Acoustic wave and machine learning approaches for the estimation of distance and road condition

Abstract

All regions where temperatures drop below freezing are experiencing difficulties with black ice. The black ice is thin ice on the road surface and is not black in itself, but it becomes visible as the color of the road surface by transmitting the road color. In other words, although indistinguishable to the human eye, the road surface is very slippery and looks like a general road. Therefore, this road surface poses a danger to all ground moving objects that accelerate or decelerate according to the friction coefficient of the road surface. When acceleration or deceleration is performed without considering the friction coefficient of these road surfaces, slippage occurs in the moving object, and the moving object can move differently from the intention. Ground-moving objects accelerate or decelerate depending on the friction coefficient of the surface of the road. In this study, in order to provide predictive information of the ground-moving object, a machine learning based solution that estimates the type of road surface via reflected acoustic signals using a pair of transmitters and receivers is provided. All substances have their own acoustic impedance, the value of which varies from substance to substance and frequency to frequency. That is, the sound wave reflected from the road surface embraces the material type features of the surface, and the characteristics differ depending on the frequency of the signal. From these facts, the information in the frequency domain is confirmed by using the short-time Fourier transform on the sound wave signal reflected on the road surface. In addition, this information is used for modeling and classification of machine learning. In addition the sensors, signal processors and machine learning algorithms are selected to perform these tasks. In order to demonstrate the effectiveness of the proposed method, we conduct an experiment to identify the test pieces made of 4 materials and 12 pieces in different sizes. Here, a method using a short-time Fourier transform and a learning based method of an artificial neural network is used. Not only that, correspondence learning with multi-modal material classification using sound waves and vision information, and a method to prevent false input of artificial intelligence through attention modules have also been proposed, and its effects is demonstrated. A method for classifying nine road surfaces using short-time Fourier transform and artificial neural network learning method based on the test piece classification method is presented. For that purpose, a method of acquiring a data set in various environments and creating a road surface classification model is described. Not only that, the application method of applying the sensor to the road infrastructure is also introduced. Due to the characteristics of the road infrastructure, a method to solve the problem caused by installing the sensor at a high position is also proposed, and the experimentally proposed method is verified. Here, a method of deceiving an artificial intelligence model learned by using the above-mentioned method of intentionally generating a fraud input and the method of atmospheric attenuation of sound wave are proposed, and its effect is also experimentally verified. An embedded hardware that can obtain road surface information is also manufactured for this experiment, and a data set is acquired for that hardware as well, and is used for training and testing of artificial intelligence. Since the short-time Fourier transform takes a long conversion time, the real-time property is diminished. While resolving these limitations, a methodology for observing specific frequency components is provided. Utilizing the fact that the cross-correlation function resembles a convolution equation, the received acoustic raw signal is used as an input to a deep convolution neural network without short-time Fourier transform to classify the road surface. Therefore, the effect is experimentally verified by using the above-mentioned obtained road surface signal. Through this proposed method, it is possible to estimate the road surface condition more than 100 times per second. For this study, more than 10,000 datasets for nine different road surfaces is acquired and efforts are made to obtain signals under various conditions in order to ensure the generality of the trained model. As a result, the obtained road surface information can be used to actively prevent accidents caused by black ice, which is a social problem. In addition, the limit of the regenerative braking system in the electric vehicle can be solved by predictive control, and the limit of the existing anti-lock braking system can be solved. In other words, by predicting the type of the surface of the road to come in advance and using the road surface information for controlling the dynamics of the vehicle, it is possible to improve the performance of the vehicle and ensure safety.