DOI:
https://doi.org/10.63385/sriic.v1i2.348Keywords:
Artificial Intelligence, Dynamic Vehicle Simulation, Machine Learning, Multimetric Evaluation, Vehicle ElectrificationAbstract
This article proposes and illustrates a multimetric evaluation template for digital twins based on machine learning in critical engineering applications using an example as a specific testbed for discussing unified assessment principles. We analyze this concept through the lens of a particular and relevant example: the performance of a hybrid vehicle during homologation tests for transient emissions. In this scenario, ML models must not only optimize efficiency but also ensure strict compliance with environmental regulations in dynamic, operating modes. The case illustrates the complexity that arises when attempting to unify requirements for accuracy, fault tolerance, adaptability, and regulatory compliance, providing a framework for exploring how a unified evaluation system can lead to a more consistent and reliable integration of ML into critical systems. Test data of emission tests on a Hybrid vehicle was used to train a Random Forest model. Different sets of input parameters illustrate some of the capabilities and limitations of using AI. Shapley values were used to discuss some of the AI model characteristics and limitations. Using as input parameters only Vehicle speed, Acceleration, and Battery State of Charge (SoC) allowed the digital twin to achieve R2 0.80. Inclusion of Internal Combustion Engine oil temperature increased the model R2 to 0.97 and curiously changed the ranking of the other input parameters. SoC, which was the most influential in previous cases, went down to almost negligible impact on the model results. Despite reaching a quite high R2 of 0.97, the model can miss important aspects of the physical system without Human expert supervision.
References
