A total of 31 varieties of Baijiu were purchased from Walmart stores (Wuhan, China). The samples were the same as “Rapid identification of high-temperature Daqu Baijiu with the same aroma type through the excitation emission matrix fluorescence of maillard reaction products” by Chen et al. [15] please refer to Table S1 (https://ars.els-cdn.com/content/image/1-s2.0-S0956713523003389-mmc3.docx ) of the published articles [15]. The authenticity of the Baijiu was confirmed through traceable codes that were affixed to the wine boxes, as well as through Walmart’s purchase records. Before conducting the analysis, the samples were stored at 4°C under controlled conditions. HBT (98%), and zinc acetate [Zn(OAc)₂, 99%] were purchased from Aladdin Biotechnology Co., Ltd. (Shanghai, China). N,N-Dimethylformamide (DMF) was provided by Sinopharm Co., Ltd. (Beijing, China). All chemicals were analytical grade and could be used directly without further purification. The ultra-pure water machine used in the laboratory was bought from Chengdu Suyuan Technology Co., Ltd. (SYS-II-10L, Chengdu, China). Its resistivity is 18.25 MΩ. The spectra of the samples were recorded on the ultraviolet spectrophotometer (UV-1800, Mapada Instruments Co., Ltd., Shanghai, China). A two-pass optical quartz colorimeter with a diameter of 10 mm.

Appropriate amounts of Zn(OAc)₂ and HBT were weighed, dissolved in DMF, and ultrasonicated for 10 min to dissolve them uniformly. Solution of Zn(OAc)₂ at 5 × 10^–5 mol/L and solution of HBT at 1 × 10^–5 mol/L were obtained. Took 100 μL each of HBT and Zn(OAc)₂, 700 μL of DMF, mixed well, and then added 100 μL of Baijiu (without pretreatment) and mixed well to make the solution to be tested (in order to get the best reaction ratio of HBT, Zn(OAc)₂ and Baijiu and got better reaction effect, the three reaction volumes were set to be the same). To ensure the reproducibility of the experimental method, 15 parallels were made for each Baijiu.

After waiting for the reaction of the test solution for 5 min, it was added into a two-pass quartz cuvette with an optical diameter of 10 mm, and the spectra were scanned and recorded with a UV spectrophotometer. Parameter settings: wavelength 260–650 nm, absorbance 0.0–0.8.

Four chemometric methods including random forest (RF), RF regression (RFR), data driven-soft independent modelling of class analogies (DD-SIMCA), and orthogonal partial least squares discriminant analysis (OPLS-DA), were applied to the classification and identification of high-temperature Daqu Baijiu with the same aroma type. RF is an integrated learning algorithm that performs the classification task by constructing multiple decision trees, each of which is trained and predicted independently, and the final result is derived by voting or averaging [16]. The optimal parameters of RF, including nTree and LVs were chosen by out-of-bag (OOB) error rate. RFR is mainly used to predict the content of Maillard products in Baijiu that have made great contribution to classification. It improves the performance and stability of the model by constructing multiple decision trees and averaging or weighted averaging the predicted results of these trees to obtain the final regression result [17]. The one-class classification method is usually a better choice for detecting multiple adulterations because it requires only real samples to build a classification model that can identify any sample that is different from this as adulterated [18, 19]. DD-SIMCA is a combination of PCA and SIMCA for developing a threshold that separates the target class from all other samples. It has certain development potential in the field of food adulteration detection [20, 21]. In this study, a DD-SIMCA model was constructed to test the identification ability of the UV-VIS sensor of HBT combined with Zn²⁺ for adulteration of Baijiu with the same aroma type and different grades. OPLS-DA is a multivariable statistical analysis method, that combines partial least squares regression and orthogonal signal correction techniques to distinguish different groups of Baijiu samples and identify the key compounds that affect the classification of Baijiu through variable importance projection values (VIP) [22, 23].

MATLAB (R2018b, MathWorks.Inc, USA) was used for RF, RFR, and DD-SIMCA. Origin (Pro 9.0, OriginLab, USA) was used to draw the UV spectrogram. OPLS-DA ran in SIMCA14.1 (Sartorius Stedim Data Analytics AB, Sweden).

