SHIP-MOTION PREDICTION: ALGORITHMS AND SIMULATION RESULTS

X. Zhao, R. Xu and C. Kwan

Intelligent Automation, Inc.
7519 Standish Place, Suite 200
Rockville, MD 20855, USA
ckwan@i-a-i.com
numerical simulations were carried out and compared with
conventional algorithms like NN, AR, and Wiener prediction.
Abstract
2. MCA based prediction algorithm
Ship-motion prediction will be very useful for several naval
operations such as aircraft landing, cargo transfer, off-loading of
small boats, and ship mating between a big transport ship and
some small ships. The prediction information will be extremely
useful in sea states above 3. Five to ten seconds ship motion
prediction can give the operator ample time to avoid serious
collisions.

MCA has similar mathematics as the Principal Component

Analysis [2], except that MCA utilize the eigenvectors
corresponding to the minor components. Specifically, the motion
data was aligned into a sequence of vectors X i , i=1 N. Then, the
eigenvectors and eigenvalues of the following autocorrelation
matrix was calculated
N

This paper summarizes the development of a high performance

ship-motion prediction algorithm using Minor Component
Analysis (MCA). Simulation results showed this method can
predict ship motion a long time ahead with consistent accuracy.
That is, the prediction error is almost the same for the 5 second and
20 second predictions. Other conventional algorithms like Neural
Network (NN), Autoregressive method (AR), Wiener prediction
were also studied for comparative purpose.
1. Introduction
There are many ship operations that require ship motion prediction
[1]. First, the landing of aircraft in carriers and helicopter in a
destroyer is an important application. In high sea states, the landing
operation may be quite dangerous. Second, side by side cargo
transfer is another application. One crane may off-load some
dangerous ammunition from a big ship to a small ship. Again the
operation can be quite dangerous in high sea states (above sea state
3). Third, in some naval operations, there are some mating
operations between a large transport ship and some small ships.
The large ship is floating far from shore and the small ships go
back and forth between the large ship and the shore. During high
sea states, it may be difficult for the small ships to go inside the
large ship.
In all of the above applications, ship motion prediction will be very
important because if one can look 5 to 10 seconds ahead of time,
then one can time the landing to avoid a serious crash in the case of
aircraft landing. For cargo transfer, the prediction information may
prevent crash of cargo that may explode. Finally, a smooth
mating between ships also requires some prediction information
of ship motion.
In this paper, a prediction algorithm based on MCA was developed
and implemented for ship-motion prediction. The algorithm takes
past motion data for training, and predicts the data in the next 10 or
20 seconds based on the recent motion data. The algorithm can
update its core from time to time to accommodate the sea state or
ship speed change. To evaluate the performance of this predictor,

,(((

X X
i

T
i

(1)

i 1

This will yield the n eigenvalues O0 , ..., On1 , and the associated
eigenvectors u 0 ,..., u n 1 . Now we choose P eigenvectors
corresponding to the P smallest eigenvalues and form a
matrix B R Pu N , which can in turn divides into 2 sub matrices,

> B1 , B2 @ ,

with B1 R pu n1 and B2 R

p u N  n1 

. Here

n1

is

usually chosen to be larger than 2/3 of N. Thus we can rewrite each

X
X i , i=1,,N. as X i 1i , where X 1i R n1 and X 2i R N  n1 .
X 2i
According to MCA procedure, the following approximate
equalities hold
B1 X 1i  B2 X 2 i | 0
1

T
2

(2)
T
2

X 2 i | ( B B2 ) B B1 X 1i

(3)

Thus Eq. (3) can be used to perform ship motion prediction. The
dimension of X 2 i will be the length of the prediction window.

X 1i is a vector containing the past data.

