非均衡小样本条件下基于SAE - ACGANs的复杂 供输机构故障诊断方法  

Fault diagnosis method for complex feeding and ramming mechanisms based on SAE - ACGANs with unbalanced limited training data

作　　者：闫啸家 梁伟阁 张钢 佘博[1] 田福庆[1] YAN Xiaojia;LIANG Weige;ZHANG Gang;SHE Bo;TIAN Fuqing(College of Weaponry Engineering,Naval University of Engineering,Wuhan 430033,China;Missile and Gun Department,Dalian Naval Academy,Dalian 116000,China)

机构地区：[1]海军工程大学兵器工程学院,武汉430033 [2]大连舰艇学院导弹与舰炮系,辽宁大连116000

出　　处：《振动与冲击》2023年第2期89-99,共11页Journal of Vibration and Shock

基　　金：国家自然科学基金(61640308);湖北省自然科学基金(2019CFB362)。

摘　　要：非均衡和小样本问题是制约深度学习技术在复杂供输机构故障诊断领域中应用效果的关键因素。为克服传统深度学习方法难以获取小样本数据内在分布和传统非均衡数据处理方法未考虑类别信息均衡化的缺陷,提出一种基于稀疏自编码辅助分类生成式对抗网络(sparse autoencoder auxiliary classifier generative adversarial networks,SAE-ACGANs)的复杂供输机构故障诊断方法。首先,对供输机构振动信号进行连续小波变换(continuous wavelet transform,CWT),得到反映信号时频特征的二维时频图;然后,运用稀疏编码器提取图像特征,并与类别信息融合为隐变量,强化其表征与图像所属类别相关特征的能力;生成器将融合隐变量映射为与真实样本分布类似的生成样本,从而扩展训练数据集;判别器从扩展数据集中挖掘有效的深度特征,并实现样本真伪和类别的判断;最后,通过优化后的生成器和判别器对抗学习训练机制,相互交替优化以达到纳什平衡,提高方法在非均衡小样本条件下样本生成质量和故障判定能力。复杂供输机构台架试验研究结果表明:SAE-ACGANs框架可以充分学习输入样本的内在分布和深度特征,相较于原ACGANs框架提升了判别器的性能,实现了模型收敛速度、训练精度和稳定性的提升;相较于传统非均衡数据处理算法,模型有效改善多数类分类偏好的影响,对于少数类故障样本的识别能力大幅提升。The problem of imbalance and limited training data is the key factor that restricts the application effect of deep learning technology in the field of fault diagnosis of complex feeding and ramming mechanisms. In order to overcome the shortcomings of traditional deep learning methods that are difficult to obtain the internal distribution of limited data and the defect that traditional unbalanced data processing methods do not consider the equalization of category information, a fault diagnosis method for complex feeding and ramming mechanisms based on the wavelet time-frequency diagram and sparse autoencoder auxiliary classifier generative adver sarial networks(SAE-ACGANs) was proposed. Firstly, the continuous wavelet transform(CWT) was performed on the vibration signal of the feeding and ramming mechanism to obtain a two-dimensional time-frequency diagram reflecting the time-frequency characteristics of the signal. Then, the sparse encoder in the model was used to extract image features which were afterwards merged with category information into hidden variables so as to strengthen the ability of latent variables to represent the characteristics related to the category of the image. By the generator, the fused latent variables were mapped to generated samples which had the sample distribution similar to the real one to expand the training data set. The discriminator mined effective depth features from the expanded data set and realized the judgment of the authenticity and category of the samples. Finally, through an adversarial learning and training mechanism, the optimized generator and discriminator alternately optimized each other to achieve the Nash balance. The method improves the sample generation quality and fault judgment ability with unbalanced limited training data. The research results of the bench tests for the complex feeding and ramming mechanism show that: the SAE-ACGANs framework can fully learn the internal distribution and depth characteristics of the input samples. Compared with the original A

关 键 词：故障诊断 生成式对抗网络(GANs) 非均衡数据 小样本 复杂供输机构 

分 类 号：TH17[机械工程—机械制造及自动化]