A tea mixing proportion automatic adjusting method, system and medium

Abstract

The application discloses a kind of tea mixing proportion automatic regulation method, system and medium, method includes: obtaining the visible light image and near infrared spectrogram of tea raw material, extract color, texture, integrity, water content and phenol amine ratio feature information by transfer learning model, generate raw material parameter information by PLS regression model;Target parameter information is generated based on the information of mixed tea, and the initial proportioning parameter is calculated and mixed by combining multi-objective optimization algorithm;Real-time acquisition stirring resistance value and stirring image in the mixing process, dynamically correct proportioning parameter according to feedback information, realize the dynamic adaptation of stirring state and target proportioning by cyclic regulation, until mixing is completed.This method solves the problem of static setting and dynamic mismatching process of proportioning parameter by multi-modal spectral feature extraction and real-time stirring feedback control, improves mixing uniformity and quality stability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of tea processing, and in particular to a tea mixing ratio automatic adjustment method and system and a medium. BACKGROUND

[0002] The existing tea mixing process mainly relies on manual sensory tea evaluation and fixed parameter mechanical stirring, but the correlation between the manual experience and the physicochemical indicators and sensory quality of tea is difficult to accurately quantify, which easily leads to insufficient matching degree of raw material parameters and target ratio; and the traditional mixing equipment lacks real-time sensing ability for dynamic changes of material bulk density in the stirring process, which easily causes problems such as poor uniformity and high tea breaking rate, especially for multi-category blended tea, which lacks self-adaptive control mechanism when the stirring resistance suddenly changes, resulting in ratio execution deviation. SUMMARY

[0003] Therefore, the purpose of the present application is to provide a tea mixing ratio automatic adjustment method and system and a medium to solve the problem of lack of self-adaptive control mechanism in the current tea mixing, which leads to tea mixing ratio deviation and poor mixing uniformity.

[0004] In order to achieve the above technical purpose, in a first aspect, the present application provides a tea mixing ratio automatic adjustment method, comprising:

[0005] Obtaining spectral image information of a plurality of tea raw materials, the spectral image information including visible light images and near-infrared spectral images;

[0006] Inputting the spectral image information into a transfer learning model for feature extraction to obtain tea raw material feature information, the tea raw material feature information including color, texture, integrity, moisture content and phenol-ammonia ratio;

[0007] Inputting the tea raw material feature information into a PLS regression model to obtain raw material parameter information;

[0008] Obtaining mixing tea information, the mixing tea information including tea categories after mixing and tea characteristics;

[0009] Generating target parameter information according to the mixing tea information;

[0010] Inputting the target parameter information and the raw material parameter information into a multi-objective optimization algorithm model for calculation to obtain initial mixing ratio parameters;

[0011] Mixing the plurality of tea raw materials according to the initial mixing ratio parameters;

[0012] Performing the following steps at a preset frequency:

[0013] Obtaining stirring information in real time during the mixing process, the stirring information including stirring resistance value and stirring image;

[0014] generate a feedback information according to the stirring information, generate an adjustment strategy according to the feedback information, and adjust the initial proportioning parameters according to the adjustment strategy, denoted as corrected proportioning parameters;

[0015] mix the multiple tea raw materials according to the corrected proportioning parameters;

[0016] until the tea raw materials are mixed.

[0017] In some embodiments, the transfer learning model is configured to be trained by the following steps:

[0018] obtain a pre-trained coffee bean grading model, and retain a bottom convolutional network structure of the coffee bean grading model, the bottom convolutional network structure comprising at least three convolution-pooling modules connected in series;

[0019] input the sample spectral image information in the sample database into the bottom convolutional network structure to obtain sample tea general features;

[0020] perform adversarial training on the sample tea general features through a gradient reversal layer, so that the sample tea general features satisfy the domain invariance;

[0021] construct a tea-specific processing layer and splice it at the output end of the bottom convolutional network structure, the tea-specific processing layer comprising a visible light branch processing layer and a near-infrared branch processing layer;

[0022] input the sample tea general features and the sample spectral image information into the tea-specific processing layer to obtain a feature output result, the feature output result comprising a first sample branch feature and a second sample branch feature;

[0023] fuse the first sample branch feature and the second sample branch feature through a cross-modal attention mechanism to obtain a sample fusion feature;

[0024] and optimize the parameters of the tea-specific processing layer through supervised learning, comprising:

[0025] minimizing the classification error of the first sample branch feature;

[0026] minimizing the spectral reconstruction error of the second sample branch feature;

[0027] splice the sample fusion feature and the sample tea general features to form sample tea raw material feature information.

[0028] In some embodiments, constructing the tea-specific processing layer and splicing it at the output end of the bottom convolutional network structure comprises:

[0029] the visible light branch processing layer and the bottom convolutional network structure are connected through a feature map channel splicing manner of the sample tea general features to obtain first branch input information;

[0030] Construct a first dilated convolutional module, the dilation rate of which is positively correlated with the average spacing of tea leaf veins in the first branch input information;

[0031] The near-infrared branch processing layer receives the sample near-infrared spectral image from the sample spectral image information, and fuses it with the common features of the sample tea leaves through convolution to obtain the second branch input information;

[0032] A second dilated convolutional module is constructed, and the kernel size of the second dilated convolutional module is matched with the bandwidth of the tea polyphenol feature absorption peak of the input information of the second branch.

[0033] In some embodiments, inputting spectral image information into a transfer learning model for feature extraction to obtain tea raw material feature information includes:

[0034] The visible light image is processed by skeletonization to obtain a binary image of leaf veins, and the leaf vein spacing feature is obtained, which is expressed by formula (1), as follows:

[0035] ;

[0036] In formula (1), It is a characteristic quantity of leaf vein spacing. For the number of effective leaf vein pairs, These are the coordinates of the first center point of the adjacent leaf veins after the skeletonization process. The coordinates of the second center point of the adjacent leaf vein after skeletonization;

[0037] The expansion rate of the first expanded convolution module is adjusted according to the leaf vein spacing feature, which is expressed by formula (2), as follows:

[0038] ;

[0039] In formula (2), Scaling factor As the reference spacing, For adjustment function, expansion rate The lower limit value, expansion rate The upper limit, Expansion rate;

[0040] The characteristic peak parameters of tea polyphenols were extracted from the near-infrared spectral image and expressed by formula (3), which is as follows:

[0041] ;

[0042] In formula (3), a characteristic peak parameter of tea polyphenols, a first half-height width boundary, a second half-height width boundary;

[0043] a kernel size of a second dilated convolution module is generated according to the characteristic peak parameter, which is represented by formula (4) as follows:

[0044] ;

[0045] In formula (4), is a kernel size, is a spectral resolution, is a floor function.

[0046] In some embodiments, the characteristic information of the tea raw material is input into the PLS regression model to obtain the raw material parameter information, which includes:

[0047] The characteristic information of the tea raw material is standardized and converted to obtain a raw material characteristic vector;

[0048] The raw material characteristic vector is projected into the PLS model latent variable space, and a plurality of latent variable scores are obtained by iterative solution;

[0049] The plurality of latent variable scores are converted into standard raw material parameters by a regression equation;

[0050] The standard raw material parameters are restored and converted to obtain the raw material parameter information.

[0051] In some embodiments, the multi-objective optimization algorithm model is configured to be constructed based on a multi-objective genetic algorithm;

[0052] The target parameter information and the raw material parameter information are input into the multi-objective optimization algorithm model for calculation to obtain initial matching parameters, which include:

[0053] An optimization objective function is constructed, which is represented by formula (5) as follows:

[0054] ;

[0055] In formula (5), is a flavor matching function, is a raw material utilization function, is a process handling function, is a matching vector, is a predicted flavor characteristic, is a target flavor characteristic, is the Euclidean distance between the predicted flavor characteristic and the target flavor characteristic, by adjusting the matching to minimize the difference between the predicted flavor characteristic and the target flavor characteristic, is a raw material cost coefficient, is a raw material proportion, is a raw material moisture content, is a maximum allowable difference of equipment, is a target raw material proportion, is a target raw material moisture content;

[0056] The multi-objective genetic algorithm is used to solve the optimization objective function, and the input parameters of the multi-objective genetic algorithm are the raw material parameter information and the target parameter information. The variable of the multi-objective genetic algorithm is defined as the mass percentage of the tea raw material in the mixing process.

[0057] The ratio set is obtained, and the ratio set includes a plurality of ratio arrays, and each ratio array has a ratio score;

[0058] The ratio array with the highest ratio score is selected as the initial ratio parameter, and the initial ratio parameter includes the feeding ratio of each tea raw material and the feeding order.

