Akademik Veri Yönetim Sistemi
Tez Türü: Doktora
Tezin Yürütüldüğü Kurum: Cranfield University, School of Aerospace, Transport and Manufacturing, Aerospace, İngiltere
Tez Danışmanı: Zeeshan A. Rana
Tezin Onay Tarihi: 2024
Tezin Dili: İngilizce
Özet:
Autonomous ground refuelling of civil aircraft is a critical operation that demands strict adherence to safety protocols and attention to detail. This is to avoid significant harm such as unintended ignition of fuel vapour, fuel spillage arising from procedural errors, leaks, aircraft tank venting or failure of pressurised fuel lines or their couplings. Additionally, the accumulation of a surface static charge on either an aircraft or its fuelling vehicle also poses risks. These hazards could lead to substantial damage to the aircraft and pose serious threats to the safety of passengers and crew. The introduction of automation and robotics has the potential to enhance the efficiency and safety of ground refuelling operations. This thesis proposes an autonomous ground refuelling framework that utilises computer vision and robotics to improve the refuelling process of civil aircraft. This thesis has shed light on the domain randomisation effects through a novel hybrid dataset of the pressurised refuelling adaptor, the ablation effects by developing the custom-designed neural network for real-time detection and localisation, the effectiveness of the vision based localisation framework by comparing it with the laser-based localisation system.
The framework utilises the Intel® RealSense™ D435 Stereo Depth Camera and the custom designed neural network to detect and localise the refuelling adaptor of the aircraft in three-dimensional space. The custom-designed neural network has been trained using the novel hybrid dataset of the pressurised refuelling adaptor incorporating real images from Boeing 737-400 aircraft and synthetic images generated by a physics-based simulator. By utilising the visual trajectory planning control framework, the position and orientation of the pressurised refuelling adaptor are procured from Boeing 737-400. Using this position and orientation information of the refuelling adaptor is used to guide the robotic manipulator to perform the refuelling operation autonomously.
The experiments demonstrate that the task-specific custom-developed neural network has reached 99.19% accuracy, 0.023% validation loss, 98.26% precision, 99.58% recall and 97.92% mAP@95% by surpassing well-known Transfer Leaning models such as EfficientNet-B0, VGG-16, ResNet-18, YOLOv4, YOLOv5s and YOLOv5x. Furthermore, it has been shown that the synthetic image dataset generated through a physics based simulator can be used in harmony with real images taken from Boeing 737-400 to train the custom-developed neural network to detect and locate the pressurised refuelling adaptor in real time. This is rather important as the development of a dataset containing only real images can be very expensive and time-consuming. The visual trajectory planning control framework has also demonstrated remarkable precision with errors as small as 0.67 mm at a distance of 57 cm from the refuelling adaptor. The integration of computer vision and robotics in the proposed framework provides increased precision, accuracy and efficiency while minimising the risks associated with human error. The framework’s potential to enhance safety and efficiency in ground refuelling operations aligns with the increasing trend towards automation and intelligent robots in various industrial circumstances.