This study develops a hybrid framework integrating a CBGRU-DAtt deep learning model with an adaptive maximum correntropy unscented Kalman filter (AMCUKF) for GPS/INS navigation. The study aims to enhance positioning accuracy during GPS outages and improve robustness against time-varying non-Gaussian noise through dynamic kernel bandwidth adjustment.
A CBGRU-DAtt model (CNN, BiGRU and dual attention) is designed to predict pseudo-GPS position increments from IMU data during outages. These predictions are fused with IMU data using a maximum correntropy unscented Kalman filter (MCUKF) to suppress non-Gaussian noise. An adaptive bandwidth strategy based on exponential weighted moving average (EWMA) is proposed to dynamically adjust the kernel bandwidth. The framework is validated on real vehicle datasets with simulated GPS outages and time-varying Laplace noise.
The proposed CBGRU-DAtt model reduces position RMSE by over 30% compared to LSTM, TCN and CNN-BiGRU during GPS outages. Integrating MCUKF further enhances robustness under non-Gaussian noise, with s = 6 achieving optimal accuracy. The adaptive AMCUKF dynamically adjusts kernel bandwidth via EWMA, reducing east velocity RMSE from 0.298 m/s to 0.201 m/s under time-varying Laplace noise, demonstrating superior adaptability and stability in complex environments.
This paper proposes a novel hybrid framework integrating CBGRU-DAtt deep learning with an adaptive maximum correntropy unscented Kalman filter (AMCUKF) for GPS/INS navigation. The key innovations include (1) a dual-attention mechanism for enhanced spatiotemporal feature extraction from IMU data, (2) integration of deep learning predictions with robust filtering to mitigate error accumulation during GPS outages and (3) an EWMA-based adaptive kernel bandwidth strategy enabling dynamic response to time-varying non-Gaussian noise. This systematic solution significantly improves navigation accuracy and robustness in complex real-world environments.
