Audio deformation based data augmentation for convolution neural network in vibration analysis

Abstract

Audio deformations in audio processing have proved ability in preserve semantic meaning for audio signal. Convolution Neural Network (CNN) is among deep learning model that requires huge dataset during training for excellence performance Thus, data augmentation (DA) method is used to overcome the problem of limited dataset number for vibration analysis. Several signal processing phases including segmentation and image converting need to be performed before the vibration signal can be used as input for CNN. In this research, audio-deformation based DA is proposed in generating the additional vibration signal dataset. The proses is start by encoding the raw vibration signal to audio signal format to enable the audio deformation process performing, then decoding back into new vibration signal. Speed and amplify transformation are selected for audio deformation process. The new vibration data set of bearing fault detection problem are used for training CNN to validate the proposed approach. The results obtained from 13 experiments setting have shown that the proposed DA able to increase the accuracy of training for CNN until 13% compared with the previous DA method.