https://doi.org/10.31449/inf.v48i13.6190 Informatica 48 (2024) 67–80 67 
Impacts of Crew Training Quality Evaluation System Based on 
GA-BP Algorithm on Transportation Safety 
 
Xicai Ma 
Modern Vocational Education Research Center, Dalian Vocational and Technical College, Dalian 116021, China 
E-mail: maxicaidl@163.com 
Keywords: shipping safety, crew education and training, genetic algorithm, back propagation, genetic algorithm-back 
propagation 
Received: May 14, 2024 
Nowadays, shipping safety accidents occur frequently. Crew education and training can reduce 
shipping accidents to a certain extent. However, currently, there is no established method to investigate 
the specific relationship between crew education and training and shipping safety. To solve the 
problem, the study utilizes a genetic algorithm (GA) with strong global search capability and 
scalability to improve the low learning efficiency, slow convergence speed, and local optimal problems 
in the back propagation (BP) neural network, resulting in the genetic algorithm-back propagation 
(GA-BP) algorithm. The relevant indicators in crew education and training are input into the 
algorithm to predict the safety of shipping. The performance comparison experiments of GA-BP , 
particle swarm optimization-back propagation algorithm particle swarm optimization-back 
propagation (APSO-BP) algorithm and traditional BP neural network are carried out. The 
experimental outcomes showed that the hit rate of GA-BP was 90%, and its mean absolute error was 
0.00152, the mean square error was 0.00323, and maximum absolute percentage error was 5.3%, 
which were better than APSO-BP and traditional BP neural network. In addition, in the empirical 
analysis, the output outcomes of psychological quality and adaptability were 0.92 and 0.93, 
respectively. The above outcomes show that the optimized GA-BP has better predictive performance, 
and the specific effects of psychological quality and adaptability in crew education and training on the 
safety of shipping can be obtained. This method not only provides data support for the improvement of 
crew education and training methods, but also provides a new method for predicting shipping safety 
Povzetek: Opisan je vpliv sistema ocenjevanja kakovosti usposabljanja posadke, ki temelji na 
algoritmu GA-BP , na varnost v transportu. Rezultati kažejo, da GA-BP izboljša napovedno točnost z 
večjo natančnostjo in manjšimi napakami v primerjavi z drugimi metodami, kar prispeva k izboljšanju 
metod usposabljanja posadke in večji varnosti v pomorstvu.
1 Introduction 
With the frequent occurrence of ship losses and pollution 
accidents in the maritime field, the causes of accidents 
have attracted the attention of relevant organizations [1]. 
After investigation by relevant organizations, it was 
found that 80% of the accidents were caused by humans. 
Therefore, the quality of crew members has attracted 
widespread attention from the international community 
[2]. The quality of the crew depends largely on the 
education and training of the crew, so the quality of the 
education and training for crew members has a significant 
impact on shipping safety [3]. However, at present, there 
is a lack of quality evaluation models for crew education 
and training. Therefore, it is necessary to explore a model 
that can accurately evaluate the quality of crew education 
and training. With the rapid development of neural 
network, back propagation (BP) neural network is widely 
used in various fields [4]. However, the traditional BP 
network has slow convergence speed and low accuracy.  
 
The genetic algorithm (GA) can be used for BP neural 
network to solve slow convergence speed [5]. Therefore, 
the research combines BP and GA to form a GA-BP 
model, and applies the GA-BP model to shipping safety 
prediction. Through this model, the evaluation indicators 
of the quality of crew education and training are 
evaluated to predict the shipping safety, and improve the 
dilemma of lacking an evaluation system for the quality 
of crew education and training. It is expected to improve 
relevant indicators through prediction results to enhance 
shipping safety. 
2 Literature review 
In recent years, with the rapid development of BP neural 
network, its application in many fields is deepening. 
Aiming at the low container transportation efficiency, 
Zhu W et al. proposed a container inter-modal 
transportation evaluation model based on BP neural 
network. The experimental outcomes showed that the 

68   Informatica 48 (2024) 67–80                                                                       X. Ma 
evaluation effect of the model was accurate, which 
effectively improved the efficiency of container 
inter-modal transportation [6]. To solve the inaccurate 
digital currency risk prediction methods, Shao et al. 
proposed an investment risk model that combined 
information entropy and BP neural network. The model 
could change the weight to carry out weighted scoring of 
digital currency to increase the accuracy. The results 
showed that the model could accurately predict the 
development prospects of digital currency, effectively 
helping investors avoid investment risks [7]. Liu et al. 
designed a BP neural network model for thermal error 
calculation of five-axis machining centers. The 
relationship between radial error and axial error during 
five-axis machining was calculated. The prediction 
accuracy of the error was high, which was more effective 
in practical applications [8]. Aiming at the insufficient 
urban evaluation model, Xi et al. proposed a new model 
by combining BP neural network and analytic hierarchy 
process. The separation index of urban solid waste was 
established, which was applied to the classification and 
evaluation of urban domestic waste. The evaluation 
outcomes showed that the model could effectively predict 
the urban garbage classification ability in the actual 
environment, providing a new way for urban evaluation 
to improve related capabilities [9]. Zhou et al. proposed a 
BP neural network based on monotonic constraints to 
solve the difficulty in controlling the phosphorus content 
at the end of the dephosphorization converter. The model 
was empirically analyzed. Phosphorus content in algae 
provided a control aid [10]. 
In addition, there are more methods applied in the 
field of shipping safety. In order to improve the shipping 
safety, Zhou et al. proposed a USV decision support 
model to improve the accuracy of collision avoidance 
decisions. The empirical test of the model showed that 
the model judgment in the collision stage was accurate, 
and the ship could take direct and effective collision 
avoidance measures, which effectively improved the 
safety of shipping [11]. Nosov et al. proposed a model 
based on navigation decision tree principle to address the 
insufficient safety of maritime navigation. A new type of 
ship control system was developed based on the 
prediction results of the model. The test results showed 
that it reduced the probability of critical situations by 
18-54%, effectively reducing emergency risks [12]. To 
solve the new challenges related to navigation safety 
caused by the low maneuverability of ships and the 
increasingly dense maritime traffic, Shilov proposed an 
operation method aimed at designing a navigation safety 
recommendation system. The recommendation system 
operation method updated the feedback of the system 
knowledge base. By comparing the knowledge before and 
after, the security of navigation was effectively improved 
[13]. Hasanspahi et al. proposed a risk assessment matrix 
based on the grounding risk assessment combined with 
factors such as ship speed, hull quality, and loading 
conditions in order to solve the problem that oil tankers 
were easily grounded in narrow waterways and other 
specific navigation areas. The matrix could effectively 
evaluate the navigation safety of oil tankers in narrow 
waterways, which had important practical significance 
[14]. The results and limitations of the above research 
content are shown in Table 1. 
 
