Alarm analysis method and system based on integrated large language model

Through the integrated alarm analysis method of large language model, the problem of low efficiency in cross-system alarm information analysis is solved, and the cross-system alarm information coordination and low-cost analysis effect is achieved.

CN118674057BActive Publication Date: 2024-12-03北京天旦网络科技发展有限公司 +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411154874.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-22
Publication Date
2024-12-03
Estimated Expiration
2044-08-22

AI Technical Summary

Technical Problem

The existing technology is difficult to effectively analyze and judge alarm information across systems, resulting in the manual summary and analysis taking a lot of manpower and time, and a single large language model cannot process complex alarm information.

Method used

The alarm analysis method based on the fusion large language model is adopted, and the alarm information coordination is achieved through access to various alarm subsystems, screening and fusion large language models, natural language descriptions are generated and comprehensive reasoning is carried out to achieve cross-system alarm information coordination.

Benefits of technology

提高了跨系统告警分析和诊断效率,降低了不同系统间告警信息交互的成本,提高了复杂告警研判速度,具备较高的实用性和低实施成本。

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118674057B_ABST
    Figure CN118674057B_ABST
Patent Text Reader

Abstract

The present invention provides an alarm analysis method and system based on a fused large language model, including step S1: connecting each alarm subsystem to a corresponding device, collecting and processing relevant information; the device includes an alarm subsystem alarm access device, an alarm interpretation model and a comprehensive reasoning model screening device, an alarm interpretation model fusion device, and an alarm analysis device; step S2: filtering out large language models that meet preset conditions and then fusing them; step S3: outputting the analysis result based on the fused large language model. The present invention realizes the conversion of alarm instructions to text descriptions and the reasoning of multiple alarm text descriptions to root causes through a large language model, realizes cross-system alarm information coordination, improves the efficiency of cross-system alarm analysis and diagnosis, and achieves the effect of low-cost alarm information interaction between different systems.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of large language models, and in particular to an alarm analysis method and system based on a fusion large language model. Background Art

[0002] The information systems of large enterprises usually contain a wide variety of subsystems that are interrelated. Generally, these different subsystems have their own independent alarm subsystems due to their suppliers or technical reasons. An abnormality at a certain local point will often cause multiple alarm subsystems associated with it to alarm. Then, the information from each subsystem is manually summarized and analyzed to find the root cause. This process consumes a lot of manpower and time.

[0003] The Large Language Model (LLM) can answer user questions inputted, but different LLMs have different focuses. For example, some LLMs are suitable for comprehensive reasoning, some are good at program error detection, and some have more network knowledge. However, complex alarm information from multiple subsystems cannot be analyzed through a single existing model.

[0004] In the Chinese patent document with publication number CN115858226A, an intelligent operation and maintenance system based on artificial intelligence is disclosed, which relates to the field of intelligent operation and maintenance systems. The intelligent operation and maintenance system based on artificial intelligence includes a data acquisition management module, an intelligent alarm analysis module, an intelligent indicator analysis module, an intelligent log analysis module, an intelligent operation decision module and an AI operation and maintenance middle platform. However, the patent document uses traditional statistical machine learning models to interpret class indicators and generate alarms; while the present invention is mainly based on alarm analysis and reasoning supported by a large language model, which is essentially different from the technical method used in the patent document.

[0005] In the Chinese patent document with publication number CN111294217A, an alarm analysis method, device, system and storage medium are disclosed, which are applied to the field of operation and maintenance monitoring, wherein the alarm analysis method includes: importing a business system topology diagram; obtaining node matching rules; after receiving an alarm message, determining the target node corresponding to the alarm message according to the alarm message and the node matching rules, and presenting the alarm message in a visual form on the business system topology diagram where the target node is located. However, the patent document performs alarm matching and analysis through the business topology diagram and pre-set rules; while the present invention performs alarm analysis and reasoning based on the support of a large language model, which is essentially different from the technical method adopted in the patent document. Summary of the invention

[0006] In view of the defects in the prior art, the purpose of the present invention is to provide an alarm analysis method and system based on a fusion large language model.

[0007] According to the present invention, an alarm analysis method based on a fusion large language model is provided, comprising:

[0008] Step S1: Connect each alarm subsystem to the corresponding device to collect and process relevant information;

[0009] The device includes an alarm subsystem alarm access device, an alarm interpretation model and a comprehensive reasoning model screening device, an alarm interpretation model fusion device and an alarm analysis and judgment device;

[0010] Step S2: select large language models that meet preset conditions and then fuse them;

[0011] Step S3: Output the analysis result based on the fused large language model.

[0012] Preferably, the step S1 includes inputting original alarm information of the subsystem and outputting natural language alarm information.

[0013] Preferably, step S1 includes the following sub-steps:

[0014] Step S1.1: The alarm information description template library module stores templates for alarm information of different types of alarm subsystems; the alarm information description template library module can customize templates according to the operation of the user system;

[0015] Step S1.2: The original alarm text generation module inputs the original alarm information of each alarm subsystem, selects the corresponding description template from the template library according to different alarm subsystem types, fills in the template and generates a natural language description of the alarm.

[0016] Preferably, step S2 includes the following sub-steps:

[0017] Step S2.1: Based on the natural language alarm information and reasoning scenario information of different alarm subsystems, select large language models that meet the preset conditions;

[0018] Step S2.2: Input a large language model that meets the preset conditions, integrate the interpretation capabilities of all large language models for different alarm sources, and output a fused alarm interpretation large language model.

[0019] Preferably, step S2.1 screens a large language model that meets preset conditions based on an alarm interpretation model selection module and an alarm comprehensive reasoning model selection module; the input of the alarm interpretation model selection module includes a natural language description of the alarm information, and the output includes the name of the large language model used to explain each alarm type; the input of the alarm comprehensive reasoning model selection module includes an alarm language description set and AlertSelectedModelSet, and the output includes the name of the large language model for alarm comprehensive reasoning.

