Smart Activity Sequence Generator in Wearable IoT

jatindeep singh, Punit Mishra

Abstract:

Abstract Sensors in activity based computing enable continuous monitoring of
numerous physiological signals when attached to the human body. This finds wide
application in areas of activity monitoring, bio-medical rehabilitation, and fitness
tracking. Primary challenges in embedded application development for smart
wearable include high energy efficiency and user compatibility. Existing algorithms
and applications are still unable to fully utilize the true power of the data being
collected. They provide lot of descriptive data analytics but lack in predictive
analysis. Energy efficiency of computing as predicted by Koomey’s is expected to
strike the second law of thermodynamics based on Launder’s Limit within few
decades. In this work an energy efficient computing technique for next generation
mobile applications is developed. Proposed Artificial Intelligence based
energy-efficient embedded algorithm that provide personalized training sequence
recommendation in order to achieve desired calorie goals. Suggested training
sequence of 6 activities fall under high, medium and low calorie burn with achieved
median for 234C:535C:688C respectively. The crux of this implementation is
Calorie Matrix Regeneration via state feedback technique using Markov Decision
Process (MDP) and Genetic Algorithm (GA). Number of generations required by
the GA to reach a suboptimal solution is optimized. While Machine learning
algorithms are written in C/C++ for effective embedded implementation, certain
computationally expensive modules like MDP and GA are coded in Python with
proposed IoT cloud based implementation thereby improving battery efficiency to
12–16 h. This implementation is first of its kind and a step ahead of available state
of the art fitness training algorithms/applications.