Database fault diagnosis large model training method and system and readable storage medium

Abstract

The invention relates to a database fault diagnosis large model training method and system and a readable storage medium, and the method comprises the steps: carrying out the structured preprocessing of a fault work order of a database, converting the processing process of the fault work order into a triple structure containing problem description, diagnosis interaction and a final solution, and storing the triple structure into a work order library; performing text vectorization processing on the triple structure to convert the triple structure into a semantic vector, and constructing a hierarchical index structure based on the triple structure and the corresponding semantic vector; a large language model is provided, data in a work order library serves as a training sample, a reinforcement learning algorithm is adopted to train the large language model, a hierarchical index structure is called in the training process to be used for calculation of instant rewards, and then model parameters are optimized according to the instant rewards. According to the method, the large language model is trained based on the retrieval enhancement generation technology, meanwhile, the reinforcement learning algorithm is adopted to carry out fine tuning optimization on the large language model, and the accuracy and generalization ability of database fault diagnosis are improved.

Description

Technical Field

[0001] The present invention relates to the field of database technology, and in particular to a method, system and readable storage medium for training a large model of database fault diagnosis. Background Art

[0002] Database systems play a vital role in modern information infrastructure and are widely used in key areas such as finance, healthcare, telecommunications, and e-commerce. However, with the continuous growth of database scale and the increasing complexity of application scenarios, fault diagnosis of database systems has become an important challenge in operation and maintenance management.

[0003] Current database fault diagnosis technologies mainly include the following: 1. Rule-based method: This method relies on expert knowledge and predefined rules to determine possible faults by analyzing database logs, system indicators, and SQL execution. The main drawbacks of this method are: 1) The formulation of rules relies on manual experience and it is difficult to cover all possible failure scenarios; 2) The rule set needs to be continuously maintained and updated to adapt to new database versions and application scenarios; 3) Faced with complex and ever-changing database failures, rule matching methods often have limitations and it is difficult to accurately locate the problem.

[0004] 2. Method based on statistical analysis: This method uses statistical modeling of historical data, such as time series analysis and anomaly detection, to discover database performance anomalies or failure modes. Its main shortcomings include: 1) Statistical methods usually require a long period of data accumulation to form a reliable anomaly detection model; 2) Statistical analysis lacks a deep understanding of the root causes of database failures and can only provide abnormal signals but cannot accurately indicate the source of the failure; 3) Faced with a complex and changing environment, the generalization ability of statistical models is limited and it is difficult to adapt to changes in different database systems.

[0005] 3. Machine learning-based methods: In recent years, machine learning has been applied to database fault diagnosis, such as using supervised learning or unsupervised learning models for anomaly detection and root cause analysis. However, this method still has the following problems: 1) Supervised learning methods rely on high-quality labeled data, but database fault data is usually difficult to obtain or has high labeling costs.

[0006] 2) Although unsupervised learning methods can discover abnormal patterns, they often lack interpretability and are difficult to guide actual operation and maintenance.

[0007] 3) The lack of effective use of domain knowledge during training limits the generalization ability of the model.

[0008] It can be seen that the common defect of the above methods is that they lack effective knowledge enhancement and adaptive capabilities in database fault diagnosis tasks. This is mainly caused by the following aspects: 1) Traditional rules and statistical methods rely on fixed experience and patterns and are difficult to dynamically adapt to complex and changing database environments; 2) Existing machine learning methods lack sufficient database domain knowledge, which makes it difficult for the model to make accurate judgments when encountering new types of faults; 3) Existing methods generally lack the ability to retrieve external knowledge bases or document information and cannot fully utilize existing database troubleshooting experience; 4) Supervised learning methods rely on a large amount of labeled data, but the high-quality labeled data available during database operation and maintenance is limited, which limits the performance of the model. Summary of the invention

[0009] The purpose of the present invention is to provide a database fault diagnosis large model training method, system and readable storage medium to overcome the shortcomings of the prior art in database fault diagnosis tasks, which lack effective knowledge enhancement and self-adaptation capabilities, and improve the accuracy and generalization ability of database fault diagnosis.

