Manifold-Consistent Network for Robust Human Activity Recognition

A novel deep learning architecture known as the Manifold-Consistent Spatio-Temporal Network (MCSTN) has been introduced to enhance sensor-driven Human Activity Recognition (HAR) in medical and healthcare monitoring, especially within the Internet of Medical Things (IoMT). Traditional models often falter due to issues like missing data, sensor malfunctions, and environmental disturbances, which compromise the assumption of pristine data. MCSTN addresses this by employing dual-level corruption modeling, which mimics flaws via physical-level corruption and diffusion-driven continuous corruption. The model achieves stable, corruption-invariant semantic representations by ensuring representation consistency across various corrupted inputs. This research is documented in the arXiv repository under the identifier 2605.00913.