Wearable-based human activity recognition: from a healthcare application to a kinetic energy harvesting approach
Registro completoMostrar el registro completo del ítem
Manjarrés Córdoba, José Elías
Wearable technology is changing society by becoming an essential component of daily life. Human activity recognition (HAR) is one of the most prominent research areas where wearable devices play a key role. The first major contribution to the field in this dissertation is a smart physical work- load tracking system that combines wearable-based HAR and heart rate tracking. The proposed system employs a concept from ergonomics, the Frimat’s method, to compute the physical workload from heart rate measurements within a specified time window. This dissertation includes a case of study where tracking of an individual over the course of 20 days corroborates the ability of the system to assess adaptation to an exercise routine. The second and third contributions of this dissertation point to KEH in wearable environments. The second contribution is an energy logger for wrist-worn systems, with the purpose of tracking energy generation in KEH systems during daily activities. Thus, it is possible to determine if the harvested energy is enough to power a conventional wearable device. The proposed system computes the harvested energy using the characteristics of the objective load, which in this case is a battery charger. I carried out experiments with multiple subjects to examine the generation capabilities of a commercial harvester under the conditions of human motion. This study provides insights of the performance and limitations of kinetic harvesters as battery chargers. The third contribution is a KEH-based HAR system using deep learning, data augmentation and transfer learning to outperform existing classification approaches in the KEH domain. The proposed architecture comprises convolutional neural networks (CNN) and long short-term memory networks (LSTM), which has been demonstrated to outperform other architectures found in the literature. Since deep learning classifiers require large amounts of data, and KEH datasets are limited in size, this thesis also includes the proposal of three data augmentation methods to synthesize KEH signals simulating new users. Finally, transfer learning is employed to build a system that maintains performance independent of device location or the subject wearing the device.