Recent Advancements in energy harvesting hardware have created an opportunity for realizing self-powered wearables for continuous and pervasive human context recognition. Unfortunately, the power requirements of the continuous sensing using different sensors, such as accelerometers, and the burdensome on-node classification and communication are relatively high compared to the amount of power that can be practically harvested, which limit the energy harvesting’s usefulness.
A generic architecture of self-powered wearable devices
This research introduces a novel approach that employs kinetic energy harvesting (KEH) and infers human context information directly from the KEH patterns without using any other sensors such as accelerometers which need continuous power to operate. The underlying idea lies in the fact that different ambient vibrations generate energy in a different way producing different energy generation patterns in the KEH circuit. Because no actual sensor such as accelerometer is needed, a significant percentage of the limited harvestable energy can be saved, making significant progress towards self-powered autonomous wearables.
Conventional accelerometer-based vibration recognition in self-powered device
Proposed architecture of KEH-Sense
In this project, we quantify the capabilities of KEH-Sense for various applications:
All of our results are based on real data collected under natural (non-laboratory) conditions. Using KEH Recorder I (presented below), we collected extensive real data to demonstrate KEH as a potential new source of information for the previously mentioned applications. The table below summaries the data collection details, the algorithms used, and the accuracy reported in each study, confirming KEH as an efficient source of information for a wide range of wearable applications.
|Application||Data Collection||Algorithm Used||Accuracy|
|Human Activity Recognition||10 subjects, 5 different activities, 2 holding positions||K-nearest neighbour (KNN)||81% for hand and 87% for waist|
|Step Counting||4 subjects, different walking scenarios||Peak detection algorithm||96%|
|Calorie Expenditure Estimation||10 subject, 2 different activites||Standard statistical regression||88% for walking and 84% for running|
|Hotword Detection||8 subjects||Decision tree classifier||73% for speaker independent and 85% for speaker dependent|
|Transport Mode Detection||3 hours of data traces for three motorized modes (car, train, bus)||Voltage peak based learning algorithm||85%|
|User Authentication||20 subjects||Multi-Step Sparse Representation Classification||93%|
|Railway Trip Tracking||4 distinct train routes in the Sydney metropolitan area (36 trips including 360 data point)||Ensemble classifier and a probabilistic-based trip inferring algorithm||97.2% for a journey of 7 stations|
|Audio Communication Receiver||Series of experiments for varying distances between the speaker and the VEH receiver (2cm-100cm) and varying data rates (5bps to 35bps)||ON-OFF keying (de)modulation scheme.||5-30 bps communication depending on the distance|
This is a second-order mass spring damping system which is used to model an inertial harvester. It produces the amount of power that could be generated by an inertial harvester for a kinetic/ acceleration input.
KEH Recorder I
KEH Recorded II
To be announced soon!