743

1 INTRODUCTION

Although the concept of artificial neural networks is

known starting from the past century, nowadays it

found a new breath due to modern computer

technology development: software/hardware

resources and capabilities. Artificial neural networks

are related to many disciplines: neurophysiology,

mathematics, statistics, physics, computer science and

engineering. They find their application in various

fields such as modeling, time series analysis, pattern

recognition, signal processing and control due to their

data learning ability.

The purpose of the current research is to review and

evaluate the usage of digital neural networks (DNN) in

the context of the maritime industry by analyzing

existing studies to propose new concepts in this matter.

2 ANALYSIS OF EXISTING MODELS AND

RESEARCH METHODOLOGY

Neural networks are one of the areas in the field of

artificial intelligence, based on attempts to reproduce

the human nervous system, namely: the ability to learn

and correct errors, which should allow simulating,

albeit rather roughly, the work of the human brain. A

neuron is a base element of DNN. The basic structural

scheme of the network’s j – neuron is presented in

figure 1.

Figure 1. Neuron model

Application Perspective of Digital Neural Networks in

the Context of Marine Technologies

V.Konon & N. Konon

National University “Odessa Maritime Academy”, Odessa, Ukraine

ABSTRACT: This study is focused on the issue of digital neural networks’ implementation in the context of

maritime industry. Various algorithms of such networks in the terms of the marine technologies have been

reviewed in the current study in order to evaluate the effectiveness of the methodology and to propose a new

concept of an artificial neural network’s application in this way. Fire-detection system simulation based on the

thermal imagers’ data input had been developed to assess the efficiency of the concept suggested with a multi-

layer perceptron (MLP) algorithm integrated into the designed 3d-model.

http://www.transnav.eu

the International Journal

on Marine Navigation

and Safety of Sea Transportation

Volume 16

Number 4

December 2022

DOI: 10.12716/1001.16.04.16

744

Different studies already proposed and/or

reviewed the concept of DNN models, which are

applied or may be applied to the maritime industry.

Autonomous ship control and vessel motion

prediction present a widely pursued area of research

for already more than 30 years. Autonomous

shipboard navigation in channels was considered in

[18] using a neural network consisting of an adaptive

critic element (ACE) and an adaptive search element

(ASE). ASE was responsible for studying the channel

area while ACE evaluated the ASE performance,

tending to predict potential failures in navigation. The

developed model showed good results through

software simulation with graphical feedback.

A combining density-based spatial clustering of

applications with noise (DBSCAN)-based Long Short-

Term Memory (LSTM) model, as an extension of

recurrent neural network (RNN), was developed for

vessel trajectory prediction, where DBSCAN was used

to cluster vessel tracks from automatic identification

system (AIS) data, and LSTM was used for training and

prediction [21]. The resulting predictions of the

proposed model showed a 2% advantage over other

models in the study. The nonlinear autoregressive

exogenous (NARX) network was used as the core of the

ship motion prediction framework in the study [7]. The

advantage of the NARX network’s long-term

prediction ability a multi-step-ahead prediction was

realized under the hybrid learning strategy and

experimental results showed the efficiency of the

NARX network in generating the data-driven model

for ship motion prediction. The use of multi-layered

perceptron (MLP) was discussed in the study [22] for

vessel motion prediction in specified conditions,

showing good agreement with the expected results. To

enhance optimization and multi-criterial automatic

design of ships in terms of fast and simultaneous

analysis of many design solutions within a similar time

interval, the work [2] presented a neural network

implementation for the assessment of ship

manoeuvrability qualities. The back-stepping method

using a neural network to approximate the nonlinear

ship course control system was used in research [23]. A

neural network was utilized to approximate nonlinear

terms in the ship model. To verify the effectiveness of

the proposed control algorithm MATLAB simulations

were carried out.

