Diagnosis of wind turbine faults with transfer learning algorithms

A framework of using transfer learning algorithms, Inception V3 and TrAdaBoost, for fault diagnosis of two wind turbine faults is presented and verified. Two failure modes, blade icing accretion and gear cog belt fracture, are analyzed using SCADA data. A new index named 'Comprehensive Index' is defined to evaluate performance of different algorithms. Traditional machine learning algorithms do not perform well for data sets that are unbalanced and follow different distributions. The former causes bias in classification and the latter leads to poor adaptability of algorithms. A novel transfer learning algorithm studied in this paper, TrAdaBoost, has been proved to have superior performance on dealing with data imbalance and different distributions. A new approach to calibrate data labels using transfer learning algorithms is also proposed, which provides important insights into unsupervised learning for wind turbine fault diagnosis. (c) 2020 Elsevier Ltd. All rights reserved.

Automated summary

This study presents a framework for fault diagnosis of wind turbine faults using transfer learning algorithms Inception V3 and TrAdaBoost, and introduces a new evaluation index 'Comprehensive Index'. Traditional machine learning algorithms perform poorly for unbalanced and differently distributed datasets, while the novel transfer learning algorithm TrAdaBoost shows superior performance in dealing with such challenges.