Anomaly Detection Method Of Complex System With Missing Data And Unbalanced Samples And Its Application

Complex system is composed of several different subsystems connected according to a certain structure.With the improvement of scientific and technological ability and cognitive level,the complexity of the system is increasing.On the one hand,the complex system has made great progress in control accuracy,operation speed and intelligence,but at the same time,it has brought the reliability problem of the complex system.Timely detection,accurate diagnosis and safe treatment of complex system abnormalities can reduce the operation risk of the system and avoid major economic losses and casualties caused by system paralysis or scrapping.At present,the rapid development of sensor information,computer and other technologies makes it possible to collect,store and process a large amount of data.These data provide more comprehensive information than before and support more accurate anomaly detection.However,it is not easy to find abnormal data.The main challenge lies in the following two points: lack of data.Data loss may change the distribution characteristics of data and bring non negligible interference to subsequent anomaly detection tasks.Sample imbalance.Sample imbalance makes the classification algorithm pay too much attention to most categories,which reduces the detection performance of abnormal samples.To solve the above problems,focusing on the anomaly detection methods of complex systems,this paper studies from two aspects: building a more reliable missing data completion method and building a more accurate pattern representation method,and proposes a data completion method based on long-term and short-term memory network and a pattern library representation method based on kernel density estimation and Gaussian mixture model.The main innovations are:1.In view of the lack of observation data in the nonlinear system and the problems of weak fitting ability and poor prediction effect in the traditional data modeling methods,this paper introduces the time series data modeling and completion method based on long short-term memory(LSTM)network,which has better data completion effect with the help of the nonlinear data modeling ability of LSTM network.2.Aiming at the switching of complex systems under various stationary working modes,a system anomaly detection method based on Gaussian mixture model is proposed.This method carries out cluster analysis and modeling based on the stationary characteristics of the system,then uses EM algorithm to solve the unknown parameters in the Gaussian mixture model,and uses the distinguishability,stability and goodness of fit to determine the optimal number of clusters,Enhance the credibility of pattern library construction.3.The optimal bandwidth theorem and bandwidth convergence theorem of multidimensional kernel density estimation are given,which provides a theoretical basis for the optimal kernel density estimation algorithm.Based on this,an anomaly detection algorithm based on optimal kernel density estimation and JS divergence distribution is proposed,which effectively improves the diagnosis performance of anomalies,especially unknown anomalies.4.The anomaly detection experiments of rolling bearing and satellite solar array drive assembly(SADA)system data are carried out to verify the anomaly detection performance of the above method in the case of incomplete data and unbalanced samples.