Predictive maintenance modelling on offshore wind turbines: combining big data with expert knowledge

Abstract

The downward price pressures on wind power in the past fifteen years have spurred research into predictive maintenance modelling on wind turbines. This research explored fault prediction modelling using unprecedented quantities of wind turbine SCADA data, work-order data and expert knowledge.

This thesis details the development of prediction models for eight fault case studies. In each case study, methodologies for applying large wind turbine operational datasets to model development were explored and implemented.

These datasets were used to develop enhanced class labelling approaches for training datasets. Developed models underwent extensive validation testing using the large number of example faults available. These approaches were consolidated into a recommended framework for developing future models.

The developed models were implemented into the operational strategies of eighteen offshore wind farms. Hundreds of instances of site feedback from site technicians were used to determine an accurate picture of performance metrics such as lead time and precision.

The research established that interpretable models, developed through observation and rule extraction approaches, can be applied to multiple wind turbines across wind farms without compromising performance. Recommended training dataset sizes for producing robust classifiers were determined from experiments on two case studies.