Fast Fault Disposal Method for Power Generation Systems Based on Speculative Retrieval Enhancement Generation

Abstract

The present invention belongs to the technical field of power generation systems. The present invention discloses a method for quickly disposing of power generation system faults based on speculative retrieval enhanced generation, which includes the following steps: S1, collecting relevant data of the power generation system and constructing a fault symptom query library based on the relevant data; S2, obtaining the current fault symptom description of the power generation system and performing a search in the fault symptom query library based on the current fault symptom description, thereby obtaining a plurality of data segments related to the current fault symptom description; S3, encoding the plurality of data segments and forming a set, and dividing the set into K clusters through a clustering algorithm, thereby obtaining M fault knowledge enhanced subsets. The present invention can consider different fault problems more comprehensively, avoid the deviation that may be caused by a single perspective, and thus significantly improve the accuracy of fault disposal decisions.

Description

Technical Field

[0001] The present invention relates to the technical field of power generation systems, and more particularly to a method for quickly handling power generation system faults based on speculative retrieval enhancement generation. Background Art

[0002] In modern power generation systems, fault diagnosis and response processing are the key to ensuring system stability and safety. With the increase in the types of renewable energy power generation, the stability of the power system has been challenged. In the face of the increase in the types of equipment and the complexity of fault modes in the power system, machine learning and artificial intelligence technologies have been introduced into the field of fault diagnosis in recent years, and have achieved certain results. However, the existing intelligent decision-making system still has some key problems, such as response delays during data processing, poor adaptability to real-time data, and the accuracy and efficiency of decision-making in complex fault scenarios still have much room for improvement.

[0003] Moreover, in the process of fault handling, it is usually necessary to retrieve and analyze a large amount of sensor data, historical fault records, operating status data and other information. This information is often redundant and the data sources are diverse. Therefore, how to intelligently and in real time process and analyze massive data in complex power generation systems to improve the decision-making efficiency and accuracy of fault handling is a key technical problem that needs to be solved in the present invention.

[0004] In view of this, the present invention proposes a method for quickly handling power generation system faults based on speculative retrieval enhanced generation to solve the above problems. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned purpose, the present invention provides the following technical solution: a method for quickly handling power generation system faults based on speculative retrieval enhanced generation, comprising the following steps:

[0006] S1. Collect relevant data of the power generation system and build a fault symptom query library based on the relevant data;

[0007] S2. Obtain a current fault symptom description of the power generation system, and search in a fault symptom query library based on the current fault symptom description, thereby obtaining multiple data fragments related to the current fault symptom description;

[0008] S3, data encoding multiple data fragments to form a set, and dividing the set into K clusters through a clustering algorithm, thereby obtaining M fault knowledge enhanced subsets;

[0009] S4. Train a lightweight language model as a fault handling draft generator, input M fault knowledge enhancement subsets into the fault handling draft generator, and then obtain M fault handling drafts;

[0010] S5. Use a general large language model as the fault handling decision model, score M fault handling drafts, and take the optimal fault handling draft as the final fault handling decision;

[0011] S6. Take the final fault handling decision as the final response strategy and execute it.

[0012] Further, the step of collecting relevant data of the power generation system and constructing a fault symptom query library based on the relevant data includes:

[0013] Collect technical documents, books, historical fault records, maintenance manuals, and fault analysis reports of the power generation system as relevant data of the power generation system;

[0014] Slice the relevant data into text blocks, encode them into vectors using the BEG model and store them in a vector database, and use the vector database as the fault diagnosis query library.

[0015] Further, the step of obtaining the current fault symptom description of the power generation system and retrieving in the fault symptom query library based on the current fault symptom description to obtain multiple data segments related to the current fault symptom description includes:

[0016] Obtain the current fault symptom description of the power generation system and retrieve the corresponding text blocks in the fault symptom query library;

[0017] Take the corresponding text blocks as multiple data segments related to the current fault symptom description.

