基于并行卷积神经网络和特征融合的小样本轴承故障诊断方法

Small sample bearing fault diagnosis method based on parallel convolution neural network and feature fusion

在风力发电机轴承故障诊断过程中,基于深度学习的故障诊断方法受限于有限的标注样本,存在模型收敛困难和识别准确率较低等问题,为此,提出了一种基于并行卷积神经网络(P-CNN)和特征融合的小样本风机轴承故障诊断方法。首先,采用集合经验模态分解(EEMD)方法,将轴承的原始振动信号分解为若干个本征模态函数(IMF)分量以及残余分量;然后,分别对其进行了短时傅里叶变换(STFT),将其转换为时频特征图,同时构建了多个相同的卷积神经网络分支,以此作为特征提取器;最后,在融合层中,将提取到的时频域特征进行了通道特征融合,作为最终分类器的输入数据,对风机轴承进行了故障识别;并采用美国凯斯西储大学不同大小的轴承数据集,对该方法的适用性和有效性进行了验证。研究结果表明:在仅含有160个样本时,基于并行卷积神经网络(P-CNN)和特征融合的诊断方法的平均准确率高达94.5%;与支持向量机(SVM)、故障网络(FaultNet)、第一层宽卷积核深度卷积神经网络(WDCNN)相比,该诊断方法具有更高的准确率和更强的鲁棒性。

In the process of bearing fault diagnosis of wind turbine,the fault diagnosis method based on deep learning is limited by limited labeled samples,which has problems such as difficulties in model convergence and low recognition accuracy.For this purpose,a parallel convolutional neural network(P-CNN)and feature fusion-based fault diagnosis method for small sample wind turbine bearings was proposed.Firstly,the vibration signal of the bearing was decomposed into several intrinsic mode functions(IMF)components and residual components by ensemble empirical mode decomposition(EEMD).Then,the short time Fourier transform(STFT)was performed on them,and they were respectively converted into time-frequency characteristic maps,and multiple identical convolutional neural network branches were constructed as feature extractors.Finally,the extracted time-frequency domain features were fused in the fusion layer and used as the input of the final classifier to achieve fault identification of wind turbine bearings,the applicability and effectiveness of this method was validated using different size bearing datasets from Case Western Reserve University.The results show that the parallel convolutional neural network(P-CNN)and feature fusion-based fault diagnosis method has an average accuracy of 94.5%when containing only 160 samples,which has higher accuracy and stronger robustness compared to support vector machine(SVM)、FaultNet and deep convolutional neural networks with wide first-layer kernel(WDCNN).