[0059] In some embodiments, the adjustment strategy includes a basic adjustment mode and an advanced adjustment mode;

[0060] The feedback information is generated according to the stirring information, the adjustment strategy is generated according to the feedback information, and the initial ratio parameter is adjusted according to the adjustment strategy, which is denoted as the corrected ratio parameter, including:

[0061] The stirring resistance value is subjected to a moving average filtering, the current bulk density characteristic value is calculated, and the agglomerate area and the broken leaf rate are obtained by image processing the stirring image;

[0062] The bulk density characteristic value, the agglomerate area and the broken leaf rate are arranged into actual indexes;

[0063] It is judged in turn whether the actual indexes meet the range of the preset index threshold value;

[0064] If one of the actual indexes exceeds the preset index threshold value, the basic adjustment mode is switched;

[0065] If two or more of the actual indexes exceed the preset index threshold value, the advanced adjustment mode is switched.

[0066] In some embodiments, the basic adjustment mode is configured to:

[0067] The deviation information of the exceeded actual index and the preset index threshold value is calculated to generate the first deviation information;

[0068] The first deviation information is converted into the corrected ratio parameter;

[0069] The advanced adjustment mode is configured to:

[0070] calculate deviation information of the actual index exceeding the preset index threshold, to generate second deviation information;

[0071] obtain an actual index category to which a maximum deviation value in the second deviation information belongs;

[0072] If the actual index category is the lumping area, the tea raw material added so far is obtained and recorded as an abnormal tea raw material, and it is determined whether the abnormal tea raw material in the initial proportioning parameter still exists in the subsequent process, and if so, the added mass and / or the added number of the abnormal tea raw material is reduced;

[0073] If the actual index category is the broken leaf rate, the torque and the rotating speed of the stirring group are adjusted according to the change trend of the broken leaf rate;

[0074] If the actual index category is the bulk density, the adding speed of the tea raw material is adjusted according to the change trend of the bulk density.

[0075] In the second aspect, the present application further provides a tea mixing and proportioning automatic adjusting system, comprising a control unit, a plurality of raw material conveying belts, a spectral image acquisition assembly, a stirring assembly and a feedback assembly, the control unit is used to execute the method of the first aspect; each raw material conveying belt is paved with a kind of tea raw material; the spectral image acquisition assembly is arranged above the raw material conveying belt, and the spectral image acquisition assembly comprises a visible light camera, a near-infrared detector and a light source, the visible light camera, the near-infrared detector and the light source are electrically connected with the control unit respectively; the stirring assembly comprises a stirring barrel and a stirring group, the stirring barrel is communicated with the plurality of raw material conveying belts, the stirring group comprises a stirring shaft, a stirring blade, an optical fiber strain gauge and a force sensor, the stirring blade is arranged along the circumference of the stirring shaft, the optical fiber strain gauge is arranged on the stirring blade, the force sensor is arranged at the input end of the stirring shaft, the force sensor is used to detect the torque of the stirring shaft, and the optical fiber strain gauge and the force sensor are electrically connected with the control unit respectively; the feedback assembly comprises a camera, and the camera is arranged above the stirring barrel.

[0076] In the third aspect, the present application further provides a computer readable storage medium, which stores computer program instructions, and the computer program instructions realize the method of the first aspect when executed by a processor.

[0077] By adopting the technical scheme, the present application has the beneficial effects compared with the prior art:

[0078] The technical solution provides a tea mixing and proportioning automatic adjustment method, a system and a medium, and the method comprises the following steps: acquiring a visible light image and a near-infrared spectrum image of tea raw materials, extracting color, texture, integrity, water content and phenol-ammonia ratio feature information through a transfer learning model, and generating raw material parameter information through a PLS regression model; target parameter information is generated based on proportioning tea information, and initial proportioning parameters are calculated and mixed by combining a multi-objective optimization algorithm; in the mixing process, stirring resistance values and stirring images are collected in real time, the proportioning parameters are dynamically corrected according to the feedback information, and dynamic adaptation of the stirring state and the target proportioning is realized through cyclic adjustment until the mixing is completed. The above technical solution solves the problems of static setting of proportioning parameters and mismatching in the dynamic mixing process through multi-modal spectrum feature extraction and real-time stirring feedback regulation, and improves the mixing uniformity and quality stability. BRIEF DESCRIPTION OF DRAWINGS

[0079] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of these drawings.

[0080] Figure 1 is a method step diagram of steps S101 to S111 of the adjustment method described in the specific embodiment;

[0081] Figure 2 is a method step diagram of steps S201 to S209 of the adjustment method described in the specific embodiment. DETAILED DESCRIPTION

[0082] The present application will be further described in detail below in combination with the drawings and embodiments. It is particularly pointed out that the following embodiments are only used to illustrate the present application, but do not limit the scope of the present application. Similarly, the following embodiments are only some embodiments of the present application, not all embodiments, and all other embodiments obtained by those skilled in the art without creative labor are within the scope of the present application.

[0083] Please refer to Figure 1 In the first aspect, the present embodiment provides a tea mixing and proportioning automatic adjustment method, comprising:

[0084] S101, acquiring spectrum image information of a plurality of tea raw materials, the spectrum image information comprising a visible light image and a near-infrared spectrum image;

[0085] S102, input the spectral image information to the transfer learning model for feature extraction to obtain tea material feature information, the tea material feature information including color, texture, integrity, water content and phenol-ammonia ratio;

[0086] S103, input the tea material feature information to the PLS regression model to obtain material parameter information;

[0087] S104, obtain the information of the blended tea, the information of the blended tea including the tea category and the tea characteristics after blending;

[0088] S105, generate target parameter information according to the information of the blended tea;

[0089] S106, input the target parameter information and the material parameter information to the multi-objective optimization algorithm model for calculation to obtain initial blending parameters;

[0090] S107, blend the plurality of tea materials according to the initial blending parameters;

[0091] The following steps are performed at a predetermined frequency:

[0092] S108, obtain stirring information in real time during the blending process, the stirring information including stirring resistance value and stirring image;

[0093] S109, generate feedback information according to the stirring information, generate adjustment strategy according to the feedback information, and adjust the initial blending parameters according to the adjustment strategy, denoted as corrected blending parameters;

[0094] S110, blend the plurality of tea materials according to the corrected blending parameters;

[0095] S111, until the blending of the tea materials is completed.

[0096] In step S101, preferably, the visible light image is a tea surface visual feature image collected by a visible light camera in the visible light waveband (400-700nm), used to represent the color, texture and morphological integrity of the tea; the near-infrared spectral image is a tea molecular vibration absorption characteristic spectrum obtained by a near-infrared detector (900-1700nm), used to analyze the internal chemical composition of the tea.

[0097] In step S102, the color, texture, and integrity of the tea leaves are analyzed through visible light images, and the water content and phenol-ammonia ratio are detected through near-infrared spectroscopy. Preferably, the transfer learning model reuses the underlying network of the coffee bean grading model to extract general features (such as edges and shapes), and adds a tea-specific network layer for analysis. The transfer learning model extracts general shape features through the pre-trained convolutional layers of the coffee bean grading model, and realizes cross-modal feature fusion through the newly added tea-specific network layer. The near-infrared response module focuses on the tea polyphenol feature peak by optimizing the convolution kernel parameters, and the cross-modal attention mechanism establishes the spatial correlation weight between visible light texture and near-infrared spectrum. Finally, a composite feature vector containing physical characteristics and chemical indicators is output. The specific content is described in the following text.

[0098] In step S103, the tea raw material feature information is converted into quantifiable process parameters through a PLS regression model. Preferably, the raw material parameter information is a linear mapping relationship between the tea raw material feature information and the quantifiable process parameters (such as water content threshold and phenolic substance proportion) established by the PLS regression model. After iterative calculation and based on the regression equation, it is converted into standard process parameters, and finally it is restored to actual process parameters through inverse transformation, realizing the mapping of multi-dimensional features to quantifiable parameters.

[0099] In step S104, the matching tea information is matched with the characteristic parameters (such as target phenol-ammonia ratio and aroma intensity) of the target tea category (such as black tea and green tea) through a pre-set database, or converted and generated by manually inputting quality requirements (such as taste thickness).

[0100] In step S105, the target parameter information is a set of process parameter target values established according to the characteristics of the matching tea. The sensory description (such as "freshness") can be converted into quantifiable physicochemical index threshold values through a fuzzy reasoning system.

[0101] In step S106, preferably, the multi-objective optimization algorithm model is based on a genetic algorithm framework. Under the condition of satisfying the raw material parameter constraints, the optimization objectives are to minimize the matching cost and maximize the quality closeness. The Pareto optimal solution set is obtained through non-dominated sorting and crowding degree calculation. The solution with the highest comprehensive weight is selected as the initial matching parameter. Preferably, the initial matching parameter includes the mass percentage of each raw material and the feeding order.

[0102] In steps S107 to S108, the stirring resistance value is preferably obtained by real-time measurement of the axial strain through the optical fiber strain gauge arranged on the stirring blade, and the stirring image can be collected by the camera on the stirring assembly for monitoring the material distribution state.

