MARLIN-AD

MARLIN-AD is a typed Python library for two coupled technical tasks:

1. anomaly detection in maritime operational data, and
2. anomaly detection in AI model behaviour.

The project is built for settings where model outputs influence safety-critical or high-cost decisions. In these settings, the relevant failure surface includes both the physical process (vessel, machinery, environment, operations) and the model layer (drift, instability, calibration, uncertainty, and data-quality degradation).

Core concept

MARLIN-AD unifies two anomaly lenses in a single workflow and API family:

• Operational anomalies (data-level): unusual patterns in AIS tracks, engine/sensor telemetry, metocean context, and operational logs.
• Model-behaviour anomalies (AI-level): changes in model-facing distributions and output behaviour that indicate reduced reliability.

A practical deployment generally needs both. A changed operating regime can induce model anomalies, while model degradation can mask or distort operational anomaly signals.

Current scope (v0.1)

Detection

• Typed detector interfaces (fit, score) returning structured DetectionResult.
• Baseline detectors:
  • IsolationForestDetector
  • RobustZScoreDetector
• Explicit score-threshold conversion helpers.

Monitoring

• Deterministic monitor interfaces (fit, evaluate) returning structured MonitorResult.
• Implemented monitors:
  • KSTestDriftMonitor
  • PredictionStabilityMonitor
  • SelfTestMonitor
  • FeatureImportanceDriftMonitor (typed scaffold for richer attribution workflows)
• Contract-first stubs for calibration/concept/uncertainty expansion.

Evaluation, orchestration, and interfaces

• Evaluation helpers for precision/recall/F1, robustness checks, and report scaffolds.
• Minimal orchestration pipelines and rule-based alert formatting.
• Dataset loading and validation utilities for maritime tabular data.
• Typed plugin and CLI entry-point structure.

Installation

Development install (recommended for research workflows)

pip install -e ".[dev,docs]"

Package install

pip install marlin-ad

Minimal CLI run

marlin-ad demo --rows 500 --seed 42

The demo performs one deterministic end-to-end pass:

• operational anomaly scoring on synthetic maritime-like telemetry,
• drift monitoring between reference and current windows.

You can also run the same pipeline on a local CSV:

marlin-ad demo --csv path/to/data.csv

List built-in datasets:

marlin-ad datasets list

Minimal Python workflow

import pandas as pd

from marlin_ad.detection import IsolationForestDetector
from marlin_ad.monitoring import KSTestDriftMonitor

reference = pd.DataFrame(
    {
        "speed_knots": [12.8, 13.2, 12.9, 13.1],
        "engine_temp_c": [79.1, 78.9, 79.3, 79.0],
    }
)
current = pd.DataFrame(
    {
        "speed_knots": [13.0, 17.8, 13.4, 13.1],
        "engine_temp_c": [79.2, 96.5, 80.1, 79.0],
    }
)

# A) Operational anomalies in current maritime data.
detector = IsolationForestDetector(random_state=42).fit(reference.to_numpy())
detection = detector.score(current.to_numpy())

# B) AI-behaviour anomalies via feature-distribution drift.
monitor = KSTestDriftMonitor(pvalue_threshold=0.01).fit(reference)
drift = monitor.evaluate(current)

print("anomaly_rate", float(detection.labels.mean()))
print("drift_alerts", drift.alerts)

Documentation

The documentation in docs/ is organized around manuscript- and operations-oriented reading paths:

• Concepts: dual-anomaly framing and architectural principles.
• Maritime data: schema assumptions, validation, and loader contracts.
• Anomaly detection: detector API and score semantics.
• Model monitoring: drift/stability/health monitoring API and interpretation.
• Explainability: optional audit-oriented explanation hooks.
• Evaluation: reproducible benchmarking and reporting checklist.
• Minimal workflow: compact runnable end-to-end example.
• Implementation status: explicit implemented-vs-scaffolded matrix (docs/implementation-status.md).

Build docs locally:

mkdocs build -f docs/mkdocs.yml

Repository layout

• src/marlin_ad/detection/ — maritime anomaly detectors
• src/marlin_ad/monitoring/ — model/self-monitoring modules
• src/marlin_ad/datasets/ — loaders and schema validation
• src/marlin_ad/evaluation/ — metrics, robustness, reporting scaffolds
• src/marlin_ad/explainability/ — adapters and report hooks
• src/marlin_ad/pipelines/ — orchestration helpers
• examples/ — executable reference examples
• docs/ — MkDocs documentation

Reproducibility baseline

Before publishing results derived from MARLIN-AD, run:

ruff check .
mypy src/marlin_ad
pytest -q

Record package version, dependency versions, random seeds, dataset versions, and split definitions.

Citation

If MARLIN-AD contributes to academic or industrial work, cite using CITATION.cff.

License

Apache-2.0 © 2026 Dr Tymoteusz Miller, kpt. Irmina Durlik.