[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Next Article in Journal
Determination of Methotrexate Using an Electrochemical Sensor Based on Carbon Paste Electrode Modified with NiO Nanosheets and Ionic Liquid
Previous Article in Journal
Investigating Sepsis-Associated Delirium Through Optical Neuroimaging: A New Frontier in Critical Care Research
You seem to have javascript disabled. Please note that many of the page functionalities won't work as expected without javascript enabled.
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Chemosensors
Volume 12
Issue 12
10.3390/chemosensors12120265
chemosensors-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Investigation of Microbial Fermentation Degree of Pu-Erh Tea Using Deep Learning Coupled Colorimetric Sensor Array via Prediction of Total Polyphenols

by
Min Liu
1,*,
Cui Jiang
1,
Md Mehedi Hassan
2,
Xinru Zhang
1,
Runxian Wang
1,
Renyong Cao
1,
Wei Sheng
2 and
Huanhuan Li
2
1
College of Tea and Food Science and Technology, Jiangsu Vocational College Agriculture and Forestry, Zhenjiang 212013, China
2
College of Food and Biological Engineering, Jiangsu University, Zhenjiang 212013, China
*
Author to whom correspondence should be addressed.
Chemosensors 2024, 12(12), 265; https://doi.org/10.3390/chemosensors12120265
Submission received: 21 October 2024 / Revised: 9 December 2024 / Accepted: 11 December 2024 / Published: 16 December 2024
(This article belongs to the Special Issue Functional Nanomaterial-Based Sensors for Food Analysis)
Browse Figures
Figure 1
<p>Schematic overview of prediction of optimum microbial fermentation of Pu −erh tea liquid via polyphenol detection using CSA and chemometrics.</p> "> Figure 2
<p>Flow chart illustrating the overall experimental procedure.</p> "> Figure 3
<p>Variation in total polyphenol content throughout the fermentation process.</p> "> Figure 4
<p>CSA images before and after the interaction of the volatile compound of Pu-erh tea liquid and their difference.</p> "> Figure 5
<p>Scatter plot of predicted concentration (mg/mL) vs. measured concentration (mg/mL) in calibration and prediction set for PLS (<b>A</b>), RF-PLS (<b>C</b>), IRIV-PLS (<b>D</b>), GA-PLS (<b>F</b>), and CARS-PLS (<b>H</b>). Spectral variables selected using RF (<b>B</b>), GA (<b>E</b>), and CARS (<b>G</b>).</p> "> Figure 6
<p>Basic setting of CNN architecture in this study (<b>A</b>). Optimized CNN model corresponds to the ideal batch_size and number of epochs for polyphenol prediction (<b>B</b>). Scatter plot of predicted vs. measured concentration (mg/mL) for CNN (<b>C</b>) in both calibration and prediction sets (<b>C</b>).</p> ">
Versions Notes

Abstract

:
The degree of tea fermentation is crucial as it ultimately indicates the quality of the tea. Hence, this study developed a total polyphenol prediction system for Pu-erh tea liquid using eight porphyrin dyes and one pH dye in a printed colorimetric sensor array (CSA) coupled with a convolutional neural network (CNN) during microbial fermentation. Firstly, the Box–Behnken sampling method was applied to optimize the degree of microbial fermentation of Pu-erh tea liquid using the response surface methodology. Under optimized conditions, the polyphenol degradation rate reached up to 66.146%. CSA images were then collected from the volatile compounds of Pu-erh tea-reacted CSA sensors. Subsequently, six chemometric approaches were comparatively investigated, and CNN achieved the best results for predicting total polyphenol content. Therefore, the results suggest that the proposed approach can be used to predict the degree of fermentation by measuring total polyphenols in microbial-fermented Pu-erh tea liquid.

1. Introduction

Tea is the most widely consumed nonalcoholic beverage globally, with an average daily intake of 500 mL per capita, due to its unique sensory features, health benefits, and sociocultural significance. Several studies have shown that tea has preventive and therapeutic effects on oxidative stress-related illnesses such as cancer, cardiovascular diseases, cognitive dysfunction, liver diseases, and type 2 diabetes, thanks to the antioxidant properties of phenolic compounds and caffeine (purine alkaloids), which are key components of tea [1]. Both in vitro and in vivo studies have demonstrated consistent antioxidant activities of tea components [2]. Pu-erh tea is a unique fermented tea made from Camellia sinensis var. assamica (Masters) Kitamura (Theaceae) in the western and southern Yunnan Mountains. It can be categorized into two types, fermented (ripened) and raw (nonfermented) Pu-erh tea, depending on the processing method [3]. Although Pu-erh tea is famous in Yunnan Province, its popularity is gradually increasing, especially in Southeast Asia and Western countries, due to its health benefits, taste, and flavor. Various microbial enzymes related to Pu-erh tea fermentation are produced by over 350 fungal strains, which significantly influence the composition of phenolic compounds [4]. Among these fungal strains, Aspergillus niger is a key species responsible for post-fermentation in Pu-erh tea [5]. During the fermentation process, microorganisms stimulate the aging of Pu-erh tea by enzymatically breaking down the main phenolic components, such as catechins (epigallocatechin gallate, gallocatechin gallate, epigallocatechin gallate, and epicatechin gallate), gallic acid, caffeine, etc. [6]. In the fermentation of Pu-erh tea, polyphenol degradation and transformation driven by microorganisms are crucial for determining the flavor, aroma, and antioxidant properties of the final product [7,8]. A high polyphenol degradation rate indicates active microbial metabolism and a well-progressed fermentation process [8]. Exploring the polyphenol degradation, therefore, not only reflects the progress of fermentation but also guides the optimization of fermentation conditions to enhance the quality of the tea. However, the degradation rate is inherently dependent on accurate measurements of total polyphenol content, as it is calculated based on the change in polyphenol levels during fermentation. Hence, the total polyphenols are a critical quality indicator of Pu-erh tea. Accurately determining the total polyphenols is essential for effectively monitoring the quality of Pu-erh tea throughout the fermentation process.
Conventional ways of assessing tea quality, such as sensory evaluation and instrumental techniques, have their limitations. Sensory evaluation, while useful, is inherently subjective, relying on expert opinions and preferences, which can introduce variability [9,10]. Instrumental methods like high-performance liquid chromatography (HPLC) and capillary electrophoresis have improved precision but are often hindered by drawbacks such as destructiveness, high cost, time-consuming sample preparation, extensive chemical use, and the need for specialized analysts [11,12,13].
In response to these challenges, colorimetric sensor arrays (CSAs) have emerged as a promising nondestructive alternative. CSAs mimic the human sensory system and can detect volatile compounds in food through color changes in color-sensitive dyes. These sensors are constructed on an inert base to immobilize color-sensitive dyes that respond to volatile compounds by changing color, making them effective for analyzing food quality [14]. To analyze CSA data, digital image processing technology is applied to extract RGB values (red, green, and blue) from the dye, which can then be used to qualitatively and quantitatively measure the presence of volatile compounds. However, CSA data are typically large and complex, containing both valuable and redundant information. This is where chemometric approaches come in, providing efficient methods for feature extraction and reducing data complexity. The combination of CSA and chemometrics has proven highly effective in analyzing volatile compounds in the food and agricultural industries [15,16,17,18].
Chemometric techniques have gained significant attention as powerful tools for analyzing spectral data. These methods enable the quantitative prediction of target compounds and improve prediction accuracy by eliminating noise, baseline drift, and redundant variables from the data. Partial Least Squares (PLS) is one of the most widely used chemometric algorithms for CSA data analysis. However, PLS can sometimes retain excessive common and redundant information, which may reduce prediction efficiency. To address this, advanced data selection techniques such as Random Frog (RF), Iterative Retained Information Variables (IRIVs), Genetic Algorithm (GA), and Competitive Adaptive Weighted Sampling (CARS) have been developed. These methods help in selecting the most relevant data for model development, thus enhancing the accuracy of the CSA approach [19]. For instance, Li et al. successfully utilized CSA combined with chemometric methods to dynamically monitor the quality change in green tea during ultrasonic-assisted fermentation [20]. However, Pu-erh tea fermentation differs significantly from green tea fermentation due to variations in microbial strains, fermentation processes, and polyphenol degradation rates. These distinctions necessitate a more specific optimization approach for Pu-erh tea fermentation. In addition, in our previous research, we developed NIR-coupled chemometric techniques to analyze total polyphenol content in Pu-erh tea [21]. While Li et al. and our previous studies demonstrate the effectiveness of combining traditional machine learning methods with NIR or CSA data for model building, such methods often involve complex processes for variable selection and model optimization. These processes can be time-consuming and require careful manual adjustments. In contrast, deep learning techniques, used in this study, particularly convolutional neural networks (CNNs), have been applied to improve the accuracy of quantifying compounds in one-dimensional data, showing great potential for enhancing the performance of CSA-based models [22]. By effectively extracting valuable data from CSA images, CNNs can significantly improve the predictive capabilities of the model.
Given these advancements, this study introduces a dynamic microbial fermentation monitoring approach specifically designed for Pu-erh tea. By integrating CSA with deep learning approaches, particularly CNNs, we present a novel, nondestructive method to predict total polyphenols during fermentation, as shown in Figure 1. The study first optimizes three experimental parameters for Pu-erh tea fermentation—liquid-to-solid ratio, inoculum quantity, and pH—using a response surface methodology. Subsequently, CSA images were obtained from the volatile compounds reacting with Pu-erh tea CSA sensors. Thereafter, various machine learning chemometric methods, including PLS, RF-PLS, IRIV-PLS, GA-PLS, and CARS-PLS, were applied to develop a polyphenol prediction model for Pu-erh tea liquid. In addition to utilizing CNN as a representative deep learning approach, we optimized its parameters to enhance model performance, enabling more accurate variable extraction from CSA images and facilitating a comprehensive comparison with traditional machine learning methods.

