Data Driven Fault Diagnosis Method Based on XGBoost Feature Extraction

Aiming at the problems that the current machine learning methods in fault diagnosis were not possible to fully exploit the hidden fault feature informations in the data and has insufficient approximation accuracy, an implicit feature information extraction method was proposed based on XGBoost. Firstly, the loss function of the XGBoost algorithm was customized according to the fault data and the fault types, and the fault splitting tree was constructed iteratively. Secondly, the leaf node position index vector of the sample in the fault tree was extracted and the feature code was reconstructed to obtain the intelligent representation of the implicit fault informations. Thirdly, based on the characterization matrix, a machine learning algorithm such as SVM was used to establish a fault diagnosis model to realize predictive diagnosis of multiple failure modes. Finally, the proposed method was validated by taking a driver fault diagnosis as an example. The results show that compared with the fault diagnosis model under the original features, the diagnostic model based on XGBoost extraction implicit features has higher accuracy and better robustness, which may give the order of importance of the feature variables. © 2020, China Mechanical Engineering Magazine Office. All right reserved.