An accurate and efficient machine fault diagnosis approach using a recurring broad learning model

Abstract

In most of previous machine fault diagnosis, the performance of traditional methods was over-dependent on high-quality feature extraction from original signals. Recently, deep learning–based fault recognition methods can successful automatically learn high-level hidden features from measured signals, but a deep neural network has too much hyperparameter tuning and a complicated architecture, so the training process is time-consuming. To address these issues, a novel machine fault diagnosis approach using a recurring broad learning model is reported in this article. This method uses a recurring broad learning model first proposed in this article, which learns high-level hidden features from sensor signals with different working status of a machine effectively and efficiently. Experimental results conducted on two public machine fault datasets verify the accuracy and efficiency of the proposed approach for machine fault diagnosis separately. Compared with many existing mainstream methods, the state-of-the-art framework presented reveals its superiority on both datasets. The accuracy achieves nearly 100% on both bearing dataset (Dataset 1) and gearbox-bearing dataset (Dataset 2). In addition, it can obviously realize faster training than most of the methods reported in this article.