Available online at www.sciencedirect.

com

Journal of the Franklin Institute 356 (2019) 10102–10122

www.elsevier.com/locate/jfranklin

Hierarchical recursive generalized extended least

squares estimation algorithms for a class of nonlinear
stochastic systems with colored noise
Longjin Wang a,∗, Yan Ji b, Lijuan Wan b, Ni Bu b
a College of Electrical and Mechanical Engineering, Qingdao University of Science and Technology, Qingdao
266061, PR China
b College of Automation and Electronic Engineering, Qingdao University of Science and Technology, Qingdao

266061, PR China
Received 20 January 2019; received in revised form 29 May 2019; accepted 22 June 2019
Available online 10 July 2019

Abstract
This paper considers the parameter identification problem of a bilinear state space system with colored
noise based on its input-output representation. An input-output representation of a bilinear state-space
system is derived for the parameter identification by eliminating the state variables in the model, and
a recursive generalized extended least squares algorithm is presented for estimating the parameters of
the obtained model. Furthermore, a three-stage recursive generalized extended least squares algorithm is
proposed for reducing the computational cost. The validity of the proposed method is evaluated through
a numerical example.
© 2019 Published by Elsevier Ltd on behalf of The Franklin Institute.

1. Introduction

Mathematical models are the basis of controller design [1–3] and system analysis [4–6].
Parameter estimation methods have been proposed for modelling linear systems [7–10] and
nonlinear systems [11] and multivariable systems [12,13], and they can be applied to many

∗ Corresponding author.
E-mail address: [email protected] (L. Wang).

https://doi.org/10.1016/j.jfranklin.2019.06.032
0016-0032/© 2019 Published by Elsevier Ltd on behalf of The Franklin Institute.
 L. Wang, Y. Ji and L. Wan et al. / Journal of the Franklin Institute 356 (2019) 10102–10122 10103

fields [14–16]. As most of natural systems offer nonlinear characteristics, nonlinear systems
are more complicated than linear systems in identification and modeling. The bilinear system
is a special class of nonlinear systems, which widely exist and its model structure includes
the products of the inputs and states. The identification of bilinear systems has been exten-
sively studied [17,18]. Recently, the identification of nonlinear systems has attracted much
research attention and many identification algorithms have been proposed [19,20]. Mao et al.
presented a novel parameter separation based identification algorithm for Hammerstein sys-
tems [21] and an adaptive filtering based multi-innovation gradient algorithm for nonlinear
systems to improve the convergence speed [22].
The least squares algorithms contain the least squares based iterative (LSI) algorithms
[23], which are suitable for the off-line parameter estimation and the recursive least squares
(RLS) algorithms [24], which are suitable for the on-line parameter estimation. Xu et al.
proposed the recursive least squares and multi-innovation stochastic gradient parameter es-
timation methods for signal modeling [25], and an iterative parameter estimation algorithm
for signal models based on measured data [26]. The computational complexity of the least
squares algorithms depends on the dimensions of the information vectors [27]. The hierarchi-
cal identification principle is presented to reduce the computational burden by decomposing
the complex systems into several subsystems with smaller dimensions and fewer parameters
[28]. Xu et al. derived the parameter estimation algorithms for dynamical response signals
based on the multi-innovation theory and the hierarchical principle [29]. Ding et al. proposed
stochastic gradient estimation algorithms for a class of bilinear-in-parameter systems by using
the hierarchical identification principle [30].
The difficulty of identifying the stochastic systems is that the information vector contains
unmeasurable terms. Fortunately, the auxiliary model identification idea has been presented to
solve this problem [31]. Using the auxiliary model, Li et al. derived the least squares iterative
identification algorithms for the bilinear systems [32], and Wang et al. a novel data filtering
based parameter identification algorithm for multiple-input multiple-output systems [33]. Many
identification algorithms have been proposed for the bilinear systems. Li et al. presented a
filtering-based maximum likelihood gradient iterative estimation algorithm for bilinear systems
with autoregressive moving average noise [34] and least squares based iterative algorithms for
a bilinear system with autoregressive noise using the data filtering technique [35]. Zhang et al.
proposed several recursive state-space identification schemes for bilinear sate-space systems,
including a state filtering-based least squares algorithm with the hierarchical identification
principle [36], a hierarchical approach for joint parameter and state estimation algorithm
[37] and a combined state and parameter estimation algorithm [38].
The state-space identification algorithm can directly provide the state-space model, but
the computational complexity increases as the dimension of the parameter vector increases.
Different from the state-space identification algorithms, this paper focuses on the recursive
identification algorithms with less computational burden based on the input-output represen-
tation of the bilinear systems. The main contributions of this paper are as follows:

• The difficulty of identifying the bilinear state-space systems is that their model structures
contain the products of the states and inputs. An input-output representation of a bilin-
ear state-space system is derived for system identification through eliminating the state
variables in the model.
• Using the auxiliary model identification idea, a RGELS algorithm is proposed for the
bilinear systems with colored noises.
10104 L. Wang, Y. Ji and L. Wan et al. / Journal of the Franklin Institute 356 (2019) 10102–10122

• Using the hierarchical identification principle, the bilinear system is decomposed into three
subsystems with fewer variables and a 3S-RGELS algorithm is presented to reduce the
computational burden.
• The computational efficiency comparison is discussed between the RGELS algorithm and
the 3S-RGELS algorithm to illustrate the high efficiency of the 3S-RGELS algorithm.

The rest of this paper is organized as follows. Section 2 describes the identification model
of a bilinear system with colored noise. Section 3 proposes a RGELS algorithm by using the
auxiliary model. Section 4 gives a 3S-RGELS algorithm by using the auxiliary model and the
hierarchical identification principle, respectively. Section 5 discusses the computational effi-
ciency of the proposed algorithms. Section 6 provides an example to evaluate the effectiveness
of the proposed algorithms. Finally, we offer some concluding remarks in Section 7.

2. System description and identification model

Let us define some notations first. “A ࣏ X” or “X := A” stands for “A is defined as X”; I n

denotes an identity matrix of size n × n; 1n denotes an n × 1 vector whose elements are all
unity; z denotes a unit forward shift operator with zx(t ) = x(t + 1) and z−1 x(t ) = x(t − 1).
Consider the bilinear state-space system:
x(t + 1) = Ax(t ) + Bx(t )u(t ) + l u(t ), (1)

y(t ) = cx(t ), (2)

where x(t ) := [x1 (t ), x2 (t ), . . . , xn (t )]T ∈ Rn is the state vector, u(t ) ∈ R is the system input,
y(t ) ∈ R is the system output, A ∈ Rn×n , B ∈ Rn×n , l ∈ Rn and c ∈ R1×n are the system
parameter matrices and vectors:
⎡ ⎤
0 1 0 ... 0
⎢ 0 0 1 ... 0 ⎥
⎢ ⎥
⎢ .. . . . .. ⎥ ∈ Rn×n ,
A := ⎢ . .. .. .. . ⎥
⎢ ⎥
⎣ 0 0 0 ... 1 ⎦
−an −an−1 −an−2 . . . −a1
⎡ ⎤
l1
⎢ l2 ⎥ 
⎢ ⎥
⎢ .. ⎥ 0
l := ⎢ . ⎥ ∈ R , B :=
n×1
∈ Rn×n ,
⎢ ⎥ b
⎣ln−1 ⎦
ln
c := [1, 0, 0, . . . , 0] ∈ R1×n , b := [−bn , −bn−1 , −bn−2 , . . . , −b1 ] ∈ R1×n .
Based on the work in [35] to eliminate the state variables in Eqs. (1) and (2), the input-output
expression of the bilinear system can be equivalently written as
[A(z) + u(t − n)B(z)]y(t ) = [C(z) + u(t − n)D(z)]u(t ), (3)
where
A(z) := 1 + a1 z−1 + a2 z−2 + · · · + an z−n , ai ∈ R,
 L. Wang, Y. Ji and L. Wan et al. / Journal of the Franklin Institute 356 (2019) 10102–10122 10105

