DEVELOPMENT OF AIRCRAFT’S AUTOPILOT SYSTEM DURING TAKE-OFF PHASE BASED ON ARTIFICIAL NEURAL NETWORK

• Main Author: Iqbal Yahya, Ahmad
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/73218
• Institution: Institut Teknologi Bandung

In aircraft flight, one of the most critical and prone-to-issues phases is the take-off phase. One way to reduce the risk level during this phase is by having a system that assists pilots during take-off. Many programs have been developed to provide such functionality, such as autopilot. However, the current autopilot programs are predominantly based on hard-coded programs, which are limited in scope and flexibility. Therefore, this research aims to explore alternative methods in developing autopilot systems, specifically through deep learning. This method will result in soft-coded programs that are more flexible and capable of broader coverage. The use of deep learning in this research is expected to generate a control model that can mimic the pilot's behavior using flight data that recorded during actual flights condition and pilot’s control inputs. The data will be obtained from Quick Access Recorder of a real world flight as well as artificially generated data from X-Plane simulation. The control model will be developed using the Boeing 747 and Cirrus Vision SF-50 aircraft. The control model will also be trained to handle crosswind disturbances during flight. The results obtained from this research include a set of control models that are capable of performing take-off procedures under crosswind conditions. The control models are even able to fly the aircraft in conditions beyond the training data, indicating that the resulting models are robust. However, it is also found that models that perform well in statistical metrics do not necessarily perform better when implemented in aircraft’s control system, as there are other factors involved. Additionally, in the development of these control models, it is observed that the type, quality, and preprocessing approach of the training dataset have the most significant impact on the model's performance.