[0010] In order to achieve the above object, the present invention provides a method for training a large model of database fault diagnosis, comprising the following steps: Performing structured preprocessing on the fault work orders in the database, converting the processing process of the fault work orders into a triple structure including problem description, diagnostic interaction and final solution, and storing it in the work order database; Performing text vectorization processing on the triple structure to convert it into a semantic vector, and constructing a hierarchical index structure based on the triple structure and the corresponding semantic vector; A large language model is provided, and the data in the work order library is used as a training sample. The large language model is trained using a reinforcement learning algorithm, and the hierarchical index structure is called during the training process for calculating the instant reward, and then the model parameters are optimized according to the instant reward.

[0011] Optionally, the structured preprocessing of the fault work order in the database includes: Extract key fields from the fault work order; A multi-round dialogue sequence is constructed based on the key fields, and the processing process of the fault work order is converted into the triple structure.

[0012] Optionally, regular expressions are used to extract key fields in the fault ticket.

[0013] Optionally, the structural preprocessing of the fault work order in the database further includes: The fault work orders are graded according to the difficulty of fault handling, and the large language model is trained in stages according to the fault levels.

[0014] Optionally, based on the triple structure and the corresponding semantic vector, the hierarchical index structure is constructed according to the database type and the fault level.

[0015] Optionally, the reinforcement learning algorithm is a GRPO algorithm, and the adopting the reinforcement learning algorithm to train the large language model includes: Sampling the diagnostic interaction from the work order library as training data, wherein each round of dialogue of the diagnostic interaction includes a user question and a standard answer; For each user question in the conversation, the hierarchical index structure is called to obtain final solutions to several similar fault tickets, and the large language model is used to generate answers for the current round; Calculate the instant reward of the answer in this round according to the final solution; Model parameters are optimized according to the immediate reward.

[0016] Optionally, the instant reward is calculated as follows: r = R sim(a，D) + R format Among them, R sim(a，D) is the similarity reward, which is used to measure the similarity between the answer in this round and the final solution. format This is a format correctness reward, which is used to evaluate whether the format of the answer in this round meets the requirements.

[0017] Optionally, when the hierarchical index structure is called to obtain multiple final solutions, the similarity reward between each final solution and the answer of the current round is calculated respectively, and the average of all similarity rewards is taken as the similarity reward for calculating the immediate reward.

[0018] Based on the same inventive concept, the present invention also provides a database fault diagnosis large model training system, comprising: A data preprocessing module is used to perform structured preprocessing on the fault work orders in the database, convert the processing process of the fault work orders into a triple structure including problem description, diagnostic interaction and final solution, and store it in the work order database; An index building module, used for performing text vectorization processing on the triple structure to convert it into a semantic vector, and building a hierarchical index structure based on the triple structure and the corresponding semantic vector; The model training module is used to provide a large language model, use the data in the work order library as training samples, adopt a reinforcement learning algorithm to train the large language model, and call the hierarchical index structure for the calculation of instant rewards during the training process, and then optimize the model parameters according to the instant rewards.

[0019] Based on the same inventive concept, the present invention also provides a readable storage medium on which a computer program is stored. When the computer program is executed, it can implement the database fault diagnosis large model training method as described above.

