基于多传感器的不平衡数据轴承故障诊断

Fault Diagnosis for Bearings Based on Unbalanced Data of Multi-Sensor

针对轴承故障数据类别不平衡导致诊断模型精度下降的问题,使用多传感器数据丰富数据特征,同时从数据增强和损失函数两方面着手,提出一种基于多传感器的不平衡数据轴承故障诊断方法。首先,设计一种多通道的带辅助分类器的生成对抗网络(ACGAN),利用独特的鉴别-分类结构生成高质量的多传感器数据以补充原始数据集;然后,提出一种改进的均值焦点损失函数(m-Focal Loss),将不平衡问题转化为样本的难易分类问题,根据难易程度进行加权,从而进一步提高诊断精度;最后,将一维卷积神经网络(1DCNN)作为分类网络,在凯斯西储大学(CWRU)轴承数据集和实验室数据集上进行消融试验。结果表明:使用多传感器数据可以有效抑制不平衡数据带来的精度下降问题,加入ACGAN所生成的样本可大大提高不平衡数据下轴承故障诊断模型的精度。

Aimed at the problem that the accuracy of diagnostic model decreases due to unbalanced bearing fault data categories,the multi-sensor data is used to enrich the data features,and a fault diagnosis method for bearings based on unbalanced data of multi-sensor is proposed from two aspects of data enhancement and loss function.Firstly,a multi-channel auxiliary classifier generative adversarial networks(ACGAN)is designed to generate high-quality multi-sensor data to supplement the original dataset by using a unique discrimination-classification structure.Then,an improved mean focal loss function(m-Focal Loss)is proposed,which transforms the unbalanced problem into difficulty classification problem of samples and then weights them according to difficulty degree,so as to further improve the diagnostic accuracy.Finally,one-dimensional convolutional neural network(1DCNN)is used as classification network to perform ablation experiments on bearing datasets and laboratory datasets from Case Western Reserve University(CWRU).The results show that the accuracy reduction caused by unbalanced data can be effectively suppressed by using multi-sensor data,and the sample generated by ACGAN can greatly improve the accuracy of bearing fault diagnosis model under unbalanced data.