2. Materials and Methods

2.1. Reagents and Organism

Phosphate buffer (pH 7.5), PDB medium, ferrous tartrate, potassium sodium tartrate, and iron (II) sulfate were sourced from Sinopharm Chemical Reagent Co., Ltd. (Shanghai, China). C2 inverse-phase silica gel plates, pH indicators, and chemo-responsive dyes (porphyrins) were sourced from Merck KGaA (Darmstadt, Germany) and Sigma-Aldrich Chemical Co., Ltd. (St. Louis, MO, USA). Purified water was obtained from a Millipore water purification system. The Aspergillus niger strain used in the fermentation process was provided by the Shanghai Biological Collection Center.

2.2. Instruments

This study employed a range of equipment, including a pH meter (pHS-25, Shanghai INESA Scientific Instrument Co., Ltd., Shanghai, China), a UV-Vis spectrophotometer (ZF-C, Shanghai Huxi Co., Ltd., Shanghai, China), a Tri-CCD camera (CMLN-13S2M/C, SONY, Tokyo, Japan), a light source (YS-L5017 type, YVSION Co., Ltd., Hangzhou, China), a water bath (HH-6, Shanghai Lichen Bangxi Instrument Technology Co., Ltd., Shanghai, China), an autoclave (DSX-18L-I, Shanghai Shen’an Medical Device Factory, Shanghai, China), a temperature shaker (ZQTY-70E, Shanghai Zhichu Instrument Co., Ltd., Shanghai, China), and a clean bench (HF-1200LC, Likang Biomedical Technology Co., Ltd., Beijing, China).

2.3. Factors of Response Surface Test on Fermentation

The liquid-to-solid ratio, inoculum quantity, and initial pH were identified as the primary factors influencing the degradation of tea polyphenols, with degradation rate serving as the evaluation criterion. A second-order equation was established using the Box–Behnken sampling method to correlate the response values with the experimental factors [22]. In accordance with the Box–Behnken design principles, a combined experimental setup was established, incorporating these primary factors and the evaluation criterion. Data analysis was conducted using Design-Expert 8.0 software. The following equation was used to construct a model that links the degradation of tea polyphenols to these influencing factors:
T e a   p o l y p h e n o l   d e g r a d a t i o n   r a t e = M 1 − M 2 M 1 × 100 %
M1 represents the initial concentration of total polyphenols in mg/mL before fermentation, while M2 represents the concentration of tea polyphenols in mg/mL after fermentation.

2.4. Sample Preparation

The Pu-erh tea extract was fermented by inoculating it with a seed solution of Aspergillus niger. The mixture was then incubated at 28 °C with continuous agitation at 120 rpm for a duration of six days. Samples were taken at specific intervals throughout the fermentation process for subsequent CSA analysis. Specifically, samples were collected at 0, 24, 48, 72, 96, and 120 h to monitor the progression of the fermentation.

2.5. Chemical Analysis of Polyphenols

The tea polyphenol content was determined using spectrophotometry, as outlined in GB/T 21733-2008 [23]. Specifically, 0.5 g of fermented Pu-erh tea liquid was mixed with 4 mL of water and 5 mL of ferrous tartrate in a test tube. The mixture was thoroughly blended, and the pH was adjusted to 7.5 using phosphoric acid. A reference solution was prepared in the same manner, substituting water for the fermented Pu-erh tea liquid. The absorbance of the tea extract was then measured at a wavelength of 540 nm using a 10 mm cuvette, with the reagent blank used as the reference. The total tea polyphenol content was calculated using the equation as below:
M = A 1 − A 2 × 1.957 × 2 × K m × 1000
The total polyphenol content in the tea sample, denoted by M (mg/kg), is calculated using several parameters: A1, the absorbance of the mixture after the reaction; A2, the absorbance of the blank; K, the dilution factor; and m, the sample mass (g).

