A novel model with the ability of few-shot learning and quick updating for intelligent fault diagnosis

• Published in: Mechanical systems and signal processing Vol. 138; p. 106608
• Main Authors: Ren, Zhijun, Zhu, Yongsheng, Yan, Ke, Chen, Kaida, Kang, Wei, Yue, Yi, Gao, Dawei
• Format: Journal Article
• Language: English
• Published: Berlin Elsevier Ltd 01.04.2020; Elsevier BV
• Subjects: Algorithms; Capsule auto-encoder; Coders; Dynamic routing; Fault diagnosis; Feature extraction; Few-shot learning; Intelligent fault diagnosis; Machine learning; Performance degradation; Quick updating; Time compression; Training; Dynamic routing; Few-shot learning; Quick updating; Intelligent fault diagnosis; Capsule auto-encoder
• ISSN: 0888-3270; 1096-1216
• DOI: 10.1016/j.ymssp.2019.106608

•A capsule auto-encoder model, which can extract vector features, is proposed.•The model can achieve good results by using only a small number of samples.•The proposed method can significantly reduce the training time.•The proposed method has good anti-noise ability under the noisy environment. Both of traditional intelligent fault diagnosis (TIFD) based on artificial features and modern intelligent fault diagnosis (MIFD) based on deep learning have made healthy progress in recent times. But, the bulk of methods neglects the actual application environments of intelligent diagnosis: (1) There are only few samples of faults, which greatly limit the popularity of TIFD and MIFD. Therefore, it is urgent to develop intelligent models with the ability of few-shot learning. (2) The performance degradation will occur when the equipment runs for a long time, which requires intelligent diagnosis models to possess the ability of quick updating. In order to remedy these shortcomings, a capsule auto-encoder model based on auto-encoder and capsule network, namely CaAE, is proposed for intelligent fault diagnosis. By constructing a capsule auto-encoder, various meaningful feature capsules are extracted from the input data, and then these capsules are fused adaptively into status capsules by the dynamic routing algorithm to represent health statuses. After that, status capsules are fed into the classifier to distinguish health statuses. The extraction of feature capsules enhances the model’s ability of mining information, which reduces the dependence of CaAE on the number of samples and compresses training time by reducing layers of the network. The loss function in the model introduces the penalty for the samples classified correctly and the constraints on the capsule auto-encoder, which make the model fast in training and excellent in feature extracting. A bearing dataset is utilized to validate the performance of the proposed CaAE. The results indicate that CaAE is suitable for few-shot learning and quick updating of intelligent models. In addition, the model can also achieve satisfactory results under noisy environment.