Stochastic simulation assessment of an automated vibration-based condition monitoring framework for wind turbine gearbox faults

Effectively monitoring the health of a wind turbine gearbox is a complex and often multidisciplinary endeavor. Recently, condition monitoring practices increasingly combine knowledge from fields like signal processing, machine learning, and mechanics. Such a diverse approach becomes necessary when dealing with the vast amount of data that is generated by the multitude of sensors that are typically placed on a wind turbine gearbox. Ideally, this approach needs to be automated and scalable as well, since it is unfeasible to perform all the necessary processing work manually in a continuous manner. This paper focuses on assessing the performance of such an automated processing framework for the case of gearbox fault detection using vibration measurements. A year of vibration measurements on a gearbox is simulated by stochastic variation of the operating conditions and the system behavior. A bearing fault is progressively introduced as to track the detection capabilities of the framework in such stochastic circumstances. The used signal model is based on previously obtained experience with experimental data sets originating from wind turbine gearboxes. The framework itself consists of multiple pre-processing steps where each step tries to deal with compensating for the external or unwanted influences such as speed variation or noise. Finally, multiple features are calculated on the pre-processed signals and trended as to see whether the processing scheme can provide any benefit compared to basic traditional statistical indicators. It is shown that the multi-step pre-processing approach is beneficial and robust for the advanced feature calculation and thus the early fault detection.