AI-Based Predictive Maintenance for Marine Auxiliary Engines

AI-Based Predictive Maintenance for Marine Auxiliary Engines

Executive summary

AI-driven predictive maintenance (PdM) for marine auxiliary engines—generators, pumps, compressors, fans, and other hotel-power equipment—uses sensor telemetry, oil/chemical analysis, operational context, and machine learning to detect incipient faults and predict remaining useful life (RUL). When implemented correctly it reduces unplanned downtime, lowers lifecycle cost, and improves safety and regulatory compliance. Major engine OEMs and class/technical bodies are actively developing condition-based and data-driven services to realize these benefits at scale. DNV+1

1. Why PdM matters for marine auxiliary engines

Auxiliary systems are mission-critical: losing hotel power, pumps or compressors can create safety risks, force port diversions, delay cargo operations, and trigger expensive emergency repairs. Traditional time-based maintenance commonly either over-services running equipment (wasting parts and labor) or fails to catch progressive failures early. PdM shifts maintenance decisions from fixed calendars to data-driven, condition-aware actions—catching faults earlier, optimizing spares logistics, and improving vessel availability. This condition-based approach is already being formalized by class and inspection bodies as an accepted alternative to purely scheduled regimes. DNV

2. What data you need (and why those signals matter)

Successful PdM starts with observability. The highest-value data streams for auxiliary engines are:

• Vibration (tri-axial accelerometers, spectral bands) — best early indicator for bearings, misalignment, shaft/gear problems.

• Oil and fluid condition (continuous monitors, lab analysis) — direct evidence of wear metals, contamination, and lubricant degradation. Continuous oil monitoring products are now mainstream for engine fleets.

• Temperature & pressure sensors — detect overheating, cooling failures, blocked strainers, or pump cavitation.

• Electrical telemetry (current, voltage, power factor, harmonics) — reveals generator and motor electrical faults or combustion anomalies on gensets.

• RPM, load %, and operational context (start/stop cycles, ambient conditions) — essential to separate operationally-driven variation from real faults.

• Event and alarm logs, maintenance work orders, and inspection reports — provide labels for supervised models and context for root-cause analysis.

OEMs that offer condition monitoring emphasize vibration + oil analytics as primary high-yield sources for auxiliary assets. Wartsila.com+1

3. Signal processing & feature engineering (brief practical guide)

Raw sensors are rarely ready to feed an ML model directly. Effective preprocessing steps that materially improve results:

• Compute time-domain metrics (RMS, crest factor, kurtosis, skewness) over sliding windows.

• Compute frequency-domain features (FFT bands, harmonic magnitudes) for rotating machinery and apply envelope analysis for bearing faults.

• Use wavelet transforms or Short-Time Fourier Transforms for transient events.

• Create operationalized features: normalized vibration per RPM, oil particle counts per milliliter and size bin, temperature deltas (bearing vs ambient).

• Aggregate features with short (minutes), medium (hours), and long (days) windows to capture both sudden faults and slow degradation trends.

Good feature engineering often lets simpler models (gradient boosting, logistic regression) outperform complex deep models in low-label regimes.

4. Modeling approaches — pick by data maturity

• No/limited labels: start with unsupervised anomaly detection — autoencoders (LSTM/Conv), isolation forests, or one-class classifiers — to flag deviations from normal operating envelopes.

• Some labeled failures: build RUL and classification models using sequence models (LSTM/GRU/Transformers) or tree-based models on engineered features (XGBoost/LightGBM). Survival analysis methods are useful when failure times are censored.

• Combined physics + data (hybrid models / digital twins): combine first-principles models (bearing life equations, thermodynamic fuel/combustion models) with ML residual predictors to improve extrapolation and interpretability. The literature shows digital-twin–based PdM produces higher fidelity predictions when physical behavior is well understood. ScienceDirect+1

5. Edge vs Cloud: practical architecture

• Edge (onboard) — perform high-frequency signal capture and lightweight inference (feature extraction, anomaly scoring) at the gateway to avoid saturating limited satellite or cellular links. Edge inference keeps latency low and maintains availability when connectivity is intermittent.

• Cloud (shore) — aggregate fleet data, run heavy re-training, cross-vessel analytics, and generate fleet-level insights and dashboards. Cloud platforms also simplify version control and model governance.

• Hybrid pattern (recommended) — do signal conditioning and initial inference on edge; transmit summaries and selected raw windows for cloud training and root-cause investigations. Many OEMs’ digital offerings adopt this hybrid model. Wartsila.com+1

6. MLOps, governance, and crew workflows

Key elements to operationalize PdM:

• Model versioning & traceability — every alert should record model id, version, input snapshot, and confidence.

• Drift monitoring — detect input distribution drift and alert for model retraining.

• Feedback loop — allow engineers to label alerts (true/false positive) and feed that data into retraining pipelines.

