Loading...

Design and Implementation of a Predictive Nonlinear Robust Controller in order to Reduce Interaction Forces in a Lower Limb Exoskeleton Robot used for Power Augmentation

Aliyari Glojeh, Alireza | 2023

99 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 56034 (08)
  4. University: Sharif University of Technology
  5. Department: Mechanical Engineering
  6. Advisor(s): Vossoughi, Gholamreza
  7. Abstract:
  8. Many workers and soldiers suffer from musculoskeletal problems due to carrying heavy loads. Using exoskeleton robots which are designed for power augmentation can be effective in preventing these disorders. Due to the interaction of these robots with human, it is necessary to design an appropriate control system for these robots, therefore, the aim of this research is to design a predictive nonlinear control system for a three degrees of freedom lower-limb Exoskeleton robot, in order to improve the performance of the robot, follow trajectory of human joints and reduce the interaction forces between human and the robot during the squatting activity. Multi-stage model predictive controller with 3 control horizons and 3 scenarios, which is optimized by the genetic algorithm is used to achieve a robust control system for exoskeleton robots with fixed uncertainties in the range of -40 to +50 percent applied to the robot parameters. Different scenarios are considered for the uncertainties of the system and seperate predictive controllers are designed for each of these scenarios. Then, the probability of occurrence of each scenario is obtained by parallel Kalman filters. Since there is no information about the trajectory of human joints along the prediction horizon, a recurrent neural network with long short-term memory known as LSTM network is used to predict the interaction forces along this horizon and a data augmention method based on separating and regenerating new time series has been used to enrich the training dataset. The designed LSTM network has appropriate performance for the first steps of the prediction horizon. Then a feed forward deep neural network with 3 hidden layers and 100 neurons in each layer has been used to approximate the behavior of the multi-stage controller and reduce the computational costs. Separate neural networks are designed for each of the scenarios and a tube-based data augmentation method has been used to enrich the training dataset. The results of the simulations indicate that the maximum of the interaction forces in neural networks-based controller has increased by 2.7 to 5.1 times compared to the multi-stage controller, but the time required to achieve these signals has decreased by 30000 times
  9. Keywords:
  10. Lower Limb Exoskeleton ; Data Augmentation ; Feed Forward Neural Network ; Long Short Term Memory (LSTM) ; Deep Neural Networks ; Multi-Stage Model Predictive ; Predictive Controller ; Squatting

 Digital Object List

 Bookmark

No TOC