B(z) := b1 z−1 + b2 z−2 + · · · + bn z−n , bi ∈ R,

C(z) := c1 z−1 + c2 z−2 + · · · + cn z−n , ci ∈ R,
D(z) := d2 z−2 + d3 z−3 + · · · + dn z−n , di ∈ R.
The parameters ai , bi , li and the coefficients ci , di have the following relations:
[cn , . . . , c2 , c1 ] := [ln + an−1 l1 + an−2 l2 + · · · + a1 ln−1 , . . . , l2 + a1 l1 , l1 ] ∈ R1×n ,
[dn , . . . , d3 , d2 ] := [bn−1 l1 + bn−2 l2 + · · · + b1 ln−1 , . . . , b1 l1 ] ∈ R1×(n−1) .
As the actual systems are usually corrupted by various stochastic noises, we introduce a noise
term ω(t ) ∈ R to Eq. (3), and then we obtain a bilinear system with colored noise
[A(z) + u(t − n)B(z)]y(t ) = [C(z) + u(t − n)D(z)]u(t ) + ω(t ), (4)
where ω(t) may be a moving average (MA) process, an autoregressive (AR) process or an
autoregressive moving average process (ARMA). Without loss of generality, we consider ω(t)
as an ARMA process. That is, E (z)ω(t ) = F (z)v(t ), where v(t ) ∈ R is a white noise process
with zero mean, E(z) and F(z) are polynomials in z−1 , and
E (z) := 1 + e1 z−1 + e2 z−2 + · · · + ene z−ne , ei ∈ R,
F (z) := 1 + f1 z−1 + f2 z−2 + · · · + fn f z−n f , f i ∈ R.
Assume that the orders n, ne and nf are all known, and y(t ) = 0, u(t ) = 0 and v(t ) = 0 as
t ≤ 0. The objective is to develop new recursive algorithms for estimating the parameters ai ,
bi , ci , di , ei and fi using the measured input-output data {u(i), y(i) : i = 1, 2, . . . , t}.
Define the parameter vector θ and the information vector ϕ (t ) as
⎡ ⎤ ⎡ ⎤
θ s1 ϕs1 (t )
θ := ⎣θ s2 ⎦ ∈ Rn0 , ϕ (t ) := ⎣ϕs2 (t )⎦ ∈ Rn0 , n0 := 4n + ne + n f − 1,
θn ϕn (t )
where
θ s := [a1 , a2 , . . . , an , b1 , b2 , . . . , bn , c1 , c2 , . . . , cn , d2 , d3 , . . . , dn ]T ∈ Rn1 , n1 := 4n − 1,
θ := [a , a , . . . , a , b , b , . . . , b ]T ∈ Rn2 , n := 2n,
s1 1 2 n 1 2 n 2

θ s2 := [c1 , c2 , . . . , cn , d2 , d3 , . . . , dn ]T ∈ Rn3 , n3 := 2n − 1,
θ n := [e1 , e2 , . . . , ene , f1 , f2 , . . . , fn f ]T ∈ Rn4 , n4 := ne + n f ,
ϕs (t ) := [−y (t − 1), −y (t − 2), . . . , −y (t − n), −u(t − n)y(t − 1),
−u(t − n)y(t − 2), . . . , −u(t − n)y(t − n),
u(t − 1), u(t − 2), . . . , u(t − n), u(t − n)u(t − 2), u(t − n)u(t − 3), . . . ,
u(t − n)u(t − n)]T ∈ Rn1 ,
ϕs1 (t ) := [−y (t − 1), −y (t − 2), . . . , −y (t − n),
−u(t − n)y(t − 1), −u(t − n)y(t − 2), . . . , −u(t − n)y(t − n)]T ∈ Rn2 ,
ϕs2 (t ) := [u(t − 1), u(t − 2), . . . , u(t − n), u(t − n)u(t − 2), u(t − n)u(t − 3), . . . ,
u(t − n)u(t − n)]T ∈ Rn3 ,
ϕn (t ) := [−ω (t − 1), −ω (t − 2), . . . , −ω (t − ne ), v(t − 1), v(t − 2), . . . , v(t − n f )]T ∈ Rn4 .
10106 L. Wang, Y. Ji and L. Wan et al. / Journal of the Franklin Institute 356 (2019) 10102–10122

According to the above definitions, Eq. (4) can be written as

y(t ) = ϕT T T
s1 (t )θ s1 + ϕs2 (t )θ s2 + ϕn (t )θ n + v(t ) (5)

= ϕT
s (t )θ s + ω(t ) (6)

= ϕT (t )θ + v(t ). (7)
Eq. (7) is the identification model of the bilinear system in (4).
In what follows, a RGELS algorithm is derived to give the parameter estimates of a bilinear
system with colored noise by using the auxiliary model identification idea. Furthermore,
a 3S-RGELS algorithm is presented to enhance the computational efficiency by using the
hierarchical identification principle. The difficulty of the generalized extended least squares
estimation algorithms is that the extended information vectors contain unmeasurable terms.
To overcome the difficulty, the unknown variables in the information vectors are replaced
with their estimates by using the auxiliary model identification idea. A simulation example is
provided to evaluate the estimation accuracy and the computational efficiency of the proposed
algorithms.

3. The recursive generalized extended least squares algorithm

In this section, a RGELS algorithm is proposed based on the input-output representation

of the bilinear system with colored noise by using the auxiliary model. In order to generate
the parameter estimates, we construct an extended information vector and replace the unmea-
surable terms with their estimates based on the auxiliary model. Then, we can derive the
RGELS algorithm for the bilinear system with colored noise.
Use the input-output data to define the stacked vector Y t and the stacked matrix Φt as
⎡ ⎤ ⎡ ⎤
y(1) ϕT (1)
⎢y(2)⎥ ⎢ T ⎥
⎢ ⎥ ⎢ϕ (2)⎥
Y t := ⎢ . ⎥ ∈ R , Φt := ⎢
t
⎢ .. ⎥ ∈ R
⎥ t×n0
.
⎣ .. ⎦ ⎣ . ⎦
y(t ) ϕT (t )
According to (7), define a quadratic criterion function:
J1 (θ) := Y t − Φt θ 2 . (8)
Minimizing J1 (θ) and letting its partial derivative with respect to θ be zero, we can derive
the following recursive least squares relations:
ˆ ) = θ(t
θ(t ˆ − 1) + P (t )ϕ (t )[y(t ) − ϕT (t )θ(t
ˆ − 1)], (9)

P −1 (t ) = P −1 (t − 1) + ϕ (t )ϕT (t ). (10)
Lemma 1. Let A ∈ Rn×n , B ∈ Rn×r , C ∈ Rr×n . Assume that A and (I + C A−1 B ) are invert-
ible, then the following equality holds,
(A + BC )−1 = A−1 − A−1 (I + C A−1 B )−1C A−1 .
 L. Wang, Y. Ji and L. Wan et al. / Journal of the Franklin Institute 356 (2019) 10102–10122 10107

Applying Lemma 1, Eqs. (9) and (10) can be equivalently rewritten as

ˆ ) = θ(t
θ(t ˆ − 1) + L (t )[y(t ) − ϕT (t )θ(t
ˆ − 1)], (11)

L (t ) = P (t − 1)ϕ (t )[1 + ϕT (t )P (t − 1)ϕ (t )]−1 , (12)

P (t ) = [I n0 − L (t )ϕT (t )]P (t − 1). (13)

However, the information vector ϕ (t ) in Eqs. (11)–(13) contains the unmeasurable terms ω(t −
ˆ )
i) (i = 1, 2, . . . , ne ) and v(t − i) (i = 1, 2, . . . , n f ), so Eq. (11) cannot give the estimate θ(t
directly. The solution is to replace the unmeasurable terms ω(t − i) and v(t − i) in ϕ (t ) with
their corresponding estimates ωˆ (t − i) and vˆ(t − i). Define the estimates of ϕn (t ) and ϕ (t )
as
ϕˆ n (t ) := [−ωˆ (t − 1), −ωˆ (t − 2), . . . , −ωˆ (t − ne ), vˆ(t − 1), vˆ(t − 2), . . . , vˆ(t − n f )]T ∈ Rn4 ,
(14)

ϕs (t )
ϕˆ (t ) := ∈ Rn0 . (15)
ϕˆ n (t )

From Eq. (6), we have ω(t ) = y(t ) − ϕT ˆ

s (t )θ s . Replacing θ s with its estimate θ s (t ), the esti-
mate of ω(t) can be computed by

ωˆ (t ) = y(t ) − ϕT ˆ
s (t )θ s (t ). (16)

From Eq. (7), we have v(t ) = y(t ) − ϕT (t )θ. Replacing ϕ (t ) and θ with their estimates ϕˆ (t )
ˆ ), respectively, the estimate of v(t ) can be computed by
and θ(t
ˆ ).
vˆ(t ) = y(t ) − ϕˆ T (t )θ(t (17)
Replacing ϕ (t ) in Eqs. (11)–(13) with its estimate ϕˆ (t ), we have
ˆ ) = θ(t
θ(t ˆ − 1) + L (t )[y(t ) − ϕˆ T (t )θ(t
ˆ − 1)], (18)

L (t ) = P (t − 1)ϕˆ (t )[1 + ϕˆ T (t )P (t − 1)ϕˆ (t )]−1 , (19)

P (t ) = [I n0 − L (t )ϕˆ T (t )]P (t − 1). (20)

Combining Eqs. (14)–(20), we can obtain the RGELS algorithm for the bilinear system with
colored noise as follows:
ˆ ) = θ(t
θ(t ˆ − 1) + L (t )[y(t ) − ϕˆ T (t )θ(t
ˆ − 1)], (21)

L (t ) = P (t − 1)ϕˆ (t )[1 + ϕˆ T (t )P (t − 1)ϕˆ (t )]−1 , (22)

P (t ) = [I n0 − L (t )ϕˆ T (t )]P (t − 1), (23)

10108 L. Wang, Y. Ji and L. Wan et al. / Journal of the Franklin Institute 356 (2019) 10102–10122


ϕs (t )
ϕˆ (t ) = , (24)
ϕˆ n (t )

ϕs (t ) = [−y (t − 1), −y (t − 2), . . . , −y (t − n), −u(t − n)y(t − 1),

−u(t − n)y(t − 2), . . . , −u(t − n)y(t − n), u(t − 1), u(t − 2), . . . ,
u(t − n), u(t − n)u(t − 2), u(t − n)u(t − 3), . . . , u(t − n)u(t − n)]T , (25)

ϕˆ n (t ) = [−ωˆ (t − 1), −ωˆ (t − 2), . . . , −ωˆ (t − ne ), vˆ(t − 1), vˆ(t − 2), . . . , vˆ(t − n f )]T , (26)

ωˆ (t ) = y(t ) − ϕT ˆ
s (t )θ s (t ), (27)

ˆ ),
vˆ(t ) = y(t ) − ϕˆ T (t )θ(t (28)

θˆ s (t ) = [aˆ1 (t ), aˆ2 (t ), . . . , aˆn (t ), bˆ 1 (t ), bˆ 2 (t ), . . . , bˆ n (t ), cˆ1 (t ), cˆ2 (t ), · · · , cˆn (t ),

dˆ2 (t ), dˆ3 (t ), · · · , dˆn (t )]T , (29)

θˆ n (t ) = [eˆ1 (t ), eˆ2 (t ), · · · , eˆne (t ), fˆ1 (t ), fˆ2 (t ), · · · , fˆn f (t )]T . (30)

The computation procedures of the RGELS algorithm in Eqs. (21)–(30) are listed in the
following.

