Food additive formula optimization method, medium and system

Abstract

The invention provides a food additive formula optimization method, medium and system, and belongs to the technical field of computational chemistry, the method comprises the following steps: firstly, establishing a food additive evaluation index system, collecting samples, measuring the content of each component to construct an initial formula component matrix, then carrying out feature extraction based on an improved convolutional neural network model, and calculating the formula of the food additive. Formula optimization is carried out in a mode of combining one-way optimization and multi-objective optimization, comprehensive evaluation is carried out by utilizing a multi-level judgment function, and finally, a final optimized formula is obtained through stability verification and fine adjustment. The whole optimization process is systematic and comprehensive, the complex interaction among the components is considered, the balance of a plurality of performance indexes is also considered, and the efficient optimization of the formula is realized. The technical problem that in the prior art, the optimization effect is not ideal due to the fact that complex nonlinear interaction among components is difficult to accurately characterize in the food additive formula optimization process is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of computational chemistry, and more particularly, relates to a method, medium, and system for optimizing food additive formulations. Background Art

[0002] Food additives are an essential and integral part of the modern food industry, and the optimization of their formulations directly affects the quality and safety of food. Traditional methods for optimizing food additive formulations mainly rely on single-factor experimental methods and orthogonal experimental methods. By setting experimental factors at different levels, the effects of the contents of various components on product performance are investigated to determine the optimal formulation. This method is simple to operate and easy to implement, and has been widely used in the food industry. With the development of the food industry, more and more new food additives have been developed, and the formulation compositions have become increasingly complex. Traditional optimization methods are no longer able to meet the needs of the modern food industry.

[0003] In recent years, some researchers have attempted to use mathematical optimization methods such as response surface methodology and genetic algorithms to optimize food additive formulations. Response surface methodology describes the relationship between formulation components and performance indicators by establishing a quadratic polynomial regression model. However, its limitation is that it can only handle simple linear or quadratic relationships and cannot accurately express complex non-linear interactions. Although genetic algorithms have the ability of global optimization, their optimization process lacks theoretical guidance, is prone to falling into local optimal solutions, and the convergence speed of the algorithm is relatively slow. Some researchers have also attempted to use neural network methods for formulation optimization, but existing neural network models often only focus on the optimization of a single performance indicator and are difficult to achieve the comprehensive optimization of multiple indicators.

[0004] Currently, the main challenge in optimizing food additive formulations lies in the complex non-linear interactions between components, which can affect various performance indicators of the product. Since traditional optimization methods cannot accurately characterize this complex interaction relationship, the optimization results are often not satisfactory, and it is even possible that the performance of the optimized formulation may decline instead. In addition, existing technologies also have problems such as a single optimization target, lack of theoretical guidance in the optimization process, and poor stability of optimization results. There is an urgent need to develop an efficient formulation optimization method that can comprehensively consider component interactions. Summary of the Invention

[0005] In view of this, the present invention provides a method, medium, and system for optimizing food additive formulations, which can solve the technical problem in the prior art that it is difficult to accurately characterize the complex non-linear interactions between components during the optimization process of food additive formulations, resulting in unsatisfactory optimization effects.

[0006] The present invention is implemented as follows: A method for optimizing the formula of food additives provided by the first aspect of the present invention includes the following steps: establishing an evaluation index system for food additives, where the evaluation index system for food additives includes sensory indexes, physical and chemical indexes, and functional indexes; collecting food additive samples, measuring the mass fractions of each component in the food additive samples, and establishing an initial formula component matrix; establishing a neural network model, inputting the initial formula component matrix into the L-th layer of the neural network model, and determining the initial convolution kernel parameters of the L-th layer based on the correlation coefficients of the indexes in the evaluation index system for food additives; obtaining the equivalent convolution kernel of the L-th layer of the neural network model according to the initial convolution kernel of the L-th layer of the neural network model, where the equivalent convolution kernel of the L-th layer of the neural network model is used to extract features of the interaction between components; constructing a single evaluation index optimization vector, and respectively performing unidirectional optimization on the initial formula component matrix based on the single evaluation index optimization vector to obtain a unidirectional optimized formula for each index in the evaluation index system for food additives; constructing a multi-objective optimization model, using the unidirectional optimized formula as a constraint condition, and performing comprehensive optimization by the Pareto optimization method.

[0007] Among them, the sensory indexes include color uniformity, odor intensity, and solubility. The measurement method of the sensory indexes includes inviting 10 trained professional reviewers to evaluate the food additive samples using a 9-point scoring method. The color uniformity is measured by visual observation under a standard light source, the odor intensity is measured immediately after opening in a sealed container, and the solubility is measured by the dissolution rate and solubility during the stirring process. The average score of the reviewers is taken as the final measurement value.

[0008] Among them, the physical and chemical indexes include pH value, viscosity, and water content. The measurement method of the physical and chemical indexes includes measuring the pH value using a precision pH meter, measuring the viscosity using an NDJ8S rotational viscometer, and measuring the water content using the vacuum drying method. Each food additive sample is measured 3 times, and the average value is taken as the final measurement value.

[0009] Among them, the functional indexes include antioxidant activity, emulsification stability, and water retention capacity. The antioxidant activity is measured by the scavenging rate of 2,2-diphenyl-1-picrylhydrazyl radical, the emulsification stability is measured by a centrifugation test, and the water retention capacity is measured by the centrifugation water retention rate.

[0010] Among them, the neural network model adopts an improved residual network structure. The neural network model includes an input layer, a feature extraction layer, a residual block layer, a fully connected layer, and an output layer. A distribution model is set before the feature extraction layer, and the distribution model includes a component balance equation, an interaction equation, a performance prediction equation, and a constraint condition equation.

[0011] Optionally, the food additive includes an emulsifier, a stabilizer, a thickener, and an antioxidant. The initial formulation of the food additive is composed of 35% by mass of monoglyceride, 25% by mass of sucrose fatty acid ester, 20% by mass of phospholipid, 10% by mass of pectin, 5% by mass of carrageenan, and 5% by mass of xanthan gum.

[0012] Optionally, the food additive is any one of a compound emulsifier for bread, a stabilizer for ice cream, a thickener for fruit juice beverage, or an antioxidant for vegetable oil. The initial formulation of the compound emulsifier for bread is composed of 33% by mass of monoglyceride, 22% by mass of sucrose fatty acid ester, 18% by mass of phospholipid, 12% by mass of carboxymethyl cellulose, 8% by mass of carrageenan, and 7% by mass of guar gum; the initial formulation of the stabilizer for ice cream is composed of 25% by mass of sodium alginate, 20% by mass of carrageenan, 15% by mass of xanthan gum, 15% by mass of guar gum, 15% by mass of carboxymethyl cellulose, and 10% by mass of pectin; the initial formulation of the thickener for fruit juice beverage is composed of 30% by mass of xanthan gum, 20% by mass of guar gum, 20% by mass of pectin, 15% by mass of carrageenan, 10% by mass of carboxymethyl cellulose, and 5% by mass of konjac gum; the initial formulation of the antioxidant for vegetable oil is composed of 35% by mass of tocopherol, 25% by mass of rosemary extract, 20% by mass of tea polyphenol, 10% by mass of citric acid, 6% by mass of ascorbyl palmitate, and 4% by mass of propyl gallate.

