Feedforward Neural Network for Orientation Estimation Under Magnetic and Acceleration Disturbances

Orientation estimation using magnetic, angular rate, and gravity (MARG) sensors is primarily affected by two error sources: external accelerations and magnetic disturbances.These disturbances hinder the accurate separation of sensor observations from their corresponding reference vectors, thereby degrading estimation accuracy.This paper presents a feedforward neural network (FFNN) architecture designed to estimate the disturbance vectors, enabling the generation of disturbance-compensated signals for improved state estimation within an Extended Kalman Filter (EKF) framework.First, the MARG sensor models are introduced and their integration into an EKF-based orientation estimation system is outlined.The proposed FFNN model for disturbance estimation is described in detail, with emphasis on its architectural design and the methodology for effective augmentation of the EKF.The practical feasibility of the approach is demonstrated in a laboratory setting using a calibrated UR5 robot.The robot performs various motions with a MARG sensor mounted on its end-effector, thereby subjecting the system to a range of disturbance magnitudes.A comprehensive evaluation is conducted across 16 distinct measurement scenarios, analyzing the influence of various disturbance combinations and intensities on estimation performance.The results indicate that the proposed FFNN-EKF approach consistently outperforms the standard EKF, with improvements in estimation accuracy ranging from 19% to 49%.This demonstrates the effectiveness of NN-based disturbance compensation in enhancing orientation estimation under real-world conditions.