Particle filtering based parameter estimation for systems with output-error type model structures
Accepted Manuscript
Particle filtering based parameter estimation for systems with output-error type model structures
Jie Ding, Jiazhong Chen, Jinxing Lin, Lijuan Wan
PII: S0016-0032(19)30309-6
DOI: https://doi.org/10.1016/j.jfranklin.2019.04.027
Reference: FI 3913
To appear in: Journal of the Franklin Institute
Received date: 16 September 2018 Revised date: 1 March 2019 Accepted date: 29 April 2019
Please cite this article as: Jie Ding, Jiazhong Chen, Jinxing Lin, Lijuan Wan, Particle filtering based parameter estimation for systems with output-error type model structures, Journal of the Franklin Institute (2019), doi:https://doi.org/10.1016/j.jfranklin.2019.04.027
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
Particle filtering based parameter estimation for systems with output-error type model structures
✩Jie Dinga,∗, Jiazhong Chena, Jinxing Lina, Lijuan Wanb
aSchool of Automation and Artificial Intelligence, Nanjing University of Posts and Telecommunications, Nanjing, 210023, China
bCollege of Automation and Electronic Engineering, Qingdao University of Science and Technology, Qingdao 266061, PR China
Abstract
The output-error model structure is often used in practice and its identification is important for analysis of output-error type systems. This paper considers the parameter identification of linear and nonlinear output-error models. A particle filter which approximates the posterior probability density function with a weighted set of discrete random sampling points is utilized to estimate the unmeasurable true process outputs. To improve the convergence rate of the proposed algorithm, the scalar innovations are grouped into an innovation vector, thus more past information can be utilized. The convergence analysis shows that the parameter estimates can converge to their true values. Finally, both linear and nonlinear results are verified by numerical simulation and engineering.
Keywords: output-error model, particle filter, parameter estimation, convergence analysis
1. Introduction
Mathematical models and model parameter estimation are the basis of system analysis and con- troller design [1–4]. These involve scalar systems [5–7] and multivariable systems [8–12]. The output- error (OE) type systems have received considerable attention due to their wide application in engineer- ing. Some discussion includes but not limited to parameter estimation, stability analysis and control
5
of OE type systems [13, 14]. The parameter estimation for OE type systems is of prime importance [15, 16]. The key issue is how to estimate the true output of OE type systems with noise-contaminated data, so that the observation errors can reduce the influence on the accuracy of parameter estimation.
Some methods involve the estimation of noise-free outputs and the bias corrected parameter estima- tion. One of the techniques is the auxiliary model idea [17, 18], which replaces the unmeasurable
10
output of the OE type system by the output of a properly designed or selected auxiliary model [19].
For example, for multivariate output-error autoregressive systems, an auxiliary model-based recursive generalized least squares algorithm was proposed by using the data filtering [20].
Recently, Jin et al. studied an auxiliary model based identification algorithm for multivariable OE- like systems with missing outputs [21]. For bias corrected parameter estimation, a bias correction or
15
bias compensation is devoted to eliminate estimation bias caused by colored noise [22]. For nonlinear systems with OE type model structures, Piga and T´oth presented a bias-corrected estimator [23].
A bias correction method was discussed by Zheng for the identification of linear dynamic errors-in- variables models [24]. Inspired by the auxiliary model idea [25], a particle filter is introduced to handle the noisy output data, which can improve the accuracy of the priori noise-free output estimates. The
20
particle filter has been successfully applied to nonlinear systems. An EM-based particle filter algorithm was developed for nonlinear parameter varying systems in [26]. Chen et al. proposed a novel particle filter based gradient iterative algorithm for the identification of dual-rate nonlinear systems [27] and
✩This work was supported by the National Natural Science Foundation of China (Nos. 61203028 and 61473158) and the Natural Science Foundation of NJUPT (No. NY217063).
∗Corresponding author
Email address: [email protected] (Jie Ding)
ACCEPTED MANUSCRIPT
investigated a stochastic gradient based particle filter algorithm for an ARX model with nonlinear communication output [28].
25
In this paper, the benefits of the particle filter are taken into account to estimate the process output. The main contributions can be phrased as follows:
• Motivated by the auxiliary model idea, a particle filtering technique is introduced to estimate the true output of generalized OE type systems. It approximates the posterior probability density function with a weighted set of discrete random sampling points, yielding much better
30
identification performance in nonlinear OE models than linear ones. The convergence properties of the proposed algorithm is analyzed in detail.
• The multi-innovation identification theory is employed to improve the convergence rate of the proposed algorithm. By grouping a number of innovations into an innovation vector, a particle filter based multi-innovation identification algorithm is proposed.
35
Briefly, the rest of this paper is organized as follows. In Section 2, an auxiliary model based recursive least squares algorithm for OE type systems is introduced. In Section 3, a particle filter is employed to estimate the noise-free output, further the parameter estimation algorithm is presented.
In Section 4, the multi-innovation identification theory is introduced to improve the convergence rate of the proposed algorithm. In Section 5, the convergence result of the proposed algorithm is analyzed.
40
Both linear and nonlinear examples are given in Section 6. Finally, some conclusions and future topics are given in Section 7.
2. Problem formulation
Consider the generalized output-error type system with the following input-output relation:
x(t) = f (x(t− 1), · · · , x(t − nx), u(t),· · · , u(t − nu)), (1)
y(t) = x(t) + v(t), (2)
where u(t) and y(t) are the input and output sequence of the system, respectively, x(t) is the noise-free process output and is unknown, v(t) is a white Gaussian noise sequence with zero mean and variance σ2, the function f (·) is a linear one with respect to x(t) and u(t), e.g.,
x(t) = B(z−1)
A(z−1)u(t) = [1− A(z−1)]x(t) + B(z−1)u(t) = f (·),
with A(z−1) and B(z−1) are polynomials in unit backward shift operator z−1:
45
A(z) = 1 + a1z−1+ a2z−2+· · · + anaz−na, B(z) = b1z−1+ b2z−2+ b3z−3+· · · + bnbz−nb,
where z−1 represents the unit backward shift operator, i.e., z−1y(t) = y(t− 1).
More generally, f (·) can be treated as a nonlinear function with a real-valued multivariate poly- nomial, which is parameterized as follows:
f (x(t−1), · · · , x(t − nx), u(t),· · · , u(t − nu))
=
nθ
X
i=1
θiψi(x(t− 1), · · · , x(t − nx), u(t),· · · , u(t − nu)),
where θi (i = 1, 2,· · · , nθ) are the unknown parameters to be estimated, while ψi (i = 1, 2,· · · , nθ) are a-priori chosen functions (e.g., the canonical polynomial basis) in the variables{x(t − 1), · · · , x(t − nx), u(t),· · · , u(t − nu)}.
