Method, system and electronic device for cross variety prediction of soy-based low moisture gel quality

Abstract

The application discloses a soybean-based low-moisture gel quality cross-species prediction method and system and electronic equipment, and belongs to the technical field of grain by-product quality inspection and food processing. The method comprises the following steps: acquiring the grain physicochemical indexes and gel product comprehensive score F of a plurality of known soybean varieties, and constructing a basic database; taking the grain physicochemical indexes as input characteristics and F as a label, training a deep neural network to obtain an existing variety model M0, and adding a variety-independent feature representation layer during training to make the output distribution of different varieties consistent; collecting the grain physicochemical indexes and F of a preset number of samples of a new target variety, and calculating a variety similarity measure D; adaptively determining the number of frozen layers of M0 according to D, adjusting the unfrozen layers and the output layer with the data of the new target variety, and obtaining a new target variety prediction model M1; and inputting the indexes of the raw materials to be inspected into M1 to output a prediction result. The application can realize cross-species quality prediction of soybean-based low-moisture gel under the condition of a small amount of new samples.

Description

Technical Field

[0001] This invention relates to the field of quality inspection of grain by-products and food processing technology. More specifically, this invention relates to a method, system, and electronic device for cross-variety prediction of the quality of soybean-based low-moisture gels, such as dried tofu, tofu jerky, vegetarian meat, or high-fiber protein gels. Background Technology

[0002] Soy-based low-moisture gels include soy-based protein products or soy products such as dried tofu, tofu jerky, vegetarian meat, or high-fiber protein gels. Dried tofu, a traditional specialty soy product in northern my country, owes its final product quality, including textural characteristics, yield, and flavor, to a large extent depending on the variety of soybeans and their inherent physicochemical properties. my country has abundant soybean germplasm resources, and different varieties exhibit significant differences in physicochemical indicators such as protein content, 11S / 7S globulin ratio, phytic acid content, and the content of minerals such as calcium and magnesium. These differences directly affect protein gel formation, network structure construction, and the final product's hardness, elasticity, and chewiness during the dried tofu processing.

[0003] However, the soybean processing industry currently faces two major challenges in raw material screening: First, there is a lack of a systematic evaluation system. Existing methods mostly rely on single empirical indicators such as protein content or small-scale trial experiments, which cannot comprehensively and quantitatively reveal the relationship between the complex characteristics of soybean seeds and the multi-dimensional quality of dried tofu products. Second, there is a lack of predictive ability. Enterprises cannot accurately predict the overall quality of finished products based solely on raw material testing data before purchasing or starting production. They can only screen through the traditional cycle of trial production and testing, which is inefficient and costly.

[0004] Existing research has attempted to link the quality of soybean raw materials and tofu products by establishing mathematical models. For example, patent CN120181635A discloses a method, device, and electronic device for predicting the quality of tofu products made from soybean raw materials. This method acquires quality index data of the soybean raw materials to be tested, inputs the quality index data into a tofu product quality prediction model, and obtains the tofu product quality prediction result. The prediction model is trained based on the quality index data of soybean raw materials as training samples and the quality index data of tofu products processed from soybean raw materials as labels. This approach improves the accuracy of tofu quality prediction to a certain extent.

[0005] However, the existing technologies mentioned above still have the following shortcomings: First, their prediction targets are the separate outputs of multi-dimensional indicators, lacking a unified, objective, and quantifiable comprehensive quality evaluation standard. When multiple quality indicators show a trade-off, decision-makers find it difficult to make a comprehensive judgment when faced with multiple independent predicted values. Second, their model training and application both rely on a large amount of historical data of the same or mixed varieties, as well as a large amount of raw material indicator input data for model prediction (input indicators > 10). When facing a completely new soybean variety, such varieties often only have a small number of samples available for testing, lack historical data accumulation, and cannot perform a large number of raw material indicators (more than 10 indicators such as protein, isoflavones, and saponins) detection for model prediction. The existing models will directly fail. If data is collected from scratch and the model is trained in the traditional way, it usually requires a data accumulation period of more than six months, which seriously restricts the raw material screening efficiency and market response speed of enterprises.

[0006] Therefore, how to establish an objective and quantifiable comprehensive evaluation standard specifically for soybean-based low-moisture gel products, and how to achieve knowledge transfer from known varieties to new varieties and quickly build high-precision prediction models based on only a very small amount of new sample data, have become technical problems that urgently need to be solved in this field. Summary of the Invention

[0007] One object of the present invention is to provide a cross-variety prediction method for the quality of soybean-based low-moisture gels, comprising the following steps:

[0008] S1. Obtain the physicochemical indicators of soybean grains and the corresponding soybean-based low moisture gel comprehensive score F for multiple known soybean varieties, and construct a multi-variety basic database. The physicochemical indicators of the grains include moisture, protein, fat, 11S globulin, 7S conglobulin, 11S / 7S ratio, water-soluble protein, protein solubility, phytic acid, lectins, calcium, and magnesium.

[0009] S2. Using the physicochemical indicators of each variety's grains as input features and F as the label, train a deep neural network to obtain the existing variety model M0. During training, add a variety-independent feature representation layer in the hidden layer. The training objective of the feature representation layer is to make the output distribution of different varieties in this layer tend to be consistent and to extract cross-variety common features.

[0010] S3. For the new target variety, collect a predetermined number of samples for grain physicochemical indicators and F, as small sample data for the new target variety, and construct a vector from the grain physicochemical indicators of the new target variety. The center vector of the p-th variety among the existing varieties Calculate Euclidean distance Where j=1,2,…,12, p=1,2,…,P, P is the total number of varieties in the multi-variety basic database, and the variety similarity measure is D=min(d p );

[0011] S4. Keep the parameters of the frozen hidden layer in M0 unchanged, adjust the parameters of the unfrozen hidden layer and the output layer with the new target variety data, introduce regularization constraints and set an early stop mechanism in the loss function to obtain the new target variety prediction model M1.

[0012] S5. For the remaining raw materials to be tested for the new target variety, input the measured physicochemical indicators of the grains into M1 and output the comprehensive quality prediction results.

[0013] Preferably, the comprehensive score F of soybean-based low-moisture gel is obtained through the following steps:

[0014] Quality index data of soybean-based low-moisture gel products obtained from multiple varieties of soybean raw materials through standardized processes were obtained, including 15 quality indicators such as moisture content, protein content, fat content, calcium content, magnesium content, hardness, adhesiveness, resilience, cohesiveness, elasticity, adhesion, chewiness, tensile strength, shear force, and overall sensory acceptability.

[0015] KMO test and Bartlett's test of sphericity were performed on the 15 quality index data. When the KMO test value was >0.6 and the significance level of Bartlett's test of sphericity was <0.05, the data were deemed suitable for principal component analysis.

[0016] The quality index data that passed the inspection were standardized using Z-score, with the mean of each quality index data being 0 and the standard deviation being 1.

[0017] Principal component analysis was performed on the standardized data to extract the top K principal components whose eigenvalues ​​were greater than 1 and whose cumulative variance contribution rate was ≥75%. The variance contribution rate λ of each principal component was then obtained. i and the factor loading matrix, where i = 1, 2, ..., K;

[0018] The scores Z of each principal component are constructed based on the factor loading matrix. i The function expression: , where a ij Let X be the loading coefficient of the i-th principal component on the j-th quality index. j Let j be the standardized value of the j-th quality indicator, where j = 1, 2, ..., 15;

[0019] The weights are determined based on the variance contribution rate of each principal component, and the original comprehensive score calculation formula is constructed as follows: ;

[0020] For all varieties of F raw Perform deviation standardization and map to the 0-1 interval: , of which F raw1 and F raw2The original comprehensive score F for all varieties is as follows: raw The minimum and maximum values.

[0021] Preferably, the number of components K is 6, and the variance contribution rates of the 6 principal components are 22.948%, 14.391%, 12.866%, 11.151%, 8.050%, and 7.473%, respectively, with a cumulative variance contribution rate of 76.879%. The factor loading matrices and functional expressions of the 6 principal components are as follows:

[0022] The score of the first principal component Z1 = 0.015X1 - 0.471X2 + 0.808X3 + 0.599X4 + 0.778X5 + 0.605X6 + 0.109X7 - 0.513X8 - 0.549X9 + 0.222X 10 -0.311X 11 +0.241X 12 -0.008X 13 +0.668X 14 -0.101X 15 ;

[0023] The score of the second principal component Z2 = 0.304X1 + 0.320X2 + 0.031X3 + 0.002X4 + 0.108X5 - 0.151X6 + 0.246X7 + 0.720X8 + 0.739X9 - 0.032X 10 +0.429X 11 +0.054X 12 +0.266X 13 -0.365X 14 +0.249X 15 ;

[0024] The score of the third principal component Z3 = 0.041X1 - 0.071X2 - 0.030X3 - 0.577X4 - 0.432X5 + 0.514X6 - 0.776X7 + 0.159X8 + 0.060X9 + 0.466X 10 -0.032X 11 +0.595X 12 +0.097X 13 +0.208X 14 +0.362X 15 ;

[0025] The score of the fourth principal component Z4 = 0.704X1 - 0.144X2 - 0.210X3 - 0.288X4 - 0.007X5 + 0.225X6 - 0.027X7 + 0.278X8 + 0.124X9 - 0.552X 10 +0.525X 11-0.346X 12 -0.064X 13 +0.148X 14 +0.263X 15 ;

[0026] The score of the 5th principal component Z5 = 0.113X1 + 0.566X2 + 0.005X3 + 0.226X4 - 0.028X5 - 0.166X6 + 0.121X7 + 0.175X8 + 0.011X9 + 0.071X 10 -0.032X 11 -0.157X 12 +0.770X 13 +0.098X 14 -0.401X 15 ;

[0027] The score of the 6th principal component Z6 = 0.305X1 - 0.171X2 - 0.510X3 + 0.043X4 + 0.070X5 + 0.134X6 + 0.262X7 + 0.033X8;

[0028] Wherein, X1 is the moisture content of soybean-based low-moisture gel, X2 is the protein content of soybean-based low-moisture gel, X3 is the fat content of soybean-based low-moisture gel, X4 is the calcium content of soybean-based low-moisture gel, X5 is the magnesium content of soybean-based low-moisture gel, X6 is the hardness of soybean-based low-moisture gel, X7 is the adhesiveness of soybean-based low-moisture gel, X8 is the resilience of soybean-based low-moisture gel, X9 is the cohesiveness of soybean-based low-moisture gel, and X... 10 X is a soybean-based low-moisture gel elasticity. 11 For soybean-based low-moisture gel adhesion, X 12 X is a soybean-based low-moisture gel with chewy properties. 13 For soybean-based low-moisture gel stretchability, X 14 For soybean-based low-moisture gel shear force, X 15 Overall sensory acceptability of soybean-based low-moisture gels;

[0029] The comprehensive score is calculated as F = (22.948Z1 + 14.391Z2 + 12.866Z3 + 11.151Z4 + 8.050Z5 + 7.473Z6) / 76.879.

