Compressive Sampling and Deep Neural Network (CS‐DNN)

The compressive sampling and sparse autoencoder-based deep neural network (CS-SAE-DNN) uses CS for the sparse time-frequency representation model to produce highly compressed vibration measurements from the high-dimensional vibration data collected for the purpose of machine condition monitoring. This chapter presents an approach that has been proposed through the design of an intelligent fault-classification method from highly compressed measurements using sparse-overcomplete features and training a deep neural network through a sparse autoencoder (CS-SAE-DNN). This method includes the extraction of overcomplete sparse representations from highly compressed measurements. It involves unsupervised feature learning with a SAE algorithm for learning feature representations in multiple stages of nonlinear feature transformation based on a DNN. Case studies of bearing datasets are used to demonstrate how CS-SAE-DNN works and to validate its efficacy compared with other state-of-the-art fault-diagnosing techniques.