Measurements using arrays of electrochemical gas sensors, combined wit
h pattern recognition methods, were used to classify wheat samples by
quality grade. The classifications corresponded closely to those made
by trained grain inspectors. Volatile compounds evolved from warmed sa
mples of grain were passed over a heated noble metal catalyst and then
into a series of electrochemical sensors. Signals from four sensors w
ere recorded for four different catalyst temperatures in order to gene
rate 16 signals for each grain odor sample. The 16 sensor signals were
treated as a 16-dimensional vector or pattern of responses that was c
haracteristic of the odor sample. The patterns for different grain odo
r samples were compared using both nearest-neighbor analysis and a com
mercial neural network simulation (NNS) program. These methods classif
ied the samples correctly by grade with an accuracy of 68% and 65%, re
spectively. After compensation for instrument parameters, the NNS scor
e improved to 83%; the nearest-neighbor analysis could not be similarl
y compensated. The robustness of the two algorithms was compared by ad
ding simulated random and systematic errors to the sensor response pat
terns. The original data were used as the training set, and the patter
ns with errors added were used as the test set. In these cases, the NN
S consistently outperformed the nearest-neighbor method at classificat
ion of the grain odor samples.