[0020] In the database fault diagnosis large model training method, system and readable storage medium provided by the present invention, the fault work orders in the database are structured preprocessed to facilitate conversion into semantic vectors and then construct a hierarchical index structure, so that the large language model can be trained based on retrieval enhancement generation technology. At the same time, the reinforcement learning algorithm is used to fine-tune and optimize the large language model, thereby improving the accuracy and generalization ability of database fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Those skilled in the art should understand that the drawings are provided for a better understanding of the present invention and do not constitute any limitation on the scope of the present invention. Figure 1 A flowchart of a method for training a large database fault diagnosis model according to an embodiment of the present invention; Figure 2 A flowchart of using the GRPO algorithm to train a large language model is provided in accordance with an embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, advantages and features of the present invention clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the accompanying drawings are in a very simplified form and use non-precise proportions, which are only used to conveniently and clearly assist in explaining the purpose of the embodiments of the present invention. In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, please refer to the accompanying drawings. It should be noted that the structure, proportion, size, etc. illustrated in the drawings of this specification are only used to match the content disclosed in the specification for people familiar with this technology to understand and read, and are not used to limit the limiting conditions for the implementation of the present invention. Any modification of the structure, change in the proportional relationship or adjustment of the size, under the same or similar conditions as the effects that can be produced by the present invention and the purposes that can be achieved, should still fall within the scope of the technical content disclosed by the present invention.

[0023] As used in the present invention, the singular forms "a", "an", and "the" include plural referents unless the context clearly indicates otherwise. As used in the present invention, the term "or" is generally used in a sense including "and / or" unless the context clearly indicates otherwise.

[0024] Please refer to Figure 1 This embodiment provides a method for training a large model for database fault diagnosis, comprising the following steps: S1. Perform structured preprocessing on the fault work orders in the database, convert the processing process of the fault work orders into a triple structure including problem description, diagnostic interaction and final solution, and store it in the work order database; S2, performing text vectorization processing on the triple structure to convert it into a semantic vector, and constructing a hierarchical index structure based on the triple structure and the corresponding semantic vector; S3. Provide a large language model, use the data in the work order library as training samples, adopt a reinforcement learning algorithm to train the large language model, call the hierarchical index structure for instant reward calculation during the training process, and then optimize the model parameters according to the instant reward.

[0025] By performing structured preprocessing on the fault work orders in the database, they are converted into semantic vectors and then a hierarchical index structure is constructed. The large language model can then be trained based on retrieval enhancement generation technology. At the same time, the reinforcement learning algorithm is used to fine-tune and optimize the large language model, thereby improving the accuracy and generalization ability of database fault diagnosis.

[0026] First, execute S1 to perform structured preprocessing on the fault work orders in the database, convert the processing process of the fault work orders into a triple structure including problem description, diagnostic interaction and final solution, and store it in the work order database.

[0027] In this embodiment, the structured preprocessing of the fault work order in the database includes: Extract key fields from the fault work order; A multi-round dialogue sequence is constructed based on the key fields, and the processing process of the fault work order is converted into the triple structure.

[0028] In some embodiments, regular expressions are used to extract key fields in the fault ticket, and the key fields include database version, error log, solution, operation timestamp, etc.

[0029] In some embodiments, the structural preprocessing of the fault work order in the database further includes: The fault work orders are graded according to the difficulty of fault handling, and the large language model is trained in stages according to the fault level. For example, the fault work orders are divided into single-round simple fault work orders (with clear error codes), multi-round compound fault work orders, and incomplete work orders with noise according to the difficulty of fault handling, and then the large language model is trained in stages based on these three fault levels, such as selecting the single-round simple fault work order for model training in the first stage, selecting the multi-round compound fault work order for model training in the second stage, and selecting the incomplete work order with noise for model training in the third stage, thereby continuously improving the accuracy and generalization ability of the database fault diagnosis of the large language model.

[0030] Then execute S2, perform text vectorization processing on the triple structure to convert it into a semantic vector, and construct a hierarchical index structure based on the triple structure and the corresponding semantic vector. In this embodiment, the text vectorization processing can adopt the well-known technology in the field, for example, using the Sentence-BERT model to generate a 768-dimensional semantic vector, and then based on the triple structure and the corresponding semantic vector, the hierarchical index structure can be constructed according to the database type and the fault level, so as to facilitate the subsequent call of the hierarchical index structure to achieve retrieval enhancement and improve the reliability and accuracy of model training.

[0031] Finally, S3 is executed to provide a large language model, use the data in the work order library as training samples, adopt a reinforcement learning algorithm to train the large language model, and call the hierarchical index structure for instant reward calculation during the training process, and then optimize the model parameters according to the instant reward.

