NoFED-Net: Nonlinear Fuzzy Ensemble of Deep Neural Networks for Human Activity Recognition

In the era of the Internet of Things (IoT), the need for human activity recognition (HAR) is growing, especially in smart-healthcare applications using on-body smart sensor devices. These devices amass data and employ various classification models to analyze and discern user activities. However, existing techniques that are susceptible to the data type, user inputs, and ensemble-based models lack the ability to correct a wrong classification made by a base classifier. Addressing the shortcomings, we propose a novel fuzzy ensemble of three deep neural networks, using three nonlinear functions to generate fuzzy scores. The proposed model works on sensor data and can adaptively penalize the activity classes when the classification is assumed to be incorrect. Besides, a novel rewarding technique is proposed that aids the ensemble to extract the correct class in adverse situations. The proposed model reports state-of-the-art accuracy when evaluated on four publicly available wearable sensor data sets. In addition, activities corresponding to real-time sensor data collected using a smartphone are predicted correctly by the proposed model, thereby establishing itself as a reliable smart-HAR model. We also discuss a possible future scope of implementing the model over the cloud for smart activity recognition.