2.6. Design of CSA

A typical CSA is constructed using color-sensitive compounds (CSCs), which must possess at least one functional group capable of reacting with volatile compounds, exhibit a strong color change upon reaction, and have good hydrophobic properties to minimize interference from water vapor in the sample solution. Considering the above points, the CSA was constructed employing one pH indicator (hydrophobic) and eight porphyrins. Table 1 provide detailed information on the 9 CSCs used in this study. A total of 2 mg of each porphyrin dye and pH indicator was dissolved in 1 mL of dichloromethane and 1 mL of absolute ethanol, respectively, inside a fume hood. Then, 1 μL of each solution was printed on C2 inverse-phase silica gel plate with a size of 3 × 3 cm using a microcapillary tube. After drying for 15 min in a fume hood, the dye became stabilized on the silica gel plate. After that, the array was placed in a nitrogen-flushed glove bag for subsequent use.

2.7. Acquisition of CSA Data

The image of the CSA sensors was captured using a Tri-CCD camera under illumination from a transmission LED light system. A 10 mL sample of microbial-fermented Pu-erh tea liquid was placed in a Petri dish, which was then covered with a lid attached to the CSA using double-sided tape. The CSA was allowed to interact with the volatile compounds from the Pu-erh tea liquid for 10 min, resulting in a color change profile for each sample. The image was then immediately captured using the Tri-CCD camera. Additionally, a control image of the same CSA was taken before it interacted with the volatile compounds. Both the pre-reaction and post-reaction images were transferred to a computer for further analysis. MATLAB R2014b was used to process the images using techniques such as morphological processing (to improve image precision), threshold binarization (for contour magnification), and median filtering (to remove noise). A 15-pixel circle was selected for each point of the CSA as a ROI (region of interest). The grayscale image was composed of three-color components. The average value of the grayscale image was mined from the ROI of every CSA point before and after the interaction with a volatile component of microbial-fermented Pu-erh tea liquid. Then, the obtained averaged grayscale value was subtracted to create color component differences (ΔR, ΔG, and ΔB). After that, ΔR, ΔG, and ΔB were normalized prior to overlaying the grayscale image to obtain the differences in images. Each dye provided three variables and nine dyes of each CSA altogether provided 27 variables (3 color components × 9 points). The extracted variables were then combined with the reference values (total polyphenol content), quantified via the spectrophotometric method in Section 2.5, to develop the chemometric model. Figure 2 shows the general process of the model built.

2.8. Establishment of the Prediction Model

To enable the rapid detection of polyphenols during fermentation, the full dataset PLS model, along with variable selection techniques such as RF, IRIV, GA, and CARS, were employed to create regression models for polyphenol prediction. For comparison, a CNN model was also developed to predict total polyphenol content, providing insights into the effectiveness of deep learning versus traditional approaches.
PLS is a statistical technique that utilizes the full dataset to address multicollinearity issues by integrating aspects of canonical principal component analysis, multiple linear regression, and correlation analysis. However, since PLS includes both relevant and irrelevant data, its predictive accuracy may be compromised [24,25,26]. Nevertheless, PLS is commonly used as a baseline model for comparison with other methods.
RF is a widely used chemometric variable selection method that operates iteratively. For regression analysis, RF was combined with PLS. In essence, RF mimics the search behavior of a group of frogs in a wetland, where each variable extraction probability is measured through a reversible jump Markov Chain Monte Carlo approach, maintaining a steady-state distribution in the model space [27].
IRIV is a variable selection approach, which follows the possible interaction amid variables by random arrangement. It can evaluate variables related to the target compound. The selected variables can be grouped as strong, weak, uninformative, or interfering variables. In every iterative round, strong and weak variables are retained until the removal of uninformative and interfering variables has been achieved [28].
Genetic Algorithm (GA) is a nature-inspired stochastic optimization technique based on the principles of natural selection and evolution. It simulates the reproduction and evolution of variables within a dataset, enabling the identification of optimal solutions with high adaptability. Through iterative processes, GA refines the variable set by selecting those that optimize the objective function, ultimately converging on the global optimum solution [29]. In this context, the Partial Least Squares (PLS) method is used to establish the analytical correction model for the filtered variables identified by GA [28].
CARS progressively filters out uninformative variables through PLS regression, aiming to retain only valuable predictors. The approach involves Monte Carlo (MC) sampling of the calibration set to create a PLS model, followed by the removal of low-coefficient variables using an exponential decreasing function (EDF). The sampling process is optimized based on the minimum RMSECV (root means square error of cross-validation) value used as a selection criterion [30].
CNN is a prominent deep learning approach that performs convolution operations on input data. It consists of multiple layers, including convolution, pooling, and fully connected layers, where weights and biases are applied. The network’s weights are updated iteratively using backpropagation and a loss function, enabling the model to learn and refine its predictions through successive iterations. The loss function determines the difference in the value between each output layer. Subsequently, the error backpropagation technique updates the weights by propagating the loss through each layer until it is minimized or falls below a specified threshold [31]. CNN integrates preprocessing, variable selection, and regression into a single approach, allowing for end-to-end training without the need for manual adjustments [22].

2.9. Model Calibration and Validation

Before constructing the model, the dataset was randomly divided into calibration and prediction sets at a 3:2 ratio, with 50 samples assigned to the calibration set and 32 to the prediction set. Model performance was evaluated using a correlation coefficient (Rc) for the calibration set and a prediction coefficient (Rp) for the prediction set, providing insight into the model’s correlation strength for each data partition. Additionally, RMSECV and the root mean square error of prediction (RMSEP) were used to assess the model’s accuracy and reliability [32]. An ideal model should exhibit higher Rc and Rp values, lower RMSECV and RMSEP values, and minimal differences between Rc and Rp, as well as RMSECV and RMSEP [31,32,33]. The models were built using MATLAB R2014b and Python 3.9 in Jupyter Notebook, running on Windows 10.

2.10. Limit of Detection

The limit of detection (LOD) is an essential metric for evaluating the sensitivity and applicability of any analytical method. In this study, the LOD for total polyphenol content in microbial-fermented Pu-erh tea liquid was determined using the following equation, as outlined by our previous study [21]:
L O D = 3 × R M S E P
where RMSEP represents the root mean square error of prediction obtained from the optimal chemometric model used to quantify total polyphenols.

3. Results and Discussion

3.1. Construction of Regression Models and Statistical Analysis

Response surface analysis was conducted using Design-Expert 8.0 to investigate the polyphenol degradation rate, including the liquid–solid ratio, inoculum volume, and pH. The significance of the model was assessed through analysis of variance (ANOVA). The results indicated that the factors influencing polyphenol degradation were ranked in the following order: liquid–solid ratio > pH > inoculum volume, as discussed in detail in our previous study [21].

