ELECTRONIC NOSE & ITS APPLICATION

4. Experimental SET-UP

 

 

4.1 The Pico-1 Electronic nose

 

                        Five semiconductors, SnO₂ based thin films sensors were utilised. Two are pure SnO₂ sensors; one is catalysed with gold, one with palladium and one with platinum. They were grown by sputtering with the RGTO technique. RGTO technique is a technique for growing SnO₂ thin films with high surface area. The surface of the film after thermal oxidation step of the RGTO technique presents porous, nano-sized agglomerates which are known to be well suited for gas absorption. A thin layer of noble metals was deposited as catalyst on three sensors to improve sensitivity and selectivity. Thin film sensor produced by sputtering is comparatively stable and sensitive. Furthermore, since the growing conditions are controllable, they can be taylored towards the particular application. Even if catalysed the sensors are not selective and therefore sensor arrays together with multivariate pattern recognition techniques are used.

 

                        The odour sampling system depends on the type of sample and on its preparation. For a simple gas mixtures one uses automated gas mixing stations consisting of certified gas bottles, switches and mass flow controllers. In the case of complex odours like food odours, the volatile fraction (the so-called headspace) is formed inside a vial where a certain amount of odour-emitting sample is put. The vapour can then be collected either by flushing a carrier inside the vial (dynamic headspace scheme) or extracted with a syringe and injected, at constant velocity, in the air flow which is used as carrier (static headspace scheme).

           

                        There are two different design considerations of designing the sensors. Those are first design consideration i.e., linear and the second design consideration i.e., parallel design. The first is comparatively less costlier than the second one. At the same time it has certain disadvantages that the distribution of the sample into each sensor element is uneven, but second consideration have this advantage. The construction of second type is much complex when compared to first.

 

            The basic schematic diagram of an electronic nose is shown in the below figure.

Fig.1 Process in electronic nose

Fig 2. Lab arrangement of electronic nose for coffee analysis

1. An auto sampler (Hs 850 CE Instruments). This device is a standard component of chromatographs; its utility is a high sample throughput and a high reproducibility due to the automation of the measurement process. It consists of a sample carousel, where the vials containing the odour-emitting sample are held; an oven, where the sample is pre-conditioned; a movable mechanic arm with syringe (A).

 

                        The electro-mechanical part of the EN used in this experiment consists of (see a scheme in fig. 2):

 

             2. A mass flow controller (B) to set the flow of the carrier gas.

 

            3. A stained steel chamber (C) which can contain up to five chemical sensors plus a humidity sensor.

 

             4. Control electronics (D) permitting to steer the system (auto sampler, mass flow controllers and sensors) via PC.

 

                        The typical measurement consists of the exposure of the sensors to a concentration step, that is a change of odour concentration from zero to c (each component of the vector stands for a gas component) and back to zero again, and of the recording of the subsequent change in resistance. The classical feature extracted from the response curve is the relative change in resistance.

 

                        A set of Mat lab functions (toolbox) has been developed for analyzing the data. The toolbox permits to perform the following tasks.

 

• Data cleaning (median filter for spikes removal, possible noise averaging) and plotting (for gaining a first impression of the response curves). Software for drift compensation is currently under study.

 

• Exploratory analysis (visual). First various plots of the response curves and of the features can be drawn for each sensor separately (univariate analysis). The most important multivariate tool for exploratory analysis is Principal  Component Analysis (PCA) (score and loading plots). PCA is implemented with a simple user interface giving the possibility of selecting the sensors and classes to be displayed and of grouping classes together. PCA also serves for feature reduction before the use of Multilayer Perceptrons (MLP).

 

• Learning with MLP. The inputs to the MLP are the projections of the data on the first m principal components (the so called PCA scores). The number of inputs m (PCA dimensions) is then a variable to be optimized. To prevent over fitting early stopping (ES) or weight decay regularization  can be used. Both a division in two sets (training and testing) or in three sets for ES (training set is subdivided in training and validation sets) is possible. The error function is minimized with the Levenberg Marquardt algorithm. Ten network initialization are usually performed and the net with the best result on the test set is held.

