Method for monitoring baking degree of large yellow tea by using functionalized dye to construct CSA coupled hyperspectral imaging technology

Abstract

The invention discloses a method for monitoring the baking degree of large yellow tea by using a CSA coupled hyperspectral imaging technology constructed by a functionalized dye, a new CSA is constructed by modifying an original dye by using PSN and MOF porous materials, and the new CSA is coupled with an HSI system so as to further improve the discrimination ability of a model. In the established baking degree model, the CNN shows the highest discrimination rate, and the discrimination rate of the CNN model based on PSN / MOF (at) N-TPP reaches 100%, which is superior to that of the CNN model based on TPP (87.5%). The result shows that the provided method can effectively improve the monitoring precision of the LYT baking degree.

Description

Technical Field

[0001] This invention belongs to the field of sensors, specifically relating to a method for monitoring the roasting degree of Huangda tea using CSA coupled hyperspectral imaging (HSI) technology constructed with functionalized dyes. Background Technology

[0002] LYT (Yellow Tea) is a high-yield variety of yellow tea, belonging to the six major tea categories. Its unique roasted rice-like aroma and health benefits have made it popular among consumers. The roasting process (commonly known as "lao Huo") is a key factor in LYT processing, enabling it to develop its distinctive aroma.

[0003] Roasting levels can be categorized as low (SF), medium (MF), and high (OF), each producing distinct aroma characteristics. Due to the formation of heterocyclic compounds in the Maillard reaction, OF-LYT exhibits a stronger roasted and nutty flavor. It is noteworthy that different consumer groups have varying preferences for roasting levels. Younger and middle-aged consumers prefer the sweeter taste of SF-LYT, while older consumers favor the smoky, even burnt, aroma of OF-LYT. Therefore, monitoring the roasting level during LYT processing is crucial for ensuring the product meets consumer demands and fostering a thriving market.

[0004] Currently, the identification of LYT roasting degree mainly relies on sensory evaluation, which lacks objectivity and cannot provide accurate, data-based evaluation results. Furthermore, gas chromatography-mass spectrometry (GC-MS) is frequently used to analyze volatile components in tea leaves for qualitative and quantitative analysis. However, this method is time-consuming and requires skilled operators. Aroma-based non-destructive testing techniques have been widely used for food flavor detection. Electronic noses (E-nose) have been successfully applied to detect the roasting degree of tea leaves. However, the results are easily affected by humidity, failing to meet the needs of online monitoring during tea processing. In contrast, colorimetric sensor arrays (CSA) have proven to be a simple, rapid, and sensitive method for vapor and gas analysis. Compared to previous studies, this method provides rapid, visual, and non-destructive assessment. Selecting dyes for CSA is a key aspect of its construction. Axial coordination of TPP materials, exhibiting a significant color response under ligand action, makes them relatively ideal dyes for CSA arrays.

[0005] Compared to monomers, nanoparticles exhibit smaller size effects, quantum tunneling effects, and surface effects. Utilizing nanostructures to surface-modify array dyes can effectively improve the sensitivity and chemical activity of the sensing array. Using porphyrin-coated single-walled carbon nanotubes in the sensing array can enhance the rapid detection of aliphatic and aromatic volatile organic compounds (VOCs) at concentrations as low as 1 ppm at ambient temperatures. In monitoring the solid-state fermentation of aged vinegar, N-TPP showed a stronger signal response to VOCs compared to TPP. In our previous study, we investigated the performance of N-TPP in assessing the fermentation quality of black tea and demonstrated that it significantly improves the accuracy of monitoring black tea fermentation compared to TPP.

[0006] Modifying array dyes with specific nanomaterials is a promising strategy to further improve the sensitivity of sensor arrays. Novel MOF nanomaterials, composed of inorganic metal nodes and organic linkers, possess a porous structure that can increase the contact time and collision probability between gas molecules and the sensor, thereby further enhancing the detection sensitivity of the sensor array. Liu et al. designed a paper-based colorimetric sensor array using Cu-MOF as the dye to distinguish wheat with different mold growth rates based on volatile mold markers. Furthermore, modifying array dyes with PSN is another method to improve sensor sensitivity. PSN encapsulates functional nanoparticles and small organic compounds, forming a rough, porous reaction surface, thereby increasing the reaction area and improving sensitivity. Lin et al. constructed a novel colorimetric sensor for the rapid and accurate detection of early mold growth in wheat by preparing PSN-modified boron dibromide (NO2 BDP) pigment.

