Machine learning for prediction of muscle activations for a rule-based controller

The inductive learning (IL) technique, radial basis function (RBF) type of artificial neural network (AMY), and the combination of IL and RBF were used to predict muscle activation patterns and sensory data based on the preceding sensory data. The input consisted of the hip and knee joint angles, horizontal and vertical ground reaction forces recorded in an able-bodied human. The output data consisted of the patterns of muscle activities. These patterns were obtained from simulation of walking with a functional electrical stimulation (FES) system. The simulation takes into account the individual biomechanical characteristics of the eventual user having spinal cord injury (SCI). The mappings were tested using numerous data from five minutes of walking previously not used for the training. We illustrate the technique by presenting the estimation of the activations of the equivalent flexor knee muscle and the knee joint sensor for four strides. The correlation is better and tracking errors are smaller when the combination of IL and RBF is used compared to the usage of IL or RBF. We show that the prediction of sensory state is achievable; thus, the delays imposed by the properties of the neuro-muscular system can be minimized.