ˆ
1. Initialize: let t = n0 , θ(0) = 1n0 /p0 , P (0) = p0 I n0 , ωˆ (t − i) = 0 (i = 1, 2, . . . , ne ), vˆ(t −
i) = 0 (i = 1, 2, . . . , n f ), and p0 is taken as a large number, e.g., p0 = 106 .
2. Collect the input-output data u(t) and y(t). Form ϕs (t ), ϕˆ n (t ) and ϕˆ (t ) by Eqs. (25),
(26) and (24), respectively.
3. Compute the gain vector L (t ) by Eq. (22). Compute the covariance matrix P (t ) by Eq. (23).
4. Update the parameter estimate θ(t ˆ ) by Eq. (21).
5. Compute ωˆ (t ) and vˆ(t ) by Eqs. (27) and (28), respectively.
6. Increase t by 1 and go to Step 2.

The proposed RGELS algorithm for the bilinear system with colored noise can be extended
to develop new identification approaches for multivariable systems with colored noise such
as signal modeling and communication networked systems [39–49].

4. The three-stage recursive generalized extended least squares algorithm

In order to enhance the computational efficiency, the bilinear system is decomposed into
three fictitious subsystems with fewer variables by using the hierarchical identification prin-
ciple, and then each parameter vector involving the fictitious subsystems is estimated by the
least squares based algorithm. The detail is as follows.
Define three intermediate variables:
y1 (t ) := y(t ) − ϕT T
s2 (t )θ s2 − ϕn (t )θ n , (31)
 L. Wang, Y. Ji and L. Wan et al. / Journal of the Franklin Institute 356 (2019) 10102–10122 10109

y2 (t ) := y(t ) − ϕT T
s1 (t )θ s1 − ϕn (t )θ n , (32)

y3 (t ) := y(t ) − ϕT T
s1 (t )θ s1 − ϕs2 (t )θ s2 . (33)
Using Eqs. (31)–(33), Eq. (4) can be decomposed into the following three subsystems:

y1 (t ) = ϕT
s1 (t )θ s1 + v(t ), (34)

y2 (t ) = ϕT
s2 (t )θ s2 + v(t ), (35)

y3 (t ) = ϕT
n (t )θ n + v(t ). (36)
Use the input-output data to define
⎡ ⎤ ⎡ ⎤
y1 (1) y2 (1)
⎢y1 (2)⎥ ⎢y2 (2)⎥
⎢ ⎥ ⎢ ⎥
Y 1,t := ⎢ . ⎥ ∈ Rt , Y 2,t := ⎢ . ⎥ ∈ Rt ,
⎣ .. ⎦ ⎣ .. ⎦
y1 (t ) y2 (t )
⎡ ⎤ ⎡ ⎤
y3 (1) ϕTs1 (1)
⎢y3 (2)⎥ ⎢ T ⎥
⎢ ⎥ ⎢ϕs1 (2)⎥
Y 3,t := ⎢ . ⎥ ∈ R , Φ1,st := ⎢
t
⎢ .. ⎥ ∈ R
⎥ t×n2
,
⎣ . ⎦. ⎣ . ⎦
y3 (t ) ϕTs1 (t )
⎡ ⎤ ⎡ ⎤
T
ϕs2 (1) ϕT n (1)
⎢ T ⎥ ⎢ T ⎥
⎢ϕs2 (2)⎥ ⎢ϕn (2)⎥
Φ2,st := ⎢ . ⎥ ∈ R , Φnt := ⎢
⎢ ⎥ t×n3
⎢ .. ⎥ ∈ R .
⎥ t×n4

⎣ . ⎦ . ⎣ . ⎦
ϕTs2 (t ) ϕT n (t )

According to Eqs. (34)–(36), define three quadratic criterion functions:

J2 (θ s1 ) := Y 1,t − Φ1,st θ s1 2 , (37)

J3 (θ s2 ) := Y 2,t − Φ2,st θ s2 2 , (38)

J4 (θ n ) := Y 3,t − Φnt θ n 2 . (39)

Minimizing the quadratic criterion functions J2 (θ s1 ), J3 (θ s2 ) and J4 (θ n ), and letting the partial
derivatives of J2 (θ s1 ), J3 (θ s2 ) and J4 (θ n ) with respect to θ s1 , θ s2 and θ n be zero, we can derive
the following recursive least squares relations:

θˆ s1 (t ) = θˆ s1 (t − 1) + Ls1 (t )[y1 (t ) − ϕT ˆ
s1 (t )θ s1 (t − 1)], (40)

Ls1 (t ) = P s1 (t − 1)ϕs1 (t )[1 + ϕT −1

s1 (t )P s1 (t − 1)ϕs1 (t )] , (41)
10110 L. Wang, Y. Ji and L. Wan et al. / Journal of the Franklin Institute 356 (2019) 10102–10122

P s1 (t ) = [I n2 − Ls1 (t )ϕT
s1 (t )]P s1 (t − 1), (42)

θˆ s2 (t ) = θˆ s2 (t − 1) + Ls2 (t )[y2 (t ) − ϕT ˆ
s2 (t )θ s2 (t − 1)], (43)

Ls2 (t ) = P s2 (t − 1)ϕs2 (t )[1 + ϕT −1

s2 (t )P s2 (t − 1)ϕs2 (t )] , (44)

P s2 (t ) = [I n3 − Ls2 (t )ϕT
s2 (t )]P s2 (t − 1), (45)

θˆ n (t ) = θˆ n (t − 1) + Ln (t )[y3 (t ) − ϕT ˆ
n (t )θ n (t − 1)], (46)

Ln (t ) = P n (t − 1)ϕn (t )[1 + ϕT −1
n (t )P n (t − 1)ϕn (t )] , (47)

P n (t ) = [I n4 − Ln (t )ϕT
n (t )]P n (t − 1). (48)
Inserting Eqs. (31) into (40), (32) into (43), and (33) into (46), we can obtain the following
relations:
θˆ s1 (t ) = θˆ s1 (t − 1) + Ls1 (t )[y(t ) − ϕT T T ˆ
s2 (t )θ s2 − ϕn (t )θ n − ϕs1 (t )θ s1 (t − 1)], (49)

θˆ s2 (t ) = θˆ s2 (t − 1) + Ls2 (t )[y(t ) − ϕT T T ˆ
s1 (t )θ s1 − ϕn (t )θ n − ϕs2 (t )θ s2 (t − 1)], (50)

θˆ n (t ) = θˆ n (t − 1) + Ln (t )[y(t ) − ϕT T T ˆ
s1 (t )θ s1 − ϕs2 (t )θ s2 − ϕn (t )θ n (t − 1)]. (51)
Similarly, the right-hand sides of Eqs. (49)–(51) contain the unknown parameter vectors θ s1 ,
θ s2 and θ n , and the unknown information vector ϕn (t ), which contains the unmeasured noise
terms ω(t − i) (i = 1, 2, . . . , ne ) and v(t − i) (i = 1, 2, . . . , n f ), so Eq. (49)–(51) cannot
give the estimates θˆ s1 (t ), θˆ s2 (t ) and θˆ n (t ) directly. The solution is to replace θ s1 , θ s2 , θ n , and
ϕn (t ) with their corresponding estimates. Replacing the unknown items ω(t − i) and v(t − i)
in ϕn (t ) with their estimates ωˆ (t − i) and vˆ(t − i), define the estimates of ϕn (t ) and ϕ (t )
as
ϕˆ n (t ) := [−ωˆ (t − 1), −ωˆ (t − 2), . . . , −ωˆ (t − ne ), vˆ(t − 1), vˆ(t − 2), . . . , vˆ(t − n f )]T ∈ Rn4 ,
(52)

⎡ ⎤
ϕs1 (t )
ϕˆ (t ) := ⎣ϕs2 (t )⎦ ∈ Rn0 . (53)
ϕˆ n (t )

From Eq. (6), we have ω(t ) = y(t ) − ϕT ˆ

s (t )θ s . Replacing θ s with its estimate θ s (t ), the esti-
mate of ω(t) can be computed by

