改进生成式对抗网络的不均衡样本转子系统故障诊断  

An Improved Fault Diagnosis Method for Imbalanced Samples Rotor Systems Based on Generative Adversarial Networks

作　　者：李金赫 段礼祥[1,2] 姜垣良 冯斌 LI Jin-he;DUAN Li-xiang;JIANG Yuan-liang;FENG Bin(College of Safety and Ocean Engineering,China Univers ity of Petroleun(Beijing);Key Laboratary of Oi and Gas Production Safety and Emergeney Technology,Ministry of Emergency Managerment;Sino-Pipelire International Company Lirnited)

机构地区：[1]中国石油大学(北京)安全与海洋工程学院 [2]应急管理部油气生产安全与应急技术重点实验室 [3]中油国际管道有限公司

出　　处：《化工自动化及仪表》2025年第2期239-249,共11页Control and Instruments in Chemical Industry

基　　金：中石油战略合作科技专项“海外长输油气管道灾害监测预警及动力设施诊断技术研究”(批准号:ZLZX2020-05)资助的课题。

摘　　要：实际工业应用中转子系统能够采集到的故障样本远少于正常工况样本。针对不同类别样本数量不均衡时传统深度学习模型会倾向于预测出样本更多的类别而忽视较少出现的类别的现状,提出基于条件深度卷积生成式对抗网络(CDCGAN)结合改进的卷积块注意力机制(CBAM)+可变性卷积的有监督二维数据生成方法(CBAM-CDCGAN),实现不均衡样本转子系统的故障诊断。首先用小波变换将得到的振动数据转换为二维时频图像;之后将改进的CBAM注意力机制与可变形卷积分别嵌入生成式对抗网络的生成器与判别器中,并利用该生成网络进行样本生成;最后将生成样本与原始样本混合,划分为训练集和测试集,通过双路径网络进行训练和测试,结果表明:在样本不均衡比为1:10时,用CBAM-CDCGAN模型生成样本后进行故障诊断,转子和轴承故障识别的准确率较不均衡时分别提升16.10%和21.28%。In practical applications,the fault samples that rotor system can collect is far less than the samples of normal working conditions.Aiming at an imbalance of the samples in different categories,traditional deep learning models tend to predict more sample categories while ignoring those that appear less frequently.Therefore,a supervised two-dimensional data generation method based on conditional deep convolutional generative adversarial network(CDCGAN)and combined with improved convolutional block attention mechanism(CBAM)plus variability convolution(CBAM-CDCGAN)was proposed to realize fault diagnosis of the rotor system with unbalanced samples.Firstly,having the obtained vibration data transformed into a two-dimensional time-frequency image by the wavelet transform;and then,having the improved CBAM attention mechanism and deformable convolution embedded into the generator and discriminator of the generative adversarial network respectively,and the generative network used to generate samples;and finally,having the generated samples mixed with the original samples,divided into the training set and test set,and trained and tested by dual-path network.The results show that,when the sample imbalance ratio is 1:10,the CBAM-CDCGAN model can be used to generate samples for fault diagnosis.Compared with the unbalanced condition,the accuracy of rotor and bearing fault recognition can be increased by 16.10%and 21.28%respectively.

关 键 词：故障诊断 CBAM-CDCGAN模型 转子系统 双路径网络 样本不均衡 生成式对抗网络 

分 类 号：TP277[自动化与计算机技术—检测技术与自动化装置] TE95[自动化与计算机技术—控制科学与工程]