There are two main advantages of using machine and deep learning based systems over the traditional hard coded systems. Second, the machine and deep learning-based systems can adapt to different circumstances through retraining based on collected data.
Introduction
Deep Learning
- Types of Deep Learning Layers
- Activation Function and Weight Initialization Methods
- Optimizers
- Training Dataset Scaling
- Transfer Learning
One of the common activation functions is the tangent hyperbolic (tanh) shown in the figure. For easier understanding of gradient descent, an illustration of gradient descent is shown in the figure.
Reinforcement Learning
A detailed diagram of the proposed K-means sea state selection method is shown in Figure 65. Structure of a PID feedback system with proposed motion prediction control.
Machine Learning-Based Mooring Line Tension Prediction System
Brief Comparison Between Conventional and Proposed Mooring Line Tension Prediction
A brief diagram of the conventional ANN-based mooring stress prediction system is shown in Figure. A brief diagram of the proposed DL-based mooring tension prediction system is shown in Figure 36. The use of the proposed K-means based seawater system state selection method instead of the manual sea state selection and the use of the proposed hybrid neural network architecture instead of the regular ANN.
The details of the proposed K-means based sea state selection method and hybrid neural network architecture are covered in later subsections.
Proposed K-Means-Based Sea State Selection Method
- Padding
- K-Means
- K-Means-Based Monte Carlo Method
- Feature Vector Generation
- Clustering of Relevant Sampled Sea States with K-Means
By using the valid sea state filled area, data loss can be prevented when using the K-means based Monte Carlo method. Fig. 40 Applying padding to the wave dispersion diagram. 44, the effectiveness of the K-means-based Monte Carlo method is clearly shown compared to the typical Monte Carlo method in the valid area of the packed sea state.
43Application of the K-means based Monte Carlo method to the filled valid sea state area. a) Typical Monte Carlo method (b) K-means based Monte Carlo method Fig. 44Comparison between the typical Monte Carlo method and the K-means based Monte.
Proposed Hybrid Neural Network Architecture
- Architecture
- Training Procedure
The proposed soft-binary attention module determines how much attention should be paid to the general Hs-focused NN model and the low-Hs-focused NN model given a joint vector of and . The architectures of the general Hs-focused NN model, the low-Hs-focused NN model, and the soft binary attention module are shown in Figs. The general Hs-focused NN model and the low-Hs-focused NN model are trained before training the proposed hybrid neural network architecture.
Set the general Hs-centered NN model and the low-Hs-centered NN model to not trainable.
Simulation and Result Discussion
- Simulation Conditions
- Overall Hs-focused NN model
- Effectiveness of Batch Normalization
- Low Hs-focused NN model
- Proposed Hybrid Neural Network Architecture
65, a structure of the PID feedback system with the proposed motion predictive control is presented. The essence of the proposed motion predictive control lies in the use of instead of. 75 where most of the peaks of the 'with the proposed motion predictive control' in , , and.
87 Example of wave motion load time history (sea state: very rough, environmental direction to a ship: 40°).
Motion Predictive Control for DPS Using Predicted Drifted Ship Position Based on Deep
PID Feed-Back System and Wind Feed-Forward System
As mentioned earlier, the conventional PID feedback system in the DPS often adopts the wind feed system. It should be noted that the terms in the square brackets denote the PID feedback system and − ̂ ( ) denotes the wind forward system. A structure of the PID feedback system with the wind feed system is shown in Fig.
64 indicates a vector of the estimated wind loads in the wave, sway and yaw directions.
Proposed Motion Predictive Control
For weight initialization in LSTM, a Xavier initialization method is used (Glorot and Bengio, 2010). As briefly introduced in the introductory chapter, the training of the proposed predictive motion control is performed using the online machine learning system, playback buffer, and real-time normalization method. In the training process of the proposed motion predictive control, online machine learning system, playback buffer and real-time normalization method play important roles.
Therefore, in the proposed motion prediction control, the DNN must make accurate predictions for the position of the relying ship under different sea environment conditions.
Numerical Modeling of Target Ship’s Behavior
- Target Ship and DPS
- Equation of Motion of Target Ship
It is well known that the training process can be very inefficient if the feature distributions in the training data are very different, therefore, normalization of the training data is considered to be one of the critical steps in neural network training (Lee, Lee and Seo, 2020a). 68 Adjusting the azimuth thrusters Table 12 Locations and thrust limits of the azimuth thrusters Azimuth. There are three types of environmental loads that contribute to the rolling motion of a ship, such as wind, current and wave loads.
The wind and current loads can be expressed as Eq. where and are wind and current loads, respectively, , , are the wave, sway and yaw coefficients relative to the wind or current direction relative to a ship's course, denotes density, is a relative wind or current speed past a ship, , are the exposed areas above the waterline for and the submerged areas for. where is the number of the regular wave components for describing irregular sea state, Redraws the real part of a complex number, QTF is the quadratic transfer function of wave driving which can be obtained from a diffraction analysis in the frequency domain of the target ship, is the direction of the regular wave component in relative to the ship's heading, , , , is the wave period, amplitude, frequency and phase delay of the regular wave component. 109) shows that the QTF is applied to each pair of the regular wave components to give each pair's contribution to the wave driving load.
Effectiveness of Proposed Algorithms
- Simulation Conditions
- Types of Deep Learning Layers
- Real-Time Normalization Method
- Replay Buffer
To train the adaptive fine-tuning systems, they are trained in 12-hour time steps in the simulation. The overall performance of the adaptive fine-tuning system is presented in terms of the station holder performance and control efficiency. The overall performance evaluation of the adaptive fine-tuning system compared to the PID controller with the fixed base gain is shown in Fig.