Define the parameter vector θ and the information vector ϕ(t) as
50
θ := [θ1, θ2,· · · , θnθ]T∈ Rnθ,
ACCEPTED MANUSCRIPT
- f(·) -+j? -
v(t)
u(t) x(t) y(t)
ppppp ppppp ppppp ppppp ppppp ppppp ppp p p p p p p p p p p p p p p p
1 N
PN j=1
ˆ xj(t)
ϕTa(t)ˆθ(t) ppp ppp -ppp ppp -ppp ppp ppp ppp
ˆ xa(t)
ˆ x(t)
Original System
Auxiliary model Particle filter
Figure 1: The output-error type system with the auxiliary model or particle filter
ϕ(t) := [ψ1(x(t− 1), · · · , x(t − nx), u(t),· · · , u(t − nu)),· · · , ψnθ(x(t− 1), · · · , x(t − nx), u(t),· · · , u(t − nu))]T∈ Rnθ, Then Eqs. (1)–(2) can be rewritten as
x(t) = ϕT(t)θ,
y(t) = x(t) + v(t). (3)
In Eq. (3), the information vector ϕ(t) is unknown since it contains variables x(t− i). Thus the standard least squares cannot be applied directly. Based on the auxiliary model idea, the unknown variables x(t− i) in ϕ(t) can be replaced with the outputs ˆxa(t− i) of an auxiliary model as shown in Figure 1, where{ϕa(t), θa(t)} are the information vector and the parameter vector of the auxiliary
55
model at time t that are defined as follows:
ϕa(t) := [ψ1(ˆxa(t− 1), · · · , ˆxa(t− nx), u(t),· · · , u(t − nu)),· · · ,
ψnθ(ˆxa(t− 1), · · · , ˆxa(t− nx), u(t),· · · , u(t − nu))]T∈ Rnθ. (4) Then the cost function can be formed as
J(θ) = Xt i=1
[y(i)− ϕTa(i)θ]2.
Minimizing J(θ) and defining the estimate of θ as ˆθ yield the following auxiliary model based recursive least squares (AM-RLS) algorithm [29]:
ˆθ(t) = ˆθ(t− 1) + P (t)ϕa(t)[y(t)− ϕTa(t)ˆθ(t− 1)], (5)
P−1(t) = P−1(t− 1) + ϕa(t)ϕTa(t), (6)
ˆ
xa(t) = ϕTa(t)ˆθ(t). (7)
To initialize this algorithm, let ˆθ(0) = 1n/p0, where 1n is an n-dimensional column vector whose elements are all 1 and p0 is a large positive number (p0= 106), and P (0) = p0I, where I denotes an
60
identity matrix of appropriate dimensions.
Remark 1. The basic idea of the auxiliary model is to construct an auxiliary model, generally, an OE-like model is the top choice. Other auxiliary models can be designed in different cases, e.g., a finite impulse response model, which was proposed in [30] to predict the noise-free fast-rate output.
An over-parameterization identification model was presented for input nonlinear OE autoregressive
65
systems in [31]. A key term separation based identification model was developed for input nonlinear OE systems in [32]. A well-designed auxiliary model can improve the accuracy of the parameter estimation.
ACCEPTED MANUSCRIPT
3. The particle filtering based RLS algorithm
The output of the auxiliary model ˆxa(t) is actually the estimate of the unmeasurable noise-free
70
process outputs x(t), inspired by this idea, a particle filter is proposed to get the similar estimate by approximating the posterior probability density function (pdf) with a weighted set of discrete random sampling points as shown in Figure 2.
3.1. Particle filtering technique
To define the noise-free process outputs, consider the evolution of the process output sequence
75
{x(t − i)} of a target given by Eq. (1). The noise-free process outputs x(t − i) can be recursively computed from noisy outputs Eq. (2). Actually, we seek the filtered estimates of x(t− i) based on the set of all available inputs and noisy outputs {u(j), y(j)} (j = 1, · · · , t − i) and estimated parameter ˆθ(t− i − 1) at time t − i.
From a Bayes perspective [33], the noise-free process outputs estimation problem is to recursively
80
calculate some degree of belief in the process output x(t− i) at time t − i with {u(1), · · · , u(t − i), y(1),· · · , y(t − i), ˆθ(t − i − 1)}. Thus, it is required to construct the posterior pdf p(x(t − i)|y(t − i),· · · , y(1), x(t − i − 1), · · · , x(1), u(t − i), · · · , u(1), ˆθ(t − i − 1)). It is assumed that the posterior pdf at previous time t− i − 1 which is also known as the prior, is available. Then the posterior pdf at time t− i may be obtained recursively in two steps: prediction and update.
85
The prediction step involves using the noise-free process output model Eq. (1) to obtain the prior pdf at time t− i via the Chapman-Kolmogorov equation
p(x(t− i)|y(t − i − 1), · · · , y(1), x(t − i − 1), · · · , x(1), u(t − i), · · · , u(1), ˆθ(t − i − 1))
= Z
p(x(t− i), x(t − i − 1)|y(t − i − 1), · · · , y(1), x(t − i − 1), · · · , x(1), u(t − i), · · · , u(1), θ(tˆ − i − 1))dx(t − i − 1)
= Z
p(x(t− i)|x(t − i − 1), y(t − i − 1), · · · , y(1), x(t − i − 1), · · · , x(1), u(t − i), · · · , u(1), θ(tˆ − i − 1))p(x(t − i − 1)|y(t − i − 1), · · · , y(1), x(t − i − 1), · · · , x(1), u(t − i), · · · , u(1), ˆθ(t− i − 1))dx(t − i − 1)
= Z
p(x(t− i)|x(t − i − 1))p(x(t − i − 1)|y(t − i − 1), · · · , y(1), x(t − i − 1), · · · , x(1), u(t− i), · · · , u(1), ˆθ(t − i − 1))dx(t − i − 1), (8) with the fact that p(x(t−i)|x(t−i−1), y(t−i−1), · · · , y(1), x(t−i−1), · · · , x(1), u(t−i), · · · , u(1), ˆθ(t−
i− 1)) = p(x(t − i)|x(t − i − 1)), where Eq. (1) is a Markov process of order one. The probabilistic model of the process output evolution p(x(t− i)|x(t − i − 1)) is defined by Eq. (1).