[0032] In this embodiment, the large language model is a large language model commonly used in the prior art, such as Qwen2.5-7B, and the reinforcement learning algorithm is, for example, a group relative policy optimization (GRPO) algorithm. The core idea of ​​GRPO is to optimize the policy model through relative rewards within the group, rather than relying on the traditional criticism model, thereby reducing the computational burden and improving training stability.

[0033] In this embodiment, Figure 2 As shown, the use of the GRPO algorithm to train the large language model includes: S31, sampling the diagnostic interaction from the work order library as training data, wherein each round of dialogue of the diagnostic interaction includes a user question and a standard answer; S32, for each user question in the dialogue, calling the hierarchical index structure to obtain final solutions to several similar fault tickets, and using the large language model to generate answers for this round; S33, calculating the instant reward for the answer in this round according to the final solution; S34. Optimize model parameters according to the instant reward.

[0034] Specifically, S31 is first executed. In each training cycle, the diagnostic interaction is firstly sampled from the work order library as training data. Each round of dialogue of the diagnostic interaction includes user questions and standard answers.

[0035] Then, S32 is executed, and for each user question in the dialogue, the hierarchical index structure is called to obtain the final solutions of several similar fault tickets; and the large language model is used to generate the answers for this round. It should be noted that when the hierarchical index structure is called to obtain the final solution, for example, the most relevant Top-5 documents are selected as the retrieval result, and the retrieval result can be used as external knowledge support.

[0036] Then execute S33 to calculate the instant reward for the answer in this round according to the final solution. In this embodiment, the instant reward is calculated as follows: r = R sim(a，D) + R format Among them, R sim(a，D) is a similarity reward, which is used to measure the similarity between the answer in this round and the final solution to ensure the reliability of the answer; R format This is a format correctness reward, which is used to evaluate whether the format of the answer in this round meets the requirements.

[0037] For example, if the answer in this round is highly matched with the retrieved final solution, the answer in this round is considered reliable and a positive reward (+1) is given; if the match is low, a negative reward (-1) is given. The specific reward value is continuously assigned based on the ratio of the similarity score to the threshold. The specific assignment method is as follows: Calculate the similarity score sim(a, D) between the answer of this round and the retrieved final solution; Set the similarity threshold T high and T low : If sim(a, D) ≥ T high , indicating that the answer in this round highly matches the final solution, and a +1 reward is given; If sim(a, D)≤T low , indicating that the answer in this round does not match the final solution, and a -1 reward is given; If T low <sim(a,D )< T high , linear interpolation can be used to calculate the reward value to ensure the continuity of the reward value.

[0038] The calculation of similarity reward is explained below through a specific example.

[0039] For example, setting the similarity threshold T high = 0.8, T low = 0.4; If sim(a, D) = 0.9, then R sim(a，D) = +1; If sim(a, D) = 0.3, then R sim(a，D) = -1 If sim(a, D) = 0.6, then R sim(a，D) = (0.6 - 0.4) / (0.8 - 0.4) * 2 - 1 = 0.

[0040] In this embodiment, the format reward is, for example, to evaluate whether the answer format conforms to the "thinking + reply" structure. If the requirements are met, a positive reward (+1) is given; otherwise, a negative reward (-1) is given, thereby guiding the model to output according to the "thinking + reply" structure, thereby improving the readability and reasoning ability of the answer.

[0041] The specific setting method is as follows: Predefine standard examples that conform to the format and train a classifier to determine whether the answer conforms to the format: Thinking part: contains reasoning steps or explanation process; Response section: Give the final answer; If the answer conforms to this format, then R format = +1, otherwise R format = -1 .

[0042] The instant reward for this round of answers can be calculated using the similarity reward and format reward obtained.

[0043] Finally, S34 is executed to optimize the model parameters according to the instant reward. It should be noted that in the reinforcement learning algorithm, in addition to the instant reward, a value model for evaluating the current state is also included. During actual training, the existing value model and instant reward can be combined to adjust the model parameters through a policy gradient method (such as a gradient descent method), thereby guiding the model to gradually optimize the reliability and readability of the answer.

