A plant meat quality prediction method and system based on iterative analysis of SPI raw materials

Abstract

The application provides a plant meat quality prediction method and system based on SPI raw material iterative analysis, and relates to the field of food science and technology, including the following steps: constructing a raw material characteristic index database; determining the quality characteristics of plant meat prepared in advance by soybean protein isolate raw material, and constructing a plant meat quality characteristic index database according to the quality characteristic determination results; based on the raw material characteristic index database and the plant meat quality characteristic index database, respectively constructing and training an elastic network regression model and a genetic algorithm-multivariate linear regression model; using the elastic network regression model and the genetic algorithm-multivariate linear regression model to preliminarily predict the quality of the plant meat, and combining a ternary group wisdom information evidence fusion method to fuse the preliminary prediction results to obtain the final plant meat quality prediction results. The application is helpful to more accurately grasp the quality of plant meat, and provides strong support for the production and optimization of plant meat.

Description

Technical Field

[0001] This invention relates to the field of food science and technology, and more specifically, to a method and system for predicting the quality of plant-based meat based on SPI (Survey Point Intake) raw material iterative analysis. Background Technology

[0002] The plant-based meat industry is booming globally, but it still faces numerous challenges in production technology, particularly in the development of high-end plant-based meat products, which remains in the experimental and exploratory stage. Currently, the quality stability of high-moisture textured protein for plant-based meat production is a significant issue. Its performance is easily affected by fluctuations in the properties of the protein raw materials, and traditional control methods, such as adjusting extrusion conditions or relying on exogenous additives, are largely experience-based and have limited effectiveness. Furthermore, even when stable textured protein is obtained, there are still many shortcomings in processing suitability, such as difficulty in flavoring due to its dense structure, difficulty in adding intramuscular fat, and poor structural stability after processing. These problems not only increase the complexity of the production process but also significantly increase production costs, greatly hindering the commercialization of high-end plant-based meat. Simultaneously, there is an urgent need to develop new raw materials to significantly improve the texture, flavor, and nutritional properties of current plant-based meat products. On the other hand, the emergence of numerous novel proteins, including those from microorganisms, the ocean, and other new resource proteins, raises questions about the impact of their raw material properties on extrusion fiberization and subsequent processing into plant-based meat quality and nutritional properties.

[0003] Currently, there is a lack of data models linking protein raw material properties and plant-based meat quality characteristics. The few existing data prediction models in the food sector typically employ simple algorithms such as support vector machines and backpropagation neural networks. Despite involving large datasets and various algorithms for model training, existing food sector models to date do not support self-optimization after training, which diminishes their practical applicability.

[0004] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0005] In view of this, the present invention provides a method and system for predicting the quality of plant-based meat based on SPI raw material iterative analysis, in order to solve the problems mentioned above.

[0006] To solve the above problems, the specific technical solution adopted by the present invention is as follows:

[0007] According to a first aspect of the present invention, a method for predicting the quality of plant-based meat based on SPI raw material iterative analysis is provided, comprising the following steps:

[0008] S1. Obtain samples of soybean protein isolate raw materials and conduct physicochemical property analysis. Based on the results of the physicochemical property analysis, construct a database of raw material characteristic indicators.

[0009] S2. The quality characteristics of plant-based meat prepared from soy protein isolate were determined, and a database of plant-based meat quality characteristic indicators was constructed based on the results of the quality characteristic determination.

[0010] S3. Based on the raw material characteristic index database and the plant meat quality characteristic index database, construct and train the elastic network regression model and the genetic algorithm-multiple linear regression model, respectively.

[0011] S4. The plant-based meat quality is initially predicted using an elastic network regression model and a genetic algorithm-multiple linear regression model. The preliminary prediction results are then fused using a ternary swarm intelligence information evidence fusion method to obtain the final plant-based meat quality prediction results.

[0012] Preferably, the construction and training of the elastic network regression model and the genetic algorithm-multiple linear regression model based on the raw material characteristic index database and the plant meat quality characteristic index database respectively includes the following steps:

[0013] S31. Preprocess the index data in the raw material characteristic index database and the plant meat quality characteristic index database respectively to obtain preprocessed raw material characteristic index data and plant meat quality characteristic index data.

[0014] S32. Using the multiple screening method, variables are screened on the pre-treated raw material characteristic index data and plant meat quality characteristic index data to obtain input variables and output variables.

[0015] S33. Use mini-batch partitioning to divide the input and output variables into datasets, and combine leave-one-out cross-validation to iteratively train the elastic network regression model and the genetic algorithm-multiple linear regression model.

[0016] S34. During the iterative training of the elastic network regression model and the genetic algorithm-multiple linear regression model, stochastic gradient descent is used to optimize the model, resulting in the optimized elastic network regression model and the genetic algorithm-multiple linear regression model.

[0017] Preferably, the preliminary prediction of plant-based meat quality using an elastic network regression model and a genetic algorithm-multiple linear regression model, and the fusion of the preliminary prediction results using a ternary swarm intelligence information evidence fusion method to obtain the final plant-based meat quality prediction result, includes the following steps:

[0018] S41. Obtain the data of soy protein isolate raw materials to be made into plant-based meat, and use the elastic network regression model and the genetic algorithm-multiple linear regression model to make multiple preliminary predictions on the quality of plant-based meat, and obtain multiple preliminary prediction results.

[0019] S42. Based on the multiple preliminary prediction results of the elastic network regression model and the genetic algorithm-multiple linear regression model, calculate the ternary swarm intelligence parameters of the cloud model, and perform two-dimensional characterization processing on the ternary swarm intelligence parameters to obtain the two-dimensional characterization results.

[0020] S43. Transform the two-dimensional representation results into evidence, and determine the conflict between the elastic network regression model and the genetic algorithm-multiple linear regression model through similarity calculation to obtain the conflict degree.

[0021] S44. Evidence fusion is performed using a quantum weighted average operator based on the degree of conflict, and the final prediction result is obtained through the DS evidence synthesis rule.

[0022] Preferably, the calculation of the ternary swarm intelligence parameters of the cloud model based on multiple preliminary prediction results from the elastic network regression model and the genetic algorithm-multiple linear regression model, and the two-dimensional representation processing of the ternary swarm intelligence parameters to obtain the two-dimensional representation result includes the following steps:

[0023] S421. Based on the multiple preliminary prediction results of the elastic network regression model and the genetic algorithm-multiple linear regression model, calculate the cloud model parameters of the elastic network regression model and the genetic algorithm-multiple linear regression model respectively. The cloud model parameters include the expected value, the entropy value and the hyperentropy value.

[0024] S422. Obtain the model validation data of the elastic network regression model and the genetic algorithm-multiple linear regression model during iterative training, and calculate the three-dimensional quality coefficients of the elastic network regression model and the genetic algorithm-multiple linear regression model, respectively. The three-dimensional quality coefficients include reliability, feature preference and robustness.

[0025] S423. Normalize the cloud model parameters and three-dimensional quality coefficients of the elastic network regression model and the genetic algorithm-multiple linear regression model, and combine the normalized cloud model parameters and three-dimensional quality coefficients to obtain the two-dimensional representation results of the elastic network regression model and the genetic algorithm-multiple linear regression model.

[0026] Preferably, the step of converting the two-dimensional representation results into evidence bodies and determining the conflict between the elastic network regression model and the genetic algorithm-multiple linear regression model through similarity calculation to obtain the conflict degree includes the following steps:

[0027] S431. Based on the preset plant-based meat quality evaluation criteria, and according to the two-dimensional representation results of the elastic network regression model and the genetic algorithm-multiple linear regression model, the basic probability allocation functions of the elastic network regression model and the genetic algorithm-multiple linear regression model are constructed respectively.

[0028] S432. Based on the cloud model forward generator algorithm, calculate the membership degree of the elastic network regression model and the genetic algorithm-multiple linear regression model to each quality level, and obtain the evidence bodies of the elastic network regression model and the genetic algorithm-multiple linear regression model respectively by correcting the basic probability assignment function based on the uncertainty probability.

[0029] S433. Based on the evidence body and three-dimensional quality coefficient of the two-dimensional representation of the elastic network regression model and the genetic algorithm-multiple linear regression model, calculate the cloud value similarity and quality similarity of the elastic network regression model and the genetic algorithm-multiple linear regression model respectively.

[0030] S434. The cloud value similarity and quality similarity are fused to obtain the comprehensive similarity, and the conflict degree between the elastic network regression model and the genetic algorithm-multiple linear regression model is calculated through the comprehensive similarity.

[0031] Preferably, the step of using a quantum weighted average operator to perform evidence fusion based on the degree of conflict and obtaining the final prediction result through the DS evidence synthesis rule includes the following steps:

[0032] S441. Based on the linear correlation method, construct a correlation model between the conflict degree and the weights in the quantum weighted average operator;

[0033] S442. Calculate the dynamic weight of each piece of evidence based on the linear correlation model, and normalize the dynamic weight based on the evidence support.

