Hybrid Deep Learning Based on GAN for Classifying BSR Noises from Invehicle Sensors

Abstract

BSR (Buzz, squeak, and rattle) noises are essential criteria for the quality of a vehicle. It is necessary to classify them to handle them appropriately. Although many studies have been conducted to classify noise, they suffered some problems: the difficulty in extracting features, a small amount of data to train a classifier, and less robustness to background noise. This paper proposes a method called transferred encoder-decoder generative adversarial networks (tedGAN) which solves the problems. Deep auto-encoder (DAE) compresses and reconstructs the audio data for capturing the features of them. The decoder network is transferred to the generator of GAN so as to make the process of training generator more stable. Because the generator and the discriminator of GAN are trained at the same time, the capacity of extracting features is enhanced, and a knowledge space of the data is expanded with a small amount of data. The discriminator to classify whether the input is the real or fake BSR noises is transferred again to the classifier; then it is finally trained to classify the BSR noises. The classifier yields the accuracy of 95.15%, which outperforms other machine learning models. We analyze the model with t-SNE algorithm to investigate the misclassified data. The proposed model achieves the accuracy of 92.05% for the data including background noise.