Ternary protein textured protein prediction method based on machine learning

Abstract

The invention provides a ternary protein textured protein prediction method based on machine learning. The method comprises the following steps: acquiring raw material characteristic parameters, process parameters and product characteristic parameters of three protein raw materials, namely vital gluten, yeast protein and soybean meal under a high-moisture extrusion condition; the method comprises the following steps: constructing a ternary protein coupling characteristic and a yeast protein threshold characteristic on the basis of an inhibition effect of yeast protein on wheat gluten cross-linking and a sensitive characteristic of yeast protein solubility on moisture; constructing a back-propagation neural network model, taking the raw material characteristic parameters, the process parameters, the ternary protein coupling features and the yeast protein threshold features as inputs of the back-propagation neural network model, taking the product characteristic parameters as outputs of the back-propagation neural network model, and performing iterative training on the back-propagation neural network model until convergence to obtain a final product. A textured protein prediction model is obtained; and performing forward quality prediction or reverse process derivation on a to-be-detected system by using the textured protein prediction model.

Description

Technical Field

[0001] This invention belongs to the field of food science and engineering technology, specifically relating to a method for predicting the structured proteins of ternary proteins based on machine learning. Background Technology

[0002] High-moisture extrusion denatures, orients, and cross-links plant proteins during continuous cooking-shearing-cooling processes, forming layered fibers similar to muscle. This is currently the most promising method for preparing textured proteins for industrialization. However, the high-temperature, high-shear "black box" environment inside the extruder means the fiber formation mechanism remains incomplete: more than ten variables, such as raw material protein content, moisture, screw speed, and die temperature, are interdependent. Even slight deviations can lead to drastic fluctuations in product quality indicators such as hardness, elasticity, and segregation, resulting in defects such as excessive stickiness, brittleness, and lack of fiber texture. When the formulation is expanded from a single soybean protein to a ternary system of wheat gluten-yeast protein-soybean meal, the inhibitory effect of yeast protein on the cross-linking of wheat gluten and its moisture-sensitive solubility further amplify the complexity of the parameter-structure mapping. This causes the experimental quantity of the traditional single-factor trial-and-error method to increase exponentially, severely restricting the rapid application of low-cost soybean meal in high-quality plant-based meat.

[0003] To reduce the number of experiments, current research commonly employs Response Surface Methodology (RSM) to establish quadratic polynomial models. Using Box-Behnken or central composite design, 17–30 sets of experiments are arranged at once. Statistical regression is used to fit the explicit equation between process variables and response values, and contour lines are plotted to find the optimal interval. This method has been validated in studies on liquid-fumigated high-moisture extrusion of plant proteins and peanut protein texturization. Taking the three factors of liquid fumigation temperature, time, and fumigation liquor concentration as an example, RSM obtained optimized conditions with a sensory score of 87.4 using only 17 sets of experiments, reducing workload by approximately 60% compared to single-factor methods. Furthermore, the quadratic equation obtained from RSM can be directly embedded into the production line PLC for rapid process window estimation, thus becoming the mainstream optimization tool in the current high-moisture extrusion field.

[0004] However, RSM, based on the assumptions of low-order polynomials and linear regression, cannot describe the nonlinear and strongly coupled mechanisms such as threshold cross-linking inhibition of yeast proteins in the ternary protein system. This results in a still relatively large sample size, requiring ≥17 groups for three factors and three levels. If raw material characteristics and interaction terms are included, the number of experiments surges to 30–50 groups, which is less efficient compared to the advantage of small sample size in machine learning. The extrapolation error of the quadratic equation in the extreme value region is often >15%. When the proportion of soybean meal is >30% or the moisture content is <50%, the predicted hardness value deviates from the measured value by more than 20%, requiring additional verification experiments for correction. Furthermore, RSM can only optimize the parameters in a positive direction for quality, and cannot deduce the target texture to the raw materials / processes in reverse like neural networks. This leads to researchers having to manually adjust the formula repeatedly, extending the development cycle. Summary of the Invention

[0005] This invention proposes a machine learning-based method for predicting the structured proteins of ternary proteins, which solves the problems of low fitting accuracy of nonlinear coupling relationships in ternary protein systems and inability to achieve reverse process derivation in existing methods.

