Comparison of the Accuracy of Ground Reaction Force Component Estimation between Supervised Machine Learning and Deep...

Area of Science:

• Biomechanics
• Sensor Technology
• Machine Learning

Background:

• Estimating Ground Reaction Force (GRF) components is crucial for biomechanical analysis.
• Pressure insole sensors offer a portable method for GRF estimation.
• Comparing various machine learning algorithms for GRF estimation is essential for practical applications.

Purpose of the Study:

• To estimate GRF components (Fx, Fy, Fz) using pressure insole sensors across six activities, including novel static and manual material handling scenarios.
• To compare the accuracy of six different methods—three Deep Learning (DL) and three Supervised Machine Learning (SML)—for GRF component estimation.
• To identify the most accurate method for GRF estimation in different activities.

Main Methods:

• Six methods were evaluated: Artificial Neural Network, Long Short-Term Memory, Convolutional Neural Network (DL), and Least Squares, Support Vector Regression, Random Forest (SML).
• Data were collected from nine subjects performing six distinct activities.
• Root Mean Square Error (RMSE) was used to quantify estimation accuracy against force plate data.

Main Results:

• The Random Forest (RF) method demonstrated the highest accuracy in estimating GRF components for static activities.
• RF achieved mean RMSE values significantly lower than reference measurements for static situations.
• Supervised Machine Learning methods, particularly RF, outperformed the tested Deep Learning methods in GRF estimation accuracy.

Conclusions:

• Pressure insole sensors coupled with Supervised Machine Learning, specifically Random Forest, provide accurate GRF component estimation.
• The study expands GRF estimation to new activities, offering valuable insights for biomechanics and ergonomics.
• RF presents a superior alternative to Deep Learning methods for GRF estimation in static and potentially other activities.