A flavor uniformity evaluation and improvement of Chinese spirit by electronic nose

Abstract Fenjiu spirits are famous for their multifarious flavors in China. Nevertheless, the uniformity and homogeneity of the products are always challenges for the producers. Three flavor measurement methods, that is, zNose™, GC‐MS, and spirit tasters, were employed and correlated for evaluation of Fenjiu flavors' difference. An automatic blending system was designed to improve the uniformity, where base liquors of Dazongjiu, Dajiu, and Daijiu, as well as pure water, were used for blending, and BPNN was employed for regression and error minimization. Results showed that zNose™ results could be correlated with GC‐MS in 95%; hence, zNose™ can replace GC‐MS in Fenjiu flavor measurement. With the zNose™ assistance, an experimental scale, fully automatic blending system could be optimized with different modern mathematical algorithms to improve the Fenjiu products uniformity. We believe that this detailed work will advance the scientific knowledge in the field and help to facilitate the uniformity and homogeneity of Fenjiu products.

blended to formulate the final Fenjiu products. Nevertheless, as an agricultural produce, the uniformity and homogeneity of Fenjiu products are always a challenge for the producers. This means, the named same product may have different flavor and quality (Jiang, Peng, & Peng, 2010;Yu, Sun, Wang, & Huang, 2018;Zhang, Rao, & Li, 2016).
The major causes may include the following: 1. The production season: Fenjiu is produced in three seasons of autumn, winter, and spring. Generally, products of different seasons possess obvious different flavors and quality; 2. The base liquors' production process: In the above production steps, each step may induce difference due to the diversity and complexity of raw materials. In addition, workers' experience, equipment, and aging condition are the potential treats cause the flavor differences of the base liquors; 3. The spirit tasters: In classifying, grading, and blending of the final Fenjiu products, spirit tasters play the key role. However, as human beings, they are individually influenced by age, health status, emotion, fatigue, and even weather condition, which may lead to inaccurate judgment (Deng, Li, Wang, Song, & Zhang, 2017;Li, Ge, & Sun, 2017). All these situations occur in the entire Chinese spirits industry, which makes the spirits flavors and quality unstable and unreliable. With the novel technologies development, it is a good time now to find a solution to improve the uniformity of Chinese spirits to benefit this old industry.
Blending is the last but most important step in spirits' production.
A sound blending should compensate the imperfection of the base liquors and improve the flavor and quality uniformity. The cost is also considered in this step where low-price base liquors should be used in priority, but good flavors and quality should be retained. However, the traditional "trial to scale" method is still widely adopted in spirits industry, and the evaluation criterion mainly depends on the spirit tasters (Huang, Sun, & Su, 2014). This method has shown its weakness in quality assurance due to the above reasons. Moreover, a good cost efficiency is hard to achieve because of the complexity and large variations of the base liquors (Peng & Pan, 2015;Yuan, Zhao, Chen, Chen, & Li, 2018). Some modern blending methods have been attempted with assistance of GC-MS and computation algorithms, that is, fuzzy logic, neural network, and objective programming (Xie & Li, 2007;Yang, Wu, Huang, & Huang, 2010;Zhang, Meng, & Li, 2011). However, fast and online measurement of flavors in blending process has not been reported in literature, and a fully automatic blending operation has not been implemented up to date. Also, those methods reported in literature are confined to unaltered base liquors and cannot adapt to varying raw materials. Furthermore, the blending effects cannot be verified online, and error cannot be modified instantly.
This study is aimed to evaluate the flavors difference and attempt a solution to improve the uniformity. Various tools will be employed, compared, and correlated for Fenjiu flavors measurements. A zNose™ is introduced for fast, nondestructive, and online flavors detection, and its accuracy will be testified. With the zNose™ assistance, an experimental scale, fully automatic blending system was designed and optimized with different modern mathematical algorithms to improve the Fenjiu products uniformity.

| Fenjiu samples
Five hundred and seventy-one samples were used (Table 1), including 90 finish products of different grades, 235 finish products of different styles, 227 base liquors, and 19 auto-blended products.
Ninety Fenjiu samples of different grades were randomly selected, and their sensory scores were given by ten expert spirit tasters in Fenjiu Group according to their enterprise standard (Table 1b).
To demonstrate the flavor difference among the same grade products, three styles products (A, B, and C) produced in the same year but in difference months were selected (Table 1c). Three different batches of the base liquors were obtained for their difference evaluation (Table 1d). These base liquors will also be used in automatic blending experiments. Nineteen samples were taken from finished liquor tank of automatic blending system.