 
Table 1: Results and limitations of the research content 
Reference Methodology Model Application Key findings Limitation 
Zhu et al. [6] 
BP neural 
network 
Transport 
accurate 
evaluation 
model 
Container 
transport 
Improve the 
efficiency of 
container 
inter-modal 
transportation 
Limited to 
container 
shipping only 
Shao et al. [7] 
Joint information 
entropy and BP 
neural network 
Investment risk 
model 
Digital currency 
risk prediction 
Accurately predict 
the development 
prospects of digital 
currency 
There is 
greater 
uncertainty 
Liu et al. [8] 
BP neural 
network 
BP neural 
network model 
for thermal 
error 
calculation 
Five-axis 
machining 
center thermal 
error calculation 
Improve the 
accuracy of error 
prediction 
Weak 
anti-interferen
ce 
Xi et al. [9] 
BP neural 
network and 
hierarchical 
analysis 
Evaluation 
model of 
household 
waste 
classification 
Urban garbage 
classification 
evaluation 
Accurately predict 
the garbage 
classification ability 
of cities in the 
actual environment 
Large 
calculation 
time 
complexity 
Zhou et al. 
[10] 
Monotonically 
constrained, BP 
neural network 
Phosphorus 
content control 
model 
Control of 
phosphorus 
content in 
Accurate control of 
the phosphorus 
content 
Only for 
phosphorus 
content 

Impacts of Crew Training Quality Evaluation System Based on… Informatica 48 (2024) 67–80 69 
dephosphorizati
on converter 
control 
Zhou et al. 
[11] 
USV 
Decision 
support model 
Shipping safety, 
collision 
avoidance 
Improve the 
shipping safety 
Information is 
difficult to 
obtain 
Nosov et al. 
[12] 
Navigation of the 
decision tree 
principle 
New ship 
control system 
Maritime 
navigation 
safety 
Effectively reduce 
the risk of 
emergency 
situations 
It's difficult to 
operate 
Shilov [13] 
Navigation 
security 
Safety 
recommended 
system 
operation 
method 
Navigation 
safety in dense 
maritime traffic 
Improve navigation 
safety 
Information 
overload 
Hasanspahi et 
al. [14] 
Stranded risk 
assessment 
Risk 
assessment 
matrix 
Navigation 
safety of oil 
tankers in 
narrow 
waterways 
Effectively assess 
navigation safety in 
narrow waterways 
Subjectivity 
and 
uncertainty 
 
The above research shows that BP neural network 
has been widely used in many fields. Methods applied to 
the field of shipping safety are also emerging one after 
another. However, the research on applying BP neural 
network algorithm to the field of shipping safety is still 
relatively lacking. Therefore, this study applies BP neural 
network to predict shipping safety risks, and improves the 
BP neural network through GA. Through a 
comprehensive evaluation, it is expected to help identify 
deficiencies in crew training and make improvements to 
improve shipping safety. 
 
3 Evaluation system of crew 
education and training quality 
based on BP neural network 
 
3.1 BP neural network and its improved 
algorithm 
With the in-depth understanding of neural network theory, 
more neural networks are applied in real life. Currently, 
BP neural network is one of the most successful and 
common neural networks [15]. Its structure is similar to a 
multi-layer perceptron, and its principle is to forward the 
input signal, and then adjust the weights by back 
propagating the error signal, thereby reducing the model 
error [16]. The main function of BP neural network is to 
classify samples and predict outcomes, etc. The 
traditional structure diagram of the BP neural network is 
shown in Figure 1 [17]. 
 

70   Informatica 48 (2024) 67–80                                                                       X. Ma 
Input Layer
Hidden Layer
Output Layer
X
1
X
2
X
3
Y
1 Y
2 Y
3
 
Figure 1: Structure diagram of traditional BP neural network 
 
As shown in Figure 1, the structure is the traditional 
BP neural network. ( 1,2, , )
n
x n k = is the input value, 
and ( 1,2, , )
n
y n k = is the output value. 
1 k
N
−
 denotes 
the number of nodes of the neurons. In the forward 
propagation of the neural network, each neuron will 
summarize the weighted information, and its expression 
is shown in equation (1). 
 
 
1
( 1) ( 1)
1
k
N
kl k j k jl
j
net O W
−
−−
=
=

 (1) 
 
In equation (1), 
kl
net represents the k aggregated 
weighted information of the 
( 1) kj
O
−
-th neuron in the 
j
-th layer . l represents the 1 k − output value of the 
1 k − -th neuron in the k -th layer 
j
. 
( 1) k jl
W
−
 is the 
connection weight from the 
j
-th neuron in the 
j
-th 
layer to the k -th neuron in the 
j
-th layer. In the 
forward propagation, in addition to summarizing the 
weighted information, each neuron will also pass the 
summed information through the activation function and 
output, and its expression is shown in equation (2). 
 
 
1
()
1 exp[ ( )]
kl s kl
kl kl
O f net
net 
==
− − −
 (2) 
 
In equation (2) 
kl
 represents the k threshold of 
the first neuron in the l -th layer. In the BP neural 
network, the mean square error is selected as the index to 
judge its performance. The algorithm back propagates the 
difference between the output outcome 
l
y and the 
expected output 
l
d to correct the weights of neurons in 
each layer. The expression of the objective function is 
shown in equation (3). 
 
