Wind turbine pitch bearing fault detection with Bayesian augmented temporal convolutional networks
Chao Zhang, Long Zhang
Abstract
There are few studies on the fault diagnosis of deep learning in real large-scale bearings, such as wind turbine pitch bearings. We present a novel fault diagnosis method, Bayesian augmented temporal convolutional network (BATCN), to filter the raw signal in wind turbine pitch bearing defect detection. This method, which employs temporal convolutional neural networks, is designed to capture the temporal dependencies of the signal, with such a focus on non-stationary relationships in the collected signals. By referring to the thoughts of Bayesian optimization, our approach can spontaneously find the best patch length that influences fault signal extraction during the filtering process, avoiding manual tuning of this hyper-parameter. This BATCN method is first performed on simulation signals and an open-source dataset of general bearings, and then validated on industrial wind turbine pitch bearings both in the lab and in the real wind farm, where the bearings have been operated for over 15 years. The results show that our method can work well for large-scale slow-speed wind turbine pitch bearings.