An Intelligent Measurement-Compensated Sliding-Window FGO Method for Robust Surface–Underwater Navigation in GNSS/LBL-Denied Conditions
Abstract
In complex surface and underwater environments, the loss of external position observations degrades strapdown inertial navigation system (SINS)/Doppler velocity log (DVL) navigation accuracy. This article develops an intelligent measurement-compensated sliding-window factor graph optimization (FGO) method for robust surface–underwater SINS/DVL navigation under global navigation satellite system (GNSS)/long baseline (LBL)-denied conditions. Within this framework, the gated recurrent unit (GRU)–convolutional neural network (CNN)–attention network with occlusion awareness (GCA-OA) learns the position residual between an externally aided FGO trajectory and its SINS/DVL-only counterpart from inertial measurements and SINS/DVL-derived states. During GNSS or LBL outages, the predicted residual directly corrects the SINS/DVL-only FGO output. The occlusion-aware strategy combines contiguous-block observation masking, supervised residual regression, and inertial measurement unit (IMU)/DVL-based consistency constraints, while sliding-window marginalization maintains bounded online computation. Experiments under surface GNSS and underwater LBL outages show that GCA-OA + FGO outperforms the evaluated outage-compensation methods in position and velocity accuracy and achieves lower empirical tail errors.