PROCEDURE FOR ESTIMATING POSTURE ANGLES IN SPACE USING AN EXTENDED KALMAN FILTER WITH ADAPTIVE NOISE MATRIX COEFFICIENTS BASED ON DEEP LEARNING

Abstract

The invention proposes a complete estimation process (including training and execution) of attitude angles in space from gyroscope and accelerometer data. The invention is characterized by a hybrid architecture, combining an extended Kalman filter (EKF) and a long-short-feedback neural network (LSTM). The EKF filter operates at a high frequency matching the sensor's sampling frequency to ensure no data omissions and improve accuracy, while the LSTM operates at a lower frequency to analyze past data sequences and provide adaptation coefficients to adjust the system noise covariance matrix Q and measurement noise R for the EKF filter. This method increases the accuracy and stability of attitude angle estimation (in positioning and navigation problems for aircraft and robots), especially under highly maneuverable motion and changing environments; while significantly reducing computational cost and angle drift even without magnetometers.

Images