Ocean Engineering 294 (2024) 116853

Contents lists available at ScienceDirect

Ocean Engineering
journal homepage: www.elsevier.com/locate/oceaneng

Research paper

Online ship motion identification modeling and its application to

course-keeping control
Yao Meng a, b, Xianku Zhang a, b, *, Xiufeng Zhang a, b, Daocheng Ma a, Yating Duan a, b
a
Navigation College, Dalian Maritime University, Dalian, 116026, China
b
Key Lab. of Marine Simulation and Control, Dalian Maritime University, Dalian, 116026, China

A R T I C L E I N F O A B S T R A C T

Keywords: Online identification modeling can lay the model foundation for ship motion control. Given that the parameters
Online parameter identification of ship motion of the ship motion model cannot be accurately determined, online parameter identification methods can quickly
model estimate model parameters using limited data. In order to improve adaptability and stability, this study in­
Course-keeping control of the ship model
troduces fuzzy fading memory into the improved unscented Kalman filter to conduct online parameter estimation
Prior model for ship motion control
of the ship motion model. The unscented Kalman filter with fuzzy forgetting factor is taken as a comparison
Improved unscented kalman filter
Closed-loop gain shaping algorithm algorithm to analyze the superiority of the proposed scheme. Taking a ship model similar to an USV as the
research plant, the identified model is converted into a control model. This study conducts research on the
organic combination of online parameter identification and ship motion control, using the closed-loop gain
shaping algorithm to perform course-keeping control. The research results indicate that the proposed online
identification method can ensure high prediction accuracy and provide an accurate control model. In the case of
unknown control models, the proposed scheme can provide a prior model foundation for motion control using
limited data. This scheme combines identification modeling with motion control, which can promote the
development and research of USV navigation controllers.

1. Introduction addressing the uncertainty of the environment and solving the chal­
lenges of mechanism modeling. Compared with traditional mechanism
1.1. Background and motivation modeling methods, it has the advantages of adaptability and univer­
sality. It can also provide prior knowledge and the model foundation for
As the intelligent industry and information technology developed, ship motion control. If the ship motion model is unknown, it is impos­
intelligent unmanned system technology was applied in many research sible to conduct relevant control research. Therefore, it is necessary to
fields. After unmanned aerial vehicles (UAVs) and unmanned vehicles develop a research system that organically combines system identifica­
gradually came into people’s lives with the support of advanced tech­ tion modeling and motion control.
nology and modern equipment, the research on USVs has developed
rapidly. USVs can achieve autonomous navigation and operation in 1.2. Related work
complex marine environments through technologies such as control,
sensing, and communication systems. Given the requirements of navi­ The application of system identification has become a new modeling
gation tasks, the shape of the USV differs from that of traditional ships. trend, which is used to establish the motion model of USV (Pei et al.,
Therefore, it is difficult to establish a motion mathematical model of 2023). Generally speaking, based on ship motion data and mathematical
USV using mechanism modeling. This is also an important challenge for models, parameter identification uses algorithms to determine the pa­
ship motion control. rameters or hydrodynamic coefficients of the model. In order to reflect
System identification techniques have become a well-defined and the dynamic characteristics of ship and prevent overfitting, the selection
well-developed research field. There are many research achievements in scheme of data sample is varied. Some studies selected a single type of
theoretical foundation, practical application, and other aspects. It is test data as the identification data sample (Jiang et al., 2021; Sutulo and
based on data samples, mediated by models, with algorithms at its core, Guedes Soares, 2014; Zhu et al., 2017), while others selected multiple

* Corresponding author. Navigation College, Dalian Maritime University, Dalian, 116026, China.
E-mail address: [email protected] (X. Zhang).

https://doi.org/10.1016/j.oceaneng.2024.116853
Received 30 October 2023; Received in revised form 8 December 2023; Accepted 20 January 2024
Available online 27 January 2024
0029-8018/© 2024 Elsevier Ltd. All rights reserved.
Y. Meng et al. Ocean Engineering 294 (2024) 116853

