An improved CNN based on attention mechanism with multi-domain feature fusion for bearing fault diagnosis

Abstract

Fault diagnosis of rotating machinery is essential in the modern industry. Although fault diagnosis methods based on deep learning have achieved high accuracy, most of them only extract features from a single domain. Methods based on a single domain are difficult to apply to environments with noise. This paper presented a diagnosis method based on the attention mechanism and the fusion of time domain and frequency domain features to improve diagnosis accuracy. The presented method contains three modules. Firstly, two shallow convolutional neural networks are employed to extract the time domain and frequency domain features from the vibration signal. Then, the attention mechanism is adopted to extract important features and perform preliminary feature fusion. Finally, a deep convolutional network is used to fuse feature further and extract high-level features. The presented method can effectively fuse multi-domain features and improve diagnosis accuracy. This paper validates the effectiveness of the proposed method through a fault diagnosis experiment. A comparative experiment illustrates that the presented method has obvious advantages in noise resistance. When the signal to noise ratio equals 0dB, the diagnosis accuracy of the presented method is up to 6.4% higher than that of the single domain method.