At time step t− i, a noisy output y(t − i) becomes available, and this may be used to update the prior (update step) via Bayes’ rule
p(x(t− i)|y(t − i), · · · , y(1), x(t − i − 1), · · · , x(1), u(t − i), · · · , u(1), ˆθ(t − i − 1))
= p(y(t− i)|x(t − i), y(t − i − 1), · · · , y(1), x(t − i − 1), · · · , x(1), u(t − i), · · · , u(1), ˆθ(t − i − 1))
×p(x(t− i)|y(t − i − 1), · · · , y(1), x(t − i − 1), · · · , x(1), u(t − i), · · · , u(1), ˆθ(t − i − 1)) p(y(t− i)|y(t − i − 1), · · · , y(1), x(t − i − 1), · · · , x(1), u(t − i), · · · , u(1), ˆθ(t − i − 1)),
= p(y(t− i)|x(t − i)) (9)
×p(x(t− i)|y(t − i − 1), · · · , y(1), x(t − i − 1), · · · , x(1), u(t − i), · · · , u(1), ˆθ(t − i − 1)) p(y(t− i)|y(t − i − 1), · · · , y(1), x(t − i − 1), · · · , x(1), u(t − i), · · · , u(1), ˆθ(t − i − 1)), where the normalizing constant
p(y(t− i)|y(t − i − 1), · · · , y(1), x(t − i − 1), · · · , x(1), u(t − i), · · · , u(1), ˆθ(t − i − 1))
ACCEPTED MANUSCRIPT
= Z
p(y(t− i)|x(t − i))p(x(t − i)|y(t − i − 1), · · · , y(1), x(t − i − 1), · · · , x(1), u(t− i), · · · , u(1), ˆθ(t − i − 1))dx(t − i),
depends on the likelihood function p(y(t−i)|x(t−i)) which can be expressed by the noisy output model in Eq. (2). In the update step Eq. (9), the noisy output y(t− i) is used to modify the prior density to obtain the required posterior density of the current process output. The recurrence relations (8) and (9) form the basis for the optimal Bayes solution. In general, due to the existence of integration, the optimal Bayes solution can be unavailable. Based on the Monte Carlo method [34], the posterior pdf can be expressed as
p(x(t− i)|y(t − i), · · · , y(1),x(t − i − 1), · · · , x(1), u(t − i), · · · , u(1), ˆθ(t − i − 1))
≈ XN j=1
ωjδ(x(t− i) − ˆxj(t− i)), (10)
where ωj is the normalized weight associated with the jth particle, δ(·) represents the Dirac delta function and ˆxj(t− i) denotes the jth particle sampled from the posterior pdf p(x(t − i)|y(t − i),· · · , y(1), x(t − i − 1), · · · , x(1), u(t − i), · · · , u(1), ˆθ(t − i − 1)) at time t − i. Once the particles and their weights have been determined, the noise-free process outputs can be computed as
ˆ
x(t− i) = XN j=1
ωjxˆj(t− i),
where ωj is the normalized weight, i.e., ωj = ωj/(PN
j=1
ωj).
However, it is difficult to sample particles from the posterior pdf directly. An alternative way is to sample from a reference distribution which is referred to as the importance density q(·). Then the regular choice of the importance density q(·) is the prior density:
q(x(t− i)|y(t − i), · · · , y(1), x(t − i − 1), · · · , x(1), u(t − i), · · · , u(1), ˆθ(t − i − 1))
= p(x(t− i)|x(t − i − 1), · · · , x(1), u(t − i), · · · , u(1), ˆθ(t − i − 1)). (11) Thus the new particles can be obtained from Eqs. (3) and (11). Next step is to update the weight associated with each new particle. Based on [26], one can get the updated weight ωj(t) as follows:
ωj(t) = p(y(t)|ˆxj(t))ωj(t− 1). (12)
According to Eq. (3), the density function p(y(t)|ˆxj(t)) can be computed by p(y(t)|ˆxj(t)) = 1
√2πσexp
−(y(t)− ˆxj(t))2 2σ2
. (13)
Then, Eq. (12) can be written as ωj(t) = 1
√2πσexp
−(y(t)− ˆxj(t))2 2σ2
ωj(t− 1). (14)
Because the variance of the particle weight will increase with time iteration, degradation may occur in the particle filter algorithm, the weights of most particles will be small enough to be neglected except for some important particles. In order to reduce the impact of degradation, three choices can be taken: increasing the number of particles, re-sampling technique, and choosing a reasonable importance density. Here, the re-sampling technique is utilized, i.e., copying and eliminating the particles ˆxj(t− i) according to the normalized weight ωj(t), then resetting the weight ωj(t) = ωj(t) = N1 yields
ˆ
x(t− i) = 1 N
XN j=1
ˆ
xj(t− i). (15)
Then the noise-free process outputs can be estimated by the following particle filter:
90
ACCEPTED MANUSCRIPT
Step 1. Initialization: Suppose the initial process output x(0) is Gaussian, then sample N initial particles {ˆxj(0)}Nj=1 from the prior density p(x(0)|ˆθ(t − 1)) and set the weight of each particle to N1.
Step 2. Importance sampling:
• Collect the input-output data {u(1), y(1), u(2), y(2), · · · , u(t), y(t)}, compute {ˆxj(t −
95
i)}Nj=1 from Eq. (11).
• Recalculate weights for each particle by Eq. (14).
Step 3. Re-sampling:
• According to the normalized weight, copy and eliminate ˆxj(t− i).
• Reset the weight ωj(t) = ωj(t) = N1.
100
Step 4. Output: Compute{ˆx(t − i)}Nj=1 by Eq. (15).
Step 5. Iteration: Let i = i− 1, if i ≥ 0 go to step 2.; otherwise, end the procedure.
3.2. Identification algorithm
In general, the estimate ˆθ(t) is considered to be much closer to the true value θ than previous estimates{ˆθ(t − 1), · · · , ˆθ(0)}. Thus, in order to estimate the noise-free process outputs at time t − i
105
more accurately, parameter estimate ˆθ(t−i−1) is utilized, and corresponding {x(1), x(2), · · · , x(t−i)}
can be updated with Eqs. (11)–(15).
With process output estimate ˆx(t− i), a novel particle filtering based recursive least squares (PF- RLS) algorithm can be presented as follows:
θ(t) = ˆˆ θ(t− 1) + L(t)[y(t) − ˆϕT(t)ˆθ(t− 1)], (16) L(t) = P (t) ˆϕ(t) = P (t− 1) ˆϕ(t)[1 + ˆϕT(t)P (t− 1) ˆϕ(t)]−1, (17)
P (t) = [In− L(t) ˆϕT(t)]P (t− 1), (18)
ˆ
ϕ(t) = [ψ1(ˆx(t− 1), · · · , ˆx(t − nx), u(t),· · · , u(t − nu)),· · · ,
ψnθ(ˆx(t− 1), · · · , ˆx(t − nx), u(t),· · · , u(t − nu))]T∈ Rnθ, (19) ˆ
x(t− i) = 1 N
XN j=1
ˆ
xj(t− i), i = 1, 2, 3, · · · , nx. (20)
The initialization of the PF-RLS algorithm is similar to that of the AM-RLS algorithm. The pseudo
110
code of the PF-RLS algorithm is described by Algorithm 1.
