Downsampling wearable sensor data packets by measuring their information value

Abstract

Long-Short Term Memory models (LSTMs) are data-driven routines that classify Human Activity Recognition (HAR) with minimum human input. The price to pay for analysing large sequences of on-body sensor measurements with LSTMs are high processing power and battery requirements. In this paper, we recognize that sensor data packets have differing information value to classify HAR and propose to quantify it with cross entropy (CrossEn), Kullback Leibler (KL) divergence and sample entropy (SampEn). Both, CrossEn and SampEn have the potential to guide dropping redundant data packets without compromising HAR. However, we do not find substantial improvements in dropping rates when downsampling by CrossEn and SampEn over computationally cheaper random and uniform alternatives. Our results show that the KL divergence, evaluated at training time is equivalent to the classification accuracy criteria that involves a testing set. The computational requirements to compute the KL in real-time could well guide sensor node design to downsample wearable measurements near the user.