소음 데이터를 이용한 딥러닝 기반의 차량 진단 기술 개발

Abstract

In this paper, we propose deep learning models for fault diagnosis and noise level estimation using vehicle noise data. First, we use two spectrograms as a feature vector by converting the input signal and the signal of the separated percussive component in the input signal. For fault diagnosis, we design a classification model. Two spectrograms are respectively fed into a series of convolutional layers that includes a convolutional block attention module(CBAM) block and max-pooling. Then, the two outputs are combined and passed through fully connected layers that is finally converted to a probability. Next, we design a regression model for noise level index estimation. We first define the noise level index using signal processing techniques and use it as a target for the deep learning model. Unlike the fault diagnosis model, the two spectrograms are combined and fed into a series of convolutional layers. Then, the output is passed through fully connected layers, and the estimated real value is rounded to the nearest integer value from 1 to 5. Experimental results showed excellent performance with an accuracy of 96 % for fault diagnosis and 86 % for noise level index estimation.