[0034] S443. Input the normalized dynamic weights into the quantum weighted average operator and combine them with the DS evidence synthesis rule to determine the final prediction result.

[0035] Preferably, the step of inputting the normalized dynamic weights into the quantum weighted average operator and determining the final prediction result in combination with the DS evidence synthesis rule includes the following steps:

[0036] S4431. Input the normalized dynamic weights into the quantum weighted average operator and calculate the probability amplitude and phase angle of each evidence body in the quantum state.

[0037] S4432. Based on the probability amplitude and phase angle of each piece of evidence in the quantum state, each piece of evidence is represented as a quantum state, and all quantum states are superimposed to obtain the superposition state of the quantum weighted average operator.

[0038] S4433. Perform quantum measurements on the superposition state of the quantum weighted average operator to obtain the preliminary fused evidence body;

[0039] S4434. The initially fused evidence is then fused a second time using the DS evidence synthesis rules to obtain the final prediction result.

[0040] According to a second aspect of the present invention, a plant-based meat quality prediction system based on SPI raw material iterative analysis is provided, the system comprising:

[0041] The raw material characteristic index acquisition module is used to acquire soybean protein isolate raw material samples and perform physicochemical property analysis. Based on the physicochemical property analysis results, a raw material characteristic index database is constructed.

[0042] The plant-based meat quality characteristic index acquisition module is used to determine the quality characteristics of plant-based meat prepared in advance using soy protein isolate raw materials, and to construct a plant-based meat quality characteristic index database based on the quality characteristic determination results.

[0043] The model building module is used to build and train elastic network regression models and genetic algorithm-multiple linear regression models based on raw material characteristic index databases and plant meat quality characteristic index databases, respectively.

[0044] The quality prediction module is used to make preliminary predictions of plant-based meat quality using an elastic network regression model and a genetic algorithm-multiple linear regression model. The preliminary prediction results are then fused using a ternary swarm intelligence information evidence fusion method to obtain the final plant-based meat quality prediction results.

[0045] According to a third aspect of the present invention, an electronic device is provided, the electronic device comprising: one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the prediction method described in any embodiment of the present invention.

[0046] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, wherein a computer program is stored therein, wherein the computer program, when executed, controls the device in which the computer-readable storage medium is located to perform the prediction method described in any embodiment of the present invention.

[0047] The beneficial effects of this invention are as follows:

[0048] 1. This invention constructs a raw material characteristic index database by obtaining soybean protein isolate raw material samples and analyzing their physicochemical properties, and constructs a quality characteristic index database by measuring the quality characteristics of plant-based meat. This provides a comprehensive and accurate data foundation for subsequent modeling. Based on these two databases, an elastic network regression model and a genetic algorithm-multiple linear regression model are constructed and trained. This fully utilizes the advantages of different models for preliminary prediction. The preliminary prediction results are then fused using a ternary swarm intelligence information evidence fusion method. By comprehensively considering multi-source information, the accuracy and reliability of plant-based meat quality prediction results are effectively improved, which helps to more accurately grasp the quality of plant-based meat and provides strong support for the production and optimization of plant-based meat.

[0049] 2. This invention utilizes an elastic network regression model and a genetic algorithm-multivariate linear regression model to perform multiple preliminary predictions, fully leveraging the advantages of different models to obtain multivariate prediction information; it calculates the ternary swarm intelligence parameters of the cloud model and performs two-dimensional representation, comprehensively considering the relationship between model characteristics and prediction results; it transforms these parameters into evidence and calculates the degree of conflict, accurately grasping the differences between models; based on the degree of conflict, it uses a quantum weighted average operator to fuse the evidence, combined with the DS evidence synthesis rule, which both enhances the fusion effect with the help of quantum computing and ensures the rationality of fusion using classical rules, ultimately obtaining accurate and reliable plant-based meat quality prediction results, providing strong guidance for plant-based meat production. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0051] Figure 1 This is a flowchart of a plant-based meat quality prediction method based on SPI raw material iterative analysis according to an embodiment of the present invention;

[0052] Figure 2 This is a flowchart of model optimization in a plant-based meat quality prediction method based on SPI raw material iterative analysis according to an embodiment of the present invention.

[0053] Figure 3 This is a flowchart of the final prediction result in a plant-based meat quality prediction method based on SPI raw material iterative analysis according to an embodiment of the present invention.

[0054] Figure 4 This is a principle block diagram of a plant-based meat quality prediction system based on SPI raw material iterative analysis according to an embodiment of the present invention;

[0055] Figure 5This is a hardware structure block diagram of the host device in a plant-based meat quality prediction method based on SPI raw material iterative analysis according to an embodiment of the present invention.

[0056] In the picture:

[0057] 1. Raw material characteristic index acquisition module; 2. Plant meat quality characteristic index acquisition module; 3. Model building module; 4. Quality prediction module. Detailed Implementation

[0058] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0059] The methods and embodiments provided in this application can be executed on a host device or a similar computing device. Taking running on a host device as an example, Figure 5 This is a hardware structure block diagram of a host device for a plant-based meat quality prediction method based on SPI raw material iterative analysis, according to an embodiment of this application. As shown in Figure 5, the host device may include one or more ( Figure 5 Only one is shown in the diagram. The processor (which may include, but is not limited to, a microprocessor (MCU) or programmable logic device (FPGA), etc.) and storage for storing data are also shown. The host device may further include transmission devices for communication functions and input / output devices. Those skilled in the art will understand that... Figure 5 The structure shown is for illustrative purposes only and does not limit the structure of the host device described above. For example, the host device may also include components that are larger than... Figure 5 The more or fewer components shown, or having the same Figure 5 The different configurations shown.

[0060] The memory can be used to store computer programs, such as application software programs and modules, like the computer program corresponding to the exception handling method in this embodiment. The processor executes various functional applications and data processing by running the computer program stored in the memory, thus implementing the above-described method. The memory may include high-speed random access memory (RAM) and non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the host device via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks (LANs), mobile communication networks, and combinations thereof.

[0061] Transmission devices are used to receive or send data over a network. Specific examples of the network described above may include a wireless network provided by a communication provider for the host device. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0062] According to an embodiment of the present invention, a method and system for predicting the quality of plant-based meat based on SPI raw material iterative analysis is provided.

[0063] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1-3 As shown, according to a first embodiment of the present invention, a method for predicting the quality of plant-based meat based on SPI raw material iterative analysis is provided, comprising the following steps:

[0064] S1. Obtain samples of soybean protein isolate raw materials and conduct physicochemical property analysis. Based on the results of the physicochemical property analysis, construct a database of raw material characteristic indicators.

[0065] It should be noted that this invention characterizes the influence of the basic physicochemical properties of SPI (soy protein isolate) on the quality of plant-based meat through approximate analysis, including water content, oil holding capacity, total sulfhydryl groups, free sulfhydryl groups, disulfide bonds, foaming properties, foam stability, solubility, emulsifying activity, emulsifying stability, hydrophobicity, turbidity, protein digestibility, and average particle size. Specifically, the physicochemical properties of SPI affect the quality of plant-based meat through intermolecular interactions (hydrogen bonds, hydrophobic interactions, disulfide crosslinking, etc.) or interfacial behavior (oil-water interface adsorption, foam / gel formation). The theoretical logic of each indicator is shown in Table 1.

[0066] Table 1. Theoretical Logic of Each Indicator

[0067]

[0068]

[0069] The detection methods for each indicator are as follows:

[0070] 1. Water-holding capacity (WHC) / Oil-holding capacity (OHC);

[0071] Place SPI (1.0 g, denoted as m0) and 10 mL of distilled water or oil in a pre-weighed centrifuge tube, incubate at 25°C for 30 min, then centrifuge at 2000 rpm for 30 min. Remove free water or oil, drain excess water or oil from the upper phase using filter paper for 10 min, then weigh the precipitate and record the weight as m1. The formulas for calculating water-holding capacity (WHC) / oil-holding capacity (OHC) are as follows:

[0072] WHC(%) = [(m1-m0)) / m0] × 100%;

[0073] OHC(%) = [(m1-m0)) / m0] × 100%.

[0074] 2. Thiol group (SHt / SHf) and disulfide bond (SS): Ellman reagent method;

[0075] Solution preparation specifically includes:

[0076] Tris-HCl buffer (0.1M, pH 8.0): Weigh 12.14g of Tris base and dissolve it in 800mL of ultrapure water. Adjust the pH to 8.0 with HCl and bring the volume to 1000mL.

[0077] (For total sulfhydryl determination) Tris-HCl buffer containing 8M urea: Add 480g of high-quality urea to the above buffer solution, stir to dissolve, and then bring the volume to 1000mL. Slight heating may be used to aid dissolution if necessary, but the temperature should not be too high (<30°C) to prevent urea decomposition. Prepare fresh before use.

