A diagnostic framework for wind turbine gearboxes using machine learning

Operation and maintenance costs of wind turbines are highly driven by gearbox failures, especially offshore were the logistics of replacements are more demanding. It is therefore very critical to foresee incipient gearbox faults before they become catastrophic failures. Wind turbine gearbox condition monitoring is usually performed using vibration signals coming from accelerometers installed on the gearbox surface. The current monitoring practice is a rule-based approach, where alarms are activated based on thresholds. However, too much manual analysis may be required for some failure modes and this can become quite challenging as the installed wind capacity grows. Also, since false alarms have to be avoided, these thresholds are set quite high, resulting in late stage diagnosis of components. Given the fact there is a large amount of historic operating data with confirmed gearbox failure incidents, this paper proposes a framework that uses a machine learning approach. Vibration signals are used from the gearbox sensors and processed in the frequency domain. Features are extracted from the processed signals based on the fault locations and failure modes, using domain knowledge. These features are used as inputs in a layer of pattern recognition models that can determine a potential component fault location and failure mode. The proposed framework is illustrated using failure examples from operating offshore wind turbines.