Learnable Residual Denoiser Enhanced Deep Sensor Fusion for Smartphone Navigation Under GNSS Outages

This study introduces an advanced sensor-fusion approach to improve localization robustness during Global Navigation Satellite Systems (GNSS) outages on smartphones. We developed a hybrid deep learning architecture combining Convolutional Neural Networks (CNN) and Gated Recurrent Units (GRU) that learns spatio-temporal patterns in inertial data to generate pseudo-GNSS positions. To improve training stability and noise robustness, we evaluate two Inertial Measurement Unit (IMU) denoising strategies before the CNN layers: (a) a Hierarchical Denoising Pipeline (HDP) and (b) a Learnable Residual Denoiser Network (LRDN). Evaluations on real-world vehicular logs under 30 seconds and 60 seconds outages demonstrate a clear performance hierarchy. The HDP-integrated CNN-GRU consistently outperforms conventional fusion and baseline deep-learning approaches by reducing error metrics and improving trajectory smoothness. The LRDN-integrated CNN-GRU, based on a learnable residual denoiser, achieves the highest gains, reducing RMSE by nearly 74% compared to the plain CNN-GRU, up to 85% compared to baseline models including standalone Extended Kalman Filter (EKF) fusion, and 52% over the HDP-enhanced CNN-GRU. This work investigates the integration feasibility of the proposed deep-learning-aided INS-GNSS navigation framework on smartphone-based sensing and processing platforms, addressing both sensor-level noise suppression and temporal sequence modeling within a unified architecture. The results demonstrate computational feasibility within the limits of commodity smartphones, confirming the practical relevance of the approach for edge-deployable navigation in GNSS-challenged environments.