ACCEPTED MANUSCRIPT
Algorithm 1 The particle filtering based RLS algorithm Data: u, y, nθ, length, N
Result: ˆθ
1: Initialization: ˆθ(0) = [1/p0,· · · , 1/p0]T ∈ Rnθ, P (0) = p0Inθ, p0 = 106, ˆx(0) = randn(N, 1), {ωj(0)}Nj=1 = N1
2: for t = 1 : length do
3: function Particle Filter(u, y, ˆθ(t− 1), t) 4: for k = 1 : t do
5: for j = 1 : N do 6: Draw ˆxj(k)∼ (11) 7: Calculate ωj(k) by (14)
8: end for
9: Normalize ωj(k)
10: [ˆxj(k), N1]=Re-sample [ˆxj(k), ωj(k)]
11: Compute ˆx(k) from (15)
12: end for
13: end function
14: Form ˆϕ(t) according to (19) 15: Calculate L(t) from (17) 16: Compute P (t) by (18) 17: Estimate ˆθ(t) by (16) 18: end for
4. The particle filtering based multi-innovation recursive least squares algorithm
The PF-RLS algorithm in Eqs. (16)–(20) can achieve good estimation accuracy but suffer long convergence rate, which needs more iterations. In order to improve its performance while keeping the same accuracy, the multi-innovation identification theory [35–38] is employed as shown in Figure 2,
115
where e(t) is stacked before parameter estimation.
- f(·) -+i? -
v(t)
u(t) x(t) y(t)
-
RLS Algorithm
Particle Filter -+i?
?
e(t),· · · , e(t − p + 1) ˆ
x(t)
-
e(t)
>
Figure 2: The output-error type system with PF-MILS
Let p represent the innovation length. Notice that the innovation e(t) = y(t)− ˆϕT(t)ˆθ(t− 1) in Eq. (16) is a scalar quantity. In order to make good use of limited information, one can expand this scalar innovation e(t)∈ R1 to an innovation vector
E(p, t) := [y(t)− ˆϕT(t)ˆθ(t− 1), y(t − 1) − ˆϕT(t− 1)ˆθ(t − 1), · · · , y(t− p + 1) − ˆϕT(t− p + 1)ˆθ(t − 1)]T
= = Y (p, t)− ˆφT(p, t)ˆθ(t− 1) ∈ Rp.
Define the information matrix ˆφ(p, t) and the stacked output vector Y (p, t) as
120
φ(p, t) := [ ˆˆ ϕ(t), ˆϕ(t− 1), · · · , ˆϕ(t− p + 1)] ∈ Rnθ×p,
ACCEPTED MANUSCRIPT
Y (p, t) := [y(t), y(t− 1), · · · , y(t − p + 1)]T∈ Rp, the innovation vector E(p, t) can be expressed as
E(p, t) = Y (p, t)− ˆφT(p, t)ˆθ(t− 1).
An improved particle filtering based multi-innovation least squares (PF-MILS) algorithm is derived as θ(t) = ˆˆ θ(t− 1) + L(t)[Y (p, t) − ˆφT(p, t)ˆθ(t− 1)], (21) L(t) = P (t) ˆφ(p, t) = P (t− 1)ˆφ(p, t)[Ip+ ˆφT(p, t)P (t− 1)ˆφ(p, t)]−1, (22)
P (t) = [Inθ− L(t)ˆφT(p, t)]P (t− 1), (23)
φ(p, t) = [ ˆˆ ϕ(t), ˆϕ(t− 1), · · · , ˆϕ(t− p + 1)], (24) ˆ
ϕ(t) = [ψ1(ˆx(t− 1), · · · , ˆx(t − nx), u(t),· · · , u(t − nu)),· · · ,
ψnθ(ˆx(t− 1), · · · , ˆx(t − nx), u(t),· · · , u(t − nu))]T∈ Rnθ, (25) ˆ
x(t− i) = 1 N
XN j=1
ˆ
xj(t− i), i = 1, 2, 3, · · · , nx. (26)
When the innovation length p = 1, E(1, t) = e(t), ˆφ(1, t) = ˆϕ(t), Y (1, t) = y(t), the PF-MILS algorithm in Eqs. (21)–(26) degrades to the PF-RLS algorithm in Eqs. (16)–(20). The pseudo code of the PF-MILS algorithm is described by Algorithm 2.
125
Algorithm 2 The particle filtering based MILS algorithm Data: u, y, nθ, length, N , p
Result: ˆθ
1: Initialization: ˆθ(0) = [1/p0,· · · , 1/p0]T ∈ Rnθ, P (0) = p0Inθ, p0 = 106, ˆx(0) = randn(N, 1), {ωj(0)}Nj=1 = N1
2: for t = 1 : length do
3: function Particle Filter(u, y, ˆθ(t− 1), t) 4: for k = 1 : t do
5: for j = 1 : N do 6: Draw ˆxj(k)∼ (11) 7: Calculate ωj(k) by (14)
8: end for
9: Normalize ωj(k)
10: [ˆxj(k), N1]=Re-sample [ˆxj(k), ωj(k)]
11: Compute ˆx(k) from (15)
12: end for
13: end function
14: Form ˆϕ(t) according to (25) 15: Form ˆφ(p, t) by (24)
16: Calculate L(t) from (22) 17: Compute P (t) by (23) 18: Estimate ˆθ(t) by (21) 19: end for
5. Convergence analysis
In this section, we present and prove the convergence property of the proposed algorithm in Eqs.
(16)–(20) according to the method in [39, 40].
ACCEPTED MANUSCRIPT
Theorem 1. For the output-error type system in Eq. (3) and the PF-RLS algorithm in Eqs. (16)–
(20),{v(t)} is a white noise sequence with zero mean and variance σ2, that is
130
(A1) E[v(t)] = 0, (A2) E[v2(t)] = σ2,
and that there exist positive constants α, β and integer t0such that for t≥ t0, the following generalized persistent excitation condition holds:
(A3) αIn6 1 t
Xt j=1
ˆ
ϕ(j) ˆϕT(j)6 βIn, a.s.
Then, the parameter estimates ˆθ(t) given by the algorithm Eqs. (16)–(20) converge to teir true value θ.
Proof of Theorem 1. Define the parameter estimation error vector
135
θ(t) = ˆ˜ θ(t)− θ. (27)
Substituting Eqs. (3) and (5) into Eq. (27) gives
θ(t) = ˆ˜ θ(t− 1) + P (t) ˆϕ(t)[y(t)− ˆϕT(t)ˆθ(t− 1)] − θ
= ˜θ(t− 1) + P (t) ˆϕ(t)[−˜y(t) + ∆(t) + v(t)], (28)
where
˜
y(t) : = ˆϕT(t)˜θ(t− 1) ∈ R,
∆(t) : = [ϕ(t)− ˆϕ(t)]Tθ∈ R.
