ORIGINAL RESEARCH (AR) DOI: 10.4274/jems.2024.

46514
Journal of ETA Maritime Science 2024;12(1):64-73

Modelling Using Neural Networks and Dynamic Position

Control for Unmanned Underwater Vehicles
Melek Ertogan1, Philip A. Wilson2
1İstanbul Technical University Faculty of Maritime, Department of Maritime Transportation and Management, İstanbul, Türkiye
2The University of Southampton, Ship Dynamics within Engineering and Physical Sciences, Southampton, United Kingdom

Abstract
Underwater construction, maintenance, and mapping use autonomous underwater vehicles (AUVs) for path planning, path following,
and target tracking operations. However, dynamic position management and localization of AUVs are critical issues. Correct localization
and dynamic position management to prevent drifts can be used to acquire information on energy efficiency, another crucial topic. In this
paper, AUV dynamic modeling using experimental data and position control is studied. The experiments were implemented on a Delphin2
scaled AUV model belonging to the Engineering and Environment Faculty, University of Southampton, UK. Hover and flight style motions
according to the different speeds of Delphin2 were implemented in the test tank. Nonlinear coupled mathematical models were studied
using shallow neural networks. The models are formed into depth-pitch and heading motion black-box models using the shallow neural
network (SNN) algorithm. Proportional integral derivative control of heading motions and depth-pitch motion simulation studies were
applied to the SNN model.
Keywords: Autonomous underwater vehicle, Shallow neural networks, Black box modelling, Dynamic position control

1. Introduction and magnetometer. The heading-pitch-depth motions of

Autonomous underwater vehicles’ (AUVs) research and Delphin2 can be measured in tank tests and underwater
application topics include localization, path following, applications.
target tracking, underwater mapping, and dynamic position Surge and sway motions of an AUV cannot be measured
management. Energy efficiency is a key issue that can be during underwater operations for this research study
resolved by combining precise localization with dynamic because an ultra-short baseline (USBL) positioning system
position control to halt drifts. cannot be used for tank tests because of the wall effect.
The goal of this study is to create and test an algorithm for An over-actuated design enables the Delphin2 AUV model to
dynamic position controllers for AUVs. The scaled Delphin2 perform various missions, from hover-style operation at zero
AUV model from the Engineering and Environment Faculty or slow speeds to flight-style operation at forward speeds up
at the University of Southampton in the UK was used for to approximately 1 m/s.
the experiments [1-6]. To provide dynamic position control Hover and flight style motions according to the different
during hover and flight-style operations underwater, the speeds of Delphin2 were implemented in the tank.
local position definition of AUVs is a crucial challenge. Nonlinear coupled mathematical models were studied using
Delphin2 has an inertial measurement unit (IMU), a experimental data. Proportional integral derivative (PID)
depth pressure sensor, a sounding altimeter, a mechanical control of heading and depth-pitch motions simulation
scanning sonar, and a global positioning system (GPS). GPS studies were performed on the nonlinear mathematical
does not work underwater or in closed areas such as tank model. Previous control studies on Delphin2, such as model
tests. The IMU includes a 3D accelerometer, gyroscope, predictive control (MPC) depth-pitch motion control [3]

Address for Correspondence: Melek Ertogan, İstanbul Technical University Faculty of Maritime, Department Received: 14.08.2023
of Maritime Transportation and Management, İstanbul, Türkiye Last Revision Received: 28.11.2023
E-mail: [email protected] Accepted: 03.01.2024
ORCID ID: orcid.org/0000-0002-9968-6254

To cite this article: M. Ertogan, and P. A. Wilson. “Modelling Using Neural Networks and Dynamic Position Control for Unmanned Underwater Vehicles.”
Journal of ETA Maritime Science, vol. 12(1), pp. 64-73, 2024.

Copyright© 2024 the Author. Published by Galenos Publishing House on behalf of UCTEA Chamber of Marine Engineers.
This is an open access article under the Creative Commons AttributionNonCommercial 4.0 International (CC BY-NC 4.0) License.

