A study on adaptation lightweight architecture based deep learning models for bearing fault diagnosis under varying working conditions. (1st December 2020)

• Record Type: Journal Article
• Authors: Wu, Jie; Tang, Tang; Chen, Ming; Wang, Yi; Wang, Kesheng
• Is Part Of: Expert systems with applications. Volume 160(2020)
• Publication Date: 2020-12-01
• Subjects: Diagnosis -- CNN -- Normalization -- Lightweight -- Transfer learning; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2020.113710
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: Bearing fault diagnosis under varying working conditions. Improving the quality of spectrums via considering practical distribution. Lightweight models implementation in industrial diagnostic scenarios. Realizing transfer learning with feature adaptation in unsupervised tasks. Abstract: Deep learning models have been widely studied in fault diagnosis recently. A mainstream application is to recognize patterns in spectrograms of faults. However, some common drawbacks still remain as following: a) Preprocess to improve the quality of spectrograms is rarely explored; b) Computing cost of a conventional CNN far exceeds the requirements of fast analysis in industry; c) Adequate labeled data cannot be acquired to train a comprehensive diagnosis model for varying working conditions. In this paper, an Adaptive Logarithm Normalization (ALN) is proposed to realize preprocess considering data distribution, it attempts to improve the quality of spectrograms via eliminating truncation phenomenon and enriching details simultaneously; Meanwhile, simplified lightweight models are built on the basis of present lightweight building blocks to reduce parameters, while maintaining high performances; Furthermore, an adaptation architecture is proposed by integrating Deep Adaptation Network (DAN) idea with simplified lightweight models, aiming at enhancing the generalization capability of models. Experiments have been carried out to implement the proposed methods with two different datasets. The overall success not only proves the methods feasible, but also indicates a possible diagnosis prospect for real industrial scenarios.