⚡ electrical engineering

Ordinal temporal learning for operational quality control of significant wave-height observations from coastal buoy networks

This paper presents an online temporal ordinal classifier that effectively performs real-time quality control on coastal buoy significant wave-height observations by integrating temporal context and residual features to accurately detect various sensor failures and support auditable operational workflows.

Original authors: Marcos Batista Figueredo, Jana Ribeiro Santana, Maria Dolores Ribeiro Orge, Leila Maria Mendes Santos, Lizandra Santos Morais, Leticia Nayara Barreto dos Santos, Everton Vitor Almeida Monville, João A

Published 2026-10-21

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Marcos Batista Figueredo, Jana Ribeiro Santana, Maria Dolores Ribeiro Orge, Leila Maria Mendes Santos, Lizandra Santos Morais, Leticia Nayara Barreto dos Santos, Everton Vitor Almeida Monville, João Antonio Assis Santana Batista, Elaine Pereira Garrido, Marta Giovanetti, Vagner Souza Fonseca

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The ocean is a restless, powerful place, and for ships navigating its edges, for ports managing their docks, and for engineers designing coastal defenses, knowing the height of the waves is a matter of safety and economics. To get this information, scientists deploy networks of floating buoys that bob in the surf, measuring the significant wave height—the average height of the highest third of the waves in a given period. These instruments send data back to shore in near real-time, creating a continuous stream of numbers that tell the story of the sea. However, these sensors are not infallible. Like any machine left to the mercy of salt spray, storms, and constant motion, they can malfunction. A sensor might freeze, drift slowly off its true reading, or suddenly spike with a value that makes no physical sense. If these bad numbers slip into the historical record, they can distort our understanding of climate trends or lead to flawed engineering calculations. If they are not caught in real-time, they could trigger unnecessary emergency responses or cause operators to ignore a genuine storm. The challenge, therefore, is to build a system that can instantly tell the difference between a truly wild ocean and a broken instrument, without needing a human to stare at a screen every second.

For decades, the standard way to handle this has been to set rigid rules. If a wave height jumps too high too fast, or if it stays exactly the same for too long, the computer flags it as a problem. These rules are transparent and easy to check, but they are also blunt instruments. They often struggle to distinguish between a genuine, rare storm surge and a sensor error, or they might miss a slow, creeping failure that doesn't break any single rule. In a new study, researchers from Brazil and Italy have proposed a different approach. Instead of relying on fixed limits, they trained a computer system to learn the rhythm of the ocean itself. This system acts like a seasoned observer who knows what the waves usually do at a specific location. It watches the current wave height and compares it to the pattern of the waves that came just before it. If the new number fits the story the ocean has been telling, the system marks it as good. If it breaks the pattern, the system doesn't just say "bad"; it assigns a level of suspicion, ranging from "suspect" to "bad," allowing operators to decide how urgently to investigate.

The researchers built this system using data from six different coastal buoys along the Brazilian coast, a region where the ocean conditions can change rapidly due to tides and river outflows. They taught the computer to look at the last week of wave data to predict what the next hour's wave height should be. When the actual measurement arrived, the system calculated the difference between what it expected and what it saw. It didn't stop there; it also checked if the sensor was behaving consistently over time, looking for slow drifts or sudden jumps that a simple rule might miss. Crucially, the system was designed to make decisions using only the information available at that exact moment, without peeking at future data. This is essential for real-time monitoring, where you cannot wait to see what happens next to decide if the current reading is trustworthy. The team tested this "ordinal" system—which means it ranks the quality of the data on a scale rather than just saying yes or no—against a variety of simulated failures, including sensors that got stuck, drifted, or burst with random noise.

