Seleção de modelos de machine learning para prever os resultados da competição Baja SAE Brasil a partir dos dados históricos e ficha técnica do veículo

Abstract

Artificial Intelligence (AI) and Machine Learning (ML) are at the forefront of current technological innovations, with applications in decision-making, product optimization, and data-driven predictions. In the context of Industry 4.0, these technologies enable detailed analysis and process automation, improving the efficiency and performance of both industrial and academic projects, such as the Baja SAE Brasil competitions. In these competitions, the challenge proposed by SAE, highlights the opportunity to enhance project performance and highlights the need to develop a predictive model based on ML to assist the Baja team of UTFPR. The focus is to predict the team’s rankings using the vehicle’s technical data, allowing for precise adjustments tailored to the project’s needs. The main objective of this work is to apply data mining methodologies to create a machine learning model capable of predicting the UTFPR Baja team’s competition results, utilizing historical data provided by SAE Brasil. The framework used was the CRISP-DM, which includes business understanding, data preparation, modeling, evaluation, and deployment phases. Data was processed and explored to identify relevant patterns and relationships, and different ML models were trained and tested to select the most suitable one for the application. Validation was performed with a test dataset, and the final model was deployed to predict the vehicle’s performance in the competition. This work resulted in the performance evaluation for 9 different machine learning models and on implementation of Gradient Boosted Trees (GBT) and Support Vector Machines (SVM) for making concise predictions about team performance in the dynamic events of the competition. The CRISP-DM methodology proved effective, and the study provides insights for future ML applications in similar contexts. Therefore, this thesis addresses the development of Machine Learning models for multi-class result prediction, considering different attributes with complex nonlinear relationships.