A generalized, outage period-aware CNN-BiLSTM architecture for real-world INS/GPS deployment under extended GNSS denials
GPS signal outages present a fundamental challenge for low-cost integrated navigation systems, leading to unbounded error growth in the Inertial Navigation System (INS). Although the traditional filtering techniques provide a substantial Performance enhancement in GPS-INS integration, it fails at signal loss conditions. While machine-learning-based (ML-based) compensation methods have been explored, they often fail to model the complex, long-term error dynamics of MEMS sensors and face limitations in real-world applicability due to high computational costs. To mitigate these limitations, a DeGIN (Deep GPS Increment Network) is proposed, which is a hybrid CNN-BiLSTM architecture for predicting GNSS position correction increments in latitude and longitude during signal outages. The CNN module learns discriminative features from raw Inertial Measurement Unit (IMU) measurements, while the BiLSTM captures long-range temporal dependencies in both forward and backward directions, addressing the limitations of strictly unidirectional recurrent models. A key contribution is the explicit inclusion of an outage timer that conditions the predictions on the elapsed duration of signal loss. In addition, a hybrid direction-aware loss function is introduced to promote physically plausible trajectory estimates. Rigorous cross-domain experiments on multiple real-world public datasets demonstrate that the proposed model achieves higher accuracy and better generalization than conventional baselines and recent deep learning approaches. The model was trained on the NaveGo benchmark dataset and validated on an unseen held-out split, as well as on three entirely separate datasets: Nav200 (collected by the authors), INSANE (a published benchmark flight dataset), and simulated data generated by our simulation framework—all without any re-training or fine-tuning. On real-world data, DeGIN achieves a 64.2–86.4% improvement in position accuracy over recent deep learning baselines, with gains reaching up to 98.92% in aerial flight outage scenarios, and reduces error by 95.2–99.9% compared to EKF dead reckoning. Additionally, the framework is accompanied by a detailed optimization pathway, supporting its suitability for deployment on resource-constrained edge hardware. Overall, the proposed solution provides a robust and deployable approach for maintaining reliable navigation in GPS-denied environments.