• Integration with CMMS/EAM — allow automatic work orders for high-confidence failures and ensure spare-parts recommendations tie into procurement.

• Human-centric alerting — alerts must be actionable: include probable cause, confidence, recommended inspection steps, and required urgency; otherwise crews will start ignoring them.

• Cybersecurity & network segregation — protect sensor networks and ensure encrypted telemetry and hardened edge devices.

DNV and other industry actors emphasize standardizing maintenance logs and high-quality labels as prerequisites for reliable ML models; joint-industry efforts are underway to improve labeled datasets and benchmarking. DNV+1

7. Explainability and trust

Operators need more than a “score.” Provide:

• Feature-importance or SHAP summaries to show which signals drove an alert.

• Waveform or spectral snapshots with the anomaly highlighted.

• Suggested next steps tied to engineering procedures (e.g., reduce load, schedule dry-dock inspection, sample oil).

Explainability reduces false-alarm fatigue and enables class and flag state audits.

8. Regulatory, class society, and contractual considerations

Switching from scheduled to condition-based maintenance touches survey and statutory requirements. To preserve class acceptance and insurance compliance you must:

• Maintain auditable records: raw feature snapshots, alerts, decisions and corrective actions.

• Agree contractual data-sharing and IP terms with OEMs/third-party analytics providers.

• Coordinate with class societies when making survey-relevant maintenance changes based on PdM outputs; several class societies already provide guidance and acceptance frameworks for CBM/PdM. DNV

9. Real-world vendor examples & industry traction

• Wartsila offers propulsion and equipment condition-monitoring services that combine vibration and oil condition analytics with operational telemetry to deliver fleet-level insights and service recommendations. This is an example of an OEM building integrated PdM services for auxiliary systems. Wartsila.com

• MAN Energy Solutions has invested in continuous oil-monitoring hardware and fleet software (Multi Fluid Monitor) to detect lubricant anomalies and support predictive workflows—illustrating how oil analytics are being productized for engines. Man ES+1

These vendor programs illustrate the industry shift: from ad-hoc monitoring to fully packaged PdM services that combine onboard sensing, edge analytics, cloud retraining, and shore-based expert insight.

10. Example pilot plan (practical, 6–12 month pilot)

Scope: 2 vessels, 4 identical auxiliary generator sets each.

Month 0–1: discovery

• Inventory existing sensors, PLCs, and data logs. Prioritize tri-axial vibration, temperature, oil sample points, and electrical metering.

Month 1–3: capture & baseline

• Install missing sensors and an edge gateway for feature extraction. Collect baseline operating data across load profiles for 3 months.

Month 3–6: model & validation

• Run unsupervised anomaly detection (autoencoder) to surface early anomalies. Begin occasional supervised models if labeled historical events exist.

• Route alerts to engineering team; capture feedback (true/false).

Month 6–12: extend & integrate

• Integrate alerts into CMMS for work-order generation on high-confidence events. Retrain with feedback; plan fleet rollout if metrics (precision, recall, MTTD) meet targets.

KPIs: alert precision (>60% initially, to be improved), mean time to detect, reduction in emergency repairs, and ROI (reduced downtime + spares savings vs investment).

11. Typical failures PdM can detect early (examples)

• Bearing defects (HF energy and crest factor rise).

• Fuel injector spray issues (increased SFOC, combustion noise signatures).

• Lubrication degradation and contamination (particle counts, wear metals).

• Pump cavitation or impeller damage (pressure fluctuations + acoustic signatures).

• Generator winding or excitation issues (current harmonics, temperature trends).

Vibration and oil analytics are repeatedly cited as high-value sensors for these failure modes. Wartsila.com+1

12. Challenges & mitigation

• Label scarcity — use unsupervised methods, simulation/digital twin-generated failure data, and cross-vessel transfer learning.

• Fleet heterogeneity — cluster assets by type and either train per-cluster models or use meta-learning techniques.

• Connectivity limits — favor edge preprocessing; transfer only compressed summaries or selected raw windows.

• Crew acceptance — involve operators early, tune thresholds, and make alerts actionable with clear instructions.

13. Future directions

• Digital twins for scenario simulation and synthetic failure generation. ScienceDirect

• Federated learning to let OEMs and operators jointly train models without sharing raw data (preserving privacy/IP).

• Tighter OEM/operator collaboration and class society frameworks for validating PdM outcomes and using PdM to inform statutory surveys.

Conclusion & recommended next steps

AI-based PdM for marine auxiliary engines is a high-impact, practical way to reduce unplanned downtime and optimize lifecycle costs—provided you pair good sensors with pragmatic ML approaches and strong operational integration. Start with a focused pilot on the highest-value auxiliary asset (gensets or seawater pumps), emphasize vibration + oil monitoring, deploy hybrid edge/cloud architecture, and embed an engineer feedback loop for continuous improvement.