[0044] Optionally, when the hierarchical index structure is called to obtain multiple final solutions, the similarity reward between each final solution and the answer of this round can be calculated separately, and the average of all similarity rewards can be taken as the similarity reward for calculating the immediate reward, so as to improve the accuracy of the model answer.

[0045] Based on the same inventive concept, the embodiment of the present invention also proposes a database fault diagnosis large model training system, including: A data preprocessing module is used to perform structured preprocessing on the fault work orders in the database, convert the processing process of the fault work orders into a triple structure including problem description, diagnostic interaction and final solution, and store it in the work order database; An index building module, used for performing text vectorization processing on the triple structure to convert it into a semantic vector, and building a hierarchical index structure based on the triple structure and the corresponding semantic vector; The model training module is used to provide a large language model, use the data in the work order library as training samples, adopt a reinforcement learning algorithm to train the large language model, and call the hierarchical index structure for the calculation of instant rewards during the training process, and then optimize the model parameters according to the instant rewards.

[0046] Based on the same inventive concept, an embodiment of the present invention further proposes a readable storage medium on which a computer program is stored. When the computer program is executed, the database fault diagnosis large model training method as described above can be implemented.

[0047] The readable storage medium can be a tangible device that can keep and store the instructions used by the instruction execution device, such as but not limited to an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device or any suitable combination thereof. The more specific example (non-exhaustive list) of the readable storage medium includes: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a convex structure in a groove on which instructions are stored, and any suitable combination thereof. The computer program described herein can be downloaded from the readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network and / or a wireless network. The network can include copper transmission cables, optical fiber transmission, wireless transmission, routers, firewalls, switches, gateway computers and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer program from the network and forwards the computer program for storage in the readable storage medium in each computing / processing device. The computer program for performing the operation of the present invention can be an assembly instruction, an instruction set architecture (ISA) instruction, a machine instruction, a machine-related instruction, a microcode, a firmware instruction, a state setting data, or a source code or object code written in any combination of one or more programming languages, including object-oriented programming languages-such as Smalltalk, C++, etc., and conventional procedural programming languages-such as "C" language or similar programming languages. The computer program can be executed completely on the user's computer, partially on the user's computer, as an independent software package, partially on the user's computer and partially on the remote computer, or completely on the remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (for example, using an Internet service provider to connect through the Internet). In some embodiments, by utilizing the state information of a computer program to personalize an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), the electronic circuit can execute computer-readable program instructions to implement various aspects of the present invention.

[0048] Here, various aspects of the present invention are described with reference to the flowchart and / or block diagram of the method, system and computer program product according to the embodiment of the present invention. It should be understood that each square frame of the flowchart and / or block diagram and the combination of the square frames in the flowchart and / or block diagram can be realized by a computer program. These computer programs can be provided to the processor of a general-purpose computer, a special-purpose computer or other programmable data processing device, so as to produce a machine, so that these programs, when executed by the processor of a computer or other programmable data processing device, produce a device for realizing the function / action specified in one or more square frames in the flowchart and / or block diagram. These computer programs can also be stored in a readable storage medium, and these computer programs make the computer, programmable data processing device and / or other equipment work in a specific way, so that the readable storage medium storing the computer program includes a manufactured product, which includes instructions for realizing various aspects of the function / action specified in one or more square frames in the flowchart and / or block diagram.

[0049] The computer program may also be loaded onto a computer, other programmable data processing apparatus, or other device so that a series of operational steps are executed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the computer program executed on the computer, other programmable data processing apparatus, or other device to implement the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0050] In summary, the embodiments of the present invention provide a method, system and readable storage medium for training a large model for database fault diagnosis. By performing structured preprocessing on the fault work orders in the database, they are converted into semantic vectors and then a hierarchical index structure is constructed. The large language model can be trained based on retrieval enhancement generation technology. At the same time, a reinforcement learning algorithm is used to fine-tune and optimize the large language model, thereby improving the accuracy and generalization ability of database fault diagnosis.

