Research Area:  Machine Learning

Abstract:

The vibration signal of rolling bearing contains abundant fault information. However, due to the existence of variable speed, it is easy to conceal the fault characteristics in the vibration signal, which makes it difficult to determine the corresponding fault type when using the traditional deep learning model for fault diagnosis. In order to solve the above problems, multiscale attention feature fusion network (MAFFN) is designed, which combines MDL module and SDAM module. Compared with the traditional attention mechanism module, the additional introduction of standard deviation pooling as a supplement can better help the network to adaptively learn and extract useful features in the input data. Besides, the fault features under different receiving domains are obtained by multiscale fusion, and the DenseNet is combined with LSTM to extract features from space and time, which effectively overcomes the problem of spectrum change and fault feature extraction caused by variable speed in multi-frequency range. Since the DenseNet takes all the previous layers as input, by combining with the SDAM module, the network can choose a layer that is more relevant to the feature, which enhances the effective propagation of features. Using the Spectra Quest and the Mendeley bearing data set, the ablation experiments and the comparative experiments verify that the MAFFN model has superior performance for rolling bearing fault diagnosis under variable speed conditions.

Keywords:  

Author(s) Name:  Jiayi Shen, Dongfang Zhao, Shulin Liu & Ze Cui

Publisher name:  Springer

DOI:  10.1007/s11760-024-03171-8

Volume Information:  Volume 18, Pages 523-535, (2024

Paper Link:   https://link.springer.com/article/10.1007/s11760-024-03171-8