Kinematic analysis of the gait for the man-machine interface of a future lower-limb exoskeleton

Robotics is increasingly being used in the medicine to solve a wide range of problems in a multitude of different fields. Robots are becoming metallic allies for the benefit of the vulnerable, and apparently their place in the healthcare industry is increasing every day. Some of the fields in which these new technological tools have been incorporated are surgery, telemedicine, sterilization, invasive medicine, and rehabilitation. In the last field, exoskeletons are used as orthoses to assist or rehabilitate individuals with locomotor impairments. One type are lower limb exoskeletons, which are used to rehabilitate gait in patients suffering from stroke, accidents, or spinal injurie. These must help the wearer walk, supporting its posture maintaining balance and guaranteeing patient compliance. Unfortunately, no exoskeletons that combine these two purposes perfectly have been designed yet. Some impose physiological movements on the joints in a passive manner, without taking the patient's will into account (Lokomat by Hokoma); some focus only on maintaining balance while walking (EKSO GT by Ekso Bionics); others are synchronised with the wearer's desired movements through electromyography (EMG) signals (HAL by Cyberdyne). Our thesis focuses on mapping the kinematics on gait and on correlating EMG signals to lower limb motion, for the future development of a man-machine interface in a haptic exoskeleton for postural rehabilitation, which simultaneously offers compliance to the patient for voluntary control of multiple joints and balance. Two experimental datasets containing joint patterns and associated EMG were used, one obtained during gait on a treadmill and the other on the ground. The kinematics analysis is carried to offer a structured environment to the experimental data, and to reconstruct certain information not readily available directly from the data. First, the balance was evaluated, recreating from the gait data the trajectories of the zero-moment point (ZMP), of the centre of mass (COM) and of the meta of the feet, using the linearized inverse pendulum model; then an estimation algorithm was used to fit the previously calculated trajectories and the joint angles from the dataset to a kinematic model. In this way, it was possible to model the kinematics that aims at both a stable gait and a physiological joint motion. Artificial neural networks were used to correlate muscular activity (patient voluntary action) and joint’s motion of lower limbs. These networks have been trained to predict the angular positions of leg joints from EMG signals. Two recurrent types of neural network have been tested for this purpose, the non-linear autoregressive network with exogenous inputs and the long and short memory neural network. The EMG signals were properly treated to eliminate the unnecessary parts, by means of the extraction of the principal muscle activation, the computation of the envelopes and the muscle synergy, to apply them to the neural networks.