Abnormal service analysis method and device and electronic equipment

Abstract

The invention discloses an abnormal service analysis method and device and electronic equipment, and the method comprises the steps: obtaining a to-be-analyzed abnormal information set, and generating a descriptive language corresponding to each piece of abnormal information in the abnormal information set; inputting the descriptive languages into a language model obtained by pre-training for scene recognition, and determining first scene information corresponding to each descriptive language; determining a corresponding scene log demand according to the first scene information; and obtaining log information from the corresponding service platform according to the scene log demand, and determining an exception analysis result of the exception information in the corresponding service scene according to the log information. According to the embodiment of the invention, problem analysis under multiple scenes can be realized, and meanwhile, an abnormal analysis result corresponding to a specific scene can be accurately given, so that analysis disorder caused by non-correspondence of models utilized by different scene problems is avoided, the analysis efficiency and applicability are improved, and the method can be widely applied to the technical field of computers.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to an abnormal business analysis method, device and electronic equipment. Background Art

[0002] At present, more and more business platforms can provide rich business services, for example: social business platforms can provide social business services; e-commerce business platforms can provide e-commerce business services, and so on. However, when using business platforms in different business scenarios, objects will encounter different platform problems, and these problems in different business platforms are usually cross-domain. It is difficult to cover all business platforms through a common log analysis model to achieve abnormal analysis and answers to all platform problems. Therefore, related technologies usually require manual collection and classification of related problems, and then subsequent identification based on different log analysis models trained on different business platforms, resulting in low efficiency and a small scope of application and poor versatility of a single model. Summary of the invention

[0003] In view of this, the embodiments of the present invention provide an abnormal business analysis method, device and electronic device, which can realize problem analysis in multiple scenarios and improve analysis efficiency and applicability.

[0004] An aspect of an embodiment of the present invention provides a method for analyzing abnormal services, including:

[0005] Obtaining a set of abnormal information to be analyzed, and generating descriptive language corresponding to each abnormal information in the abnormal information set;

[0006] Inputting the descriptive language into the pre-trained language model for scene recognition, and determining the first scene information corresponding to each descriptive language;

[0007] Determine a corresponding scene log requirement according to the first scene information;

[0008] Obtain log information from the corresponding business platform according to the scenario log requirements, and determine the exception analysis results of the exception information in the corresponding business scenario based on the log information.

[0009] Another aspect of an embodiment of the present invention provides an abnormal service analysis device, including:

[0010] The first module is used to obtain a set of abnormal information to be analyzed and generate a descriptive language corresponding to each abnormal information in the abnormal information set; wherein each abnormal information comes from a different business platform;

[0011] The second module is used to input the descriptive language into the pre-trained language model for scene recognition, and determine the first scene information corresponding to each descriptive language;

[0012] The third module is used to determine the corresponding scene log requirement according to the first scene information;

[0013] The fourth module is used to obtain log information from the corresponding business platform according to the scenario log requirements, and determine the abnormality analysis results of the abnormal information in the corresponding business scenario according to the log information.

[0014] Optionally, the first module is specifically used for:

[0015] In response to the problem feedback request of each business platform, trigger the network hook corresponding to the problem feedback request to obtain abnormal information from each business platform to which the problem feedback request belongs through the network hook; wherein the abnormal information is used to represent the real-time data when the abnormal problem occurs in each business platform;

[0016] Collect historical solutions of various business platforms in solving platform problems;

[0017] Generate descriptive language for exception information and historical resolution.

[0018] Optionally, the first module is further used for:

[0019] Extract key information from abnormal information obtained from various business platforms;

[0020] Calculate the number of abnormal business platforms where the same key information appears at the same time;

[0021] When the number of abnormal platforms meets the threshold condition of aggregate feedback, the abnormal information of each business platform is aggregated to obtain aggregate feedback information.

[0022] Optionally, the second module is specifically used for:

[0023] Mapping and associating the selected multiple scene names with the scene description words corresponding to each scene name to construct a scene mapping table;

[0024] Using the language model, each descriptive language in the descriptive language set is identified and converted to obtain keywords of the descriptive language;

[0025] According to the degree of association between the keyword and each scene description word in the scene mapping table, one or more candidate scene names corresponding to each descriptive language are determined as the first scene information.

[0026] Optionally, the second module is specifically used for:

[0027] Collect experience documents used to solve business anomalies and perform data cleaning on the experience documents;

[0028] Input the cleaned experience documents into the language model to obtain a question-answering dataset that meets the preset format requirements;

[0029] Filter and format the question-answering dataset to build a pre-training dataset;

[0030] Adjust the parameters of the language model according to the pre-training dataset and add information of the experience document to the language model.

[0031] Optionally, the third module is specifically used for:

[0032] Determine one or more scene identifiers from a scene mapping table according to the first scene information;

[0033] Summarizing the scene log requirements corresponding to the scene identifiers determined from the first scene information;

[0034] Among them, scenario log requirements are used to characterize the log information to be collected by each business platform.

[0035] Optionally, the fourth module is specifically used for:

[0036] Obtain log information from the corresponding business platform according to the scenario log requirements; the log information includes the log collection timestamp, business object account and abnormal log segment;

[0037] Perform regular matching on abnormal log segments of log information according to the preset analysis script to determine the target log line;

[0038] Perform exception analysis on the target log line based on the empirical data of the exception analysis library to determine the exception analysis results for the corresponding business scenario.

[0039] Optionally, the fourth module is further used for:

[0040] In general scenarios, extract information from the target log line to obtain the business platform version information and business object account to determine the abnormal analysis results of the business object in general scenarios;

[0041] In the audio and video usage scenario, extract information from the target log line to obtain the interface operation information and audio device fault information of the business object to determine the abnormal analysis results of the business object in the audio and video usage scenario;

[0042] In the shared screen viewing scenario, information is extracted from the target log line to obtain shared operation information and virtual location information of the business object to determine the abnormal analysis result of the business object in the shared screen viewing scenario.

[0043] Optionally, the abnormal service analysis device further includes a fifth module, and the fifth module is specifically configured to:

[0044] Vectorize the descriptive language corresponding to each abnormal information in the abnormal information set to obtain a problem vector;

[0045] Vectorize the experience documents on solving exceptions to obtain an experience vector set;

[0046] Perform vector matching on the question vector according to the experience vector set, and obtain multiple experience vectors whose approximation is higher than the target threshold as context information of the question vector;

[0047] The question vector and context information are combined to generate descriptive language corresponding to the exception information.

[0048] Optionally, the above-mentioned vectorization processing of the descriptive language corresponding to each abnormal information in the abnormal information set to obtain the problem vector specifically includes:

[0049] According to a preset field set, serializing each field in the field set in the descriptive language to obtain a structured descriptive language;

[0050] Each field is processed in blocks to obtain block values ​​of multiple sub-blocks;

[0051] Determine the peak block of the current field according to the block values ​​of each sub-block, and obtain the character string of the peak block;

[0052] Extract keywords from the string of the peak block, and use the extracted keywords to compress the usage frequency of each word in the string to obtain a normalized string;

[0053] The structured descriptive language is updated according to the normalized character string;

[0054] After vectorizing the updated structured descriptive language, a problem vector is obtained.

[0055] Optionally, the abnormal service analysis device further includes a sixth module, and the sixth module is specifically configured to:

[0056] Structural decomposition of the abnormal analysis results is performed to determine multiple information components in the abnormal analysis results;

[0057] Map each information component to obtain the descriptive language corresponding to the abnormal analysis result.

[0058] Optionally, the abnormal service analysis device further includes a seventh module, and the seventh module is specifically used for:

[0059] Obtain exception analysis results in multiple business scenarios in parallel;

[0060] According to the timestamp information of each anomaly analysis result, each anomaly analysis result is sorted and summarized, and according to the business scenario of each anomaly analysis result, multiple anomaly analysis result sets are constructed; wherein each anomaly analysis result in each anomaly analysis result set is in the same business scenario.

[0061] Optionally, the fourth module is further used to:

[0062] Match the scenario log requirements with the candidate input templates cached in the preset cache rule library to determine the target input template;

[0063] According to the target input template, historical cache information corresponding to the descriptive language is obtained from the preset cache result library;

[0064] Use the historical cache information as the request result of the scene log requirement, and output the request result as log information.

[0065] Optionally, a seventh module is further included, for:

[0066] Get the template to be updated;

[0067] A tag list is obtained by traversing the template to be updated; wherein each tag in the tag list is used to represent the

[0068] The field type of the scene recognition result;

[0069] Generate a key name of a label template vector according to the label list;

[0070] According to the key name, multiple values ​​corresponding to the key name are obtained from the template to be updated to construct multiple key-value pairs; wherein the key of the key-value pair is used to represent the label of the scene recognition result, and the value of the key-value pair is used to represent the field information corresponding to the label;

[0071] Update each of the key-value pairs to the scene label template library.

[0072] Another aspect of an embodiment of the present invention provides an electronic device, including a processor and a memory;

[0073] The memory is used to store programs;

[0074] The processor executes the program to implement the above-mentioned abnormal business analysis method.

[0075] Another aspect of an embodiment of the present invention provides a computer-readable storage medium, wherein the storage medium stores a program, and the program is executed by a processor to implement the above-mentioned abnormal business analysis method.

[0076] The embodiment of the present invention also discloses a computer program product or a computer program, which includes a computer instruction stored in a computer-readable storage medium. A processor of a computer device can read the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device executes the above-mentioned abnormal business analysis method.

[0077] The embodiment of the present invention first obtains a set of abnormal information to be analyzed, and generates a descriptive language corresponding to each abnormal information in the abnormal information set; then the descriptive language is input into a pre-trained language model for scene recognition, and the first scene information corresponding to each descriptive language is determined, and then the corresponding scene log requirement is determined according to the first scene information. Finally, log information is obtained from the corresponding business platform according to the scene log requirement, and the abnormal analysis result of the abnormal information in the corresponding business scene is determined according to the log information; the embodiment of the present invention uses a pre-trained language model to perform scene recognition on the abnormal information sent in different business scenes, and obtains relevant log content based on the recognized scene. By analyzing the relevant log lines, accurate analysis of various problems is achieved, and a language model can be used to achieve cross-scene abnormal analysis. At the same time, the abnormal analysis result corresponding to the specific scene can be accurately given, avoiding analysis confusion caused by the inconsistency of the models used for different scene problems, thereby improving analysis efficiency and applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0079] Figure 1 A schematic diagram of an implementation environment of an abnormal business analysis method provided by an embodiment of the present invention;

[0080] Figure 2 A schematic diagram of a process for abnormal business analysis provided by an embodiment of the present invention;

[0081] Figure 3 A schematic diagram of a process for obtaining an abnormal information set provided by an embodiment of the present invention;

[0082] Figure 4 A flow chart of language model tuning provided by an embodiment of the present invention;

[0083] Figure 5 A schematic diagram of a multi-scenario script analysis process provided by an embodiment of the present invention;

[0084] Figure 6 A schematic diagram of a process for generating an experience vector set provided by an embodiment of the present invention;

[0085] Figure 7 A schematic diagram of a flow chart of recognition conversion of a language model provided by an embodiment of the present invention;

[0086] Figure 8 A flowchart of another abnormal business analysis method provided by an embodiment of the present invention;

[0087] Fig. 9 A modular schematic diagram of another abnormal service analysis device provided by an embodiment of the present invention;

[0088] Fig.10 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention;

[0089] Fig.11 A schematic diagram of a computer system structure block diagram of an electronic device implementing an embodiment of the present application is provided in accordance with an embodiment of the present invention. DETAILED DESCRIPTION

[0090] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0091] It should be understood that in the description of the embodiments of the present invention, the meaning of multiple (or multiple) is more than two, greater than, less than, and exceeding are understood to exclude the number itself, and above, below, and within are understood to include the number itself. If there is a description of "first", "second", etc., it is only used for the purpose of distinguishing technical features, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features.

