Quantitative artificial neural network for electronic noses

This paper reports a quantitative artificial neural network (ANN) to implem ent an electronic nose (enose). A new approach was proposed by the combinat ion of ANN with fundamental aspects of analytical chemistry, especially wit h the concept of relative error (RE) in quantitative analysis. Thus, both t he qualitative and quantitative requirements for ANN in implementing enose can be satisfied. Converging criterion while training the ANN can be set ac cording to RE function (RE-Func) designed in this work. Fast converging spe ed and good prediction accuracy could be promised with the use of RE-Func. In addition, transform functions in logarithmic, sigmoid and their combined forms to pre-process training data sets were evaluated. Also training meth ods, such as order of training data magnitude and treatment of data passed the RE requirement checking in last iteration, were optimized. The enose wa s constructed to response quantitatively towards alcohol vapor within conce ntration range of 0.001-1 mg/l in the presence of petroleum gas and water v apor. The prediction error was <10%. No qualitative mistake of prediction w as observed for samples of alcohol and petroleum vapors, or for their mixtu res. (C) 2000 Elsevier Science B.V. All rights reserved.