A novel fault diagnosis framework based on adaptive VAEGAN and optimal data selection for imbalanced data

Rolling bearings generally operate under normal conditions, making it difficult to collect fault data. This results in an imbalanced data distribution, which adversely affects the accuracy of rolling bearing fault diagnosis. However, current research on data augmentation still remains to be improved...

Full description

Saved in:
Bibliographic Details
Published in:Measurement : journal of the International Measurement Confederation Vol. 256; p. 118117
Main Authors: Hou, Yandong, Cai, Xiaoao, Chen, Zhengquan, Huang, Huige, Zhai, Xiaodong
Format: Journal Article
Language:English
Published: Elsevier Ltd 01.12.2025
Subjects:
ADS
AVAEGAN
Fault diagnosis
Imbalanced data
Rolling bearing
ADS
Fault diagnosis
AVAEGAN
Imbalanced data
Rolling bearing
ISSN:0263-2241
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Rolling bearings generally operate under normal conditions, making it difficult to collect fault data. This results in an imbalanced data distribution, which adversely affects the accuracy of rolling bearing fault diagnosis. However, current research on data augmentation still remains to be improved in adversarial intensity of the generative model training, and the selection of optimal data. To address this issue, this paper proposes a novel framework for data augmentation and fault diagnosis, called the Adaptive Variational Autoencoding Generative Adversarial Network and Adaptive Data Selector (AVAEGAN-ADS). Firstly, an adaptive training module is designed. By setting a conversion threshold to control the difference of loss values between the variational autoencoder and the discriminator, the difference of their learning ability is ensured to be within a controllable range, which improves the adversarial intensity of the model training. Secondly, the ADS is constructed by combining sensitivity and comprehensiveness of the two evaluation metrics Maximum Mean Discrepancy (MMD) and Wasserstein Distance (WD), enabling identification of optimal generated data to reconstruct a balanced data. Finally, Multi-scale Convolutional Neural Network and Long Short-Term Memory (MCNN-LSTM) is trained using the balanced data for fault diagnosis. Through the above data augmentation and optimal data selection, the accuracy of fault diagnosis is significantly improved. Experimental validation is conducted on different bearing datasets, including the HENU, CWRU and NASA datasets. AVAEGAN-ADS achieves accuracy of 97.14%, 97.52% and 98.17% on these datasets, demonstrating its effectiveness. •A novel fault diagnosis framework AVAEGAN-ADS for imbalanced data.•The adaptive training module controls the training of the model.•The optimal selection of data is realized through the adaptive data selector.•The fault diagnosis accuracy is greatly improved after data augmentation.
ISSN:0263-2241
DOI:10.1016/j.measurement.2025.118117