64
 Journal of ETA Maritime Science 2024;12(1):64-73

and sliding mode control heading motion control [6], were systems is presented, along with recommendations for the
implemented. limitations of USBL regarding ice area docking and shallow
The feedback signals were collected in the tank tests from coast applications. The drift in the location estimation that
the sensors, a sounding altimeter, a pressure depth sensor, results from DVL-based navigation makes long-range AUV
and an IMU, according to set values on the actuator, including navigation over 300 m considerably more challenging. To
the vertical and horizontal tunnel thrusters, the vertical and navigate underwater, a DVL is rarely used alone; instead, it
horizontal control surfaces as the tails, and the propeller. The is integrated with other sound sensors [15]. According to
input-output test data were used to form nonlinear coupled Chen et al. [16] and Paull et al. [17], survey research was
mathematical models. The models were formed in two conducted to investigate the mapping, navigation, and
groups: altitude-pitch and heading motion black-box models localization of AUVs.
using a shallow neural network (SNN) algorithm [7]. This The accuracy of AUV localization during tank tests, for
nonlinear coupled mathematical model was used to develop example, to verify control algorithms, is compromised
a dynamic position control design. The five actuator control by wall effects when using the USBL/LBL measurement
signals must be calculated online by a controller system system. Additionally, the AUV’s localization range, when
using 3D depth-pitch-heading motions for the over-actuated using eye feedback, is extremely small, as is its frequency.
Delphin2. AUV motion drifts could be caused by dead reckoning
techniques. The combination of DVL and IMU for
2. Literature Review underwater vehicles was investigated using multisensor
Sensor data fusion for navigation AUVs has received Kalman filtering [18].
much attention in the literature because their localization There are fewer dynamic position control studies than
remains a challenge. Underwater navigation techniques navigation studies on AUVs. However, there have been few
most commonly employed include long baseline (LBL) with experimental studies on dynamic position control. Dynamic
IMU, USBL, and Doppler velocity log (DVL) sensors. positioning systems of remotely operated vehicles and AUVs
Because of wall effects, USBL-based echo-sounding are explained comparatively [19]. Simulation studies on
communication cannot be used in a tank or on a shallow dynamic position control algorithms have been conducted
coast. Therefore, rangefinder sensors, such as a laser-based [20,21]. In addition, experimental data were designed and
vision feedback sensor, and sonar systems were used to verified for the 4-DOF and 6-DOF advanced controllers of
calculate AUV’s localization in a tank in the literature. With AUVs [22]. The system identification (SI) method was used
the aid of an IMU and a laser-based vision system, online with experimental data for the surge and yaw modeling
localization of AUVs via Kalman filtering was established of AUVs [23]. Using the USBL measurement system, we
[8,9]. These tests were performed in a tank. However, the demonstrated a novel optimization-based method with
laser-based vision feedback measurement range was only simulation results for dynamically placing a fully actuated
approximately 30 cm, and the computer vision feedback AUV [24]. The neural network model reference adaptive
was seldom performed. After the localization study, online control serves as the dynamic loop in the proposed dynamic
identification and visual control were studied [10]. Another position approach, whereas the nonlinear MPC serves as
laser-based rangefinder study was applied in a tank, and the the kinematic loop.
range was approximately 5 m [11].
The position of the AUV in artificially structured 3. Materials and Methods
environments, such as tanks, harbors, marinas, and 3.1. AUV Characteristics
maritime platforms, was determined using a mechanical The Delphin series for AUVs was initiated in 2007 as a
scanning sonar with an IMU [12]. The experimental results collaboration between the University of Southampton
of the simultaneous localization and mapping measurement and the National Oceanography Centre [1]. Delphin2 was
method are presented in this thesis. To address motion- developed as a scaled model of Autosub6000 [25]. With
induced distortion caused by high measurement error or control surfaces and a propeller, its body is designed like
lengthy scan periods, a different localization of AUV based a torpedo. Because of the inefficiency of these types of
on mechanically scanned imaging sonar was investigated rear control surfaces, a typical design flaw is that the
[13]. AUV is unable to maintain sufficient control at zero or
For open-water applications, acoustic-based USBL and LBL slow speeds. Thus, the design is expanded to include two
communication systems are used to determine the location horizontal and two vertical thrusters. Delphin2 serves
of AUVs. According to Plueddemann et al. [14], a comparative as a research platform for control system development
discussion of USBL- and LBL-based AUV navigation and for studying the performance of over-actuated AUVs.

65
Modelling Using Neural Networks and Dynamic Position Control for Unmanned Underwater Vehicles

A Delphin2 AUV with actuators is shown in Figure 1. This forward speeds of u = {0, 0.26, 0.6, 1.0} m/s or u = {0, 0.42,
AUV has been successfully demonstrated during hover- 0.82, 1.03} m/s for operating on the water surface. However,
style motions [3,4]. these speeds were calculated according to the operation
time and measured distance regardless of the drifts of the
Delphin2 AUV.
Accordingly, they are referred to as zero, low, mid, and high-
speed cases. The thrusters function well at low and zero
speeds. However, the tails work effectively at high speeds.
In addition, both thrusters and tails cannot work fully
effectively at mid-speeds. The effectiveness of the actuators
at mid-speeds is determined with weighting functions, and
the illustration is shown in Figure 2. ​ωth​ ​, a tunnel thruster
weight, and ​ωs​​, a control surface tail weight.

