prompt is used to prompt the vertical question and answer model to evaluate the correctness of whether the answer result meets the standard.

[0088] In step a, the service question obtained in step S3 and the first answer result are sequentially taken as retrieval vectors to retrieve the first vector similarity graph and the fourth vector similarity graph, respectively, to generate two traversal results of the first similar text and the first evaluation text. Specifically, first, the first similar text, i.e., text segment 1-1, is obtained by traversing and screening the first vector similarity graph based on the service question, i.e., question Q, using the method in step S2, then the text segment 1-1 and the question Q are filled into the first preset prompt to obtain prompt 1-1, prompt 1-1 is: [answer the question according to {question Q}, {text segment 1-1}, without adding other sentences.], and then prompt 1-1 is input into the vertical question and answer model to obtain the first answer result, i.e., answer 1-1 in Figure 2

[0089] Then, the first evaluation text, i.e., text segment 1-2, is obtained by traversing and screening the fourth vector similarity graph based on the answer 1-1 using the method in step S2, the answer 1-1 and the text segment 1-2 are filled into the second preset prompt to obtain prompt 2-1, prompt 2-1 is: [for {answer 1-1}, please analyze {text segment 1-2} from the aspects of rationality and normativity. If it meets the rationality and normativity, please return 1; if it does not meet the rationality and normativity, please return 0.]. Finally, prompt 2-1 is input into the vertical question and answer model for evaluation to obtain the first evaluation result, i.e., answer 1-2.

[0090] ​The text segment 1-1 as a solution to a business question in a technical Q&A database is input into a vertical Q&A model together with the business question, which can assist in guiding the generation of an answer result, thereby improving the quality and accuracy of the answer result. Based on the answer 1-1 traversing the fourth vector similarity graph, it can be judged and evaluated whether the answer result meets the national or industry standards according to the relevant data in the standard specification and strong text database, to assist the operation and maintenance personnel in determining the executability of the answer result. When the answer 1-2 is zero, it indicates that the answer 1-1 cannot be used as a reference solution to the business question, and further retrieval of the theoretical knowledge text database is required.

[0091] b, using the method of step S2, traversing the second vector similarity graph based on the business question obtained in step S3 to obtain a second similar text, filling the second similar text and the business question into the first preset prompt, inputting the filled first preset prompt into the vertical Q&A model to generate a second answer result, using the method of step S2 to traverse and screen the fourth vector similarity graph based on the second answer result to obtain a second evaluation text, filling the second answer result and the second evaluation text into the second preset prompt, inputting the filled second preset prompt into the vertical Q&A model to obtain a second evaluation result for judging whether the second answer result meets the standards in the standard database, if not, returning to step S2, otherwise executing step S5;

[0092] In step b, the business question and the second answer result are sequentially used as retrieval vectors to retrieve the second vector similarity graph and the fourth vector similarity graph, respectively, to generate two retrieval results, namely the second similar text and the second evaluation text. Specifically, using the method of step S2, the second vector similarity graph is traversed and screened based on the business question to obtain the second similar text, i.e., the text segment 2-1. The text segment 2-1 and the question Q are filled into the first preset prompt to obtain prompt 1-2, which is: [answer the question according to {question Q}, {text segment 2-1}, without adding other statements.]. Then, prompt 1-2 is input into the vertical Q&A model to obtain the second answer result, i.e., answer 2-1.

[0093] Using the method of step S2, the fourth vector similarity graph is traversed and screened based on the answer 2-1 to obtain the second evaluation text, i.e., the text segment 2-2. The answer 2-1 and the text segment 2-2 are filled into the second preset prompt to obtain prompt 2-2, which is: [for {answer 2-1}, please analyze from the aspects of rationality and normativity according to {text segment 2-2}. If it meets the rationality and normativity, return 1; if it does not meet the rationality and normativity, return 0.]. Finally, prompt 2-2 is input into the vertical Q&A model for evaluation to obtain the second evaluation result, i.e., answer 2-2.

[0094] Text segment 2-1 is a text database combined with theoretical knowledge and practical skills. The knowledge retrieval of the business problem is input into the vertical question and answer model together with the business problem, which can assist in guiding the generation of the answer result, thereby improving the quality and accuracy of the answer result. Based on the traversal of the fourth vector similarity graph of the answer 2-1, the related data in the standard specification and strong text database can be used to judge whether the second answer result meets the national or industry standard, helping the operation and maintenance personnel to determine the executability of the answer result. When the answer 2-2 is zero, it means that neither the technical question and answer text database nor the text database combined with theoretical knowledge and practical skills can find the solution and knowledge related to the business problem, and the operation and maintenance personnel need to consider solving the problem from other directions.

