An intelligent failure feature learning method for failure and maintenance data management of wind turbines

Abstract

This paper introduces an intelligent feature learning framework for the failure and maintenance data management of the wind energy sector. The framework employs Bidirectional Encoder Representations from Transformers and the Conditional Random Field model to intelligently identify failures in wind turbines. Additionally, a transfer training model is constructed to infer offshore wind turbine failures based on knowledge learned from onshore devices, which can address the insufficient knowledge of the offshore sector. The accuracy of the feature learning is enhanced by creating an adaptive resampling mechanism to detect features of rare failures often overlooked by high-frequency ones. Two failure and maintenance datasets, LGS-Onshore and LGS-Offshore, are collected and analysed to recognise differences in failure and maintenance between onshore and offshore wind turbines. The results demonstrate that this innovative data analysis framework outperforms existing methods, contributing to the wind energy sector's data foundation by providing essential datasets and new insights into wind farm operation and maintenance.