Method and system for determining frozen quality of meat based on multi-scale information of ice crystals

Abstract

The application discloses a method and system for determining frozen quality of meat based on ice crystal multi-scale information, and belongs to the technical field of food freezing processing and quality control. The method obtains multi-scale information such as spatial distribution, time evolution and structure morphology of ice crystals in meat during the freezing process of meat through microscopic imaging and dynamic monitoring technology, integrates raw material components and freezing process parameters, and establishes an ice crystal multi-scale information database; meat quality indexes such as thawing loss rate and muscle fiber fracture index are determined, an ice crystal morphology-texture damage relationship model, a frozen quality score model and an ice crystal growth kinetics prediction model under different freezing conditions are constructed through correlation analysis, and the frozen quality grade of meat is output based on the model, and the optimal freezing process and the optimal frozen shelf life are determined. The application realizes accurate prediction of the frozen quality of meat from the perspective of ice crystal multi-scale information, and can dynamically control the freezing process parameters, thereby providing technical support for quality control and cold chain logistics management of meat processing.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of food freezing processing and quality control. More particularly, the present application relates to a method and system for determining the freezing quality of meat based on ice crystal multi-scale information. BACKGROUND

[0002] In the field of meat freezing processing, freezing quality evaluation and determination of optimal freezing process parameters are key technical links to ensure product quality. Traditional freezing quality evaluation mainly relies on sensory evaluation, single thawing loss rate determination, texture analysis and other methods, which have obvious limitations: first, sensory evaluation is greatly influenced by subjective experience and has poor repeatability; second, thawing loss rate and other indicators only reflect the final result of freezing quality degradation and cannot reveal the dynamic damage mechanism of ice crystal formation process on muscle tissue; third, existing technologies only extract single morphological parameters (such as size) of ice crystals, lack multi-scale integrated analysis of ice crystal temporal evolution characteristics (such as growth rate, nucleation rate) and spatial distribution characteristics (such as density, uniformity), and cannot accurately quantify the internal relationship between ice crystals and meat freezing quality. In terms of prediction of freezing quality and optimal freezing process parameters, existing models are mostly based on temperature, time and other macroscopic parameters, ignoring the key influence of ice crystal multi-scale information on meat quality: the temporal scale characteristics of ice crystals determine their growth rate and nucleation efficiency, the spatial scale characteristics reflect the uniformity of muscle tissue damage, and the structural scale characteristics depict the complexity of ice crystal morphology. The lack of these multi-scale information makes it difficult for traditional models to distinguish the differential effects of ice crystal evolution on freezing quality and final shelf life under different freezing processes, and the prediction results deviate from the actual results by more than 30%.

[0003] The solution to the above problems faces multiple technical challenges. First, the observation of ice crystal multi-scale information requires high-precision microscopic imaging technology, while the preparation process of frozen samples easily introduces ice crystal recrystallization and other artifacts, affecting the authenticity of the observation results. Second, the quantitative analysis of ice crystal multi-scale information involves multiple disciplines such as image processing, time series analysis, and morphology, and requires standardized extraction methods for ice crystal size, spatial distribution, growth dynamics, and structural complexity. Third, the ice crystal growth kinetics model needs to consider the coupling effects of thermodynamics, mass transfer and other multi-physical fields, and traditional models cannot meet the prediction accuracy requirements under complex conditions. In addition, the correlation mechanism between ice crystal multi-scale information and meat quality damage involves microscopic processes such as biological macromolecular conformation change and cell structure damage, and its quantitative research needs interdisciplinary theoretical support.

[0004] Limitations of the prior art urgently require the development of a frozen meat quality prediction method and system based on ice crystal multiscale information to break through the experience dependence of traditional methods and realize the precise transformation from micro-ice crystal regulation to macro-quality control. The core of this technical path lies in establishing the quantitative relationship between ice crystal multiscale information and meat quality damage, constructing a prediction model that integrates ice crystal growth kinetics and machine learning, and thus realizing the dynamic optimization of the freezing process and the precise prediction of frozen quality, optimal freezing process and optimal frozen shelf life. SUMMARY

[0005] An object of the present application is to provide a method and system for predicting the frozen quality of meat based on ice crystal multiscale information, overcoming the defects in the prior art that the evaluation of frozen meat quality relies on a single indicator, the utilization of ice crystal information is insufficient, and the prediction accuracy of the optimal frozen shelf life is low.