[0078] Ellman's reagent (DTNB) solution (10mM): Weigh 39.6 mg of DTNB and dissolve it in 10 mL of Tris-HCl buffer (0.1 M, pH 8.0). Store protected from light and refrigerate at 4°C. It can be used for one week.

[0079] Reducing agent solution (1MDTT): Weigh 154 mg of DTT and dissolve it in 1 mL of ultrapure water. Prepare fresh before use, or aliquot and store at -20°C.

[0080] SPI sample solution: Accurately weigh a certain amount of SPI powder and prepare a solution with a concentration of approximately 1-3 mg / mL using the appropriate buffer solution.

[0081] For the determination of free thiol groups: Prepare with urea-free Tris-HCl buffer (0.1M, pH 8.0).

[0082] For measuring total thiol: Prepare with Tris-HCl buffer containing 8M urea.

[0083] To ensure complete protein dissolution, stir at low speed on a magnetic stirrer for at least 30 minutes, then centrifuge (10,000 rpm, 10 min, 4°C) to collect the supernatant as the test sample and record the actual protein concentration of the supernatant.

[0084] Measurement steps:

[0085] (1) Determination of free thiol groups: Take 1 mL of SPI sample solution (dissolved in urea-free buffer) using the sample tube. Take 1 mL of the corresponding buffer (urea-free Tris-HCl buffer) using the blank tube. Add 50 μL of 10 mM MDTNB solution to both the sample tube and the blank tube. Vortex to mix and react at room temperature in the dark for 30 minutes. Use a spectrophotometer at a wavelength of 412 nm, zeroing with the blank tube, to measure the absorbance value (A1) of the sample tube.

[0086] (2) Determination of total sulfhydryl groups: Take 1 mL of SPI sample solution (dissolved in buffer containing 8 M urea) and add 20 μL of 1 MdTT solution (final concentration approximately 20 mM). Vortex to mix and react at room temperature in the dark for 1 hour to ensure that disulfide bonds are fully reduced. Use a desalting column to load the entire reaction solution onto a desalting column (such as a GE Healthcare PD-10 column) that has been pre-equilibrated with Tris-HCl buffer containing 8 M urea. Elute with equilibration buffer, collect the first protein peak (pale yellow opalescent solution), and remove DTT and urea. Use the desalted or dialyzed protein solution as a new sample. Sample tube: Take 1 mL of the treated protein solution. Blank tube: Take 1 mL of buffer for elution / dialysis (Tris-HCl buffer containing 8 M urea). Add 50 μL of 10 mMdTNB solution to each, vortex to mix, and react in the dark for 30 minutes. At a wavelength of 412 nm, the absorbance value (A2) of the sample tube was measured using a blank tube as the zeroing point.

[0087] Data processing and computation:

[0088] Molar absorptivity: TNB 2- The molar absorptivity at 412 nm is ε = 14150 M. -1 cm -1 .

[0089] Formula for calculating thiol concentration: Thiol concentration (μmol / g) = (A × D × V × 10) 6) / (ε×C×I);

[0090] In the formula, A represents the measured absorbance value (A1 or A2); D represents the dilution factor (the sample is diluted after the addition of DTNB. In this example: 1 mL sample + 0.05 mL DTNB, D = 1.05 / 1 = 1.05); V represents the total volume of the reaction system (mL) (1.05 mL in this example); 10 6 ε represents the coefficient that converts molar number to micromoles (μmol); ε represents the molar absorptivity (14150 M). -1 cm -1 C represents the sample protein concentration (g / L) (Note: mg / mL must be converted to g / L, for example, 2 mg / mL = 2 g / L); I represents the cuvette optical path (cm) (1 cm).

[0091] Free thiol groups (μmol / g) = (A1 × 1.05 × 1.05 × 10) 6 ) / (14150×C×1)≈(A1×77.9) / C;

[0092] Total thiol groups (μmol / g) = (A² × 77.9) / C;

[0093] C represents the protein sample concentration (g / L) used in the DTNB reaction.

[0094] Disulfide bond content (μmol / g) = (total thiol groups - free thiol groups) / 2';

[0095] 3. Emulsifying activity (EAI) / Emulsifying stability (ESI);

[0096] 3.75 mL of soybean oil and 21.25 mL of 3% (w / v) protein solution were added and homogenized at 10,000 rpm for 1 min. At 0 and 10 min after homogenization, a 200 μL sample of the emulsion was transferred from the bottom of the container into 1.8 mL of 0.1% sodium dodecyl sulfate (SDS), and its absorbance was measured at 500 nm. Emulsion activity was measured immediately after emulsion formation (t = 0 min), and emulsion stability was estimated over a period of up to 10 min. Emulsion activity and emulsion stability were calculated using the following formula:

[0097] ;

[0098] ;

[0099] In the formula, A0 represents the absorbance of the emulsion immediately after homogenization and dilution, N represents the dilution factor (N=100), and C represents the weight of protein per volume (g / mL). This indicates the oil volume fraction of the emulsion, (A0-A...10 () represents the absorbance change between 0 and 10 min, where t = 10 min.

[0100] 4. Foaming properties (FC) and foaming stability (FS);

[0101] The SPI sample (20 mL 3% w / v dispersed in distilled water) was homogenized at 13500 rpm for 1 min, and the mixture was immediately transferred to a graduated cylinder. The volumes were recorded before homogenization and at 0 and 30 min. FC and FS were calculated as follows:

[0102] ;

[0103] ;

[0104] Where V0 = volume of the solution after homogenization at 0 min; V1 = initial solution volume; V 30 = The volume of the solution after 30 minutes of homogenization.

[0105] 5. Solubility and turbidity;

[0106] Prepare a 5% w / v SPI emulsion, centrifuge the solution (13000×g, 10 min), and measure the protein content in the supernatant using a nitrogen analyzer. Solubility is expressed as the percentage of protein content in the supernatant relative to the total protein content of the initial solution.

[0107] ;

[0108] 6. Hydrophobic;

[0109] The surface hydrophobicity (H0) of different SPI samples was measured using 1-anilino-8-naphthalenesulfonic acid (ANS) as a hydrophobic fluorescent probe. 4 mL of protein solutions at different concentrations (0.2–1.0 mg / mL) were mixed with 20 µL (8 mM) of freshly prepared ANS solution. The mixture was vortexed and stored in the dark at room temperature. Fluorescence (FI) of each sample at 280 nm (excitation) and 350 nm emission was recorded using a fluorometer. H0 was calculated by linear regression of the initial slope with FI and protein concentration (mg / mL).

[0110] 7. Turbidity;

[0111] Disperse a 0.5% w / v SPI solution in buffer (PBS solution, pH 7.0) with magnetic stirring for 60 minutes. After centrifugation (10000×g, 10 minutes recommended), collect the supernatant and measure its absorbance (OD) at 600 nm using a 1 cm cuvette. 600 The OD value, obtained by zeroing with the appropriate buffer solution, is the indicator for measuring turbidity.

[0112] 8. Protein digestibility;

[0113] A multi-stage digestion experiment was conducted on plant-based meat protein using an in vitro simulated digestion system. The specific procedure was as follows: 1.00 g of sample was accurately weighed and subjected to simulated digestion in the oral cavity, stomach, and intestines sequentially. After each stage of digestion was completed, the mixture was centrifuged at 3000 rpm for 10 min. The precipitate was collected, freeze-dried, and the protein content of the original sample before digestion and the undigested residue at each stage was determined using the Kjeldahl method. The formula for calculating the in vitro protein digestibility is:

[0114]

[0115] In the formula: D represents protein digestibility, W1 and W2 represent the mass of the sample before digestion (g) and the mass of the undigested residue (g), respectively; V1 and V2 correspond to the protein content (mass fraction, %) of the sample before digestion and the undigested residue.

[0116] 9. Average particle size;

[0117] The SPI sample (1 g) was dispersed in 100 mL of distilled water until it was clear and transparent. The particle size distribution was then measured using an S3500 laser particle size analyzer. To avoid sample swelling, measurements were performed immediately after sample preparation.

[0118] S2. The quality characteristics of plant-based meat prepared from soy protein isolate were determined, and a database of plant-based meat quality characteristic indicators was constructed based on the results of the quality characteristic determination.

[0119] Specifically, the methods for determining the quality characteristics of plant-based meat include:

[0120] (1) Texture characteristics: The plant-based meat samples were fixed on the platform, and the texture properties of the samples were analyzed using a texture analyzer. Following the AOAC 991.43 standard, a TA.XTplus texture analyzer with a P / 36R probe (cylindrical, contact area 36 mm²) was used. The TPA mode was selected for testing. When the compression ratio exceeded 75%, the spherical probe caused a stress concentration effect (stress deviation >15%). The probe descent speed was 2.0 mm / s, the testing speed was 1.0 mm / s, the ascent speed was 2.0 mm / s, the compression ratio was 40%, and the interval between two compressions was 4.0 s. Hardness, elasticity, cohesiveness, adhesiveness, chewiness, and resilience were recorded to analyze the texture characteristics of the samples.