[0020] Preferably, the step S2.1 includes the following sub-steps:

[0021] Step S2.1.1: Collect alarm natural language descriptions and construct an alarm language description set; add command prompts to the alarm natural language description text, wherein the command prompts include requiring the model to output an alarm type; construct an alarm analysis instruction set; each element in the alarm analysis instruction set is an analysis instruction;

[0022] Step S2.1.2: Input each element in the alarm language description set to all candidate large language models one by one, and calculate the corresponding Perplexity; based on the model output content, extract the alarm type, aggregate the Perplexity of the model output by the alarm type and model name, and calculate the average Perplexity to obtain the AlertModelPerplexity set; each element in the AlertModelPerplexity set is a triple consisting of "alarm type, model name, average perplexity";

[0023] Step S2.1.3: Input each element in the alarm analysis instruction set to all candidate large language models one by one, the element including the analysis instruction, record the return value, and obtain the AlertModelResponse set; each element in the AlertModelResponse set is a triple consisting of "analysis instruction, model name, alarm explanation description";

[0024] Step S2.1.4: Combine the instruction of each element in AlertModelResponse with the return statement part to form a scoring instruction, so that all candidate models give rationality scores and averages, and obtain an AlertModelRespScore set; each element in the AlertModelRespScore set is a triple of alarm type, model name and rationality average score;

[0025] Step S2.1.5: Calculate the average of "1 / average perplexity" and "average rationality score" for the records with the same alarm type and model name in AlertModelPerplexity and AlertModelRespScore. This average is called the "comprehensive rationality score" of the model. For a certain alarm type, the model corresponding to the record with the highest "comprehensive rationality score" is called the preferred model of the alarm type. Each alarm type and its preferred model constitute a tuple. Multiple tuples constitute an alarm type preferred model set, which is recorded as AlertSelectedModelSet.

[0026] Step S2.1.6: Input each alarm in the alarm language description set into the model that best suits the alarm and output the corresponding description, and combine the descriptions to obtain a SelectedModelRespSet set; each element in the SelectedModelRespSet set is a tuple of an alarm time and an alarm explanation description;

[0027] Step S2.1.7: Construct a diagnosis description set, use the DBSCAN algorithm to cluster the elements in SelectedModelRespSet with the occurrence time as the feature, cluster the alarms that occur within the preset time into a cluster, combine the elements in the same cluster and add the task description to obtain the reasoning instructions of the alarm comprehensive reasoning model;

[0028] Step S2.1.8: Input each element in the diagnosis description set into all candidate large language models one by one, record the return value, and obtain the DiagnoseModelResp set; each element in the DiagnoseModelResp set is a triple of reasoning instruction, model name and reasoning conclusion;

[0029] Step S2.1.9: Combine the reasoning instruction of each element in DiagnoseModelResp with the reasoning conclusion part to form a reasoning scoring instruction, so that all candidate models give a rationality score and give an average, and obtain a DiagnoseModelRespScore set; each element in the DiagnoseModelRespScore set is a triple of reasoning instruction, model name and reasoning rationality score;

[0030] Step S2.1.10: Calculate the mean of the reasoning rationality scores for the records with the same model name in DiagnoseModelRespScore, and select the model with the highest mean as the alarm comprehensive reasoning model.

[0031] Preferably, the step S2.2 includes the following sub-steps:

[0032] Step S2.2.1: construct a unified vocabulary of multiple alarm interpretation models;

[0033] Step S2.2.2: construct and initialize an untrained fusion model; the architecture of the model conforms to the decoder part of the transformer; sort the words in VocubCommon and start encoding from 0, and use the ordered sequence of words obtained as the vocabulary of the model;

[0034] Step S2.2.3: prepare training samples, select the most suitable model for each alarm instruction in the alarm analysis instruction set according to the relationship between the alarm type and the model in the AlertSelectedModelSet set, denoted as AlertModel_i, calculate the weight vector of the instruction through the model, denoted as AlertWeightVec_i; each dimension of the weight vector AlertWeightVec_i corresponds to a value in the vocabulary of AlertModel_i, adjust the numerical order in AlertWeightVec_i according to the order of VocubCommon, fill the missing part of the vector with 0, and the new vector is denoted as AlertCommonWeightVec_i, which is called the alarm explanation description vector; generate a vector for each data in the alarm analysis instruction set, and construct a training set as TrainSet; each element in the training set is a tuple of an analysis instruction and an alarm explanation description vector;

[0035] Step S2.2.4: training the fusion model; the training includes sending the analysis instruction part of each sample in TrainSet into the initialized untrained fusion model to obtain pred_vec, calculating the cross entropy of pred_vec and the alarm explanation description vector of the sample as the loss, and updating the initialized untrained fusion model based on the Adam algorithm to obtain the trained model AlertCommonModel.

[0036] Preferably, step S3 includes inputting alarm signals of each alarm subsystem, interpreting the alarm signals and comprehensively reasoning about the alarm causes, and outputting analysis and judgment results.

[0037] Preferably, step S3 includes the following sub-steps:

[0038] Step S3.1: converting the alarm into an alarm explanation description; receiving the alarm information and inputting it into a fusion model to obtain an alarm explanation description; the fusion model is a trained model;

[0039] Step S3.2: constructing the alarm explanation description within the preset time window into reasoning instructions;

[0040] Step S3.3: Input the reasoning instruction into the alarm comprehensive reasoning model and use the output as alarm analysis information.

[0041] According to the present invention, an alarm analysis system based on a fusion large language model is provided, comprising:

[0042] Module M1: connect each alarm subsystem to the corresponding device, collect and process relevant information;

[0043] The device includes an alarm subsystem alarm access device, an alarm interpretation model and a comprehensive reasoning model screening device, an alarm interpretation model fusion device and an alarm analysis and judgment device;

[0044] Module M2: Filter out the large language models that meet the preset conditions and then fuse them;

[0045] Module M3: Outputs the analysis and judgment results based on the fused large language model.

[0046] Preferably, the module M1 includes inputting original alarm information of the subsystem and outputting natural language alarm information.

[0047] Preferably, the module M1 includes the following submodules:

[0048] Module M1.1: The alarm information description template library module stores templates for alarm information of different types of alarm subsystems; the alarm information description template library module can customize templates according to the operation of the user system;

[0049] Module M1.2: The original alarm text generation module inputs the original alarm information of each alarm subsystem, selects the corresponding description template from the template library according to the different alarm subsystem types, and generates a natural language description of the alarm after filling in the template.

[0050] Preferably, the module M2 includes the following submodules:

[0051] Module M2.1: Based on the natural language alarm information and reasoning scenario information of different alarm subsystems, select the large language models that meet the preset conditions;

[0052] Module M2.2: Input the large language model that meets the preset conditions, integrate the interpretation capabilities of all large language models for different alarm sources, and output a fused alarm interpretation large language model.

[0053] Preferably, the module M2.1 screens a large language model that meets preset conditions based on the alarm interpretation model selection module and the alarm comprehensive reasoning model selection module; the input of the alarm interpretation model selection module includes the natural language description of the alarm information, and the output includes the name of the large language model used to explain each alarm type; the input of the alarm comprehensive reasoning model selection module includes the alarm language description set and AlertSelectedModelSet, and the output includes the name of the large language model for alarm comprehensive reasoning.

