A connectionist approach to fuel cell sensor array processing for gas discrimination

Abstract Fuel cell sensors are in widespread use throughout the world, applied primarily in the areas of evidential breath testing and environmental monitoring. A problem which affects the adoption of such transducers however lies in their specificity. Although the fuel cell does have some inherent specificity it will, for example in the case of breath alcohol measurements, also respond to other alcohols and aldehydes if they have a significant presence in the test sample. This paper presents methods which improve transducer specificity through the incorporation of discrimination algorithms based on artificial neural networks. The use of such algorithms in conjunction with a single cell has facilitated the detection of the presence of contaminants in a sample (for example methanol in ethanol) at levels as low as 5% and also allows the identification of a single pure gas from a set of known gases. The results obtained are improved however when neural networks are used in conjunction with an array of sensors where each cell in the array is operated under different conditions. Exploitation of the non-linear temperature effects exhibited by an array of cells eases the task of the network and facilitates improved performance. The paper compares both supervised and unsupervised network paradigms together with a range of signal pre-processing techniques applied to both single and multiple cell systems.