[0121] (2) Texture: The plant-based meat sample was cut into rectangles 10 mm long, 10 mm wide, and 5 mm high. Then, the sample was compressed to 75% of its original thickness in both the vertical and parallel directions at a speed of 1.0 mm / s. The longitudinal shear force (FL) and the transverse shear force (FC) were defined as the maximum forces required to cut the sample in the vertical and parallel directions, respectively. The ratio of the longitudinal shear force to the transverse shear force was the degree of fibrosis (FD), i.e., texture.

[0122] S3. Based on the raw material characteristic index database and the plant meat quality characteristic index database, construct and train the elastic network regression model and the genetic algorithm-multiple linear regression model, respectively.

[0123] It should be noted that the model construction process is as follows:

[0124] 1. Clearly define related variables (ensure comprehensive variable coverage);

[0125] Independent variables (SPI raw material properties): 14 physicochemical indicators, divided into 3 categories:

[0126] Functional properties: oil retention, foaming properties, foam stability, emulsifying activity, emulsifying stability;

[0127] Structural characteristics: total thiol groups, free thiol groups, disulfide bonds, hydrophobicity, average particle size;

[0128] Basic properties: water content, solubility, turbidity, and protein digestibility.

[0129] Dependent variable (plant-based meat quality characteristics): 7 textural indicators, focusing on key commercial quality:

[0130] Macro structure: organization (a core indicator that measures the feel of the fabric);

[0131] Mechanical properties: hardness (chewing texture), elasticity (resilience), cohesion (structural integrity), adhesiveness (processability), chewiness (overall taste), resilience (taste persistence).

[0132] 2. Pre-screening based on correlation (reducing redundant variables and improving model efficiency);

[0133] Using bicorrelation analysis and threshold screening, key associated variables are accurately identified. The steps are as follows:

[0134] 1) Pearson linear correlation analysis: Calculate the linear correlation coefficient r between each physicochemical index and texture index of SPI, and screen variables with |r|>0.5 (significant linear correlation), such as disulfide bond and degree of texture r=0.68, hydrophobicity and hardness r=0.62.

[0135] 2) Spearman rank correlation analysis: For non-linear associations (such as the association between oil holding capacity and adhesiveness, which increases first and then decreases), calculate the rank correlation coefficient ρ, and screen variables with |ρ|>0.45 (significant non-linear association), such as oil holding capacity and adhesiveness with ρ=0.51.

[0136] 3) Grey relational analysis (GRA, suitable for small samples);

[0137] Step 1: Determine the reference sequence (texture index, such as texture degree Y) and the comparison sequence (SPI physicochemical index X1-X). 14 );

[0138] Step 2, Dimensionless data (initialization: X') i (k)=X i (k) / X i (Eliminating the influence of dimensions)

[0139] Step 3, calculate the correlation coefficient:

[0140] ;

[0141] In the formula, X' represents the correlation coefficient. i Y(k) represents the value of the i-th comparison sequence at the k-th data point, and Y(k) represents the value of the reference sequence at the k-th data point. This represents the resolution coefficient.

[0142] Step 4: Calculate the correlation degree;

[0143] ;

[0144] Filtering correlation r i Variables with a value greater than 0.6 (key related variables).

[0145] 4) Final variable selection: Take the intersection of the above three analysis results to obtain 8-10 key SPI indicators (such as disulfide bonds, hydrophobicity, oil holding capacity, protein digestibility, etc.), which will be used as model input variables (redundant variables, such as foam stability r, will be removed). i =0.48, was removed).

[0146] 3. Construction of the dual-association model (elastic network regression model + genetic algorithm - multiple linear regression model);

[0147] (1) Elastic network regression association model Y (handling multivariate collinearity), the model form is:

[0148] ;

[0149] In the formula, m represents the number of key SPI indicators after screening (8-10), j represents the index value, β0 represents the intercept, and β j Let X represent the regression coefficient, ε be the random error (following N(0,σ²)), and X be the regression coefficient. j This represents the specific measured value of the j-th key physicochemical index of soy protein isolate (SPI) after variable screening, with L1+L2 regularization applied: .

[0150] Address collinearity among SPI indices (e.g., emulsifying activity versus emulsifying stability r=0.85).

[0151] Model training: Through grid search, α∈[0.001,0.1], l1_ratio∈[0,1], traverse the parameter combinations, and combine leave-one-out cross-validation to select the parameter combination with the smallest validation MSE (e.g., α=0.01, l1_ratio=0.6).

[0152] The model was trained using full data from 11 different SPIs to obtain the final regression coefficient β. j Clarify the contribution direction of each SPI index to texture (β) j Positive indicates promotion, negative indicates inhibition.

[0153] (2) GA-MLR correlation model (optimized linear regression coefficients);

[0154] The expression for the MLR base model Y is:

[0155] ;

[0156] In the formula, β0 represents the intercept, β j X represents the regression coefficient. j This represents the specific measured value of the j-th key physicochemical indicator of soy protein isolate (SPI) obtained after variable screening.

[0157] The GA optimization process is as follows:

[0158] 1) Population initialization: Randomly generate 50 sets of regression coefficients (β0, β1, ..., β...). m ), as the initial population;

[0159] 2) Fitness function: Fitness is 1 / MSE (the smaller the MSE, the higher the fitness).

[0160] 3) Genetic manipulation:

[0161] Selection: Roulette wheel selection, selecting 20 parent individuals based on fitness percentage;

[0162] Crossover: Single-point crossover, where the parent coefficients exchange some genes with a probability of 0.8 (e.g., β1 and β2 are exchanged).

[0163] Mutation: A coefficient (e.g., β) is randomly adjusted with a probability of 0.05. j →β j ×(1+0.1×rand(-1,1))).

[0164] 4) Iteration Termination: After 50 iterations, the individual with the highest fitness is selected as the optimal regression coefficient to obtain the GA-MLR association model.

[0165] 4. Model validation and optimization (ensuring the reliability of the association);

[0166] Validation metrics: Calculate the full sample of 11 SPI types:

[0167] Coefficient of determination R 2 (The closer to 1 the better, R should be...) 2 >0.8);

[0168] Root mean square error (RMSE) (the smaller the better, RMSE < 5% × mean of texture index).

[0169] Mean absolute error (MAE) (the smaller the better, MAE < 3% × mean of texture index).

[0170] As a preferred embodiment, the construction and training of the elastic network regression model and the genetic algorithm-multiple linear regression model based on the raw material characteristic index database and the plant meat quality characteristic index database respectively includes the following steps:

[0171] S31. Preprocess the index data in the raw material characteristic index database and the plant meat quality characteristic index database respectively to obtain preprocessed raw material characteristic index data and plant meat quality characteristic index data.

[0172] S32. Using the multiple screening method, variables are screened on the pre-treated raw material characteristic index data and plant meat quality characteristic index data to obtain input variables and output variables.

[0173] S33. Use mini-batch partitioning to divide the input and output variables into datasets, and combine leave-one-out cross-validation to iteratively train the elastic network regression model and the genetic algorithm-multiple linear regression model.

[0174] S34. During the iterative training of the elastic network regression model and the genetic algorithm-multiple linear regression model, stochastic gradient descent is used to optimize the model, resulting in the optimized elastic network regression model and the genetic algorithm-multiple linear regression model.

[0175] It should be noted that, considering the relatively small sample size of the 11 soy protein isolates (SPIs), a four-stage iterative learning process was designed: "data preprocessing - hierarchical initialization - iterative iteration - validation feedback". The specific steps are as follows:

[0176] 1. Data preprocessing stage (ensuring data reliability);

[0177] (1) Data cleaning: For 14 SPI physicochemical indicators (water content, oil holding capacity, total thiol, free thiol, disulfide bond, foaming property, foam stability, solubility, emulsifying activity, emulsifying stability, hydrophobicity, turbidity, protein digestibility, average particle size) and 7 plant meat texture indicators (texture, hardness, elasticity, cohesiveness, adhesiveness, chewiness, resilience), outliers were removed: The improved Z-score method was used (the threshold was adjusted in combination with the measurement error range of the food testing instrument. For example, if the particle size analyzer error is ±3%, the sample with an absolute Z-score value >2.5 for the indicator is judged as an outlier. The conventional Z-score threshold is ±3). Outliers were filled by "adjacent sample interpolation method" (to avoid insufficient data due to sample deletion).

