Fault diagnosis based on feature enhancement multiscale network under nonstationary conditions

Abstract

Convolution neural network (CNN) is widely used in rotating machinery fault diagnosis. However, in real industries, the rotating machinery often operates under changing speed and heavy background noise conditions. As a result, the fault-related information from collected signals is submerged by interference pulse, and most existing CNN-based diagnosis methods can hardly extract enough discriminative features. To tackle the above issues, this paper proposes a feature enhancement multiscale network (FEMN) for health state prediction. First, the convolution local attention mechanism is introduced to adaptively extract discriminative features. Next, to fully utilize features from intermediate layers, the ADD module is leveraged to intelligently integrate the feature information from each two CLAMs. Besides, the multiscale feature enhancement module is used to filter the noise interference and extract multiscale features, and the boundary feature enhancement module is applied to focalize the distribution of fault-related features. Finally, the FEMM is constructed based on the above contributions. Experimental results on the motor and bearing dataset under nonstationary conditions demonstrate the FEMN outperforms five state-of-the-art methods.