[0007] To date, there are no reports on using CSA (Chemical Acid Saturation) to assess the roasting degree of tea. It is worth noting that the changes in aroma quality during roasting are dynamic and complex, placing higher demands on the sensitivity of dye responses. The feasibility of improving the performance of CSA in monitoring LYT roasting degree through nano-sizing and functionalization remains to be explored. Using color information solely based on the CSA sensor to assess material properties often fails to provide comprehensive data, while acquiring spectral data from CSA can effectively improve model performance. Previous studies have shown that combining CSA with the HSI (High-Speed ​​Integrated Stabilization) system can more accurately assess green tea quality. The discrimination effect of CSA is affected by the dye's sensitivity and the dimensionality of information acquired; therefore, in our previous study, we constructed a CSA using self-assembled N-TPP as a dye to assess the fermentation degree of black tea. Although the model discrimination rate of CSA improved due to the nano-effect of N-TPP, the model performance using only the color information obtained from the CSA system was still insufficient. Therefore, we coupled CSA with the HSI system to improve the accuracy of the data dimensionality. In this experiment, we modified the original dye using PSN and MOF porous materials to construct a new CSA, which was then coupled with the HSI system to further improve the model's discriminative power. Summary of the Invention

[0008] This invention addresses the problems existing in the prior art by providing a method for monitoring the roasting degree of Huangda tea using a CSA-coupled hyperspectral imaging technique constructed with functionalized dyes. Specifically, it proposes a novel method for determining the roasting degree of Huangda tea using a new colorimetric sensor based on nano-modification and PSN / MOF porous material-modified TPP dye coupled with HSI. Among the established roasting degree models, CNN shows the highest discrimination rate. CNN models based on PSN / MOF@N-TPP all achieve a discrimination rate of 100%, which is superior to the TPP-based CNN model (87.50%). The results indicate that the proposed method can effectively improve the monitoring accuracy of LYT roasting degree.

[0009] This invention utilizes functionalized dyes to construct a method for monitoring the roasting degree of Huangda tea using CSA coupled hyperspectral imaging technology, comprising the following steps:

[0010] S1: Dissolve the functionalized dye in NN-dimethylacetamide solution to prepare a 2 mg / mL solution, sonicate for 30 minutes, and then store in the dark;

[0011] S2: A C2 reverse-phase silica gel plate (MerckKGAA, Frankfurt, Germany) was selected as the solid substrate for the visualization sensor array. 5 μL of the functionalized dye solution prepared in S1 was added dropwise to the C2 reverse-phase silica gel plate, and dried in a fume hood for 14 minutes to obtain CSA for subsequent experiments. In the specific experimental process, this invention fabricated a 4×6 visualization sensor array, obtaining a total of 24 array points (…). Figure 1 The first row has 4 TPP dye dots, the second row has 4 N-TPP dye dots, the third row has 4 PSN@TPP dye dots, the fourth row has 4 PSN@N-TPP dye dots, the fifth row has 4 MOF@TPP dots, and the sixth row has 4 MOF@N-TPP dye dots.

[0012] S3: Yellow tea samples with different roasting levels and the CSA obtained in S2 were placed in dry petri dishes and reacted in an oven at 65℃ for 6 minutes. Six to eight parallel experiments were conducted for each tea sample. Hyperspectral images of the CSA before and after the reaction were acquired using a hyperspectral information acquisition system. During system operation, the colorimetric sensor array was placed on a stage that moved laterally in parallel at a speed of 2.44 mm / s and scanned within the 550-900 nm wavelength range. The images obtained after scanning were corrected. Finally, ENVI 4.7 (ITT Visual Information Solutions, Boulder, CO, USA) was used to extract the spectral information of each pixel of the stained points in the images, and the average spectral value was obtained.

[0013] S4: Employing a Convolutional Neural Network (CNN) algorithm, the weight parameters and bias terms of the convolution kernels are first initialized, and the number of convolutions is set to 500. The spectral data obtained in S3 is input, and features are extracted using convolution operations until the model training is complete. The trained CNN model can then be used to predict and classify new input data.

[0014] The functionalized dyes were prepared by the following method:

[0015] Step 1: Synthesis of N-TPP

[0016] TPP was dissolved in DMAC, then the surfactant polyethylene glycol-600 (PEG600) was added. The mixture was stirred for 15 minutes and allowed to stand in a cool place to obtain 7×10⁻⁶ ppm. -4 mol / L N-TPP solution;

[0017] The TPP is selected from compounds with the following structures:

[0018]

[0019] Step 2: PSN Synthesis

[0020] PSNs were synthesized using a reverse microemulsion method. Briefly, 65 μL of TEOS, 70 μL of 28 wt% ammonia, 9.50 mL of cyclohexane, and 500 μL of ultrapure water were mixed, followed by the addition of 200 μL of the surfactant Triton X-100 and stirring overnight. Then, 65 μL of TEOS and 50 μL of LAPTS were added to the reaction mixture, and stirring was repeated overnight. The resulting white powder was collected by centrifugation, washed, and dried to obtain PSNs.