3. Simulation Studies
Ship motion prediction

The ship motion data was provided by JJMA, Inc. which has
simulation software fully tested and evaluated by experienced
Naval officers. We used the DDG-51 destroyer in all the data files.
The ship is moving at a low speed (10-12 knots), heading 15
degrees against the wave direction; Ship motion status (surge,
sway, heave, roll, pitch, yaw) were sampled at 8Hz rate at sea state
3 which has a wave height of about 10 ft.
To implement the above algorithm, the ship motion data were first
down-sampled to 2 points per second. No information was lost
because the ship moved rather slowly. The data were then
separated into two segments: the first 5000 points of data were
used for MCA training and the rest for testing. During the training,

9

,&$663

the size of the data window was 800 points, and each vector was
obtained by shifting the window by 10 points. So altogether there
were 420 vectors used for training. The minor components were
selected in a sense that their energy added up to no more than 1.5%
of the total energy. For the testing part, 750 past values were
collected for predicting the next 50 points. So this is a 25 second
prediction (longer prediction is also possible with longer training
data and less accuracy). In the next several figures, the predicted
and the real ship motions (surge, sway, heave, roll, pitch, and yaw)
are plotted versus time. Although we predicted 50 points ahead of
time, the plots only show the 40th point (20 seconds ahead) in the
prediction window. The green lines are the real motions; the red
ones are the predicted.

Predicted ship motion

Real ship motion

0.15
0.1
0.05
0
-0.05
-0.1
-0.15

100

200

300

400

500

600

Fig. 5 MCA prediction of sway 20 seconds ahead

Predicted ship motion
Real ship motion

1.5
1
Predicted ship motion
Real ship motion

0.5
0

0.5

-0.5
0

-1
-0.5

-1.5
0

100

200

300

400

500

600

-1

100

200

300

400

500

Fig. 1 MCA prediction of roll 20 seconds ahead

It is seen that the 20th second prediction data using MCA matches
quite well with the real ship motion for all the six degree-offreedom (DOF) in a decoupled manner [3]. It can predict the next
50 points simultaneously, which can be used to improve the
prediction accuracy by using the mean value of several predictions
for the same future point. And within the prediction window (in
this case 50 points), the error keeps at the same level (Table 1), i.e.,
the prediction error for the last point is not necessary worse than
the first point and the prediction error is small.

Predicted ship motion

Real ship motion

Fig. 6 MCA prediction of heave 20 seconds ahead

600

-1

-2

-3
0

100

200

300

400

500

600

Fig. 2 MCA prediction of pitch 20 seconds ahead

0.3

Comparative studies

Several other conventional prediction methods were utilized to

compare the MCA predictor, e.g. NN, AR, and Wiener predictor.
A brief description of those predictors is listed below.

Predicted ship motion

Real ship motion

0.2

Neural network (NN)

0.1

T1

V (x )

-0.1

T2
V (x )

-0.2

.
.

-0.3
0

100

200

300

400

500

600

Fig. 3 MCA prediction of yaw 20 seconds ahead

.
.

.
.

V (x )

T N2

V(x)

Input layer
Predicted ship motion
Real ship motion

0.5

Output layer

Hidden layer

0.4

Fig. 7 Ship motion forecast using three-layer NN

0.3
0.2
0.1
0
-0.1
-0.2
-0.3
-0.4
-0.5
0

100

200

300

400

500

Fig. 4 MCA prediction of surge 20 seconds ahead

600

Three-layer neural network shown in Fig. 7 was used for

prediction. For a given input vector x in R N1 , the NN has an output
given by
N2

N1

yi wijV v jk xk  T vj  T wi ,
i = 1,..., N 3 (4)
j 1
k
1

with V (x) the activation function such as a sigmoid function,

v jk

the first-to-second layer interconnection weights, and wij the

9

second-to-third layer interconnection weights. T vm ,T wm , m 1, 2...,

are called the threshold offsets and the number of neurons in layer l
is Nl, with N 2 the number of hidden-layer neurons. In MATLAB
code, function newff(IND,[100 M],{'tansig' 'purelin'},'trainrp') was
used for training and testing.

where K is the size of the prediction window and xiT , xi P are the
true and predicted values of ship motion parameters. Second, we
compared the computational complexity and/or computation time
needed for each scheme. Third, the robustness of the two most
accurate methods was compared. That is, the prediction
performance with respect to noise and disturbances.

