Log anomaly detection method and device, equipment, computer program product and medium

By building a multi-dimensional text graph and gated graph neural network, combining self-supervised learning and large language models, anomaly information reports are generated, which solves the problems of low accuracy and high cost of existing log anomaly detection methods, and achieves efficient logging Anomaly detection and system security improvement.

CN118916234APending Publication Date: 2024-11-08CHINA MOBILE (SUZHOU) SOFTWARE TECH CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410979169.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-22
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

Existing log anomaly detection methods are difficult to provide effective exception information reporting, and the detection accuracy is low. They rely on the experience of operation and maintenance personnel, and are costly. It is difficult to capture the discontinuous and long-distance semantic relationships in log data, and the log analysis is generalized. Poor performance affects detection accuracy.

Method used

By constructing a multi-dimensional text graph, combining sequence, grammar, and semantic features, a gated graph neural network is used to perform feature extraction and abnormal detection, and a self-supervised learning strategy is introduced, and an exception information report is generated using a large language model to provide exception handling suggestions.

Benefits of technology

It improves the accuracy and robustness of log abnormality detection, reduces the pressure on operation and maintenance personnel, responds quickly to system problems, reduces the cost of log data labeling, and enhances system security.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118916234A_ABST
    Figure CN118916234A_ABST
Patent Text Reader

Abstract

The invention provides a log anomaly detection method, device and equipment, a computer program product and a medium, and relates to the technical field of security. The method comprises the following steps: performing anomaly detection on a to-be-processed log to obtain detection result information; under the condition that the detection result information shows that the to-be-processed log is abnormal, iteration processing is conducted on the detection result information, an abnormal information report is generated, and the abnormal information report comprises abnormal processing suggestion information for the to-be-processed log. According to the scheme, the problem that an existing log anomaly detection method is difficult to provide an effective anomaly information report is solved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of security technology, and in particular to a log anomaly detection method, device, equipment, computer program product and medium. Background Art

[0002] With the rapid development of cloud computing, computer software systems rely on a distributed environment centered on large server clusters, resulting in a rapid increase in the number of servers, larger and larger systems, and increasingly more difficult system maintenance, and the pressure of operation and maintenance. Along with this come various vulnerabilities and failures, including external failures such as malicious attacks and internal failures such as memory overloads, which in turn lead to system crashes or delays that may cause significant economic losses. Therefore, how to accurately detect these anomalies is crucial to building a secure and reliable system, and system log data is the most common and main data source for anomaly detection models in the system, playing an important role in computer security. Therefore, log anomaly detection is very important for system security.

[0003] However, the existing log anomaly detection method only analyzes whether the log data is abnormal. After learning of the anomaly, the operation and maintenance personnel need to rely on their own experience or investigate and analyze the specific abnormal situation to find a solution. Summary of the invention

[0004] The purpose of the present invention is to provide a log anomaly detection method, device, equipment, computer program product and medium, which solves the problem that the existing log anomaly detection method is difficult to provide effective anomaly information report.

[0005] To achieve the above object, an embodiment of the present invention provides a log anomaly detection method, comprising:

[0006] Perform anomaly detection on the logs to be processed and obtain detection result information;

[0007] In the case where the detection result information indicates that the log to be processed is abnormal, the detection result information is iteratively processed to generate an abnormal information report, wherein the abnormal information report includes abnormal processing suggestion information for the log to be processed.

[0008] To achieve the above object, an embodiment of the present invention provides a log anomaly detection device, comprising:

[0009] Anomaly detection module, used to perform anomaly detection on the logs to be processed and obtain detection result information;

[0010] The report generation module is used to iteratively process the detection result information and generate an abnormal information report when the detection result information indicates that the log to be processed is abnormal, and the abnormal information report includes abnormal processing suggestion information for the log to be processed.

[0011] To achieve the above objective, an embodiment of the present invention provides a device, including a processor and a transceiver, wherein the processor is configured to:

[0012] Perform anomaly detection on the logs to be processed and obtain detection result information;

[0013] In the case where the detection result information indicates that the log to be processed is abnormal, the detection result information is iteratively processed to generate an abnormal information report, wherein the abnormal information report includes abnormal processing suggestion information for the log to be processed.

[0014] To achieve the above-mentioned purpose, an embodiment of the present invention provides a device, including a transceiver, a processor, a memory, and a program or instruction stored in the memory and executable on the processor; when the processor executes the program or instruction, the above-mentioned log anomaly detection method is implemented.

[0015] To achieve the above object, a computer program product is provided, including computer instructions, which, when executed by a processor, implement the steps of the above log anomaly detection method.

[0016] To achieve the above objective, an embodiment of the present invention provides a readable storage medium on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps in the above log anomaly detection method are implemented.

[0017] The beneficial effects of the above technical solution of the present invention are as follows:

[0018] The method of the embodiment of the present invention, after performing anomaly detection to obtain detection result information, can iteratively process the detection result information, generate an anomaly information report, and provide anomaly processing suggestion information for the log to be processed in a normal information report. In this way, it can assist relevant operators to accurately and effectively identify log anomaly information, speed up the decision-making process, and thus quickly respond to and solve system problems, reduce the pressure on operation and maintenance personnel, improve system security, and have broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Flow chart of a log anomaly detection method according to an embodiment of the present invention;

[0020] Figure 2 This is an example diagram of sentence dependency analysis in an embodiment of the present invention;

[0021] Figure 3 It is an iterative optimization flow chart of an embodiment of the present invention;

[0022] Figure 4 It is an iteration prompt diagram of an embodiment of the present invention;

[0023] Figure 5 Schematic diagram of the overall structure of the log anomaly detection system according to an embodiment of the present invention;

[0024] Figure 6 This is a structural diagram of a log anomaly detection module according to an embodiment of the present invention;

[0025] Figure 7 A structural diagram of a log anomaly detection device according to an embodiment of the present invention;

[0026] Figure 8 is a structural diagram of a device according to an embodiment of the present invention;

[0027] Fig. 9 This is a structural diagram of a device according to another embodiment of the present invention. DETAILED DESCRIPTION

[0028] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0029] It should be understood that the references to "one embodiment" or "an embodiment" throughout the specification mean that the specific features, structures, or characteristics associated with the embodiment are included in at least one embodiment of the present invention. Therefore, the references to "in one embodiment" or "in an embodiment" appearing throughout the specification do not necessarily refer to the same embodiment. In addition, these specific features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0030] In various embodiments of the present invention, it should be understood that the size of the serial numbers of the following processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0031] Additionally, the terms "system" and "network" are often used interchangeably herein.