[0054] Preferably, the module M2.1 includes the following submodules:

[0055] Module M2.1.1: Collect alarm natural language descriptions and construct an alarm language description set; add command prompts to the alarm natural language description text, wherein the command prompts include the alarm type required to be output by the model; construct an alarm analysis instruction set; each element in the alarm analysis instruction set is an analysis instruction;

[0056] Module M2.1.2: Input each element in the alarm language description set to all candidate large language models one by one, and calculate the corresponding Perplexity; based on the model output content, extract the alarm type, aggregate the Perplexity of the model output by the alarm type and model name, and calculate the average Perplexity to obtain the AlertModelPerplexity set; each element in the AlertModelPerplexity set is a triple consisting of "alarm type, model name, average perplexity";

[0057] Module M2.1.3: Input each element in the alarm analysis instruction set to all candidate large language models one by one, the element includes the analysis instruction, records the return value, and obtains the AlertModelResponse set; each element in the AlertModelResponse set is a triple consisting of "analysis instruction, model name, alarm explanation description";

[0058] Module M2.1.4: Combine the instruction of each element in AlertModelResponse with the return statement part to form a scoring instruction, so that all candidate models give a rationality score and an average value, and obtain an AlertModelRespScore set; each element in the AlertModelRespScore set is a triple of alarm type, model name and average rationality score;

[0059] Module M2.1.5: Calculate the average of "1 / average perplexity" and "average rationality score" for the records with the same alarm type and model name in AlertModelPerplexity and AlertModelRespScore. This average is called the "comprehensive rationality score" of the model. For a certain alarm type, the model corresponding to the record with the highest "comprehensive rationality score" is called the preferred model of the alarm type. Each alarm type and its preferred model constitute a tuple. Multiple tuples constitute the alarm type preferred model set, which is recorded as AlertSelectedModelSet.

[0060] Module M2.1.6: Input each alarm in the alarm language description set into the model that best suits the alarm and output the corresponding description, and combine the descriptions to obtain the SelectedModelRespSet set; each element in the SelectedModelRespSet set is a tuple of the alarm time and the alarm explanation description;

[0061] Module M2.1.7: Construct a diagnostic description set, use the DBSCAN algorithm to cluster the elements in SelectedModelRespSet based on the occurrence time, cluster the alarms that occur within the preset time into a cluster, combine the elements in the same cluster and add the task description to obtain the reasoning instructions of the alarm comprehensive reasoning model;

[0062] Module M2.1.8: Input each element in the diagnosis description set into all candidate large language models one by one, record the return value, and obtain the DiagnoseModelResp set; each element in the DiagnoseModelResp set is a triple of reasoning instruction, model name and reasoning conclusion;

[0063] Module M2.1.9: Combine the reasoning instruction of each element in DiagnoseModelResp with the reasoning conclusion part to form a reasoning scoring instruction, so that all candidate models give a rationality score and give an average, and obtain a DiagnoseModelRespScore set; each element in the DiagnoseModelRespScore set is a triple of reasoning instruction, model name and reasoning rationality score;

[0064] Module M2.1.10: Calculate the mean of the reasoning rationality scores for the records with the same model name in DiagnoseModelRespScore, and select the model with the highest mean as the alarm comprehensive reasoning model.

[0065] Preferably, the module M2.2 includes the following submodules:

[0066] Module M2.2.1: Construct a unified vocabulary for multiple alarm interpretation models;

[0067] Module M2.2.2: Build and initialize an untrained fusion model; the architecture of the model conforms to the decoder part of the transformer; sort the words in VocubCommon and start encoding from 0, and use the ordered sequence of words obtained as the vocabulary of the model;

[0068] Module M2.2.3: Prepare training samples, select the most suitable model for each alarm instruction in the alarm analysis instruction set according to the relationship between the alarm type and the model in the AlertSelectedModelSet set, denoted as AlertModel_i, and calculate the weight vector of the instruction through the model, denoted as AlertWeightVec_i; each dimension of the weight vector AlertWeightVec_i corresponds to a value in the vocabulary of AlertModel_i, adjust the numerical order in AlertWeightVec_i according to the order of VocubCommon, fill the missing part of the vector with 0, and the new vector is denoted as AlertCommonWeightVec_i, which is called the alarm explanation description vector; generate a vector for each data in the alarm analysis instruction set, and construct a training set as TrainSet; each element in the training set is a tuple of an analysis instruction and an alarm explanation description vector;

[0069] Module M2.2.4: training fusion model; the training includes sending the analysis instruction part of each sample in TrainSet into the initialized untrained fusion model to obtain pred_vec, calculating the cross entropy of pred_vec and the alarm explanation description vector of the sample as the loss, and updating the initialized untrained fusion model based on the Adam algorithm to obtain the trained model AlertCommonModel.

[0070] Preferably, the module M3 includes inputting alarm signals of each alarm subsystem, interpreting the high-information signals and comprehensively reasoning about the alarm causes, and outputting the analysis and judgment results.

[0071] Preferably, the module M3 includes the following submodules:

[0072] Module M3.1: converting alarms into alarm explanation descriptions; receiving alarm information and inputting it into a fusion model to obtain alarm explanation descriptions; the fusion model is a trained model;

[0073] Module M3.2: Construct the alarm explanation description within the preset time window into reasoning instructions;

[0074] Module M3.3: Input the reasoning instructions into the alarm comprehensive reasoning model and use the output as alarm analysis information.

[0075] Compared with the prior art, the present invention has the following beneficial effects:

[0076] 1. The present invention realizes the conversion of alarm instructions into text descriptions and the reasoning of multiple alarm text descriptions to root causes through a large language model, realizes cross-system alarm information coordination, improves the efficiency of cross-system alarm analysis and diagnosis, and achieves the effect of low-cost alarm information interaction between different systems.

[0077] 2. The device proposed in the present invention can select appropriate LLMs for each alarm subsystem according to the alarm content, and then fuse a dedicated analysis model through distillation technology, and finally realize alarm analysis through the fusion of a large language model, which has high practicality.

[0078] 3. The present invention improves the speed of analyzing and judging complex alarms, and can be used to automatically generate fusion models for different information systems, with low implementation costs and good practicality.

[0079] Other beneficial effects of the present invention will be explained in the specific implementation manner through the introduction of specific technical features and technical solutions. Through the introduction of these technical features and technical solutions, those skilled in the art should be able to understand the beneficial technical effects brought about by the technical features and technical solutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0081] Figure 1 The present invention is a flow chart of the method.

[0082] Figure 2 The figure is a flow chart of the operation of the alarm access device of the alarm subsystem in the embodiment of the present invention.

[0083] Figure 3 The present invention is a flowchart of the operation of the alarm explanation model and the comprehensive reasoning model screening device in the embodiment of the present invention.

