Hierarchical Multiscale Dense Networks for Intelligent Fault Diagnosis of Electromechanical Systems

Deep learning, which is characterized by its powerful feature extraction capabilities, has been widely used in the field of mechanical fault diagnosis. Traditional deep learning models usually perform feature extraction at a single-scale level, which prevents them from extracting multiscale discriminative information from mechanical vibration signals. In this article, we put forward a hierarchical multiscale dense network (HMSDN) for the fault recognition of electromechanical systems. The architecture is proposed with the aim of learning the inherent and multiscale feature information that is essential for the fault identification of mechanical signals under nonstationary conditions. The major contributions can be summarized into two points. On the one hand, due to the various modes of mechanical signals, we embed a hierarchical procedure into the CNN structure so that the structure is incorporated with multiscale learning ability. On the other hand, in consideration of the fact that the signal boasts various information delivered from different transmission paths, the signal is complex and nonstationary. Hence, a multiscale dense connection structure is designed to learn discriminative features of the measured signals. The performance of the proposed method is evaluated on two electromechanical datasets. The results reveal that the proposed approach can achieve state-of-the-art performance in comparison with some competitive approaches.

Automated summary

Deep learning has been widely used in mechanical fault diagnosis due to its powerful feature extraction capabilities. However, traditional deep learning models lack the ability to extract multiscale discriminative information from mechanical vibration signals. In this article, a hierarchical multiscale dense network (HMSDN) is proposed to address this issue. The HMSDN incorporates a hierarchical procedure into the CNN structure and uses a multiscale dense connection structure to learn discriminative features from measured signals. Experimental results on two electromechanical datasets demonstrate that the proposed method achieves state-of-the-art performance.