[0006] To address the aforementioned technical problems, this invention provides a machine learning-based method for predicting ternary protein tissues, comprising the following steps: Step S1: Obtain the raw material characteristic parameters, process parameters, and product characteristic parameters of three protein raw materials—gluten, yeast protein, and soybean meal—under high moisture extrusion conditions; Step S2: Based on the inhibitory effect of yeast protein on gluten cross-linking and the sensitivity of yeast protein solubility to moisture, construct ternary protein coupling features and yeast protein threshold features; the ternary protein coupling features include at least the temperature-ternary protein ratio coupling coefficient, the moisture-ternary protein cofactor, and the protein hydrophobicity coupling index; the yeast protein threshold features include at least the yeast protein moisture threshold feature and the yeast protein ratio threshold feature. Step S3: Construct a backpropagation neural network model. Use the raw material characteristic parameters, process parameters, ternary protein coupling characteristics, and yeast protein threshold characteristics as inputs to the backpropagation neural network model, and use the product characteristic parameters as outputs. Iteratively train the backpropagation neural network model until it converges to obtain a textured protein prediction model. Step S4: Use the described organized protein prediction model to perform positive quality prediction or reverse process derivation of the test system.

[0007] Preferably, the raw material characteristic parameters in step S1 include at least protein content, carbohydrate content, and total sulfhydryl content; the process parameters include at least moisture content, barrel temperature, and screw speed; and the product characteristic parameters include at least fiber content, elasticity, and chewiness.

[0008] Preferably, the formula for calculating the temperature-ternary protein ratio coupling coefficient in step S2 is: ; In the formula, This is the temperature-ternary protein ratio coupling coefficient; The barrel temperature; The ratio of gluten powder; This refers to the yeast protein ratio; This refers to the proportion of soybean meal.

[0009] Preferably, the calculation formula for the water-ternary protein cofactor is as follows: ; In the formula, It is a water-trin cofactor; Moisture content; This refers to the yeast protein ratio; The ratio of gluten powder; The proportion of soybean meal; The screw rotation speed; The formula for calculating the protein hydrophobic coupling index is as follows: ; In the formula, The protein hydrophobic coupling index; The hydrophobicity coefficient of wheat gluten; The hydrophobicity coefficient of soybean meal; The hydrophobicity coefficient of yeast protein; This refers to the barrel temperature.

[0010] Preferably, the value of the yeast protein moisture threshold feature in step S2 is determined by the following rules: when the moisture content M ≥ 18%, the value is 1, and when the moisture content M < 18%, the value is 0. The value selection rule for the yeast protein proportion threshold feature is as follows: when When the value is 1, The value is 0; where, This refers to the yeast protein ratio. The ratio of gluten powder, This refers to the proportion of soybean meal.

[0011] Preferably, the backpropagation neural network model in step S3 uses a hybrid activation function, which consists of a ReLU activation function and an ELU activation function, with a ReLU activation function to ELU activation function ratio of 6:4.

[0012] Preferably, the backpropagation neural network model uses a weighted mean square error loss function, with different weights set for different product characteristic parameters.

[0013] Preferably, in step S3, a constraint mechanism is applied to the weight update of the backpropagation neural network model, wherein the constraint mechanism requires that the absolute value of the input layer weights directly connected to the ternary protein coupling feature be kept above 0.25.

[0014] Preferably, the positive quality prediction in step S4 is as follows: inputting the raw material characteristic parameters, process parameters, ternary protein coupling characteristics and yeast protein threshold characteristics of the system to be tested into the textured protein prediction model, and outputting the predicted values ​​of the product characteristic parameters; the reverse process derivation is as follows: based on the target product characteristic parameters, the required raw material ratio or process parameters are deduced in reverse through the textured protein prediction model.

[0015] Preferably, the method further includes dynamic adjustment of a temperature threshold based on a yeast protein ratio threshold: when At that time, the barrel temperature threshold was 140℃; when At that time, the barrel temperature threshold was adjusted to 150℃; among which, This refers to the yeast protein ratio. The ratio of gluten powder, This refers to the proportion of soybean meal.

[0016] The beneficial effects of the present invention include at least the following: 1. The two experimentally verified biochemical laws of yeast protein inhibiting gluten crosslinking and yeast protein solubility-moisture sensitivity are quantified into ternary protein coupling characteristics and yeast protein threshold characteristics and input into the model, so that the network can obtain physicochemical priors directly related to structure formation during the training phase, which significantly reduces the requirement for sample size. 2. The constructed BPNN model can effectively identify the nonlinear relationship between variables in the process of tissue protein extrusion, and realize multi-path and multi-directional parameter prediction. All three types of models can achieve a rapid decrease in loss within a short training period. 3. Due to the synergistic effect of the interaction between gluten, yeast protein, and soybean meal, traditional experiments cannot quantify the contribution of each protein individually. However, this model, through dedicated feature engineering and structural optimization, has for the first time achieved a precise decomposition of the ternary relationship, an effect that cannot be extended to binary system technical solutions. Attached Figure Description

