An Weighted CNN Ensemble Model with Small Amount of Data for Bearing Fault Diagnosis

Abstract

Rolling bearings are undoubtedly one of the key components in rotating machines. Bearing failure can adversely affect not only on mechanical failures, but also on operation schedules, and production processes. Therefore, the fault diagnosis of the bearing is a very important field. Recently, many machine learning and deep learning researches have been conducted and deep learning approaches usually have shown better results. Although, deep learning approaches have shown good results, most of open source bearing data was created in a slightly different environment from the actual factory environment by using clear data generated by the simulator so that it usually shows great results. Also, deep learning approaches need huge amount of data to train the model. Therefore, in this paper, we provided a method to address these issues. Firstly, we added Gaussian noise to CWRU (Case Western Reserve University) dataset to set it closer to the actual factory condition. Secondly, we adopted a model that obtains higher stability and accuracy than a normal CNN (Convolutional Neural Network) by constructing the weighted arithmetic mean CNN ensemble model. As a result of the accuracy and F-1 score analysis, the proposed model showed better result than the simple CNN and CNN ensemble averaging model.