Abstract

Human Activity Recognition (HAR) has gained significant attention with the rapid development of artificial intelligence (AI) and Internet of Things (IoT) technologies, particularly for its potential in smart monitoring and assistive systems. While existing HAR frameworks demonstrate promising performance, they often suffer from high computational complexity and suboptimal feature representation. In this study, we propose a novel framework, termed HAMEFE-SARNN, which integrates ensemble feature engineering with a self-attention-based recurrent neural network for efficient HAR. The proposed approach first applies min–max normalization to standardize input data, followed by a hybrid feature selection strategy combining multiple filter-based methods (Variance Threshold, Mutual Information, Chi-square, ANOVA, and L1-based selection) and a wrapper-based Recursive Feature Elimination (RFE) technique to identify the most informative features. Subsequently, a bidirectional long short-term memory network with a self-attention mechanism (BiLSTM-SA) is employed to effectively capture temporal dependencies and emphasize relevant patterns in activity sequences. The model is evaluated on a benchmark HAR dataset, demonstrating improved performance compared to existing approaches. Although the proposed framework is validated on standard datasets, it provides a robust and efficient foundation that can be integrated into real-time smart monitoring systems, with potential applications in assisting vulnerable and disabled individuals. Future work will focus on validating the model in real-world environments and multimodal sensing scenarios.