A new ensemble convolutional neural network with diversity regularization for fault diagnosis. (January 2022)

• Record Type: Journal Article
• Authors: Wen, Long; Xie, Xiaotong; Li, Xinyu; Gao, Liang
• Is Part Of: Journal of manufacturing systems. Volume 62(2022)
• Page Start: 964
• Page End: 971
• Publication Date: 2022-01
• Subjects: Fault diagnosis -- Ensemble learning -- Deep learning -- Generalization ability; Manufacturing processes -- Periodicals; Production engineering -- Data processing -- Periodicals; Robots, Industrial -- Periodicals; Production, Technique de la -- Informatique -- Périodiques; Robots industriels -- Périodiques; Electronic journals; 670.42
• DOI: 10.1016/j.jmsy.2020.12.002
• Languages: English
• ISSNs: 0278-6125

Abstract

Highlights: This paper proposed a new diversity regulation to improve the generalization ability of DL method. An improved snapshot ensemble convolutional neural network is proposed based on the diversity regulation. The proposed ISECNN is conducted on the CWRU and MFPT bearing dataset, and the results validates its performance. Abstract: Fault diagnosis is an essential technique to ensure the safety in modern industry. With the development of smart manufacturing, deep learning (DL) has been widely used to handle with massive mechanical data in fault diagnosis. However, the individual DL method suffers from the low generalization ability. In this research, a new improved snapshot ensemble Convolutional Neural Network (ISECNN) is proposed in order to obtain a stable and well-performed DL based fault diagnosis method. ISECNN applies the diversity regularization to generate several local minima and keeps their diversity during the training process, as the increasing of the diverse would promote the generalization ability of the group of local minima. Then, ISECNN combines all the local minima to form the ensemble method. The proposed ISECNN has been conducted on two famous bearing datasets. The prediction accuracy and the standard deviation are applied as the criterion. The experimental results show that ISECNN can increase the generalization ability without decreasing the prediction accuracy. ISECNN is also compared with traditional DL and machine learning methods, and the results validate the potential performance of ISECNN in the fault diagnosis field.