Improved deep domain adversarial neural network with joint maximum mean discrepancy for bearing multi-condition fault diagnosis

The non-Gaussian and other complex forms of data distribution for the speed and load of different bearings are due to various nonlinear factors. When processing data, marginal distribution only focuses on the distribution differences on a single layer and cannot fully capture the joint distribution differences between multiple layers. To solve this problem, this paper proposes a bearing multi-condition fault diagnosis method based on joint adversarial adaptation network. First, joint distribution adaptation is designed to transfer data, establish connections between distribution constraints between layers, form a unified joint distribution, and calculate the joint distribution difference between each layer of the network. Then, the adversarial training method is adopted in the forward propagation process, the target domain data enters the domain classifier through the feature extractor, where it is confused with the source domain. Finally, a dropout layer is added to the feature extractor part to modify the neurons of the hidden layer. The weight update no longer depends on the joint action of fixed relationship nodes, which prevents the fault feature from working only under specific nodes. To have a better generalization performance of the model, different bearing speeds and loads are used for experiments. The experimental results show that the proposed method improves the accuracy by 14.27% and 8.07% respectively compared with other methods.