Increasing efficiency of two-phase flowmeters using frequency-domain feature extraction and neural network in the detector output spectrum

In this paper, three different regimes including annular, stratified and homogeneous in the range of 5%-90% void fraction, were simulated by Mont Carlo N-Particle (MCNP) Codes. In simulated structure, a cesium 137 source and two Nal detectors were used to record received photons. In this study, the Fast Fourier Transform (FFT) was applied to the registered signals from two detectors in order to analyze in the frequency domain. Several features of signals in the frequency domain were extracted using average value of fast Fourier transform, the amplitude of dominant frequency, kurtosis, Standard Deviation (STD), RMS (Root Mean Square) and Variance. The same features were extracted from analyzed signals of both detectors in order to find the best separation patterns. These extracted features were used as inputs of artificial neural networks (ANNs) to increase the efficiency of two-phase flowmeters. Two multi-layer perceptrons (MLP) neural networks were implemented in MATLAB software in order to classify flow regimes (annular, stratified and homogeneous) and predict the void fraction. All of the training and testing data were obtained correctly and the mean relative error percentage of the predicted void fraction was 1/15 %.