The Deep Transfer Learning for Sensor-Based Human Activity Recognition Using Class Augmentation

Sensor modality diversity as well as data annotation shortage are common challenges for sensing-based human activity recognizing (HAR) in smartphone settings. In light of this finding, we developed a graph-inspired machine learning strategy for the sensor-based HAR challenges, and we used this strategy to the construction of a deep learning transfer learning machine that provides a preliminary answer to each of these two difficult issues. To accomplish sensorbased HAR tasks, we introduce the Hara-ResTLCNN, a multilayer residue structural integrated graph convolutional neural network (CNN). The results of the experiment on the PAMAP2 as well as mHealth sets of data show that our ResTLCNN is successful at collecting the properties of actions, with median precision of
$98.18{{\%}}$
as well as
$99.07{{\%}}$
, respectively, similar to existing sensor-based HAR models. Moreover, the ResTLCNN model exhibits superior transferability as well as few-shot learning efficiency in the deep learning transfer trials. Th