基于LSGAN-SqueezeNet的轴承故障诊断研究

A study on bearing fault diagnosis based on LSGAN-SqueezeNet

在滚动轴承的故障诊断过程中,需要大量的故障样本对模型进行训练,但由于工程实际环境较为复杂,难以采集到足量的轴承故障样本。提出基于最小二乘生成对抗网络(least squares generative adversarial networks,LSGAN)结合SqueezeNet在样本数据量不足条件下滚动轴承故障诊断的方法。该方法首先利用滚动轴承的原始样本训练LSGAN,生成不同状态下的轴承信号样本,采用SqueezeNet对真实样本与生成样本进行模型训练,实现样本数据量不足条件下的轴承故障诊断。通过凯斯西储大学及德国帕德博恩大学的滚动轴承试验数据对所提轴承故障诊断方法进行验证。试验结果表明,真实样本结合生成样本训练模型,对故障的诊断准确率均达到99.7%,该方法具有较高的诊断精度。实现了样本数据量不足条件下故障诊断的可行性及验证其泛化性,且有效提高了故障诊断的准确率、降低了训练成本。

In the process of rolling bearing fault diagnosis,a large number of fault samples are needed to train amodel.It is difficult to collect a sufficient number of bearing fault samples due to the complex engineering environment.A new method of rolling bearing fault diagnosis based on the least squares generated adversary networks(LSGAN)and the SqueezeNet under the condition of insufficient sample data was proposed.Firstly,LSGAN was trained by using the original samples of rolling bearings to generate bearing signal samples in different states.Real samples and generated samples were trained by the SqueezeNet,and the bearing fault diagnosis was realized under the condition of insufficient sample data.The proposed method was verified by the rolling bearing experimental data of Case Western Reserve University and Paderborn University.The experimental results show that accuracy of the fault diagnosis is 99.7%.The LSGAN-SqueezeNet model has high diagnostic accuracy.It not only realizes the feasibility of fault diagnosis and verifies its generalization,but also effectively improves the accuracy of fault diagnosis and reduces the training cost.