Recently, great results have been achieved by processing data with deep learning techniques, and, specifically, by using convolutional neural networks (CNN) with images as input. This neural network’s great performance for reading, processing and extracting the most important features of two dimensional data have highly contributed to its popularity. However, even in scenarios where input data isn’t formatted as an image, many transformation methods have helped apply CNNs to other data types. Time series is one of these data structures that can be modeled to approach the problem from a computer vision perspective.

Spectrograms are one of the most popular representations for signals, in which time series carry information with time and frequency as magnitude dimensions. Though spectrograms are graphical representations of frequency spectrum over time, some nuances exist between these graphical representations and pictures taken with a camera, or paintings. For example, when observing a landscape image, near pixels normally belong to the same object. However, in spectrograms, local relationships are represented using a different domain. Overall, this slight concept complicates the local feature extraction feeding two-dimensional CNN layers with spectrograms, as they have non-local relationships, unlike pictures. We can only take advantage of 2D CNN considering visual representations with inherent spatial invariance as they efficiently provide the best input for a convolutional layer.

Data scientists commonly work with sets of data representing temporal series, e.g. the historic evolution of a trend, a magnitude measured by a sensor or daily tracking of any source rich for analysis. Moreover, this information is usually composed of multiple variables, as different aspects of the scenario are measured. **Recurrence plots** are an advanced technique for visually representing multivariate non-linear data. In essence, this refers to a graph representing a matrix, where elements correspond to those times at which the data recurs to a certain state or phase. Recurrent behavior, such as periodicities or irregular cyclicities, is a fundamental property of deterministic dynamical systems, like non-linear or chaotic systems. As higher dimensional datasets can’t be pictured easily, they can only be visualized by projection onto 2D or 3D sub-spaces. Recurrence plots enables the visualisation of the mm-dimensional phase space through a two dimensional representation of its recurrence. This recurrence of a certain state at time ii at a different time jj is marked within a 2D squared matrix and can be mathematically expressed as: