Anomaly Detection In Artificial Intelligent For IT Operations

IT system is infrastructure of social economy,and its’ operation and maintenance are crucial to the stable operation of socio-economic activities.Artificial Intelligent for IT operations integrates big data and machine learning methods into operation and maintenance of IT.It can discover anomalies and locate the cause of faults from massive operation and maintenance data quickly,and has become a research hotspot in the industrial and academic fields in recent years.Time series detection discover anomalies from massive operation and maintenance data and is an important component of Artificial Intelligent for IT operations.This thesis studies the anomaly detection of operation and maintenance metric time series and its research scenario and dataset are both from actual production environments.Specifically,(1)this thesis designs and implements an anomaly detection scheme for practical operation and maintenance data.By exploring a large amount of metric time series coming from a real production environment,this thesis summarizes those time series into typical waveform patterns;designs a waveform classification algorithm to classify metric series;designs corresponding anomaly detection algorithms for each type of time series.On this real dataset,the F1 score of the time series classification method is 0.94,and the F1 score of the overall anomaly detection scheme is 0.872.(2)this thesis proposes a Binary Mixed Detection(BMD)algorithm based on binary method for Pseudo-Constant metrics.This thesis finds a considerable number of Pseudo-Constant metrics in real production datasets.Based on this finding,this thesis proposes BMD algorithm for these metrics,BMD algorithm can shorten detection time under the condition of maintaining detection performance.When anomaly probability is0.01,the detection time of the BMD algorithm is 0.15 times that of the Individual Method.When anomaly probability is 0.02,the detection time of the BMD algorithm is 0.73 times that of the Individual Method.(3)this thesis build models to evaluate BMD algorithm’s performance.Firstly,this thesis builds a model for the number of vector addition and detection of the BMD algorithm and then obtains the running time of the BMD algorithm,and verifies the model through experiments.In addition,this thesis analyzes the false positive rate and false negative rate of the BMD algorithm,and obtains the applicable conditions of the BMD algorithm.