[0013] Wherein, after obtaining the optimization result of the multi-objective optimization model, it further includes making a decision by using the fuzzy comprehensive evaluation method, calculating the component compatibility coefficient of the food additive sample in combination with the eigenvector, constructing a third decision function, and substituting the component compatibility coefficient, the comprehensive performance score, and the functional characteristic score into the third decision function to obtain the total score of the optimized formulation.

[0014] Wherein, after selecting the formulation with the highest total score of the optimized formulation as the optimal formulation, it further includes verifying the stability of the optimal formulation, storing it at 25 degrees Celsius for 180 days, measuring all the indicators in the evaluation index system every 30 days, and fine-tuning the optimal formulation by using the neural network model according to the stability verification result to obtain the final optimized formulation.

[0015] The second aspect of the present invention provides a computer-readable storage medium, in which program instructions are stored. When the program instructions run on a computer, they are used to execute the above-mentioned method for optimizing the food additive formulation.

[0016] The third aspect of the present invention provides a food additive formula optimization system, which includes the above-mentioned computer-readable storage medium. The system can be any one of a computer, a server, and a single-chip microcomputer. The computer-readable storage medium is set inside the system, and a microprocessor for executing the program instructions stored in the computer-readable storage medium is set inside the system.

[0017] Compared with the prior art, the present invention provides a food additive formula optimization method, medium and system. The present invention proposes a food additive formula optimization method based on an improved neural network. By establishing a complete evaluation index system and combining it with an improved convolutional neural network model, the method realizes the accurate characterization and prediction of the complex interaction between formula components. The method adopts a convolutional kernel parameter initialization strategy based on correlation analysis to improve the accuracy of feature extraction; by introducing equivalent convolutional kernels, the expression ability of the model for component interaction is enhanced.

[0018] The method of the present invention innovatively combines unidirectional optimization and multi-objective optimization, which not only ensures the optimization effect of each performance index but also realizes the comprehensive balance of multiple indicators. By constructing a multi-level decision function system, the comprehensive evaluation of the formula performance is realized; the fuzzy comprehensive evaluation method is used for decision analysis to improve the reliability of the optimization result. Especially in the optimization process, the component compatibility is considered and a variance penalty term is introduced, which effectively improves the stability of the optimized formula.

[0019] Essentially, the present invention solves the core problem that it is difficult to characterize the interaction between components in the optimization of food additive formulas, which mainly benefits from the following aspects: First, the improved neural network model can automatically learn and extract the non-linear relationship between components; second, the multi-level optimization strategy ensures the comprehensiveness and reliability of the optimization process; finally, the component compatibility and stability requirements are considered, making the optimization result more practical. In summary, the present invention solves the technical problem in the prior art that it is difficult to accurately characterize the complex non-linear interaction between components in the process of optimizing food additive formulas, resulting in unsatisfactory optimization effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a flowchart of the method of the present invention.

[0021] Figure 2 It is a graph showing the change trend of color uniformity with the number of iterations during the unidirectional optimization of the compound emulsifier for bread in Example 2.

[0022] Figure 3 It is a radar chart comparison diagram of various performance indicators of the compound emulsifier for bread before and after optimization in Example 2.

[0023] Figure 4It is a graph showing the emulsion stability and the change in ice crystal size of the ice cream stabilizer during low-temperature storage in Example 3.

[0024] Figure 5 It is a heat map showing the correlation between the component contents and performance indicators in Example 3. Detailed implementation manners

[0025] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0026] As Figure 1 shown, it is a flowchart of a method for optimizing a food additive formula provided in the first aspect of the present invention. This method includes the following steps: S1. Establish a food additive evaluation index system. The food additive evaluation index system includes sensory indexes, physical and chemical indexes and functional indexes. The sensory indexes include color uniformity, odor intensity, and solubility. The physical and chemical indexes include pH value, viscosity, and water content. The functional indexes include antioxidant activity, emulsion stability, and water retention capacity; S2. Collect the food additive samples, measure the mass fractions of the components in the food additive samples, and establish an initial formula component matrix; S3. Establish a neural network model, input the initial formula component matrix into the L-th layer of the neural network model. The initial convolution kernel parameters of the L-th layer are determined based on the correlation coefficients of the indexes in the food additive evaluation index system; S4. Obtain the equivalent convolution kernel of the L-th layer of the neural network model according to the initial convolution kernel of the L-th layer of the neural network model. The equivalent convolution kernel of the L-th layer of the neural network model is used to extract features of the interaction between components; S5. Measure the values of the indexes in the food additive evaluation index system of the food additive samples to obtain an initial performance parameter matrix, and use the initial performance parameter matrix as the input of the (L + 1)-th layer of the neural network model; S6. Construct a single evaluation index optimization vector, establish an optimization objective function for each index in the food additive evaluation index system, and generate the single evaluation index optimization vector; S7. Based on the single evaluation index optimization vector, perform unidirectional optimization on the initial formula component matrix respectively to obtain unidirectional optimized formulas for each index in the food additive evaluation index system; S8. Construct a first decision function, substitute the measured values of the sensory indexes and physical and chemical indexes in the food additive evaluation index system into the first decision function to obtain a comprehensive performance score; S9. Construct a second determination function, substitute the measured values of the functional indicators in the food additive evaluation index system into the second determination function to obtain the functional characteristic scores; S10. Construct a multi-objective optimization model, use the unidirectional optimized formula as a constraint condition, and perform comprehensive optimization using the Pareto optimization method; S11. Design a formula optimization plan based on the orthogonal test method, adjust the mass fractions of each component on the basis of the initial formula component matrix, and generate multiple groups of optimized formulas by the way of directional expansion considering the jitter results of establishing independent optimization objective functions for each evaluation index; S12. Prepare food additive samples of the multiple groups of optimized formulas, and obtain the measured values of all the indicators in the food additive evaluation index system of the food additive samples; S13. Input the measured values of the food additive samples into the neural network model to obtain the feature vectors of the food additive samples; S14. Use the fuzzy comprehensive evaluation method to make a decision on the optimization results of the multi-objective optimization model, and calculate the component compatibility coefficient of the food additive samples in combination with the feature vectors; S15. Substitute the component compatibility coefficient, the comprehensive performance score and the functional characteristic score into the third determination function to obtain the total score of the optimized formula; S16. Perform normalization processing on the total scores of the optimized formulas of the multiple groups of optimized formulas, and select the formula with the highest total score of the optimized formula as the optimal formula; S17. Obtain the stability verification results of the optimal formula, specifically store it at 25 degrees Celsius for 180 days, and measure all the indicators in the evaluation index system every 30 days; S18. According to the stability verification results, use the neural network model to fine-tune the optimal formula to obtain the final optimized formula.

[0027] The determination method of the sensory indicators includes inviting 10 trained professional reviewers to evaluate the food additive samples using a 9-point scoring method. The color uniformity is determined by visual observation under a standard light source, the odor intensity is determined immediately after opening in a sealed container, the solubility is determined by the dissolution rate and solubility during the stirring process, and the arithmetic mean of the reviewers' scores is taken as the final measured value.

[0028] The determination method of the physical and chemical indicators includes using a precision pH meter to measure the pH value, using an NDJ8S rotary viscometer to measure the viscosity, and using the vacuum drying method to measure the water content. Each food additive sample is measured 3 times repeatedly, and the average value is taken as the final measured value.

[0029] The food additives include emulsifiers, stabilizers, thickeners, and antioxidants. The initial formulation of the food additives consists of 35 mass percentages of monoglyceride, 25 mass percentages of sucrose fatty acid ester, 20 mass percentages of phospholipid, 10 mass percentages of pectin, 5 mass percentages of carrageenan, and 5 mass percentages of xanthan gum.

