Raw-Data Driven Functional Data Analysis with Multi-Adaptive Functional Neural Networks for Ergonomic Risk...

Area of Science:

• Occupational Health and Safety
• Biomedical Engineering
• Machine Learning

Background:

• Workplace injuries from manual lifting tasks necessitate accurate ergonomic risk classification.
• Classifying multi-modal time-series data (facial landmarks, ECG, EDA) presents challenges like temporal dynamics and class imbalance.

Purpose of the Study:

• To develop and evaluate a novel deep learning model, the Multi-Adaptive Functional Neural Network (Multi-AdaFNN), for classifying ergonomic risk levels in manual lifting.
• To address challenges in multi-modal time-series data classification, including feature fusion and class imbalance.

Main Methods:

• Proposed the Multi-Adaptive Functional Neural Network (Multi-AdaFNN) integrating functional data analysis and deep learning.
• Evaluated the model on five configurations: facial landmarks, bio-signals, full fusion, reduced facial landmarks, and reduced facial landmarks with bio-signals.
• Utilized 100 independent stratified splits and weighted cross-entropy loss for robust evaluation and class imbalance management.

Main Results:

• The full fusion of facial landmarks and bio-signals achieved the highest classification accuracy and robustness.
• Adaptive basis functions within the Multi-AdaFNN identified critical lifting task phases for risk prediction.
• The Multi-AdaFNN demonstrated efficacy and transparency in multi-modal ergonomic risk assessment.

Conclusions:

• The Multi-AdaFNN framework effectively classifies ergonomic risk using multi-modal data, outperforming single-modality approaches.
• The model's ability to identify critical temporal phases enhances understanding of injury risk factors.
• This approach holds significant potential for real-time monitoring and proactive injury prevention in industrial settings.