[0018] Further, the step of encoding multiple data segments and forming a set, and dividing the set into K clusters through a clustering algorithm to obtain M fault knowledge enhanced subsets includes:

[0019] Set the set of multiple data segments as , map multiple data segments to vectors , , and then the vectors of multiple data segments form a set , where represents encoding the data segment into a semantic vector using the Sentence - BERT algorithm ;

[0020] Based on the K - means algorithm, divide the set V of clusters into k clusters , and the clustering process of the set V satisfies the minimization of the objective function , and its objective function is:

[0021] ,

[0022] In the formula, is the minimization of the objective function, is the j-th cluster, is the cluster center of, is the i-th vector in the set V;

[0023] Randomly select a data segment from each of the k clusters, and then obtain M fault knowledge enhanced subsets. Among them, the preset M fault knowledge enhanced subsets are , then, the th fault knowledge enhanced subset is expressed as:

[0024] ,

[0025] In the formula, represents a data segment randomly selected in the cluster .

[0026] Furthermore, the step of training the lightweight language model as a fault handling draft generator and inputting the M fault knowledge enhanced subsets into the fault handling draft generator to obtain M fault handling drafts includes:

[0027] Train a lightweight language model as a fault handling draft generator. Given a fault event , enhance it through the answer A of the fault knowledge enhanced subset S to generate a reason E, that is, obtain as the training data T, where Q is the given query, ;

[0028] Fine-tune the lightweight pre-trained language model using the loss function. The loss function L is:

[0029] ,

[0030] In the formula, is the probability that the fault handling draft generator generates the answer and the reason under the condition of the given query and the fault knowledge enhanced subset ;

[0031] Parallelly input all the fault knowledge enhanced subsets into the fault handling draft generator. Then, for the fault handling drafts generated by each fault knowledge enhanced subset and their corresponding reasoning processes are:

[0032] ,

[0033] In the formula, is the fault handling draft generated for the j-th fault knowledge enhancement subset, represents the reason for the generated j-th fault handling draft, represents the fault handling draft generator;

[0034] Finally, M fault handling drafts , The expression of is:

[0035] .

[0036] Furthermore, the step of using the general large language model as the fault handling decision model and scoring the M fault handling drafts, and taking the optimal fault handling draft as the final fault handling decision includes:

[0037] Using the general large language model as the fault handling decision model and respectively scoring the M fault handling drafts in three dimensions of consistency, reliability, and historical fault record consistency, and then obtaining the consistency scores, reliability scores, and historical fault record consistency scores of the M fault handling drafts ;

[0038] Respectively determine the minimum passing score thresholds for the three dimensions of consistency, reliability, and historical fault record consistency according to the percentiles of the scores of the M fault handling drafts. The minimum passing score thresholds for the three dimensions are respectively expressed as , and . Represent the scores of the three dimensions as , and respectively. The weights corresponding to the scores of the three dimensions are respectively expressed as ;

[0039] Calculate the optimal fault handling draft through the scores of the three dimensions and the minimum passing score thresholds of the three dimensions. The calculation formula is:

[0040] ,

[0041] In the formula, is the optimal fault handling draft, represents the score of the consistency of the j-th fault handling draft, represents the score of the reliability of the j-th fault handling draft, represents the score of the historical fault record consistency of the j-th fault handling draft, represents the weight corresponding to the score of the consistency, The weight corresponding to the score indicating reliability The weight corresponding to the score indicating the consistency of historical fault records. The optimal fault handling draft is when meeting: and and The fault handling draft with the highest score under the circumstances.

[0042] Furthermore, the scoring expression for the consistency of the j-th fault handling draft is:

[0043] ,

[0044] In the formula, is the scoring function for consistency;

[0045] The scoring expression for the reliability of the j-th fault handling draft is:

[0046] ,

[0047] In the formula, is the scoring function for reliability;

[0048] The scoring expression for the consistency of historical fault records of the j-th fault handling draft is:

[0049] ,

[0050] In the formula, is the scoring function for the consistency of historical fault records, is the similar case of historical fault records.

