The use of the chemometric techniques to exploit the selectivity of gas sen
sors, although current in the case of multisensors systems, is tested here
with only one gas sensor. The method consists in measuring the electric con
ductance of the sensitive element, a tin dioxide bar, at various temperatur
es. Although these gas sensors follow non linear laws, the modelization alg
orithm Partial Least Square (PLS) is preferred to neural networks. A compar
ison of the two algorithms is made, as well as a study of the portability o
f the model and stability of the prediction. Lastly, the concept is tested
on an autonomous prototype in real conditions.