Research On Time Series Classification Algorithm Based On Convolutional Neural Network And Recurrent Neural Network

Time series classification(TSC)is an important and challenging problem in data mining.The essence of the task is to assign a predefined label to a specified time series sample.At present,domestic and foreign research scholars have proposed some TSC methods,but due to the characteristics of time series data,some traditional methods have some difficulties in dealing with this problem.For the classification of univariate time series,the main work of this thesis is:(1)The research background and significance of TSC are explained,and the related knowledge and concepts of time series data mining,convolutional neural network and recurrent neural network are introduced.(2)This thesis proposes a TSC model based on SE-Res Net.Among the neural network-based methods,the Res Net-based method has the deepest network structure,but the accuracy of the network still needs to be improved.In order to further improve the accuracy of this network,this thesis applies the idea of SENet network to Res Net network,and proposes a method based on SE-Res Net.This method conducts experiments on 44 UCR data sets,and compares it with Res Net method and other existing TSC methods.Finally,we use Class Activation Map to visualize the Coffee dataset,and provide interpretability for the classification process of the network.(3)This thesis proposes a TSC model based on CNN-LSTM.At present,some traditional TSC methods rely too much on the similarity measurement and artificial feature selection between sequences,and do not consider the order and different timescale features of the sequences themselves.This paper uses the multi-scale convolutional network module to extract different time-scale features of sequences,and uses LSTM network to extract time dependent features between sequences,and then uses the two extracted features as the basis for classification.Finally,experiments were performed on 15 UCR data sets.The results show that the model proposed in this thesis is more advanced than other traditional classification methods.