An Electronic Nose Based on Coated Piezoelectric Quartz Crystals to Certify Ewes’ Cheese and to Discriminate between Cheese Varieties

An electronic nose based on coated piezoelectric quartz crystals was used to distinguish cheese made from ewes’ milk, and to distinguish cheese varieties. Two sensors coated with Nafion and Carbowax could certify half the ewes’ cheese samples, exclude 32 cheeses made from cow’s milk and to classify half of the ewes’ cheese samples as possibly authentic. Two other sensors, coated with polyvinylpyrrolidone and triethanolamine clearly distinguished between Flamengo, Brie, Gruyère and Mozzarella cheeses. Brie cheeses were further separated according to their origin, and Mozzarella grated cheese also appeared clearly separated from non-grated Mozzarella.


Introduction
Food product authenticity is of great concern both for consumers and manufacturers. Fraud, namely where products hold a certificate of origin, must be detected. There is a need for low cost, portable devices that would allow inspectors to use them in non-laboratory environments, to avoid the delay of sending samples for analysis. In the absence of such a device, an inexpensive technique that does not require dedicated facilities would also be of great help.

OPEN ACCESS
Piezoelectric quartz crystals are inexpensive, as they are standard electronic components, and show a remarkable sensitivity to mass. When coated with a compound that interacts with an analyte, forming a product with different mass, they can be used as chemical sensors. An array of crystals with different coatings can be comparable to a human nose with different receptors. The cost of such a device increases as the frequency of several crystals must be read simultaneously, but even so, they are much cheaper than a gas chromatograph.
The application of electronic noses to dairy products is not new [1], but Ampuero et al. [1] have referred to the difficulties found due to matrix complexity. In fact, most of the work has been carried out by GC/MS, also called an electronic nose in some reports [2][3][4], although no similarity with a human nose can be established. Mass spectrometry without prior chromatographic separation has also been used for dairy products [5][6][7][8]. These bulky and expensive instruments allowed the identification of aroma compounds, only possible with electronic noses if analytes are limited in number and after extensive training [9]. Another possibility consists of the evaluation of dairy products by a sensory panel, which is a subjective method, very dependent on the expertise of the panel members and influenced by human limitations.
Until now, electronic noses used in cheese evaluation have been either commercial [10][11][12][13] or self-assembled in the laboratory [14,15], and have used sensors based on two different types of transducers: metal oxide semiconductors (MOS) [10][11][12][13][14], or bulk acoustic wave sensors (BAW) [15]. Data processing from all sensors was mandatory, and usually no attempt was made to reduce the number of sensors, or to identify those that were really needed, even when a single target compound was defined [15]. In this work we have tried to limit the number of sensors, and to make the discrimination of cheese by milk type or variety as simple as possible. Sensors have been limited to two for each problem. This limitation on the number of sensors made the analysis simpler and users do not need to have any chemometric know-how.
The volatile compounds in the cheese samples were extracted by the static headspace method. Solid phase microextraction (SPME) allowed many volatile compounds to be collected without the use of any solvent.
Cheeses selected for e-nose analysis included some of the world's best known varieties including soft cheeses, such as Camembert and Brie, fresh cheeses, semi-hard and hard cheeses, such as Gruyère, Grana Padano, Gouda, and Manchego, and many others. Cheeses of the same variety produced in different regions were also included in the sample set.
This work showed that an electronic nose with just four sensors could be used as a first screening method to distinguish between cheese made from ewe's milk and cheese produced from cows', or mixed milk (two sensors required), as well as several varieties of cheeses, such as Flamengo, Brie, Gruyère, and Mozzarella (two other sensors needed). The Brie cheeses from different origins were also distinguished, and, among the Mozzarellas, grated cheese and goat's cheese were also separated.
Compounds used to test the sensors' sensitivity were: 2,3-butanedione (Aldrich B85307)  Table 1 lists the 57 analysed cheeses. A code composed of two letters and a number is used to identify the cheeses throughout the paper. A constant nitrogen flow was maintained through the system. The flow was controlled by a flowmeter (Cole Parmer), placed upstream from the distribution valve. Between the flow controller and the distribution valve, a homemade oven with a septum allowed introduction of the SPME fibre, and the thermal desorption of the cheese sample volatile compounds.
The frequencies of oscillation of the sensors were simultaneously monitored and stored on a PC at intervals of 1 second, using a Counter/Timer device PXI 1033, from National Instruments, and software written in LabView.