[0006] In order to achieve these objects and other advantages of the present application, according to one aspect of the present application, a method for determining the frozen quality of meat based on ice crystal multiscale information is provided, comprising the following steps:

[0007] Step one, obtain ice crystal multiscale information (morphology, spatial distribution, equivalent diameter, roundness, fractal dimension, etc.) of frozen livestock and poultry meat samples at different stages and under different conditions through microscopic imaging technology, integrate frozen process parameters, meat raw material component data and ice crystal multiscale information, and establish an ice crystal multiscale information database;

[0008] Step two, detect the quality indicators of the frozen meat samples after thawing, correlate them with the corresponding ice crystal multiscale information, and construct an ice crystal multiscale information-texture damage relationship model;

[0009] Step three, score and grade the frozen quality of meat according to the integrated data in step one and the ice crystal multiscale information-texture damage relationship model constructed in step two;

[0010] Step four, based on the ice crystal multiscale information obtained in step one and the ice crystal multiscale information-texture damage relationship model obtained in step two, establish an ice crystal growth kinetics prediction model, optimize the freezing time and establish a frozen optimal quality prediction model, and the frozen optimal quality prediction model determines the optimal freezing process and optimal frozen shelf life corresponding to the optimal frozen quality through the frozen quality score of step three.

[0011] Preferably, the ice crystal multiscale information includes spatial scale characteristics, time scale characteristics and structural scale characteristics; wherein the spatial scale characteristics include ice crystal equivalent diameter, distribution density and regional aggregation degree; the time scale characteristics include ice crystal growth rate and nucleation rate; the structural scale characteristics include ice crystal roundness, stretchability and fractal dimension.

[0012] Preferably, the ice crystal multi-scale information database in step one is established by the following process:

[0013] Select different types of meat samples, and determine the original component data, including water content, protein content, and fat content.

[0014] Use a scanning electron microscope to perform dyeing and freeze section imaging on the frozen meat samples. Collect ice crystal images during the initial nucleation period, rapid growth period, and stable period of the samples during freezing, covering a temperature gradient of -5℃ to -80℃ and a freezing rate of 0.1℃ / min to 100℃ / min. The thickness of the freeze section is controlled to be 10-20μm.

[0015] Extract the spatial scale, time scale, and structural scale characteristics of the ice crystals, including equivalent diameter, ice crystal growth rate, ice crystal roundness, ice crystal stretch, and ice crystal fractal dimension, using image processing algorithms. Establish a three-dimensional ice crystal multi-scale information database based on freezing process parameters, raw material components, and freezing time.

[0016] Preferably, the specific method for constructing the ice crystal multi-scale information-texture damage relationship model in step two is as follows: measure the thawing loss rate and muscle fiber breakage index of the thawed frozen meat; analyze the correlation between the ice crystal multi-scale information (equivalent diameter, ice crystal growth rate, ice crystal roundness, ice crystal stretch, and ice crystal fractal dimension) and the texture damage indicators using Pearson or Spearman correlation coefficients; based on the correlation analysis results and data characteristics, select multiple linear regression, random forest, or support vector machine algorithms to construct the ice crystal multi-scale information-texture damage relationship model.

[0017] Preferably, the specific method for meat freezing quality scoring in step three is as follows:

[0018] Non-dimensionalize the original component data of meat, including water content, protein content, and fat content, the ice crystal multi-scale information (equivalent diameter, ice crystal growth rate, ice crystal roundness, ice crystal stretch, and ice crystal fractal dimension), and the freezing temperature, freezing rate, and freezing time of the freezing process, a total of 11 parameters. Based on the ice crystal multi-scale information-texture damage relationship model established in step two, select the ice crystal multi-scale information features that are significantly correlated with the thawing loss rate and muscle fiber breakage index through correlation analysis.

[0019] Determine the weight coefficients of each parameter using the analytic hierarchy process, and establish the evaluation model using multiple linear regression. The freezing quality score is calculated according to the following formula:

[0020] ;

[0021] where w i is the weight coefficient of each parameter, and x iFor the standardized parameter value, b is the bias term, S is the frozen quality score, the range of S is 0-100 points, and 80≤S≤100 is excellent, 60≤S<80 is good, 40≤S<60 is general, and 0≤S<40 is poor.

[0022] Preferably, the method for establishing the ice crystal growth kinetics prediction model in step four is:

[0023] The freezing temperature, freezing rate and temperature-time curve of the meat sample are collected in real time, the characteristic parameters (such as the starting temperature, ending temperature, duration or platform slope of the phase change platform period) representing the phase change process are extracted based on the temperature-time curve, the freezing time and the original component data of the frozen meat sample are combined with the initial ice crystal multi-scale information, the classical nucleation theory, the KGT model and the Langer model are fused, and a compound kinetics model is constructed; then, an ice crystal growth prediction model is constructed by combining the LSTM-XGBoost hybrid machine learning algorithm, and the dynamic evolution result of the ice crystal multi-scale information is output.