[0032] In the embodiments provided in the present application, it should be understood that "B corresponding to A" means that B is associated with A, and B can be determined according to A. However, it should also be understood that determining B according to A does not mean determining B only according to A, and B can also be determined according to A and / or other information.

[0033] Due to the large differences in log data formats, high heterogeneity and complexity, and imbalance between normal and abnormal log data, log anomaly detection has certain difficulties. The following technical problems exist in related technologies:

[0034] (1) Existing log anomaly detection methods only analyze whether the log data is abnormal. After learning of the anomaly, operation and maintenance personnel need to rely on their own experience or investigate and analyze the specific abnormal situation to find a solution.

[0035] (2) Existing log anomaly detection methods only focus on keywords in log data and the relationship between keywords. However, since log data is short text with sparse features, existing log anomaly detection methods are difficult to effectively solve the problem of keyword polysemy, resulting in low performance of anomaly detection models.

[0036] (3) Most of the existing log anomaly detection methods are based on deep learning methods, such as the BiGRU neural network model based on the attention mechanism and the temporal convolutional network (TCN). Most of the deep learning-based methods only capture local features in a small window, which makes it difficult to effectively capture the non-continuous and long-distance semantic relationships between words in the log data, and it is difficult to distinguish the contribution of different words in the log data, resulting in poor detection accuracy.

[0037] (4) Most of the existing log anomaly detection methods use log parsing methods for preprocessing. Log parsing refers to preprocessing semi-structured log data and converting it into specific event templates. However, considering that computer software systems are constantly updated, log types are updated quickly, and log data vocabulary is constantly increasing, log templates also need to be constantly changed. However, the existing log templates have poor generalization and are difficult to apply to all types of logs. In addition, feature extraction is not comprehensive. The small number of features in the log template is difficult to cover all the information in the log data, which can easily affect the accuracy of subsequent detection and result in low detection accuracy. Moreover, the deviations or errors caused by log parsing in the preprocessing stage often cause anomaly detection to drop by an order of magnitude in performance.

[0038] (5) The accuracy of existing log anomaly detection methods is heavily dependent on the amount and accuracy of log data. However, the cost of collecting and annotating large amounts of log data is very high, resulting in high costs for log anomaly detection.

[0039] like Figure 1 As shown, a log anomaly detection method according to an embodiment of the present invention includes:

[0040] Step 101: perform anomaly detection on the log to be processed and obtain detection result information.

[0041] Through the detection result information obtained in this step, it can be known whether the log to be processed is abnormal, so as to determine whether the system (here, the system that generates the log to be processed) has an abnormality.

[0042] Step 102, when the detection result information indicates that the log to be processed is abnormal, iteratively process the detection result information to generate an abnormal information report, wherein the abnormal information report includes abnormal processing suggestion information for the log to be processed.

[0043] It should be noted that traditional log anomaly detection can only detect whether the log is abnormal. In the implementation of the present invention, when a log anomaly is detected, the detection result information can be further iteratively processed and summarized into an abnormal information report. The abnormal information report also provides abnormal processing suggestion information for the log to be processed, which is provided to the operator so that he can quickly resolve the fault and maintain system security.

[0044] In this embodiment, after performing anomaly detection to obtain detection result information, the detection result information can be iteratively processed to generate an anomaly information report, and anomaly processing suggestion information for the log to be processed can be given in the normal information report. In this way, relevant operators can be assisted to accurately and effectively identify log anomaly information, speed up the decision-making process, and thus quickly respond to and solve system problems, which can reduce the pressure on operation and maintenance personnel, improve system security, and has broad application prospects.

[0045] In some embodiments, step 101, performing anomaly detection on the log to be processed to obtain detection result information, may specifically include the following steps:

[0046] Step 1011: construct a text graph corresponding to the log to be processed, wherein the text graph includes at least two of a sequence graph, a syntax graph, and a semantic graph.

[0047] In this step, the text graph corresponding to the log to be processed can be constructed based on the mapping method of the three dimensional features of sequence, grammar, and semantics. It should be noted that the sequence graph can describe the local word co-occurrence through point-by-point mutual information, the grammar graph can describe the grammatical rule attributes through word dependencies, and the semantic graph can extract semantic information through topic information and topic keywords, so as to capture the contextual information of different language attributes from multiple attributes, solve the problems of sparse text log features, polysemy of keywords, and difficulty in effectively capturing non-continuous and long-distance context information, and improve the anomaly detection performance.

[0048] Here, we take the example of constructing a text graph based on the three dimensions of sequence, syntax, and semantics. The text graph can be expressed as T = (G1, G2, G3). Among them, T is a third-order text graph tensor, G i (i=1,2,3) is the i-th text graph in T. Gi Defined as G i =(V i ,E i ), where V i is a node set, E i is an edge set. The node set contains node types and topic types. The node type is the word in the log, and the topic type can be the topic word extracted by the Bi-Term topic model. It should be noted that each G i Both involve a feature matrix X i and the adjacency matrix A i , where the feature matrices of all graphs are packaged into a text graph tensor γ = (X1, X2, X3) and initialized uniformly using the distributed word representation model Global Vector (which can be used to obtain word vector representations so that vectors contain as much semantic and grammatical information as possible); the adjacency matrices of all graphs are packaged into an adjacency matrix tensor Λ = (A1, A2, A3).

[0049] Step 1012: extract features from the text graph to obtain node feature information corresponding to the text graph.

[0050] Specifically, feature extraction of text graphs can be implemented based on gated graph neural networks. Feature extraction obtains node feature information, that is, log text representation.

[0051] Step 1013: Generate a target text graph (G a ).

[0052] In some optional embodiments, step 1013 may specifically include: determining the weight corresponding to each of the text graphs; performing weighted calculation according to the node feature information and the weight corresponding to each of the text graphs to obtain the target text graph (G a ).

