The Research On Classification Of Multivariable Time Series Data Streams Considering New Classes

With the wide application of different types of sensors in the network,the precise classification of multivariable time series data stream has become a research hotspot in the field of data mining and machine learning.Because the characteristics of data flow in the network change with the change of time and network environment,there will be new categories of data flow,which leads to the decline of classification accuracy.At present,there are a lot of classification methods for multivariable data streams,but they do not take into account the generation of new classes.However,in real life,a lot of application data will have new data categories,timely detection of new categories,can improve the classification accuracy,and has great significance for real life.In the current research of data stream classification,there are still three problems that are not effectively solved:(1)The new class detection problem in multivariate data streams needs to be studied;(2)Less consideration of the interaction between features within streams and between features of different streams in multivariable data streams are given to multivariate time series data streams;(3)In the new class detection method for univariate data streams,the existing method assumes that there is only one new class at a time,but the reality is that multiple new categories may appear at the same time.Therefore,for the above three problems,the main task of this paper is to study the problem of new class detection in data streams by introducing special feature extraction methods and model updating methods based on clustering,and apply this method to multivariable time series data streams.This thesis will study one of important problems in stream mining,that is,the research on classification of multivariate time series data streams considering new classes or CMCNC.In order to measure the interaction between multivariate time series data streams more intuitively and quickly,this thesis uses the method of finding motifs in bioinformatics to extract the motifs and timing relationships.After vectorization,the input is Classification and new class detection are performed in the random forest classification model.When the model is updated,all new class instances are clustered using the feature weights and the k-means method to achieve the purpose of correctly detecting the new class.The experimental results show that compared with the existing classification methods,the classification method proposed in this thesis not only takes into account the time series relationship between multivariable time series data streams,but also achieves the purpose of new class detection,and can improve the accuracy of new class detection.Eventually,the classification method has achieved remarkable results.