Sparse Model-Driven Deep Learning for Weak Fault Diagnosis of Rolling Bearings

This chapter focuses on the sparse model-driven deep learning for weak fault diagnosis of rolling bearings. Sparse representation is an effective approach to digging the fault features of vibration signals, but it is not liable to reliably extract the fault features while maintaining good generalization. Therefore, this chapter proposes a novel end-to-end Deep Network-based Sparse Denoising (DNSD) framework based on the model-data-collaborative linkage framework. First, a global differentiable framework is built based on sparse representation, and then a deep network is introduced to compute key parameters. Furthermore, a multi-mode data set with fault prior information is developed as one part participating in the construction of the framework. DNSD will be trained as a denoising autoencoder, and reconstruction loss will be used to update the parameters of the network and sparse theory. The simulation and experiment of some bearing race faults verify that the proposed DNSD framework has superiority and stronger robustness in bearing fault feature extraction.