[0092] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present invention. The appearance of the phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0093] Before further describing the embodiments of the present invention in detail, the nouns and terms involved in the embodiments of the present invention are described. The nouns and terms involved in the embodiments of the present invention are applicable to the following interpretations:

[0094] Artificial Intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines so that machines have the functions of perception, reasoning and decision-making.

[0095] Artificial intelligence technology is a comprehensive discipline that covers a wide range of fields, including both hardware-level and software-level technologies. Basic artificial intelligence technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operating / interactive systems, and mechatronics. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning. The display device including an image acquisition component shown in the present invention mainly involves computer vision technology, machine learning / deep learning, autonomous driving, smart transportation, and other directions.

[0096] Machine Learning (ML) is a multi-disciplinary subject that involves probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It specializes in studying how computers simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are spread across all areas of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and teaching learning.

[0097] Deep learning, the concept of deep learning originated from the study of artificial neural networks. Multilayer perceptrons with multiple hidden layers are a type of deep learning structure. Deep learning combines low-level features to form more abstract high-level representations of attribute categories or features to discover distributed feature representations of data. Deep learning is a new field in machine learning research. Its motivation is to establish and simulate neural networks that analyze and learn in the human brain, and to interpret data such as images, sounds, and text by imitating the mechanisms of the human brain. Commonly used deep learning models include: Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), Deep Neural Networks (DNN), Deep Belief Networks (DBNs) and other neural networks. There are two ways for data to propagate in a neural network. One is along the path from input to output, which is called forward propagation, and the other is from output back to input, which is called backpropagation. In the forward propagation process, the input information is processed layer by layer by the neural network and transmitted to the output layer. The error between the output value and the expectation is described by the loss function, and then it enters the backward propagation. The partial derivative of the loss function with respect to the weight of each neuron is calculated layer by layer to form the weight gradient data of the loss function with respect to the weight vector. This data serves as the basis for updating the weight parameters, and the training of the neural network is completed in the process of continuously updating the weight parameters.

[0098] Natural Language Processing (NLP) is a discipline that studies language issues in human-computer interaction. The key to processing natural language is to make computers "understand" natural language, so natural language processing is also called natural language understanding (NLU), also known as computational linguistics. On the one hand, it is a branch of language information processing, and on the other hand, it is one of the core topics of artificial intelligence (AI). Based on the above theoretical foundations, as well as the research and progress of artificial intelligence technology, artificial intelligence technology has been studied and applied in many fields, such as common smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, unmanned driving, automatic driving, drones, robots, smart medical care, smart customer service, etc. It is believed that with the development of technology, artificial intelligence technology will be applied in more fields and play an increasingly important role.

[0099] At present, more and more business platforms can provide rich business services, for example: social business platforms can provide social business services; e-commerce business platforms can provide e-commerce business services, and so on. However, when using business platforms in different business scenarios, objects will encounter different platform problems, and these problems in different business platforms are usually cross-domain. It is difficult to cover all business platforms through a common log analysis model to achieve abnormal analysis and answers to all platform problems. Therefore, related technologies usually require manual collection and classification of related problems, and then subsequent identification based on different log analysis models trained on different business platforms, resulting in low efficiency and a small scope of application and poor versatility of a single model.

[0100] To this end, an embodiment of the present invention first obtains a set of abnormal information to be analyzed, and generates a descriptive language corresponding to each abnormal information in the abnormal information set; then the descriptive language is input into a pre-trained language model for scene recognition, and the first scene information corresponding to each descriptive language is determined, and then the corresponding scene log requirement is determined according to the first scene information. Finally, log information is obtained from the corresponding business platform according to the scene log requirement, and the abnormal analysis result of the abnormal information in the corresponding business scenario is determined according to the log information; the embodiment of the present invention uses a pre-trained language model to perform scene recognition on the abnormal information sent in different business scenarios, and obtains relevant log content based on the recognized scene. By analyzing the relevant log lines, accurate analysis of various problems is achieved, and a language model can be used to achieve cross-scenario abnormal analysis. At the same time, the abnormal analysis results corresponding to the specific scenario can be accurately given, avoiding analysis confusion caused by the inconsistency of the models used for different scenario problems, thereby improving analysis efficiency and applicability.

[0101] It is understandable that the language model provided in the embodiment of the present invention can be applied to any computer device with data processing and computing capabilities, and this computer device can be various types of terminals or servers. When the computer device in the embodiment of the present invention is a server, the server is an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. Optionally, the terminal is a smart phone, a tablet computer, a laptop computer, a desktop computer, etc., but is not limited to this.

[0102] It should be further explained that the terminals involved in the embodiments of the present invention include but are not limited to smart phones, computers, intelligent voice interaction devices, smart home appliances, vehicle terminals, aircraft, etc. The embodiments of the present invention can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, smart transportation, assisted driving, etc.

[0103] In some possible implementations, a computer program capable of implementing the abnormal business analysis method provided in an embodiment of the present invention may be deployed and executed on a computer device, or on multiple computer devices located at one location, or on multiple computer devices distributed at multiple locations and interconnected by a communication network. Multiple computer devices distributed at multiple locations and interconnected by a communication network may constitute a blockchain system.

[0104] The embodiment of the present invention also discloses a computer program product or a computer program, wherein the computer program product or the computer program includes a computer instruction, and the computer instruction is stored in a computer-readable storage medium. A processor of a computer device can read the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device executes Figure 2 The method shown.

[0105] like Figure 1 As shown, Figure 1 Schematic diagram of an implementation environment provided by an embodiment of the present invention. Figure 1 The implementation environment includes at least one server 100 and a terminal. For example, the terminal includes terminal 201, terminal 202 and terminal 203. The server 100 is connected to multiple terminals by wired or wireless means. Each terminal is installed with applications of one or more business platforms. During the application process of the application, if an abnormal situation occurs, the terminal sends the abnormality-related information to the server 100. The abnormality-related information includes the application scenario, log data, problem type or solution corresponding to the abnormal situation, and receives the abnormality analysis results or processing methods fed back by the server 100. The server 100 stores, classifies and applies the abnormality-related information uploaded by the applications corresponding to the multiple business platforms. For example, the language model in the server 100 is trained according to the abnormality-related information uploaded by the applications corresponding to the multiple business platforms, and the unknown abnormal information is judged and analyzed according to the classified business information.

[0106] like Figure 2 , which is a flowchart of a method for analyzing abnormal business provided by an embodiment of the present invention. The execution subject of the method for analyzing abnormal business is the server 200, and includes steps S100 to S400:

[0107] S100: Acquire a set of abnormal information to be analyzed, and generate descriptive language corresponding to each abnormal information in the abnormal information set; wherein each abnormal information comes from a different business platform.

[0108] Business platform refers to a carrier that implements specific functions through software programs. For example, business platforms include but are not limited to instant messaging platforms, shopping platforms, game platforms, current news platforms, test training platforms, industry information publishing platforms or video sharing platforms. Abnormal information refers to feedback information generated when a certain set function cannot be implemented during the operation of the business platform. For example, abnormal information includes but is not limited to abnormal display of pictures on the industry information publishing platform, deterioration of video resolution on the video sharing platform, the game screen on the game platform lags behind the voice corresponding to the game screen, and the price displayed on the shopping platform is not the discounted price. The descriptive language corresponding to the abnormal information refers to the textual description of the performance or method corresponding to the abnormal information. For example, the resolution of the pictures uploaded by the industry information publishing platform is too high and cannot be updated adaptively, the video uploaded by the video sharing platform is too long, the processor or display configuration of the game platform does not meet the standards, and the shopping platform has not updated the discount method.

[0109] Specifically, the server receives the abnormal information to be analyzed sent by multiple business platforms of several terminals, forms a set of the abnormal information to be analyzed, makes the abnormal information in the same set close or similar, and generates corresponding descriptive language based on the abnormal information in the set of abnormal information to be analyzed with relevant professional knowledge. It should be noted that the abnormal information set contains the business platform identification information of the source of each abnormal information.

[0110] S200: Input the descriptive language into the pre-trained language model for scene recognition, and determine the first scene information corresponding to each descriptive language.

[0111] It should be noted that the language model has been trained before use; the input of the language model is the descriptive language, and the output of the language model is one or more scenarios that match the descriptive language to the required degree. The specific structure of the language model can adopt an existing model, and this embodiment does not make any specific restrictions.

[0112] Specifically, the server first inputs each descriptive language into a trained language model for scene recognition to obtain one or more scene information corresponding to each descriptive language, such as a scene name or a scene ID, and then aggregates the one or more scene information corresponding to each descriptive language to obtain the first scene information corresponding to each descriptive language.

[0113] The embodiment of the present invention combines a pre-trained large language model solution for problem analysis and troubleshooting, which helps to improve the contextual interactive understanding ability of problem troubleshooting. In the process of preliminary scene recognition using the large language model, based on the recognition training of various historical scenes in the pre-training process, the abnormal information contained in the current real-time problem feedback request can be quickly identified and analyzed. This process does not require the abnormal information to be compared and identified with a large amount of scene content one by one, thereby improving the recognition efficiency and reducing the dependence on a large amount of original data sets.

[0114] S300. Determine a corresponding scenario log requirement according to the first scenario information.

[0115] The first scenario information includes scenario category information determined according to a descriptive language, wherein the descriptive language is determined according to the abnormal information fed back by each business platform. The scenario log requirement is used to characterize the log information to be collected by each business platform. During the project development process, the business platform develops multiple functions according to the needs, each function has multiple scenarios, and the complete implementation of each scenario requires certain operations. The scenario log records the operations corresponding to multiple scenarios at different time points. Therefore, when the scenario category information is determined, the corresponding scenario log requirement can be determined based on the scenario category information in the first scenario information.