NN application perspectives were also reflected in

the mooring lines’ condition monitoring. In this way,

the work [12] focused on the development of a DNN-

based damage detection approach with floater

responses for mooring lines with even local damage in

the tension leg platform. Simulation data was used for

DNN training, validation, and testing. Another study

[9] used RNN for effective analysis of the time-series

response data employed for damage detection. The

results of the RNN-based catenary mooring line

damage detection approach proposed confirmed the

efficiency of the RNN model in comparison with

results obtained in [12]. Another research [3]

highlighted the issue of the mooring lines’ complex

behavior and a Radial Basis Function (RBF) neural

network was proposed for damage with a modelling

method based on Rod theory and the Finite Element

Method (FEM). In order to improve the modelling

accuracy, boundary conditions uncertainty was

applied using Submatrix Solution Procedure (SSP) and

a round-off error is removed by SSP. The described

method showed 65% higher performance than the

Fuzzy method and 13% percent higher performance

than the conventional one.

The use of neural networks is also being actively

integrated into the development of autonomous ship

berthing control. In the following works [3-6] an

automatic berthing controller from the neural network

technology was designed via virtual window theory

algorithms, nonlinear programming, as well as the PD

hybrid control. The effectiveness of the model was

verified from the free-running model test. In [16], the

artificial neural network algorithm was used to

establish an automatic berthing model and the berthing

process of the ship was divided into two stages: the

proceeding and deceleration stage, and the turning and

berthing stage. The paper [7] highlights four major

challenges for the implementation of ANN controller

for ship berthing. The problem of auto-berthing control

was considered in [17], where a neural network (NN)

adaptive approach based on the navigation dynamic

deep-rooted information (DRI) was proposed to

resolve the uncertainties caused by unknown ship

dynamics and external disturbances. The dynamic

surface control (DSC) and the minimum learning

parameter (MLP) techniques were used to minimize

the computational load of the adaptive NN control

scheme. Simulations were carried out on a vessel to

verify the effectiveness of the proposed model. The

study [13] is devoted to the development of an

automatic ship`s berthing system, by using the RNN

for a non-linear optimal feedback controller. The

Proposed system was evaluated through extensive

computer simulations and an actual experiment was

carried out. MLP in conjunction with the back-

propagation algorithm was applied in a study [1] to

develop a feed-forward controller using a non-

simplified mathematical model for modelling an

automated ship berthing controller.

Several works, such as [10, 14, 19], present the

models related to fire-detection systems implementing

various DNN algorithms. Research [10] describes a

real-time fire system based on grey-fuzzy algorithms.

A fire detection algorithm, which uses a pyroelectric

infrared motion sensor (PIR) and DNN is proposed in

[19]. Data collected from the sensor was processed for

further DNN training. According to the described

study, such a model showed efficient results during

tests and was capable to detect not only fire but also

human motion.

In the study [14] fire datasets from laboratory

experiments are presented by authors for further

processing of these datasets using machine learning

techniques.

The current study is focused on the DNN

application in shipboard fire-detection systems,

namely, in the processing of thermal data received

from thermal imaging devices. Cargo fire simulation

inside the container vessel’s cargo hold was modelled

for further experiments and data collection within the

task. As the cargo holds configuration, especially cross-

decks’ arrangement, may differ depending on a

vessel’s design, and there is a possibility, that some

difficulties may be encountered at the imagers’

allocation stage (e.g. blind zones), an artificial neural

network may provide a solution for containers’

745

thermal condition control over the cargo hold,

estimating areas of an issue. The allocation diagram

with positions of thermal imagers, correspondent

numbers of containers, bulkheads and blind zones is

presented in figure 2. Simulation has been carried out

using instruments of Unity IDE and C# programming

language capabilities.

As the DNN for data processing, a multi-layer

perceptron has been chosen to evaluate the efficiency

of the formulated idea in the first approximation. In

this way, such an evaluation could be stated as an

objective within the purpose of the current research.

Figure 2. Allocation diagram sample

2.1 DNN implementation

Multi-layer perceptron could be applied for solving a

wide variety of tasks, herewith its training may be

carried out by means of error back-propagation

algorithm [8, 20]. Such an algorithm in its turn is based

on the error-correction learning rule. This method of

training supposes forward and backward passes

through all the DNN’s layers (see fig. 3).

Figure 3. Forward and backward pass

During the forward pass an input data vector is sent

to the sensor nodes with further spreading among the

network, resulting in a set of output signals, which may

be interpreted as a network reaction onto the initial

input. An error signal is generated by the deduction of

actual output from the desired one, setting up all the

synaptic weights in accordance with the error-

correction rule during the backward pass.

