Drift counteraction with multiple self-organising maps for an electronic nose

In this paper, a new mSom neural network methodology has been developed and applied to improve the classification of odour classes sensed by a multisensor system as an electronic nose subjected to drift. The mSom network proved to be a suitable technique to recognise the response patterns of a chemical sensor array for its means of counteracting the parameter drift problem. This neural architecture involves the use of multiple self-organising maps. Each map approximates the statistical distribution of a single odour set and it is able to adapt itself to changes of input probability distribution due to drift effects by means of repetitive self-training processes based on its experience. The new mSom algorithm proposed here allows to carry out autonomously the needed retraining processes once the input probability distribution changes. At this aim, the analysis of the function dependent on the Euclidean distance between the input data vectors and map codebook vectors is performed also with the use of smoothing filters during the network testing phase (network performance).