Figure 1. Delphin2 AUV [3]

AUV: Autonomous Underwater Vehicles

AUV Delphin2 has a mid-body diameter of 0.26 m and a

length of 1.96 m, giving it a torpedo-like shape. The over-
actuated Delphin2 has 7 actuators to be fully controlled.
These are the propeller, the vertical and horizontal tails
acting as control surfaces, the front and back vertical
thrusters, and the front and rear horizontal thrusters. 3D
depth-pitch-heading motion coupled control using five
actuators was the focus of the research study. Consequently, Figure 2. Actuator weighting functions [6]
it may effectively perform a range of tasks, such as survey
flights at a surge speed of approximately 0-1 m/s and zero-
3.2. Nonlinear Dynamic Modeling
speed hovering. In the event of a system failure, an AUV
Before control applications on the Delphin2 AUV model,
can autonomously return to the water’s surface because of
dynamic modeling of the model was investigated; hence,
its small positive buoyancy. Delphin2 is typically ballasted
a modeling strategywas chosen: a) nonlinear modeling
to be buoyant at 6 N. A surge-sway motion measurement
is somewhat complex but might better capture actual
system is not present on the AUV [2].
dynamics. The coefficients can be determined using data
The Delphin2 AUV has a pressure transducer rated from from scaled-model testing or full-scale sea trials when the
0 to 5 bar to measure its depth below the free surface. model’s structure is readily accessible for the specific type
For surface operation, the GPS offers a current position at of sample AUV. b) Linearized models such as state-space
a sample rate of one hertz (Hz). There is no underwater models or transfer functions may be adequate for the initial
positioning option for the Delphin2 AUV. The direction and control design. The simplicity of these linear models is one
turning rate are provided at a sample rate of 20 Hz via the of their advantages. However, the simulation results and
Xsens 4th generation MTi-30 IMU. The dynamics model is actual system reactions could vary significantly.
used to estimate the forward and sway velocities used in
AUV dynamics can be modeled using three different
the control systems at 20 Hz to comply with the IMU. Using
approaches: white box modeling, which employs the Navier-
acoustic backscatter techniques, the altimeter and scanning-
Stokes equations to characterize fluid structure interactions;
sonar track the distance between the AUV and the seabed.
gray box modeling, which combines experimental data with
Delphin2 is equipped with two analog color charge-coupled
partial theoretical structures; and black box modeling, which
device cameras, one facing ahead and the other below [2,6].
uses only experimental data. White box modeling promises
During the testing, propeller demands of uprop = {0, 10, 16, accurate predictions of AUV motions, but due to its time-
22} were employed. These values, which roughly translate consuming nature, it is not useful for control design. The
to 2.4, 4.5, and 6.15 rev/s, agree with prop. The motor term “System Identification-SI” approach refers to studies
control board needs these set points to operate the motor. including gray and black box modeling in transdisciplinary
For fully submerged operation, they roughly correspond to disciplines. The SI method has proven to be very accurate

66
 Journal of ETA Maritime Science 2024;12(1):64-73

when compared with both empirical and theoretical

methodologies [26]. There are few studies on AUV motion
modeling using the SI technique for controller design.
The feedback signals were collected with the sensors of
Delphin2, while the actuators were sending various signals
in the tank experiments. These input-output data were
used for nonlinear modeling of the AUV. Nonlinear coupled
mathematical modeling based on the black-box model
method was studied. Nonlinear depth-pitch motions and Figure 4. Input-output data for depth-pitch dynamics
heading motions were modeled using SNNs to develop a
nonlinear coupled control algorithm. Two datasets-training and test data-are prepared and used
SNNs typically have fewer hidden layers, whereas deep in SNN modeling. Test data are used to gage the model’s
neural networks can have dozens or even hundreds of layers. performance after it has been developed using training
There are several layers in the network structure, including data. Several acronyms of input and output data are
an input layer, hidden layers, and an output layer. While the employed in the following figures. The AUV’s propulsion
hidden layer routes inputs to the output layer, the input data are displayed as “uprop”. During the testing, propeller
layer handles intermediate calculations. When input values demands of uprop = {0, 10, 16, 22} were employed. These
are applied, the primary goal of the network is to generate values, which roughly translate to 2.4, 4.5, and 6.15 rev/s,
the desired output. For neural network training, the well- agree with prop. The horizontal control surfaces as tail
liked supervised learning method back propagation (BP) are abbreviated as “tailH”, their unit is degree. The vertical
is propesed [27]. Due to the typical BP algorithm’s slower control surfaces as tail are indicated as “tailV”. The altimeter
convergence and longer training times, BP with adaptive measures the vertical distance of the AUV from the tank
learning rate and momentum term (BPALM) is advocated bottom, and this distance is called the “altitude”. In addition,
[28]. Because of the shorter training time, the BPALM, the vertical thrusters located at the rear and front are
which is based on the standard BP, adjusts its learning rate abbreviated as “thrVrear” and “thrVfront”, respectively. The
and momentum rate at each iteration. Traingdx in Matlab horizontal thrusters located at the rear and front are shown
is used to implement the BPALM method [29]. An output as “thrHrear” and “thrHfront”, respectively. All thrusters’
layer of linear neurons follows one or more hidden layers units are rpm.
of sigmoid neurons in feedforward networks. Many layers The experimental data used in the SNN modeling includes
of neurons with nonlinear transfer functions within the the depth, pitch, and heading motions of the AUV in hover-
network enable the learning of both nonlinear and linear and flight-style conditions at various speeds. In addition,
interactions between input and output vectors. The linear data on the actuators of the AUV, such as the horizontal-
output layer indicates that the network may produce values vertical thrusters and control surfaces (tails), were obtained.
outside the 1 to +1 range. The feedforward network models Depth-pitch motion and heading motion were modeled as
were trained using the BPALM technique [7]. decoupled. The depth-pitch and heading motion training
In Figure 1, the actuators used as input data for modeling data of the AUV given in Figures 5-11 are data of the same
heading dynamics SNNs and depth-pitch dynamics SNNs mission motions. In addition, the depth-pitch and heading
are displayed. Figures 3 and 4, respectively, show the motion validation/test data of the AUV given in Figures 6-12
input-output data for heading dynamics and depth-pitch are data of the same mission motions.
dynamics.

