Modeling and control of parachute cargo landing systems using deep learning methods

In this thesis, the development of a parachute-controlled descent system for cargo delivery in military operations and the study of its simulation in a computer environment were investigated. The system was first mathematically modeled, and the guidance controls were addressed by transferring the model to the MATLAB Simulink environment. The developed controllers were used to control the system at the desired yaw angle and yaw rates. In the next stage, efforts were focused on hardware and software integration tests for the system. At this stage, the system was modeled in the Gazebo environment and the necessary software for ROS integration was developed. The Pixhawk autopilot card was used for data collection, communication, and motor controls. Here, the open-source software known as PX4 was integrated into the system's simulation environment and the first flight tests were carried out in Gazebo. The simulation results were used as input in system identification and model predictive control studies. In the system identification studies, the performance of the NARX neural network was obtained and analyzed parametrically. As a result of the review, it was seen that a one-layer model consisting of 5 neurons was sufficient. Finally, the developed algorithms were tested with real flight data using simulation data. Flight tests were carried out using two different environments and system variables. By adding online training methods to the controller algorithm, it was aimed to keep the developed system model continuously updated. In this direction, the trained model has been able to achieve an average estimation of 12 degrees in attack angle on flight data for different environmental conditions and system weight. Real-time controller studies were carried out and the system's descent to the desired target position was achieved with an error of 3m in windless environment and 7m in windy environment.