Comparison of Neural Network Types and Learning Methods for Self Commissioning of Speed Sensorless Controlled Induction Machines

Speed sensorless control of induction machines at zero speed is so far only possible using signal injection methods and exploiting non-fundamental wave effects. When applying such methods the resulting control signal shows a heavy dependence on the machine's operating point, i.e., the flux and the load level. To achieve speed sensorless control around zero speed it is thus necessary to identify and eliminate the flux/load dependence. In this paper different neural network approaches are tested with respect to their ability to perform an autonomous self commissioning of the control. The multi layer perceptron (MLP), the functional link neural network (FNL), as well as the time delayed neural network (TDL) are all trained using the backpropagation algorithm. To avoid the necessity of a speed sensor during commissioning, a modified flux observer is applied to deliver the reference values of the training data. A machine with closed rotor slots was chosen for this investigation because this type of machine is considered the most difficult for zero speed sensorless control. The results show that for this specific problem, the MLP shows the best performance followed by the FNL whereas the TDL is only applicable using an extensive amount of training data.