Figure 5. SNN model training data for depth-pitch motion

Figure 3. Input-output data for heading dynamics SNN: Shallow neural network

67
Modelling Using Neural Networks and Dynamic Position Control for Unmanned Underwater Vehicles

Figure 6. SNN model test data for depth-pitch motion

Figure 9. Comparison between SNN output and experimental pitch
SNN: Shallow neural network
data for training and validation
SNN: Shallow neural network

Figure 7. Comparison between SNN output and experimental

altitude data for training and validation Figure 10. Correlation coefficients, R, of the SNN model for training
and validation pitch angle data
SNN: Shallow neural network
SNN: Shallow neural network

Figure 8. Correlation coefficients, R, of NNM for training and

validation altitude data Figure 11. SNN model training data for heading motion
SNN: Shallow neural network

Figures 5 and 6 show the depth-pitch motion modeling variables is measured by the R-value, which also indicates
training and validation/test data, respectively. If the the direction of the association. The values fall between -1
horizontal control surface, tailH, is sent saturated limit and +1. This coefficient range can be interpreted as negligible
signals, it causes large amplitudes, as seen in Figure 5, with if it falls between 0 and 0.09, weak if it falls between 0.1
approximately 500 sampled data. and 0.39, moderate if it falls between 0.4 and 0.69, strong
The correlation coefficient (R-value) is used to assess the if it falls between 0.7 and 0.89, and very strong if it falls
outcomes of SNN model applications created with the aid between 0.9 and 1.00 [30,31]. In the application results, the
of Matlab software. The linear link between two continuous R-value is expressed as a proportional percentage.

68
 Journal of ETA Maritime Science 2024;12(1):64-73

the proportional KP coefficient should first be adjusted

according to the gain rate of the system. The KP coefficient
should be increased until the response of the system
reaches the reference signal. In the second step, if there is
a steady-state error in the system response, the integrative
KI coefficient should be increased/decreased until this
steady-state error is eliminated. However, the disadvantage
of increasing the integration coefficient is that delays and
oscillations in the system response increase. In the third
step, the derivative KD coefficient may need to be adjusted to
optimally adjust the system response time and oscillation.
Figure 12. SNN model validation data for heading motion
SNN: Shallow neural network There are two operating conditions in which depth-
pitch control is used. The first involves modifying the
Altitude and pitch motions of SNN depth-pitch model type equation using a hover-style control method (at
outputs have high R-values, approximately 90%. These zero speed). Using vertical thrusters, the AUV depth and
comparisons between the SNN output and experimental pitch may be changed. The second is a flight-style control
test data and the R-values are shown in Figures 7-10. approach that uses vertical thrusters for depth control
and horizontal control surfaces (tails) for pitch motion
Training and validation data for SNN heading motion
reduction control. When operating across a wide range
modeling are given in Figures 10-12. A comparison between
of speeds, the horizontal aft surfaces,​​𝝎​ ​​, and the vertical
the SNN output and experimental heading data for training s
tunnel thrusters, ​𝝎​ ​ ​​, offer a transition between the two
and validation is given in Figure 13. R-values, of SNN th
control techniques. The integration (I) coefficient was the
heading modeling training and validation outputs are 78%
most effective in depth control based on simulation studies.
and 87%, respectively.
When dealing with a significant shift in depth demand, an
4. Dynamic Position Control integrator for the generalized thrust control law may grow
unreasonably large while the AUV is diving. The impact
The PID controller is the most common type of closed-loop
of the integral windup phenomenon is minimized using
control system. These controllers continually monitor and
a conditional integration technique. The pitch for depth
modify a system’s output to maintain a specified set point.
for hover-style and flight-style operations, proportional
The comparator loops back to the system output, y(t),
(P), integral (I)-derivative (D), and PI-D controls are used.
and compares it to the set point, r(t), to produce the error
It implemented a PI-D strategy instead of a PID to avoid a
signal, e(t)=r(t)-y(t). The closed-loop control reduces this
spike in the derivative term when changing the demand
error signal as much as possible before using it to produce
[2,6]. The Enhanced Differentiator (ED) approach was used
the control signal u(t). The most general mathematical
to differentiate between variables and error variables [33].
representation of the entire control function, Equations
If the ED algorithm is used for derivation, PID control can
1-4, can be represented as the sum of the three individual
be used.
contributions [26].