[0053] Here, considering the sequence graph G seq 、Grammar diagram G syn , semantic graph G sem The importance is not the same. In order to measure the different contributions of each graph G, an attention mechanism can be introduced to learn the weight of the graph. The attention mechanism allows the model to give different weights to specific inputs, and finally obtain the target text graph G. a , and its calculation formula is as follows:

[0054]

[0055] u t =tanh(W t G t +b t )

[0056]

[0057] Among them, G t represents the node feature information corresponding to the t-th text graph (t=1, 2, 3), β t Represents each t-th text graph G t The corresponding weight (β t To measure G t For generating the target text graph G a The importance of u t Indicates that G t After being fed into the full embedding layer, G t The hidden representation of W t represents the trainable weights, b t Represents a trainable bias parameter.

[0058] In this way, the attention mechanism can be used to measure the importance of sequence graphs, grammatical graphs, and semantic graphs of different dimensions, so as to better describe the contextual information of words, help the model better understand log data, improve model classification accuracy, and avoid misjudgment of word meaning.

[0059] Step 1014, using the Readout function, obtain the feature representation (h a ).

[0060] It should be noted that the final text graph (i.e., the target text graph G) is obtained in step 1013. a ), the nodes in the target text graph have been fully updated. Here, a Readout function can be used to obtain the node representation (i.e., the feature representation h corresponding to the node) a ), improve the contribution of keywords. This process can be expressed by the following formula:

[0061] x i =σ(f1(x i ))⊙tanh(f2(x i ))

[0062]

[0063] Among them, f1 and f2 are multilayer perceptrons (MLP). The former is a soft attention weight that determines which nodes are related to the current task; the latter is a nonlinear feature change. Maxpooling is used to select the keywords that contribute most to the current graph task.

[0064] x1,x2,…x nrepresents the keywords obtained through the multilayer perceptron, n represents the total number of keywords, x i represents the i-th keyword, Represents the average weighted word.

[0065] Step 1015: Obtain the detection result information according to the feature representation corresponding to the node in the target text graph.

[0066] In some embodiments, before the above step 1015, the method further includes:

[0067] (i) Preprocessing the log sample data in a preset manner to obtain target log sample data, wherein the preset manner includes at least one of the following: a masked language modeling method, a document rotation method, and a center word prediction method.

[0068] It should be noted that in log anomaly detection (that is, anomaly classification task), different attribute information of log data is obtained by combining sequence, grammar, semantics and other features. Therefore, three auxiliary tasks, namely masked language modeling, document rotation and prediction of central words, can be selected here to preprocess the log sample data. The specific introduction is as follows:

[0069] (1) Masked language modeling method: Randomly use special symbols to mask (MASK) some word tokens. The task is to predict the masked real token based on the context word information that is not masked.

[0070] (2) Document rotation method: a random word token in the text log is used as the rotation point, and this token is used as the starting word. The original document rotation task is to restore the document before rotation. The document rotation task in the embodiment of the present invention can be to determine whether the text log is rotated.

[0071] (3) Center word prediction method: Take a piece of text with a certain window size, and the task is to predict the center word based on the surrounding words.

[0072] (ii) Based on the target log sample data, an initial model is trained to obtain a target model, where the target model is used to obtain the detection result information.

[0073] It should be noted that in the text log (i.e. log sample data) log = (x1, x2, ... x n ) can be randomly used to generate the converted text log (that is, the target log sample data) log′=(x1′,x′2,...x′) before being sent to the model. n) and its label, and each time log and log′ are sent to the network as a set of data for training. Among them, the target log sample data (log′) does not participate in the classification prediction and keyword extraction of log anomaly detection, but only participates in three auxiliary tasks (i.e., judging whether log rotation is required, generating mask words, and predicting central words).

[0074] In this embodiment, in order to solve the problem of log data labeling, in log anomaly detection, three methods of masked language modeling, document rotation and prediction of central words are used to strengthen model training, solve the problems of log data labeling and dependence on large amounts of log data, and improve model performance.

[0075] Since the supervised anomaly prediction task depends on the quantity and accuracy of log data, and the cost of collecting and annotating a large amount of log data is very high, the embodiment of the present invention introduces a self-supervised learning strategy, which autonomously learns the intrinsic structural information of log data through three specific auxiliary tasks (including masked language modeling, document rotation, and prediction of central words) to improve the performance of the main task of anomaly classification and keyword extraction.

[0076] It should be noted that the training method of self-supervised learning is supervised learning, that is, learning the input-output mapping function from labeled supervised data, but the label of the input data is generated by itself. The core idea is to use part of the information of the input data to predict another part of the information in some form.

[0077] Optionally, in step 1011, constructing a text graph corresponding to the log to be processed includes at least two of the following:

[0078] (i) extracting the sequence features of the log to be processed and constructing the sequence graph (G seq ).

[0079] It should be noted that the sequence graph can describe the local co-occurrence properties between words, so it can be used in text classification based on graph convolutional neural networks.

[0080] Here, pointwise mutual information (PMI) can be used to describe the sequence information, that is, the edge weight of two word nodes. The larger the value of PMI, the higher the frequency of shared occurrence of the two words. The calculation formula of PMI is:

[0081] in, Represents word pair (x i ,x j ) appears in the log text log, W(i,j) represents the probability that the sliding window contains word x at the same time i and xj , |W| represents the total number of words in the sliding window. Representation word x i The probability of occurrence in log, W(x i ) indicates that the sliding window contains word x i The number of j ) represents word x j The probability of appearing in log. It should be noted that in this embodiment, only the edges with positive PMI are retained.

[0082] (ii) extracting the grammatical features of the log to be processed and constructing the grammar graph (G syn ).

[0083] It should be noted that grammatical analysis is the process of parsing natural language using grammatical rules. Here, we can specifically use Stanford CoreNLP (a Java-based natural language processing tool that supports word segmentation, syntactic component analysis, dependency syntactic analysis and other functions in multiple languages) to generate dependency analysis to describe the grammatical features of the sentence. Dependency analysis is used to obtain the dependency relationship between words. In a sentence, different word dependencies can be extracted. Each word dependency is a directed triple consisting of a relation, a governor, and a dependent word. For example, Figure 2 The following is an illustration of the word dependency relationship in the sentence “Error has been observed at the following sites.”

