Use of an artificial neural network for current derivative estimation

The Fundamental PWM technique for saliency tracking based sensorless (encoderless) motor control uses current derivative measurements to calculate the rotor position. However parasitic impedances in the drive, motor and cabling cause high frequency oscillations in the current, when the inverter's IGBTs switch. This prevents the immediate measurement of the current derivative when a new voltage is imposed on the motor and has led to an enforced minimum PWM vector time restriction that allows the oscillations in the current response to decay sufficiently before current derivative measurements are made. In this work a new method is proposed to reduce this minimum PWM vector time restriction by estimating the current derivative in the presence of such oscillations using a neural network. Training of the neural network is performed off-line with the neural network configuration (weights and biases) being stored on removable storage media. This can reduce the training burden by allowing network configurations to be saved and recalled, potentially offering a 'plug and play' solution for previously encountered drive setups. An additional benefit of the proposed solution is that the current derivative is estimated from data captured using standard industrial current sensors instead of dedicated current derivative sensors. The proposed method and its implementation are discussed and on-line experimental results are presented which validate the feasibility and performance of the proposed technique.