[0051] Furthermore, the step of taking the final fault handling decision as the final response strategy and executing it includes:

[0052] Automatically execute the final response strategy or hand it over to the operator for execution. During the execution process, monitor the recovery of the current fault symptom description in real time and make corresponding adjustments.

[0053] The technical effects and advantages of the method for quickly handling power generation system faults based on speculative retrieval enhancement generation of the present invention:

[0054] By introducing a multi-perspective data sampling strategy, multiple fault handling drafts can be generated from different fault knowledge enhancement subsets, and each fault handling draft is inferred based on different perspectives of fault scenarios. Compared with traditional methods, this diversified draft generation method can consider different fault problems more comprehensively, avoid the biases that may be caused by a single perspective, and thus significantly improve the accuracy of fault handling decisions. By fine-tuning the fault handling draft generator model, the present invention can quickly and concurrently generate multiple fault handling drafts, and uniformly verify these drafts by using a relatively large fault handling validator model. By using a divide-and-conquer strategy, it avoids the inefficient operations of repeatedly retrieving and verifying a large amount of redundant information in traditional methods. This process not only improves the processing speed of decision-making but also significantly reduces the system latency, and remarkably enhances the timeliness of fault response. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 FIG. is a schematic flowchart of a method for quickly handling power generation system faults based on speculative retrieval enhancement according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0056] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0057] Embodiment 1

[0058] Please refer to Figure 1 As shown, the method for quickly handling power generation system faults based on speculative retrieval enhancement in this embodiment includes the following steps:

[0059] S1. Collect relevant data of the power generation system and construct a fault symptom query library based on the relevant data;

[0060] S2. Obtain the current fault symptom description of the power generation system and perform a search in the fault symptom query library based on the current fault symptom description, and then obtain multiple data segments related to the current fault symptom description;

[0061] S3. Encode the multiple data segments and form a set, and divide the set into K clusters through a clustering algorithm to obtain M fault knowledge enhancement subsets;

[0062] S4. Train a lightweight language model as a fault handling draft generator, and input the M fault knowledge enhancement subsets into the fault handling draft generator to obtain M fault handling drafts;

[0063] S5. Use a general large language model as a fault handling decision model, score M fault handling drafts, and take the optimal fault handling draft as the final fault handling decision;

[0064] S6. Take the final fault handling decision as the final response strategy and execute it.

[0065] Further, the step of collecting relevant data of the power generation system and constructing a fault symptom query library based on the relevant data includes:

[0066] Collect technical documents, books, historical fault records, maintenance manuals, and fault analysis reports of the power generation system as relevant data of the power generation system;

[0067] Slice the relevant data into text blocks, encode them into vectors using the BEG model and store them in a vector database, and use the vector database as the fault diagnosis query library.

[0068] Further, the step of obtaining the current fault symptom description of the power generation system and retrieving in the fault symptom query library based on the current fault symptom description to obtain multiple data segments related to the current fault symptom description includes:

[0069] Obtain the current fault symptom description of the power generation system and retrieve the corresponding text blocks in the fault symptom query library;

[0070] Take the corresponding text blocks as multiple data segments related to the current fault symptom description;

[0071] Specifically, the fault symptom description includes factors such as real-time sensor data of the control system, equipment operation status description, historical fault records, temperature, etc.; based on the current fault symptom description, text similarity retrieval (retrieval methods include keyword-based, BGE retrieval model, etc.) is performed on relevant data (document data) such as sensor data, equipment operation status description, technical documents, books, historical fault records, maintenance manuals, and fault analysis reports in the fault symptom query library, and multiple data segments related to the current fault symptom description or scenario can be obtained.