ωˆ (t ) = y(t ) − ϕT ˆ
s (t )θ s (t ). (54)
 L. Wang, Y. Ji and L. Wan et al. / Journal of the Franklin Institute 356 (2019) 10102–10122 10111

From Eq. (7), we have v(t ) = y(t ) − ϕT (t )θ. Replacing ϕ (t ) and θ with their estimates ϕˆ (t )
ˆ ), respectively, the estimate of v(t ) can be computed by
and θ(t
ˆ ).
vˆ(t ) = y(t ) − ϕˆ T (t )θ(t (55)

Replacing ϕn (t ), θ n and θ s2 in Eq. (49) with their estimates ϕˆ n (t ), θˆ n (t − 1) and θˆ s2 (t − 1),

replacing ϕn (t ), θ n and θ s1 in Eq. (50) with their estimates ϕˆ n (t ), θˆ n (t − 1) and θˆ s1 (t ), and
replacing ϕn (t ), θ s1 and θ s2 in Eq. (51) with their estimates ϕˆ n (t ), θˆ s1 (t ) and θˆ s2 (t ), we
have
θˆ (t ) = θˆ (t − 1) + L (t )[y(t ) − ϕˆ T (t )θˆ (t − 1) − ϕT (t )θˆ (t − 1) − ϕT (t )θˆ (t − 1)],
s1 s1 s1 n n s1 s1 s2 s2
(56)

θˆ s2 (t ) = θˆ s2 (t − 1) + Ls2 (t )[y(t ) − ϕˆ T ˆ T ˆ T ˆ
n (t )θ n (t − 1) − ϕs1 (t )θ s1 (t ) − ϕs2 (t )θ s2 (t − 1)], (57)

θˆ n (t ) = θˆ n (t − 1) + Ln (t )[y(t ) − ϕT ˆ T ˆ ˆT
s1 (t )θ s1 (t ) − ϕs2 (t )θ s2 (t ) − ϕ
ˆ
n (t )θ n (t − 1)]. (58)
Therefore, we can obtain the 3S-RGELS algorithm for the bilinear system with colored noise
as follows:
θˆ (t ) = θˆ (t − 1) + L (t )[y(t ) − ϕˆ T (t )θˆ (t − 1) − ϕT (t )θˆ (t − 1) − ϕT (t )θˆ (t − 1)],
s1 s1 s1 n n s1 s1 s2 s2
(59)

θˆ s2 (t ) = θˆ s2 (t − 1) + Ls2 (t )[y(t ) − ϕˆ T ˆ T ˆ T ˆ
n (t )θ n (t − 1) − ϕs1 (t )θ s1 (t ) − ϕs2 (t )θ s2 (t − 1)], (60)

θˆ n (t ) = θˆ n (t − 1) + Ln (t )[y(t ) − ϕT ˆ T ˆ ˆT
s1 (t )θ s1 (t ) − ϕs2 (t )θ s2 (t ) − ϕ
ˆ
n (t )θ n (t − 1)], (61)

Ls1 (t ) = P s1 (t − 1)ϕs1 (t )[1 + ϕT −1

s1 (t )P s1 (t − 1)ϕs1 (t )] , (62)

P s1 (t ) = [I n2 − Ls1 (t )ϕT
s1 (t )]P s1 (t − 1), (63)

Ls2 (t ) = P s2 (t − 1)ϕs2 (t )[1 + ϕT −1

s2 (t )P s2 (t − 1)ϕs2 (t )] , (64)

P s2 (t ) = [I n3 − Ls2 (t )ϕT
s2 (t )]P s2 (t − 1), (65)

Ln (t ) = P n (t − 1)ϕn (t )[1 + ϕˆ T ˆ n (t )]−1 ,

n (t )P n (t − 1)ϕ (66)

P n (t ) = [I n4 − Ln (t )ϕˆ T
n (t )]P n (t − 1), (67)

ϕs (t ) = [−y (t − 1), −y (t − 2), . . . , −y (t − n), −u(t − n)y(t − 1),

−u(t − n)y(t − 2), . . . , −u(t − n)y(t − n), u(t − 1), u(t − 2), . . . , u(t − n),
10112 L. Wang, Y. Ji and L. Wan et al. / Journal of the Franklin Institute 356 (2019) 10102–10122

u(t − n)u(t − 2), u(t − n)u(t − 3), . . . , u(t − n)u(t − n)]T , (68)

ϕs1 (t ) = [−y (t − 1), −y (t − 2), . . . , −y (t − n),

−u(t − n)y(t − 1), −u(t − n)y(t − 2), . . . , −u(t − n)y(t − n)]T , (69)

ϕs2 (t ) = [u(t − 1), u(t − 2), . . . , u(t − n),

u(t − n)u(t − 2), u(t − n)u(t − 3), . . . , u(t − n)u(t − n)]T , (70)

ϕˆ n (t ) = [−ωˆ (t − 1), −ωˆ (t − 2), . . . , −ωˆ (t − ne ), vˆ(t − 1), vˆ(t − 2), . . . , vˆ(t − n f )]T , (71)

ωˆ (t ) = y(t ) − ϕT ˆ
s (t )θ s (t ), (72)

ˆ ),
vˆ(t ) = y(t ) − ϕˆ T (t )θ(t (73)

θˆ s (t ) = [aˆ1 (t ), aˆ2 (t ), . . . , aˆn (t ), bˆ 1 (t ), bˆ 2 (t ), . . . , bˆ n (t ),

cˆ1 (t ), cˆ2 (t ), . . . , cˆn (t ), dˆ2 (t ), dˆ3 (t ), . . . , dˆn (t )]T , (74)

θˆ s1 (t ) = [aˆ1 (t ), aˆ2 (t ), . . . , aˆn (t ), bˆ 1 (t ), bˆ 2 (t ), . . . , bˆ n (t )]T , (75)

θˆ s2 (t ) = [cˆ1 (t ), cˆ2 (t ), . . . , cˆn (t ), dˆ2 (t ), dˆ3 (t ), . . . , dˆn (t )]T , (76)

θˆ n (t ) = [eˆ1 (t ), eˆ2 (t ), . . . , eˆne (t ), fˆ1 (t ), fˆ2 (t ), . . . , fˆn f (t )]T . (77)

The computational procedures of the 3S-RGELS algorithm in Eqs. (59)–(77) are listed in the
following.

1. Initialize: let t = n0 , θˆ s1 (0) = 1n2 /p0 , P s1 (0) = p0 I n2 , θˆ s2 (0) = 1n3 /p0 , P s2 (0) =
p0 I n3 , θˆ n (0) = 1n4 /p0 , P n (0) = p0 I n4 , ωˆ (t − i) = 0 (i = 1, 2, . . . , ne ), vˆ(t − i) = 0 (i =
1, 2, . . . , n f ), and p0 is taken as a large number, e.g., p0 = 106 .
2. Collect the input-output data u(t) and y(t). Form ϕs1 (t ), ϕs2 (t ) and ϕˆ n (t ) by Eqs. (69),
(70) and (71), respectively.
3. Compute the gain vector Ls1 (t ) and the covariance matrix P s1 (t ) by Eqs. (62) and (63).
Compute the gain vector Ls2 (t ) and the covariance matrix P s2 (t ) by (64) and (65). Compute
the gain vector Ln (t ) and the covariance matrix P n (t ) by Eqs. (66) and (67).
4. Update the parameter estimates θˆ s1 (t ), θˆ s2 (t ) and θˆ n (t ) by Eqs. (59), (60) and (61),
respectively.
5. Compute ωˆ (t ) and vˆ(t ) by Eqs. (72) and (73), respectively.
6. Increase t by 1 and go to Step 2.
 L. Wang, Y. Ji and L. Wan et al. / Journal of the Franklin Institute 356 (2019) 10102–10122 10113

Table 1
The computational efficiency of the RGELS algorithm.

Variables Computational sequences Multiplications Additions

ˆ )
θ(t ˆ ) = θ(t
θ(t ˆ − 1) + L (t )e(t ) n0 n0
ˆ − 1)
e(t ) := y(t ) − ϕˆ T (t )θ(t n0 n0
L (t ) L (t ) = ξ(t )[1 + ϕˆ T (t )ξ(t )]−1 2n0 n0
ξ(t ) := P (t − 1)ϕˆ (t ) n02 (n0 − 1)n0
P (t ) P (t ) = P (t − 1) − L (t )ξ T (t ) n02 n02
ωˆ (t ) ωˆ (t ) = y(t ) − ϕT ˆ
s (t )θ s (t ) n1 n1
vˆ(t ) ˆ )
vˆ(t ) = y(t ) − ϕˆ T (t )θ(t n0 n0
Sum 2n02 + 5n0 + n1 2n02 + 3n0 + n1
Total flops N1 := 4n02 + 8n0 + 2n1

5. The computational efficiency

The computational efficiency is usually counted by the flop. Here, an addition, a mul-
tiplication, a subtraction, a division all is a flop. In general, a division is considered as a
multiplication and a subtraction is considered as an addition. Thus, the computational amount
of an identification algorithm can be expressed by additions and multiplications.
The multiplications and additions of the RGELS and 3S-RGELS algorithms at each step
are shown in Tables 1–2. The total flop numbers of the RGELS and 3S-RGELS algorithms at
each step are N1 = 4n02 + 8n0 + 2n1 and N2 = 4n22 + 4n32 + 4n42 + 11n4 + 13n1 , respectively.
For comparison, the flops of the RGELS and 3S-RGELS algorithms are listed in Table 3,
where n0 = 4n + n4 − 1, n1 = 4n − 1, n2 = 2n, n3 = 2n − 1. Taking n = 4, n4 = 8 gives an
example of the flop comparison in Table 4.
The flop difference between the RGELS algorithm and the 3S-RGELS algorithm is
N1 − N2 = 8n2 n3 + 8n2 n4 + 8n3 n4 − 3n1 − 3n4
= 8n2 n3 + 8n1 n4 − 3n1 − 3n4
= 8n2 n3 + 2n1 n4 + 3n1 (n4 − 1) + 3n4 (n1 − 1) > 0.
Then, we can draw some conclusions.

