LSTM-Kalman Filter-Based Multi-Sensor Signal Fusion for UAV Altitude Prediction in Non-Gaussian Environments
Nuo Li, Qiang Miao · Traitement du signal · 2025
To address altitude estimation inaccuracies in Unmanned Aerial Vehicles (UAVs) under non-Gaussian noise and intermittent sensor failures, this paper proposes a Long Short-Term Memory (LSTM)-Kalman cooperative architecture that establishes symbiotic interaction between deep feature extraction and physical filtering.The core innovation lies in bidirectional cyclic learning: LSTM layers distill temporal noise patterns while Kalman modules inject state-space constraints through differentiable projection.A manifold interpolation mechanism resolves multi-rate signal mismatches, utilizing LSTM-derived coherence weights to guide Lie group synchronization for phase distortion suppression.The framework incorporates a fractal-aware decoupling network where LSTM cells generate adaptive masks, dynamically separating Gaussian/non-Gaussian components to reconstruct Kalman gain rules.Experimental validation demonstrates the architecture's superiority in balancing physical consistency and learning capability, providing a novel paradigm for robust navigation signal fusion under complex noise conditions.