types of test data (Perera et al., 2016; Wang et al., 2020; Zhang et al., and performed online ship motion state prediction.
2021). Each data sample selection scheme has its own advantages. Un­ An accurate ship motion model can reflect the motion state of the
like parameter identification modeling methods, nonparametric identi­ full-scale ship. The study on the maneuverability of full-scale ship is
fication modeling requires a large amount of ship motion data and does crucial to the design of autopilots and other controllers. However,
not require a known motion mathematical model framework. Based on a because the main particulars of ships are unknown, it is a challenge to
large amount of ship motion data, the dynamic characteristics of the ship quickly and accurately obtain a ship motion model. Sun et al. (2019)
and the mapping relationships between various state variables are ob­ established and identified the USV motion model and proposed an
tained. Furthermore, in order to obtain an accurate ship motion pre­ adaptive sliding mode trajectory tracking controller for USV. Abrougui
diction model, multiple types of test data are often taken as training et al. (2021) identified the forward resistance and turning resistance
data. System identification modeling, as a data-driven modeling models and proposed an autopilot scheme for self-steering an USV.
method, requires high-quality data samples. Thus, the specific data se­ Wang et al. (2022) utilized elastic net regression to identify the yaw
lection scheme needs to be determined based on simulation cases. motion model and design a path-following controller based on the
The mathematical model of ship motion plays a connecting role in AMBS-P algorithm. However, the algorithms used in these studies are all
parameter identification modeling. Abkowitz, MMG, Fossen, Norrbin, offline. In the research of ship motion control, closed-loop gain shaping
and Nomoto mathematical models were widely used. The first three algorithm (CGSA), fuzzy controllers, adaptive controllers, sliding mode
models contain many hydrodynamic coefficients, i.e., the three models controllers, and model predictive control algorithms were applied to the
are relatively complex. The Norrbin and Nomoto mathematical models research of course-keeping control, path following, and stabilization of
were simplified forms of the Abkowitz model. They have the advantages the ship or USV (Cao and Zhang, 2023; Gao and Zhang, 2022; Guo et al.,
of few parameters and clear physical significance. Furthermore, the 2023; Han and Zhang, 2022; Li et al., 2022; Zhang et al., 2022). CGSA
Nomoto model is a single-input, single-output model. In the existing (Han and Zhang, 2020), as a concise and efficient ship motion control
research on ship motion control, most of the research adopted the algorithm, has been applied to the motion control of USV in practice.
Nomoto and 3 degrees of freedom (DOF) motion models. The Nomoto
model was widely used in the development of ship autopilots and 1.3. Contributions and structure
research on course-keeping control (Gao and Zhang, 2022). The 3 DOF
ship motion mathematical model was often used in research on ship path A high-precision parameter estimation algorithm of ship motion
following (Guo et al., 2023). The identification results of the 4 DOF ship model can effectively improve the performance of ship motion control
motion mathematical model (including roll motion) and the heave-pitch and is more suitable for practical applications. When the main particu­
motion model can be applied to the research of stabilization (Cao and lars of the ship are unknown or the mechanism modeling is inaccurate,
Zhang, 2023), such as fin stabilizers and rudder resistance roll. the purpose of online identification modeling is to provide an accurate
Nonparametric identification modeling has low requirements for ship ship motion model. Given the adaptability and time consumption of
motion models. It can be applied to the design of adaptive controllers. online identification modeling, this study utilizes the fuzzy fading
Algorithms have a significant impact on the final results in both memory (FFM) method to modify the improved UKF (IUKF) for online
system identification modeling and ship motion control. Both classical identification research of the three-dimensional ship motion state-space
and intelligent identification algorithms were widely used in parameter model. In addition, considering the close connection between identifi­
identification research. Due to some limitations of least squares (LS) and cation modeling and ship motion control, this study uses online identi­
Kalman filtering (KF), Zhang et al. (2015), Song et al. (2021), and Xie fication modeling to provide a prior model foundation for ship motion
et al. (2019) modified their error innovations to improve the accuracy control. There is relatively little research on the organic combination of
and convergence performance of LS. The improved algorithm was ship motion identification modeling and control. The main contributions
applied to the parameter estimation of the ship motion model. The final of this study are summarized as follows:
parameter estimation and prediction results showed higher accuracy
compared to the original LS. Sabet et al. (2014) and Deng et al. (2021) (1) The stability of UKF is improved to ensure non-negativity of the
utilized unscented Kalman filter (UKF) to identify an autonomous un­ state error covariance matrix during the filtering process. Using a
derwater vehicle hydrodynamic model. However, the above schemes are ship model similar to an USV as the research plant, the accuracy
still offline identification modeling research. In order to meet the re­ and generalization of the online parameter identification results
quirements of online identification, H. Wang et al. (2022) conducted a are verified using zigzag test data different from the identification
second-order nonlinear response mathematical model of the Mariner samples.
using nonlinear Gaussian filtering. The final results indicated that the (2) Given that the accumulation of old data has a significant impact,
algorithm can effectively improve parameter identification accuracy this study uses FFM to update the forgetting factor to improve the
and reduce computational complexity. Zheng et al. (2022) utilized UKF estimation accuracy of ship motion model parameters. The UKF
and rolling wavelet denoising to estimate the parameters of a nonlinear with a fuzzy forgetting factor (FFMUKF) is used as a comparison
ship motion model. Classical algorithms consume less time and can algorithm. The comparison results provide a superiority analysis.
perform online parameter estimation in certain cases. As an intelligent (3) The online identification results are combined with the ship
identification method, support vector machine (SVM) (Wang et al., motion control. CGSA is utilized to conduct the course-keeping
2015) and its improved form (Xu et al., 2020) were proved to have ac­ control of the ship model. The algorithm is consistent with the
curate parameter estimation performance. However, as the size of data practical application background.
increases, intelligent algorithms have the problem of high model
training time consumption. Machine learning and deep learning have The layout of the paper is mainly divided into three parts: (1)
shown strong advantages in nonparametric identification research modeling, (2) identifying, and (3) control. The work of this study is
(Zhang et al., 2023). They have strong feature learning and data pro­ organized based on the above layout. The ship motion mathematical
cessing capabilities. SVM, neural networks (NN), Gaussian process model and the preliminary work of the proposed algorithm are intro­
regression, locally weighted learning, and their improved forms were duced in Section 2. Section 3 utilizes the proposed algorithm to perform
applied to ship motion state and attitude prediction. The final prediction online parameter identification of the established model and provides
results demonstrated their generalization (Bai et al., 2019; Ouyang and generalization verification results. Using the online identification results
Zou, 2021; Silva and Maki, 2022; Wang et al., 2020). In order to improve from Section 3, Section 4 conducts research on the course-keeping
the efficiency of prediction algorithms, Xue et al. (2022), Pei et al. control of the ship model under wind and wave interferences. The
(2023), and Chen et al. (2023) improved machine learning algorithms main research results and analysis of this study are presented in Section

2
Y. Meng et al. Ocean Engineering 294 (2024) 116853

5. online identification modeling and motion control of the ship model.

Most ships and USVs are generally underactuated due to the lack of
2. Model and identification algorithm thrusters. Therefore, we only convert the equation of the yaw rate in Eq.
(3). The converted expression is shown in Eq. (5). The first row in Eq. (5)
2.1. Ship motion mathematical model is the ship response mathematical model. The second and third rows are
the transfer functions of the ship response mathematical model. The
The earth-fixed and body-fixed coordinate systems are generally transformation of the model parameters is shown in Eq. (6).
used to describe 6 DOF ship motion. Currently, most control algorithms ⎫
are based on a 3 DOF ship motion mathematical model (Guo et al., 2023; T1 T2 r̈ + (T1 + T2 )ṙ + r = Kδ + KT3 δ̇ ⎪
⎪
⎪
⎪
⎪
Zhang et al., 2022). Therefore, the motion model established in this Gψ′ =
K ⎪
⎬
study does not consider the effects of roll, pitch, and heave motion. The s(Ts + 1) (5)
⎪
⎪
Abkowitz model has strict mathematical derivation, but the physical K(T3 s + 1) ⎪
⎪
⎪
⎪
significance of hydrodynamic coefficients still needs to be studied, and Gψ″ =
s(T1 s + 1)(T2 s + 1)
⎭
the model is particularly complex. Neglecting high-order variables in the
model, the linear model has certain accuracy while also processing a21 b11 − a11 b21 1
⎫
⎪
⎪
physical significance (Abkowitz, 1964). The simplified model is mainly K=
a11 a22 − a12 a21
, T1 T2 =
a11 a22 − a12 a21 ⎪
⎪
⎪
⎬
used to study the coupling effect between sway velocity (v) and yaw rate a11 +a22 b21 (6)
(r). The expression of the 3 DOF simplified motion mathematical model T1 + T2 = − , KT3 = ⎪
a11 a22 − a12 a21 a11 a22 − a12 a21 ⎪
⎪
⎪
is shown in Eq. (1). ⎪
⎭
T = T1 + T2 − T3
⎫
(m − Xu˙ )u̇ = Xu Δu ⎬
mv̇ + mVr + mxG ṙ = Yv v + Yr r + Yv˙v̇ + Yr˙ṙ + Yδ δ (1) where T, T1 , T2 , and T3 are time coefficients, K is gain coefficient, r̈ is the
⎭
Izz v̇ + mxG v̇ + mxG Vr = Nv v + Nr r + Nv˙v̇ + Nr˙ṙ + Nδ δ second derivative of yaw rate, δ̇ is the rate of rudder angle. Based on the
identified parameters in Eq. (3), Eqs. (5) and (6) are obtained. Eq. (5) is
where m is the ship mass, Xu , Xu̇ , Yv … and Nδ are hydrodynamic co­ used as the model of ship motion control. Gψ′ and Gψ″ are the first-order
efficients, V is resultant velocity, u is surge velocity, Δu = u − V, Izz is
and second-order transfer functions of response mathematical model,
the moment of inertia, xG is the longitudinal coordinate of the gravity
respectively.
center, δ is rudder angle, which is the input variable of the ship motion
model. Remark 1. There are more parameters to be identified in Eq. (3) than
The control model of ship motion usually needs to convert the in Eq. (5), which proves the effectiveness of the proposed algorithm
maneuvering dynamics model into the form of a state equation. Convert from the perspective of identification complexity. Furthermore, there is
Eq. (1) into a state equation, as shown in Eq. (2). still a lot of extensive research work on parameter identification in Eq.
⎡ ⎤⎡ ⎤ ⎡ ⎤⎡ ⎤ (3). In addition, the research and development of autopilots are mostly
m − Xu˙ 0 0 u̇ Xu 0 0 Δu based on a response mathematical model. Therefore, this study selects
⎣ 0 m − Yv˙ mxG − Yr˙ ⎦ ⎣ v̇ ⎦ ⎣
= 0 Yv Yr − mV ⎦ ⎣ v ⎦
models that can demonstrate the efficiency of the proposed identifica­
0 mxG − Nv˙ Izz − Nr˙ ṙ 0 Nv Nr − mxG V r
⎡ ⎤ tion algorithm and meet the requirements of course-keeping control.
0
+ ⎣ Yδ ⎦δ
Nδ 2.2. Fuzzy fading memory
(2)
The convergence and accuracy of classical algorithms are affected by
In Eq. (2), the equation of surge velocity is decoupled from the
the forgetting factor. The main role of the forgetting factor is to deter­
equations of sway velocity and yaw rate. Therefore, the coupling
mine the weights of the old and new data. When the forgetting factor f is
equations of sway velocity and yaw rate are obtained. This study con­
smaller, the weight of new data will increase, and the state tracking
ducts parameter identification for this coupled equation, which is suit­
ability will become stronger, but the convergence will become poorer,
able for the research of course-keeping control. Eq. (2) is decoupled and
and vice versa. Therefore, in some cases, it is difficult to accurately and
organized to meet the research of parameter identification modeling.
quickly estimate the parameters of a ship motion mathematical model
The decoupled expression is shown in Eq. (3). This model can also add
with a constant forgetting factor (L. Wang et al., 2022).
nonlinear Norrbin forces and a deceleration function of surge velocity to
In order to make up for the shortcomings of the constant forgetting
conduct ship trajectory tracking control under wind and wave interfer­
factor, FFM is applied to the dynamic updating of the forgetting factor to
ence. Eq. (3) is discretized using the Euler forward difference, as shown
perform online parameter identification modeling. This FFM is similar to
in Eq. (4).
an event-triggering mechanism, which dynamically tunes the f based on
⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤
v̇ a11 a12 0 v b11 the errors and the error rates. The inference rules are as follows: (1)
⎣ ṙ ⎦ = ⎣ a21 a22 0 ⎦⎣ r ⎦ + ⎣ b21 ⎦δ (3) When |e| is large, fuzzy control is used to dynamically reduce f to
ψ̇ 0 1 0 ψ 0 improve the estimation ability of algorithm, and vice versa. (2) When eė
⎫ <0, |e| has a decreasing trend. And if |ė| is large, fuzzy control is used to
v(k + 1) = v(k) + h(a11 v(k) + a12 r(k) + b11 δ(k)) ⎬ dynamically increase f. (3) When eė >0, |e| has an increasing trend. And
r(k + 1) = r(k) + h(a21 v(k) + a22 r(k) + b21 δ(k)) (4)
⎭
ψ (k + 1) = ψ (k) + hr(k)
Table 1
Fuzzy inference rule.
where a11 , a12 , a21 , …, and b21 are the parameters to be identified, which
are converted from hydrodynamic coefficients. v̇, ṙ, and ψ̇ are sway ac­ e NB NM ZO PM PB
ė f
celeration, yaw acceleration, and yaw rate. ψ is yaw angle, h is sample
interval. NB LL LM BM BM M
NM LM LM BM M M
Nomoto proposed a second-order response mathematical model
ZO LM M BB M LM
based on Eq. (2). This enables Eq. (2) to be linked to the design of course- PM M M BM LM LM
keeping controllers and autopilot. This research mainly studies the PB M BM BM LM LL