Consequently, it is spreading in the direction opposite

to one of the synaptic connections.

A schematic diagram of the network being used

during the current experiment is presented in figure 4.

Figure 4. Multi-layer perceptron model

The input layer of the DNN used in the research

consists of an array with a length of fourteen elements.

These elements are seven maximum temperature

values of containers in the simulated thermal imagers’

field of view (FOV), allocated in accordance with the

defined pattern, and their correspondent locations (i.e.

seven temperatures plus seven container numbers,

distributed in a sequence from left to right and from up

to down as shown in fig. 2).

To normalize data input for further processing next

equation has been used:

min 2 1

max min

( ) ( )

x x d d

−  −

−

(1)

where

ni – normalized data,

x – data to be normalized,

[xmin; xmax] – interval of x values,

[d1; d2] – interval of normalized values.

Actual data processing is carried out inside of

hidden layers. There are four layers of this type that

were set for the current task. Quantities of layers and

neurons are defined in an experimental way to show

the most accurate outputs.

Sigmoidal nonlinearity has been used as an

activation function and it is determined as a logistic

one:

1 exp( )

+−

(2)

where

yj – neuron’s output;

vj – weighted sum of all synaptic inputs for neuron j

plus neuron’s threshold value.

For the training purpose, the output data was set as

an array with a length of two elements: the maximum

temperature value in the scene and its location. The

minimum temperature value for the experiment was

defined as the value of zero degrees Celsius and

considered as sufficient within the framework of the

task. The DNN was integrated into the 3d simulation

model in order to perform the experiment.

746

2.2 Simulation modelling

The designed simulation consists of 3d-container

models (see figure 5) with a fire source randomly

located inside one of them. Each side of the container is

divided into a specified quantity of segments. The color

of each segment depends on the current temperature

value inherent to it. Temperature intervals for the color

palette may be set manually by the operator. Detailed

heat exchange simulation between the segments is a

matter of separate research and a linear dependency at

each frame is considered as sufficient within the set

task.

Figure 5. Container 3D-model

To achieve the set objective, the next initial

conditions have been applied to the model in order to

select only necessary and sufficient data for the

fulfillment of the experiment:

− only 40-feet containers are used in the simulation;

− cargo containers are considered empty to avoid any

additional issues, which may be assumed as

unnecessary within the stated task, except the one

carrying the fire source;

− containers’ quantity of 40 units is considered

sufficient to provide a thorough examination of the

simulation process;

− thermal imager’s allocation should ensure the most

effective coverage over the row-tier area;

− thermal imagers are allocated in the cross-deck

areas of the cargo hold (see figure 6).

It should be clear, that the color palette of the

temperature distribution, shown in figure 6, is

demonstrated for better visualization of the process

and it does not fully reflect a real thermal image.

Figure 6. Allocated camera view sample

3 RESULTS OVERVIEW

During the experiment, it was found, that readings of

at least three thermal imagers in the vicinity of the fire

source are necessary to obtain adequate results. Their

referent positions are also important in this matter and

the most accurate results have been shown when the

desired container had been enclosed in a triangle with

vertices inside of the sensors’ FOV.

As the DNN’s output values belong to the interval

[0; 1], their standard form may be obtained by

expressing x from (1):

1 max min

min

( ) ( )

n d x x

−  −

−

(3)

Training of the MLP was carried out using about

seventy-five data sets received from the developed 3d-

model (three data sets of various temperature

distributions at different time points per twenty-five

containers).

The results of five measurement series are brought

to the table 1. Presented cases have been chosen as

those which fairly reflect the outputs of the experiment

in terms of summarizing the unique conditions for the

placement of ignition sources in the context of data

processing within the framework of the task. In four of

five presented cases a temperature error does not

exceed the value of five degrees Celsius (corresponds

to about 90% of all cases, i.e. including those not shown

in table 1), and, in general, values do not exceed the

limit of ten degrees. As the importance of referent

positions has already been mentioned, in this way it

should be noted that only in three of five presented

series the DNN was able to accurately define desired

fire source position located inside of blind zones

(corresponds to about 71% of all cases, i.e. including

those not shown in table 1). In the rest of series,

position’s determination accuracy of blind zone

container did not exceed the value of two units in any

direction.

