A Deep Learning Approach for Fault-Tolerant Data Fusion Applied to UAV Position and Orientation Estimation

Abstract

This work introduces a novel fault-tolerance technique for data fusion in Unmanned Aerial Vehicles (UAVs), designed to address sensor faults through a deep learning-based framework. Unlike traditional methods that rely on hardware redundancy, our approach leverages Long Short-Term Memory (LSTM) networks for state estimation and a moving average (MA) algorithm for fault detection. The novelty of our technique lies in its dual strategy: utilizing LSTMs to analyze residuals and detect errors, while the MA algorithm identifies faulty sensors by monitoring variations in sensor data. This method allows for effective error correction and system recovery by replacing faulty measurements with reliable ones, eliminating the need for a fault-free prediction model. The approach has been validated through offline testing on real sensor data from a hexarotor UAV with simulated faults, demonstrating its efficacy in maintaining robust UAV operations without resorting to redundant hardware solutions.