Baijiu contains nearly two thousand organic compounds such as aldehydes, ketones, acids, esters, alcohols, ethers, and heterocycles. Among these organic substances, there are chromophore groups such as –COOH, –CHO, –C=N that can produce ultraviolet or visible absorption, and there are also chromophore groups such as –OH, –OR, –NO₂ that have no absorption band in the ultraviolet and visible regions, but can undergo n-π conjugations when connected with chromophore groups, which can enhance the absorption peak of chromophore groups or redshift them. The UV-VIS spectrum of Baijiu is caused by the valence electron transition after absorption of ultraviolet or visible light by unsaturated groups in these organic molecules. These organic flavor substances only account for about 2% of Baijiu, and although the content is very small, the content of flavor substances in different flavors of Baijiu is significantly different. Based on this, the UV-VIS absorption spectrum plays a certain role in the identification of Baijiu flavor types.

Figure 1 shows the ultraviolet absorption spectra of three different flavor types of Baijiu Figure 1A: sauce aroma type (11 kinds); Figure 1B: mixed aroma type (10 kinds); Figure 1C: strong aroma type (10 kinds). The results showed that there was only one absorption band (250–300 nm) in all three aroma types, and this band was mainly the absorption peak produced by furfural and other substances. Furfural, also known as 2-furanaldehyde, is a scientific name is α-furanaldehyde, which is derived from the substitution of the hydrogen atom on the furan 2 positions by the aldehyde group. It is chemically active. When continuous wavelength ultraviolet light irradiates the Baijiu sample, the electron absorption energy in the unsaturated bond in the organic compound transitions from the low energy level to the high energy level, forming strong characteristic absorption at 276 nm [24]. Furfural and other organic substances containing unsaturated bonds in Baijiu are mainly produced by the Maillard reaction and microbial metabolism. Affected by factors such as grain quality, environmental water source, fermentation temperature and time, and microbial flora, furfural content in Baijiu with different flavor types is bound to be different [25–27]. The concentration of compounds is directly proportional to the absorbance. According to the spectral results, the content of furfural is the highest in sauce aroma type, followed by that of mixed aroma type and that of strong aroma type is the lowest. The content of flavor substances is related to the unique brewing process of sauce aroma type Baijiu, which is one of the important bases to distinguish sauce aroma type from other flavors.

OPLS-DA uses partial least squares regression to establish a relationship model between supervised variables and sample classes. To achieve the prediction of the sample category. R²Y and Q² are the predictive parameters of the model, which can be considered effective models when they are > 0.5. The closer their values are to 1, the better the fit of the model, which is stable and reliable, and the higher the accuracy. In this study, OPLS-DA was used to classify different grades of Baijiu with the same aroma type. Figure 1D shows that Maotai alcohol and Wuling classic Piaoxiang wine can be separated from other sauce-flavor wines. The brewing technology of Maotai alcohol and Wuling Piaoxiang wine produced since 2009 is quite different from that of traditional Maotai Baijiu. The production cycle of traditional Maotai Baijiu is 7 months, the steamed wine is stored in the warehouse for more than 4 years, and then nearly 200 wines of different wine ages are carefully mixed in the way “the wine is blended with more wine”, which is a special technique that involves mixing Baijiu of different quality, taste, style, and function in a certain proportion to coordinate and balance the “color, aroma, taste and style” of the Baijiu so that the Baijiu can achieve the expected style and quality standards. After five years, Maotai Baijiu has the characteristics of outstanding sauce flavor, mellow body, long aftertaste, and lasting aroma in the empty glass, which is a typical Chinese sauce aroma style [28]. Although Maotai alcohol is also a member of the Maotai family and is of the same origin as Maotai, it continues to innovate on the basis of adhering to the traditional brewing process, adopting the Dragon Boat Festival harvest, double ninth feeding, nine cooking, eight fermentation, and seven distilling, and the production cycle is as long as 5 years. The quality of the sauce wine is more prominent, and the wine body is softer and comfortable [29]. Wuling wine is made of glutinous red sorghum grown in southern Sichuan, with solid-state fermentation and solid-state distillation as the raw material. With the essence of the production process, the original wine is produced according to three typical types of sauce flavor, mellow sweet flavor, and cellar bottom flavor, and different rounds of wine are stored for a long time (more than 3 years). Wuling wine is different from Maotai Baijiu in that it has made innovative improvements in every key point of brewing technology. For example, Wuling wine is 5 degrees higher than Maotai Baijiu in the stage of high-temperature accumulation and high-temperature makes Daqu, enriching the microbial system that can withstand high-temperature and produce fragrance [30]. Figure 1E results showed that Baiyunbian 20 years 42° and Yuquan square bottle wine can be significantly distinguished from other mixed aroma wines. Yuquan wine, because of its unique two-step brewing process of thick sauce and fragrance, has significant differences in flavor content and type from other mixed aroma wines, so the classification and discrimination results are the best. The classification effect of the strong aroma type is the best (Figure 1F), and all 10 types of Baijiu can be distinguished. In summary, UV absorption spectra combined with a stoichiometric pattern recognition algorithm cannot completely distinguish different grades of Baijiu with the same aroma type, so a new strategy needs to be sought.

