Batch-normalized deep neural networks for achieving fast intelligent fault diagnosis of machines

Abstract

Numerous researches have been conducted on developing effective intelligent fault diagnosis systems. As a commonly used deep learning technique, stacked autoencoders (SAEs) have shown the ability of automatic feature extraction and classification. However, the traditional SAEs have two deficiencies: (1) The multi-layer structure and too many epoch number always require plenty of time for training. (2) The internal covariate shift problem exists in deep networks, leading to that it is hard to train the model with saturating nonlinearities. To overcome the aforementioned deficiencies, a recently developed optimization method called batch normalization is introduced into deep neural networks (DNNs). The method is employed in every layer of DNNs to obtain a steady distribution of activation values during training. Besides, it applies normalization technique on every mini-batch training. As a result, it offers an easy starting condition for training, and the training epoch number can also be reduced. Thus, fault features can be extracted rapidly in an elegant way. A bearing and a gearbox datasets are adopted to conform the effectiveness of the proposed method. The experimental results show that the proposed method can not only solve the two deficiencies of SAEs, but also achieve a superior performance to the existing methods.