摘要
针对基于数据驱动的柴油机故障诊断方法在训练数据匮乏时易过拟合、准确率低的问题,提出一种基于人工数据融合的数据增强方法,实现训练数据的增广。该方法将Wasserstein距离与梯度惩罚法引入辅助分类生成对抗网络(auxiliary classifier generative adversarial network,ACGAN),解决原始ACGAN训练不稳定的问题;将优化前后的ACGAN生成的两种人工数据按比例引入原始训练集中,从强化原有数据和优化诊断网络判定范围两个角度对训练集进行数据增强。经柴油机故障诊断试验验证,采用该方法对训练集进行数据增强后,在不同故障类型下的诊断准确率均有明显提高,且效果优于其他对比方法。
Here,aiming at the problem of data-driven diesel engine fault diagnosis method being easy to over-fit and having low accuracy due to lack of training data,a data augmentation method based on artificial data fusion was proposed to realize augmentation of training data.In this method,Wasserstein distance and gradient penalty method were introduced into the auxiliary classifier generative adversarial network(ACGAN)to solve the problem of the original ACGAN’s training being unstable.The two kinds of artificial data generated by ACGAN before and after optimization were introduced into the original training set in proportion,and the training set was augmented from two aspects of strengthening original data and optimizing judgment range of diagnosis network.Diesel engine fault diagnosis tests showed that diagnostic accuracies under different fault types are obviously improved by using the proposed method to augment training set’s data;the proposed method’s effect is better compared with other methods mentioned here.
作者
黄盟
毕晓阳
杨晓
李鑫
汤代杰
毕凤荣
HUANG Meng;BI Xiaoyang;YANG Xiao;LI Xin;TANG Daijie;BI Fengrong(State Key Lab of Engines,Tianjin University,Tianjin 300072,China;College of Mechanical Engineering,Hebei University of Technology,Tianjin 300401,China)
出处
《振动与冲击》
EI
CSCD
北大核心
2023年第13期278-286,共9页
Journal of Vibration and Shock
基金
河北省高等学校科学技术研究项目(QN2022159)。
