Estimation of multi-component mixture proportions using regression machine analysis of ultra-wideband spectroscopic measurements

Abstract

Ultra-wideband signals are used to examine multiple-constituent fluid mixtures in a semi-open system. A feedforward neural network operates on an array of easily computed signal properties, plus the weight and temperature of the fluid samples, to provide an estimate of the constituent proportions. The average performance of the neural network is tested by artificially increasing the test data sample size and repeatedly training neural networks of the same topology. Networks of differing topologies are compared. Statistical analysis is performed on these results and the 95% confidence interval of the data prediction is shown. The 95% accuracy averages around ± 6.9 percentage points for both oil and water.