Adversarial Signal Augmentation for CNN-LSTM to Classify Impact Noise in Automobiles

Seok-Jun Bu, Hyung-Jun Moon, Sung-Bae Cho

Abstract: The classification of impact noise on vehicle steering gear mainly addresses the issue of modeling the transient and impulsive signals. In particular, variations between the steering systems arising from the differences in manufacturing processes according to the vehicle types extremely limit the conventional deep acoustic models. Focusing on the fact that the major hurdles addressed can be mitigated by generating and modeling the virtual impact noise, we propose an adversarial signal augmentation method for the vehicle noise modeling. The impact noise is represented based on the Fourier transform and the variance between vehicle types is alleviated using a generative adversarial network with an auxiliary classifier in order to improve the generalization performance of the model. Experiments with the dataset of 134,400,000 time-series collected from a global motor corporation show that the proposed method has more than 3% of accuracy improvement against the conventional approaches.