System identification of biped robot
based on dynamic fuzzy neural network and improved
RBF neural network
*
Xiaoguang Wu and Tianci Zhang, Lei Wei, Ping Xie, Yihao Du
Institute of Electrical Engineering
Yanshan University
Qinhuangdao, Hebei 066004, China
wuxiaoguang@ysu.edu.cn Tianci_zhang@sina.com
*
This work was supported by the National Natural Science Foundation of China (NSFC) under Grants 61503325/F2014203246, Postdoctoral Science Foundation
of China under Grants 2015M581316 and the self-managed project of YANSHAN University under Grant 13LGB009.
Abstract - Aiming at the identification problem in continuous
and discrete hybrid dynamical system of biped robots, we present
a joint identification method based on dynamic fuzzy neural
network (DFNN) and improved radial basis function neural
network (RBFNN) with expectation-maximization clustering
algorithm. First we apply the improved RBFNN to identify the
swing stage in biped robot walking, and then use the DFNN to
identify collision stage in biped motion, further more, achieve the
whole identification of the biped robot’s motion trajectory. The
simulation results show that the method has fast learning speed,
high identification precision, can effectively identify the biped
robot system, thus it’s applicable to hybrid dynamic systems with
continuous and discrete properties.
Index Terms - system identification; RBF neural network; dynamic
fuzzy neural network
I. INTRODUCTION
In the past two or three decades, many scholars have a
profound research on various types of robots. Among many
types of robots, the biped robot is the closest to the human
[1]
.
Unlike the wheeled, tracked robot, the biped robot having a
highly nonlinear, hybrid dynamic characteristic with
continuous and discrete makes it difficult to establish its
accurate mathematical model. Therefore, using the method of
system identification, as an effective means to replace the
robot real system, has become a hot spot in the research of the
robot.
At present, the existing robot identification methods
include continuous system identification and target trajectory
tracking, iterative system identification
[2]
. So the existing
methods are not suitable for the hybrid system of the biped
robot. Fortunately, the swing phase and the collision stage are
coupled with each other, which provides the possibility for the
overall identification of the hybrid system of the biped robot.
In many methods of system identification, neural network is
widely used in the identification of nonlinear systems
[3-5]
. The
neural network having good nonlinear mapping ability, self
learning ability and parallel information processing
capabilities, provides a new perspective to solve the uncertain
nonlinear system identification and control problems. In
various kinds of neural networks, radial basis function (RBF)
neural network comparing with other networks have the
advantages of simple topology and high efficiency, which are
widely used in nonlinear complex systems. Theory and
practice prove that the RBF network can approximate the
single valued function with any precision
[6]
. The biggest
feature of the dynamic fuzzy neural network(DFNN) is that
the adjustment of parameters and the identification of the
structure are carried out simultaneously, and the learning
speed is fast, the network generalization ability is strong. So
they provide a powerful tool for the identification of nonlinear
dynamic systems.
In the past research, scholars have proved the feasibility
of hybrid dynamic system identification. We combine the two
methods, namely we use the method of identify iterative
system to establish identification model of the collision
process, apply the method of identify continuous system to
achieve trajectory tracking and finally accomplish the whole
trajectory identification.
II. MATIERIALS AND METHODS
A. Radial basis function neural network model
Radial basis function(RBF) neural network is composed
of three layers, namely input layer, hidden layer and output
layer
[7]
.The input layer node only transmits the input signal to
the hidden layer, and the hidden layer nodes are formed by the
RBF such as the Gauss function, and the output node is
usually a simple linear function. In the RBF network, the input
variables of the low dimensional model are mapped to the
hidden layer space in high dimension, and the hidden layer
unit selects the basis function to be transformed. Assuming the
input is
nmnn
xxxX ,...,,
21n
, the actual output
is
, then the input and output relations of
the three layer network can be expressed as:
(1) ),,(
1
I
i
inijjkj
XXy
978-1-5090-4102-2/16/$31.00 ©2016 IEEE
Proceedings of the IEEE
International Conference on Information and Automation
Ningbo, China, August 2016