[0030] Preferably, the adaptive freezing layer number determination method for the variety similarity metric D is as follows:

[0031] The existing variety model M0 contains a total of L hidden layers, which are numbered sequentially from the input layer to the output layer as the 1st hidden layer to the Lth hidden layer.

[0032] The number of frozen layers t is determined based on the variety similarity metric D:

[0033] When D ≤ the first distance threshold, t = L-1, freeze the parameters of the first hidden layer to the (L-1)th hidden layer, and only adjust the parameters of the Lth hidden layer and the output layer;

[0034] When D > the first distance threshold and ≤ the second distance threshold, t = L-2, freeze the parameters of the first hidden layer to the (L-2)th hidden layer, and adjust the parameters of the (L-1)th hidden layer, the Lth hidden layer, and the output layer.

[0035] When D > the second distance threshold, t = L-3, freeze the parameters of the first hidden layer to the (L-3)th hidden layer, and adjust the parameters of the (L-2)th hidden layer, the (L-1)th hidden layer, the Lth hidden layer, and the output layer.

[0036] Wherein, the first distance threshold is less than the second distance threshold, and the first and second distance thresholds are preset values ​​based on the Euclidean distance distribution among varieties in the multi-variety basic database.

[0037] Preferably, the regularization constraint introduced in the loss function is L2 regularization, and the expression for the loss function Loss is: Loss = MSE + α × ||W|| 2 Where MSE is the mean squared error between the predicted composite score and the actual composite score, α is the L2 regularization coefficient, and ||W|| 2 The regularization coefficient α is the sum of squares of the L2 norm of the weight parameters of the adjusted layer. The regularization coefficient α is adaptively set according to the number of samples n of the small sample data of the new target variety: α=β / n, where β is the preset basic regularization constant.

[0038] The early termination mechanism is introduced as follows: the small sample data of the new target variety is divided into an adjustment set and a validation set according to a preset ratio. After each iteration, the prediction mean square error on the validation set is calculated. When the prediction mean square error on the validation set no longer decreases in consecutive m iterations, the training is terminated early, where m is the preset number of patience rounds.

[0039] Preferably, the training objective of the feature representation layer in S2 is achieved by maximizing the mean difference, specifically:

[0040] When training an existing variety model M0, the loss function Loss is composed of the prediction error loss. pred Distribution difference loss mmd Weighted composition: Loss = Loss pred +γ×Loss mmd , where γ is the weighting coefficient of the distribution difference loss;

[0041] Loss mmd The calculation method is as follows: Loss mmd =‖μ u-μ v || 2 , where μ u Let μ be the mean vector output by variety u in the feature representation layer. v Let v be the mean vector output by variety v in the feature representation layer. Iterate through all pairwise combinations of varieties in the multi-variety basic database, sum and average them. Variety u and variety v are any two different varieties in the multi-variety basic database.

[0042] Preferably, after obtaining small sample data of the new target variety in S3 and before adjusting the parameters in S4, a variety selection step is also included:

[0043] The physicochemical indicators of the grains from a small sample of the new target variety are input into the existing variety model M0. The output vector of the feature representation layer is extracted, and the cosine similarity between the output vector and the output center vector of each existing variety in the multi-variety basic database is calculated. The top Q existing varieties with the highest cosine similarity are selected as the preferred existing variety set, where Q is the preset number of preferred varieties. When the maximum cosine similarity in the preferred existing variety set is less than the preset similarity threshold, the variety selection is abandoned, and the parameters of the unfrozen hidden layer and the output layer are adjusted using all existing variety data.

[0044] In S4, except for the above-mentioned rollback scenario, only the data corresponding to the preferred existing variety set is used to adjust the parameters of the unfrozen hidden layer and the output layer.

[0045] Preferably, after outputting the overall quality prediction results, the method also includes:

[0046] The prediction results are compared with the preset suitability threshold. When the prediction result is greater than or equal to the suitability threshold, the soybean raw material is determined to be suitable for soybean-based low-moisture gel processing. When the prediction result is less than the suitability threshold, the soybean raw material is determined to be unsuitable for soybean-based low-moisture gel processing, and an evaluation report is output.

[0047] A cross-variety prediction system for soybean-based low-moisture gel quality is provided, configured to execute the aforementioned cross-variety prediction method for soybean-based low-moisture gel quality. The system includes:

[0048] The data acquisition module is configured to acquire the physicochemical indicators of soybean grains and the corresponding soybean-based low-moisture gel comprehensive score F of multiple known soybean varieties, as well as the physicochemical indicators of soybean grains of new target varieties.

[0049] The comprehensive evaluation module is configured to construct a mathematical model for the comprehensive evaluation of the processing suitability of soybean-based low-moisture gel based on the quality index data of soybean-based low-moisture gel products through principal component analysis, and calculate the comprehensive score F of soybean-based low-moisture gel.

[0050] The existing variety training module is configured to train a deep neural network as an existing variety model using the physicochemical indicators of known soybean grains as input features and F as the label. During training, a variety-independent feature representation layer is added to the hidden layer. The training objective of this layer is to make the output distribution of different varieties in this layer tend to be consistent.

[0051] The transfer learning module is configured to calculate the variety similarity metric D between the new target variety and each variety in the existing varieties. Based on D, the number of frozen hidden layers in the existing variety model is adaptively determined. The parameters of the frozen hidden layers are kept unchanged. The parameters of the unfrozen hidden layers and the output layer are adjusted with the data of the new target variety. The loss function introduces regularization constraints and sets an early stopping mechanism to obtain the prediction model of the new target variety.

[0052] The quality prediction module is configured to input the physicochemical indicators of the raw materials to be tested for the new target variety into the new target variety prediction model and output the comprehensive quality prediction results.

[0053] An electronic device is provided, including a processor and a memory storing a computer program, wherein the processor executes the computer program to implement the aforementioned cross-variety prediction method for soybean-based low-moisture gel quality.

[0054] The present invention has at least the following beneficial effects:

[0055] First, this invention constructs a comprehensive evaluation mathematical model for the suitability of soybean-based low-moisture gel processing using principal component analysis. After KMO and Bartlett tests, the 15 multidimensional quality indicators of soybean-based low-moisture gel products are objectively quantified into a single comprehensive score F in the range of 0 to 1, with the variance contribution rate of the top K principal components whose eigenvalues ​​are greater than 1 and whose cumulative variance contribution rate is not less than 75% as weights. This solves the problem of the lack of a unified, objective, and quantifiable comprehensive quality evaluation standard for soybean-based low-moisture gel products in the prior art, provides standardized label data for the training of the prediction model, and improves the objectivity of model training.

[0056] Secondly, this invention adds a variety-independent feature representation layer to the hidden layer of the existing variety model M0, uses the maximum mean difference as a distribution difference metric to constrain the output distribution of different varieties in this layer to tend to be consistent, and adaptively determines the number of the model's frozen hidden layers based on the variety similarity metric D calculated based on the Euclidean distance between the new target variety and the existing varieties. The model adaptation can be completed with only a preset number of new target variety samples, which solves the problem of model failure caused by lack of historical data when facing new varieties in the existing technology, and shortens the traditional modeling process that requires a large amount of historical data accumulation to rapid modeling under the condition of a limited number of samples.

[0057] Third, this invention uses the cosine similarity output by the feature representation layer to select varieties before parameter adjustment. It only uses existing variety data with similar feature patterns to the new target variety for parameter adjustment and sets a similarity threshold backoff mechanism to avoid screening failure in extreme cases. At the same time, it introduces L2 regularization constraints and early stopping mechanism into the loss function. The regularization coefficient is adaptively set according to the number of samples, which improves the efficiency and stability of parameter adjustment under conditions with few samples.

[0058] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0059] Figure 1 This is a schematic diagram of the overall process of the cross-variety prediction method for soybean-based low-moisture gel quality of the present invention.