[0024] Preferably, the compound kinetics model comprises:

[0025] Nucleation stage: based on the classical nucleation theory, the critical radius of ice crystal r* = 2σT m / (ΔHΔT) is calculated, wherein σ is the interface energy, the value range is 0.02-0.05 J / m 2 ; ΔH is the latent heat of phase change, J / m 3 ; ΔT is the supercooling degree; T m is the freezing point of the meat sample;

[0026] Growth stage: KGT model is used to describe the ice crystal growth rate , wherein Δμ is the driving force, the value range is 10 3 ~10 5 J / mol; β, γ are fitting parameters;

[0027] Competitive growth: Langer model is introduced to analyze the ice crystal tip growth speed , wherein D is the diffusion coefficient, the value range is 10 -9 ~10 -8 m 2 / s; ΔT0 is the characteristic supercooling degree.

[0028] Preferably, the machine learning algorithm process is:

[0029] The following parameters are used as input features: ice crystal critical radius, ice crystal growth rate, freezing temperature, freezing rate, characteristic parameters representing the phase transition process, freezing time, original component data of frozen meat including moisture content, protein content, fat content, and initial ice crystal multi-scale information obtained by microscopic imaging technology including ice crystal equivalent diameter, ice crystal growth rate, ice crystal roundness, ice crystal stretchability, and ice crystal fractal dimension;

[0030] The LSTM neural network is used to process time series data, and the ice crystal multi-scale information prediction value is output;

[0031] The XGBoost algorithm is used to establish a multi-factor correlation model, and the ice crystal multi-scale information prediction value is output;

[0032] In the model training stage, the initial weight ratio w1 and w2 of the LSTM and XGBoost models are determined by the Bayesian optimization algorithm; in the real-time prediction stage, based on the feedback of the prediction error in the sliding window, the incremental learning or periodic weight fine-tuning strategy is adopted to dynamically update w1 and w2;

[0033] The final prediction value of the ice crystal multi-scale information = w1 x LSTM output value + w2 x XGBoost output value.

[0034] Preferably, the determination method of the optimal freezing process and the optimal frozen shelf life in step four is:

[0035] Step a, obtain the original component data of the meat sample and the initial ice crystal multi-scale information from the ice crystal multi-scale information database, and set the initial freezing temperature, rate and time;

[0036] Step b, according to the method of step three, the initial freezing parameters, original component data and initial ice crystal multi-scale information are dimensionless processed, the significant correlation ice crystal multi-scale information is screened, the weight coefficient is determined, and the initial freezing quality skin roller S0 is calculated;

[0037] Step c, use the real-time ice crystal growth kinetics prediction model to predict the change of ice crystal multi-scale information under the current freezing time:

[0038] If S0 is greater than or equal to 80, the current freezing condition and freezing time are maintained if S0 is greater than or equal to 80; if 60 is less than 80, the freezing rate is fine-tuned, and the freezing time is maintained; if 40 is less than 60, the freezing parameters are adjusted according to the preset rules: the freezing time is shortened preferentially, and the freezing temperature and rate are adjusted in a stepwise manner within a set range based on the feedback of the ice crystal growth prediction model, the score S is recalculated until S is greater than or equal to 60; if S0 is less than 40, the freezing process is reset when the quality is poor: in the database or the preset process library, the freezing process parameter combination that matches the current raw material components and has a historical score greater than or equal to 60 is selected as the new initial process, and the optimization process of steps b to c is repeated until the score is improved to the general and above range; the shelf life data of the meat under different freezing quality scores are collected, and a correlation model of the freezing quality score and the shelf life is established by using the Cox proportional hazards model, and the optimal freezing shelf life is determined by substituting the score obtained by dynamic optimization into the model.

[0039] The application also provides a system for determining the freezing quality of meat based on ice crystal multi-scale information, comprising:

[0040] An ice crystal multi-scale information acquisition module: an image processing unit and a dynamics analysis unit are integrated, which is used to receive the ice crystal images and dynamic data obtained by external microscopic imaging equipment and sensors, and extract the ice crystal spatial scale, time scale and structure scale information;

[0041] A three-dimensional database module: storing raw material component data, freezing process parameters and ice crystal multi-scale information, establishing an ice crystal multi-scale information database with multi-scale index;

[0042] A texture damage analysis module: connected to a thawing loss rate detection unit and a muscle fiber fracture index detection unit, performing Pearson / Spearman correlation analysis of ice crystal multi-scale information and meat quality indicators, and constructing and outputting an ice crystal multi-scale information-texture damage relationship model;