[0084] Figure 4 The figure is a flow chart of the operation of the alarm interpretation model fusion device in the embodiment of the present invention.

[0085] Figure 5 The figure is a flow chart of the operation of the alarm analysis and judgment device in the embodiment of the present invention. DETAILED DESCRIPTION

[0086] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0087] Reference Figure 1 and Figure 2 As shown, an alarm analysis method based on the fusion of a large language model includes:

[0088] The alarm access device of the alarm subsystem is operated; the device inputs the original alarm information of the subsystem, outputs the natural language alarm information, and has an alarm information description template library module and an original alarm text generation module.

[0089] The alarm information description template library module stores templates for alarm information of different types of alarm subsystems, and there are some commonly used templates of alarm subsystem types; at the same time, templates can be quickly customized according to the special circumstances of the user system. For description templates, for example:

[0090] Example 1.1:

[0091] Type: HTTP service

[0092] Application: Ngix

[0093] Time: 2023-11-10 23:59:31

[0094] URL: http: / / malldemo.com / order

[0095] HTTP error code: 500

[0096] Other additional information:…

[0097] Example 1.2:

[0098] Type: Database Service

[0099] Application: MySQL

[0100] Time: 2023-11-10 23:59:22

[0101] Error code: 1002 SQLSTATE: HY000 (ER_NO)

[0102] Statement: update ...

[0103] Example 1.3:

[0104] Type: Network

[0105] Time: 2023-11-10 23:59:21

[0106] Node information: 172.16.235.22

[0107] Error message: High Retransmission

[0108] The original alarm text generation module inputs the original alarm information of each alarm subsystem, selects the corresponding description template from the template library in the above step according to the different alarm subsystem types, fills in the template and generates a natural language description of the alarm.

[0109] Reference Figure 3 As shown, the alarm explanation model and comprehensive reasoning model screening device are operated. The device can automatically target the natural language alarm information and reasoning scenario information of different alarm subsystems, and then screen out the most suitable large language model for each. The device has an "alarm explanation model" selection module and a "alarm comprehensive reasoning model" selection module.

[0110] The input of the "alarm interpretation model" selection module is the natural language description of the alarm information obtained by the "alarm subsystem alarm access device" in the above steps, and the output is the name of the most suitable large language model for explaining each alarm type. The specific workflow is as follows:

[0111] Collect the natural language description of the alarm and construct an alarm language description set, recorded as AlertDescSet. The elements in the set are, for example, Example 1.1, Example 1.2, and Example 1.3. Add command prompts to the natural language description text of the alarm and construct an alarm analysis instruction set, recorded as AlertInstructSet. The elements in the alarm analysis instruction set are "analysis instructions", for example:

[0112] Example 2.1:

[0113] The following is a warning message, please analyze it in simple language

[0114] Type: HTTP service

[0115] Application: Ngix

[0116] Time: 2023-11-10 23:59:31

[0117] URL: http: / / malldemo.com / order

[0118] HTTP error code: 500

[0119] Other additional information:…

[0120] For all candidate large language models, input each element in AlertDescSet one by one and calculate the corresponding Perplexity. Perplexity represents perplexity, which is an indicator used to evaluate the quality of models in the field of large language models. The larger the indicator, the worse the model (the greater the distance from the real human language), and the larger the indicator, the better the model (the smaller the distance from the real human language).

[0121] For the same "alarm type, model name" Perplexity, the average value is calculated to obtain the AlertModelPerplexity set. AlertModelPerplexity represents the alarm model perplexity set. Each element in the set is a triple of (alarm type, model name, average Perlexity). Among the elements of the set, the smaller the average perplexity, the better the model and the corresponding type of alarm description information.

[0122] For example (each line is an element in the collection):

[0123] (HTTP service, Model 1, 35.3)

[0124] (HTTP service, Model 2, 21.2)

[0125] (SQL service, model 1, 15)

[0126] (SQL service, model 2, 27)

[0127] For all candidate large language models, input each element in AlertInstructSet one by one, record the return value, and obtain the AlertModelResponse set. Each element in the set is a triple (analysis instruction, model name, alarm explanation description), where the analysis instruction contains the alarm that needs to be described.

[0128] For example (each line is an element in the collection):

[0129] (Instruction 1, Model 1, the error is an HTTP error, error code 500 indicates an internal server error...)

[0130] (Instruction 1, Model 2, this error is an HTTP error, HTTP errors can be caused by a variety of reasons...)

[0131] (Instruction 2, Model 1, the error is a database error, the database instance is Mysql, there is a problem in the sql statement...)

[0132] (Instruction 2, Model 2, this error Mysql server alerts, from its error code we can know...)

[0133] The "instruction" and "return statement" of each element in AlertModelResponse are combined to form a "scoring instruction", requiring all candidate models to give a rationality score and give an average to obtain the AlertModelRespScore set; AlertModelRespScore represents the scoring set of the alarm information model description. Each element of the set is a triple (alarm type, model name, and average rationality score), where the third item is a score for the rationality of the information described by the model, with a direct range of 0 to 10 points, and 10 points being the most reasonable.

[0134] Examples of “scoring instructions” are:

[0135] "Are the following warning information and conclusions correct? Please give a score between 0-10, with 10 being the most reasonable.

[0136] Type: HTTP service

[0137] Application: Ngix

[0138] Time: 2023-11-10 23:59:31

[0139] URL: http: / / malldemo.com / order

[0140] HTTP error code: 500

[0141] Other additional information:…

[0142] Conclusion: This error is an HTTP error, and error code 500 indicates an internal server error..."

[0143] For example, the elements in AlertModelRespScore are (one element per line):

[0144] (HTTP service, model 1, 3.33)

[0145] (HTTP service, model 2, 8)

[0146] (SQL Service, Model 1, 7)

[0147] (SQL Service, Model 2, 2)

[0148] Calculate the average of "average perplexity" and "average rationality score" for the records with the same "alarm type" and "model name" in AlertModelPerplexity and AlertModelRespScore. The "alarm type" and "model name" corresponding to the element with the highest average among the elements with the same "alarm type" and "model name" is the model that best suits the alarm type. This set is called AlertSelectedModelSet, where each element is (one element per line):

[0149] (HTTP service, model 3)

[0150] (SQL Service, Model 7)

[0151] The input of the "alarm comprehensive reasoning model" selection module is the AlertDescSet and AlertSelectedModelSet obtained above, and the output is the name of the most suitable large language model for alarm comprehensive reasoning. The specific workflow is as follows:

[0152] Input each alarm in AlertDescSet into the model that best suits the alarm (the most suitable alarm model can be obtained by querying AlertSelectedModelSet). The model will output the corresponding description. After combining the descriptions, we get the SelectedModelRespSet set. Each element in the set is a tuple (alarm time, alarm explanation description), where the alarm explanation description is the explanation description given by the "preferred model" of the alarm type.

