A Convolutional Attention Mechanism-based Capsule Network scheme for Gearbox fault diagnosis using Two directions signals and Noise Environment

Recently, intelligent fault diagnosis methods have received increasing research attention with achieving success. However, in the real situation, non-stationary operational conditions and environment noises may influence intelligent fault diagnosis methods accuracy due to the variations of different distributions between the training and test data domains. In this paper, a scheme of combined use of two-directional vibration signals and seeding into a convolutional attention mechanism capsule network (CA-CN) is proposed for fault diagnosis of the gearbox and different noise modes. The superior attention mechanism, which may exclude concentrate on relevant information, is incorporated in Capsule Network. It enables an enhanced model with the ability to tolerate data differences between domains further. As a result, the data's key features can be obtained through the model even though the data differences between source and target domains exist. The proposed method's accuracy has been verified by experimental data with gearbox using two directions signals compared with a convolutional neural network (CNN) deep learning methods. The results show that the introduced scheme is superior to other classification accuracy methods with higher accuracy and strong tolerance with noises.