A Hybrid Deep Learning Approach with MAOA Optimization for Enhanced Human Activity Recognition

Human Activity Recognition (HAR) plays a crucial role in improving human-computer interaction and personal health monitoring. Recent advancements in Deep Learning (DL) improve HAR performance by capturing intricate patterns in sensor data, but manual hyperparameter tuning remains a significant challenge. This process often leads to inefficiencies, including overfitting or underfitting, which negatively impact a model’s ability to generalize across diverse datasets. To address these issues, this paper introduces a DL-based approach that integrates Bidirectional Long Short-Term Memory (BiLSTM), Temporal Convolutional Network (TCN), and Convolutional Block Attention Module (CBAM) for enhanced feature extraction and sequence modeling. The Modified Arithmetic Optimization Algorithm (MAOA) is employed to automate hyperparameter tuning, improving model performance and generalization. The proposed model achieves 98.00% accuracy on the KU-HAR dataset and 99.69% on the mHEALTH dataset. Additionally, 5-fold cross-validation is applied to ensure robust evaluation, demonstrating the model’s efficiency and adaptability across varying data complexities. Our approach offers a scalable, automated solution to HAR that can be extended to other domains requiring accurate and efficient activity classification.