The XCNN flow meter - a combined cross-correlation and neural network model

In this report we propose the XCNN flow meter model, which consists of an integration of a cross-correlator (XC) of pressure measurements and an ensemble of neural network (NN) estimators. Since pressure information does not only travel with the fluid, like for example particles, bubbles, eddies and, to a big extent, temperature, but also through the fluid, the transit time of a pressure disturbance estimated by cross-correlation needs to be corrected to take into account the propagation velocity of pressure differentials in the fluid. This correction is performed by the neural network models, which in this case are simple single input single output three layer feed-forward neural networks. Instead of a single neural network an ensemble is used to reduce the variance of the estimate. The proposed method involves several stages where pressure transmitter data is first filtered, then fed to the cross-correlator whose result is interpolated and filtered again before being fed to the ensemble of neural networks, which produce the final flow estimate. An average accuracy of 0.29% (with 0.18 standard deviation) of a reference ultrasonic meter has been obtained on experimental measurements performed at Tecnatom s.a. This report marks the conclusion of the Virtual Sensors for Feedwater Flow Measurement project at the HRP, which run in the 2001-2003 period. (Author)