4.1. Mariner Class Vessel with Simulation Data
Firstly, the Mariner class vessel is taken as the research object and the effectiveness of LGPR is evaluated by using the simulation data. The simulation tests are performed in MATLAB R2020a. The main parameters of the ship are given in [
25].
The datasets are generated by carrying out simulations of maneuvers with a nonlinear whole-ship model, where the hydrodynamic derivatives are taken from Chislett and Tejsen [
25]. The training datasets are collected from simulated 5°/5°, 10°/10° and 35°/5° zigzag maneuvers at a sampling time interval of 0.5 s, and the sample size of each maneuver is 400. The parameters in LGPR are configured as follows: 0.9 for the crossover possibility, 0.1 for the mutation possibility, 30 for the evolutional generation, and 3 for the number of clusters
k. Meanwhile, a classic GPR without clustering analysis process and a neural network (NN) based on a back-propagation algorithm are also used to model and predict the ship steering motion for comparison purposes. NNs are a widely used nonparametric modeling method due to its universal approximation ability. In this study, the NN consists of three layers: an input layer, a hidden layer with 10 hidden nodes, and an output layer. The Levenberg–Marquardt algorithm is used as the optimization method for tuning the weights and the bias values in the NN, the mean-square error is utilized as the loss function, the hyperbolic tangent function is adopted as the activation function for the input layer and the hidden layer, and the linear transfer function is adopted as the activation function for the output layer. The maximum training iterations and the learning rate are set as 1000 and 0.1, respectively.
In LGPR, the k-means clustering analysis algorithm is applied to divide the training dataset automatically [
26], and the min–max normalization is carried out on the training dataset before performing the clustering analysis. In respect of the clustering analysis, the results are presented in
Figure 3 and
Table 1. The yellow dotted line in
Figure 3 represents the borders of the clusters.
The hyperparameters tuned and optimized by GA are presented in
Table 2, where
and
are the length scales of the covariance function of the input variables, i.e., yaw rate and rudder angle, respectively;
is the change magnitude of the output of the covariance function; and
is the standard deviation of the noise contained in the target value
y.
The generalization ability of the identified response model is validated by predicting the maneuvers which are excluded in the identification procedure, including 10°/5° and 15°/15° zigzag maneuvers, and random maneuvers. For the predicted maneuvers, the results are presented in
Figure 4,
Figure 5 and
Figure 6, where ‘RHM’ denotes the results simulated by the hydrodynamic model (here the whole-ship model) for ease of exposition.
Figure 7 presents the results of the comparison among GPR, LGPR and the NN in respect of the time consumed for prediction, while the results of the comparison in respect of the root-mean-square error (RMSE) values of the yaw rate are given in
Table 3, where
(deg/s) is calculated as
where
is the RHM value and
is the value predicted by the identified response model.
Figure 4a shows that in the predicted 10°/5° zigzag maneuver, the predicted yaw rates by GPR, LGPR and the NN are smaller than the RHM in the early stage. After 100 s, the predicted yaw rates by the three methods begin to fit well with the RHM. It can be seen from
Figure 4b that the accumulative prediction errors of the heading angle by the three methods are rather large, and those of the NN are the most noticeable.
Figure 5 shows that in the prediction of 15°/15° zigzag maneuver, the results of the yaw rate predicted by the three methods are smaller than the RHM during 0–50 s, while after 50 s the prediction results for the yaw rate are larger than the RHM, leading to the accumulative prediction errors for the heading angle. The heading angles predicted by the three methods are larger than the RHM later in the prediction, and those predicted by LGPR and the NN exhibit more noticeable deviations.
The response model is usually applied to develop the controller of ship steering motion. Generally, the rudder angle during the motion control operation is not as regular as that in the standard zigzag maneuvers, but rather random. Therefore, it is necessary to assess the prediction accuracy of the identified response model for the simulation of random rudder angle, as shown in
Figure 6a. In this study, the predictions of the random maneuver for short time (0–5 s), medium time (0–100 s) and long time (0–600 s) are performed, and the results are shown in
Figure 6b–d.
As can be seen from
Figure 6b, in the prediction of the short time random maneuver, the heading angles predicted by GPR and the NN show acceptable accuracy, while those predicted by LGPR exhibit rather distinct deviations during 3–5 s.
Figure 6c shows that in the prediction of the medium time random maneuver, the results of the three methods fit well with the RHM in the first 30 s. After 30 s the prediction results begin to show errors. The performances of the three methods are similar and GPR gives slightly better results after 70 s. One can find from
Figure 6d that in the prediction of the long time random maneuver, the prediction results of GPR and LGPR are in accord with the RHM overall, while those of the NN exhibit quite large deviations. These results indicate that the performances of the three methods are similar in the predictions of short time and medium time random maneuvers, while in the long time prediction the performance of the NN is not satisfactory, indicating that the nonparametric response model identified by GPR has better generalization ability.