[0103] In steps S109 to S111, the feedback information is generated by fusing the bulk density change trend corresponding to the stirring resistance value and the uniformity index obtained by analyzing the stirring image, and the adjustment strategy is used to calculate the compensation amount of the proportioning parameter based on the fuzzy PID controller. For example, when the bulk density rises beyond the threshold value, the proportion of the fluffy raw material is increased.

[0104] The mixing completion condition is determined by the stirring time threshold value and the uniformity index. Preferably, when the uniformity change rate is less than 1% in three consecutive sampling periods, a termination signal is triggered.

[0105] The tea mixing and proportioning automatic adjustment method proposed in this embodiment significantly improves the proportioning accuracy and process stability by fusing multi-modal spectral feature extraction and dynamic process feedback mechanism, and constructing a closed-loop system from raw material analysis to mixing and control. Based on the cross-modal feature fusion of visible light and near-infrared spectrum, the general morphological feature extraction ability of coffee bean grading is reused through the transfer learning model, and the chemical composition is accurately analyzed by combining the tea special network layer; the multi-dimensional features are mapped to quantifiable process parameters by using PLS regression model, and the initial proportioning is generated by combining multi-objective optimization algorithm; in the mixing process, the proportioning parameters are dynamically corrected by real-time monitoring data of stirring resistance and material distribution.

[0106] This method overcomes the misjudgment of single modal physicochemical indicators by multi-spectral joint analysis, enhances the adaptability of the model to different raw materials through the transfer learning mechanism, balances the cost and quality targets through the non-dominated sorting algorithm, and effectively solves the process difficulties of mismatch between static proportioning and dynamic rheological characteristics through the closed-loop feedback system. At the same time, the fuzzy reasoning is used to convert the sensory description into quantifiable parameters, which opens up the collaborative optimization path between artificial experience and machine decision-making. Finally, the utilization rate of raw materials, quality consistency and process controllability in the tea mixing process are simultaneously improved, forming a full-link intelligent solution from feature modeling, parameter optimization to execution and control.

[0107] Please refer to Figure 2 In some embodiments, the transfer learning model is configured to be trained by the following steps:

[0108] S201, obtaining a pre-trained coffee bean grading model, and retaining the bottom convolutional network structure of the coffee bean grading model, the bottom convolutional network structure comprising at least three convolution-pooling modules connected in series;

[0109] S202, inputting sample spectral image information in a sample database into the bottom convolutional network structure to obtain sample tea general features;

[0110] S203, performing adversarial training on the sample tea general features through a gradient reversal layer, so that the sample tea general features satisfy the domain invariance;

[0111] S204, a tea special processing layer is constructed and spliced at the output end of the bottom convolutional network structure, the tea special processing layer includes a visible light branch processing layer and a near-infrared branch processing layer;

[0112] S205, the sample tea general feature and the sample spectrum image information are input into the tea special processing layer to obtain a feature output result, the feature output result includes a first sample branch feature and a second sample branch feature;

[0113] S206, the first sample branch feature and the second sample branch feature are fused through a cross-modal attention mechanism to obtain a sample fusion feature;

[0114] And the parameters of the tea special processing layer are optimized through supervised learning, including:

[0115] S207, minimizing the classification error of the first sample branch feature;

[0116] S208, minimizing the spectrum reconstruction error of the second sample branch feature;

[0117] S209, the sample fusion feature and the sample tea general feature are spliced to form a sample tea raw material feature information.

[0118] In this embodiment, the bottom convolutional network is used to extract field-independent general features, supporting the reusability of the transfer learning model; the tea special processing layer is used to capture tea-specific features, realizing color, texture, and phenolic amine ratio extraction; and the cross-modal attention mechanism is used for feature fusion.

[0119] In step S201, the coffee bean grading model is trained by an image dataset (containing visible light and near-infrared spectrum images of different varieties of coffee beans) in a coffee bean grading scene. Preferably, the bottom convolutional network structure of the model is composed of three levels of convolution-pooling modules in series, each level of module extracts general morphological features such as edges and shapes through a 3x3 convolution kernel, and realizes feature dimension reduction through a maximum pooling operation, finally forming a general representation ability of plant tissue structure. The model uses a cross-entropy loss function to optimize the parameters, and uses coffee bean category labels as a supervision signal to complete pre-training.

[0120] In step S202, the sample spectrum image information is composed of visible light images and near-infrared spectrum images of tea samples, and after being input into the bottom convolutional network of the pre-trained model, the sample tea general features are extracted through the hierarchical activation response of the convolution kernel, the feature dimension is aligned with the original coffee bean feature space, and the visual attributes shared across species such as color distribution and edge contour are included.

[0121] In step S203, the gradient inversion layer multiplies the feature gradient by a negative coefficient through the back propagation stage to drive the sample tea general features extracted by the bottom layer network in the adversarial training process to retain the general representation ability of plant cell structure and fiber direction, and to be unable to distinguish the input source as coffee beans or tea leaves through the domain classifier, thereby realizing domain invariance. Specifically, the domain classifier is constructed by using a fully connected layer, and the domain specificity is eliminated when the bottom layer network parameter is updated through the binary classification cross-entropy loss.

[0122] In step S204, the training of the tea special processing layer does not need gradient inversion, and the target is to strengthen the tea specificity features (such as the “brown rice color” texture of Longjing tea) and to ensure that the near-infrared features strictly correspond to the real chemical composition.

[0123] Preferably, the visible light branch processing layer of the tea special processing layer adopts a residual network structure, and the capturing ability of the tea surface color gradient, leaf vein texture and other physical properties is strengthened through a jump connection; the near-infrared branch processing layer is constructed by a one-dimensional convolution kernel, and is slid along the spectral wavelength dimension to focus on the feature absorption peak of tea polyphenols, amino acids and other chemical components. The visible light branch processing layer and the near-infrared branch processing layer are independently initialized parameters, and only the special layer weight is updated during training to maintain the domain invariance of the bottom layer network.

[0124] In step S205, the visible light image in the sample spectral image information is input into the visible light branch processing layer, and the first sample branch feature (such as color RGB distribution, leaf edge integrity) representing the physical property is output after being extracted by the residual module; the near-infrared spectral image is input into the near-infrared branch processing layer, and the second sample branch feature (such as phenolic substance absorption intensity, water response curve) representing the chemical composition is output after one-dimensional convolution operation. The first sample branch feature, the second sample branch feature and the sample tea general feature jointly constitute a multi-level representation.

[0125] In step S206, the cross-modal attention mechanism is used to quantify the correlation strength of the visible light and the near-infrared feature. Preferably, the cross-modal attention mechanism establishes the correspondence between the texture details and the chemical indicators by calculating the correlation weight matrix of the spatial position of the visible light feature map and the wavelength dimension of the near-infrared spectrum. Specifically, the first sample branch feature is taken as a query vector, and the second sample branch feature is taken as a key-value pair, and after point product attention calculation, the near-infrared feature is weighted and fused to form a sample fusion feature containing physical properties and chemical indicators.

[0126] In step S207, the classification error of the first sample branch feature can be minimized by using a cross-entropy loss function, and the supervision signal comes from the tea appearance grade label (such as special grade, first grade), which drives the visible light branch to accurately identify the color uniformity, leaf damage degree and other physical properties affecting the grading.

[0127] In step S208, the minimization of the spectral reconstruction error of the second sample branch feature can be achieved by a mean square error loss function, which reconstructs the near-infrared feature output by the dedicated layer into the original spectral curve through the deconvolution network, forces the model to retain the spectral response mode corresponding to the true chemical composition, and avoids the loss of chemical information in the adversarial training process.

[0128] In step S209, the sample fusion feature and the sample tea leaf general feature are spliced through the channel dimension to form a dimensionally expanded sample tea leaf raw material feature information, wherein the sample tea leaf general feature provides a cross-species morphological benchmark, and the sample fusion feature carries tea leaf specific information, both of which cooperatively support the parameter mapping of the subsequent PLS regression model to achieve complementary enhancement of the feature space.

[0129] In this embodiment, the bottom convolutional network of the pre-trained coffee bean grading model is reused to extract cross-species general features, and after the gradient inversion layer is used for adversarial training to eliminate the domain difference, the visible light texture and near-infrared spectral features are captured by the tea-specific processing layer composed of a residual network and a one-dimensional convolution, respectively. The cross-modal attention mechanism is used to establish the correlation between physical properties and chemical indicators, and finally a multi-dimensional raw material representation is formed by feature splicing. The bottom convolutional network migration strategy effectively reuses the general representation ability of plant tissues, reducing the sample demand for tea feature training; the domain invariance maintained by the adversarial training ensures that the general features are not disturbed by the switching of raw material categories; the special processing layer strengthens the analysis accuracy of tea surface morphology and internal composition through the residual structure and one-dimensional convolution, respectively; the cross-modal attention mechanism solves the heterogeneous fusion problem of visible light and near-infrared features, and establishes a precise mapping relationship between color texture and phenol-ammonia ratio, moisture content; multi-level supervised learning simultaneously ensures the accuracy of physical property recognition and the fidelity of chemical composition representation. The overall scheme reduces the data labeling cost while achieving efficient extraction and precise correlation of multi-source features in the tea blending process, providing reliable feature engineering support for subsequent ratio optimization.

