Details

Autor(en) / Beteiligte

• Sun, Hongchun
• Wang, Changdong
• Cao, Xu

Titel

An adaptive anti-noise gear fault diagnosis method based on attention residual prototypical network under limited samples

Ist Teil von

• Applied soft computing, 2022-08, Vol.125, p.109120, Article 109120

Ort / Verlag

Elsevier B.V

Erscheinungsjahr

2022

Quelle

Elsevier ScienceDirect Journals

Beschreibungen/Notizen

• Deep learning networks are widely used to realize the intelligent diagnosis of gear faults. However, the problem of the insufficient number of typical fault samples and strong noise often occur in practical applications. Thence, this paper proposes an adaptive anti-noise gear fault diagnosis method based on attention residual prototypical network (ARPN) under the limited sample. In order to maximize the characterization of implicit classification information under fewer samples, frequency slice wavelet transform is applied to convert the vibration signal into a time–frequency image. Then, the Bayesian optimization algorithm is introduced to automatically adjust hyperparameters to meet different application conditions. And the feature embedding stage combined with the improved Non-Local-Pooling-Attention module is constructed to capture the effective feature information better under strong noise conditions. Finally, the internal principle of the proposed model is analyzed based on the visualization process. Meanwhile, the initial application of an interpretable deep learning network in the classification of gear health status has been realized. The ARPN is verified on the Connecticut standard gear data set and the data set collected actually by the laboratory. The results show that the diagnostic accuracy of the ARPN is higher and has stronger recognition ability under different loads. •A novel adaptive-parameter-adjustment gear fault diagnosis method is proposed.•The NLPA module is developed to capture remote dependent features in limited noisy samples.•The improved prototypical network is applied to learn the internal relationships of fewer samples.•The proposed method has superior performance in generalization and robustness.