[0153] For example (each line represents one element):

[0154] (2023-11-10 23:58:22, this error is an HTTP error, error code 500 indicates an internal server error...)

[0155] (2023-11-10 23:58:24, this error is a Mysql warning, in the sql statement...)

[0156] …

[0157] Next, we construct a "diagnosis description set": we cluster the elements in SelectedModelRespSet based on the occurrence time using the DBSCAN algorithm, cluster the alarms that occur within a specific time interval (such as 5 seconds) into a cluster, combine the elements in the same cluster and add the "task description" to obtain the "reasoning instructions" of the alarm comprehensive reasoning model. The following is an input example of the elements in a cluster (the first line is the "task description"):

[0158] Please infer the possible root cause of the following alarm information:

[0159] 2023-11-10 23:58:22, This error is an HTTP error. Error code 500 indicates an internal server error...

[0160] 2023-11-10 23:58:24, Mysql warning of this error, in the sql statement...

[0161] 2023-11-10 23:58:25, this error is a warning for cache Redis...

[0162] …”

[0163] A "diagnostic description set" is composed of multiple input instructions and is recorded as DiagnoseDescSet.

[0164] For all candidate large language models, input each element in DiagnoseDescSet one by one, record the return value, and obtain the DiagnoseModelResp set. Each element in the set is a triple (inference instruction, model name, inference conclusion), for example (each line is an element in the set):

[0165] (Inference instruction 1, model 1, including multiple alarms for HTTP, SQL, and Redis, among which HTTP received the alarm first...)

[0166] (Inference instruction 2, model 1, multiple warnings may be caused by incorrect SQL statements...)

[0167] (Inference instruction 1, model 2, each warning is explained below, followed by analysis...)

[0168] (Inference instruction 2, model 2, Redis exception may be the root cause of all exceptions, analysis is as follows...)

[0169] Combine the "reasoning instruction" and "reasoning conclusion" of each element in DiagnoseModelResp to form a "reasoning scoring instruction", requiring all candidate models to give a reasonableness score and give an average, and obtain the DiagnoseModelRespScore set, where each element is a triple (reasoning instruction, model name, reasoning reasonableness score). For example, a "reasoning scoring instruction":

[0170] “Below is the anomaly description and the root cause reasoning of the model (Model 1). Please give a reasonable score on a scale of 0-10, with higher scores being more reasonable.

[0171] Exception description:

[0172] 2023-11-10 23:58:22, This error is an HTTP error. Error code 500 indicates an internal server error...

[0173] 2023-11-10 23:58:24, Mysql warning of this error, in the sql statement...

[0174] 2023-11-10 23:58:25, this error is a warning for cache Redis...

[0175] Root cause reasoning:

[0176] Contains multiple alarms for HTTP, SQL, and Redis, among which HTTP received the alarm first..."

[0177] Elements in DiagnoseModelRespScore, for example (one element per line):

[0178] (Inference instruction 1, model 1, 7.23)

[0179] (Inference instruction 1, model 2, 5.35)

[0180] (Inference instruction 2, model 1, 2.21)

[0181] …

[0182] The average of the "reasoning rationality score" of the records with the same "model name" in DiagnoseModelRespScore is calculated, and the model with the highest average is selected as the "alarm comprehensive reasoning model".

[0183] Reference Figure 4 As shown, the alarm interpretation model fusion device is run. The input of the device is the large language models of multiple different alarm subsystems output above, and the interpretation capabilities of these large language models for different alarm sources are integrated to output a fused alarm interpretation large language model.

[0184] The specific steps are as follows:

[0185] Construct a unified vocabulary of multiple "alarm explanation models". If "model 1" is selected as the explanation model of "subsystem 1", the vocabulary is Vocub1; if "model 2" is selected as the explanation model of "subsystem 2", the vocabulary is Vocub2; ...; if "model n" is selected as the explanation model of "subsystem n", the vocabulary is Vocubn. Then the unified vocabulary is:

[0186] VocubCommon = Sort(Vocub1 ∪Vocub2…∪Vocubn)

[0187] Then train the fusion model:

[0188] Build and initialize an untrained fusion model, denoted as AlertCommonInitModel. The architecture of this model conforms to the decoder part of the transformer. Sort the words in VocubCommon (a unified vocabulary of multiple alarm interpretation models) and encode them from 0. The ordered sequence of words is used as the vocabulary of AlertCommonInitModel.

[0189] Prepare training samples.

[0190] According to the relationship between the alarm type and the model in the AlertSelectedModelSet set, select the most suitable model for each alarm instruction in the AlertInstructSet set (recorded as AlertModel_i), and calculate the weight vector of the instruction (recorded as AlertWeightVec_i) through the model. Each dimension of the weight vector AlertWeightVec_i corresponds to a value in the vocabulary of AlertModel_i. The numerical order in AlertWeightVec_i is adjusted according to the order of VocubCommon. Since the length of VocubCommon is greater than or equal to Vocub_i, the missing part of the vector is filled with 0. The new vector is recorded as AlertCommonWeightVec_i, which is called "alarm explanation description vector". According to the above steps, the vector of each data in AlertInstructSet is generated, and the training set is constructed as TrainSet, in which each element is a tuple (analysis instruction, alarm explanation description vector).

[0191] Train the fusion model:

[0192] Send the "analysis instruction" part of each sample in TrainSet to AlertCommonInitModel to obtain pred_vec, calculate the cross entropy between pred_vec and the sample's "alarm explanation description vector" as the loss, and use the Adam algorithm to update the AlertCommonInitModel model. The model obtained after training is recorded as AlertCommonModel.

[0193] Among them, pred_vec is the prediction vector, which represents the vector composed of the probability distribution of the next word considered by the model. Each dimension in the vector corresponds to a word, and its value represents the probability that the next word is that word.

[0194] Reference Figure 5As shown, the alarm analysis device is operated. The input of the device is the alarm signal of each alarm subsystem, and then the high-tech signal is interpreted, the root cause of the alarm is comprehensively inferred, and the analysis result is output.

[0195] The specific steps are as follows:

[0196] Convert the alarm into an "alarm explanation description". Receive the alarm information and input it into the "fusion model", i.e. AlertCommonModel, to obtain the alarm explanation description.

[0197] The "alarm explanation description" within a specific time window (such as within 5 consecutive seconds) is constructed as an "inference instruction". For example, the reasoning instruction is:

[0198] "The following is a description of the anomaly. Please provide the root cause reasoning.