[0116] In a specific embodiment, the game platform is provided with an interface display module, an operation module, and a prop equipment module, etc. The interface display module is used to display the dynamic images generated during the game, the operation module is used to receive the operation instructions of the business object and generate certain actions based on the operation instructions, and the prop equipment module is provided with a variety of equipment, and different equipment has different specific functions. When the business object purchases a certain equipment, the character in the game will have the corresponding special function. The scene log demand corresponding to the interface display module is the scene log information corresponding to the screen display period, the scene log demand corresponding to the operation module is the scene log information corresponding to the business object input operation, and the scene log demand corresponding to the prop equipment module is the scene log information corresponding to the character image or skill change after the equipment is selected.

[0117] S400: Obtain log information from a corresponding business platform according to scenario log requirements, and determine an exception analysis result of the exception information in a corresponding business scenario according to the log information.

[0118] Log information is used to represent the record information of the operations or actions performed by the execution subject at multiple time points. It should be noted that log information includes log code information, which belongs to the computer program script language and has limited readability. The abnormal analysis results are used to represent abnormal problems and corresponding solutions.

[0119] Specifically, since the exception information set contains the business platform identification information of the source of each exception information, the descriptive language corresponding to each exception information is also accompanied by the business platform identification information. Similarly, the first scenario information determined according to the descriptive language and the scenario log requirements determined according to the first scenario information are both accompanied by the business platform identification information. First, determine the corresponding business platform according to the scenario log requirements, then select the log information corresponding to the scenario log requirements from the log of the corresponding business platform, and then use the log information analysis tool to analyze the corresponding log information obtained to determine the exception analysis results under the corresponding business scenario. It should be noted that the log information analysis tool can use one or more existing analysis tools, and this embodiment does not make specific restrictions.

[0120] Optionally, step S100 specifically includes steps S110 to S130:

[0121] S110. In response to the problem feedback request from each business platform, trigger the network hook corresponding to the problem feedback request to obtain abnormal information from each business platform to which the problem feedback request belongs through the network hook; wherein the abnormal information is used to represent the real-time data when abnormal problems occur in each business platform.

[0122] Web hook: A behavior that changes a web application by means of a custom callback function. Through web hooks, you can customize some behavior notifications to the specified uniform resource locator (URL). When the request data is completed, the web hook will respond to the object with the corresponding request data through the uniform resource locator set by the object. For example, when the request is updated and the request update reminder in the object settings is triggered, the system will make a post (POST) request to the callback URL customized by the object on the web hook page. Web hooks include but are not limited to the API (Application Programming Interface) given by the server or the API gateway on the cloud.

[0123] Specifically, when an abnormality occurs in the operation of the business platform on the terminal, the terminal sends a problem feedback request to the server through a preset feedback channel; the server responds to the problem feedback request and triggers the network hook corresponding to the problem feedback request to obtain real-time data from the business platform when an abnormal problem occurs on the business platform through the network hook. It should be noted that the preset feedback channel includes but is not limited to the "feedback platform" contained in the business platform itself or a third-party service channel.

[0124] S120. Collect historical solutions of various business platforms for solving platform problems.

[0125] The historical solution refers to the set of valid operations performed by the business platform from the time when the abnormal information appears to the time when the abnormal information is eliminated. The server collects the historical solutions of the business platform when solving platform problems according to the actual application, and this embodiment does not impose specific restrictions. For example, when the business platform observes the processing status of the abnormal problem, the server collects the valid operations of the business platform from the abnormal start state to the abnormal end state; or, the server directly receives the abnormal problem and its corresponding complete solution from the business platform.

[0126] S130. Generate descriptive language for exception information and historical solutions.

[0127] Exception information represents real-time data when an exception occurs on each business platform, and historical solutions represent solutions to eliminate the exception. Therefore, exception information and historical solutions may contain code programs, etc., which have limited readability. In order to improve the subsequent processing efficiency, the embodiment of the present invention generates descriptive language for exception information and historical solutions here, which improves readability. Specifically, the exception information or historical solutions in the embodiment of the present invention can be determined to obtain the corresponding descriptive language based on industry-related knowledge such as the program code script and program language interpretation of the business platform.

[0128] In some embodiments, a pre-trained language model can be used to input the abnormal information set into the language model to obtain the corresponding descriptive language in a conversational form. For example:

[0129] When the program crashes for user 12345, the developer sends a request to the language model: "Get the Android crash log for account 12345"

[0130] Language model return result: "bugly link: https: / / www.AA.com /

[0131] User 12345 crashed at 2023-08-10 17:08:37, and the log recovery command has been issued to you

[0132] The log retrieval time is 2023-08-10 16:08:37~2023-08-10 18:08:37,

[0133] After the log is uploaded successfully, you can continue the conversation on this platform and analyze the scenes you are interested in in the log. For usage, please refer to the "Common Usage Methods" of this platform.

[0134] Similarly, when other users also encounter the same program crash, they can send a corresponding request to the language model to obtain corresponding return information, and the return information is the descriptive language described in the embodiment of the present invention.

[0135] The above examples illustrate the process of finding the corresponding analysis scenario ID list through the language model when users give some questions, and then the application uses these scenario ID lists to do actual analysis and finally obtain descriptive language.

[0136] In addition, in the problem-assisted analysis strategy, analyzing the scenario is one of the steps. In addition, general solutions can be found from the general knowledge base and directly output as descriptive language.

[0137] In addition, since the sources of user feedback issues of different software are different, in order to make the language model adapt to the differences brought by different sources and achieve better module reuse, the data processing results can be processed uniformly at the back end. The main working steps are as follows:

[0138] 1) Software developers build different "user feedback service platforms" based on their own needs, or directly use third-party services to collect user feedback. Receive and store the user feedback results. On this basis, the received user feedback results are further sent to the developer's customized webhook URL.

[0139] 2) The webhook can be an API given by the backend, or an API gateway on the cloud, which ultimately forwards the request to the backend service in the present invention.

[0140] 3) The backend service processes user feedback and extracts the following user on-site information:

[0141] a) User account and contact information;

[0142] b) A detailed description of the feedback issue;

[0143] c) The specific time when the feedback problem occurred, or the time range;

[0144] d) Picture or video description of the feedback issue.

[0145] It should be noted that the data collection and processing in this application should be strictly in accordance with the requirements of relevant national laws and regulations when applied in examples, and the informed consent or separate consent of the personal information subject should be obtained, and subsequent data use and processing should be carried out within the scope of authorization of laws and regulations and the personal information subject.

[0146] Optionally, obtaining a set of abnormal information to be analyzed in step S100 specifically includes steps S140 to S160:

[0147] S140. Extract key information from the abnormal information obtained from each business platform.

[0148] Abnormal information includes but is not limited to object site information, which includes but is not limited to object account and contact information, detailed description of feedback problem, specific time of feedback problem or time range of feedback problem, picture or video description of feedback problem, etc. Key information refers to useful information for judging and handling abnormal problems, etc. Key information is determined according to actual application, and this embodiment does not make specific restrictions.

[0149] Specifically, the server extracts key information from the abnormal information obtained in the business platform according to the demand. It should be noted that this embodiment does not limit the specific key information extraction method, for example, the key information is extracted by NLP keyword extraction.

[0150] S150. Calculate the number of abnormal business platforms where the same key information appears simultaneously.

[0151] The same key information of business platforms indicates that the corresponding abnormal problems are the same or similar, and the corresponding solutions may also be the same or similar. Counting the number of abnormal platforms of business platforms with the same key information at the same time makes it easier to screen out outstanding abnormal information and its corresponding abnormal business platforms, which is beneficial to subsequent data analysis. The more abnormal platforms of business platforms with the same key information at the same time, the more outstanding the abnormal problems corresponding to the same key information are, and the more attention needs to be paid to analysis.

[0152] S160: When the number of abnormal platforms meets the threshold condition of aggregate feedback, the abnormal information of each business platform is aggregated to obtain aggregate feedback information.

[0153] Threshold conditions refer to the boundary conditions that trigger aggregate feedback. For example, threshold conditions include but are not limited to reaching a preset number of times per unit time. Aggregate feedback information refers to the summary information corresponding to the abnormal information of each business platform. When the number of abnormal platforms meets the threshold conditions of aggregate feedback, it indicates that the abnormal problems corresponding to the same key information appearing at the same time are prominent. The abnormal information of each business platform is aggregated to obtain aggregate feedback information, which is convenient for subsequent data preview and data processing.

[0154] In a specific embodiment, the threshold condition refers to the occurrence of object feedback exceeding a specified threshold (e.g., 5 cases) within a short period of time (e.g., 0.5 hours), and these feedbacks reflect the same type of problem. For example, during the period of 16:00-16:30, 10 people reported through various channels that they could not hear other people's voices on the audio and video sub-channels of the channels of instant messaging products. Then, the audio and video sub-channels of the channels of various instant messaging products that they could not hear other people's voices are summarized to obtain aggregate feedback information.

[0155] Collect aggregate feedback information so that it can be processed first in the subsequent log analysis process, thereby improving the analysis efficiency of most objects for current emergency faults.

[0156] In a specific embodiment, please refer to Figure 3 The method shown completes the collection of abnormal information sets and the acquisition of descriptive language. Specifically, the method of the embodiment of the present invention includes the following steps:

[0157] T301, the business object feedbacks abnormal information through the terminal's business platform APP, official website and other channels;

[0158] T302, the business platform background receives and stores abnormal information;

[0159] T303, the business platform triggers the network hook to send the abnormal information to the server through the network hook;

[0160] T304, the server receives the exception information and pre-processes the exception information, for example, performs data cleaning or keyword extraction. The pre-processed exception information can be used to summarize and obtain aggregate feedback information (as shown in step T306). The pre-processed exception information can also be used for language model recognition (as shown in step T305).

[0161] T307, obtains abnormal analysis results and other feedback problem interaction results.

[0162] Optionally, step S200 specifically includes steps S210 to S230:

[0163] S210: Map and associate the selected multiple scene names with the scene description words corresponding to each scene name to construct a scene mapping table.

[0164] The scene name represents the distinguishing name of the scene, the scene description word represents the descriptive language corresponding to the scene, and the scene mapping table is a corresponding mapping table between the scene name and the scene description word, and the scene name and the scene description word in the scene mapping table correspond to each other. Specifically, first determine the scene name, then determine the scene description word corresponding to the scene name, and then map and associate the scene name with the corresponding scene description word to construct the scene mapping table.

