An integrated INS/GNSS system with an attention-based hierarchical LSTM during GNSS outage

Navigating drones and estimating their future states for a prolonged duration are challenging without the Global Navigation Satellite System (GNSS) network. Our study proposes an Attention-based Hierarchical Long Short-Term Memory (AHLSTM) model to navigate the drone in GNSS-denied environments. Micro-Electro-Mechanical Sensors (MEMS) can provide navigation systems with inexpensive, accurate measurements. However, these accumulated errors of the sensors could introduce various uncertainties and noises in the state estimation of the drone. We suggested an architecture that consists of a set of Hierarchical LSTMs and an eventual attention mechanism to achieve multi-stage predictions in the long term, whose output is based on two consecutive layers of LSTMs. The proposed performance of the algorithm is evaluated using experimental data obtained from flight tests. The results show that using the suggested model leads to a 70% improvement in the long-term prediction of position and velocity compared to similar methods.