[0178] Furthermore, this invention dynamically adjusts the threshold based on the sensitivity of the indicator:

[0179] For highly sensitive indicators (screened through grey relational analysis, such as disulfide bonds and hydrophobicity, with a correlation degree > 0.7), a "low threshold + soft threshold processing" approach is adopted (to retain more detailed data).

[0180] For low-sensitivity indicators (such as turbidity and foam stability, with a correlation degree <0.5), a "high threshold + hard threshold processing" method is used (to filter redundant noise).

[0181] (2) Data standardization: All indicators are processed using Min-Max standardization, and the formula is:

[0182] ;

[0183] In the formula, X represents the original data. max X min These are the maximum and minimum values ​​of the indicator, respectively, to ensure that the weights of indicators with different dimensions (such as hydrophobicity: 0-100, average particle size: 100-500nm) are balanced for model training.

[0184] 2. The iterative learning initialization phase (adapting to small sample characteristics) specifically includes the following steps;

[0185] (1) Model parameter initialization, including:

[0186] Elastic network regression: The initial range of the regularization parameter α is set to [0.001, 0.1] (determined through preliminary experiments to avoid underfitting due to excessively large α), and the initial value of the L1 / L2 weight ratio I1_ratio is set to 0.5;

[0187] Genetic Algorithm - Multiple Linear Regression (GA-MLR): Initial population size is set to 50 (to balance search efficiency and diversity), crossover probability is 0.8, mutation probability is 0.05 (to avoid premature convergence), and the upper limit of the number of iterations is 50.

[0188] (2) Data set partitioning: Leave-one-out cross-validation method is adopted (for the small sample scenario of 11 types of SPI, conventional random partitioning will lead to insufficient training set): In each iteration, 1 type of SPI data is used as the validation set, and the remaining 10 types are used as the training set. The cycle is repeated 11 times to ensure that each sample participates in the validation and improves the generalization of the model.

[0189] 3. Iterative training phase (core process): Taking single-leave-one-out cross-validation as an example, the iterative steps are as follows:

[0190] (1) Input training data: Input the physicochemical data (independent variables) of the 10 SPIs in the current batch and the corresponding plant flesh texture data (dependent variables) into the model;

[0191] (2) Forward prediction calculation: The predicted value of texture is calculated by elastic network / GA-MLR model, and the initial error (mean square error MSE) is obtained by comparing it with the actual value.

[0192] (3) Local parameter update: Adjust the model parameters based on the current training set gradient (Elastic Network updates regression coefficients through gradient descent, and GA-MLR updates regression coefficients in the population through selection / crossover / mutation);

[0193] (4) Iteration termination judgment: The batch of iterations shall be stopped if any of the following conditions are met: the MSE decreases by <0.001 for 5 consecutive iterations (error convergence); the number of iterations reaches 30 (avoid over-iteration);

[0194] (5) Validation set evaluation: Input the batch of validation set (1 type of SPI data) into the optimized model, calculate the validation MSE, and record the current model parameters and errors.

[0195] 4. Global Iterative Optimization Phase (Integrating Results from 11 Iterations): Collect the model parameters and validation MSE from 11 leave-one-out cross-validation iterations, and select the 3 sets of parameters with the smallest validation MSE as candidate parameter sets;

[0196] The candidate parameter set is weighted and fused (the weight is 1 / validation MSE, and the smaller the error, the larger the weight), to obtain the final iterative optimization model, ensuring the stability of the model across the entire sample range.

[0197] Furthermore, for 11 SPI small sample scenarios, this invention adopts an SGD (Stochastic Gradient Descent) optimization model of "parameter initialization - batch gradient calculation - adaptive update - convergence judgment", the specific steps of which are as follows:

[0198] 1. Initialize SGD parameters (adapting to food model characteristics).

[0199] Basic parameter settings, specifically including:

[0200] Initial learning rate η0: set to 0.01 (determined through preliminary experiments to avoid parameter oscillation due to excessively large η0, and slow convergence due to excessively small η0);

[0201] Momentum coefficient μ: set to 0.9 (to balance local optimization and global search, and avoid gradient vanishing);

[0202] Convergence threshold ε: set to 0.001 (convergence is determined if the change in MSE is less than this value);

[0203] Maximum number of iterations: set to 100 (to avoid infinite iterations).

[0204] Parameter gradient initialization: The initial gradients of the regression coefficients β of the elastic network and the population regression coefficients of GA-MLR are set to 0 (to ensure the accuracy of the first update direction).

[0205] 2. Training data batch partitioning (adapting to small samples);

[0206] Mini-batch partitioning is adopted: Since the sample size of 11 SPIs is small, 8 types are randomly selected each time as a mini-batch (to balance batch size and gradient stability; conventional mini-batch is suitable for large samples, but this scheme adjusts the batch size to adapt to small samples). The remaining 3 types are used as the gradient validation set, and a total of 2 mini-batch partitions are made (if the last batch has less than 8 types, samples from the previous batch are added to avoid data waste).

[0207] 3. SGD gradient calculation and parameter update, specifically including the following steps;

[0208] (1) Taking the optimization of regression coefficient β in elastic network regression as an example, the steps are as follows:

[0209] Calculate the loss function: Use MSE as the loss function L(β), the formula is:

[0210] ;

[0211] In the formula, y i Indicates the actual texture value, The predicted value is represented by α, which is the regularization parameter to ensure that the parameters do not diverge. n represents the total number of SPI samples trained in the current batch of models. Its function is as follows: the core of the loss function (MSE) is the average error of all samples. n, as the denominator, ensures that the loss value is not affected by the number of samples (e.g., the total error of 10 samples needs to be divided by 10, and 11 samples need to be divided by 11), making the loss values ​​of different batches comparable. i represents the index of the SPI sample currently participating in training. By traversing i, the error between the actual value and the predicted value of all training samples can be calculated, thus obtaining the overall loss (MSE), providing a basis for subsequent parameter updates. l_ratio represents the weight ratio of L1 regularization and L2 regularization, which can solve the collinearity problem between SPI indicators mentioned in the document (e.g., the correlation coefficient r=0.85 between emulsification activity and emulsification stability). L1 regularization selects core indicators, L2 regularization reduces the interference of collinearity on parameters, and l1_ratio controls the balance weight of the two. This is just an example. β j This represents the regression coefficient corresponding to the j-th key SPI physicochemical index.

[0212] Calculate the gradient: Take the partial derivative (gradient) of the loss function L(β) with respect to β. The formula is:

[0213] ;

[0214] In the formula, Let α represent the predicted value, X represent the regularization parameter, X represent the SPI physicochemical data matrix, and sign(β) represent the sign function of β (adapted to L1 regularization).

[0215] Parameter update: Introducing a momentum term to avoid local optima, the update formula is as follows:

[0216] ;

[0217] In the formula, v t Let v represent the momentum at time t. t-1 Let η represent the momentum at time t-1. t Let represent the learning rate at time t, initially set to 0.01, μ represent the momentum coefficient, and β represent the learning rate at time t. t β represents the regression coefficients at time t (the current iteration). Specifically, it represents the new regression coefficient values ​​corresponding to each key SPI physicochemical indicator in the elastic network regression model after the t-th iteration update, and is the final target value for this parameter update; t-1 The regression coefficients at time t-1 (the previous iteration) represent the final values ​​of the regression coefficients corresponding to the key SPI physicochemical indicators in the elastic network regression model before the model is updated for the tth iteration. These coefficients are the initial base values ​​for this parameter update.

[0218] Convergence criterion and termination: If the change in loss over three consecutive iterations |L t -L t-1 |<ε (ε=0.001), and verify the loss L val If the value steadily decreases, stop SGD optimization; if the number of iterations reaches 100 and still fails to converge, force a stop and output the current optimal β (to avoid excessive consumption of computational resources).

[0219] Furthermore, since different physicochemical indicators of SPI have different weights on texture (e.g., the influence of disulfide bonds on texture degree is 3 times that of turbidity), this invention uses grey relational analysis to determine the weights of each texture indicator (e.g., texture degree 0.3, hardness 0.25, elasticity 0.2, and others 0.25; this is just an example), and incorporates these weights when constructing the loss function.

[0220] ;

[0221] In the formula, w k L represents the weight of the k-th texture index. k Let MSE represent the k-th texture index.

[0222] The learning rate is dynamically adjusted based on the contribution of each parameter to the loss:

[0223]

[0224] In the formula, η0 represents the initial learning rate, and L weighted Let β represent the loss function. j Let represent the regression coefficient of the j-th SPI index, and γ represent the adjustment coefficient, set to 0.5. A small learning rate is used for parameters with high contribution (such as β corresponding to disulfide bonds) to avoid oscillations, and a large learning rate is used for parameters with low contribution (to accelerate convergence). This reduces parameter update imbalance by 40% (the gradient variance of each parameter decreases from 0.05 to 0.03). The model's prediction R for key texture indicators (organization degree) is [value missing]. 2 Increased to 0.88 (the standard SGD is 0.80).