The results showed that this learning-based approach worked remarkably well. When tested on data from the six buoys, the system achieved an AUC of 0.85 while keeping false alarms low. More importantly, it proved that a single model trained on data from all six locations could be applied to any of them without needing to be re-tuned for each specific site. This is a significant step forward, as it means one system could potentially monitor an entire network of buoys, regardless of whether they are in a calm bay or a rough open coast. The system was particularly good at catching sudden, dramatic errors, such as a sensor that suddenly stopped working or sent a value that was clearly impossible. However, the study also found that the system needed time to catch slower, more subtle problems. If a sensor began to drift slowly away from the truth, the system could not flag it immediately; it needed to accumulate evidence over several hours before it was confident enough to raise an alarm. This mirrors how a human expert might behave, noticing a small oddity but waiting to see if it persists before calling it a failure.

The researchers also compared their new system to the traditional rule-based methods used in operational archives. They found that the old rules, which rely on fixed thresholds, often struggled to balance catching errors with avoiding false alarms. In some cases, the rigid rules would flag a genuine storm as a sensor error, while in others, they would miss a subtle malfunction. The new system, by learning the local behavior of the waves, was better at distinguishing between a true environmental extreme and a broken sensor. When the researchers ran the system on a continuous stream of real data, simulating a live monitoring environment, it detected abrupt failures almost instantly. For the slower, creeping failures, it took longer, but it eventually caught them by tracking the accumulation of small inconsistencies. The study concluded that this approach offers a practical, auditable way to improve the quality of ocean data. It does not replace human experts but provides a smart, automated layer that can triage the data, sending the most suspicious readings to human review while letting the clear, good data flow through for immediate use. By turning a binary "pass or fail" decision into a nuanced assessment of risk, the system helps ensure that the stories we tell about the ocean are based on reliable numbers.

Technical Summary: Ordinal Temporal Learning for Operational Quality Control of Significant Wave-Height Observations

Problem Statement
Reliable sensor data from coastal buoy networks is critical for navigation, port operations, and environmental monitoring. However, moored instruments are susceptible to various failure mechanisms—including drift, saturation, freezing, and communication errors—that produce anomalous or physically inconsistent observations. Traditional quality control (QC) often relies on deterministic rules (e.g., QARTOD procedures) based on physical limits and rate-of-change thresholds. While transparent, these methods struggle to distinguish genuine environmental variability from sensor faults across different hydrodynamic regimes and often require region-specific threshold tuning. Furthermore, existing machine learning approaches for oceanographic QC frequently focus on forecasting accuracy rather than operational fault detection, or they fail to address the specific engineering requirements of one-sided temporal decision-making, ordinal severity assessment, and cross-station transferability in a streaming context.

Methodology
The study proposes an online temporal ordinal classifier designed to assign real-time quality flags to significant wave-height ( $H_s$ ) observations. The system operates under strict one-sided temporal constraints, using only the current observation and its preceding context to make decisions.

Architecture: The core consists of a univariate temporal predictor (three LSTM layers with decreasing width: 128, 64, 32 units) trained to predict the next $H_s$ value based on a 168-hour (7-day) lookback window.
Feature Engineering: The classifier ingests a feature vector derived strictly from information available at time $t$ $t$ . This includes:
- Instantaneous indicators: One-sided second differences, deviations from trailing medians/MAD, rate of change, and out-of-range flags.
- Accumulated evidence: Exponentially Weighted Moving Averages (EWMA) of residuals, two-sided Cumulative Sum (CUSUM) statistics, local linear slopes, and sign-persistence counters to detect slow drifts.
Ordinal Decision Head: Instead of binary classification, the model outputs probabilities for three ordered classes: GOOD < SUSPECT < BAD. This is achieved via two sigmoid units estimating $Pr(Q_t \geq SUSPECT)$ and $Pr(Q_t = BAD | Q_t \geq SUSPECT)$ , trained with a weighted ordinal loss function.
Training and Data: The model was trained on pooled records from six public Brazilian Coastal Monitoring System (SiMCosta) buoys. Training targets combined "strong labels" (derived from synthetic corruption of seven failure families: additive spikes, multiplicative shifts, level shifts, drift, stuck sensors, burst noise, and missing/sentinel values) and "weak labels" (statistical outliers on uncorrupted data).
Evaluation Protocols: The system was evaluated using:
1. Pooled testing: Chronological hold-out partitions on the six-buoy dataset.
2. Leave-one-buoy-out (LOBO) transfer: Training on five buoys and testing on the sixth to assess generalization without station-specific tuning.
3. Streaming simulation: A sequential experiment on real BA-1 data with injected failure episodes to measure onset latency and episode coverage under continuous error feedback.
4. Comparator benchmarks: Comparison against fixed-threshold detectors (e.g., LSTM-Peak, Isolation Forest, VAE-LSTM) and a reference statistical protocol.