[0060] Figure 2 This is a schematic diagram of the training process and network structure of the existing variety model M0 of the present invention;

[0061] Figure 3 This is a schematic diagram of the cross-variety transfer learning process of the present invention;

[0062] Figure 4 This is a schematic diagram of the cross-variety prediction system for soybean-based low-moisture gels of the present invention. Detailed Implementation

[0063] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0064] It should be noted that, unless otherwise specified, the experimental methods described in the following embodiments are all conventional methods, and the reagents and materials described are all commercially available unless otherwise specified. In the description of this invention, the orientation or positional relationship indicated by the terms is based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing this invention and simplifying the description. It does not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0065] This invention provides a cross-variety prediction method for the quality of soybean-based low-moisture gels, comprising the following steps:

[0066] S1. Obtain the physicochemical indicators of soybean grains and their corresponding soybean-based low-moisture gel electrophoresis comprehensive scores (F) from multiple known soybean varieties to construct a multi-variety basic database. The physicochemical indicators of the grains include moisture, protein, fat, 11S globulin, 7S paraglobulin, 11S / 7S ratio, water-soluble protein, protein solubility, phytic acid, lectins, calcium, and magnesium. Specifically, the known soybean varieties should cover the main cultivated varieties in Northeast China, the Huang-Huai-Hai Plain, and Southern China, with more than 20 varieties, such as 34, to ensure sufficient variety diversity and representativeness in the basic database. The determination methods for the physicochemical indicators of the grains are as follows: moisture content can be determined according to the method specified in GB 5009.3-2016; protein content can be determined according to the method specified in GB 5009.5-2016 using a protein conversion factor of 5.71; and fat content can be determined according to GB... The Soxhlet extraction method specified in GB 5009.6-2016 is used to determine the content of 11S globulin and 7S conglobulin. The content can be determined by SDS-PAGE electrophoresis. The 11S / 7S ratio is obtained by dividing the 11S globulin content by the 7S conglobulin content. The content of water-soluble protein can be determined according to the method specified in NY / T1205-2006. The protein solubility is the percentage obtained by dividing the water-soluble protein content by the total protein content. The phytic acid content can be determined by colorimetric method. The lectin content can be determined by enzyme-linked immunosorbent assay (ELISA). The calcium and magnesium contents can be determined by inductively coupled plasma mass spectrometry (ICP-MS) specified in GB 5009.268-2016. The above-mentioned measuring equipment can be selected from laboratory analytical equipment such as Kjeldahl nitrogen analyzer, Soxhlet extraction device, electrophoresis apparatus, ELISA reader, and ICP-MS. The comprehensive score F for soybean-based low-moisture gel is calculated using the mathematical model for comprehensive evaluation of the processing suitability of soybean-based low-moisture gel, described later. This comprehensive score F is a single quantifiable evaluation value obtained by reducing the dimensionality of the multidimensional quality indicators of soybean-based low-moisture gel products through principal component analysis. The higher the comprehensive score F, the better the overall quality of the soybean-based low-moisture gel processed from that soybean variety. Each record in the multi-variety basic database contains the sample measurements and mean values ​​of 12 physicochemical indicators of a known soybean variety, as well as the corresponding comprehensive score F for that variety.

[0067] S2. Using the physicochemical indices of each variety's grains as input features and F as the label, train a deep neural network to obtain the existing variety model M0. During training, add a variety-independent feature representation layer to the hidden layers. The training objective of the feature representation layer is to make the output distribution of different varieties in this layer tend to be consistent, extracting common features across varieties. Specifically, the deep neural network can adopt a multilayer perceptron structure. The number of nodes in the input layer can be set to 12, corresponding to the 12 grain physicochemical indices, and the number of nodes in the output layer can be set to 1, outputting the predicted value of the comprehensive score F. The number of hidden layers can be set to 3 to 5 layers, for example, 4 layers, with the number of nodes in each hidden layer set to 64, 128, 64, and 32 respectively. The activation function can be the ReLU function. The optimizer used during training can be the Adam optimizer, with an initial learning rate of 0.001, 500 training epochs, and a batch size of 16. The training data can be a multi-variety basic database divided into training and testing sets according to a preset ratio (e.g., 8:2). Variety-independent feature representation layers can be located in the penultimate or penultimate hidden layer of the deep neural network. The concept of convergence means constraining the differences in output distributions of different varieties in the feature representation layer, enabling the model to learn common quality mapping patterns independent of variety origin, rather than completely eliminating variety differences. This constraint can be achieved using distribution difference measures such as maximum mean difference. After training, the trained deep neural network is saved as an existing variety model M0. The existing variety training module can be built based on deep learning frameworks such as PyTorch or TensorFlow.

[0068] S3. For the new target variety, collect a predetermined number of samples for grain physicochemical indicators and F, as small sample data for the new target variety, and construct the grain physicochemical indicators of the new target variety as a vector x=(x1,…,x…). 12 ), where c is the center vector of the p-th variety among the existing varieties. p =(c p1 ,…,c p12 ), calculate Euclidean distance Where j=1,2,…,12, p=1,2,…,P, P is the total number of varieties in the multi-variety basic database, and the variety similarity measure is D=min(d p Specifically, the preset number can be set according to the availability of samples of the new target variety and the modeling accuracy requirements. For example, it can be set to 5 to 10, where x1 to x2 in vector x are the samples. 12 The measurements, in order, are for moisture, protein, fat, 11S globulin, 7S conglobulin, 11S / 7S ratio, water-soluble protein, protein solubility, phytic acid, lectins, calcium, and magnesium. (Central vector c) p =(c p1 ,…,c p12The similarity metric D is calculated by taking the average of 12 physicochemical indicators of the p-th variety from existing varieties, where j = 1, 2, ..., 12, p = 1, 2, ..., P, and P is the total number of existing varieties in the multi-variety database. The physical meaning of the variety similarity metric D is the shortest distance between the new target variety and each of the existing varieties in the original physicochemical indicator space. A smaller D value indicates that the physicochemical characteristics of the new target variety are closer to those of an existing variety, requiring more hidden layers of the existing variety model M0 to be frozen in subsequent transfer learning, thus preserving more existing knowledge. Conversely, a larger D value indicates a greater difference in physicochemical characteristics between the new target variety and existing varieties, requiring fewer layers to be frozen, allowing more parameters to be adjusted to adapt to the specific characteristics of the new variety.

[0069] S4. Keeping the parameters of the frozen hidden layers in M0 unchanged, adjust the parameters of the unfrozen hidden layers and the output layer using the new target variety data. Introduce regularization constraints and set an early stopping mechanism in the loss function to obtain the new target variety prediction model M1. Specifically, the adaptive determination of the number of frozen layers is achieved through the following rules: The existing variety model M0 contains a total of L hidden layers, numbered sequentially from the input layer to the output layer as the 1st to the Lth hidden layers. The first distance threshold and the second distance threshold are preset values ​​based on the Euclidean distance distribution between varieties in the multi-variety basic database. The first distance threshold can be set to the 25th percentile of the Euclidean distance between all pairs of varieties in the multi-variety basic database, and the second distance threshold... The value can be set to the 75th percentile. When D is less than or equal to the first distance threshold, the number of frozen layers t is equal to L-1. The parameters of the first to L-1 hidden layers remain unchanged, and only the parameters of the Lth hidden layer and the output layer are adjusted. When D is greater than the first distance threshold and less than or equal to the second distance threshold, the number of frozen layers t is equal to L-2. The parameters of the first to L-2 hidden layers remain unchanged, and the parameters of the L-1, Lth, and output layers are adjusted. When D is greater than the second distance threshold, the number of frozen layers t is equal to L-3. The parameters of the first to L-3 hidden layers remain unchanged, and the parameters of the L-2, L-1, Lth, and output layers are adjusted. The optimizer used during parameter adjustment can be the Adam optimizer, and the learning rate can be set to 0.0001, lower than the initial learning rate in step S2, to finely adjust the existing model with a small step size. The regularization constraint can be L2 regularization. The early termination mechanism can be achieved by dividing the small sample data of the new target variety into an adjustment set and a validation set according to a preset ratio, and using the fact that the error of the validation set no longer decreases as the termination condition.

[0070] S5. For the remaining raw materials to be tested for the new target variety, input the measured physicochemical indicators of the grains into M1 and output the comprehensive quality prediction results. Specifically, the measurement method is the same as in step S1. The prediction results output by the model are values ​​in the range of 0 to 1. The closer to 1, the better the predicted comprehensive quality of soybean-based low-moisture gel. The closer to 0, the worse the predicted comprehensive quality of soybean-based low-moisture gel.

[0071] Using the above method, even with only a preset number of new target variety samples, the model freezing strategy can be adaptively determined through variety similarity measurement and transfer learning can be performed to quickly obtain a high-precision prediction model applicable to the new target variety. This allows the modeling process, which originally required a large amount of historical data accumulation, to be completed under conditions of a small number of samples, thereby improving the raw material screening efficiency of soybean-based low-moisture gel production enterprises when facing new soybean varieties.

[0072] In another technical solution, the comprehensive score F of soybean-based low-moisture gel is calculated using the following mathematical model for comprehensive evaluation of the processing suitability of soybean-based low-moisture gel:

[0073] Quality index data for soybean-based low-moisture gel products obtained from multiple soybean varieties using standardized processes were obtained. These included 15 quality indicators: moisture content, protein content, fat content, calcium content, magnesium content, hardness, adhesiveness, resilience, cohesiveness, elasticity, adhesion, chewiness, tensile strength, shear force, and overall sensory acceptability. Specifically, the standardized process refers to using a uniform processing flow for soybean-based low-moisture gels, including steps such as soaking, grinding, boiling, coagulation, curdling, and pressing. Process parameters for each step, such as soaking time, grinding mesh size, boiling temperature and time, type and amount of coagulant, and pressing pressure and time, were kept constant to eliminate the impact of process differences on product quality. The moisture content of the soybean-based low-moisture gel can be determined according to the method specified in GB 5009.3-2016. Hardness, adhesiveness, resilience, cohesiveness, elasticity, adhesion, and chewiness can be obtained by TPA testing on 2 cm cube soybean-based low-moisture gel samples using a texture analyzer with a P / 36R probe. The compression ratio for the TPA test can be set to 40%, and the test rate can be set to 1 mm / s. Tensile properties can be measured using an A / TG tensile clamp on a texture analyzer. Shear force can be measured using an A / CKB shear probe on a texture analyzer. Overall sensory acceptability can be obtained by a trained sensory evaluation team using a 9-point scale, with a scoring standard of 1 point for extremely dislike and 9 points for extremely like. Protein content can be determined according to the method specified in GB 5009.5-2016 with a protein conversion factor of 6.25. Fat content can be determined according to the Soxhlet extraction method specified in GB 5009.6-2016. Calcium and magnesium content can be determined according to the inductively coupled plasma mass spectrometry method specified in GB 5009.268-2016.