 
2
1
()
2
ll
l
E d y =−

 (3) 
 
In equation (3) 
ll
dy − represents the difference 
between the output outcome and the expected output, that 
is, the back propagated error signal. The BP network uses 
the gradient descent method to correct the neuron weights, 
and its expression is shown in equation (4). 
 
 klj
klj
E
W
w


 = −

 (4) 
 
In equation (4), 

 represents the learning step size. 
klj
W 
 indicates the weight correction between the 
j
 
neuron and the k neuron in the l layer. The 
relationship 
klj
W 
 between it and the neuron output is 
calculated using partial derivatives, and its specific 
expression is shown in equation (5). 
( 1) ( 1)
( 1)
k j k j
klj kj
klj k j klj klj
net net
EE
W
w net w w
   
++
+


 = − = − =
   
 (5) 
In equation (5) 
kj

 is the weight correction value 
between the 1 k + and 
j
 neurons. 
( 1)
kj
kj
E
net

+

=−

. 
The first part of equation (5) is solved, and the solution 
process is shown in equation (6). 

Impacts of Crew Training Quality Evaluation System Based on… Informatica 48 (2024) 67–80 71 
 
( 1)
1
()
k
N
kj
kh khj kl
h
klj klj
net
O W O
ww
+
=


==

 (6) 
 
In equation (6) 
k
N represents the number of nodes 
of the k neurons in the first layer. In the right side of 
equation (6), 
kl
O represents the output value of the k 
neuron in the l layer. Equation (7) can be obtained from 
equation (5) and (6). 
 
 klj kj kl
klj
E
WO
w
  

 = − = −

 (7) 
 
In equation (7), 
kj

 is shown in equation (8). 
 
( 1)
( 1) ( 1) ( 1)
kj
kj
k j k j k j
O
EE
net O net

+
+ + +


= − = −
  
 (8) 
 
In equation (8)
( 1)
( 1)
( 1)
'( )
kj
kj
kj
O
f net
net
+
+
+

=

. The 
expression for derivation is shown in equation (9). 
 
()
( ) 2
'( ) ( )[1 ( )]
[1 ]
x
x
e
f x f x f x
e


−−
−−
= = −
+
 (9) 
 
The calculation process shown in equation (10) can 
be obtained by equation (9). 
 
( 1)
( 1) ( 1) ( 1) ( 1)
( 1)
'( ) ( )(1 ) (1 )
kj
k j k j k j k j
kj
O
f net f net f O O
net
+
+ + + +
+

= = − = −

(10) 
 
When 
( 1) kl
O
+
 is a hidden layer node, the expression 
of 
( 1) kj
E
net
+


 and 
kj

 is shown in equation (11). 
22
2
( 2)
( 1) ( 1)
11 ( 1) ( 2)
( 1) ( 1) ( 1) ( 1)
1
(1 )
kk
k
NN
kh
k h k jh
hh k j k h
N
kj k j k j k h k jh
h
net
EE
w
net net O
O O w


++
+
+
++
== ++
+ + + +
=
 
= = −

  



=−




(11) 
When 
( 1) kl
O
+
 is an output layer node, the 
expression of 
( 1) kj
E
O
+


 and 
kj

 is shown in equation 
(12). 
( 1)
( 1) ( 1)
( ) (1 ) ( ) (1 )
jj
kj
kj j j k j k j mj mj mj
E
yd
O
d y O O d O O O 
+
++
 
=−




= − − = − −

(12) 
Through the calculation, the weight adjustment 
equation of the algorithm can be obtained, as shown in 
equation (13). 
 
2
( 1) ( 1) ( 1) ( 1)
1
( ) (1 ) i+1 layer is the output layer
(1 ) ( ) i+1 layer is a hidden layer
k
j j j j kl
N
klj
k j k j k h k jh kl
h
d y y y O
W
O O w O


+
+ + + +
=
−− 

= 
−



(13) 
 
Similarly, the threshold adjustment equation of the 
BP neural network can be obtained, as shown in equation 
(14). 
 
2
( 1) ( 1) ( 1) ( 1)
1
( ) (1 ) i+1 layer is the output layer
(1 ) ( ) i+1 layer is a hidden layer
k
j j j j
N
kj
k j k j k h k jh
h
d y y y
O O w



+
+ + + +
=
−− 

=

−



(14) 
 
Although the current BP neural network has 
extensive applications, it still has many shortcomings, 
such as long training time and easy to fall into local 
minima [18]. Therefore, the research uses GA to optimize 
the BP neural network. The GA can optimize the weights 
and thresholds of the BP network by using its strong 
search ability to increase its convergence speed and 
optimize its performance [19]. The optimization process 
of the GA is actually to iterate the individuals in the 
algorithm, judge the fitness of the individuals after 
iteration, and select the individual with the best fitness for 
output [20]. The specific process of the GA after 
optimization is shown in Figure 2. 
 

72   Informatica 48 (2024) 67–80                                                                       X. Ma 
Set genetic 
operation 
parameters
Start Variation
Production 
initial 
population
Calculate 
individual 
fitness
Choice
Whether the 
conditions are 
met
Overlapping
Output optimal 
solution
Y
N
 
Figure 2: The specific process diagram of the genetic algorithm 
 
In Figure 2, the first step is to initialize the 
population, generate different individuals, and then use 
the fitness function to judge the fitness of the individual 
to determine its adaptability to the environment. 
Individuals with better environmental fitness are selected 
to enter the next generation for iteration, so that the 
algorithm can gradually obtain the optimal solution in the 
iterative process. Individuals are selected from the 
population to perform crossover and mutation operations 
according to the corresponding probability to generate 
new individuals and enter the next generation. Finally, 
whether the output outcomes meet the requirements is 
determined. If they meet the requirements, the algorithm 
ends and the optimal individual is output. Otherwise, the 
process is returned to continue execution. The 
preprocessing method of the GA-BP is generally to 
normalize the data, and map the data to the [0,1] interval, 
thus increasing the convergence speed of the algorithm. 
The normalization is calculated, as described in equation 
(15). 
 ( min) / (max min) yx = − − (15) 
 
In equation (15) 
x
 indicates the input value. min 
indicates the minimum value in the input value. 
max
 
indicates the maximum value in the input value. Equation 
(15) converts the data into intervals [0,1] to reduce the 
training time of the algorithm and improve its 
computational efficiency. The shipping safety is very 
important, which is of great practical significance to 
accurately predict its safety. The GA-BP algorithm 
obtained by optimizing the BP neural network through 
the GA has better prediction performance. Therefore, 
using the GA-BP to predict the shipping safety can 
improve the prediction accuracy. The specific prediction 
flowchart is shown in Figure 3. 
 
