Multivariate CNN Model for Human Locomotion Activity Recognition with a Wearable Exoskeleton Robot.

Area of Science:

• Robotics
• Machine Learning
• Biomechanics

Background:

• Wearable lower limb robots assist with mobility.
• Accurate identification of user locomotion is crucial for effective robot control.
• Existing methods for activity recognition have limitations.

Purpose of the Study:

• To develop and evaluate a novel convolutional neural network (CNN) architecture for identifying user locomotion activities.
• To compare the proposed CNN model against existing CNN and hybrid models (CNN-LSTM, LSTM-CNN).
• To assess the model's performance using data from wearable lower limb robots and electromyograms (EMGs).

Main Methods:

• Developed novel single and multi-head convolutional neural network (CNN) architectures.
• Collected prospective data from 500 healthy adults performing five locomotion activities across three terrains.
• Utilized multivariate signals from electromyograms (EMGs) and wearable lower limb robot sensors (hip angles, velocities, roll, pitch, yaw).

Main Results:

• The proposed deeper CNN architecture significantly outperformed three competing models (DDLMI, DeepConvLSTM, LSTM-CNN).
• Achieved a superior F-measure of 96.17% compared to 90.68%, 94.41%, and 95.57% for the other models.
• Demonstrated efficient inference speed of 1.14 seconds.

Conclusions:

• The novel CNN architecture is highly effective for recognizing locomotion activities in users of wearable lower limb robots.
• The model's ability to leverage diverse sensor data (kinematic and EMG) contributes to its superior performance.
• This advancement holds potential for improving human-robot interaction and rehabilitation technologies.