Research on Ship Track and Navigation Behavior Characteristics
Based on Deep Learning
Yue Yang, Tianyi Liu and Xiaolong Wu
Dalian Naval Academy, Liaoning, China
Keywords: Neural Network, Deep Learning, Ship Track Characteristics, AIS System.
Abstract: Along with the expansion of China's Marine interests and the improvement of comprehensive national
strength, Marine navigation safety and ship abnormal navigation behaviour problems are increasingly acute.
In this paper, a deep learning CALS algorithm based on neural network is proposed to analyse more than
AIS ship sailing track data in the Chinese sea area in 2021, and build ship sailing track prediction and early
warning models. By exploring the navigation characteristics of various types of ships in the Chinese sea area
and the characteristics of navigation behaviour at different time and space scales, this paper is to solve the
problem of abnormal trajectory detection and early warning of ships, and provide information support for safe
navigation and military mission decision-making.
1 INTRODUCTION
With the advent of the era of intelligent big data, the
statistical analysis and mining technology of big data
play an important role. With the continuous
advancement of Chinese-style modernization in the
process of building a maritime power and the
deepening application of Automatic Identification
System (AIS) in maritime safety and communication
between ships and ports, ships and ships, etc., The
historical accumulation of ship track time series data
in the China Sea area has exploded, which provides
strong data support for mining and analyzing the
distribution characteristics of various ship tracks and
the characteristics of navigation behavior at different
time and space scales.
With the continuous progress of China's
comprehensive national strength and the continuous
expansion of overseas interests, the mission of the
People's Navy has gradually changed from offshore
defense to far sea defense. In order to better explore
the navigation characteristics of different types of
ships and navigation behavior characteristics at
different time and space scales in the sea of China,
and make use of the above characteristics to support
military mission decision-making and ensure
navigation safety, this paper starts from the analysis
of ship navigation trajectory and ship behavior. It uses
AIS data of all ships in the sea of China in 2021.
Based on LSTM long and short-term memory
network machine learning algorithm, the innovative
neural network structure uses the real-time sailing
trajectory data of nearby ships to analyze the
abnormal sailing position, tracking anomaly, heading
anomaly, speed anomaly and other abnormal
situations. The new algorithm desire to screen out the
ships that exceed the historical path range and may
have suspicious sailing behaviors. So, it can provide
early warning for military mission decision-making
and safe navigation.
2 DATA PROCESSING
In this paper, about
9
3 10
AIS data of 13 types of
vessels, such as fishing vessels and pilot vessels,
under 11 conditions such as sailing and losing control
in the Chinese sea area in 2021 are collected. After
decoding the original AIS message information, two
tables of Pos and ShipInfo are obtained,
corresponding to ship dynamic information and static
information respectively. The generation of ship
navigation trajectory adopts the method of connecting
ship dynamic information in AIS message
information to a single ship according to the time
series, and records the real-time transmission of
navigation-related information.
370
Yang, Y., Liu, T. and Wu, X.
Research on Ship Track and Navigation Behavior Characteristics Based on Deep Learning.
DOI: 10.5220/0012284200003807
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 2nd International Seminar on Artificial Intelligence, Networking and Information Technology (ANIT 2023), pages 370-374
ISBN: 978-989-758-677-4
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.
2.1 Data Cleaning and Preprocessing
According to the data format requirements of AIS,
data that is not within a reasonable range in ship
dynamic information is excluded. The constraints are
shown in Table 1.
Table 1. AIS message data format table.
message
data within
MMSI
[200000000,900000000]
Lat
[-181000000,181000000]
Lon
[-91000000,91000000]
Sog
[0,52576]
Cog
[0,35990]
Hdg
[0,35990]
Rot
[-1200,1200]
In this paper, the single sailing track of a ship is
the basic research unit. So it is necessary to divide the
original track data. Based on information such as time
interval of adjacent track points, maximum ship
speed, longitude and latitude of track points, flight
segments are divided automatically. The specific
algorithm steps are as follows:
Step1. The AIS message information after
excluding the cross-border data is sorted according to
the ship MMSI number and the priority of the data
update time.
Step2. Traverse the track points according to the
order of the ship MMSI number, until the track point
i
X
is found, and the time interval between it and the
adjacent track point
1i
X
is greater than the set
threshold T. And its sailing speed is 0, or the sailing
state is anchored or berthed and the sailing speed is
close to 0, the track point
i
X
is marked as the track
division point.