[0130] In some embodiments, the tea-specific processing layer is constructed and spliced at the output end of the bottom convolutional network structure, which includes:

[0131] The visible light branch processing layer is connected to the bottom convolutional network structure through channel splicing of the feature map of the sample tea leaf general feature, to obtain first branch input information;

[0132] A first dilated convolution module is constructed, and the dilated rate of the first dilated convolution module is positively correlated with the average distance of the tea leaf veins of the first branch input information;

[0133] The near-infrared branch processing layer receives the sample near-infrared spectral image in the sample spectral image information, and fuses the sample near-infrared spectral image with the sample tea leaf general feature to obtain second branch input information;

[0134] A second expansion convolution module is constructed, and a size of a convolution kernel of the second expansion convolution module matches a tea polyphenol characteristic absorption peak bandwidth of the second branch input information.

[0135] In the embodiment, the visible light branch processing layer and the feature map channel splicing manner of the bottom layer convolutional network structure can be understood as splicing the sample tea leaf general features (including color distribution, edge contour, and other cross-species visual attributes) output by the bottom layer convolutional network along the channel dimension with the initial input features of the visible light branch (local texture extracted by shallow layer convolution of the original visible light image), to form the first branch input information with expanded dimension. By fusing the sample tea leaf general features of the bottom layer network and the detailed information of the shallow layer network, the analysis capability of the visible light branch on the microscopic texture of the tea leaf surface, such as the vein direction, the hair density, and other attributes requiring cross-level feature collaborative judgment, is enhanced.

[0136] The first branch input information refers to the multi-dimensional feature tensor after channel splicing, which contains both the cross-species general features (such as plant cell arrangement pattern) extracted by the bottom layer network and the tea-specific texture details (such as leaf edge serration morphology) captured by the shallow layer convolution of the visible light image, providing multi-scale input for the subsequent expansion convolution module.

[0137] The first expansion convolution module is a processing unit adopting an expansion convolution structure, and the expansion rate is set according to the statistical value of the average distance of the leaf veins in the tea sample library. Specifically, the median distance corresponding to the pixel number can be taken as the expansion rate reference value by measuring the leaf vein distance distribution of the typical tea sample. By inserting a gap in the conventional convolution kernel, the receptive field is expanded to match the spatial periodicity of the leaf vein structure, thereby enhancing the extraction accuracy of the tea texture features (such as the distribution of “brown rice” patches in Longjing tea).

[0138] The second branch input information refers to the feature tensor obtained by channel weighted fusion of the sample near-infrared spectrum image and the sample tea leaf general features after passing through a one-dimensional convolution layer. The one-dimensional convolution kernel slides along the spectral wavelength dimension to extract the characteristic absorption peak response of tea polyphenols, amino acids, and other substances, and then adaptively weights the prior knowledge of plant tissue structure (such as cell wall light transmittance) in the sample tea leaf general features, to form the second branch input that fuses the chemical properties and morphological basis.

[0139] The second expansion convolution module refers to a convolution processing unit designed for near-infrared features, and the size of the convolution kernel is determined according to the characteristic absorption peak bandwidth of tea polyphenols in a specific wavelength range. By setting the kernel size according to the number of wavelength points corresponding to the half-width of the absorption peak, it is ensured that the convolution operation can completely cover the feature peak range, thereby improving the quantization accuracy of tea polyphenol content and oxidation degree. By performing local feature aggregation in the spectral dimension, noise interference is eliminated and the feature response of chemical components is strengthened.

[0140] This embodiment concatenates the general features of sample tea leaves output from the bottom-level network with the visible light shallow texture features to form the first branch input information. This first branch input information is then enhanced by a first dilated convolution module whose dilation rate is adapted to the leaf vein spacing. Near-infrared spectroscopy is fused with general features via one-dimensional convolution to form the second branch input information. A second dilated convolution module, whose kernel matches the absorption peaks of tea polyphenols, strengthens chemical composition analysis. This method achieves accurate identification of leaf vein direction and trichome density by channel-based feature fusion across layers. The positive correlation between dilation rate and leaf vein spacing allows the convolution receptive field to adaptively match the periodicity of the leaf structure, improving the accuracy of texture feature extraction. The matching design of the convolution kernel size and feature absorption peak bandwidth in the near-infrared branch ensures the accuracy of quantifying the degree of tea polyphenol oxidation and eliminates spectral noise interference. This embodiment, through dual-branch differential processing and physical property-driven parameter optimization, simultaneously improves the resolution quality of surface morphology and internal components during tea blending, establishing a reliable feature engineering foundation for ratio optimization.

[0141] In some embodiments, inputting spectral image information into a transfer learning model for feature extraction to obtain tea raw material feature information includes:

[0142] The visible light image is processed by skeletonization to obtain a binary image of leaf veins, and the leaf vein spacing feature is obtained, which is expressed by formula (1), as follows:

[0143] ;

[0144] In formula (1), It is a characteristic quantity of leaf vein spacing. For the number of effective leaf vein pairs, These are the coordinates of the first center point of the adjacent leaf veins after the skeletonization process. The coordinates of the second center point of the adjacent leaf vein after skeletonization;

[0145] The expansion rate of the first expanded convolution module is adjusted according to the leaf vein spacing feature, which is expressed by formula (2), as follows:

[0146] ;

[0147] In formula (2), Scaling factor As the reference spacing, For adjustment function, expansion rate The lower limit value, expansion rate The upper limit, Expansion rate;

[0148] The tea polyphenol characteristic peak parameter of the near-infrared spectrum image is extracted, and is expressed by formula (3) as follows:

[0149] ;

[0150] In formula (3), is the tea polyphenol characteristic peak parameter, is the first half-height width boundary, is the second half-height width boundary;

[0151] The kernel size of the second dilated convolution module is generated according to the characteristic peak parameter, and is expressed by formula (4) as follows:

[0152] ;

[0153] In formula (4), is the kernel size, is the spectral resolution, is the floor function.

[0154] In this embodiment, the skeletonization processing of the visible light image to obtain the vein binary image can be understood as that the vein area is gradually thinned to a single-pixel-width connected curve through a morphological thinning algorithm, the redundant information of the leaf pulp is eliminated while the topological structure of the vein is retained, and a binary image only representing the vein direction is formed. The edge pixels of the vein are removed through an iterative erosion operation until the thinning cannot continue, so that the subsequent vein spacing calculation is only carried out on the geometric relationship between the main vein and the side vein. The vein spacing feature quantity represents the average Euclidean distance between the center lines of adjacent parallel veins, and is calculated by formula (1), wherein the number of effective vein pairs needs to meet that the center point connecting line does not cross multiple veins and the angle with the main vein is less than a threshold value, and the vein pairs meeting the spatial distribution rule are selected after detecting the vein direction through the Hough transform.

[0155] In formula (2), the scaling factor is a dynamic parameter for adjusting the proportion relationship between the dilated rate and the vein spacing, and different value intervals are set according to the differences between tea varieties, and satisfy , for example, when processing the Biluochun with fine and delicate vein structure, the scaling factor is used to suppress the excessive amplification of the dilated rate, and the scaling factor is used for Pu'er tea with thick veins to strengthen the receptive field expansion; the reference spacing is obtained by statistically analyzing the median of the vein spacing of a typical tea sample, and is used as a unified reference scale between different varieties; the adjustment function limits the dilated rate to be in the range of , so as to ensure the effectiveness of the dilated convolution and avoid feature extraction failure caused by extreme values.

[0156] In formula (3), the tea polyphenol characteristic peak parameter By positioning the typical absorption peak at 1450 nm in the near-infrared spectrum, the first half-height width boundary and the wavelength difference between the second half-height width boundary are determined. The first half-height width boundary and the second half-height width boundary are determined by using a Gaussian fitting method, and the two boundary points are obtained by drawing a horizontal tangent line at 50% of the peak intensity and intersecting the spectral curve.

[0157] In formula (4), the spectral resolution is determined by the hardware parameters of the spectrometer, and represents the wavelength interval between adjacent sampling points. The down rounding operation can ensure that the convolution kernel size is an odd number, so that the convolution operation has a symmetric center point. For example, when and , the calculation result is . This embodiment makes the convolution kernel width completely cover the characteristic peak range, and strengthens the extraction of the spectrum features related to the oxidation degree of tea polyphenols.