[0074] KMO and Bartlett's test of sphericity were performed on the 15 quality indicators. When the KMO value was greater than 0.6 and the Bartlett's test of sphericity significance was less than 0.05, the data was deemed suitable for principal component analysis. Specifically, the KMO test measures the partial correlation between the quality indicators, indicating whether the data is suitable for principal component analysis. The Bartlett's test of sphericity tests whether the correlation matrix between the quality indicators is an identity matrix; when the significance level is less than 0.05, the null hypothesis that the quality indicators are independent is rejected.

[0075] The quality index data that passed the test were standardized using Z-score. The mean of each quality index data was 0 and the standard deviation was 1. Specifically, the standardization method was: X'=(X-μ) / σ, where X is the original measured value, μ is the mean of the quality index among all varieties, and σ is the standard deviation of the quality index among all varieties.

[0076] Principal component analysis was performed on the standardized data to extract the top K principal components whose eigenvalues ​​were greater than 1 and whose cumulative variance contribution rate was not less than 75%. The variance contribution rate λ of each principal component was then obtained. i And the factor loading matrix, where i=1,2,…,K; specifically, the criterion for eigenvalues ​​greater than 1 follows the Kaiser criterion, indicating that the variance explained by the principal component is greater than the average variance of the individual original indices;

[0077] The scores Z of each principal component are constructed based on the factor loading matrix. i The function expression: , where a ij Let X be the loading coefficient of the i-th principal component on the j-th quality index. j Let Z be the standardized value of the j-th quality indicator, where j = 1, 2, ..., 15; specifically, each principal component score Z... i It is a linear combination of the standardized values ​​of 15 quality indicators, with a loading factor α. ij This reflects the magnitude and direction of the contribution of the j-th quality index to the i-th principal component;

[0078] The weights are determined based on the variance contribution rate of each principal component, and the original comprehensive score calculation formula is constructed as follows: Specifically, the formula uses the variance contribution rate of each principal component as a weight, performs a weighted sum of the principal component scores, and then normalizes the sum by dividing by the sum of the variance contribution rates.

[0079] For all varieties of F raw Perform deviation standardization and map to the 0-1 interval: , of which F raw1 and F raw2 The original comprehensive score F for all varieties is as follows: raw The minimum and maximum values; the closer the standardized F value is to 1, the better the overall quality of the soybean-based low-moisture gel obtained from the processing of this soybean variety.

[0080] By constructing the above-mentioned mathematical model for the comprehensive evaluation of the suitability of soybean-based low-moisture gel processing, the 15 dispersed quality indicators that originally required manual comprehensive judgment were objectively reduced in dimensionality and quantified into a single comprehensive score F in the range of 0 to 1 through principal component analysis. This provides a unified quantitative comparison benchmark for the suitability of soybean-based low-moisture gel processing for different soybean varieties, and provides objective and standard training labels for subsequent prediction models.

[0081] In another technical solution, the number of principal components K is specifically 6, and the variance contribution rates of the 6 principal components are 22.948%, 14.391%, 12.866%, 11.151%, 8.050%, and 7.473% respectively, with a cumulative variance contribution rate of 76.879%.

[0082] Specifically, the factor loading matrices and functional expressions of the six principal components are as follows, where X1 represents the moisture content of the soybean-based low-moisture gel, X2 represents the protein content of the soybean-based low-moisture gel, X3 represents the fat content of the soybean-based low-moisture gel, X4 represents the calcium content of the soybean-based low-moisture gel, X5 represents the magnesium content of the soybean-based low-moisture gel, X6 represents the hardness of the soybean-based low-moisture gel, X7 represents the adhesiveness of the soybean-based low-moisture gel, X8 represents the resilience of the soybean-based low-moisture gel, X9 represents the cohesiveness of the soybean-based low-moisture gel, and X... 10 X is a soybean-based low-moisture gel elasticity. 11 For soybean-based low-moisture gel adhesion, X 12 X is a soybean-based low-moisture gel with chewy properties. 13 For soybean-based low-moisture gel stretchability, X 14 For soybean-based low-moisture gel shear force, X 15 Overall sensory acceptability of soybean-based low-moisture gels;

[0083] The score of the first principal component Z1 = 0.015X1 - 0.471X2 + 0.808X3 + 0.599X4 + 0.778X5 + 0.605X6 + 0.109X7 - 0.513X8 - 0.549X9 + 0.222X 10 -0.311X 11 +0.241X 12 -0.008X 13 +0.668X 14 -0.101X 15 From the perspective of loading coefficients, the first principal component has a large positive loading on fat content, calcium content, magnesium content, hardness and shear force, and a large loading on protein content, resilience and cohesion, reflecting the contrast between textural strength and mineral element content and protein network softness.

[0084] The score of the second principal component Z2 = 0.304X1 + 0.320X2 + 0.031X3 + 0.002X4 + 0.108X5 - 0.151X6 + 0.246X7 + 0.720X8 + 0.739X9 - 0.032X 10 +0.429X 11 +0.054X 12 +0.266X 13 -0.365X 14 +0.249X 15 ;

[0085] The score of the third principal component Z3 = 0.041X1 - 0.071X2 - 0.030X3 - 0.577X4 - 0.432X5 + 0.514X6 - 0.776X7 + 0.159X8 + 0.060X9 + 0.466X 10 -0.032X 11 +0.595X 12 +0.097X 13 +0.208X 14 +0.362X 15 ;

[0086] The score of the fourth principal component Z4 = 0.704X1 - 0.144X2 - 0.210X3 - 0.288X4 - 0.007X5 + 0.225X6 - 0.027X7 + 0.278X8 + 0.124X9 - 0.552X 10 +0.525X 11 -0.346X 12 -0.064X 13 +0.148X 14 +0.263X 15 ;

[0087] The score of the 5th principal component Z5 = 0.113X1 + 0.566X2 + 0.005X3 + 0.226X4 - 0.028X5 - 0.166X6 + 0.121X7 + 0.175X8 + 0.011X9 + 0.071X 10 -0.032X 11 -0.157X 12 +0.770X 13 +0.098X 14 -0.401X 15 ;

[0088] The score of the 6th principal component Z6 = 0.305X1 - 0.171X2 - 0.510X3 + 0.043X4 + 0.070X5 + 0.134X6 + 0.262X7 + 0.033X8 + 0.165X9 + 0.435X 10 +0.042X 11 +0.005X 12 +0.026X 13 -0.454X 14 +0.089X 15 ;

[0089] The comprehensive score calculation formula is: F=(22.948Z1+14.391Z2+12.866Z3+11.151Z4+8.050Z5+7.473Z6) / 76.879.

[0090] The loading coefficients and variance contribution rates mentioned above were calculated using principal component analysis based on quality index data from 34 soybean varieties. These varieties cover the main cultivated varieties in Northeast China, the Huang-Huai-Hai Plain, and Southern China, and are widely representative. When processing soybean-based low-moisture gels for each variety, the aforementioned standardized process was followed to ensure the comparability of quality data between different varieties. By providing the specific number of principal components, variance contribution rates, loading matrices, and comprehensive scoring formulas, those skilled in the art can directly reproduce the mathematical model for the comprehensive evaluation of the suitability of soybean-based low-moisture gel processing without having to repeat the complete principal component analysis process.

[0091] In another technical solution, the adaptive freezing layer number for the variety similarity metric D is determined as follows:

[0092] The existing variety model M0 contains L hidden layers, numbered sequentially from the input layer to the output layer as the 1st to the Lth hidden layers. Specifically, the value of L depends on the network structure set when training the existing variety model M0 in step S2, for example, it can be set to 4. The first distance threshold and the second distance threshold are preset values ​​based on the Euclidean distance distribution between varieties in the multi-variety basic database. The first distance threshold can be set to the 25th percentile of the Euclidean distance between all pairs of varieties in the multi-variety basic database, and the second distance threshold can be set to the 75th percentile. For example, if the 25th percentile of the Euclidean distance between pairs of varieties in the multi-variety basic database is 3.5 and the 75th percentile is 7.2, then the first distance threshold is 3.5 and the second distance threshold is 7.2.

[0093] The number of frozen layers t is determined based on the variety similarity metric D:

[0094] When D ≤ the first distance threshold, t = L-1, freeze the parameters of the first hidden layer to the (L-1)th hidden layer, and only adjust the parameters of the Lth hidden layer and the output layer; specifically, when D ≤ the first distance threshold, it indicates that the physicochemical characteristics of the new target variety are very close to those of an existing variety. At this time, the number of frozen layers t is equal to L-1, the parameters of the first hidden layer to the (L-1)th hidden layer remain unchanged, and only the parameters of the Lth hidden layer and the output layer are adjusted.

