Optimal Design of Traction Electric Motors by a New Adaptive Pareto Algorithm

This paper deals with an adaptive multi-objective optimisation process for the design of electric motors by means of supervised learning techniques. The process is based on a multi-objective optimisation approach which involves many objective functions (mass, rotor inertia, average total losses), constraints (geometrical feasibility, current and voltage limit, torque-speed profile, thermal constraints) and a large number of design variables (rotor, stator and winding parameters) that describe the physical properties of the motor. The design of computer experiments is based on a Low Discrepancy Sequence (LDS).The electric motor performance indices are evaluated through multiphysics simulations (electromagnetic, thermal) carried out by Motor-CAD(TM) software. Artificial Intelligence (AI) is adopted to approximate the physical model behaviour. Artificial Neural Networks (ANN) are exploited, in this way a large set of design variables combinations can be investigated with a reasonable computational effort. Since the multi-objective optimisation of electric motor is particularly complex, a special algorithm had to be developed to find Pareto-Optimal solutions. A new Adaptive Pareto Algorithm allows to reduce the computational cost and to achieve an even distribution of the optimal solutions in the Pareto optimal front. The algorithm is particularly effective when it is integrated with a global approximation model, the infill operation allows to improve the metamodel approximation. The derived electric motors show the best compromise performances among millions of possible configurations.