超算系统节点串口缓存启动日志的异常检测研究OA

As vital infrastructure for national scientific and technological innovation,supercomputing systems are widely applied in scientific research,engineering design,and other fields.As the scale of supercomputing systems continues to expand,abnormalities and failures in a vast number of components have become the norm.Currently,various machine learning and deep learning methods have been exten-sively employed in log anomaly detection tasks,achieving favorable results.The serial port buffer start-up logs of supercomputing systems record critical information during the system boot process and serve as an important basis for debugging and operation.However,existing log anomaly detection methods primarily target semi-structured log data and struggle to effectively process the unstructured serial port buffer startup logs in supercomputing systems.To address this issue,we propose a semi-supervised log anomaly detection method named COMLogNet.This method is composed of a BiLSTM module as its core,integrating a stacked and weighted structure of variational autoencoder(VAE)and isolation forest.The BiLSTM module incorporates both intra-block and inter-block attention mechanisms,along with semantic block preprocessing and a word vector-inverse document frequency feature extraction method,supplemented by a multi-line log anomaly detection mechanism,thereby achieving efficient log anomaly detection.Experimental results demonstrate that this method outperforms existing mainstream methods in anomaly detection performance while incurring lower time overhead,offering significant guidance and practical value for enhancing the operational maintenance efficiency and availability of supercomputing systems.