​u​(t)​ = ​u​ ​(t)​+ ​u​​(t)​+ ​u​ ​(​t​)​​ (1)

P I D

​u​ (​ t)​ = ​K​ ​* e​(​t​)​​ (2)

P P

t
​u​(​ t)​ = ​K​​* ​∫ ​e​​(𝝉)​d𝝉​ (3)
I I 0

_(​t​)​
de​
​u​ (​ t)​ = ​K​ ​* ​dt ​​ (4)
D D

It is possible to find many methods in the literature

regarding the adjustment of KP, KI, and KD coefficients. Figure 13. Comparison between SNN output and experimental
The main adjustment method is the Ziegler-Nichols rule heading data for training and validation
[32]. In practically adjusting these control coefficients, SNN: Shallow neural network

69
Modelling Using Neural Networks and Dynamic Position Control for Unmanned Underwater Vehicles

Due to their effectiveness and energy efficiency, the 1 to 0. In this case, the total force calculated in the first
rudders (the vertical control surfaces for the Delphin2 layer required for the vehicle to dove in the hover-style
AUV) are frequently used for heading control during high- condition multiplied by the thruster weight coefficient can
speed operation. However, when operating in a low-speed be met by vertical thrusters in the flight-style condition. In
regime where the control surfaces lose their effectiveness addition, as the AUV’s speed increases, the control surface,
or when coping with significant heading mistakes, ws, increases from 0 to 1. To calculate the horizontal control
additional pressures from the horizontal tunnel thrusters surface deflection for compensating pitch motion, the PID
are necessary. The AUV model can be maintained at the control equations are calculated in two layers. In the first
desired equilibrium with the help of the proportional (P)- layer calculation, the pitch bias, i.e., the reference value,
derivative (D) (PD) function. Two cascade modules make is obtained by PID calculation, depending on the depth
up the heading control; the first module establishes a error of the vehicle. In the second layer, the deflection of
generalized moment for following heading demand, and the the horizontal control surfaces is calculated using the PID
second module divides the generalized moment between calculation depending on the pitch error and multiplied by
the thrusters and rudders [34]. The horizontal thrusters ws according to the speed of the vehicle. The heading control
and vertical control surfaces of the Delphin2 AUV model are problem includes two cascade modules. In the first module,
responsible for distributing the generalized yaw moment the PID controller determines the total moment depending
[6]. The control signal can be applied in two ways for depth on the heading error. In the second module, the controller
control, as Equations 5 and 6. The definitions of symbols allocates the total moment between the horizontal thrusters
are as follows; ​u​ ​, total control signal, ​u​ ​, total control and the vertical control surfaces based on the AUV speed.
c_total c_prev
signal calculated one step earlier, ​∆ u​, calculated according In addition, the moment sharing between front and rear
to PD control. Conversely, ​u​​control signal can be calculated horizontal thrusters should be calculated based on the
c
according to PID control. changing center of rotation of the vehicle [6].
In addition, the control was applied for altitude referenced
​u​ ​ = ​u​ ​+ ∆ u​ (5) 0.5 m, flight-style operation. The control signal, according
c_total c_prev

