A novel convolutional network with a self-adaptation high-pass filter for fault diagnosis of wind turbine gearboxes

Abstract

Abstract The fault diagnosis of a wind turbine gearbox is helpful for reducing the operating costs and risks of wind power systems. However, existing machine-learning-based gearbox fault diagnosis methods have two shortcomings: (a) data samples of gearbox faults are always scarce; and (b) due to the complex structure of gearboxes, the collected vibration signals often contain a large amount of low-frequency noise, which is detrimental to both feature extraction and fault diagnosis. To solve the above two problems, a combination of deep convolutional generative adversarial networks (DCGANs) and a convolutional network with a high-pass filter (CNHF) is proposed in this paper. Among them, the DCGAN combined with one-dimensional (1D) vibration data converted to a grayscale map is used to expand the fault data to solve the problem of a lack of fault data samples. The CNHF is realized by adding an adaptive high-pass filter to the conventional convolutional layer, and the threshold of the high-pass filter is adaptively set by the 1D convolution according to different data characteristics, thus greatly filtering out the interference of low-frequency noise and realizing the accurate diagnosis of faults. Experiments are performed on a drivetrain dynamics simulator rig to verify the efficacy of the proposed method.