[0030] Each element in the initial formulation component matrix represents the mass percentage of each component in the total formulation. The dimension of the initial formulation component matrix is n×m, where n represents the number of formulation components and m represents the number of batches of food additive samples.

[0031] The neural network model adopts an improved residual network structure. The neural network model includes an input layer, a feature extraction layer, a residual block layer, a fully connected layer, and an output layer. A distribution model is set before the feature extraction layer, and the distribution model includes a component balance equation, an interaction equation, a performance prediction equation, and a constraint condition equation.

[0032] The component balance equation is used to calculate that the sum of the mass percentages of each component is 100 mass percentages. The input parameters of the component balance equation include the mass percentages of each component, and the output result of the component balance equation is the normalized component ratio.

[0033] The interaction equation is used to describe the interaction between components. The input parameters of the interaction equation include the component pair interaction coefficient matrix, and the output result of the interaction equation is the component interaction intensity value.

[0034] The performance prediction equation is used to predict the formulation performance index. The input parameters of the performance prediction equation include the component ratio and the component interaction intensity value, and the output result of the performance prediction equation is the predicted performance index value.

[0035] The initial performance parameter matrix is composed of the measured values of each evaluation index. The dimension of the initial performance parameter matrix is p×q, where p represents the number of evaluation indexes and q represents the number of batches of food additive samples.

[0036] The input parameters of the first decision function include the measured values of sensory indexes and physical and chemical indexes after normalization processing. The output result of the first decision function is a comprehensive performance score between 0 and 1.

[0037] The input parameters of the second decision function include the measured values of functional indexes after normalization processing. The output result of the second decision function is a functional characteristic score between 0 and 1.

[0038] The specific implementation manners of the above steps are described in detail below. The specific implementation manner of step S1 is to establish a comprehensive food additive evaluation index system, which includes three categories: sensory indexes, physical and chemical indexes, and functional indexes. The sensory index evaluation adopts a 9-point scoring standard. The scoring standard for color uniformity is as follows: 9 points indicate that the color is completely uniform, 7 points indicate that it is basically uniform, 5 points indicate slight non-uniformity, 3 points indicate obvious non-uniformity, and 1 point indicates serious non-uniformity; the scoring standard for odor intensity is as follows: 9 points indicate that the odor is moderate, 7 points indicate that the odor is relatively moderate, 5 points indicate that the odor is slightly strong or slightly weak, 3 points indicate that the odor is significantly too strong or too weak, and 1 point indicates that the odor is extremely inappropriate; the scoring standard for solubility is as follows: 9 points indicate rapid and complete dissolution, 7 points indicate relatively fast dissolution, 5 points indicate normal dissolution, 3 points indicate slow dissolution, and 1 point indicates difficult dissolution. The physical and chemical indexes are measured by precision instruments. The qualified range of pH value is 6.0 to 8.0, the qualified range of viscosity is 500 to 2000 millipascal seconds, and the qualified range of water content is 2% to 5%. In the functional indexes, the antioxidant activity is measured by the scavenging rate of 2,2-diphenyl-1-picrylhydrazyl radical, the emulsion stability is measured by a centrifugation test, and the water retention capacity is measured by the centrifugation water retention rate. The purpose of this step is to establish a scientific and reasonable evaluation system to provide a basis for subsequent optimization.

[0039] The specific implementation manner of step S2 is to collect food additive samples and conduct component analysis. First, the content of monoglyceride is measured by high performance liquid chromatography, the content of sucrose fatty acid ester is measured by gas chromatography, the content of phospholipid is measured by thin layer chromatography, the content of pectin is measured by spectrophotometry, and the content of carrageenan and xanthan gum is measured by gravimetry. The measurement results are recorded according to the mass percentage to construct an initial formulation component matrix, and each element in the matrix represents the percentage content of the corresponding component in the total formulation. This step adopts the principle of chromatographic analysis, and the purpose is to accurately obtain the content data of each component to provide data support for subsequent modeling.

[0040] The specific implementation manner of step S3 is to construct the L-th layer structure of the neural network model. This layer adopts an improved convolutional neural network structure. First, the Pearson correlation coefficient between each index in the food additive evaluation index system is calculated as the correlation coefficient, and the correlation coefficient matrix is used as the basis for the initial convolutional kernel parameters. The size of the convolutional kernel is set to 3×3, the stride is 1, and the padding method is same padding. This step adopts the principle of deep learning. By correlation analysis, the initial convolutional parameters are determined, and the purpose is to establish a network structure capable of extracting component features.

[0041] The specific implementation of step S4 is to calculate the equivalent convolution kernel based on the initial convolution kernel of the L-th layer. The initial convolution kernel is transformed into the frequency domain using the fast Fourier transform method, and convolution operations are performed in the frequency domain to obtain the frequency domain representation of the equivalent convolution kernel. Then, the equivalent convolution kernel in the spatial domain is obtained through the inverse Fourier transform. This step uses signal processing principles with the aim of improving the efficiency and accuracy of feature extraction.

[0042] The specific implementation of step S5 is to measure the numerical values of various indicators of the food additive sample. For sensory indicators, the average value is taken after 10 reviewers independently score; for physical and chemical indicators, each sample is measured 3 times and the average value is taken; for functional indicators, they are measured according to the standard method. All the measurement results are organized into an initial performance parameter matrix, which serves as the input for the (L + 1)-th layer of the neural network model. This step uses statistical analysis methods with the aim of obtaining accurate performance parameter data.

[0043] The specific implementation of step S6 is to construct an optimization vector for a single evaluation indicator. For sensory indicators, the goal is to maximize the score value; for physical and chemical indicators, the goal is to make the measured value close to the middle value of the standard range; for functional indicators, the goal is to maximize the activity value, stability value, and water retention rate. The normalization processing method is used to unify the optimization goals of each indicator into the range of 0 to 1. This step uses the multi-objective optimization principle with the aim of clarifying the optimization directions of each indicator.

[0044] The specific implementation of step S7 is to perform unidirectional optimization. The gradient descent method is used to adjust the content of each component in the initial formulation component matrix, with a step size of 0.1% each time, until the optimization goal of a single indicator is achieved or the maximum number of iterations, 100 times, is reached. This step uses the optimization algorithm with the aim of obtaining the optimal formulation for a single indicator.

[0045] The specific implementation of step S8 is to construct a first decision function. Using the weighted summation method, the weight of the sensory indicator is set to 0.6, and the weight of the physical and chemical indicator is set to 0.4. The measured values of each indicator are multiplied by the corresponding weights after normalization processing and then summed to obtain the comprehensive performance score. This step uses the multi-criteria decision-making method with the aim of quantitatively evaluating the comprehensive performance of the formulation.

[0046] The specific implementation of step S9 is to construct a second decision function. Using the fuzzy comprehensive evaluation method, the weights of the three functional indicators of antioxidant activity, emulsification stability, and water retention ability are set to 0.4, 0.3, and 0.3 respectively. A membership function is established to calculate the functional characteristic score. This step uses the principles of fuzzy mathematics with the aim of evaluating the functional characteristics of the formulation.

[0047] The specific implementation of step S10 is to construct a multi-objective optimization model. Using the formulation obtained from single-objective optimization as the initial solution, the non-dominated sorting genetic algorithm is adopted for multi-objective optimization. The population size is set to 50, the number of generations is set to 100, the crossover probability is 0.8, and the mutation probability is 0.1. This step uses the principle of evolutionary computation, aiming to find the balanced solutions of multiple optimization objectives.

