Research And Implementation Of KPI Anomaly Detection Based On Variational Self Coding

Currently,mainstream network monitoring and anomaly detection mostly rely on systems and human intervention,but with the continuous expansion of the Internet scale and the complexity of business models,this method gradually fails to meet the requirements.With the development of the field of deep learning,in order to ensure the normal operation of business,large-scale Internet communication enterprises have begun to conduct real-time and close monitoring of key performance time series KPI data to achieve intelligent operation and maintenance.The information carried by the data transmitted and stored in the key business performance indicators has hidden value.It is assumed that there are exceptions that have not been detected and cannot be processed in a timely manner.The continuous accumulation of exceptions can cause incalculable losses.Especially in the face of sudden and aggressive exceptions,there is an urgent need for efficient exception detection methods to achieve real-time early warning of abnormal traffic events,timely and accurately locate the location of abnormal events,and analyze them,Find out the causes of exceptions to prevent them from further expanding the harm to services provided by critical businesses.This work combines a variational self encoder with a short-term and short-term memory network that incorporates an attention mechanism,and selects an end to end training method to learn abnormal scores.This allows the representation learning operation and the abnormality detection operation to be completed at the same stage,avoiding intermittent optimization of abnormal scores,resulting in inefficient use of data and unsatisfactory abnormal score results.The KPI anomaly detection framework Ve Lm Net based on variational self coding proposed in this paper consists of three parts:an anomaly score network,a standard score generator,and a loss function.Make full use of the very small number of anomalies carried by the KPI data itself to process the data,significantly deviating the abnormal score of the abnormal data from the standard abnormal score,and taking into account the accuracy and timeliness of the abnormal detection results.Through the experimental verification framework,the accuracy and timeliness of the abnormal detection can be considered,confirming its feasibility.Based on the experimental verification results of the above framework,this paper designs and implements a key performance indicator anomaly detection system with Ve Lm Net as the core.The system has the following four logical function modules: data acquisition module,data preprocessing module,exception detection module,and exception display notification module.Through an introduction to the system architecture design,module introduction,functional implementation process,and other aspects,the underlying logic of the system design is detailed from the overall to the partial.