Loading...

Prediction of Knee Joint Angle in a Lower-Limb Exoskeleton Robot Using sEMG signal and Deep Neural Networks

Haghgoo Daryakenari, Farshad | 2024

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 57459 (08)
  4. University: Sharif University of Technology
  5. Department: Mechanical Engineering
  6. Advisor(s): Vossoughi, Gholamreza
  7. Abstract:
  8. One of the major limitations for individuals with spinal cord injuries is the loss or weakening of their motor abilities. These individuals participate in physical therapy sessions, where they repeatedly perform movements with the assistance of a physical therapist to gradually regain their lost motor function. However, traditional physical therapy methods impose limitations on both the patient and the therapist, highlighting the need for new alternative methods. One suitable alternative for aiding motor function is wearable rehabilitation robots. The aim of this research is to design an algorithm for implementation on the knee joint of a lower-limb exoskeleton robot to assist individuals with knee impairments. More specifically, the work conducted in this research lays the initial foundation required for developing an intelligent robot to assist individuals with impairments by predicting the knee joint angle of a healthy person. To develop a robot capable of predicting future knee angles to track the user, the first step is to gather user data, followed by creating a suitable mapping to convert this data into the knee joint angle. Initially, electromyography (EMG) signals were considered as a source of sufficient information. However, further analyses revealed that raw EMG data alone is not adequate for predicting the user’s knee angle. As a result, the EMG data was preprocessed and used alongside IMU signals or knee joint angles as inputs to the network to achieve better prediction accuracy. A Transformer autoencoder neural network was employed to create the mapping. For training the network, data from two datasets from the University of Georgia and the University of Twente were used. Finally, to generalize the results, new data was collected for a user in the Movafaghian laboratory, and the network's prediction results, with and without transfer learning, were evaluated on this user. The network's performance was evaluated both quantitatively and qualitatively over two prediction horizons: a short horizon of 25 to 30 milliseconds and a long horizon of 250 milliseconds. It was shown that the Transformer network could predict the knee angle of a new user, unseen by the network, with an error of 2.1% over the short horizon and 3.2% over the long horizon, using processed electromyography and IMU data as input. The results demonstrated an improvement in error by 26.7% and 30.4% for the short and long horizons, respectively, compared to the case where only electromyography data was used as input. It was also shown that by replacing IMU data with knee angle data, the prediction error was reduced to 0.7% for the short horizon, whereas the error increased to 3.75% for the long horizon. Finally, to evaluate the network's practical performance, data was collected in the Movafaghian laboratory with a new user by adding interaction force through the attachment of a concentrated weight to the user's leg. Without retraining the network, an error of 2.9% for the short horizon and 24.9% for the long horizon was observed. After retraining the network, the prediction error decreased to 1.3% for the short horizon and 12.17% for the long horizon
  9. Keywords:
  10. Deep Learning ; Transformer Network ; Electromyogram Signal ; Lower Limb Exoskeleton ; Exoskeleton ; Inertial Measurement Unite (IMU) ; Angle Prediction ; Rehabilitation Exoskeleton Robot

 Digital Object List

 Bookmark

No TOC