A new Neural Network architecture for Time Series Classification

Abstract

Time Series Classification (TSC) is an important and challenging problem for many subject-matter domains and applications. It consists in assigning a class to a specific time series, recorded from sensors or live observations over time. TSC finds application in different fields, such as finance, medicine, robotics and physics. It can be used mainly for: Failure prediction, Anomaly detection, Pattern recognition and Alert generation. Here we present a new Neural Networks architecture, called Convolutional Echo State Network (CESN), to detect patterns and classify the univariate and multivariate time series. This architecture results from the combination of the Convolutional Neural Networks (CNNs) and the Echo State Networks (ESNs). CESN results are declared to be appropriate for the TSC tasks, both univariate and multivariate TS, while demonstrating a higher accuracy and sensitivity compared to previous tests with other existing algorithms. We applied this technique to the inertial sensors of a falling detection device.