3
Y. Meng et al. Ocean Engineering 294 (2024) 116853

if |ė| is large, fuzzy control is used to dynamically decrease f. The ⎫

established fuzzy inference rule is shown in Table 1. This study uses χ(i) ̂
k|k = X k|k , i = 0
⎪
⎪
⎪
⎪
⎬
Table 2 to clearly describe the identification modeling process of the √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
χ(i) ̂
k|k = X k|k + (n + λ)Pk , i = 1, 2…, n (9)
proposed scheme. √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ ⎪
⎪
⎪
(n + λ)Pk , i = nx + 1, nx + 2…, n ⎪
(i) ̂ k|k
χk|k =X − ⎭
2.3. Improved unscented Kalman filter
⎫
λ = α2 (nx + κ) − n ⎪
⎪
UKF is a joint implementation of a hidden Markov model and ⎪
⎪
⎪
λ ⎪
Bayesian theorem, which can smooth, filter, and predict states through W(i)
m = ,i = 0 ⎪
⎪
⎪
⎪
n+λ ⎬
observation information, state transitions, and observation models. It
λ ( ) (10)
has been proven to be superior to extended KF in the cases of state W(i)
c = + 1 − α2 + β , i = 0 ⎪ ⎪
n+λ ⎪
⎪
tracking and parameter estimation. In order to compensate that UKF is ⎪
⎪
⎪
⎪
easy to generate singular values, this study uses IUKF to ensure the (i)
Wc = Wm =(i) 1 ⎪
, i = 1, 2…, 2n ⎪
⎭
stability of the algorithm. This algorithm utilizes QR decomposition to 2(n + λ)
solve the square root calculation of the covariance matrix, which ensures ( ) ⎫
numerical stability during the calculation process. In order to further ̂ (i) = f χ(i) , uk
X k+1|k k|k ⎪
⎪
⎪
improve the convergence performance, FFM is used to reduce the impact ⎪
⎪
⎪
∑
2n ⎪
⎪
of the accumulation of old data, which makes IUKF accurately estimate ̂ k+1|k =
X W(i) ̂(i) ⎪
⎪
m X k+1|k ⎪
⎬
parameters. The derivation process of IUKF with fuzzy forgetting factor i=0
(11)
is shown below. The model of a discretized nonlinear system is shown in {[√̅̅̅̅̅̅̅̅̅ ( ) √̅̅̅̅̅̅ ]} ⎪ ⎪
⎪
P−k+1|k W(i) ̂ 1:2n − X
̂ k+1|k , Qk ⎪
Eq. (7). = qr c X k+1|k ⎪
⎪
⎪
⎪
⎪
} { } ⎪
⎪
X k+1 = f (Xk , uk ) + ωk − ̂0 ̂ k+1|k , W (0) ⎭
(7) Pk+1|k = cholupdate PX , X k+1|k − X c
Zk+1 = h(Xk+1 ) + vk+1
( ) ⎫
where X is the state variable of the system, f( ⋅) is a three-dimensional ̂ (i) = h χ(i)
Z ⎪
k+1|k k|k ⎪
⎪
⎪
ship motion state-space model, u is the input variable of the system, ω ⎪
⎪
⎪
∑
2n ⎪
is process noise, Z is the observation vector of the system, h( ⋅) is the ⎪
(i)
̂ k+1|k =
Z W(i) ̂ ⎪
m Z k+1|k ⎪
⎪
⎪
observation equation of the system, v is the observation noise. i=0 ⎪
⎪
⎪
{[√̅̅̅̅̅̅̅̅̅ ( ) √̅̅̅̅̅ ]} ⎪
Firstly, initialize the state variables and covariance matrix P, as ⎬
P−z = qr W(i) ̂ 1:2n − Z
Z ̂ k+1|k , Rk (12)
shown in Eq. (8). c k+1|k
⎪
⎪
{0
} ⎪ ⎪
⎪
⎪
X̂ 0 = E(X0 ) ̂
PZ = cholupdate Pk+1|k , Z − Ẑ k+1|k , W (0) ⎪
⎪
⎪
(8)
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ k+1|k c
[ ]̅ ⎪
⎪
⎪
P0 = E (X0 − X ̂ 0 )(X0 − X ̂ 0 )T ∑
2n [ i ][ i ]T ⎪
⎪
⎪
⎪
PXZ = (i) ̂
Wc X k+1|k − X̂ k+1|k Z ̂ ̂ ⎪
⎭
k+1|k − Z k+1|k
Then, using symmetric point selection, select Sigma points χ to i=0

approximate the statistical characteristics of the nonlinear system. The

method of symmetrically taking Sigma points is shown in Eq. (9). n is the where α, κ, and β are scale factors representing different meanings,
dimension of the system state variables. The calculation of the scale ̂ k+1|k is the estimated system state variable of step k + 1 when iteration
X
factor λ is shown in Eq. (10). The time prediction update is shown in Eq. ̂ (i) is the estimated value in step k of the i-th Sigma point, W
to step k, X k+1|k
(11). The measurement update is shown in Eq. (12).

