Dual-Phase Neural Networks for Feature Extraction and Ensemble Learning for Recognizing Human Health Activities

Dual-Phase Neural Networks for Feature Extraction and Ensemble Learning for Recognizing Human Health Activities

The integration of smart devices into healthcare has led to the creation of vast amounts of sensor data, which are crucial for advancing various healthcare applications such as elderly care, lifestyle enhancement, and health monitoring. Human Activity Recognition (HAR), which relies on these data, is essential for the success of these applications. While Deep Learning (DL) methods, particularly Convolutional Neural Networks (CNN) and Machine Learning (ML), have been somewhat successful in HAR, they often face performance limitations. These limitations arise from the challenges of extracting complex features from sensor-based HAR data and dealing with noise. Current methods often rely on a single-phase feature extraction process. In contrast, adopting a multi-phase feature extraction approach, which rigorously performs feature extraction across multiple distinct phases, could more effectively address these challenges. To overcome these challenges, we introduce a novel hybrid framework named Dual-Phase Fused Neural Networks with Ensemble Learning (DP-FusedNN-EL), designed to achieve robust feature extraction and enhanced human activity recognition tasks. This model operates in two main stages: dual-phase feature extraction and classification. Initially, it employs two neural networks for feature extraction: a novel Dual-Head Fused CNN for local features and a CNN combined with a Stacked Bidirectional Gated Recurrent Unit and Attention network for local-global features. Subsequently, it utilizes a Dual-Phase Ensemble Learning model for classification, aiming to reduce overfitting by leveraging the strengths of local-global features. We evaluated our DP-FusedNN-EL model on several HAR datasets, achieving remarkable performance with accuracies ranging from 87.47% to 99.66%. These results significantly outperform existing models, demonstrating the effectiveness of the DP-FusedNN-EL model in HAR tasks.

The accessible file is the submitted version of the paper. Please refer to the URL for the published version.