[0021] Step 3: Synthesis of MOFs

[0022] At room temperature, 11.57 g of 2-methylimidazole and 2.03 g of Zn(CH3COO)2 powder were mixed and dissolved in 93.44 mL of methanol. The mixture was stirred for 10 minutes, allowed to stand, centrifuged, and the precipitate was washed three times with methanol and dried to obtain MOF.

[0023] Step 4: Preparation of Functionalized Dyes

[0024] To remove impurities from the production process, the framework material was placed in a reactor and calcined at 600°C for 4 hours. Subsequently, the dye and framework material were dissolved in DMAC at a mass ratio of 2:1 and mixed to obtain a 2 mg / mL stock solution. PEG-600 was then mixed with the stock solution at a volume ratio of 1:9 (1000 rpm, 50°C, 2 h) to obtain the functionalized dye.

[0025] The framework material is selected from PSN or MOF; the dye is selected from TPP or N-TPP. The functionalized dyes obtained by this invention can be abbreviated as PSN@TPP, MOF@TPP, PSN@N-TPP, and MOF@N-TPP, depending on the selection of the framework material and dye. The obtained functionalized dyes were sealed and stored in a refrigerator at 4°C for 12 hours.

[0026] In this study, CSA was prepared by self-assembling and nano-manufacturing of TPP dyes and then modified with PSN or MOF. The results showed that nano-manufactured dyes modified with PSN or MOF improved the response of LYT roasting, with MOF / PSN@dyes exhibiting the most significant performance improvement. Furthermore, we compared the performance of LYT roasting degree discrimination models using ELM, LSSVM, and CNN algorithms, finding that CNN was the best performing model. The CNN model achieved a discrimination rate of 88.75% on the N-TPP prediction set, a 2.5% improvement compared to the TPP model. In addition, the prediction sets for PSN@TPP and MOF@TPP also showed significant improvements, reaching 97.5% and 98.75%, respectively. Most notably, both CNN models based on PSN@N-TPP and MOF@N-TPP achieved a discrimination rate of 100%. The proposed method effectively improves the monitoring accuracy of LYT roasting degree, thus providing a new approach for intelligent processing of LYT. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating the structure and reaction process of the colorimetric sensor of this invention.

[0028] Figure 2 These are RGB response difference images and bubble diagrams of LYT samples to three roasting degrees of TPP and N-TPP dyes.

[0029] Figure 3 These are scanning electron microscope images of (A) N-MnTPPⅠ, (B) N-MnTPPⅡ, (C) PSN, and (F) MOF;

[0030] (D) TEM images of PSN@N-MnTPPⅠ and (G) MOF@N-MnTPPⅠ; (E) EDS images of PSN and (H) MOF.

[0031] Figure 4 The Euclidean distances of the dye dots in each array to the LYT samples at different roasting degrees are: (A) SF, (B) MF, (C) OF; and the average spectral curves based on the responses of the four dyes to the LYT samples are: (D) FeTPPⅡ, (E) CoTPP, (F) MnTPPⅠ, (G) MnTPPⅡ. In Figures AC, A, B, C, and D represent FeTPPⅡ, CoTPP, MnTPPⅠ, and MnTPPⅡ, respectively. Different lowercase letters indicate significant differences in the reaction intensities of different dyes (P<0.05). In Figures AG, 1-6 represent TPP, N-TPP, TPP@PSN, TPP@MOF, N-TPP@PSN, and N-TPP@MOF, respectively.

[0032] Figure 5 It involves performing confusion matrix analysis on the prediction set of the material model.

[0033] Figure 6 These are photos of LYT-SF, LYT-MF, and LYT-OF samples.

[0034] Figure 7 This is a structural diagram of porphyrin materials.

[0035] Figure 8 These are RGB differential response images of the LYT-SF, LYT-MF, and LYT-OF samples. Detailed Implementation

[0036] This invention proposes a highly sensitive dye-based CSA to determine the roasting degree of LYT. The specific method is as follows: (1) Select four types of TPP based on the response to the aroma of LYT; (2) Prepare and characterize self-assembled nano-sized and PSN / MOF modified TPP; (3) Extract the feature information of CSA using hyperspectral imaging technology (HSI) and construct a discrimination model for evaluating the roasting degree of LYT; (4) Compare the effects of ELM, LSSVM and CNN algorithms on the model performance.

[0037] 1. Materials and Methods

[0038] 1.1. Sample collection and sensory evaluation

[0039] A total of 240 commercially available LYT samples with three different roasting levels (SF, MF, OF) were purchased from eight companies in the Dabie Mountains region of Anhui Province (116.33°E, 31.40°N) (Table 1). Ten samples were obtained for each roasting level, and eight parallel samples were obtained for each sample from the same production batch. All samples were processed and prepared in accordance with the Yellow Tea Processing Technical Specification (GB / T 39592-2020). The collected samples were stored in a cold storage at 4°C for later use.