[0199] Exception description:

[0200] 2023-11-10 23:58:22, This error is an HTTP error. Error code 500 indicates an internal server error...

[0201] 2023-11-10 23:58:24, Mysql warning of this error, in the sql statement...

[0202] 2023-11-10 23:58:25, this error is a warning for cache Redis..."

[0203] Input the reasoning instruction into the "alarm comprehensive reasoning model", and the output is the alarm analysis information. Analysis information is as follows:

[0204] "According to the alarm description, multiple types of alarms, including HTTP, Redis and SQL, occurred in a short period of time. Among them, the SQL alarm occurred earlier, the Redis alarm was that the key value was not hit, and the HTTP alarm occurred last. It is inferred that the Redis key value was not loaded accurately due to the SQL exception, which finally caused the HTTP alarm..."

[0205] The present invention also provides an alarm analysis system based on a fused large language model. The alarm analysis system based on a fused large language model can be implemented by executing the process steps of the alarm analysis method based on a fused large language model, that is, those skilled in the art can understand the alarm analysis method based on a fused large language model as a preferred implementation of the alarm analysis system based on a fused large language model.

[0206] Specifically, an alarm analysis system based on a fusion large language model includes:

[0207] Module M1: connect each alarm subsystem to the corresponding device, collect and process relevant information;

[0208] The device includes an alarm subsystem alarm access device, an alarm interpretation model and a comprehensive reasoning model screening device, an alarm interpretation model fusion device and an alarm analysis and judgment device;

[0209] Module M2: Filter out the large language models that meet the preset conditions and then fuse them;

[0210] Module M3: Outputs the analysis and judgment results based on the fused large language model.

[0211] The module M1 includes inputting the original alarm information of the subsystem and outputting the natural language alarm information.

[0212] The module M1 includes the following submodules:

[0213] Module M1.1: The alarm information description template library module stores templates for alarm information of different types of alarm subsystems; the alarm information description template library module can customize templates according to the operation of the user system;

[0214] Module M1.2: The original alarm text generation module inputs the original alarm information of each alarm subsystem, selects the corresponding description template from the template library according to the different alarm subsystem types, and generates a natural language description of the alarm after filling in the template.

[0215] The module M2 includes the following submodules:

[0216] Module M2.1: Based on the natural language alarm information and reasoning scenario information of different alarm subsystems, select the large language models that meet the preset conditions;

[0217] Module M2.2: Input the large language model that meets the preset conditions, integrate the interpretation capabilities of all large language models for different alarm sources, and output a fused alarm interpretation large language model.

[0218] The module M2.1 screens a large language model that meets preset conditions based on the alarm interpretation model selection module and the alarm comprehensive reasoning model selection module; the input of the alarm interpretation model selection module includes the natural language description of the alarm information, and the output includes the name of the large language model used to explain each alarm type; the input of the alarm comprehensive reasoning model selection module includes the alarm language description set and AlertSelectedModelSet, and the output includes the name of the large language model for alarm comprehensive reasoning.

[0219] The module M2.1 includes the following submodules:

[0220] Module M2.1.1: Collect alarm natural language descriptions and construct an alarm language description set; add command prompts to the alarm natural language description text, wherein the command prompts include the alarm type required to be output by the model; construct an alarm analysis instruction set; each element in the alarm analysis instruction set is an analysis instruction;

[0221] Module M2.1.2: Input each element in the alarm language description set to all candidate large language models one by one, and calculate the corresponding Perplexity; based on the model output content, extract the alarm type, aggregate the Perplexity of the model output by the alarm type and model name, and calculate the average Perplexity to obtain the AlertModelPerplexity set; each element in the AlertModelPerplexity set is a triple consisting of "alarm type, model name, average perplexity";

[0222] Module M2.1.3: Input each element in the alarm analysis instruction set to all candidate large language models one by one, the element includes the analysis instruction, records the return value, and obtains the AlertModelResponse set; each element in the AlertModelResponse set is a triple consisting of "analysis instruction, model name, alarm explanation description";

[0223] Module M2.1.4: Combine the instruction of each element in AlertModelResponse with the return statement part to form a scoring instruction, so that all candidate models give a rationality score and an average value, and obtain an AlertModelRespScore set; each element in the AlertModelRespScore set is a triple of alarm type, model name and average rationality score;

[0224] Module M2.1.5: Calculate the average of "1 / average perplexity" and "average rationality score" for the records with the same alarm type and model name in AlertModelPerplexity and AlertModelRespScore. This average is called the "comprehensive rationality score" of the model. For a certain alarm type, the model corresponding to the record with the highest "comprehensive rationality score" is called the preferred model of the alarm type. Each alarm type and its preferred model constitute a tuple. Multiple tuples constitute the alarm type preferred model set, which is recorded as AlertSelectedModelSet.

[0225] Module M2.1.6: Input each alarm in the alarm language description set into the model that best suits the alarm and output the corresponding description, and combine the descriptions to obtain the SelectedModelRespSet set; each element in the SelectedModelRespSet set is a tuple of the alarm time and the alarm explanation description;

[0226] Module M2.1.7: Construct a diagnostic description set, use the DBSCAN algorithm to cluster the elements in SelectedModelRespSet based on the occurrence time, cluster the alarms that occur within the preset time into a cluster, combine the elements in the same cluster and add the task description to obtain the reasoning instructions of the alarm comprehensive reasoning model;

[0227] Module M2.1.8: Input each element in the diagnosis description set into all candidate large language models one by one, record the return value, and obtain the DiagnoseModelResp set; each element in the DiagnoseModelResp set is a triple of reasoning instruction, model name and reasoning conclusion;

[0228] Module M2.1.9: Combine the reasoning instruction of each element in DiagnoseModelResp with the reasoning conclusion part to form a reasoning scoring instruction, so that all candidate models give a rationality score and give an average, and obtain a DiagnoseModelRespScore set; each element in the DiagnoseModelRespScore set is a triple of reasoning instruction, model name and reasoning rationality score;

[0229] Module M2.1.10: Calculate the mean of the reasoning rationality scores for the records with the same model name in DiagnoseModelRespScore, and select the model with the highest mean as the alarm comprehensive reasoning model.

