A hybrid denoising model using deep learning and sparse representation with application in bearing weak fault diagnosis

Abstract

Sparse representation is a practical approach for mining fault information from vibration signals, for which the capabilities of superior fault feature extraction and strong generalization are generally not available simultaneously. Therefore, a novel end-to-end deep network-based sparse denoising (DNSD) framework based on a model-data-collaborative linkage framework is proposed in this paper. First, a globally differentiable sparse model is established, and a deep neural network is introduced to learn the hyper-parameters, such as the sparsity per layer in sparse representation models, which is a data-driven process without prior knowledge. In this framework, a multi-modal dataset simulated fault signal was established based on the bearing fault mechanism. The dataset was used for model training as part of the framework. DNSD is trained in the form of a denoising autoencoder, and the reconstruction loss updates the parameters of the network and the sparse theory. Both the simulation and analysis of bearing race fault measurements verify that the proposed DNSD framework is superior and robust in bearing fault feature extraction.