Abstract:
Bearing fault can seriously reduce industrial production efficiency and even endanger the safety of people's lives and property. Monitoring the operating conditions of bearings conducting fault diagnosis are of great significance for ensuring the safe operation of the production process. This paper proposes a Multi-Scale and Attentive Convolutional Neural Network (MACNN) based on attention mechanism for bearing faults classification. Firstly, one-dimensional bearing signals are input into a convolution layer, followed by a maximum pooling layer to suppress noise and reduce information redundancy. Then, four MACNN modules are taken as input, each of which adopts a parallel network structure of ordinary convolution and void convolution. Under the premise without increasing model parameters, the model's receptive field is expanded to extract more fault features for improving the accuracy. In addition, the attention mechanism module is connected at the end of each MACNN model and the feature information is further extracted by using the ability of automatic extraction of important features. The average pooling layer is used in the network structure to prevent the overfitting and the full connection layer from the output of experimental classification results. Furthermore, the Boundary Equilibrium Generative Adversarial Networks (BEGAN) model is adopted to enhance fault data, change the proportion of unbalanced data sets, increase the number of dataset samples, reduce overfitting of MACNN model, and improve diagnostic accuracy. Finally, the experimental results on the Paderborn University Dataset show that MACNN can achieve better performance in feature extraction and fault classification, outperforming the state-of-the-art methods.