Real-time multi-domain optimization controller for multi-motor electric vehicles using automotive-suitable methods and heterogeneous embedded platforms

Abstract

In this Thesis, an elaborate control solution combining Machine Learning and Soft Computing techniques has been developed, targeting a chal lenging vehicle dynamics application aiming to optimize the torque distribution across the wheels with four independent electric motors.The technological context that has motivated this research brings together potential -and challenges- from multiple dom ains: new automotive powertrain topologies with increased degrees of freedom and controllability, which can be approached with innovative Machine Learning algorithm concepts, being implementable by exploiting the computational capacity of modern heterogeneous embedded platforms and automated toolchains. The complex relations among these three domains that enable the potential for great enhancements, do contrast with the fourth domain in this context: challenging constraints brought by industrial aspects and safe ty regulations. The innovative control architecture that has been conce ived combines Neural Networks as Virtual Sensor for unmeasurable forces , with a multi-objective optimization function driven by Fuzzy Logic , which defines priorities basing on the real -time driving situation. The fundamental principle is to enhance vehicle dynamics by implementing a Torque Vectoring controller that prevents wheel slip using the inputs provided by the Neural Network. Complementary optimization objectives are effici ency, thermal stress and smoothness. Safety -critical concerns are addressed through architectural and functional measures.Two main phases can be identified across the activities and milestones achieved in this work. In a first phase, a baseline Torque Vectoring controller was implemented on an embedded platform and -benefiting from a seamless transition using Hardware-in -the -Loop - it was integrated into a real Motor -in -Wheel vehicle for race track tests. Having validated the concept, framework, methodology and models, a second simulation-based phase proceeds to develop the more sophisticated controller, targeting a more capable vehicle, leading to the final solution of this work. Besides, this concept was further evolved to support a joint research work which lead to outstanding FPGA and GPU based embedded implementations of Neural Networks. Ultimately, the different building blocks that compose this work have shown results that have met or exceeded the expectations, both on technical and conceptual level. The highly non-linear multi-variable (and multi-objective) control problem was tackled. Neural Network estimations are accurate, performance metrics in general -and vehicle dynamics and efficiency in particular- are clearly improved, Fuzzy Logic and optimization behave as expected, and efficient embedded implementation is shown to be viable. Consequently, the proposed control concept -and the surrounding solutions and enablers- have proven their qualities in what respects to functionality, performance, implementability and industry suitability.The most relevant contributions to be highlighted are firstly each of the algorithms and functions that are implemented in the controller solutions and , ultimately, the whole control concept itself with the architectural approaches it involves. Besides multiple enablers which are exploitable for future work have been provided, as well as an illustrative insight into the intricacies of a vivid technological context, showcasing how they can be harmonized. Furthermore, multiple international activities in both academic and professional contexts -which have provided enrichment as well as acknowledgement, for this work-, have led to several publications, two high-impact journal papers and collateral work products of diverse nature.