LEE- Uçak ve Uzay Mühendisliği Lisansüstü Programı
Bu topluluk için Kalıcı Uri
Gözat
Yazar "Koyuncu, Emre" ile LEE- Uçak ve Uzay Mühendisliği Lisansüstü Programı'a göz atma
Sayfa başına sonuç
Sıralama Seçenekleri
-
ÖgeData-driven delay estimation and anomaly detection: A study on European and Turkish air traffic(Graduate School, 2023-05-18) Aksoy, Muhammet ; Koyuncu, Emre ; 511201136 ; Aeronautical and Astronautical EngineeringAir traffic networks represent highly complex and interconnected physical systems. Unlike other transportation networks, air traffic is very heavily regulated and physically constrained. Although the airways and airspaces are somehow more flexible compared to land based transportation systems, the fact that aircrafts can only positioned on and operated by airports make them quite dependent on the operations of the airports. Air traffic is regulated to ensure safety, while also maintaining the throughput of travel from one location to another. While these regulations does a decent job on keeping the air travel safe and systematical, they fall short when there are disruptions among the network that hinders the air traffic. There are numerous reasons for disruptions in air transportation; weather conditions, accidents, capacity constraints, personnel strikes etc. Yet their negative effect to the air traffic is mostly the same: introducing delays. Due to the connected nature of the air traffic and airports, when a delay generating event occurs at one place, the other members of the network could experience the similar effects, if not at a larger degree. This delay propagation means there is a ripple effect through the network which can snowball the delay generations and cause very large congestions. To relieve the effects of delay generating events, air traffic federators regulate the air traffic in a reactionary way. This may include reducing the capacity on certain airports or airways, giving NOTAMs, holding aircrafts on the ground or in the air (with hold patterns). Since all these actions are \emph{reactionary}, they are set in place after the delays already propagates through the network since it is trivial to asses and quantify the propagations in a large and complex network system. This study hypothesis that if the air traffic network can be modeled so that the propagations can be accurately calculated, it becomes possible to take proactive actions instead of reactive ones. Proactive actions are significantly more important when there is a risk of snowballing and propagation. It allows to take action when the ill effects are still contained on fewer members with smaller intensities. This paves the way for a more effective and less costly approach. Hence, the study proposes a method with 3 main parts; first one is to model the air traffic network so that propagations can be quantified, second one is to estimate the parameters of this model to keep a short-sighted vision into the upcoming network state and third one is to come up with a comprehensive action generating model to find optimal proactive actions that can keep the delay spreading at minimum and improve system resiliency. The air traffic modeling part is done via adopting compartmental model from epidemiology. This model explains the tranmission of disease within a population. When it is applied to the physical network system, instead of disease and humans, the delay amount and aircrafts is used. Additionally with the meta population model, instead of considering aircrafts one by one, airports can be used as they are focused points of aircraft populations. By linking transmit rate to the flight frequency between airports and the recovery rate to the delay handling characteristics of the airport, The parameter estimation part is done by calculating the historic recovery rates of the airports and then using deep learning inference to predict the next time step's recovery rates. The other parameters of the air traffic model, such as the traffic flow, is already known before hand (flight plans). Therefore through the estimation of recovery rate the network state of the upcoming states can be accurately predicted. This prediction can then be fed to the action generating algorithm to make the most informed decision. The action generating algorithm therefore must fundamentally be a deterministic state to action mapper. Reinforcement learning approach is utilized to train this state to action mapper to make it capable of generating optimal decisions under a sufficiently large spectrum of conditions. The final part of this study concerns with anomalous flight detection in air traffic as these types of flights are one of the sources of disruptions in an air traffic network. Although flight paths naturally diverge from one another, they still adhere to a set of patterns that have been tested in various environments and are optimized for them. These patterns may or may not be simple, depending on a number of factors, such as airspace use, the cognitive complexity of controllers, the weather, and NOTAMs. It is a challenging task to accurately classify flights just by their trajectories into a desired set of categories based solely on its statistical properties because of the high variance. For this purpose, the study incorporates a statistical approach that takes into account the time-based characteristics of the flight trajectories to determine whether they are abnormal or not. This statistical method with LSTM autoencoders makes it possible to train the model with historical data and quickly predict the flight class, taking into account the time-based characteristics of a flight trajectory. LSTM autoencoders can capture the class of a flight with relatively shorter time windows (16 second intervals). Therefore the air space can be periodically sweeped for anomalies while the network model and action algorithm runs in parallel. The obtained results demonstrate that the suggested architecture is quite capable of classifying abnormal flight trajectories as it successfully detects simulated fighter aircraft trajectories in airspaces with high commercial flight density. With the applications of deep learning and reinforcement learning, this whole methodology ensembles is largely data-driven, however the introduction of the compartmental model from epidemiology lays out a strong and accurate mathematical formula to support these data-centric approach. As the results suggests, The whole network's resiliency, i.e. its ability to keep delays from spreading and absorbing them, significantly increases when the optimal actions are reflected on the parameters. Additionally with the help of unsupervised learning, anomalous flights are also detected and represented as a disruption source to the network. Possible biases and shortcomings due to the data-driven approach is recognized throughout the study yet the overall method is deemed to be of significant importance in terms of managing resiliency through air traffic networks.
