Robust decision making for reactive collision avoidance of maritime autonomous surface ships

Lee, Paul

Thesis

Robust decision making for reactive collision avoidance of maritime autonomous surface ships

Download PDF

Creator

Lee, Paul

Rights statement

Strathclyde Thesis Copyright

Awarding institution

University of Strathclyde

Date of award

2024

Thesis identifier

T17068

Person Identifier (Local)

202074883

Qualification Level

Doctoral (Postgraduate)

Qualification Name

Doctor of Philosophy (PhD)

Department, School or Faculty

Department of Naval Architecture, Ocean and Marine Engineering

Abstract

The maritime industry is expected to utilize maritime autonomous ships (MASSs) to realise new sustainability levels, but their ability to make robust decisions without human intervention during safety-critical operations, such as collision avoidance, remains a challenge. This research aims to propose a holistic methodology that enhances the robustness of decision making for the reactive collision avoidance of a MASS from design to operation. In the first phase, the risk-informed design of the collision avoidance system (CAS) is conducted to reduce critical internal collision risk indices (CRIs). In the second phase, a digital twin environment is developed to simulate critical collision scenarios pertaining to internal CRIs-based dynamic safety domain, propeller-rudder coupled control, noisy perception sensor, critically proximate and unclassified static or dynamic obstacles. In the third phase, a decision-making agent is optimised for the reactive collision avoidance operation within the above scenarios in a deep reinforcement learning (DRL)-based training framework. In the fourth phase, a classification agent is optimised for the identification of static or dynamic obstacles based on real-time perception sensor data in a deep learning training framework. In the fifth phase, the robustness of decision-making is quantified, verified, and validated within and outside the investigated scenarios. Fault tree analysis, deep deterministic policy gradient algorithm, and feedforward neural networks are employed for the risk analysis, DRL-, and DL-based training frameworks, respectively. A high-speed single screw container ship and a light detection and ranging (LIDAR) sensor are considered for the reference ship. The results demonstrate that a risk-informed design conducted from a cost-benefit perspective leads up to 91% reduction of the CAS probability of failure from its baseline configuration. Robustness of decision-making agent for the reactive collision avoidance is verified even when the safety domain and LIDAR noise is increased by 20% and 132% from its maximum trained value, respectively, and the size of the critically proximate static obstacle and size of the static or dynamic obstacle is increased by 200% and decreased by 88.5% from its trained value, respectively. The ability of the decision-making agent to prioritise safety over efficiency depending on the size of the safety domain and LIDAR noise level, conduct sophisticated propeller-rudder coupled control manoeuvres when encountered with critically proximate static obstacles, and conduct efficient or navigational rules-compliant evasive manoeuvres when encountered with static or dynamic obstacles are noted. The accuracy of the classification agent is verified above 98.12% even in different LIDAR configurations. The novelty of this research stems from the introduction of a framework that enhancing the safety against the unknown-unknowns by increasing the robustness against critical CRIs, such as system probability of failure, uncertainty of sensor measurements, collision-inducing obstacles, and limited information. This research contributes towards the development of safe maritime autonomous systems that are capable of making robust decisions under safety-critical scenarios.

Advisor / supervisor