On-line implementation of an adaptive speed filter and its experimental demonstration

Abstract

Sensorless speed estimation in induction machines is important for numerous applications like speed control and fault detection. Sensors are expensive and they are not reliable enough to be used in rugged industrial environments. In this work, a previously developed neural network speed filter is implemented for on-line induction motor speed estimation. The speed filter is constructed using a combination of five neural networks. A neural networks framework developed in this work is used to construct the speed filter. The filter uses the three motor terminal voltages, the line currents, and the RMS of on-line current as inputs to estimate the speed. The data are preprocessed by a set of LabVIEW modules before they are sent to the neural networks. The preprocessed data are used by the neural networks to compute the induction motor speed. The output from the neural networks is then scaled to obtain the motor speed estimate. The filter is implemented and tested using both off-line and on-line collected data. The filter is also tested with unbalanced power supply and faulty motors to study its generalization capability. The filter had an average estimation error between 0.1% to 0.3% for the data collected off-line. For the data obtained from on-line setup, the average estimation at steady state is 0.15%. This research demonstrates the feasibility of using adaptive file-based software sensors instead of hardware sensors thereby significantly reducing implementation costs and improving overall system robustness. The neural networks framework developed in this work adds flexibility and scalability in further improving the developed induction motor speed filter.