-
ÖgeNovel polynomial shaping method for impact time and angle guidance law designs: Bézier curve approach(Graduate School, 2023-05-15) Çatak, Akın ; Koyuncu, Emre ; 511201107 ; Aeronautical and Astronautical EngineeringOne of the topics that intensive research is conducted on in aviation and space technologies is the development of algorithms that will direct an aircraft from one point to another. While it is possible to plan the route of the aircraft and develop control algorithms that will follow that route for vehicles with low speeds such as Unmanned Air Vehicles, guidance laws are used for vehicles with high speed and maneuverability such as missiles. If the requirements from engagement are to increase the damage to be given to the target or to make a salvo attack with multiple missiles,advanced guidance laws should be designed. The engagement requirements in the literature have been examined and it has been seen that impact time and impact angle of the missile are the two most important constraints. When the solution methods proposed for these problems are examined from literature, it is seen that analytical results are obtained using polynomial shaping methods in addition to control methods such as nonlinear control and optimal control. Within these studies, the studies created with polynomial shaping methods have been examined in depth as they form the basis of this study. As a result, the goal is to shape the Line of Sight as a polynomial and to obtain a guidance law that can control the impact time and angle together with the help of the obtained polynomial. B\'ezier curves, a polynomial shaping method, were used to shape the Line of Sight angle polynomial. In this method, polynomials are calculated as linear combinations of control points and basis functions. After analyzing engagement dynamics and making various adjustments, the resulting guidance law requires the second derivative of the LOS angle polynomial. Therefore, the curve to be created must be a curve that can be differentiated twice. By definition, this curve yields results for knot values between zero and one. Therefore, the curve to be created is arranged to provide the total impact time. The final value of the curve should be equal to the desired Line of Sight angle. The generated Line of Sight curve is in a nonlinear relation with the range and range rate values in the proposed guidance law. Therefore, one control point of the B\'ezier the curve is left free and calculated using an offline algorithm. The algorithm finds the parameter that minimizes the impact time error time using the upper and lower limits given as initial values with iterative simulations. With good initial estimations, the algorithm converges to the result in a few steps. The B\'ezier curve is completed after the determination of the final control point.Thus, the proposed method is completed with the completed curve. The proposed guidance law has been tested in several cases. In the first two cases impact time and angle capabilities are tested, it was observed that the impact time could be achieved for certain intervals for each engagement.The reason for this fact was stated to be the convergence of the look angle values to 90 degrees. The performance of the designed method was demonstrated in a salvo attack scenario where missiles with different initial flight path angles and at different initial ranges simultaneously approached the target at the same final line of sight angle. It was also demonstrated through simulation that the method developed for moving targets works with decent performance if the target acceleration value is known. In a cost analysis comparing the proposed method with a method in the literature, the proposed method was found to be more efficient. The proposed guidance law however, has limitations. The offline nature of the algorithm means that disturbances and uncertainties may affect its performance. Additionally, the initial guesses for the algorithm and the tuning parameter are areas for further development. In the future works of this areas are possible in addition to extensions to the 3D implementations. A guidance law that can adjust the impact time is also designed by shaping the range using B\'ezier curves. This guidance law does not require an offline algorithm .Two cases were conducted to test the impact time performance of the given method. The first one evaluates the impact time intervals to be achieved while the second one for testing the salvo attack performances. The simulations shows the effectiveness of the proposed law with high impact time intervals. Finally, a new method is proposed based on one of the guidance laws from the recent literature. In this method, Bézier curves are used to add the ability of controlling the impact time of the Impact Vector Guidance law which has only impact angle capabilities. In the proposed method, the missile follows a virtual target instead of the real target and the virtual target coincides with the real target at the desired impact time. The virtual target is designed to move on a Bézier curve, created using the boundary conditions of the engagement. To formulate the position of the virtual target along the curve, an S-shaped (sigmoid) function is proposed. This is necessary to remove the sudden jumps in the movement of the virtual target and ensures that it coincides with the real target at the end of the engagement. The alpha variable, which is specific to each case, is used in the design phase of the S-shaped function and is determined by an offline algorithm before the engagement. The lower value for alpha is fed to algorithm, then the algorithm increases it until the impact time is within the given tolerance. The alpha determines the saddle point of the S-shaped function. This method, which is still in the earlier phases, shows satisfactory results. In future studies, obtaining the online counterpart of the algorithm is possible. As an another future study, the offline algorithm can be used as an initial guesses for nonlinear receding horizon control method to deal with moving targets while compensating the disturbances and uncertainties.