Using Eqs. (6) and (28), we have
˜θT(t)P−1(t)˜θ(t) ={˜θ(t − 1) + P (t) ˆϕ(t)[−˜y(t) + ∆(t) + v(t)]}TP−1(t){˜θ(t − 1) +P (t) ˆϕ(t)[−˜y(t) + ∆(t) + v(t)]}
= ˜θT(t− 1)P−1(t− 1)˜θ(t − 1) − [1 − ˆϕT(t)P (t) ˆϕ(t)]˜y2(t)
+2[1− ˆϕT(t)P (t) ˆϕ(t)]˜y(t)[∆(t) + v(t)] + ˆϕT(t)P (t) ˆϕ(t)[v2(t) + ∆2(t) + 2∆(t)v(t)]
= ˜θT(t− 1)P−1(t− 1)˜θ(t − 1) + 2[1 − ˆϕT(t)P (t) ˆϕ(t)]˜y(t)[∆(t) + v(t)]
+ ˆϕT(t)P (t) ˆϕ(t)[v2(t) + ∆2(t) + 2∆(t)v(t)]. (29) Note that 1− ˆϕT(t)P (t) ˆϕ(t) = [1 + ˆϕT(t)P (t) ˆϕ(t)]−1≥ 0, and {v(t)} is a white Gaussian noise with zero mean independent of the input signal{u(t)}. Referring to the methods in [39, 40], we can know
140
that the particle filter algorithm is convergent, i.e., when t → ∞, ˆx(t) → x(t). Then we can obtain that when t→ ∞, ˆϕ(t) → ϕ(t), i.e., ∆(t) is bounded with ∆2(t) ≤ ε < ∞. Define a non-negative definite function
V (t) := E[˜θT(t)P−1(t)˜θ(t)].
Since ˜θT(t− 1)P−1(t− 1)˜θ(t − 1), ˜y(t), ˆϕT(t)P (t) ˆϕ(t) and ∆(t) are uncorrelated with v(t), taking the expectation of both sides of Eq. (29) and using (A1) and (A2) get
145
V (t)≤ V (t − 1) + 0 + E{ ˆϕT(t)P (t) ˆϕ(t)[v2(t) + ∆2(t)]}
≤ V (t − 1) + E[ ˆϕT(t)P (t) ˆϕ(t)](σ2+ ε)
≤ V (0) + E[
Xt j=1
ϕˆT(j)P (j) ˆϕ(j)](σ2+ ε). (30)
From Eq. (6), we can obtain
P−1(t− 1) = P−1(t)− ˆϕ(t) ˆϕT(t)
ACCEPTED MANUSCRIPT
= P−1(t)[In− P (t) ˆϕ(t) ˆϕT(t)].
Taking the determinant of both sides of above equation gives:
|P−1(t− 1)| = |P−1(t)||In− P (t) ˆϕ(t) ˆϕT(t)|
=|P−1(t)|[1 − ˆϕT(t)P (t) ˆϕ(t)].
or ˆ
ϕT(t)P (t) ˆϕ(t) = |P−1(t)| − |P−1(t− 1)|
|P−1(t)| .
Replacing t with j and summing for j from j = 1 to t yield Xt
j=1
ϕˆT(j)P (j) ˆϕ(j) = Xt j=1
|P−1(j)| − |P−1(j− 1)|
|P−1(j)|
≤ ln |P−1(t)| − ln |1
p0In| = ln |P−1(t)| + n ln p0. (31) Using Eq. (6) and (A3), we get
150
P−1(t) = Xt j=1
ˆ
ϕ(j) ˆϕT(j) + P−1(0)≤ βtIn+ In/p0= (βt + 1/p0)In, P−1(t)≥ αtIn+ In/p0 = (αt + 1/p0)I,
and
(αt + 1/p0)n ≤ |P−1(t)| ≤ (βt + 1/p0)n. Substituting above condition into Eq. (31) gives
Xt j=1
ϕˆT(j)P (j) ˆϕ(j) = ln|P−1(t)| + n ln p0 ≤ n ln(βt + 1/p0) + n ln p0.
According to the definition of V (t), we have
V (t) = E[˜θT(t)P−1(t)˜θ(t)]≥ (αt + 1/p0)E[k˜θ(t)k2].
Since V (0) = E[˜θT(0)P−1(0)˜θ(0)] = n/p20, from Eq. (30), we have
(αt + 1/p0)E[k˜θ(t)k2]≤ V (t) ≤ V (0) + E[
Xt j=1
ˆ
ϕT(j)P (j) ˆϕ(j)](σ2+ ε)
≤ n/p20+ [n ln(βt + 1/p0) + n ln p0](σ2+ ε).
Taking the limit of both sides of the above equation gives
155
tlim→∞E[k˜θ(t)k2]≤ limt
→∞
n/p20+ [nln(βt + 1/p0) + nlnp0](σ2+ ε)
(αt + 1/p0) = 0.
This means that the parameter estimation errorkˆθ(t) − θk converges to zero with the increase of t.
ACCEPTED MANUSCRIPT
6. Case studies
Case study 1. Consider the following linear two-order output-error system:
y(t) = X4 i=1
θiψi(x(t− 1), x(t − 2), u(t), u(t − 1), u(t − 2)) + v(t), ϕ(t) = [−x(t − 1), −x(t − 2), u(t − 1), u(t − 2)]T,
θ = [θ1, θ2, θ3, θ4]T= [0.412, 0.309, 0.6804, 0.6303]T.
Here, {u(t)} is taken as a persistent excitation signal sequence with zero mean and unit variance, and{v(t)} as a white noise sequence with zero mean and variance σ2 = 0.502 with the corresponding
160
noise-to-signal ratios (NSR): δns= 67.86%. In simulation, we choose 50 particles.
The AM-RLS algorithm in Eqs. (4)–(7) and the PF-RLS algorithm in Eqs. (16)–(20) are applied to estimate the parameters θi of this system. Further, the PF-MILS algorithm in Eqs. (21)–(26) with the innovation length p = 2 and 3 is applied to estimate the parameters of this system. The parameter estimates and their errors are shown in Tables 1 to 3 and the estimation errors δ :=kˆθ(t) − θk/kθk
165
versus t are shown in Figures 3 to 4.