[0043] A quality score calculation engine: based on the analytic hierarchy process to determine the weight coefficient, and through a multiple linear regression model, the three-dimensional database data and the texture damage relationship model are processed in a dimensionless manner, the freezing quality score is calculated in real time, and the quality grade is divided;

[0044] A freezing process and shelf life optimization system:

[0045] An ice crystal growth prediction sub-module: a composite dynamics algorithm integrating nucleation theory, KGT model and Langer model is used to predict the evolution of ice crystal multi-scale information in real time combined with LSTM-XGBoost machine learning;

[0046] A process parameter control sub-module: dynamically adjusting the freezing temperature, rate and time according to the quality score;

[0047] Shelf life prediction submodule: map the frozen quality score S to the optimal frozen shelf life through the Cox proportional hazards model;

[0048] Wherein, the Cox proportional hazards model is defined as: h(t|S)=h0(t)·exp(βS); wherein h(t|S) is the risk function under the damage score S, h0(t) is the baseline risk function, and β is the regression coefficient of the score S.

[0049] The present application at least includes the following beneficial effects:

[0050] 1. Multi-scale information integration: Break through the limitation of traditional single ice crystal morphology parameter, integrate spatial scale, time scale, structure scale information, more comprehensively reveal the correlation mechanism of ice crystal and meat frozen quality;

[0051] 2. Accurate quality evaluation: Establish a quantitative frozen quality score system to realize the grading determination of meat frozen quality, and solve the subjective problem of traditional sensory evaluation;

[0052] 3. Dynamic prediction and regulation: Fusion of physical kinetics model and machine learning algorithm, realize real-time prediction of ice crystal multi-scale information, and dynamically optimize frozen process parameters according to quality score, improve the prediction accuracy of optimal frozen quality and time length;

[0053] 4. Industrial adaptability: The established ice crystal multi-scale information database covers multiple varieties of meat, wide temperature range and variable speed conditions, which can provide data support for process optimization of meat freezing processing.

[0054] Other advantages, objects and features of the present application will be partly embodied by the following description, and partly understood by those skilled in the art through research and practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 The flow chart of the method for determining the frozen quality of meat based on ice crystal multi-scale information of the present application.

[0056] Figure 2 The contrast chart of the appearance state of chicken breast under different frozen temperatures in the example.

[0057] Figure 3 The contrast chart of the microstructure of muscle fibers under different frozen temperatures in the example.

[0058] Figure 4 The contrast chart of ice crystal imaging under different frozen temperatures in the example.

[0059] Figure 5 The contrast chart of ice crystal equivalent diameter analysis under different frozen temperatures in the example Figure 1 .

[0060] Figure 6 Comparison of ice crystal equivalent diameter analysis at different freezing temperatures in the examples Figure 2 .

[0061] Figure 7 Comparison of ice crystal equivalent diameter analysis at different freezing temperatures in the examples Figure 3 .

[0062] Figure 8 Comparison of ice crystal equivalent diameter analysis at different freezing temperatures in the examples Figure 4 .

[0063] Figure 9 Comparison of ice crystal relative area analysis at different freezing temperatures in the examples.

[0064] Figure 10 Comparison of characteristic parameters such as meat freezing rate and freezing time at different freezing temperatures in the examples. DETAILED DESCRIPTION

[0065] The application will be further described in conjunction with specific embodiments to enable those skilled in the art to implement the application according to the description.

[0066] It should be understood that the terms such as "have", "contain" and "include" used herein do not exclude the presence or addition of one or more other elements or combinations thereof.

[0067] It should be noted that the experimental methods in the following embodiments are conventional methods unless otherwise specified, and the reagents and materials can be obtained from commercial channels unless otherwise specified.

[0068] The application provides a method for determining the freezing quality of meat based on ice crystal multi-scale information, comprising the following steps:

[0069] Step one, obtain ice crystal morphology images of frozen meat samples at different periods and under different conditions through microscopic imaging and dynamic monitoring technology, extract ice crystal multi-scale information therefrom, integrate frozen process parameters, raw material component data of meat samples and ice crystal multi-scale information, and establish an ice crystal multi-scale information database;

[0070] Step two, detect meat quality indicators after thawing of the frozen meat sample, correlate them with corresponding ice crystal multi-scale information, and construct an ice crystal multi-scale information-texture damage relationship model;

[0071] Step three, score and divide the quality grade of the meat freezing quality according to the integrated data in step one and the ice crystal multi-scale information-texture damage relationship model constructed in step two;

[0072] Step four, based on the ice crystal multi-scale information obtained in step one, the ice crystal multi-scale information-texture damage relationship model obtained in step two, the ice crystal growth kinetics prediction model is established, the freezing time is optimized and the frozen optimal quality prediction model is established, the frozen optimal quality prediction model determines the optimal freezing process and the optimal freezing shelf life corresponding to the optimal freezing quality through the freezing quality score of step three.