Table 2
Identification modeling process of the proposed scheme.

4
Y. Meng et al. Ocean Engineering 294 (2024) 116853

is the weight vector of the Sigma points, P−k+1|k is the upper triangular application background. The basic framework and structure of the
matrix of the state covariance matrix estimated in step k, qr(⋅) is an proposed scheme are shown in Fig. 1.
orthogonal trigonometric decomposition, cholupdate(⋅) is a Cholesky
decomposition process, Q is the covariance matrix of process noise, R is 3. Online parameter identification and verification
the covariance matrix of observation noise, Z ̂ k+1|k is the estimated
3.1. Online parameter identification
observed variable of step k + 1 when iteration to step k. PZ is the
observation update covariance matrix.
Based on the ship motion model and proposed scheme mentioned
Finally, the filtering results are updated, as shown in Eq. (13). To
above, this study conducts online parameter identification modeling
avoid the covariance matrix from stabilizing, a designed fuzzy forgetting
research on a ship model similar to an USV. Research on online identi­
factor F is added to Eq. (13) during the iteration process.
fication modeling can lay the model foundation for autonomous control
⎫
PXZ ⎪ of USV. We simplify the 3 DOF Abkowitz model and obtain a three-
Kk = ⎪
⎪
PZ PZ T ⎪
⎪
⎪ dimensional ship motion state-space model, which can be further
( )⎪
⎬ derived as a second-order response mathematical model. The motion
̂ ̂ ̂
X k+1|k+1 = X k+1|k + K k Zk+1 − Z k+1|k (13)
⎪
⎪ model is then used as the identification framework to further obtain the
⎪
U = K k PZ ⎪
⎪ response model. We also apply the response model to the study of
{ }⎪
⎪
⎭
Pk+1|k+1 = F⋅cholupdate Pk+1|k , U, − 1 course-keeping control. The motion data of the ship model is collected
from the SIMMAN 2020 workshop (SIMMAN, 2020). The main partic­
where, K k is Kalman gain matrix, U is the process matrix. ulars of the ship model are shown in Table 3.
Introduce the idea of LS with a forgetting factor into IUKF. A fuzzy Different identification modeling schemes vary in the selection of
forgetting factor is used to directly process the covariance matrix to identification data samples. For example, some studies select a single
prevent it from stabilizing. IUKF introduces a square root covariance test data sample, while others select a combination of multiple test data
matrix and a fuzzy forgetting factor mechanism to conduct online samples. In general, it is important to select suitable data samples for the
identification research. Its iteration foundation is still UKF. The stability identification scheme. This study selects part of the − 20◦ /-20◦ zigzag
analysis of IUKF requires the use of math definitions and decision the­
orems, such as stochastic processes. Liu and Xie (2010) provided suffi­ Table 3
cient conditions for the stability of UKF and validated it using simulation Main particulars of the ship model.
cases. Kulikov and Kulikova (2018) analyzed the stability of UKF for Parameters Values
estimating continuous discrete stochastic systems. The stochastic sta­
Length of waterline (Lwl/m) 6.1355
bility of UKF was verified by Hu et al. (2015). The proposed online Length between perpendiculars (Lpp/m) 6.0702
identification modeling scheme has certain adaptability and consumes Breadth (B/m) 0.8498
less time. The identified results provide a prior model foundation for Draft (d/m) 0.2850
motion control. Based on the identified results, the proposed control Block coefficient (CB/-) 0.6510
Displacement (∇/m3) 0.9565
scheme is simple and effective and can meet the practical engineering

Fig. 1. Basic framework and structure of the proposed scheme.

5
Y. Meng et al. Ocean Engineering 294 (2024) 116853

test data as identification sample. The number of identification sample is fuzzy domain of f to [0.95,1]. The − 10◦ /-10◦ zigzag test data, as
585. The sample interval is 0.08 s. The ship model tests were conducted generalization validation data, is completely different from the identi­
by MARIN. The initial velocity of the ship model was 2 m/s. Propeller fication sample.
presented and working at constant RPM. The speed of the propeller was UKF and EKF are both classical identification modeling algorithms
approximately 642 revolutions per minute. These test data were and the two most widely used filtering algorithms. Both can be applied
collected in deep water. During the process of collecting data from the to online identification modeling research. The performance of the two
tank, it was affected by the environment and sensors. The collected data algorithms has been compared in previous research. UKF has some ad­
have some noise interference (Ouyang and Zou, 2021). Therefore, no vantages compared to EKF (Sabet et al., 2018; Deng et al., 2021; S. Wang
white noise interference is added to the test data. The proposed algo­ et al., 2022; Zheng et al., 2022). In order to meet the requirements of
rithm utilizes mathematical reasoning and iterative recursion to conduct online identification modeling, a fuzzy forgetting factor is added to
parameter estimation. During the iteration process, there is only math­ improve identification accuracy and adaptability. The comparison al­
ematical calculation, so the algorithm consumes less time overall and gorithm, FFMUKF, is used to analyze the prediction results. To maintain
can meet the requirements of online identification. The forgetting factor uniformity, the settings of FFMUKF are the same as the proposed
can be dynamically adjusted when the environment changes so that the scheme. The prediction results of the two online identification schemes
proposed algorithm can identify accurate model parameters. The pro­ are listed. Online identification results of − 20◦ /-20◦ zigzag test data are
posed algorithm sets Q = diag(0.02,0.02,0.01, 01×6 )2, R = diag(0.02, shown in Fig. 2. It can be seen from Eq. (2)~(4) that the parameters of
0.02,0.01)2. The setting of noise matrix is small, which can improve the the identification model have no specific meaning and are the combi­
stability and accuracy of filter estimation (Xie et al., 2019). When the nation of multiple hydrodynamic coefficients. Only the parameter
value of noise matrix is small, the accuracy of state tracking is relatively identification results of the proposed scheme are shown in Table 4.
high. Therefore, set the fuzzy domain of e and ė to [− 0.5,0.5] and the Approximately half of the − 20◦ /-20◦ zigzag test data is taken as an

Fig. 2. Online prediction results of − 20◦ /-20◦ zigzag test.