• N1 > N2 means that the 3S-RGELS algorithm is more flop-efficient than the RGELS algo-
rithm.
• The 3S-RGELS algorithm saves 57.47% = N1N−N 1
2
computational cost compared with the
RGELS algorithm according to Table 4.

6. Example

Consider the following bilinear system with colored noise:

[A(z) + u(t − n)B(z)]y(t ) = [C(z) + u(t − n)D(z)]u(t ) + ω(t ),
F (z)
ω(t ) = v(t ),
E (z)
A(z) = 1 + a1 z−1 + a2 z−2 = 1 + 0.91z−1 + 0.63z−2 ,
B(z) = b1 z−1 + b2 z−2 = 0.2z−1 − 0.18z−2 ,
10114 L. Wang, Y. Ji and L. Wan et al. / Journal of the Franklin Institute 356 (2019) 10102–10122

Table 2
The computational efficiency of the 3S-RGELS algorithm.

Variables Computational sequences Multiplications Additions

θˆ s1 (t ) θˆ s1 (t ) = θˆ s1 (t − 1) + Ls1 (t )e(t ) n2 n2
e1 (t ) := y(t ) − ϕˆ T ˆ T ˆ
n (t )θ n (t − 1) − ϕs1 (t )θ s1 (t − 1) − n0 n0
ϕT (t ) ˆ s2 (t − 1)
θ
s2
Ls1 (t ) L (t ) = ξ (t )[1 + ϕˆ T (t )ξ (t )]−1
s1 s1 s1 s1 2n2 n2
ξ s1 (t ) := P s1 (t − 1)ϕˆ s1 (t ) n22 n22 − n2
P s1 (t ) P s1 (t ) = P s1 (t − 1) − Ls1 (t )ξ s1 T (t ) n22 n22
θˆ s2 (t ) ˆθ s2 (t ) = θˆ s2 (t − 1) + Ls2 (t )e(t ) n3 n3
e2 (t ) := y(t ) − ϕˆ T ˆ T ˆ
n (t )θ n (t − 1) − ϕs1 (t )θ s1 (t ) − n0 n0
T ˆ
ϕs2 (t )θ s2 (t − 1)
Ls2 (t ) L (t ) = ξ (t )[1 + ϕˆ T (t )ξ (t )]−1
s2 s2 s2 s2 2n3 n3
ξ s2 (t ) := P s2 (t − 1)ϕˆ s2 (t ) n32 n32 − n3
P s2 (t ) P s2 (t ) = P s2 (t − 1) − Ls2 (t )ξ s2 T (t ) n32 n32
θˆ n (t ) ˆθ n (t ) = θˆ n (t − 1) + Ln (t )e(t ) n4 n4
en (t ) := n0 n0
y(t ) − ϕˆ T ˆ T ˆ T ˆ
n (t )θ n (t − 1) − ϕs1 (t )θ s1 (t ) − ϕs2 (t )θ s2 (t )
Ln (t ) T
L (t ) = ξ (t )[1 + ϕˆ (t )ξ (t )] −1 2n4 n4
n n n n
ξ n (t ) := P n (t − 1)ϕˆ n (t ) n42 n42 − n4
P n (t ) P n (t ) = P n (t − 1) − Ln (t )ξ nT (t ) n42 n42
ωˆ (t ) ωˆ (t ) = y(t ) − ϕT ˆ
s (t )θ s (t ) n1 n1
vˆ(t ) T ˆ
vˆ(t ) = y(t ) − ϕˆ (t )θ(t ) n0 n0
Sum 2n22 + 2n32 + 2n42 + 6n4 + 7n1 2n22 + 2n32 + 2n42 +
5n4 + 6n1
Total flops N2 := 4n22 + 4n32 + 4n42 + 11n4 + 13n1

Table 3
The flop amounts of the algorithms.

Algorithms Multiplications Additions Flops

RGELS 2n02 + 5n0 + n1 2n02 + 3n0 + n1 N1 := 4n02 + 8n0 + 2n1
3S-RGELS 2n22 + 2n32 + 2n42 + 6n4 + 7n1 2n22 + 2n32 + 2n42 + 5n4 + 6n1 N2 := 4n22 + 4n32 + 4n44 + 11n4 + 13n1

Table 4
The comparison of the flop amounts with n = 4, n4 = 8.

Algorithms Multiplications Additions Flops

RGELS 1188 1182 2330
3S-RGELS 507 484 991

C(z) = c1 z−1 + c2 z−2 = 0.8z−1 − 2.3z−2 ,

D(z) = d2 z−2 = 0.16z−2 ,
E (z) = 1 + e1 z−1 = 1 − 0.17z−1 ,
F (z) = 1 + f1 z−1 = 1 + 0.1z−1 .
 L. Wang, Y. Ji and L. Wan et al. / Journal of the Franklin Institute 356 (2019) 10102–10122 10115

Table 5
The RGELS estimates and their errors with σ 2 = 2.02 .

t a1 a2 b1 b2 c1 c2 d2 e1 f1 δ (%)
100 0.94511 0.68054 0.21753 −0.16403 1.06025 −2.28314 0.11286 −0.11267 0.07945 10.36202
200 0.92659 0.68918 0.20194 −0.15722 1.10432 −2.30620 0.05459 −0.12157 0.14470 12.41566
500 0.92111 0.65839 0.21959 −0.14413 0.82198 −2.23270 0.17130 −0.20284 0.07804 3.57147
1000 0.91464 0.64820 0.21364 −0.15650 0.79828 −2.33718 0.16286 −0.16108 0.17643 3.39516
2000 0.91070 0.64067 0.19203 −0.17973 0.83206 −2.38696 0.12458 −0.17259 0.13553 3.93510
3000 0.91429 0.63283 0.19629 −0.18241 0.80760 −2.33598 0.13915 −0.16729 0.12384 1.81737
4000 0.91610 0.63485 0.19863 −0.18013 0.81038 −2.30848 0.13787 −0.15379 0.13579 1.76663
True values 0.91000 0.63000 0.20000 −0.18000 0.80000 −2.30000 0.16000 −0.17000 0.10000

Table 6
The RGELS estimates and their errors with σ 2 = 5.02 .

t a1 a2 b1 b2 c1 c2 d2 e1 f1 δ (%)
100 0.81264 0.61253 0.20329 −0.18788 1.72532 −2.58738 0.11402 −0.13753 −0.14166 37.22107
200 0.85015 0.66291 0.21990 −0.15281 1.54126 −2.41774 −0.00163 −0.24036 −0.10668 29.69316
500 0.86728 0.63159 0.21249 −0.13919 0.83701 −2.24159 0.19634 −0.45116 −0.26470 17.44107
1000 0.86805 0.63075 0.21305 −0.14955 0.78625 −2.46788 0.16398 −0.26842 −0.00266 8.40152
2000 0.88851 0.63457 0.18421 −0.17767 0.88857 −2.54402 0.10121 −0.26612 −0.00620 11.23870
3000 0.89898 0.62732 0.19284 −0.18464 0.83106 −2.42313 0.11946 −0.25313 −0.00438 7.00449
4000 0.90180 0.62886 0.19450 −0.18126 0.83331 −2.35282 0.10905 −0.21188 0.03969 4.05450
True values 0.91000 0.63000 0.20000 −0.18000 0.80000 −2.30000 0.16000 −0.17000 0.10000

Table 7
The 3S-RGELS estimates and their errors with σ 2 = 2.02 .

t a1 a2 b1 b2 c1 c2 d2 e1 f1 δ (%)
100 0.92286 0.66520 0.22533 −0.19691 1.04532 −2.25978 0.30166 0.45587 0.67146 33.17410
200 0.90932 0.67552 0.22022 −0.16284 1.07254 −2.30180 0.19364 0.11312 0.37836 17.98354
500 0.91460 0.65222 0.22952 −0.14492 0.81340 −2.25483 0.21843 −0.20421 0.07003 3.76189
1000 0.91063 0.64498 0.22062 −0.15370 0.79601 −2.34535 0.20063 −0.15439 0.18028 4.01494
2000 0.91369 0.64154 0.20022 −0.17629 0.83262 −2.38068 0.16101 −0.17159 0.13972 3.57417
3000 0.91709 0.63457 0.20224 −0.18012 0.80945 −2.33396 0.16344 −0.18883 0.10269 1.52344
4000 0.91731 0.63625 0.20364 −0.17828 0.81156 −2.30918 0.15514 −0.16557 0.12279 1.10438
True values 0.91000 0.63000 0.20000 −0.18000 0.80000 −2.30000 0.16000 −0.17000 0.10000

