LAUSR

Abstract This paper aims to study the new methodology of nonlinear system identification problem of one–stage spur gearbox. Consider the influence of the complicated factors, the model of a one–stage spur gearbox obviously possesses the nonlinear dynamic characteristics. To identify the nonlinear system, a pseudo–linear neural network (PNN) based on the idea of gained–scheduling techniques and extended linearization model is presented. The PNN consists of multilayer architectures, in which the first hidden layers and output layers are the same as those used in a multilayer perceptron (MLP). The second hidden layers are composed of neurons with multiplication function, whose outputs are the products of the inputs. In order to improve the identification effectiveness, an extended Kalman filter (EKF) algorithm is employed to train the weights of the PNN, and model verification of a class of nonlinear system is fulfilled via computer simulation. The PNN identification approach of a one–stage spur gearbox in no–load and load condition are based on the experimental data, which are subdivided into a training set and a testing set. The training set is used for identification the PNN model of displacement and acceleration, while the testing set is applied to examine the classification accuracy of the proposed PNN model. To compare the identification precision, the diagonal recurrent neural network (DRNN) and MLP based on EKF methods are applied to model the same gearbox, respectively. The final results of identification and comparison of the one–stage spur gearbox demonstrate that the proposed PNN paradigm has the applicability and good performance.