Unsupervised Machine Learning Method for Classification of Weather Precipitation for Autonomous Vehicles applications

With recent advances in the development of autonomous vehicles (AVs), more attention has been paid to the performance of intelligent navigation systems. AVs are able to recognize its surroundings using a set of sensors classified as advanced driving assistance systems (ADAS). Among these devices are radars, Lidars, cameras and ultrasonic sensors, which present functioning problems when exposed to severe climatic stressors such as rain, snow and fog. The present work is part of a Canada Research Chair Program in Adaptive Aerodynamics specifically the Weather on Wheels (WoW) project in collaboration with ACE at Ontario Tech University with the objective of mitigating negative effects of adverse weather conditions on the operation of AVs. As part of the efforts made so far, a method based on unsupervised machine learning has been developed to identify key features of real precipitation events for realistic weather simulations in both CFD and climatic wind tunnels. For this purpose, the project relies on a test vehicle to collect weather data at GM's McLaughlin Advanced Technology Track (MATT) in Oshawa, Ontario. Particle size distributions (PSD) are recorded and fitted using a theoretical model proposed by the authors. The aim is to create a data matrix containing the parameters of the model for each recorded sample. Then, Principal Component Analysis (PCA) is used to reduce the dimensions of the generated matrix and a K-means model is applied to cluster the data points based on their overall proximity to each other. Statistical analysis of data contained in the clusters provides information on recurrent types of precipitation in the region where the tests were conducted. Also, by reversing the process, it is possible to generate synthetic PSD that summarize the characteristics of each identified cluster. More recently, by investigating temperature and relative humidity values recorded during the tests, correlations between the shape of the particle size distribution and the type of snow encountered are being mapped out. Given that whether the snow is dry or wet affects properties such as density and sticking efficiency, developing a method that enables this type of classification in real time represents an important advance in ice load prediction models for vehicles and structures. With future experimental campaigns being prepared, a snow accretion model for moving vehicles is envisioned to ensure passenger safety in a future where AVs will be a major presence on the roads.