Identification of a biomechanical system using neural networks

This paper describes the modelling of lower-limb dynamics of paraplegics using different neural network structures. A state-space model in form of a Recurrent Neural Network (RNN) and an input-output model involving a Multi-Layer Perceptron (MLP) have been applied to the identification of knee-joint dynamics under electrical stimulation of the quadriceps muscle group. A comparison of these black-box modelling techniques shows that both approaches are suitable for this application in order to achieve an approximation of the nonlinear system. The identification by means of the RNN is described in detail as it represents a new approach for the modelling of this class of system. Advantages of RNNs in comparison to MLPs such as simple structure selection, are highlighted. Additionally, this paper presents a very efficient and novel second-order training technique for RNNs based on the Levenberg-Marquardt optimisation method