Drone Attitude and Position Prediction via Stacked Hybrid Deep Learning Model for Massive MIMO Applications

Abstract

This paper presents a novel stacked hybrid deep learning model for real-time prediction of drone attitude and position, specifically designed to support applications in massive Multiple-Input Multiple-Output (MIMO) systems. The proposed model integrates Long Short-Term Memory (LSTM) networks, autoencoders, and attention mechanisms into a stacked architecture, enabling it to capture both short- and long-term dependencies as well as spatial features from complex drone flight data, including sensor readings and environmental conditions. This approach addresses limitations in traditional control methods, which struggle to maintain stability under noisy, dynamic environments. Extensive experiments were conducted to assess the model's accuracy and robustness across diverse flight scenarios. When compared to the Extended Kalman Filter (EKF), a traditional method for attitude estimation, the stacked hybrid model demonstrated superior performance, achieving significantly lower Mean Absolute Error (MAE) across all predicted parameters. Specifically, the stacked model achieved an average MAE of 0.0680, outperforming the EKF's MAE in predicting latitude, longitude, height, pitch, roll, and speed. The model's use of attention mechanisms further enhances its adaptability, allowing it to dynamically focus on the most relevant features within high-dimensional input data, improving prediction reliability in real-time scenarios.