A black tea fermentation process real-time detection system

Abstract

The present application belongs to the technical field of black tea detection equipment, and particularly relates to a real-time detection system for fermentation process of black tea. The detection system improves the matching degree between the image features of fermented leaves and the fermentation degree and the accuracy of the model, and most accurately ensures the discrimination efficiency and accuracy of the moderate stage of black tea fermentation. The present application comprises the establishment of a digital classification label of fermentation degree and the establishment of a fermentation degree discrimination model, and the steps are as follows: 1) the establishment of a digital classification label of fermentation degree; 2) the establishment of a fermentation degree discrimination model; 3) morphological recognition of fermented leaves through the fermentation degree discrimination model, so as to determine which stage the fermentation degree of the current fermented leaves is in.

Description

Technical Field

[0001] This invention belongs to the technical field of black tea testing equipment, specifically relating to a real-time detection system for the fermentation process of black tea. Background Technology

[0002] The primary processing steps for black tea are withering, rolling, fermentation, and drying. Fermentation in black tea is mainly the enzymatic oxidation of polyphenols, forming theaflavins, thearubigins, and theabrownins. This process results in a green-yellow-red color change in the tea leaves and is crucial for the characteristic red liquor and red leaves of black tea. Insufficient or excessive fermentation severely affects the quality of the tea. Currently, controlling the degree of fermentation relies on human experience or manual control based on simple process parameters such as temperature, humidity, and time. However, because the color and aroma indicators change little after fermentation reaches a certain stage, coupled with the non-uniform color change of the leaves, especially the slow change in leaf color in the later stages of fermentation, manual judgment becomes highly arbitrary, leading to inconsistent tea quality.

[0003] Currently, black tea processing has achieved continuous production, but due to the lack of precise process monitoring and control, it is impossible to obtain real-time quality information of the finished products, resulting in unstable product quality and inconsistent flavor. This is also a bottleneck for the upgrading of tea processing towards standardization, automation, and intelligence.

[0004] In recent years, domestic and international research has been conducted on the regulation of tea product quality and physicochemical components. However, due to the wide variety and large differences in the content of quality components in tea, and the fact that the formation of quality components is a complex process, differences in internal substrates or external conditions can lead to completely different evolutionary pathways and degrees. It is very difficult to clarify the laws and mechanisms of their influence on quality. Therefore, it is difficult to make quality decisions and regulate tea production using tea quality components in the short term.

[0005] With the development of machine vision technology, it has been widely used in the fields of quality control, defect detection, and grading in food, beverages, and pharmaceuticals. It provides unparalleled quality, precision, and efficiency for food safety, ensuring high-level food inspection and compliance with food safety standards. It also provides a good and feasible opportunity to improve the standardization level and product quality of tea production.

[0006] Chinese invention patent CN104155299A discloses a method and apparatus for determining the fermentation mode of black tea based on hue histograms. It extracts the hue histogram features of the tea product through a computer image acquisition and analysis module. The main peak migrates from right to left during fermentation, and the optimal point for light fermentation is when the hue angle corresponding to the main peak reaches its minimum. The optimal fermentation mode is determined when the main peak reaches its leftmost point 0.5–1.5 hours later; or when the main peak shifts to the right and stabilizes, and the height difference ΔH between the main peak and the minimum hue angle is greater than or equal to 0.3%.

[0007] Chinese invention patent CN104297160B discloses a method and device for judging the fermentation mode of black tea. It sets a standard image and threshold T for the fermentation mode using a computer, and collects and analyzes the histograms of the R, G, and B color components of the image of the black tea product during the fermentation process in real time. It calculates the Manhattan distance between the image and the standard image. When the mean value of the Manhattan distance D is less than the set threshold T, the fermentation mode is judged to be appropriate.

[0008] The inventions published with CN104155299A and CN104297160B require the fermented black tea sample to be evenly spread on the sample pool and image acquisition box in order to acquire and analyze images. They cannot automatically acquire image information in real time under high humidity and high fog conditions, and their application in online rapid non-destructive testing in production lines is still limited.

[0009] Chinese invention patent CN109002855A discloses a method for identifying the fermentation degree of black tea based on a convolutional neural network. The method involves selecting sample images of black tea fermentation captured by a Supervisory Control and Data Acquisition (SCADA) system in a black tea production line as the training set for a CNN; secondly, performing a CNN training process; and thirdly, identifying and determining the fermentation degree of the black tea at the output layer of the CNN. This invention does not mention how to obtain images of fermented leaves.

