A deep contractive auto-encoding network for machinery fault diagnosis

In the field of mechanical fault diagnosis, methods based on traditional machine learning algorithms are commonly used. However, these methods require a lot of manual design of fault features, which needs domain knowledge that makes these methods difficult to generalize. This paper presents a deep contractive auto-encoding network (DCAEN) for machinery fault diagnosis, which is constructed by contractive auto-encoder (CAE), i.e., an unsupervised learning way for feature extraction. In addition, we add a sparsity constraint to the loss function of CAE for removing the redundant information. With the help of deep learning, this model can automatically extract features from raw data without artificially constructing features. In order to validate the superiority of the model, we conduct experiments and compare the model to existing methods with a rolling bearing dataset.