Step3. Repeat Step2 until the end of the current
traversed ship MMSI number track sequence. The
track data corresponding to the current ship MMSI
number is cut according to the marked track division
points, and the start and end point of the track are
determined according to the time sequence. The
segment data is identified and presented.
Step4. Switch to the next ship MMSI number, and
repeat Step2 and Step3 until all ship MMSI numbers
are traversed. And the ship track data set is finally
generated.
The data sample after cleaning is shown in Table
2.
Table 2. Sample data after cleaning.
MMSI
Lat
Sog
Cog
Hdg
200018895
22.00362
1955
25640
25400
200018895
22.002265
2418
33680
33600
200018895
22.007665
2624
34370
34300
...
...
...
...
...
200018895
22.016293
3550
32480
31700
200018895
22.022635
3704
34250
34200
200018895
22.029445
3807
34410
34500
2.2 Data Normalization
In order to improve model training efficiency and
prediction accuracy, eliminate prediction errors
caused by dimensional disunity, maximum-minimum
normalization is adopted to normalize track data. The
value of processed data is between [0,1]. The
normalization formula for mapping the actual value
i
of the attribute
to
i
in the interval [0,1] is as
follows.
min
max min
i
i

(1)
In (1)
min
represents the minimum value of the
attribute
and
max
represents the maximum value
of the attribute
.
Normalization processing can effectively
eliminate the problem that the feature weights caused
by different dimensions are not in line with the
reality. At the same time, it can amplify the feature
differences, reduce the data scale and the training
amount of the model. In the follow training part, in
order to better visualize the results or compare the
errors, it is often necessary to carry out anti-
normalization operations, and the formula is as
follows.
max min min
ii

(2)
3 SELECTION OF FEATURE
PARAMETER
In AIS message information, there are seven track-
related parameters. According to the requirements of
the algorithm, five parameters such as longitude (Lon,
degree), latitude (Lat, degree), speed to ground (Sog,
mm/sed), course to the ground (Cog, 1/100 degree)
and ship heading (Hdg, 1/100 degree) were selected.
And the consistency of the five parameters was tested
through Kendall’s consistency test. The analysis
process was shown in Figure 1.
Research on Ship Track and Navigation Behavior Characteristics Based on Deep Learning
371
Figure 1. Kendall’s consistency checking process.
The analysis results are shown in Table 3.
Table1. Kendall’s analysis results.
name
rank
mean
value
Kendall’s
coefficient
2
P
value
Lat
1.038
0.875
34978.682
0.000
Lon
2.047
Sog
3.154
Cog
4.628
Hdg
4.134
Significance level is 1%.
Table 3 shows the results of Kendall's model test,
including rank mean value, Kendall coordination
coefficient W,
2
value and significance P value.
The results of Kendall coefficient consistency test
show that the significance P value of the overall data
is 0.000, which presents significance at the level 1%
and rejects the null hypothesis, so the data presents
consistency. Meanwhile, the Kendall’s coordination
coefficient value of the model is 0.875, which is
within the interval of [0.8,1.0]. Therefore, the degree
of correlation is almost identical. Based on the test
results, the rationality of the selection of track related
feature parameters is verified.
4 CONSTRUCTION OF CALS
ALGORITHM MODEL
This paper starts from long short-term memory
LSTM network machine learning algorithm, the basic
idea is as follows. CNN-Attention-LSTM neural
network structure was used to train the ship trajectory
prediction model. Real-time AIS data was used to
predict the ship trajectory. Sliding window was used
to model the distribution of the difference sequence.
And the difference between the predicted value and
the actual value was converted into a unified anomaly
score. According to the anomaly score, we can
determine whether the trajectory is abnormal and
determine the type of abnormal trajectory. The
process is shown in Figure 2.
Figure 2. Flow chart of CALS algorithm.
In this paper, the CNN model is combined with
LSTM to extend the application scenario of the
algorithm, so that it can support long input sequences,
which can be read by the CNN model as block or
subsequence information. The working mode of
CNN-LSTM is that the sub-sequence information is
passed into the CNN model as input information, and
then LSTM summarizes and processes the sub-
sequence information before output. The LSTM
based on one-dimensional convolutional neural
network enhances the processing power of LSTM
convective data and improves the learning efficiency.
At the same time, considering the AIS data set has
significant time characteristics, time attention
mechanism is introduced on the basis of CNN-LSTM.
