Accelerometer time series augmentation through externally driving a non-linear dynamical system

Coupled non-linear dynamical systems have gained attention not only as a ubiquitous occurrence in natural and artificial scenarios, but also as a basis for atypical computation paradigms. This paper introduces an approach to time series data augmentation involving driving a single low-dimensional entity, namely the Rossler system, with a physically-recorded sensor signal, and leveraging its responses to enhance the performance of a conventional classifier. A representative internet of things application in agriculture, namely cattle behavior recognition using a triaxial accelerometer, is investigated via a publicly-available dataset. Numerical simulations and experiments with an analog electronic circuit reveal that diversified responses to the external input are attainable, and the additional time series obtained from the driven system enhance the behavior classification accuracy. The advantage, down to the combined effects of its dynamical response and a static non linearity transforming the driving signal, is appreciable both when using a small multi-layer perceptron network operating on elementary features and, albeit to a lesser extent, when feeding the time series directly to a convolutional neural network. One possibility is that the driven system translates non-linear dynamical features into linear signal properties that can be more easily extracted. Some considerations about the engineering implementation using either analog hardware or programmable logic on an edge device are given.