[0017] Figure 1 This is a flowchart of a method according to an embodiment of the present invention; Figure 2 This is a structural diagram of the backpropagation neural network (BPNN) in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the training process of the backpropagation neural network (BPNN) in an embodiment of the present invention. Figure 4 This refers to the model training loss result of the raw material characteristics in the embodiments of the present invention; Figure 5 These are the predicted results of raw material properties in the embodiments of the present invention; Figure 6 This refers to the model training loss result of the process parameters in this embodiment of the invention; Figure 7 These are the predicted results of process parameters in the embodiments of the present invention; Figure 8 This refers to the model training loss result of product characteristics in this embodiment of the invention; Figure 9 These are the predicted results of product characteristics in embodiments of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0019] like Figure 1 As shown, this embodiment of the invention provides a machine learning-based method for predicting ternary protein organization, comprising the following steps: Step S1: Obtain the raw material characteristic parameters, process parameters, and product characteristic parameters of three protein raw materials—gluten, yeast protein, and soybean meal—under high moisture extrusion conditions.

[0020] Specifically, the raw material characteristic parameters include at least protein content, carbohydrate content, and total sulfhydryl content. Among them, protein content is a fundamental indicator for evaluating the structural building potential of raw materials, and it has a significant negative correlation with its corresponding carbohydrate content, with the two being distributed in a substitutional manner in the raw material composition; the total sulfhydryl content is closely related to the ability of proteins to form disulfide bonds during extrusion, and can indirectly characterize its network building ability.

[0021] Process parameters include at least moisture content, barrel temperature, and screw speed. Moisture content, screw speed, and barrel temperature all play a crucial role and are the most important thermo-mechanical control factors in high-moisture extrusion. They directly affect the system's heat input, shear environment, and material flow behavior, and are key control points that determine the efficiency and quality of product structure construction.

[0022] Product characteristic parameters include at least fiber content, elasticity, and chewiness. Fiber content, as a core structural indicator, reflects the degree of formation of meat-like fiber structures in textured proteins. A higher fiber content indicates a clearer fiber arrangement and more fiber filaments in the extrudate, resulting in a stronger meat-like quality and a closer resemblance to the structural characteristics of real meat. Chewiness and elasticity reflect the mechanical response and texture characteristics of textured proteins. In addition, product characteristic parameters may also include L-value, a-value, etc. The value is used to characterize the color attributes of a product and is related to sensory acceptance and market application.

[0023] By acquiring multi-dimensional parameters of textured proteins, they were categorized into three key parameter classes: raw material characteristics, process parameters, and product characteristics, totaling 27 feature indicators. The characteristics of textured proteins obtained from different raw materials under different process conditions through high-moisture extrusion were mapped one-to-one, constructing a total of 77 valid datasets. The data were divided into training and testing sets in an 8:2 ratio for the construction and validation of the prediction model's dataset.

[0024] Step S2: Based on the inhibitory effect of yeast protein on gluten cross-linking and the sensitivity of yeast protein solubility to moisture, construct ternary protein coupling features and yeast protein threshold features.

[0025] Because of the significant differences in thermal stability and interactions among gluten, yeast protein, and soybean meal, conventional binary features cannot capture ternary coupling relationships. Therefore, the following ternary-specific coupling features were designed: 1. Characteristics of ternary protein coupling Ternary protein coupling characteristics include at least the temperature-ternary protein ratio coupling coefficient, the moisture-ternary protein synergist, and the protein hydrophobicity coupling index.

[0026] (1) Temperature-Ternary Protein Ratio Coupling Coefficient The formula for calculating the temperature-ternary protein ratio coupling coefficient is as follows: ; In the formula, This is the temperature-ternary protein ratio coupling coefficient; The barrel temperature; The ratio of gluten powder; This refers to the yeast protein ratio; This refers to the proportion of soybean meal.

[0027] Wheat gluten is the main cross-linking agent, and excessive aggregation is likely to occur at high proportions. Yeast protein is the cross-linking regulator; the higher the proportion, the stronger the inhibitory effect. This coefficient is used to capture the balance between the cross-linking tendency of wheat gluten and the inhibitory effect of yeast protein at high temperatures. For example, when the proportion of wheat gluten is 60% and the proportion of yeast protein is 10%, the coefficient can quantify the degree to which yeast protein alleviates wheat gluten aggregation at 140℃.

[0028] (2) Water-Ternary Protein Cofactor The formula for calculating the water-ternary protein cofactor is: ; In the formula, It is a water-trin cofactor; Moisture content; This refers to the yeast protein ratio; The ratio of gluten powder; The proportion of soybean meal; This represents the screw rotation speed.

[0029] The solubility of yeast protein is sensitive to moisture and needs to be matched with the screw speed to ensure uniform mixing of materials. This factor reflects the effect of moisture on the solubility of yeast protein plus the effect of shear force on the mixing degree of ternary protein, and is directly related to whether yeast protein can be effectively dispersed in the gluten-soybean meal system.

