The paper identifies suitable learning algorithm for online flux estimation for sensorless vector controlled Induction Motor (IM) drives. The sensorless vector controlled IM drive requires accurate estimation of flux and speed. The speed estimation depends on the motor flux, which has to be measured or estimated. The flux measurement is difficult and expensive and hence generally estimated. Conventional voltage model equations for flux estimation are computationally rigorous and encounter major drawbacks at low frequencies/speed. Neural Network (NN)-based estimator provides an alternate solution for online flux estimation. The online flux estimator requires the NN model to be accurate, simple and structurally compact to ensure faster execution time in real implementation for effective control. This in turn, to a large extent, depends on the type of neural learning algorithms used to train the neural architecture. The Single Neuron Cascaded Neural architecture is trained offline using three types of learning algorithms to efficiently model the online flux estimator. Their performance is compared in terms of accuracy, epochs needed for training and structural compactness. The suitable learning algorithm for offline training of online NN-based flux estimation in sensorlessvector controlled IM drives is identified and the promising results obtained are presented.

Advances in digital technology have made the vector control realizable by industries for high performance variable speed control applications. Various vector controlled techniques for Induction Motor (IM) drives were proposed in the literature. In particular, sensorless vector control is an emerging area. The speed sensor, which is expensive and fragile, requires extra attention from failures under hostile environment and needs special enclosures and cabling is not needed for sensorless closed loop control of IM drives. This leads to cheaper and more reliable control.

The performance of sensorless vector controlled IM drive depends to a large extent on the knowledge of motor speed and flux. Measurement of flux using sensor is difficult and expensive. Hence flux is estimated using either current model or voltage model equations of IM drive. In sensorless control of IM drives, current model is not used because it demands the drive speed as one of its input, and hence only voltage model is preferred and used for flux estimation. The voltage model-based estimator also suffers from the problems of pure integrator and variation of stator resistance especially at low speed/frequencies (Holtz and Quan, 2003; and Bose, 2005). Several techniques were proposed in the literature to overcome the problems of pure integrator (Bose

and Patel, 1997; Hu and Wu, 1998; and Kevin et al., 1998). Stator resistance varies with temperature during motor operation and more predominant at low frequencies/speed. Numerous methods for online Rs estimation were proposed in the literature (Karanayil et al., 2007; and Jaalam et al., 2011). But the additional Rs estimator would increases the complexity of the drive system.

Neural Network (NN)-based estimators provide an alternate solution for flux estimation. It dispenses the direct use of complex mathematical model of the machine and hence overcomes the problems of integrator. The nonlinear dynamic system mapping capability of neural network was well proven in the literature (Narendra and Parthasarathy, 1990). They can be trained to be adaptive for parameter variations. Several NN methods were reported for flux estimation. Programmable-cascaded low pass filter was realized as a recurrent NN whose weights were obtained through a polynomial-NN (Da Silva et al., 1999). Single Layer Feed-Forward Neural Network (SLFF-NN) trained using input/output data was proposed for rotor flux estimation (Shady et al., 2009). It is shown to improve the performance of the drive at very low and near zero speed, provide immunity to motor parameter variations, remove low-pass filter/ integrator and reduce the error. The multilayer feed-forward neural architecture with two hidden layers was proposed for stator flux estimation (Himavathi et al., 2007). The heuristic design methodology for multilayer feed-forward NN-based flux estimator was proposed (Venkadesan et al., 2010). A compact NN model with desired accuracy assumes importance in real implementation of online flux estimator to ensure faster estimation for effective control. Single Neuron Cascaded (SNC) NN model was identified and showed to provide distinctly compact NN model for online flux estimation (Muthuramalingam et al., 2010).

Electrical and Electronics Engineering Journal, Learning Algorithms, Single neuron cascaded neural architecture, Artificial
neural network, Rotor flux estimators, Sensorless operation, Vector-controlled IM drives.