[0010] The above methods for identifying the degree of fermentation in black tea offer promising applications for online quality monitoring in black tea production and for comprehensively improving the automation level of tea processing equipment. However, none of the above invention patents fully explain how to automatically acquire high-quality fermented leaf images in real time under high humidity and high fog conditions during production. Furthermore, the efficiency, stability, and accuracy of the black tea fermentation degree identification model established based on image classification according to fermentation time still need to be verified and addressed. Summary of the Invention

[0011] The purpose of this invention is to overcome the shortcomings of the prior art and provide a real-time detection system for the fermentation process of black tea. This system improves the matching degree between fermented leaf image features and the degree of fermentation and the accuracy of the model, thus ensuring the efficiency and accuracy of the determination of the appropriate fermentation stage of black tea.

[0012] Therefore, the present invention adopts the following technical solution:

[0013] A real-time detection system for the fermentation process of black tea, characterized by the establishment of digital classification labels for the degree of fermentation and the establishment of a fermentation degree discrimination model, the steps of which are as follows:

[0014] 1) Establishment of digital classification labels for fermentation degree

[0015] a) Conduct batch processing experiments of black tea using a specified number of tea tree varieties, collect tea samples at different fermentation stages, and record human sensory experience information at the same time as collecting samples, including fermented leaf color, fermentation time and degree of fermentation.

[0016] b) Conduct sensory evaluation on tea samples with different fermentation processes, and assess the key stages of tea quality changes during the fermentation process through taste, aroma and leaf residue, so as to divide the fermentation process into six stages: light, slightly light, close to moderate, moderate, close to over, and over.

[0017] c) Verify the results of step b): Detect the catechin content of tea samples at different fermentation stages. It was found that the slope of the change curve at each stage changed in three segments, that is, the catechin content can be divided into three stages of decline as the fermentation process progresses: a rapid decline stage, a slowing decline stage, and a slow decline stage. The sensory quality of tea is best when the catechin content is 12% to 15%. At this time, two nodes where the degree of fermentation changes during the slow decline of catechins were identified. These two nodes include the time from under-fermentation to moderate fermentation and the time from moderate fermentation to over-fermentation. The comparison showed that it was the same as the sensory evaluation results in step b).

[0018] d) Integrate the above information through statistical analysis to establish a digital classification label for the degree of fermentation. Assign values ​​to the fermentation process. The larger the value, the heavier the degree of fermentation. Fermentation starts at kneading and ends at the degree of fermentation at the beginning of kneading. Assign values ​​to the light, slightly light, near moderate, moderate, near over, and over ranges respectively.

[0019] 2) Establishment of a fermentation degree discrimination model

[0020] a) Based on the digital classification labels of fermentation degree established in step 1), the original images of fermenting leaves collected during the fermentation process are divided into six datasets according to the ratio of training set: test set: validation set = 7:2:1: light, light, moderate, moderate, close to excessive, and excessive fermentation degree.

[0021] b) Perform histogram equalization, geometric data augmentation, and color space conversion on the images in each dataset to improve image contrast and brightness and model accuracy. Use the converted image data as one of the final input data for the deep learning model.

[0022] c) Build a MobileNetV3 model based on PyTorch

[0023] The initial part consists of one convolutional layer, which extracts features through a 3x3 convolution, i.e., a convolutional layer, a BN layer, and an h-switch activation layer, which is the same in both Large and Small sizes.

[0024] Middle section: Includes a network structure containing blocks of convolutional layers;

[0025] The final part: the Squeeze operation is omitted, Avg Pooling is brought forward, and 1x1 convolutions are used directly instead of fully connected layers to output the categories; among them, Deptwise Convolution layers and Inverted Residuals structures are added, so that less information is lost after the high-dimensional information is activated by the ReLU activation function, and the h-swish activation function is used instead of the swish function in the structure.

[0026] By inputting image information into MobileNetV3, the intermediate layer feature maps and deep layer feature maps are extracted and superimposed as image patches in the Vision Transformer model. Position embedding is added to the embedding of the image patch, and spatial / positional information is preserved globally through different strategies.

[0027] After the model is built, the data in step b) is used as the input of the initial model to train the model, and finally the trained weights are used to perform the final result inference.