Strat
Initialization 
parameters
Input samples
Calculate the 
output of each layer
calculation error
Calculate error 
signal
Update weight
Judge whether the 
data is input
Judge whether the 
algorithm meets the 
end condition
End
GA optimal 
individual 
decoding
Y
Y
N
N
 
Figure 3: Flowchart of GA-BP prediction algorithm 
 
As shown in Figure 3, the specific steps of the 
GA-BP prediction algorithm are as follows. Firstly, the 
structure of the neural network is determined in 
combination with the sample set in practical applications. 
Then the parameters of the BP are initialized by using the 
optimal individual code output by GA. The next step is to 

Impacts of Crew Training Quality Evaluation System Based on… Informatica 48 (2024) 67–80 73 
calculate the output of each level of the neural network 
through the input samples, and compare the actual output 
outcomes with the expected output outcomes to obtain 
the error. The error is converted into an error signal and 
back propagated, with each neuron weighted and updated. 
Finally, it is judged whether the data input is completed. 
After the input is completed, whether the algorithm ends 
is judged. If the end condition is met, the algorithm ends. 
Otherwise, the output of each level is calculated from the 
input sample and the algorithm is restarted until the end 
condition is reached. The GA-BP can solve the problem 
that the training falls into the local minimum due to the 
large difference in initial values of the BP, greatly 
strengthening the performance of the BP, and better 
predicting the shipping safety. 
 
 
 
 
3.2 Construction of quality evaluation 
System for crew education and training 
In the crew education and training, there are many factors 
that affect the shipping safety more or less. To predict the 
shipping safety, it is necessary to build an evaluation 
index system for the education and training system of 
seafarers, that is, to analyze the training system and find a 
suitable index system to comprehensively evaluate it. On 
the basis of analyzing the factors that affect the overall 
security of the system, a scientific and reasonable 
evaluation system is established. For the composition of 
the system, it can be represented by equation (16). 
 
 ( , ) S E R = (16) 
 
Equation (16) S represents the system. E 
represents the element set of the system. R represents 
the relationship between the elements in the system. 
Evaluating the system from different perspectives will 
generate different evaluation indicators. In the research 
on the impact of the crew education and training system 
on shipping safety, the evaluation indicators for the 
impact of the crew education and training system on 
shipping safety are shown in Table 2. 
 
 
 
 
 
 
 
 
Table 2: Evaluation indicators for the impact of crew education and training systems on shipping safety 
Evaluation system Primary index Indicator code 
Indicator 
code 
Shipping safety evaluation 
system 
Crew factor 
Family environment background Z1 
education level Z2 
Psychological quality Z3 
Physical quality Z4 
Adaptability Z5 
Education and training 
institutions 
Meet the requirements of the competent 
authority 
Z6 
Teacher level Z7 
Condition of facilities and equipment Z8 
Training scale Z9 
Training cycle Z10 
Operation of quality system Z11 
market demand Z12 
Information integration processing Z13 
Crew entry standard settings Z14 
Shipping company 
Feedback of new ship technology Z15 
Demand for production safety Z16 
Crew management level Z17 
Competent authority 
Evaluation of crew training effect Z18 
Daily supervision service Z19 
 
From Table 2, the research selects evaluation 
indicators from four major directions: crew factors, 
education and training institutions, shipping companies, 
and competent authorities. Among the above evaluation 

74   Informatica 48 (2024) 67–80                                                                       X. Ma 
indicators, because the background factors of the crew 
have already occurred, it is difficult to change and plays 
an important part in crew training, so the study selects 
them as the evaluation indicators. The crew factors are 
further subdivided into five specific evaluation indicators: 
crew family environment, cultural level, psychological 
quality, physical quality, and adaptability. In addition to 
crew factors, another major category of evaluation 
indicators that has a significant impact on the training 
quality of the training system is education and training 
institutions. For education and training institutions, they 
are the main body of crew training, which play an 
important role in the crew training. The indicators of 
education and training institutions are mainly divided into 
nine aspects, namely meeting the objective requirements 
of the competent authority, the number of teachers for 
education and training, the condition of training 
equipment and facilities, training scale, training cycle, the  
 
operation of the training quality system, market demand, 
information integration and processing, and crew entry 
standard. The other two categories of evaluation 
indicators are shipping companies and competent 
authorities. Shipping company indicators are further 
subdivided into information feedback on new ship 
technologies, demand for production and safety, and 
internal management. These three evaluation indicators 
are mainly conducted on land. The indicators of the 
competent authority are divided into crew training effects 
and daily supervision services. These two evaluation 
indicators have the function of testing the outcomes and 
process of crew training. The effectiveness of crew 
training can be comprehensively evaluated through 19 
evaluation indicators. The training effect of crew 
members has a significant impact on shipping safety. 
Therefore, based on the 19 evaluation indexes, this study 
establishes a quality evaluation system model for crew 
safety education and training based on GA-BP. The 
transportation safety model based on GA-BP algorithm is 
shown in Figure 4. 
 
