Direct Torque Control of Electric Motors with Neural Network Implementation

In this paper, we propose a novel approach implementing neural networks to replace conventional stator flux observers for motor direct torque control. The system architecture typically involves training a multi-layer perceptron (MLP) or recurrent neural network (RNN) using historical motor operational data, where input parameters may include stator currents, voltages, and rotor speed measurements. This method significantly enhances system performance and stability through adaptive flux estimation, while reducing sensor dependency and associated costs. The neural network implementation enables more accurate torque estimation via real-time weight adjustments and backpropagation algorithms, achieving superior control precision compared to traditional PI controllers. Key functions include flux linkage prediction using sigmoid activation functions and gradient descent optimization for minimum estimation error. This innovative approach demonstrates substantial potential for motor control applications, establishing a foundation for future research in AI-driven industrial automation systems.