Time Series Classification Using Recurrence Plots and Deep Convolutional Neural Networks

Abstract

we addressed the problem of time series classification using deep learning methods, where the raw signal is not processed directly. We first transformed the one-dimensional signal into a two-dimensional representation using recurrence plots, which capture when the system revisits similar states and how frequently these returns occur. This transformation is not always intuitive, but it simplifies subsequent processing, since the resulting representation can be treated as an image and processed using convolutional neural networks. We used a standard convolutional architecture consisting of several convolutional layers, followed by pooling layers for dimensionality reduction and a fully connected layer that produces the final prediction. The model does not rely on manually designed features, as features are learned during training. However, the final performance strongly depends on how well the chosen signal representation preserves the relevant structure of the original time series.