3.1. GC-MS Analysis Results
From the comparison between samples chromatograms, substantial differences were found. The main difference between 12-months and 24-months ripened grated PR lies in the amount of fatty acids that characterize this product. They are acetic acid, butanoic acid, hexanoic acid, octanoic acid, n-decanoic acid, and their presence is much greater in 24-months PR. In Figure 2
, histograms for each of the aforementioned compounds are shown: results are presented in terms of mean ± standard deviation of the mean with an arbitrary unit. This result is widely confirmed in the literature. Indeed, it is well known that these fatty acids are the results of fermentation processes, especially in butter and seasoned cheese. Some studies revealed that the amount of acetic acid and butanoic acid doubles in the period between 12 and 24 months [51
]. The same trend was observed in the other compounds, since biochemical processes that lead to their formation are very similar.
Differences were also found in comparing samples with different percentages of rind and the same rind working process; the same trend is valid both for 12-months and 24-months ripened PR. It turned out that in increasing the quantity of rind, the presence of three compounds increases. Besides butanoic and hexanoic acid, 2-nonanone has the same behavior. It is a member of the class of methyl ketones and it can be found in several foods, such as milk and cheese [53
]. It is produced by the oxidative degradation of fatty acids [54
]. These results suggest that both the fermentation and the degradation happen closer to the rind than in the central part of the cheese.
3.2. S3 Analysis Results
Once data were acquired, sensors responses were checked first. Since the first measures of each session were very different from the others, they were discarded. Consequently, there is a different number of replicas for each sample. Most likely, experimental conditions of first acquisitions were not the same as the following measures in terms of the temperature of the auto-sampler oven that the vials were put in, as explained in Section 2.3
. S3 Analysis. In Table 2
, a detailed description of the number of samples that were considered for the following analysis is shown.
The choice to extract ΔR/R0
as a feature was made after viewing the sensors responses. In Figure 3
, the resistance value, as a function of time during the injection phase, is presented for four sensors that represent the four types of MOX in the S3. They are CuO, SnO2
Au+Au-Nanowire and TGS2602. Since the starting point is equal for all the measures, the variation of normalized resistance exhibit that all the sensors are capable of distinguishing samples with different concentrations of rind (samples colors: red for 100% rind, green for 0%, blue for 18%, cyan for 26% and black for 45%). In addition, they show also the ability to recognize the two ripening degrees, characterized with a solid line for 24-months samples and with a dotted line for the others. Finally, responses to the working processes are highlighted using a thicker line for WR samples as compared to SR ones.
TSG2602 and SnO2Au+Au nanowires seem to be the best MOX to identify ripening and rind working process at fixed concentration, since minimum resistance varies mostly for samples with the same rind percentage. Conversely, CuO and SnO2Au (RGTO) responses are more useful to recognize “pure” samples (0% and 100% of rind) from mixtures, since in the first case ΔR is bigger.
In Figure 4
, boxplots of TGS2602 response that include ΔR/R0
for each sample are shown. This sensor represents the general trend that can be observed in all the sensors. Obviously, since different sensing materials are used, there are differences in the highlighted groups that overlap. On the left part of the figure, there are 24-months seasoned samples, in the upper part, grated cheese with SR, while in the lower, PR with WR. In the right part, there are 12-months ripened samples, and they follow the same trend. The first boxplot is relative to samples of 0% rind; its ΔR/R0
is different with respect to all the other groups, but it is more similar to WR grated PR, both at seasoning stage. This result reflects the fact that they are characterized by a greater amount of humidity.
After checking the general sensors performances, JB-test was applied to the dataset. Only 4 of the eight parameters followed a normal distribution (p
< 0.05); they correspond to the features extracted by the two tin oxide nanowires and RGTO sensors. This was the main reason for choosing PLS. In Figure 5
, PLS score plot was made, considering the first two latent variables (LV) for a total explained variance equal to 99.95% (99.87% for LV1 and 0.08% for LV2). The plot measures are divided by seasoning degree. It can be observed that the 24-months class is in the central part of the graph, while the other one is divided in the left and right part.
For this reason, classification techniques were used in a hierarchical way. In addition, another motive for this choice was to simplify classification models, since this is a 15-class problem. Hence, in the first step, classifiers were used to distinguish the seasoning degree; in a second step, for each ripening state the different working processes were discriminated; finally, ring percentage was taken into account. In Figure 6
, a scheme of the steps is shown.
Regarding ANNs structures, three different ones were considered, one for each step. In the first case, a two-layers architecture with 3 neurons in the input layer and 1 in the output layer was considered. For the second stage, the same number of layers was used, but two neurons were put in the first one. Finally, the third ANN had the same structure as the previous ones, but with 6 neurons in the input layer. For all the neurons, hyperbolic tangent sigmoid transfer function was chosen.
In Table 3
, overall classification rate of the two classifiers is put side by side. In general, ANN classification rates are better than those of PLS-DA. Indeed, ANN is able to recognize correctly all the samples based on seasoning and rind working processes. PLS-DA performances are lower, although it can reach good classification rates. The distinction between rind percentage shows that both classifiers can classify samples with SR better than those with WR. A possible explanation for this result could be the different amount of humidity: WR samples have a higher content of humidity because of water treatment and this could cause the occupation of the adsorption sites by water molecules instead of the ones that characterize the volatile fingerprint of the samples.
To the author’s knowledge, only few researches have been carried out regarding rind composition of grated cheeses. The preparation of the samples in some works was carried out through the grating process, although the aim was to classify the different varieties of cheeses, like Swiss [55
] or Emmental cheese [56
]. For the latter, it has been tried unsuccessfully to find the “rind-taste” off-flavor [57
]; in this case, the lack of positive results could be due to non-volatile compounds that change only the taste but not the aroma. As regards PR, cheese aroma authenticity and rind percentage recognition have been achieved with an electronic nose equipped with SnO2
and ZnO sensors made at SENSOR Laboratory of University of Brescia [58
]. In this case, the tool was also able to distinguish samples with little differences in terms of rind percentage, such as 18% and 19%. However, unlike this study, only one type of ripening was considered (12-months) and the different working processes were not taken into consideration. Finally, the comparison with this study allows for the assessment of the utility of S3 for fraud detection, since the results point in the same direction.