Modeling And Controlling Of Cortical Signals And Pneumatic Artificial Muscles Using The Echo State Networks

Neural Networks have been used widely for modeling and controlling of nonlinear systems for their good performance, while the recurrent neural networks(RNN) take obvious advantages.Echo State Networks is a kind of newly developed RNN which has novel learning method and good short term memories. The basic principle of it is constructing a dynamic reservoir using a great number of sparsely connected neurons for storing information. The training method of it is calculating linear recurrent weights between the internal state space and the outputs in order to minimize the square error. The content of this paper is to model and control nonlinear systems using ESN, specially applying it in the modeling of cortical signals and the pneumatic artificial muscles.First this paper researches the decoding method of getting movement information from the neural activities of a monkey using ESN. The data are movement trajectories of 8 different directions and the simultaneously recorded neural signals. First, we build the model for predicting the movement trajectory. For every direction, we pick up several trails for training and use the rest trails for testing and the test results reveal the feasibility of this method. Then using the same data we adopt ESN with leaky integrator neurons to predict the moving direction of the monkey's arm.We also apply this network in modeling and controlling of McKibben artificial muscle, which is also referred to as Pneumatic Artificial Muscle (PAM). First, we collect relevant data by doing experiments on PAM, then using these data we research on modeling of it using ESN offline. At last, we solve the displacement tracking problem of PAM using ESN based internal model control system on Matlab/simulink and the simulation shows satisfactory results. We then apply this method in the experimental installation of PAM and successfully realized displacement tracking.