[0084] Here, we focus on word pairs (control words, dependent words). We can filter out word pairs where both the control words and the dependent words are stop words (such as prepositions, articles, etc.). The word pairs that are retained have at least one word of a specified part of speech (such as nouns, verbs, adjectives, etc.). syn In , the edge weight of each pair of words can be defined as E(governor,dependent)=1, for example, E(Error,observed)=1, E(following,sites)=1, etc.

[0085] (III) extracting the semantic features of the log to be processed and constructing the semantic graph (G sem ).

[0086] Here, in order to enrich the contextual information of words, the correlation between words and topics can be used to obtain semantic features. It should be noted that the topic model is a commonly used model for automatically extracting document topic information in natural language processing (NLP). Common algorithms include latent Dirichlet allocation (LDA), lightning search algorithm (LSA), etc. Considering that a single log data is a short text, the use of the above-mentioned traditional topic model will result in poor results because the words in the log text are too sparse. Therefore, the embodiment of the present invention can specifically use the Bi-Term Topic Model (BTM) to analyze log topics. BTM uses the biterms word co-occurrence pattern to model short texts, which can obtain more representative global topics. The overall process of BTM mainly includes:

[0087] For each topic z, extract the topic-word distribution And the topic distribution θ~Dir(α), where α and β are obtained by Dirichlet prior; for each biterm, extract a topic z~Multi(θ) and two words By formula: We can get a bigram (ω i ,ω j ); the distribution of k specified topics in the log text generated by BTM, and the topic-word matrix of the word distribution under each topic, and the weight of each topic-word in the matrix is ​​used to define the edge weight between the word node and the topic node: E i,j =word-topic i,j . Among them, i represents a word and j represents a topic.

[0088] Optionally, step 1012, extracting features from the text graph to obtain node feature information corresponding to the text graph, includes: extracting features from the text graph based on a gated graph neural network to obtain node feature information corresponding to the text graph.

[0089] Here, after constructing the multidimensional text graph, it can be input into the gated graph neural network (GGNN) to obtain the log text representation (that is, the node feature information corresponding to the text graph). The gated graph neural network based on self-supervised multidimensional dynamics does not need to convert log data into event templates, thereby avoiding the impact of errors in the log parsing stage on subsequent model classification.

[0090] It should be noted that GGNN is able to learn local and global structures from the graph by leveraging interactions between nodes. That is, in GGNN, a node receives messages from neighboring nodes and then merges these messages into itself, thereby updating its own representation information. As the number of GGNN layers increases, this message passing and merging process is repeated to recursively update the node representation. Since the graph layer only works on first-order neighbor nodes, the layer can be stacked t times to achieve high-order feature interactions. This allows the embedded learning node to contain information about t-hop neighbors.

[0091] In this embodiment, a method based on a gated graph neural network is used for feature extraction. Compared with the deep learning method, the embodiment of the present invention can effectively distinguish the importance of words in the log data by encoding the syntactic structure of the log data.

[0092] Optionally, in step 1015, obtaining the detection result information according to the feature representation corresponding to the node in the target text graph includes at least one of the following:

[0093] (i) performing anomaly prediction according to feature representations corresponding to nodes in the target text graph, obtaining a category label, and determining the category label as the detection result information, wherein the category label is used to indicate whether the log to be processed is abnormal or normal.

[0094] It should be noted that in step 1014, the node representation (i.e., the feature representation h corresponding to the node) is obtained through the readout layer (through the Readout function). a ), we need to perform the abnormal classification task, that is, to determine whether the text log (log to be processed) is abnormal. As a binary classification task, h a The predicted category label is sent to the Softmax layer and obtained by minimizing the loss function. The loss function adopts the cross entropy loss function, which is defined as follows:

[0095]

[0096] Among them, W represents weight, b represents bias, and L represents loss function; Represents the one-hot label of the i-th element.

[0097] (ii) extracting keywords based on the feature representations corresponding to the nodes in the target text graph, obtaining abnormal keywords, and adding the abnormal keywords to the detection result information.

[0098] That is to say, the log anomaly detection method of the embodiment of the present invention can perform a keyword extraction task in addition to the anomaly classification task, which can output anomaly keywords, such as inputting the keywords ranked in the top N in descending order of importance. It should be noted that h aIt can describe the contribution of keywords and the relevance between keywords and topics. a After inputting into the Sigmoid function, the nodes sorted by weight can be obtained, thereby obtaining the most important N words. For example, N is set to 10, that is, the first 10 keywords sorted from high to low by weight are taken as abnormal keywords, which can be used when generating abnormal information reports.

[0099] Optionally, in step 102, the iterative processing of the detection result information to generate an abnormal information report includes:

[0100] Step 1021, initializing the target information to the log to be processed.

[0101] It should be noted that, at the beginning of the iterative processing, the log to be processed may be used as the target information, and then, with the subsequent iterative process, the target information will be further updated.

[0102] Step 1022: Based on the large language model, obtain a target log report corresponding to the target information.

[0103] It should be noted that in the field of natural language processing, models such as BERT and GPT are trained based on a large amount of text data, which can learn rich context representations and generate human-like language. The text summarization task is fine-tuned based on these models to obtain excellent performance. Since pre-trained models such as BERT still require a large amount of labeled data to complete downstream tasks, large language models (LLM) such as GPT-4, LLaMA and BloomZ adopt the framework of context learning, which can fine-tune the model in the absence of a large amount of annotated data. Therefore, the target log report corresponding to the target information can be obtained based on the large language model. Specifically, the large language model (LLM) in the embodiment of the present invention can be implemented using ChatGPT, or different large language models such as GPT-4, LLaMA and BloomZ can be used. ChatGPT is based on the GPT-3.5 model and uses training data in dialogue mode to promote users to have good human-computer interaction actions. In the following embodiments, ChatGPT is used as an example for explanation.

[0104] Step 1023 , performing a similarity evaluation between the target log report and a preset reference log report to obtain a target evaluation value corresponding to the target log report.

[0105] In this step, in order to evaluate the LogSUM generated by the large language model ChatGPT iteration LLM(also known as the target log report) and the reference log exception summary LogSUM REF The similarity between the generated result and the manual standard can be measured by Rouge (Recall-Oriented Understudy for Gisting Evaluation). Rouge-N measures the similarity between the generated result and the manual standard word by word, and evaluates the quality of the generated result by counting the basic units that overlap between the two, including grammar, word sequence, and word pair. N is the n-gram, and here we use 1-gram, that is, Rouge-1. The calculation formula is as follows:

[0106]

[0107] Among them, Count match (gram n ) represents the number of n-grams that match between the candidate result and the reference standard, Count(gram n ) represents the number of n-grams in the reference standard.

