Integrated Data Augmentation for Accelerometer Time Series in Behavior Recognition: Roles of Sampling, Balancing, and Fourier Surrogates

Abstract

The behavioral monitoring of farmed animals such as cattle is a fundamental element of precision farming in which it enables unobtrusive ongoing health monitoring. This application presents two ubiquitous challenges typical of sensing applications of the Internet of Things: limited dataset size and dataset imbalance. Recently, data augmentation has emerged as a way of addressing their negative influences on the training process without overburdening the data acquisition phase. However, there remains no consensus regarding which methods should be applied to time series and in what combination. Here, we present the first comprehensive analysis that synergistically combines multiple approaches. These approaches are benchmarked on a dataset of triaxial accelerometer time series, which were acquired from six freely roaming cows through a collar-mounted sensor and labeled by experienced human observers according to five behaviors. Our results indicate that integrating data augmentation with the training process can substantially improve the time-series classification performance while retaining a fixed convolutional neural network architecture. The improvement is maximized when the dataset is balanced by applying a suitable sampling scheme and the negative influence of data duplication is reduced via generating synthetic time series with Fourier surrogates. With the proposed approach, the overall accuracy is improved from 90% to 96%, and the classification accuracy of an under-represented behavior, namely, grazing, is elevated from 45% to 91%. This work provides a direction toward a general methodology, motivating research on other datasets and applications.