Memory-Efficient AI Model for Virtual Voltage Vectors on Low-Cost Controllers in Asymmetrical Six-Phase Induction Machine Drives

Introduction: Asymmetrical six-phase induction machines (A6PIM) inherently excite the secondary x–y subspace, producing circulating currents that do not contribute to torque, degrade current quality, and increase losses. In recent years, virtual voltage vector (VVV) modulation has been adopted to generate variable duty-cycle switching patterns; however, under fixed switching frequency operation, it typically relies on high-dimensional lookup tables (LUTs) to map the reference voltage (Vref, θref) into duty cycles. Although effective, this approach becomes impractical for low-cost processors due to excessive memory. This work addresses that bottleneck by replacing the LUT with an AI model that reproduces the original LUT duty cycles, thereby preserving the model behavior.

Methods: A sector-based VVV formulation is adopted in the α–β plane (24 sectors, 15° each). For each sector, a fixed set of active vectors is used with a symmetric pulse mid-sequence, and duty cycles are computed to match the α–β reference while enforcing zero average x–y voltage (and thus eliminating x–y current injection). An offline dataset is generated over the operating framework in (Vref, θref), and a compact regression model is trained to approximate the duty-cycle map (Vref, θref)↦{dk}. The resulting model is exported in a lightweight format suitable for DSP implementation.

Results: The AI model reproduces the LUT duty-cycle outputs with a very small error, while reducing memory requirements from approximately 175 MB to about 39 kB. Since the duty cycles (and corresponding switching commands) are replicated, the control VVV switching pattern and switching frequency are preserved. Real-time feasibility is compatible with low-cost control hardware such as the TMS320F28379D.

Conclusions: The main contribution is a memory-efficient AI model that replaces the VVV duty-cycle LUT while reproducing the original duty cycles, preserving the control modulation structure and switching frequency, and enabling real-time implementation on low-cost controllers for A6PIM.