EMG-Informed Estimation of Human Walking Dynamics for Assistive Robots

Abstract

The demand for gait rehabilitation is increasing globally, and conventional rehabilitation practices cannot cope with this increase. Robotic-based rehabilitation and assistive robots are alternative solutions for gait rehabilitation, but challenges remain to bring this technology into the clinic. Assistive robots would ideally provide a personalized level of assistance based on an individual’s physical and neurological condition, biomechanics, and muscular fatigue. An assistive robot should also produce a smooth movement based on the user's motion intention. Thus, a prediction of motion intention and corresponding adjustment of the robot actuator forces are the fundamental requirements for a controller of an assistive robot. Electromyography (EMG) signals have been used widely for motion intention estimation. However, most EMG-based models are subject or task-specific, requiring complex calibration. Creating an accurate, EMG-based motion estimation model which is generalizable across individuals and experimental conditions is a major challenge and was the goal of this thesis. The chosen application was to predict the motion and moments of the ankle joint during a range of different walking conditions. As such, a set of experiments was designed to collect motionrelated data from 10 individuals during a wide range of activities. Initially, an artificial neural network was designed to predict ankle moment during constant speed walking based on a list of input time series, including the EMG signals of four muscles from each leg and ankle kinematics. The results helped find the list of most important input time series and the length of information required for ankle moment prediction at each step. Next, a machine learning approach was explored, including feature extraction and selection from the input time series. The selected list of features optimized the model training process and was generalizable across individuals to estimate the ankle moment during constant-speed walking. Exploring the influence of the training dataset on model predictions at various walking speeds was the focus of the next step. It was discovered that training the model on acceleration data from 0.5 m/s to 2.5 m/s enabled the model to predict ankle moment during walking at any speed in this range. Random forest, backpropagation neural networks, and linear regression were compared as potential predictive models, with the random forest having the best predictions across walking speeds. In addition to making the model compatible with a range of activities, the desire was to update the model parameters based on the error between the model output and target value regardless of the training dataset. An adaptive model was developed and implemented to predict ankle angle during walking at four different speeds and three inclines to achieve this. The base model was initially trained on data from level ground walking on one participant at 1 m/s. The simplicity of the model structure made it possible to update the parameters whenever there was an error between the predicted and actual ankle angle with less than a 30 ms time delay. The RMSE of the model for all of the test conditions was less than 5 degrees across the cohort of ten individuals (including nine unseen individuals). Continuous and accurate prediction of joint kinematics under different walking conditions and multiple individuals promises a stable and reliable control for wearable assistive robots, thus achieving the goal of the thesis.