基于RFE-BXGBoost的轴承套圈沟道表面缺陷识别方法

Identification of bearing ring groove defects based on RFE-BXGBoost algorithm

轴承套圈是轴承部件的重要组成部分,其表面缺陷影响轴承的服役期限。为了解决轴承沟道表面缺陷难以被准确识别的问题,提出了一种基于特征递归消除的贝叶斯极度梯度提升树(RFE-BXGBoost)的轴承套圈沟道表面缺陷识别模型(方法)。首先,基于特征衍生的思想,对轴承沟道的时域、频域等特征进行了提取,使用了极度梯度提升树(XGBoost)作为基于特征递归消除(RFE)的基学习器,对影响轴承沟道表面缺陷最佳特征子集进行了选择,并过滤了冗余特征;然后,利用基于贝叶斯优化的XGBoost模型组成弱分类器,为了降低模型预测结果的方差,使用有放回随机抽样法,对基分类器进行了选取;最后,根据抽样结果,利用投票法获得了最终的表面缺陷识别结果,并使用轴承套圈沟道实测数据集进行了模型预测性能的测试。实验结果表明:基于RFE-BXGBoost的表面缺陷识别模型的识别准确率为0.90,F1-score为0.879,优于仅使用自适应提升法(Adaboost)、随机森林、梯度提升树的表面缺陷识别结果。研究结果表明:该表面缺陷识别模型对复杂零部件和系统的表面缺陷识别有一定的效果。

The bearing ring groove is an important part of bearings,and its service life is affected by surface defects.The bearing groove surface defects were difficult to identify,therefore an identification model named recursive feature elimination-Bayesian extreme gradient boosting tree(RFE-BXGBoost)was proposed for bearing ring groove surface defects.Firstly,feature derivatization was used to extract the features of the time,frequency,etc.The extreme gradient boosting tree(XGBoost)was used as a base learner for recursive feature elimination(RFE)to achieve feature selection and eliminate the redundant features from the original dataset.Then,according to the results of feature selection in the original dataset,the hyperparameters of XGBoost were used as variables to be optimized by using Bayesian optimization algorithm.Moreover,in order to decrease the variance of the model,random sampling with replacement was used for the XGBoost that had been optimized by the Bayesian optimization algorithm.Finally,the identified results were obtained by using voting method based on the predicted results of XGBoost under random sampling with replacement method.The RFE-BXGBoost was applied to the test dataset of bearing ring grooves.The experiment result shows that the surface defect recognition model based on RFE-BXGBoost has high identification performance,its accuracy is 0.90,and its F1-score is 0.879.The RFE-BXGBoost has higher effectiveness compared with other popular algorithms such as adaptive boosting(Adaboost),random forest,gradient boosting decision tree.The result shows that the RFE-BXGBoost has certain reference value for identifying the surface defects of bearing ring groove.