Automated Learning of In-vehicle Noise Representation with Triplet-Loss Embedded Convolutional Beamforming Network

Seok-Jun Bu, Sung-Bae Cho

Abstract: In spite of various deep learning models devised, it is still a challenging task to classify in-vehicle noise because of the reverberation and the variance in the low-frequency band generated from the narrow interior space. Considering the impulsive characteristics of the vehicle noise and the multi-channel sampling environment at the same time, it is essential to automatically learn the disentangled noise representation as well as parameterize the conventional beamforming operation. We propose a method to overcome the above two major hurdles by parameterizing a beamforming operation based on convolutional neural network. Moreover, we improve the structure of the beamforming network by explicitly learning of the distance between vehicle noises within the triplet network framework. Experiments with the dataset consisting of a total 241,958,848 time-series collected by a global motor company show that the proposed model improves the classification accuracy by 5% compared to the latest deep acoustic models. The detailed analysis shows that the proposed method can potentially compensate for the disjoint issues between the learning and validation vehicle types.