SDA: Regularization with Cut-Flip and Mix-Normal for machinery fault diagnosis under small dataset. (May 2021)

• Record Type: Journal Article
• Authors: Lv, Haixin; Chen, Jinglong; Zhang, Tianci; Hou, Rujie; Pan, Tongyang; Zhou, Zitong
• Is Part Of: ISA transactions. Volume 111(2021)
• Page Start: 337
• Page End: 349
• Publication Date: 2021-05
• Subjects: Deep learning -- Small dataset -- Data augmentation -- Regularization method -- Batch Normalization; Engineering instruments -- Periodicals; Engineering instruments; Periodicals; Electronic journals; 629.805
• DOI: 10.1016/j.isatra.2020.11.005
• Languages: English
• ISSNs: 0019-0578

Abstract

Abstract: Data-driven intelligent diagnosis model plays a key role in the monitoring and maintenance of mechanical equipment. However, due to practical limitations, the fault data is difficult to obtain, which makes model training unsatisfactory and results in poor testing performance. Based on the characteristics of 1-D mechanical vibration signal, this paper proposes Supervised Data Augmentation (SDA) as a regularization method to provide more effective training samples, which includes Cut-Flip and Mix-Normal. Cut-Flip is used directly on the raw sample without parameter selection. Mix-Normal mixes the data and labels of a random sample with a random normal sample at a certain ratio. The proposed SDA is verified on two bearing datasets with some popular intelligent diagnosis networks. Besides, we also design a Batch Normalization CNN (BNCNN) to learn the small dataset. Results show that SDA can significantly improve the classification accuracy of BNCNN by 10%–30% under 1-8 samples of each class. The proposed method also shows a competitive performance with existing advanced methods. Finally, we further discuss each data augmentation method through a series of ablation experiments and summarize the advantages and disadvantages of the proposed SDA. Highlights: SDA is proposed to augment small dataset with low calculation cost and stable improvement effect. Cut-Flip expands the boundary of samples, Mix-Normal further improves the generalization ability of model. The effectiveness of SDA is verified on two datasets with literature diagnosis methods. SDA has the potential to be a universal preprocessing method under small dataset.