As shown in
Table 3, the values of
Rr of the NN are larger than those of GPR and LGPR in the predictions of 10°/5° zigzag and random maneuvers, while in the prediction of 15°/15° zigzag maneuver, the
Rr values of the three methods are close.
Figure 7 demonstrates that LGPR is the least time-consuming of the three methods. These results reveal that LGPR is efficient, while the loss of prediction accuracy is not distinct. The improvement in computational efficiency and the guarantee of prediction accuracy can be attributed to the introduction of the clustering analysis method. The prediction result of the predictive sample is obtained on the cluster for which the center is closest to that of the predictive sample, rather than on the whole training dataset, thereby the computational burden is significantly reduced.
4.2. KVLCC2 Tanker Model with Experimental Data
To further verify the robustness of LGPR, the nonparametric modeling of the ship response model is conducted by using the experimental data of the KVLCC2 tanker model. The detailed information about this ship and the experimental data are given in SIMMAN 2008 Workshop [
27].
The training datasets are collected from the 10°/5°, 10°/10°, 30°/5° and 35°/5° zigzag maneuvers at a sampling time interval of 0.05 s, and the whole training datasets include 1288 samples. The parameters set in LGPR are the same as those set in the previous subsection. In respect of the clustering analysis, the results are presented in
Figure 8 and
Table 4. The yellow dotted line in
Figure 8 represents the borders of the clusters. The hyperparameters tuned and optimized by GA are presented in
Table 5.
Similar to the previous subsection, the maneuvers which are excluded in the identification procedure are predicted by the identified models, including 20°/10° (S), 20°/10° (P) and 20°/5° zigzag maneuvers, where ‘(S)’ and ‘(P)’ stand for the zigzag maneuvers started with the rudder turning to the starboard and port sides, respectively. GPR and the NN are used for comparison purposes and the structure of the NN is the same as that used in the previous subsection. The results of the predicted maneuvers are shown in
Figure 9,
Figure 10 and
Figure 11; the results of the comparison among GPR, LGPR and the NN in respect of time consumed for prediction are presented in
Figure 12; and the results of the comparison of RMSE values of yaw rate are given in
Table 6.
According to
Figure 9a and
Figure 11a, in the predictions of 20°/10° (S) and 20°/5° zigzag maneuvers, the yaw rate predicted by GPR, LGPR and the NN are smaller than the experimental data in the early stage of prediction. As the prediction progresses, the results of the predicted yaw rate become larger than the experimental data, leading to the deviations in the predictions of the heading angle.
Figure 9b and
Figure 11b show that the heading angle predicted by LGPR exhibits rather large deviations in the later stage of prediction, and the heading angle predicted by GPR shows higher accuracy than LGPR and the NN.
Figure 10b shows that in the prediction of 20°/10° (P) zigzag maneuver, the heading angle predicted by the NN exhibits distinct errors for the valley points. The prediction deviations of the nonparametric model identified by the NN may be attributed to the limited training samples. In general, a larger sample size is required by the NN to identify the mapping relationship between the input and output data [
28], while GPR is able to establish a rather robust model with a small sample size [
23].
From
Figure 12 and
Table 6, it can be seen that LGPR is able to give high computational efficiency and acceptable prediction accuracy when compared with GPR and the NN. Moreover,
Rr of the predicted maneuvers given in
Table 6 are of the order of 10
−1, which are larger than those given in
Table 3. This indicates that modeling based on the experimental data is more difficult than that based on the clean simulation data, since the experimental data contain noise caused by environmental interferences and measurement uncertainties.
In the above study of the two cases, the results predicted by the nonparametric response model exhibit rather distinct accumulative errors in the later stage of prediction. On the one hand, this may be attributed to the modeling errors of the identification algorithm. On the other hand, it may be attributed to the mathematical structure of the black box model (Equation (5)). In this study, the yaw acceleration is determined by yaw rate and rudder angle, which are derived from the Nomoto model. However, based on the mathematical model of 3-DoF ship maneuvering motion, the yaw acceleration is considered to be determined by surge speed, sway speed, yaw rate, and rudder angle [
15,
16,
18,
19,
22]. Equation (5) can be regarded as a simplified black box model in comparison with that used in the modeling of 3-DoF ship maneuvering motion. Therefore, the identification algorithm cannot capture the complete dynamic characteristics of ship steering motion from the training dataset constructed based on Equation (5). Nevertheless, the prediction accuracy of the nonparametric response model is acceptable, and the computational efficiency is expected to be higher than that of the nonparametric model of 3-DoF ship maneuvering motion because of neglecting the motion variables surge speed and sway speed.