Enhancing a Data-Centric Framework for PredictiveMaintenance of Wind Turbines

Abstract

Predictive maintenance of wind turbines is a machine learning task aimed at minimizing repair costs and improving efficiency in the wind turbine and renewable energy industry. Existing machine learning solutions often fail to meet real-world deployment requirements due to fragmented pipelines, lack of domain integration, and reliance on black-box models. Zephyr, a data-centric machine learning framework, addresses these challenges by enabling Subject Matter Experts (SMEs) to incorporate their domain knowledge into the prediction process, and to leverage automated tools for labeling, feature engineering, and prediction tasks without requiring extensive technical knowledge. However, the current version of Zephyr still has limitations, including usability gaps and a reliance on external tools for certain steps. Case studies with real-world data from the renewable energy company Iberdrola demonstrate Zephyr’s potential to integrate domain expertise into wind turbine predictive maintenance (thus streamlining the process) but also expose a sub-optimal user experience. This thesis explores gaps in the current state of the Zephyr framework and proposes refinements to enhance its usability. Key improvements include the consolidation of current tooling and relevant external libraries into a single API, state management with careful logging and exception handling, and improved support for model evaluation. These enhancements aim to support seamless end-to-end predictive modeling workflows, and to provide a more refined and flexible user experience for the Zephyr user base.