Fully interpretable neural network for locating resonance frequency bands for machine condition monitoring. (1st April 2022)

• Record Type: Journal Article
• Authors: Wang, Dong; Chen, Yikai; Shen, Changqing; Zhong, Jingjing; Peng, Zhike; Li, Chuan
• Is Part Of: Mechanical systems and signal processing. Volume 168(2022)
• Publication Date: 2022-04-01
• Subjects: Sparsity measure -- Fully interpretable neural network -- Machine condition monitoring -- Cyclo-stationarity; Structural dynamics -- Periodicals; Vibration -- Periodicals; Constructions -- Dynamique -- Périodiques; Vibration -- Périodiques; Structural dynamics; Vibration; Periodicals; 621
• DOI: 10.1016/j.ymssp.2021.108673
• Languages: English
• ISSNs: 0888-3270

Abstract

Graphical abstract: Proposed fully interpretable neural network architecture for machine condition monitoring. Highlights: A fully interpretable neural network for machine health monitoring is proposed. A fully interpretable neural network architecture consists of four physical hidden layers. The first to third hidden layers exhibit the cyclo -stationarity of repetitive transients. Sparsity measures are used to characterize the cyclo -stationarity of repetitive transients. An iterative optimization strategy is developed to optimize the proposed neural network. Abstract: In recent years, various neural networks have been developed to process vibration signals for machine condition monitoring. Nevertheless, the physical interpretation of neural networks is still on-going and not fully explored. This paper aims to design a fully interpretable neural network for machine condition monitoring from the aspects of signal processing and physical feature extraction. The main idea of the fully interpretable neural network is to extend the uninterpretable structure of extreme learning machine (ELM) to an interpretable structure for machine condition monitoring. From the aspect of signal processing, wavelet transform, square envelope and Fourier transform are incorporated into the input layer of the original ELM to extract repetitive transients, localize informative frequency bands for an enhancement of a signal-to-noise ratio, and realize square envelope spectra for exhibiting cyclo -stationarity of repetitive transients. Hence, the first to four layers of the proposed network are physically interpretable. From the aspect of physical feature extraction, popular sparsity measures are innovatively incorporated into all random nodes in the single-hidden layer of the original ELM to interpret the use of all hidden nodes in the fifth layer of the proposed network to characterize cyclo -stationarity of repetitive transients. The significance of this paper is to show that signal processing algorithms and physical feature extraction can be reformulated as the architecture of an interpretable neural network to automatically localize informative frequency bands for machine condition monitoring. This paper attempts to inspire researchers in the field of signal processing and machine learning to think about the design of more advanced interpretable neural networks for machine condition monitoring.