基于改进SAGGAN模型的齿轮故障分类方法研究  

Gear fault classification method based on improved SAGGAN model

作　　者：刘洋[1,2,3,4] 但斌斌[1,2,3] 易灿灿[1,2,3] 严旭果[1,2,3] 薛家成 LIU Yang;DAN Binbin;YI Cancan;YAN Xuguo;XUE Jiacheng(Key Laboratory of Metallurgical Equipment and Control Ministry of Education,Wuhan University of Science and Technology,Wuhan 430080,China;Hubei Key Laboratory of Mechanical Transmission and Manufacturing Engineering,Wuhan University of Science and Technology,Wuhan 430080,China;Institute of Precision Manufacturing,Wuhan University of Science and Technology,Wuhan 430080,China;Central Research Institute,Baosteel Group Corporation(Wuhan Iron and Steel Limited Technology Center),Wuhan 430080,China)

机构地区：[1]武汉科技大学冶金装备及其控制教育部重点实验室,湖北武汉430080 [2]武汉科技大学机械传动与制造工程湖北省重点实验室,湖北武汉430080 [3]武汉科技大学精密制造研究院,湖北武汉430080 [4]宝钢股份中央研究院(武钢有限技术中心),湖北武汉430080

出　　处：《机电工程》2024年第12期2185-2194,共10页Journal of Mechanical & Electrical Engineering

基　　金：国家自然科学基金资助项目(52205537)。

摘　　要：针对齿轮故障样本获取困难,导致深度学习驱动故障分类模型的可靠性和准确性不足这一问题,提出了一种基于改进自注意力门单元生成对抗网络(SAGGAN)的半监督齿轮故障分类模型。首先,为增强改进SAGGAN模型的特征表示能力,提升齿轮故障的半监督分类效果,在自注意力生成对抗网络(SAGAN)的基础上,引入了门控通道转换模块(GCT)、改进自注意力门控模块(SAG)和预训练的Inception V3分支;然后,使用齿轮故障实验装置采集齿轮断齿、磨损、周节误差和正常四种状态下的振动信号,并将数据划分为训练集、验证集与测试集;最后,将计算结果与现有的半监督分类方法:TripleGAN、Bad-GAN、Reg-GAN、SF-GAN进行了对比,并对改进模块进行了消融实验研究。研究结果表明:在标签样本为40、60、80、100时,改进SAGGAN模型的整体分类准确率分别为89%、90%、92%、94.25%,远高于其他四种方法,特别在只有少量标签样本情况下的优越性更为明显。以上结果揭示了改进的SAGGAN模型在齿轮故障分类领域中的实用性和优越性。It was difficult to acquire gear fault samples,and it compromised the reliability and accuracy of deep learning-driven fault classification models,therefore a semi-supervised gear failure classification model called improved self-attention and gate unit generated adversarial network(SAGGAN)built upon the improvements to the self-attention mechanism was proposed.Firstly,in order to enhance the feature representation capabilities of the proposed SAGGAN model and consequently improve the semi-supervised classification performance for gear failures,the enhancements were made to the existing self-attention generative adversarial network(SAGAN)framework by incorporating gated channel transformation(GCT),refining self-attention gate modules(SAG),and integrating pre-trained Inception V3 branches.Then,the vibration signals were collected from a gear failure experimental apparatus,capturing data across four states:gear breakage,wear,pitch error,and normal operation.The collected data was then partitioned into training,validation,and test sets for further analysis.Finally,the performance of the proposed SAGGAN model was compared against existing semi-supervised classification methods such as TripleGAN,Bad-GAN,Reg-GAN,and SF-GAN.Additionally,a study on the effectiveness of the enhancement modules was conducted through ablation experiments.The research results indicate that the improved SAGGAN model achieves significantly higher overall classification accuracy,particularly demonstrating superiority when the number of labeled samples is limited.Specifically,at label sample sizes of 40,60,80,and 100,the overall classification accuracies of the improved SAGGAN model are respectively 89%,90%,92%,and 94.25%,which surpasses the performance of the other four methods.This suggests that the improved SAGGAN model can effectively enhance classification performance,especially in scenarios with a limited number of labeled samples.The above results reveal the practicality and superiority of the improved SAGGAN model in the field of gear fa

关 键 词：齿轮故障 模式分类 自注意力门单元生成对抗网络 半监督学习 自注意力生成对抗网络 门控通道转换模块 自注意力门控模块 

分 类 号：TH132.41[机械工程—机械制造及自动化]