[0095] When D > the first distance threshold and ≤ the second distance threshold, t = L-2, freeze the parameters of the first hidden layer to the (L-2)th hidden layer, and adjust the parameters of the (L-1)th hidden layer, the Lth hidden layer, and the output layer; specifically, when D is greater than the first distance threshold and less than or equal to the second distance threshold, it indicates that there is a moderate degree of difference between the new target variety and the existing varieties. At this time, the number of frozen layers t is equal to L-2, the parameters of the first hidden layer to the (L-2)th hidden layer remain unchanged, and the parameters of the (L-1)th hidden layer, the Lth hidden layer, and the output layer are adjusted.

[0096] When D > the second distance threshold, t = L-3, freeze the parameters of the first hidden layer to the (L-3)th hidden layer, and adjust the parameters of the (L-2)th hidden layer, the (L-1)th hidden layer, the Lth hidden layer, and the output layer. Specifically, when D is greater than the second distance threshold, it indicates that the new target variety is significantly different from the existing varieties. At this time, the number of frozen layers t is equal to L-3, the parameters of the first hidden layer to the (L-3)th hidden layer remain unchanged, and the parameters of the (L-2)th hidden layer, the (L-1)th hidden layer, the Lth hidden layer, and the output layer are adjusted.

[0097] The design of this adaptive freezing strategy is based on the following observations: the first few hidden layers of a deep neural network typically learn general low-level feature representations of the data, which have strong commonalities among different varieties and can be preserved; while the hidden layers closer to the output layer typically learn specific features related to a particular variety, which need to be adjusted according to the new variety data. The higher the variety similarity, the closer the physicochemical properties of the new variety are to those of existing varieties, the more common features there are, and the more hidden layer parameters can be preserved; the lower the variety similarity, the greater the difference, and the more layers need to be adjusted. Through a three-level adaptive freezing strategy based on the variety similarity metric D, flexible adaptation to different degrees of variety differences is achieved without over-design, avoiding the underfitting or overfitting problems that may be caused by a uniform freezing strategy.

[0098] In another technical solution, the regularization constraint introduced into the loss function is L2 regularization, and the expression for the loss function Loss is: Loss = MSE + α × ‖W‖ 2 Where MSE is the mean squared error between the predicted composite score and the actual composite score, α is the L2 regularization coefficient, and ||W|| 2 The regularization coefficient α is the sum of squared L2 norms of the adjusted layer weight parameters. It is adaptively set based on the sample size n of the small sample data for the new target variety: α = β / n, where β is a preset base regularization constant. Specifically, α is the L2 regularization coefficient, controlling the penalty strength of the regularization term. 2 The regularization coefficient α is the sum of squares of the L2 norm of all weight parameters in the adjusted layer, i.e., the cumulative sum of the squares of each weight parameter in the adjusted layer. The regularization coefficient α is adaptively set according to the number of samples n of the small sample data of the new target variety. β is a preset basic regularization constant, which can be set to a value in the range of 0.01 to 0.1, for example, it can be set to 0.05. The principle of adaptive design is as follows: when the number of samples n is small, the risk of overfitting is high, and the regularization coefficient α automatically increases to apply a stronger penalty; when the number of samples n is large, the risk of overfitting decreases, and the regularization coefficient α automatically decreases. For example, when n=5, α=0.05 / 5=0.01; when n=10, α=0.05 / 10=0.005.

[0099] The early termination mechanism is introduced as follows: The limited sample data of the new target variety is divided into an adjustment set and a validation set according to a preset ratio. After each iteration, the mean squared error of the prediction on the validation set is calculated. Training is terminated early when the mean squared error of the prediction on the validation set no longer decreases in m consecutive iterations, where m is the preset number of iterations. Specifically, the limited sample data of the new target variety is divided into an adjustment set and a validation set according to a preset ratio, which can be set to 7:3, i.e., 70% of the data is used for parameter adjustment and 30% for validation. After each parameter iteration, the current model is used to predict the validation set data, the mean squared error of the prediction on the validation set is calculated, and the trend of the mean squared error of the prediction on the validation set is recorded. Training is terminated early when the mean squared error of the prediction on the validation set no longer decreases in m consecutive iterations. m is the preset number of training iterations, which can be set to 10. The criterion for no further decrease is that the minimum value of the validation set prediction mean square error in the last m iterations has not been refreshed, that is, the validation set prediction mean square error in the most recent m iterations is greater than or equal to the previously reached minimum value. When training is terminated early, the model parameters revert to the parameter state corresponding to the iteration with the lowest validation set prediction mean square error. The maximum number of iterations for parameter adjustment can be set to 200 iterations. If the maximum number of iterations is reached but early stopping is not triggered, the parameters obtained in the last iteration are used as the final model parameters. The risk of overfitting under few-sample conditions is suppressed by L2 regularization constraints, and the optimal training stopping point is automatically determined by the early stopping mechanism, thus avoiding overtraining while ensuring prediction accuracy.

[0100] In another technical solution, the training objective of the feature representation layer in step S2 is achieved through maximizing the mean difference, specifically:

[0101] When training an existing variety model M0, the loss function Loss is composed of the prediction error loss. pred Distribution difference loss mmd Weighted composition: Loss = Loss pred +γ×Loss mmd , where γ is the weighting coefficient of the distribution difference loss;

[0102] Loss mmd The calculation method is as follows: Loss mmd =‖μ u -μ v || 2 , where μ u Let μ be the mean vector output by variety u in the feature representation layer. v Let v be the mean vector output by variety v in the feature representation layer. Iterate through all pairwise combinations of varieties in the multi-variety basic database, sum and average them. Variety u and variety v are any two different varieties in the multi-variety basic database.

[0103] Specifically, in training existing variety models During the process, the model's loss function Loss consists of two weighted parts: Among them, Loss pred To predict error loss, mean squared error (MSE) can be used, calculated as the mean of the squared differences between the predicted and true values ​​of all training samples. mmd The distribution difference loss is used to measure the differences in the output distribution of different varieties in the feature representation layer. This is the weighting coefficient for the distribution difference loss, used to adjust the relative importance of the distribution alignment task to the prediction task. It can be set to a value in the range of 0.1 to 1.0, for example, it can be set to 0.5.

[0104] Loss mmd The calculation method is as follows: Loss mmd =‖μ u -μ v || 2 For any two different varieties in the multi-variety basic database and varieties , respectively All samples and varieties All samples are input into the model, the output vector of the feature representation layer is extracted, and the variety is calculated. The mean vector output by the feature representation layer and varieties The mean vector output by the feature representation layer Calculate the square of the Euclidean distance between these two mean vectors: Iterate through all pairwise combinations in the multi-variety basic database, calculate the squared distance for each pair, sum the squared distances of all combinations, divide by the total number of combinations, and take the average to obtain the loss. mmd For example, if there are multiple varieties in the basic database If there are varieties, then the total number of pairwise combinations is Loss mmd Equal to all combinations Sum divided by ;

[0105] During training, Loss pred This enables the model to learn from grain physicochemical indicators to comprehensive scoring. The accurate mapping relationship, Loss mmd By encouraging the distributions of different varieties' outputs in the feature representation layer to converge, and jointly optimizing these two objectives, a model for existing varieties can be trained. While learning the quality prediction capability, common feature representations across varieties were extracted. By introducing the maximum mean difference as a measure of distribution difference, the goal of aligning the distributions of different varieties was explicitly incorporated into the loss function of model training. This allows the feature representation layer to naturally learn quality-related features that are invariant to different varieties during training, laying the foundation for subsequent cross-variety transfer learning.

[0106] In another technical solution, after obtaining small sample data of the new target variety in S3 and before parameter adjustment in S4, a variety selection step is also included:

[0107] The physicochemical indicators of the grains from a small sample of the new target variety are input into the existing variety model M0. The output vector of the feature representation layer is extracted, and the cosine similarity between the output vector and the output center vector of each existing variety in the multi-variety basic database is calculated. The top Q existing varieties with the highest cosine similarity are selected as the preferred existing variety set, where Q is the preset number of preferred varieties. When the maximum cosine similarity in the preferred existing variety set is less than the preset similarity threshold, the variety selection is abandoned, and the parameters of the unfrozen hidden layer and the output layer are adjusted using all existing variety data.

[0108] In S4, except for the above-mentioned rollback scenario, only the data corresponding to the preferred existing variety set is used to adjust the parameters of the unfrozen hidden layer and the output layer;

[0109] Specifically, the physicochemical indicators of the grains from a small sample of the new target variety are input into the existing variety model M0. During the forward propagation of the existing variety model, the output values ​​of the output layer are not taken; instead, the output vector of the feature representation layer is extracted. This output vector is an intermediate representation of the model after extracting abstract features from the input data and before making the final quality prediction. It contains the distribution of the variety in the model's feature space. When the maximum cosine similarity of the selected existing variety set is less than a preset similarity threshold, it indicates that the new target variety differs significantly from all existing varieties in the feature space, and the significance of variety selection is limited. In this case, variety selection is abandoned, and in step S4, all existing variety data are used to analyze the unfrozen hidden layer and... The output layer is adjusted for parameters, and the similarity threshold can be set to 0.5. When the maximum cosine similarity is greater than or equal to 0.5, it indicates that there are existing varieties that are close to the feature patterns of the new target variety. In step S4, only the data corresponding to the preferred set of existing varieties is used to adjust the parameters of the unfrozen hidden layer and the output layer. By using the feature representation layer of the existing variety model M0 to screen the similarity between the new target variety and the existing varieties in the feature space before transfer learning, the data used for transfer learning is limited to the range of existing varieties that are close to the feature patterns of the new target variety, which improves the efficiency of parameter adjustment for new varieties. At the same time, the similarity threshold backoff mechanism avoids invalid screening in extreme cases.