[0028] d) Collect the HSV feature values ​​of images from each fermentation stage and the area of ​​connected components retained after bud and leaf morphology processing. Analyze the data collected by the two methods, and obtain the final color and morphological features through Gaussian filtering. Based on these features, data modeling can be performed, and the results can be predicted based on the model.

[0029] e) By encapsulating the deep learning model and the traditional model established in d), a fermentation degree discrimination model is obtained;

[0030] 3) Morphological identification of fermented leaves is performed using a fermentation degree discrimination model to determine the current fermentation stage of the leaves.

[0031] Preferably, in step 3), the fermentation degree discrimination model is used to perform morphological identification on the fermentation leaves, and the reliability of the current fermentation degree of the fermentation leaves can be output simultaneously.

[0032] Preferably, the original image of the fermenting leaves during the fermentation process is obtained by a real-time image acquisition device for the fermenting leaves; the real-time image acquisition device for the fermenting leaves includes a housing for forming a sealed space, a camera for image acquisition is arranged inside the housing, a defogging window is arranged on the bottom surface of the housing at the lens end of the camera, and a defogging component is arranged inside the defogging window; a sample tray for holding the fermenting leaves is arranged directly below the defogging window, and the sample tray is placed on a base; a straight sleeve-shaped light shield is also vertically mounted on the base, the cavity of the light shield forming a mounting cavity for accommodating the sample tray and allowing the bottom end of the housing to extend into it, and an adjustable brightness light source is arranged inside the light shield; the base, light shield, sample tray and housing are all located inside the outer casing, which serves as a sealed cavity; wherein:

[0033] The sample disk is made of honeycomb ceramic and is prepared using the following method:

[0034] a) Preparation of adhesive

[0035] Mix agar powder and water at a weight ratio of 0.25:100 and heat to dissolve to obtain a solution; after cooling, add 0.25% corn flour, 0.25% sucrose and 0.1% yeast powder per 1 kg of the solution by weight, place at 35-37℃ for 4-5 hours, stirring 2-3 times in between, and pass through a 60-mesh sieve for later use.

[0036] b) Thoroughly mix 45% by weight of Qimen porcelain clay powder, 25% by weight of natural slate She inkstone powder, 15% by weight of activated carbon powder, and 15% by weight of diatomaceous earth powder. Then add a binder and stir to form a slurry, and shape the slurry into a blank. After the blank is shaped, let it stand at 35-37℃ for 6-7 hours, then transfer it into a kiln and fire it at 80-130℃ for 2 hours. Gradually increase the temperature from 130℃ to 450℃ within 2 hours and fire it for 10 hours. Then let it cool naturally to obtain the finished product.

[0037] Preferably, in step b), the Qimen porcelain clay powder is sieved through an 80-mesh sieve, the natural slate She inkstone powder is sieved through a 100-mesh sieve, the activated carbon powder is sieved through a 100-mesh sieve, and the diatomaceous earth powder is sieved through an 80-mesh sieve.

[0038] Preferably, the material of the light shield is the same as that of the sample tray.

[0039] Preferably, the adjustable brightness light source is a moisture-proof and fog-proof surface light source with a size of 32cm×32cm and a central opening size of 7.5cm×7.5cm.

[0040] Preferably, the sample tray has an external diameter of 30cm × height of 7.5cm and an internal diameter of 29.5cm × height of 7cm.

[0041] Preferably, the outer dimensions of the light shield are 35cm in diameter × 28cm in height, and the inner dimensions are 34.5cm in diameter × 28cm.

[0042] Preferably, the defogging window is made of single-sided ITO coated glass with a length of 55mm × width of 50mm × thickness of 1.1mm. An electrode with a resistance of 9 to 10Ω is installed on the inner side of the defogging window, and the entire electrode is coated with conductive silver paste. The electrode is electrically connected to the power supply outside the housing. The electrode constitutes the defogging assembly.

[0043] Preferably, the housing is mounted on a lifting bracket located inside the outer casing.

[0044] The beneficial effects of this invention are as follows:

[0045] (1) Based on large sample image information of different tea tree varieties and different processing periods, the model has higher accuracy and precision.