[0095] S5, filling the business problem, the first answer result or the second answer result to the third preset prompt, inputting the filled third preset prompt into the vertical question and answer model to obtain a third evaluation result for judging whether the business problem contains a fault and corresponding fault information, judging whether the equipment has a fault according to the third evaluation result, if not, returning to step S3, otherwise outputting the third evaluation result and executing step S6;

[0096] The business problem, the first answer result or the second answer result is filled into the third preset prompt to obtain prompt 3-1, prompt 3-1 is: [Please analyze whether {question Q}, {answer 1-1} or {answer 2-1} contains fault or defect phenomenon, accident information. If it contains, please return 1 and output the fault or defect phenomenon, accident information, without adding other statements. If it does not contain, please return 0.]. Input prompt 3-1 into the vertical question and answer model to obtain the third evaluation result, i.e. answer 3-1.

[0097] The vertical question and answer model generates answer 3-1 according to the business problem, answer 1-1 / answer 2-1, which can assist the operation and maintenance personnel to determine the fault or defect state of the current described equipment. When the answer 3-1 is zero, it means that the business problem does not contain fault information.

[0098] S6, using the method of step S2 to traverse the third vector similarity graph based on the fault information in the third evaluation result obtained in step S5 to obtain a third similar text, filling the business problem, the third evaluation result, the third similar text into the fourth preset prompt, inputting the filled fourth preset prompt into the vertical question and answer model to obtain a third answer result, filling the business problem, the third evaluation result, the third similar text into the fifth preset prompt, inputting the filled fifth preset prompt into the vertical question and answer model to obtain a fourth answer result;

[0099] In step S6, the fault information is used as a search vector to search and traverse the third vector similarity graph, generating a third similar text as the search result. Specifically, the method of step S2 is used to traverse the third vector similarity graph based on the fault information obtained in step S5 to obtain the third similar text, namely text segment 4-1. Question Q, answer 3-1, and text segment 4-1 are filled into the fourth preset prompt to obtain prompt 4-1. Prompt 4-1 is: [For {question Q}, {answer 3-1}, please analyze the possible causes of the fault or defect in question Q based on {text segment 4-1} and provide treatment measures. Do not add other statements. ] Prompt 4-1 is input into the vertical question-answering model to obtain the third answer result, namely answer 4-1.

[0100] Fill the fifth preset prompt with question Q, answer 3-1, and text snippet 4-1 to obtain prompt 5-1. Prompt 5-1 reads: [For {question Q}, {answer 3-1}, please analyze the clauses involved in the occurrence of the fault or defect in question Q based on {text snippet 4-1} and provide countermeasures. Do not add other statements.] Enter prompt 5-1 into the vertical question-answering model to obtain the fourth answer, answer 5-1.

[0101] Traversing the third vector similarity graph based on answer 3-1 can determine whether there have been similar failure cases before. The vertical question-answering model generates answer 4-1 by inputting question Q, answer 3-1, and text snippet 4-1. The large model can then output the cause of the failure and corresponding countermeasures, providing operations personnel with reliable solutions.

[0102] S7, filling the business question, the first answer result or the second answer result, the third answer result, the fourth answer result, the first evaluation result or the second evaluation result into the sixth preset prompt, and inputting the filled sixth preset prompt into the vertical question-answering model to obtain the optimal business solution corresponding to the business question;

[0103] Fill question Q, answer 1-1 or answer 2-1, answer 1-2 or answer 2-2, answer 4-1, and answer 5-1 into the sixth preset prompt to obtain prompt 6-1, which is:

[0104] [Based on {Question Q}, {Answer 1-1} or {Answer 2-1}, {Answer 1-2} or {Answer 2-2}, {Answer 4-1}, {Answer 5-1}, answer business question Q. Please answer strictly according to the information provided and do not add other statements;

[0105] If {Answer 2-2} contains "0", please output "No relevant information found";

[0106] If {answer 1-2} or {answer 2-2} contains "1", {answer 3-1} contains "0", please output according to the content of {question Q}, {answer 1-1} or {answer 2-1} in the format of question, answer;

[0107] If {answer 1-2} or {answer 2-2} contains "1", {answer 3-1} contains "1", please output according to the content of {question Q}, {answer 1-1} or {answer 2-1}, {answer 1-2} or {answer 2-2}, {answer 4-1}, {answer 5-1} in the format of question, answer, phenomenon, possible cause, treatment measure, involved clause, coping measure.

