Research On Anomaly Localization Of Multi-Dimensional Monitoring Indicators In Artificial Intelligence IT Operations

In the era of big data explosion,in order to solve the problems of larger cluster size,more complex monitoring visualization and intelligentization in traditional IT operations,artificial intelligence IT operations emerged as the times require.In the field of artificial intelligence IT operations,the anomaly localization of multi-dimensional monitoring indicators is a very important issue.A good anomaly localization method can quickly and accurately detect the moment when the anomaly occurs,and analyze the specific dimension that causes the anomaly at that moment,and reduce the loss caused by the anomaly to a certain extent.Therefore,this thesis focuses on the anomaly localization of multi-dimensional monitoring indicators in artificial intelligence IT operations.According to the process of anomaly localization,the anomaly detection and root cause analysis are deeply studied.Firstly,the related theories and techniques involved in the design and optimization of the algorithm are introduced,including three aspects: 1.Introduction to recurrent neural network and long short-term memory;2.Research on variational auto-encoder in depth generation models and its application in anomaly detection,and analyze the advantages of using reconstruction probability to perform anomaly detection.3.The upper-boundary algorithm in game decision-making is introduced and the parameters in the algorithm are simulated.,the Monte Carlo tree search algorithm used in huge search space is studied.Secondly,in the aspect of anomaly detection,this paper designs a time series anomaly detection algorithm based on variational recurrent neural network model and semi supervised learning.The algorithm combines the variational auto-encoder with the long short-term memory,and effectively solves the limitations of anomaly detection algorithm based on variational auto-encoder in time series anomaly detection and the difficulty of obtaining tags in actual operation scene.Thirdly,this paper makes a specific symbol definition for the root cause analysis of multidimensional monitoring indicators.Aiming at the shortcomings of existing algorithms,this thesis proposes a new measurement index,which considers both the root cause contribution and the abnormal similarity,and can well measure the influence of the element sets in each dimension on the total monitoring indicators.And using this new metric,the Monte Carlo tree search algorithm is optimized.In addition,referring to a frequent item set algorithm in data mining,this thesis proposes a layered pruning strategy to further reduce the search space.Finally,this thesis uses the data set in the actual operation scenario to analyze the anomaly detection algorithm and root cause analysis algorithm proposed in this thesis.In terms of anomaly detection,the accuracy and recall rate of the algorithm on the three data sets exceeds 0.8,which proves the feasibility of the anomaly detection algorithm proposed in this thesis in the actual operation scenario.In terms of root cause analysis,compared to the current root cause analysis algorithm,the algorithm achieves higher F-score in 20 different types of abnormal root causes,and has made great progress in terms of effect and robustness.And shorten the time of root analysis from less than 1 hour of manual positioning to less than 60 seconds.