[0040] Samples collected from the experiment were used for sensory evaluation. The sensory evaluation panel consisted of seven professionally trained tea tasters, and the evaluation method referenced the scoring coefficients for yellow tea evaluation factors in the national standard (GB / T 23776-2018). The panel members categorized the roasting degree of the samples into three levels: LYT-SF, LYT-MF, and LYT-OF. The review panel further divided the collected samples into three groups: SF (T1-T10), MF (T11-T20), and OF (T21-T30), as follows... Figure 6 As shown. In terms of color, the LYT-SF sample is yellowish-brown, while the LYT-MF and LYT-OF samples are dark brown, slightly darker than LYT-SF. The differences in shape and color are not significant.

[0041] 1.2. Preparation and self-assembly characterization of functionalized dyes

[0042] 1.2.1. Experimental Reagents

[0043] TPP, tetraethyl orthosilicate (TEOS), polyoxyethylene (10) isooctylphenyl ether (Triton X-100), and 3-aminopropyltriethoxysilane (APTS) were purchased from Sigma-Aldrich Chemicals Ltd. (Shanghai, China). NN-dimethylacetamide (DMAC), 28 wt% ammonia, cyclohexane, Zn(CH3COO)2, 2-methylimidazole, and methanol were purchased from Sinopharm Chemical Reagent Co., Ltd. (Shanghai, China).

[0044] 1.2.2. Synthesis of N-TPP

[0045] The N-TPP dispersion was prepared according to our previous experimental method. TPP was dissolved in DMAC, then the surfactant polyethylene glycol-600 (PEG600) was added. The mixture was stirred for 15 minutes and allowed to stand in a cool place to obtain a dispersion of 7 × 10⁻⁶ N-TPP. -4 mol / L N-TPP solution.

[0046] 1.2.3. PSN Synthesis

[0047] PSNs were synthesized using a reverse microemulsion method. Briefly, 65 μL of LTEOS, 70 μL of 28 wt% ammonia, 9.50 mL of cyclohexane, and 500 μL of ultrapure water were mixed, followed by the addition of 200 μL of the surfactant Triton X-100 and stirring overnight. Then, 65 μL of LTEOS and 50 μL of LAPTS were added to the reaction mixture, and stirring was repeated overnight. The resulting white powder was collected by centrifugation, washed, and dried to obtain PSNs.

[0048] 1.2.4. Synthesis of MOFs

[0049] At room temperature, 11.57 g of 2-methylimidazole and 2.03 g of Zn(CH3COO)2 powder were mixed and dissolved in 93.44 mL of methanol. The mixture was stirred for 10 minutes, allowed to stand, centrifuged, and the precipitate was washed three times with methanol and dried to obtain MOF.

[0050] 1.2.5. Preparation and Characterization of Functionalized Dyes

[0051] To remove impurities from the production process, PSN and MOF were calcined at 600°C for 4 hours. Then, TPP, MOF, and PSN were dispersed in DMAC and mixed to obtain a 2 mg / mL solution. 200 μL of PEG-600 was then added, and the mixture was stirred at 50°C for 2 hours to produce PSN / MOF@TPP. PSN / MOF@N-TPP was also produced using the same method. The composite dyes were sealed and stored in a refrigerator at 4°C for 12 hours.

[0052] 5 μL of the sonicated solution was transferred onto a single-crystal silicon wafer, dried, and then firmly attached to the sample tray. The external appearance of the sample was then characterized using an S-4800 SEM and EDS (Hitachi High-Tech Co., Ltd., Japan). During testing with an HT-7700 TEM (Hitachi High-Tech Co., Ltd., Japan), 20 μL of the sonicated solution was aspirated onto a paraffin film, covered with a copper mesh for 5 minutes, dried, and observed using an electron microscope.

[0053] 1.3. Construction of CSA

[0054] This invention prepares a 4×6 visual sensor array, constructing a total of 24 array points. The first row has 4 TPP dye points, the second row has 4 N-TPP dye points, the third row has 4 PSN@TPP dye points, the fourth row has 4 PSN@N-TPP dye points, the fifth row has 4 MOF@TPP points, and the sixth row has 4 MOF@N-TPP dye points.

[0055] 1.4. Collection of reaction and response information

[0056] Figure 1 The reaction system and reaction data acquisition for CSA are shown. 3 grams of LYT sample and prepared CSA were placed in a dry petri dish and reacted in an oven at 65°C for 6 minutes. Eight parallel experiments were conducted for each tea sample. The response data of CSA were collected using a hyperspectral imaging system (HSI) and a flatbed scanner.

