Comparison of Fisher's linear discriminant to multilayer perceptron networks in the classification of vapors using sensor array data

Abstract

Two different classification methods, Fisher's linear discriminant (FLD) and a multilayer perceptron neural network (MLP), were directly compared with respect to their abilities to differentiate response patterns arising from arrays of chemical vapor detectors. The algorithms were compared in five different types of tasks that had been selected because they produced classification problems of varying character and difficulty. In one task, an array of 20 compositionally distinct carbon black–polymer composite vapor detectors was exposed to P/P⁰ = 0.0075 1-propanol and P/P⁰ = 0.0083 2-propanol, where P and P⁰ are the partial pressure and standard vapor pressure, respectively, of a given analyte. The second task consisted of classification of a mixture of P/P⁰ = 0.011 1-propanol and P/P⁰ = 0.0090 2-propanol versus a mixture of P/P⁰ = 0.0090 1-propanol and P/P⁰ = 0.011 2-propanol. A third task consisted of multiple concentrations of three hydrocarbons, and a fourth task involved clustering two hydrocarbons in the presence of a variable background composition. An additional dataset was generated by exposing an array of five thin-film metal-oxide sensors to the headspace of seven different coffee blends. In each case, the MLP and FLD techniques were compared using the 5-sensor subset of the 20 available sensors that proved optimal for that dataset. The FLD and MLP algorithms yielded comparable performance on straightforward classification tasks, whereas the MLP technique yielded better performance on tasks that involved non-linear classification boundaries. In addition, for the four datasets produced by the carbon black–polymer composite detector array, the performance of each possible 5-sensor subset was evaluated using both signal processing approaches. The performance of the best 5-sensor subset selected with MLP was found to be slightly better than the performance of the FLD-selected subsets, and the performance of the median 5-sensor subset using MLP was nearer to that of the optimal subset than the median sensor array selected by FLD. In one case, the optimal test set performance distribution was found to be significantly better with MLP than with FLD: MLP had a clear advantage (86% versus 57% correct classification rate) when applied to the "coffees" dataset, and this trend is likely applicable to other multi-cluster classification tasks that consisted of non-Gaussian shaped data in lower-dimensional spaces.

Additional details