HBT is a compound with typical ESIPT properties. During proton transfer, enol-type and keto-type tautomerism occur, showing dual-wavelength absorption in the UV-VIS spectrum. As shown in Figure 2, there are two strong absorption peaks and one weak absorption peak in the absorption spectrum of HBT, which are about 287 nm, 335 nm, and 400 nm respectively. In general, HBT is dominated by a stable enol form, and the two strong absorption peaks at 287 nm and 335 nm are represented by π-π* transitions of carbon-carbon double bonds in the enol form [31]. In addition, Zheng et al. [32] concluded that the absorption peak of ketone configuration should be located in the long band at 400 nm by studying the spectral rules of HBT under different polar conditions. According to literature reports [19], Zn²⁺ was easy to coordinate with compounds such as pyrazines containing nitrogen and oxygen, but there was no obvious optical signal change. In this study, the combination of Zn²⁺ and HBT could significantly amplify the spectral signal. Figure 2A shows that after the addition of Zn²⁺ to HBT, the double absorption peak of the enol configuration near 287 nm and 335 nm is significantly reduced, while the absorption of the ketone configuration near 400 nm is enhanced. It was concluded that Zn²⁺ and nitrogen on the 2-molecule HBT heterocyclic ring and oxygen on the hydroxyl group form a complex centered on Zn²⁺, which hindered the ESIPT process, resulting in significant changes in UV absorption of the two configurations (Figure 3). After Baijiu was added to the mixture of HBT+Zn²⁺, the absorption peak of the enol configuration near 335 nm increased significantly, while the absorption peak of the ketone configuration near 400 nm disappeared. Since Baijiu is mostly water and ethanol, the comparison of alcohol and water with different degrees of alcohol showed that the absorption peak of the enol configuration near 335 nm was also significantly increased, while the absorption of the ketone configuration near 400 nm was weakened to different degrees (Figure 2B), which was different from the absorbance presented when Baijiu was added. It might be that water and alcohol formed intermolecular hydrogen bonds with HBT, and the pyrazines and other compounds containing nitrogen and oxygen in Baijiu competed with HBT to coordinate Zn²⁺, which affected the ESIPT process under multiple actions, and finally significantly amplified the spectral signal [19, 31]. It is similar to the UV-VIS spectra shown in Figure 1, all Baijiu has a response under this method, but the difference in the spectrum is very subtle (Figure S1). Therefore, it was difficult to directly identify each group of Baijiu samples from the peak shape and peak position of the spectrum, and chemical information hidden in the UV-VIS spectra should be extracted by the stoichiometric method. The classification model was established to identify 31 kinds of Baijiu samples.

RF is a powerful ensemble learning algorithm based on multiple decision trees, each of which is trained and predicted independently of each other. By averaging the prediction results of multiple decision trees, the risk of overfitting can be effectively reduced, and the overall prediction accuracy can be improved. In order to better evaluate the recognition ability of HBT_Zn²⁺ UV-VIS sensor for high-temperature Daqu Baijiu, a discrimination model of high-temperature Daqu Baijiu was established by using a chemometrics RF algorithm. First, we used the RF constructed by the “TreeBagger” function in Matlab2018b to model the sensor data and set “oobvarimp” in this function to enable the variable importance assessment of samples outside the bag, in which the number of decision trees nTree was optimized from 1 to 100 according to the sensor data for cross-validation. Then, based on the optimized nTree, samples accounting for 80% of the total sample volume were randomly selected as the training set, and the remaining samples were used as the prediction set to build an RF model. Finally, classification and prediction were made according to the accuracy of the training set and prediction set of the model. It could be seen from Table 1 that in the identification of different brands of Baijiu with the same aroma type, the result of the sensor system was better than the original spectrum of Baijiu, compared with the single UV-VIS spectrum of Baijiu, the identification accuracy of this method was improved from 62.37% to 100%. It was successfully used in the identification of high-temperature Daqu Baijiu.

