Improved capsule network method for rolling bearing fault diagnosis

Abstract:

Rolling bearing noise environment is complex, low signal-to-noise ratio and the vibration signal is nonstationary and nonlinear, the accuracy of traditional diagnostic methods is low in noisy environment, thus an improved capsule network feature extraction structure and the calculation method of back propagation loss value is proposed. The Inception structure of multi-scale convolutional kernel and spatial attention mechanism are applied to extract features instead of the single convolutional layer of capsule network, so as to obtain prominent feature data in key areas under different scales. Vector neurons are constructed with capsule structure, and digital capsules of classification structure are obtained through dynamic routing, so as to realize fault diagnosis. The loss calculation of training process adopts the method of combining interval loss and reconstruction loss, and a more reasonable calculation process of back propagation is constructed by adjusting the proportional coefficient of the two. To verify the actual diagnostic effect of the model, the experiment was carried out by adding gaussian white noise of different amplitude energy and adjusting the signal-to-noise ratio by using the experimental data of four rotating speeds and corresponding four load conditions in the bearing data set of Caesar western reserve university. The double convolutional layer capsule network was compared with the traditional convolutional neural network. The results show that the method can get good diagnostic results in noisy environment and has obvious advantages over other diagnostic methods in noise resistance.