So that the neural network can make more reasonable
use of information sources and better explore data
characteristics when making predictions. The
structural relationship of the three is shown in Figure
3.
ANIT 2023 - The International Seminar on Artificial Intelligence, Networking and Information Technology
372
Figure 3. CNN-Attention-LSTM diagram.
5 TEST RESULTS AND
ANALYSIS
5.1 Reading and Processing of Real-
Time AIS Data
This method is implemented based on python
language. Taking a total of 1,048,576 AIS message
information of ships in the sea area of China in 2021
as an example, after data cleaning, model
construction, anomaly detection, type identification
and other steps, the method simulates real-time AIS
data push through Socket technology to realize real-
time detection of ships with abnormal track.
In order to reduce repeated reading of the model,
reduce performance waste, and improve prediction
speed, this paper uses Socket technology to create
threads to monitor real-time data, form resident
services, and use sockets to pass parameters. The
specific algorithm steps are as follows.
Step1: Listen for client connection requests, and
use sockets to pass data to threads after successfully
establishing a connection.
Step2: Start the thread, process the data and return
the result to the client.
Step3: Close the connection, the thread suspends,
and continue listening for client connection requests.
The specific implementation effect is shown in
Figure 4.
Figure 4. Socket monitors real-time data effects.
5.2 Prediction Model Evaluation
Two commonly used standards of absolute mean
error (MAE) and root mean square error (RMSE) are
selected for the evaluation indicators of the prediction
model. They can measure the deviation between the
observed value and the true value and reflect the
actual situation of the predicted value error. The
smaller is the value, the higher is the accuracy of the
model.
2
1
1
n
ii
i
RMSE y y
n

(3)
1
1
n
ii
i
MSE y y
n

(4)
The error values of the three prediction models
are shown in Table 4.
Table 4. Error values of the three prediction models.
Model
CNN-
Attention-
LSTM
CNN-LSTM
LSTM
RMSE
0.021401891
0.027483279
0.033624116
MSE
0.017074395
0.022646103
0.031701226
It is not difficult to see from the Table 4 that CNN-
Attention-LSTM model has a great reduction in
absolute mean error and root mean square error
compared with CNN-LSTM model and LSTM
model. It can be seen that the CNN-Attention-LSTM
model used in this paper has high accuracy.
Research on Ship Track and Navigation Behavior Characteristics Based on Deep Learning
373
6 CONCLUSION
From the perspective of Marine security, the research
on ship track and navigation behavior characteristics
in China's sea area conducted in this paper has
performed well in the actual test. It reached the
expected goal, and can be applied in the field of ship
trajectory anomaly warning, providing technical
support for commanders and combatants to judge and
make decisions on the enemy situation and the safe
navigation of Chinese ships.
ACKNOWLEDGMENTS
This work was financially supported by Dalian Naval
Academy research and innovation team fund
DJYKYKT2021-018 ， and student research fund
DJYKYKT2022-003.
REFERENCES
Wenyao Ma. Consistency Detection of ship abnormal
behavior [D]. Dalian Maritime University, 2018.
Bochen Yang. Research and application of ship trajectory
analysis based on AIS [D]. University of Electronic
Science and Technology of China, 2018.
Shuang Li. Fishing boats suspicious behavior detection
based on AIS [D]. Dalian Ocean University, 2022.
DOI:10.27821/d.cnki.gdlhy.2022.000214.
Jianguo Kong, Yabin Li. Research on track prediction
based on CNN-LSTM-attention model [J]. Aeronautical
Computing Technology, 2019, 53(01):1-5.
Chengyong Liu, Wenjie Qiao. Research on ship track
prediction model based on LSTM and attention
mechanism [J]. Navigation of China, 2021, 44(04):94-
100+106.
Yuling Jiang, Zhennan Xiong, Jihong Tang. Ship trajectory
clustering algorithm based on trajectory segment
DBSCAN [J]. Navigation of China, 2019, 42(03):1-5.
Jingyi Xie. Research on ship track prediction method based
on navigation strategy learning [D]. University of
Chinese Academy of Sciences (national center for space
science), 2021.
DOI:10.27562/d.cnki.gkyyz.2021.000030.
Yingjie Yan. Transformer state anomaly detection based on
sliding window and clustering algorithm [J]. High
Voltage Technology, 2016, 42(12): 4020-4025. DOI:
10.13336/j.1003-6520.hve.20161128041.
ANIT 2023 - The International Seminar on Artificial Intelligence, Networking and Information Technology
374