Machine Learning Method for Road Vehicle Collected Data Analysis

Abstract

A major challenge encountered in the development of systems exposed to weather stressors, such as autonomous vehicles (AVs) and unmanned aerial vehicles (UAVs), is to ensure their proper functioning under adverse rain or snow conditions. Since the sensing of the surroundings by these vehicles relies on optical sensors such as lidars and cameras, it is essential to ensure the robustness of these systems from the early stages of the project. In this respect, experiments in climatic wind tunnels provide a solution for simulating the operating conditions in which the autonomous vehicles will be confronted. This work proposes a method based on field measurements and unsupervised machine learning to faithfully reproduce in controlled environments real weather conditions captured during wintertime in Ontario, Canada. The purpose of this paper is not to investigate correlations between observed weather conditions and the characteristics of the precipitation encountered, but rather to establish a consistent method based on outdoor disdrometer data to identify critical parameters to be simulated in climatic wind tunnels. To achieve this goal, weather data such as temperature, relative humidity, and droplet size distribution (DSD) were recorded at GM's McLaughlin Advanced Technology Track (MATT) using an FD70 disdrometer and WXT530 weather transmitter, both manufactured by Vaisala, installed on a car provided by the Automotive Center of Excellent (ACE) team of Ontario Tech University. The implementation of the proposed method allowed the identification of precipitation clusters characterized by parameters of a theoretical model for particle size distributions fitted to the collected data.