REPOSITORY > RESULTS

Abstract

Widespread accessibility of wearable sensing devices opens many possibilities for tracking
the users who wear them. Possible applications range from measuring their exercise patterns
and checking on their health, to determining their location. In this work we will use
the term context-recognition for all these tasks.
A common problem when using such context-recognition systems is their impact on the
battery life of the sensing device. It is easy to imagine that an application that monitors
users’ habits using all the sensors in a smartphone (accelerometer, GPS, Wi-Fi etc.) will
quickly drain the phone’s battery, making it useless in practice.
While many methods for reducing the energy-consumption of a context-recognition
system already exist, most of them are specialized. They work either in a specific domain
or can only optimize the energy consumption of specific sensors. Adapting these methods
to another domain can be laborious and may require a lot of expert knowledge and
experimentation.
We developed three novel methods for generating good energy-efficient solutions that
are independent of the domain and can easily be used to optimize a wide range of contextrecognition
tasks. The first method, Setting-to-Context Assignment (SCA) changes the
sensing settings – which sensors to use, with what frequency, what duty cycles to use etc.
– depending on what context was last detected. To do so, we developed a mathematical
model that can predict the performance of any given setting-to-context assignment. The
SCA method then uses the NSGA-II algorithm to search the space of possible assignments,
finding the best ones. The second method, Duty-Cycle-Assignment (DCA) works in a
similar way to the SCA method, but is specialized in optimizing only duty-cycling, i.e.,
periodically turning the sensors on and off. Finally, the Cost-Sensitive Decision-Tree (CSDT)
method adapts sensing settings directly to the sensor data. It works by using a costsensitive
decision tree that was adapted for context-recognition tasks. All three methods
were then combined in three different ways in order to join their individual strengths. Two
of these combinations adapt to both the current context and to the sensor data.
Our methodology was tested on four different real-life datasets, and on a family of
artificial datasets. Doing so, we proved that it works under many different conditions,
with different possible settings, sensors and problem domains. We also showed that our
methods compare favourably against other state-of-the-art methods.
Instead of returning only one solution, all of our methods can find different trade-offs
between the classification quality and the energy consumption. Returning these tradeoffs
can help the system designer to pick the most suitable for their system. For each
of the tested datasets we found many energy-efficient solutions. For example, for the
Commodity12 dataset, we reduced the energy consumption from 123 mA to 29 mA in
exchange for less than 1 percentage point of accuracy. For another example, we were able
to use only 5% of the available data in the SHL dataset (by using a lower frequency, dutycycling
and a subset of sensors) and in exchange sacrifice only 5 percentage points of the
accuracy.