Table 1: The AM-RLS estimates and errors (σ2= 0.502, δns= 67.86%)
t θ1 θ2 θ3 θ4 δ (%)
100 0.23728 0.34316 0.67433 0.41086 26.64187
200 0.37396 0.32041 0.69209 0.53047 10.18674
500 0.40668 0.35113 0.69840 0.55136 8.61788
1000 0.40755 0.34643 0.69120 0.60374 4.46389
2000 0.40395 0.32467 0.69038 0.61313 2.50211
3000 0.40711 0.31837 0.68724 0.61693 1.73159
True values 0.41200 0.30900 0.68040 0.63030
Table 2: The PF-RLS estimates and errors (σ2= 0.502, δns= 67.86%)
t θ1 θ2 θ3 θ4 δ (%)
100 0.21370 0.30012 0.70090 0.41462 27.69758
200 0.39000 0.29028 0.70566 0.55396 8.05393
500 0.41494 0.33589 0.70509 0.56456 7.09380
1000 0.41183 0.33827 0.69393 0.61096 3.54478
2000 0.40604 0.31989 0.69183 0.61716 2.01602
3000 0.40861 0.31474 0.68824 0.61995 1.37588
True values 0.41200 0.30900 0.68040 0.63030
0 500 1000 1500 2000 2500 3000
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
t
δ
PF−RLS AM−RLS
Figure 3: The AM-RLS and PF-RLS estimation errors with σ2= 0.502
ACCEPTED MANUSCRIPT
Table 3: The PF-MILS estimates and errors (σ2= 0.502, δns= 67.86%)
p t θ1 θ2 θ3 θ4 δ (%)
1 100 0.21370 0.30012 0.70090 0.41462 27.69758
200 0.39000 0.29028 0.70566 0.55396 8.05393
500 0.41494 0.33589 0.70509 0.56456 7.09380
800 0.41672 0.33558 0.69159 0.60154 3.86475
2 100 0.24936 0.30815 0.68445 0.45138 22.79525
200 0.40375 0.29457 0.69600 0.56782 6.26967
500 0.41746 0.34262 0.70448 0.56640 7.19303
800 0.41184 0.33439 0.68374 0.62011 2.59784
3 100 0.30909 0.26896 0.72887 0.55025 13.64422
200 0.43326 0.27162 0.73050 0.62831 6.22629
500 0.42389 0.32863 0.71296 0.56153 7.49176
800 0.40247 0.32123 0.68433 0.63107 1.50952
True values 0.41200 0.30900 0.68040 0.63030
0 100 200 300 400 500 600 700 800
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
PF−RLS ( MILS,p=1)
PF−MILS,p=2 PF−MILS,p=3
t
δ
Figure 4: The PF-MILS estimation errors with different innovation length (σ2= 0.502)
c
g
(((((((((((((
((((((((((((( c
DD DD
DD DD DD
DD DD
DD
"!
# h
p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p pα Ball- Beam
Level Arm
Servo Motor
Figure 5: A ball and beam system
Case study 2. Consider an experiment setup of a ball and beam system in Figure 6, where the input u(t) is angle of the beam α and the output y(t) is position of the ball. The ball and beam system consists of a horizontal beam which can pivot about one end, a servo motor whose shaft is attached to the other end of the beam and a ball which can freely roll on top of the beam [41]. Due to the
170
connection of the beam with a level arm, the beam can move up and down according to the change of the position angle of the motor. Thus the system can be considered as the following linear third-order OE system:
y(t) = X6 i=1
θiψi(x(t− 1), x(t − 2), x(t − 3), u(t), u(t − 1), u(t − 2), u(t − 3)) + v(t),
ACCEPTED MANUSCRIPT
ϕ(t) = [−x(t − 1), −x(t − 2), −x(t − 3), u(t − 1), u(t − 2), u(t − 3)]T,
where {u(t), y(t)} are the input and output data downloaded from DaISy [42]. The first 900 data is employed to estimate the parameters θi of this system with σ2 = 0.502 by applying PF-RLS and
175
PF-MILS algorithms, while the last 100 data can be used to test and verify. Substituting the estimate parameters and the last 100 input data into above OE system, the estimate of the last 100 output data can be obtained. Then the estimation errors are shown in Figures 6–7.
910 920 930 940 950 960 970 980 990 1000
−2
−1.5
−1
−0.5 0 0.5 1 1.5 2
t
δ
True value Prediction Estimation errors
Figure 6: The PF-RLS estimation errors (σ2= 0.502)
910 920 930 940 950 960 970 980 990 1000
−2
−1.5
−1
−0.5 0 0.5 1 1.5 2
t
δ
True value Prediction Estimation errors
Figure 7: The PF-MILS estimation errors with innovation length p = 3 (σ2= 0.502)
Case study 3. Consider the following nonlinear output-error system:
y(t) = X5 i=1
θiψi(x(t− 1), x(t − 2), u(t), u(t − 1)) + v(t), ϕ(t) = [x(t− 1), x(t − 2), x2(t− 1), u(t − 1), u2(t− 1)]T,
θ = [θ1, θ2, θ3, θ4, θ5]T= [−0.5, 0.3, 0.2, 0.15, −1.4]T,
where the initialization of simulation is the same as that of Case study 1. Applying the AM-RLS
180
algorithm, PF-RLS algorithm, and PF-MILS algorithm with the innovation length p = 2 and 3 to estimate the parameters θi of this system. The parameter estimates and their errors are shown in Tables 4 to 6 and the estimation errors δ :=kˆθ(t) − θk/kθk versus t are shown in Figures 8 to 9.
