TL;DR
This paper tackles the challenge of accurately modeling ship maneuvering for autonomous docking. They propose using a global optimization algorithm (CMA-ES) to identify system parameters for a complex ship dynamics model (MMG) using free-running model test data, achieving better results than traditional methods with less data.
What Problem Does This Paper Address?
Autonomous docking and berthing of ships is a critical but complex task. To achieve this, we need highly accurate predictive models of how a ship will maneuver, especially in the challenging low-speed, high-interaction environment near a dock. Traditional methods for creating these models, like captive model tests (CMT), are time-consuming and can struggle with scale effects. System identification (SI) offers an alternative by learning model parameters from observed ship motion, but previous SI efforts have largely focused on simpler maneuvers like turning and zig-zag, not the intricate dynamics of berthing. This paper aims to bridge this gap by exploring how to effectively use SI with a sophisticated ship maneuvering model (MMG) for berthing scenarios, using readily available free-running test data.
Methods
The core of this research lies in applying the Covariance Matrix Adaptation Evolution Strategy (CMA-ES), a powerful global optimization algorithm, to tune the parameters of the Mathematical Maneuvering Group (MMG) model. The MMG model is a "model-rich" representation that captures various forces acting on a ship.
Here's what's novel:
- Berthing-Specific MMG: The authors extend the standard MMG model with sub-models specifically designed to handle low-speed and berthing maneuvers, which involve complex control inputs and ship motions (like reverse propeller thrust and large drift angles).
- System Identification (SI) with CMA-ES: Instead of relying solely on expensive captive model tests, they use CMA-ES to optimize the MMG model's parameters by minimizing the difference between simulated ship trajectories and actual free-running model test data.
- Objective Function Exploration: They investigate different choices for the objective function (what CMA-ES tries to minimize), comparing functions that focus on position, velocity, or a combination of both.
- Data Set Generation: They utilize free-running model test data, which is more readily available than captive model test data, and explore different combinations of maneuvers (random, turning, zig-zag) for training the model.
Datasets
The study uses data generated from free-running model tests of a VLCC model ship (Esso Osaka) conducted in the Inukai pond at Osaka University.
- Training Data: Three sets of training data were used, comprising different combinations of maneuvers:
- Train-R: Random maneuvers.
- Train-TR: Random and turning maneuvers.
- Train-TZR: Random, turning, and zig-zag maneuvers.
- Test Data: Data from random, turning, zig-zag, and berthing (starboard and port side) maneuvers were used for evaluation.
This approach differs from traditional methods that rely heavily on captive model tests. The use of random maneuvers in training is particularly noteworthy for its potential to create more robust models.
Key Insights and Ideas
- Random Maneuvers are Key: Using random maneuvers for training proved to be the most effective strategy for developing an accurate berthing model. This is likely because random inputs better cover the state-space and lead to more robust parameter estimation.
- Position and Velocity Matter: Objective functions that consider both the ship's position and velocity (like J2 in the paper) yielded better results than those focusing solely on position or velocity. This makes intuitive sense for berthing, where precise stopping and orientation are crucial.
- CMA-ES is Effective for Complex Models: The CMA-ES algorithm successfully navigated the complex, multi-modal landscape of the MMG model's parameters, finding good solutions even with a "model-rich" representation.
- Reduced Data Needs: The proposed SI approach requires significantly less data compared to traditional captive model tests, making it more practical and cost-effective.
- Overcoming Scale Effects: By using free-running model test data (or potentially full-scale data), the method can inherently account for scale effects, which is a major advantage over CMT.
Key Results
- The proposed system identification method using CMA-ES achieved better agreement with free-running model test data than the traditional EFD (Experimental Fluid Dynamics) model, especially for berthing maneuvers.
- The best performing models were achieved using the J2 objective function (position and velocity) with either Train-R (random only) or Train-TR (random and turning) data sets.
- The optimal model showed improved accuracy across various test data sets (random, turning, zig-zag, and berthing) compared to the EFD model.
- The method required fewer data amounts than a CMT-based scheme.
- The optimized model demonstrated numerical stability even when tested with control inputs (propeller revolution) exceeding the training data limits.
Limitations and Future Work
The authors acknowledge several limitations:
- Modular Structure Compromised: The MMG model's modularity is somewhat reduced. If the ship design changes, the entire data set may need to be reacquired, unlike CMT where only specific parameters might need re-evaluation.
- Parameter Interdependence: The optimization process can lead to parameter interdependence ("cancellation effect"), meaning obtained parameters might not be interpretable or usable independently.
- Stochastic Search Dependency: CMA-ES is a stochastic algorithm, meaning results can vary with the random seed. Multiple trials are necessary to ensure robustness.
- Model Complexity: While the MMG model is "model-rich," it still might not fully capture all complex hydrodynamic phenomena in extreme berthing scenarios (e.g., large drift angles affecting rudder force).
- Data Amount Dependency: The study used a constant amount of training data. Further investigation is needed to understand how varying the data quantity impacts accuracy and robustness.
Future work could involve:
- Improving the MMG model's degree of freedom to better capture complex hydrodynamics.
- Investigating the impact of different amounts of training data.
- Combining SI with direct CFD simulations for even more robust model construction.
- Exploring the application of this method to full-scale ship data.
Summary updated