[0165] It should be noted that the scene mapping table may also include other relevant information, such as scene ID, etc. Refer to Table 1. In Table 1, the scene mapping table includes scene name, scene ID and scene description. Both scene name and scene ID can be used to represent scene distinguishing information. Scene name and scene ID can be used in different occasions. For example, scene name can be used to display to business objects for viewing, so that business objects can understand the meaning of the scene, and scene ID is used for machine language learning, which facilitates machine recognition of normalized data content and improves learning efficiency. The scene mapping table can also include general scripts, which are used to analyze specific scene content. For example, it can be used to obtain general basic information such as App version and object account information.

[0166] Table 1

[0167]

[0168] S220 , identifying and converting each descriptive language in the descriptive language set by using a language model to obtain keywords of the descriptive language.

[0169] The keywords of the descriptive language represent the key information of the descriptive language. The scenes corresponding to the descriptive language can be identified through the keywords of the descriptive language with high accuracy and speed. Specifically, each descriptive language in the descriptive language set is input into the trained language model for recognition conversion, and useless information is filtered out to obtain the keywords of the descriptive language.

[0170] S230: Determine one or more candidate scene names corresponding to each descriptive language as first scene information according to the degree of association between the keyword and each scene description word in the scene mapping table.

[0171] The degree of association between the keyword and the scene description word represents the correlation between the keyword and the scene description word. The greater the degree of association between the keyword and the scene description word in the scene mapping table, the closer the descriptive language corresponding to the keyword is to the scene description word in the scene mapping table, that is, the higher the matching degree between the scene corresponding to the descriptive language of the abnormal information and the scene corresponding to the scene description word in the scene mapping table.

[0172] Specifically, the server first calculates the correlation between the keyword and each scene description word in the scene mapping table, then sorts the correlation degrees, and uses the candidate scene names corresponding to several scene description words with greater correlation degrees as the first scene information.

[0173] Optionally, step S200 specifically further includes steps S201 to S204, where steps S201 to S204 represent a training and tuning process of a language model:

[0174] S201. Collect experience documents for solving business anomalies, and perform data cleaning on the experience documents.

[0175] Experience documents refer to files containing abnormal problems and solutions. The specific method of collecting experience documents for solving business anomalies and performing data cleaning on the experience documents is not specifically limited in this embodiment. For example, experience documents for solving business anomalies uploaded by business objects are directly received, or abnormal problems submitted by business objects through the business platform and corresponding solutions taken by the business platform are collected.

[0176] S202: Input the cleaned experience document into the language model to obtain a question-answering dataset that meets the preset format requirements.

[0177] The experience document includes abnormal problems and solutions to abnormal problems, and may also include information such as channel sources. The training and optimization of the language model only needs to select useful information from the experience document. Therefore, the experience document needs to be cleaned to reduce the amount of calculation and improve the accuracy. The data cleaning method is determined according to the actual application, and this embodiment does not make specific requirements, such as merging polysemous words or removing duplicate words. The question and answer dataset contains multiple abnormal problems and their corresponding solutions. The format requirements of the question and answer dataset are determined according to the actual application, and this embodiment does not make specific requirements. For example, chatgpt (ChatGenerative Pre-trained Transformer) is used to generate a question and answer dataset with preset format requirements.

[0178] Specifically, the experience documents are first cleaned, and then the cleaned experience documents are input into the language model and converted through a conversion tool to obtain a question-answering dataset that meets the preset format requirements.

[0179] S203: Filter and format the question-answering dataset to construct a pre-training dataset.

[0180] The question-and-answer dataset may contain invalid information, so it is necessary to filter out invalid information to reduce the amount of calculation and improve accuracy. Similarly, the question-and-answer dataset is formatted to unify the format of the question-and-answer dataset to coordinate the differences of the question-and-answer datasets corresponding to different business platforms and improve batch processing capabilities.

[0181] Specifically, the question-answering dataset is first filtered, and then the filtered question-answering dataset is formatted to obtain a filtered and standardized question-answering dataset in a unified format to establish a pre-training dataset.

[0182] S204: Adjust parameters of the language model according to the pre-training data set, and add information of the experience document to the language model.

[0183] Language models include but are not limited to LLM (Large Language Model l, large-scale pre-trained language model). The information of the experience document includes the attribute information of the experience document, and different attribute information corresponds to different result formats; for example, the information of the experience document includes but is not limited to scenario information, and the scenario information includes but is not limited to multiple scenarios such as historical question vectors, general document vectors, and log scenario vectors.

[0184] Specifically, the pre-trained data set is input into the language model, and the language model adjusts its own parameters according to the pre-trained data set. In addition, the information of the experience document is input into the language model to adjust the input and output format information of the language model to obtain the expected answer. For example, based on the LangChain application framework, a large language model (LLM) is developed. LangChain can directly work with OpenAI's text-davinci-003, gpt-3.5-turbo model and various open source language models of Hugging Face, such as Google's flan-t5 and other models. By using LangChain, various applications of LLM that are more powerful and efficient can be developed. The LLMs that come with langchain are all based on the default prompt template, and the business object only needs to submit the results of the LLM vector matching. Exemplarily, in the problem analysis scenario, multiple scenarios such as historical problem vectors, general document vectors, and log scenario vectors are involved. Different formats of replies need to be given according to different scenarios, so as to customize the LLM prompt. By implementing the langchain LLM interface, the language model corresponding to the LLM request is determined according to the customized prompt to obtain the expected answer. Adjust the language model parameters based on historical experience documents to improve the accuracy of the model.

[0185] In a specific embodiment, see Figure 4 ,The tuning process of the language model is as follows:

[0186] T401, first collect historical development experience documents as experience documents for solving business exceptions.

[0187] T402, then perform data cleaning on the experience documents and the cleaned data;

[0188] T403, submit the cleaned data to chatgpt, so that it can return the question-answering dataset in a fixed format to us as required;

[0189] T404, filtering invalid information in the question-answering dataset and formatting it to form a pre-training dataset;

[0190] T405, fine-tuning the open source model using the processed pre-training dataset;

[0191] T406, add development experience documentation to optimize parameters to achieve the expected results.

[0192] Optionally, step S300 specifically further includes steps S310 to S320:

[0193] S310. Determine one or more scene identifiers from a scene mapping table according to the first scene information.

[0194] The first scene information includes one or more candidate scene names corresponding to the descriptive language. The scene mapping table includes but is not limited to the scene name, scene description word and scene ID, etc. There is a one-to-one correspondence between the scene name, scene description word and scene ID, etc. The scene identifier is used to distinguish different scene categories. The scene identifier is determined according to the actual application. This embodiment does not make specific restrictions. For example, the scene ID is used as the scene identifier.

[0195] Specifically, one or more scene names are first extracted from the first scene information, and then one or more scene identifiers corresponding to the one or more scene names are screened out from the scene mapping table.

[0196] S320: Summarize the scenario log requirements corresponding to the scenario identifiers determined from the first scenario information; wherein the scenario log requirements are used to represent the log information to be collected by each business platform.

[0197] The scene category information is represented by the scene identifier. There is a one-to-one correspondence between the scene identifier and the scene log. During the implementation of the scene corresponding to the scene identifier, certain scene log information will be generated. The scene log requirements can be determined based on information such as the functions and operations implemented by the scene.

[0198] Specifically, after determining the scene identifier obtained from the first scene information, the corresponding scene log requirements can be determined according to each scene identifier, and then one or more scene log requirements corresponding to one or more scene identifiers are summarized to facilitate obtaining multiple scene log requirements at one time and reduce data acquisition time.

[0199] Optionally, step S400 specifically further includes steps S410 to S430:

[0200] S410. Obtain log information from the corresponding business platform according to the scenario log requirements; wherein the log information includes a log collection timestamp, a business object account, and an abnormal log segment.

[0201] Log information is reflected in the form of log lines. The log information corresponding to the scenario log requirements contains one or more log lines. Each log line contains a log collection timestamp. Each log line also includes any one or more of the business object account, abnormal phenomenon or abnormal operation. Among them, the log corresponding to the abnormal phenomenon or abnormal operation can be understood as an abnormal log segment.

[0202] Specifically, the server determines the corresponding business platform according to the scenario log requirements, and obtains log information that meets the scenario requirements from the determined business platform. For example, the scenario log requirements are the abnormal log information of the video sharing platform in the video freeze scenario and the abnormal log information of the test training platform in the download scenario. The server obtains the log information during the video freeze period from the video sharing platform and the log information during the download period from the test training platform according to the scenario log requirements.

[0203] S420: Perform regular expression matching on the abnormal log segment of the log information according to a preset analysis script to determine the target log line.

[0204] Regular matching refers to the use of regular expressions to match strings in the analysis script. Regular expressions are a language used to describe text patterns. They can be used for operations such as matching, searching, and replacing strings. Regular expressions are text patterns composed of ordinary strings and metacharacters. For example, in a shell script, you can use tools such as grep, sed, and awk to perform regular matching operations: use the grep command to find lines containing a certain pattern, and use the sed command to replace a string of a certain pattern. It should be noted that the preset analysis script is determined according to the actual application, and this embodiment does not make specific restrictions. The target log line represents the key log line of the abnormal log segment, and the key log line contains the characterization information of the abnormal analysis result.

[0205] Specifically, the server performs regular matching on the abnormal log segment of the log information according to a preset analysis script, and selects key log lines from the abnormal log segment as target log lines.

[0206] It should be noted that after determining the target log line from the abnormal log segment, the target log line can be further analyzed to obtain the corresponding descriptive language to increase readability. In a specific embodiment, referring to Table 2, the first column in Table 2 is the target log line, and the second column is the descriptive language corresponding to the first column.

[0207] Table 2

[0208]

[0209]

[0210] S430: Perform an exception analysis on the target log line according to the experience data of the exception analysis library to determine the exception analysis result under the corresponding business scenario.

[0211] The exception analysis library contains exception logs corresponding to multiple exception information of multiple business platforms and their corresponding exception analysis results. It should be noted that the exception analysis library is established based on actual application experience and cases, and is continuously updated during use. The exception analysis results include exception information and its corresponding solutions.

[0212] Specifically, the server matches the target log line with the exception logs in the exception analysis library, and uses the exception analysis results corresponding to one or more exception logs whose matching degree meets the preset requirements as the exception analysis results under the corresponding business scenario, or uses the exception analysis results corresponding to the exception log with the highest matching degree as the exception analysis results under the corresponding business scenario.

[0213] It should be noted that the anomaly analysis library of the embodiment of the present invention can be understood as an AI knowledge base, which is information stored as structured and unstructured data, and can be used to analyze problems or infer the causes of problems. It is used to further train the language model mentioned above. The AI ​​knowledge base of the embodiment of the present invention can convert words, sentences or text blocks in structured or unstructured documents into numerical vectors through embedding, and complete the storage of data in the AI ​​knowledge base based on the numerical vectors. Furthermore, the AI ​​knowledge base can continuously update and iterate its own data based on the results of previous anomaly analyses, so that the anomaly logs stored in the library are more abundant and complete, and when a new anomaly analysis demand is initiated next time, the matching accuracy of the anomaly analysis can be improved.