[0046] (2) The fermentation time required for optimal fermentation varies significantly among different tea varieties and fresh leaves of varying tenderness. Under the same fermentation conditions, some varieties can ferment to the optimal level in 120–150 minutes, while others require 180–240 minutes. Classifying the fermentation process solely based on fermentation time leads to large model errors. This invention integrates human sensory experience information for judging the degree of fermentation, tea sensory evaluation information, catechin residue data, and statistical analysis. By assigning values ​​to the fermentation process, a digital classification label for the degree of fermentation is established, improving the matching degree between fermented leaf image features and the degree of fermentation, and enhancing the accuracy of the model.

[0047] (3) The buds, leaves and tender stems of fresh leaves are subjected to different pressures during the kneading process. The order of cell breakage from high to low is tender stems, buds and leaves. Therefore, the order of fermentation process from fast to slow is tender stems, buds and leaves. Among them, the fermentation of tender stems accounts for about 50 to 60% of the fermentation process time, the fermentation of buds accounts for about 85 to 95% of the fermentation process time, and some areas of leaves always show a green or yellow-green state throughout the fermentation process. Therefore, in order to improve the accuracy of the fermentation degree discrimination model, this invention constructs fermented leaf characteristics from the perspectives of the overall color and individual differences of tender stems, buds and leaves. It is important to extract detailed features such as saturation and brightness of leaves and stems. Through large sample feature statistical analysis, a micro morphological feature group is formed, and a fermentation degree discrimination model is established based on this.

[0048] (4) The fermentation of black tea is mainly a process in which polyphenols are enzymatically oxidized to form theaflavins, thearubigins, and theabrownins. The color of the tea leaves shows a non-uniform trend of green-yellow-red. During a certain period after the fermentation of black tea reaches a certain level, especially during the moderate fermentation stage, the color index changes little, but the brightness and saturation index changes significantly. The fermentation degree discrimination model established by this invention based on color, brightness, and saturation improves the accuracy of the discrimination of the moderate fermentation stage.

[0049] (5) Based on the above solution, the present invention also provides a real-time image acquisition device for fermented leaves. This image acquisition device employs a sealed outer casing, ensuring the airtightness of the internal detection environment and guaranteeing the formation of a high-humidity, high-fog scene. This helps to achieve a stable observation environment for fermented leaves, providing a fundamental guarantee for the accuracy of image information acquisition. Furthermore, by using a defogging component in conjunction with a sealed housing, the camera is isolated from other areas within the outer casing, solving the problem of camera interference and image quality issues caused by fog in high-humidity, high-fog scenes, and ensuring real-time continuous image acquisition. Simultaneously, the sample tray of the present invention uses honeycomb ceramic material, resulting in a rough, porous surface that allows moisture from the fermentation environment to penetrate and continuously provide the necessary moisture for fermentation. Moreover, the present invention uses Qimen porcelain clay powder combined with natural slate and Shexian inkstone powder to form the sample tray, preventing the formation of condensed water droplets that would cause color distortion in the fermented leaves. Ultimately, this further ensures the comprehensiveness and stability of the image information, achieving multiple benefits.

[0050] (6) Moisture-proof and fog-proof surface light source, which can be used well in humid fermentation conditions, can provide uniform lighting within the shooting field of view, and provide effective high-quality images.

[0051] (7) The aforementioned honeycomb ceramics can be used to form sample trays and as raw materials for light shields. Conventional honeycomb ceramics, to ensure high mechanical strength, chemical stability, and thermal shock resistance, generally require the addition of metal oxides and high-temperature firing (above 1500℃). During firing, the boundaries of the raw material particles melt and bond, resulting in smooth pore walls and surfaces. Although this provides good air permeability, it easily causes reflection and water vapor condensation. Practice shows that the honeycomb ceramic material of this invention has a rough, matte surface that does not reflect light from the light source and effectively blocks external sunlight and artificial light, ensuring uniform light within the field of view. It also avoids the mold growth caused by long-term use of wooden materials and the water droplet condensation and reflection problems caused by metal materials, demonstrating significant effectiveness. Attached Figure Description

[0052] Figure 1 This is a simplified structural diagram of a real-time image acquisition device for fermentation leaves.

