Abstract:  In the process of bearing fault diagnosis, there are some problems, such as lack of sufficient fault samples and different signal distribution under off-design conditions. Although the application of intelligent fault diagnosis method based on machine learning and deep learning method has made many achievements, this method still faces some challenges in the application process, which hinders the application of intelligent fault diagnosis method in actual industrial scenes. Therefore, an bearing intelligent fault diagnosis method based on improved joint distribution adaptation(BIFD-IJDA)was proposed. Firstly, the vibration signal was decomposed and reconstructed by wavelet packet transform, and statistical parameters of construction signals were calculated to build original features set. Then, a transferable feature selection method based on feature importance and Kullback-Leibler (KL) divergence was designed to quantitatively evaluate each statistical feature. Next, the improved joint distribution adaptation method was used to adapt the distribution of the feature sets in the source and target domains to reduce the distribution differences between domains. Finally, the fault diagnosis model trained by the source domain feature samples was employed to predict the fault category of the target domain samples, and the fault diagnosis experiments under different working conditions were carried out using the bearing fault data from Case Western Reserve University test rig and mechanical fault simulation(MFS) test rig. The experimental results show that the maximum fault diagnosis accuracies of the fault diagnosis method under the two bearing fault data are respectively 100% and 96.29%, which is significantly better than other comparison models. The results show that it has the potential to be applied in actual industrial scenarios.
Key words:  bearing intelligent fault diagnosis variable working condition; insufficient number of fault samples; improved joint distribution adaptation; transferable feature; neighborhood preserving embedding(NPE); transfer component analysis(TCA)