LSTM을 이용한 항법시스템의 확장칼만필터 공분산 튜닝

Abstract

Noise covariances in the EKF (Extended Kalman Filter) act as tuning parameters due to modeling errors, such as the error generated in the process of linearizing the nonlinear system equation. Filter designers traditionally prefer the trial-and-error method. However, because the noise covariances require retuning in concert with changes in trajectories or changes in system equations, human tuning is very time consuming. Accordingly, this study proposes an automation method for real-time parameter tuning using an LSTM (Long Short-Term Memory) neural network. The proposed method consists of two stages: 1) generating residual and state error covariances from an EKF using fixed-noise covariances and 2) producing a state estimate from an EKF using tuned time-varying noise covariances obtained from an LSTM neural network. This scheme is applied to navigation systems to calculate accurate velocity, attitude, and IMU bias estimates under challenging maneuvering situations. Using simulation, we showed that the proposed method outperforms the conventional EKF of navigation systems in suppressing estimation error in different trajectory scenarios.