[0108] In addition, the existing large language model can analyze the possible causes of the failure or defect phenomenon in the question, give treatment measures, and the clauses involved in the occurrence of the failure or defect phenomenon, and give coping measures. However, the answer given is not comprehensive and accurate enough, and multiple rounds of question and answer are needed to obtain a solution. The vertical question and answer model of the present application can give a professional solution at one time through self-evaluation prompt optimization algorithm. According to the output of the optimal business solution, the operation and maintenance personnel can obtain the corresponding knowledge based on the knowledge database, whether the equipment is faulty and the cause of the fault, the standards involved in the fault, and other related content, such as the following business problem and its optimal business solution:

[0109] Business problem: What is the reason for the main transformer heavy gas, pressure release and differential protection action trip?

[0110] Optimal business solution (including phenomenon, possible cause, treatment measure, involved clause and / or coping measure):

[0111] Phenomenon: Main transformer heavy gas, pressure release and differential protection action trip.

[0112] Possible cause: It is caused by foreign matter in the high-voltage bushing of the transformer, which causes a short-circuit fault in the near zone.

[0113] Treatment measures: a) Check the transformer bushing burst. b) After confirming that the breakers on both sides of the transformer have tripped, check that the oil cooler fan and submersible pump of the forced oil cooling transformer have stopped running. c) Carefully check and verify the transformer differential protection action information, and check the operation of other equipment protection action signals, primary and secondary circuits, DC power supply system and station power system. d) If the station power system is completely powered off, it should be restored as soon as possible. e) According to the dispatching instructions or the provisions of the "Special Procedures for Substation On-site Operation", adjust the transformer neutral point operation mode. f) Check if the running transformer is overloaded, and according to the load condition, put in the cooler. If the transformer is running over load, report to the on-duty control personnel to transfer the load. g) Check the action of the standby automatic switching device. If the standby automatic switching device acts correctly, exit the device according to the dispatching instructions. If the standby automatic switching device does not act correctly, check if the standby automatic switching device acts on the circuit breaker, and according to the dispatching instructions, exit the standby power supply automatic switching device, and immediately close the circuit breaker after the standby automatic switching device acts. Restore the load of the power-off bus. h) Quickly isolate the fault transformer. i) Record the transformer protection action time and primary and secondary equipment inspection results and report. j) Arrange safety measures for fault transformer maintenance test in advance.

[0114] Related clauses: "State Grid Corporation of China Eighteen Major Power Grid Anti-accident Measures (2018 Revised Edition) and Compilation Instructions" 5.1.3.4 "Periodically clean up floating objects, plastic greenhouses, color steel buildings, kites and tall trees in and around the substation. Special inspection should be carried out before and after the strong wind to prevent foreign objects from floating and causing equipment short circuit."

[0115] Response measures: State Grid Inspection

[2015] No. 376 "State Grid Corporation of China Notice on the issuance of sixteen measures to prevent substation full stop (trial) "12.3 "Periodically clean up flammable and explosive materials in and around the substation to prevent accidents. Periodically clean up floating objects, plastic greenhouses, color steel buildings, kites, etc. in and around the substation. Special inspection should be carried out before the strong wind to prevent foreign objects from floating and causing equipment short circuit. Burn, burn straw, blasting operations, dust emission and other behaviors near the substation should be stopped in time to prevent dust pollution from causing short circuit."

[0116] The embodiment of the application also provides a substation equipment intelligent operation and maintenance device based on improved RAG, comprising:

[0117] User interaction module: used for inputting business problems and processing the business problems to generate data recognizable by a large model;

[0118] Professional knowledge base construction module: used for collecting professional knowledge data in the substation operation and maintenance field and classifying, and constructing different vector similarity graphs based on the classified professional knowledge data;

[0119] Vertical large model construction module: used to build a vertical question-answering model based on the classified technical question-answering database obtained from the large language model and professional knowledge base construction module;

[0120] Search algorithm building module: used to generate a spectral clustering BFS search algorithm based on the spectral clustering method to improve the breadth-first search algorithm and obtain a spectral clustering BFS search algorithm for similar text retrieval based on the retrieval vector and the vector similarity graph obtained by the professional knowledge base building module;

[0121] Knowledge Reasoning Module: This module is used to use the business questions collected by the user interaction module as retrieval vectors to traverse the vector similarity graph corresponding to the knowledge database using the algorithm of the search algorithm building module. Based on the traversal results, business questions, and vertical question-answering models, it generates answers to the business questions. It uses the answers as retrieval vectors to search the vector similarity graph corresponding to the standard database. Based on the retrieval results, the answers, and the vertical question-answering models, it obtains an evaluation result for determining whether the answers meet the standards in the standard database.

