System identification of electronic nose data from cyanobacteria experiments

Abstract

Linear black-box modeling techniques are applied to both steady state and dynamic data gathered from two electronic nose experiments involving cyanobacteria cultures. Analysis of the data from a strain identification experiment shows that very simple low order MISO black box model structures are able to produce very high success rates (up to 100%) when identifying the toxic strain of cyanobacteria. This is comparable with the best success rates for steady state data reported elsewhere using artificial neural networks. Analysis of data from a growth phase identification experiment using MIMO black-box models produces success rates of 82.3% for steady state data and 76.6% for dynamic data. This compares poorly with the best performing nonlinear artificial neural networks, which obtained a 95.1% success rate on the same data. This demonstrates the limitations of these linear techniques when applied to more difficult problems.