[0048] The specific implementation of step S11 is to design a formulation optimization plan. Using the orthogonal experiment method, select the L16(45) orthogonal array. Based on the initial formulation, adjust the contents of each component at 5 levels, and considering the jitter range of evaluation indicators, generate 16 groups of optimized formulations. This step uses the experimental design method, aiming to systematically explore the formulation space.

[0049] The specific implementation of step S12 is to prepare samples of the optimized formulation. Using the homogenization and emulsification process, control the temperature at 60 to 70 degrees Celsius, the homogenization pressure at 25 MPa, and the homogenization time at 10 minutes. After preparation, measure all evaluation indicators. This step uses the principle of food technology, aiming to obtain actual performance data.

[0050] The specific implementation of step S13 is feature extraction. Input the measured values of the samples into the trained neural network model. After multiple convolutional and pooling operations, extract the feature vectors of the samples. This step uses the principle of deep learning feature extraction, aiming to obtain the feature representations of the samples.

[0051] The specific implementation of step S14 is to conduct decision analysis. Using the fuzzy comprehensive evaluation method, establish a judgment matrix, calculate the weight vector, and combine the feature vectors to calculate the component compatibility coefficient. This step uses the decision analysis method, aiming to evaluate the overall performance of the formulation.

[0052] The specific implementation of step S15 is to construct a third decision function. Using the weighted geometric mean method, set the weights of the component compatibility coefficient, the comprehensive performance score, and the functional characteristic score to 0.3, 0.4, and 0.3 respectively, and calculate the total score of the optimized formulation. This step uses the comprehensive evaluation method, aiming to obtain the final score of the formulation.

[0053] The specific implementation of step S16 is normalization processing. Using the maximum-minimum normalization method, map the total scores of all formulations to the interval of 0 to 1, and select the formulation with the highest score as the optimal formulation. This step uses the principle of data normalization, aiming to achieve fair comparison of formulations.

[0054] The specific implementation of step S17 is to conduct stability verification. Store at a constant temperature of 25 degrees Celsius for 180 days, and measure all evaluation indicators once every 30 days according to the standard method, and record the data change trend. This step uses the stability test method, aiming to verify the long-term stability of the formulation.

[0055] The specific implementation of step S18 is formula fine-tuning. According to the stability verification results, the backpropagation algorithm of the neural network model is used to fine-tune the content of the key components affecting stability in the optimal formula, and the adjustment range does not exceed 5% of the original content to obtain the final optimized formula. This step adopts the principle of machine learning optimization, aiming to improve the stability of the formula.

[0056] The specific representation of the initial formula component matrix is as follows: ; In the formula, represents the mass percentage of the th component in the th batch, is the number of components, is the number of sample batches; , indicating that the sum of the mass percentages of all components in each batch is 100.

[0057] The specific representation of the component balance equation is as follows: ; In the formula, is the mass percentage of the th component; and are respectively the minimum and maximum allowable mass percentages of the th component; and are penalty coefficients, and the value range is from 1.0 to 10.0.

[0058] The specific representation of the interaction equation is as follows: ; In the formula, is the component pair interaction coefficient; is the three-component interaction coefficient; these coefficients are obtained through orthogonal experiments.

[0059] The specific representation of the performance prediction equation is as follows: ; In the formula, is the individual performance contribution function of the th component; is the weight coefficient; and are adjustment coefficients, and the range is from 0.1 to 1.0; is the error term, which follows the normal distribution .

[0060] The calculation method of the initial convolution kernel parameters of the L-th layer of the neural network is as follows: ; In the formula, is the Pearson correlation coefficient between evaluation indicators, and is calculated by the following formula: ; In the formula, and are respectively the measured values of the -th and -th evaluation indicators in the -th sample; and are the average values of the corresponding indicators.

[0061] The calculation process of the equivalent convolution kernel is as follows: ; In the formula, and respectively represent two-dimensional fast Fourier transform and inverse transform; is the input feature matrix.

[0062] The construction method of the single evaluation index optimization vector is as follows: ; In the formula, is the measured value of the -th evaluation indicator; and are respectively the minimum and maximum allowable values of this indicator; is the target value; is the adjustment coefficient, and the range is from 0.01 to 0.1.

[0063] The specific representation of the first decision function is as follows: ; In the formula, and are respectively the normalized values of sensory indicators and physical and chemical indicators; and are the corresponding weights; and are respectively the numbers of sensory indicators and physical and chemical indicators.

[0064] The specific representation of the second decision function is as follows: ; In the formula, is the normalized value of the function indicator; is the corresponding weight; is the number of function indicators.

[0065] The specific representation of the third determination function is as follows: ; In the formula, is the component compatibility coefficient; , , are the weight coefficients, and their sum is 1; is the penalty coefficient, with a range of 0.1 to 1.0; is the sensitivity coefficient, with a range of 1 to 10; is the variance of the sample performance index.

[0066] As an optional implementation method 1 is the formulation optimization of the compound emulsifier for bread. The initial formulation consists of 33 mass percentages of monoglyceride, 22 mass percentages of sucrose fatty acid ester, 18 mass percentages of phospholipid, 12 mass percentages of carboxymethyl cellulose, 8 mass percentages of carrageenan, and 7 mass percentages of guar gum. Through the established food additive evaluation index system, it is measured that in the sensory indexes of the initial formulation, the color uniformity is 6.8 points, the odor intensity is 7.2 points, and the solubility is 6.5 points; in the physical and chemical indexes, the pH value is 6.8, the viscosity is 1200 mPa·s, and the water content is 3.8%; in the functional indexes, the antioxidant activity is 68.5%, the emulsification stability is 82.3%, and the water retention capacity is 75.6%. After neural network model analysis and multi-objective optimization, the final optimized formulation is 35 mass percentages of monoglyceride, 20 mass percentages of sucrose fatty acid ester, 20 mass percentages of phospholipid, 10 mass percentages of carboxymethyl cellulose, 8 mass percentages of carrageenan, and 7 mass percentages of guar gum. The indexes of the optimized formulation are significantly improved. Among them, the color uniformity is increased to 8.5 points, the odor intensity is increased to 8.3 points, and the solubility is increased to 8.0 points; the pH value is adjusted to 7.2, the viscosity is reduced to 1000 mPa·s, and the water content is reduced to 3.2%; the antioxidant activity is increased to 75.8%, the emulsification stability is increased to 89.6%, and the water retention capacity is increased to 82.3%. After 180 days of stability verification, the change range of each index is within 5%, indicating that the optimized formulation has good application effects.