[0030] (3) Protein hydrophobic coupling index The formula for calculating the protein hydrophobic coupling index is: ; In the formula, The protein hydrophobic coupling index; The hydrophobicity coefficient of wheat gluten; The hydrophobicity coefficient of soybean meal; The hydrophobicity coefficient of yeast protein; This refers to the barrel temperature.

[0031] Based on the amino acid composition, the hydrophobicity of the three proteins was ranked as follows: gluten (0.8) > soybean meal (0.6) > yeast protein (0.4). Hydrophobicity directly affects the hydrophobic interaction between protein molecules. This index is used to quantify the promoting effect of temperature on the hydrophobic aggregation of ternary proteins.

[0032] 2. Yeast protein threshold characteristics Based on preliminary experiments, yeast proteins were found to have dual threshold characteristics. Yeast protein threshold features include at least yeast protein moisture threshold features and yeast protein proportion threshold features.

[0033] (1) Characteristics of moisture threshold of yeast protein The value of the moisture threshold characteristic of yeast protein is determined by the following rules: when the moisture content M ≥ 18%, the value is 1, and when the moisture content M < 18%, the value is 0.

[0034] Yeast proteins form a hydrophobic core when the moisture content is less than 18%, causing a sharp drop in solubility. This characteristic is used to capture the minimum moisture conditions under which yeast proteins can exert their regulatory effects.

[0035] (2) Yeast protein ratio threshold characteristics The rule for determining the value of the yeast protein proportion threshold feature is as follows: when When the value is 1, The value is 0; where, This refers to the yeast protein ratio. The ratio of gluten powder, This refers to the proportion of soybean meal.

[0036] When the proportion of yeast protein exceeds 15%, the inhibitory effect on cross-linking of wheat gluten reaches saturation. This characteristic quantifies the critical proportion of yeast protein's regulatory effect, which cannot be precisely determined by traditional experiments. Because yeast protein is less hydrophobic than wheat gluten and soybean meal, and can form weak interactions with both through hydrogen bonds, it is introduced into the coupling coefficient. The study aimed to quantify its inhibitory effect on excessive aggregation of gluten.

[0037] Step S3: Construct a backpropagation neural network model. Use the raw material characteristic parameters, process parameters, ternary protein coupling characteristics, and yeast protein threshold characteristics as inputs to the backpropagation neural network model, and use the product characteristic parameters as outputs. Iteratively train the backpropagation neural network model until it converges to obtain a textured protein prediction model.

[0038] Backpropagation Neural Networks (BPNNs) are artificial neural networks widely used for prediction and classification problems. BPNNs are trained through two stages: forward propagation and backward propagation. The output is then propagated forward to minimize prediction error. Figure 2 As shown, a BPNN typically consists of an input layer, one or more hidden layers, and an output layer. The input layer receives external data and passes it to the hidden layers in the network. The neurons in the hidden layers perform weighted summations of the inputs and transform them using a non-linear activation function to output the result. The number of layers and neurons significantly impacts the model's expressive power and complexity. The neurons in the output layer convert the outputs of the hidden layers into the final prediction result.

[0039] like Figure 3 The diagram shows the training process of a BPNN. In a BPNN, when the input is passed through the network, the weighted input for a neuron in the l-th layer is: ; in, This represents the weighted sum of the l-th layer; Here is the weight matrix for the l-th layer; The activation value of the (l-1)th layer; This is a bias term.

[0040] Through activation function Perform a nonlinear transformation to obtain the output of this layer: ; in, This is the output of the l-th layer.

[0041] The key to BPNN lies in its backpropagation algorithm, which adjusts the weights and biases in the network based on the output error, thereby gradually reducing the error. During backpropagation, the error of the output layer is first calculated, and then the error is propagated back to the input layer of the network layer by layer using a chain rule.

[0042] Assuming the target output is The network output is The loss function typically uses mean squared error (MSE). For the error of the output layer: ; in, For the error of the output layer; The gradient of the loss function with respect to the activation values; This is the derivative of the activation function.

[0043] The error in the hidden layer can be calculated using backpropagation: ; in, The error of the l-th layer; The transpose weights of the (l+1)th layer; This is the derivative of the activation function for that layer.

[0044] After calculating the error through backpropagation, the weights and biases of the network are updated using gradient descent. The update formulas for the weights and biases of the l-th layer are as follows: ; ; in, For learning rate, and These are the gradients of the loss function with respect to the weights and biases, respectively.

[0045] Common loss functions for BPNNs include mean squared error (MSE) and cross-entropy loss. For regression problems, mean squared error (MSE) is often used as the loss function. ; Where N is the number of samples; It is the true label of the i-th sample; It is the network's predicted output.

[0046] The training process of BPNN involves iteratively optimizing the loss function, using backpropagation to adjust the network weights in each iteration until the error converges. To accelerate the training process, optimization algorithms such as batch gradient descent, stochastic gradient descent, or mini-batch gradient descent are typically used.