[0072] Further, the step of encoding multiple data segments into a set, dividing the set into K clusters through a clustering algorithm, and obtaining M fault knowledge enhanced subsets includes:

[0073] Set the set of multiple data segments as , map multiple data segments to vectors , , and then the vectors of multiple data segments form a set , where, It represents encoding data segments into semantic vectors using the Sentence-BERT algorithm ;

[0074] Based on the K-means algorithm, clustering the set V into k clusters , and the clustering process of the set V satisfies the objective function to be minimized, and its objective function is:

[0075] ,

[0076] where, is the minimization of the objective function, is the j-th cluster, is the cluster center of, is the i-th vector in the set V;

[0077] Randomly select a data segment from each of the k clusters, and then obtain M fault knowledge enhanced subsets. Among them, the preset M fault knowledge enhanced subsets are , then, the -th fault knowledge enhanced subset is expressed as:

[0078] ,

[0079] where, represents a data segment randomly selected in the cluster .

[0080] Specifically, use the sentence-bert algorithm to encode the retrieved data segments into vectors, then cluster the retrieval results based on the K-Means clustering algorithm, and each cluster represents a different fault analysis perspective; then, randomly select a data segment from each cluster to generate multiple fault knowledge enhanced subsets; through the multi-perspective sampling strategy, the system efficiently clusters the vectors of multiple data segments to form the set V to reduce redundancy and ensure multi-perspective fault analysis, ensuring the diversity of different fault characteristics; among them, the role of multi-perspective sampling: these data segments may come from different technical materials or different monitoring devices, sensors or historical fault records, with a large amount of redundancy; in order to avoid information redundancy and improve the diversity of data, a multi-perspective sampling strategy is introduced.

[0081] Furthermore, the steps of training the lightweight language model as a fault handling draft generator and inputting the M fault knowledge enhanced subsets into the fault handling draft generator to obtain M fault handling drafts include:

[0082] Train a lightweight language model as a fault handling draft generator, given a fault event , enhanced by the answer A of the fault knowledge enhancement subset S, to generate the reason (reasoning process) E, that is, to obtain as the training data T, Q is the given query, , where the reasoning process E extracts key information from the fault knowledge enhancement subset S and concisely explains why the answer is reasonable for the question; the length of the reasoning should be kept short (the number of words does not exceed H, and H is set by relevant personnel according to the actual situation), and ensure that its meaning is consistent with the content of the fault knowledge enhancement subset S; use ChatGpt (or other large artificial intelligence models) based on the thinking chain of prompts technology to automatically generate the reasoning E of each triple, that is, to obtain as the training data T;

[0083] Fine-tune the lightweight pre-trained language model using the loss function. The loss function L is:

[0084] ,

[0085] In the formula, is the probability that the fault handling draft generator generates the answer and the reason under the condition of the given query and the fault knowledge enhancement subset ; train the model by maximizing this probability so that it can more accurately predict the answer and the reason;

[0086] Input all the fault knowledge enhancement subsets into the fault handling draft generator in parallel. Then, for each fault handling draft generated by the fault knowledge enhancement subset and its corresponding reasoning process are:

[0087] ,

[0088] In the formula, is the fault handling draft generated by the j-th fault knowledge enhancement subset, , and each fault handling draft represents a potential fault handling solution; represents the reason for the j-th fault handling draft generated, , represents the fault handling draft generator;

[0089] Finally, obtain M fault handling drafts , The expression of is:

[0090] .

[0091] Specifically, the draft generation process is parallelized, which not only improves efficiency but also avoids positional deviations in long texts and token overflow in model inputs caused by long texts.

[0092] Further, the step of using a general large language model as a fault handling decision model and scoring M fault handling drafts, and taking the optimal fault handling draft as the final fault handling decision includes:

[0093] Using a general large language model (such as ChatGPT, etc.) as a fault handling decision model and respectively scoring M fault handling drafts in three dimensions of consistency, reliability, and historical fault record consistency, and then obtaining the consistency scores, reliability scores, and historical fault record consistency scores of M fault handling drafts ;

[0094] Respectively determining the minimum passing score thresholds for the three dimensions of consistency, reliability, and historical fault record consistency according to the percentiles of the scores of M fault handling drafts. The minimum passing score thresholds for the three dimensions are respectively expressed as , and . Express the scores of the three dimensions as , and respectively, and the weights corresponding to the scores of the three dimensions are respectively expressed as ;

