Data Augmentation Using Empirical Mode Decomposition on Neural Networks to Classify Impact Noise in Vehicle
Gue-Hwan Nam, Seok-Jun Bu, Namu Park, Jae-Yong Seo, Hyeon-Cheol Jo, Won Tae Jeong
Abstract
In a vehicle, impact noise may occur during steering action due to clearance between parts of steering systems. Via structural path the noise is perceived by the drivers' ears and it can be the cause of a repair campaign. It is importatnt to know where the collision occurs to modify the parts causing impact noise. In this paper, we performed data augmentation using Empirical Mode Decomposition (EMD) method that decomposes the original signal into a finite number of intrinsic mode functions (IMFs). The IMFs were decomposed by descending order from high frequency to low frequency, and we add the residue each time one IMF is separated. After the data augmentation, the data were trained using the neural network model CNN-LSTM. The proposed method showed better classification performance than other classification methods. It seems that proposed method takes advantage of the impact noise characteristics concentrated at low frequency range.