3.2. Analysis of Response Surface Outcomes

The response surface and contour plots were generated using Design-Expert 8.0 based on the analysis of model variance. These plots illustrate the effects of the liquid–solid ratio, pH, and inoculum quantity on the polyphenol degradation rate (Table 2). The degradation rate of polyphenols was significantly affected by the interaction between pH (X2) and inoculum quantity (X3), with a p value of less than 0.05. According to the results, the optimal conditions for maximizing the degradation rate under the combined influence of pH, inoculum quantity, and solid–liquid ratio were determined to be pH 4.818, inoculum quantity 8.814%, and a solid–liquid ratio of 1:20 (g/mL). At optimal conditions, the model predicted a maximum polyphenol degradation rate of 66.146%. For further studies aimed at controlling the Pu-erh tea fermentation process, the ideal parameters were identified as pH = 5, inoculum quantity = 9%, and liquid–solid ratio = 1/20 g/mL.

3.3. Dynamic Variations in Total Polyphenol Content Under Fermentation

Under the optimal fermentation conditions, the dynamic changes in total polyphenol content are shown in Figure 3. As seen from Figure 3, the total polyphenol contents decreased steadily with the progression of fermentation. Initially, the content was at its highest, approximately 5 mg/L. By day 2, it had slightly declined to around 4.34 mg/L. From day 3 onward, the polyphenol content exhibited a consistent downward trend, reaching its lowest level of about 2.61 mg/L on day 6. This decline is likely attributed to the enzymatic activity of polyphenol oxidase and peroxidase, which facilitate the oxidation and polymerization of tea polyphenols. These enzymatic reactions result in the formation of volatile compounds that contribute to the distinctive flavors and characteristics of the tea.

3.4. CSA Responses

To dynamically monitor the changes in volatile compounds released during fermentation, we utilized a CSA to capture the odorous characteristics of microbial-fermented Pu-erh tea and converted them into visual data. Figure 4 shows the original CSA image, after the interaction of volatile compounds of Pu-erh tea liquid, as the final image of the CSA and the image of their differences from day one to day six. Overall, a total of 82 CSA-RGB feature extraction datasets were collected from 82 samples for further analysis. As can be seen from Figure 4, consistent homologous color features were observed across duplicate experiments, confirming the reproducibility of the CSA system. Furthermore, the unique difference images corresponding to each fermentation day highlight the CSA’s sensitivity to the volatile compounds in Pu-erh tea liquid. Consequently, the use of appropriate chemometric methods holds promise for the in-depth identification of characteristic variables.

3.5. Evaluation of the Performance of Prediction Models

3.5.1. PLS Model Results

First, the PLS was employed on the entire variables of the CSA dataset obtained from fermented Pu-erh tea liquid to predict total polyphenol content. It can be seen from Table 3 that when the number of PCs was nine, the model performs optimally. Figure 5A illustrates the PLS model results for polyphenol prediction. The model yielded an Rc of 0.8497 and an RMSECV of 0.417 mg/mL for the calibration set, while the prediction set achieved an Rp of 0.8632 and an RMSEP of 0.480 mg/mL. Despite these results, the relatively high RMSEP and low Rp values indicate the presence of irrelevant information, which compromised the model’s performance.

3.5.2. RF-PLS Model Results

Following the PLS results, the RF algorithm was implemented to eliminate irrelevant variables, resulting in the development of the RF-PLS model. In this work, RF parameters of the variable selection probability threshold, PCs, variable numbers in the initial model, and the run numbers were set to 0.2, 10, 2, and 10,000, respectively. Variables selected from CSA data by the RF are shown in Figure 5B. In total, RF selected 16 variables from the CSA dataset collected on fermented Pu-erh tea liquid to establish the RF-PLS model for tea polyphenols. The combined scatter plot of the RF-PLS model for total polyphenols is presented in Figure 5C. The model achieved an Rc of 0.8580, Rp of 0.8233, and RMSEP of 0.547 mg/mL with eight PCs (Table 3). In comparison with the full variable PLS model, the Rc of the RF-PLS was improved, but its Rp was not significantly increased, prompting the exploration of alternative variable selection algorithms to further improve prediction accuracy.

3.5.3. IRIV-PLS Model Results

To improve the model performance, IRIV was further employed for variable selection, leading to the development of an IRIV-PLS model for predicting total polyphenol content. The IRIV algorithm was configured with key parameters, including an exponential decreasing function, binary matrix sampling, and a best-to-worst model ratio of k sub-models set to 50, 1000, and 0.1, respectively, leading to the selection of 14 variables from the CSA dataset. The PLS model, built with six PCs, achieved moderate success with an Rc of 0.8268 and Rp of 0.8201 (Figure 5D, Table 3). However, the IRIV technique did not significantly improve the robustness of the PLS model, suggesting that alternative variable selection algorithms could be explored to further enhance prediction accuracy.

3.5.4. GA-PLS Model Results

Subsequently, GA was employed to further refine the variable selection process and identify the most informative variables from the CSA dataset, enabling the construction of a more optimized model. Figure 5E illustrates the GA variable selection process. First, the target for optimization was set to predict tea polyphenol content, and the variables were based on the CSA data of Pu-erh tea fermentation liquid, which were converted into binary data. The GA then generated an initial random population, using RMSECV as the fitness function to guide the optimization process. Through this approach, 10 key variables were selected from the CSA dataset for building the GA-PLS model. The GA-PLS model demonstrated strong performance, yielding an Rc of 0.8596, RMSECV of 0.405 mg/mL, Rp of 0.8566, and RMSEP of 0.474 mg/mL, using seven PCs (Figure 5F, Table 3). Nevertheless, the high Rp and relatively low RMSEP values highlight GA’s effectiveness in selecting relevant variables, leading to improved predictive accuracy for tea polyphenol content.

3.5.5. CARS-PLS Model Results

The purpose of using CARS was to identify key variables that contain essential information about tea polyphenols, enabling the development of a more precise and reliable prediction model. The CSA dataset initially contained 27 variables, of which 15 significant variables were selected using CARS. The CARS parameters were configured with a 5-fold cross-validation and 50 Monte Carlo sampling runs. Figure 5G illustrates the CARS variable selection process. The red asterisk line in Figure 5G(I) marks the number of sample runs associated with the lowest RMSECV value, while Figure 5G(II) shows the regression coefficient path during variable selection, and Figure 5G(III) depicts the number of variables selected by CARS. As shown in Figure 5G(I), the number of selected variables drops sharply in the initial phase of extraction based on EDF. In the second phase, this decline slows considerably, as expected in Figure 5G(II), demonstrating the typical two-step screening behavior of the EDF process (Figure 5G(III)). As sampling runs progressed from 1 to 25, a large number of irrelevant variables associated with tea polyphenols were filtered out, resulting in a decline in RMSECV. Beyond 25 sampling runs, the RMSECV stabilizes until certain important variables are removed, which then leads to a rise in RMSECV (Figure 5G(I)). Accordingly, the model was built using PCs = 9, based on the lowest RMSECV value, as shown in Table 3. Figure 5H highlights the performance of the CARS-PLS model, where the training set produced an Rc of 0.8567 mg/mL and RMSECV of 0.409 mg/mL, while the prediction set achieved an Rp of 0.8503 mg/mL and RMSEP of 0.439 mg/mL. The superior Rp and lower RMSEP values, compared to other models (PLS, RF-PLS, IRIV-PLS, and GA-PLS), indicated that CARS-PLS effectively filtered out noninformative variables, thus improving prediction accuracy for tea polyphenols.