[0095] Calculating the optimal fault handling draft through the scores of the three dimensions and the minimum passing score thresholds of the three dimensions. The calculation formula is:

[0096] ,

[0097] where, is the optimal fault handling draft, represents the score of the jth fault handling draft for consistency, represents the score of the jth fault handling draft for reliability, represents the score of the jth fault handling draft for historical fault record consistency, represents the weight corresponding to the score of consistency, represents the weight corresponding to the score of reliability, represents the weight corresponding to the score of historical fault record consistency. The optimal fault handling draft is the fault handling draft with the highest score under the conditions of satisfying: and and .

[0098] Further, the scoring expression for the consistency of the j-th fault handling draft is:

[0099] ,

[0100] wherein, is the scoring function for consistency, representing a function for calculating the matching degree between the fault handling draft and the current fault scenario. Among them, the consistency score is used to score the consistency of the fault handling draft and the reasoning process (reasons) under the current fault symptom description corresponding to the current fault scenario;

[0101] The scoring expression for the reliability of the j-th fault handling draft is:

[0102] ,

[0103] wherein, is the scoring function for reliability. Among them, the reliability score is used to score the rationality of the draft through the fault handling decision model to ensure that the generated fault handling draft can be effectively executed in actual fault handling;

[0104] The scoring expression for the consistency of the j-th fault handling draft with historical fault records is:

[0105] ,

[0106] wherein, is the scoring function for the consistency of historical fault records, is a similar case of historical fault records. Among them, the consistency of historical fault records is used to measure the consistency between the fault handling draft and historical fault records. The scoring function for the consistency of historical fault records matches and compares according to the fault handling draft and historical fault records, and outputs a score for the consistency of historical fault records. In addition, historical fault records can also increase attention and be used as a separate input variable;

[0107] Specifically, the scoring function is scored by the fault handling decision model. The scoring function for consistency includes: evaluating the consistency between the fault handling draft, the reasons for the fault handling draft and the current fault symptoms. The judgment criterion is whether the reasoning process of the draft matches the current fault symptom description, and a consistency score is given. The scoring range is [0 - 10];

[0108] The scoring process of the scoring function for consistency is: Fault symptoms: {input the content of the fault symptom description};

[0109] Fault handling draft: {input the content of the fault handling draft};

[0110] Reasons for the fault handling draft: {input the content of the reasons for the fault handling draft};

[0111] A consistency score of [0 - 10] can be obtained.

[0112] The scoring function for reliability includes:

[0113] Evaluate the consistency between the fault handling draft, the reasons for the fault handling draft, and the current fault symptoms. The judgment criterion is whether the draft is reasonable and can be effectively executed in actual fault handling, and a reliability score is given, with the scoring range being [0 - 10];

[0114] The scoring process of the scoring function for reliability is as follows: Fault symptoms: {Input the description content of the fault symptoms};

[0115] Fault handling draft: {Input the content of the fault handling draft};

[0116] Reasons for the fault handling draft: {Input the reasons content of the fault handling draft};

[0117] A reliability score of [0 - 10] can be obtained.

[0118] The scoring function for the consistency of historical fault records includes: Evaluate the consistency between the fault handling draft, the reasons for the fault handling draft, the current fault symptoms, and the historical fault records. The judgment criterion is to measure the consistency between the fault handling draft and the historical fault records, and a historical fault record consistency score is given, with the scoring range being [0 - 10];

[0119] The scoring process of the scoring function for the consistency of historical fault records is as follows:

[0120] Fault symptoms: {Input the description content of the fault symptoms};

[0121] Historical fault records: {Input the content of the historical fault records};

[0122] Fault handling draft: {Input the content of the fault handling draft};

[0123] Reasons for the fault handling draft: {Input the reasons content of the fault handling draft};

[0124] A historical fault record consistency score of [0 - 10] can be obtained.

