Deep convolutional generative adversarial network with semi-supervised learning enabled physics elucidation for extended gear fault diagnosis under data limitations
Deep convolutional generative adversarial network with semi-supervised learning enabled physics elucidation for extended gear fault diagnosis under data limitations
复制标题
DOI：
10.1016/j.ymssp.2022.109772
复制
发表时间：
2023-02
影响因子：
8.4
通讯作者：
K. Zhou;Edward J. Diehl;Jiong Tang
中科院分区：
文献类型：
--
作者：
K. Zhou;Edward J. Diehl;Jiong Tang
利用振动测量对齿轮系统进行故障检测和诊断，对保证齿轮系统的可靠性和安全性起着重要作用。计算智能通过各种代理模型利用分类，最近取得了一定程度的成功。然而，重大挑战依然存在。代理模型的建立通常需要大量的训练数据，这些训练数据具有与明确已知的齿轮故障状态相对应的特定标签，这在实际应用中可能是不可用的。由于高成本，可用数据的大小和相应的标签可能非常有限，这阻碍了以期望的可靠性诊断看不见的/意外的故障。在这项研究中，我们合成了一个深度卷积生成对抗网络（DCGAN）来应对这一挑战。这种新的方法遵循半监督学习的概念，其性能显着提高，通过引入额外的廉价的未标记的数据。DCGAN通过合理设计其结构实现了诊断器和生成器之间的对抗效应，从而在标记数据较少的情况下实现了高准确度的诊断。更重要的是，通过充分利用未标记数据中指向各种不可见故障的丰富故障特征，可以通过DCGAN中独特的半监督学习策略隐式地阐明不可见和已知故障之间的潜在物理学的内在相关性。因此，可以实现对训练数据集中已知故障以外的未知故障的扩展诊断能力，具有实际意义。系统的案例研究，从实验室规模的齿轮系统获得的实验数据进行验证新的诊断框架。
Fault detection and diagnosis of gear systems using vibration measurements play an important role in ensuring their functional reliability and safety. Computational intelligence, leveraging upon classification through various surrogate models, has recently demonstrated certain level of success. Major challenge however remains. The establishment of surrogate models generally requires large size of training data with specific labels corresponding to explicitly known gear fault conditions, which may not be available in practical applications. Both the size of available data and the respective labels may be quite limited due to the high cost, which hinders the diagnosis of unseen/unexpected faults with desired reliability. In this research we synthesize a deep convolutional generative adversarial network (DCGAN) to tackle this challenge. This new approach follows the semi-supervised learning concept, the performance of which is significantly enhanced by introducing additionally the inexpensive unlabeled data. The balanced adversarial effect between the discriminator and generator in DCGAN is realized by appropriately designing their architectures, which as a result can enable the high accuracy of diagnosis with scarce labeled data. More importantly, by taking full advantage of the rich fault signatures in the unlabeled data that point to the diverse unseen faults, the intrinsic correlation of underlying physics between the unseen and known faults can be implicitly elucidated via unique semi-supervised learning strategy featured in DCGAN. Therefore, the extended capability in diagnosing the unseen faults that are beyond the known faults in training dataset can be realized, which bears practical significance. Systematic case studies using experimental data acquired from a lab-scale gear system are carried out to validate the new diagnosis framework.