| In the field of time series data,deep learning technology has been used to complete the task of classification,Compared with the traditional machine learning method,its classification effect is greatly improved.However,when the size of training data is small,the model is easy to over fit,resulting in serious performance degradation.Therefore,it is necessary to expand the size of data through data enhancement to reduce the possibility of over fitting.The existing data enhancement technology can expand the time series data,but it can’t guarantee that the expanded data and the original sample keep the same distribution,and the generation countermeasure network can generate new data with the same distribution based on the original data distribution.The existing generate adversarial network is not designed for time series data,so it needs to be improved according to local conditions in order to achieve good results.TS-GAN model improves the network architecture and loss function of the original GAN model according to the characteristics of time series data and the requirements of classification tasks.Through the analysis of JS loss function leading to the instability of the model,TS-GAN chooses Wasserstein distance to measure the distance between the generated distribution and the real distribution,which needs to meet the Lipschitz continuity condition.In order to limit the gradient parameter explosion,a penalty mechanism is added to the loss function.Time series is one-dimensional data.In view of the relatively stable generation characteristics of convolution layer,G network abandons the full link layer and uses one-dimensional deconvolution network to generate data,so as to enhance the generation ability of G network characteristics.The attention mechanism is introduced into the D-Network to improve its ability to judge the similarity of time series,and then to urge the G-Network to generate a better distribution.In addition,in order to prevent non convergence during data enhancement,the D-Network in TS-GAN model needs to use instance standardization,and the update times of G-Network and D-Network also need to be adjusted.In the experiment,GAN and TS-GAN are used to expand the training set from UCR time series data set respectively,and FCN classification model is trained based on the enhanced data set,then the test sets of each data set are used to verify the classification effect.Experiments show that the classification accuracy and F1 score of the data set expanded by TS-GAN are increased by 4.19% and 5.01% respectively,while the performance of the data set expanded by GAN network can not be improved,but the corresponding indexes are decreased by 1.9% and 1.94% respectively.This shows that TS-GAN can overcome the limitation of the extended time series of GAN and improve the classification effect of UCR data set significantly,which shows that TS-GAN model has better data enhancement ability. |
