Abstract

A Convolutional Neural Network is a deep learning model that is an active research topic and is being applied extensively to analyse vibration data for condition monitoring. However, existing CNN architectures for automated fault diagnosis have some limitations, such having too few layers or converting the raw vibration data into a two-dimensional form, etc. To address these limitations, this paper develops a one-dimensional CNN architecture with three feature extraction layer groups (CNN-Three) for automated fault diagnosis. The developed CNN-Three architecture uses one-dimensional raw vibration data as an input to train the developed model. A wide convolutional filter in the first feature extraction layer group is used to cover a longer length of the time series inputs and suppress noise effects. Then, multilayer narrow convolutional filters size corresponding to the second and third feature extraction layer groups are used to extract more detailed features and improve the network performance. The effectiveness of the developed CNN-Three architecture is evaluated through analysis of simulated and experimental vibration data. The results demonstrate that the CNN-Three architecture achieves higher diagnostic accuracy and outperforms three recent CNN architectures reported in the literature.