All samples were supplied by Shanxi Xinghuacun Fenjiu Group
Limited Liability. Mean value of three measurements was used for analysis.

| Flavor measurements
Ten state-level Chinese spirit tasters were selected from the tasters' team in Fenjiu Group to form a panelist, with five males and five females. Sensory scores of 1-100 in six aspects were graded according to Fenjiu enterprise standard which is listed in Table 2. The final grades were the average of ten results.
An ISQ GC-MS (Trace1310-ISQ LT, Thermo Fisher Scientific) was used to identify the flavor chemicals in Fenjiu finished samples, as well as in the base liquors. The parameters were set as follows: The column temperature was 40°C at the beginning, raised to 250°C at a rate of 40°C/min, then raised to 300°C at a rate of 6.5°C/min, and held for 18 min. The flow rate of carrying helium gas was 1.3 ml/min at the beginning, raised to 2 ml/min at a speed of 1 ml/min, and then kept for 15 min. The PTV sampling temperature was 40°C at the beginning, raised to 320°C at a rate of 14.5°C/min, held for 27 min, then raised to 350°C at a rate of 14.5°C/min, and held for 3 min. An EI ion source was adapted, and its temperature was set to 230°C.
A zNose™ (Electronic Sensor Technology) was used for fast flavor detection, which can detect a flavor in 1 min and send the detected signals to a PC instantly. The signals were illustrated in the form of multipeaks, which was similar to a GC, and could also form a fingerprint for a specific flavor. Its operation parameters were set as follows: senor detection temperature at 60°C, sensor baking temperature 150°C, and column temperature was raised from 40 to 180°C at a rate of 10°C/min. The flow rate of helium gas was 0.03 ml/s. Details of zNose™ can be found in literatures (Gan, Man, Tan, NorAini, & Nazimah, 2005;Li, Wang, Raghavan, & Vigneault, 2011).

| Mathematical methods
Pearson correlation coefficient (SPSS 19.0, IBM) was analyzed for the peak areas of GC-MS and zNose™. A t test was conducted with r and p values reported; multilinear line regression (MLR) was employed to model the sensory scores with zNose™ peaks; analysis of means (ANOM) was applied to base liquors where means (MN) and coefficient variable (CV) were used to select the representative peaks for automatic blending.
Principle component analysis (PCA) of zNose™ peaks was conducted for differentiation of the Fenjiu samples. The new PCs were analyzed and displayed in three-dimensional scores plots in this study. The main purpose of PCA is dimension reduction to obtain a few new variables that can represent the original variables without loss of information as much as possible (Ezhilan, Nesakumar, & Babu, 2018;Ickes & Cadwallader., 2018;Jiang, Wang, Wang, & Cheng, 2017;Verma & Panda, 2018;Zhang et al., 2017). Further, analysis of means (ANOM) was conducted for significant difference from the overall mean. Mean value (MN), standard deviation (SD), and variable coefficient (CV) were used as the criteria (Equation 1), where CV <0.15 was considered as a small variation, 0.16-0.35 as medium, and larger than 0.36 as strong variation (Brown, 2011). Large CV reflected a large difference among different base liquors.
x: represented the peak areas, N: the total peak numbers.
A three-layered feed-forward back propagation neural network (BPNN) algorithm (MATLAB, 7.0) was constructed with an architecture of 3 × 5 × 3, by trial and error. It was first used to build mathematical models between spirit tasters and zNose™ and then to adjust the proportions of base liquors in automatic blending operation. BPNN is a supervised learning algorithm which needs training and validation. Weights of neurons are adjusted to minimize the mean square error between the predicted and desired output values. This method is considered to be best suited for agricultural products where nonlinear relationship and multivariable problems without regularity need to be addressed.

| Automatic blending system
To improve the uniformity of Fenjiu products, an automatic blending system was designed. Its schematic diagram is shown in Figure 1. LabVIEW data acquisition card (NISB6001, NI) and software were used for the control unit. MATLAB was employed for BPNN calculation, and the results were transferred to LabVIEW for process control. The details of the system can be found in our previous study (Deng, 2018).
A Fenjiu B-style product was selected as the "Target Fenjiu" in this study. Its flavor was detected with zNose™, and the flavor fingerprint was stored in the flavor characteristic map. In blending process, all base liquors and blending Fenjiu flavors were detected in turn. The control objective was to approach the target fingerprint flavor through online adjustment of the base liquors' flow rates.
When the target flavor was approached, the blending process finished. The blending flow chart is shown in Figure 2. Three common peaks of target Fenjiu and base liquors detected with zNose™ were selected for BPNN algorithm.

| Correlation analysis of GC-MS, spirit tasters, and zNose™
The correlation of three flavor measurement methods (GC-MS, spirit tasters, and zNose™) would be analyzed through the different grades' 90 Fenjiu samples (Table 1). Different regression models would be built for the purpose of mutual replacement of each other, as only zNose™ could be used for fast flavor detection in automatic blending system. The same experiments were conducted for all other samples with similar results, but the report is neglected.

| Spirit tasters versus zNose™
The detected peak areas by zNose™ were also correlated with the sensory scores. MLR and BPNN were employed to build regres-

| Flavor difference analysis of Fenjiu products of three styles
The flavor difference of Fenjiu products would be investigated. The Fenjiu samples of three styles (A, B, and C) produced in the same year but in different months were included (Table 1). Although zNose™, spirit tasters, and GC-MS analysis were all conducted in our experiments, only zNose™ data are reported here, as it had been proved to have acceptable correlation with the other two methods.
The nine peak area values detected by zNose™ from A, B, and C were analyzed with PCA, and variation explanation rates are shown in Table 5. It can be observed that the first three variables can reflect the most difference of the initial nine variables. Score plots were graphed with PC1, PC2, and PC3 as three axes and shown in Figure 5. A sphere center was constructed where the summed distances to the center from each point was the minimum. A spherical surface was further graphed around the sphere center with a fixed distance (Fan et al., 2018 Hence, the uniformity and homogeneity of the same style's Fenjiu products need to be improved for better products.

