IJCAS

Magnetic, angular rate, and gravity (MARG) sensor-based orientation estimation is commonly implemented using extended Kalman filter (EKF) frameworks, where robustness against dynamic motions and magnetic disturbances is ensured through online EKF covariance updates. This paper proposes an alternative solution involving a neural network (NN) with an appropriate topology, which is trained to provide MARG correction signals to the EKF, thereby delivering disturbance-suppressed measurements for improved orientation estimation. First, the EKF-based orientation estimation is revisited, and the main sources of error are highlighted. Then, a universal laboratory framework is introduced, which performs various motions with a calibrated UR5 cobot and records the raw measurements from a MARG sensor attached to its end-effector, along with ground truth pose data. Using a comprehensive database consisting of 16 different scenarios, the derivation of target signals for error-corrected MARG data is presented. Next, three shallow NN architectures—feed forward NN (FFNN), cascade forward NN (CFNN), and focused time-delay NN (FTDNN)—are trained for these targets, and their performances are evaluated using Pearson’s correlation metrics as a function of the number of hidden neurons and input channel combinations. Finally, each NN-EKF combination is assessed, and it is found that the FFNN offers the most suitable topology in terms of both performance and computational cost. The proposed FFNN-EKF approach enhances orientation estimation quality by a significant 44.5% based on the mean squared quaternion error metrics, even in highly disturbed environments. The performance of the FFNN-EKF is also compared to other common methods, demonstrating that the proposed approach outperforms the benchmark filters.

Keywords: Disturbance compensation, Kalman filter, neural network, orientation estimation, sensor fusion.