Chemometrics methods for the identification and the monitoring of an odour in the environement with an electronic nose

The purpose of the paper is to briefly review some researches regarding the adaptation of the electronic nose principle to recognise some malodour sources in the environment, if possible directly in the field, and to monitor the odour intensity continuously. Research aims at improving the portability and the user-friendliness of the instrument, together with testing what kind of signal may be used to monitor the odour. A laboratory-made electronic nose, constituted of an array of tin-oxide sensors, is used in different configurations. The ambient air is either sampled around environmental sources (landfill, urban waste composting facilities,...), or directly transferred into the sensor chamber in the field. Two main options are considered: firstly identifying the source of odour in the background and among interfering odours and, secondly, when the malodour is recognised, trying to monitor it continuously in order, for example, to assess the nuisance or to control an odour abatement system. Chemometrics methods are generally used for both purposes. They provide quick answers and allow to evaluate the relationships between variables and between observations at a glance. They are applied on the sensor signals, eventually preprocessed by a suitable algorithm. Non-supervised analyses, such as Principal Component Analysis (PCA), provide basically a performance evaluation of the system during the development phase. On the contrary, supervised analyses, such as Discriminant Analysis (DA), or some Neural Networks algorithms are quite appropriate to make a reliable recognition in real time, when the system is developed. To predict the odour intensity, different techniques are tested: either using only one of the sensor elements, or applying different chemometrics techniques, such as Multilinear Regression (MLR) on the original measured sensor signals, Principal Component Regression (PCR), or Partial Least Squares regression (PLS). The latter seems to be the most adapted model for the intensity prediction.