[0067] As an alternative embodiment, 2 is the optimization of the formula of the ice cream stabilizer. The initial formula composition is 25 mass percentages of sodium alginate, 20 mass percentages of carrageenan, 15 mass percentages of xanthan gum, 15 mass percentages of guar gum, 15 mass percentages of carboxymethyl cellulose, and 10 mass percentages of pectin. Among the sensory indexes of the initial formula, the color uniformity is 7.0 points, the odor intensity is 6.8 points, and the solubility is 6.2 points. Among the physical and chemical indexes, the pH value is 7.0, the viscosity is 1500 millipascal seconds, and the water content is 4.2%. Among the functional indexes, the antioxidant activity is 62.3%, the emulsification stability is 78.5%, and the water retention capacity is 72.8%. Through the optimization method of the present invention, analyzed by a neural network model, after unidirectional optimization and multi-objective optimization, the final optimized formula is 28 mass percentages of sodium alginate, 18 mass percentages of carrageenan, 15 mass percentages of xanthan gum, 15 mass percentages of guar gum, 14 mass percentages of carboxymethyl cellulose, and 10 mass percentages of pectin. The optimized formula shows better stability under low-temperature conditions. The color uniformity is increased to 8.6 points, the odor intensity is increased to 8.2 points, and the solubility is increased to 8.4 points. The pH value is adjusted to 7.3, the viscosity is increased to 1800 millipascal seconds, and the water content is reduced to 3.8%. The antioxidant activity is increased to 70.5%, the emulsification stability is increased to 85.6%, and the water retention capacity is increased to 80.2%. The stability verification shows that the performance of this formula remains good under freezing conditions.

[0068] As an alternative embodiment, 3 is the formulation optimization of thickeners for fruit juice beverages. The initial formulation consists of 30 mass percentages of xanthan gum, 20 mass percentages of guar gum, 20 mass percentages of pectin, 15 mass percentages of carrageenan, 10 mass percentages of carboxymethyl cellulose, and 5 mass percentages of konjac gum. Among the sensory indicators of the initial formulation, the color uniformity is 6.5 points, the odor intensity is 7.0 points, and the solubility is 6.8 points. Among the physical and chemical indicators, the pH value is 6.5, the viscosity is 800 millipascal seconds, and the water content is 4.5%. Among the functional indicators, the antioxidant activity is 58.6%, the emulsification stability is 75.2%, and the water retention capacity is 70.5%. Through the optimization by the method of the present invention, the neural network model is used to analyze the interactions between components, and combined with the results of single-direction optimization and multi-objective optimization, the final optimized formulation is determined to be 32 mass percentages of xanthan gum, 18 mass percentages of guar gum, 22 mass percentages of pectin, 15 mass percentages of carrageenan, 8 mass percentages of carboxymethyl cellulose, and 5 mass percentages of konjac gum. The optimized formulation shows excellent thickening effect and thermal stability in the fruit juice system, and all indicators are significantly improved. The color uniformity is increased to 8.3 points, the odor intensity is increased to 8.5 points, and the solubility is increased to 8.2 points. The pH value is adjusted to 6.8, the viscosity is increased to 1000 millipascal seconds, and the water content is reduced to 4.0%. The antioxidant activity is increased to 65.8%, the emulsification stability is increased to 82.3%, and the water retention capacity is increased to 78.6%. After stability verification, this formulation shows stability under different acidity conditions.

[0069] As an alternative embodiment, 4 is the formulation optimization of antioxidants for vegetable oils. The initial formulation consists of 35 mass percentage of tocopherol, 25 mass percentage of rosemary extract, 20 mass percentage of tea polyphenols, 10 mass percentage of citric acid, 6 mass percentage of ascorbyl palmitate, and 4 mass percentage of propyl gallate. In the sensory indexes of the initial formulation, the color uniformity is 7.2 points, the odor intensity is 6.5 points, and the solubility is 6.6 points. In the physical and chemical indexes, the pH value is 6.2, the viscosity is 600 mPa·s, and the water content is 2.8%. In the functional indexes, the antioxidant activity is 72.5%, the emulsion stability is 70.8%, and the water retention capacity is 65.3%. By using the optimization method of the present invention, the synergistic effect of each component is analyzed through a neural network model, and the final formulation obtained after optimization is 38 mass percentage of tocopherol, 22 mass percentage of rosemary extract, 20 mass percentage of tea polyphenols, 10 mass percentage of citric acid, 6 mass percentage of ascorbyl palmitate, and 4 mass percentage of propyl gallate. The optimized formulation exhibits excellent antioxidant performance, with the color uniformity increased to 8.8 points, the odor intensity increased to 8.4 points, and the solubility increased to 8.5 points. The pH value is adjusted to 6.5, the viscosity is adjusted to 650 mPa·s, and the water content is reduced to 2.5%. The antioxidant activity is significantly increased to 85.6%, the emulsion stability is increased to 78.5%, and the water retention capacity is increased to 72.6%. The stability verification shows that the antioxidant effect of this formulation is persistent and stable under normal temperature storage conditions.

[0070] The second aspect of the present invention provides a computer-readable storage medium, in which program instructions are stored. When the program instructions run on a computer, they are used to execute the above-mentioned food additive formulation optimization method.

[0071] The third aspect of the present invention provides a food additive formulation optimization system, which includes the above-mentioned computer-readable storage medium. The system can be any one of a computer, a server, and a single-chip microcomputer. The computer-readable storage medium is set inside the system, and a microprocessor for executing the program instructions stored in the computer-readable storage medium is set inside the system.

[0072] Specifically, the principle of the present invention is as follows: The technical principle of the present invention is based on deep learning and multi-objective optimization theory. In terms of deep learning, an improved convolutional neural network structure is adopted to realize the extraction and representation of complex features through multiple non-linear transformations. The initialization of the convolutional kernel parameters is based on the correlation analysis between evaluation indexes. This method can introduce prior knowledge into the model and improve the pertinence of feature extraction. The introduction of equivalent convolutional kernels further enhances the expression ability of the model, enabling it to capture the high-order interaction effects between components.

[0073] In terms of the optimization strategy, the present invention adopts a combination of unidirectional optimization and multi-objective optimization. In the unidirectional optimization stage, the gradient descent method is used to iteratively optimize to make each performance index reach the expected goal. In the multi-objective optimization stage, the Pareto optimization method is adopted to achieve overall balance on the basis of ensuring the performance of each index. This optimization strategy not only considers the optimization effect of a single index but also takes into account the trade-off between multiple indexes, meeting the actual application requirements.

[0074] The construction of the decision function system is based on the multi-level evaluation theory. The first decision function evaluates the sensory and physical-chemical indexes by weighted summation, the second decision function uses the geometric mean method to evaluate the functional indexes, and the third decision function comprehensively considers the component compatibility and performance stability. This multi-level evaluation system can comprehensively reflect the various performances of the formula and provide a reliable basis for the optimization decision.

[0075] A specific embodiment 1 of the present invention is provided below, and the specific implementation manners of each step in this embodiment 1 are described in detail as follows.

[0076] The specific implementation manner of step S1 is to establish an evaluation index system for food additives, which includes sensory indexes, physical-chemical indexes and functional indexes. Specifically, when implementing, first, based on the analytic hierarchy process, the hierarchical structure of the evaluation index system is constructed, and each index is stratified according to its importance. The sensory indexes include color uniformity, odor intensity, and solubility, and each index adopts a 9-point scoring standard, and the higher the scoring value, the better the performance. The scoring results of the reviewers are standardized through the following equation: ; In the formula, is the standardized score of the th reviewer for the th index; is the original scoring value; and are the minimum and maximum scoring values of this index respectively. The consistency of the reviewers' scores is tested by the Kendall's coefficient of concordance, and the calculation formula of the coefficient of concordance is: ; In the formula, is the sum of the squared deviations of the scores of each reviewer; is the number of reviewers; is the number of evaluation indexes. When is greater than 0.7, it is considered that the scoring results have good consistency. The physical-chemical indexes are measured by standard analysis methods, and each sample is measured 3 times repeatedly, and the relative standard deviation of the measurement results should be less than 5%. When measuring the functional indexes, the free radical scavenging rate method is used for the antioxidant activity, and the calculation formula of the scavenging rate is: ; In the formula, is the free radical scavenging rate; is the absorbance after adding the sample; is the absorbance of the blank control. The emulsion stability was determined by a centrifugation test, and the stability index calculation formula is: ; In the formula, is the stability index; is the height of the emulsion layer after centrifugation; is the total height before centrifugation. The water retention capacity was determined by the centrifugal water retention rate, and the calculation formula is: ; In the formula, is the water retention rate; is the mass of the sample before centrifugation; is the mass of the water lost after centrifugation. The evaluation index system established in this step provides a quantitative basis for subsequent optimization.

