Multiscale dynamic fusion prototypical cluster network for fault diagnosis of planetary gearbox under few labeled samples

• Published in: Computers in industry 2020-12, Vol.123, p.103331, Article 103331
• Main Authors: Li, Biao, Tang, Baoping, Deng, Lei, Yu, Xiaoxia
• Format: Article
• Language: English
• Subjects: Distance metric; Fault diagnosis; Few labeled samples; Joint learning schema; Multiscale dynamic fusion; Planetary gearbox
• ISSN: 0166-3615; 1872-6194
• DOI: 10.1016/j.compind.2020.103331

•Multiscale dynamic fusion module is designed to extract more discriminative features.•Prototype fuzzy c-means cluster algorithm is proposed to obtain refined prototypes.•Joint learning schema is adopted to jointly train the entire network. Deep learning has been widely applied for fault diagnosis of planetary gearbox because of its powerful feature extraction and nonlinear capabilities. However, due to the limited labeled samples of planetary gearbox in some industrial application scenarios, traditional deep learning methods based on sufficient labeled data cannot achieve satisfactory results. In this work, a novel deep learning method named multiscale dynamic fusion prototypical cluster network (MFPCN) is proposed to address this issue. Firstly, multiscale dynamic fusion module is designed to dynamically modulate multiscale features through a gate mechanism for obtaining more discriminative features under few labeled samples. After that, prototype fuzzy c-means cluster algorithm is proposed to refine the position of class prototypes in high-dimensional space by merging unlabeled samples information, making it more representative of the same class features, which provides a more accurate distance metric benchmark for prototypical nearest neighbor classifier. Finally, joint learning scheme is constructed by combining prototypical nearest neighbor classifier and global classifier to jointly train the entire network. Extensive experiments and the engineering application show that the proposed method is more effective in the case of few labeled samples than other methods.