A neuro-wavelet based approach for diagnosing bearing defects. (October 2020)

• Record Type: Journal Article
• Authors: Gharesi, Niloofar; Arefi, Mohammad Mehdi; Razavi-Far, Roozbeh; Zarei, Jafar; Yin, Shen
• Is Part Of: Advanced engineering informatics. Volume 46(2020)
• Publication Date: 2020-10
• Subjects: Variational Mode Decomposition -- Wavelet Neural Network -- Wavelet Transform -- Machine Learning -- Generalized Discriminant Analysis -- Manifold Techniques -- CWRU Dataset; Computer-aided engineering -- Periodicals; Engineering -- Data processing -- Periodicals; 620.00285
• DOI: 10.1016/j.aei.2020.101172
• Languages: English
• ISSNs: 1474-0346

Abstract

Highlights: Designing a precise fault diagnosis system with an improved accuracy of 99.48%. Providing exhaustive experimentation of 27 integrated diagnostic schemes. A comparative study between variational mode decomposition and wavelet based methods. Examining projective methods and manifold modeling to remove redundant features. Abstract: In recent years advanced signal processing techniques are used increasingly to excavate the nonstationary vibration signals and extract elemental-fault information. However, managing and analyzing a multicomponent signal mixed with background noise using only a single analysis tool is not a simple task and may lead to low diagnostic accuracy and a delayed diagnosis. This paper introduces a novel intelligent neuro-wavelet based system with high diagnostic accuracy based on nonrecursive variational mode decomposition (VMD) and wavelet-based neural network, which mainly consists of three steps (i.e. feature extraction (FE), dimension reduction (DR), and fault classification). Firstly, the vibration signals are segmented and processed by a novel nonrecursive VMD, which can decompose the nonstationary signals into a series of discrete modes adaptively, to extract informative features from vibration signals. Multi-Class generalized discriminant analysis is then used in the second step that aims to reduce the dimension of the feature set and improve the computational burden by selecting meaningful information and removing redundant features. In the next step, the obtained features vector is fed to a state-of-the-art hierarchical multi-resolution classifier, so-called wavelet neural network (WNN), which possesses the advantages of both wavelet transform and artificial neural networks for the decision-making. Additionally, to evaluate the information extraction capability of VMD, the subsequent DR method and the calculation accuracy of WNN, other state-of-the-art techniques are used in this work. In this regard, the superiority of the proposed approach is also confirmed through an experimental comparison with published works in the literature.