Data Reconstruction Methods in Multi-Feature Fusion CNN Model for Enhanced Human Activity Recognition.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Human Activity Recognition (HAR) is crucial for digital healthcare, including exercise monitoring and elderly care.
• Traditional HAR methods using accelerometer data require extensive preprocessing, such as noise reduction and manual feature extraction.
• Deep learning models for HAR often struggle with noise and limited feature extraction from one-dimensional accelerometer data.

Purpose of the Study:

• To develop a robust and accurate HAR method using accelerometer data.
• To overcome limitations of existing deep learning approaches, including noise sensitivity and complex preprocessing.
• To enhance feature extraction by transforming time-series signals into 2D representations.

Main Methods:

• A multi-input, two-dimensional Convolutional Neural Network (CNN) architecture was proposed.
• Three distinct data reconstruction methods were employed to transform time-series signals into 2D representations.
• Features from these reconstructed images were fused to improve feature extraction capabilities, validated on a custom HAR dataset without complex preprocessing.

Main Results:

• The proposed multi-input CNN method significantly outperformed single-reconstruction methods and raw one-dimensional data models.
• Accuracy, precision, and recall improved by 16.64%, 13.53%, and 16.3%, respectively, compared to a 1D baseline.
• The model demonstrated superior robustness against noise and effectively captured latent patterns through feature fusion.

Conclusions:

• Multi-input CNNs with reconstructed data offer an effective strategy for improving human activity recognition.
• This approach provides a practical and efficient solution by streamlining feature extraction.
• The method enhances performance and is well-suited for real-world digital healthcare applications.