6
Y. Meng et al. Ocean Engineering 294 (2024) 116853

Table 4 different from the identification sample as the generalization validation

Parameter identification results of a three-dimensional ship motion state-space data. The online prediction results of the − 10◦ /-10◦ zigzag test data are
model. shown in Fig. 3. To clearly describe the prediction results of rudder and
Model Parameters Values Model Parameters Values yaw angles, Fig. 3 (c) and (d) show the results of two algorithms
a11 − 1.882 a12 − 0.1232
respectively. The prediction effects of the two algorithms are shown in
a21 − 7.422 a22 − 0.8129 Table 6.
b11 0.0099 b21 0.1657 The − 10◦ /-10◦ zigzag test data is different from the − 20◦ /-20◦
zigzag test data, both in terms of input incentives and maneuverability
requirements. In order to clearly demonstrate the generalization of the
identification sample for online model parameter estimation. The rest is
identified model, simulation validation of the − 10◦ /-10◦ zigzag test is
taken as a validation sample. In the iterative updating process, the
conducted using the estimated parameters. The simulation experiment is
forgetting factor is adaptively updated based on the ship motion data.
conducted based on the dynamic characteristics of ship model. How­
This enables the proposed algorithm to obtain accurate model parame­
ever, the servo motor of the ship model in the tank is different from that
ters. It can be seen from Fig. 2 that the online prediction results of sway
of the full-scale and virtual ships. The simulation verification results are
velocity, yaw rate, and yaw angle have a high coincidence degree with
compared with the tank test data. For the proposed scheme, it can be
the tank test data for the proposed scheme. This also demonstrates the
seen from Fig. 3 that the predicted trends of the three state variables are
effectiveness of the proposed online identification algorithm. FFMUKF
basically consistent with the tank test data. However, there are some
has a good estimation effect in some stages. However, it cannot guar­
deviations between the prediction results of FFMUKF and tank test data.
antee a good prediction effect over the whole period. In order to quantify
As can be seen from Table 6, the prediction accuracy of state variables is
the prediction accuracy, mean square error (MSE) is used to evaluate the
higher than that of the − 20◦ /-20◦ zigzag test data. Compared with
deviation between the predicted motion state and the tank test data.
FFMUKF, the prediction accuracy of the proposed scheme is improved
Mean absolute error (MAE) is utilized to evaluate the deviation between
by about twice on average.
the predicted yaw angle and the tank test data. Table 5 shows the pre­
In terms of effectiveness and generalization verification, the pro­
diction effects of the two algorithms. As can be seen in Table 5, the
posed scheme has good prediction results. Online estimation of ship
proposed scheme can improve the prediction accuracy of yaw angle by
motion model parameters can lay the foundation of the model and
24.8% compared to FFMUKF. For other motion states, the proposed
provide prior knowledge for ship motion control. The ship motion model
scheme also has a significant improvement in prediction accuracy. On­
is crucial for ship motion control. We utilize FFM to dynamically adjust
line identification modeling requires real-time performance. For the
the forgetting factor according to ship motion data. This can improve the
proposed algorithm, each iteration process consumes less time. The
estimation accuracy of the ship motion model parameters. In addition,
proposed algorithm consumes about 0.377 s under the premise of
the proposed algorithm consumes less time during the iteration process.
selected identification samples.
On the premise of selected identification samples, the proposed algo­
Remark 2. When the ship is small and fast, the steerage will be good. rithm has certain advantages in terms of time consumption.
The ship model or USV is more flexible than a large ship. Their yaw
Remark 3. When a ship (USV) has unknown design parameters, main
angle changes quickly. The yaw rate of the ship model is larger than the
particulars, or inaccurate model parameters solved using empirical
sway velocity. The maximum is 6.2◦ /s. Therefore, the MSE of the yaw
formulas, online parameter identification modeling can obtain accurate
rate is relatively larger compared to the MSE of the sway velocity.
model parameters using limited data with less time consumption. This
study verifies the effectiveness and generalization of the model param­
eters obtained through online estimation using two different zigzag test
3.2. Parameter generalization verification
data, which lays the model foundation for the course-keeping control in
Section 4.
The zigzag test, as a standard test for studying ship maneuverability
and seakeeping performance, is widely used in ship factory inspection
and maneuvering dynamics research (Park et al., 2022). Both the ship 4. Application for course-keeping control of ship model
model and USV have the characteristics of sensitivity and low inertia.
Because zigzag test requires frequent steering, the test data have strong The identification modeling results in Section 3 can be applied to
nonlinearity and coupling between motion states. In addition, the zigzag ship motion control. With the rapid development of USV, it can be
test data contains many ship dynamic features, which is often used as applied to various marine operation scenarios, such as water rescue,
training samples in ship motion identification modeling. The input in­ ocean survey, etc. (He et al., 2023). However, as an emerging type of
centives of − 20◦ /-20◦ and − 10◦ /-10◦ zigzag tests are different. The ship, USV has a small hull size. Its mathematical model needs further
− 20◦ /-20◦ zigzag test has stronger nonlinearity than the − 10◦ /-10◦ exploration. The research plant of this study is similar to USV. Its
zigzag test. During the data collection process, each set of tank test data identification modeling and motion control work can be extended to the
was influenced by different factors, such as the instability of sensors, etc. research of USV motion modeling and control.
Therefore, there are some differences between the two sets of test data.
The proposed scheme can adaptively adjust the forgetting factor based 4.1. Control model
on changes in tank test data. In order to verify the generalization of
online estimated parameters, we select − 10◦ /-10◦ zigzag test data The research on the course-keeping controller is the core of autopi­
lots. The design of the autopilots is based on a response mathematical
Table 5 model. The model obtained in Section 3 is a premise for deriving a
Evaluation criteria of prediction results for − 20◦ /-20◦ zigzag test. response model. Based on the parameters of the identified model in
Algorithms Evaluation criteria Motion states Values Section 3, the maneuverability indices are converted using Eq. (6) to
obtain a second-order response mathematical model. The specific values
FFMUKF MSE v 0.0152
r 1.28 of the maneuverability indices are shown in Table 7 for the ship model.
MAE ψ 4.43 The Nomoto standard test method is a calculation method that directly
The proposed scheme MSE v 0.0037 obtains the K and T of the response model based on graphical analysis of
r 0.71 zigzag test data curve. However, due to the influence of ship shape and
MAE ψ 3.55
type or environmental interference, this calculation method has some

7
Y. Meng et al. Ocean Engineering 294 (2024) 116853

Fig. 3. Online prediction results of − 10◦ /-10◦ zigzag test.