[0073] In the technical solution, the device for obtaining ice crystal images can select a scanning electron microscope or a low-temperature confocal microscope. Before slicing the frozen meat sample, a frozen slicer is used, the blade of which is a steel blade, and the sample table is made of metal. During the slicing of the frozen sample, the sample is placed in a frozen embedding agent, which can be a water-soluble glue with specific components. The thickness of the slice is controlled between 10 microns and 20 microns. Before slicing, the sample needs to be frozen in a programmed cooling box, and the cooling source of the cooling box can be selected from mechanical refrigeration or liquid nitrogen. When collecting images, the sample chamber of the microscope is in a vacuum or low-temperature nitrogen protection environment. The imaging system is connected to a computer, and the computer is installed with an image acquisition card and image analysis software. The image processing algorithm can be realized through software, which is used to extract the equivalent diameter, roundness and other spatial and structural parameters of ice crystals from the original images. The extracted ice crystal characteristic data, the corresponding frozen process parameters (such as temperature from -5℃ to -80℃, freezing rate from 0.1℃ / min to 100℃ / min) and the component data of the meat raw material (moisture, protein and fat content, which can be measured by near-infrared spectrometer or conventional chemical method) are stored in a database constructed by a relational database management system, and the database server can be installed on a local computer or an internal network server in the laboratory.

[0074] The equipment for detecting the quality of thawed meat includes a centrifuge and a texture analyzer. The method for measuring the thawing loss rate is to dry the surface moisture of the thawed frozen meat sample with filter paper and weigh it, and calculate the weight loss percentage. The determination of the muscle fiber breakage index can use the texture analyzer to perform shear force test, or another optical microscope to observe the section and count the proportion of broken muscle fibers. The above detection process is completed in the physical and chemical analysis area of the laboratory. The obtained quality index data and the corresponding ice crystal multi-scale information in the database are imported into the statistical analysis software. First, the correlation between the ice crystal equivalent diameter, roundness and other characteristics and the thawing loss rate, muscle fiber breakage index is analyzed through the Pearson correlation coefficient calculation module in the software. Then, based on the correlation analysis result, the modeling algorithm is selected. For example, a random forest regression algorithm can be run using the statistical analysis software or another computer installed with Python programming environment and Scikit-learn machine learning library. The algorithm takes the selected ice crystal multi-scale information as input features and the thawing loss rate or muscle fiber breakage index as output target for training, and finally generates an ice crystal multi-scale information-texture damage relationship model file that can predict the degree of texture damage.

[0075] The computer calls the raw material component data, ice crystal multi-scale information and freezing process parameters of the sample to be scored from the database, a total of 11 items. First, the 11 parameters are standardized (dimensionless) by program code, for example, using the Z-score standardization method. The determination of the weight uses the analytic hierarchy process, which can be realized by running MATLAB software or AHPy library in Python on the computer, and the weight coefficients are calculated by experts scoring the two-by-two comparison of parameters. Subsequently, a multivariate linear regression algorithm is used to establish an evaluation model. The model multiplies the standardized parameter values by their respective weight coefficients and adds a bias term obtained by regression analysis, to finally calculate a frozen quality score S between 0 and 100. According to the size of S value, the system automatically divides it into four levels: 80≤S≤100 is excellent, 60≤S<80 is good, 40≤S<60 is general, and 0≤S<40 is poor. The score result can be displayed on the computer screen or transmitted to the downstream control system.

[0076] By systematically collecting the multi-scale information of ice crystal's spatial distribution, time evolution and structural morphology, and correlating with the characteristics of meat raw materials and freezing process parameters, a more comprehensive bottom layer data is provided for the analysis of frozen quality. Further, by establishing a quantitative relationship model between ice crystal information and measured texture damage indicators, it is possible to predict macro meat quality damage based on micro ice crystal state. Based on this, the standardized scoring system can objectively and quantitatively evaluate the meat quality under different freezing conditions, reducing the deviation of subjective judgment. Finally, combined with the ice crystal growth kinetics prediction model, this scheme can provide decision-making basis for different meats under a specific freezing process to find a freezing time that takes into account quality and efficiency, thereby realizing fine control of the freezing process.

[0077] In another technical solution, the ice crystal multi-scale information includes spatial scale features, time scale features and structural scale information; wherein the spatial scale features include ice crystal equivalent diameter, distribution density, regional aggregation degree; the time scale features include ice crystal growth rate, nucleation rate; the structural scale features include ice crystal roundness, stretchability, fractal dimension.

