Among various uses of exoskeleton robots, the rehabilitation of stroke patients is a more recent application. There is, however, considerable environmental uncertainty in such systems including uncertain robot dynamics, unwanted user reflexes, and, most importantly, uncertainty in user intended trajectory. Hence, it is challenging to develop transparent, stable, and wide-scale exoskeleton robots for rehabilitation. This paper proposes an adaptive fuzzy impedance controller (AFIC) and a convolutional neural network (CNN) which uses electromyographic (EMG) signals for early detection of human intention and better integration with a lower limb exoskeleton robot. Specifically, the primary purpose of the AFIC is to manage the mechanical interaction between human, robot, and environment and to deal with uncertainties in internal control parameters. CNN uses EMG signals, inertial measurement units, foot force sensing resistors, joint angular sensors, and load cells to deal with signal uncertainties and noise through automatic feature processing in order to detect user’s desired joint angles with high accuracy. EMG is particularly effective here since it reflects the human intention to move faster than the other mechanical sensors. In the experimental procedure, signals were sampled at 500 Hz as two healthy individuals walked normally at 0.3, 0.4, 0.5, and 0.6 m/s for eight minutes while wearing a robot with zero inertia. Approximately 70% of the data is used for training and 30% for testing the network. The estimated angle from the trained network is then used as the desired angle in the AFIC loop, which controls the robot online as the desired trajectory. Pearson correlation coefficient and normalized root mean square error are computed to evaluate the accuracy and robustness of the proposed angle estimation with CNN and AFIC algorithms. Experimental results show that the proposed approach successfully obtains the torque of the robot joints despite uncertainties in changing the walking speed.

Keywords