A hybrid CNN-SVM approach with dynamic time warping for time series classification

Recently, convolutional neural networks (CNNs) have been successfully applied in many fields, including time series classication. However, in earlier years an important role played methods based on a dynamic time warping (DTW) distance. Therefore some researchers tried to combine the advantages of CNNs and DTW and create hybrid methods. In this work, we demonstrate a new approach which consists of three main stages. In the initial phase, a conventional CNN is trained. Then, based on weights from the rst convolutional layer of the learned network, DTW distances are calculated. Finally, such extracted features are used for the training of a traditional classier like Support Vector Machine (SVM). The use of SVM instead of another neural network should reduce the whole calculation time and ensure the more explainable model. We present initial results of this method for selected datasets.