International Journal of Control, Automation and Systems 16(5) (2018) 2225-2234 ISSN:1598-6446 eISSN:2005-4092

http://dx.doi.org/10.1007/s12555-017-0596-y http://www.springer.com/12555

Data Filtering Based Multi-innovation Gradient Identification Methods

Abstract: With the development of industry information technology, many researchers pay attention to the estima-
tion problems of feedback nonlinear systems increasingly. In this paper, a filtering based multi-innovation stochastic
gradient algorithm is derived for Hammerstein equation-error autoregressive systems by using the hierarchical tech-
nique. The parameter estimates accuracy can be improved with the innovation length increasing. These algorithms
are easy to implement on-line. The simulation results verify the effectiveness of the proposed algorithm.

Keywords: Feedback nonlinear system, filtering, gradient search, multi-innovation identification, parameter esti-
mation.

1. INTRODUCTION tion for its ability to separate the useful signal from the
noisy measurements, and has been used in system iden-
System identification can be applied to many engineer- tification [24, 25]. Wang et al. proposed a data filter-
ing areas [1–5], and many identification methods have ing based least mean square algorithm and a data filter-
been reported for time-varying systems [6], multivariable ing based multi-innovation least mean square algorithm
systems [7], state space systems [8, 9] and nonlinear sys- by employing the data filtering technique [26]. Ma et al.
tems [10, 11]. The identification methods can be roughly presented a decomposition-based recursive least squares
classified into the iterative methods [12,13], the maximum method for multivariate autoregressive systems using the
likelihood methods [14], the recursive methods [15–17] multi-innovation identification theory [27].
and so on. Recently, there is an increasing amount of work This paper applies the hierarchical identification prin-
on the identification of nonlinear systems. ciple and the filtering technique to closed-loop (or feed-
Nonlinear systems include several basic model struc- back) nonlinear systems with autoregressive noise. Xu et
tures: the input nonlinear systems, the output nonlinear al. studied the multi-innovation gradient parameter esti-
systems. Input nonlinear systems have many basic model mation algorithms [28]. Based on the work in [28], this
structures. Shen et al. proposed a gradient based re- paper gives the detailed formula derivation and the simu-
cursive identification method for input nonlinear systems lations of the filtering based hierarchical multi-innovation
[18]. Yu, Zhang and Xie studied the parameter estima- stochastic gradient algorithm for feedback nonlinear sys-
tion problem for Hammerstein systems using a blind iden- tems. The difficulty of identification lies in that the system
tification based deterministic estimation algorithm [19]. contains colored noise, the product of nonlinear structure
Pan et al. presented a filtering based multi-innovation and linear structure, etc. To solve these problems, this
extended stochastic gradient algorithm for multivariable paper presents the filtering based hierarchical stochastic
control systems [20]. There are many studies on out- gradient algorithm.
put nonlinear systems [21, 22]. Hu et al. derived a least This paper is organized as follows. Section 2 derives
squares parameter estimation algorithm for feedback non- the identification problem for feedback nonlinear sys-
linear systems using the hierarchical identification princi- tems. Section 3 describes a filtering based hierarchical
ple [23]. stochastic gradient algorithm. Section 4 presents a filter-
The data filtering technique has received much atten- ing based hierarchical multi-innovation stochastic gradi-

Manuscript received September 30, 2017; revised January 31, 2018 and March 7, 2018; accepted April 1, 2018. Recommended by Associate
Editor Xiaojie Su under the direction of Editor Hamid Reza Karimi. This work was supported by the 111 Project (No. B12018), the
Jiangsu Province Industry University Prospective Joint Research Project (BY2015019-29), the Fundamental Research Funds for the Central
Universities (JUSRP51733B) and the National Natural Science Foundation of China (No. 61472195).

Bingbing Shen, Feng Ding, and Ling Xu are with the Key Laboratory of Advanced Process Control for Light Industry (Ministry of Ed-
ucation), School of Internet of Things Engineering, Jiangnan University, Wuxi 214122, P. R. China (e-mails: bbshen12@126.com, fd-
ing@jiangnan.edu.cn, lingxu0848@163.com). Feng Ding is with the College of Automation and Electronic Engineering, Qingdao University
of Science and Technology, Qingdao 266061, P. R. China. Feng Ding and Tasawar Hayat are with the NAAM Research Group, Department
of Mathematics, Faculty of Science, King Abdulaziz University, Jeddah, Saudi Arabia (e-mail: fmgpak@gmail.com).
* Corresponding author.