to Equation 5 and 6, was applied separately in the depth

​u​​ = ​u​ ​​ (6)
c PID

The controller includes a two-layer PID controller so that the

AUV vehicle can descend to the given reference depth while
diving at zero speed, i.e., in a hover-style state. In the first
layer, the total PD control signal, which calculates the total
force required for the vehicle to dive to the reference depth,
is calculated depending on the depth error. In the second
layer, the dynamic change in the vehicle’s longitudinal
rotation center relative to the initial rotation center due to
pitch error is calculated with PID control. Force allocation
of the vertical thrusters is achieved with the PID control Figure 14. Altitude outputs of the control result data during altitude
referenced 0.5 m, hover-style operation
calculated in the second layer [6].
The control signal, according to Equation 5 and 6, was
applied separately in the depth control of the AUV, while
the altitude referenced 0.5 m, hover-style operation. A
comparison of the two simulation results is shown in Figure
14. The result showed that the total PD control signal should
be applied for depth control because when the vehicle
reference altitude comes, the signal to the thrusters is not
reset. The PID control signal method was applied in pitch
motion control, as shown in Figure 15.
It is not possible to use the hover-style control approach
at fast forward speeds. This is because as forward speed Figure 15. Pitch angle outputs of PID control result data during
increases, thruster performance diminishes. As the AUV’s altitude referenced 0.5 m, hover-style operation
speed increases, the thruster weight, wth, decreases from PID: Proportional integral derivative

70
 Journal of ETA Maritime Science 2024;12(1):64-73

control of the AUV, while the altitude referenced 0.5 m, The depth of the AUV was measured using both an echo-
flight-style operation. A comparison of the two simulation sounding altimeter and a pressure sensor. The depth
results is shown in Figure 16. The outcome demonstrated pressure sensor was negatively affected by the vertical
that the total PD control signal should be used for depth tunnel thrusters, and its signal was very unstable and
control because the signal to the thrusters is not reset when noisy. The altimeter output signal was more stable than
the vehicle reference altitude arrives, as shown in Figure the pressure depth sensor output; therefore, the altimeter
17. The PID control signal method was applied to pitch and output signal data were used for depth-pitch control.
heading motion control. The other simulation results are The depth-pitch and PD heading motion operations were
shown in Figures 18-20. performed for each mission operation in the tank tests.
Feedback signals were collected in the tank tests from
5. Conclusions and Proposed Methods
5.1. Conclusions
An over-actuated design enabled the AUV model to perform
various missions, spanning from hover-style operation at
zero or slow speeds to flight-style operation at forward
speeds up to approximately 1 m/s. The current heading
and turning rates were obtained from the IMU at a sample
rate of 20 Hz. Surge and sway motions used in the control
systems were estimated as dead reckoning using the
dynamics model at 20 Hz to comply with the IMU, whereas
the control systems were implemented at 5 Hz so that they
Figure 18. Pitch angle outputs of PID control result data during
could synchronize to the sensor and actuator interfacing altitude referenced 0.5 m, flight-style operation
nodes. However, the localization calculation based only on
PID: Proportional integral derivative
the dynamic model, without DVL measuring the speed of
the AUV, caused drifts, and this case negatively affects the
running control algorithms.

Figure 19. Heading angle outputs of PID control result data during
altitude referenced 0.5 m, flight-style operation
Figure 16. Altitude outputs of the control result data during altitude PID: Proportional integral derivative
referenced 0.5 m, flight-style operation

Figure 20. Horizontal thruster front outputs of PID control result

Figure 17. Vertical thruster front outputs of the control result data data during altitude referenced 0.5 m, flight-style operation
during altitude referenced 0.5 m, flight-style operation PID: Proportional integral derivative