[0108] Step 1024: Update the target information according to the target evaluation value corresponding to the target log report.

[0109] Step 1025, repeatedly executing the steps of obtaining the target log report corresponding to the target information based on the large language model, to updating the target information according to the target evaluation value corresponding to the target log report, until the maximum number of iterations is reached.

[0110] Step 1026: determine the target log report corresponding to the target evaluation value with the largest value as the abnormal information report.

[0111] In this embodiment, based on a large language model and abnormal keywords, an optimal log abnormality summary can be obtained through iterative updates of an automatic evaluation algorithm.

[0112] In some optional embodiments, the detection result information includes at least one abnormal keyword; the above step 1024 may specifically include at least one of the following:

[0113] (1) when the target evaluation value corresponding to the target log report is greater than or equal to a threshold, updating the target information according to the target log report and the log to be processed;

[0114] (2) When the target evaluation value corresponding to the target log report is less than the threshold, the target information is updated according to the target log report, the log to be processed and the at least one abnormal keyword.

[0115] Exemplarily, at least one abnormal keyword may be sorted from high to low according to weight, and the top N abnormal keywords are selected to update the target information.

[0116] Here, the above iterative process is explained in detail with examples:

[0117] like Figure 3 As shown, in the first iteration, the text log log = (x1, x2, ... x n ) is directly input into the large language model ChatGPT to obtain the target log report (LogSUM LLM ), using Rouge-1 to evaluate LogSUM LLM and reference log reports (LogSUM REF )’s similarity:

[0118] If the similarity is greater than the set threshold Th, then LogSUM is determined LLM For a good summary, add this judgment to the iteration prompt;

[0119] If the similarity is less than the set threshold Th, then LogSUM is determined LLM For the poor summary, it is also added to the iteration prompt, and from at least one abnormal keyword (key1, key2, ...key 10 ), add one keyword key at a time according to the weight from high to low n Merge into the iteration prompt until the iteration ends.

[0120] Finally, the target log report with the highest evaluation (ie, the highest similarity evaluation score) during the iteration is determined as the final abnormal information report for output.

[0121] The iterative optimization algorithm in the above iterative process is specifically shown in the following Algorithm 1:

[0122]

[0123]

[0124] In this iterative optimization algorithm, based on a large language model and abnormal keywords, the summarized abnormal information is automatically evaluated and iteratively updated, and ultimately the optimal log abnormality summary can be obtained.

[0125] It should be noted that the ability of the large language model ChatGPT to summarize abnormal information in log data can be optimized through iteration prompts. After the summary generated during the iteration process is compared with the reference summary for similarity, the obtained good and bad evaluations are added to the iteration prompts to provide further guidance for ChatGPT to ensure that the next generated summary is closer to a good summary while avoiding bad summaries. Specifically, if the score is less than the set threshold during the similarity comparison process, one keyword will be selected from the sorted keywords and merged into the iteration prompt to ensure that ChatGPT's summary can grasp the "key points" and does not deviate from expectations.

[0126] like Figure 4 As shown in the figure, the information input to ChatGPT is accessed in the form of a string through the Application Programming Interface (API) and is divided into three roles: system, user, and assistant. The system message initiates the conversation and provides a description of the task. The user message instructs the assistant and serves as the user's input. Finally, the assistant component generates a response for the model.

[0127] Specifically, first describe the task and initiate a conversation, define the role of ChatGPT as a system operator, and define the task of summarizing abnormal information in log data; then, add the summary after similarity evaluation to the iteration prompt in the role of user + assistant, and each iteration, good and bad summaries will be updated, so that ChatGPT can be optimized from summaries with different evaluations; if the similarity evaluation is a bad summary, the abnormal keywords will also be included in the prompt; at the end of the iteration, the final query part provides specific optimization rules to generate a query that is consistent with a good summary and avoids a bad summary.

[0128] In a specific embodiment of the present invention, a log anomaly detection system executes the log anomaly detection method provided by the embodiment of the present invention. The log anomaly detection system includes a log anomaly detection module and a log anomaly information summary module. The log anomaly detection module uses a gated graph neural network, and a self-supervised learning strategy can be introduced in the log anomaly detection module to improve model performance. The log anomaly information summary module can generate an anomaly information report including anomaly processing suggestion information for the log to be processed, so as to help operation and maintenance or developers effectively identify log anomaly information, speed up decision-making, and quickly respond to solve system problems. Figure 5As shown, the overall process of the log anomaly detection system mainly includes the following steps: preprocessing the obtained logs to be processed (i.e. text log1, log2, etc.), converting them into matrices using Word2Vec and inputting them into the log anomaly detection module. The log anomaly detection module then builds a graph based on the three dimensions of sequence, syntax, and semantics, and then uses the gated graph neural network to detect whether the logs to be processed are abnormal, and extract abnormal keywords. Afterwards, if the log anomaly detection module determines that the logs to be processed are abnormal, the logs to be processed are input into the log anomaly information summary module. Specifically, the logs to be processed that are predicted to be abnormal can be input into a large language model, and based on the extracted abnormal keywords, accurate log anomaly information (i.e., anomaly information report) can be summarized and output using an iterative optimization method.

[0129] like Figure 6 As shown, in another specific embodiment of the present invention, given a text log log = (x1, x2, ... x n ), which contains n words. First, extract the sequence, grammatical, and semantic features in the log to construct the sequence graph G seq 、Grammar diagram G syn , semantic graph G sem ; Then, based on the gated graph neural network, the nodes are updated and the graph attention is introduced to calculate the target text graph G a , the node representation is obtained through the readout layer (that is, the feature representation h corresponding to the node a ); Finally, h a Two branches are sent in: one branch is a classification predictor, which is used to predict whether the log is abnormal and obtain a binary classification label The other branch is a keyword extractor, which finds the TOP N nodes, that is, abnormal keywords, according to the node weight value sorting. The results of both branches will be used to generate abnormal information reports.

