Lateral walking gait phase recognition for hip exoskeleton by denoising autoencoder-LSTM

Area of Science:

• Biomechanics
• Robotics
• Machine Learning

Background:

• Lateral resistance walking effectively strengthens hip abductor muscles.
• Accurate recognition of lateral walking gait is crucial for exoskeleton-assisted rehabilitation and exercise.
• Existing gait recognition methods may lack the accuracy and robustness required for real-time exoskeleton control.

Purpose of the Study:

• To propose and evaluate a novel Denoising Autoencoder-LSTM (DAE-LSTM) algorithm for lateral walking gait recognition.
• To compare the performance of DAE-LSTM against traditional machine learning models.
• To assess the algorithm's accuracy, recognition time, and robustness for exoskeleton applications.

Main Methods:

• Collected Inertial Measurement Unit (IMU) data from ten subjects across three speeds and strides.
• Implemented a Denoising Autoencoder-LSTM (DAE-LSTM) model for gait phase recognition.
• Compared DAE-LSTM with Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Neural Networks (NN) models.

Main Results:

• DAE-LSTM achieved an average cross-subject recognition accuracy of 90.2%, surpassing other models.
• The algorithm demonstrated high accuracy (>90%) even with significant noise (SNR > 100:1).
• DAE-LSTM's average recognition time per frame was 0.383 ms, meeting practical requirements.

Conclusions:

• The proposed DAE-LSTM algorithm provides accurate and robust lateral walking gait recognition.
• This algorithm meets the performance requirements for real-time application in exoskeleton systems.
• DAE-LSTM offers a promising solution for enhancing exoskeleton-assisted rehabilitation and training exercises.