Zephyr: a Data-Centric Framework for Predictive Maintenance of Wind Turbines

Abstract

Because wind turbines often operate through harsh weather events, under variable operating conditions, and in difficult-to-access locations, turbine maintenance is often challenging and costly. In this thesis, we present Zephyr, a flexible machine learning framework for predictive maintenance of wind energy assets. Manual analysis of wind turbine data is difficult and time-consuming due to its volume, variety, and, most importantly, the need for quick detection of issues. Machine learning (ML) methods are able to automate large-scale data analysis. However, the enormous amount of contextual information required to actually understand the data impedes the ability of ML frameworks to provide actionable insights. To this end, Zephyr enables Subject Matter Experts (SMEs) to incorporate their knowledge at various stages of ML model development. The Zephyr framework consists of a signal-processing-based featurization library, a data labeling algorithm – which helps analyze operational data and maintenance events in order to create labels for machine learning problems – and a set of automated machine learning pipelines for predicting outcome types. SMEs incorporate their expertise by providing labeling functions, bands for frequency domain-based featurization, and several other inputs in an intuitive way. We demonstrate the efficacy of this framework through two case studies involving maintenance operation data from wind turbines. Moreover, we show that ML performance can increase when involving domain expertise by a value as high as 48%.