[0057] Images of the colorimetric sensor array before and after the reaction were acquired using an HP Scanjet 4890 flatbed scanner (HP Shanghai, China). During image acquisition, the image was placed face down on the scanner glass at a resolution of 600 dpi before scanning. The acquired images were saved in JPEG format. The images before and after the reaction were input into MATLAB 2014a (Mathworks, Natick, MA, USA) to collect the CSA response information, obtain the RGBHSVLab color difference, and extract variables.

[0058] ΔR=R a -R b

[0059] ΔG=G a -G b

[0060] ΔB=B a -B b

[0061] ΔH=H a -H b

[0062] ΔS=S a -S b

[0063] ΔV=V a -V b

[0064] ΔL=L a -L b

[0065] Δa=a a -a b

[0066] Δb=b a -b b

[0067] Where: a represents after the reaction, b represents before the reaction; the Euclidean distance of the RGB values ​​of each staining point is used as the reaction intensity index, and the calculation formula is as follows:

[0068]

[0069] Hyperspectral images of the CSA before and after the reaction were acquired using a hyperspectral information acquisition system. The system consisted of a visible-near-infrared spectrometer, a CCD camera, a light source, and a motorized transport stage. During system operation, the colorimetric sensor array was placed on a stage that moved laterally in parallel at a speed of 2.44 mm / s and scanned within the 550-900 nm wavelength range. The acquired images after scanning underwent correction processing to eliminate the effects of illumination inhomogeneities and dark current noise.

[0070] R = (ID) / (WD)

[0071] Where: R is the corrected hyperspectral image, I is the original hyperspectral image, and W and D are standard white with a reflectance close to 100% and standard black with a reflectance close to 0%, respectively.

[0072] Finally, ENVI 4.7 (ITT Visual Information Solutions, Boulder, CO, USA) was used to extract the spectral information of each pixel of the stained points in the image, and the average spectral value was obtained.

[0073] 1.5. Multivariate Statistical Analysis

[0074] 1.5.1. Extreme Learning Machine

[0075] ELM is a nonlinear machine learning algorithm based on a single-hidden-layer feedforward neural network. After setting the number of nodes in the hidden layer, the weighting coefficients connecting the input layer and the hidden nodes are randomly assigned. Then, an activation function is selected to obtain the corresponding output weights. After calculating the weights, the ELM model is established, and these weights can be used to predict and classify new input data.

[0076] 1.5.2. Least Squares Support Vector Machine

[0077] LSSVM is a nonlinear intelligent learning algorithm that transforms the input space into a high-dimensional space through a nonlinear transformation defined by an inner product function, and then derives the optimal classification surface in this space. In this experiment, the LSSVM algorithm based on the RBF kernel function is applied, and the parameters gam(γ) and sig are optimized using a grid search method. 2 (σ 2 The optimization problem of LSSVM is solved by minimizing the sum of the regularization term and the error term. Based on the solution to the optimization problem, after calculating the model's weight vector and bias term, an LSSVM model is built to predict and classify new input data.

[0078] 1.5.3. Convolutional Neural Networks

[0079] CNN is a deep learning artificial neural network machine learning algorithm. It has a deep structure with multiple layers of non-linear mapping, including convolutional layers, fully connected layers, and softmax layers. Convolutional operations obtain a feature data matrix and map it to the original data. Then, the resulting vector features of each object are passed to the fully connected layer, and the softmax layer outputs the feature output indicating the baking degree of the LYT samples.

[0080] 1.5.4. Confusion Matrix Model

[0081] The confusion matrix is ​​calculated by comparing the true values ​​with the model's position and classification.

[0082] Accuracy, precision, and sensitivity are the main metrics for evaluating the confusion matrix of a model, and their calculation methods are as follows:

[0083] Accuracy = (TP + TN / (TP + FP + FN))

[0084] Precision = TP / (TP + TN)

[0085] Semsitiity = TP / TP + FN

[0086] When the true value is positive, the model's perceived positive value is denoted as TP, and the model's perceived negative value is denoted as FN; when the true value is negative, the model's perceived positive value is denoted as FP, and the model's perceived negative value is denoted as TN. In the confusion matrix, the three baking levels (LF, MF, and OF) are labeled as categories 1, 2, and 3, respectively.

[0087] 1.6. Software

[0088] Statistical analyses were performed in Origin 2021 (Origin Lab Corp., Massachusetts, USA). Data variable filtering and model building were performed using MATLAB 2014a. Principal component analysis (PCA), hierarchical cluster analysis (HCA), and significance analysis of the data were performed in SPSS 26.0 (IBM, USA). The significance of differences between treatments was tested using Duncan's multiple range test.

