Hybrid data augmentation method for combined failure recognition in rotating machines

Rotating machines are frequently subject to a wide range of rough conditions, resulting in mechanical failures and performance degradation. Thus, it is important to apply proper failure detection and recognition techniques, such as machine learning algorithms, to prevent these issues early. In indus...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Journal of intelligent manufacturing Ročník 34; číslo 4; s. 1795 - 1813
Hlavní autoři: Martins, Dionísio H. C. S. S., de Lima, Amaro A., Pinto, Milena F., Hemerly, Douglas de O., Prego, Thiago de M., e Silva, Fabrício L., Tarrataca, Luís, Monteiro, Ulisses A., Gutiérrez, Ricardo H. R., Haddad, Diego B.
Médium: Journal Article
Jazyk:angličtina
Vydáno: New York Springer US 01.04.2023
Springer Nature B.V
Témata:
Advanced manufacturing technologies
Algorithms
Business and Management
Classification
Control
Data augmentation
Failure
Failure detection
Machine learning
Machines
Manufacturing
Mechatronics
Performance degradation
Processes
Production
Random noise
Recognition
Robotics
Rotating machinery
Rotating machines
Support vector machines
Combined failures recognition
Data augmentation
Imbalance
Rotating machines
Misalignment
Predictive maintenance
ISSN:0956-5515, 1572-8145
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Rotating machines are frequently subject to a wide range of rough conditions, resulting in mechanical failures and performance degradation. Thus, it is important to apply proper failure detection and recognition techniques, such as machine learning algorithms, to prevent these issues early. In industrial environments, little data exists regarding failure conditions, which hinders the training stage of the classification algorithms responsible for classifying the failures. Therefore, this work proposes a hybrid method of data augmentation to increase the number of minority class instances in order to improve classifier performance. The approach combines the synthetic minority over-sampling and the additive white Gaussian noise techniques to create a set of artificial signals. The results show that the proposal is able to achieve better results than applying those techniques separately and also when using an undersampling strategy. For comparison purposes, four machine learning classification methods were analyzed alongside our data augmentation proposal, namely, support vector machines, K -nearest neighbors, random forest and stacked sparse autoencoder. The proposed hybrid data augmentation method associated with stacked sparse autoencoder outperformed the other models obtaining an accuracy of 100% and a processing time of 0.13 s.
Bibliografie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:0956-5515
1572-8145
DOI:10.1007/s10845-021-01873-1