Table 1. Processing results

________________________________________________

№ Parameters Measurement spots Fire DNN's

Source output

________________________________________________

1 Imager's 1 3 4

number

Temperature, 150.38 78.56 150.11 165.00 167.33

°C

Container 0 10 16 8 8

number

2 Imager's 4 7 6

number

Temperature, 127.57 126.78 64.76 138.00 133.02

°C

Container 16 34 27 32 26

number

3 Imager's 6 5 7

number

Temperature, 98.76 80.18 80.00 112.00 103.79

°C

Container 30 23 37 39 30

number

4 Imager's 3 2 1

number

Temperature, 110.34 109.98 101.75 120.00 124.13

°C

Container 12 5 2 4 4

number

5 Imager's 7 6 4

number

Temperature, 180.31 180.45 181.10 214.00 218.38

°C

Container 34 27 18 26 26

number

________________________________________________

747

4 CONCLUSIONS AND DISCUSSION

In the current research various artificial neural

networks’ algorithms were reviewed in the terms of

their application in maritime industry: navigation,

mooring operations, motion prediction, fire-fighting

systems and etc. in an efficient way, and a wide variety

of such an algorithms found their implementation into

marine technologies: multi-layer perceptron, recurrent

neural networks, DNNs on fuzzy algorithms etc. In this

way, the usage of multi-layer perceptron in the concept

of shipboard fire-fighting system based on thermal

imagers’ data has been proposed, performing the

simulation modelling for the experiment.

The main objective of the current research is to

evaluate the effectiveness of designed model in the first

approximation by carrying out the fire source

determination with its temperature value within the

equipment’s limits considering the blind zones of the

allocated imagers’ FOVs. Experiment results have

demonstrated that the temperatures of the determined

by MLP fire sources may be obtained with the

sufficient accuracy (less than ± 10°C) with position

prediction not exceeding the value of two units in any

direction (of a “row-tier” area). In order to enhance the

presented concept, namely, to increase the

performance accuracy and to provide its application

for various loading conditions and cargo holds’

configurations more data should be collected and

processed in order to train the DNN sufficiently.

Taking into account safety and economic issues

inherent to the conducting of such experiments, in

order to improve the proposed concepts’ performance,

the enhancing of the designed 3d-model is considered

necessary.

REFERENCES

1. Abdelali K, Yousra M, Khalifa M, Mostafa R. Artificial

neural network and mathematical modeling of automatic

ship berthing, Commun. Math. Biol. Neurosci. 2022;

Article ID 113. https://doi.org/10.28919/cmbn/7727

2. Abramowski T. Application of artificial neural networks to

assessment of ship manoeuvrability qualities. Polish

Maritime Research. 2008; 15(2) 1521.

https://doi.org/10.2478/v10012-007-0059-0

3. Ahmed Y. A., Hasegawa K. Automatic Ship Berthing using

Artificial Neural Network Based on Virtual Window

Concept in Wind Condition. IFAC Proceedings Volumes.

2012; 45(24), 286–291. https://doi.org/10.3182/20120912-3-

bg-2031.00059

4. Ahmed Y. A., Hasegawa K. Automatic ship berthing using

artificial neural network trained by consistent teaching

data using nonlinear pro-gramming method. Engineering

Applications of Artificial Intelligence. 2013; 26(10), 2287–

2304. https://doi.org/10.1016/j.engappai.2013.08.009

5. Ahmed Y. A., Hasegawa K. Implementation of Automatic

Ship Berthing using Artificial Neural Network for Free

Running Experiment. IFAC Proceedings Volumes. 2013;

46(33), 25–30. https://doi.org/10.3182/20130918-4-jp-

022.00036

6. Ahmed Y. A., Hasegawa K. Consistently Trained Artificial

Neural Network for Automat-ic Ship Berthing Control.

TransNav, the Interna-tional Journal on Marine

Navigation and Safety of Sea Transportation. 2015; 9(3),

417–426. https://doi.org/10.12716/1001.09.03.15

7. Ahmed Y. A., Hannan M. A., Siang K. H. Arti-ficial Neural

Network controller for automatic ship berthing:

challenges and opportunities. Ma-rine Systems &

Ocean Technology. 2020; 15(4), 217–242.

https://doi.org/10.1007/s40868-020-00089-x

8. Haykin Simon, “Neural networks and learning machines,”

—3rd ed, Rev. ed of: Neural net-works. 2nd ed., 1999.