[0214] Therefore, after the embodiment of the present invention realizes the rapid scenario recognition of abnormal problems based on the language model as mentioned above, it completes the iterative update of historical recognition results based on the AI ​​knowledge base, and then completes a whole set of improvements from problem analysis to solution supplementation, which greatly improves the rapid and accurate abnormal business analysis in dealing with multiple scenarios and multiple problems.

[0215] Optionally, the business scenario includes a general scenario, an audio and video usage scenario, or a shared screen viewing scenario, and step S430 specifically further includes steps S401 to S403:

[0216] S401. In a general scenario, extract information from a target log line to obtain business platform version information and a business object account to determine an abnormality analysis result of the business object in the general scenario.

[0217] Refer to Table 1. The general scenario represents the basic scenario applicable to various business scenarios. The target log corresponding to the general scenario contains the general basic information of different business scenarios. It should be noted that in the general scenario, the information contained in the target log line includes but is not limited to time information, business platform information, business platform version information, business object account information and other abnormal analysis results.

[0218] Specifically, in a general scenario, information is extracted from the target log line corresponding to the general scenario according to the general scenario script analysis template to obtain the business platform version information and the business object account number, so as to determine the abnormal analysis results of the business object in the general scenario, such as whether the business platform version information is the latest version, whether the business object account number exists, etc. It should be noted that the general scenario script analysis template includes but is not limited to the general scenario description, the general scenario analysis script code, etc.

[0219] S402: In the audio and video usage scenario, information is extracted from the target log line to obtain interface operation information and audio device fault information of the business object, so as to determine an abnormal analysis result of the business object in the audio and video usage scenario.

[0220] Referring to Table 1, the interface operation information represents the operation information input by the business object received by the business platform through the interface, such as click or slide operation, etc. The audio device fault information represents the abnormal information of audio playback, such as abnormal information of the audio device or audio file encoding error. It should be noted that the abnormal analysis results in the audio and video usage scenario include but are not limited to interface operation information and audio device fault information.

[0221] Specifically, in the audio and video usage scenario, information is extracted from the target log line corresponding to the audio and video usage scenario according to the audio and video usage scenario script analysis template to obtain the interface operation information and audio device fault information of the business object, so as to determine the abnormal analysis results of the business object in the audio and video usage scenario, such as whether the interface operation of the business object is correct and whether the audio device fault has an output line fault. It should be noted that the audio and video usage scenario script analysis template includes but is not limited to the audio and video usage scenario description, the audio and video usage scenario analysis script code, etc.

[0222] S403: In a shared screen viewing scenario, extract information from the target log line to obtain shared operation information and virtual location information of the business object to determine an abnormality analysis result of the business object in the shared screen viewing scenario.

[0223] Referring to Table 1, the sharing operation information represents the operation information of executing sharing received by the business object from the business platform, such as initiating or interrupting the sharing screen operation. The virtual location information of the business object represents the setting location information of the business object on the business platform, such as the voice room of the shared room.

[0224] Specifically, in the shared screen viewing scenario, information is extracted from the target log line corresponding to the shared screen viewing scenario according to the shared screen viewing scenario script analysis template to obtain shared operation information and virtual location information of the business object, so as to determine the abnormal analysis results of the business object in the shared screen viewing scenario, such as whether the shared operation process is correct and whether the virtual location of the business object is at the preset location. It should be noted that the shared screen viewing scenario script analysis template includes but is not limited to the shared screen viewing scenario description, shared screen viewing scenario analysis script code, etc.

[0225] It should be noted that the business scenario is determined according to the application scenario of the business platform, and this embodiment does not impose any specific restrictions. For example, the business scenario may also include a video playback scenario, an image storage scenario, and the like.

[0226] In a specific embodiment, see Figure 5 ,The multi-scenario script analysis process is as follows:

[0227] T501, the server first receives the abnormal information sent by different business objects through different business platforms, and aggregates the abnormal information to form a list of objects and their abnormal problems;

[0228] T502, obtain the problem scene ID list, i.e., the scene mapping table, and determine the scene identifier corresponding to the object according to the abnormal information and the scene mapping table;

[0229] T503, traverse the list of feedback objects and obtain the identification information of each object;

[0230] T504-T505, guide the object to upload the operation log to the server through identification information;

[0231] T506, obtaining log information corresponding to the abnormal problem from the operation log according to the scenario identifier, analyzing the log information corresponding to the abnormal problem to obtain a scenario script of the abnormal problem, and obtaining a scenario analysis script library;

[0232] T507-T511, compare and analyze the scenario scripts of abnormal problems with the scenario analysis script library to obtain script analysis results for different scenarios (such as general script analysis T5071, scenario A script analysis T5072, and scenario B script analysis T5073), summarize the script analysis results of different scenarios and determine whether additional manual analysis is needed. If the summarized log scenario analysis results can cover all problems, manual analysis is not required, and the analysis results are sent to the corresponding business platform to notify the object; if the summarized log scenario analysis cannot cover all problems, create the corresponding scenario, write the corresponding scenario analysis description and analysis script code, and add them to the scenario analysis script library.

[0233] Optionally, the abnormal business analysis method further includes steps S510 to S540:

[0234] S510: vectorize the descriptive language corresponding to each abnormal information in the abnormal information set to obtain a problem vector.

[0235] Text vectorization refers to representing text information as a vector that can express the semantics of the text, using a numerical vector to represent the semantics of the text. Word embedding is a method of converting words in a text into digital vectors. Common text vectors and word embedding methods include the One Hot Model, Bag of Words Model, TF-IDF, N-Gram, Word2vec, Doc2vec, etc. In the training of big data language models, converting text information into vector information for training can greatly reduce computing time and save computing resources.

[0236] Specifically, the descriptive language corresponding to each abnormal information in the abnormal information set is subjected to text formatting and text semantic segmentation using a vectorization processing method to obtain a problem vector.

[0237] S520: Vectorize the experience document for solving the exception to obtain an experience vector set.

[0238] The experience document for solving anomalies contains abnormal problems and solutions corresponding to the abnormal problems. The experience vector set represents a vector set of multiple abnormal problems and their corresponding solutions. Similarly, a vectorization method is used to vectorize each abnormal problem and its corresponding solution in the experience document one by one to obtain an experience vector set.

[0239] See also Figure 6,The generation process of the experience vector set can be realized through the interaction process of the object, business platform, problem analysis platform and vector processing platform. Specifically, the generation process of the experience vector set is as follows: S01, the state flow of the abnormal problem fed back by the business platform has been solved; S02, the problem analysis platform observes the state of problem processing; S03, when the state flows to solved, the problem analysis platform obtains the cause and solution of the abnormal problem; S04, the business platform feeds back the cause and solution of the abnormal problem to the problem analysis platform, thereby obtaining the abnormal problem history document; S05, the object uploads the experience document to the problem analysis platform; S06, the problem analysis platform pre-processes the collected problem history experience documents and the uploaded experience documents, such as text format processing or semantic segmentation; S07, sends the pre-processed experience documents to the vector processing platform; S08, the vector processing platform vectorizes the experience documents and stores them in the vector library to form an experience vector set.

[0240] S530 , performing vector matching on the question vector according to the experience vector set, and obtaining a plurality of experience vectors whose approximation is higher than a target threshold as context information of the question vector.

[0241] Vector matching refers to the degree of similarity between vectors. In this embodiment, text vector matching refers to calculating text similarity, which is calculated by converting text into vector form. The similarity calculation method includes but is not limited to the distance or angle between vectors, and the similarity is measured by the size of the distance or angle. Vector matching methods include but are not limited to cosine similarity, Euclidean distance, jaccard similarity, etc.

[0242] Specifically, a vector matching method is used to perform vector matching on multiple experience vectors in the experience vector set and the question vector, and multiple experience vectors with approximations higher than the target threshold are obtained as context information of the question vector. Among them, high approximation can be understood as an approximation that exceeds a preset threshold. It can be understood that the higher the approximation, the higher the consistency between the experience vector and the question vector, and the context information corresponding to the experience vector is closer to the real question vector.

[0243] S540: Combine the question vector and the context information to generate a descriptive language corresponding to the abnormal information.

[0244] The vector information after the question vector and context information are combined is more comprehensive, which can increase the feasibility of the results. Specifically, the server first combines the question vector and context information to form new vector information, and then uses the inverse process of vectorization to convert the new vector information into text information to obtain the descriptive language corresponding to the exception information, thereby increasing the readability of the exception information.

[0245] In some embodiments, the above step S510 may include the following steps:

[0246] S5101, according to a preset field set, serialize each field in the descriptive language that is in the field set to obtain a structured descriptive language; S5102, block each field to obtain block values ​​of multiple sub-blocks; S5103, determine the peak block of the current field according to the block values ​​of each sub-block, and obtain the character string of the peak block; S5104, extract keywords from the character string of the peak block, and use the extracted keywords to compress the usage frequency of each word in the character string to obtain a normalized character string; S5105, update the structured descriptive language according to the normalized character string; S5106, vectorize the updated structured descriptive language to obtain the problem vector.

[0247] Specifically, the embodiment of the present invention vectorizes the historical experience of problems, which may include the following processing steps: 1. Checking the status of problem processing on the problem platform, and when the status flows to solved, obtaining the cause and solution of the problem; 2. Using natural language processing technology to perform semantic segmentation on the obtained information; 3. Vectorizing the processed text and storing it in the vector library.

[0248] The above vectorization process is described in detail as follows:

[0249] 1. Serialize the target field extracted from the question into JSON, for example:

[0250] "bug_id": bug ID,

[0251] "project_id": project ID,

[0252] "status": problem defect handling status,

[0253] "url": defect link URL,

[0254] "title": defect title,

[0255] "description": Full description of the defect: such as description of the problem, code stack when the crash occurs, etc.

[0256] "platform": terminal type,

[0257] "handler": person who handles the situation;

[0258] 2. Vectorize the serialized results through a unified interface. Specifically, serialize a series of strings through JSON to obtain a JSON sequence. Divide the JSON sequence into blocks, and then vectorize each block.

[0259] In the above-mentioned vectorization process, if a JSON dimension of a document or question is too large, it will greatly affect the block result. A typical usage scenario is, for example, in a certain abnormal question, the title has 1 field and the content has 500 fields. At this time, the JSON document of the content dimension is obviously much larger than the title dimension, resulting in a large difference in blocks, which ultimately leads to a large number of errors in the vectorization results. Therefore, the embodiment of the present invention performs normalization processing on dimensions that will generate a large number of blocks. The normalization process is to segment the value of the field and terminate the process when a block size is basically consistent with other dimensions.