Key Contributions

Online Temporal Architecture: A system that combines instantaneous and accumulated evidence for graded quality assessment, strictly adhering to one-sided temporal constraints (no future data leakage).
Ordinal Severity Assessment: A shift from binary rejection to a three-tiered operational workflow (GOOD, SUSPECT, BAD), allowing for tiered triage where "SUSPECT" observations are flagged for review rather than immediate removal.
Cross-Station Transferability: Demonstration that a single model trained on pooled multi-station data can transfer to unseen coastal buoys with distinct wave regimes without local recalibration.
Operational Validation: An engineering assessment of the system under continuous streaming conditions, explicitly measuring detection latency for different failure modes (abrupt vs. slow-evolving).

Results

Performance on Controlled Failures: On the pooled controlled-failure test set, the classifier achieved a binary F1 score of 0.853 (GOOD vs. non-GOOD). The class-specific F1 scores were 0.977 (GOOD), 0.634 (SUSPECT), and 0.614 (BAD). Precision was 0.890 and recall was 0.819.
Transferability: In the LOBO transfer evaluation, the model achieved a mean binary F1 of 0.856 across six unseen stations. Performance varied by station (ranging from 0.779 to 0.940), indicating that while transferability is feasible, station-specific environmental factors influence performance.
Streaming Detection:
- Abrupt failures (spikes, sentinel codes, multiplicative scaling) were detected promptly at $t+1$ .
- Slow failures (drift, stuck sensors) exhibited near-zero immediate onset detection rates (0.011 and 0.048, respectively) because initial perturbations fell within natural ambient variance. Detection of these required accumulated temporal evidence (EWMA/CUSUM).
- The system maintained a false-alarm rate of 2.03% on clean data and achieved a mean episode coverage of 47.85%.
Ablation and Comparators: A univariate predictor (using only $H_s$ ) performed as well as or better than multivariate alternatives. Local threshold calibration (using robust Student-t dispersion) outperformed fixed constants inherited from other regions. The proposed ordinal classifier outperformed several benchmark detectors (e.g., Isolation Forest, VAE-LSTM) in the specific context of wave-height QC.

Significance and Claims
The paper positions the proposed system as an auditable, transferable triage layer for operational ocean-observation workflows. It claims to solve the engineering problem of converting streaming observations and prediction residuals into reliable, actionable fault assessments without relying on future data or station-specific tuning.

The authors emphasize that the system does not replace operational acceptance procedures or sensor-maintenance diagnostics but complements them by:

Transforming rigid binary rejection into a tiered workflow (GOOD for routine use, SUSPECT for review, BAD for removal).
Providing a transparent decision process where the provenance of each flag (statistical vs. temporal) is preserved.
Identifying the operational boundary of the method: it is highly effective for abrupt anomalies but requires accumulated evidence for slow drifts, necessitating a distinction between instantaneous onset rates and episode coverage in performance metrics.

The study concludes that the system is a viable tool for real-time engineering monitoring in heterogeneous coastal networks, supported by evidence of cross-station generalization and robustness under sequential error feedback, while acknowledging limitations regarding reliance on controlled corruption models and the assumption that historical baselines are largely correct.

Technical Summary: Ordinal Temporal Learning for Operational Quality Control of Significant Wave-Height Observations

More like this