Five different brands of instant starch noodles with braised beef were obtained at the commercial market in China. The five brands were marked as C, D, G, L, and Z, respectively. Each brand contained three categories: seasoning powder (P), seasoning mixture sauce (S), and the mixture of powder and sauce (M). Table 1 summarizes detailed information about the tested seasoning samples.

The PEN3 portable e-nose (Airsense Analytics GmbH., Schwerin, Germany) was applied to classify and characterize the instant starch noodles seasonings. The apparatus is composed of a sampling unit and a gas detection system consisting of 10 metal oxide semiconductor (MOS) sensors with varying sensitivity to each volatile compound (15). Prior to analysis, the e-nose was preheated and calibrated with pure air.

Based on the preliminary experimental results, 7 g of seasoning powder, 4 g of seasoning mixture sauce, and 11 g of the mixture of powder and sauce were put into three 150 mL beakers, respectively. The seasoning solution was prepared after pouring 100 mL boiling water into the beaker and stirring for 2 min. A volume of 10 mL sample solution was put into a 30 mL sealed glass vial and incubated for 30 min at 40°C. The chamber and injection flow rate was 400 mL/min. The measurement time, sample interval time, flush time, and automatic zero adjustment time were 60, 1, 300, and 5 s. Each sample was measured 10 times.

Instant starch noodles seasonings were analyzed using ASTREE e-tongue (Alpha M. O. S. Co., Toulouse, France). The instrument consists of an automatic sampler, chemical sensor array, and chemometric software package (16). The sensor array includes seven different chemical sensors and a reference electrode (Ag/AgCl) (17).

According to preliminary experiments, 6 g of seasoning powder, 8 g of seasoning mixture sauce, and 14 g of the mixture of powder and sauce were put into three 500 mL beakers, respectively. The seasoning solution was prepared after pouring 400 mL boiling water into the beaker and stirring for 2 min. The seasoning solution was first filtered using a piece of gauze and the filtrate was then filtered through a 0.45 μm nylon syringe filters. Then, 80 mL of filtrate of each sample was analyzed using the e-tongue. Before the test begins, 0.01 mol/L HCl, NaCl, and sodium glutamate were applied to activate, calibrate, and diagnose the chemical sensors (18). The e-tongue parameters were as follows: acquisition time of 120 s, clean time of 10 s, acquisition period of 1 s, stirring rate of 3 r/s. Each sample was measured in three times.

Data fusion involves different levels of data abstraction (19), with low-level fusion being the most common method and giving good results. In low-level fusion, the e-nose and e-tongue signals were simply concatenated before model construction. In the resulting data matrix, the number of rows is equivalent to the number of samples, and the number of columns is equivalent to the number of signals obtained from e-nose and e-tongue (20). We used this low-level fusion to combine e-nose and e-tongue data for the combined analyses discussed below.

The individual e-nose and e-tongue data and their fusion were applied to PCA, CA, and MLPN. PCA is the most popular feature extraction method in pattern recognition. It is used to differentiate samples and extract the most important information from data. PCA reduces the dimensionality of datasets while retaining most of the information of the original data (21). PCA interprets the data in two or three dimensions by the first two or three PCs, which explain the largest amount of variance in the data (22).

Cluster analysis is an unsupervised classification technology which can cluster samples according to their similarity in multidimensional space (23). It identifies degrees of homogeneity of samples that are similar, while samples placed in different clusters reflect dissimilarity (24). A hierarchical CA method was applied to classify 60 samples based on the response signals of 17 sensors (7 e-tongue sensors and 10 e-nose sensors), with the squared Euclidean distance taken as a measure of similarity and the Ward’s method taken as the amalgamation rule.

Multilayer perceptron neural networks analysis is a feed-forward supervised classification method consisting of an input layer, several hidden layers, and an output layer (22). It can be described as a non-linear projection vector from input to output. The sensors and different brands of instant starch noodles seasonings were used as the input layer and output layer, respectively. Training dataset used 10 sensors as inputs, eight neurons in the first hidden layer and six neurons in the second hidden layer. The four groups of different drop heights were used as the outputs, so there were four neurons in the output layer. The activation function of the hidden layer is a hyperbolic tangent function, and the activation function of the output layer is a Softmax function (25). Besides, to achieve a good generalization capability, the dataset used for adjusting the neural networks was again divided into a training set and a testing set. Again the 60 samples (12 samples for each group) were divided into two groups: 9 samples (75 percent of samples) for the training, and 3 samples (25 percent of samples) for the testing.

Principal component analysis and MLPN were performed using Statistical Product and Service Solutions (SPSS) 20.0. CA was conducted using Minitab 16.2.

The typical e-nose response curves of CP, CS, and CM are shown in Figures 1A–C. The response signal was denoted by relative resistivity (G/G0), where G and G0 are the conductivities of the sensors when exposed to sample gas and pure air, respectively (26, 27). The e-nose response curves in the three figures exhibited similar temporal changes. The response values of sensors W1W, W2W, and W5S reached peak value at the 5th second and reduced quickly anaphase, finally, it reached a stable value. The response values of W1S rose rapidly from 0 to 10 s, then increased more slowly slightly, and finally remained high for the rest of the measurement. The response values of W3S, W2S, and W6S was appearing to increase slightly. In contrast, W1C, W3C, and W5C reduced slightly, where G/G0 values were generally less than one since they are negative metal-oxide sensors (28). Overall, the response signals of W1S, W2S, W5S, and W1W remained at a high level (Figures 1A–C). The results suggested that aromatic compounds, nitrogen oxides, terpenes, and sulfur-containing compounds were the main constituents in the aroma of instant starch noodles seasonings (29).