The model parameters in the simulation should ensure the stability of the bilinear system.
Under the persistent excitation condition, the convergence of the identification algorithm is
not affected by the initial value, and we set the initial value as a small positive real vector.
In the simulation, the input {u(t)} is taken as an uncorrelated persistent excitation sequence
with zero mean and unit variance, {v(t )} as a white noise sequence with zero mean and
different variances σ 2 = 2.02 and σ 2 = 5.02 , respectively. Applying the RGELS algorithm
in Eqs. (21)–(30) and the 3S-RGELS algorithm in Eqs. (59)–(77) to identify this example
ˆ ) − θ/θ versus the data length
system. The parameter estimates and the errors δ(t ) := θ(t
t are shown in Tables 5–8.
10116 L. Wang, Y. Ji and L. Wan et al. / Journal of the Franklin Institute 356 (2019) 10102–10122

Table 8
The 3S−RGELS estimates and their errors with σ 2 = 5.02 .

t a1 a2 b1 b2 c1 c2 d2 e1 f1 δ (%)
100 0.94423 0.68433 0.21031 −0.14794 1.42882 −2.25379 0.01260 −0.10989 0.05696 24.28368
200 0.93020 0.71276 0.19330 −0.15498 1.50399 −2.27921 −0.09176 −0.12016 0.14382 28.00145
500 0.91934 0.66352 0.20342 −0.13392 0.81344 −2.14437 0.18167 −0.21729 0.05550 6.66906
1000 0.91653 0.65568 0.20759 −0.15083 0.78792 −2.41034 0.15299 −0.17798 0.15797 4.88558
2000 0.91613 0.64993 0.18005 −0.17938 0.87526 −2.51834 0.07307 −0.18858 0.12220 9.26268
3000 0.92135 0.63670 0.19085 −0.18505 0.81729 −2.39303 0.10824 −0.16768 0.12887 4.18211
4000 0.92460 0.63733 0.19303 −0.18117 0.82417 −2.32341 0.10117 −0.14811 0.14875 3.26344
True values 0.91000 0.63000 0.20000 −0.18000 0.80000 −2.30000 0.16000 −0.17000 0.10000

For model validation, we use the RGELS estimates and the 3S-RGELS estimates to con-
struct the estimated model, respectively, that is

Eˆ (z ) Eˆ (z )
yˆ(t ) = y(t ) − [Aˆ (z) + u(t − n)Bˆ (z )] y(t ) + [Cˆ (z) + u(t − n)Dˆ (z )] u(t ). (78)
Fˆ (z ) Fˆ (z )
Eˆ (z) Eˆ (z)
Define yˆ f (t ) := Fˆ (z)
y(t ) and uˆ f (t ) := Fˆ (z)
u(t ), and then Eq. (78) can be equivalent to

yˆ(t ) = y(t ) − [Aˆ (z) + u(t − n)Bˆ (z)]yˆ f (t ) + [Cˆ (z) + u(t − n)Dˆ (z)]uˆ f (t ).
Using the RGELS parameter estimates in Table 5 at t = 4000 to construct the RGELS esti-
mated model
yˆ1 (t ) = y(t ) − [Aˆ 1 (z) + u(t − n)Bˆ 1 (z)]yˆ f 1 (t ) + [Cˆ1 (z) + u(t − n)Dˆ 1 (z)]uˆ f 1 (t ),
Eˆ1 (z) Eˆ1 (z)
yˆ f 1 (t ) = y(t ), uˆ f 1 (t ) = u(t ),
Fˆ1 (z) Fˆ1 (z)
Aˆ 1 (z) = 1 + 0.91610z−1 + 0.63485z−2 ,
Bˆ 1 (z) = 0.19863z−1 − 0.18013z−2 ,
Cˆ1 (z) = 0.81038z−1 − 2.30848z−2 ,
Dˆ 1 (z) = 0.13787z−2 ,
Eˆ1 (z) = 1 − 0.15379z−1 ,
Fˆ1 (z) = 1 + 0.13579z−1 .
Using the 3S-RGELS parameter estimates in Table 7 at t = 4000 to construct the 3S-RGELS
estimated model
yˆ2 (t ) = y(t ) − [Aˆ 2 (z) + u(t − n)Bˆ 2 (z)]yˆ f 2 (t ) + [Cˆ2 (z) + u(t − n)Dˆ 2 (z)]uˆ f 2 (t ),
Eˆ2 (z) Eˆ2 (z)
yˆ f 2 (t ) = y(t ), uˆ f 2 (t ) = u(t ),
Fˆ2 (z) Fˆ2 (z)
Aˆ 2 (z) = 1 + 0.91731z−1 + 0.63625z−2 ,
Bˆ 2 (z) = 0.20364z−1 − 0.17828z−2 ,
Cˆ2 (z) = 0.81156z−1 − 2.30918z−2 ,
Dˆ 2 (z) = 0.15514z−2 ,
 L. Wang, Y. Ji and L. Wan et al. / Journal of the Franklin Institute 356 (2019) 10102–10122 10117

Fig. 1. The estimation errors δ versus t with σ 2 = 2.02 .

Fig. 2. The estimation errors δ versus t with σ 2 = 5.02 .

Eˆ2 (z) = 1 − 0.16557z−1 ,

Fˆ2 (z) = 1 + 0.12279z−1 .

In order to validate these estimated models, we use the rest 100 data from t = 4001 to
t = 4100 to compute the predicted output yˆi (t ). The actual output y(t), the predicted output
yˆi (t ) and their error yˆi (t ) − y(t ) are shown in Figs. 3 and 4 for the RGELS and 3S-RGELS
algorithms. Figs. 3 and 4 show that the predicted outputs are very close to the true outputs
10118 L. Wang, Y. Ji and L. Wan et al. / Journal of the Franklin Institute 356 (2019) 10102–10122

Fig. 3. The predicted output yˆ1 (t ), the actual output y(t) and their error yˆ1 (t ) − y(t ) versus t based on the RGELS
estimates.

Fig. 4. The predicted output yˆ2 (t ), the actual output y(t) and their error yˆ2 (t ) − y(t ) versus t based on the 3S-RGELS
estimates.

of the bilinear system. This demonstrates that the estimated models capture the dynamics of
the system.
From Tables 1–8 and Figs. 1–4, we can draw the following conclusions.
 L. Wang, Y. Ji and L. Wan et al. / Journal of the Franklin Institute 356 (2019) 10102–10122 10119

• The estimation errors δ given by the RGELS algorithm and the 3S-RGELS algorithm
become smaller and smaller as t increases – see Tables 5–8.
• Compared with the RGELS algorithm, the 3S-RGELS algorithm can not only enhance the
parameter estimation accuracy, but also reduce the computational amount effectively – see
Tables 1–4 and Figs. 1 and 2.
• The outputs of the estimated models approach those of the actual system – see Figs. 3
and 4.
• The RGELS and 3S-RGELS algorithms are both effective for identifying the bilinear sys-
tems with colored noises.

7. Conclusions

This paper proposes a recursive generalized extended least squares algorithm for the bi-
linear system with colored noise using the auxiliary model identification idea. Moreover,
a three-stage recursive generalized extended least squares algorithm is presented to reduce
the computational burden using the auxiliary model identification idea and the hierarchical
identification principle, respectively. Compared with the RGELS algorithm, the 3S-RGELS
algorithm requires less computational cost and gives more accurate parameter estimation. The
proposed three-stage recursive generalized extended least squares algorithm for bilinear state
space systems in this paper can combine the iterative methods [50–53], the least squares meth-
ods [54–57] and the statistical tools [58–64] to study the parameter estimation algorithms of
linear and nonlinear systems with colored noise [65–71] and can be applied to other fields
[72–83] such as parameter identification [84,85].

Acknowledgements

This work was supported by the National Natural Science Foundation of China (No.
61304093), the Natural Science Foundation of Shandong Province (No. ZR201702170236)
and the Postdoctoral Science Foundation of Shandong Province (No. 2017M622153).