[0110] In another technical solution, after outputting the comprehensive quality prediction results, it also includes:

[0111] The prediction results are compared with the preset suitability threshold. When the prediction result is greater than or equal to the suitability threshold, the soybean raw material is determined to be suitable for soybean-based low-moisture gel processing. When the prediction result is less than the suitability threshold, the soybean raw material is determined to be unsuitable for soybean-based low-moisture gel processing, and an evaluation report is output.

[0112] Specifically, the predicted results are compared with a preset suitability threshold. This threshold can be set based on the minimum acceptable standard for soybean-based low-moisture gel quality in actual production, for example, 0.4. When the predicted result is greater than or equal to 0.4, the soybean raw material is deemed suitable for soybean-based low-moisture gel processing; when the predicted result is less than 0.4, the soybean raw material is deemed unsuitable for soybean-based low-moisture gel processing. The judgment result can be output in the form of an evaluation report. The evaluation report can include basic information about the soybean raw material variety, measured values ​​of 12 grain physicochemical indicators, specific values ​​of the comprehensive quality prediction results, and the suitability judgment conclusion. The evaluation report can be generated automatically. Text files, spreadsheets, or structured data records can be used. The suitability threshold can be adjusted according to the actual needs of soybean-based low-moisture gel manufacturers. For example, companies pursuing high-quality products can set the suitability threshold to a higher value, such as 0.6. The suitability threshold does not have to be a single fixed value, but can be set according to different product positioning. For example, 0.6 and above can be set as excellent, 0.4 to 0.6 as qualified, and below 0.4 as unqualified. By quantitatively comparing the comprehensive quality prediction results output by the model with the preset suitability threshold, the prediction results are transformed into clear processing suitability conclusions, enabling companies to make screening decisions quickly and objectively in the raw material procurement process.

[0113] The present invention also provides a cross-variety prediction system for soybean-based low-moisture gel quality. The system is configured to execute any of the aforementioned cross-variety prediction methods for soybean-based low-moisture gel quality. The system includes a data acquisition module, a comprehensive evaluation module, an existing variety training module, a transfer learning module, and a quality prediction module. The modules can communicate with each other through a data bus and control signals.

[0114] The data acquisition module is configured to acquire the physicochemical indicators of soybean grains and the corresponding soybean-based low-moisture gel comprehensive score F of multiple known soybean varieties, as well as to acquire the physicochemical indicators of new target varieties.

[0115] The data acquisition module can provide a data input interface, which can be in the form of file import, database connection or manual input interface;

[0116] The comprehensive evaluation module is configured to use the quality index data of soybean-based low-moisture gel products to construct a mathematical model for the comprehensive evaluation of the processing suitability of soybean-based low-moisture gels through principal component analysis, and to calculate the comprehensive score F of soybean-based low-moisture gels.

[0117] The comprehensive evaluation module can include a data preprocessing submodule, a testing submodule, a principal component analysis submodule, and a scoring calculation submodule. The functions of these submodules can be implemented using programming languages ​​such as Python in conjunction with scientific computing libraries such as NumPy and scikit-learn.

[0118] The existing variety training module is configured to use the physicochemical indicators of known soybean varieties as input features and F as the label to train a deep neural network as an existing variety model. During training, a variety-independent feature representation layer is added to the hidden layer. The training objective of this layer is to make the output distribution of different varieties in this layer tend to be consistent. The existing variety training module can be built based on deep learning frameworks such as PyTorch or TensorFlow. The feature representation layer can be a fully connected layer, and its number of nodes can be set to 16 or 32.

[0119] The transfer learning module is configured for computation.

[0120] This invention also provides an electronic device, including a processor and a memory storing a computer program. When the processor executes the computer program, it implements any of the aforementioned methods for predicting soybean-based low-moisture gel quality across varieties. The processor can be a central processing unit, a graphics processing unit, a tensor processor, or a neural network processor, or a combination thereof. When the neural network structure is large or the training data is abundant, a graphics processing unit can be preferentially selected to utilize its parallel computing capabilities to accelerate the training and inference process of the deep neural network. The memory can include random access memory and non-volatile memory. The non-volatile memory stores the computer program, a multi-variety basic database, and pre-trained parameters of existing variety models and parameters of new target variety prediction models. The processor and the memory are connected via a communication bus. The electronic device may also include a communication interface for receiving externally input grain physicochemical index data and outputting comprehensive quality prediction results. The communication interface can be a wired network interface or a wireless network interface. The computer program can be stored in a non-transitory computer-readable storage medium, which can be a USB flash drive, a portable hard drive, a read-only memory, a random access memory, a magnetic disk, or an optical disk.

[0121] When a processor executes a computer program, the following process is performed:

[0122] We obtained the physicochemical indices of soybean grains and the corresponding soybean-based low-moisture gel composite score F for multiple known soybean varieties, and constructed a multi-variety basic database. Using the physicochemical indices of each variety as input features and F as the label, we trained a deep neural network to obtain the existing variety model M0. During training, we added a variety-independent feature representation layer in the hidden layer. The training objective of this layer is to make the output distribution of different varieties in this layer tend to be consistent.

[0123] For a new target variety, collect a predetermined number of samples to obtain the physicochemical properties of the grains and F, and calculate the variety similarity measure D between the new target variety and existing varieties.

[0124] The number of frozen hidden layers in M0 is determined by D adaptively. The parameters of the frozen hidden layers remain unchanged. The parameters of the unfrozen hidden layers and the output layer are adjusted with the data of the new target variety. The loss function is subject to regularization constraints and an early stop mechanism is set to obtain the prediction model M1 for the new target variety.

[0125] Input the physicochemical indicators of the raw materials to be tested for the new target variety into M1, and output the comprehensive quality prediction results. By programming the cross-variety prediction method for soybean-based low-moisture gel quality and executing it on electronic devices, the method can be automated on conventional computing hardware, thus lowering the barrier to entry.

[0126] <Example 1>

[0127] Construction of a multi-variety basic database

[0128] This embodiment uses 34 soybean varieties as the research object to construct a multi-variety basic database. These 34 varieties cover the main cultivated varieties in Northeast China, the Huang-Huai-Hai Plain, and Southern China. After collecting grain samples from each variety, 12 physicochemical indicators were measured: moisture content was determined according to GB 5009.3-2016; protein content was determined according to GB 5009.5-2016 with a protein conversion factor of 5.71; fat content was determined according to GB 5009.6-2016 using Soxhlet extraction; 11S globulin and 7S conglobulin contents were determined by SDS-PAGE electrophoresis, with the 11S / 7S ratio being the 11S globulin content divided by the 7S conglobulin content; water-soluble protein content was determined according to NY / T 1205-2006; protein solubility was the percentage obtained by dividing the water-soluble protein content by the total protein content; phytic acid content was determined by colorimetric method; lectin content was determined by enzyme-linked immunosorbent assay (ELISA); and calcium and magnesium content were determined according to GB... 5009.268-2016 Determination by inductively coupled plasma mass spectrometry.

[0129] The coefficients of variation for 12 physicochemical indicators of grains from 34 varieties ranged from 5.92% to 22.37%, indicating significant differences in physicochemical characteristics among varieties and sufficient diversity in the database. Specifically, moisture content ranged from 5.81 g / 100g to 9.45 g / 100g, protein content from 31.45 g / 100g to 41.83 g / 100g, fat content from 15.08 g / 100g to 21.79 g / 100g, 11S globulin content from 102.93 mg / g to 175.99 mg / g, and 7S paraglobulin content... The range is 77.23 mg / g to 149.62 mg / g, the 11S / 7S ratio ranges from 1.01 to 2.03, the water-soluble protein content ranges from 13.3 g / 100g to 33.4 g / 100g, the protein solubility ranges from 36% to 86%, the phytic acid content ranges from 5.25 mg / g to 10.19 mg / g, the lectin content ranges from 7.15 mg / g to 11.19 mg / g, the calcium content ranges from 1210 mg / kg to 2990 mg / kg, and the magnesium content ranges from 1920 mg / kg to 2640 mg / kg.

[0130] Each variety was processed into dried tofu using a standardized process (as an example of soybean-based low-moisture gel). The standardized process included a soaking time of 12 hours, a grinding mesh of 100 mesh, a boiling temperature of 95°C for 10 minutes, magnesium chloride as the coagulant at 0.3% of the soybean milk mass, a pressing pressure of 0.05 MPa, and a pressing time of 30 minutes. Fifteen quality indicators were measured in the resulting dried tofu product, including moisture content, protein content, fat content, calcium content, magnesium content, hardness, adhesiveness, resilience, cohesiveness, elasticity, adhesion, chewiness, tensile strength, shear force, and overall sensory acceptability. Twelve physicochemical indicators of the seeds and fifteen quality indicators of the soybean-based low-moisture gel for each variety were recorded in a database, forming the multi-variety basic database of this embodiment.

[0131] <Example 2>

[0132] Construction of the PCA Comprehensive Evaluation Model

[0133] Based on the data of 15 quality indicators of 34 varieties of soybean-based low-moisture gel products in the multi-variety basic database constructed in Example 1, a mathematical model for comprehensive evaluation of the processing suitability of soybean-based low-moisture gel was constructed.

[0134] KMO test and Bartlett's test of sphericity were performed on the 15 quality index data. In this example, the KMO test value was 0.650, which is greater than 0.6; the significance level of the Bartlett's test of sphericity was less than 0.001, which is much less than 0.05. Both tests passed, confirming that the data is suitable for principal component analysis.