[0077] The specific implementation of step S2 is to collect food additive samples and determine the content of each component to construct an initial formulation component matrix. When sampling, the stratified random sampling method is used, and the sampling amount of each batch of samples is not less than 100 grams. The component content is determined by chromatographic analysis methods, specifically including high performance liquid chromatography, gas chromatography and thin layer chromatography. The initial formulation component matrix is constructed through the following equation based on the component content determination results: ; In the formula, represents the mass percentage of the th component in the th batch, is the number of components, is the number of sample batches. The precision of the determination of each component content is evaluated by the relative standard deviation, and the calculation formula is: ; In the formula, is the relative standard deviation; is the measured value; is the average value; is the number of repeated determinations. When is less than 2%, the measurement result is considered reliable. The accuracy of the component content determination is evaluated by the spike recovery rate, and the calculation formula is: ; In the formula, is the recovery rate; is the measured concentration of the spiked sample; is the measured concentration of the unspiked sample; is the theoretical spike amount. When the recovery rate is between 95% and 105%, the measurement result is considered accurate. This step provides the basic data for subsequent modeling.

[0078] The specific implementation of step S3 is to construct the structure of the L-th layer of the neural network model. This layer adopts an improved convolutional neural network structure. First, calculate the Pearson correlation coefficient between evaluation indicators and construct the initial convolutional kernel parameter matrix: ; In the formula, is the correlation coefficient between evaluation indicators and is calculated through the following formula: ; In the formula, and are respectively the measured values of the -th and the -th evaluation indicators in the -th sample; and are the average values of the corresponding indicators. The convolutional kernel size is set to 3×3, the stride is 1, and the padding method is same padding. The feature map calculation formula for the convolutional operation is: ; In the formula, is the output feature map; are the convolutional kernel parameters; is the input feature; and are the height and width of the convolutional kernel; and are the strides; is the bias term. This step establishes a network structure capable of extracting component features.

[0079] The specific implementation of step S4 is to calculate the equivalent convolutional kernel of the L-th layer. Using the fast Fourier transform method, the initial convolutional kernel is transformed to the frequency domain for operation. The calculation process of the equivalent convolutional kernel is: ; In the formula, and respectively represent the two-dimensional fast Fourier transform and the inverse transform; is the input feature matrix. The frequency domain transformation adopts the discrete Fourier transform, and the calculation formula is: ; In the formula, is the spatial domain signal; is the frequency domain signal; and are the signal dimensions; is the imaginary unit. This step improves the computational efficiency of feature extraction.

[0080] The specific implementation of step S5 is to measure the numerical values of the evaluation indexes of the sample and construct an initial performance parameter matrix. When measuring the sensory indexes, the reliability of the scoring results of the reviewers is verified through analysis of variance, and the calculation formula is: ; In the formula, is the mean square between groups; is the mean square within groups. When the value is greater than the critical value, it is considered that there are significant differences in the scoring results. The measurement uncertainty evaluation formula for the physical and chemical indexes is: ; In the formula, is the combined standard uncertainty; is the sensitivity coefficient; is the standard uncertainty. This step obtains reliable performance parameter data.

[0081] The specific implementation of step S6 is to construct an optimization vector for a single evaluation index and establish an optimization objective function in the form of an exponential function: ; In the formula, is the measured value of the th evaluation index; and are the minimum and maximum allowable values of this index respectively; is the target value; is the adjustment coefficient, and the range is from 0.01 to 0.1. The gradient calculation formula of the optimization objective function is: ; This step establishes the optimization objectives of each index.

[0082] The specific implementation of step S7 is to perform unidirectional optimization and adjust the formula using the gradient descent method. The update formula for each iteration is: ; In the formula, is the content of the th component at the th iteration; is the learning rate, and the initial value is set to 0.1. The convergence criterion is: ; In the formula, is the convergence threshold, which is set to 0.001. This step obtains the optimized formula for a single index.

[0083] The specific implementation of step S8 is to construct a first decision function in the form of weighted summation: ; In the formula, and are the normalized values of sensory indicators and physicochemical indicators respectively; and are the corresponding weights. The weight coefficients are determined by the analytic hierarchy process, and the calculation formula for the consistency ratio of the judgment matrix is: ; In the formula, is the consistency index; is the random consistency index. When , it is considered that the weight distribution is reasonable. This step realizes the comprehensive performance evaluation.

[0084] The specific implementation of step S9 is to construct a second judgment function in the form of weighted geometric mean: ; In the formula, is the normalized value of the function index; is the corresponding weight. The normalization of the index adopts the fuzzy membership function, and the calculation formula is: ; In the formula, and are the lower limit value and the upper limit value. This step evaluates the functional characteristics of the formula.

[0085] The specific implementation of step S10 is to construct a multi-objective optimization model using the Pareto optimization method. First, establish the dominance relation judgment criterion: ; In the formula, and are two solutions; is the th objective function. The update of the non-dominated solution set adopts the fast non-dominated sorting algorithm, and the crowding degree calculation formula is: ; In the formula, is the crowding degree of the th solution; is the number of objective functions. This step obtains the balanced solution of multi-objective optimization.

[0086] The specific implementation of step S11 is to adopt the orthogonal experiment to design the formula optimization scheme. The significance of each factor is evaluated by variance analysis, and the calculation formula is: ; In the formula, is the sum of squares of factor A; is the experimental observation value; is the number of repetitions; is the total number of tests. When the value is greater than the critical value of the significance level 0.05, the factor is considered significant. The jitter range of the formula is determined by the response surface method, and the quadratic regression equation is: ; where is the regression coefficient; is the random error. This step generates multiple groups of optimized formulas.

[0087] The specific implementation of step S12 is to prepare samples of the optimized formula and conduct performance measurements. The samples are prepared using the precise weighing method, and the precision of the balance is 0.0001 grams. The calculation formula for the shear rate during the homogenization process is: ; where is the shear rate; is the rotor radius; is the rotational speed; is the clearance. This step obtains the actual performance data.

[0088] The specific implementation of step S13 is feature extraction. The calculation formula for the forward propagation of the neural network is: ; where is the activation value of the th layer; is the weight matrix; is the bias vector; is the activation function. The rectified linear unit function is used as the activation function: ; This step extracts the feature vector of the sample.

[0089] The specific implementation of step S14 is to calculate the component compatibility coefficient. The calculation process of fuzzy comprehensive evaluation is: ; where is the fuzzy relation matrix; is the weight vector; is the fuzzy composition operator. The calculation formula for the compatibility coefficient is: ; where is the correlation coefficient between components. This step evaluates the overall performance of the formula.

[0090] The specific implementation of step S15 is to construct the third decision function: ; In the formula, , , are weight coefficients; is a penalty coefficient; is a sensitivity coefficient; is the variance of the performance index. The calculation formula of the variance is: ; The total score of the formula is obtained in this step.