Z1
Z2
Z3
.
.
.
Z19
Evaluating indicator
BP neural network prediction 
model
GA searching for the best 
individual
Shipping safety
Optimal individual 
decoding input
 
Figure 4: Schematic diagram of GA-BP algorithm for predicting shipping safety 
 
As shown in Figure 4, GA is used to find the optimal 
individual sample in the sample set. Then the sample 
code is used to initialize the parameters of the BP neural 
network. The 19 evaluation indicators of the crew 
education and training system are input into the BP after 
parameter initialization. The positive and negative 
propagation of the BP neural network is used to adjust the 
weights and thresholds of each neuron until the algorithm 
is finally completed. By inputting 19 evaluation 
indicators from the crew education and training system to 
predict shipping safety, the impact of crew education and 
training quality on shipping safety can be determined. 
Because the 19 evaluation indicators of the crew 
education and training system can also evaluate the 
quality of crew education and training, the relationship 
between the quality of crew education and training and 
shipping safety can be analyzed. 
 
 
 
 
 
 
4 Algorithm performance 
comparison and model empirical 
analysis 
The parameters in the experiment are set as follows. The 
population size of the GA is set to 100. The number of 
iterations is set to 500. The crossover probability is set to 
0.5 and the variation probability is 0.005. The learning 
rate in the BP neural network is set to 0.005, and the 

Impacts of Crew Training Quality Evaluation System Based on… Informatica 48 (2024) 67–80 75 
momentum coefficient is set to 0.7. The number of 
particles in the APSO algorithm is set to 10, and the 
number of iterations is set to 500. The potential depth is 
10, and the learning factor is 1 2 2 CC == . The spectrum 
width is 0.99. The data set in the experiment originated 
from the data set in the traveling salesman problem (TSP). 
In the performance analysis of the GA-BP, the research 
mainly analyzes the error curve, the prediction outcome 
and the comparison of the expected output. The research 
compares the performance of the GA-BP, tradition BP 
and the APSO-BP, and uses the same function to train the 
three algorithms. The number of iterations and the mean 
square error (MSE) are shown in Figure 5. 
 
0 100 200 300 400 500
0
10
-4
10
-3
10
-2
10
-1
10
0
0 100 200 300 400 500
0
10
-4
10
-3
10
-2
10
-1
10
0
BP
GA-BP
APSO-BP
Goal
(a) The first function training (a) The first function training
500 epochs 500 epochs 500 epochs 500 epochs
Mean square error Mean square error
Mean square error Mean square error
(b) The second function training (b) The second function training
BP
GA-BP
APSO-BP
Goal
 
Figure 5: Iteration times and error curve 
 
Figure 5 (a) and Figure 5 (b) display the number of 
iterations and MSE of the two functions used by the BP, 
GA-BP, and APSO-BP algorithms. In Figure 5(a), among 
the three algorithms, GA-BP had the fastest convergence 
speed and reached a stable state when the number of 
iterations was 130. At this time, the MSE of the GA-BP 
algorithm was the lowest, which was 0.00323. The BP 
neural network had the lowest convergence speed and 
reached a stable state when the number of iterations was 
300. At this time, the MSE of the BP neural network was 
the highest, which was 0.00973. In Figure 5(b), the 
training times of the three algorithms were roughly 
consistent with the error curve. Among the three 
algorithms, the GA-BP algorithm had the fastest 
convergence speed and the lowest MSE after stabilization. 
There was still a gap between the MSE and the target 
value. Therefore, the performance of the algorithm can be 
optimized. The lower the MSE, the higher the accuracy of 
the algorithm. Therefore, from the MSE dimension, the 
performance of the GA-BP is best. This algorithm has 
higher accuracy in predicting data sets. The results show 
that the strong search ability of GA improves the weight 
and threshold of BP, improves convergence speed, and 
makes the MSE of GA-BP algorithm smaller than that of 
BP. In addition to MSE, the MAE and the maximum 
absolute percentage error (MAPE) can also be used as the 
index for assessing the performance of the algorithm. In 
the training of the three algorithms, the MAE and MAPE 
are shown in Figure 6. 
 
0 100 200 300 400 500
0
10
-4
10
-3
10
-2
10
-1
10
0
BP
GA-BP
APSO-BP
Goal
(a) Iterations - Mean absolute error (a) Iterations - Mean absolute error
500 epochs 500 epochs
Mean absolute error Mean absolute error
0 100 200 300 400 500
4.5
5.0
5.5
6.0
6.5
7.0
BP
GA-BP
APSO-BP
Goal
(b) Iterations -Maximum absolute percentage 
error
(b) Iterations -Maximum absolute percentage 
error
500 epochs 500 epochs
Maximum absolute percentage error(%) Maximum absolute percentage error(%)
 
Figure 6: MAE and MAPE curves 
 
Two sub-graphs in Figure 6 show the MAE and 
MAPE of the three algorithms during the training process, 
respectively. From Figure 6(a), the GA-BP had the fastest 
convergence speed among the three algorithms, and the 
MAE tended to be stable at 0.00152 when the iteration 
was 135. The convergence speed of the APSO-BP 
algorithm was second only to that of the GA-BP. For the 
APSO-BP, when the number of trainings was 180, the 
MAE tended to be stable, which was 0.00635. When the 
number of trainings was 250, the MAE tended to be 
stable, which was 0.00817. From Figure 6(b), the GA-BP 
algorithm had the fastest convergence speed among the 

76   Informatica 48 (2024) 67–80                                                                       X. Ma 
three algorithms. The MAPE tended to be stable at 4.8% 
when the training times were 135 times. The convergence 
speed of the APSO-BP was the second. For the GA-BP, 
when the training times were 210, the MAPE tended to be 
stable, at 5.3%. The BP neural network had the slowest 
convergence speed. When the training times were 310, 
the MAPE tended to be stable at 6.1%. The results show 
that the GA-BP algorithm can solve the local minimum 
value, thus reducing the MAE value and MAPE value of 
the BP. From the above outcomes, it can be concluded 
that among the three algorithms, the GA-BP algorithm 
has better performance than the other two algorithms, but 
its MAE and MAPE have not yet reached the target value, 
and further improvements are necessary. To further 
compare the errors of each algorithm, the three 
algorithms after training are applied to the same training 
samples. The comparison of the prediction errors is 
shown in Figure 7. 
 
