New domain adaptation method in shallow and deep layers of the CNN for bearing fault diagnosis under different working conditions

• Published in: International journal of advanced manufacturing technology Vol. 124; no. 11-12; pp. 3701 - 3712
• Main Authors: Jin, Tongtong, Yan, Chuliang, Chen, Chuanhai, Yang, Zhaojun, Tian, Hailong, Guo, Jinyan
• Format: Journal Article
• Language: English
• Published: London Springer London 01.02.2023; Springer Nature B.V
• Subjects: Adaptation; Artificial neural networks; CAE) and Design; Computer-Aided Engineering (CAD; Domains; Engineering; Fault diagnosis; Industrial and Production Engineering; Industrial applications; Kernels; Machining; Mechanical Engineering; Media Management; Multilayers; Original Article; Rotating machinery; Working conditions; Deep learning; Fault diagnosis; Domain adaptation; Multilayer adaptation convolutional neural network; Maximum mean discrepancy
• ISSN: 0268-3768; 1433-3015
• DOI: 10.1007/s00170-021-07385-9

Deep learning models are widely used in fault diagnosis to learn hierarchical representations from collected signals. However, most of the models depend considerably on the assumption that training (source domain) and test (target domain) data sets are from the same feature distribution. This assumption is difficult to meet in practical scenarios of industrial applications because the working conditions of rotating machinery change with different machining tasks and labelled data with fault information are difficult and expensive to collect. Therefore, a novel fault diagnosis method, called multilayer adaptation convolutional neural network (MACNN), is constructed to solve the above-mentioned problems. The method regards raw temporal signals as input and uses wide kernels following a multiscale convolutional module to capture low-frequency features at multiple scales in shallow layers. Then, small convolutional kernels are used to implement multilayer nonlinear mapping in deep layers. Adaptive batch normalisation and multi-kernel maximum mean discrepancy are combined to reduce the feature distribution discrepancy in shallow and deep layers of the model, respectively, which improves the domain adaptation capability of the model. The proposed method is validated through 12 fault diagnosis experiments. The average 99.21% diagnosis precision demonstrates the reliability and stability of the method under different working loads.