Classifying Chaotic Time Series Using Gramian Angular Fields and Convolutional Neural Networks

Chaotic time series is derived from nonlinear dynamical systems which form the basis of real-life time series. This work focuses on classifying the chaotic time series by converting it into an image and using the image as input to a Convolutional Neural Network. The time series is converted to Gramian Angular Field images. Three chaotic time series—Hénon, Ikeda, and Tinkerbell—are used in this work. Four CNNs are trained—three binary classifiers for each time series and a general classifier. The best output from the binary classifiers is selected and compared with the results of the general classifier. The final output is decided by the similarity of the two answers. This method has shown accuracy over 99.5% which strongly supports the potential for classification based on time-series imaging techniques.