Initial Velocity Prediction Based on Radial Basis
Function (RBF) Neural Network
Nan Hu, Weiqun Yuan, Rong Xu, Ducheng Zheng, Ping Yan, Ying Zhao and Wenping Cheng
Institute of Electrical Engineering, Chinese Academy of Sciences
Address : No.6 North two Zhongguancun, Beijing, China
hunan@mail.iee.ac.cn
Abstract
In this paper, an initial velocity prediction strategy in pulsed current driven system applying RBF neural
network is proposed. Firstly, an RBF neural network model is built including an input layer, a hidden
layer, and an output layer. Then it is trained with the preprocessed data set which includes four
features armature mass, acceleration distance, current time integration and initial velocity of armature.
By setting the target variables as initial velocity we trained a model predicting this variable. When the
number of training samples reaches 400, the average prediction error on the test set reaches a
minimum of 18.3 m/s(1.12%) and the prediction costs time 0.3 seconds, which shows the high
accuracy and efficiency of this method. As it is based on a machine learning framework, which is easy
to apply and flexible to expand, it has an undoubtable great advantage over the traditional numerical
calculation with long calculation time and complicated structure.
Keywords: initial velocity prediction, Radial Basis Function, neural network, machine learning framework
1. Introduction
In pulsed current driven system, the pulse power supply network consists of a plurality of pulse power
supply modules in parallel. The discharge timing of each module can be adjusted to change the
current waveform which directly affects the initial velocity [1]. Therefore, precise speed control through
current is the advantage of electromagnetic launch over the traditional emission method. It is
absolutely meaningful to find a method to predict the initial velocity, as it is the basics of speed control.
The initial velocity can be calculated by simulation of the launch process. Develop a transient electric
circuit model for a pulsed current driven system and the transient performance of the system can be
simulated through Micro-Cap VI, an electrical circuit analysis software package [2]. Also, researchers
improved the calculation accuracy of the rails’ inductance and the force distribution upon the rails [3,4].
However, there are different degrees of simplification of working conditions and calculations in all
above methods. It is often one-way coupling rather than two-way coupling that is used in these
simulations. And the simulation time could be hours if more factors are considered.
In this paper, a new prediction strategy based on RBF neural network is presented. An RBF neural
network is built, and then it is trained with the preprocessed data obtained in a simulation system. The
accuracy of the simulation system has been verified experimentally [5]. The method is easy to apply
and expand as it is based on the machine learning framework. Once the model has completed training,
each initial velocity prediction can be completed in constant time, which has undoubted advantages
considering about the complexity of numerical simulation.
2. RBF Neural Network Building and Training
Fig. 1. The structure of the RBF neural network
The structure of the RBF neural network is shown in Fig.1. The first layer is the input layer. An input
vector x is used as input to all radial basis functions. The second layer is the hidden layer with a non-