[0091] The specific implementation of step S16 is normalization processing and selection of the optimal formula. The min-max normalization method is adopted: ; In the formula, is the score after normalization. The optimal formula is determined in this step.

[0092] The specific implementation of step S17 is to conduct stability verification. The Arrhenius formula is used to evaluate stability: ; In the formula, is the reaction rate constant; is the pre-exponential factor; is the activation energy; is the gas constant; is the absolute temperature. The long-term stability of the formula is verified in this step.

[0093] The specific implementation of step S18 is formula fine-tuning. The backpropagation algorithm is used to update the network parameters: ; In the formula, is the error function; is the weight; is the error term; is the activation value. The final optimized formula is obtained in this step.

[0094] To better understand and implement the present invention, Example 2 of a specific application scenario of the present invention is provided below: A certain food research institute conducts research on the formula optimization of compound emulsifiers for bread: The researchers first carry out market research and systematically analyze the current situation of the use of compound emulsifiers and their performance requirements in the bread processing process. After a large amount of literature review and preliminary laboratory research, the following formula optimization objectives are determined: improving the uniformity and solubility of the emulsifier, improving its pH value and viscosity, and enhancing the antioxidant performance and emulsification stability. To achieve this goal, a complete evaluation index system is established, including three categories: sensory indexes, physical and chemical indexes, and functional indexes. Table 1 lists the evaluation indexes and their evaluation criteria.

[0095] Table 1 Evaluation Index System of Compound Emulsifiers for Bread

[0096] After determining the evaluation index system, the research team designed an initial formula. Based on the functional characteristics and usage experience of each component, monoglyceride, sucrose fatty acid ester, and phospholipid were selected as the main emulsifying components, and carboxymethyl cellulose, carrageenan, and guar gum were selected as the stabilizing components. The specific composition of the initial formula was: monoglyceride 33 mass percentage, sucrose fatty acid ester 22 mass percentage, phospholipid 18 mass percentage, carboxymethyl cellulose 12 mass percentage, carrageenan 8 mass percentage, and guar gum 7 mass percentage.

[0097] When determining the component content of the collected samples, the following methods were used: The contents of monoglyceride and sucrose fatty acid ester were determined by high performance liquid chromatography. The chromatographic conditions were a C18 column, the mobile phase was a methanol-water system, the flow rate was 1.0 mL / min, and the column temperature was 30 °C. The phospholipid content was determined by thin layer chromatography, and the developing agent was a chloroform-methanol-water system. The contents of carboxymethyl cellulose, carrageenan, and guar gum were determined by spectrophotometry, and the measurement wavelengths were 620 nm, 540 nm, and 490 nm respectively. Each sample was measured 3 times repeatedly, and the relative standard deviation of the measurement results was less than 2%. The initial formula component matrix was constructed based on the measurement results as follows: 。

[0098] When establishing the neural network model, the correlation between each evaluation index was first analyzed by Pearson correlation coefficient. The calculated correlation coefficient matrix was: 。

[0099] Based on the correlation coefficient matrix, an initial convolution kernel was constructed, and an equivalent convolution kernel was obtained through fast Fourier transform. The specific parameter settings of the neural network model are shown in Table 2.

[0100] Table 2 Parameter Settings of Neural Network Model

[0101] The samples were comprehensively measured for performance, and each index was measured 3 times repeatedly, and the average value was taken as the final result. Table 3 lists the performance measurement results of the initial formula.

[0102] Table 3 Performance Measurement Results of Initial Formula

[0103] An optimization objective function was constructed for each evaluation index. Taking the optimization of color uniformity as an example: 。

[0104] Among them, the target value of 8.5 points is determined based on market research and expert opinions. Similarly, the optimization objective functions for other indicators are constructed. The optimization process adopts a strategy combining unidirectional optimization and multi-objective optimization. Table 4 records some of the iterative results during the unidirectional optimization process.

[0105] Table 4 Iterative Records of the Unidirectional Optimization Process (Taking Color Uniformity as an Example)

[0106] As Figure 2 shown, it shows the changing trend of color uniformity with the number of iterations during the unidirectional optimization process of the compound emulsifier for bread. When performing multi-objective optimization, the Pareto optimization method is adopted, and the population size is set to 50 and the number of generations of evolution is set to 100. The key parameter settings during the optimization process are shown in Table 5.

[0107] Table 5 Parameter Settings for Multi-Objective Optimization

[0108] The final formula obtained through optimization is: monoglyceride 35 mass percentage, sucrose fatty acid ester 20 mass percentage, phospholipid 20 mass percentage, carboxymethyl cellulose 10 mass percentage, carrageenan 8 mass percentage, guar gum 7 mass percentage. Comprehensive performance tests are carried out on the final formula, and the results are shown in Table 6.

[0109] Table 6 Performance Test Results of the Optimized Formula

[0110] To verify the stability of the optimized formula, a 180-day storage test is carried out. Under the condition of 25 degrees Celsius, all indicators are measured every 30 days, and the results show that the change range of each indicator is within 5%, proving that the optimized formula has good stability. As Figure 3 shown, it is a radar chart comparison of various performance indicators of the compound emulsifier for bread before and after optimization, visually showing the optimization effect.

[0111] The traditional optimization of the compound emulsifier formula for bread mainly adopts the single-factor test method and the orthogonal test method. The single-factor test method examines the influence of the content of a single component on the performance under the condition that other factors are fixed. This method ignores the interaction between components. Although the orthogonal test method considers the interaction between components, it can only analyze linear interaction relationships and cannot characterize complex non-linear interactions. In addition, these traditional methods usually only focus on single or a few performance indicators and are difficult to achieve the comprehensive optimization of multiple indicators.

[0112] In Example 2 of the present invention, an improved neural network model is used for formula optimization, which has the following advantages: First, the non-linear relationship between components is automatically extracted by deep learning methods, overcoming the problem that traditional methods are difficult to characterize complex interactions; Second, a strategy combining unidirectional optimization and multi-objective optimization is adopted, which not only ensures the optimization effect of a single index but also achieves the comprehensive balance of multiple indexes; Finally, by introducing component compatibility evaluation and stability verification, the practicality of the optimization results is improved. The optimized formula not only significantly improves various performance indexes but also has good stability, laying a foundation for the industrial application of compound emulsifiers for bread.

[0113] A specific Example 3 of the present invention is provided below. The R & D center of a dairy enterprise conducts research on the formula optimization of ice cream stabilizers: First, through market research and production practice, the researchers found that ice cream is prone to problems such as large ice crystals and uneven texture during freezing storage and transportation, which are mainly related to the stabilizer formula. To solve this problem, the research team established a complete evaluation index system and conducted systematic formula optimization research. Table 7 lists the specific content of the evaluation index system.

[0114] Table 7 Evaluation Index System

[0115] The selection of the initial formula is based on literature research and the results of preliminary experiments. Sodium alginate, carrageenan, xanthan gum, guar gum, carboxymethyl cellulose, and pectin are selected as the main components. The composition of the initial formula is: 25 mass percentages of sodium alginate, 20 mass percentages of carrageenan, 15 mass percentages of xanthan gum, 15 mass percentages of guar gum, 15 mass percentages of carboxymethyl cellulose, and 10 mass percentages of pectin. As Figure 4 shown, it shows the emulsion stability and the change of ice crystal size of the ice cream stabilizer during low-temperature storage.

[0116] Weigh each component by a precision balance and determine the component content by standard analysis methods to construct the initial formula component matrix: .