References

[1] C.L. Zhang, J.M. Li, Adaptive iterative learning control for nonlinear pure-feedback systems with initial state
error based on fuzzy approximation, J. Frankl. Inst. 351 (3) (2014) 1483–1500.
[2] H.G. Zhang, L.L. Cui, Y.H. Luo, Near-optimal control for nonzero-sum differential games of continuous-time
nonlinear systems using single-network ADP, IEEE Trans. Cybern. 43 (1) (2013) 206–216.
[3] H.G. Zhang, Y.H. Luo, D.R. Liu, Neural-network-based near-optimal control for a class of discrete-time affine
nonlinear systems with control constraints, IEEE Trans. Neural Netw. 20 (9) (2009) 1490–1503.
[4] L. Xu, The parameter estimation algorithms based on the dynamical response measurement data, Adv. Mech.
Eng. 9 (11) (2017) 1–12, doi:10.1177/1687814017730003.
[5] L. Xu, F. Ding, Parameter estimation for control systems based on impulse responses, Int. J. Control Automat.
Syst. 15 (6) (2017) 2471–2479.
[6] X. Zhang, F. Ding, L. Xu, E.F. Yang, Highly computationally efficient state filter based on the delta operator,
Int. J. Adapt. Control Signal Process. 33 (6) (2019) 875–889.
[7] L. Xu, A proportional differential control method for a time-delay system using the taylor expansion approxi-
mation, Appl. Math. Comput. 236 (2014) 391–399.
[8] L. Xu, Application of the Newton iteration algorithm to the parameter estimation for dynamical systems, J.
Comput. Appl. Math. 288 (2015) 33–43.
[9] L. Xu, L. Chen, W.L. Xiong, Parameter estimation and controller design for dynamic systems from the step
responses based on the Newton iteration, Nonlinear Dyn. 79 (3) (2015) 2155–2163.
10120 L. Wang, Y. Ji and L. Wan et al. / Journal of the Franklin Institute 356 (2019) 10102–10122

[10] L. Xu, The damping iterative parameter identification method for dynamical systems based on the sine signal
measurement, Signal Process. 120 (2016) 660–667.
[11] L. Xu, F. Ding, Y. Gu, A. Alsaedi, T. Hayat, A multi-innovation state and parameter estimation algorithm for
a state space system with d-step state-delay, Signal Process. 140 (2017) 97–103.
[12] Q.Y. Liu, F. Ding, Auxiliary model-based recursive generalized least squares algorithm for multivariate out-
put-error autoregressive systems using the data filtering, Circuits Syst. Signal Process. 38 (2) (2019) 590–610.
[13] Q.Y. Liu, F. Ding, L. Xu, E.F. Yang, Partially coupled gradient estimation algorithm for multivariable equa-
tion-error autoregressive moving average systems using the data filtering technique, IET Control Theory Appl.
13 (5) (2019) 642–650.
[14] K.E. Liu, Z.J. Ji, G.M. Xie, L. Wang, Consensus for heterogeneous multi-agent systems under fixed and switch-
ing topologies, J. Frankl. Inst. 352 (9) (2015) 3670–3683.
[15] H.G. Zhang, C.B. Qin, B. Jiang, Y.H. Luo, Online adaptive policy learning algorithm for H∞ state feedback
control of unknown affine nonlinear discrete-time systems, IEEE Trans. Cybern. 44 (12) (2014) 2706–2718.
[16] C.J. Li, G.P. Liu, Consensus for heterogeneous networked multi-agent systems with switching topology and
time-varying delays, J. Frankl. Inst. 355 (10) (2018) 4198–4217.
[17] S. Gibson, A. Wills, B. Ninness, Maximum-likelihood parameter estimation of bilinear systems, IEEE Trans.
Automat. Control 50 (10) (2005) 1581–1596.
[18] X. Zhang, F. Ding, F.E. Alsaadi, T. Hayat, Recursive parameter identification of the dynamical models for
bilinear state space systems, Nonlinear Dyn. 89 (4) (2017) 2415–2429.
[19] Y.J. Wang, F. Ding, A filtering based multi-innovation gradient estimation algorithm and performance analysis
for nonlinear dynamical systems, IMA J. Appl. Math. 82 (6) (2017) 1171–1191.
[20] Y.J. Wang, F. Ding, Iterative estimation for a non-linear IIR filter with moving average noise by means of the
data filtering technique, IMA J. Math. Control Inf. 34 (3) (2017) 745–764.
[21] Y.W. Mao, F. Ding, A novel parameter separation based identification algorithm for hammerstein systems, Appl.
Math. Lett. 60 (2016) 21–27.
[22] Y.W. Mao, F. Ding, E.F. Yang, Adaptive filtering based multi-innovation gradient algorithm for nonlinear systems,
Int. J. Adapt. Control Signal Process. 31 (10) (2017) 1388–1400.
[23] Y.J. Wang, F. Ding, Filtering-based iterative identification for multivariable systems, IET Control Theory Appl.
10 (8) (2016) 894–902.
[24] F. Ding, H.B. Chen, L. Xu, J.Y. Dai, Q.S. Li, T. Hayat, A hierarchical least squares identification algorithm for
hammerstein nonlinear systems using the key term separation, J. Frankl. Inst. 355 (8) (2018) 3737–3752.
[25] L. Xu, F. Ding, Recursive least squares and multi-innovation stochastic gradient parameter estimation methods
for signal modeling, Circuits Syst. Signal Process. 36 (4) (2017) 1735–1753.
[26] L. Xu, F. Ding, Iterative parameter estimation for signal models based on measured data, Circuits Syst. Signal
Process. 37 (7) (2018) 3046–3069.
[27] L. Xu, W.L. Xiong, A. Alsaedi, T. Hayat, Hierarchical parameter estimation for the frequency response based
on the dynamical window data, Int. J. Control Automat. Syst. 16 (4) (2018) 1756–1764.
[28] L. Xu, F. Ding, Q.M. Zhu, Hierarchical newton and least squares iterative estimation algorithm for dynamic
systems by transfer functions based on the impulse responses, Int. J. Syst. Sci. 50 (1) (2019) 141–151.
[29] L. Xu, F. Ding, Parameter estimation algorithms for dynamical response signals based on the multi-innovation
theory and the hierarchical principle, IET Signal Process. 11 (2) (2017) 228–237.
[30] F. Ding, X.H. Wang, Hierarchical stochastic gradient algorithm and its performance analysis for a class of
bilinear-in-parameter systems, Circuits Syst. Signal Process. 36 (4) (2017) 1393–1405.
[31] Y.J. Wang, F. Ding, The auxiliary model based hierarchical gradient algorithms and convergence analysis using
the filtering technique, Signal Process. 128 (2016) 212–221.
[32] M.H. Li, X.M. Liu, Auxiliary model based least squares iterative algorithms for parameter estimation of bilinear
systems using interval-varying measurements, IEEE Access 6 (2018) 21518–21529.
[33] Y.J. Wang, F. Ding, Novel data filtering based parameter identification for multiple-input multiple-output systems
using the auxiliary model, Automatica 71 (2016) 308–313.
[34] M.H. Li, X.M. Liu, et al., Filtering-based maximum likelihood gradient iterative estimation algorithm for bilinear
systems with autoregressive moving average noise, Circuits Syst. Signal Process. 37 (11) (2018) 5023–5048.
[35] M.H. Li, X.M. Liu, The least squares based iterative algorithms for parameter estimation of a bilinear system
with autoregressive noise using the data filtering technique, Signal Process. 147 (2018) 23–34.
[36] X. Zhang, F. Ding, L. Xu, E.F. Yang, State filtering-based least squares parameter estimation for bilinear systems
using the hierarchical identification principle, IET Control Theory Appl. 12 (12) (2018) 1704–1713.
 L. Wang, Y. Ji and L. Wan et al. / Journal of the Franklin Institute 356 (2019) 10102–10122 10121

