In the automotive industry, the growing demand for energy efficiency and reduced CO2 emissions makes flex-fuel vehicles a promising alternative, despite the challenges in optimizing their efficiency and minimizing emissions. This study proposed a methodology based on machine learning to estimate in-wheel efficiency by CO2 emissions, using algorithms such as Decision Tree, Random Forest and Multilayer Perceptron in a vehicle diagnostics system for TinyML. Decision Tree stood out for its shortest inference time (4 us), lowest energy consumption (248.02 mW) and average absolute error of 0.30, while Random Forest had the shortest compilation time of 46 s and the lowest RAM usage of 23496 bytes. MLP Float32, on the other hand, had the highest accuracy, with an MAE of 0.27. These results indicate that, although there are trade-offs between inference time, energy consumption and accuracy, the Decision Tree and Random Forest models show particular promise for embedded systems where energy efficiency and resource use are crucial.

Translated with DeepL.com (free version)