A novel method based on nonlinear auto-regression neural network and convolutional neural network for imbalanced fault diagnosis of rotating machinery. (September 2020)

• Record Type: Journal Article
• Authors: Zhou, Quan; Li, Yibing; Tian, Yu; Jiang, Li
• Is Part Of: Measurement. Volume 161(2020)
• Publication Date: 2020-09
• Subjects: Fault diagnosis -- Rotating machinery -- Convolutional neural network -- Data imbalance -- Nonlinear auto-regressive neural network; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2020.107880
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: Nonlinear auto-regression neural network (NARNN) is applied to expand the minority fault samples. Convolutional neural network (CNN) is applied to automatically extract features and achieve fault diagnosis. A novel method based on CNN and NARNN for imbalanced fault diagnosis of rotating machinery is presented. The effectiveness of the proposed method has been respectively verified by a bearing dataset and a gear dataset. The experimental results demonstrate its superiority over the other traditional methods under imbalanced datasets. Abstract: Although the diagnosis methods of rotating machinery based on convolutional neural network (CNN) have achieved great success, they generally assume the number of normal and fault samples is the same. However, it's difficult to obtain adequate fault samples. Moreover, CNN cannot well handle the imbalanced fault diagnosis. Nonlinear auto-regressive neural network (NARNN) has strong prediction ability and can expand the small number of fault samples. Thus, a novel fault diagnosis approach combining CNN with NARNN has been proposed. First, NARNN is applied to expand the small number of samples. Thereby, the sample sizes of different health conditions are equal. Subsequently, continuous wavelet transform is employed to convert the 1-dimensional vibration signals into 2-dimensional time-frequency images. Finally, CNN is established to automatically learn the characteristics and achieve fault identification. Through the comparative experiments, the superiority of the proposed method has been validated based on the two datasets with different imbalanced levels.