[0078] In the technical solution, the device for acquiring and analyzing spatial scale features mainly relies on an image processing system. The system can be installed with professional image analysis software. The software can call built-in or custom algorithm modules. The object material for processing is the microscopic image of frozen meat sample, which is obtained by imaging devices such as scanning electron microscope. The computer is usually placed on the data processing workstation in the laboratory, connected with the microscope through data line. Its working process is that the image processing software reads the original ice crystal gray or binary image, identifies the outline of each independent ice crystal through edge detection algorithm. Then, the software calculates the ice crystal projection area and perimeter based on the outline pixel points. The ice crystal equivalent diameter is obtained by equating the ice crystal area to the circular area and calculating the diameter. The distribution density is calculated by counting the number of ice crystals in unit image area. The regional aggregation degree is evaluated by spatial statistical method (such as nearest neighbor index) to evaluate the aggregation degree of ice crystal distribution.

[0079] The device for analyzing time scale characteristics is also based on an image processing system, but needs to process time series images. The microscope control system with time lapse shooting function and the corresponding time sequence analysis software module can be configured. The processed material is a set of continuous images taken at fixed time intervals, reflecting the growth process of ice crystals in the same field of view. These devices and modules are integrated in the image acquisition and analysis computer. The working process is that the analysis software first registers the images at different time points to ensure the consistency of the observation field. For ice crystal growth rate, the software calculates the rate of change of the size (such as equivalent diameter) of a specific ice crystal over time by tracking its size at different time points. For nucleation rate, the software identifies the number of newly appearing ice crystal nuclei by comparing images at adjacent time points, and calculates the rate of new nucleus generation per unit time, per unit sample volume or area. The parameter setting is based on the sampling frequency of the actual freezing process.

[0080] The device for analyzing structural scale characteristics relies on a morphological analysis module in the image processing software. This module can be integrated into the software of the same image analysis computer mentioned above. The processed material is a single or specific time point ice crystal binary image. The working process is that the software first identifies the contour of each ice crystal in the image. Ice crystal roundness is obtained by calculating the area and perimeter of the contour, and the value closer to 1 indicates that the shape is closer to a circle. Ice crystal elongation is obtained by calculating the aspect ratio of the minimum circumscribed rectangle of the ice crystal, which is used to describe the elongation of the shape. Ice crystal fractal dimension can be calculated using the box counting method, which is used to quantify the complexity and irregularity of the ice crystal contour. These calculations are automatically completed by the built-in geometric and morphological functions of the software.

[0081] By clearly defining the three dimensions of spatial scale, time scale and structural scale of ice crystal information and their specific characteristic parameters, a complete and operable feature system is provided for quantitative description of ice crystals. The spatial scale parameter can objectively represent the size and distribution state of the ice crystal, the time scale parameter can reflect the dynamic process of ice crystal formation and growth, and the structural scale parameter can reveal the geometric complexity of the ice crystal shape. The combination of information of the three scales can more comprehensively represent the characteristics of the ice crystal population formed in the freezing process than a single scale, thereby laying a solid data foundation for subsequent establishment of an accurate correlation model between ice crystal characteristics and meat texture damage. This multi-scale description method helps to better understand the different paths of the influence of different freezing conditions on the microstructure of meat.

[0082] In another technical solution, the establishment process of the ice crystal multi-scale information database in step one is:

[0083] Select different types of meat samples, and measure their original component data, including water content, protein content, and fat content;

[0084] The frozen meat sample is cut and imaged by scanning electron microscope. The ice crystal images of initial nucleation, rapid growth and stable stages are collected under the temperature gradient of -5℃ to -80℃ and the freezing rate of 0.1℃ / min to 100℃ / min. The thickness of the frozen section is controlled to be 10-20μm.

[0085] The ice crystal multi-scale information is extracted by the image processing algorithm and the kinetic analysis model. The three-dimensional ice crystal multi-scale information database is established according to the freezing process parameters, raw material components and freezing time.

[0086] In the technical scheme, the experimental object can be selected from commercially available pork longissimus dorsi, beef eye muscle or chicken breast. The equipment for determining the components can be a near-infrared spectrum rapid analyzer, or a conventional oven, a Kjeldahl nitrogen determination instrument and a Soxhlet extraction device. These devices are placed in the physical and chemical analysis area of the laboratory. The working process is as follows: first, a representative sample is taken from the meat sample to be tested. If a near-infrared spectrometer is used, the meat sample is cut into uniform slices or crushed and filled into a sample cup, which is placed on the sample table of the instrument for scanning. The built-in calibration model of the instrument directly outputs the percentage of water, protein and fat. If a conventional method is used, the water content is obtained by placing a known weight of meat sample in a 105℃ oven and drying to constant weight; the protein content is obtained by determining the total nitrogen by the Kjeldahl method and then converting it; and the fat content is obtained by Soxhlet extraction using anhydrous ether or petroleum ether as the solvent. The test results are recorded in the form of an electronic table.