8
Y. Meng et al. Ocean Engineering 294 (2024) 116853

Table 6 model, this study uses the proposed identification algorithm to obtain a
Evaluation criteria of prediction results for − 10◦ /-10◦ zigzag test. control model. This solves the challenge of difficult parameter deter­
Algorithms Evaluation criteria Motion states Values mination in ship motion modeling. Given the efficiency of the algorithm
and the background of engineering applications, this study uses CGSA to
FFMUKF MSE v 0.0046
r 0.83 conduct research on course-keeping control based on an identified
MAE ψ 3.69 model.
The proposed scheme MSE v 0.0016
r 0.49 Remark 4. In Section 3, the parameters of the ship motion model were
MAE ψ 1.95 identified using the proposed identification algorithm. The effectiveness
and generalization of the obtained model were verified. In this section,
the model obtained from online identification is applied to the study of
Table 7 course-keeping control. This integrates identification modeling and ship
Identified maneuverability indices of the ship model. motion control into an organic system, forming a complete system
modeling and control research.
Parameters Estimated values

K (1/s) 0.3867 CGSA utilizes the sensitivity function and mending sensitivity func­
T1 (s) 3.9688 tion to form a simplified H∞ robust control algorithm. Firstly, the al­
T2 (s) 0.4094 gorithm utilizes maximum singular value, bandwidth frequency, closing
T3 (s) 0.6960
slope, and closed-loop spectral peak to construct a mending sensitivity
T (s) 3.6822
function (Yan et al., 2020). Then, the mending sensitivity function is
indirectly constructed using the sensitivity function. Finally, the
errors to a certain extent. The results obtained by this method can only course-keeping controller is derived.
be used as references. The K and T are 0.576 and 4.816 using this Based on the introduction of control algorithms and ship control
method, respectively. This study has demonstrated the accuracy of the models, the first-order CGSA is used for controller design. CGSA has the
identified model by utilizing its effectiveness and generalization results features of simple parameter tuning, easy adjustment, and strong
in Section 3. The estimated parameters are close to those obtained by the robustness. The proposed CGSA sets the closing slope and Tw as − 20 dB/
Nomoto standard test method, which proves the accuracy of the iden­ dec and 3. The singular value curve of the closed-loop spectrum is
tification results to some extent. approximately the spectrum curve of a first-order inertial system with a
When large ships or USVs navigate, they are disturbed by the marine maximum singular value of 1. The expression is shown in Eq. (16). Tw
environment. In order to be close to the marine environment, the wind has a certain inhibitory effect on the wave interference.
and wave interference models are added to the identified response Gψ′ Kc
1
model. It is very difficult to accurately describe the wind interference on = (16)
Tw s + 1 1 + Gψ′ Kc
a ship. In most studies, an empirical formula is used to obtain the rudder
angle signal. The wave interference is a random process, and it is also
where 1/ Tw is the bandwidth frequency, Kc is the CGSA controller.
difficult to accurately solve the interference on ships. In general, the
In order to make the control process concise, we select the controller
second-order transfer function driven by white noise is used to simulate
derived from the first order response model. Then, the controller is used
wave interference, which is concise and convenient (Gao and Zhang,
to control the identified second-order response model. This reasoning
2022; Yan et al., 2020). Based on the empirical formula for calculating
method has a high coincidence degree with the simplicity of CGSA.
the rudder angle summarized by scholars and the transfer function
According to Eq. (16), the controller is shown in Eq. (17).
model of wave, the wind and wave interferences are added to the
response mathematical model obtained from the online identification. 1 T
Kc = + s (17)
The rudder angle produced by the wind interference is expressed using KTw KTw
Eq. (14). The transfer function of wave interference is shown in Eq. (15).
( )
Vw √̅̅̅̅̅̅̅̅̅̅̅̅ Remark 5. Fan et al. (2023), Gao and Zhang (2022), and Yan et al.
δ w = Kw sin(q) (14)
Vs (2020) proved the stability of CGSA by using the Nyquist stability cri­
terion. Therefore, the stability analysis of the proposed control algo­
0.2386s
Gwave = (15) rithm is not discussed in this study. In addition, CGSA has achieved good
s2 + 0.3593s + 0.8067
research results in the fields of trajectory tracking and stabilization
where δw is rudder angle, Kw is the leeway coefficient, with a value range control.
of [± 0.5, ± 1], Vw is the wind velocity, Vs is the velocity of ship, q is
windward angle, Gwave is the transfer function of wave. 4.3. Control result

The online identification parameters have good prediction results in

4.2. Control method generalization verification. They are not significantly different from
those obtained by the Nomoto standard test method. This indicates that
There are more control methods, such as PID, adaptive control, and the identified model can be further used for course-keeping control. This
sliding mode control algorithms in ship motion control research. How­ research is very important for the autonomous navigation system of the
ever, the most common method in industry is still the PID at present. USV. The main research objective is to organically combine online
This also proves that the control method has both theoretical and identification modeling with ship motion control. For ship course-
practical significance. Zhang and Jia (1999) proposed CGSA based on keeping control, online identification modeling provides a model foun­
H∞ robust control theory, which essentially uses the results of dation. CGSA provides the algorithm foundation. After determining the
mixed-sensitivity algorithms to design robust controllers. All parameters ship motion model and control algorithm in Sections 4.1 and 4.2, this
have engineering significance. The proposed algorithm is relatively section conducts research on the course-keeping control of the ship
concise and easy to achieve information symmetry. In practical engi­ model.
neering applications, the algorithm has been applied to the study of In order to demonstrate the course-keeping control of the ship model
course-keeping control for USV. Considering the unknown ship motion in wind and wave, this study adds interferences to the control model

9
Y. Meng et al. Ocean Engineering 294 (2024) 116853

based on Eqs. (14) and (15). The initial course and target course are set
to 0◦ and 60◦ . According to Eq. (14), the rudder angle is 2.23◦ under
wind scales of Beaufort No. 5. The wind direction is 0◦ . The rudder angle
is taken as 2◦ . The model of wave interference is shown in Eq. (15). The
response mathematical model does not consider the main particulars of
the ship. After setting the parameters of the controller and environment
interferences, the course-keeping control effect of the ship model is
shown in Fig. 4. The input rudder angle is shown in Fig. 5.
In order to demonstrate the effectiveness of the proposed algorithm,
the duration of the simulation experiment is 200 s. The error points take
4 s as the sampling interval to clearly show the control error in Fig. 4. In
addition, the line segment connected by the error points is the error of
the entire control process. As can be seen from Fig. 4, after entering the
steady-state process, the rise time of the proposed algorithm is 24 s, and
the deviation between the course-keeping control result and the target
course is 1.46◦ at most. From the perspective of control effect, the
proposed algorithm can not only satisfy certain control accuracy but also
satisfy the practical application of USV. It can be seen from Fig. 5 that
the maximum rudder angle is less than 25◦ . After stabilizing the input
Fig. 5. Rudder angle of the proposed control scheme.
rudder angle, the variation range of the rudder angle is [− 1.09◦ ,
− 3.78◦ ], with a variation amplitude of 2.69◦ . The amplitude change of
models, this study uses CGSA to conduct the course-keeping control of
the rudder angle is small, which can achieve the goal of saving energy
the ship model. During the steady-state process, the MAE between the
consumption. In order to ensure that the rudder angle meets the re­
control result and the target course is 0.86◦ . Based on the results ob­
quirements of maritime practice, this study limits the maximum rudder
tained in Sections 3 and 4, the feasibility of the proposed scheme that
angle and the rate of rudder angle. Although there is a certain control
combines online identification and course-keeping control is confirmed.
deviation, the proposed scheme is still guaranteed to meet the re­
quirements of navigation practice. The average rudder angle during the
entire control process is − 1.70◦ . 5. Conclusions

