%0 Journal Article
%T 基于StackLog的日志异常检测
Log Anomaly Detection Based on StackLog
%A 徐克
%A 胡标
%A 刘军
%J Computer Science and Application
%P 382-393
%@ 2161-881X
%D 2025
%I Hans Publishing
%R 10.12677/csa.2025.154111
%X 日志信息是一种记录了系统运行状态的重要数据，是进行问题诊断、性能监控以及故障排除的主要依据。当系统出现故障时，通过详细系统地分析日志文本信息，研究人员可以快速准确地定位到系统出现问题的地方。通过分析现有的日志异常检测方法，本文发现当前用于日志异常检测的Text-CNN存在日志文本词向量维度急剧降低导致信息损失的问题。为了更充分地利用日志文本转换的词向量中携带的信息，本文提出了一种基于StackLog的日志异常检测方法。该方法采用堆叠卷积层逐步降低词向量维度的策略，尽可能保留词向量所携带的信息，并通过引入自注意力机制提升模型的检测能力。在HDFS和BGL两个公开数据集上进行对比实验验证了该模型在日志异常检测任务中的有效性。
Log information is an important data that records the operating status of a system, and is the main basis for problem diagnosis, performance monitoring, and troubleshooting. When the system malfunctions, researchers can quickly and accurately locate the problem by analyzing the log text information in detail and systematically. By analyzing existing log anomaly detection methods, this paper finds that the current Text-CNN used for log anomaly detection has the problem of information loss caused by a sharp decrease in the dimensionality of log text word vectors. In order to fully utilize the information carried in the word vectors of log text conversion, this paper proposes a log anomaly detection method based on StackLog. This method adopts a strategy of gradually reducing the dimensionality of word vectors by stacking convolutional layers, preserving the information carried by word vectors as much as possible, and improving the detection ability of the model by introducing self attention mechanism. Comparative experiments were conducted on two publicly available datasets, HDFS and BGL, to validate the effectiveness of the model in log anomaly detection tasks.
%K 深度学习，
%K 日志异常检测，
%K StackLog，
%K BERT
Deep Learning
%K Log Anomaly Detection
%K StackLog
%K BERT
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=113156