[0158] In this embodiment, the visible light image is skeletonized to generate a leaf vein binary image, and the leaf vein spacing feature quantity is calculated to dynamically adjust the dilation rate of the first dilation convolution module. Meanwhile, the tea polyphenol characteristic peak parameter in the near-infrared spectrum is extracted to generate the convolution kernel size of the second dilation convolution module. Further, the effective leaf vein pairs are screened through morphological thinning and Hough transform, so as to ensure that the leaf vein spacing feature quantity accurately represents the geometric structure of the leaf, and to make the dilation rate adaptively adjusted according to the density of the leaf vein by combining the difference setting of the scaling factor and the reference spacing of different varieties. The near-infrared branch uses Gaussian fitting to locate the half-height width boundary of the characteristic peak, and uses the odd size convolution kernel designed by the spectral resolution and the down rounding operation to completely cover the tea polyphenol absorption peak range, thereby enhancing the quantization precision of the oxidation degree. In this embodiment, the morphological features of the leaf vein and the spectral chemical properties are integrated into the convolution parameter optimization, so as to realize the double-dimensional accurate analysis of the visual texture and the internal composition of the tea raw material, and to realize the accurate extraction of the tea features by adaptively adjusting the dilation rate and the convolution kernel size, thereby providing high-robustness feature engineering support for the blending process.

[0159] In some embodiments, the tea raw material feature information is input into a PLS regression model to obtain raw material parameter information, including:

[0160] The tea raw material feature information is standardized and converted to obtain a raw material feature vector;

[0161] The raw material feature vector is projected into a PLS model latent variable space, and a plurality of latent variable scores are obtained by iterative solving;

[0162] The plurality of latent variable scores are converted into standard raw material parameters through a regression equation;

[0163] The standard raw material parameters are reduced and converted to obtain raw material parameter information.

[0164] In this embodiment, the conversion of multi-source heterogeneous features to standardized parameters is realized by a PLS regression model. Specifically, the standardization conversion of tea raw material feature information can be understood as aligning the distribution centers and unifying the dimensions of heterogeneous data such as leaf vein distance features extracted from visible light images and tea polyphenol characteristic peak parameters extracted from near-infrared spectra, so as to form a raw material feature vector with a mean of 0 and a variance of 1. This process eliminates the feature scale deviation caused by differences in light source intensity or fluctuations in spectrometer sensitivity, for example, mapping millimeter-level values of leaf vein distance and nanometer-level values of spectral wavelength to the same order of magnitude, to ensure the stability of subsequent model calculations.

[0165] Projecting the raw material feature vector to the PLS model latent variable space can be understood as finding the maximum covariance direction between the original feature space and the standard raw material parameter space through the partial least squares regression algorithm, and constructing a latent variable axis that can explain the relationship between the feature variables and the target parameters. The iterative solution process is realized by extracting orthogonal latent variables one by one: first, calculate the covariance matrix of the feature vector and the target parameter to determine the first maximum covariance direction as the first latent variable axis; calculate the residual matrix after projecting the original data along the axis, and repeat the above steps until the number of preset latent variables is met, and finally obtain a plurality of mutually independent latent variable scores. Each latent variable score represents the projection intensity of the raw material feature in a certain covariance direction, for example, the first latent variable may reflect the comprehensive correlation between leaf vein density and tea polyphenol content.

[0166] Converting multiple latent variable scores to standard raw material parameters through a regression equation can be understood as establishing a linear combination relationship between the latent variables and the standardized target parameters (such as moisture content, polyphenol oxidase activity, etc.), and solving the regression coefficients of each latent variable using the least squares method to form a parameter prediction equation. The reduction and conversion of standard raw material parameters means that the prediction results are de-normalized, and the prediction values are restored to the original physical dimension by multiplying the standard deviation of the target parameter and adding the mean value, for example, the normalized moisture content prediction value is restored to the percentage form. The above steps effectively solve the multicollinearity problem between image and spectral features through latent variable space dimension reduction and covariance maximization projection, and improve the robustness of the prediction of key parameters of mixed raw materials.

[0167] The embodiment eliminates the dimensional difference between the vein spacing and the tea polyphenol characteristic peak by standardized conversion, constructs a characteristic vector with a mean of zero and a uniform variance, then iteratively solves the maximum covariance direction, projects the characteristics to the orthogonal latent variable space, extracts the comprehensive latent variable score representing the correlation between leaf vein density and chemical composition, then maps the latent variable to the standard raw material parameter based on the regression equation, and restores the physical dimension through reduction and conversion. This method effectively suppresses the multicollinearity interference of image texture features and spectral absorption peak parameters by maximizing the projection of covariance, and ensures the interpretability of the predicted values of parameters such as moisture content and polyphenol oxidase activity in the original dimension by using standardization and reverse standardization processing. The standardized conversion reduces the influence of light source fluctuation and instrument deviation on feature extraction, the latent variable space dimension reduction strengthens the weight of key associated features, and finally improves the robustness and adaptability of the prediction of raw material parameters of different varieties of tea leaves.

[0168] In some embodiments, the multi-objective optimization algorithm model is configured to be constructed based on a multi-objective genetic algorithm;

[0169] The target parameter information and the raw material parameter information are input into the multi-objective optimization algorithm model for calculation to obtain initial blending parameters, including:

[0170] An optimization objective function is constructed, which is represented by formula (5) as follows:

[0171] ;

[0172] In formula (5), is a flavor matching function, is a raw material utilization function, is a process handling function, is a blending vector, is a predicted flavor feature, is a target flavor feature, is the Euclidean distance between the predicted flavor feature and the target flavor feature, by adjusting the blending to minimize the difference between the predicted flavor feature and the target flavor feature, is a raw material cost coefficient, is a raw material proportion, is a raw material moisture content, is a maximum allowed difference of equipment, is a target raw material proportion, is a target raw material moisture content;

[0173] A multi-objective genetic algorithm is used to solve the optimization objective function, and the input parameters of the multi-objective genetic algorithm are the raw material parameter information and the target parameter information. The variable definition of the multi-objective genetic algorithm is the mass percentage of tea raw materials in the blending process;

[0174] obtaining a ratio set, the ratio set comprising a plurality of ratio arrays, each ratio array having a ratio score;

[0175] selecting a ratio array with the highest ratio score as an initial ratio parameter, the initial ratio parameter comprising a proportion of each tea material and a sequence of feeding.

[0176] In this embodiment, the optimization objective function is used to balance the complex relationship between flavor matching degree, material cost efficiency and process feasibility. In formula (5), the target flavor characteristics According to the tea category preset, the flavor matching function By calculating the Euclidean distance square term of the predicted flavor characteristics and the target flavor characteristics, the significant difference is amplified to strengthen the flavor deviation sensitivity, so that the algorithm preferentially adjusts the ratio vector to narrow the flavor difference.

[0177] The material cost coefficient is determined by the material grade and dynamically updated according to the current year's material purchase price, and the material utilization function By maximizing the weighted sum of the material cost coefficient and the proportion of the material, the proportion of low-cost high-grade materials is promoted, and the cost coefficient is dynamically updated according to the current purchase price to reflect market fluctuations.

[0178] The water content of the material is obtained by near-infrared detection; the maximum difference allowed by the equipment is determined by equipment test (such as the power- water content gradient relationship of the mixer), and the process handling function By constraining the absolute difference of the product of the water content and the proportion of different materials to be no more than the maximum difference allowed by the equipment, the processing feasibility of the mixed material in the mixer and other equipment is ensured, and the difference threshold is obtained by testing the gradient relationship between the power and the water content of the equipment.

[0179] It should be noted that increasing the proportion of low-cost materials may increase the flavor deviation and the water content difference may limit the ratio combination.

[0180] Multi-objective genetic algorithm is an intelligent algorithm that simulates biological evolution mechanism to solve the above optimization objectives. The input parameters include the water content, cost coefficient in the raw material parameter information and the target flavor characteristics in the target parameter information. The variable is defined as the mass percentage of each tea raw material in the blending process. The algorithm maps the blending vector to the chromosome individual through gene coding, generates a diverse blending set by using crossover and mutation operations, and evaluates the blending score of each blending array by non-dominated sorting and crowding degree calculation to reflect its convergence and distribution on the Pareto frontier. When selecting the blending array with the highest blending score as the initial blending parameter, the combination optimality of the feeding ratio and the process adaptability of the feeding order need to be met at the same time, for example, feeding high-moisture raw materials first can reduce the equipment load.

[0181] It should be particularly noted that the target flavor characteristics are pre-defined by tea categories, such as the fresh fragrance characteristics of Longjing tea or the aged alcohol characteristics of Pu'er tea, which are quantified by standardized flavor indicators. The water content of the raw material is obtained in real time by near-infrared spectroscopy detection, avoiding the time delay of the traditional drying method. The relationship curve between the water content gradient and the stirring torque is determined by equipment testing to establish the maximum difference allowed by the equipment.

