Aiming at the problem of low accuracy of traditional fault diagnosis models facing massive fault data, a joint denoising method of local mean decomposition and fixed point algorithm was proposed to eliminate the noise in bearing vibration signals. Secondly, in order to solve the problem of difficult extraction of sensitive features from original signals, a dimension reduction method based on kernel principal component analysis is proposed. Thirdly, a fault diagnosis model based on the improved eXtreme Gradient Boosting decision tree was constructed. The GS-PSO algorithm was used to optimize the performance of SVM, and then the residual error of classification model was modified by using the improved XGBoost idea to improve the classification accuracy of the model. Spark- Big data platform was used to carry out scientific calculation by parallel processing technology. Finally, the rolling bearing data provided by CWRU is used for training and simulation, which proves that the established model can realize the identification and diagnosis of different types of rolling bearings and ensure the accuracy of diagnosis results. Through comparative analysis of four different fault diagnosis models, the results show that the model is feasible and advantageous.