[0225] Specifically, this invention combines the natural fluctuation characteristics of SPI raw materials (such as the ±5% fluctuation of free thiol groups in SPI during actual production) to design a gradient jump mechanism:

[0226] If the loss reduction rate is less than 0.1% for 5 consecutive iterations (indicating a local optimum), perform fluctuation simulation on the SPI physicochemical data of the current training set:

[0227] ;

[0228] In the formula, The natural fluctuation coefficient of the j-th index is represented (determined through industry data, such as free thiols 8=0.05, average particle size 6=0.03), and rand(-1,1) is a random number between [-1,1].

[0229] Using the fluctuated X j 'Recalculate the gradient to achieve directed jumps (rather than random jumps) and escape local optima. The technical effect is: the probability of the model getting stuck in local optima is reduced from 35% to 10%, the global optimum acquisition rate is increased to 90% (compared to 65% for conventional SGD), and the validation MSE is reduced by 20%-25%.'

[0230] Conventional SGD has strong local search capabilities but weak global search capabilities, while GA has strong global search capabilities but low local accuracy. This invention combines GA-SGD for hybrid optimization:

[0231] Phase 1 (GA Global Search): The regression coefficients are optimized using GA in the first 20 iterations: 10 sets of candidate parameters are selected by selection (retaining the 20% of parameters with the smallest loss), crossover (generating offspring parameters by single-point crossover), and mutation (randomly adjusting 1-2 parameters).

[0232] Phase 2 (SGD Local Optimization): For each set of candidate parameters, iterate SGD 30 times to calculate the local optimal loss;

[0233] Phase 3 (Parameter Fusion): Select the three sets of parameters with the smallest local optimum loss, and obtain the final parameters by weighted averaging (weight is 1 / loss).

[0234] The technical benefits are: a 50% improvement in optimization efficiency (reducing from 100 iterations in conventional SGD to 50 iterations), and an increase in model prediction accuracy R0. 2 It remains stable above 0.85 (the typical SGD fluctuates between 0.78 and 0.83).

[0235] S4. The plant-based meat quality is initially predicted using an elastic network regression model and a genetic algorithm-multiple linear regression model. The preliminary prediction results are then fused using a ternary swarm intelligence information evidence fusion method to obtain the final plant-based meat quality prediction results.

[0236] As a preferred embodiment, the preliminary prediction of plant-based meat quality using an elastic network regression model and a genetic algorithm-multiple linear regression model, and the fusion of the preliminary prediction results using a ternary swarm intelligence information evidence fusion method to obtain the final plant-based meat quality prediction result, includes the following steps:

[0237] S41. Obtain the data of soy protein isolate raw materials to be made into plant-based meat, and use the elastic network regression model and the genetic algorithm-multiple linear regression model to make multiple preliminary predictions on the quality of plant-based meat, and obtain multiple preliminary prediction results.

[0238] S42. Based on the multiple preliminary prediction results of the elastic network regression model and the genetic algorithm-multiple linear regression model, calculate the ternary swarm intelligence parameters of the cloud model, and perform two-dimensional characterization processing on the ternary swarm intelligence parameters to obtain the two-dimensional characterization results.

[0239] As a preferred embodiment, the calculation of the ternary swarm intelligence parameters of the cloud model based on multiple preliminary prediction results of the elastic network regression model and the genetic algorithm-multiple linear regression model, and the two-dimensional representation processing of the ternary swarm intelligence parameters to obtain the two-dimensional representation result includes the following steps:

[0240] S421. Based on the multiple preliminary prediction results of the elastic network regression model and the genetic algorithm-multiple linear regression model, calculate the cloud model parameters of the elastic network regression model and the genetic algorithm-multiple linear regression model respectively. The cloud model parameters include the expected value, the entropy value and the hyperentropy value.

[0241] It should be noted that the expected value reflects the central tendency of the prediction results. For each model, the arithmetic mean of multiple preliminary predictions is calculated as the expected value. The entropy value represents the degree of uncertainty in the prediction results. The entropy value can be approximately estimated by calculating the standard deviation of the prediction results. The hyperentropy value reflects the uncertainty of the entropy value and is usually estimated using empirical formulas or expert knowledge. A common method is to calculate the hyperentropy value by setting a proportionality coefficient related to the entropy value based on the magnitude of the entropy and the distribution characteristics of the data.

[0242] S422. Obtain the model validation data of the elastic network regression model and the genetic algorithm-multiple linear regression model during iterative training, and calculate the three-dimensional quality coefficients of the elastic network regression model and the genetic algorithm-multiple linear regression model, respectively. The three-dimensional quality coefficients include reliability, feature preference and robustness.

[0243] It's important to note that reliability reflects the stability and accuracy of the model's predictions. Reliability can be determined by calculating the variance of the model's prediction error on the validation set. Feature preference represents the model's dependence on different input features. Feature preference can be measured by calculating the sum of the absolute values ​​of the coefficients of each feature in the model. Robustness reflects the model's ability to resist noise and outliers in the input data. Robustness can be measured by adding a certain proportion of noise or outliers to the validation data and then calculating the change in the model's predictive performance.

[0244] S423. Normalize the cloud model parameters and three-dimensional quality coefficients of the elastic network regression model and the genetic algorithm-multiple linear regression model, and combine the normalized cloud model parameters and three-dimensional quality coefficients to obtain the two-dimensional representation results of the elastic network regression model and the genetic algorithm-multiple linear regression model.

[0245] The normalized cloud model parameters (expected value, entropy value, hyperentropy value) and three-dimensional quality coefficients (reliability, feature preference, robustness) are combined. Vector concatenation can be used to combine the normalized parameters and coefficients of the elastic network regression model into a vector; similarly, a vector can be formed for the genetic algorithm-multiple linear regression model. These two vectors can serve as two-dimensional representations of the elastic network regression model and the genetic algorithm-multiple linear regression model, respectively, for subsequent evidence fusion.

[0246] S43. Transform the two-dimensional representation results into evidence, and determine the conflict between the elastic network regression model and the genetic algorithm-multiple linear regression model through similarity calculation to obtain the conflict degree.

[0247] In a preferred embodiment, the step of converting the two-dimensional representation results into evidence and determining the conflict between the elastic network regression model and the genetic algorithm-multiple linear regression model through similarity calculation to obtain the conflict degree includes the following steps:

[0248] S431. Based on the preset plant-based meat quality evaluation criteria, and according to the two-dimensional representation results of the elastic network regression model and the genetic algorithm-multiple linear regression model, the basic probability allocation functions of the elastic network regression model and the genetic algorithm-multiple linear regression model are constructed respectively.

[0249] It should be noted that the evidence theory requires a clearly defined identification framework (i.e., all possible quality outcomes). Therefore, it is necessary to first classify the quality grades of plant-based meat according to industry or company standards, for example:

[0250] The identification framework Ω = {Ω1 (Excellent, texture ≥ 85%, hardness 30-40N), Ω2 (Good, 70% ≤ texture < 85%, hardness 40-50N), Ω3 (Medium, 55% ≤ texture < 70%, hardness 50-60N), Ω4 (Poor, texture < 55%, hardness > 60N)}; a center value needs to be determined for each grade as the benchmark for subsequent calculation of membership degree.

[0251] Furthermore, for the elastic network regression model, based on its two-dimensional representation results (including cloud model parameters and three-dimensional quality coefficients) and combined with the plant-based meat quality evaluation criteria, a base probability is assigned to each quality grade. The base probability reflects the reliability of the model's prediction belonging to that quality grade.

[0252] For example, when an elastic network regression model predicts the sensory quality of a plant-based meat sample, based on the analysis of its two-dimensional representation results, if the characteristics of the sample match the description of the excellent grade well, then the excellent grade is assigned a high base probability, such as 0.6; it matches the good grade to a certain extent, and is assigned 0.3; and it matches the medium and poor grades to a low extent, and is assigned 0.05 and 0.05 respectively.

[0253] Similarly, a basic probability allocation function is constructed for the genetic algorithm-multivariate linear regression model, and based on its two-dimensional representation results and quality evaluation criteria, the corresponding basic probability is assigned to each quality level.

[0254] S432. Based on the cloud model forward generator algorithm, calculate the membership degree of the elastic network regression model and the genetic algorithm-multiple linear regression model to each quality level, and obtain the evidence bodies of the elastic network regression model and the genetic algorithm-multiple linear regression model respectively by correcting the basic probability assignment function based on the uncertainty probability.

[0255] It should be noted that the cloud model forward generator algorithm can transform qualitative concepts into quantitative values. For each quality level, a certain number of cloud droplets are generated through the forward generator using the expectation, entropy, and hyperentropy parameters of the cloud model. By calculating the membership degree of each cloud droplet to that quality level, and then statistically averaging the membership degrees of all cloud droplets, the overall membership degree of the model to that quality level is obtained.