[0089] 2. Results and Discussion

[0090] 2.1 Selection of High-Sensitivity Dyes

[0091] We initially selected six dyes with high chemical responsiveness and enhanced response values ​​after nano-sizing: FeTPPⅠ, FeTPPⅡ, FeTPPⅢ, CoTPP, MnTPPⅠ, and MnTPPⅡ. Figure 7 From these dyes, more sensitive dyes were further screened to prepare CSA. The RGB difference obtained from the images before and after the reaction was calculated as Euclidean distance and used to plot a bubble diagram. Figure 2 As shown, the N-TPP array spots are relatively brighter in color compared to TPP. Combined with the bubble size in the bubble diagram, this indicates that N-TPP reacts more strongly to the LYT sample. This observation is due to the larger reaction surface area of ​​the nano-dye array, which improves the dye's reaction sensitivity. However, the color difference between the FeTPPI and FeTPPII array spots and the nano-array spots is relatively small, indicating that their reaction sensitivity was not effectively improved by nano-sizing. Conversely, the nano-scaled FeTPPⅡ, CoTPP, MnTPPⅠ, and MnTPPⅡ array spots are brighter than the original dye spots, indicating that the sensitivity of these four dyes was significantly improved after nano-sizing. Furthermore, according to the bubble diagram, these array dyes exhibit different reactions to different LYT roasting levels, suggesting that CSAs constructed based on these dyes may be able to distinguish between different LYT roasting levels. Therefore, FeTPPⅡ, CoTPP, MnTPPⅠ, and MnTPPⅡ were selected as the gas-sensitive materials for the experiment. All of these materials were commercially available, and their structural formulas are shown in [reference needed]. Figure 7.

[0092] 2.2. Performance of the colorimetric sensor used to monitor the baking degree of LYT

[0093] 2.2.1. Characterization of Nanomaterials

[0094] Besides the magnetic FeTPPⅡ and CoTPP, N-MnTPPⅠ, N-MnTPPⅡ, PSN, and MOF were all characterized by SEM. N-MnTPPⅠ and N-MnTPPⅡ exhibited similar shapes, stacked in elongated bulk form with relatively uniform dimensions of approximately 200 nm. PSN had a capsule-like appearance with a rough surface and dimensions of approximately 300 nm. EDS( Figure 3 E) correspondingly shows the presence of Si and O elements, indicating that these particles are composed of silicon dioxide. The MOF particles have a regular dodecahedral shape, a uniform size distribution between 100 and 200 nanometers, and a relatively rough surface. EDS( Figure 3 H) confirmed the presence of Zn, Au, and O elements in the particles. PSN@MnTPPⅠ and MOF@MnTPPⅠ were characterized by TEM. PSN@MnTPPⅠ particles were relatively uniform in size and had a rough, porous surface. MOF@TPPⅠ particles were morphologically uniform with low internal contrast, indicating the formation of transparent cavities within the crystal. This suggests that the dye was successfully modified by PSN and MOF, encapsulated within the cavities of their respective capsule-type PSN and dodecahedral MOF.

[0095] 2.2.2. Performance Characterization of Colorimetric Sensors

[0096] like Figure 4 As shown in Figure A, N-FeTPPⅡ exhibits higher reactivity compared to FeTPPⅡ. After modification with functionalized materials, MOF@FeTPPⅡ and PSN@FeTPPⅡ show significantly improved reactivity compared to FeTPPⅡ. Furthermore, the reactivity of MOF@N-FeTPPⅡ and PSN@N-FeTPPⅡ is much higher than that of N-FeTPPⅡ. Compared to CoTPP, N-CoTPP modified with PSN and MOF exhibits higher reactivity and demonstrates superior performance. Moreover, from... Figure 4 As can be seen from A, B, and C, the array dyes exhibit different reaction intensities to the three roasting degrees, with the highest reaction observed in the OF sample. This indicates that the array dyes can distinguish the roasting degree of LYT.

[0097] The RGB difference response images of the LYT-SF, LYT-MF, and LYT-OF samples provide similar conclusions. Figure 8The brighter colors of the PSN@dye array spots and MOF@dye array spots indicate that the dyes modified with PSN and MOF have higher sensitivity. Finally, the same array spots exhibit different levels of brightness at the three roasting degrees. The sensing array shows a brighter color after reacting with the LYT-OF sample, indicating that the sensing array can distinguish between the three roasting degrees.

[0098] like Figure 4 As shown in DG, the reflectance spectra of various dyes and their nano-sized and functionally modified dye dots exhibit similar trends, yet differences still exist within a certain range. For example... Figure 4 D. In the 400-600 nm range, the reflectance spectra of MOF@FeTPP II (A-4) and MOF@N-FeTPPII (A-6) almost overlap, with MOF@ exhibiting the highest reflectance. This observation indicates that MOF@ dyes possess the strongest responsiveness. Similarly, in Figure 4 Among the dyes, MOF@N-TPP (B-6) also has the highest reflectance. The reflectance spectra of N-FeTPP II (A-2) and PSN@ dyes (A-3, A-5) almost overlap, indicating similar reaction sensitivities, but the reaction is stronger than that of FeTPP II. Figure 4 F and G also showed similar results in the 500-800 nm range. Overall, the modification of MOFs significantly improved the reactivity sensitivity of the dyes, resulting in optimal performance.