[0053] The actual correspondence between the reference numerals and component names in this invention is as follows:

[0054] 10-Camera housing; 11-Defog window; 20-Camera; 30-Defog assembly

[0055] 40 - Sample tray; 50 - Base; 60 - Light shield; 70 - Adjustable brightness light source

[0056] 80-Outer casing 81-Lifting bracket Detailed Implementation

[0057] For ease of understanding, this section combines... Figure 1 The specific structure and operation of the present invention are further described below:

[0058] A real-time monitoring system for the fermentation process of black tea includes a hardware system and a software system, wherein:

[0059] The hardware system is a real-time image acquisition device for fermentation leaves, and its specific structure is as follows: Figure 1 As shown, the system includes an adjustable lifting bracket 81 and a base 50. A camera 20 with a lens is fixed inside the housing 10, while a defogging assembly 30 is installed at the defogging window 11 at the bottom of the housing 10 to ensure a fog-free field of view for the camera 20 lens during operation. An adjustable brightness light source 70 is installed at the lower end of the adjustable lifting bracket 81. A sample tray 40 is placed on the base 50. Both the sample tray 40 and the adjustable brightness light source 70 are housed within a light shield 60. See the attached diagram for details. Figure 1 As shown.

[0060] For the adjustable lifting bracket 81, a corresponding linear motion mechanism such as a vertical electric rail, a vertical cylinder, or a rack and pinion mechanism can be selected to drive the housing 10 to produce a vertical movement and cause the housing 10 to move closer to or further away from the sample tray 40. Similarly, the lines transmitting the corresponding power and information naturally pass through the housing 10 or the outer cover 80 to connect to external equipment; such as the power supply for the camera 20, data transmission lines, and the voltage-adjustable power adapter connecting the defogging assembly 30, etc., can all be connected and fixed and pass through the housing 10 or even extend outside the outer cover 80.

[0061] In practical operation, the defogging window 11 is crucial for achieving a fog-free field of view. Therefore, the defogging window 11 of this invention uses a single-sided ITO coated glass with a length of 55mm, a width of 50mm, and a thickness of 1.1mm. The inner surface of this single-sided ITO coated glass is fitted with electrodes with a resistance of 9-10Ω, and the entire electrode is coated with CB-813 conductive silver paste to avoid interference with the imaging quality of the camera 20 caused by fog in extremely humid and foggy environments. For the adjustable brightness light source 70, a moisture-proof and fog-proof surface light source with dimensions of 32cm × 32cm and a central opening size of 7.5cm × 7.5cm can be selected. Similarly, the sample tray 40 has an external diameter of 30cm × height of 7.5cm and an internal diameter of 29.5cm × height of 7cm; the light shield 60 has an external diameter of 35cm × height of 28cm and an internal diameter of 34.5cm × 28cm.

[0062] Based on the above structure, this invention also provides a method for preparing honeycomb ceramic, which is used to prepare sample disk 40 and light shield 60. The specific preparation method is as follows:

[0063] a) Preparation of binder: Mix agar powder and water at a weight ratio of 0.25:100 and heat to dissolve. After cooling, add 0.25% corn flour, 0.25% sucrose and 0.1% yeast powder per 1 kg of solution. Place at 35-37℃ for 4-5 hours, stirring 2-3 times in between. Pass through a 60-mesh sieve for later use.

[0064] b) Thoroughly mix 45% by weight of Qimen porcelain clay powder (passed through 80 mesh), 25% by weight of natural slate She inkstone powder (passed through 100 mesh), 15% by weight of activated carbon powder (passed through 100 mesh), and 15% by weight of diatomaceous earth powder (passed through 80 mesh). Add binder and stir to form a slurry. After shaping the slurry, let it stand at 35-37℃ for 6-7 hours, then transfer it into a kiln and fire it at 80-130℃ for 2 hours. Gradually increase the temperature from 130℃ to 450℃ within 2 hours and fire for 10 hours. Allow it to cool naturally to obtain the finished product.

[0065] The software system includes the establishment of digital classification labels for fermentation degree and the establishment of fermentation degree discrimination model. Specifically, it includes digital classification labels for fermentation degree based on original image data of different bud and leaf parts during the fermentation process, artificial sensory experience data on fermentation degree and tea quality, quality component data, etc.; fermentation degree discrimination model developed based on PyTorch's MobileNetV3+Vision Transformer; and fermentation degree discrimination visualization software.

