A Lesson on Operationalizing Machine Learning for Predictive Maintenance of Gas Turbines

OEMs, service providers and end-users are moving from preventative to predictive maintenance to minimize the risk of unwanted power plant shut-downs and to maximize profitability. Digital Twin and Machine Learning (ML) are important techniques in this transformation as it complements and improves the traditional expert-based knowledge systems. There is a continued trend to use data-driven, so-called black-box, ML techniques as an improvement over traditional statistical approaches. However, these ML approaches suffer from low interpretability or explainability, making it hard to trust how or why a certain anomaly in the system is detected, limiting the trust in the prediction and making it much less likely to identify the real original cause of the problem. In this paper, we present our lesson learnt from operationalizing ML in a real-world use case that studied data from the 1.85 MWe OPRA OP16 all radial single-shaft gas turbine. We comment on the unforeseen obstacles we uncovered during our ML anomaly detection application and juxtapose them with the high potential value that our novel ML applications and explanation method can provide. ML may be enticing for the predictive maintenance of gas turbines but our lesson makes it clear that operationalizing ML goes beyond merely algorithm specifics. In line with the nature of the Digital Twin, it requires careful consideration of the specialized IT system supporting the algorithm, and the specific process it supports and in which it is deployed.