[0230] The module M2.2 includes the following submodules:

[0231] Module M2.2.1: Construct a unified vocabulary for multiple alarm interpretation models;

[0232] Module M2.2.2: Build and initialize an untrained fusion model; the architecture of the model conforms to the decoder part of the transformer; sort the words in VocubCommon and start encoding from 0, and use the ordered sequence of words obtained as the vocabulary of the model;

[0233] Module M2.2.3: Prepare training samples, select the most suitable model for each alarm instruction in the alarm analysis instruction set according to the relationship between the alarm type and the model in the AlertSelectedModelSet set, denoted as AlertModel_i, and calculate the weight vector of the instruction through the model, denoted as AlertWeightVec_i; each dimension of the weight vector AlertWeightVec_i corresponds to a value in the vocabulary of AlertModel_i, adjust the numerical order in AlertWeightVec_i according to the order of VocubCommon, fill the missing part of the vector with 0, and the new vector is denoted as AlertCommonWeightVec_i, which is called the alarm explanation description vector; generate a vector for each data in the alarm analysis instruction set, and construct a training set as TrainSet; each element in the training set is a tuple of an analysis instruction and an alarm explanation description vector;

[0234] Module M2.2.4: training fusion model; the training includes sending the analysis instruction part of each sample in TrainSet into the initialized untrained fusion model to obtain pred_vec, calculating the cross entropy of pred_vec and the alarm explanation description vector of the sample as the loss, and updating the initialized untrained fusion model based on the Adam algorithm to obtain the trained model AlertCommonModel.

[0235] The module M3 includes inputting the alarm signals of each alarm subsystem, interpreting the high-tech signals and comprehensively reasoning the alarm causes, and outputting the analysis results.

[0236] The module M3 includes the following submodules:

[0237] Module M3.1: converting alarms into alarm explanation descriptions; receiving alarm information and inputting it into a fusion model to obtain alarm explanation descriptions; the fusion model is a trained model;

[0238] Module M3.2: Construct the alarm explanation description within the preset time window into reasoning instructions;

[0239] Module M3.3: Input the reasoning instructions into the alarm comprehensive reasoning model and use the output as alarm analysis information.

[0240] Those skilled in the art know that, in addition to realizing the system and its various devices, modules, and units provided by the present invention in a purely computer-readable program code, it is entirely possible to realize the same functions in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered as a hardware component, and the devices, modules, and units included therein for realizing various functions can also be regarded as structures within the hardware component; the devices, modules, and units for realizing various functions can also be regarded as both software modules for realizing the method and structures within the hardware component.

[0241] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. An alarm analysis method based on a fusion large language model, characterized in that: include: Step S1: Connect each alarm subsystem to the corresponding device to collect and process relevant information; The device includes an alarm subsystem alarm access device, an alarm interpretation model and a comprehensive reasoning model screening device, an alarm interpretation model fusion device and an alarm analysis and judgment device; Step S2: select large language models that meet preset conditions and then fuse them; Step S2.1: Based on the natural language alarm information and reasoning scenario information of different alarm subsystems, select the large language models that meet the preset conditions; based on the alarm interpretation model selection module and the alarm comprehensive reasoning model selection module, select the large language model that meets the preset conditions; the input of the alarm interpretation model selection module includes the natural language description of the alarm information, and the output includes the name of the large language model used to explain each alarm type; the input of the alarm comprehensive reasoning model selection module includes the alarm language description set and AlertSelectedModelSet, and the output includes the name of the large language model of alarm comprehensive reasoning; Step S2.1.1: Collect alarm natural language descriptions and construct an alarm language description set; add command prompts to the alarm natural language description text, wherein the command prompts include requiring the model to output an alarm type; construct an alarm analysis instruction set; each element in the alarm analysis instruction set is an analysis instruction; Step S2.1.2: Input each element in the alarm language description set to all candidate large language models one by one, and calculate the corresponding Perplexity; based on the model output content, extract the alarm type, aggregate the Perplexity of the model output by the alarm type and model name, and calculate the average Perplexity to obtain the AlertModelPerplexity set; each element in the AlertModelPerplexity set is a triple consisting of "alarm type, model name, average perplexity"; Step S2.1.3: Input each element in the alarm analysis instruction set to all candidate large language models one by one, the element including the analysis instruction, record the return value, and obtain the AlertModelResponse set; each element in the AlertModelResponse set is a triple consisting of "analysis instruction, model name, alarm explanation description"; Step S2.1.4: Combine the instruction of each element in AlertModelResponse with the return statement part to form a scoring instruction, so that all candidate models give rationality scores and averages, and obtain an AlertModelRespScore set; each element in the AlertModelRespScore set is a triple of alarm type, model name and rationality average score; Step S2.1.5: Calculate the average of "1 / average perplexity" and "average rationality score" for the records with the same alarm type and model name in AlertModelPerplexity and AlertModelRespScore. This average is called the "comprehensive rationality score" of the model. For a certain alarm type, the model corresponding to the record with the highest "comprehensive rationality score" is called the preferred model of the alarm type. Each alarm type and its preferred model constitute a tuple. Multiple tuples constitute an alarm type preferred model set, which is recorded as AlertSelectedModelSet. Step S2.1.6: Input each alarm in the alarm language description set into the model that best suits the alarm and output the corresponding description, and combine the descriptions to obtain a SelectedModelRespSet set; each element in the SelectedModelRespSet set is a tuple of an alarm time and an alarm explanation description; Step S2.1.7: Construct a diagnosis description set, use the DBSCAN algorithm to cluster the elements in SelectedModelRespSet with the occurrence time as the feature, cluster the alarms that occur within the preset time into a cluster, combine the elements in the same cluster and add the task description to obtain the reasoning instructions of the alarm comprehensive reasoning model; Step S2.1.8: Input each element in the diagnosis description set into all candidate large language models one by one, record the return value, and obtain the DiagnoseModelResp set; each element in the DiagnoseModelResp set is a triple of reasoning instruction, model name and reasoning conclusion; Step S2.1.9: Combine the reasoning instruction of each element in DiagnoseModelResp with the reasoning conclusion part to form a reasoning scoring instruction, so that all candidate models give a rationality score and give an average, and obtain a DiagnoseModelRespScore set; each element in the DiagnoseModelRespScore set is a triple of reasoning instruction, model name and reasoning rationality score; Step S2.1.10: Calculate the mean of the reasoning rationality scores of the records with the same model name in DiagnoseModelRespScore, and select the model with the highest mean as the alarm comprehensive reasoning model; Step S2.2: input a large language model that meets the preset conditions, integrate the interpretation capabilities of all large language models for different alarm sources, and output a fused alarm interpretation large language model; Step S3: Output the analysis result based on the fused large language model.

2. According to claim 1, the alarm analysis method based on the fusion of large language models is characterized in that: The step S1 includes inputting original alarm information of the subsystem and outputting natural language alarm information.

3. The alarm analysis method based on the fusion of large language models according to claim 2 is characterized in that: The step S1 comprises the following sub-steps: Step S1.1: The alarm information description template library module stores templates for alarm information of different types of alarm subsystems; the alarm information description template library module can customize templates according to the operation of the user system; Step S1.2: The original alarm text generation module inputs the original alarm information of each alarm subsystem, selects the corresponding description template from the template library according to different alarm subsystem types, fills in the template and generates a natural language description of the alarm.