⃝ICROS,
c KIEE and Springer 2018
2226 Bingbing Shen, Feng Ding, Ling Xu, and Tasawar Hayat

ent algorithm. Three examples are given in Section 5 to problem more difficult. In order to solve this problem, we
illustrate the effectiveness of the proposed algorithms. Fi- decompose the model into two subsystems by using the
nally, concluding remarks are offered in Section 6. hierarchical identification principle.
Remark 1: From (4), we can see that the output y(t)
2. PROBLEM FORMULATION contains autoregressive noise, which results in biased es-
timates. In this work, we eliminate the colored noise by
Let us define some notation. “X := A” or “A =: X” using the filtering technique. The details are as follows.
means “A is defined as X”; I represents an identity ma-
trix of appropriate sizes; the superscript T represents the 3. THE FILTERING BASED HIERARCHICAL
matrix transpose; sgn[X] is a sign function that extracts the STOCHASTIC GRADIENT ALGORITHM
sign of a real number X; the norm of a matrix X is defined
as ∥X∥2 := tr[XX T ]. In general, C(z) is the polynomial, and can be expressed
Mathematically, a feedback nonlinear dynamic rational as C(z) := 1 + c1 z−1 + c2 z−2 + · · · + cnc z−nc . Define the fil-
model is defined as tered input as rf (t) := C(z)r(t), the variable as gf (y(t)) :=
na nb
C(z)g(y(t)) and the filtered output as yf (t) := C(z)y(t).
y(t) + ∑ ai y(t − i) = ∑ bi u(t − i) + w(t), (1) Multiplying both sides of (5) by C(z) yields
i=1 i=1 na nb
nc yf (t) + ∑ ai yf (t − i) = ∑ bi [rf (t) − gf (y(t))γ ] + v(t).
w(t) + ∑ ci w(t − i) = v(t), (2) i=1 i=1
i=1 (6)
u(t) = r(t) − ȳ(t), (3)
Define the information vectors φ f (t), φ r (t), φ s (t), φ n (t)
where y(t) ∈ R is the measured output, ȳ(t) ∈ R is the and the identification matrix Gf (t) as
output of the nonlinear block, r(t) ∈ R is reference input φ f (t) := [−yf (t − 1), −yf (t − 2), · · · , −yf (t − na )]T ,
of the system, u(t) ∈ R is the control input of the model,
φ r (t) := [rf (t − 1), rf (t − 2), · · · , rf (t − nb )]T ,
v(t) ∈ R is a stochastic white noise with zero mean and
variance σ 2 . φ s (t) := [−y(t − 1), −y(t − 2), · · · , −y(t − na ),
Assume that the degrees na , nb and nc are known and u(t − 1), u(t − 2), · · · , u(t − nb )]T ,
y(t) = 0, r(t) = 0 and v(t) = 0 for t ⩽ 0. Define the φ n (t) := [−w(t − 1), −w(t − 2), · · · , −w(t − nc )]T ,
parameter vectors a, b and γ as a := [a1 , a2 , · · · , ana ]T ,
Gf (t) := [−gf (y(t − 1)), −gf (y(t − 2)), · · · ,
b := [b1 , b2 , · · · , bnb ]T , and γ := [γ1 , γ2 , · · · , γnγ ]T . Vectors
a and b are the parameter vectors of the linear subsystems − gf (y(t − nb ))]T .
and γ is the parameter vector of the nonlinear part. The Then from (6), we can obtain the identification model:
nonlinear part g(·) of the system can be represented by a
na nb
linear combination of γnγ nonlinear basis function, yf (t) = − ∑ ai yf (t) + ∑ bi rf (t − i)
i=1 i=1
nγ
nb
ȳ(t) = g(y(t)) = ∑ γ j g(y(t − j)) = g(y(t))γ . (4) − ∑ bi gf (y(t − i))γ + v(t)
j=1 i=1

In practice, many industrial processes have nonlinear =φ Tf (t)a + φ Tr (t)b + bT Gf (t)γ + v(t). (7)
feedback. These processes can be expressed as feedback Apparently, the right-hand side of (7) contains the product
nonlinear systems. For example, a feedback nonlinear sys- of b and γ . In order to get the unique parameter estimates,
tem whose feedback consists of a columbic friction ele- we use the standard constraint on γ . In general, we make
ment was introduced in [29]. ∥γ ∥ = 1 and the first entry of γ is positive, i.e., γ 1 > 0.
From (1)–(4), we have According to the hierarchical identification principle,
na nb Equation (7) can be decomposed into two subsystems,
y(t) = − ∑ ai y(t − i) + ∑ bi u(t − i) + w(t) they contain the parameter vectors γ , θ := [a, b]T , c :=
i=1 i=1 [c1 , c2 , · · · , cnc ]T and ϑ := [a, b, c, γ ]T , respectively. Rewrite
na nb
(2) as
= − ∑ ai y(t − i) + ∑ bi r(t − i)
i=1 i=1 w(t) = φ Tn (t)c + v(t), (8)
nb nγ
− ∑ bi ∑ γ j g(y(t − j)) + w(t). (5) and the generalized information vectors ϕ f (γ ,t), ψ f (b,t)
i=1 j=1 and the fictitious output y1f (t) as
[ ]
The model in (1) contains the product of two parameter φ f (t)
ϕ f (γ ,t) := ∈ Rna +nb ,
vectors and colored noise, which makes the identification φ r (t) + Gf (t)γ