Ensemble adaptive convolutional neural networks with parameter transfer for rotating machinery fault diagnosis

Abstract

There exist many rotating machinery parts, and many types of failure modes, including single failure modes and compound failure modes. This brings high requirements on the performance and generalization ability of fault diagnosis methods. Compared with single fixed model, ensemble model can gather the strengths of others to achieve more accurate identification performance and stronger generalization ability. Based on this, a novel method called ensemble adaptive batch-normalized convolutional neural networks is proposed for rotating machinery fault diagnosis. Firstly, batch normalization and exponentially decaying learning rate are applied to basic convolutional neural network to address internal covariate shift problem, and achieve better diagnostic results and faster convergence speed. Secondly, a series of adaptive batch-normalized convolutional neural networks with different properties are designed. Thirdly, K-fold cross validation is utilized to train all models and parameter transfer is adopted to save computing time. Finally, a new combination strategy is proposed to efficiently ensemble the diagnosis results of all models. The proposed method is demonstrated by practical locomotive bearing dataset and extensive experiments.