Coastal buoys get a three-level wave-height quality flag: one temporal model transfers across stations at a mean binary F1 of 0.856

Significant wave-height observations from coastal buoy networks can now carry an hourly three-level quality flag rather than only a keep-or-remove verdict. Operational quality control has relied mainly on deterministic tests for physical limits, climatology, rate of change, spikes and local dispersion, whose performance depends on thresholds and reference distributions that may not transfer between regions. On the pooled controlled-failure test set from six Brazilian Coastal Monitoring System buoys, the model reached a binary F1 of 0.853, BAD-class F1 of 0.614, precision of 0.890, recall of 0.819 and BAD-class AUPRC of 0.648; the confusion matrix shows most errors fell between adjacent severity levels, with only four GOOD observations assigned BAD and 53 BAD observations assigned GOOD.

One trained model can serve buoys it never saw without station-specific retuning. A model developed for one station may not retain its performance at another, and cross-station transfer is a recurring requirement in marine sensor quality control. Under leave-one-buoy-out validation, training on five buoys and evaluating on the sixth, the mean binary F1 across the six unseen stations was 0.856, ranging from 0.779 on ES-1 to 0.940 on BA-1.

Splitting the decision into current residual and accumulated evidence lets slowly evolving faults be caught in later hours. An instantaneous residual is insensitive to slow drift, because the initial deviation lies within the natural variability of the preceding sea state. In a streaming simulation over 8592 hours of the real BA-1 series, onset detection at t+1 was 1.000 for missing or sentinel codes, 0.838 for level shifts and 0.747 for multiplicative anomalies, but 0.011 for drift and 0.048 for stuck sensors; the false-alarm rate outside injected episodes was 2.03% and mean episode coverage was 47.85%.

The demonstrated scope is significant wave-height series from six buoys of the Brazilian Coastal Monitoring System, with decisions using only the current observation and the preceding 168 hours of univariate context and returning GOOD, SUSPECT or BAD. Evaluation covers pooled testing, leave-one-buoy-out transfer, seven controlled failure families and hourly streaming simulation, and the quantitative scores are conditional on the synthetic failure scenarios. The model is intended as hourly online triage, kept alongside the centred-window statistical reference protocol with both labels and their provenance preserved in the archive.

Quantitative results are conditional on seven controlled corruption families, and the scores depend on how faithfully those scenarios represent real sensor behaviour. The streaming evaluation treats the historical baseline as uncorrupted, so genuine anomalies within it are scored as false alarms, biasing the reported false-alarm rate upward rather than understating it. No hardware-level latency benchmark is included, so the evidence supports algorithmic online operation rather than a millisecond guarantee. The study does not test whether bathymetry, energy spectra, sensor configuration or regional climate explains the variation between stations.