[0260] Normalization means changing the value of the field to ensure that the value of each JSON field has a basically consistent effect on the vector result.

[0261] The vectorization process combined with normalization processing in an embodiment of the present invention is described in detail below:

[0262] 1. Get descriptive language;

[0263] 2. Generate JSON structured data A based on the descriptive language;

[0264] 3. Separate each field of JSON structured data A into separate blocks;

[0265] 4. Compare the block sizes of each field;

[0266] 5. When there is a peak value of the block size of a single field, extract the value a_value of the peak field a; extract the key words in the a_value string; according to the extraction results of the key words, compress the usage frequency of each word, and finally obtain a new result a_value_normalize; in this way, complete the data processing of all fields;

[0267] 6. The normalized result obtained by the above data processing replaces the original JSON structured data A, and A is divided into blocks again as a whole;

[0268] 7. Finally, the block results are vectorized to obtain the final vectorized data.

[0269] Among them, when determining whether there is a peak in the block size of a single field, the embodiment of the present invention can perform unsupervised classification on all fields using the block size dimension. The final unsupervised classification result is that one or several blocks with similar sizes will form a group, and see whether the difference in block size between each group is >300%. If so, it can be determined that there is a peak in block size.

[0270] In addition, when determining whether the normalization of a certain field is completed, the embodiment of the present invention can continuously segment the content and rearrange the results according to the number of occurrences of each segment. When the segment sizes of the most frequently extracted key segmentations in a certain scene are similar, the analysis is considered to be completed.

[0271] Based on this, the embodiment of the present invention can ensure that the sizes of the blocks are balanced during the vectorization processing, thereby improving the accuracy of the vectorization processing.

[0272] In a specific embodiment, see Figure 7 , the recognition conversion process of the language model is as follows:

[0273] T701, first vectorize the descriptive language;

[0274] T702, then use the matching algorithm to perform similarity search in the generated vector library;

[0275] If T703-T705 can match the text vector in the library, it will obtain the top k pieces of data with the highest matching degree and combine them with the context and question into the input language form for feedback;

[0276] T706-T707, if no match is found, return empty.

[0277] It should be noted that the input and output of the language model can be in the form of prompts. Prompts are a type of natural language input that instructs the AI ​​model what actions it should take or what outputs it should generate when performing a specific task. Therefore, the natural language understanding capabilities of the language model can be used to obtain answers.

[0278] Optionally, the abnormal business analysis method further includes steps S610 to S620:

[0279] S610: Structurally decompose the abnormal analysis result to determine multiple information components in the abnormal analysis result.

[0280] It should be noted that the abnormality analysis result is determined based on the log information corresponding to the abnormal information. Since the readability of the log information is poor, if the log information corresponding to the abnormality analysis result is directly displayed, it is difficult for ordinary business objects to understand.

[0281] The information components represent the key factors that constitute the information. Specifically, the server first performs structural decomposition on the log information corresponding to the abnormal analysis result to obtain multiple information components in the abnormal analysis result. It should be noted that the rules for structural decomposition are determined according to the actual application, for example, the decomposition is performed according to the execution step node that causes the abnormality.

[0282] S620: Map each information component to obtain a descriptive language corresponding to the abnormal analysis result.

[0283] Mapping refers to the process of determining another factor based on the "corresponding" relationship between the elements of the two sets of elements and one of the elements. Therefore, before mapping, the corresponding relationship between the two elements needs to be constructed. In this embodiment, the corresponding relationship between the two elements refers to the corresponding relationship between the log information and the descriptive language, such as a mapping table between the log information and the descriptive language.

[0284] Specifically, the server maps various information components according to the corresponding relationship between the log information and the descriptive language, obtains the corresponding phonetic description, and uses it as the descriptive language corresponding to the abnormal analysis result.

[0285] Optionally, the abnormal business analysis method further includes steps S710 to S720:

[0286] S710. Acquire abnormality analysis results in multiple business scenarios in parallel.

[0287] In a certain period of time, a certain business object may generate abnormal information in multiple business scenarios on one or more business platforms, corresponding to abnormal analysis results in multiple business scenarios. Specifically, the server obtains abnormal analysis results in multiple business scenarios in parallel to improve the speed and completeness of data acquisition.

[0288] S720. Sort and summarize the anomaly analysis results according to the timestamp information of the anomaly analysis results, and construct multiple anomaly analysis result sets according to the business scenarios in which the anomaly analysis results are located; wherein the anomaly analysis results in each anomaly analysis result set are in the same business scenario.

[0289] Specifically, the server first sorts and summarizes the anomaly analysis results in chronological order according to the timestamp information of the anomaly analysis results, and then classifies the sorted anomaly analysis results according to the category of the business scenario. See Table 3, which contains business scenarios such as general scenarios and audio and video usage scenarios. The anomaly analysis results of each business scenario are arranged in chronological order. Sorting and classifying the anomaly analysis results of business objects in multiple business scenarios facilitates the display and reading of the anomaly analysis results.

[0290] Table 3

[0291]

[0292]

[0293] In a specific embodiment, see Figure 8 The process of abnormal business analysis is as follows: S11. The operation object summarizes the abnormal problem feedback and outputs the descriptive language; S12. The operation object sends the summarized problems and descriptive language to the development object, and the development object records the abnormal problems; S13. The operation object sends the summarized problems and descriptive language to the pre-trained language model; S14. After the pre-trained language model analyzes, the common causes of the abnormal problems are fed back to the operation object; S15. The pre-trained language model identifies the scene of the abnormal problem and maps a single problem to a series of scene IDs; S16. The pre-trained language model sends the logs of the abnormal scene in batches to the log analysis tool; S17. The pre-trained language model sends the common causes of abnormal problems to the development object; S18, returns the log analysis results through webhook to feedback to the development object; S19, the development object conducts a comprehensive analysis based on the common causes of abnormal problems and the returned log analysis results; S20, the development object sends a query to the pre-trained language model to repair the abnormal problem; S21, the development object receives the repair plan; S22, the development object determines the supplementary information based on the comprehensive analysis results and the repair plan, and sends the supplementary information to the pre-trained language model so that the pre-trained language model can be further improved and trained; S23, the development object sends the final abnormal analysis results to the operation object. The above abnormal business analysis process continuously updates and improves the training data set through the operation object and the development object, so as to make the abnormal business analysis method more accurate.

[0294] In some embodiments, the above step S400 obtains log information from the corresponding business platform according to the scenario log requirement, and further includes:

[0295] S4001, matching the scenario log requirements with the candidate input templates cached in the preset cache rule library to determine the target input template;

[0296] S4002, according to the target input template, obtaining historical cache information corresponding to the descriptive language from a preset cache result library;

[0297] S4003, use the historical cache information as the request result of the scene log requirement, and output the request result as log information.

[0298] Specifically, in the process of analyzing abnormal problems, an embodiment of the present invention provides an optional intent recognition method. The role of intent recognition here is to determine what operation to perform next based on problem experience, such as: analyzing the cause of the crash, analyzing the program principle, and querying and analyzing the logs uploaded by the user.

[0299] Intent recognition methods include intent recognition without a cache strategy and intent recognition with a cache strategy.

[0300] For intent recognition without a cache strategy, an embodiment of the present invention may provide an LLM prompt design template. By using the prompt template and providing input content {input}, the formatted JSON result may be extracted through the trained language model and provided to the target object.

[0301] For example, the content of a prompt template can be:

[0302] Intent type: ask_question;

[0303] Intent description: obtain certain knowledge and information, seek advice and help, inquire about certain issues, inquire about the causes and solutions of problems;

[0304] Return format: {{"action":"ask_question","question":"","platform":"","crash":""}};

[0305] Parameter Description:

[0306] - "action": intent type, value "ask_question"

[0307] - "question": the specific content of the user's query or information

[0308] - "platform": the device platform type that the user is querying questions or obtaining information on, ranging from "android", "ios" and "general". If the device platform type cannot be identified in the user input, the default value is "general"

[0309] - "crash": whether the problem reported by the user is flash back, crash, or crash. If yes, the value is "1", otherwise the value is "0". The default value is "0";

[0310] At this time, if the target object inputs "webview cross-process rendering flow" into the language model, the language model can recognize {{"action":"ask_question","question":"webview cross-process rendering flow","platform":"general","crash":"0"}} according to the above template.

[0311] Next, the embodiment of the present invention may also use precise intent recognition with a cache strategy to reduce the number of times the above-mentioned AI interface is used, reduce costs, and solve the time-consuming problem of AI interface requests.

[0312] First, the embodiment of the present invention constructs a cache rule library, as shown in Table 4 below:

[0313] Table 4

[0314]

[0315] The embodiment of the present invention can provide some template inputs, and different users can click on the account or time in these template inputs to change to their real content. Therefore, the structured results of these queries are also consistent, and only the account or time information in the results needs to be replaced.

[0316] The cache rule base is used to decide what input to cache. Therefore, a regular expression rule is provided in the "hit rule" to extract different variables in the input template in the "variable extraction rule".

[0317] The embodiment of the present invention will also build a cache result library, and the cache result is used to cache the result returned by the large language model last time. It is convenient for direct reuse in future requests. Exemplary:

[0318]

[0319] In the above case, each time you input and obtain logs, you can use the previous JSON result, but replace the "u in":"123456789" field in the result. When the user's input matches the rule 111122, the previous result will be used as long as it is within the validity period, and the user's u in input will be extracted into the cache and returned.

[0320] In some embodiments, the abnormal business analysis method further includes a step of correcting the legitimacy of the intended result based on the label rule, specifically including:

[0321] G100, obtain the scene recognition result of the language model, and generate a vector to be verified for the scene recognition result;

[0322] G200, according to the vector to be verified, matches the corresponding label template vector from the scene label template library;

[0323] G300, calculates the vector distance between the vector to be checked and the label template vector;

[0324] G400, determines the legitimacy verification result of the scene recognition result according to the vector distance;

[0325] G500, when the legitimacy check result of the scene recognition result is illegal, calculate the distance between each label dimension in the vector to be checked and the corresponding dimension in the label template vector to determine the illegal vector dimension;

[0326] G600, determine the label to be updated according to the illegal vector dimension;

[0327] G700, according to the tag to be updated, extracts the field information corresponding to the tag to be updated from the descriptive language to replace the scene recognition result.