0 4 8 12 16 20 24
-0.04
-0.03
-0.02
-0.01
0
0.01
0.02
Delay
Prediction error
28 32 36 40
c
APSO-BP
GA-BP
0.03
0.04
BP
 
Figure 7: Prediction error of three algorithms 
 
Figure 7 shows the prediction errors of the three 
algorithms in the training samples. In Figure 7, GA-BP 
algorithm predicted 40 samples more accurately. It is a 
prediction hit when the prediction error was within ±
0.015. Therefore, the hit rate of GA-BP algorithm was the 
highest among the three algorithms, which was 90%. The 
maximum error value of GA-BP algorithm was 0.019, 
which was lower than the maximum error value of 0.039 
for BP and 0.028 for APSO-BP. The GA-BP has the 
highest hit rate and the lowest maximum error value 
among the three algorithms. From the perspective of the 
prediction error dimension, the performance of the 
GA-BP is the best among the three algorithms. The 
results show that applying the scalability of GA can 
reduce the error value in BP neural network. To compare 
the actual performance of the three algorithms, in 
addition to applying the three algorithms to the training 
samples for prediction, they are also applied to the test 
samples for prediction. The predicted outcomes are 
compared with the expected outcomes. The predicted and 
expected outcomes are shown in Figure 8. 
 
0 4 8 12 16 20 24
0.35
0.40
0.45
0.50
0.55
0.60
0.65
Delay
Function output
28 32 36 40
APSO-BP predictive output
GA-BP predictive output
0.70
BP predictive output
Expected output
（a ）Training sample
0 4 8 12 16 20 24
0.35
0.40
0.45
0.50
0.55
0.60
0.65
Delay
Function output
28 32 36 40
APSO-BP predictive output
GA-BP predictive output
0.70
BP predictive output
Expected output
（b ）Test sample
 
Figure 8: Predicted outcomes and expected outcomes of three algorithms 
 
Figure 8 (a) and (b) are the scatter plots of the 
prediction outcomes and expected outcomes of the three 
algorithms in training samples and testing samples. In 
Figure 8 (a), among the three algorithms, the prediction 
outcome of GA-BP algorithm was the closest to the 
expected outcome. The difference between the predicted 
outcomes and the expected outcomes of the GA-BP 
algorithm was relatively stable, which showed that the 
GA-BP algorithm had the best prediction performance in 
the training samples among the three algorithms. From 

Impacts of Crew Training Quality Evaluation System Based on… Informatica 48 (2024) 67–80 77 
Figure 8 (b), the prediction effect of the GA-BP 
algorithm was the best in the testing samples, the 
APSO-BP algorithm was the second, and the BP was the 
worst. It shows that GA uses the objective function, 
thereby improving the prediction accuracy of the GA-BP. 
The GA-BP algorithm has the best prediction 
performance among the three algorithms in the testing 
sample. Summarizing the performance of the above three 
algorithms in PE, MSE, MAE, and MAPE, GA-BP 
algorithm has the best performance. Using this algorithm 
to evaluate the quality of crew training and teaching has 
higher accuracy, which can better predict the shipping 
safety. After the GA-BP model is completed, the model is 
empirically analyzed to find the factors that have the 
greatest impact on shipping safety among the factors 
affecting crew education and training. The model analysis 
outcomes are shown in Table 3. 
 
 
Table 3: Shipping safety assessment outcomes 
Primary index Indicator Code 
Indicator 
code 
Output 
outcomes 
Crew factor 
Family environment background Z1 0.85 
education level Z2 0.91 
Psychological quality Z3 0.92 
Physical quality Z4 0.90 
Adaptability Z5 0.93 
Education and 
training institutions 
Meet the requirements of the competent 
authority 
Z6 0.68 
Teacher level Z7 0.76 
Condition of facilities and equipment Z8 0.55 
Training scale Z9 0.91 
Training cycle Z10 0.93 
Operation of quality system Z11 0.75 
market demand Z12 0.63 
Information integration processing Z13 0.66 
Crew entry standard settings Z14 0.77 
Shipping company 
Feedback of new ship technology Z15 0.56 
Demand for production safety Z16 0.63 
Crew management level Z17 0.72 
Competent authority 
Evaluation of crew training effect Z18 0.48 
Daily supervision service Z19 0.53 
 
The scores of indicators Z2, Z3, Z4, Z5, Z9, and Z10 
were all above 0.90. It can be seen that these indicators 
are most relevant to shipping safety. In the subsequent 
crew education and training, more attention should be 
paid to these indicators. By strengthening these indicators, 
the overall quality of the crew is improved, thereby 
reducing shipping risks and improving shipping safety. 
5 Discussion 
This study conducted performance comparison 
experiments on GA-BP algorithm, BP, and APSO-BP 
algorithm. The experimental results showed that the MSE 
value of the GA-BP algorithm reached stable at 0.00323 
after 130 iterations, However, the MSE value of the BP 
neural network did not reach the stable value until the 
iterations were 300, and the value was the highest at 
0.00973 among the three algorithms. Due to the 
advantage of GA, the weights and thresholds in the neural 
network algorithm are adjusted and the convergence rate 
is optimized. The mean squared error value of GA-BP 
algorithm was less than that of BP. This result is similar 
to the study conducted by Wei et al. [21]. Later, the MAE, 
MAPE and error values of the three algorithms were 
compared. The results showed that the GA-BP algorithm 
converged the fastest, and the MAE, MAPE and 
maximum error value of the algorithm were 0.00152, 
4.8% and 0.019 respectively. The convergence rate of 
APSO-BP algorithm was less than that of GA-BP 
algorithm, and the MAE, MAPE and error value of this 
algorithm were 0.00635, 5.3% and 0.028, respectively. 
However, the BP showed the slowest convergence rate, 
and the maximum MAE, MAPE and error values were 
0.00817, 6.1% and 0.039, respectively. The reason for 
this result may be that the GA in GA-BP algorithm can 
solve the problem that the initial value difference of BP 
causes the local minimum value of training, thus reducing 
the MAE, MAPE, and error values of the algorithm. The 
result coincides with results conducted by Wang [22]. 
The study also compared the prediction results and 
expected results of the three algorithms in practical 
application. The comparison results showed that the 
difference between the predicted results and the expected 