4. The alarm analysis method based on the fusion of large language models according to claim 1 is characterized in that: The step S2.2 includes the following sub-steps: Step S2.2.1: construct a unified vocabulary of multiple alarm interpretation models; Step S2.2.2: construct and initialize an untrained fusion model; the architecture of the model conforms to the decoder part of the transformer; sort the words in VocubCommon and start encoding from 0, and use the ordered sequence of words obtained as the vocabulary of the model; Step S2.2.3: prepare training samples, select the most suitable model for each alarm instruction in the alarm analysis instruction set according to the relationship between the alarm type and the model in the AlertSelectedModelSet set, denoted as AlertModel_i, and calculate the weight vector of the instruction through the model, denoted as AlertWeightVec_i; each dimension of the weight vector AlertWeightVec_i corresponds to a value in the vocabulary of AlertModel_i, adjust the numerical order in AlertWeightVec_i according to the order of VocubCommon, fill the missing part of the vector with 0, and the new vector is denoted as AlertCommonWeightVec_i, which is called the alarm explanation description vector; generate a vector for each data in the alarm analysis instruction set, and construct a training set as TrainSet; each element in the training set is a tuple of an analysis instruction and an alarm explanation description vector; Step S2.2.4: training the fusion model; the training includes sending the analysis instruction part of each sample in TrainSet into the initialized untrained fusion model to obtain pred_vec, calculating the cross entropy of pred_vec and the alarm explanation description vector of the sample as the loss, and updating the initialized untrained fusion model based on the Adam algorithm to obtain the trained model AlertCommonModel.

5. The alarm analysis method based on the fusion of large language models according to claim 4 is characterized in that: The step S3 includes inputting the alarm signals of each alarm subsystem, interpreting the high-information signals and comprehensively reasoning the alarm causes, and outputting the analysis and judgment results.

6. The alarm analysis method based on the fusion of large language models according to claim 5 is characterized in that: The step S3 comprises the following sub-steps: Step S3.1: converting the alarm into an alarm explanation description; receiving the alarm information and inputting it into a fusion model to obtain an alarm explanation description; the fusion model is a trained model; Step S3.2: constructing the alarm explanation description within the preset time window into reasoning instructions; Step S3.3: Input the reasoning instruction into the alarm comprehensive reasoning model and use the output as alarm analysis information.

7. An alarm analysis system based on a fusion large language model, characterized in that: include: Module M1: connect each alarm subsystem to the corresponding device, collect and process relevant information; The device includes an alarm subsystem alarm access device, an alarm interpretation model and a comprehensive reasoning model screening device, an alarm interpretation model fusion device and an alarm analysis and judgment device; Module M2: Filter out the large language models that meet the preset conditions and then fuse them; Module M2.1: Based on the natural language alarm information and reasoning scenario information of different alarm subsystems, select the large language models that meet the preset conditions; based on the alarm interpretation model selection module and the alarm comprehensive reasoning model selection module, select the large language model that meets the preset conditions; the input of the alarm interpretation model selection module includes the natural language description of the alarm information, and the output includes the name of the large language model used to explain each alarm type; the input of the alarm comprehensive reasoning model selection module includes the alarm language description set and AlertSelectedModelSet, and the output includes the name of the large language model of alarm comprehensive reasoning; Module M2.1.1: Collect alarm natural language descriptions and construct an alarm language description set; add command prompts to the alarm natural language description text, wherein the command prompts include the alarm type required to be output by the model; construct an alarm analysis instruction set; each element in the alarm analysis instruction set is an analysis instruction; Module M2.1.2: Input each element in the alarm language description set to all candidate large language models one by one, and calculate the corresponding Perplexity; based on the model output content, extract the alarm type, aggregate the Perplexity of the model output by the alarm type and model name, and calculate the average Perplexity to obtain the AlertModelPerplexity set; each element in the AlertModelPerplexity set is a triple consisting of "alarm type, model name, average perplexity"; Module M2.1.3: Input each element in the alarm analysis instruction set to all candidate large language models one by one, the element includes the analysis instruction, records the return value, and obtains the AlertModelResponse set; each element in the AlertModelResponse set is a triple consisting of "analysis instruction, model name, alarm explanation description"; Module M2.1.4: Combine the instruction of each element in AlertModelResponse with the return statement part to form a scoring instruction, so that all candidate models give a rationality score and an average value, and obtain an AlertModelRespScore set; each element in the AlertModelRespScore set is a triple of alarm type, model name and average rationality score; Module M2.1.5: Calculate the average of "1 / average perplexity" and "average rationality score" for the records with the same alarm type and model name in AlertModelPerplexity and AlertModelRespScore. This average is called the "comprehensive rationality score" of the model. For a certain alarm type, the model corresponding to the record with the highest "comprehensive rationality score" is called the preferred model of the alarm type. Each alarm type and its preferred model constitute a tuple. Multiple tuples constitute the alarm type preferred model set, which is recorded as AlertSelectedModelSet. Module M2.1.6: Input each alarm in the alarm language description set into the model that best suits the alarm and output the corresponding description, and combine the descriptions to obtain the SelectedModelRespSet set; each element in the SelectedModelRespSet set is a tuple of the alarm time and the alarm explanation description; Module M2.1.7: Construct a diagnostic description set, use the DBSCAN algorithm to cluster the elements in SelectedModelRespSet based on the occurrence time, cluster the alarms that occur within the preset time into a cluster, combine the elements in the same cluster and add the task description to obtain the reasoning instructions of the alarm comprehensive reasoning model; Module M2.1.8: Input each element in the diagnosis description set into all candidate large language models one by one, record the return value, and obtain the DiagnoseModelResp set; each element in the DiagnoseModelResp set is a triple of reasoning instruction, model name and reasoning conclusion; Module M2.1.9: Combine the reasoning instruction of each element in DiagnoseModelResp with the reasoning conclusion part to form a reasoning scoring instruction, so that all candidate models give a rationality score and give an average, and obtain a DiagnoseModelRespScore set; each element in the DiagnoseModelRespScore set is a triple of reasoning instruction, model name and reasoning rationality score; Module M2.1.10: Calculate the mean of the reasoning rationality scores for the records with the same model name in DiagnoseModelRespScore, and select the model with the highest mean as the alarm comprehensive reasoning model; Module M2.2: Input the large language model that meets the preset conditions, integrate the interpretation capabilities of all large language models for different alarm sources, and output a fused alarm interpretation large language model; Module M3: Outputs the analysis and judgment results based on the fused large language model.