-
ÖgeOptimization based-control of cooperative and noncooperative multi aircraft systems( 2020) Başpınar, Barış ; Koyuncu, Emre ; 625456 ; Uçak ve Uzay MühendisliğiIn this thesis, we mainly focus on developing methods that ensure autonomous control of cooperative and noncooperative multi-aircraft systems. Particularly, we focus on aerial combat, air traffic control problem, and control of multiple UAVs. We propose two different optimization-based approaches and their implementations with civil and military applications. In the first method, we benefit from hybrid system theory to present the input space of decision process. Then, using a problem specific evaluation strategy, we formulate an optimization problem in the form of integer/linear programming to generate optimal strategy. As a second approach, we design a method that generates control inputs as continuous real valued functions instead of predefined maneuvers. In this case, we benefit from differential flatness theory and flatness-based control. We construct optimization problems in the form of mixed-integer linear programming (MILP) and non-convex optimization problem. In both methods, we also benefit from game theory when there are competitive decision makers. We give the details of the approaches for both civil and military applications. We present the details of the hybrid maneuver-based method for air-to-air combat. We use the performance parameters of F-16 to model the aircraft for military applications. Using hybrid system theory, we describe the basic and advanced fighter maneuvers. These maneuvers present the input space of the aerial combat. We define a set of metrics to present the air superiority. Then, the optimal strategy generation procedure is formulated as a linear program. Afterwards, we use the similar maneuver-based optimization approach to model the decision process of the air traffic control operator. We mainly focus on providing a scalable and fully automated ATC system and redetermining the airspace capacity via the developed ATC system. Firstly, we present an aircraft model for civil aviation applications and describe guidance algorithms for trajectory tracking. These model and algorithms are used to simulate and predict the motion of the aircraft. Then, ATCo's interventions are modelled as a set of maneuvers. We propose a mapping process to improve the performance of separation assurance and formulate an integer linear programming (ILP) that benefits from the mapping process to ensure the safety in the airspace. Thereafter, we propose a method to redetermine the airspace capacity. We create a stochastic traffic environment to simulate traffics at different complexities and define breaking point of an airspace with regards to different metrics. The approach is validated on real air traffic data for en-route airspace, and it is shown that the designed ATC system can manage traffic much denser than current traffic. As a second approach, we develop a method that generates control inputs as continuous real valued functions instead of predefined maneuvers. It is also an optimization-based approach. Firstly, we focus on control of multi-aircraft systems. We utilize the STL specifications to encode the missions of the multiple aircraft. We benefit from differential flatness theory to construct a mixed-integer linear programming (MILP) that generates optimal trajectories for satisfying the STL specifications and performance constraints. We utilize air traffic control tasks to illustrate our approach. We present a realistic nonlinear aircraft model as a partially differentially flat system and apply the proposed method on managing approach control and solving the arrival sequencing problem. We also simulate a case study with a quadrotor fleet to show that the method can be used with different multi-agent systems. Afterwards, we use the similar flatness-based optimization approach to solve the aerial combat problem. In this case, we benefit from differential flatness, curve parametrization, game theory and receding horizon control. We present the flat description of aircraft dynamics for military applications. We parametrize the aircraft trajectories in terms of flat outputs. By the help of game theory, the aerial combat is modeled as an optimization problem with regards to the parametrized trajectories. This method allows the presentation of the problem in a lower dimensional space with all given and dynamical constraints. Therefore, it speeds up the strategy generation process. The optimization problem is solved with a moving time horizon scheme to generate optimal combat strategies. We demonstrate the method with the aerial combats between two UAVs. We show the success of the method through two different scenarios.