Self-learning based intelligent control of ship manoeuvring in narrow waters

Published statistics indicate that human error is a very significant factor attributed to ship accidents. The intelligent control approach developed and examined is thought to be capable of reducing the burden of ship operators and to provide assistance with effective decision making in the context of course keeping, track-keeping and ship collision avoidance. In each case the back-propagation method is used to provide real-time upgrade of the control system parameters. A single-input multi-output control strategy is adopted to cope with the large inertia of real ships and hence the slow responsiveness to rudder changes. For course-keeping and track-keeping an on-line trained ‘neurofuzzy’ based control scheme is proposed. For anti-collision control three aspects must be addressed. First fuzzy set interpretation is used to decide whether and when anti collision action is required. The membership functions associated with this step are associated with the traditional marine navigational concept of Time to Closest Point of Approach (TCPA) and a new introduced concept designated Relative Safety Time (RST). There is no attempt to address any psychological factors that may influence the human decision making process. To determine the RST capability required for the anti-collision model an innovative Adaptive Neurofuzzy Inference System (ANFIS) based network is proposed and applied. In this case the learning process is based on off-line training data and a hybrid-learning algorithm that combines the least squares and the back-propagation methods. Finally to identify an optimal anti-collision action a genetic algorithm (GA) approach is developed based on alternative ‘trial’ avoidance manoeuvres. Each of the developed controllers is examined using the Japanese Mathematical Manoeuvring Group (MMG) ship model for different ship types. The characteristics of the MMG model for each representative ship type is pre-adjusted to provide known turning circle characteristics. Where meaningful the new controllers are compared directly with previously developed neural network based controllers.