Multi-Sensor Data Fusion and Augmentation for Imbalanced Fault Diagnosis of Bearings

Stable operation of bearing is required, while on the other side, it can lead to insufficient fault data collection and imbalanced data, which further deteriorates the performance of deep learning(DL) methods in the fault diagnosis of bearings. In this paper, a novel method combining multi-sensor data fusion and augmentation for bearing imbalanced fault diagnosis is proposed. First, the limited real fault samples are fed into the one-dimensional Wasserstein generative adversarial network with gradient penalty (1D-WGAN-GP) to generate the fake samples for augmenting the fault data. Then, the features of augmented data from different sensors are extracted and fused by using the proposed multi-branch one-dimensional convolutional neural network (1D-MCNN). Finally, imbalanced fault diagnosis of bearings is achieved based on the fused features. The performance of the proposed method is verified experimentally. The results show that, compared to existing methods the proposed method can be used to fuse fault features from multiple sensors, and effectively enhance the fault data, thereby achieving excellent accuracy and stability under the imbalanced data of bearing.