[0051] In addition, it should be recognized that although the present invention has been disclosed as a preferred embodiment, the above embodiment is not intended to limit the present invention. For any technician familiar with the art, without departing from the scope of the technical solution of the present invention, the technical content disclosed above can be used to make many possible changes and modifications to the technical solution of the present invention, or modified into equivalent embodiments of equivalent changes. Therefore, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the content of the technical solution of the present invention still belongs to the scope of protection of the technical solution of the present invention.

Claims

1. A method for training a large model for database fault diagnosis, characterized in that: The following steps are involved: Performing structured preprocessing on the fault work orders in the database, converting the processing process of the fault work orders into a triple structure including problem description, diagnostic interaction and final solution, and storing it in the work order database; Performing text vectorization processing on the triple structure to convert it into a semantic vector, and constructing a hierarchical index structure based on the triple structure and the corresponding semantic vector; A large language model is provided, and the data in the work order library is used as a training sample. The large language model is trained using a reinforcement learning algorithm, and the hierarchical index structure is called during the training process for calculating the instant reward, and then the model parameters are optimized according to the instant reward.

2. The method for training a large database fault diagnosis model according to claim 1, characterized in that: The structured preprocessing of the fault work order in the database includes: Extract key fields from the fault work order; A multi-round dialogue sequence is constructed based on the key fields, and the processing process of the fault work order is converted into the triple structure.

3. The method for training a large database fault diagnosis model according to claim 2, characterized in that: Regular expressions are used to extract key fields in the fault ticket.

4. The method for training a large database fault diagnosis model according to claim 1, characterized in that: The structured preprocessing of the fault work order of the database also includes: The fault work orders are graded according to the difficulty of fault handling, and the large language model is trained in stages according to the fault levels.

5. The method for training a large database fault diagnosis model according to claim 4, characterized in that: Based on the triple structure and the corresponding semantic vector, the hierarchical index structure is constructed according to the database type and the fault level.

6. The method for training a large database fault diagnosis model according to claim 1, characterized in that: The reinforcement learning algorithm is the GRPO algorithm, and the use of the reinforcement learning algorithm to train the large language model includes: Sampling the diagnostic interaction from the work order library as training data, wherein each round of dialogue of the diagnostic interaction includes a user question and a standard answer; For each user question in the conversation, the hierarchical index structure is called to obtain final solutions to several similar fault tickets, and the large language model is used to generate answers for the current round; Calculate the instant reward of the answer in this round according to the final solution; Model parameters are optimized according to the immediate reward.

7. The method for training a large database fault diagnosis model according to claim 6, characterized in that: The instant reward is calculated as follows: r= R sim(a，D) + R format Among them, R sim(a，D) is the similarity reward, which is used to measure the similarity between the answer in this round and the final solution. format This is a format correctness reward, which is used to evaluate whether the format of the answer in this round meets the requirements.

8. The method for training a large database fault diagnosis model according to claim 7, characterized in that: When the hierarchical index structure is called to obtain multiple final solutions, the similarity reward between each final solution and the answer of the current round is calculated respectively, and the average of all similarity rewards is taken as the similarity reward for calculating the immediate reward.

9. A database fault diagnosis large model training system, characterized in that: include: A data preprocessing module is used to perform structured preprocessing on the fault work orders in the database, convert the processing process of the fault work orders into a triple structure including problem description, diagnostic interaction and final solution, and store it in the work order database; An index building module, used for performing text vectorization processing on the triple structure to convert it into a semantic vector, and building a hierarchical index structure based on the triple structure and the corresponding semantic vector; The model training module is used to provide a large language model, use the data in the work order library as training samples, adopt a reinforcement learning algorithm to train the large language model, and call the hierarchical index structure for the calculation of instant rewards during the training process, and then optimize the model parameters according to the instant rewards.

10. A readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed, it can implement the database fault diagnosis large model training method according to any one of claims 1-8.

Images