71
Modelling Using Neural Networks and Dynamic Position Control for Unmanned Underwater Vehicles

the sensors, altimeter, pressure depth sensor, and IMU, Acknowledgments

according to set values on the actuator, including the vertical The authors would like to thank Sophia SCHILLAI for her
and horizontal tunnel thrusters, the vertical and horizontal help in providing the AUV model test data.
control surfaces as the tails, and the propeller. The input-
output test data were used to form nonlinear coupled Authorship Contributions
mathematical models. The models were formed in two Concept design: M. Ertogan, Data Collection or Processing:
groups as depth-pitch motion and heading motion black- M. Ertogan, and P. A. Wilson, Analysis or Interpretation: M.
box models using the SNN algorithm. The SNN R-value, of Ertogan, and P. A. Wilson, Literature Review: M. Ertogan,
the depth-pitch motion SNN model and the SNN R-value, of and P. A. Wilson, Writing, Reviewing and Editing: M. Ertogan,
the heading motion SNN model were approximately 0.90 and P. A. Wilson.
and 0.80, respectively. This nonlinear coupled mathematical Funding: This study was supported by The Department of
model was used to develop the control design. Science Fellowships and Grant Programs (TÜBİTAK)-2219
The control process was applied for altitude-referenced International Program. Application no. 1059B191501331,
0.5 m, hover-style, and flight-style operations. The control ID: 841558.
signal, according to the total PD and PID control, was applied
References
separately in the depth control of the AUV. The outcome
[1] J. Liu, et al. Design and Control of a Flight-Style AUV with Hovering
demonstrated that the total PD control signal should be Capability: In Proceedings of the International Symposium on
used for depth control because the signal to the thrusters Unmanned Unethered Submersible Technology (UUST 2009),
is not reset when the vehicle reference altitude arrives. pages 1-9, Durham NH, USA. Autonomous Undersea Systems
The PID control signal method was applied to pitch and Institute (AUSI), 2009.
heading motion control. Good performance was obtained [2] A. B. Phillips, et al. “Delphin2: an over actuated autonomous
underwater vehicle for manoeuvring research.” Transactions
in the simulation studies. To achieve better results in
of The Royal Institution of Naval Architects Part A. International
every mission, there is a need to study an advanced control Journal of Maritime Engineering, vol. 155, pp. 171-180, 2013.
algorithm in the future.
[3] L. V. Steenson, Experimentally Verified Model Predictive Control
5.2. Proposed Methods of a Hover-Capable AUV. PhD thesis, University of Southampton,
UK, 2013.
The Delphin2 software’s dead reckoning technique results
[4] L. V. Steenson, et al. “Model predictive control of a hybrid
in AUV motion drifts. The DVL/IMU measuring system can
autonomous underwater vehicle with experimental verification,”
be employed because of less drift. Because of wall effects, Proceedings of the Institution of Mechanical Engineers, Part M:
USBL and LBL measurement equipment do not function Journal of Engineering for the Maritime Environment, vol. 228,
properly in tanks. In addition, the AUV’s localization range pp. 166-179, Feb 2014.
(about 30 cm) when using vision feedback from a laser and [5] S. M. Schillai, S. R. Turnock, E. Rogers, and A. B. Phillips,
camera is relatively small. “Evaluation of terrain collision risks for flight style
autonomous underwater vehicles,” Published in: 2016 IEEE/
To locate the AUV during tank testing, it has been OES Autonomous Underwater Vehicles (AUV), Tokyo, Japan,
recommended to integrate two or three echo-sounding 2016, pp. 311-318.
altimeters with the IMU and a pressure depth sensor on the [6] K. Tanakitkorn, Guidance, Control and Path Planning for
AUV. An AUV can measure the x- and y- axes positions when Autonomous Underwater vehicles. Ph.D. thesis, University of
two echo-sounding altimeters are added. The frequency Southampton, UK, 2017.
range of the altimeter is 1-4 Hz. If a magnetometer [7] M. Ertogan, P. A. Wilson, G. T. Tayyar, and S. Ertugrul, “Optimal
displays large drifts during tank tests, three-echo-sounding trim control of a high-speed craft by trim tabs/interceptors Part
I: Pitch and surge coupled dynamic modelling using sea trial
altimeters can also identify the yaw angle. Furthermore,
data,” Ocean Engineering, vol. 130, pp. 300-309, Jan 2017.
DVL is more expensive than an altimeter that emits an echo
[8] C. G. Karras, D. J. Panagou, and K. J. Kyriakopoulos, “Target-
[35].
referenced localization of an underwater vehicle using a laser-
In addition, an echo-sounding altimeter would be used based vision system,” in IEEE, OCEANS, Boston, MA, USA, 2006.
to maintain distance control on AUVs in real-world [9] G. C. Karras, and K. J. Kyriakopoulos, “Localization of an
circumstances, such as during the flight of AUVs in shallow underwater vehicle using an IMU and a laser-based vision
coastal and under-ice areas, as well as during AUV docking, system,” in IEEE, Mediterranean Conference on Control and
Automation, Athens-Greece, 2007.
because the USBL measurement system might not work
[10] G. C. Karras, S. G. Loizou, and K. J. Kyriakopoulos, “Towards semi-
properly close to shore. Both stationary and moving targets
autonomous operation of under-actuated underwater vehicles:
might aid in avoiding AUV collisions. sensor fusion, on-line identification and visual servo control.”
Autonomous Robots, vol. 31, pp. 67-86, May 2011.