[0047] In response to the complexity of the ternary protein system, this embodiment of the invention makes a ternary-specific adjustment to the BPNN structure: the core change in the ternary protein data is that the feature dimension is expanded from 2 raw materials + 3 processes + 2 couplings to 3 raw materials + 3 processes + 3 couplings, and the data distribution shows a double peak of gluten powder-dominated type and yeast protein-regulated type, and the proportion of extreme value samples is increased.

[0048] The backpropagation neural network model uses a hybrid activation function, which consists of ReLU and ELU activation functions, with a ReLU to ELU activation function ratio of 6:4.

[0049] In ternary systems, a high proportion of yeast protein can lead to new extreme values ​​such as supersolubility and low elasticity. For example, when yeast protein content is 20% and moisture content is 19%, the solubility can reach 55%, which is much higher than the maximum value of 48% in binary systems. Increasing the ELU ratio can better capture these new extreme values ​​and avoid prediction bias. This adjustment is based on the bimodal distribution characteristics of ternary data.

[0050] The backpropagation neural network model uses a weighted mean square error loss function, with different weights set for different product characteristic parameters.

[0051] Based on the priority of the impact of yeast protein on product parameters, the weights were redistributed: the weight of the core parameter protein solubility was set to 1.6; the weight of elasticity and chewiness was set to 1.0; and the weight of color L value was set to 0.7.

[0052] In this embodiment of the invention, a constraint mechanism is applied to the weight update of the backpropagation neural network model. The constraint mechanism requires that the absolute value of the input layer weights directly connected to the ternary protein coupling features remain above 0.25. Yeast protein is the regulatory core of the ternary system, and the influence of its features on the prediction results cannot be weakened. This constraint avoids ignoring the key role of yeast protein during network learning and is designed based on the regulatory mechanism of the ternary system.

[0053] The BPNN model is constructed using Python's PyTorch library, and model parameters are flexibly configured through configuration files. The number of neurons in the input layer matches the number of input features in the dataset, while the number of neurons in the output layer matches the number of output features in the prediction task. In this model, the number of hidden layers is set to 2, with each layer containing 64 neurons, ensuring both the model's expressive power and effectively avoiding overfitting and excessive computational cost. The number of neurons in the hidden layers can be adjusted to 1.3 times the input dimension.

[0054] To further prevent overfitting, a Dropout layer was added after each hidden layer of the model, with a Dropout ratio set to 0.2, which means that 20% of the neuron connections are randomly dropped during each training session to reduce the dependence on the training data.

[0055] The optimizer chosen is Adam, which combines the advantages of momentum and adaptive learning rate to accelerate model training and improve performance.

[0056] During training, the batch size was set to 32, and the number of training epochs was 100. Smaller batches help reduce memory consumption and speed up training, while larger batches help improve training stability. A higher number of training epochs allows the model to fully adjust the weights, thus better fitting the data.

[0057] After training, the model and its related parameters will be saved according to the `model_path` parameter in the configuration file, so that the trained model can be used for target prediction later. At the same time, loss curves, evaluation metrics, and other files generated during training will also be saved in the specified output path for further evaluation of the model's performance.

[0058] Mean squared error (MSE), as a fundamental regression indicator, can effectively quantify the absolute deviation between predicted and actual values. However, a single indicator is insufficient to comprehensively evaluate the engineering applicability of a model. Therefore, this invention constructs a multi-dimensional evaluation system encompassing accuracy, stability, and interpretability. Coefficient of determination It is a core indicator in statistics used to quantify the explanatory power of regression models for the variability of observed data, reflecting the goodness of fit between model predictions and actual observed values. Its formula is as follows: ; in, For the first i One observation value; For the first i One predicted value; The average of the observed values; The total number of samples.

[0059] Difference rate It is a stability metric for measuring the generalization performance of a model. It assesses the risk of overfitting by comparing the error difference between the training and test sets. Its formula is as follows: ; in, The mean squared error of the training set. To test the mean square error of the set.

[0060] The Pearson correlation coefficient is a parameter that measures the strength of the linear correlation between predicted and actual values, revealing the consistency of trends between variables. Its formula is as follows: ; in, For predicted values, The mean of the predicted values, For the true value, This represents the average of the true values.

[0061] Step S4: Use the organized protein prediction model to perform positive quality prediction or reverse process derivation of the test system.

[0062] (1) Positive quality prediction Positive quality prediction involves inputting the raw material characteristic parameters, process parameters, ternary protein coupling characteristics, and yeast protein threshold characteristics of the system to be tested into the texturized protein prediction model, and outputting the predicted values ​​of the product characteristic parameters.