[0256] The process involves modifying the previously constructed basic probability assignment function based on the calculated membership degrees. For example, if the membership degree of the elastic network regression model for a certain quality level is low, it indicates that the model's prediction certainty for that level is not high. Therefore, the uncertainty probability is appropriately increased, and the basic probabilities for other quality levels are adjusted accordingly, so that the sum of all probabilities is 1. After membership degree calculation and basic probability assignment function modification, the basic probabilities, uncertainty probabilities, and relevant model information (such as model name, key parameters in the two-dimensional representation results, etc.) of the model for each quality level are combined to form the evidence body of the model. For example, the evidence body of the elastic network regression model includes its modified basic probabilities and uncertainty probabilities for each quality level, the model name, and the expected values ​​in the cloud model parameters.

[0257] S433. Based on the evidence body and three-dimensional quality coefficient of the two-dimensional representation of the elastic network regression model and the genetic algorithm-multiple linear regression model, calculate the cloud value similarity and quality similarity of the elastic network regression model and the genetic algorithm-multiple linear regression model respectively.

[0258] It should be noted that cloud value similarity is calculated using either the Euclidean distance method or the cosine similarity method. Taking the Euclidean distance method as an example, the Euclidean distance between the parameter vectors of the cloud models of the two models is calculated, and then the cloud value similarity is obtained through a certain transformation (such as taking the reciprocal or performing normalization).

[0259] Among them, the quality similarity is calculated using the correlation coefficient method or the Manhattan distance method. Taking the correlation coefficient method as an example, the correlation coefficient between the three-dimensional quality coefficients of the two models is calculated as the quality similarity.

[0260] S434. The cloud value similarity and quality similarity are fused to obtain the comprehensive similarity, and the conflict degree between the elastic network regression model and the genetic algorithm-multiple linear regression model is calculated through the comprehensive similarity.

[0261] Specifically, a weighted fusion method is used to combine cloud value similarity and quality similarity into a comprehensive similarity. The weights of cloud value similarity and quality similarity are determined based on the degree of influence of cloud model parameters and three-dimensional quality coefficients on the model prediction results.

[0262] S44. Evidence fusion is performed using a quantum weighted average operator based on the degree of conflict, and the final prediction result is obtained through the DS evidence synthesis rule.

[0263] As a preferred embodiment, the step of using a quantum weighted average operator to perform evidence fusion based on the degree of conflict and obtaining the final prediction result through the DS evidence synthesis rule includes the following steps:

[0264] S441. Based on the linear correlation method, construct a correlation model between the conflict degree and the weights in the quantum weighted average operator;

[0265] ;

[0266] Bel(B) = 1 - Bel(A);

[0267] In the formula, Bel(A) and Bel(B) represent the final weights of the elastic network regression model (denoted as Model A) and the genetic algorithm-multiple linear regression model (denoted as Model B) in the quantum weighted average operator, respectively. These are the core parameters used for evidence fusion, determining the weight ratio of the evidence from the two models during fusion. λ represents the weight offset adjustment coefficient, Bel(A) + Bel(B) = 1, and both are ∈ [0,1]. Bel0(A) and Bel0(B) represent the basic weights of the elastic network regression model and the genetic algorithm-multiple linear regression model, respectively, and K represents the degree of conflict between the two models.

[0268] S442. Calculate the dynamic weight of each piece of evidence based on the linear correlation model, and normalize the dynamic weight based on the evidence support.

[0269] S443. Input the normalized dynamic weights into the quantum weighted average operator and combine them with the DS evidence synthesis rule to determine the final prediction result.

[0270] In a preferred embodiment, the step of inputting the normalized dynamic weights into the quantum weighted average operator and determining the final prediction result in conjunction with the DS evidence synthesis rule includes the following steps:

[0271] S4431. Input the normalized dynamic weights into the quantum weighted average operator and calculate the probability amplitude and phase angle of each evidence body in the quantum state.

[0272] It should be noted that the normalized dynamic weights reflect the relative importance of different pieces of evidence in the fusion process. For each piece of evidence, its probability amplitude is related to the normalized dynamic weight. Generally, the square of the probability amplitude corresponds to a certain probability. Assuming the normalized dynamic weight of the evidence is w, in a simple quantum model, the probability amplitude a can be expressed as: The phase angle is used to reflect the interaction and potential interference effects between pieces of evidence. The phase angle is determined based on the similarity or other correlation information between the pieces of evidence. For example, the phase angle can be calculated based on the correlation coefficient between the pieces of evidence. If the correlation coefficient between two pieces of evidence is high, it indicates a strong correlation, and the phase angle may be set relatively close; conversely, if the correlation coefficient is low, the phase angle may differ significantly.

[0273] S4432. Based on the probability amplitude and phase angle of each piece of evidence in the quantum state, each piece of evidence is represented as a quantum state, and all quantum states are superimposed to obtain the superposition state of the quantum weighted average operator.

[0274] It should be noted that, based on the calculated probability amplitude α and phase angle θ, each piece of evidence can be represented as a quantum state. .in E represents evidence body E i The corresponding quantum state basis vectors are an abstract representation of the evidence body in quantum space. This representation combines the classical information of the evidence body (such as weights and features) with quantum properties (probability amplitude and phase angle), providing a foundation for subsequent quantum computing and fusion. By superimposing the quantum states corresponding to all evidence bodies, the superposition state of the quantum weighted average operator is obtained.

[0275] S4433. Perform quantum measurements on the superposition state of the quantum weighted average operator to obtain the preliminary fused evidence body;

[0276] It should be noted that, in the context of evidence fusion, measuring the superposition state of a quantum weighted average operator means obtaining a definite body of evidence information from the superposition state. This is achieved using a projective measurement method, selecting a complete set of projective operators {P}. j}, where Pj corresponds to different propositions or states. For superposition states... The probability of obtaining measurement result j by performing projection measurement is: Based on the measurement results, the corresponding evidence body is taken as the preliminarily fused evidence body E. initial .

[0277] S4434. The initially fused evidence is then fused a second time using the DS evidence synthesis rules to obtain the final prediction result.

[0278] Specifically, the DS evidence synthesis rule is a method for integrating information from multiple pieces of evidence. It synthesizes the basic probability assignment functions of different pieces of evidence to obtain a more comprehensive and reliable basic probability assignment function. If multiple pieces of evidence exist after preliminary fusion, a step-by-step synthesis approach is used. First, two pieces of evidence are synthesized to obtain an intermediate result. Then, this intermediate result is synthesized with a third piece of evidence, and so on, until all relevant pieces of evidence have been synthesized. The class with the highest probability corresponding to the final basic probability assignment function is the final prediction result. The final result is analyzed. If the highest probability value is significantly higher than the probability values ​​of other classes, the prediction result is relatively reliable. If the probability values ​​of multiple classes are close, further data or model analysis may be needed, or other methods may be considered to assist decision-making. Finally, the final prediction result is output.

[0279] It should be noted that the application value of this invention can be reflected in the following aspects:

[0280] Raw material screening: Companies can test key indicators of SPI such as disulfide bonds, hydrophobicity, and solubility, and then use the results to predict the quality of plant-based meat and quickly screen SPI raw materials suitable for high-end plant-based meat production (e.g., if a texture degree >2 is required, the disulfide bond content must be >50μmol / g; this is just an example).

[0281] Process optimization: If the predicted texture does not meet the standard (e.g., the hardness is too low), the SPI index that needs to be improved can be deduced through the correlation model (e.g., increase hydrophobicity) to guide the pretreatment of raw materials (e.g., appropriate heat treatment to improve SPI hydrophobicity).

[0282] Intelligent manufacturing: The correlation model is embedded into the detection system of the production line to collect SPI physicochemical data in real time, automatically predict the texture, realize closed-loop control of "raw materials-quality", and reduce production costs by 15%-20%.

[0283] like Figure 4 As shown, according to a second embodiment of the present invention, a plant-based meat quality prediction system based on SPI raw material iterative analysis is provided, the system comprising:

[0284] Module 1 for acquiring raw material characteristic indicators is used to acquire soybean protein isolate raw material samples and perform physicochemical property analysis. Based on the physicochemical property analysis results, a database of raw material characteristic indicators is constructed.

[0285] The plant-based meat quality characteristic index acquisition module 1 is used to determine the quality characteristics of plant-based meat prepared in advance using soy protein isolate raw materials, and to construct a plant-based meat quality characteristic index database based on the quality characteristic determination results.

[0286] Model building module 2 is used to build and train elastic network regression model and genetic algorithm-multiple linear regression model based on raw material characteristic index database and plant meat quality characteristic index database, respectively.