[0182] This embodiment solves the problems of flavor deviation caused by the increase of the proportion of low-cost raw materials and the risk of equipment overrun caused by too large water content difference through multi-objective collaborative optimization, while ensuring that the flavor approaches the target value. It dynamically adapts to the fluctuations in raw material costs and changes in process constraints, ultimately achieving the triple optimization of economy, quality stability and production feasibility.

[0183] In some embodiments, the adjustment strategy includes a basic adjustment mode and an advanced adjustment mode.

[0184] The feedback information is generated according to the stirring information, the adjustment strategy is generated according to the feedback information, and the initial blending parameter is adjusted according to the adjustment strategy, denoted as the corrected blending parameter, which includes:

[0185] The stirring resistance value is subjected to a moving average filter to calculate the current bulk density characteristic value, and the stirring image is subjected to image processing to obtain the clumping area and the broken leaf rate.

[0186] The bulk density characteristic value, the clumping area and the broken leaf rate are arranged into actual indicators.

[0187] The actual indicators are sequentially judged whether they meet the range of the preset indicator threshold.

[0188] If one of the actual indicators exceeds the preset indicator threshold, the basic adjustment mode is switched.

[0189] If two or more of the actual indicators exceed the preset indicator threshold, the advanced adjustment mode is switched.

[0190] In the embodiment, the sliding average filtering of the stirring resistance value can be understood as a window weighted average processing of the continuously collected stirring resistance value, eliminating abnormal fluctuations caused by instantaneous mechanical vibration or sensor noise, ensuring the stability of the calculation of the bulk density characteristic value, and the window length is dynamically adjusted according to the stirring period. The bulk density characteristic value refers to the material unit volume mass inversely deduced from the filtered resistance value, which characterizes the mixing uniformity and fluffy state, and the inversion model used is established based on the stirring torque-density calibration curve, which can be obtained by pressure sensor and weighing module calibration.

[0191] The caking area refers to the pixel ratio of the agglomerated area extracted after the stirring image is binarized and segmented, reflecting the degree of moisture absorption or electrostatic adsorption of the material. Morphological opening operation is used for image processing to eliminate small particle noise; the broken leaf rate refers to the ratio of broken leaves calculated after edge detection and skeleton analysis, which is determined by the leaf aspect ratio and connected domain area threshold, and is used to evaluate the influence of mechanical shearing on the integrity of the raw material.

[0192] The adjustment range of the basic regulation mode is calculated according to the deviation degree and the preset response coefficient; the advanced regulation mode uses a machine learning model based on historical regulation records to predict parameter combinations, such as simultaneously adjusting the moisture content gradient of the raw material and the stirring speed, and the model training data comes from multiple index abnormal cases in the process database and the corresponding regulation effect. When switching the regulation mode, it is necessary to ensure the priority and conflict resolution of parameter adjustment, for example, when the broken leaf rate and the caking area are both over-standard, the speed is reduced first instead of increasing the dispersant to avoid further deterioration of the broken leaves.

[0193] The preset index threshold is the allowable fluctuation range of the material state set according to the historical process data and equipment performance, which is used to judge whether the bulk density, caking and broken leaf conditions in the stirring process are in the acceptable interval, for example, the threshold of the bulk density characteristic value is determined by the material flowability test, and the threshold of the caking area is set based on the screening efficiency.

[0194] The dynamic calibration of the preset index threshold is realized by combining offline simulation and online monitoring, for example, the bulk density threshold range is corrected every quarter according to the seasonal moisture change of the raw material, to ensure the adaptability of the regulation strategy to the fluctuation of production conditions.

[0195] The embodiment monitors stirring resistance and material image in real time, adopts sliding average filtering and image processing to extract bulk density, caking area and broken leaf rate, constructs multi-dimensional indexes, dynamically distinguishes abnormal degree based on preset threshold, triggers basic adjustment mode when single index exceeds limit, and corrects proportioning according to deviation; when multiple indexes exceed limit, the advanced adjustment mode is switched to, a machine learning model based on historical data is used to predict parameter combination, and key process variables are optimized in coordination. Through dual-dimensional perception of physical signals and visual features, the material rheological characteristics are accurately identified; the mode grading strategy takes into account single deviation fast response and complex condition collaborative regulation; sliding filtering and morphological processing suppress noise interference and improve caking and broken leaf identification accuracy; the historical data driven parameter prediction mechanism excavates multi-index correlation rules to avoid secondary problems; dynamic calibration threshold adapts to seasonal changes of raw materials, and the adaptability of the system is enhanced. Finally, the closed-loop fine regulation and control of the material state in the mixing process is realized, and the regulation efficiency and process stability are improved.

[0196] In some embodiments, the basic adjustment mode is configured to:

[0197] deviation information of the actual index exceeding the preset index threshold is calculated to generate first deviation information;

[0198] The first deviation information is converted into a corrected proportioning parameter;

[0199] The advanced adjustment mode is configured to:

[0200] deviation information of the actual index exceeding the preset index threshold is calculated to generate second deviation information;

[0201] The actual index category to which the maximum deviation value in the second deviation information belongs is obtained;

[0202] If the actual index category is caking area, the tea raw materials added so far are obtained, denoted as abnormal tea raw materials, it is judged whether the abnormal tea raw materials still exist in the subsequent process in the initial proportioning parameter, and if so, the added mass and / or added times of the abnormal tea raw materials are reduced;

[0203] If the actual index category is broken leaf rate, the torque and speed of the stirring group are adjusted according to the change trend of the broken leaf rate;

[0204] If the actual index category is bulk density, the adding speed of the tea raw materials is adjusted according to the change trend of the bulk density.

[0205] In the embodiment, the preset index threshold is obtained by offline simulation of stirring conditions under different raw material proportioning, the allowed fluctuation range is set in combination with equipment operating parameters and process standards, and the dynamic calibration is regularly carried out according to the changes of raw material characteristics.

[0206] The first deviation information refers to the absolute difference or relative proportion of a single actual indicator (bulk density, caking area or broken leaf rate) deviating from the preset indicator threshold. The deviation proportion is linearly or non-linearly converted by the offset degree of the current actual value relative to the upper and lower limits of the preset indicator threshold. For example, when the caking area exceeds the upper limit of the preset indicator threshold, the deviation proportion is calculated as the percentage of the actual over-limit area to the maximum allowed area.

[0207] The correction ratio parameter is generated according to the product of the preset response coefficient and the first deviation information. The response coefficient is obtained by regression analysis of the adjustment effect of the same type of deviation in historical process data. For example, the increase in the proportion of fluffy raw materials corresponding to each unit of caking area deviation.

[0208] In the advanced adjustment mode, the deviation degree of multiple over-limit indicators needs to be integrated when generating the second deviation information. The actual indicator category to which the maximum deviation value belongs is determined by comparing the normalized weights of each indicator deviation. The weights are dynamically allocated according to the influence coefficient of the indicator on the mixing quality. For example, the influence weight of broken leaf rate on the integrity of finished products is higher than that of bulk density.

[0209] The determination of abnormal tea raw materials is based on the correlation analysis of the initial mixing ratio parameters and the current added raw materials. If the subsequent raw material addition plan still exists in the initial mixing ratio parameters, the remaining addition quality is redistributed under the condition of satisfying the total mixing ratio constraint through a linear programming model, and the addition times of high-moisture or easily caking raw materials are preferentially reduced.

[0210] The broken leaf rate change trend is calculated by the gradient value of the broken leaf rate through a time series sliding window. If the gradient is continuously positive, the combination of stirring torque and speed is reduced. The adjustment amplitude is dynamically set according to the device performance curve and the broken leaf rate-shear force correlation model.

[0211] The bulk density change trend is determined by the first derivative of the bulk density characteristic value. If the density continuously rises, the addition speed of low-density raw materials is accelerated. The adjustment step of the addition speed is determined by the material flowability test marked by the feeding rate-density response relationship.

[0212] The embodiment realizes fine control of the tea blending process through hierarchical regulation architecture of basic regulation mode and advanced regulation mode, combined with physical index dynamic analysis and process parameter adaptive correction. The basic regulation mode is based on linear or nonlinear proportional conversion of the first deviation information, and uses the response coefficient of historical process data regression analysis to quickly generate correction blending parameters, ensuring efficient correction ability when a single index exceeds the limit. The advanced regulation mode adopts a multi-index collaborative discrimination mechanism with normalized weight dynamic allocation, identifies the core abnormal factors through maximum deviation value classification, triggers the linear programming model to optimize the raw material addition plan for caking area abnormalities, adjusts the stirring equipment parameters for broken leaf rate changes, and dynamically adjusts the addition speed for bulk density fluctuations, forming a multi-dimensional process parameter collaborative optimization strategy. This method balances regulation efficiency and accuracy through a hierarchical response mechanism, uses multi-source data analysis and process constraint condition fusion to suppress local parameter conflicts while maintaining overall blending stability, effectively reducing the incidence of abnormal conditions such as caking and broken leaves, and improving the robustness of the blending process and the consistency of the finished product quality.