Remark 6. Under harsh sea conditions, it is difficult to collect motion A study on the organic combination of online identification modeling
data for both large ships and USVs. It is also difficult to conduct research and course-keeping control of the ship model is conducted. When the
on ship motion identification modeling under harsh sea conditions. Due state tracking error is large, this FFM will be triggered to dynamically
to the unknown design parameters and main particulars of the ship tune the forgetting factor. The IUKF enhances the stability and accuracy
model or USV, mechanism modeling is difficult to meet the requirements of the original algorithm. Based on the motion data of the ship model,
of simulation and control. In addition, given the influence of the shape FFM, and IUKF, the parameters of the three-dimensional ship motion
and environment, there are certain differences in the mechanism state-space model are estimated online. To verify the superiority of the
modeling between the USV and large ship. Therefore, the nominal model proposed scheme, we use FFMUKF as a comparative algorithm. The
is not easy to solve. The control law and effect of the nominal model are comparison results demonstrate the effectiveness and generalization of
not given in this study. The purpose of online identification modeling in the proposed scheme. In the prediction results of the proposed scheme,
this study is to provide an accurate ship motion model for the afore­ the MSE of the predicted motion state is below 0.71◦ . The MAE of the
mentioned situations. predicted yaw angle is below 3.55◦ . The proposed algorithm consumes
This study utilizes the IUKF algorithm to perform online parameter less time per iteration, with a total iteration time of 0.377 s. Using the
estimation of a three-dimensional ship motion state-space model. The identified motion model, wind and wave interference models, and
identified model is converted into a response mathematical model based CGSA, research on the course-keeping control of the ship model is per­
on Eq. (6). Based on the control model and established interference formed. This CGSA can be applied to practical engineering applications.
After entering steady state, the maximum deviation between the course-
keeping control result and the target course is 1.46◦ . The maximum
input rudder angle is less than 25◦ . Given the challenges of mechanism
modeling, system identification modeling can provide a prior model
foundation for motion control of USV.
This study is initial research that combines online identification
modeling and course-keeping control. Although some research results
were obtained in this study, there are still some limitations. In future
work, we will focus on improving the classical identification algorithm
using multi-innovation or nonlinear innovation theory. In addition, we
will also further improve the accuracy of the control algorithm and
develop it for the practical application of USV in our laboratory. We will
also further conduct more complex ship motion identification modeling
research and apply it to ship motion control. If we can collect ship
motion data under harsh sea conditions, we will also conduct ship mo­
tion modeling and control under harsh sea conditions. This is of great
significance for future research on USVs.

CRediT authorship contribution statement

Fig. 4. Control effect of the ship model course. Yao Meng: Conceptualization, Methodology, Software, Validation,

