摘要
在实际工况中,发电机中传感器采集到的故障样本数据有限,使用基于深度学习的方法进行故障诊断存在过拟合问题导致模型泛化能力较差以及诊断精度不高。为了解决这个问题,采用样本扩充的思路,提出了一种改进的辅助分类器条件深度卷积生成对抗网络(MACDCGAN)的故障诊断方法。通过对采集的一维时序信号进行小波变换增强特征,构建简化结构参数的条件深度卷积生成对抗网络模型生成样本,并在模型中采用Wasserstein距离优化损失函数解决训练过程中存在模式崩塌和梯度消失的缺点;通过添加一个独立的分类器来改进分类模型的兼容性,并在分类器中引入学习率衰减算法增加模型稳定性。试验结果表明,该方法可以有效地提高故障诊断的精度,并且验证了所提模型具有良好的泛化性能。
Under actual working conditions,fault sample data collected by sensors in generator are limited,using deep learning-based method for fault diagnosis has overfitting problems to result in poorer model generalization ability and low diagnostic accuracy.Here,to solve these problems,using a sample expansion idea,an improved fault diagnosis method called modified auxiliary classifier conditional deep convolutional generative adversarial network(MACDCGAN)was proposed.Wavelet transform was performed for the collected one-dimensional time series signals to enhance features.A conditional deep convolutional generative adversarial network model for simplifying structural parameters was constructed to generate samples.Wasserstein distance optimization loss function was used in the model to solve shortcomings of pattern collapse and gradient disappearance in training process.An independent classifier was added to improve the compatibility of the classification model,and the learning rate decay algorithm was introduced into the classifier to increase model stability.The experimental results showed that the proposed method MACDCGAN can effectively improve the accuracy of fault diagnosis;its good generalization performance is verified.
作者
曹洁
尹浩楠
王进花
CAO Jie;YIN Haonan;WANG Jinhua(College of Electrical&Information Engineering,Lanzhou University of Technology,Lanzhou 730050,China;Gansu Provincial Engineering Research Center for Manufacturing Information,Lanzhou 730050,China)
出处
《振动与冲击》
EI
CSCD
北大核心
2024年第11期227-235,共9页
Journal of Vibration and Shock
基金
国家自然科学基金项目(62063020
61763028)
国家重点研发计划项目(2020YFB1713600)
甘肃省自然科学基金项目(20JR5RA463)。
