Neural Network Aided Drone Estimation in GPS-Denied Environments

While drones typically navigate using GPS, there are circumstances which cause GPS signals to be interrupted. Data from inertial measurement units (IMU) can be used to estimate position and attitude, but the noise and bias present in these measurements typically lead to estimations that suffer from drift. Neural networks have shown the ability to learn and recognize many complex patterns in a variety of applications. The goal is to use the capabilities of neural networks to develop an estimation method which gives better position and attitude predictions than traditional mathematical models when only IMU measurements are available. This project covers the development, testing, and results of various neural network-based estimation architectures. Architectures attempted range from pure neural network models with Long Short-Term Memory (LSTM) recurrent neural networks as the basis to architectures that attempt to fuse the benefits of neural networks and physics based mathematical models. The varying results of the architectures attempted sheds light on the potential applications of neural networks for guidance, navigation, and control.