Autoregressive model (AR)

a) Prediction accuracy comparison

Based on the past n points in the ship motion data, an AR predictor
predicts the next value of ship motion based on the past n data
sample
y k a1 yk 1  a2 yk  2  ...  an yk  n
(5)
where yk 1 , yk  2 ,..., yk n are the past data samples. The key here

Predictio
n method

is to obtain the AR coefficients A(k ) > a1 (k ) a2 (k ) .... an (k )@T

The on-line scheme to update A(k) is given by

MCA

yk 1
y
A(k ) A(k  1)  ge(k ) k  2
(6)
....

yk  n
where e( k ) y ( k )  y ( k ) is the prediction error and g is gain

NN

factor which should be a small value. If g is too large, instability

may occur. In our simulations, we set g = 0.005. To generate many
future predictions, several AR predictors are needed with the first
one performing a one-step prediction, the second one performing a
two-step prediction, and so on and so forth

Wiener

Wiener Predictor

An FIR Wiener predictor of order p-1 for multi-step linear

prediction has the form
p 1

x ( n  D )

w(l ) x(n  l )

(7)

l 0

where x ( n  D ) is the predicted value of the D

th

point ahead and

x( n  l ) for l =0,1,...,p-1 are the past p data values, w(l ) are the
coefficients of the prediction filter. The Wiener prediction problem
requires that we find the filter coefficients that minimize the meansquare error
[ E{| e(n) |2 } E{| x(n  D )  x (n  D ) |2 }
(8)
th
where x( n  D ) is the true value of the D data point ahead. The
necessary and sufficient condition is the derivative of [ with
respect to w* ( k ) be equal to zero for k=0, 1... p-1, thus we can
form the following Wiener-Hopf equation:
p 1

w(l )r (k  l )
x

rx (k  D ); k

0,1, ..., p  1

(9)

AR

Modal
order
(past data
number)
40 points
100 points
800 points
40 points
100 points
800 points
40 points
100 points
800 points
40 points
100 points
800 points

RMS
error for
0.5s ahead
(m)
0.1507
0.1245
0.0466
0.1484
0.4437
0.5747
0.0576
0.1050
0.0453
0.0817
0.0604
0.0175

RMS
error for
5s ahead
(m)
0.4419
0.4678
0.0538
0.3303
0.5531
0.5287
0.9026
1.0726
0.1207
0.3843
0.3469
0.1248

RMS error
for
20s ahead
(m)
Impossible
0.7022
0.0540
1.2900
1.0095
0.4724
1.6180
1.7495
0.1412
Impossible
0.5792
0.2743

Table 1 Root-Mean-Square error comparison for the four

prediction algorithms. The best performers are marked in pink and
the runner-ups are in yellow.
Table 1 summarizes the initial results on the RMS errors of the
four prediction methods, each having three model orders and
predicting 0.5s, 5s and 20s motion ahead. It is seen that the MCA
approach with model order of 800 gives the best performance for
predicting ship motion 5s and 20s ahead of time. It also yields a
good performance for predicting 0.5s ahead of time. The error
within the prediction window (1 ~40 points) maintains at roughly
the same low level. This is because MCA treat the whole
prediction window as a missing part of the signal space, whereas
the minor components act as noise space. The orthogonal
relationship between noise space and signal space lead the
prediction to be as accurate as possible for the whole window from
a mean square point of view. The runner-up is the AR model with
an order of 800, which predicts both short term and long term ship
motion quite accurately. The 800th order Wiener filtering approach
has the best accuracy in predicting 1 point ahead, but is weak in the
long term prediction due to the loosened correlation.

l 0

Once the weights w(l ) are obtained, we can proceed to do the

prediction based on Eq. (7).
In this study, we did three comparisons. First was the prediction
accuracy. The Root-Mean-Square error was used to assess the
performance of different methods based on using the same past
data points to predict the same future values. Mathematically,
RMS is defined as
K

(x
RMS

T
i

i 1

 xiP )

(10)

b) Computational complexity and/or computation time comparison

Table 2 summarizes the comparison between various prediction
algorithms from the computational complexity point of view. For
each prediction, 10 points and 40 points (5 s and 20 s, respectively)
of prediction were performed. The offline training time and the online prediction time (500 predictions were made) and
computational complexity were compared with MATLAB program
on a Pentium 1.6 GHz PC. It is seen that the training time for MCA
is short; and it is also the fastest method among the four prediction
methods. NN approach requires long training time. AR model is
fast in short term predictions. However, both its training time and
prediction time are proportional to the prediction data length.