[0063] (2) Reverse process derivation Reverse process derivation involves using a textured protein prediction model to deduce the required raw material ratios or process parameters based on the target product's characteristic parameters.

[0064] The method in this embodiment of the invention further includes dynamic adjustment of the temperature threshold based on the yeast protein ratio threshold: when At that time, the barrel temperature threshold was 140℃; when At that time, the barrel temperature threshold was adjusted to 150℃; in, This refers to the yeast protein ratio. The ratio of gluten powder, This refers to the proportion of soybean meal.

[0065] This rule solves the industry problem of inconsistent temperature thresholds in ternary systems. When the proportion of yeast protein is low, gluten dominates cross-linking, and the temperature threshold remains at 140℃; when the proportion of yeast protein is high, yeast protein inhibits cross-linking, and the temperature threshold can be increased to 150℃.

[0066] The final method for closed-loop regulation of ternary proteins was obtained: Step a: Determine the target product parameters, such as solubility S≥50%, elasticity E=2.8-3.2; Step b: Input the initial raw material ratio, such as gluten powder = 55%, yeast protein = 10%, soybean meal = 35%, and process parameters. The model predicts that the solubility S = 47% (not up to standard) and the elasticity E = 3.0 (up to standard). Step c: Adjust parameters based on quantitative relationship: Keep the barrel temperature T=130℃ and moisture content M=19% unchanged, and increase the yeast protein ratio from 10% to 15%; Step d: The model's second prediction yielded a solubility of S=50.5% (meeting the standard). Simultaneously, based on the temperature threshold adjustment rules, the temperature threshold was automatically adjusted to 150℃. The temperature can be appropriately increased to 145℃ to further optimize the taste. Step e: The production line executes the adjusted parameters, with an adjustment error ≤ 2.5%.

[0067] Example 1: Prediction of Raw Material Properties Raw material property prediction model: Inputs are process parameters (Y1-Y3), including moisture content, barrel temperature, and screw speed; and product properties (Z1-Z6), including fiber content, elasticity, chewiness, and L-value. 、a value, The output is the raw material characteristics (X1-X3), including protein content, carbohydrate content, and total sulfhydryl content.

[0068] like Figure 4 The figure shows the training loss results of the raw material characteristic prediction model. As can be seen from the figure, the raw material characteristic prediction model exhibits delayed response characteristics. The loss value remains above 0.05 for the first 20 epochs, which may be due to the nonlinear coupling effect of process-product parameters on raw material characteristics. After 20 epochs, the convergence still maintains a relatively fast trend. The difference rate Δ between the final test set loss MSE=0.027 and the training set loss MSE=0.031 is 12.9%. The phase delay phenomenon of its convergence trajectory reveals the parameter sensitivity characteristics unique to backpropagation.

[0069] Protein content is a fundamental indicator for evaluating the potential of raw materials for network construction. It also has a significant negative correlation with the corresponding carbohydrate content. The two are distributed in a substitute manner in the composition of raw materials. Therefore, this study only shows the protein content to represent the overall compositional characteristics of this type of nutrient. The total sulfhydryl content is closely related to the ability of protein to form disulfide bonds during extrusion, and can indirectly characterize its network construction ability.

[0070] The raw material characteristic prediction model performs exceptionally well in reverse reasoning. Figure 5 It can be seen that the average coefficient of determination R² of each parameter in the raw material characteristic prediction model is 0.94, which shows high reliability and provides a new idea for reducing economic cost losses and labor-intensive experimental iterations in industrial production.

[0071] Example 2: Prediction of process parameters Process parameter prediction model: The inputs are raw material characteristics (X1-X3) and product characteristics (Z1-Z6), and the output is process parameters (Y1-Y3).

[0072] like Figure 6 The figure shows the training loss results of the process parameter prediction model. As can be seen, the model achieves ultra-fast convergence in the first 20 epochs, with the MSE decreasing from 0.38 to 0.04 at a rate of 0.017 epochs. However, it exhibits periodic fluctuations in the 20-100 epoch range, with an amplitude of ±0.015. This oscillating convergence characteristic may be related to the gradient saturation effect of discrete process parameters. Despite local perturbations, the model ultimately achieves excellent performance with a test set loss MSE of 0.026 and a training set loss MSE of 0.029, along with a difference rate Δ of 10.3%, confirming the strong representational ability of the parameter mapping relationship.

[0073] In terms of process parameters, moisture content, screw speed, and barrel temperature all play a crucial role and are the most important thermo-mechanical control factors in high-moisture extrusion. They directly affect the system's heat input, shear environment, and material flow behavior, and are key control points that determine the efficiency and quality of product structure construction.

[0074] like Figure 7 As shown, the process parameter prediction model exhibits good engineering applicability. The average coefficient of determination (R²) between the model's predicted and measured values ​​on the test set reaches 0.97, fully meeting the requirements for production process control and demonstrating the model's reliability.

