Improved Human Activity Recognition Using Stacked Sparse Autoencoder (SSAE) Algorithm

This study aims to enhance the performance of Human Activity Recognition (HAR) systems by implementing the Stacked Sparse Autoencoder (SSAE) algorithm combined with Support Vector Machine (SVM). The objective is to enhance the classification accuracy of human activities using sensor data. The materials for this study include a dataset collected from wearable devices equipped with accelerometers and gyroscopes. These devices generate time-series data representing a range of activities, such as walking, running, sitting, and standing. The raw data were preprocessed through normalization and segmented into fixed time windows to ensure uniformity and reliability for analysis. The methods utilized involve employing SSAE for automated feature extraction. The SSAE algorithm extracts hierarchical and abstract features from sensor data, enabling the model to learn complex patterns that traditional methods might overlook. The extracted features are then input into the SVM classifier to perform activity classification. SSAE was trained using unsupervised learning techniques, followed by supervised fine-tuning with labeled datasets. The results demonstrate that the SSAE-SVM model achieves superior performance compared to traditional SVM. The SSAE-SVM achieved 89% accuracy, 87% precision, 89% sensitivity, and 88% F1 score, significantly outperforming the traditional SVM’s 37% accuracy, 75% precision, 37% sensitivity, and 36% F1 score. These findings underscore the potential of SSAE in enhancing HAR systems by effectively extracting features from sensor data. Future research should focus on the real-time implementation of SSAE, leveraging diverse sensor modalities, and exploring its applicability in broader fields, such as predictive maintenance and personalized health monitoring.