[0213] In a second aspect, the embodiment also provides a tea blending ratio automatic regulation system, comprising a control unit, a plurality of raw material conveying belts, a spectral image acquisition component, a stirring component, and a feedback component. The control unit is used to execute the method of the first aspect. Each raw material conveying belt is laid with a kind of tea raw material. The spectral image acquisition component is arranged above the raw material conveying belt. The spectral image acquisition component includes a visible light camera, a near-infrared detector, and a light source. The visible light camera, the near-infrared detector, and the light source are electrically connected with the control unit. The stirring component includes a stirring barrel and a stirring group. The stirring barrel is in communication with the plurality of raw material conveying belts. The stirring group includes a stirring shaft, stirring blades, an optical fiber strain gauge, and a force sensor. The stirring blades are arranged along the circumference of the stirring shaft. The optical fiber strain gauge is arranged on the stirring blades. The force sensor is arranged at the input end of the stirring shaft. The force sensor is used to detect the torque of the stirring shaft. The optical fiber strain gauge and the force sensor are electrically connected with the control unit. The feedback component includes a camera. The camera is arranged above the stirring barrel.

[0214] In the embodiment, preferably, the raw material conveying belt is designed with a flexible low-friction surface material and a multi-section buffer structure. The surface is covered with anti-slip texture and configured with an elastic support layer to reduce mechanical impact through an adaptive pressure dispersion mechanism. The conveying process matches the brittleness threshold of different raw materials through a segmented speed control algorithm, isolates the contact between raw materials using independent channels, avoids extrusion deformation, and suppresses the conveying belt vibration frequency outside the inherent frequency safety range of the material. In combination with the edge guide bar to guide the raw materials to be distributed centrally, preventing morphological damage caused by lateral collision. Different types of tea raw materials are conveyed through multiple raw material conveying belts. The running speed and addition amount of each conveying belt are dynamically regulated by the control unit according to the blending parameters.

[0215] The spectral image acquisition component realizes multi-dimensional feature extraction of tea raw materials through the cooperation of a visible light camera and a near-infrared detector. The visible light camera is used to capture visual features such as the surface morphology, color, and lump distribution of the raw materials. The near-infrared detector analyzes the internal properties of the raw materials, such as moisture content and chemical composition, based on different wavelength reflection characteristics. The light source uses a specific wavelength combination of the lighting module to enhance the distinguishability of the feature signals.

[0216] The optical fiber strain gauge in the stirring component reflects the stirring resistance and material rheological properties in real time by detecting the micro-deformation data of the stirring blade in the material flow field and combining the input torque of the stirring shaft measured by the force sensor. The optical fiber strain gauge uses a distributed layout strategy to capture the force differences at different positions of the blade, and the force sensor eliminates mechanical transmission errors through a dynamic calibration algorithm.

[0217] The camera in the feedback component can be a spectral camera, and the number of cameras can be multiple, covering different visual angle areas at the top of the stirring barrel. The camera analyzes the bulk density distribution and leaf suspension state of the material through multi-spectral feature fusion, and the time sequence images collected by the camera are used to construct a dynamic three-dimensional feature field of the stirring process.

[0218] The control unit establishes a correlation model of raw material properties, stirring state, and product quality by integrating the multi-source data of the spectral image acquisition component, the mechanical feedback of the stirring component, and the visual information of the feedback component, and realizes closed-loop optimization of the blending parameters based on the method of the first aspect. The wavelength range of the light source of the spectral image acquisition component is customized according to the spectral absorption characteristics of tea raw materials, and the characteristic wavelength band of the near-infrared detector is determined through principal component analysis and calibration tests of key chemical components of tea. The sensitivity threshold of the optical fiber strain gauge is obtained through baseline calibration in the stirring empty state, ensuring that the deformation signal is linearly related to the actual rheological resistance of the material. The multi-spectral camera of the feedback component eliminates the imaging distortion caused by the rolling of the material in the stirring barrel through multi-view image stitching and feature alignment algorithms, and its spectral channel configuration is complementary to that of the spectral image acquisition component, enhancing the quantitative evaluation capability of the mixing uniformity.

[0219] The embodiment realizes fine proportioning control of tea mixing through multi-component cooperation. The raw material conveying belt adopts flexible low-friction surface material and multi-section buffer structure design, combines segmented speed control algorithm and independent channel isolation mechanism, and effectively maintains the integrity of the physical form of the raw materials during the conveying process. The spectral image acquisition component extracts multi-dimensional features through a visible light camera and a near-infrared detector, combines a customized light source to enhance signal recognition, and accurately obtains the apparent characteristics and inherent properties of the raw materials. The stirring component analyzes the rheological properties of the material flow and the change of stirring resistance in real time based on the mechanical feedback data of distributed optical fiber strain gauges and dynamic calibration force sensors. The feedback component uses a multispectral camera cluster to cover multiple visual angle areas of the stirring barrel, evaluates the mixing uniformity through multispectral feature fusion and three-dimensional dynamic feature field construction technology. The control unit integrates spectral features, mechanical parameters and visual feedback information, establishes a raw material characteristics-stirring state-finished product quality correlation model, forms a regulation and control system with proportioning parameter closed-loop optimization as the core, realizes the comprehensive improvement of lump inhibition, broken leaf rate control and ingredient uniformity in the tea mixing process, and ensures the consistency of finished product quality.

[0220] In a third aspect, the embodiment also provides a computer-readable storage medium having stored thereon computer program instructions, which when executed by a processor, implement the method of the first aspect.

[0221] The computer program involved in the embodiment can be stored in a computer device readable storage medium, including but not limited to magnetic disk, magnetic tape, magnetic card, floppy disk, flash memory, optical disc, optical card, read-only memory (ROM), random access memory (RAM), erasable programmable ROM (EPROM) and electrically erasable programmable ROM (EEPROM) and the like, and also includes other biological, physical or chemical structures that can realize similar or equivalent functions as the above-mentioned storage media, such as DNA, RNA, protein and the like units with information storage ability. In specific embodiments, the storage medium can be one of the above-mentioned medium types, or a combination of the above-mentioned medium types. In different embodiments, the computer program involved in the embodiment can be centrally stored in a single medium, or can be distributedly stored in multiple media. The memory containing the computer device readable storage medium can be a non-volatile memory or a random access memory. These computer device readable storage media can be built-in in the device, or connected with the device as an external device or part of the external device. In some embodiments, the memory with the computer device readable storage medium is deployed locally; in other embodiments, the memory can also be deployed remotely from the processor, such as network attached storage accessed via RF circuit or external port and communication network, wherein the communication network can be Internet, one or more intranets, local area network (LAN), wide area network (WAN), storage area network (SAN) and the like, or appropriate combination thereof, as long as the access of the computer device to the memory can be realized. In addition, the computer program involved in the embodiment can be stored in plaintext / ciphertext form, or can be designed as training data, and integrated and reorganized in the parameter state of the deep neural network or other machine learning model by means of model training.

[0222] Compared with the prior art, the above technical scheme has the following beneficial effects:

[0223] The tea mixing proportion automatic adjusting method provided by the application constructs a closed-loop control system through multi-modal spectral feature extraction and dynamic feedback mechanism, which significantly improves the proportioning accuracy and process stability. Based on the cross-modal feature fusion of visible light image and near-infrared spectrum, the transfer learning model is used to reuse the general morphological feature extraction ability of coffee bean grading, combined with the tea special network layer to realize the accurate analysis of chemical composition, and overcome the misjudgment of single modal physicochemical indicators; through the PLS regression model, the multi-dimensional heterogeneous characteristics are mapped to the standard process parameters, and the quantitative correlation between the raw material characteristics and the mixing demand is established; based on the multi-objective optimization algorithm, the Pareto optimal proportioning parameters are generated among the cost, quality and process constraints, and the economic efficiency and the target flavor fitting degree are balanced; in the mixing process, through the real-time monitoring data of stirring resistance and material distribution, the hierarchical adjustment strategy is used to dynamically correct the proportioning parameters, and the mismatch problem of static proportioning and dynamic rheological characteristics is solved. The method enhances the adaptability of the model to different raw materials through the transfer learning mechanism, the closed-loop feedback system suppresses the abnormal working conditions such as caking and broken leaves, the fuzzy reasoning module converts the sensory description into quantifiable indicators, realizes the collaborative optimization of artificial experience and machine decision, and finally synchronously improves the raw material utilization rate, quality consistency and process controllability in the tea mixing process.

[0224] The above only describes some embodiments of the application, and does not limit the protection scope of the application, and any equivalent device or equivalent process transformation obtained by using the content of the specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the application.