[0075] Example 3: Product Characteristic Prediction Product characteristic prediction model: The inputs are raw material characteristics (X1-X9) and process parameters (Y1-Y3), and the output is product characteristics (Z1-Z6).

[0076] like Figure 8 The figure shows the training loss results of the process parameter prediction model. As can be seen from the figure, the product characteristic prediction model exhibits a typical two-stage convergence pattern. In the first 20 epochs, the training / test set loss drops sharply from about 0.400 to about 0.050 at a rate of 0.018 epochs, a decrease of 87.5%. Then it enters a stable optimization phase, slowly converging to a stable state at a rate of 0.0005 epochs. The test set MSE=0.031, the training set MSE=0.035, and the difference rate Δ=11.4%. This highly synchronous convergence trend, with a correlation r=0.96, confirms the robustness of the positive prediction model.

[0077] Among the product characteristics, select fiber content, chewiness, elasticity, etc. , b-value is used as an output indicator. Among them, fiber density, as a core structural indicator, reflects the degree of formation of meat-like fiber structures in textured proteins. A higher value indicates a clearer fiber arrangement and more fiber filaments in the extrudate, resulting in a stronger meat-like ability and a closer resemblance to the structural characteristics of real meat. Chewability and elasticity reflect the mechanical response and mouthfeel characteristics of textured proteins; L, , The value is used to characterize the color attributes of a product, which is related to sensory acceptance and market application.

[0078] like Figure 9 As shown, the key parameters of the product characteristic prediction model have strong characterization capabilities. The prediction error of the model in the steady state is less than 5%, and the peak prediction accuracy for the transient process reaches 99.7%. The average coefficient of determination R² of each parameter of the product characteristic prediction model is 0.95, which reflects the reliability of the prediction model and provides a clear direction for parameter adjustment for process optimization.

[0079] Example 4: Closed-loop regulation of ternary proteins This example demonstrates the specific application of the ternary protein closed-loop regulation method: Step 1: Determine the target product parameters as follows: solubility S≥50%, elasticity E=2.8-3.2.

[0080] Step 2: Input the initial raw material ratio (gluten powder = 55%, yeast protein = 10%, soybean meal = 35%) and process parameters (barrel temperature T = 130℃, moisture content M = 19%). The model predicts that the solubility S = 47% (not up to standard) and the elasticity E = 3.0 (up to standard).

[0081] Step 3: Based on the quantitative relationship of solubility control, while keeping the barrel temperature T=130℃ and moisture content M=19% constant, increase the yeast protein ratio from 10% to 15%. At this point... =15 / (55+35)=0.167≥0.15, triggering the yeast protein ratio threshold feature.

[0082] Step 4: The model's secondary prediction yields a solubility S of 50.5% (meets the standard). Simultaneously, based on the dynamic adjustment rule of the temperature threshold, since... If the value is ≥0.15, the temperature threshold will be automatically adjusted to 150℃ (originally 140℃). The temperature can be appropriately increased to 145℃ to further optimize the taste.

[0083] Step 5: The production line executes the adjusted parameters, and actual production verification shows that the control error is ≤2.5%, which is far lower than the 12% error of the traditional trial and error method.

[0084] In terms of model training performance, the model can achieve a rapid decrease in loss within a short training period. The final mean square error (MSE) of the test set is less than 0.04, the training-test error difference rate (Δ) is controlled within 13.0%, and the convergence trajectory correlation coefficient (r) is greater than 0.95, demonstrating high fitting accuracy and engineering application potential.

[0085] In terms of predictive performance, the predictions for raw materials, processes, and product characteristics all demonstrated good reliability. Specifically, the average R² of the process parameter prediction model reached 0.97, the prediction error of the product characteristic prediction model under steady-state conditions was less than 5%, and the peak accuracy reached 99.7%. The raw material characteristic prediction model showed stability in reverse reasoning, with an average R² of 0.94, demonstrating strong reverse reasoning capabilities.

[0086] Compared to binary systems, ternary systems offer higher predictive accuracy: the ternary system has R² values ​​of 0.99 (solubility) and 0.97 (elasticity), higher than the binary system's 0.98 and 0.95. In terms of industrial applications, the ternary system reduces production setup time from 24 hours to 1.5 hours, lowers raw material loss by 18%, compared to only 15% for the binary system, and increases product qualification rate to 98%, compared to 95% for the binary system.

[0087] In summary, the method provided in this embodiment of the invention can effectively identify the nonlinear relationships between variables during the extrusion process of textured proteins, and realize multi-path and multi-directional parameter prediction. It shows good application prospects in intelligent modeling and process control of high-moisture extrusion production of textured proteins.

