WMNN: Wearables-Based Multi-Column Neural Network for Human Activity Recognition.

Area of Science:

• Biomedical Engineering
• Computer Science

Background:

• Human activity recognition (HAR) in e-health is gaining interest.
• Current HAR methods often overlook crucial biomechanical data like muscle activity, limiting comprehensive motion analysis.
• Evaluating the correctness of motion tasks requires more than just spatial information.

Purpose of the Study:

• To develop an advanced HAR system that integrates multi-sensor data and deep learning.
• To address the limitations of existing HAR technologies by incorporating biomechanical insights.
• To improve the accuracy and comprehensiveness of motion task evaluation in e-health applications.

Main Methods:

• A Wearables-based Multi-column Neural Network (WMNN) was designed for HAR.
• The system utilizes multi-sensor fusion and deep learning techniques.
• The Tai Chi Eight Methods were used as a case study, analyzing both postures and muscle activity strengths.

Main Results:

• The WMNN achieved high accuracy rates of 96.9% for training and 92.5% for testing.
• The system successfully analyzed 144 distinct postures and their associated muscle activities.
• A human-machine interface (HMI) was developed to provide users with actionable motion feedback.

Conclusions:

• The proposed HAR technique significantly enhances users' self-training efficiency.
• This approach offers a more comprehensive evaluation of motion tasks by considering biomechanical data.
• The findings contribute to the advancement of HAR technologies in e-health and beyond.