[0099] 2.3. LYT Roasting Degree Discrimination Model Based on Spectral Information

[0100] Using a random partitioning method, the 240 samples were divided into training and prediction sets in a 2:1 ratio. Table 2 shows the results for all models. TPP exhibited the lowest model performance among all dyes. Compared to TPP, all three models based on N-TPP achieved better discrimination rates, according to the confusion matrix ( Figure 5 The accuracy and sensitivity of the N-TPP-based model (D, E, F) have been improved.

[0101] Both PSN@TPP and MOF@TPP models improved the discrimination rate compared to TPP. Among them, the CNN model achieved the best performance, increasing the prediction set discrimination rates of PSN@TPP and MOF@TPP to 93.75% and 96.25%, respectively. This is based on the confusion matrices of the three PSN@TPP models (…). Figure 5In the ELM model (G, H, I), SF samples were incorrectly classified as both MF and OF, a discrepancy likely due to misclassification during training. In the LSSVM model, samples of all three roasting levels exhibited misclassification between adjacent roasting levels. However, in the CNN model, except for two MF samples misclassified as SF, the recall rates for both MF and OF samples were 100%. The deep structure of multi-layered nonlinear mappings and the ability to approximate complex functions endow the CNN algorithm with powerful expressiveness and learning capabilities.

[28] Therefore, CNN demonstrates better training performance compared to ELM and LSSVM. The discrimination rates of PSN@N-TPP and MOF@N-TPP models are significantly improved compared to N-TPP, exhibiting the best performance among all dyes. The discrimination results of ELM, LSSVM, and CNN for dyes are similar to those of PSN@TPP and MOF@TPP, with the CNN model still showing the best performance. Notably, the CNN models based on PSN@N-TPP and MOF@N-TPP achieve a discrimination rate of 100% in both the calibration and prediction sets, and a recall rate of 100% for SF, MF, and OF in the prediction set. Figure 5 The L and R values ​​accurately distinguish the three roasting degrees of LYT. The Euclidean distance intensity index and hyperspectral reflectance images show strong differences in response to different materials, similar to the results of the qualitative discrimination model. Compared to TPP, N-TPP exhibits higher model discrimination power, likely due to its nanoscale effect. Compared to TPP, PSN / MOF@TPP shows better model discrimination results because the porous nanomaterial improves the adsorption performance of the array dyes when bound to VOCs. Compared to the PSN@N-TPP and MOF@N-TPP models, the PSN@N-TPP and MOF@N-TPP models show better model discrimination results, likely due to the better optical properties of the N-TPP-modified dyes.

[0102] 2.4. Discussion

[0103] High-temperature roasting is a crucial step in LYT processing, imparting its unique aroma and smooth texture. Therefore, accurate monitoring of the roasting degree is essential. However, current methods for identifying the roasting degree of LYT primarily rely on the experience of evaluators, lacking objectivity and accurate databases. Guo et al. used GC-MS to analyze the volatile components of LYT at different roasting degrees, finding significant differences in their aromas. The distinctive crispy rice flavor of OF-LYT is due to its high proportion of heterocyclic and aromatic compounds. However, GC-MS technology is time-consuming, labor-intensive, and expensive. Non-destructive testing techniques have the potential to meet the requirements of intelligent and standardized LYT processing, ensuring its flavor. A summary is shown in Table 3, by Song et al.

[10] The roasting process of Tieguanyin tea was successfully monitored using an electronic nose, and a radio frequency model for distinguishing roasting levels was established. However, the electronic nose is susceptible to humidity, has low sensitivity in detecting low concentrations of volatile substances, and struggles to differentiate between similar compounds. In contrast, CSA (Combined Straightening and Analyzing) technology represents a rapid, visual, and non-destructive method for monitoring tea quality and has been successfully applied in recent years.