Sensitivity Evaluation
A flask, with a screw cap with a septum and connected to a smaller flask, was flushed with nitrogen. The compounds under study, usually found in the cheese bouquet, were introduced in both flasks and thermostated in a bath, at 20, or 25 °C, temperatures at which the vapour pressure of the compounds under study were known. An opened tube, immersed into the liquid of the second flask, allowed equilibration at atmospheric pressure, while trapping undesired soluble volatiles. Samples of first flask headspace were withdrawn with a syringe. The exact concentration of the compound in the syringe could be calculated, after which the sample was injected through the injection port of the electronic nose, and the responses of all the five sensors were recorded. By injecting known volumes, a calibration graph could be plotted, and the slope of the linear portion of the calibration curve (sensitivity), could be obtained after least squares fitting.

Cheese Analysis
The SPME fibre was cleaned in a homemade oven, at ~230 °C, and the complete desorption of compounds was confirmed by analysing the frequency of oscillation of the quartz crystals.
Cheese (2.0 g) was weighed in a 10 mL vial, which was then closed with a silicon septum coated with Teflon, and a removable centre crimp seal. After storage at 4 °C for 24 h, the vials were immersed in a water bath at 30 °C, and a SPME fibre was introduced into the vial headspace for exactly 30 minutes.
Meanwhile, nitrogen was continuously flowing through all the sensors (total flow of 30 mL/mim), and baseline frequencies were recorded. The SPME fibre was then inserted into the oven, and the compounds were desorbed and flushed to the five crystals. The frequencies of the five sensors were simultaneously displayed on the PC monitor, and saved to disc at 1 s intervals. The minimum frequency values were recorded for each sensor, and the difference to the corresponding baseline frequencies computed.
The frequencies of the piezoelectric crystals decreased due to the compounds interacting with the coatings after which they increased again, and reached the baseline values, as soon as desorption from the fibre, as well as from the coatings of the piezoelectric quartz crystals, was completed. The electronic nose and fibre were then ready for a new analysis. The reported frequency shift values for each cheese and each sensor was the mean of four or five replicate analyses. Table 2 shows the frequency decrease for the piezoelectric quartz crystals due to coating. Assuming that the Sauerbrey equation [16] could be applied for these coatings, i.e., that they were all thin rigid films, which is very unlikely, at least for the polymeric coatings, the amount of coating was estimated to range from 9 to 113 μg. The coatings were selected keeping in mind that they must physically adsorb the volatile compounds reported to be responsible for the cheese bouquets. The interaction between the sensor coatings and the volatile compounds must not involve the establishment of chemical covalent bonds, as this would prevent sensor reversibility. Selectivity is not an issue in electronic noses, but different sensitivities to the target compounds from the different sensors is mandatory.  Figure 3 shows the sensitivity of the first five sensors to 14 compounds that, according to the literature, appear in the bouquet of many cheeses. These were selected because they were known to appear in high quantities in some of the selected cheeses, although many other compounds could have been tested, as more than 600 volatile compounds have been identified in cheeses [17]. Sensor 3, coated with Carbowax, was the most sensitive to 3-(methylthio)propioanaldehyde, as well as to acetic acid, 2-nonanone, and 2-undecanone. This last compound was also detected with similar sensitivity using sensor 1, coated with 1,10-decanedithiol. This sensor was also the most sensitive to 2,3-butanedione, 2-butanone. 2-heptanone, butyric acid, 2-pentanone, isovaleric acid, isovaleraldehyde and ethyl hexanoate (sensitivity very similar to sensor 5), although very low sensitivity for the first two compounds was observed. Butyric acid is one of the most important short chain free fatty acids in ewes' cheese [18], and could also have been detected with high sensitivity by sensor 3 (Carbowax coated). Sensor 5, coated with poly(dimethylsiloxane) -PDMS, was the most sensitive to butyraldehyde, although overall sensitivity to this compound was low. Sensor 4, coated with Tenax, was the most sensitive to dimethyl disulfide.

