Fault detection of offshore wind turbine gearboxes based on deep adaptive networks via considering Spatio-temporal fusion

Abstract

To fully use the limited monitoring data with fault information for anomaly detection of the wind turbine gearbox (WTG) for operational state-based maintenance strategy optimization, reliability improvement, and cost savings of a wind turbine, we propose a combined operating state prediction method based on deep learning, fuzzy synthesis, and deep domain adaptive networks. After having filtered the monitoring data with fault information of 11 WTGs from offshore and onshore wind farms for nearly two years, we calibrate the operational state based on deep learning and fuzzy synthesis. Accordingly, the time-frequency domain analysis on condition monitoring system (CMS) data verifies the calibration results. Correspondingly, three different domain distribution distances are used to narrow the discrepancy in the data distribution of each WTG in the deep domain adaptation network. The results demonstrate that the proposed state calibration method can detect potential fault information in advance. The deep domain adaptation network can reduce the data distribution discrepancy between onshore and offshore WTGs, with the classification accuracy reaching above 0.9. This method can be used as a basis for the universality of monitoring data between different types of WTG, both offshore and onshore.