9

Wiener filter approach does not require offline training, yet it is

very slow in motion prediction, especially for long term.

AR
Wiener

prediction time
(500
predictions)

0.5

Heave [m]

Heave [m]

Offline
training time

Predicted ship motion

Real ship motion

0.5

-0.5

-0.5

-1

-1

-1.5

-1.5

-2

50

100

150

200

250
Data points

300

350

400

450

-2

500

50

100

150

200

(AR_0%)

10 points
40 points
10 points
40 points
10 points
40 points
10 points
40 points

40 seconds
40 seconds
1735 seconds
3467 seconds
11.4 seconds
47 seconds
Not required
Not required

1.5

0.09 seconds
0.09 seconds
3.44 seconds
3.68 seconds
1.6 seconds
6.4 seconds
843.3 seconds
1054 seconds

-0.5

-1

-1.5

150

200

250
Data points

300

350

400

450

-2

500

50

100

Predicted ship motion

Real ship motion

Usually in the ship motion measurement, the sensors such as

accelerometer and gyroscope data may be noisy due to sensor
noise and external interference, etc. To test the robustness of the
prediction methods, a zero-mean Gaussian random noise with
various standard deviations was added to the ship motion data.
Here we compare the AR and MCA approaches. Table 3 shows the
RMS error for these two approaches when predicting the heave
motion. Figure 8 shows the prediction results

0.14
0.31
0.55

0.056
0.24
0.51

300

350

400

450

500

Predicted ship motion

Real ship motion

0.5
Heave [m]

Heave [m]

250
Data points

0
-0.5

-0.5

-1

-1

-1.5

-1.5

50

100

150

200

250
Data points

300

350

400

450

500

-2

50

100

150

200

250
Data points

300

350

400

450

500

(MCA_20%)

(AR_20%)

MCA model (800

points)
RMS error (m)

200

1.5

c) Robustness comparison

AR model (800
points)
RMS error (m)

150

2
1.5

-2

500

(MCA_10%)

0.5

Noise level
(standard deviation
= peak amplitude
percentage)
0%
10%
20%

(AR_10%)

Table 2 Comparison of computational time for the four prediction

algorithms

450

-1

100

400

-0.5

-1.5

50

350

Predicted ship motion

Real ship motion
1

0.5

300

1.5

Predicted ship motion

Real ship motion
1

0.5

-2

250
Data points

(MCA_0%)

Heave [m]

NN

Prediction
length

1.5

Predicted ship motion

Real ship motion

Heave [m]

Prediction
method
(800 data
points)
MCA

1.5

Fig. 8 Heave motion prediction (red) and actual value (green) at

noisy environment. (AR_0%) means AR model prediction for 0%
Gaussian noise, etc
4. Conclusions

A new algorithm based on MCA for ship motion prediction has

been presented. Based on our evaluations with neural network,
autoregressive and Wiener method, MCA based algorithm yields
better performance and is also suitable for real-time
implementation.
Acknowledgment

This research was supported by the Office of Naval Research

under contract N00014-02-M-0252.
References

Table 3 Comparison of RMS error of AR and MCA model with the

presence of Gaussian noise
It can be seen that both AR model and MCA degrade a bit with
noise added. With 20% Gaussian noise added, the peak value
prediction exhibits relative large error. However, they are quite
robust in predicting the ship motion with noise level less than 10%.

[1] C. Kwan et al., Phase 1 final report to the ONR, January, 2003.
[2] S. Haykin, Neural Networks, Prentice Hall, 1989.
[3] J. C. Chung, Z. Bien and Y. S. Kim, A note on ship-motion
prediction based on wave-excitation input estimation, IEEE
J. Oceanic Engineering, vol. 15, pp244 -250, 1990

9