Research On KPI Anomaly Detection For Intelligent Operation And Maintenance Under Cloud Environment

In the cloud environment,the system architecture and application services are increasing,in order to ensure the reliability and stability of the systems and services in the cloud environment,the monitoring and anomaly detection of large-scale data becomes particularly important.In the anomaly detection of KPI(Key Performance Indicator)data,the traditional anomaly detection method faces the challenge of high dimension of monitoring data instances and the constant change of KPI data characteristics in the cloud environment.If the anomaly detection is carried out to the KPI data directly,will give the anomaly detection system a huge overhead.The complexity of the user's operating habits in the cloud environment makes the given KPI data characteristics change between several types.The traditional anomaly detection model adjustment method makes the operators consume huge human cost.In order to solve the problem of KPI data anomaly detection in large scale and changing data characteristics under the cloud environment,this paper build an intelligent KPI data anomaly detection model based on unsupervised learning and reinforcement learning.(1)Firstly,in order to solve the problem of large data analysis overhead brought by the large scale of monitoring data,by analyzing the correlation between different data instances,using the principal component analysis method to select a small part of the data to express the information contained in the whole data,thus reducing the number of repeated information,and using the fuzzy clustering the selected information.The data instance is classified,which makes the data analysis work can be carried out for the specific and smaller data,and the anomaly detection algorithm can adapt the anomaly detection scene and reduce the overhead of scene selection.(2)In order to cope with the change of data characteristics under the cloud environment,this paper first uses the difference technology to judge the relationship between the global fluctuation trend and the local fluctuation trend of the monitored data,and identify the data characteristics of the current monitoring data.Aiming at the anomaly detection model of the change of data characteristics,in this paper,based on the reinforcement learning technology,the artificial adjustment process of the anomaly detection model is transformed into an automatic Markov decision process,which is optimized by setting dynamic reward function of the different anomaly detection algorithms,and the iterative process of the parameter adjustment is reduced.The automatic adjustment of anomaly detection models in the face of changes in data characteristics is presented.