Adaptive EKF With DNN-Predicted Weight Factor for Outlier-Alleviated Signal Estimation in Laser Galvanometers

Achieving accurate position estimation is essential to ensure high stability and accuracy in laser galvanometer technologies. The widely used approaches for estimation are the nonlinear Kalman filters, particularly the extended Kalman filters (EKFs). However, precision position estimation remains a significant challenge due to measurement inaccuracies and the presence of signal outliers. These challenges cannot be effectively addressed using the classical EKF with the covariance of static measurement noise and the need for adaptive methods to achieve accuracy. The problem is the factor in adaptive EKF (AEKF) often remains near its maximum value limiting outlier reduction and dynamic noise adaptability. In this article, a deep neural network (DNN) is designed to predict an adaptive weight factor dynamically utilizing residual. First, the residual dataset is utilized as input to train the DNN model for predicting the adaptive weight factor. Second, the predicted weight factor is employed to dynamically adjust the covariance of measurement noise. Third, the updated measurement noise covariance is used to recalibrate the Kalman gain, thereby mitigate the outliers and improve the estimation accuracy without ground truth. The proposed deep neural network–based adaptive extended Kalman filter (DNNAEKF) is evaluated against classical extended Kalman filter (CEKF), hybrid adaptive extended Kalman filter (HAEKF), adaptive unscented Kalman filter (AUKF), and innovation AEKF (IAEKF) through simulation and experiment. The results demonstrate that DNNAEKF significantly improves precision position estimation for the laser galvanometer while maintaining reliable performance under outliers.