| A JAVA program (ENBrain) was developed to implement artificial neural networks in electronic noses (e-noses) for odour assessment. Three-layer feed-forward neural networks with sigmoid function neurons were built up by ENBrain. The Back-propagation algorithm was applied to train the networks. Networks with different number of hidden units and initial weights were experimented. N-butanol was used as the test odorant and measured with an Alpha MOS Fox 3000 e-nose equipped with 12 metal oxide sensors. Each n-butanol sample was analysed with 4 replications, and a total of 80 data vectors were collected. Among these 80 data vectors, 32 ones were used for training the neural networks and the other 48 were used for testing the neural networks. Odour intensities of n-butanol samples were assessed by human panels using an eight-point and a 0--200 intensity scale. PCA was performed on training and testing data sets. For training set, all data vectors are linearly separable, while for testing set, four n-butanol samples are more difficult to be linearly separated.; High accuracy (100%) was achieved for predicting odour concentrations and intensities during training. With the increase in the number of hidden units, an increase in accuracy and decreases in the SSE and MSE were observed from 1 to 14 hidden units and little changes were observed after the number of hidden units reached 14.; Trained networks were tested over the testing data set. High coefficient of determination (R2 = 0.89) was obtained between network predicted and actual odour concentrations. A high correlation (R2 > 0.95) was obtained between network predicted and human panel assessed odour intensities in both the eight-point and 0--200 intensity scales. |