[0125] Furthermore, the step of taking the final fault handling decision as the final response strategy and executing it includes:

[0126] Automatically execute the final response strategy or hand it over to the operator for execution. During the execution process, monitor the recovery situation of the currently described fault symptoms in real time and make corresponding adjustments.

[0127] In this embodiment, the large language model can analyze and analogize similar abnormal situations according to the types of faults occurring in the power generation system, forming a closed loop of "quick response - parallel analysis - joint decision-making"; in the initial stage of the fault occurrence, the large language model parallelly analyzes and matches the type of the fault and the fault elimination measures, effectively shortening the overall response time for troubleshooting; by utilizing the large language model, useful and readable fault assistance decision-making schemes can be generated, greatly reducing the time for fault assistance decision-making and the time cost required for manual fault analysis; by enabling the large language model to perform speculative retrieval to enhance the analysis of problems from multiple perspectives, the risk of misjudgment from a traditional single perspective can be reduced; through the realization of information intercommunication during the analysis process of the large language model, comprehensive and accurate assistance decision-making schemes can be formed by integrating data; by clustering the retrieved fault data to form multiple fault knowledge enhancement subsets from different perspectives, the repeated calculation of redundant data can be avoided, and it is ensured that valuable fault information can be obtained from multiple perspectives. Furthermore, this multi-perspective sampling technology improves the comprehensiveness and accuracy of the diagnosis results; by adopting a parallel processing method to generate multiple draft disposal schemes, and verifying and scoring them by the general large language model from three dimensions: consistency, reliability score, and historical fault record consistency; different from the linear reasoning and manual intervention of traditional methods, the parallel draft generation method of the present invention not only improves the response speed of fault handling, but also ensures that the final response strategy scheme has higher accuracy and reliability.

[0128] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed in the present invention can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraint conditions of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0129] In several embodiments provided by the present invention, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only an example, and there can be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces, and the indirect couplings or communication connections of devices or units can be in an electrical, mechanical, or other forms.

[0130] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention.

[0131] Finally, the above are only preferred embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for quickly handling power generation system faults based on speculative retrieval enhanced generation, characterized in that: The following steps are involved: S1. Collect relevant data of the power generation system and build a fault symptom query library based on the relevant data; S2. Obtain a current fault symptom description of the power generation system, and search in a fault symptom query library based on the current fault symptom description, thereby obtaining multiple data fragments related to the current fault symptom description; S3, data encoding multiple data fragments to form a set, and dividing the set into K clusters through a clustering algorithm, thereby obtaining M fault knowledge enhanced subsets; The step S3 includes: setting a set of multiple data segments as , based on the sentence-BERT algorithm, multiple data fragments are Mapping to vector , , and then the vectors of multiple data fragments constitute a set ,in, Indicates that the data fragment is converted using the Sentence-BERT algorithm Encoded as semantic vector ; Based on the K-means algorithm, the set V is clustered into k clusters. , the clustering process of set V makes the objective function Minimize, its objective function is: ， In the formula, is the objective function, is the jth cluster, yes The cluster center of is the i-th vector in the set V; A data segment is randomly selected from each of the k clusters, and then M fault knowledge enhancement subsets are obtained, where the preset generated M fault knowledge enhancement subsets are , then, Fault Knowledge Enhanced Subset It is expressed as: ， In the formula, Indicates that in clustering A randomly selected data segment from ; S4. Train a lightweight language model as a fault handling draft generator, input M fault knowledge enhancement subsets into the fault handling draft generator, and then obtain M fault handling drafts; S5. Use the general large language model as the fault handling decision model, score the M fault handling drafts, and use the optimal fault handling draft as the final fault handling decision; S6. The final fault handling decision is taken as the final response strategy and executed.

2. The method for rapid handling of power generation system faults based on speculative retrieval enhanced generation according to claim 1, characterized in that: The step of collecting relevant data of the power generation system and building a fault symptom query library based on the relevant data includes: Collect technical documents, books, historical fault records, maintenance manuals, and fault analysis reports of the power generation system as relevant data of the power generation system; The relevant data are divided into text blocks, encoded into vectors using the BEG model and stored in a vector database, which is used as a fault diagnosis query library.

