Method of state identification of rolling bearings based on deep domain adaptation under varying loads. Issue 3 (1st May 2020)

• Record Type: Journal Article
• Authors: Kang, Shouqiang; Chen, Weiwei; Wang, Yujing; Na, Xiaodong; Wang, Qingyan; Mikulovich, Vladimir Ivanovich
• Is Part Of: IET science, measurement & technology. Volume 14:Issue 3(2020)
• Page Start: 303
• Page End: 313
• Publication Date: 2020-05-01
• Subjects: fault diagnosis -- rolling bearings -- belief networks -- Gaussian processes -- vibrations -- convolution -- feature extraction -- production engineering computing; frequency‐domain amplitudes -- rolling bearings -- joint distribution adaptation -- labelled source domain -- unlabelled target domain -- labelled vibration data -- data distribution -- multiple‐state identification method -- deep domain adaptation method -- deep belief network -- convolutional Gaussian–Bernoulli DBN -- deep generalised feature extraction; Measurement -- Periodicals; Electrical engineering -- Periodicals; Electronics -- Periodicals; Nanotechnology -- Periodicals; Electromagnetism -- Periodicals; Medical instruments and apparatus -- Periodicals; 621.3
• DOI: 10.1049/iet-smt.2019.0043
• Languages: English
• ISSNs: 1751-8822

Abstract

Abstract : Large amounts of labelled vibration data of rolling bearings are difficult to acquire in full during operating conditions under varying loads. Moreover, a large divergence in data distribution exists between source and target domains for the same state. A multiple‐state identification method for rolling bearings under varying loads is proposed. The deep domain adaptation method integrates the convolutional and pooling theory with the deep belief network (DBN) that enables the construction of a convolutional Gaussian–Bernoulli DBN, which is used to extract the deep generalised features from the frequency‐domain amplitudes of the rolling bearings. The weighted mixed kernel is then used instead of the single kernel to improve the joint distribution adaptation, which is used to process the features of both the labelled source domain and the unlabelled target domain for domain adaptation, and reduce the distribution divergence. Finally, the k ‐nearest neighbour algorithm is used for identification. Experimental results show that the proposed method can make full use of unlabelled data, mine the deep features of vibration signals, and reduce the divergence between data of the same state. In resolving the multiple‐state identification of rolling bearings under varying loads, a higher accuracy is attained in the identification.