Abstract
This article presents a rolling bearing fault severity identification methodology that aims to adaptively extract fault severity features and intelligently identify the fault severity. The presented method is developed based on a deep sparse auto-encoder network trained with noise added sample expansion. A sparse auto-encoder is a learning algorithm that is capable of performing unsupervised learning of the inner structure and characters of source data. However, even though the original shallow structure of the sparse auto-encoders is capable of extracting the features, they do not have the capability for fault and severity classification. By stacking multiple sparse auto-encoders with a classifier layer, a deep sparse auto-encoder network with the ability of fault severity feature extraction and intelligent severity identification can be obtained. After unsupervised layer-wise self-learning and supervised fine-tuning, the designed deep sparse auto-encoder network can perform severity identification with automatically extracted fault features. Also, to overcome the issue of overfitting caused by limited number of training samples and stacked structure with numerous neurons and layers in a deep sparse auto-encoder network, Gaussian noises are added into the training samples to train the deep sparse auto-encoder network. Using deep sparse auto-encoder network for rolling bearing fault severity identification, the overfitting phenomenon can be retrained to increase the robustness of the network. The performance of the presented method is validated using bearing fault datasets that contain different levels of bearing severity and bearing life stages and compared with other methods. In comparison with recent solutions developed using deep learning approaches, the advantages of the presented method are that it does not require any pre-processing of the vibration data and provides a more robust diagnostic performance.
Keywords
Get full access to this article
View all access options for this article.