[0122] Fault reasoning module: used to generate a third evaluation result based on the business question, all answer results obtained by the knowledge reasoning module, and the vertical question-answering model to determine whether the business question contains a fault and the corresponding fault information. Based on the third evaluation result, it is determined whether the device has a fault. If not, the process returns to step S3; otherwise, it executes step S6.

[0123] Standard output module: used to use the third evaluation result output by the fault reasoning module as a retrieval vector to search the vector similarity graph corresponding to the fault database using the algorithm of the search algorithm building module, and generate a third answer result including the cause of the fault based on the search result, the third evaluation result, and the vertical question-answering model. At the same time, based on the search result, the business questions collected by the user interaction module, the third evaluation result, and the vertical question-answering model, generate a fourth answer result indicating the standard involved in the fault;

[0124] Optimal answer generation module: used to generate the optimal answer solution for the business problem based on the knowledge database to provide corresponding knowledge, determine whether the equipment is faulty and the cause of the fault, and indicate the standards involved in the fault based on all the answer results and evaluation results obtained by the knowledge reasoning module, fault reasoning module and standard output module, the business problems collected by the user interaction module and the vertical question and answer model.

[0125] like Figure 1 As shown in the figure, the search algorithm building module, knowledge reasoning module, fault reasoning module, standard output module, and optimal answer generation module together constitute the reasoning enhancement module, which improves the model output accuracy and work efficiency.

[0126] The embodiment of the present application further provides an electronic device, which comprises a processor and a memory, and the memory stores a computer program, and the computer program is executed by the processor to realize the improved RAG-based intelligent operation and maintenance method for power transformation equipment.

[0127] The embodiment of the present application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to realize the improved RAG-based intelligent operation and maintenance method for power transformation equipment. Specifically, the storage medium can be a general storage medium, such as a mobile disk, a hard disk, etc., and the computer program on the storage medium can be executed to perform the embodiment of the improved RAG-based intelligent operation and maintenance method for power transformation equipment.

[0128] The above merely describes the preferred embodiment of the present application, but should not be used to limit the present application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. An intelligent operation and maintenance method for substation equipment based on improved RAG, characterized in that: The steps include: S1: Collect and classify knowledge data in the substation operation and maintenance field, obtain a vertical question-answering model based on the classified data and a large language model, and construct vector similarity graphs based on the classified data. The classified professional knowledge data includes a knowledge database, a fault database and a standard database; S2, improving the breadth-first search algorithm based on the spectral clustering method to obtain a spectral clustering BFS search algorithm for performing similar text retrieval on the vector similarity graph in step S1 based on the retrieval vector; S3, collect business questions, and use the vertical question-answering model obtained in step S1 to determine whether the business question is a substation operation and maintenance business. If so, execute step S4, otherwise collect business questions again; S4, using the business question obtained in step S3 as a retrieval vector, using the algorithm in step S2 to traverse the vector similarity graph corresponding to the knowledge database to obtain a corresponding traversal result, generating an answer result for the business question based on the obtained traversal result, the business question, and the vertical question-answering model, using the answer result as a retrieval vector to search the vector similarity graph corresponding to the standard database, and obtaining an evaluation result for determining whether the answer result meets the standard in the standard database based on the retrieval result, the answer result, and the vertical question-answering model; S5: Based on the business question, the answer obtained in step S4, and the vertical question-answering model, a third evaluation result is generated to determine whether the business question contains a fault and fault information. Based on the third evaluation result, a determination is made as to whether the device has a fault. If not, the process returns to step S3; otherwise, the process proceeds to step S6. S6: Using the third evaluation result as a retrieval vector, the algorithm of step S2 is used to search the vector similarity graph corresponding to the fault database to obtain corresponding search results. Based on the obtained search results, the third evaluation result, and the vertical question-answering model, a third answer result including the cause of the fault is generated. Simultaneously, based on the search results, the business question, the third evaluation result, and the vertical question-answering model, a fourth answer result is generated to indicate the criteria involved in the fault. S7, generating an optimal solution to the business problem based on the answer results and evaluation results obtained in steps S4, S5, and S6, the business problem, and the vertical question-answering model; The step S2 comprises the following steps: i, calculate the normalized Laplacian matrix based on the degree matrix and similarity matrix of the vector similarity graph; ii. Perform eigendecomposition on the normalized Laplace matrix to obtain the first eigenvalue and the first eigenvector, and use the K-means algorithm to cluster the first eigenvector and divide it into several clusters; iii. Calculate the similarity between the search vector and the cluster center of each cluster and each node in the cluster using a similarity calculation method, and use a breadth-first search algorithm based on a second similarity threshold to search the vector similarity graph and generate similar texts.