| Flavor difference analysis of base liquors
Consistent quality of the product would be more recognized, while the diversity and complexity of raw materials, workers' experience, equipment, and aging condition cause the flavor differences of the base liquors. Three base liquors, called Dazongjiu, Dajiu, and Daijiu (Table 1), were used for blending Fenjiu B-style products; hence, they were detected with zNose™. Their respective nine peak areas were analyzed with PCA to investigate the difference (Figure 6), as well as ANOM for analysis of MN and CV.
The difference of Dazongjiu samples produced in three batches was evaluated with PCA ( Figure 6a)   Analysis of means showed that by means of MN and SD, C7, C9, and C13 had large variance (Table 7). This implies that the three peaks can be used in automatic blending system. If their difference could be minimized, the difference existed in different Fenjiu products would be decreased in a significant degree.

| Automatic blending
The automatic blending system would be used to improve the uni-

| System operation
In the above Section 3.3, ANOM analysis of Dazongjiu, Dajiu, and Daijiu showed that they had different variances but with the three common, with large variated peaks of C7, C9, and C13. Hence, these three peaks would be used as control parameters in the automatic blending system to improve the product uniformity. Less parameters also simplified the system and speed up the control. The steps were as in Figure 2, and the operation repeated until predefined error was reached. Here, the 5% error was used as it was defined by Fenjiu Group in artificial blending operation. Total 19 blended products were obtained for test.

| zNose™ test of blended products
The zNose™ detected nine peak areas of the 19 blended products and shown in Figure 7 with standard deviation. Among them, C7, C9, C10, and C13 had large variations.
A PCA was again conducted for the 19 blended products together with 85 Fenjiu B samples. The first three variables explained 90.34% difference and almost covered all the initial information of nine peaks. PC1, PC2, and PC3 contributed to 70.70%, 12.04%, and 7.64%, respectively. Figure 8 showed the three-dimension score plots of PCA. A sphere surface was graphed whose center was used as the reference point of Fenjiu B. Those inside the sphere were homogeneous samples, and those outside were heterogeneous samples. The results showed that within the 19 blended products, 16 were inside and three were outside. Hence, the dispersion ratio was 15.79%, which was much lower than the 42%, 56%, 40%, and 50% of March, June, October, and December samples. The lower dispersion ratio means less difference to the "Target Fenjiu." The three heterogeneous points might be the error caused by the blending system.
If not counting the three heterogeneous points and comparing the 16 samples with "Target Fenjiu," it can be observed that those peaks of C7, C9, and C13 which had large variation previously now have low variation, and the variation is below ±5%.

| GC-MS measurements of the blended products
GC-MS was also used to check the 16 samples. Total 26 flavor chemicals designated by Fenjiu Group were used for analysis and shown in Table 8. It can be observed that the variation was small for every flavor chemical, which indicated that the automated blending system can improve the blending results in terms of flavor quality.

| Sensory evaluation score of the blended products
Sensory scores were used to evaluate the 16 samples and to compare with 85 finish samples (Table 9). The sensory score of the blended products was 81.5, higher than other products. At the same time, the variance was lower than those of other products. So, the blended products had high flavor evaluation and consistency.  For different style products, which were produced in the same year but in different months, zNose™ were used to detect their difference. PCA showed that their dispersion ratio was all more than 40%. Results showed that Fenjiu A-style, B-style, and C-style, although considered as the same products by the producer, have large difference among them.
Three base liquors of Fenjiu B-style were detected by zNose™.
PCA was used for their difference analysis, and ANOM was used for character peaks selection. All three base liquors have variation in their own group. Three peaks of C7, C9, and C13 were found to be the common and large variables by comparison of their MN and CV, and selected as the character peaks for automatic blending. An automatic blending system was designed for Fenjiu uniformity improvement. zNose™ was employed for fast online flavor detection; C7, C9, and C13 were selected from nine peaks as the character peaks. Three base liquors of Dazongjiu, Dajiu, and Daijiu, as well as pure water, were used to blend a "Target" Fenjiu. Based on BPNN, the "Target" Fenjiu was obtained with an error within 5%.
The blended product results were evaluated by zNose™, GC-MS, and spirit tasters all showed improved of uniformity.

ACK N OWLED G M ENTS
This research was supported by Jiangsu province union innovation funds-prospective joint research project (BY2016022-10) and Open

Fund of Beijing Advanced Innovation Center for Food Nutrition and
Human Health (20181039).

CO N FLI C T O F I NTE R E S T
The authors have no conflicts of interest.

E TH I C A L A PPROVA L
This study does not involve any human or animal testing.