Discriminative manifold random vector functional link neural network for rolling bearing fault diagnosis

• Published in: Knowledge-based systems 2021-01, Vol.211, p.106507, Article 106507
• Main Authors: Li, Xin, Yang, Yu, Hu, Niaoqing, Cheng, Zhe, Cheng, Junsheng
• Format: Article
• Language: English
• Subjects: Discriminative manifold random vector functional link neural network; Fault diagnosis; Iterative methods; Machine learning; Manifolds (mathematics); Neural networks; Roller bearings; Rolling bearing; Similarity; Soft label matrix; Within-class similarity graph
• ISSN: 0950-7051; 1872-7409
• DOI: 10.1016/j.knosys.2020.106507

Random vector functional link neural network (RVFLNN) is an effective and powerful neural network model, and it has been commonly used for various engineering applications. In this paper, an improved RVFLNN model called discriminative manifold RVFLNN (DMRVFLNN) is proposed and applied for rolling bearing fault diagnosis, which has the following outstanding characteristics. First, DMRVFLNN relaxes the strict one-hot label matrix into a soft label matrix to fully exploit the interclass discriminative information and simultaneously enlarge the distances between interclass samples. Second, a within-class similarity graph based on manifold learning is designed for DMRVFLNN to force interclass samples to keep close together as far as possible. In the construction process of the within-class similarity graph, cosine distance is used to measure the distances between samples to enhance its robustness. Moreover, we design an effective update method solution for DMRVFLNN with a closed-form solution in each iteration. Two rolling bearing datasets are used to verify the fault diagnosis performance, and the experimental results prove the effectiveness and superiority of the proposed method for rolling bearing fault diagnosis. •A novel DMRVFLNN model is proposed for rolling bearing fault diagnosis.•DMRVFLNN uses a soft label matrix to enlarge the distances of interclass samples.•DMRVFLNN uses a manifold graph to enhance the compactness of within-class samples.•We devise an iterative update method to solve the DMRVFLNN model.•Two fault datasets of bearing are used to verify the superiority of DMRVFLNN.