Odour classification

An early application of a pattern recognition method to the prediction of olfactory qualities from physicochemical data was reported by Schiffman C2623. A set of 39 odorants was represented by 25 parameters (including Raman spectral information). Non-linear mapping revealed correlations of the chemical structure and smell. 

A similar study deals with the classification of trigerainally active compounds C2881. A set of 12 molecular descriptors has been selected that are most useful for a semi quantitative description of the activities of 47 compounds. 

Pattern recognition techniques have been applied by Me Gill and Kowalski to determine the intrinsic dimensionality of smell C197, 1981. 

A set of 43 features was derived from physical and chemical data CIR-, UV-, NMR-spectra, molecular weight, melting point, boiling point, density, specific rotation, solubility in water and alcohol) for each of a total of 47 compounds. Feature selection and eigenvector plots indicate that there are only two meaningful axes in the 43-dimensional data space. The axes were found to relate to the molecule s electron donor ability and its directed dipole. 

---

HARPER, R., BATESMITH, E C. LAND, D.C. 1968 Odour description and odour classification. J A Churchill Ltd., London. 

Odour has three properties, viz. character, intensity and persistence. All three are subjective and can only be measured in sensory terms. Most correlation work has been done on character since it is, superficially, the easiest to measure. However, the description of an odour is associative, since we have no hard reference points. Difficulties in finding correlations can arise from the use of odour classification systems. For instance, in our fragrance work, we found that the use of the term fruity to describe an odour family led to confusion, since the criteria for a molecule to possess an apple odour are not the same as those for pear. We classify the two together because of the similarity of the botanical sources, but this is not necessarily related to the odour properties. To study structure-odour correlations, we must therefore ensure that we are using meaningful parameters. 

Fig. 5.6. Indicator functions for the odour classification task. The same network is trained to 2 odours, A and B from which indicator functions (I.F.) are constructed. Shown is the indicator function response for both A and B when odour A and B are presented. In each case the indicator function for the learnt odour is far higher than that for the distractor odour.

Harper, R., E. C. Bate Smith, and D. G. Land Odour Description and Odour Classification — A Multidisciplinary Examination. London J. and A. Churchill Ltd. 1968. 

Water Reclamation Works by their very nature can, at times be the source of unpleasant odorous emission. The odour-intensive compounds (osmogenes) which make up these emissions are believed to arise mainly as the decomposition products of carbohydrates and proteins. The breakdown of this waste material proceeds by aerobic and anaerobic processes at various stages of the treatment plant. Atmospheric pollution of this nature frequently results in complaints from members of the public either resident, or perhaps employed in the vicinity of such works. In order to confirm or deny that a reclamation works is responsible for a particular nuisance and, if possible to identify the causal agents it was decided that the Authority should have the capability of analysing for odorous and other polluting constituents of the atmosphere. This paper describes the progress made towards this objective and summarises the experience gained with a procedure in use. There are two principle approaches available for the analytical classification of odorous emissions -... 

People use classifications all the time. For example, many types of wild mushrooms are edible, but many others are poisonous—even deadly How can you tell which is which Poisonous and deadly mushrooms have characteristics that distinguish them from edible ones, such as odour, colour, habitat, and shape of roots. It is not always easy to distinguish one type of mushroom from another the only visible difference may be the colour of the mushroom s spores. Therefore, you should never try to eat any wild mushrooms without an expert s advice. 

Figure 3B.5 shows the odour profiles of three wines as obtained using the OAV classification scheme. The example illustrates the differences between three types of wine in terms of volatile components, which have been classified into seven aroma series. 

Chastrette, M., Zakarya, D. and Peyraud, J.F. (1994). Structure-Musk Odour Relationships for Indan and Tetralins Using Neural Network. On the Contribution of Descriptors to Classification. Eur.J.Med.Chem., 29, 343-348. 

The preface to the T edition, also intended as a summary to guide the reader through the book, has in the majority retained its relevance for the present edition. The already extensive survey of our field of work is complemented by a number of new topics. Prof. W. Grosch provides the reader with a comprehensive survey of aroma analysis with a special emphasis on key odourants. Contributors from multinational food companies introduce a focus on final products in the section on applications. Additionally, the sector on non-natural flavors has been expanded to include the current state of the European chemical group classifications. 

Although no specific receptors have been identified yet, it is believed that about 100-300 receptors classes exist. This makes it difficult to predict odour sensations from the chemical structure of an odorant and to establish an objective classification system for odorants. 

A value judgement of lAQ can be given in several ways. One can make a classification (e.g. yes/no), such as ASHRAE 62-1989 [31] uses (is the air acceptable or not), resulting in a percentage of dissatisfied, or one can use a fist of descriptors to describe a chemical substance. The latter is mainly used in the food and perfume industry, from which many classification systems of odours have been developed. 

The artificial intelligence systems to which sensor arrays are coupled supply the closest likeness to the human olfactory system. Some of the recent theories on olfaction require that the human nose has only relatively few types of receptor, each with low specificity. The activation of differing patterns of these receptors supplies the brain with sufficient information for an odour to be described, if not recognized. As a consequence of this belief, the volatile chemical-sensing systems commercially available only contain from 6 to 32 sensors, each having relatively low specificity. Statistical methods such as principal component analysis, canonical discriminant analysis and Euclidian distances are used for mapping or linked to artificial neural nets as an aid to classification of the odour fingerprints . 

Apart from the use as a Quality Control device, other areas in which an artificial odour-sensing system could be utilized include all those in which classification of odour is required for example, human body odour, malodours and malodour counteractancy. Another area in which the new instrumentation could be utilized to advantage includes perfume substantivity, or diffusion from a substrate. For example, it could be used to measure levels of perfume in the air from a hard surface cleaner when used on a ceramic tile, or odour from human skin after spraying with a cologne, and so on. 

A range of other statistical techniques can be used in the formulation of a classification model. Since a detailed description of these is outside the scope of this chapter, those which have been used in the study of odour are listed below ... 

---