Extending the battery lifetime of wearable sensors with embedded machine learning

Abstract

Smart health home systems and assisted living architectures rely on severely energy-constrained sensing devices, such as wearable sensors, for the generation of data and their reliable wireless communication to a central location. However, the need for recharging the battery regularly constitutes a maintenance burden that hinders the long-term cost-effectiveness of these systems, especially for health-oriented applications that target people in need, such as the elderly or the chronically ill. These sensing systems generate raw data that is processed into knowledge by reasoning and machine learning algorithms. This paper investigates the benefits of embedded machine learning, i.e. executing this knowledge extraction on the wearable sensor, instead of communicating abundant raw data over the low power network. Focusing on a simple classification task and using an accelerometer-based wearable sensor, we demonstrate that embedded machine learning has the potential to reduce the radio and processor duty cycle by several orders of magnitude; and, thus, substantially extend the battery lifetime of resourceconstrained wearable sensors.

A Comprehensive Study of Activity Recognition Using Accelerometers

“…However, it should be noted that this was a biaxial rather than triaxial accelerometer, and that a fairly limited subset of features (no spectral features) were used, so it is difficult to draw a solid conclusion from this single study. More recent works also demonstrate that simple classification tasks can be effectively conducted at very low sampling frequency and resolution, increasing the battery lifetime of wearable sensors by more than an order of magnitude [49]. Khan et al [50] performed a comprehensive study on optimising the sampling frequency of accelerometers in the context of human activity recognition.…”

Model-Based Design of Energy-Efficient Human Activity Recognition Systems with Wearable Sensors

“…Other approaches, orthogonal to those aforementioned, study the on-board calculation of the feature extraction stage of activity recognition systems. In [14][15][16][17][18][19], it was shown that calculating features on the wireless sensor nodes could reduce the energy consumption of their wireless transceivers or flash memories due to the reduced amounts of data.…”

“…[28][29][30]. Since this stage often drastically reduces the data rate, as a relatively small number of features is calculated from a large number of samples of a window, it was subject to a lot of research to perform this stage as near to the sensor as possible, i.e., on board of a wireless sensor node [14][15][16]18,19], or even on-sensor [17,21]. However, each application-specific setup varies in sensor sampling frequency, sliding window size, sliding window overlap, and number of features and their computational effort in calculating them.…”

“…However, each application-specific setup varies in sensor sampling frequency, sliding window size, sliding window overlap, and number of features and their computational effort in calculating them. Thus, the amount of reduced communications and added computational effort changes for each setup and has to be traded off for each particular configuration [19,21].…”

Downsampling wearable sensor data packets by measuring their information value

“…Recent work investigated downsampling data packets as a method for reducing the transmission data rate without compromising machine learning classification accuracy. In [16], large amounts of redundant accelerometer data packets were found, which motivated the approach in [17] to uniformly downsampling accelerometer data in both sampling frequency and bit depth. A reduction of recorded measurements by three orders of magnitude improved battery life from weeks to years without compromising classification accuracy [17].…”