EXPERIMENTS OF RECURRENT FUZZY NEURAL NETWORKS – BASED SUPERVISORY CONTROL ON AN LIQUID FLOW CONTROL SYSTEM

Abstract

Recurrent fuzzy neural networks (RFNNs) has been successfully verified by many studies on simulation. However, the experimental controls on actual devices are still limited. There exist even some opinions that with a slow online training algorithm, it is difficult for RFNNs to guarantee the signal communication. This study conducts and experiments with a RFNNs – based supervisory control technique on the RT020 liquid flow control system of the Gunt-Hamburg, Germany. The RFNN controller parameter updating algorithm uses Jacobian information provided from a non-parameter model identifier, also using another RFNN. Experiments on the RT020 show that the RFNN controller has contributed to reduce the settling time, from about 12 seconds down to 8±0.5 seconds without stady-state error, and negligible overshoot. Besides, with external factor affected on the RT020 system by forcibly reducing the pump power, experiments have proven that the RFNN controller is effective in bring the system response back to the reference value quickly and stably.