Context-Aware Adaptive Filtering for GNSS/INS Integration via Multiscale Time–Frequency Feature Fusion

Reliable navigation in complex urban environments remains a major challenge for Global Navigation Satellite System (GNSS)/Inertial Navigation System (INS) integration because of signal blockage, multipath, and non-line-of-sight (NLOS) propagation. This paper presents a context-aware GNSS/INS framework that integrates fine-grained environment recognition with semantic-guided adaptive filtering. First, an eight-class GNSS environmental taxonomy is established to characterize representative signal degradation mechanisms ranging from open-sky to deep-indoor environments. Then, a Progressive Time-Frequency Interaction Network (PTFINet) is developed to capture multi-scale time-frequency characteristics of GNSS observations through dual-branch feature extraction and progressive interaction fusion, enabling robust environment recognition under rapid transitions. Finally, an Environment Semantic Adaptive Kalman Filter (ESAKF) is proposed for real-time adaptive measurement updating. It extends conventional innovation-based adaptive filtering in two steps: 1) the measurement-noise covariance is recursively estimated from innovation statistics using an exponentially weighted forgetting-factor scheme instead of a sliding-window estimator, thereby improving responsiveness while reducing memory cost; and 2) the estimated covariance is further regularized using environment-dependent covariance lower-bound constraints derived from the recognized context, thereby preventing noise underestimation under severely degraded conditions. Vehicle experiments show that PTFINet achieves recognition accuracies of 0.9594, 0.9278, and 0.9509 on three dynamic transition routes. For GNSS/INS integration, validation on two real-world field tests demonstrates that the proposed method achieves more accurate and robust navigation performance than representative baseline methods.