72
 Journal of ETA Maritime Science 2024;12(1):64-73

[11] C. Cain, and A. Leonessa, “Laser based rangefinder for underwater [23] N. Miskovic, Z. Vukic, I. Petrovic, and M. Barisic, Distance keeping
applications.” in Proceedings of the American Control Conference for underwater vehicles – tuning Kalman Filters using self-
American Control Conference, (ACC), 2012 Canada, pp. 6190- oscillations, IEEE, OCEANS 2009-EUROPE, 11-14 May 2009.
6191, 2012. [24] J. Gao, C. Liu, and A. Proctor, “Nonlinear model predictive dynamic
[12] D. R. Romagos, Underwater SLAM for structured environments positioning control of an underwater vehicle with an onboard
using an imaging sonar. Ph.D. thesis, Department of computer USBL system.” Journal of Marine Science and Technology, vol. 21,
Engineering, University of Girona, Spain, 2008. pp. 57-69, Aug 2015.
[13] A. Burguera, Y. Gonzalez, and G. Oliver, “The UspIC: performing [25] S. D. McPhail, “Autosub6000: A Deep Diving Long Range AUV,”
scan matching localization using an imaging sonar,” Sensors, vol. Journal of Bionic Engineering, vol. 6, pp. 55-62, Mar 2009.
12, pp. 7855-7885, Jun 2012. [26] L. Ljung, System identification: theory for the user, Prentice Hall,
[14] A. J. Plueddemann, A. L. Kukulya, R. Stokey, and L. Freitag, Second Edition, 1999.
“Autonomous underwater vehicle operations beneath coastal [27] D. E. Rumelhart, C. E. Hinton, and R. J. Williams, “Learning
sea ice,” IEEE/ASME Transactions on Mechatronics, vol. 17, pp. Internal Representations by Error Propagation,” in Parallel
54-64, Feb 2012. Distributed Processing: Explorations in the Microstructures of
[15] D. Bandara, Z. Leong, H. Nguyen, S. Jayasinghe, and A. L. Forrest, Cognition, vol. 1, Cambridge, MA: MIT Press, 1986.
“Technologies for underwater-ice AUV navigation.” 2016 IEEE/ [28] C. C., Yu, and B. D. Liu, “A backpropagation algorithm with
OES, Autonomous Underwater Vehicles (AUV), Tokyo, Japan, adaptive learning rate and momentum coefficient.” in
2016. Proceedings of the 2002 International Joint Conference on Neural
[16] L. Chen, S. Wang, K. McDonald-Maier, and H. Huosheng, “Towards Networks. IJCNN’02 (Cat. No. 02CH37290), IEEE, 2002. vol. 2, pp.
autonomous localization and mapping ofAUVs: a survey,” 1218-1223, 2002.
International Journal of Intelligent Unmanned Systems, vol. 1, pp. [29] H. Demuth, and M. Beale, Neural Network Toolbox for use with
97-120, May 2013. Matlab. User’s Guide Version 4, 2004.
[17] L. Paull, S. Saeedi, M. Seto, and L. Howard Li, “AUV navigation and [30] M. G. Kendall, The Advanced Theory of Statistics, 4th Ed.,
localization: a review,” IEEE Journal of Oceanic Engineering, vol. Macmillan, 1979.
39, pp. 131-149, Jan 2014.
[31] J. Frost, Introduction to statistics: an intuitive guide for analyzing
[18] C.-M. Lee, P.-M. Lee, S.-W. Hong, and S.-M. Kim, “Underwater data and unlocking discoveries, Statistics by Jim Publishing,
navigation system based on inertial sensor and doppler velocity 2020.
log using indirect feedback Kalman Filter,” International Journal
[32] K. Ogata, Modern Control Engineering, Fifth Edition. Pearson
of Offshore and Polar Engineering, vol. 15, pp. 8895, Jun 2005.
Education, Inc., publishing as Prentice Hall, New Jersey, 2010.
[19] J. Davis, I. Tena, “Dynamic positioning of underwater vehicles
[33] Y. X. Su, C. H. Zheng, P. C. Mueller, and B. Y. Duan, “A simple
(tethered or not),” Dynamic Positioning Conference, Dynamic
improved velocity estimation for low-speed regions based on
Positioning Committee, Marine Technology Society, pp. 1-15,
position measurement only,” IEEEE Transactions on Control
2008.
Systems Technology, vol. 14, pp. 937-942, Sep 2006.
[20] S. Liu, D. Wang, E. K. Poh, and Y. Wang, “Dynamic positioning of
[34] T. I. Fossen, T. Johansen, “A survey of control allocation methods
AUVs in shallow water environment: obeserver and controller
for ships and underwater vehicles,” in 2006 14th Mediterranean
design,” in Proceedings of the IEEE/ASME, International
Conference on Control and Automation, IEEE, 2006, pp. 1-6.
Conference on Advanced Intelligent Mechatronics, Monterey,
California, pp. 705-710, 2005. [35] M. Ertogan, G. T. Tayyar, P. A. Wilson, and S. Ertugrul, “Marine
measurement and real-time control systems with case studies,”
[21] A. P. Aguiar, and A. M. Pascoal, “Dynamic positioning and way-
Ocean Engineering, vol. 159, pp. 457-469, 2018.
point tracking of underactuated AUVs in the precence of ocean
currents,” in Proceedings of the 41st IEEE, Conference on Decision
and Control Las Vegas, Nevada USA, pp. 2105-2110, 2002.
[22] S. Zhao, Advanced control of autonomous underwater vehicles,
PhD thesis, The University of Hawai, USA, 2004.

73