A physics-informed machine learning framework for predictive maintenance applied to turbomachinery assets

The paper presents an overview of an analytics framework for predictive maintenance service boosted by Machine Learning and asset knowledge, applied to turbomachinery assets. Optimization of the maintenance scenario is performed through a risk model that assesses online health status and probability of failure, by detecting functional anomalies or aging phenomena and evaluating their impact on asset serviceability. Turbomachinery domain knowledge is used to create physics-based models, to configure a severity assessment layer and to properly map maintenance actions to anomaly types. The implemented analytics framework is able also to fore-cast engine behaviour over the future in order to optimize asset operation and maintenance, minimizing downtime and residual risk. Predictive cap-abilities are optimized thanks to the hybrid approach, where physics-based knowledge empowers long term prediction accuracy while data-driven analytics ensure fast-events prognostics. Accuracy of the hybrid approach improves maintenance optimization, allowing activities to be planned properly and in early advance with respect to outage execution