ACCEPTED MANUSCRIPT
Table 4: The AM-RLS estimates and errors (σ2= 0.502, δns= 37.17%)
t θ1 θ2 θ3 θ4 θ5 δ (%)
100 -0.22690 0.35098 0.27919 0.21101 -1.14691 25.30094
200 -0.49324 0.19968 0.18242 0.25654 -1.17595 17.45332
500 -0.51464 0.29607 0.23221 0.18240 -1.25324 10.04853
1000 -0.56426 0.27038 0.21346 0.15035 -1.30870 7.56562
2000 -0.56155 0.28366 0.20970 0.13949 -1.34659 5.48694
3000 -0.55420 0.28657 0.20407 0.13662 -1.36205 4.48587
True values -0.50000 0.30000 0.20000 0.15000 -1.40000
Table 5: The PF-RLS estimates and errors (σ2= 0.502, δns= 37.17%)
t θ1 θ2 θ3 θ4 θ5 δ (%)
100 -0.37668 0.34197 0.19781 0.16892 -1.21374 14.83952
200 -0.47321 0.31026 0.20143 0.18435 -1.30245 6.98325
500 -0.52697 0.28898 0.19662 0.15502 -1.31754 5.70338
1000 -0.50050 0.30593 0.20022 0.14267 -1.35316 3.10884
2000 -0.50764 0.30282 0.19782 0.13835 -1.36573 2.41781
3000 -0.49841 0.30274 0.19669 0.13676 -1.36790 2.27899
True values -0.50000 0.30000 0.20000 0.15000 -1.40000
0 500 1000 1500 2000 2500 3000
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
t
δ
PF−RLS AM−RLS
Figure 8: The AM-RLS and PF-RLS estimation errors with σ2= 0.502
Table 6: The PF-MILS estimates and errors (σ2= 0.502, δns= 37.17%)
p t θ1 θ2 θ3 θ4 θ5 δ (%)
1 100 -0.37668 0.34197 0.19781 0.16892 -1.21374 14.83952
200 -0.47321 0.31026 0.20143 0.18435 -1.30245 6.98325
500 -0.52697 0.28898 0.19662 0.15502 -1.31754 5.70338
800 -0.50186 0.30598 0.19837 0.15646 -1.35503 2.98583
2 100 -0.48167 0.35824 0.20636 0.13164 -1.27974 8.86523
200 -0.47763 0.32396 0.19821 0.18846 -1.34415 4.90178
500 -0.50411 0.30290 0.19392 0.18663 -1.42355 2.87956
800 -0.48745 0.31068 0.19316 0.18217 -1.40813 2.45120
3 100 -0.48952 0.31011 0.19160 0.11431 -1.40197 2.57014
200 -0.45495 0.30202 0.19441 0.16137 -1.40670 3.07860
500 -0.47557 0.29433 0.19168 0.16144 -1.44023 3.21900
800 -0.47751 0.29145 0.19571 0.15538 -1.41487 1.89389
True values -0.50000 0.30000 0.20000 0.15000 -1.40000
ACCEPTED MANUSCRIPT
0 100 200 300 400 500 600 700 800
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
PF−RLS ( MILS,p=1) PF−MILS,p=2
PF−MILS,p=3
t
δ
Figure 9: The PF-MILS estimation errors with different innovation length (σ2= 0.502)
Table 7: The iterations of PF-MILS with different innovation length at the same estimation error
Case studies Case study 1 Case study 3
δ (%) 2.6 2.46
p 1 2 3 1 2 3
Iterations 1319 794 744 1805 760 632
Remark 2: At the same estimation error, the PF-RLS needs much more iterations than PF-MILS.
Therefore, the PF-MILS can improve convergence rate at the same convergence accuracy. More
185
detailed results are as follows.
• For output-error type systems, the parameter estimates of the AM-RLS and PF-RLS algorithms can converge to their true values very fast as shown in Tables 1–2 and 4–5. The PF-RLS algorithm performs slightly better than AM-RLS does in the linear output-error system as shown in Figure 3. The main advantages of the particle filter are in the application of strong nonlinear and
190
non-Gaussian systems, so the PF-RLS performs obviously better than AM-RLS in the nonlinear output-error system as shown in Figure 8.
• With different innovation length p, the PF-MILS algorithm with larger p can lead to better convergence rate and accuracy as shown in Figures 4 and 9. It is easy to find that the PF-MILS needs fewer iterations than the PF-RLS to achieve the same convergence accuracy as shown in
195
Table 7.
• The proposed algorithms have very small estimation errors as shown in Figures 6 and 7. Thus they are widely provided with practical value in engineering.
7. Conclusions
By employing the particle filtering technique and the multi-innovation identification theory, two
200
particle filter based algorithms are proposed for output-error type systems. The limitation exists in multi-input and multi-output (MIMO) systems, since the variables of the MIMO system will be of high dimension, yielding to heavy computation burden. But it is still a good choice for single-input and single-output systems to get accurate models. From simulation results, we can find that the convergence of the particle filter is slightly better than the auxiliary model method in linear systems,
205
showing the obvious advantages in nonlinear system. Furthermore, the convergence analysis shows that the parameter estimation error consistently converges to zero. The simulation results verify that the proposed algorithms have effective convergence rate and high accuracy. The propoased algorithm can coombine the hierarchical principle [43, 44] and the iterative methods [45, 46] to study the parameter identificatin problems of multivariable systems [47, 48], bilinear systems [49–51] and nonlinear systems
210
[52], and can be applied to otther fields [53–57].
ACCEPTED MANUSCRIPT
References
[1] L. Xu, A proportional differential control method for a time-delay system using the Taylor expansion approximation, Appl.
Math. Comput. 236 (2014) 391-399.
[2] L. Xu, Application of the Newton iteration algorithm to the parameter estimation for dynamical systems, J. Comput. Appl.
215
Math. 288 (2015) 33-43.
[3] 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.
[4] L. Xu, The damping iterative parameter identification method for dynamical systems based on the sine signal measurement, Signal Process. 120 (2016) 660-667.
220
[5] F. Ding, State filtering and parameter estimation for state space systems with scarce measurements, Signal Process. 104 (2014) 369-380.
[6] 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. Franklin Inst. 356 (3) (2019) 1623-1639.
[7] Y. Gu, Y. Chou, J. Liu, Y. Ji, Moving horizon estimation for multirate systems with time-varying time-delays, J. Franklin
225
Inst. 356 (4) (2019) 2325-2345.
[8] J.L. Ding, Recursive and iterative least squares parameter estimation algorithms for multiple-input-output-error systems with autoregressive noise, Circuits Syst. Signal Process. 37 (5) (2018) 1884-1906.
[9] J.L. Ding, The hierarchical iterative identification algorithm for multi-input-output-error systems with autoregressive noise.
Complexity, 2017, 1-11. Article ID 5292894. https://doi.org/10.1155/2017/5292894
230
[10] F. Ding, Hierarchical parameter estimation algorithms for multivariable systems using measurement information, Inf. Sci. 277 (2014) 396-405.
[11] Z.N. Zhang, F. Ding, X.G. Liu, Hierarchical gradient based iterative parameter estimation algorithm for multivariable output error moving average systems, Comput. Math. Appl. 61 (3) (2011) 672-682.
[12] H.Q. Han, L. Xie, et al., Hierarchical least squares based iterative identification for multivariable systems with moving average
235
noises, Mathematical and Computer Modelling 51 (9-10) (2010) 1213-1220.
[13] J. Ma, S. Xu, Y. Li, Y. Chu, Z. Zhang, Neural networks-based adaptive output feedback control for a class of uncertain nonlinear systems with input delay and disturbances, J. Franklin Inst. 355 (13) (2018) 5503-5519.
[14] Y.J. Wang, F. Ding, M. Wu, Recursive parameter estimation algorithm for multivariate output-error systems, J. Franklin Inst.
355 (12) (2018) 5163-5181.
240
[15] 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.
[16] M.H. Li, X.M. Liu, The least squares based iterative algorithms for parameter estimation of a bilinear system with autore- gressive noise using the data filtering technique, Signal Process. 147 (2018) 23-34.