3.5.6. CNN Model Results

To effectively extract valuable variables, a deep learning CNN was applied. The CNN’s architecture is presented in Figure 6A, while Figure 6B outlines how tuning the epoch and batch size influenced model performance, reflected by changes in the Rp value (indicated by a shift from blue to red). In CNN, epoch and batch size are key hyperparameters for building the model [31]. To optimize these parameters, several analyses were conducted, and the Rp value was determined through adjusting epoch and batch size, as exhibited in Figure 6B. After optimization, the CNN model was configured with 80 epochs and a batch size of 6. The final results were impressive, with the model achieving an Rp of 0.8724, RMSEP of 0.396 mg/mL, RMSECV of 0.383 mg/mL, and Rc of 0.8735 mg/mL, as shown in Figure 6C and Table 3. The results indicated that the CNN model provided the most accurate and reliable predictions, outperforming the other models tested.

3.5.7. Overview of Modeling Results

The Box–Behnken design was initially utilized in combination with the response surface methodology to optimize fermentation parameters of Pu-erh tea liquid. Subsequently, to create the effective predictive model for polyphenols during microbial fermentation, six advanced chemometric approaches were systematically applied. These methods included PLS, RF-PLS, IRIV-PLS, GA-PLS, CARS-PLS, and CNN. The goal was to determine the most effective model for capturing the complex relationship between spectral data and polyphenol content, while also identifying the most relevant variables and enhancing prediction accuracy during the fermentation process.
Firstly, PLS was employed, utilizing all variables to assess their correlation with the reference values for total polyphenols. The Rp value (0.8632) was higher than the Rc value (0.8497) in the PLS model with an RMSEP of 0.480 mg/mL, likely due to the presence of irrelevant variables in the CSA dataset, which might affect the prediction capability of the model (Table 3). Following this, RF-PLS was introduced as a variable selection method based on the PLS results. RF selected 16 informative variables, representing 59% of the total dataset. RF is designed to explore a broad solution space by randomly selecting different combinations of variables (or “jumping” to different points in the search space). This randomness leads to a wide search for potentially useful variables, but it can also result in the selection of too many variables, including those that may not necessarily improve model accuracy [27]. Although this method showed a slightly higher RMSEP (0.547 mg/mL) compared to PLS, the difference between the Rc (0.8580) and Rp (0.8233) values was notably larger, likely caused by the exclusion of essential variables, leading to instability in the prediction results.
Next, the IRIV-PLS method was applied to enhance prediction accuracy by selecting the most informative variables. This approach used 14 variables and reduced the number of PCs to six. IRIV aims to select strong variables while filtering out weak or interfering variables [28]. Although the Rc and Rp values did not surpass the PLS and RF-PLS models, the difference between Rc (0.8268) and Rp (0.8201) was smaller, indicating better performance. However, while the RMSEP of 0.527 mg/mL was an improvement over RF-PLS, it was still higher than the PLS model.
To further enhance the prediction model, GA-PLS was applied, which uses a Genetic Algorithm to select variables through natural selection and fitness-based survival. This method resulted in better Rc and Rp values but did not significantly lower the prediction error. GA iteratively selects variables that maximize the model’s predictive accuracy. It progressively filters out irrelevant variables, leading to a more optimized set of predictors [34]. The GA-PLS model was built using ten variables and seven PCs, but while it showed improvement, the RMSEP remained higher than desirable for precise predictions. CARS-PLS was then applied as a more refined variable selection method. Improved Rc (0.8567), Rp (0.8503), and RMSEP (0.439 mg/mL) were obtained with nine PCs. CARS enhances variables selection by iteratively eliminating variables with smaller mean regression coefficients, using adaptive reweighted sampling and an exponential decline function. This process focuses on retaining the most informative variables, improving model accuracy [35]. While CARS-PLS performed better than GA-PLS, the relatively high prediction error hindered the model’s ability to make highly accurate predictions of polyphenols.
Unlike the above machine learning methods, deep learning CNN models do not require an extra variable selection process. Instead, the CNN establishes a direct relationship between the CSA data and the target variable through end-to-end learning. The CNN model, built using the full CSA dataset, achieved the highest Rc (0.8735) and Rp (0.8724) values compared to all other models, with 80 epochs and a batch size of 6. Despite using all variables, CNN achieved the lowest RMSEP (0.396 mg/mL), likely due to its ability to preprocess data and perform automatic feature extraction through deep learning. Ultimately, CNN outperformed RF-PLS, IRIV-PLS, GA-PLS, and CARS-PLS in terms of accuracy and predictive performance. The CNN’s strength lies in its ability to extract and transform informative variables into higher-level representations without needing additional variable selection steps. Moreover, by applying CNN directly to raw CSA data, the model was less affected by spectral preprocessing, resulting in superior accuracy.
All models’ prediction efficiencies were evaluated using RMSEP, RMSECV, Rp, and Rc values, with model accuracy improving in the following order: PLS < RF-PLS < IRIV-PLS < GA-PLS < CARS-PLS < CNN. Based on these results, CNN was chosen as the best model for predicting total polyphenol content, which is crucial for controlling the degree of Pu-erh tea fermentation during microbial processing.
Finally, the CNN model’s outcome was used to determine the LOD for total polyphenol content, which was calculated to be 1.188 mg/mL. In addition, Table S1 shows a comparison of our work with standard methods for polyphenol detection [36,37]. As can be seen in Table S1, the developed CSA method demonstrates comparable performance with the findings reported in the literature. Specifically, our approach not only ensures relatively accurate detection of total polyphenol content but also offers exceptional flexibility and cost-efficiency. Furthermore, this approach enables nondestructive testing, eliminating the need for pretreatment or reagents, making it an ideal choice for real-time quality assessment in tea fermentation.

4. Conclusions

A novel system was designed to determine the dynamic change in microbial-fermented Pu-erh tea through the determination of total polyphenol content based on the response surface results of experimental parameters, such as liquid–solid ratio, inoculum quantity, and pH, using a CSA and a chemometric approach. Initially, the CSA was designed from eight porphyrin dyes and one pH indicator and used for the detection of volatile tea compounds. Specifically, images of the CSA were collected before and after the reaction with volatile compounds present in microbial-fermented Pu-erh tea liquid. The RGB values (red, green, and blue) extracted from these images were then used to develop a chemometric model for total polyphenols. Subsequently, six chemometric approaches, PLS, RF-PLS, IRIV-PLS, GA-PLS, CARS-PLS, and CNN, were applied to comparatively assess total polyphenol content. Among them, the deep learning CNN approach achieved optimum results compared with other chemometric approaches. It resulted in an Rc of 0.8735, Rp of 0.8724, and RMSEP of 0.396 mg/mL. The above results suggested that a combination of CSA and CNN could be employed for monitoring of the dynamic change in microbial-fermented Pu-erh tea through the determination of total polyphenol content as a nondestructive method. Overall, our approach offers an efficient, nondestructive, and real-time method to track the changes in polyphenol content, which can lead to improved control over fermentation conditions and ultimately enhance the sensory qualities and health benefits of the final tea product.