[0135] Z-score standardization was performed on the 15 quality indicators to make the mean of each indicator 0 and the standard deviation 1. Principal component analysis was then performed on the standardized data to extract principal components with eigenvalues ​​greater than 1. A total of 6 principal components were extracted, with variance contribution rates of 22.948%, 14.391%, 12.866%, 11.151%, 8.050%, and 7.473%, respectively. The cumulative variance contribution rate was 76.879%, which exceeds the 75% information retention standard.

[0136] The function expressions corresponding to the factor loading matrices of the six principal components are as follows:

[0137] Z1=0.015X1-0.471X2+0.808X3+0.599X4+0.778X5+0.605X6+0.109X7-0.513X8-0.549X9+0.222X 10 -0.311X 11 +0.241X 12 -0.008X 13 +0.668X 14 -0.101X 15 ;

[0138] Z2=0.304X1+0.320X2+0.031X3+0.002X4+0.108X5-0.151X6+0.246X7+0.720X8+0.739X9-0.032X 10 +0.429X 11 +0.054X 12 +0.266X 13 -0.365X 14 +0.249X 15 ;

[0139] Z3=0.041X1-0.071X2-0.030X3-0.577X4-0.432X5+0.514X6-0.776X7+0.159X8+0.060X9+0.466X 10 -0.032X 11 +0.595X 12 +0.097X 13 +0.208X 14 +0.362X 15 ;

[0140] Z4=0.704X1-0.144X2-0.210X3-0.288X4-0.007X5+0.225X6-0.027X7+0.278X8+0.124X9-0.552X 10 +0.525X 11 -0.346X 12 -0.064X 13 +0.148X 14 +0.263X 15 ;

[0141] Z5=0.113X1+0.566X2+0.005X3+0.226X4-0.028X5-0.166X6+0.121X7+0.175X8+0.011X9+0.071X 10 -0.032X 11 -0.157X 12 +0.770X 13 +0.098X 14 -0.401X 15 ;

[0142] Z6=0.305X1-0.171X2-0.510X3+0.043X4+0.070X5+0.134X6+0.262X7+0.033X8+0.165X9+0.435X 10 +0.042X 11 +0.005X 12 +0.026X 13 -0.454X 14 +0.089X 15 ;

[0143] Z6=0.305X1-0.171X2-0.510X3+0.043X4+0.070X5+0.134X6+0.262X7+0.033X8+0.165X9+0.435X 10 +0.042X 11 +0.005X 12 +0.026X 13 -0.454X 14 +0.089X 15 ;

[0144] Among them, X1 to X 15 The standardized values ​​for soybean-based low-moisture gels are, in order: moisture content, protein content, fat content, calcium content, magnesium content, hardness, adhesiveness, resilience, cohesiveness, elasticity, adhesion, chewiness, tensile strength, shear strength, and overall sensory acceptability.

[0145] The weights are determined based on the variance contribution rate of each principal component, and the original comprehensive score calculation formula is constructed as follows: F raw =(22.948Z1+14.391Z2+12.866Z3+11.151Z4+8.050Z5+7.473Z6) / 76.879, calculate the 34 varieties.

[0146] <Example 3>

[0147] Training of existing variety model M0

[0148] Based on the 12 physicochemical indicators of grains from 34 varieties in Example 1 as input features, and using the comprehensive score F of each variety calculated in Example 2 as the label, the existing variety model M0 is trained. The deep neural network adopts a 4-layer hidden layer structure, with the number of nodes in each hidden layer set to 64, 128, 64, and 32, respectively. The activation function is the ReLU function, and the output layer has 1 node and uses a linear activation function. A variety-independent feature representation layer is added at the 3rd hidden layer (64 nodes). The loss function Loss is the loss due to prediction error. pred Distribution difference loss mmd Weighted composition: Loss l =Loss pred +0.5×Loss mmd The optimizer used was the Adam optimizer, with an initial learning rate of 0.001, 500 training epochs, and a batch size of 16. The training data was divided into training and test sets in an 8:2 ratio.

[0149] After training, the predicted values ​​of the existing variety model M0 on the test set showed good consistency with the actual comprehensive score F. The model learned the cross-variety common mapping rules from grain physicochemical indicators to soybean-based low-moisture gel comprehensive score. At this time, the mean values, center vectors and corresponding comprehensive scores F of 12 grain physicochemical indicators of 34 varieties in the multi-variety basic database have been recorded, the parameters of the existing variety model M0 have been saved, and the construction of the multi-variety basic database is complete.

[0150] When faced with a new target variety, the physicochemical indicators and comprehensive score F of the grains of a predetermined number of samples are collected. The physicochemical indicators of the new target variety are constructed into a vector x, and the Euclidean distance between x and the center vector of existing varieties is calculated to obtain the variety similarity measure D. Based on D, the number of frozen layers t of M0 is adaptively determined: if D is less than or equal to the first distance threshold (e.g., set to 3.5), the first 3 layers are frozen, and only the 4th hidden layer and the output layer are adjusted; if D is greater than the first distance threshold and less than or equal to the second distance threshold (e.g., set to 7.2), the first 2 layers are frozen, and the 3rd and 4th hidden layers and the output layer are adjusted; if D is greater than the second distance threshold... The value is set to the first layer before freezing. The second, third, and fourth hidden layers and the output layer are adjusted. When adjusting the parameters, the learning rate is reduced to 0.0001. L2 regularization is introduced into the loss function (the basic regularization constant β is set to 0.05), and the early stopping mechanism is enabled (the number of patient rounds m is set to 10). After the adjustment, the new target variety prediction model M1 is obtained. The 12 grain physicochemical indicators of the remaining raw materials to be tested for the new target variety are input into M1, and the comprehensive quality prediction result of soybean-based low moisture gel of the raw material can be output. The prediction result is compared with the preset suitability threshold of 0.4 to determine whether the raw material is suitable for soybean-based low moisture gel processing.

[0151] It should be noted that the above embodiments are only described using dried tofu as an example of a soybean-based low-moisture gel product. Those skilled in the art should understand that the technical solution of the present invention is also applicable to other soybean-based low-moisture gel products, such as dried tofu, vegetarian meat, and high-fiber protein gels. Only the standardized processing parameters and quality index system need to be adjusted according to the specific product.

[0152] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A cross-variety prediction method for the quality of soybean-based low-moisture gels, characterized in that, Includes the following steps: S1. Obtain the physicochemical indicators of soybean grains and the corresponding soybean-based low moisture gel comprehensive score F for multiple known soybean varieties, and construct a multi-variety basic database. The physicochemical indicators of the grains include moisture, protein, fat, 11S globulin, 7S conglobulin, 11S / 7S ratio, water-soluble protein, protein solubility, phytic acid, lectins, calcium, and magnesium. S2. Using the physicochemical indicators of each variety's grains as input features and F as the label, train a deep neural network to obtain the existing variety model M0. During training, add a variety-independent feature representation layer in the hidden layer. The training objective of the feature representation layer is to make the output distribution of different varieties in this layer tend to be consistent and to extract cross-variety common features. S3. For the new target variety, collect a predetermined number of samples for grain physicochemical indicators and F, as small sample data for the new target variety, and construct a vector from the grain physicochemical indicators of the new target variety. The center vector of the p-th variety among the existing varieties Calculate the Euclidean distance Where j=1,2,…,12, p=1,2,…,P, P is the total number of varieties in the multi-variety basic database, and the variety similarity measure is D=min(d p ); S4. Determine the number of frozen layers in M0 based on the variety similarity metric D. Keep the parameters of the frozen hidden layers in M0 unchanged. Adjust the parameters of the unfrozen hidden layers and the output layer using the new target variety data. Introduce regularization constraints and set an early stop mechanism in the loss function to obtain the new target variety prediction model M1. S5. For the remaining raw materials to be tested for the new target variety, input the measured physicochemical indicators of the grains into M1 and output the comprehensive quality prediction results.

2. The method for predicting the quality of soybean-based low-moisture gels across varieties as described in claim 1, characterized in that, The overall score F of soybean-based low-moisture gel is obtained through the following steps: Quality index data of soybean-based low-moisture gel products obtained from multiple varieties of soybean raw materials through standardized processes were obtained, including 15 quality indicators such as moisture content, protein content, fat content, calcium content, magnesium content, hardness, adhesiveness, resilience, cohesiveness, elasticity, adhesion, chewiness, tensile strength, shear force, and overall sensory acceptability. KMO test and Bartlett's test of sphericity were performed on the 15 quality index data. When the KMO test value was >0.6 and the significance level of Bartlett's test of sphericity was <0.05, the data were deemed suitable for principal component analysis. The quality index data that passed the inspection were standardized using Z-score, with the mean of each quality index data being 0 and the standard deviation being 1. Principal component analysis was performed on the standardized data to extract the top K principal components whose eigenvalues ​​were greater than 1 and whose cumulative variance contribution rate was ≥75%. The variance contribution rate λ of each principal component was then obtained. i and the factor loading matrix, where i = 1, 2, ..., K; The scores Z of each principal component are constructed based on the factor loading matrix. i The function expression: , where a ij Let X be the loading coefficient of the i-th principal component on the j-th quality index. j Let j be the standardized value of the j-th quality indicator, where j = 1, 2, ..., 15; The weights are determined based on the variance contribution rate of each principal component, and the original comprehensive score calculation formula is constructed as follows: ; For all varieties of F raw Perform deviation standardization and map to the 0-1 interval: , of which F raw1 and F raw2 The original comprehensive score F for all varieties is as follows: raw The minimum and maximum values.