10
Y. Meng et al. Ocean Engineering 294 (2024) 116853

Visualization, Writing – original draft. Xianku Zhang: Funding acqui­ Jiang, Y., Wang, X., Zou, Z., Yang, Z., 2021. Identification of coupled response models for
ship steering and roll motion using support vector machines. Appl. Ocean Res. 110
sition, Methodology, Project administration, Supervision, Writing – re­
https://doi.org/10.1016/j.apor.2021.102607.
view & editing. Xiufeng Zhang: Methodology, Supervision, Writing – Kulikov, G.Y., Kulikova, M.V., 2018. Stability analysis of Extended, Cubature and
review & editing. Daocheng Ma: Methodology, Supervision, Writing – Unscented Kalman Filters for estimating stiff continuous–discrete stochastic systems.
review & editing. Yating Duan: Supervision, Writing – review & Automatica 90, 91–97. https://doi.org/10.1016/j.automatica.2017.12.055.
Li, J., Zhang, G., Shan, Q., Zhang, W., 2022. A novel cooperative design for USV-UAV
editing. systems: 3D mapping guidance and adaptive fuzzy control. IEEE Transactions on
Control of Network Systems 10 (2), 564–574. https://doi.org/10.1109/
tcns.2022.3220705.
Declaration of competing interest Liu, T., Xie, Y., 2010. Stability analysis of UKF and its application in relative navigation.
Journal of Astronautics 31 (3), 739–747. https://doi.org/10.3873/j.issn.1000-
The authors declare that they have no known competing financial 1328.2010.03.019.
Ouyang, Z., Zou, Z., 2021. Nonparametric modeling of ship maneuvering motion based
interests or personal relationships that could have appeared to influence on Gaussian process regression optimized by genetic algorithm. Ocean Eng. 238
the work reported in this paper. https://doi.org/10.1016/j.oceaneng.2021.109699.
Park, J., Lee, D., Park, G., Rhee, S.H., Seo, J., Yoon, H.K., 2022. Uncertainty assessment
of outdoor free-running model tests for maneuverability analysis of a damaged
Data availability surface combatant. Ocean Eng. 252 https://doi.org/10.1016/j.
oceaneng.2022.111135.
Data will be made available on request. Pei, T., Yu, C., Zhong, Y., Lian, L., 2023. Adaptive event-triggered mechanism-based
online system identification framework for marine craft. Ocean Eng. 278 https://doi.
org/10.1016/j.oceaneng.2023.114572.
Acknowledgement Perera, L.P., Oliveira, P., Guedes Soares, C., 2016. System identification of vessel steering
with unstructured uncertainties by persistent excitation maneuvers. IEEE J. Ocean.
Eng. 1–14. https://doi.org/10.1109/joe.2015.2460871.
Much appreciations to each reviewer for their valuable comments
Sabet, M.T., Pouria, S., Zarini, M., 2014. Extended and Unscented Kalman filters for
and suggestions to improve the quality of this note. This work is partially parameter estimation of an autonomous underwater vehicle. Ocean Eng. 91,
supported by National Key R&D Program of China (Grant 329–339. https://doi.org/10.1016/j.oceaneng.2014.09.013.
No.2022YFB4301402), National Science Foundation of China (Grant Sabet, M.T., Daniali, H.M., Fathi, A., Alizadeh, E., 2018. Identification of an autonomous
underwater vehicle hydrodynamic model using the extended, cubature, and
No. 52171291), National Science Foundation of China (Grant transformed unscented kalman filter. IEEE J. Ocean. Eng. 43 (2), 457–467. https://
No.51679024), Dalian Innovation Team Support Plan in the Key doi.org/10.1109/joe.2017.2694470.
Research Field (Grant No.2020RT08), Doctoral Research Initial Fund SIMMAN, 2020. Simulation Workshop for Ship Maneuvering. Available online: https://
www.simman2020.kr/.
Project of Liaoning Province (No. 2021-BS-078), the University 111 Silva, K.M., Maki, K.J., 2022. Data-Driven system identification of 6-DoF ship motion in
Project of China (Grant No. B08046), International cooperation training waves with neural networks. Appl. Ocean Res. 125 https://doi.org/10.1016/j.
program for innovative talents of Chinese Scholarships Council (Grant apor.2022.103222.
Song, C., Zhang, X., Zhang, G., 2021. Nonlinear identification for 4 DOF ship
No. CSC [2022] 2260) and Fundamental Research Funds for the Central maneuvering modeling via full-scale trial data. IEEE Trans. Ind. Electron.
Universities (Grant NO.3132021139). The authors would like to thank 1829–1835. https://doi.org/10.1109/tie.2021.3062255.
anonymous reviewers for their valuable comments to improve the Sun, X., Wang, G., Fan, Y., 2019. Model identification and trajectory tracking control for
vector propulsion unmanned surface vehicles. Electronics 9 (1). https://doi.org/
quality of this article. 10.3390/electronics9010022.
Sutulo, S., Guedes Soares, C., 2014. An algorithm for offline identification of ship
References manoeuvring mathematical models from free-running tests. Ocean Eng. 79, 10–25.
https://doi.org/10.1016/j.oceaneng.2014.01.007.
Wang, L., Li, S., Liu, J., Wu, Q., 2022. Data-driven model identification and predictive
Abkowitz, M.A., 1964. Lectures on ship hydrodynamics–steering and manoeuvrability.
control for path-following of underactuated ships with unknown dynamics. Int. J.
Report Hy-5, Hydro-and aerodynamic laboratory. Lyngby 35 (18).
Nav. Archit. Ocean Eng. 14 https://doi.org/10.1016/j.ijnaoe.2022.100445.
Abrougui, H., Nejim, S., Hachicha, S., Zaoui, C., Dallagi, H., 2021. Modeling, parameter
Wang, S., Wang, L., Im, N., Zhang, W., Li, X., 2022. Real-time parameter identification of
identification, guidance and control of an unmanned surface vehicle with
ship maneuvering response model based on nonlinear Gaussian Filter. Ocean Eng.
experimental results. Ocean Eng. 241 https://doi.org/10.1016/j.
247 https://doi.org/10.1016/j.oceaneng.2021.110471.
oceaneng.2021.110038.
Wang, X., Zou, Z., Yu, L., Cai, W., 2015. System identification modeling of ship
Bai, W., Ren, J., Li, T., Chen, C.L.P., 2019. Grid index subspace constructed locally
manoeuvring motion in 4 degrees of freedom based on support vector machines.
weighted learning identification modeling for high dimensional ship maneuvering
China Ocean Eng. 29 (4), 519–534. https://doi.org/10.1007/s13344-015-0036-9.
system. ISA Trans. 86, 144–152. https://doi.org/10.1016/j.isatra.2018.11.001.
Wang, H., Wang, C., Gao, H., Xu, S., 2022. Vehicle traction force control based on the
Cao, T., Zhang, X., 2023. Nonlinear decoration control based on perturbation of ship
road adhesion coefficient estimation BY FFUKF. Chin. J. Theor. Appl. Mech. 54 (7),
longitudinal motion model. Appl. Ocean Res. 130 https://doi.org/10.1016/j.
1866–1879. https://doi.org/10.6052/0459-1879-22-211.
apor.2022.103412.
Wang, Z., Xu, H., Xia, L., Zou, Z., Soares, C.G., 2020. Kernel-based support vector
Chen, L., Yang, P., Li, S., Liu, K., Wang, K., Zhou, X., 2023. Online modeling and
regression for nonparametric modeling of ship maneuvering motion. Ocean Eng. 216
prediction of maritime autonomous surface ship maneuvering motion under ocean
https://doi.org/10.1016/j.oceaneng.2020.107994.
waves. Ocean Eng. 276 https://doi.org/10.1016/j.oceaneng.2023.114183.
Xie, S., Chu, X., Liu, C., Liu, J., Mou, J., 2019. Parameter identification of ship motion
Deng, F., Levi, C., Yin, H., Duan, M., 2021. Identification of an Autonomous Underwater
model based on multi-innovation methods. J. Mar. Sci. Technol. 25 (1), 162–184.
Vehicle hydrodynamic model using three Kalman filters. Ocean Eng. 229 https://doi.
https://doi.org/10.1007/s00773-019-00639-y.
org/10.1016/j.oceaneng.2021.108962.
Xu, H., Hassani, V., Guedes Soares, C., 2020. Truncated least square support vector
Fan, J., Zhang, X., Cao, T., Ma, D., 2023. Ship longitudinal stabilization control based on
machine for parameter estimation of a nonlinear manoeuvring model based on PMM
CGSA and nonlinear switch modification. Ocean Eng. 286 https://doi.org/10.1016/
tests. Appl. Ocean Res. 97 https://doi.org/10.1016/j.apor.2020.102076.
j.oceaneng.2023.115707.
Xue, Y., Chen, G., Li, Z., Xue, G., Wang, W., Liu, Y., 2022. Online identification of a ship
Gao, S., Zhang, X., 2022. Course keeping control strategy for large oil tankers based on
maneuvering model using a fast noisy input Gaussian process. Ocean Eng. 250
nonlinear feedback of swish function. Ocean Eng. 244 https://doi.org/10.1016/j.
https://doi.org/10.1016/j.oceaneng.2022.110704.
oceaneng.2021.110385.
Yan, Z., Zhang, X., Zhu, H., Li, Z., 2020. Course-keeping control for ships with nonlinear
Guo, Q., Zhang, X.K., Han, X., 2023. Robust adaptive event-triggered path following
feedback and zero-order holder component. Ocean Eng. 209 https://doi.org/
control for autonomous surface vehicles in shallow waters. Ocean Eng. 286 https://
10.1016/j.oceaneng.2020.107461.
doi.org/10.1016/j.oceaneng.2023.115571.
Zhang, G., Zhang, X., Pang, H., 2015. Multi-innovation auto-constructed least squares
Han, X., Zhang, X., 2020. Trajectory tracking control based on closed-loop gain shaping
identification for 4 DOF ship manoeuvring modelling with full-scale trial data. ISA
algorithm. J. Harbin Eng. Univ. 41 (11), 1629–1635. https://doi.org/10.11990/
Trans. 58, 186–195. https://doi.org/10.1016/j.isatra.2015.04.004.
jheu.201909019.
Zhang, H., Zhang, X., Bu, R., 2022. Sliding mode adaptive control for ship path following
Han, X., Zhang, X., 2022. Tracking control of ship at sea based on MPC with virtual ship
with sideslip angle observer. Ocean Eng. 251 https://doi.org/10.1016/j.
bunch under Frenet frame. Ocean Eng. 247 https://doi.org/10.1016/j.
oceaneng.2022.111106.
oceaneng.2022.110737.
Zhang, M., Taimuri, G., Zhang, J., Hirdaris, S., 2023. A deep learning method for the
He, Z., Fan, Y., Wang, G., Qiao, S., 2023. Finite time course keeping control for
prediction of 6-DoF ship motions in real conditions. Proc. IME M J. Eng. Marit.
unmanned surface vehicles with command filter and rudder saturation. Ocean Eng.
Environ. https://doi.org/10.1177/14750902231157852.
280 https://doi.org/10.1016/j.oceaneng.2023.114403.
Zhang, X., Jia, X., 1999. New control algorithm on closed-loop gain shaping and its
Hu, G., Gao, S., Zhong, Y., Gao, B., 2015. Stochastic stability of the derivative unscented
application. Acta Electron. Sin. 27 (11), 133–135.
Kalman filter. Chin. Phys. B 24 (7). https://doi.org/10.1088/1674-1056/24/7/
070202.

11
Y. Meng et al. Ocean Engineering 294 (2024) 116853

Zhang, X., Meng, Y., Liu, Z., Zhu, J., 2021. Modified grey wolf optimizer-based support Zhu, M., Hahn, A., Wen, Y.-Q., Bolles, A., 2017. Identification-based simplified model of
vector regression for ship maneuvering identification with full-scale trial. J. Mar. Sci. large container ships using support vector machines and artificial bee colony
Technol. 27 (1), 576–588. https://doi.org/10.1007/s00773-021-00858-2. algorithm. Appl. Ocean Res. 68, 249–261. https://doi.org/10.1016/j.
Zheng, J., Yan, D., Yan, M., Li, Y., Zhao, Y., 2022. An unscented kalman filter online apor.2017.09.006.
identification approach for a nonlinear ship motion model using a self-navigation
test. Machines 10 (5). https://doi.org/10.3390/machines10050312.

12