Is this project an undergraduate, graduate, or faculty project?
Graduate
Project Type
group
Campus
Daytona Beach
Authors' Class Standing
Shlok Misra, Graduate Student Tanish Jain, Graduate Student
Lead Presenter's Name
Shlok Misra
Faculty Mentor Name
Dr. Dothang Truong
Abstract
Unstabilized approaches are a major hazard for general aviation aircraft. An unstabilized approach can lead to runway excursions, structural damage on touchdown, or even Controlled Flight into Terrain (CFIT). The Aircraft Owners and Pilots Association reported that 3,257 general aviation accidents from 2009-2019 occurred during the landing phase of a flight. The advancement of machine learning technology offers the opportunity to develop low-cost and easily adaptable technology. This research is aimed at developing machine learning-based predictive warnings for pilots to abort an unstabilized approach and execute a go-around maneuver. As the first step, we collected feature-rich flight data which could be useful for making predictions of unstabilized approaches. The data utilized for the model preparation was derived from the Flight Data Monitoring (FDM) program of a Part-141 Flight Training Organization. As a first step of preprocessing, we decided to extract only the variables that would be determining factors when predicting approach stability, based on the developed criteria. Additionally, we structured the data by separating it into matrices corresponding to exactly one flight - defined as the period from one take-off through the subsequent landing - determined by the change of altitude, airspeed, and engine power variables. We will use deep neural networks to train our machine learning model to predict unstabilized approaches. Since the data is structured with data points corresponding to every second of the flight, i.e., it is time-series, we will use a Recurrent Neural Network which is specifically adept at modeling time-series data. To develop our model, we will use an 85% training set, a 5% development set, and a 10% testing set split for our complete dataset comprising approximately 42,000 flights. The deep neural network architecture will be designed using the Tensorflow 2 framework. The model developed in this project will be a low-cost, objective decision-making aid for pilots that will improve general aviation safety. The model will be integrable into avionics systems that are used by general aviation pilots, such as the Garmin G1000®, for their aircraft.
Did this research project receive funding support (Spark, SURF, Research Abroad, Student Internal Grants, Collaborative, Climbing, or Ignite Grants) from the Office of Undergraduate Research?
No
Machine Learning-based Live Predictive Warnings for Unstabilized Approaches in Aircraft
Unstabilized approaches are a major hazard for general aviation aircraft. An unstabilized approach can lead to runway excursions, structural damage on touchdown, or even Controlled Flight into Terrain (CFIT). The Aircraft Owners and Pilots Association reported that 3,257 general aviation accidents from 2009-2019 occurred during the landing phase of a flight. The advancement of machine learning technology offers the opportunity to develop low-cost and easily adaptable technology. This research is aimed at developing machine learning-based predictive warnings for pilots to abort an unstabilized approach and execute a go-around maneuver. As the first step, we collected feature-rich flight data which could be useful for making predictions of unstabilized approaches. The data utilized for the model preparation was derived from the Flight Data Monitoring (FDM) program of a Part-141 Flight Training Organization. As a first step of preprocessing, we decided to extract only the variables that would be determining factors when predicting approach stability, based on the developed criteria. Additionally, we structured the data by separating it into matrices corresponding to exactly one flight - defined as the period from one take-off through the subsequent landing - determined by the change of altitude, airspeed, and engine power variables. We will use deep neural networks to train our machine learning model to predict unstabilized approaches. Since the data is structured with data points corresponding to every second of the flight, i.e., it is time-series, we will use a Recurrent Neural Network which is specifically adept at modeling time-series data. To develop our model, we will use an 85% training set, a 5% development set, and a 10% testing set split for our complete dataset comprising approximately 42,000 flights. The deep neural network architecture will be designed using the Tensorflow 2 framework. The model developed in this project will be a low-cost, objective decision-making aid for pilots that will improve general aviation safety. The model will be integrable into avionics systems that are used by general aviation pilots, such as the Garmin G1000®, for their aircraft.