Set the plant performance history data with the new initial positions from the training dataset←.
Simulation and Result Discussion
- Simulation Under One Environmental Condition
- Simulation Under Two Different Sequential Environmental Conditions
Reinforcement Learning-Based Adaptive PID Controller for DPS
Target Ship and DPS
- PID Control in DPS
- Hydrodynamics Associated with a Drifting Motion of a Ship
The main dimension of the target ship and its coordinate system are shown in table 19 and fig. The ship's global and local positions are in meters and denoted as ( , ) and ( , ) respectively. 82, the flow of the control system is as follows: First, calculate from and.
This subchapter is intended to provide insight into the environmental loads that contribute to the ship's drift motion, and how the DP thrust counteracts the environmental loads.
Proposed Adaptive Fine-Tuning System for PID Gains in DPS
Update-gate for the integral of the errors determines how much error to update for the integral. While a ship is driven a lot due to the very large wave drift load, the integral of the error also increases drastically. Due to the drastically increased integral of the error over many drifting motions, the large rebound ship motions are caused.
The large rebound motions of the ship, caused by the increased integral of the fault, are illustrated in Figure 2.
Simulation Results
- Effectiveness of the Proposed Adaptive Fine-Tuning System
- Overall Performance Assessment
First, the effects of using the adaptive P, D gains are presented with ship motion and gain time histories in Figure. Second, the effects of the adaptive I gain and the update gate on the integral of the errors are analyzed in In Figure 90, the gain histories are normalized by dividing by , and the training of the adaptive fine-tuning systems is performed for 12 hour time steps in the simulation.
For the training of the adaptive fine-tuning system, training is given for 12-hour time steps in each environmental condition separately.
Discussion
The basic architecture of the neural network used in this study is illustrated in Fig. The number of hidden layers and the size of the neural network will be addressed in Chapter 6.3.4. The loss function history and plant performance history with initial positions of the ship from the training dataset with respect to the number of hidden layers are shown in Fig.
The significance of the application of the BN is shown and discussed in the next subchapter.
Application of Recent Developments in Deep Learning To ANN-based Automatic Berthing
Mathematical Model of Ship Maneuvering
- Mathematical Model for Ship-Maneuvering Problem
- Modeling of Propeller and Rudder
95, and are the actual positions of the ship in meters, and and are the normalized positions of , based on the length of the ship. In the ship maneuvering problem, planar motion includes only pitch, sway, and yaw motions, which are considered as motions with six degrees of freedom. Due to the kinematics of the rudder and engine mechanical system, time delays may occur between control commands and the resulting mechanical response.
First, the equation that takes into account the time delay effect for the RPS of the propeller is Eq.
Artificial Neural Network and Important Factors in Training the Network
- Artificial Neural Network
- Optimizer
- Input Data Scaling
- Number of Hidden Layers
- Overfitting Prevention
The overall flowchart of the automatic shipyard with a neural network is shown in fig. 98, the left side of the dotted line is the preparation step of the neural network model, and the right side is a processing loop with the trained neural network. The number of hidden layers is closely related to the system capacity of the neural network model.
The poor performance with the overfitting problem is most evident when new input data is extrapolated rather than interpolated by the neural network.
Application of Recent Developments in Deep Learning to Automatic Berthing
100, and the details of the training models related to the input data scaling method are listed in Table 27. Then, the effect of the number of hidden layers is analyzed based on the history of the loss function and the history of the anchoring performance with initial positions from the training data. The loss function and anchoring performance histories, and the details of the training models with respect to the use of BN, are shown in Fig.
The machine learning-based mooring tension prediction system mainly consists of the proposed K-means-based sea state selection method and the proposed hybrid neural network architecture.
Simulation and Result Discussion
Conclusion
Motion Predictive Control for DPS Using Predicted Drifted Ship Position Based on Deep
The conventional PID feedback system for the DPS is integrated with the proposed motion predictive control to improve the station keeping performance by reducing the magnitude of the ship motion movements. The proposed predictive motion control is designed based on DL layer, dropout, online machine learning system and playback buffer. Finally, the PID feedback system with the wind feedback system and the PID feedback system with the proposed predictive motion control were compared in simulations.
The simulation results showed that the proposed motion predictive control resulted in better station keeping performance than the wind feed-forward system with the same thrust consumption.
Reinforcement Learning-Based Adaptive PID Controller for DPS
However, the proposed real-time normalization method continuously updates the normalization factors with the incoming data to the replay buffer, which allows the normalizer to properly and adaptively normalize training data with new data distribution from varying environmental conditions. In the simulation results, it was shown that the proposed adaptive fine-tuning system could learn the efficient adaptive gain tuning strategy, which results in better station keeping performance without deteriorating control efficiency. The discussion section discusses the advantages and limitations of the proposed system, as well as the scope for further research.
Application of Recent Developments in Deep Learning to ANN-Based Automatic Berthing
1971) 'The Terebel dynamic positioning system - Results of five years of field work and experiments', i Proceedings of the Annual Offshore Technology Conference, s. 2015) 'Bølgefiltrering og dynamisk positionering af marinefartøjer. The Society of Naval Architects of Korea, 54(2), s. 2009) 'Autotuning Aspects for Dynamic Positioning Systems', i 8. IFAC International Conference on Maneuvering and Control of Marine Craft. 2005) 'Multivariable PID tuning of dynamic ship positioning control systems', Journal of Marine Engineering and Technology, 4(2), s. 2007) 'Design of hybrid controller for dynamic positioning from calm to extreme sea conditions', Automatica.
Springer Science and Business Media LLC, 8(3–4), p. 2006) 'Refining the DP Observation Circle', in MTS Dynamic Position Conference.