[0130] The log anomaly detection method of this embodiment is based on a self-supervised multi-dimensional dynamic gated graph neural network, which can realize automatic log anomaly detection, improve the accuracy and robustness of log anomaly detection, adapt to complex and changeable systems, and avoid security risks caused by low accuracy; it can also summarize abnormal information based on a large language model and output an abnormal information report, which includes abnormal handling suggestions, and can help operation and maintenance personnel speed up decision-making and quickly respond to and solve system problems, thereby strengthening the monitoring of the system operation status.

[0131] like Figure 7 As shown, a log anomaly detection device according to an embodiment of the present invention includes:

[0132] Anomaly detection module 710, used to perform anomaly detection on the log to be processed and obtain detection result information;

[0133] The report generation module 720 is used to iteratively process the detection result information and generate an abnormal information report when the detection result information indicates that the log to be processed is abnormal, and the abnormal information report includes abnormal processing suggestion information for the log to be processed.

[0134] In this embodiment, after performing anomaly detection to obtain detection result information, the detection result information can be iteratively processed to generate an anomaly information report, and anomaly processing suggestion information for the log to be processed can be given in the normal information report. In this way, relevant operators can be assisted to accurately and effectively identify log anomaly information, speed up the decision-making process, and thus quickly respond to and solve system problems, which can reduce the pressure on operation and maintenance personnel, improve system security, and has broad application prospects.

[0135] Optionally, the anomaly detection module 710 includes:

[0136] A first construction submodule is used to construct a text graph corresponding to the log to be processed, wherein the text graph includes at least two of a sequence graph, a syntax graph, and a semantic graph;

[0137] A feature extraction submodule is used to extract features from the text graph to obtain node feature information corresponding to the text graph;

[0138] A first generating submodule, used for generating a target text graph according to the node feature information;

[0139] The first acquisition submodule is used to obtain the feature representation corresponding to the node in the target text graph through the Readout function;

[0140] The second acquisition submodule is used to acquire the detection result information according to the feature representation corresponding to the node in the target text graph.

[0141] Optionally, the first building submodule includes:

[0142] A first construction unit is used to construct the sequence graph by extracting sequence features of the log to be processed;

[0143] A second construction unit is used to construct the syntax graph by extracting syntax features of the log to be processed;

[0144] The third construction unit is used to construct the semantic graph by extracting semantic features of the log to be processed.

[0145] Optionally, the feature extraction submodule includes:

[0146] A feature extraction unit is used to extract features from the text graph based on a gated graph neural network to obtain node feature information corresponding to the text graph.

[0147] Optionally, the first generating submodule includes:

[0148] A first processing unit, used to determine the weight corresponding to each of the text graphs;

[0149] The second processing unit is used to perform weighted calculation according to the node feature information and the weight corresponding to each of the text graphs to obtain the target text graph.

[0150] Optionally, the second acquisition submodule includes:

[0151] An exception prediction unit, used to perform an exception prediction according to the feature representation corresponding to the node in the target text graph, obtain a category label, and determine the category label as the detection result information, wherein the category label is used to indicate whether the log to be processed is abnormal or normal;

[0152] The first extraction unit is used to extract keywords according to the feature representation corresponding to the node in the target text graph, obtain abnormal keywords, and add the abnormal keywords to the detection result information.

[0153] Optionally, the report generating module 720 includes:

[0154] A first processing submodule, used for initializing the target information as the log to be processed;

[0155] A report acquisition submodule, used for acquiring a target log report corresponding to the target information based on a large language model;

[0156] A first evaluation submodule, configured to perform a similarity evaluation between the target log report and a preset reference log report to obtain a target evaluation value corresponding to the target log report;

[0157] A first updating submodule, configured to update the target information according to a target evaluation value corresponding to the target log report;

[0158] A second processing submodule is used to repeatedly execute the steps of obtaining a target log report corresponding to the target information based on the large language model, and updating the target information according to a target evaluation value corresponding to the target log report, until a maximum number of iterations is reached;

[0159] The third processing submodule is used to determine the target log report corresponding to the target evaluation value with the largest value as the abnormal information report.

[0160] Optionally, the detection result information includes at least one abnormal keyword; and the first updating submodule includes:

[0161] A first updating unit, configured to update the target information according to the target log report and the log to be processed when the target evaluation value corresponding to the target log report is greater than or equal to a threshold;

[0162] A second updating unit is configured to update the target information according to the target log report, the log to be processed and the at least one abnormal keyword when the target evaluation value corresponding to the target log report is less than the threshold.

[0163] Optionally, the device further comprises:

[0164] A preprocessing module, used to preprocess the log sample data in a preset manner to obtain target log sample data, wherein the preset manner includes at least one of the following: a masked language modeling manner, a document rotation manner, and a center word prediction manner;

[0165] The model training module is used to train an initial model based on the target log sample data to obtain a target model, and the target model is used to obtain the detection result information.

[0166] It should be noted here that the above-mentioned log anomaly detection device provided in the embodiment of the present invention can implement all the method steps implemented in the above-mentioned log anomaly detection method embodiment, and can achieve the same technical effect. The parts and beneficial effects that are the same as those in the method embodiment will not be described in detail here.

[0167] like Figure 8 As shown, a device 800 according to an embodiment of the present invention includes a processor 810 and a transceiver 820, wherein the processor 810 is configured to:

[0168] Perform anomaly detection on the logs to be processed and obtain detection result information;

[0169] In the case where the detection result information indicates that the log to be processed is abnormal, the detection result information is iteratively processed to generate an abnormal information report, wherein the abnormal information report includes abnormal processing suggestion information for the log to be processed.

[0170] In this embodiment, after performing anomaly detection to obtain detection result information, the detection result information can be iteratively processed to generate an anomaly information report, and anomaly processing suggestion information for the log to be processed can be given in the normal information report. In this way, relevant operators can be assisted to accurately and effectively identify log anomaly information, speed up the decision-making process, and thus quickly respond to and solve system problems, which can reduce the pressure on operation and maintenance personnel, improve system security, and has broad application prospects.

[0171] Optionally, when performing anomaly detection on the log to be processed and obtaining detection result information, the processor 810 is specifically configured to:

[0172] Constructing a text graph corresponding to the log to be processed, wherein the text graph includes at least two of a sequence graph, a syntax graph, and a semantic graph;

[0173] Performing feature extraction on the text graph to obtain node feature information corresponding to the text graph;

[0174] Generate a target text graph according to the node feature information;

[0175] Obtain the feature representation corresponding to the node in the target text graph through the Readout function;

[0176] The detection result information is obtained according to the feature representation corresponding to the node in the target text graph.

