Abstract: In practical engineering applications,the time of rolling bearings fault state is very short.Due to the cost,it is unrealistic to work in the fault state for a long time,which will cause the imbalance of fault diagnosis data set.That is the normal samples are far more than the fault samples,and it will greatly affect the accuracy and stability of fault diagnosis.To solve this problem,a fault diagnosis method for bearing imbalance data set was proposed based on Wasserstein distance conditional gradient penalty adversarial generation net Conditional Wasserstein Generative Adversarial Network with gradient penalty (CWGAN-GP),which could stably generate high-quality samples.In the process of fault diagnosis,the quality of the generated samples was evaluated first,and then the unbalanced data set was gradually expanded and balanced.Experiments showed that the proposed method could generate generated samples that were highly similar to the real samples,and the accuracy of fault diagnosis was effectively improved as the unbalanced data set was gradually balanced.In addition,CWGAN-GP model performed better than other generation models in sample generation.

Key words: fault diagnosis, imbalanced data set, gradient penalty, generative adversarial net, rolling bearing

摘要： 在实际工程应用中,滚动轴承在大多数时间都工作在正常状态下,故障状态时间很短。由于成本,让其长时间工作在故障状态是不现实的。这将造成故障诊断数据集的不平衡,即正常的样本远远多于故障的样本,而这会极大地影响故障诊断结果的准确性和稳定性。针对该问题,提出一种基于Wasserstein距离条件梯度惩罚生成对抗网络(CWGAN-GP)的轴承不平衡数据集故障诊断方法,该方法能够稳定地生成高质量的样本。在故障诊断过程中,首先对生成样本的质量进行评估,然后对不平衡数据集进行逐步扩充与平衡。实验表明,该方法能够生成与真实样本高度相似的生成样本,并随着不平衡数据集被逐渐平衡,故障诊断的准确率也得到有效的提高。此外,CWGAN-GP模型在样本生成方面比其他生成模型具有更好的表现。

关键词: 故障诊断, 不平衡数据集, 梯度惩罚, 生成对抗网络, 滚动轴承