[0328] Specifically, in the embodiment of the present invention, the normal intent recognition result can be expressed in JSON string format, where the key-value is different in different systems and even different intents. The key-value here is defined by the platform developer when designing the promt of the big prediction model, but the existing method mainly relies on the big model itself, which may be wrong. The following is a common error case:

[0329] 1. Missing fields

[0330] The input content is: Get the log of Android user 4e65ecd95b3f082ce280f815ab6c40a5 at 2023-08-1600:23:04;

[0331] Intent recognition results: The results returned by the large model may be as shown in Table 5 below, with the uin field missing. The main reason for this is that the results of the large model may not fully meet expectations, even if the developer has adjusted the prompt countless times.

[0332] Table 5

[0333]

[0334] 2. The result field value is inaccurate, as shown in Table 6

[0335] Input content: Who is the owner of GuildMediaMemberListView?

[0336] Intent recognition result: The code name GuildMediaMemberListView is not found in the results returned by the large model, so the description field is empty.

[0337] Table 6

[0338]

[0339] 3. The intention action is inaccurate, as shown in Table VII;

[0340] The input content is: What problems did you encounter during the development of the post viewer?

[0341] The intent recognition result is: the large model occasionally returns the intent action query_bug_history (find similar tickets), but the correct intent should be ask_question (ask a question)

[0342] Table 7

[0343]

[0344] To summarize the above problems: Intent recognition requires the big model to extract information from the user's input according to the platform's prompt content requirements and return the result in JSON format. However, due to its own reasons, the big model returns an uncertain result that may not meet the requirements.

[0345] In summary, the intent results returned by the large model have the following disadvantages:

[0346] 1. The intended result is random. The key of a field or even multiple fields may be wrong, and the value may also be wrong.

[0347] 2. The existing error correction logic is basically code-compatible retry, which cannot solve the problem of the current network request.

[0348] 3. The main method for judging the accuracy of the results is through value correlation detection, which easily leads to the expansion of a single dimension of the vector and poses a huge challenge to the setting of the threshold standard for correlation judgment.

[0349] Therefore, an embodiment of the present invention provides a correction scheme for the legality of the intended result. Legality refers to whether the result complies with the prompt design specification. The legality correction scheme specifically includes the following steps:

[0350] 1. Create label rules: This stage is carried out simultaneously with the prompt stage in the design of LLM. The intent recognition, legality recognition and correction scheme of this solution requires data preparation in the early stage. In order to achieve the goal of the solution, a concept of intent result label is first proposed. See Table 8 below for details. The labels here can be additionally defined. Multiple labels may be matched in one input. The following lists the common centralized types, which are not limited here.

[0351] Table 8

[0352]

[0353]

[0354] The legality verification process of the present invention includes a feature extraction process for generating label rules, and a feature matching process for updating the legality verification by combining the actual input and the intended result.

[0355] First of all, the feature extraction process based on label rules mainly consists of several steps:

[0356] 1. Generate the data of the first dimension of the vector. This dimension is used to store all the key names of a certain intent result, such as action, question, uin, etc.

[0357]

[0358] Among them, MaxHash is all hash (a m ), the maximum value in m The value of the first dimension is the mth key name, and the value of the other dimensions is the maximum value of the hash results of all known tags. The calculation method of Dimen is:

[0359] sum=hash("action")+hash("question")+hash("platform")+hash("description")

[0360] max=max_of(hash("action"),hash("question"),hash("platform"),hash("description"))

[0361] Dimen=sum / max

[0362] 2. Traverse the tag list of the generated intent results and generate the tag results of the "Analyze Log" intent.

[0363] Exemplarily, the identification results of the relevant content are as follows:

[0364] "action":None

[0365] "question": Complete match input, URL hyperlink

[0366] "platform": mobile platform enumeration

[0367] "description": URL hyperlink.

[0368] 3. Generate vector values ​​for the intent template. Generate vector values ​​for the label list generated in 2 according to the method in 1. m The mth label under the action dimension.

[0369] For example, when Dimen(action)=0

[0370] Dimen(question)=(hash("Complete match input")+hash("URL hyperlink")) / max_hash

[0371] Dimen(platform)=hash("mobile platform enumeration") / max_hash=1

[0372] When Dimen(description)=hash("URL hyperlink") / max_hash=1,

[0373] Assume that the value of Dimen(question) is 0.7, and the result of hashing all keys is 0.66.

[0374] The final generated vector is: [0.66, 0, 0.7, 1, 1].

[0375] 4. Verify the legitimacy of user input

[0376] The formula for judging the legitimacy of LLM's intended result is Dimen(analyze_by_log template)=Dimen(LLM intended result), and the results of all dimensions must be completely consistent.

[0377] For example, the #input is:

[0378] Analyze the crash information in the log: https: / / AA.com / test8

[0379] The intent result returned by LLM is:

[0380] {{

[0381] "action":"analyze_by_log",

[0382] "question":"Analysis of the crash information in the log: https: / / AA.com / test8",

[0383] "platform":"general",

[0384] "description":"Analyze the crash information in the log"

[0385] }};

[0386] At this time, the distance metric in the actual LLM result is calculated to be Dimen = [0.66, 0, 0.7, 1, 0.5], and it can be found that the results are inconsistent in the last dimension.

[0387] 5. Intent results self-update

[0388] When the intent analysis results of the large language model do not meet expectations, this solution uses a label-based self-update mechanism to correct the results. For example, in 4, it is found that the description dimension is deviated. Then find out from the rule template label list of analyze_by_log that the label type of the "desctiption" field is "URL hyperlink", so extract "https: / / AA.com / test8" with the label "URL hyperlink" from the input "Flashback information in the analysis log: https: / / AA.com / test8" and replace it in the result.

[0389] The following is a description of the steps of updating rules based on tags in an embodiment of the present invention:

[0390] H100, obtain the template to be updated;

[0391] H200, traverse the template to be updated to obtain a tag list; wherein each tag in the tag list is used to characterize the field type of the scene recognition result;

[0392] H300, generates the key name of the label template vector according to the label list;

[0393] H400, according to the key name, obtain multiple values ​​corresponding to the key name from the template to be updated, and construct multiple key-value pairs; wherein the key of the key-value pair is used to represent the label of the scene recognition result, and the value of the key-value pair is used to represent the field information corresponding to the label;

[0394] H500, updates each key-value pair to the scene tag template library.

[0395] Specifically, the feature matching process of the legality verification update in the embodiment of the present invention is:

[0396] When it is determined that the label rule needs to be updated, start generating a standard label template for the intent result; then extract the prompt template uploaded by the user, traverse the prompt template to obtain a list of all the user's intents; then generate the first dimension value of the result vector based on the list of all key names of the intent; sort by key name, traverse all key-value pairs in a single intent, generate a label list for the value, and hash the label list of the value to generate the nth dimension result in the result vector; finally, persist the intent rules (including the label list and the corresponding multi-dimensional vector value).

[0397] The following further describes the specific steps for verifying the legitimacy of the intent result when there is no need to update the label rules:

[0398] First, generate a label vector for the current latest input intent recognition result, find the same data as the current intent result from the intent template, and calculate the distance between the two; then determine whether the two match based on the calculated distance. If they match, the intent recognition result is considered legal; if they do not match, traverse the results of each dimension of the label vector of the current intent result one by one, calculate the distance between the result of each dimension and the corresponding object in the intent template, and finally obtain a list of dimensions that do not meet the legality requirements and the corresponding key names; find the label list corresponding to each key name, and then extract the label in the actual input to update the result.

[0399] In summary, in the abnormal business analysis process of the present invention, since the language model performs intent analysis on the received abnormal data set in the early stage, some inaccurate intent recognition problems (such as missing fields, inaccurate field values, inaccurate intent actions, etc.) will appear, so the above-mentioned method for verifying the legitimacy of the intent result is proposed. In the process of the embodiment of the present invention to verify the legitimacy of the intent result, since the presentation of the intent result is in the form of a key-value pair, the embodiment of the present invention mainly detects the correlation of each value in the key-value pair, specifically calculating the distance between the data corresponding to the intent result and the intent template, and using this as the correlation detection result, and based on the result of the correlation detection, the accuracy of the intent recognition result made by the language model is determined. Through the method of the embodiment of the present invention, it is helpful to improve the accuracy of the results returned by the language model, and then improve the accuracy of the final abnormal business analysis results. In addition, before performing the legitimacy check, the embodiment of the present invention can construct a label rule, which can be updated and supplemented according to the previous recognition results of the intent recognition, and can also be expanded and supplemented through professional opinions in related fields, so that the object coverage of the legitimacy check is wider and more practical.

[0400] See also Fig. 9Another aspect of the embodiment of the present invention provides an abnormal service analysis device 900, comprising:

[0401] The first module 901 is used to obtain a set of abnormal information to be analyzed, and generate a descriptive language corresponding to each abnormal information in the abnormal information set; wherein each abnormal information comes from a different business platform;

[0402] The second module 902 is used to input the descriptive language into the pre-trained language model for scene recognition, and determine the first scene information corresponding to each descriptive language;

[0403] The third module 903 is used to determine the corresponding scene log requirement according to the first scene information;

[0404] The fourth module 904 is used to obtain log information from the corresponding business platform according to the scenario log requirements, and determine the abnormality analysis result of the abnormal information in the corresponding business scenario according to the log information.

[0405] See also Fig.10 , another aspect of an embodiment of the present invention provides an electronic device, including a processor and a memory;

[0406] The memory is used to store programs;

[0407] The processor executes the program to implement the above-mentioned abnormal business analysis method.

[0408] On the other hand, Fig.10 As shown, an embodiment of the present invention further provides an electronic device 1000, which may include at least one processor 1010 and may also include at least one memory 1020. The memory 1020 may be used to store at least one program, and the processor 1010 may execute the program to execute the aforementioned abnormal business analysis method; take a processor 1010 and a memory 1020 as an example.

[0409] The processor 1010 and the memory 1020 may be connected via a bus or other means.

[0410] The memory 1020 can be used as a non-transient computer-readable storage medium, which can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory 1020 may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory 1020 may include a memory remotely arranged relative to the processor, and these remote memories may be connected to the device via a network. Examples of the above-mentioned network may include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0411] The system embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separated, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment.

[0412] Specifically, Fig.11 The structure block diagram of a computer system for implementing an electronic device according to an embodiment of the present application is schematically shown.

[0413] It should be noted that Fig.11 The computer system 1100 of the electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0414] like Fig.11 As shown, the computer system 1100 includes a central processing unit 1101 (CPU), which can perform various appropriate actions and processes according to the program stored in the read-only memory 1102 (ROM) or the program loaded from the storage part 1108 to the random access memory 1103 (RAM). Various programs and data required for system operation can also be stored in the random access memory 1103. The central processing unit 1101, the read-only memory 1102 and the random access memory 1103 can be connected to each other through a bus 1104. An input / output interface 1105 (Input / Output interface, i.e., I / O interface) can also be connected to the bus 1104.

