A class alignment method based on graph convolution neural network for bearing fault diagnosis in presence of missing data and changing working conditions. (August 2022)

• Record Type: Journal Article
• Authors: Kavianpour, Mohammadreza; Ramezani, Amin; Beheshti, Mohammad T.H.
• Is Part Of: Measurement. Volume 199(2022)
• Publication Date: 2022-08
• Subjects: Fault diagnosis -- Missing data -- Auto-regressive moving average filter -- Subdomain adaptation -- Graph convolution neural network; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2022.111536
• Languages: English
• ISSNs: 0263-2241

Abstract

Abstract: Bearing fault diagnosis in real-world applications has challenges such as insufficient labeled data, changing working conditions of the rotary machinery, and missing data due to multi-rate sampling of sensors. Despite the numerous applications of conventional deep learning (DL) and domain adaptation methods in bearing fault diagnosis, these methods face challenges. Domain adaptation techniques neglect alignment across subdomains with the same class, and DL techniques do not consider data relationships and interdependencies. To tackle these challenges, this paper introduces a novel semi-supervised method based on ARMA graph convolution, adversarial adaptation, and multi-layer multi-kernel local maximum mean discrepancy (MK-LMMD). Structural information of data is extracted using ARMA graph convolution, adversarial adaptation is employed to decrease structural distribution discrepancy in the domains, and MK-LMMD is used to align the classes. Additionally, ARMA graph convolution and MK-LMMD can aid in reducing distribution discrepancy caused by missing data and changing working conditions. Highlights: A new approach for diagnosing bearing faults has been presented in the context of missing data and changing operating conditions. The structural information is extracted via ARMA-based graph convolution. MK-LMMD is used to align classes and an adversarial domain discriminator is used to reduce structural distribution differences. The proposed method's performance is verified using various percentages of labeled data.