The radar plots of e-nose sensor responses are presented in Figures 1D–F. There were no significant differences among the 10 sensors in seasoning powder, seasoning mixture sauce, and the mixture of powder and sauce (except GM), indicating that all seasonings had similar aromas. The W1S, W2S, W5S, and W1W sensors had stronger responses than other sensors. These results were in accordance with the typical e-nose response curves (Figures 1A–C). The GP sample had higher values of W1S, W2S, and W5S than other samples, while LP had lower values of W1S, W2S, and W5S than other samples (Figure 1D). The DS sample had higher values of W1S and W2S than other samples (Figure 1E). Similarly, GM exhibited the highest response values in the sensors of W1S, W2S, W5S, W1W, and W2W (Figure 1F). The results suggested that different brands of instant starch noodles seasonings had significant difference in aromas.

The e-tongue sensor responses of CP, CS, and CM are shown in Figures 2A–C. For the seasoning powder, the signals of HA and GA increased from 0 to 10 s, while the signals of ZZ, BB, and JE decreased during this time period. These signals started to stabilize after approximately 30 s, and fully stabilized at 120 s. For the seasoning mixture sauce, the signals of HA, ZZ, and CA increased from 0 to 30 s, and other response sensors had no significant changes. These results suggested that HA, ZZ, BB, and GA are the main contributors to taste profiles of instant starch noodles seasonings.

The radar plots of e-tongue sensor responses are presented in Figures 2D–F, which show similar overall trends. Instant starch noodles seasonings of different brands cannot be distinguished by simply observing the response curves and radar charts of the samples. It is necessary to use multivariate statistical analysis to process the e-tongue data. Therefore, we used the stable signals at 120 s for PCA analysis.

The e-nose PCA plots are presented in Figure 3. The seasoning powder samples of the same brand were located together. PC1, PC2, and PC3 explained a combined 88.96% of the total variation. For the seasoning mixture sauce, the first three PCs explained 98.84% of the total variance. PC1 explained 84.24% of the total variance, and PC2 explained 13.43% of the total variance, PC3 explained 1.17% of the total variance. For the mixture of powder and sauce, PC1 accounted for 64.50% of the total variance. PC2 accounted for 22.07% of the total variance and PC3 accounted for 7.33% of the total variance. In total, the first three PCs explained over 85.00% of the total variance in e-nose data.

In multidimensional space, the samples were clustered into five non-overlapping groups. Samples with more similar flavor would appear closer to each other. The CP and DP samples were located very close to each other, suggesting that CP and DP had similar aromas. The characteristic aroma attributes of CS, DS, and GS were also similar. The results showed that principal components analysis of e-nose data can be used to classify different brands of instant starch noodles seasonings.

The e-tongue PCA plots are presented in Figure 4. For seasoning powder, 91.49% of the total variance was explained by PC1 and 7.80% of the total variance was explained by PC2. Similarly, for seasoning mixture sauce, the first two PCs explained 98.50% of the total variance among samples. For the mixture of powder and sauce, the first two PCs explained 98.63% of the total variation. Thus, for all three types of samples, PC1 and PC2 explained almost all the total variance.

As shown in Figures 4A–C, the CP, CS, and CM samples were distributed mainly on the positive side of PC1, while the rest of samples were mainly located in the negative side for PC1. The results indicated that the taste profiles of CP, CS, and CM were strikingly different from other samples. The possible reasons were that different manufacturers used different materials and production processes, which may cause the chemical ingredients and tastes of instant starch noodles seasonings to differ among brands. In general, the PCA results suggested that e-tongue can properly characterize different brands of instant starch noodles seasonings (30).

Combination of e-nose and e-tongue data has been applied in various foods analysis and quality control (31). PCA and CA were applied on the combined e-nose and e-tongue dataset to discriminate between different brands of instant starch noodles seasoning (Figures 5, 6). For the seasoning powder, seasoning mixture sauce, and the mixture of powder and sauce, the first three PCs explained 86.96, 98.09, and 94.38% of the total variance, respectively. As shown in Figure 5, all instant starch noodles seasonings were grouped into five different clusters, with no overlap among groups. Combining e-nose and e-tongue data resulted in much better clustering than occurred when analyzing either e-nose or e-tongue data alone.

Relationships between samples were visualized with a dendrogram based on a nearest neighbor method that used Euclidean distances. The combined e-nose and e-tongue data were analyzed using CA, for comparison to PCA. As shown in Figure 6, all seasoning powder samples of the same brand clustered together, with five distinct groups representing the five brands of samples. The seasoning mixture sauce and the mixture of powder and sauce similarly clustered by brand. CA gave us a preliminary classification even though the group divisions were different at different distances. The results of CA were in agreement with those from PCA, providing enough information to discriminate different brands of instant starch noodles seasonings.

Table 2 shows the identification results of MLPN models. For the seasoning powder and seasoning mixture sauce, the identification rate of the combined data was 100% in both the training and test groups. For the mixture of powder and sauce, the identification rate of combined data was 100% in the training group. MLPN modeling demonstrated that the combination of e-nose and e-tongue data enabled better discrimination than individual e-nose and e-tongue data alone. These indicated excellent performance of classification of samples into their correct treatment groups by using the MLPN. This improvement resulted from the complementarity of e-nose and e-tongue data. The e-nose and e-tongue were sensitive to aroma and taste of instant starch noodles seasonings. This complementary behavior of cross-sensitivity plays an important role in multisensor systems (22).