[0066] For example, in the embodiment:

[0067] The specific development process for digital classification labels for fermentation level is as follows:

[0068] (1) Sixty batches of black tea processing experiments were conducted using 10 tea tree varieties. 450 tea samples with different fermentation processes were collected. At the same time as the samples were collected, artificial sensory experience information such as the color of the fermented leaves, fermentation time and degree of fermentation were recorded.

[0069] (2) Sensory evaluation was conducted on 450 tea samples with different fermentation processes. The key stages of tea quality changes during the fermentation process were assessed by taste, aroma and leaf residue. The fermentation process was divided into six stages: light, light, close to moderate, moderate, close to over, and over.

[0070] (3) The catechin content of 450 tea samples with different fermentation processes was detected. It was found that the catechin content decreased with the progress of fermentation, which could be divided into three stages. The slopes of the change curves for each stage were 0.024, 0.017, and 0.006, respectively. The rapid decrease stage accounted for about 55% of the fermentation process, with a catechin residue of about 22%. Then, the decrease slowed down, accounting for about 22% of the fermentation process, with a catechin residue of about 15%. The slow decrease stage accounted for about 23% of the fermentation process, with a catechin residue of about 12%. Since the sensory quality of tea is better when the catechin residue is in the range of 12% to 15%, two nodes were found before and after the slow decrease stage of catechin (moderate fermentation stage) where the degree of fermentation (insufficient-moderate, moderate-excessive) changed rapidly, which is consistent with the sensory evaluation results.

[0071] (4) By integrating the information of human sensory experience in judging the degree of fermentation, the sensory evaluation information of tea, the data of catechin residue, and statistical analysis, a digital classification label for the degree of fermentation is established. The fermentation process is assigned a value of 45 to 100. The larger the value, the heavier the degree of fermentation. Fermentation begins at rolling and ends at approximately 45. The degree of fermentation is assigned a value of 45 to 65 for the lighter range, 65 to 75 for the slightly lighter range, 75 to 80 for the range close to moderate, 80 to 85 for the moderate range, 85 to 90 for the range close to the excessive range, and 90 to 100 for the excessive range.

[0072] The development process for the fermentation degree discrimination model and visualization software is as follows:

[0073] (5) Based on the developed digital classification labels for fermentation degree, the 68,000 original images of fermenting leaves collected during the fermentation process were divided into six datasets according to the ratio of training set: test set: validation set = 7:2:1: light, light, moderate, moderate, close to excessive, and excessive fermentation degree.

[0074] (6) Perform histogram equalization, geometric data augmentation, and color space conversion on the images of each dataset to improve the contrast and brightness of the images and the accuracy of the model. Use the converted image data as one of the final input data for the deep learning model.

[0075] (7) Build a MobileNetV3 model based on PyTorch.

[0076] The initial part consists of one convolutional layer, which extracts features through a 3x3 convolution, i.e., a convolutional layer, a BN layer, and an h-switch activation layer, which is the same in both Large and Small sizes.

[0077] The middle section consists of multiple convolutional layers, forming a network structure with multiple blocks (MobileBlocks) containing convolutional layers. The number of layers and parameters differ between the Large and Small versions.

[0078] The final section omits the Squeeze operation, advances Avg Pooling, and directly uses 1x1 convolutions instead of fully connected layers to output categories, reducing computational cost. It incorporates a DeptwiseConvolution layer, which significantly reduces computation and parameter count, and an Inverted Residuals structure, minimizing information loss after ReLU activation. Furthermore, it uses the h-swish activation function instead of the swish function to further reduce computation and improve performance, enhancing model efficiency without increasing network complexity. By inputting image information into MobileNetV3, intermediate and deep feature maps are extracted and superimposed as image patches in the Vision Transformer model. Position embedding is incorporated into the image patch embedding, and different strategies are used to preserve spatial / positional information globally.

[0079] After the model is established, the data in step (6) is used as the input of the initial model to train the model, and finally the trained weights are used to perform the final result inference.

[0080] (8) The HSV feature values ​​of images from each fermentation stage and the area of ​​connected components retained after bud and leaf morphology processing were collected. The data collected by the two methods were analyzed, and the final color and morphological features were obtained through Gaussian filtering. Data modeling was performed based on these features, and the results could be predicted based on the model.

[0081] (9) The deep learning model and the traditional model established in step (8) are encapsulated in software.

