<![CDATA[Human Activity Recognition (HAR) has emerged as a critical component of wearable sensing and intelligent healthcare systems, necessitating robust and computationally efficient deep learning architectures. This study presents an empirical experimental evaluation of three deep learning models 1D Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and a hybrid CNN–LSTM using the UCI HAR dataset under a subject-independent protocol. Raw inertial signals were segmented, normalized, and processed through standardized preprocessing pipelines to ensure reproducibility. Performance was assessed using accuracy, precision, recall, F1-score, cross-validation stability, and computational efficiency metrics. Results indicate that the CNN–LSTM architecture achieves the highest test accuracy (94.87%) and demonstrates improved robustness with lower variance and reduced sensitivity to signal perturbations compared to standalone models. Computational analysis confirms that the hybrid configuration maintains feasible inference latency for real-time applications despite moderate increases in parameter size. The findings validate the effectiveness of integrated spatiotemporal feature learning and provide a reproducible benchmark for future research on deep learning–based HAR systems in wearable and IoT contexts.]]>
Copyrights © 2026