3. The method for rapid handling of power generation system faults based on speculative retrieval enhanced generation according to claim 2, characterized in that: The step of obtaining the current fault symptom description of the power generation system and searching in a fault symptom query library based on the current fault symptom description to obtain a plurality of data segments related to the current fault symptom description includes: Obtaining a current fault symptom description of the power generation system, and retrieving a text block corresponding to the current fault symptom description in a fault symptom query library; The corresponding text blocks are used as multiple data segments related to the current fault symptom description.

4. The method for rapid handling of power generation system faults based on speculative retrieval enhanced generation according to claim 1, characterized in that: The step of training a lightweight language model as a fault handling draft generator, inputting M fault knowledge enhancement subsets into the fault handling draft generator, and then obtaining M fault handling drafts includes: Train a lightweight language model as a fault handling draft generator. Given a fault event , the answer A of the fault knowledge enhancement subset S is enhanced to generate reason E, that is, As training data T, where Q is a given query, ; The lightweight pre-trained language model is fine-tuned using the loss function, and the loss function L is: ， In the formula, In a given query and fault knowledge enhanced subset The fault handling draft generator generates answers under the conditions and reasons The probability of Enhance the subset of all fault knowledge The fault handling draft generator is input in parallel. Then, the fault handling draft generated for each fault knowledge enhancement subset and its corresponding reasoning process are as follows: ， In the formula, The fault handling draft generated for the jth fault knowledge enhanced subset, represents the reason for generating the jth fault handling plan, Represents a fault handling draft generator; Finally, M fault handling drafts are obtained , The expression is: 。 5. The method for rapid handling of power generation system faults based on speculative retrieval enhanced generation according to claim 4, characterized in that: The steps of using the general large language model as the fault handling decision model, scoring M fault handling drafts, and taking the optimal fault handling draft as the final fault handling decision include: The general large language model is used as the fault handling decision model and M fault handling drafts are respectively The consistency, reliability and consistency of historical fault records are scored, and then M fault handling drafts are obtained. Consistency score, reliability score and historical failure record consistency score; According to the percentiles of the scores of the M fault handling drafts, the minimum qualified score thresholds of the three dimensions of consistency, reliability and historical fault record consistency are determined respectively. The minimum qualified score thresholds of the three dimensions are expressed as , and , the scores of the three dimensions are expressed as , and , the weights corresponding to the scores of the three dimensions are expressed as ; The optimal fault handling draft is calculated by the scores of the three dimensions and the minimum qualified score thresholds of the three dimensions. The calculation formula is: ， In the formula, is the optimal fault handling protocol, represents the consistency score of the jth fault handling protocol, represents the reliability score of the jth fault handling plan, represents the score of the consistency of historical fault records of the jth fault handling draft, The weight corresponding to the consistency score, The weight corresponding to the reliability score, The weight corresponding to the score of the consistency of historical fault records is represented by the optimal fault handling draft when: and and The highest scoring draft fault disposal in the case.

6. The method for rapid handling of power generation system faults based on speculative retrieval enhanced generation according to claim 5, characterized in that: The scoring expression for the consistency of the j-th fault handling protocol is: ， In the formula, is the scoring function for consistency; The scoring expression of the reliability of the j-th fault handling plan is: ， In the formula, is the scoring function for reliability; The scoring expression for the consistency of the historical fault records of the j-th fault handling draft is: ， In the formula, is the scoring function for the consistency of historical fault records, Similar cases recorded for historical failures.

7. The method for rapid handling of power generation system faults based on speculative retrieval enhanced generation according to claim 1, characterized in that: The step of taking the final fault handling decision as the final response strategy and executing it includes: The final response strategy is automatically executed or handed over to the operator for execution. During the execution process, the recovery status of the current fault symptom description is monitored in real time and corresponding adjustments are made.

Images