Combination of signal injection and neural networks for sensorless control of inverter fed induction machines

Abstract

For mechanical sensorless control of inverter-fed induction machines, a satisfactory performance at low speed down to zero fundamental frequency can so far only be achieved by evaluating inherent saliencies of the induction machine. Similar to other sensorless methods based on signal injection, the resulting control signals of the indirect flux detection method by on-line reactance measurement is influenced by every saliency, for example, the saturation based, the slotting, and the anisotropy saliency as well as by load and flux level. Since these influences are extremely dependent on the machine design, they can hardly be calculated in advance and removed by filtering or digital signal processing. However, the possibility of utilizing a neural network for learning the individual dependencies and removing the unwanted influences can provide a very satisfactory result. Since the easy implementation of a neural network does only use a small amount of calculation power, the algorithms can be implemented even in low-cost signal processors. Measurements on mechanical sensorless controlled induction machines present adequate results up to about rated load, depending on the transient electrical behaviour, and with this on the design parameters of the induction machine.