[0177] Optionally, when constructing the text graph corresponding to the log to be processed, the processor 810 is specifically configured to perform at least two of the following:

[0178] The sequence graph is constructed by extracting sequence features of the log to be processed;

[0179] The syntax graph is constructed by extracting syntax features of the log to be processed;

[0180] The semantic graph is constructed by extracting the semantic features of the log to be processed.

[0181] Optionally, when the processor 810 performs feature extraction on the text graph to obtain node feature information corresponding to the text graph, it is specifically configured to:

[0182] Based on a gated graph neural network, feature extraction is performed on the text graph to obtain node feature information corresponding to the text graph.

[0183] Optionally, when generating a target text graph according to the node feature information, the processor 810 is specifically configured to:

[0184] Determine the weight corresponding to each of the text graphs respectively;

[0185] A weighted calculation is performed according to the node feature information and the weight corresponding to each of the text graphs to obtain the target text graph.

[0186] Optionally, when the processor 810 obtains the detection result information according to the feature representation corresponding to the node in the target text graph, it is specifically configured to perform at least one of the following:

[0187] According to the feature representation corresponding to the node in the target text graph, anomaly prediction is performed to obtain a category label, and the category label is determined as the detection result information, where the category label is used to indicate whether the log to be processed is abnormal or normal;

[0188] According to the feature representation corresponding to the node in the target text graph, keyword extraction is performed to obtain abnormal keywords, and the abnormal keywords are added to the detection result information.

[0189] Optionally, when the processor 810 iteratively processes the detection result information to generate an abnormal information report, it is specifically configured to:

[0190] Initialize the target information as the log to be processed;

[0191] Based on the large language model, obtaining a target log report corresponding to the target information;

[0192] Performing a similarity evaluation on the target log report and a preset reference log report to obtain a target evaluation value corresponding to the target log report;

[0193] Update the target information according to the target evaluation value corresponding to the target log report;

[0194] Repeating the steps of obtaining a target log report corresponding to the target information based on the large language model, and updating the target information according to a target evaluation value corresponding to the target log report, until a maximum number of iterations is reached;

[0195] The target log report corresponding to the target evaluation value with the largest value is determined as the abnormal information report.

[0196] Optionally, when the detection result information includes at least one abnormal keyword, and the target information is updated according to the target evaluation value corresponding to the target log report, the processor 810 is specifically configured to perform at least one of the following:

[0197] When the target evaluation value corresponding to the target log report is greater than or equal to a threshold, updating the target information according to the target log report and the log to be processed;

[0198] When the target evaluation value corresponding to the target log report is less than the threshold, the target information is updated according to the target log report, the log to be processed and the at least one abnormal keyword.

[0199] Optionally, before acquiring the detection result information according to the feature representation corresponding to the node in the target text graph, the processor 810 is further configured to:

[0200] Preprocessing the log sample data in a preset manner to obtain target log sample data, wherein the preset manner includes at least one of the following: a masked language modeling manner, a document rotation manner, and a center word prediction manner;

[0201] Based on the target log sample data, an initial model is trained to obtain a target model, and the target model is used to obtain the detection result information.

[0202] It should be noted here that the above-mentioned device provided in the embodiment of the present invention can implement all the method steps implemented in the above-mentioned log anomaly detection method embodiment, and can achieve the same technical effect. The parts and beneficial effects that are the same as those in the method embodiment will not be described in detail here.

[0203] A device according to another embodiment of the present invention, such as Fig. 9 As shown, it includes a transceiver 910, a processor 900, a memory 920, and a program or instruction stored in the memory 920 and executable on the processor 900; when the processor 900 executes the program or instruction, the above-mentioned log anomaly detection method is implemented.

[0204] The transceiver 910 is used to receive and send data under the control of the processor 900.

[0205] Among them, Fig. 9 In the embodiment, the bus architecture may include any number of interconnected buses and bridges, specifically one or more processors represented by processor 900 and various circuits of memory represented by memory 920 are linked together. The bus architecture may also link together various other circuits such as peripherals, voltage regulators, and power management circuits, which are well known in the art and are therefore not further described herein. The bus interface provides an interface. The transceiver 910 may be a plurality of components, namely, a transmitter and a receiver, providing a unit for communicating with various other devices on a transmission medium. For different devices, the user interface 930 may also be an interface capable of externally and internally connecting required devices, and the connected devices include but are not limited to a keypad, a display, a speaker, a microphone, a joystick, and the like.

[0206] The processor 900 is responsible for managing the bus architecture and general processing, and the memory 920 can store data used by the processor 900 when performing operations.

[0207] The present application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the above Figure 1 The various processes of the method embodiment shown can achieve the same technical effect, and will not be described again here to avoid repetition.

[0208] A computer-readable storage medium according to an embodiment of the present invention stores a program or instruction thereon, and when the program or instruction is executed by a processor, the steps in the log anomaly detection method described above are implemented, and the same technical effect can be achieved. To avoid repetition, it is not repeated here. The computer-readable storage medium is, for example, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0209] It should be further explained that the terminals described in this specification include but are not limited to smart phones, tablet computers, etc., and many of the functional components described are called modules in order to more particularly emphasize the independence of their implementation methods.

[0210] In the embodiment of the present invention, module can be implemented with software so that it can be executed by various types of processors. For example, an executable code module of an identification can include one or more physical or logical blocks of computer instructions, for example, it can be constructed as an object, process or function. Nevertheless, the executable code of the identified module does not need to be physically located together, but can include different instructions stored in different positions, and when these instructions are logically combined together, it constitutes a module and realizes the specified purpose of the module.

[0211] In fact, executable code module can be a single instruction or many instructions, and can even be distributed on a plurality of different code segments, distributed among different programs, and distributed across a plurality of memory devices. Similarly, operating data can be identified in the module, and can be implemented and organized in the data structure of any appropriate type according to any appropriate form. The operating data can be collected as a single data set, or can be distributed in different locations (including on different storage devices), and can only be present on a system or network as an electronic signal at least in part.