 

Decomposition of the global learning tasks in successive classification subtasks (hierarchical classification). First the classification between the more istinct clusters is performed, then the finer differences are determined in subsequent steps. This is particularly useful when dealing with a big number of classes and a small number of data. Ensembles of MLPs based on output coding decomposition have also been studied. Work is in progress on the topics of boosting and bagging for increased classification accuracy.

4.2 The measurements

 

                   Measurements were done on ground coffee. Two groups of coffees were analyzed. The first one consists of 6 single varieties (SV) and the blend Italian Certified Espresso (ICE) for reference (this group will be called SV) and the second one of 7 blends, including the ICE, see tables I, II. The fourth row of the tables contains a brief characterization of the coffees, where the commercial value is indicated with + and -.Two grams of ground coffee are introduced into a vial with a volume of 20cm3 which is crimped with seal and septa. The vial is then left in an incubation oven at 50◦C for 30 minutes in order to generate the aroma. Ten vials for every coffee type of the first group and 12 vials for every coffee type of the second group were prepared. Three successive extractions were performed from the same vial. All together there are 10 • 7 • 3 = 210 measurements for the first group and 12 • 7 • 3 = 252 measurements for the second group. While the data set is not big for machine learning standards, where it is usual to have hundreds of examples for each class, this is a considerable dataset to be collected with an E-Nose, where complete datasets normally don’t exceed 100-200 examples (while it is rather common to have less then 10 instances for each class).

 

 

 

 

 

 

 

 

Table I. First group of coffee: Single varieties + ice

 

# coffee	Name	Type Quality (+/-)
1.	ICE	Blend, +
2.	Brazil	Arabic natural, +
3.	Ethiopia	Arabic washed, +
4.	Rio Minas	Arabic natural with defects, -
5.	Guatemala	Arabic washed, +
6.	Peru	Arabic natural,-
7.	Cameron	Arabic,-

 

 

TABLE II. The second group of coffees: blends.

 

# coffee	Name	Note, Quality
1.	ICE	Reference, +
2.	ICE, more toasted	Strong, +
3.	ICE, without natural	Study, +
4.	Robusta	Bad,
5.	ICE def#1	Unripe, -
6.	ICE def#2	Rancid, -
7.	Commercial	Arabic + Robusta +-

 

 

                        Experimental parameters like samples’ conditioning temperature and fluxes were optimised to reduce the sensor stress and to increase the measurement rate while still reaching sensor’s steady state conditions (which are believed to be more reproducible). The time interval between the extractions sufficient for the headspace to reach equilibrium conditions was found to be 40 min.

Fig. 4

 

                        An external view of an Electronic Nose interfaced with PC is shown in the above figure.

 

                        As for the sensorial analysis, the panels (formed respectively by 18 and 14 judges) judged the final product (cups of espresso coffee) according to 10 quantitative descriptors (colour intensity, cream texture, olfactory intensity, roasted, body, acidity, bitterness, astringency, global positive odour and global negative odour) and 4 qualitative descriptors (attractiveness, finesse, balance and richness). Each descriptor is given a mark from one to nine. One sample for every coffee type (plus a random repetition per group) is tasted. In the quantitative analysis the panel is given a reference for adjusting its judgements, while this is not the case for the qualitative analysis which should provide a personal, ”hedonic” impression. Since the qualitative values are not calibrated, their spread is considerable. Therefore, for every coffee type, the mean over the 4 qualitative descriptors and over the panellists is considered as a reliable global parameter characterizing the sensorial appeal of a coffee. This quantity is pictorially termed Hedonic Index (HI). The two averages help to reduce the uncertainty (standard deviation) by a factor √N, where N is the number of sensorial measurements, i.e. N = judges • qualitative descriptors. For the SV group the standard deviation of the HI is σ mean = 0.2). The detailed procedures adopted for testing the Espresso in this study are described in.

Page   1   2   3