Includes bibliographical references and index. ISBN-13:

978-0-13-147139-9 ISBN-10: 0-13-147139-2

9. Kanghyeok L., Minwoong C., Seungjun K., Do H. S.

Damage detection of catenary mooring line based on

recurrent neural networks. Ocean Engineering. 2021; 227,

108898, ISSN 0029-8018.

https://doi.org/10.1016/j.oceaneng.2021.108898

10. Kuo H. C., Chang, H. K. A real-time shipboard fire-

detection system based on grey-fuzzy algorithms. Fire

Safety Journal. 2003; 38(4), 341–363.

https://doi.org/10.1016/s0379- 7112(02)00088-7

11. Li G, Kawan B, Wang H, Zhang H. Neural-network-based

modelling and analysis for time series prediction of ship

motion. Ship Technolo-gy Research. 2017; 64(1), 3039.

https://doi.org/10.1080/09377255.2017.1309786

12. Minwoong C., Seungjun K., Kanghyeok L., Do H.S.

Detection of damaged mooring line based on deep neural

networks. Ocean Engineering. 2020; 209, 107522, ISSN

0029-8018. https://doi.org/10.1016/j.oceaneng.2020.10752

13. Mizuno N., Kuboshima R. Implementation and

Evaluation of Non-linear Optimal Feedback Control for

Ship’s Automatic Berthing by Re-current Neural

Network. IFAC-PapersOnLine. 2019; 52(21), 91–96.

https://doi.org/10.1016/j.ifacol.2019.12.289

14. Nazir A, Mosleh H, Takruri M, Jallad A-H, Alhebsi H.

Early Fire Detection: A New Indoor Laboratory Dataset

and Data Distribution Anal-ysis. Fire. 2022; 5(1):11.

https://doi.org/10.3390/fire5010011

15. Neumann, T. (2017). Fuzzy routing algorithm in

telematics transportation systems. Communications in

Computer and Information Science. 715, 494-505

doi:10.1007/978-3-319-66251-0_40

16. Qiang L., Bi-Guang H. Artificial Neural Net-work

Controller for Automatic Ship Berthing Using Separate

Route. Journal of Web Engineer-ing. 2020;

https://doi.org/10.13052/jwe1540-9589.19788

17. Qiang Z., Guibing Z., Xin H., Renming Y. Adaptive neural

network auto-berthing control of marine ships. Ocean

Engineering. 2019; 177, 40–48.

https://doi.org/10.1016/j.oceaneng.2019.02.031

18. Starnenkovich M. An application of artificial neural

networks for autonomous ship navigation through a

channel. Vehicle Navigation and In-formation Systems

Conference. 1991;

https://doi.org/10.1109/vnis.1991.205794

19. Xavier K.L.B.L., Nanayakkara V.K. Develop-ment of an

Early Fire Detection Technique Us-ing a Passive Infrared

Sensor and Deep Neural Networks. Fire Technol. 2022;

58, 3529–3552. https://doi.org/10.1007/s10694-022-01319-x

20. Xiao Perry, “Practical Java Programming”. Indi-anapolis,

IN: John Wiley & Sons, Inc., 2019, ISBN: 978-1-119-56001-

21. Yang C-H, Lin G-C, Wu C-H, Liu Y-H, Wang Y-C, Chen

K-C. Deep Learning for Vessel Tra-jectory Prediction

Using Clustered AIS Data. Mathematics. 2022;

10(16):2936. https://doi.org/10.3390/math10162936

22. Zăgan R., Chiţu M. G., Manea E. Ship Manoeu-vrability

Prediction Using Neural Networks Analysis. Advanced

Materials Research. 2014; 1036, 946–951.

https://doi.org/10.4028/www.scientific.net/amr.1036.946

23. Zhang Q., Jiang N., Hu Y., Pan D. Design of Course-

Keeping Controller for a Ship Based on Backstepping and

Neural Networks. Interna-tional Journal of E-Navigation

and Maritime Economy. 2017; 7, 34–41.

https://doi.org/10.1016/j.enavi.2017.06.004