[0117] As Figure 5 shown, it is the correlation heat map of the formula component content and performance indexes, reflecting the influence degree of each component on different performance indexes. The parameter settings for establishing the neural network model are shown in Table 8 below: Table 8 Neural Network Model Parameters

[0118] The performance measurement results of the initial formula are shown in Table 9 below: Table 9 Initial Formula Performance Table

[0119] The single - objective optimization process uses an improved gradient descent method. Part of the record of the optimization process is shown in Table 10 below: Table 10 Record Table of the Optimization Process

[0120] The multi - objective optimization uses an improved NSGA - II algorithm, and the parameter settings are shown in Table 11 below: Table 11 Parameter Setting Table

[0121] The final optimized formula is: sodium alginate 28 mass percentage, carrageenan 18 mass percentage, xanthan gum 15 mass percentage, guar gum 15 mass percentage, carboxymethyl cellulose 14 mass percentage, pectin 10 mass percentage. The performance test results of the optimized formula are shown in Table 12 below: Table 12 Test Result Table

[0122] The stability verification is carried out at - 18 degrees Celsius. All indicators are measured every 30 days for 180 days. The test results show that during the entire verification period, the change range of each indicator is controlled within 5%, and the ice crystal particle size remains below 50 microns, proving that the optimized formula has good low - temperature stability.

[0123] Compared with the traditional orthogonal test method, the improved neural network model adopted in Example 3 of the present invention has the following advantages: First, by automatically extracting the non - linear interactions between components through deep learning, it avoids the limitations of manually designing interaction terms in traditional methods; Second, by adopting a multi - objective optimization strategy and considering multiple performance indicators simultaneously, it realizes the overall optimization of the formula; Finally, through the low - temperature stability verification, it ensures the reliability of the optimized formula in practical applications. The experimental results show that the optimized stabilizer formula not only significantly improves the texture and taste of ice cream, but also has excellent low - temperature stability, meeting the requirements of industrial production.

[0124] It should be noted that the detailed explanations of the variables involved in the present invention are shown in Table 13 and Table 14 below.

[0125] Table 13 Variable Explanation Table (First Part)

[0126] Table 14 Variable Explanation Table (Second Part)

[0127] As described above, it is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered by the protection scope of the present invention.

Claims

1. A method for optimizing a food additive formula, characterized in that: The method comprises the following steps: establishing a food additive evaluation index system, wherein the food additive evaluation index system includes sensory indexes, physical and chemical indexes and functional indexes; collecting food additive samples, determining the mass fraction of each component in the food additive samples, and establishing an initial formula component matrix; Establishing a neural network model, inputting the initial formula component matrix into the Lth layer of the neural network model, wherein the initial convolution kernel parameters of the Lth layer are determined based on the correlation coefficients of the various indicators in the food additive evaluation index system; obtaining an equivalent convolution kernel of the Lth layer of the neural network model according to the initial convolution kernel of the Lth layer of the neural network model, wherein the equivalent convolution kernel of the Lth layer of the neural network model is used to extract features of the interactions between the components; Constructing a single evaluation index optimization vector, and based on the single evaluation index optimization vector, performing unidirectional optimization on the initial formula component matrix respectively to obtain a unidirectional optimization formula for each indicator in the food additive evaluation index system; A multi-objective optimization model is constructed, the one-way optimization formula is used as a constraint, and the Pareto optimization method is used for comprehensive optimization.

2. The method for optimizing food additive formula according to claim 1, characterized in that: The sensory indicators include color uniformity, odor intensity, and solubility. The method for measuring the sensory indicators includes inviting 10 trained professional reviewers to evaluate the food additive samples using a 9-point scoring method. The color uniformity is measured by visual observation under a standard light source, the odor intensity is measured immediately after opening a sealed container, and the solubility is measured by the dissolution rate and solubility during stirring. The arithmetic mean of the reviewers' scores is taken as the final measured value.

3. The method for optimizing food additive formula according to claim 1, characterized in that: The physical and chemical indicators include pH value, viscosity, and water content. The method for measuring the physical and chemical indicators includes using a precision pH meter to measure the pH value, using a rotational viscometer to measure the viscosity, and using a vacuum drying method to measure the water content. Each food additive sample is measured repeatedly 3 times, and the average value is taken as the final measurement value.

4. The method for optimizing food additive formula according to claim 1, characterized in that: The functional indicators include antioxidant activity, emulsification stability, and water retention capacity. The antioxidant activity is determined by the 2,2-diphenyl-1-picrylhydrazyl free radical scavenging rate, the emulsification stability is determined by a centrifugation test, and the water retention capacity is determined by a centrifugal water retention rate.

5. The method for optimizing food additive formula according to claim 1, characterized in that: The neural network model adopts an improved residual network structure, and the neural network model includes an input layer, a feature extraction layer, a residual block layer, a fully connected layer, and an output layer. A distribution model is set before the feature extraction layer, and the distribution model includes a component balance equation, an interaction equation, a performance prediction equation, and a constraint condition equation.

6. The method for optimizing food additive formula according to claim 1, characterized in that: After obtaining the optimization result of the multi-objective optimization model, it also includes using a fuzzy comprehensive evaluation method to make a decision, calculating the component compatibility coefficient of the food additive sample in combination with the characteristic vector, constructing a third judgment function, substituting the component compatibility coefficient, the comprehensive performance score and the functional characteristic score into the third judgment function, and obtaining the total score of the optimized formula.

7. The method for optimizing food additive formula according to claim 1, characterized in that: After selecting the formula with the highest total score of the optimized formula as the optimal formula, the method also includes conducting stability verification on the optimal formula, storing it at 25 degrees Celsius for 180 days, measuring all indicators in the evaluation index system every 30 days, and fine-tuning the optimal formula using the neural network model based on the stability verification results to obtain the final optimized formula.

8. The method for optimizing food additive formula according to claim 1, characterized in that: The food additive is any one of a composite emulsifier for bread, a stabilizer for ice cream, a thickener for fruit juice drinks, or an antioxidant for vegetable oil. The initial formula of the composite emulsifier for bread is composed of 33 mass percent of monoglyceride, 22 mass percent of sucrose fatty acid ester, 18 mass percent of phospholipids, 12 mass percent of carboxymethyl cellulose, 8 mass percent of carrageenan, and 7 mass percent of guar gum; the initial formula of the stabilizer for ice cream is composed of 25 mass percent of sodium alginate, 20 mass percent of carrageenan, 15 mass percent of xanthan gum, 15 mass percent of guar gum, and 12 mass percent of carboxymethyl cellulose. The composition of the thickener for fruit juice beverage is 15 mass percent of xanthan gum, 10 mass percent of guar gum, 20 mass percent of pectin, 15 mass percent of carrageenan, 10 mass percent of carboxymethyl cellulose, and 5 mass percent of konjac gum; the composition of the antioxidant for vegetable oil is 35 mass percent of tocopherol, 25 mass percent of rosemary extract, 20 mass percent of tea polyphenols, 10 mass percent of citric acid, 6 mass percent of ascorbyl palmitate, and 4 mass percent of propyl gallate.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, they are used to execute the method for optimizing a food additive formula according to any one of claims 1 to 8.

10. A food additive formula optimization system, characterized in that: The system comprises the computer-readable storage medium as claimed in claim 9, wherein the system is any one of a computer, a server, and a single-chip microcomputer, the computer-readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing program instructions stored in the computer-readable storage medium.

Images