Comparative studies of deep learning neural network architectures in fault diagnosis of rubber vibration isolators

Adnan, Zainalfirdaus; Chang, Vui Chee; Ho, Jee-Hou; Chai, Ai Bao; How, Ho Cheng; Chai, Tong Yuen; Kamaruddin, Shamsul

doi:10.22059/ceij.2025.379435.2103

Comparative studies of deep learning neural network architectures in fault diagnosis of rubber vibration isolators

Articles in Press

Authors

¹ Department of Mechanical, Materials and Manufacturing Engineering, Faculty of Science and Engineering, University of Nottingham Malaysia, Selangor, Malaysia.

² Department of Mechanical, Materials and Manufacturing Engineering, Faculty of Science and Engineering, University of Nottingham Malaysia, Semenyih, Malaysia.

³ Department of Computer Science, Faculty of Information and Communication Technology, Universiti Tunku Abdul Rahman, Malaysia

⁴ Technology and Engineering Division, Malaysian Rubber Board, Kuala Lumpur, Malaysia.

10.22059/ceij.2025.379435.2103

Abstract

Automating fault diagnosis of machine components is crucial as it prevents unexpected downtime of a system that affects the operation and safety of the users. Deep learning architectures such as convolutional neural network (CNN) and long short-term memory network (LSTM) have been proven as prominent in training of sequential data due to their robustness in classifying time series sequences and achieving state-of-the-art performance for effective fault diagnosis in structural health monitoring (SHM) systems. In this study, hybrid CNN-LSTM and U-Net (a CNN-based model arranged in U-shaped architecture), are employed to detect different levels of cracks in rubber vibration isolators. Cracks were induced at the interface between the steel and rubber to simulate a faulty scenario similar to a mechanical failure in industrial practice. The vibration of experimental platform supported by rubber isolators was induced by a motor driving an eccentric disk with varying speeds. Results revealed that the proposed U-Net architecture could achieve the best overall accuracy with decent computational time for training and classification. In addition, influence of data segmentation on classification accuracy, often overlooked in the literature, was also investigated in this work. Findings showed that cleaner raw signals could be less prone to classification accuracy fluctuations.

Keywords