Recurrent Neural Network based Soft Sensor for flow estimation in Liquid Rocket Engine Injector calibration

Abstract

Injector is the critical element in the Liquid Rocket Engine (LRE), to ensure proper mixing of propellants (fuel and oxidizer) in the thrust chamber for achieving the optimum thrust. LRE injector is calibrated in order to deliver required flow rates of propellants by sizing the orifices through simple injector water calibration (IWC) techniques. In LRE-IWC process, a huge 6” turbine flow meter (TFM) is employed for the flow-rate measurement. In order to achieve and maintain the required accuracy and precision in the LRE-IWC process, periodical calibration of TFM is mandatory. It involves tremendous time, cost and human effort. Soft sensors can provide an economical and effective alternative solution for TFM flow-rate measurement. The objective of the proposed work is to develop and implement a recurrent neural network based soft sensor (RNN-SS) for TFM flow-rate measurement. In the LRE-IWC process, experimental flow trials were carried out for different flow patterns, and the necessary measurement data were generated for the soft sensor design. The designed RNN-SS was trained by tuning various hyper parameters to replace the TFM, using three related measurement parameters acquired during the experimental trials. The precise TFM flow-rate estimation was achieved by the designed RNN-SS, with a worst-case mean absolute percentage error (MAPE) of 1.91% for the experimental flow patterns considered, with good repeat-ability. The proposed RNN-SS model for TFM flow-rate estimation gives a MAPE of 0.58%, for the required flow-pattern which is well suited for practical use.