Research On KPI Prediction And Anomaly Detection Model For Intelligent Operation And Maintenance Of Data Centers

With the increasing scale and complexity of data centers,intelligent operation and maintenance(O&M)has become one of the important technologies to ensure high performance and availability of data centers.Key Performance Indicators(KPI)data,which reflects the health status of the data center system,plays a crucial role in intelligent O&M,and accurate KPI timeseries prediction and KPI anomaly detection are essential for improving the reliability of data center systems.Therefore,this paper focuses on KPI time-series prediction and KPI anomaly detection in the context of intelligent O&M for data centers,and conducts research on KPI data.The main work of the paper is as follows:(1)Propose a KPI prediction model based on contrastive learning and frequency domain enhancement.The KPI prediction is more difficult because KPI data presents more shape transformations,oscillations,amplitude,and phase deviations.In order to more fully obtain the temporal characteristics of KPI data,this model obtains the overall representation of the time series based on contrastive learning methods and improves the use of frequency domain information through a frequency domain attention module.In the experiment,we selected API call data from a data center and two publicly available KPI datasets.The results showed that the model proposed in this paper is superior to traditional time series prediction methods in terms of prediction accuracy and has strong competitiveness in KPI time series prediction.(2)Propose a KPI anomaly detection model based on HBOS-COPOD.Because anomalies are usually a very small proportion in real datasets,the distribution of data is highly imbalanced,and the performance of unsupervised models in such scenarios can be unstable.In order to alleviate this problem,this model integrates two unsupervised anomaly detection models,HBOS and COPOD,both of which are suitable for large datasets.COPOD effectively addresses the problem of HBOS’s inability to consider inter-feature correlations,and uses the SUOD framework to improve the efficiency of multi-model anomaly detection.Experimental results show that the KPI anomaly detection model based on HBOS-COPOD proposed in this paper can better detect anomalies in KPI data,help operators discover anomalies in a timely manner,and avoid further impact.(3)Practice and deployment of API call anomaly detection module.Based on the research on KPI prediction and KPI anomaly detection algorithms mentioned earlier,and relying on an actual project platform,the anomaly detection model was deployed using the Flask framework.Through this module,operators can monitor API call anomalies in real-time,promptly identify and resolve issues to ensure system stability and reliability.