[0287] The quality prediction module 3 is used to make preliminary predictions of plant-based meat quality using an elastic network regression model and a genetic algorithm-multiple linear regression model. The preliminary prediction results are then fused using a ternary swarm intelligence information evidence fusion method to obtain the final plant-based meat quality prediction results.

[0288] According to a third embodiment of the present invention, an electronic device is provided, the electronic device comprising: one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the steps in any of the above method embodiments.

[0289] According to a fourth embodiment of the present invention, a computer-readable storage medium is provided, wherein a computer program is stored in the computer-readable storage medium, wherein the computer program, when running, controls the device where the computer-readable storage medium is located to perform the steps in any of the above method embodiments.

[0290] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.

[0291] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, optical storage, etc.) containing computer-usable program code.

[0292] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the quality of plant-based meat based on SPI (Split Injection) raw material iterative analysis, characterized in that, Includes the following steps: S1. Obtain samples of soybean protein isolate raw materials and conduct physicochemical property analysis. Based on the results of the physicochemical property analysis, construct a database of raw material characteristic indicators. S2. The quality characteristics of plant-based meat prepared from soy protein isolate were determined, and a database of plant-based meat quality characteristic indicators was constructed based on the results of the quality characteristic determination. S3. Based on the raw material characteristic index database and the plant meat quality characteristic index database, construct and train the elastic network regression model and the genetic algorithm-multiple linear regression model, respectively. S4. The plant-based meat quality is initially predicted using the elastic network regression model and the genetic algorithm-multiple linear regression model. The initial prediction results are then fused using the evidence fusion method of ternary swarm intelligence information to obtain the final plant-based meat quality prediction results. The process of using an elastic network regression model and a genetic algorithm-multiple linear regression model to make preliminary predictions of plant-based meat quality, and then fusing the preliminary prediction results with a ternary swarm intelligence information evidence fusion method to obtain the final plant-based meat quality prediction results, includes the following steps: S41. Obtain the data of soy protein isolate raw materials to be made into plant-based meat, and use the elastic network regression model and the genetic algorithm-multiple linear regression model to make multiple preliminary predictions on the quality of plant-based meat, and obtain multiple preliminary prediction results. S42. Based on the multiple preliminary prediction results of the elastic network regression model and the genetic algorithm-multiple linear regression model, calculate the ternary swarm intelligence parameters of the cloud model, and perform two-dimensional characterization processing on the ternary swarm intelligence parameters to obtain the two-dimensional characterization results. S43. Transform the two-dimensional representation results into evidence, and determine the conflict between the elastic network regression model and the genetic algorithm-multiple linear regression model through similarity calculation to obtain the conflict degree. S44. Evidence fusion is performed using a quantum weighted average operator based on the degree of conflict, and the final prediction result is obtained through the DS evidence synthesis rule. Based on the multiple preliminary prediction results of the elastic network regression model and the genetic algorithm-multiple linear regression model, the ternary swarm intelligence parameters of the cloud model are calculated, and the ternary swarm intelligence parameters are processed into a two-dimensional representation to obtain the two-dimensional representation result, including the following steps: S421. Based on the multiple preliminary prediction results of the elastic network regression model and the genetic algorithm-multiple linear regression model, calculate the cloud model parameters of the elastic network regression model and the genetic algorithm-multiple linear regression model respectively. The cloud model parameters include the expected value, the entropy value and the hyperentropy value. S422. Obtain the model validation data of the elastic network regression model and the genetic algorithm-multiple linear regression model during iterative training, and calculate the three-dimensional quality coefficients of the elastic network regression model and the genetic algorithm-multiple linear regression model, respectively. The three-dimensional quality coefficients include reliability, feature preference and robustness. S423. Normalize the cloud model parameters and three-dimensional quality coefficients of the elastic network regression model and the genetic algorithm-multiple linear regression model, and combine the normalized cloud model parameters and three-dimensional quality coefficients to obtain the two-dimensional representation results of the elastic network regression model and the genetic algorithm-multiple linear regression model.

2. The method for predicting the quality of plant-based meat based on SPI raw material iterative analysis according to claim 1, characterized in that, The construction and training of the elastic network regression model and the genetic algorithm-multiple linear regression model based on the raw material characteristic index database and the plant meat quality characteristic index database respectively includes the following steps: S31. Preprocess the index data in the raw material characteristic index database and the plant meat quality characteristic index database respectively to obtain preprocessed raw material characteristic index data and plant meat quality characteristic index data. S32. Using the multiple screening method, variables are screened on the pre-treated raw material characteristic index data and plant meat quality characteristic index data to obtain input variables and output variables. S33. Use mini-batch partitioning to divide the input and output variables into datasets, and combine leave-one-out cross-validation to iteratively train the elastic network regression model and the genetic algorithm-multiple linear regression model. S34. During the iterative training of the elastic network regression model and the genetic algorithm-multiple linear regression model, stochastic gradient descent is used to optimize the model, resulting in the optimized elastic network regression model and the genetic algorithm-multiple linear regression model.

3. The method for predicting the quality of plant-based meat based on SPI raw material iterative analysis according to claim 1, characterized in that, The process of transforming the two-dimensional representation results into evidence and determining the conflict degree between the elastic network regression model and the genetic algorithm-multiple linear regression model through similarity calculation includes the following steps: S431. Based on the preset plant-based meat quality evaluation criteria, and according to the two-dimensional representation results of the elastic network regression model and the genetic algorithm-multiple linear regression model, the basic probability allocation functions of the elastic network regression model and the genetic algorithm-multiple linear regression model are constructed respectively. S432. Based on the cloud model forward generator algorithm, calculate the membership degree of the elastic network regression model and the genetic algorithm-multiple linear regression model to each quality level, and obtain the evidence bodies of the elastic network regression model and the genetic algorithm-multiple linear regression model respectively by correcting the basic probability assignment function based on the uncertainty probability. S433. Based on the evidence body and three-dimensional quality coefficient of the two-dimensional representation of the elastic network regression model and the genetic algorithm-multiple linear regression model, calculate the cloud value similarity and quality similarity of the elastic network regression model and the genetic algorithm-multiple linear regression model respectively. S434. The cloud value similarity and quality similarity are fused to obtain the comprehensive similarity, and the conflict degree between the elastic network regression model and the genetic algorithm-multiple linear regression model is calculated through the comprehensive similarity.

4. The method for predicting the quality of plant-based meat based on SPI raw material iterative analysis according to claim 1, characterized in that, The process of fusing evidence based on the degree of conflict using a quantum weighted average operator and obtaining the final prediction result through the DS evidence synthesis rule includes the following steps: S441. Based on the linear correlation method, construct a correlation model between the conflict degree and the weights in the quantum weighted average operator; S442. Calculate the dynamic weight of each piece of evidence based on the linear correlation model, and normalize the dynamic weight based on the evidence support. S443. Input the normalized dynamic weights into the quantum weighted average operator and combine them with the DS evidence synthesis rule to determine the final prediction result.

5. The method for predicting the quality of plant-based meat based on SPI raw material iterative analysis according to claim 4, characterized in that, The process of inputting the normalized dynamic weights into the quantum weighted average operator and combining them with the DS evidence synthesis rule to determine the final prediction result includes the following steps: S4431. Input the normalized dynamic weights into the quantum weighted average operator and calculate the probability amplitude and phase angle of each evidence body in the quantum state. S4432. Based on the probability amplitude and phase angle of each piece of evidence in the quantum state, each piece of evidence is represented as a quantum state, and all quantum states are superimposed to obtain the superposition state of the quantum weighted average operator. S4433. Perform quantum measurements on the superposition state of the quantum weighted average operator to obtain the preliminary fused evidence body; S4434. The initially fused evidence is then fused a second time using the DS evidence synthesis rules to obtain the final prediction result.

6. A plant-based meat quality prediction system based on SPI raw material iterative analysis, used to implement the plant-based meat quality prediction method based on SPI raw material iterative analysis as described in any one of claims 1-5, characterized in that, The system includes: The raw material characteristic index acquisition module is used to acquire soybean protein isolate raw material samples and perform physicochemical property analysis. Based on the physicochemical property analysis results, a raw material characteristic index database is constructed. The plant-based meat quality characteristic index acquisition module is used to determine the quality characteristics of plant-based meat prepared in advance using soy protein isolate raw materials, and to construct a plant-based meat quality characteristic index database based on the quality characteristic determination results. The model building module is used to build and train elastic network regression models and genetic algorithm-multiple linear regression models based on raw material characteristic index databases and plant meat quality characteristic index databases, respectively. The quality prediction module is used to make preliminary predictions of plant-based meat quality using an elastic network regression model and a genetic algorithm-multiple linear regression model. The preliminary prediction results are then fused using a ternary swarm intelligence information evidence fusion method to obtain the final plant-based meat quality prediction results.

7. An electronic device, characterized in that, The electronic device includes: one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the prediction method according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform the prediction method according to any one of claims 1 to 5.

Images