78   Informatica 48 (2024) 67–80                                                                       X. Ma 
results was the smallest, followed by APSO-BP algorithm, 
while the BP had the largest difference between the 
expected results. The GA utilized objective function to 
reduce error factor, thus improving the prediction 
accuracy of GA-BP algorithm. This result is similar to the 
Zhang result, as shown in [23]. Finally, An empirical 
analysis was conducted on the education and training 
quality evaluation model based on the GA-BP algorithm. 
By scoring the indicators, the factors that have the 
greatest impact on crew education, training, and shipping 
safety are identified. The experimental results showed 
that the educational level, psychological quality, physical 
quality, adaptability, training scale and training cycle had 
the greatest influence on shipping safety. Therefore, the 
shipping safety can be improved by strengthening the 
training of the above indicators. This result is identical to 
that of the results cunducted by Nopas [24]. From the 
above results, it can be seen that the overall prediction 
performance of GA-BP algorithm is significantly better 
than other algorithms. The influence index of crew 
training quality on shipping safety can be accurately 
obtained through this algorithm. 
6 Conclusion 
With the increase of maritime business, more maritime 
accidents have occurred, and shipping safety has been 
widely concerned. There is a correlation between the 
shipping safety and the quality of crew education and 
training, but the relationship is not particularly clear. At 
present, there is also a lack of research on this issue. To 
solve the problem, the GA was used to adjust the 
parameters of the BP to obtain the optimized GA-BP 
algorithm. The relevant factors in the crew education and 
training were calculated by this algorithm, and the 
correlation between the GA-BP algorithm and the 
shipping safety was obtained. To test the performance of 
the GA-BP algorithm, it was compared with tradition BP 
and APSO-BP algorithm from multiple dimensions. The 
hit rate of GA-BP algorithm was 90%, the maximum 
error value was 0.019, the MAE was 0.00152, the MSE 
was 0.00323, and the MAPE was 4.8%, which were 
better than APSO-BP algorithm and BP. This outcome 
shows that the GA-BP has the best performance, which 
has higher accuracy in predicting samples. In addition, 
the study also conducted an empirical analysis for the 
model. The outcomes showed that the output outcomes of 
education level, psychological quality, physical quality, 
adaptability, training quality, and training cycle were 0.91, 
0.92, 0.90, 0.93, 0.91, and 0.93, respectively. The 
outcomes illustrate that these indicators in crew training 
that have the greatest impact on shipping safety. By 
improving these indicators, shipping safety can be 
improved. This study demonstrates that in actual crew 
training, the education level, physical quality, adaptability, 
and psychological quality of crew members should be 
improved, and educational institutions should pay 
attention to the quality and duration of training. By 
improving the comprehensive abilities of crew members, 
shipping safety can be improved. Although the GA-BP 
algorithm has high prediction accuracy, it is far from the 
expected goal. The follow-up research direction is to 
establish a neural network model more suitable for 
shipping safety prediction. 
 
Funding 
The research is supported by: 2021 Basic Scientific 
Research Project of Higher Education of the Educational 
Department of Liaoning Province (general project): 
Research on the training path of High-quality seafarers 
with Chinese characteristics under the standard of STCW 
Convention. 
References 
[1] B. Debeelde, E. Tanghe, M. Yusuf, D. Plets, W. 
Joseph, “Radio channel modeling in a ship hull: path 
loss at 868 MHz and 2.4, 5.25, and 60 GHz,” IEEE 
and Wireless Propagation Letters, vol. 20, no. 4, pp. 
597-601, 2021. 
https://doi.org/10.1109/LAWP.2021.3058439 
[2] I. Subotič, “3G, HSPA AND 4G LTE network 
reliability and achievable internet speeds on airborn 
aircraft up to 10.000 FT above ground level,” 
Informatica, vol. 47, no. 5, pp. 95-110, 2023. 
https://doi.org/10.31449/inf.v47i5.4513 
[3] S. Choi, and J. Kim, “Meta-analysis on the 
effectiveness of airline crew training and 
education,” Journal of Tourism Management 
Research, vol. 24, no. 3, pp. 749-766, 2020. 
https://doi.org/10.18604/tmro.2020.24.3.34 
[4] X. Zhu, J. Li, Q. Liu, W. Yu, S. Li, J. Zhao, Y. Dong, 
Z. Zhang, H. Zhang, and S. Lin, “Use of a BP neural 
network and meteorological data for generating 
spatiotemporally continuous LAI time series,” IEEE 
Transactions on Geoscience and Remote Sensing, 
vol. 99, pp. 1-14, 2021. 
https://doi.org/10.1109/TGRS.2021.3095535 
[5] Q. Kuang, “Image pattern recognition algorithm based 
on improved genetic algorithm,” Journal of Physics 
Conference Series, vol. 182, no. 3, pp. 032038, 2021. 
https://doi.org/10.1088/1742-6596/1852/3/032038 
[6] W. Zhu, H. Wang, and X. Zhang, “Synergy evaluation 
model of container multimodal transport based on 
BP neural network,” Neural Computing and 
Applications, vol. 33, no. 2, pp. 4087-4095, 2021. 
https://doi.org/10.1007/s00521-020-05584-1 
[7] B. Shao, C. Ni, J. Wang, and Y. Wang, “Research on 
venture capital based on information entropy, BP 
neural network and CVaR model of digital currency 
in Yangtze River delta,” Procedia Computer Science, 
vol. 187, pp. 278-283, 2021. 
https://doi.org/10.1016/j.procs.2021.04.063 
[8] Y. Liu, X. Wang, X. Zhu, and Y. Zhai, “Thermal error 
prediction of motorized spindle for five-axis 
machining center based on analytical modeling and 

