The rapid expansion of the air transport industry, particularly at regional airports, necessitates the development of cost-effective trajectory prediction systems to manage increasing traffic demands. Also, a more precise planning at airports, e.g. for less resource intensive turnaround processes, is needed due to cost pressure. This relies on early knowledge of deviations from original flight schedule (scheduled time of arrival, STA). Early knowledge about changes in estimated time of arrival (ETA) is required for timely adjustment of the planning. So, accurate prediction of aircraft trajectories and ETA is a key operational requirement for regional and smaller airports that often lack advanced decision-support systems. In such environments, even small timing deviations can propagate through arrival and departure sequences (e.g., due to tight turnaround windows), affecting scheduling efficiency, runway utilization and turnaround processes. This work investigates the adaptation of a Social LSTM network (originally designed for pedestrian-motion prediction) to model aircraft lateral trajectory evolution and ETA forecasting. The proposed framework learns “interaction patterns between nearby aircraft”. In a second version, it learns as well the influence of meteorological conditions represented by Feature Joint Cubes. Experimental results on ADS-B data from the NRW region demonstrate that the Social LSTM reduces mean positional and ETA errors compared with a baseline Vanilla LSTM, particularly in complex traffic or weather situations.
The brushless DC (BLDC) machines are widely used for unmanned aerial vehicles (UAVs) and small aircraft as an electric propulsion motor. Switch faults in BLDC motors can lead to significant performance degradation, operational disruptions, and increased maintenance costs, particularly in small aircraft applications. This study examines the application of machine learning techniques for detecting switch faults in BLDC propulsion motor drive systems, with a focus on small aircraft, such as UAVs. The proposed approach involves data collection from the propulsion motor drive under both normal and faulty conditions. To simplify the data set and measurement process, only battery output current has been measured to detect an inverter switch fault based on the convolutional neural network (CNN) algorithm. To compare the detection performances, several machine learning algorithms, including CNN, have been employed, and the accuracy of fault detection and classification has been measured. This paper highlights the potential of integrating advanced data analytics and machine learning to improve the fault diagnosis processes of electric propulsion motor drive systems.
Unmanned Aerial Vehicles (UAVs), or drones, bring many opportunities to replace humans in 3D missions: missions which are dull, dirty or dangerous. One example is guiding animals out of dangerous situations, for example close to a forest fire or an accident where toxic chemicals leak into the environment. Since animals do not understand that the drones are intended to help them, they will seek to evade them and this can be used to herd them to safety. Automated shepherding has, to our knowledge, not yet reached the maturity level required to perform such an animal rescue mission. The goal of this paper is to review, select and adapt algorithms in pursuit of this purpose. We examined swarm UAV deployment covering: optimal distribution around the target animal, maintaining it when performing the required task and how to actively evade obstacles en route. This led to the conclusion, that of the presented algorithms, the Inverse Cumulative Density Function (ICDF) is the most suitable for the optimal dispersion of UAVs, while A*, Greedy Best First Search (GBFS) and the potential field algorithm are the most appropriate algorithms to determine suitable route trajectories during such guidance.
Conventional trajectory generation remains inadequate for high-accuracy aircraft guidance because Dubins- and clothoid-based turn transitions contain discontinuities in curvature or curvature derivative. As a result, even when high-order guidance laws are used to reduce tracking error, the reference trajectory itself does not provide suitable higher-order feedforward commands. Existing solutions are also unsatisfactory because they either recover higher-order smoothness only approximately through post-processing, rely on numerical optimization, which is unsuitable for deterministic airborne systems, or do not actively account for aircraft dynamic constraints. This paper closes these gaps with a closed-form method for generating three-dimensional turn trajectories as ninth-order Bézier curves. The curve is constructed analytically from smoothness requirements and aircraft dynamic constraints, yielding a high-order reference trajectory that is both dynamically consistent and deterministic. The methods presented are integrated into an operational trajectory generation system and validated analytically and through flight testing on a 700 kg medical evacuation eVTOL aircraft.