[0104] However, CSA alone sometimes fails to provide comprehensive information. Multi-sensor data fusion can effectively improve model performance. In our previous study, we combined CSA and HSI systems to evaluate green tea quality. The multi-sensor fusion model outperformed the single-sensor model in both the calibration and prediction sets. Furthermore, the discrimination effect of CSA is affected by the dye sensitivity and the dimensionality of information acquisition. Therefore, in our previous study, we constructed a CSA using self-assembled N-TPP as the dye to evaluate the fermentation of black tea. Although the model discrimination rate of CSA improved due to the nano-effect of N-TPP, the model performance based solely on color information obtained from the CVS system was still unsatisfactory. Therefore, we coupled CSA with the HSI system to improve the accuracy of data dimensionality. In the experiment, we constructed a new CSA by modifying the dye with PSN and MOF porous materials to further improve the model's resolution. Coupled with the HSI system, the discrimination rate of the LYT roasting degree discrimination model based on hyperspectral information reached 100% in both the calibration and prediction sets, and the discrimination results were similar to those of color information (Table 4).

[0105] 3. Conclusion

[0106] In this study, CSAs were constructed by self-assembling TPP dyes into nanoparticles and modifying them with PSN and MOF. The roasting degree of Huangda tea was determined by extracting spectral data using hyperspectral imaging (HSI). The results showed that nano-sized dyes modified with PSN / MOF improved the response to LYT roasting, with MOF / PSN@ dyes exhibiting the most significant performance improvement. Specifically, the prediction sets of PSN@TPP and MOF@TPP were significantly improved, reaching 93.75% and 96.25%, respectively. Most notably, both CNN models based on PSN@N-TPP and MOF@N-TPP achieved a 100% discrimination rate. The proposed method effectively improves the monitoring accuracy of LYT roasting degree, thus providing a new approach for intelligent processing of LYT.

[0107] Table 1. Company Information for LYT-SF, LYT-MF, and LYT-OF Samples

[0108]

[0109] Table 2. Modeling results based on sensor spectral information

[0110]

[0111] nn: Number of neurons in the hidden layer; γ: Regularization parameter; σ 2 : Kernel function sig 2 MaxEpochs: Input feature variable processing.

[0112] Table 3. Comparison between previous studies and our research

[0113]

[0114] Table 4. Modeling results based on sensor color information

[0115]

[0116] nn: Number of neurons in the hidden layer; γ: Regularization parameter; σ 2 : Kernel function sig 2 MaxEpochs: Input feature variable processing.

Claims

1. A method for monitoring the roasting degree of Huangda tea using CSA coupled hyperspectral imaging technology with functionalized dyes, characterized in that... Includes the following steps: S1: Dissolve the functionalized dye in NN-dimethylacetamide solution to prepare a 2 mg / mL solution, sonicate for 30 minutes, and then store in the dark; S2: Select the C2 reversed silica gel plate as the solid substrate for visualizing the sensor array. Add 5μL of the functionalized dye solution prepared in S1 to the C2 reversed silica gel plate and dry it in a fume hood to obtain CSA for subsequent experiments. S3: Place Huangda tea samples with different roasting degrees and CSA obtained in S2 in a dry petri dish and react in an oven at 65℃ for 6 minutes. Set up 6-8 parallel experiments for each tea sample. Collect hyperspectral images of CSA before and after the reaction using a hyperspectral information acquisition system to obtain the average spectral data of the corresponding samples. S4: Using the CNN algorithm, first initialize the weight parameters and bias terms of the convolution kernel, and set the number of convolutions to 500. Input the spectral data obtained in S3, and use convolution operations to extract features until the model training is complete. Then the trained CNN model can be used to predict and classify new input data.

2. The method according to claim 1, characterized in that: The functionalized dyes were prepared by the following method: Step 1: Synthesis of N-TPP TPP was dissolved in DMAC, then the surfactant polyethylene glycol-600 was added. The mixture was stirred for 15 minutes and allowed to stand in a cool place to obtain 7×10⁻⁶ ppm. -4 mol / L N-TPP solution; Step 2: PSN Synthesis PSN was synthesized using a reverse microemulsion method; Step 3: Synthesis of MOFs At room temperature, 11.57 g of 2-methylimidazole and 2.03 g of Zn(CH3COO)2 powder were mixed and dissolved in 93.44 mL of methanol. The mixture was stirred for 10 minutes, allowed to stand, centrifuged, and the precipitate was washed three times with methanol and dried to obtain MOF. Step 4: Preparation of Functionalized Dyes To remove impurities from the production process, the frame material was placed in a reactor and calcined at 600°C for 4 hours. Subsequently, the dye and framework material were dissolved in DMAC and mixed to obtain a stock solution of 2 mg / mL. PEG-600 was mixed with the original solution to obtain a functionalized dye.

3. The method according to claim 2, characterized in that: The TPP is selected from compounds with the following structures:

4. The method according to claim 2 or 3, characterized in that: The framework material is selected from PSN or MOF; the dye is selected from TPP or N-TPP.

5. The method according to claim 4, characterized in that: In step 4, the mass ratio of the dye to the frame material is 2:

1.

6. The method according to claim 2, characterized in that: In step 4, the volume ratio of PEG-600 to the original solution is 1:9.

Images