Supplementary Materials

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/chemosensors12120265/s1, Table S1: Comparison of determination of tea polyphenols content in tea.

Author Contributions

Conceptualization, M.L.; Methodology, M.L. and H.L.; Software, W.S.; Investigation, M.L.; Writing—original draft, M.L., C.J., M.M.H., X.Z., R.W., R.C., W.S. and H.L.; Writing—review & editing, M.L., C.J., M.M.H., X.Z., R.W., R.C., W.S. and H.L.; Supervision, R.C. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by [The Jiangsu Vocational and Technical College of Agriculture and Forestry Fund Cultivation General Project] grant number [2024kj34], [Industrial Technology System Integration Innovation Center] grant number [JATS [2023] 338], and [the Jiangsu Modern Agricultural (Tea) Industrial Technology System Yixing Promotion Demonstration Base] grant number [JATS [2023] 048].

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

The raw data supporting the conclusions of this article will be made available by the authors on request.

Conflicts of Interest

The authors disclose no conflicts of interest.

References

  1. Ma, B.; Wang, J.; Xu, C.; Wang, Z.; Yin, D.; Zhou, B.; Ma, C. Interrelation analysis between phenolic compounds and in vitro antioxidant activities in Pu-erh tea. LWT 2022, 158, 113117. [Google Scholar] [CrossRef]
  2. Zhao, C.; Tang, G.; Cao, S.; Xu, X.; Gan, R.; Liu, Q.; Mao, Q.; Shang, A.; Li, H. Phenolic profiles and antioxidant activities of 30 tea infusions from green, black, oolong, white, yellow and dark teas. Antioxidants 2019, 8, 215. [Google Scholar] [CrossRef] [PubMed]
  3. Ma, D.; Pang, Y.; Xie, R.; Luo, J.; Xiao, S.; Wang, J.; Yang, R.; Wang, B. Unveiling metabolite network dynamics during Pu-erh tea storage via non-targeted metabolomics. LWT 2024, 209, 116789. [Google Scholar] [CrossRef]
  4. Xu, J.; Wei, Y.; Li, F.; Weng, X.; Wei, X. Regulation of fungal community and the quality formation and safety control of Pu-erh tea. Compr. Rev. Food Sci. Food Saf. 2022, 21, 4546–4572. [Google Scholar] [CrossRef] [PubMed]
  5. Hlebová, M.; Hleba, L.; Medo, J.; Kováčik, A.; Čuboň, J.; Ivana, C.; Uzsáková, V.; Božik, M.; Klouček, P. Antifungal and synergistic activities of some selected essential oils on the growth of significant indoor fungi of the genus Aspergillus. J. Environ. Sci. Health A 2021, 56, 1335–1346. [Google Scholar] [CrossRef] [PubMed]
  6. Armstrong, L.; do Carmo, M.; Wu, Y.; Esmerino, L.; Azevedo, L.; Zhang, L.; Granato, D. Optimizing the extraction of bioactive compounds from pu-erh tea (Camellia sinensis var. assamica) and evaluation of antioxidant, cytotoxic, antimicrobial, antihemolytic, and inhibition of α-amylase and α-glucosidase activities. Food Res. Int. 2020, 137, 109430. [Google Scholar] [CrossRef]
  7. Dong, C.; An, T.; Zhu, H.; Wang, J.; Hu, B.; Jiang, Y.; Yang, Y.; Li, J. Rapid sensing of key quality components in black tea fermentation using electrical characteristics coupled to variables selection algorithms. Sci. Rep. 2020, 10, 1598. [Google Scholar] [CrossRef] [PubMed]
  8. Ge, J.; Wang, Y.; Li, L.; Shen, S.; Deng, W.; Zhang, Z.; Ning, J. Intelligent evaluation of black tea fermentation degree by FT-NIR and computer vision based on data fusion strategy. LWT 2020, 125, 109216. [Google Scholar]
  9. Li, T.; Lu, C.; Wei, Y.; Zhang, J.; Shao, A.; Li, L.; Wang, Y.; Ning, J. Chemical imaging for determining the distributions of quality components during the piling fermentation of Pu-erh tea. Food Control 2024, 158, 110234. [Google Scholar] [CrossRef]
  10. Li, T.; Lu, C.; Huang, J.; Chen, Y.; Zhang, J.; Wei, Y.; Wang, Y.; Ning, J. Qualitative and quantitative analysis of the pile fermentation degree of Pu-erh tea. LWT 2023, 173, 114327. [Google Scholar] [CrossRef]
  11. Ma, C.; Zhou, B.; Wang, J.; Ma, B.; Lv, X.; Chen, X.; Li, X. Investigation and dynamic changes of phenolic compounds during a new-type fermentation for ripened Pu-erh tea processing. LWT 2023, 180, 114683. [Google Scholar] [CrossRef]
  12. Ma, Y.; Jiang, B.; Liu, K.; Li, R.; Chen, L.; Liu, Z.; Xiang, G.; An, J.; Luo, H.; Wu, J.; et al. Multi-omics analysis of the metabolism of phenolic compounds in tea leaves by Aspergillus luchuensis during fermentation of pu-erh tea. Food Res. Int. 2022, 162, 111981. [Google Scholar] [CrossRef] [PubMed]
  13. Zhang, X.; Zhou, Q.; Liu, Z.; Qing, X.; Zheng, J.; Mu, S.; Liu, P. Comparison of three second-order multivariate calibration methods for the rapid identification and quantitative analysis of tea polyphenols in Chinese teas using high-performance liquid chromatography. J. Chromatogr. A 2020, 1618, 460905. [Google Scholar] [CrossRef]
  14. Chen, Y.; Li, Y.; Lin, L.; Liao, Y.; Fang, H.; Wang, T. Intelligent identification of picking periods of Lu’an Guapian tea by an indicator displacement colorimetric sensor array combined with machine learning. Food Res. Int. 2024, 195, 114960. [Google Scholar] [CrossRef] [PubMed]
  15. Yang, X.; Bi, Z.; Yin, C.; Zhang, S.; Song, D.; Huang, H.; Li, Y. A colorimetric sensor array based on peroxidase activity nanozyme for the highly efficient differential sensing of tea polyphenols and Tieguanyin adulteration. Food Chem. 2024, 432, 137265. [Google Scholar] [CrossRef] [PubMed]
  16. Deng, S.; Zhou, X.; Dong, H.; Xu, Y.; Gao, Y.; Wang, B.; Liu, X. Mellow and Thick Taste of Pu−Erh Ripe Tea Based on Chemical Properties by Sensory−Directed Flavor Analysis. Foods 2022, 11, 2285. [Google Scholar] [CrossRef]
  17. Lin, Y.; Ma, J.; Cheng, J.; Sun, D. Visible detection of chilled beef freshness using a paper-based colourimetric sensor array combining with deep learning algorithms. Food Chem. 2024, 441, 138344. [Google Scholar] [CrossRef]
  18. Zhang, X.; Huang, X.; Aheto, J.; Ren, Y.; Wang, L.; Yu, S.; Wang, Y. Comparable analysis of flavor compounds and quality assessment of fermented bean curd using HS-SPME-GC/MS and colorimetric sensor array. Food Biosci. 2024, 60, 104291. [Google Scholar] [CrossRef]
  19. Xu, X.; Zhang, W.; Huang, J.; Xu, H. Colorimetric sensors for detection of organophosphorus pesticides in food: From sensing strategies to chemometrics driven discrimination. Trends Food Sci. Technol. 2024, 152, 104683. [Google Scholar] [CrossRef]
  20. Li, H.; Hu, Y.; Ma, S.; Haruna, S.A.; Chen, Q.; Zhu, W.; Xia, A. Porphyrin and pH sensitive dye-based colorimetric sensor array coupled with chemometrics for dynamic monitoring of tea quality during ultrasound-assisted fermentation. Microchem. J. 2024, 197, 109813. [Google Scholar] [CrossRef]
  21. Liu, M.; Wang, R.; Shi, D.; Cao, R. Non-destructive prediction of tea polyphenols during Pu-erh tea fermentation using NIR coupled with chemometrics methods. J. Food Compos. Anal. 2024, 131, 106247. [Google Scholar] [CrossRef]
  22. Teng, Y.; Chen, Y.; Chen, X.; Zuo, S.; Li, X.; Pan, Z.; Shao, K.; Du, J.; Li, Z. Revealing the adulteration of sesame oil products by portable Raman spectrometer and 1D CNN vector regression: A comparative study with chemometrics and colorimetry. Food Chem. 2024, 436, 137694. [Google Scholar] [CrossRef] [PubMed]
  23. GB/T 21733-2008; Method for Determination of Tea Polyphenols in Tea Liquid. Standardization Administration of China: Beijing, China, 2008.
  24. El Maouardi, M.; De Braekeleer, K.; Bouklouze, A.; Vander Heyden, Y. Comparison of Near-Infrared and Mid-Infrared spectroscopy for the identification and quantification of argan oil adulteration through PCA, PLS-DA and PLS. Food Control 2024, 165, 110671. [Google Scholar] [CrossRef]
  25. Shen, G.; Kang, X.; Su, J.; Qiu, J.; Liu, X.; Xu, J.; Shi, J.; Mohamed, S. Rapid detection of fumonisin B1 and B2 in ground corn samples using smartphone-controlled portable near-infrared spectrometry and chemometrics. Food Chem. 2022, 384, 132487. [Google Scholar] [CrossRef] [PubMed]