[0415] The following components connected to the input / output interface 1105 may include an input section 1106 such as a keyboard, a mouse, etc.; an output section 1107 such as a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 1108 such as a hard disk, etc.; and a communication section 1109 which may include a network interface card such as a local area network card, a modem, etc. The communication section 1109 may perform communication processing via a network such as the Internet. A drive 1110 may also be connected to the input / output interface 1105 as needed. A removable medium 1111, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., may be installed on the drive 1110 as needed, so that a computer program read therefrom may be installed into the storage section 1108 as needed.

[0416] The embodiment of the present invention also discloses a computer program product or a computer program, which may include a computer instruction, and the computer instruction is stored in a computer-readable storage medium. A processor of a computer device may read the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device executes the above-mentioned abnormal business analysis method.

[0417] According to an embodiment of the present application, the processes described in each method flow chart can be implemented as a computer software program. For example, an embodiment of the present application includes a computer program product, which includes a computer program carried on a computer readable medium, and the computer program contains program code for executing the method shown in the flow chart. In such an embodiment, Fig.11 The computer program may be downloaded and installed from a network through the communication section 1109, and / or installed from the removable medium 1311. When the computer program is executed by the central processor 1101, various functions defined in the system of the present application are executed.

[0418] In some selectable embodiments, the function / operation mentioned in the block diagram may not occur in the order mentioned in the operation diagram. For example, depending on the function / operation involved, the two boxes shown in succession can actually be executed substantially simultaneously or the boxes can sometimes be executed in reverse order. In addition, the embodiment presented and described in the flow chart of the present invention is provided by way of example, for the purpose of providing a more comprehensive understanding of technology. The disclosed method is not limited to the operation and logic flow presented herein. Selectable embodiments are expected, wherein the order of various operations is changed and the sub-operation of a part for which is described as a larger operation is performed independently.

[0419] In addition, although the present invention is described in the context of functional modules, it should be understood that, unless otherwise specified, one or more of the functions and / or features described may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in separate physical devices or software modules. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding the present invention. More specifically, in view of the properties, functions, and internal relationships of the various functional modules in the device disclosed herein, the actual implementation of the module will be understood within the conventional skills of the engineer. Therefore, those skilled in the art can implement the present invention set forth in the claims without excessive experimentation using ordinary techniques. It is also understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of the present invention, which is determined by the full scope of the appended claims and their equivalents.

[0420] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0421] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0422] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0423] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0424] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0425] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

[0426] The above is a specific description of the preferred implementation of the present invention, but the present invention is not limited to the described embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A method for analyzing abnormal business, characterized in that: include: Acquire a set of abnormal information to be analyzed, and generate descriptive language corresponding to each abnormal information in the set of abnormal information; Inputting the descriptive language into a pre-trained language model for scene recognition, and determining first scene information corresponding to each of the descriptive languages; Determining a corresponding scenario log requirement according to the first scenario information; Obtaining log information from the corresponding business platform according to the scenario log requirements; Determine an exception analysis result of the exception information in a corresponding business scenario according to the log information.

2. The abnormal business analysis method according to claim 1, characterized in that: The obtaining of the set of abnormal information to be analyzed and generating descriptive language corresponding to each abnormal information in the set of abnormal information includes: In response to the problem feedback request of each of the business platforms, trigger the network hook corresponding to the problem feedback request, so as to obtain the abnormal information from each of the business platforms to which the problem feedback request belongs through the network hook; wherein the abnormal information is used to characterize the real-time data when the abnormal problem occurs in each of the business platforms; Collect historical solutions of each of the business platforms in solving platform problems; Descriptive language of the exception information and the historical resolution is generated.

3. The abnormal business analysis method according to claim 2, characterized in that: The step of obtaining the abnormal information set to be analyzed also includes: Extracting key information from the abnormal information obtained from each of the business platforms; Calculate the number of abnormal business platforms where the same key information appears at the same time; When the number of abnormal platforms meets the threshold condition of aggregate feedback, the abnormal information of each of the service platforms is aggregated to obtain aggregate feedback information.

4. The abnormal business analysis method according to claim 1, characterized in that: The step of inputting the descriptive language into a pre-trained language model for scene recognition to determine the first scene information corresponding to each of the descriptive languages ​​includes: Mapping and associating the selected multiple scene names with the scene description words corresponding to each of the scene names to construct a scene mapping table; Recognizing and converting each of the descriptive languages ​​in the descriptive language set by using the language model to obtain keywords of the descriptive language; According to the degree of association between the keyword and each of the scene description words in the scene mapping table, one or more candidate scene names corresponding to each descriptive language are determined as the first scene information.

5. The abnormal business analysis method according to claim 4, characterized in that: The step of inputting the descriptive language into a pre-trained language model for scene recognition and determining the first scene information corresponding to each of the descriptive languages ​​further includes: Collect experience documents used to solve business anomalies and perform data cleaning on the experience documents; Inputting the cleaned experience document into the language model to obtain a question-answering dataset that meets the preset format requirements; Performing data filtering and formatting on the question-answering dataset to construct a pre-training dataset; The parameters of the language model are adjusted according to the pre-training data set, and information of the experience document is added to the language model.

6. The abnormal business analysis method according to claim 4, characterized in that: In the scene mapping table, each scene is correspondingly configured with a scene identifier, each scene identifier is configured with a corresponding scene log requirement, and determining the corresponding scene log requirement according to the first scene information includes: Determine one or more scene identifiers from the scene mapping table according to the first scene information; Summarizing the scene log requirements corresponding to the scene identifier determined from the first scene information; The scenario log requirement is used to characterize the log information to be collected by the business platform.

7. The abnormal business analysis method according to claim 1, characterized in that: The obtaining corresponding log information from each of the business platforms according to the scenario log requirements, and determining the abnormality analysis result of the abnormal information in the corresponding business scenario according to the log information, includes: Obtain corresponding log information from each of the business platforms according to the scenario log requirements; wherein the log information includes a log collection timestamp, a business object account, and an abnormal log segment; Perform regular matching on the abnormal log segment of the log information according to a preset analysis script to determine the target log line; An exception analysis is performed on the target log line according to the experience data of the exception analysis library to determine the exception analysis result under the corresponding business scenario.

8. The abnormal business analysis method according to claim 7, characterized in that: The business scenario includes a general scenario, an audio and video usage scenario, or a shared screen viewing scenario, and performing an exception analysis on the target log line according to the experience data of the exception analysis library to determine the exception analysis result under the corresponding business scenario also includes: In the general scenario, extract information from the target log line to obtain business platform version information and business object account number to determine an abnormality analysis result of the business object in the general scenario; In the audio and video usage scenario, information is extracted from the target log line to obtain interface operation information and audio device fault information of the business object, so as to determine an abnormal analysis result of the business object in the audio and video usage scenario; In the shared screen viewing scenario, information is extracted from the target log line to obtain shared operation information and virtual location information of the business object, so as to determine an abnormal analysis result of the business object in the shared screen viewing scenario.

9. The abnormal business analysis method according to any one of claims 1 to 8, characterized in that: The method further comprises: Vectorizing the descriptive language corresponding to each abnormal information in the abnormal information set to obtain a problem vector; Vectorize the experience documents on solving exceptions to obtain an experience vector set; Performing vector matching on the question vector according to the experience vector set, and obtaining a plurality of experience vectors whose approximation is higher than a target threshold as context information of the question vector; The question vector and the context information are combined to generate a descriptive language corresponding to the abnormal information.

10. The abnormal business analysis method according to claim 9, characterized in that: The vectorization of the descriptive language corresponding to each abnormal information in the abnormal information set to obtain a problem vector includes: According to a preset field set, serializing each field in the field set in the descriptive language to obtain a structured descriptive language; Each field is processed in blocks to obtain block values ​​of multiple sub-blocks; Determine a peak block of the current field according to the block values ​​of each sub-block, and obtain a character string of the peak block; Extracting keywords from the character string of the peak block, and compressing the usage frequency of each word in the character string with the extracted keywords to obtain a normalized character string; updating the structured descriptive language according to the normalized character string; After vectorization processing is performed on the updated structured descriptive language, the problem vector is obtained.

11. The abnormal business analysis method according to any one of claims 1 to 10, characterized in that: The method further comprises: The abnormal analysis result is structurally decomposed to determine a plurality of information components in the abnormal analysis result; and each of the information components is mapped to obtain a descriptive language corresponding to the abnormal analysis result.

12. The abnormal business analysis method according to any one of claims 1 to 10, characterized in that: The acquiring log information from the corresponding business platform according to the scenario log requirement also includes: Matching the scenario log requirement with each candidate input template cached in a preset cache rule library to determine a target input template; According to the target input template, acquiring historical cache information corresponding to the descriptive language from a preset cache result library; The historical cache information is used as a request result of the scene log requirement, and the request result is output as the log information.

13. The abnormal business analysis method according to any one of claims 1 to 10, characterized in that: The method further comprises: Obtaining a scene recognition result of the language model, and generating a vector to be verified of the scene recognition result; According to the vector to be verified, a corresponding label template vector is obtained by matching from a scene label template library; Calculating the vector distance between the vector to be checked and the label template vector; Determining a legitimacy verification result of the scene recognition result according to the vector distance; When the legitimacy check result of the scene recognition result is illegal, calculating the distance between each label dimension in the vector to be checked and the corresponding dimension in the label template vector to determine the illegal vector dimension; Determine a label to be updated according to the illegal vector dimension; According to the tag to be updated, field information corresponding to the tag to be updated is extracted from the descriptive language to replace the scene recognition result.

14. The abnormal business analysis method according to claim 13, characterized in that: The method further comprises the step of updating the scene label template library, the step comprising: Get the template to be updated; Traversing the template to be updated to obtain a tag list; wherein each tag in the tag list is used to characterize a field type of the scene recognition result; Generate a key name of a label template vector according to the label list; According to the key name, multiple values ​​corresponding to the key name are obtained from the template to be updated to construct multiple key-value pairs; wherein the key of each key-value pair is used to represent a label of a scene recognition result, and the value of the key-value pair is used to represent field information corresponding to the label; Update each of the key-value pairs to the scene label template library.

15. An abnormal business analysis device, characterized in that: include: The first module is used to obtain a set of abnormal information to be analyzed, and generate a descriptive language corresponding to each abnormal information in the abnormal information set; wherein each abnormal information comes from a different business platform; The second module is used to input the descriptive language into the pre-trained language model to perform scene recognition and determine the first scene information corresponding to each of the descriptive languages; A third module is used to determine the corresponding scene log requirement according to the first scene information; The fourth module is used to obtain corresponding log information from each of the business platforms according to the scenario log requirements, and determine the abnormality analysis results of the abnormal information in the corresponding business scenario according to the log information.

Images