A Backpropagation Multidimensional Taylor Network Model for Measured 3-D Data Fitting

To solve the nonlinear fitting problem of 3-D data, an improved multidimensional Taylor network (MTN) structure based on the error backpropagation (BP) algorithm is proposed in this article. Compared to the traditional MTN, the so-called BP-MTN has three main characteristics: 1) an additional full connection layer is added after the same dimension output layer of the traditional MTN to describe the dynamic evolution of the nonlinear system state; 2) adding activation function to improve the ability of network fitting; and 3) the weights of the BP-MTN are trained using the BP algorithm. Compared to the traditional artificial neural networks, the improved MTN has a simpler structure without complex calculations, thus can ensure better real-time operation performance. For verification, the measured 3-D nonlinear torque characteristics of a 5-kW switched reluctance motor in an electric vehicle (EV) application are modeled with the proposed BP-MTN algorithm. Experimental results show that BP-MTN performs well in terms of model training time and fitting capability.