  26. Shen, Z.; Xie, H.; Zhang, J.; Li, M.; Wang, B.; Wu, Y.; Yu, H.; Nie, X.; Hao, J.; Jia, J.; et al. Rapid evaluation of the quality of Epimedium with different processing degrees by E-eye and NIR spectroscopy combined with machine learning. Microchem. J. 2024, 205, 111181. [Google Scholar] [CrossRef]
  27. Wu, X.; Zeng, S.; Fu, H.; Wu, B.; Zhou, H.; Dai, C. Determination of corn protein content using near-infrared spectroscopy combined with A-CARS-PLS. Food Chem. X 2023, 18, 100666. [Google Scholar] [CrossRef] [PubMed]
  28. Gao, S.; Xie, W. SSC and pH prediction and maturity classification of grapes based on hyperspectral imaging. Smart Agric. Technol. 2024, 8, 100457. [Google Scholar] [CrossRef]
  29. Alhijawi, B.; Awajan, A. Genetic algorithms: Theory, genetic operators, solutions, and applications. Evol. Intel. 2024, 17, 1245–1256. [Google Scholar] [CrossRef]
  30. Li, H.; Zhu, L.; Li, N.; Liu, Z.; Wang, L.; Chitrakar, B.; Xu, D.; Cui, Z.; Tang, Y.; Hu, L.; et al. NIR spectroscopy for quality assessment and shelf-life prediction of kiwifruit. Postharvest Biol. Technol. 2024, 218, 113201. [Google Scholar] [CrossRef]
  31. Cacciari, I.; Ranfagni, A. Hands-On Fundamentals of 1D Convolutional Neural Networks—A Tutorial for Beginner Users. Appl. Sci. 2024, 14, 8500. [Google Scholar] [CrossRef]
  32. Nunekpeku, X.; Zhang, W.; Gao, J.; Adade, S.; Li, H.; Chen, Q. Gel strength prediction in ultrasonicated chicken mince: Fusing near-infrared and Raman spectroscopy coupled with deep learning LSTM algorithm. Food Control 2025, 168, 110916. [Google Scholar] [CrossRef]
  33. Li, H.; Sheng, W.; Adade, S.; Nunekpeku, X.; Chen, Q. Investigation of heat-induced pork batter quality detection and change mechanisms using Raman spectroscopy coupled with deep learning algorithms. Food Chem. 2024, 461, 140798. [Google Scholar] [CrossRef]
  34. Li, H.; Nunekpeku, X.; Zhang, W.; Adade, S.; Ahmad, W.; Sheng, W.; Chen, Q. Quantitative prediction of minced chicken gel strength under ultrasonic treatment by NIR spectroscopy coupled with nonlinear chemometric tools evaluated using APaRPs. Food Chem. 2025, 463, 141373. [Google Scholar] [CrossRef]
  35. Zhang, Y.; Zareef, M.; Rong, Y.; Lin, H.; Chen, Q.; Ouyang, Q. Application of colorimetric sensor array coupled with chemometric methods for monitoring the freshness of snakehead fillets. Food Chem. 2024, 439, 138172. [Google Scholar] [CrossRef] [PubMed]
  36. Ding, M.; Yang, H.; Xiao, S. Rapid, direct determination of polyphenols in tea by reversed-phase column liquid chromatography. J. Chromatogr. A. 1999, 849, 637–640. [Google Scholar] [CrossRef] [PubMed]
  37. Luo, X.; Xu, L.; Huang, P.; Wang, Y.; Liu, J.; Hu, Y.; Wang, P.; Kang, Z. Nondestructive Testing Model of Tea Polyphenols Based on Hyperspectral Technology Combined with Chemometric Methods. Agriculture 2021, 11, 673. [Google Scholar] [CrossRef]
Chemosensors 12 00265 g001
Figure 1. Schematic overview of prediction of optimum microbial fermentation of Pu −erh tea liquid via polyphenol detection using CSA and chemometrics.
Figure 1. Schematic overview of prediction of optimum microbial fermentation of Pu −erh tea liquid via polyphenol detection using CSA and chemometrics.
Chemosensors 12 00265 g001
Chemosensors 12 00265 g002
Figure 2. Flow chart illustrating the overall experimental procedure.
Figure 2. Flow chart illustrating the overall experimental procedure.
Chemosensors 12 00265 g002
Chemosensors 12 00265 g003
Figure 3. Variation in total polyphenol content throughout the fermentation process.
Figure 3. Variation in total polyphenol content throughout the fermentation process.
Chemosensors 12 00265 g003
Chemosensors 12 00265 g004
Figure 4. CSA images before and after the interaction of the volatile compound of Pu-erh tea liquid and their difference.
Figure 4. CSA images before and after the interaction of the volatile compound of Pu-erh tea liquid and their difference.
Chemosensors 12 00265 g004
Chemosensors 12 00265 g005
Figure 5. Scatter plot of predicted concentration (mg/mL) vs. measured concentration (mg/mL) in calibration and prediction set for PLS (A), RF-PLS (C), IRIV-PLS (D), GA-PLS (F), and CARS-PLS (H). Spectral variables selected using RF (B), GA (E), and CARS (G).
Figure 5. Scatter plot of predicted concentration (mg/mL) vs. measured concentration (mg/mL) in calibration and prediction set for PLS (A), RF-PLS (C), IRIV-PLS (D), GA-PLS (F), and CARS-PLS (H). Spectral variables selected using RF (B), GA (E), and CARS (G).
Chemosensors 12 00265 g005
Chemosensors 12 00265 g006
Figure 6. Basic setting of CNN architecture in this study (A). Optimized CNN model corresponds to the ideal batch_size and number of epochs for polyphenol prediction (B). Scatter plot of predicted vs. measured concentration (mg/mL) for CNN (C) in both calibration and prediction sets (C).
Figure 6. Basic setting of CNN architecture in this study (A). Optimized CNN model corresponds to the ideal batch_size and number of epochs for polyphenol prediction (B). Scatter plot of predicted vs. measured concentration (mg/mL) for CNN (C) in both calibration and prediction sets (C).
Chemosensors 12 00265 g006
Table 1. Porphyrin materials and pH indicators used in this study.
Table 1. Porphyrin materials and pH indicators used in this study.
NumberColor-Sensitivity MaterialsFormula
15, 10, 15, 20-Tetraphenyl-21H, 23H-porphine iron (III) chlorideC44H28ClFeN4
25, 10, 15, 20-Tetraphenyl-21H, 23H-porphine manganese (III) chlorideC44H28ClMnN4
32, 3, 7, 8, 12, 13, 17, 18-Octaethyl-21H, 23H-porphine manganese (III) chlorideC44H28ClMnN4
45, 10, 15, 20-Tetrakis (4-methoxyphenyl)-21H, 23H-porphine iron (III) chlorideC44H36ClFeN4O4
55, 10, 15, 20-Tetrakis (pentafluorophenyl)-21H, 23H-porphyrin iron (III) chlorideC44H24ClFeF20N4
65, 10, 15, 20-Tetraphenyl-21H, 23H-porphineC44H30N4
75, 10, 15, 20-Tetraphenyl-21H, 23H-porphine copper (II)C44H28CuN4
85, 10, 15, 20-Tetrakis (4-methoxyhenyl)-21H, 23H-porphine cobalt (II)C48H38N4O4
9Bromophenol blueC19H10Br4O5S
Table 2. Variance analysis results for regression models [21]. Reprinted with permission from Elsevier, copyright 2024.
Table 2. Variance analysis results for regression models [21]. Reprinted with permission from Elsevier, copyright 2024.
SourceSum of SquaresDegree of FreedomAverage VarianceF Valuep ValueSignificance Level
Model806.22989.5833.37p < 0.01**
X123.33123.338.69p < 0.05*
X21.8411.840.68----
X3354.901354.90132.22p < 0.01**
X1X25.4215.422.02----
X1X30.2210.220.083----
X2X320.24120.247.54p < 0.05*
X12157.661157.6658.74p < 0.01**
X22161.711161.7160.25p < 0.01**
X3242.52142.5215.84p < 0.01**
Residual18.7972.68------
Lack of fit2.4930.830.20----
Pure error16.3044.07------
X1: pH, X2: inoculation volume, X3: solid–liquid ratio, *: significant, **: extremely significant.
Table 3. Comparison of model results for screening of polyphenols in Pu-erh tea during the fermentation process based on colorimetric sensor array.
Table 3. Comparison of model results for screening of polyphenols in Pu-erh tea during the fermentation process based on colorimetric sensor array.
ModelNumber of VariablesOptimum PCCalibration SetPrediction Set
RcRMSECV (mg/mL)RpRMSEP (mg/mL)
PLS2790.84970.4170.86320.480
RF-PLS1680.85800.4120.82330.547
IRIV-PLS1460.82680.4580.82010.527
GA-PLS1070.85960.4050.85660.474
CARS-PLS1590.85670.4090.85030.439
CNN27--0.87350.3830.87240.396
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Liu, M.; Jiang, C.; Hassan, M.M.; Zhang, X.; Wang, R.; Cao, R.; Sheng, W.; Li, H. Investigation of Microbial Fermentation Degree of Pu-Erh Tea Using Deep Learning Coupled Colorimetric Sensor Array via Prediction of Total Polyphenols. Chemosensors 2024, 12, 265. https://doi.org/10.3390/chemosensors12120265

