A Hybrid attention-LSTM framework for user-level drift identification in GNSS-derived localization

User-level localization outputs derived from GNSS often exhibit residual localization drift in dense urban environments where signal degradation coexists with legitimate dynamic movements, creating ambiguity between localization errors and true motion. To address this post-fusion, coordinate-level issue, we propose a threshold-agnostic drift identification paradigm using a hybrid BERT–LSTM architecture. The method synergizes Attention’s contextual awareness for spatiotemporal pattern extraction from 30-step sequences with LSTM’s sequential dynamics modeling, eliminating reliance on empirical kinematic thresholds. Validation across 654,855 real-world points from consumer-device deployments demonstrates a 92.2% F1-score in drift-detection accuracy, surpassing pure BERT by 5.9% and LSTM baselines by 18.0%, while reducing false positives by 38.7% versus standalone transformers. The framework successfully preserves trajectory integrity amid rapid transitions where residual coordinate drift coexists with legitimate dynamics, overcoming limitations of context-insensitive Transformers and limited-context LSTMs. The approach operates directly on post-fusion coordinate sequences rather than raw GNSS observations. This study establishes a robust foundation for reliable user-level localization assessment in challenging environments where multipath effects and dynamic movements interact.