A methodology to develop and manage data-driven models for marine engine long-term health prognosis

Area of Science:

• Marine Engineering
• Data-Driven Modeling
• Prognostics and Health Management (PHM)

Background:

• Ship machinery requires robust Prognostics and Health Management (PHM) for operational reliability.
• Existing PHM methods often lack comprehensive data-driven model management strategies.
• Exhaust valve wear is a critical degradation factor in marine engines.

Purpose of the Study:

• To propose a novel methodology for developing and managing data-driven models for ship machinery PHM.
• To investigate exhaust valve wear degradation in a four-stroke marine engine.
• To introduce an integrated approach combining physics-based digital twins and data-driven model management.

Main Methods:

• Generated simulated datasets using a physics-based digital twin integrated with stochastic degradation models.
• Developed Health Indicator (HI) construction and forecast sub-models using Multi-Layer Perceptron and Bayesian Neural Networks.
• Implemented data-driven model management using error and uncertainty metrics for sub-model re-training.

Main Results:

• Achieved significant improvements in forecast accuracy, with R-squared values increasing from 0.24 to 0.89 (Case 1) and 0.26 to 0.94 (Case 2).
• Demonstrated the effectiveness of integrating thermodynamic and stochastic degradation models for marine engine digital twins.
• Validated the contribution of data-driven model management to enhancing PHM system performance.

Conclusions:

• The proposed methodology offers a robust framework for developing and managing data-driven PHM models for marine engines.
• The integration of digital twins with stochastic degradation models and advanced data-driven techniques is crucial for accurate machinery health prediction.
• This work facilitates the wider adoption of advanced PHM systems in the maritime industry.