Impacts of Crew Training Quality Evaluation System Based on… Informatica 48 (2024) 67–80 79 
BP neural network,” Journal of Mechanical Science 
and Technology, vol. 35, no. 1, pp. 281-292, 2021. 
https://doi.org/10.1007/s12206-020-1228-7 
[9] H. Xi, Z. Li, J. Han, D. Shen, N. Li, Y. Long, Z. Chen, 
L. Xu, X. Zhang, D. Niu, H. Liu, “Evaluating the 
capability of municipal solid waste separation in 
China based on AHP-EWM and BP neural 
network,” Waste Management, vol. 139, pp. 
208-216, 2022. 
https://doi.org/10.1016/j.wasman.2021.12.015 
[10] K. X. Zhou, W. H. Lin, J. K. Sun, J. Zhang, D. 
Zhang, X. Feng, Q. Liu, “Prediction model of 
end-point phosphorus content for BOF based on 
monotone-constrained BP neural network,” Journal 
of Iron and Steel Research International, vol. 29, no. 
5, pp. 751-760, 2022. 
https://doi.org/10.1007/s42243-021-00655-6 
[11] J. Zhou, F. Ding, J. Yang, Z. Pei, C. Wang, and A. 
Zhang, “Navigation safety domain and collision risk 
index for decision support of collision avoidance of 
USVs,” International Journal of Naval Architecture 
and Ocean Engineering, vol. 13, no. 3, pp. 340-350, 
2021. https://doi.org/10.1016/j.ijnaoe.2021.03.001 
[12] P. Nosov, S. Zinchenko, A. Ben, Y. Prokopchuk, P. 
P. Mamenko, I. Popovych, V. Moiseienko, and D. 
Kruglyj, “Navigation safety control system 
development through navigator action prediction by 
data mining means,” Eastern-European Journal of 
Enterprise Technologies, vol. 2, no. 110, pp. 55-68, 
2021. 
https://doi.org/10.15587/1729-4061.2021.229237 
[13] N. Shilov, “Recommender system for navigation 
safety: requirements and methodology,” TransNav 
the International Journal on Marine Navigation and 
Safety of Sea Transportation, vol. 14, no. 2, pp. 
405-410, 2020. 
https://doi.org/10.12716/1001.14.02.18 
[14] N. Hasanspahi, V. Frani, I. Rudan, and L. Maglic, 
“Analysis of navigation safety regarding tankers in 
narrow waterways,” Journal of Maritime & 
Transportation Science, vol. 55, no. 1, pp. 201-217, 
2019. https://doi.org/10.18048/2018.00.13 
[15] Y. Zhou, Y. Wang, X. Wen, and C. Han. 
“Application of BP neural network in efficiency 
prediction of oilfield mechanized mining system,” 
Journal of Failure Analysis and Prevention, vol. 22, 
no. 2, pp. 658-665, 2022. 
https://doi.org/10.1007/s11668-022-01360-6 
[16] S. Xu, Z. Lin, G. Zhang, T. Liu, X. Yang, “A fast yet 
reliable noise level estimation algorithm using 
shallow CNN-based noise separator and BP 
network,” Signal Image and Video Processing, vol. 
14, no. 1, pp. 1-8, 2020. 
https://doi.org/10.1007/s11760-019-01608-z 
[17] H. Xi, Z. Li, J. Han, D. Shen, N. Li, Y. Long, Z. 
Chen, L. Xu, X. Zhang, D. Niu, H. Liu, “Evaluating 
the capability of municipal solid waste separation in 
China based on AHP-EWM and BP neural 
network,” Waste Management, vol. 139, pp. 
208-216, 2022. 
https://doi.org/10.1016/j.wasman.2021.12.015 
[18] X. Zhang, “Analysis of financial market trend based 
on autoregressive conditional heteroscedastic model 
and BP neural network prediction,” Journal of 
Intelligent and Fuzzy Systems, vol. 39, no. 4, pp. 
5845-5857, 2020. 
https://doi.org/10.3233/JIFS-189060 
[19] S. Samanta, A. Ojha, B. Das, and S. K. Mondal, “A 
profit maximisation solid transportation problem 
using genetic algorithm in fuzzy environment,” 
Fuzzy Information and Engineering, vol. 4, pp. 1-18, 
2021. 
https://doi.org/10.1080/16168658.2021.1915454 
[20] W. F. Li, “Reactive power optimization of a power 
system based on genetic algorithm,” Electric 
Switchgear, vol. 52, no. 4, pp. 47-49, 2014. 
https://doi.org/10.3969/j.issn.1004-289X.2014.04.01
4 
[21] W. Wei, R. Cong, Y. Li, A. A. Daniel, C. Yang, and 
Z. Chen, “Prediction of tool wear based on GA-BP 
neural network,” Proceedings of the Institution of 
Mechanical Engineers, Part B: Journal of 
Engineering Manufacture, vol. 236, no. 12, pp. 
1564-1573, 2022. 
https://doi.org/10.1177/09544054221078144 
[22] L. Wang, and X. Bi, “Risk assessment of knowledge 
fusion in an innovation ecosystem based on a 
GA-BP neural network,” Cognitive Systems 
Research, vol. 66, no. 8, pp. 201-210, 2021. 
https://doi.org/10.1016/j.cogsys.2020.12.006 
[23] H. Zhang, and Z. Tian, “Failure analysis of corroded 
high-strength pipeline subject to hydrogen damage 
based on FEM and GA-BP neural network,” 
International Journal of Hydrogen Energy, 47(7): 
4741-4758, 2022. 
https://doi.org/10.1016/j.ijhydene.2021.11.082 
[24] D. Nopas, C. Ueangchokchai, and W. Meepan, “The 
study of the training platform towards thailand's 
international airlines cabin crew during the 
pandemic of COVID-19,” Higher Education Studies, 
vol. 13, no. 3, pp. 45-53, 2023. 
https://doi.org/10.5539/hes. v13n3p45 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

80   Informatica 48 (2024) 67–80                                                                       X. Ma