[17] D.Q. Wang, L.W. Li, Y. Ji, Y.R. Yan, Model recovery for Hammerstein systems using the auxiliary model based orthogonal
245
matching pursuit method, Appl. Math. Model. 54 (2018) 537-550.
[18] D.Q. Wang, Y.R. Yan, Y.J. Liu, et al., Model recovery for Hammerstein systems using the hierarchical orthogonal matching pursuit method, J. Comput. Appl. Math. 345 (2019) 135-145.
[19] F. Ding, F.F. Wang, L. Xu, et al., Parameter estimation for pseudo-linear systems using the auxiliary model and the decom- position technique, IET Control Theory Appl. 11 (3) (2017) 390-400.
250
[20] Q.Y. Liu, F. Ding, Auxiliary model-based recursive generalized least squares algorithm for multivariate output-error autore- gressive systems using the data filtering, Circuits Syst. Signal Process. 38 (2) (2019) 590-610.
[21] Q. Jin, Z. Wang, X. Liu, Auxiliary model-based interval-varying multi-innovation least squares identification for multivariable OE-like systems with scarce measurements, J. Process Control 35 (2015) 154-168.
[22] C. Mosquera, R. Lopez, F. Perez, On the bias of the modified output error algorithm, IEEE Trans. Autom. Control 43 (9)
255
(1998) 1261-1262.
[23] D. Piga, R. T´oth, A bias-corrected estimator for nonlinear systems with output-error type model structures, Automatica 50 (9) (2014) 2373-2380.
[24] W.X. Zheng, A bias correction method for identification of linear dynamic errors-in-variables models, IEEE Trans. Autom.
Control 47 (7) (2002) 1142-1147.
260
[25] F. Ding, G.J. Liu, X.P. Liu, Parameter estimation with scarce measurements, Automatica 47 (8) (2011) 1645-1655.
[26] J. Deng, B. Huang, Identification of nonlinear parameter varying systems with missing output data, AIChE J. 58 (11) (2012) 3454-3467.
[27] J. Chen, J. Li, Y. Liu, Gradient iterative algorithm for dual-rate nonlinear systems based on a novel particle filter, J. Franklin Inst. 354 (2017) 4425-4437.
265
ACCEPTED MANUSCRIPT
[28] J. Chen, Y. Liu, et al., Gradient-based particle filter algorithm for an ARX model with nonlinear communication output, IEEE Trans. Sys. Man Cybern.: Syst. (2018), doi:10.1109/TSMC.2018.2810277.
[29] 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.
[30] F. Ding, T. Chen, Identification of dual-rate systems based on finite impulse response models, Int. J. Adapt. Control Signal
270
Process. 18 (7) (2004) 589-598.
[31] V. Stojanovic, N. Nedic, Robust identification of OE model with constrained output using optimal input design, J. Franklin Inst. 353 (2) (2016) 576-593.
[32] H.B. Chen, Y.S. Xiao, et al., Hierarchical gradient parameter estimation algorithm for Hammerstein nonlinear systems using the key term separation principle, Appl. Math. Comput. 247 (2014) 1202-1210.
275
[33] D. Simon, Optimal State Estimation: Kalman, H∞, and Nonlinear Approaches, Hoboken, NJ, USA: Wiley, 2006.
[34] A. Doucet, S. Godsill, C. Andrieu, On sequential Monte Carlo Sampling methods for Bayesian filtering, Stat. Comput. 10 (3) (2000) 197-208.
[35] 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.
280
[36] 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.
[37] 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.
[38] L. Xu, The parameter estimation algorithms based on the dynamical response measurement data, Adv. Mech. Eng. 9 (11)
285
(2017) 1-12. doi: 10.1177/1687814017730003
[39] F. Ding, System Identification — Performances Analysis for Identification Methods, Science Press, Beijing, 2014.
[40] F. Ding, Y. Gu, Performance analysis of the auxiliary model-based least-squares identification algorithm for one-step state- delay systems, Int. J. Comput. Math. 89 (15) (2012) 2019-2028.
[41] Z. Liang, Energy-based balance control approach to the ball and beam system, Int. J. Control 82(6) (2009) 981-992.
290
[42] DaISy, STADIUS’s Identification Database, http://homes.esat.kuleuven.be/ smc/daisy/daisydata.html.
[43] 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.
[44] 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 Autom. Syst. 16 (4) (2018) 1756-1764.
295
[45] L. Xu, F. Ding, Parameter estimation for control systems based on impulse responses, Int. J. Control Autom. Syst. 15 (6) (2017) 2471-2479.
[46] L. Xu, F. Ding, Iterative parameter estimation for signal models based on measured data, Circuits Syst. Signal Process. 37 (7) (2018) 3046-3069.
[47] Z.W. Ge, F. Ding, L. Xu, A. Alsaedi, T. Hayat, Gradient-based iterative identification method for multivariate equation-error
300
autoregressive moving average systems using the decomposition technique, J. Franklin Inst. 356 (3) (2019) 1658-1676.
[48] J. Pan, H. Ma, X. Jiang, et al., Adaptive gradient-based iterative algorithm for multivariate controlled autoregressive moving av- erage systems using the data filtering technique, Complexity 2018, Article ID 9598307. https://doi.org/10.1155/2018/9598307 [49] X. Zhang, L. Xu, et al., Combined state and parameter estimation for a bilinear state space system with moving average noise,
J. Franklin Inst. 355 (6) (2018) 3079-3103.
305
[50] X. Zhang, F. Ding, et al., Recursive parameter identification of the dynamical models for bilinear state space systems, Nonlinear Dyn. 89 (4) (2017) 2415-2429.
[51] 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.
[52] Y.J. Wang, F. Ding, A filtering based multi-innovation gradient estimation algorithm and performance analysis for nonlinear
310
dynamical systems, IMA J. Appl. Math. 82 (6) (2017) 1171-1191.
[53] 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 Autom. Syst. 16 (6) (2018) 2878-2887.
[54] X.Y. Li, H.X. Li, B.Y. Wu, Piecewise reproducing kernel method for linear impulsive delay differential equations with piecewise constant arguments, Appl. Math. Comput. 349 (2019) 304-313.
315
[55] Y. Cao, H. Lu, T. Wen, A safety computer system based on multi-sensor data processing, Sensors 19 (4) (2019).
https://doi.org/10.3390/s19040818
[56] Y. Cao, Y. Zhang, T. Wen, P. Li, Research on dynamic nonlinear input prediction of fault diagnosis based on frac- tional differential operator equation in high-speed train control system, Chaos 29 (1) (2019), Article Number: 013130.
https://doi.org/10.1063/1.5085397
320
[57] T. Wang, L. Liu, J. Zhang, E. Schaeffer, Y. Wang, A M-EKF fault detection strategy of insulation system for marine current turbine, Mech. Syst. Signal Process. 115 (2019) 269-280. January 2019