A NEURAL-NETWORK APPROACH TO ANALYZING MULTICOMPONENT MIXTURES

This paper describes an approach for determining the individual analyte concentrations in a mixture of analytes using sensor fusion. Sensor fusion here is considered to be a process of gathering information from many different sources and combining these data to acquire pertinent information. Sundgren and coworkers used sensor fusion and an artificial neural network to learn the relationships between the outputs from an array of partially selective sensors and the individual gas concentrations in a gaseous mixture. The research reported in this paper uses sensor fusion and simulated, partially selective sensors. Through the use of simulated sensors, this research has investigated the advantages and limitation of sensor fusion and artificial neural networks. A three-layer back-propagation network was trained on a subset of component space and it was tested for its ability to generalize, where component space is defined as all possible combinations of input analyte concentration. A learning rate of 0.1 was used and no momentum terms were used. This research has shown that sensor fusion with an artificial neural network can accurately map sensor outputs to the actual input analyte concentrations. This research has also shown that increasing the number of sensors used improves the neural network performance. The neural network approach was found to be, under most circumstances, as good as or better than a mathematical technique using a curve fitting method of addressing the problem. The neural network technique was also found to be significantly taster than the mathematical technique and to possess good noise rejection attributes.