GeoClock: Physics-Guided Causal Temporal Learning for GNSS Fault Detection and Classification Under Domain Shift

This study evaluates GeoClock under a raw-hour-disjoint protocol designed to avoid reuse of nominal observation epochs across source training and validation. The model maps variable-cardinality GNSS observations to 80 physical statistical features and processes causal 60-epoch windows using a multi-task temporal convolutional network. BDS_2I, BDS_7I, BDS_6I, and GPS_1C provide source training, validation, and a source lockbox; all GPS_2W hourly files are held out. Across five paired masked-pretraining seeds with seed-specific source-selected thresholds, direct GPS_2W transfer has a mean AUROC of 0.999377 ± 0.001246 and recall of 0.999990 ± 0.000022, but a mean FPR of 0.537431 ± 0.337510. An independently acquired cross-site and cross-domain validation set from Site B (20 one-hour injected files spanning four signal domains and five fault realizations per domain; 14,180 windows and 81 events) yields a frozen direct-transfer AUROC of 0.999983 and an FPR of 0.1108. Under the stated 240 s clean-start assumption, the FPR falls to 0.000545 with recall 0.9721. This experiment evaluates geographically distinct acquisition conditions using the same frozen model and source-selected threshold; because the fault patterns remain injected, it does not establish field integrity or generalization to arbitrary sites. Five-seed GRU and TCN comparisons, residual-threshold diagnostics, feature-family ablations, and component ablations do not resolve threshold portability. A radial ECEF sensitivity analysis shows severe degradation at 3 and 10 m coordinate offsets. The results support controlled recovery of the stated injected observables and expose calibration and position-reference limitations; they do not establish field integrity.