[0212] When a module can be implemented by software, considering the level of existing hardware technology, a person skilled in the art can build a corresponding hardware circuit to implement the corresponding function of the module that can be implemented by software without considering the cost. The hardware circuit includes a conventional very large scale integration (VLSI) circuit or gate array and existing semiconductors such as logic chips, transistors, or other discrete components. The module can also be implemented by a programmable hardware device, such as a field programmable gate array, a programmable array logic, a programmable logic device, etc.

[0213] The above exemplary embodiments are described with reference to the accompanying drawings, and many different forms and embodiments are feasible without departing from the spirit and teachings of the present invention. Therefore, the present invention should not be constructed as a limitation of the exemplary embodiments proposed herein. More specifically, these exemplary embodiments are provided so that the present invention will be perfect and complete, and the scope of the present invention will be conveyed to those who are familiar with the technology. In these figures, the component sizes and relative sizes may be exaggerated for clarity. The terms used here are only based on the purpose of describing specific exemplary embodiments and are not intended to be limiting. As used herein, unless the text clearly indicates otherwise, the singular forms "one", "an" and "the" are intended to include these multiple forms. It will be further understood that the terms "including" and / or "comprising" when used in this specification indicate the presence of the features, integers, steps, operations, components and / or components, but do not exclude the presence or increase of one or more other features, integers, steps, operations, components, components and / or their groups. Unless otherwise indicated, when stated, a range of values ​​includes the upper and lower limits of that range and any subranges therebetween.

[0214] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A log anomaly detection method, characterized in that: include: Perform anomaly detection on the logs to be processed and obtain detection result information; In the case where the detection result information indicates that the log to be processed is abnormal, the detection result information is iteratively processed to generate an abnormal information report, wherein the abnormal information report includes abnormal processing suggestion information for the log to be processed.

2. The method according to claim 1, characterized in that The abnormality detection is performed on the log to be processed to obtain the detection result information, including: Constructing a text graph corresponding to the log to be processed, wherein the text graph includes at least two of a sequence graph, a syntax graph, and a semantic graph; Performing feature extraction on the text graph to obtain node feature information corresponding to the text graph; Generate a target text graph according to the node feature information; Obtain the feature representation corresponding to the node in the target text graph through the Readout function; The detection result information is obtained according to the feature representation corresponding to the node in the target text graph.

3. The method according to claim 2, characterized in that The step of constructing a text graph corresponding to the log to be processed includes at least two of the following: The sequence graph is constructed by extracting sequence features of the log to be processed; The syntax graph is constructed by extracting syntax features of the log to be processed; The semantic graph is constructed by extracting the semantic features of the log to be processed.

4. The method according to claim 2, characterized in that: The extracting features of the text graph to obtain node feature information corresponding to the text graph includes: Based on a gated graph neural network, feature extraction is performed on the text graph to obtain node feature information corresponding to the text graph.

5. The method according to claim 2, characterized in that: The step of generating a target text graph according to the node feature information includes: Determine the weight corresponding to each of the text graphs respectively; A weighted calculation is performed according to the node feature information and the weight corresponding to each of the text graphs to obtain the target text graph.

6. The method according to claim 2, characterized in that The acquiring the detection result information according to the feature representation corresponding to the node in the target text graph includes at least one of the following: According to the feature representation corresponding to the node in the target text graph, anomaly prediction is performed to obtain a category label, and the category label is determined as the detection result information, where the category label is used to indicate whether the log to be processed is abnormal or normal; According to the feature representation corresponding to the node in the target text graph, keyword extraction is performed to obtain abnormal keywords, and the abnormal keywords are added to the detection result information.

7. The method according to claim 1, characterized in that The iterative processing of the detection result information to generate an abnormal information report includes: Initialize the target information as the log to be processed; Based on the large language model, obtaining a target log report corresponding to the target information; Performing a similarity evaluation on the target log report and a preset reference log report to obtain a target evaluation value corresponding to the target log report; Update the target information according to the target evaluation value corresponding to the target log report; Repeating the steps of obtaining a target log report corresponding to the target information based on the large language model, and updating the target information according to a target evaluation value corresponding to the target log report, until a maximum number of iterations is reached; The target log report corresponding to the target evaluation value with the largest value is determined as the abnormal information report.

8. The method according to claim 7, characterized in that The detection result information includes at least one abnormal keyword; and updating the target information according to the target evaluation value corresponding to the target log report includes at least one of the following: When the target evaluation value corresponding to the target log report is greater than or equal to a threshold, updating the target information according to the target log report and the log to be processed; When the target evaluation value corresponding to the target log report is less than the threshold, the target information is updated according to the target log report, the log to be processed and the at least one abnormal keyword.

9. The method according to claim 2, characterized in that: Before acquiring the detection result information according to the feature representation corresponding to the node in the target text graph, the method further includes: Preprocessing the log sample data in a preset manner to obtain target log sample data, wherein the preset manner includes at least one of the following: a masked language modeling manner, a document rotation manner, and a center word prediction manner; Based on the target log sample data, an initial model is trained to obtain a target model, and the target model is used to obtain the detection result information.

10. A log anomaly detection device, characterized in that: include: Anomaly detection module, used to perform anomaly detection on the logs to be processed and obtain detection result information; The report generation module is used to iteratively process the detection result information and generate an abnormal information report when the detection result information indicates that the log to be processed is abnormal, and the abnormal information report includes abnormal processing suggestion information for the log to be processed.

11. A device, characterized in that: include: transceivers and processors; The processor is used to: Perform anomaly detection on the logs to be processed and obtain detection result information; In the case where the detection result information indicates that the log to be processed is abnormal, the detection result information is iteratively processed to generate an abnormal information report, wherein the abnormal information report includes abnormal processing suggestion information for the log to be processed.

12. A device comprising: A transceiver, a processor, a memory, and a program or instruction stored in the memory and executable on the processor; characterized in that when the processor executes the program or instruction, the log anomaly detection method according to any one of claims 1 to 9 is implemented.

13. A computer program product, characterized in that The method comprises computer instructions, which, when executed by a processor, implement the steps of the log anomaly detection method according to any one of claims 1 to 9.

14. A readable storage medium having a program or instruction stored thereon, characterized in that: When the program or instruction is executed by the processor, the steps in the log anomaly detection method according to any one of claims 1 to 9 are implemented.