In order to further verify the contribution weight of Maillard reaction products in Baijiu to this classification and discrimination method, the absorbance data of high-temperature Daqu Baijiu under this method and the concentration of Baijiu flavor compounds identified by HS SPME-GC-MS (headspace solid phase microextraction-gas chromatogra-phy-mass spectrometry) were regression analyzed by RFR algorithm. Table 2 shows that this method can accurately predict the contents of Maillard reaction products such as pyrazines and furfural, and the deviation between the real value and the predicted value is less than 1.8746% ± 1.4515%. Further indicating that the Maillard reaction products such as pyrazines were markers affecting the classification of Baijiu.

Method validation is generally evaluated by the performance index and stability index of the model. Model performance is the accuracy of the model prediction. To ensure the high stability and reproducibility of the method, we made 15 batches of each Baijiu sample using the same experimental conditions to obtain a 465 (samples 15 × 31) × 300 (data points) data matrix, which was then cross validated to assess its performance. It divided the data set into subsets, used 20% of the samples as the validation set and the rest as the training set, repeated the process 95 times, and finally averaged the results of the 95 experiments to evaluate the performance of the model. From the results, the prediction accuracy of the four models reached 100%. The regression model relied on the root mean square error (RMSE) to predict, and Table 2 shows that the RMSE of the predicted compound content is less than 1%. The performance of the model was stable, and the prediction accuracy was high, which further indicated the high precision of the experiment and the good data quality.

In order to further explore the recognition ability of HBT_Zn²⁺ UV-VIS sensor system for adulteration of the same aroma type Baijiu, a DD-SIMCA model was constructed. First of all, the ultraviolet sensing spectral data was imported into MATLAB (Baijiu with higher prices in the same aroma type and brand were used as real samples, and others were used as adulterated samples for identification), then run DDSGUI in the command window, input the spectral data of real samples and adulterated samples successively in the pop-up window, and finally obtained the identification diagram of the three flavors. After beautification and modification of the format, the DD-SIMCA results of three fragrances were obtained, as shown in Figure 4. It shows the DD-SIMCA acceptance diagram of the same aroma type adulteration of three kinds of Baijiu. Overall, it could be seen that the same aroma type low-grade Baijiu adulteration of high-grade Baijiu could be accurately identified. The green line was the discriminant line, which referred to the threshold value belonging to the real sample. The real sample was marked as a green dot on the acceptance diagram, which was in the lower left corner of the green discriminant line. Other samples, known as adulterated samples, were marked with red dots in the upper right corner of the green line. Figure 4A shows that MT1 (Maotaifeitian 2019) and MT2 (Maotaifeitian 2018) are real samples, and other sauce aroma type Baijiu are adulterated samples. Figure 4B shows that LJ1 (Qinghualang) is the true sample and other sauce aroma type Baijiu is the adulterated sample. In Figure 4C, BYB1 (Baiyunbian 20 year-53%vol) is the real sample, and the other mixed aroma type wines are the adulterated samples. In Figure 4D, KZJ1 (Kouzijiao 20 year) is the true sample, and the other mixed aroma type wines are the adulterated samples. In Figure 4E, F, and G, LZ1 (Guojiao 1573), WLY1 (Wuliangyebadai), and YH1 (Yanghe Mengzhilan_M9) are taken as real samples respectively, and other strong aroma types are adulterated samples. According to the results, the results showed that the identification accuracy of sauce aroma type could reach 100%. The discriminant accuracy rate of mixed aroma type was above 90%. Strong aroma type was greater than 97%. The reason was that the Baijiu of the same aroma type and brand was similar in brewing process and flavor, and the error of human operation in the experiment might have caused misjudgment. In general, the accuracy of the model was between 90% and 100%, indicating that the array sensor had a good recognition effect on the adulteration of the same aroma type Baijiu.