Frequency-Domain Data Augmentation of Vibration Data for Fault Diagnosis using Deep Neural Networks

Abstract

This paper proposes a data augmentation method for vibration data-based fault diagnosis using deep neural networks. The proposed method is devised to deal with the practical problem in applying trained models to facilities, where frequency-domain features of data vary according to the change in the working environment of the facilities. In the proposed method, training data are augmented by scaling the frequency-domain features of raw training data by small amounts generated by a normal distribution. The proposed method is implemented to preserve the symmetricity of the positive and negative frequency-domain components and return the real part of the complex inverse transformed data as final augmented data. The advantage of the proposed method is verified by simulation, where the operating conditions of training and test data differ. Moreover, it is shown that the proposed method can improve the accuracy of models better compared to a time-domain data augmentation using similar random scaling.