2. The intelligent operation and maintenance method for substation equipment based on improved RAG according to claim 1 is characterized in that: In step S1, the method for constructing the vector similarity graph includes the following steps: ①, pre-process the data in each database in step S1 to obtain the corresponding text database; ②, segment the text database obtained in step ① and vectorize it to obtain the corresponding vector feature data; ③. Use the similarity calculation method to calculate the similarity between any two vector feature data of each text database in step ②. Take the vector feature data as nodes and construct undirected edges between the nodes based on the preset first similarity threshold to obtain the corresponding vector similarity graph. At the same time, construct edge weights based on the similarity between the two nodes corresponding to the undirected edges.

3. The intelligent operation and maintenance method for substation equipment based on improved RAG according to claim 1 is characterized in that: The knowledge database includes a technical question and answer database and a theoretical knowledge database.

4. An intelligent operation and maintenance system for substation equipment based on improved RAG, characterized in that: include: Professional knowledge base construction module: used to collect and classify professional knowledge data in the field of substation operation and maintenance, and construct different vector similarity graphs based on the classified professional knowledge data. The classified professional knowledge data includes a knowledge database, a fault database, and a standard database; Vertical large model construction module: used to build a vertical question-answering model based on the classified professional knowledge data obtained from the large language model and professional knowledge base construction module; User interaction module: used to collect business questions and use the vertical question-answering model obtained by the vertical large model construction module to determine whether the business question is related to substation operation and maintenance. If so, the knowledge reasoning module is activated; otherwise, business questions are collected again; Search algorithm building module: used to improve the breadth-first search algorithm based on the spectral clustering method, and obtain the spectral clustering BFS search algorithm for similar text retrieval based on the vector similarity graph obtained by the professional knowledge base building module based on the retrieval vector; The construction of the spectral clustering BFS search algorithm includes the following steps: i. Calculate the normalized Laplace matrix based on the degree matrix and similarity matrix of the vector similarity graph obtained by the professional knowledge base construction module; ii. Perform eigendecomposition on the normalized Laplace matrix to obtain the first eigenvalue and the first eigenvector, and use the K-means algorithm to cluster the first eigenvector and divide it into several clusters; iii. Calculate the similarity between the search vector and the cluster center of each cluster and each node in the cluster using a similarity calculation method, and use a breadth-first search algorithm to search the vector similarity graph based on a second similarity threshold to generate similar texts; Knowledge Reasoning Module: This module is used to use the business questions collected by the user interaction module as retrieval vectors to traverse the vector similarity graph generated by the professional knowledge base construction module using the algorithm of the search algorithm construction module to obtain corresponding traversal results. Based on the obtained traversal results, business questions, and vertical question-answering models, an answer result for the business question is generated. The answer result is used as a retrieval vector to search the vector similarity graph corresponding to the standard database. Based on the retrieval results, the answer result, and the vertical question-answering model, an evaluation result is obtained to determine whether the answer result meets the standards in the standard database. Fault reasoning module: This module is used to generate a third evaluation result based on the business question obtained by the user interaction module, the answer obtained by the knowledge reasoning module, and the vertical question-answering model to determine whether the business question contains a fault and the corresponding fault information. Based on the third evaluation result, the module determines whether the device has a fault. If not, the user interaction module is restarted; otherwise, the standard output module is started. Standard output module: used to use the third evaluation result output by the fault reasoning module as a retrieval vector to search the vector similarity graph corresponding to the fault category database using the algorithm of the search algorithm building module to obtain corresponding search results, generate a third answer result including the cause of the fault based on the obtained search results, the third evaluation result, and the vertical question-answering model, and generate a fourth answer result indicating the standard involved in the fault based on the search results, the business questions collected by the user interaction module, the third evaluation result, and the vertical question-answering model; Optimal answer generation module: used to generate the optimal solution to business problems based on the answer results and evaluation results obtained by the knowledge reasoning module, fault reasoning module and standard output module, the business problems collected by the user interaction module and the vertical question-answering model.

5. An electronic device comprising a processor and a memory, characterized in that: The memory stores a computer program, and when the computer program is executed by the processor, the computer program implements the intelligent operation and maintenance method of substation equipment based on the improved RAG according to any one of claims 1 to 3.

6. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the intelligent operation and maintenance method for substation equipment based on the improved RAG as claimed in any one of claims 1 to 3 is implemented.

Images