Research On Time Series Classification Based On Feature Extraction

One of the main goals of time series analysis is time series classification,which has attracted substantial interest in the area of data mining.A time series is a set of ordered observations on a quantitative characteristic of a phenomenon at equally spaced time points.According to the dimension of time series observation,it can be categorized into univariate time series and multivariate time series.Generally,time series refers to univariate time series without special explanation.Time series are often warping and unequal in length,which makes it difficult to deal with time series classification problems by using traditional machine learning clas-sification methods.At present,most existing time series classification methods fall into similarity-based methods and feature-based methods.The similarity-based meth-ods focus on the distortion invariance,displacement invariance and amplitude invari-ance between time series data.The feature-based methods include the discovery of the discriminative subsequence and the feature transformation.It is difficult to make breakthroughs in the research of similarity measurement,and the time consumption of searching discriminative subsequences is unacceptable.The feature extraction meth-ods have achieved good performance in many time series classification tasks,however,which always need hand-crafted features.To deal with the problems of high time complexity of searching discriminative subsequence,the limited classification accuracy of single discriminative subsequence and cumbersome hand-crafted features in the time series classification,the methods are proposed in the paper.The main contributions of the paper are listed as follows:Firstly,a time series classification method based on discriminative subsequence discovery is proposed.In this method,time series are segmented and k-Shape is used to cluster the segmented subsequences.Then,discriminative subsequences of two classes of time series are found in the clustering results,which are used as the basis for classifi-cation.Furthermore,time complexity of this method is given in the experiment,and the effectiveness and applicability of the proposed method are verified on UCR datasets.Secondly,a Multi-Grained Ensemble Method for time series classification(MEGoT)is proposed,which can make use of the variation of multi-grained data at the same time.In MEGoT,unstable base learners(Neural Networks)are assigned different weights to combine the ensemble.Different learners represent the learning features of different subsequences in time series,which can discover the discriminative regions,providing interpretability for classification.In the experiments,empirical evaluations are con-ducted and comparisons with various existing methods on 25 benchmark datasets.The final results show that dividing samples into smaller granularity is able to improve the diversity of ensemble.Furthermore,MGEoT can discover the discriminative regions in time series,which may be neglected in the global methods.Thirdly,a time series classification method based on multi-task learning is pro-posed.In this method,Convolutional Neural Network(CNN)and the discriminator in Generative Adversarial Network(GAN)share the first convolution layer for extract-ing richer features,which can improve the generalization performance.This method is evaluated on UCR data sets and compared with 12 existing methods.Furthermore,this model is compared with LSTM,GRU and CNN on the electrical equipment load data,which is from electrical equipment of industry in a certain area.The final results show that our model has a higher equipment identification accuracy than other deep learning models.