[37] X. Zhang, F. Ding, L. Xu, A. Alsaedi, T. Hayat, A hierarchical approach for joint parameter and state estimation
of a bilinear system with autoregressive noise, Mathematics 7 (4) (2019) Article Number: 356, doi:10.3390/
math7040356.
[38] X. Zhang, L. Xu, F. Ding, T. Hayat, Combined state and parameter estimation for a bilinear state space system
with moving average noise, J. Frankl. Inst. 355 (6) (2018) 3079–3103.
[39] P.C. Gong, W.Q. Wang, X.R. Wan, Adaptive weight matrix design and parameter estimation via sparse modeling
for MIMO radar, Signal Process. 139 (2017) 1–11.
[40] P.C. Gong, W.Q. Wang, F.C. Li, H. Cheung, Sparsity-aware transmit beamspace design for FDA-MIMO radar,
Signal Process. 144 (2018) 99–103.
[41] J. Pan, H. Ma, X. Jiang, et al., Adaptive gradient-based iterative algorithm for multivariate controlled autoregres-
sive moving average systems using the data filtering technique, Complexity 2018 (2018) Article ID 9598307,
doi:10.1155/ 2018/ 9598307.
[42] J. Pan, W. Li, H.P. Zhang, Control algorithms of magnetic suspension systems based on the improved double
exponential reaching law of sliding mode control, Int. J. Control Automat. Syst. 16 (6) (2018) 2878–2887.
[43] X.K. Wan, Y. Li, C. Xia, M.H. Wu, J. Liang, N. Wang, A T-wave alternans assessment method based on least
squares curve fitting technique, Measurement 86 (2016) 93–100.
[44] N. Zhao, Joint optimization of cooperative spectrum sensing and resource allocation in multi-channel cognitive
radio sensor networks, Circuits Syst. Signal Process. 35 (7) (2016) 2563–2583.
[45] J. Ding, J.Z. Chen, J.X. Lin, G.P. Jiang, Particle filtering-based recursive identification for controlled auto-
regressive systems with quantised output, IET Control Theory Appl. 13 (14) (2019) 2181–2187.
[46] N. Zhao, M.H. Wu, J.J. Chen, Android-based mobile educational platform for speech signal processing, Int. J.
Electr. Eng. Educ. 54 (1) (2017) 3–16.
[47] N. Zhao, Y. Liang, Y. Pei, Dynamic contract incentive mechanism for cooperative wireless networks, IEEE
Trans. Veh. Technol. 67 (11) (2018) 10970–10982.
[48] X.L. Zhao, F. Liu, B. Fu, F. Na, Reliability analysis of hybrid multi-carrier energy systems based on entropy-
-based markov model, Proc. Inst. Mech. Eng. Part O-J. Risk Reliab. 230 (6) (2016) 561–569.
[49] X.L. Zhao, Z.Y. Lin, B. Fu, L. He, F. Na, Research on automatic generation control with wind power participation
based on predictive optimal 2-degree-of-freedom PID strategy for multi-area interconnected power system,
Energies 11 (12) (2018) Article Number: 3325, doi:10.3390/en11123325.
[50] F. Ding, X.P. Liu, G. Liu, Gradient based and least-squares based iterative identification methods for OE and
OEMA systems, Digit. Signal Process. 20 (3) (2010) 664–677.
[51] F. Ding, X.G. Liu, J. Chu, Gradient-based and least-squares-based iterative algorithms for hammerstein systems
using the hierarchical identification principle, IET Control Theory Appl. 7 (2) (2013) 176–184.
[52] F. Ding, Two-stage least squares based iterative estimation algorithm for CARARMA system modeling, Appl.
Math. Model. 37 (7) (2013) 4798–4808.
[53] F. Ding, J. Pan, A. Alsaedi, T. Hayat, Gradient-based iterative parameter estimation algorithms for dynamical
systems from observation data, Mathematics 7 (5) (2019) Article Number: 428, doi:10.3390/math7050428.
[54] F. Ding, Decomposition based fast least squares algorithm for output error systems, Signal Process. 93 (5)
(2013) 1235–1242.
[55] F. Ding, Coupled-least-squares identification for multivariable systems, IET Control Theory Appl. 7 (1) (2013)
68–79.
[56] Y.J. Liu, D.Q. Wang, et al., Least squares based iterative algorithms for identifying box-jenkins models with
finite measurement data, Digit. Signal Process. 20 (5) (2010) 1458–1467.
[57] F. Ding, Y.J. Liu, B. Bao, Gradient based and least squares based iterative estimation algorithms for multi-input
multi-output systems, Proc. Inst. Mech. Eng., Part I: J. Syst. Control Eng. 226 (1) (2012) 43–55.
[58] C.C. Yin, Y.Z. Wen, Y.X. Zhao, On the optimal dividend problem for a spectrally positive levy process, Astin
Bull. 44 (3) (2014) 635–651.
[59] C.C. Yin, Y.Z. Wen, Optimal dividend problem with a terminal value for spectrally positive levy processes,
Insur. Math. Econ. 53 (3) (2013) 769–773.
[60] C.C. Yin, J.S. Zhao, Nonexponential asymptotics for the solutions of renewal equations, with applications, J.
Appl. Probab. 43 (3) (2006) 815–824.
[61] C.C. Yin, C.W. Wang, The perturbed compound poisson risk process with investment and debit interest,
Methodol. Comput. Appl. Probab. 12 (3) (2010) 391–413.
[62] C.C. Yin, Y.Z. Wen, Exit problems for jump processes with applications to dividend problems, J. Comput. Appl.
Math. 245 (2013) 30–52.
10122 L. Wang, Y. Ji and L. Wan et al. / Journal of the Franklin Institute 356 (2019) 10102–10122

[63] Y.Z. Wen, C.C. Yin, Solution of Hamilton-Jacobi-Bellman equation in optimal reinsurance strategy under dy-
namic VaR constraint, Journal of Function Spaces (2019) Article Number: 6750892, doi:10.1155/ 2019/ 6750892.
[64] L. Wang, H. Liu, L.V. Dai, Y.W. Liu, Novel method for identifying fault location of mixed lines, Energies 11
(6) (2018) Article Number: 1529, doi:10.3390/en11061529.
[65] J. Ding, J.Z. Chen, J.X. Lin, L.J. Wan, Particle filtering based parameter estimation for systems with output-error
type model structures, J. Frankl. Inst. 356 (10) (2019) 5521–5540.
[66] F.Y. Ma, Y.K. Yin, M. Li, Start-up process modelling of sediment microbial fuel cells based on data driven,
Math. Probl. Eng. (2019) ArticleNumber: 7403732, doi:10.1155/ 2019/ 7403732.
[67] Z.Y. Sun, D. Zhang, Q. Meng, C.C. Chen, Feedback stabilization of time-delay nonlinear systems with contin-
uous time-varying output function, Int. J. Syst. Sci. 50 (2) (2019) 244–255.
[68] F. Yang, P. Zhang, X.X. Li, The truncation method for the cauchy problem of the inhomogeneous Helmholtz
equation, Appl. Anal. 98 (5) (2019) 991–1004.
[69] H. Ma, J. Pan, L. Lv, et al., Recursive algorithms for multivariable output-error-like ARMA systems, Mathe-
matics 7 (6) (2019) Article Number: 558, doi:10.3390/math7060558.
[70] L.J. Wan, F. Ding, Decomposition- and gradient-based iterative identification algorithms for multivariable sys-
tems using the multi-innovation theory, Circuits Syst. Signal Process. 38 (7) (2019) 2971–2991.
[71] J. Pan, X. Jiang, X.K. Wan, W. Ding, A filtering based multi-innovation extended stochastic gradient algorithm
for multivariable control systems, Int. J. Control Automat. Syst. 15 (3) (2017) 1189–1197.
[72] M.H. Wu, X. Li, C. Liu, M. Liu, N. Zhao, J. Wang, X.K. Wan, Z.H. Rao, L. Zhu, Robust global motion
estimation for video security based on improved k-means clustering, J. Ambient Intell. Humaniz. Comput. 10
(2) (2019) 439–448.
[73] X.K. Wan, H. Wu, F. Qiao, F. Li, Y. Li, Y. Yan, J. Wei, Electrocardiogram baseline wander suppression based
on the combination of morphological and wavelet transformation based filtering, Comput. Math. Methods Med.
(2019) Article ID 7196156, doi:10.1155/ 2019/ 7196156.
[74] L. Feng, Q.X. Li, Y.F. Li, Imaging with 3-d aperture synthesis radiometers, IEEE Trans. Geosci. Remote Sens.
57 (4) (2019) 2395–2406.
[75] B. Fu, C.X. Ouyang, C.S. Li, J.W. Wang, E. Gul, An improved mixed integer linear programming approach
based on symmetry diminishing for unit commitment of hybrid power system, Energies 12 (5) (2019) Article
Number: 833, doi:10.3390/en12050833.
[76] W.X. Shi, N. Liu, Y.M. Zhou, X.A. Cao, Effects of postannealing on the characteristics and reliability of
polyfluorene organic light-emitting diodes, IEEE Trans. Electron Devices 66 (2) (2019) 1057–1062.
[77] T.Z. Wu, X. Shi, L. Liao, C.J. Zhou, H. Zhou, Y.H. Su, A capacity configuration control strategy to alleviate
power fluctuation of hybrid energy storage system based on improved particle swarm optimization, Energies 12
(4) (2019) Article Number: 642, doi:10.3390/en12040642.
[78] Y. Cao, Y.K. Sun, G. Xie, T. Wen, Fault diagnosis of train plug door based on a hybrid criterion for IMFs
selection and fractional wavelet package energy entropy, IEEE Trans. Veh. Technol. 68 (8) (2019) 7544–7551.
[79] N. Liu, S. Mei, D. Sun, W. Shi, J. Feng, Y.M. Zhou, F. Mei, J. Xu, Y. Jiang, X.A. Cao, Effects of charge
transport materials on blue fluorescent organic light-emitting diodes with a host-dopant system, Micromachines
10 (5) (2019) Article Number: 344, doi:10.3390/mi10050344.
[80] Y. Cao, Z. Wang, F. Liu, P. Li, G. Xie, Bio-inspired speed curve optimization and sliding mode tracking control
for subway trains, IEEE Trans. Veh. Technol. 68 (7) (2019) 6331–6342.
[81] Y. Cao, H. Lu, T. Wen, A safety computer system based on multi-sensor data processing, Sensors 19 (4) (2019),
doi:10.3390/s19040818.
[82] Y. Cao, Y. Zhang, T. Wen, P. Li, Research on dynamic nonlinear input prediction of fault diagnosis based on
fractional differential operator equation in high-speed train control system, Chaos 29 (1) (2019) Article Number:
013130, doi:10.1063/1.5085397.
[83] Y. Cao, P. Li, Y. Zhang, Parallel processing algorithm for railway signal fault diagnosis data based on cloud
computing, Fut. Gen. Comput. Syst. 88 (2018) 279–283.
[84] Y. Gu, J. Liu, X. Li, Y. Chou, Y. Ji, State space model identification of multirate processes with time-delay
using the expectation maximization, J. Frankl. Inst. 356 (3) (2019) 1623–1639.
[85] Y. Gu, Y. Chou, J. Liu, Y. Ji, Moving horizon estimation for multirate systems with time-varying time-delays,
J. Frankl. Inst. 356 (4) (2019) 2325–2345.