[0082] (10) The software performs morphological identification on fermented leaves. For fermented leaves with good tenderness such as single bud, one bud and one leaf, and one bud and one leaf, and fermented leaves with slightly lower tenderness such as one bud and two leaves, and one bud and two leaves, different models are used to predict the degree of fermentation. Based on the comprehensive results, the software determines the current stage of fermentation and outputs the reliability of the degree of fermentation.

[0083] Of course, those skilled in the art will recognize that the present invention is not limited to the details of the exemplary embodiments described above, but also includes the same or similar structures that can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0084] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

[0085] The technologies, shapes, and structures not described in detail in this invention are all known technologies.

Claims

1. A method for real-time detection of the fermentation process of black tea, characterized in that... The process includes establishing digital classification labels for fermentation degree and building a fermentation degree discrimination model, with the following steps: 1) Establishment of digital classification labels for fermentation degree a) Conduct batch processing experiments of black tea using a specified number of tea tree varieties, collect tea samples at different fermentation stages, and record human sensory experience information at the same time as collecting samples, including fermented leaf color, fermentation time and degree of fermentation. b) Conduct sensory evaluation on tea samples with different fermentation processes, and assess the key stages of tea quality changes during the fermentation process through taste, aroma and leaf residue, so as to divide the fermentation process into six stages: light, slightly light, close to moderate, moderate, close to over, and over. c) Verify the results of step b): Detect the catechin content of tea samples at different fermentation stages. It was found that the slope of the change curve at each stage changed in three segments, that is, the catechin content can be divided into three stages of decline as the fermentation process progresses: a rapid decline stage, a slowing decline stage, and a slow decline stage. The sensory quality of tea is best when the catechin content is 12%~15%. At this time, two nodes where the degree of fermentation changes during the slow decline of catechins were identified. These two nodes include the time from under-fermentation to moderate fermentation and the time from moderate fermentation to over-fermentation. The comparison showed that it was the same as the sensory evaluation results in step b). d) Integrate the human sensory experience information for judging the degree of fermentation, the sensory evaluation information of tea, the data on catechin residues, and statistical analysis to establish a digital classification label for the degree of fermentation. Assign values ​​to the fermentation process, with larger values ​​indicating a heavier degree of fermentation. Fermentation begins at the rolling stage and ends at the fermentation stage as the starting value. Assign values ​​to light, slightly light, near moderate, moderate, near over, and over ranges respectively. 2) Establishment of a fermentation degree discrimination model e) Based on the digital classification labels of fermentation degree established in step 1), the original images of fermenting leaves collected during the fermentation process are divided into six datasets according to the ratio of training set: test set: validation set = 7:2:1: light, light, close to moderate, moderate, close to excessive, and excessive fermentation degree. f) Perform histogram equalization, geometric data augmentation, and color space conversion on the images of each dataset to improve image contrast and brightness and model accuracy. Use the images of each dataset after histogram equalization, geometric data augmentation, and color space conversion as one of the final input data for the deep learning model. g) Build a MobileNetV3 model based on PyTorch The initial part consists of one convolutional layer, which extracts features through a 3x3 convolution, i.e., a convolutional layer, a BN layer, and an h-switch activation layer, which is the same in both Large and Small sizes. Middle section: Includes a network structure containing blocks of convolutional layers; The final part: the Squeeze operation is omitted, Avg Pooling is brought forward, and 1x1 convolutions are used directly instead of fully connected layers to output the categories; among them, Deptwise Convolution layers and Inverted Residuals structures are added, so that less information is lost after the high-dimensional information is activated by the ReLU activation function, and the h-swish activation function is used instead of the swish function in the structure. By inputting image information into MobileNetV3, the intermediate layer feature maps and deep layer feature maps are extracted and superimposed as image patches in the Vision Transformer model. Position embedding is added to the embedding of the image patch, and spatial / positional information is preserved globally through different strategies. After the model is built, the data in step f) is used as the input of the initial model to train the model and obtain a deep learning model. Finally, the trained weights are used to perform the final result inference. h) Collect the HSV feature values ​​of images from each fermentation stage and the area of ​​connected components retained after bud and leaf morphology processing. Analyze the data collected by the two methods, and obtain the final color and morphological features through Gaussian filtering. Based on these features, data modeling is performed to obtain a traditional model, which can be used to predict the results. i) By encapsulating the deep learning model and the traditional model established in h), a fermentation degree discrimination model is obtained; 3) Morphological identification of fermented leaves is performed using a fermentation degree discrimination model to determine the current fermentation stage of the leaves.

