Neural Network IPM Motor Control for Sensorless Overmodulation

The patent replaces traditional PI control methods and lookup tables with a neural network-based control system. The neural network is trained offline using approximate dynamic programming to learn optimal control strategies across the entire operating range including overmodulation and six-step regions, then deployed for real-time control without requiring complex online computations or sensor feedback, thus achieving high accuracy while maintaining implementation simplicity.

The patent performs preliminary training of the neural network offline using approximate dynamic programming across a wide range of operating conditions including linear modulation, overmodulation, and six-step regions. This preliminary action embeds adaptive knowledge into the neural network weights, allowing the system to maintain stability during operation while automatically adapting to varying conditions without requiring real-time parameter updates or retraining.

The patent implements a self-service control approach where the neural network uses easily measurable quantities (phase currents and voltages) as inputs and directly outputs optimal switching commands. The neural network internally compensates for uncertainties and eliminates the need for external position sensors by learning the relationship between electrical measurements and optimal control actions during offline training, thus achieving sensorless precise control.

The patent employs a dynamic neural network control strategy that automatically adapts to different modulation regions (linear, overmodulation, and six-step) based on real-time operating conditions. The neural network learns optimal control policies for each region during offline training using approximate dynamic programming, enabling seamless transitions and maximizing energy efficiency and voltage utilization across the entire operating range without requiring complex region detection and switching logic.