A multi-strategy optimized deep learning approach for INS/GNSS error compensation during GNSS Outages

The integrated navigation system, consisting of an Inertial Navigation System (INS) and a Global Navigation Satellite System (GNSS), is vulnerable to interference in complex environments, which can lead to GNSS signal loss and significantly degrade navigation accuracy. To improve the navigation performance of land vehicles during GNSS signal interruptions, this paper proposes an INS/GNSS integrated navigation system method that incorporates deep learning. The method enhances the Long Short-Term Memory (LSTM) model by integrating Convolutional Neural Networks (CNN), an Attention Mechanism, and a Meta-heuristic Global Optimization Algorithm (PSA), thereby improving system performance. When the GNSS signal is available, the deep learning module trains the mapping between IMU-measured acceleration and specific force, and INS-derived attitude and velocity, and the GNSS velocity and position data. In the event of GNSS signal loss, the pre-trained deep learning model predicts velocity and position, which are then fused with the INS data to replace the GNSS measurements. This effectively suppresses the accumulation of system errors. Experimental results on the dataset show that when GNSS is interrupted for 50 seconds, the proposed method improves positioning accuracy by 72.85% over INS. In real-world road tests, accuracy improves by 93.30% on straight sections and 75.83% on turns, effectively suppressing error accumulation during GNSS outages.