基于集成GAN的旋转机械故障诊断方法研究  

Research on The Method of Rotating Machinery Fault Diagnosis Based on Integrated GAN

作　　者：高玉才 付忠广[1] 王诗云 谢玉存 杨云溪 GAO Yu-cai;FU Zhong-guang;WANG Shi-yun;XIE Yu-cun;YANG Yun-xi(Key Laboratory of Power Station Energy Transfer Conversion and System,North China Electric Power University,Ministry of Education,Beijing 102206,China)

机构地区：[1]华北电力大学电站能量传递转化与系统教育部重点实验室,北京102206

出　　处：《汽轮机技术》2022年第3期213-216,共4页Turbine Technology

基　　金：国家自然科学基金(50776029,51036002)。

摘　　要：旋转机械运行工况不稳定、故障类型多样,振动信号中所含的信号成份较为复杂,传统的故障诊断方法难以满足实际需求,利用大数据和神经网络技术实现旋转机械的故障诊断已经成为当前的研究热点。针对神经网络过拟合现象,提出一种基于生成对抗网络(GAN)的旋转机械故障诊断集成方法。首先利用小波变换将振动信号转换为小波时频图,然后搭建5组生成对抗网络分别对应旋转机械的5种运行状态,利用对抗训练机制进行预训练,减轻网络过拟合,再利用所有训练样本数据分别对每组判别器进行增强训练,最后综合各判别器的输出结果预测旋转机械的运行状态。实验表明,本文方法能够准确识别旋转机械的常见故障类型。The operation conditions of rotating machinery are unstable,fault types are diverse,and the signal components contained in the vibration signal are complex,so the traditional fault diagnosis methods can hardly meet the actual needs.For the phenomenon of neural network overfitting,this paper proposes an integrated method of rotating machinery fault diagnosis based on generative adversarial network(GAN).Firstly,wavelet transform is used to convert the vibration signal into wavelet time-frequency diagram,then five groups of adversarial networks are built corresponding to the five operating states of rotating machines,pre-training is performed by using the adversarial training mechanism to reduce network overfitting,then all the training sample data are used to enhance the training of each discriminator,and finally the output results of each discriminator are integrated to predict the operating state of rotating machines.The experiments show that the method in this paper can accurately identify the common types of faults in rotating machinery.

关 键 词：旋转机械 故障诊断 小波变换 生成对抗网络 集成方法 

分 类 号：TH17[机械工程—机械制造及自动化] TK14[动力工程及工程热物理—热能工程]