[0087] The program cooling box is used to simulate and control the freezing process, the freezing sectioning machine is used to prepare the observation sample, and the scanning electron microscope is used for high-resolution imaging. The program cooling box and the freezing sectioning machine are usually placed in the sample preparation room, and the scanning electron microscope is placed in an independent electron microscope room with anti-vibration and clean requirements. The working process is as follows: first, the meat sample cut into small pieces (such as a 5mm cube) is placed into the program cooling box. According to the experimental design, the temperature gradient of the box is set, for example, from 4℃ to -30℃ at a rate of 5℃ / min, covering the above-mentioned temperature range and rate range. At the key stages of the freezing process (such as the initial nucleation stage, the rapid growth stage and the stable stage), the sample is taken out. The sample block taken out is immediately sectioned using the freezing sectioning machine, and the sectioning thickness is controlled to be 15 microns. Before sectioning, the sample can be embedded with a freezing embedding agent (such as a sodium carboxymethyl cellulose solution with a specific concentration) to provide support. The prepared frozen section is transferred to the sample table of the scanning electron microscope, and the sample table has been pre-cooled. In the vacuum chamber of the electron microscope, the sample is treated with gold or carbon spraying to enhance the electrical conductivity, and then observed and digital images are collected under a specific accelerating voltage (for example, 5kV) and magnification (for example, 2000 times).

[0088] The image processing software first pre-processes the collected ice crystal microscopic images, such as noise reduction, contrast enhancement and binarization, to separate the ice crystals from the background structure. Then, the image analysis algorithm library is called to automatically identify and label each independent ice crystal region. For each ice crystal region, the software calculates its basic geometric parameters such as area and perimeter, and further calculates spatial and structural scale characteristic values such as equivalent diameter, roundness and stretchiness. For time series images, additional time scale characteristics such as growth rate and nucleation rate are calculated. These calculated characteristic values, together with the corresponding freezing process parameters (temperature, rate), raw material component data (moisture, protein, fat content) and freezing time point of the image, are structuredly stored. The storage carrier can be a relational database (such as MySQL or SQLite), where each record represents an observed sample, and the fields contain all the above parameters, thus forming a three-dimensional ice crystal information database that can be queried and indexed by process, component and time.

[0089] Through standardized sample pre-treatment, freezing control and section imaging process, high-quality image data reflecting the state of ice crystals under different freezing conditions can be obtained, providing a reliable source for subsequent analysis. The internal component data of the raw material, the externally applied freezing process parameters and the micro-ice crystal multi-scale information data extracted from the image are systematically associated and structured into a database, breaking down the barriers between different sources of data. This three-dimensional database not only can save experimental data completely, but more importantly, provides an efficient data organization and management foundation for exploring the complex relationship between freezing process, raw material characteristics and ice crystal formation, making large-scale correlation statistical analysis and machine learning model training possible.

[0090] In another technical solution, the specific method for constructing the ice crystal multi-scale information-texture damage relationship model in step two is: measuring the thawing loss rate and muscle fiber breakage index of the thawed sample; analyzing the correlation between the ice crystal equivalent diameter, ice crystal growth rate, ice crystal roundness, ice crystal stretchiness and ice crystal multi-scale information and the texture damage index through Pearson or Spearman correlation coefficient; according to the correlation analysis results and data characteristics, selecting multiple linear regression, random forest or support vector machine algorithm to construct the ice crystal multi-scale information-texture damage relationship model.

[0091] In the technical solution, the device for determining the thawing loss rate can be an electronic balance with an accuracy of 0.01 g, a low-speed centrifuge, and a culture dish or beaker for weighing. The device for determining the muscle fiber breakage index can be a texture analyzer equipped with a standard shear probe or a puncture probe, or an optical microscope combined with an image analysis system. These devices are placed in the texture analysis area of the laboratory. The experimental object is a meat sample that has been frozen and thawed, such as a meat piece weighing about 50 g per piece. The working process is as follows: first, the thawed meat sample is dried with filter paper to remove surface moisture, and weighed on the balance to record W1. Then, the meat sample is placed in a centrifuge tube and centrifuged at 2000 r / min for 10 min. After taking out, the exudate is again dried and weighed to record W2. The thawing loss rate is calculated according to the formula (W1-W2) / W1x100%. For the muscle fiber breakage index, if a texture analyzer is used, the meat sample is cut into standard size and placed on the test platform. The shear probe is used to press at a constant speed, and the maximum shear force value is recorded as the evaluation basis. If the microscope method is used, the meat sample is made into tissue sections, and multiple fields of view are randomly selected under the microscope. The proportion of intact and broken muscle fibers is counted by image analysis software.