Claims

1. A method for automatically adjusting the proportion of tea ingredients, characterized in that, The method comprises: obtaining spectral image information of a plurality of tea raw materials, the spectral image information comprising a visible light image and a near-infrared spectrum image; inputting the spectral image information into a transfer learning model for feature extraction to obtain tea raw material feature information, the tea raw material feature information comprising color, texture, integrity, water content and phenol-ammonia ratio; the transfer learning model is configured to be obtained by training the following steps: obtain a pre-trained coffee bean grading model, and retain the bottom convolutional network structure of the coffee bean grading model, the bottom convolutional network structure comprising at least three convolution-pooling modules connected in series; input sample spectral image information in a sample database into the bottom convolutional network structure to obtain sample tea general features; perform adversarial training on the sample tea general features through a gradient reversal layer to make the sample tea general features satisfy the field invariance; construct a tea special processing layer and splice it at the output end of the bottom convolutional network structure, the tea special processing layer comprising a visible light branch processing layer and a near-infrared branch processing layer; input the sample tea general features and sample spectral image information into the tea special processing layer to obtain feature output results, the feature output results comprising first sample branch features and second sample branch features; fuse the first sample branch features and the second sample branch features through a cross-modal attention mechanism to obtain sample fusion features; and optimize the parameters of the tea special processing layer through supervised learning, comprising: minimizing the classification error of the first sample branch features; minimizing the spectral reconstruction error of the second sample branch features; splice the sample fusion features and the sample tea general features to form sample tea raw material feature information; input the tea raw material feature information into a PLS regression model to obtain raw material parameter information; obtain information of a blended tea, the information of the blended tea comprising a tea category and tea characteristics after blending; generate target parameter information according to the information of the blended tea; input the target parameter information and the raw material parameter information into a multi-objective optimization algorithm model for calculation to obtain initial blending parameters; blend a plurality of tea raw materials according to the initial blending parameters; perform the following steps at a predetermined frequency: obtain stirring information in real time during the blending process, the stirring information comprising stirring resistance values and stirring images; generate feedback information according to the stirring information, generate an adjustment strategy according to the feedback information, and adjust the initial blending parameters according to the adjustment strategy, denoted as corrected blending parameters; blend a plurality of tea raw materials according to the corrected blending parameters; until the tea raw materials are blended.

2. The tea blending automatic adjusting method according to claim 1, characterized in that, constructing a tea special processing layer and splicing it at the output end of the bottom convolutional network structure comprises: the visible light branch processing layer and the bottom convolutional network structure are connected through a feature map channel splicing method of the sample tea general features to obtain first branch input information; construct a first dilated convolution module, the dilated rate of the first dilated convolution module being positively correlated with the average distance of tea veins of the first branch input information; The near-infrared branch processing layer receives a sample near-infrared spectral image in sample spectral image information, and fuses the sample near-infrared spectral image with the sample tea general characteristics through convolution to obtain second branch input information; A second expansion convolution module is constructed, and a convolution kernel size of the second expansion convolution module matches a tea polyphenol characteristic absorption peak bandwidth of the second branch input information.

3. The tea blending automatic adjusting method according to claim 2, characterized in that, The spectral image information is input into a transfer learning model for feature extraction to obtain tea material characteristic information including: The visible light image is skeletonized to obtain a vein binary image, and a vein spacing characteristic quantity is obtained, which is represented by formula (1) as follows: ； In formula (1), is an inter-vein distance feature quantity, is the number of effective vein pairs, is the first center point coordinate of the adjacent vein after the skeletonization processing, is the second center point coordinate of the adjacent vein after the skeletonization processing. The expansion rate of the first expansion convolution module is adjusted according to the vein spacing characteristic quantity, which is represented by formula (2) as follows: ； In equation (2), is a scaling factor, is a reference distance, is an adjustment function, is an expansion rate is a lower limit value, is an upper limit value, is an expansion rate, is an expansion rate; Tea polyphenol characteristic peak parameters of the near-infrared spectral image are extracted, which are represented by formula (3) as follows: ； In Equation (3), is a tea polyphenol characteristic peak parameter, is a first half-height width boundary, is a second half-height width boundary; The convolution kernel size of the second expansion convolution module is generated according to the characteristic peak parameters, which is represented by formula (4) as follows: ； In equation (4), is the size of the convolution kernel, is the spectral resolution, is the floor function.

4. The tea blending automatic adjusting method according to claim 1, characterized in that, The tea material characteristic information is input into a PLS regression model to obtain material parameter information including: The tea material characteristic information is standardized to obtain a material characteristic vector; The material characteristic vector is projected into a PLS model latent variable space, and a plurality of latent variable scores are obtained by iterative solution; The plurality of latent variable scores are converted into standard material parameters by a regression equation; The standard material parameters are restored and converted to obtain the material parameter information.

5. The method for automatically adjusting the tea mixture ratio according to claim 1, characterized in that, The multi-objective optimization algorithm model is configured to be constructed based on a multi-objective genetic algorithm; The target parameter information and the material parameter information are input into the multi-objective optimization algorithm model for calculation to obtain initial blending parameters including: An optimization objective function is constructed, which is represented by formula (5) as follows: ； In formula (5), is a flavor matching function, is a raw material utilization function, is a process handling function, is a ratio vector, is a predicted flavor feature, is a target flavor feature, is an Euclidean distance between the predicted flavor feature and the target flavor feature, by adjusting the ratio to minimize the difference between the predicted flavor feature and the target flavor feature, is a raw material cost coefficient, is a raw material proportion, is a raw material moisture content, is a maximum allowable difference of equipment, is a target raw material proportion, is a target raw material moisture content; The multi-objective genetic algorithm is used to solve the optimization objective function, the input parameters of the multi-objective genetic algorithm are the material parameter information and the target parameter information, and the variables of the multi-objective genetic algorithm are defined as the mass percentages of tea materials in the blending process; A blending set is obtained, the blending set includes a plurality of blending arrays, and each blending array has a blending score; The blending array with the highest blending score is selected as the initial blending parameters, and the initial blending parameters include the feeding proportions and the feeding sequences of each tea material.

6. The method for automatically adjusting the tea mixture ratio according to claim 1, wherein The adjustment strategy includes a basic adjustment mode and an advanced adjustment mode; Feedback information is generated according to the stirring information, an adjustment strategy is generated according to the feedback information, and the initial blending parameters are adjusted according to the adjustment strategy, which are denoted as corrected blending parameters including: The stirring resistance value is subjected to a sliding average filtering to calculate a current bulk density characteristic value, and the stirring image is subjected to image processing to obtain a caking area and a broken leaf rate; The bulk density characteristic value, the caking area and the broken leaf rate are arranged into actual indexes; It is sequentially judged whether the actual indexes meet the range of a preset index threshold value; If one of the actual indexes exceeds the preset index threshold value, the basic adjustment mode is switched to. If two or more actual indexes exceed the preset index threshold, the advanced adjustment mode is switched.

7. The method for automatically adjusting tea mixture ratio according to claim 6, characterized in that, The basic adjustment mode is configured to: calculate the deviation information of the actual indexes exceeding the preset index threshold to generate first deviation information; convert the first deviation information into the modified proportioning parameters; The advanced adjustment mode is configured to: calculate the deviation information of the actual indexes exceeding the preset index threshold to generate second deviation information; obtain the actual index category to which the maximum deviation value in the second deviation information belongs; If the actual index category is the agglomerate area, the tea raw material added so far is recorded as abnormal tea raw material, and it is determined whether the abnormal tea raw material in the initial proportioning parameters still exists in the subsequent process, if so, the added mass and / or the added times of the abnormal tea raw material is reduced; If the actual index category is the broken leaf rate, the torque and the rotating speed of the stirring group are adjusted according to the change trend of the broken leaf rate; If the actual index category is the bulk density, the adding speed of the tea raw material is adjusted according to the change trend of the bulk density.

8. A tea mixing automatic adjusting system, characterized in that, Comprise: a control unit for executing the method of any one of claims 1-7; a plurality of raw material conveying belts, each of which is provided with a kind of tea raw material; a spectral image acquisition assembly arranged above the raw material conveying belt, the spectral image acquisition assembly comprising a visible light camera, a near-infrared detector and a light source, the visible light camera, the near-infrared detector and the light source being electrically connected to the control unit respectively; a stirring assembly comprising a stirring barrel and a stirring group, the stirring barrel being in communication with a plurality of raw material conveying belts, the stirring group comprising a stirring shaft, stirring blades, an optical fiber strain gauge and a force sensor, the stirring blades being arranged along the circumference of the stirring shaft, the optical fiber strain gauge being provided on the stirring blades, the force sensor being arranged at the input end of the stirring shaft, the force sensor being used to detect the torque of the stirring shaft, the optical fiber strain gauge and the force sensor being electrically connected to the control unit respectively; a feedback assembly comprising a camera, the camera being arranged above the stirring barrel.

9. A computer readable storage medium having stored thereon computer program instructions, characterized in that, The computer program instructions, when executed by a processor, implement the method of any one of claims 1 to 7. The computer program instructions, when executed by a processor, implement the method of any one of claims 1 to 7.

Images