Estimation of Upper Limb Dynamic Interaction Force Based on Multimodal Information.

Area of Science:

• Biomechanics and Robotics
• Machine Learning in Healthcare
• Human Movement Analysis

Background:

• Accurate estimation of interaction forces is crucial for understanding upper limb dynamics.
• Existing methods may lack precision in capturing complex, real-time movements.
• Dynamic upper limb elevation presents unique challenges for force estimation.

Purpose of the Study:

• To investigate and develop a robust method for estimating interaction forces during dynamic upper limb elevation.
• To evaluate the efficacy of a deep learning approach combining Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) networks.
• To compare the performance of different data modalities and machine learning models for this task.

Main Methods:

• Data acquisition included electromyographic (EMG) signals from the forearm, joint angles, and synchronized interaction force data.
• A hybrid deep learning model, CNN-LSTM, was developed for predictive dynamic force estimation.
• Comparative analysis involved using EMG signals versus EMG-IMU signals and comparing CNN-LSTM with Support Vector Regression (SVR).

Main Results:

• The CNN-LSTM model demonstrated effective characterization and estimation of dynamic interactive forces.
• Multimodal data, specifically EMG-IMU signals, yielded improved estimation performance compared to EMG-only signals.
• The CNN-LSTM model outperformed the SVR model in the dynamic force estimation task.

Conclusions:

• The proposed CNN-LSTM deep learning methodology is highly effective for estimating dynamic interaction forces in upper limb elevation.
• The integration of multimodal data significantly enhances the accuracy of dynamic force estimation.
• This approach offers a promising tool for applications in biomechanics, rehabilitation, and human-robot interaction.