Preprocessing of matrix QCM sensors data for the classification by means of neural network

An experimental comparison of linear and non-linear pre-processing methods for olfactory data is made. The original data are formed by 280 values of reaction from six quartz crystal microbalance (QCM) sensors taken at 1 s intervals. Data vectors are processed with the non-linear maximum filter or by the linear averaging filter and are used as inputs of a feedforward neural network using the back-propagation learning rule, one hidden layer containing from 5 to 15 neurons. The filter window size is 5–10. arning set is composed of 60 sensor reactions for six types of cologne. The neural network correctly classifies 82–86% of independent examples. The usage of the maximum filter with a small window size allows an increase of the classification rate by 3–5%. The best results (86%) are obtained when only first 50 measurements of sensor reaction are used.