Data-Driven Approaches for the design of Traction Electrical Motors

In the last decades, the automotive industry has fostered the integration of electrical machines into vehicle's powertrain. The legislation regarding the environmental pollution is becoming more and more stringent, till to the point in which the classical internal combustion engines will not be compliant anymore. Consequently, Battery Electric Vehicles (BEVs), Hybrid Electric Vehicles (HEVs), and Plug-in Hybrid Electric Vehicles (PHEVs) are designed depending on the level of integration of the electrical machine in the powertrain. This trend has increased the necessity to design effective electrical machines that are not only efficient but also cost-effective and powerful. Their design is therefore challenging and often requires extensive multi-physics considerations. For example, high Torque requires higher current, but this also leads to more complex cooling system. At the same time, electrical machines should be compact enough to fit inside the vehicles, ensuring excellent structural safety. The conflicting nature of these objectives adds complexity to the design process. Due to this challenging environment, a multi-objective design approach is essential to achieve the best trade-off for specific applications. The main tool used is the Finite Element Analysis (FEA). By means of this, it is possible to evaluate electromagnetic, structural, and thermal performance of the design. However, even if this tool represents the state of art for multi-physics analysis, it is very computationally intensive. Therefore, it is limiting the exploring capabilities of the design space. In this context, Machine Learning can provide a large help by training regression models able to speed up the design process, allowing to use FEA only on the best promising configurations. This thesis investigates the possibility of applying SVR and Neural Network to effectively predict torque and torque ripple, starting from an initial IPM (Internal Permanent Magnets) geometry defined by 8 input features. To train the models a dataset composed by 5000 samples have been created using FEA, in order to have the labels to link the geometry to the final prediction. The models are trained on a subset of the whole dataset to learn the statistics of the training set. Then this knowledge it is assessed on the test set, where the predictions can be compared with the real labels. The final outcome is that both SVR and Neural Network can be effective tools in predicting the aforementioned outputs, providing, in most of the cases, a limited error range. Neural Network showed slightly superior performances with the main drawback of a longer training time compared to SVR. Lastly, these models can be implemented in an optimization strategy, like differential evolution, to identify the non-dominated solutions belonging to the Pareto front. Given the slightly superior performances in terms of time, the SVR model has been implemented in the aforementioned differential evolution strategy. The current preliminary implementation needs more accurate validation but has already shown promising results in terms of speed of convergence, providing a good exploration of the design space.