AMA Style

Liu M, Jiang C, Hassan MM, Zhang X, Wang R, Cao R, Sheng W, Li H. Investigation of Microbial Fermentation Degree of Pu-Erh Tea Using Deep Learning Coupled Colorimetric Sensor Array via Prediction of Total Polyphenols. Chemosensors. 2024; 12(12):265. https://doi.org/10.3390/chemosensors12120265

Chicago/Turabian Style

Liu, Min, Cui Jiang, Md Mehedi Hassan, Xinru Zhang, Runxian Wang, Renyong Cao, Wei Sheng, and Huanhuan Li. 2024. "Investigation of Microbial Fermentation Degree of Pu-Erh Tea Using Deep Learning Coupled Colorimetric Sensor Array via Prediction of Total Polyphenols" Chemosensors 12, no. 12: 265. https://doi.org/10.3390/chemosensors12120265

APA Style

Liu, M., Jiang, C., Hassan, M. M., Zhang, X., Wang, R., Cao, R., Sheng, W., & Li, H. (2024). Investigation of Microbial Fermentation Degree of Pu-Erh Tea Using Deep Learning Coupled Colorimetric Sensor Array via Prediction of Total Polyphenols. Chemosensors, 12(12), 265. https://doi.org/10.3390/chemosensors12120265

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop