Improved Hierarchical Adaptive Deep Belief Network for Bearing Fault Diagnosis

Abstract

Rotating machinery plays a vital role in modern mechanical systems. Effective state monitoring of a rotary machine is important to guarantee its safe operation and prevent accidents. Traditional bearing fault diagnosis techniques rely on manual feature extraction, which in turn relies on complex signal processing and rich professional experience. The collected bearing signals are invariably complicated and unstable. Deep learning can voluntarily learn representative features without a large amount of prior knowledge, thus becoming a significant breakthrough in mechanical fault diagnosis. A new method for bearing fault diagnosis, called improved hierarchical adaptive deep belief network (DBN), which is optimized by Nesterov momentum (NM), is presented in this research. The frequency spectrum is used as inputs for feature learning. Then, a learning rate adjustment strategy is applied to adaptively select the descending step length during gradient updating, combined with NM. The developed method is validated by bearing vibration signals. In comparison to support vector machine and the conventional DBN, the raised approach exhibits a more satisfactory performance in bearing fault type and degree diagnosis. It can steadily and effectively improve convergence during model training and enhance the generalizability of DBN.