2. The method for real-time detection of the fermentation process of black tea according to claim 1, characterized in that: In step 3), the fermentation leaves are morphologically identified by the fermentation degree discrimination model, and the reliability of the current fermentation degree of the fermentation leaves can be output simultaneously.

3. A method for real-time detection of the fermentation process of black tea according to claim 1 or 2, characterized in that: The original images of the fermenting leaves during the fermentation process are obtained by a real-time image acquisition device for the fermenting leaves; the real-time image acquisition device for the fermenting leaves includes a housing (10) for forming a sealed space, a camera (20) for image acquisition is arranged inside the housing (10), a defogging window (11) is arranged on the bottom surface of the housing (10) at the lens end of the camera (20), and a defogging component (30) is arranged inside the defogging window (11); a sample tray for holding the fermenting leaves is arranged directly below the defogging window (11). 40), the sample tray (40) is placed on the base (50); a straight sleeve-shaped light shield (60) is also vertically mounted on the base (50), the cylindrical cavity of the light shield (60) forms a mounting cavity for accommodating the sample tray (40) and allowing the bottom end of the housing (10) to extend into it, and an adjustable brightness light source (70) is arranged inside the light shield (60); the base (50), the light shield (60), the sample tray (40) and the housing (10) are all located inside the outer cover (80) which serves as a sealed cavity; wherein: The sample disk (40) is made of honeycomb ceramic and is prepared using the following method: a) Preparation of adhesive Mix agar powder and water at a weight ratio of 0.25:100 and heat to dissolve to obtain a solution; after cooling, add 0.25% corn flour, 0.25% sucrose and 0.1% yeast powder per 1 kg of the solution by weight, place at 35~37℃ for 4~5 hours, stirring 2~3 times in between, and pass through a 60-mesh sieve for later use. b) Thoroughly mix 45% by weight of Qimen porcelain clay powder, 25% by weight of natural slate She inkstone powder, 15% by weight of activated carbon powder, and 15% by weight of diatomaceous earth powder. Then add a binder and stir to form a slurry, and shape it into a blank. After the blank is shaped, let it stand at 35~37℃ for 6~7 hours, then transfer it into a kiln and fire it at 80~130℃ for 2 hours. Gradually increase the temperature from 130℃ to 450℃ within 2 hours and fire it for 10 hours. Then let it cool naturally to obtain the shaped product.

4. The method for real-time detection of the fermentation process of black tea according to claim 3, characterized in that: In step b), the Qimen porcelain clay powder is sieved through an 80-mesh sieve, the natural slate She inkstone powder is sieved through a 100-mesh sieve, the activated carbon powder is sieved through a 100-mesh sieve, and the diatomaceous earth powder is sieved through an 80-mesh sieve.

5. The method for real-time detection of the fermentation process of black tea according to claim 3, characterized in that: The material of the light shield (60) is the same as that of the sample tray (40).

6. The method for real-time detection of the fermentation process of black tea according to claim 3, characterized in that: The adjustable brightness light source (70) is a moisture-proof and fog-proof surface light source with a size of 32cm×32cm and a central opening size of 7.5cm×7.5cm.

7. The method for real-time detection of the fermentation process of black tea according to claim 3, characterized in that: The sample tray (40) has an external dimension of 30cm in diameter and 7.5cm in height, and an internal dimension of 29.5cm in diameter and 7cm in height.

8. The method for real-time detection of the fermentation process of black tea according to claim 3, characterized in that: The light shield (60) has an external dimension of 35cm in diameter and a height of 28cm, and an internal dimension of 34.5cm in diameter and 28cm in height.

9. The method for real-time detection of the fermentation process of black tea according to claim 3, characterized in that: The defogging window (11) is made of single-sided ITO coated glass with a length of 55mm, a width of 50mm, and a thickness of 1.1mm. An electrode with a resistance of 9~10Ω is installed on the inner side of the defogging window (11), and the entire electrode is coated with conductive silver paste. The electrode is electrically connected to the power supply outside the housing (10). The electrode constitutes the defogging assembly (30).

10. The method for real-time detection of the fermentation process of black tea according to claim 3, characterized in that: The housing (10) is mounted on the lifting bracket (81) located inside the outer casing (80).

Images