多尺度卷积神经网络小样本轴承故障辨识方法

Small Sample Bearing Fault Identification Method Using Novel Multi-scale Convolutional Neural Network

针对基于深度学习的故障辨识方法工程应用准确率受制于样本数量与质量的问题,提出一种多尺度卷积神经网络(novel multi‑scale convolutional neural network,简称NMS‑CNN)故障辨识方法。首先,对滚动轴承的振动信号进行快速傅里叶变换(fast Fourier transform,简称FFT)获取其频域数据;其次,利用多尺度卷积提取频域数据中的多粒度敏感特征,并使用实例归一化技术(instance normalization,简称IN)对特征图进行归一化;然后,采取注意力机制对多尺度特征进行自适应加权并进一步使用卷积提取深层抽象特征;最后,使用softmax分类器完成故障辨识任务。经过实验验证,所提方法能够在较少训练样本下完成故障辨识任务,并且其抗噪性与泛化性均优于其他智能故障辨识算法。

The existing deep learning fault diagnosis methods are difficult to obtain high fault classification accuracy with fewer training samples.A novel multi-scale convolutional neural network(NMS-CNN)fault identification method is proposed.First,fast Fourier transform(FFT)on the original vibration signal of the rolling bearing is performed to obtain its frequency domain information.Secondly,it is transmitted into a multi-scale convolutional neural network to extract the multi-granularity information in the data and use instance normalization(IN)for feature standardization.Then,the attention mechanism is used to adaptively weight the multi-scale features and further use convolution to extract deep abstract features.Finally,the softmax classifier is used to complete the fault identification task.After experimental verification,this method can complete the fault identifi-cation task excellently with fewer training samples.Moreover,its noise resistance and generalization are better than other mainstream intelligent fault identification algorithms.