Abstract: In order to improve the accuracy of fault diagnosis of rolling bearing and increase the anti-noise performance of the diagnosis model, an improved stacked sparse auto-encoder (ISSAE) network combined with extreme gradient boosting (XGBoost) algorithm was proposed. ISSAE stacked multiple SAE to enhance the ability of auto-encoder network to extract deep features of data and ameliorated the loss function of SAE to improve the anti-noise performance. First, the bearing measurable signal was input into the ISSAE optimized by Adam algorithm; the network reconstructed the input signal, learned and extracted the intrinsic eigenvalue of measurable signal by itself. Then, the eigenvalues were input into the XGBoost model to train the fault diagnosis classifier model which adjusted parameters by grid search. Finally, the test set of bearing fault was input into the trained ISSAE-XGBoost network to complete the automatic identification of fault types. The effectiveness and applicability of the proposed algorithm were verified by different bearing experimental data on several experimental platforms. The results show that comparing with SSAE-XGBoost and ISSAE-SVM algorithms, the proposed method has a higher recognition rate and stronger applicability. The recognition rate can reach more than 99% in case of large sample size, and has higher recognition accuracy even when sample size is small. The loss function is improved in the network, so that the model can suppress the disturbance of small disturbance and maintain high fault diagnosis accuracy for the noisy measurement signal.

Key words:  rolling bearing; fault diagnosis; improved stacked sparse auto-encoder (ISSAE); extreme gradient boosting (XGBoost)