Results and Discussion
Sensitivities were not only defined by the coatings applied to the sensors, but also by the ability of the fibre to extract the compounds from the headspace. The SPME fibre used was coated with Carboxen-PDMS and has been shown to be capable of extracting alkanes, esters, methyl ketones, carboxylic acids, as well as sulphur compounds [19]. Its affinity to each individual compound was a major influence on the responses obtained with the sensors and therefore overall sensitivity to the different compounds depended both on the coating sensor and on the fibre. Figure 4 shows the dendrogram obtained based on data from sensor 2 and sensor 3, with Chebyshev distances and complete linkage. It can be seen that a first group contained five of the ten ewe cheeses, and that the other five ewe cheeses (four Manchego cheeses and one Seia cheese) belonged to another group, where one cheese made of a mixture of ewes', goats' and cows' milk, goat cheese, buffalo cheese and some cows' milk cheeses could also be found. Most of the cows' milk cheeses (32) were excluded from these groups. Figure 5 shows a plot of the frequency decreases observed for sensor 3 vs. sensor 2, where the groups containing ewes' cheese have been marked. It can be concluded that ewe cheeses are characterized by low frequency shift responses on both sensors. Analytical signals of less than 30 on sensor 3 and 20 on sensor 2 could be undoubtedly assigned as belonging to ewes' cheese. Cheeses giving rise to signals between 30 and 60 on sensor 3 and between 18 and 45 on sensor 2 could also be from ewes' cheese, but classification was not solid.
Outside these limits, no cheese made from ewes' milk could be found.
The distinction between ewe cheeses and cheeses made from cows' milk, and especially from cheeses made from a mixture of milks is very important from a commercial point of view, as the addition of cows' milk to ewes' milk is common. This work allowed us to identify half of the ewe cheeses, but left the other half in a grey area, where the origin could not be confirmed. All cheeses outside these two groups could however be confirmed as non-ewes' cheese. We are only aware of one other study [14], based on an electronic nose composed of chemical sensors, where cheeses made from goats' milk, were effectively separated from cows' milk cheeses. Although in this study all the milk was from the same origin.
In the present study, cheese samples were commercial, from a variety of different sources, and the milk came from many different sources. Cheese manufacturing conditions were also different, and ripening stage was unknown. No other information on the cheeses besides that displayed on the label was available. A recent work [13] where ewes' milk cheeses have been analyzed

Ethyl hexanoate
by an electronic nose and GC-MS, showed that significant differences in the bouquet arose due to ripening time and manufacturing technique. Climate, geology, forage and breed also influence milk quality [20], and it is expected therefore, that the problem of detecting fraud in ewes' cheese to be less demanding than the scenario presented here, as information on cheese variety, origin and ripening time would be available to inspectors. The five sensors used initially to distinguish cheeses made from ewes' milk from cheeses made from other milks did not allow the cheese variety to be distinguished. Therefore, 31 new samples of some of the varieties previous analyzed were in contact with five new sensors. The sensitivity of the new sensors to some of the compounds commonly found in cheeses was not studied, but they were chosen based on previous knowledge that existed in our laboratory. For instance, Mn-pht coating was known to respond to alkanes [21], and TEA to sulfur compounds [22]. Figure 6 shows the result of a cluster analysis made with data obtained with sensor 10 and sensor 8, also based on Chebyshev distance and complete linkage. Brie and Flamengo cheeses, except for Brie coeur de lion, which had a different origin, were clearly separated. Mozzarella cheeses belonged to a single group, with Mozzarella grated cheeses also separated. Gruyère cheeses were also distinguished. There was only one goats' cheese (a Gouda cheese) which was also distinguished by sensors 8 and 10. Figure 7 shows a plot of data obtained with sensor 10 vs. data obtained with sensor 8. Groups have been marked on the plot. Mozzarella cheeses were characterized by very high responses from sensor 10. Among cheeses with moderate responses from sensor 10, Brie and Flamengo cheeses were characterized by low responses on sensor 8 (less than 30), Gruyère cheeses by responses between 40 and 50, and the goats' cheese by very high signals (>65).

Conclusions
Two sensors coated with Nafion (sensor 2) and Carbowax (sensor 3) were able, in the absence of any other information, to certify that labelled ewe cheeses were effectively produced from ewes' milk (responses of less than 30 on sensor 3 and 20 on sensor2), to assign them for possible certification after further analysis by other methods (responses between 30 and 60 on sensor 3 and between 18 and 45 on sensor 2) or to exclude this hypothesis. Two sensors, sensor 8 and sensor 10, coated with polyvinylpyrrolidone and triethanolamine, respectively, were able to distinguish between Mozzarella cheeses (responses higher than 75 on sensor 10), Flamengo and Brie (responses lower than 27 on sensor 8 and between 62 and 70 on sensor 10), and Gruyère (bellow 75 on sensor 10, and between 42 and 46 on sensor 8). These four sensors will be used, in the future, to detect frauds and to confirm authenticity of cheeses with a certificate of origin.