An enhancement denoising autoencoder for rolling bearing fault diagnosis

•A novel data preprocessing method is proposed when there is not enough data for the model.•Adjust the regularization parameters appropriately as the number of layers changes.•Improve the parameter norm penalty by combining with elastic net regularization. Denoising autoencoders can automatically le...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Measurement : journal of the International Measurement Confederation Jg. 130; S. 448 - 454
Hauptverfasser: Meng, Zong, Zhan, Xuyang, Li, Jing, Pan, Zuozhou
Format: Journal Article
Sprache:Englisch
Veröffentlicht: Elsevier Ltd 01.12.2018
Schlagworte:
Denoising autoencoder
Elastic net
Fault diagnosis
Hyperparameter
Regularisation
Fault diagnosis
Denoising autoencoder
Hyperparameter
Elastic net
Regularisation
ISSN:0263-2241, 1873-412X
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:•A novel data preprocessing method is proposed when there is not enough data for the model.•Adjust the regularization parameters appropriately as the number of layers changes.•Improve the parameter norm penalty by combining with elastic net regularization. Denoising autoencoders can automatically learn in-depth features from complex data and extract concise expressions, which are used in fault diagnosis. However, they still have many drawbacks: (1) unsatisfactory results when the input data is not substantial; (2) difficulty in optimising the hyperparameter; (3) inability of existing regularisation methods to combine the advantages of L1 and L2 regularisation. To overcome the aforementioned challenges, here, a new data preprocessing method was proposed to obtain the training data. By reusing the data points between the adjacent samples, the fault identifying rate was significantly improved. Considering the different resilience of each layer after regularisation, the proposed method could alter the hyperparameter by changing the unit numbers of each layer. For a better sparse representation, the norm penalty combined L1 and L2 norm penalties, motivated by the elastic net. Comparison with a normal denoising autoencoder verified the superiority of the proposed method.
ISSN:0263-2241
1873-412X
DOI:10.1016/j.measurement.2018.08.010