Ensemble learning with temporal fusion for human activity recognition using model-based augmentation

Abstract

Human Activity Recognition (HAR) is crucial for biomedical signal processing as it provides contextual information that enhances the interpretation and accuracy of physiological data. By identifying and classifying human activities, HAR enables the differentiation between normal and abnormal physiological states. This research addresses key challenges in HAR, including the lack of data diversity, the presence of noise, and imbalanced datasets, which prevent the development of robust and accurate HAR models. To overcome these challenges, we propose a novel ensemble framework, MK-WaveNet, designed to improve the generalization and accuracy of HAR systems. Our framework integrates a model-based data augmentation backbone, termed the Inverse Wavelet Augmented Convolutional Network (IWACN), to enhance feature representation and mitigate data imbalance issues. Additionally, a novel Minkowski distance-based temporal fusion algorithm is introduced to effectively merge the outputs of ensemble models, enhancing the overall performance of the framework. Comprehensive experiments conducted on four benchmark datasets (UCI-HAR, PAMAP2, DAPHNET, and MobiAct) and different backbones demonstrate the efficacy of the proposed framework in addressing the challenges of HAR. MK-WaveNet achieved F1-scores of 98.32% on UCI-HAR, 92.21% on PAMAP2, 94.83% on DAPHNET, and 98.86% on MobiAct. The results demonstrate that our proposed framework, alongside its backbones, sets a new benchmark by outperforming existing state-of-the-art models across all considered datasets.