3. The method for predicting the quality of soybean-based low-moisture gels across varieties as described in claim 2, characterized in that, The number of principal components, K, is 6. The variance contribution rates of the 6 principal components are 22.948%, 14.391%, 12.866%, 11.151%, 8.050%, and 7.473%, respectively, with a cumulative variance contribution rate of 76.879%. The factor loading matrices and functional expressions of the 6 principal components are as follows: The score of the first principal component Z1 = 0.015X1 - 0.471X2 + 0.808X3 + 0.599X4 + 0.778X5 + 0.605X6 + 0.109X7 - 0.513X8 - 0.549X9 + 0.222X 10 -0.311X 11 +0.241X 12 -0.008X 13 +0.668X 14 -0.101X 15 ; The score of the second principal component Z2 = 0.304X1 + 0.320X2 + 0.031X3 + 0.002X4 + 0.108X5 - 0.151X6 + 0.246X7 + 0.720X8 + 0.739X9 - 0.032X 10 +0.429X 11 +0.054X 12 +0.266X 13 -0.365X 14 +0.249X 15 ; The score of the second principal component Z2 = 0.304X1 + 0.320X2 + 0.031X3 + 0.002X4 + 0.108X5 - 0.151X6 + 0.246X7 + 0.720X8 + 0.739X9 - 0.032X 10 +0.429X 11 +0.054X 12 +0.266X 13 -0.365X 14 +0.249X 15 ; The score of the third principal component Z3 = 0.041X1 - 0.071X2 - 0.030X3 - 0.577X4 - 0.432X5 + 0.514X6 - 0.776X7 + 0.159X8 + 0.060X9 + 0.466X 10 -0.032X 11 +0.595X 12 +0.097X 13 +0.208X 14 +0.362X 15 ; The score of the fourth principal component Z4 = 0.704X1 - 0.144X2 - 0.210X3 - 0.288X4 - 0.007X5 + 0.225X6 - 0.027X7 + 0.278X8 + 0.124X9 - 0.552X 10 +0.525X 11 -0.346X 12 -0.064X 13 +0.148X 14 +0.263X 15 ; The score of the 5th principal component Z5 = 0.113X1 + 0.566X2 + 0.005X3 + 0.226X4 - 0.028X5 - 0.166X6 + 0.121X7 + 0.175X8 + 0.011X9 + 0.071X 10 -0.032X 11 -0.157X 12 +0.770X 13 +0.098X 14 -0.401X 15 ; The score of the 6th principal component Z6 = 0.305X1 - 0.171X2 - 0.510X3 + 0.043X4 + 0.070X5 + 0.134X6 + 0.262X7 + 0.033X8 + 0.165X9 + 0.435X 10 +0.042X 11 +0.005X 12 +0.026X 13 -0.454X 14 +0.089X 15 ; Wherein, X1 is the moisture content of soybean-based low-moisture gel, X2 is the protein content of soybean-based low-moisture gel, X3 is the fat content of soybean-based low-moisture gel, X4 is the calcium content of soybean-based low-moisture gel, X5 is the magnesium content of soybean-based low-moisture gel, X6 is the hardness of soybean-based low-moisture gel, X7 is the adhesiveness of soybean-based low-moisture gel, X8 is the resilience of soybean-based low-moisture gel, X9 is the cohesiveness of soybean-based low-moisture gel, and X... 10 X is a soybean-based low-moisture gel elasticity. 11 For soybean-based low-moisture gel adhesion, X 12 X is a soybean-based low-moisture gel with chewy properties. 13 For soybean-based low-moisture gel stretchability, X 14 For soybean-based low-moisture gel shear force, X 15 Overall sensory acceptability of soybean-based low-moisture gels; The comprehensive score is calculated as F = (22.948Z1 + 14.391Z2 + 12.866Z3 + 11.151Z4 + 8.050Z5 + 7.473Z6) / 76.

879.

4. The method for predicting the quality of soybean-based low-moisture gels across varieties as described in claim 1, characterized in that, The method for determining the adaptive number of frozen layers for the variety similarity metric D is as follows: The existing variety model M0 contains a total of L hidden layers, which are numbered sequentially from the input layer to the output layer as the 1st hidden layer to the Lth hidden layer. The number of frozen layers t is determined based on the varietal similarity metric D: When D ≤ the first distance threshold, t = L-1, freeze the parameters of the first hidden layer to the (L-1)th hidden layer, and only adjust the parameters of the Lth hidden layer and the output layer; When D > the first distance threshold and ≤ the second distance threshold, t = L-2, freeze the parameters of the first hidden layer to the (L-2)th hidden layer, and adjust the parameters of the (L-1)th hidden layer, the Lth hidden layer, and the output layer. When D > the second distance threshold, t = L-3, freeze the parameters of the first hidden layer to the (L-3)th hidden layer, and adjust the parameters of the (L-2)th hidden layer, the (L-1)th hidden layer, the Lth hidden layer, and the output layer. Wherein, the first distance threshold is less than the second distance threshold, and the first and second distance thresholds are preset values ​​based on the Euclidean distance distribution among varieties in the multi-variety basic database.

5. The method for predicting the quality of soybean-based low-moisture gels across varieties as described in claim 4, characterized in that, The regularization constraint introduced in the loss function is L2 regularization, and the expression for the loss function Loss is: Loss = MSE + α × ‖W‖ 2 Where MSE is the mean squared error between the predicted composite score and the actual composite score, α is the L2 regularization coefficient, and ||W|| 2 The regularization coefficient α is the sum of squares of the L2 norm of the weight parameters of the adjusted layer. The regularization coefficient α is adaptively set according to the number of samples n of the small sample data of the new target variety: α=β / n, where β is the preset basic regularization constant. The early termination mechanism is introduced as follows: the small sample data of the new target variety is divided into an adjustment set and a validation set according to a preset ratio. After each iteration, the prediction mean square error on the validation set is calculated. When the prediction mean square error on the validation set no longer decreases in consecutive m iterations, the training is terminated early, where m is the preset number of patience rounds.

6. The method for predicting the quality of soybean-based low-moisture gels across varieties as described in claim 1, characterized in that, The training objective of the feature representation layer in S2 is achieved through maximizing the mean difference, specifically: When training an existing variety model M0, the loss function Loss is composed of the prediction error loss. pred Distribution difference loss mmd Weighted composition: Loss = Loss pred +γ×Loss mmd , where γ is the weighting coefficient of the distribution difference loss; Loss mmd The calculation method is as follows: Loss mmd =‖μ u -μ v || 2 , where μ u Let μ be the mean vector output by variety u in the feature representation layer. v Let v be the mean vector output by variety v in the feature representation layer. Iterate through all pairwise combinations of varieties in the multi-variety basic database, sum and average them. Variety u and variety v are any two different varieties in the multi-variety basic database.

7. The method for predicting the quality of soybean-based low-moisture gels across varieties as described in claim 1, characterized in that, After obtaining small sample data of the new target variety in S3 and before parameter adjustment in S4, the process also includes a variety selection step: The physicochemical indicators of the grains from a small sample of the new target variety are input into the existing variety model M0. The output vector of the feature representation layer is extracted, and the cosine similarity between the output vector and the output center vector of each existing variety in the multi-variety basic database is calculated. The top Q existing varieties with the highest cosine similarity are selected as the preferred existing variety set, where Q is the preset number of preferred varieties. When the maximum cosine similarity in the preferred existing variety set is less than the preset similarity threshold, the variety selection is abandoned, and the parameters of the unfrozen hidden layer and the output layer are adjusted using all existing variety data. In S4, except for the above-mentioned rollback scenario, only the data corresponding to the preferred existing variety set is used to adjust the parameters of the unfrozen hidden layer and the output layer.

8. The method for predicting the quality of soybean-based low-moisture gels across varieties as described in claim 1, characterized in that, After outputting the overall quality prediction results, the following is also included: The prediction results are compared with the preset suitability threshold. When the prediction result is greater than or equal to the suitability threshold, the soybean raw material is determined to be suitable for soybean-based low-moisture gel processing. When the prediction result is less than the suitability threshold, the soybean raw material is determined to be unsuitable for soybean-based low-moisture gel processing, and an evaluation report is output.

9. A cross-variety prediction system for the quality of soybean-based low-moisture gels, characterized in that, The system is configured to perform the cross-variety prediction method for soybean-based low-moisture gel quality as described in any one of claims 1 to 8, and comprises: The data acquisition module is configured to acquire the physicochemical indicators of soybean grains and the corresponding soybean-based low-moisture gel comprehensive score F of multiple known soybean varieties, as well as the physicochemical indicators of soybean grains of new target varieties. The comprehensive evaluation module is configured to construct a mathematical model for the comprehensive evaluation of the processing suitability of soybean-based low-moisture gel based on the quality index data of soybean-based low-moisture gel products through principal component analysis, and calculate the comprehensive score F of soybean-based low-moisture gel. The existing variety training module is configured to train a deep neural network as an existing variety model using the physicochemical indicators of known soybean grains as input features and F as the label. During training, a variety-independent feature representation layer is added to the hidden layer. The training objective of this layer is to make the output distribution of different varieties in this layer tend to be consistent. The transfer learning module is configured to calculate the variety similarity metric D between the new target variety and each variety in the existing varieties. Based on D, the number of frozen hidden layers in the existing variety model is adaptively determined. The parameters of the frozen hidden layers are kept unchanged. The parameters of the unfrozen hidden layers and the output layer are adjusted with the data of the new target variety. The loss function introduces regularization constraints and sets an early stopping mechanism to obtain the prediction model of the new target variety. The quality prediction module is configured to input the physicochemical indicators of the raw materials to be tested for the new target variety into the new target variety prediction model and output the comprehensive quality prediction results.

10. An electronic device comprising a processor and a memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the cross-variety prediction method for soybean-based low-moisture gel quality according to any one of claims 1 to 8.

Images