A Dynamic Channel Model using CNN-Based Human Activity Recognition for WBSNs

Abstract

Wireless Body Sensor Networks enable continuous health monitoring but struggle with mobility-induced signal variations that impair communication reliability. The Dynamic Channel Model framework introduces a novel approach by seamlessly integrating Convolutional Neural Network-based Human Activity Recognition with channel modeling, unlike prior static models like Random Waypoint or simplified mobility frameworks that lack real-time adaptability. DCM uses CNNs to classify five activities like standing, sitting, walking, running, and jumping, enabling precise prediction of sensor movements across 21 links (navel, chest, head, upper arm, elbow, wrist, and ankle), validated by simulations and the MHEALTH data set in NS-3. The framework uses a log-distance path loss model with activity specific parameters which achieves a significant improvement in channel estimation accuracy, outperforming RandomWayPoint by 15.3%, Static Channel Model by 22.4%, Gaussian Mixture Model by 9.2%, and Machine Learning Path Loss by 6.3%, dynamically adjusting transmission parameters to effectively mitigate the effects of mobility. The proposed scheme improves energy efficiency and reliability, enabling robust healthcare applications such as realtime fall detection for the elderly and performance monitoring.