A hybrid attention improved ResNet based fault diagnosis method of wind turbines gearbox. (July 2021)

• Record Type: Journal Article
• Authors: Zhang, Kai; Tang, Baoping; Deng, Lei; Liu, Xiaoli
• Is Part Of: Measurement. Volume 179(2021)
• Publication Date: 2021-07
• Subjects: Wind turbines -- ResNet -- Attention mechanism -- Fault diagnosis -- Wavelet transform; WTs Wind turbines -- TML Traditional machine learning -- CNN Convolutional neural networks -- ResNet Residual network -- WPT Wavelet packet transformation -- HA-ResNet Hybrid attention improved resnet -- ReLU Rectified linear unit -- BN Batch normalization -- BW Band weights -- DDS Drivetrain diagnostic simulator -- WDCNN Convolutional neural networks with wide first-layer kernels -- SGDR Stochastic gradient descent with warm restarts -- t-SNE t-distributed stochastic neighbor embedding -- CMS Condition monitoring system; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2021.109491
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: The fault diagnosis framework combines the wavelet transform and the ResNet. Frequency attention highlights weak but crucial features in wavelet coefficients. Channel attention improves the representation of the learned feature-map. Hybrid attention-based method boosts ResNet's performance in fault diagnosis. Abstract: It is significant to boost the performance of fault diagnosis of wind turbine gearboxes. In this paper, a hybrid attention improved residual network (HA-ResNet) based method is proposed to diagnose the fault of wind turbines gearbox by highlighting the essential frequency bands of wavelet coefficients and the fault features of convolution channels. First, the paper performed wavelet packet transformation (WPT) on the raw signal and improved the ResNet by the band attention to highlight features of wavelet coefficients. Second, a fault diagnosis framework based on channel attention is designed to effectively improve the nonlinear feature extraction ability of deep convolutional networks. The proposed method is verified by a simulation dataset of the drivetrain diagnostic simulator (DDS) and the measured data from a wind farm. The results illustrate the superior performance of the HA-ResNet based fault diagnosis method for time–frequency feature extraction of vibration signals, frequency band information enhancement, and recognition accuracy improvement.