Fiche individuelle

This paper proposes a neural current controller for decentralized non-sinusoidal permanent magnet multiphase machines. This architecture has many advantages, such as increased reliability and simplified motor design, but requires new algorithms for decentralized control. The neural network current controller features a so-called hybrid operation, with one part trained offline and a second part working online. Using the machine’s model, the network is first trained offline based on multi-agent reinforcement learning (RL) and is then implemented in the real machine. In a second step, an online calibration using Least Mean Square (LMS) to take account of the uncertainties of the model is applied only to the output layer to reduce computational burden. This solution combines the speed of offline (forward) and the adaptability of online (feedback) training. This proposal has been validated in simulation and tested on a highly non-sinusoidal 7-phase test bench to demonstrate the feasibility of the proposed approach.