[0088] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; only preferred embodiments of the present invention are illustrated. The descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. As long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0089] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the appended claims.

Claims

1. A machine learning-based method for predicting ternary protein tissues, characterized in that, The method includes the following steps: Step S1: Obtain the raw material characteristic parameters, process parameters, and product characteristic parameters of three protein raw materials—gluten, yeast protein, and soybean meal—under high moisture extrusion conditions; Step S2: Based on the inhibitory effect of yeast protein on gluten cross-linking and the sensitivity of yeast protein solubility to moisture, construct ternary protein coupling features and yeast protein threshold features; the ternary protein coupling features include at least the temperature-ternary protein ratio coupling coefficient, the moisture-ternary protein cofactor, and the protein hydrophobicity coupling index; the yeast protein threshold features include at least the yeast protein moisture threshold feature and the yeast protein ratio threshold feature. Step S3: Construct a backpropagation neural network model. Use the raw material characteristic parameters, process parameters, ternary protein coupling characteristics, and yeast protein threshold characteristics as inputs to the backpropagation neural network model, and use the product characteristic parameters as outputs. Iteratively train the backpropagation neural network model until it converges to obtain a textured protein prediction model. Step S4: Use the described organized protein prediction model to perform positive quality prediction or reverse process derivation of the test system.

2. The method for predicting ternary protein organization based on machine learning according to claim 1, characterized in that: The raw material characteristic parameters mentioned in step S1 include at least protein content, carbohydrate content, and total sulfhydryl content; the process parameters include at least moisture content, barrel temperature, and screw speed; and the product characteristic parameters include at least fiber content, elasticity, and chewiness.

3. The method for predicting ternary protein organization based on machine learning according to claim 1, characterized in that: The formula for calculating the temperature-ternary protein ratio coupling coefficient in step S2 is as follows: ； In the formula, This is the temperature-ternary protein ratio coupling coefficient; The barrel temperature; The ratio of gluten powder; This refers to the yeast protein ratio; This refers to the proportion of soybean meal.

4. The method for predicting ternary protein organization based on machine learning according to claim 1, characterized in that: The formula for calculating the water-ternary protein co-factor is as follows: ； In the formula, It is a water-trin cofactor; Moisture content; This refers to the yeast protein ratio; The ratio of gluten powder; The proportion of soybean meal; The screw rotation speed; The formula for calculating the protein hydrophobic coupling index is as follows: ； In the formula, The protein hydrophobic coupling index; The hydrophobicity coefficient of wheat gluten; The hydrophobicity coefficient of soybean meal; The hydrophobicity coefficient of yeast protein; This refers to the barrel temperature.

5. The method for predicting ternary protein organization based on machine learning according to claim 1, characterized in that: The value of the yeast protein moisture threshold characteristic in step S2 is determined by the following rules: when the moisture content M ≥ 18%, the value is 1, and when the moisture content M < 18%, the value is 0. The value selection rule for the yeast protein proportion threshold feature is as follows: when When the value is 1, The value is 0; where, This refers to the yeast protein ratio. The ratio of gluten powder, This refers to the proportion of soybean meal.

6. The method for predicting ternary protein organization based on machine learning according to claim 1, characterized in that: The backpropagation neural network model described in step S3 uses a hybrid activation function, which consists of a ReLU activation function and an ELU activation function, with a ReLU activation function to ELU activation function ratio of 6:

4.

7. The method for predicting ternary protein organization based on machine learning according to claim 1, characterized in that: The backpropagation neural network model uses a weighted mean square error loss function, with different weights set for different product characteristic parameters.

8. The method for predicting ternary protein organization based on machine learning according to claim 1, characterized in that: In step S3, a constraint mechanism is applied to the weight update of the backpropagation neural network model. The constraint mechanism requires that the absolute value of the input layer weights directly connected to the ternary protein coupling feature be kept above 0.

25.

9. The method for predicting ternary protein organization based on machine learning according to claim 1, characterized in that: The positive quality prediction in step S4 is as follows: inputting the raw material characteristic parameters, process parameters, ternary protein coupling characteristics and yeast protein threshold characteristics of the system to be tested into the textured protein prediction model, and outputting the predicted values ​​of the product characteristic parameters; the reverse process derivation is as follows: based on the target product characteristic parameters, the required raw material ratio or process parameters are deduced in reverse through the textured protein prediction model.

10. A method for predicting ternary protein organization based on machine learning according to any one of claims 1-9, characterized in that: The method also includes dynamic adjustment of a temperature threshold based on a yeast protein ratio threshold: when At that time, the barrel temperature threshold was 140℃; when At that time, the barrel temperature threshold was adjusted to 150℃; among which, This refers to the yeast protein ratio. The ratio of gluten powder, This refers to the proportion of soybean meal.

Images