Corrosion severity mapping: a machine learning approach

This thesis, realized during my Internship in Toyota Motor Europe from April to October 2023, consists in the developement of a continent-wide corrosion map through the usage of large-scale available weather data. As first step, the corrosion literature is revised, identifying the main causes and drivers of corrosion, then, a brief investigation on Machine Learning (ML) applied to corrosion processes is realized. The challenges and the corrosion-prevention practices of the automotive sector related to corrosion are explained. The existing corrosion standards commonly used in the industry are described and compared. After the review of the corrosion topic and the related ML applications, a data analysis step is applied to data gathered on test vehicles by Toyota Motor Europe during the years 2021/2022/2023. Two corrosion sensors, the Luna Acuity LS and the Aircorr O, are compared and benchmarked. A set of Machine Learning models is optimized to regress the corrosion current values (and the thickness loss rate) from the dataset, taking weather data as input. While the regressor performance shows to be acceptable, the conversion of its predictions to corresponding ISO-9223:2012 corrosion category results in a low performance of the model in predicting the true categories. A binary classifier is trained, validated and tested to distinguish between high corrosion and low corrosion scenarios, obtaining a good overall performance. Finally, a visualization strategy is developed as a dynamic map allowing for finer insights on the locations of interest.