The advances in MEMS technology development allow for small and thus unobtrusive designs of wearable sensor platforms for human activity recognition. Multiple such sensors attached to the human body for gathering, processing, and transmitting sensor data connected to platforms for classification form a heterogeneous distributed cyber-physical system (CPS). Several processing steps are necessary to perform human activity recognition, which have to be mapped to the distributed computing platform. However, the software mapping is decisive for the CPS’s processing load and communication effort. Thus, the mapping influences the energy consumption of the CPS, and its energy-efficient design is crucial to prolong battery lifetimes and allow long-term usage of the system. As a consequence, there is a demand for system-level energy estimation methods in order to substantiate design decisions even in early design stages. In this article, we propose to combine well-known dataflow-based modeling and analysis techniques with energy models of wearable sensor devices, in order to estimate energy consumption of wireless sensor nodes for online activity recognition at design time. Our experiments show that a reasonable system-level average accuracy above 97% can be achieved by our proposed approach.
Wearable sensor data based human activity recognition using deep learning: A new approach
