数据不平衡故障诊断：一种预训练数据增强方法

Fault Diagnosis for Imbalanced Data:a Pre-Trained Data Enhancement Method

中文关键词:  航空发动机  样本生成  数据稀缺性  数据不平衡  生成对抗网络

英文关键词:aeroengine  sample generation  data scarcity  data imbalance  GAN

中文摘要:

      针对实际航空发动机滚动轴承故障诊断应用中健康-故障数据不平衡的问题，提出一种结合梯度惩罚与辅助分类器的Wasserstein生成对抗网络（Wasserstein Generative Adversarial Network Gradiend Penalty,WGANGP）的增强诊断方法——预训练数据增强-WGANGP（PDA-WGANGP）。首先，模拟实际情况，利用健康数据和少量故障数据对网络进行预训练；其次，将训练好的网络结构和参数作为判别器和分类器的前端特征提取层；最后，通过引入残差网络，构建一个全新的生成器，从而稳定地生成高品质的故障样本，平衡测试数据集，完成网络结构的优化训练。通过对滚动轴承开展不平衡数据下的诊断应用与验证，证明了PDA-WGANGP在样本生成和高效诊断中的可行性，以及相较于同类方法的诊断性能优越性。

英文摘要:

      To solve the problem of health-fault data imbalance in the real applications of aeroengine rolling bearing fault diagnosis,a rolling bearing enhanced diagnosis method combining gradient punishment and auxiliary classifier Wasserstein generative adversarial network(WGANGP) is proposed:pre-trained data augment-WGANGP (PDA-WGANGP).Firstly,the actual situation is simulated,and the network is pre-trained by using health data and a small amount of fault data.Secondly,the trained network structure and parameters are used as the front-end feature extraction layer of discriminator and classifier.Finally,a new generator is constructed by introducing the residual network,so as to stably generate high-quality fault samples,balance the test data set,and complete the optimization training of network structure.Through the diagnostic application and verification of rolling bearing under imbalanced data,it is proved that PDA-WGANGP is feasible in sample generation and diagnosis,and its performance is superior to other widely used methods.