[0092] The correlation analysis is performed in statistical analysis software. The software can be a commercial software with complete statistical functions, or an open-source R language, Python environment combined with SciPy, Pandas, etc. This computer is usually used as a data analysis workstation and placed in the laboratory or computer room. The working process is as follows: first, the thawing loss rate and muscle fiber breakage index data obtained by measurement are arranged into a structured data table with the ice crystal multi-scale information (such as equivalent diameter, roundness, growth rate, etc.) of the corresponding samples extracted from the database, and imported into the statistical analysis software. In the software, the correlation analysis function is called. For data conforming to normal distribution, the Pearson correlation coefficient algorithm can be selected; for non-normal distribution or ordinal data, the Spearman ordinal correlation coefficient algorithm can be selected. The software will calculate the correlation coefficient value and its corresponding p value between each ice crystal feature and each texture damage index. Generally, the correlation coefficient with a p value less than 0.05 is considered statistically significant, and these significantly correlated ice crystal features are selected as candidate input variables for the next model construction.

[0093] The device for building the prediction model can be the same data analysis computer, or a higher performance server on which a machine learning algorithm library needs to be deployed. The algorithm environment that can be selected includes the scikit-learn library of Python, the caret package of R, or other specialized machine learning platforms. The working process is that, based on the significant correlation of the ice crystal feature variables screened out in feature two and the corresponding texture damage index (drip loss rate or muscle fiber breakage index) data, the data set is divided into a training set and a test set at a certain ratio (such as 7:3). The modeling algorithm is selected according to the linear separability of the data, the sample size and other characteristics. For example, if the relationship is roughly linear and the collinearity between features is not strong, a multiple linear regression algorithm can be selected. If there is a complex nonlinear relationship between variables and indicators, a random forest regression algorithm or a support vector machine regression algorithm can be selected. After the algorithm is selected, the training set data is used to train the model, and the model is optimized by adjusting the algorithm parameters (such as the number of trees of the random forest, the kernel function and the penalty coefficient of the support vector machine). After the trained model is verified by the test set data, and the evaluation indicators such as root mean square error or coefficient of determination reach a satisfactory level, it is fixed as an ice crystal multi-scale information-texture damage relationship model that can be used. The model can input new ice crystal multi-scale information to predict the possible degree of texture damage.

[0094] By systematically statistically correlating the macroscopically measurable meat quality damage indicators with the microscopic ice crystal features, the key ice crystal feature parameters that most affect the meat quality damage can be scientifically identified, avoiding the blindness of empirical judgment. Based on the data analysis results, a machine learning model is selected and trained to establish a quantitative and calculable prediction relationship from ice crystal multi-scale information to texture damage degree. This data-driven method makes the evaluation and prediction of frozen meat quality no longer rely solely on the final product test, but can be predicted in advance based on the state of ice crystal formation, providing a quantitative tool and basis for understanding the frozen damage mechanism and optimizing the process.

[0095] In another technical solution, the specific method of the frozen quality score in step three is:

[0096] The original component data of the meat, including moisture content, protein content, fat content, ice crystal multi-scale information, and 11 parameters of frozen temperature, freezing rate, and freezing time of the freezing process, are subjected to dimensionless processing. Based on the ice crystal multi-scale information-texture damage relationship model established in step two, the ice crystal multi-scale information features significantly related to the drip loss rate and muscle fiber breakage index are screened out through correlation analysis;

[0097] The weight coefficients of each parameter are determined by the analytic hierarchy process, and a multivariate linear regression method is used to establish an evaluation model. The frozen quality score is calculated according to the following formula:

[0098] ;

[0099] wherein w i is the weight coefficient of each parameter, x i is the normalized parameter value, b is the bias term, S is the frozen damage degree quality score, and the range of S is 0-100 points, wherein 80≤S≤100 is high quality, 60≤S<80 is good, 40≤S<60 is general, and 0≤S<40 is poor.

[0100] In the technical solution, a data analysis software (such as Python environment and Pandas, NumPy library) and a trained “ice crystal multi-scale information-texture damage relationship model” are deployed on the server. The server is usually placed in a laboratory room or a data center. The working process is as follows: the system first calls the original data of the sample to be scored from the ice crystal multi-scale information database. The 11 parameters include: moisture content (%), protein content (%), fat content (%), ice crystal equivalent diameter (μm), ice crystal roundness (dimensionless), ice crystal stretchability (dimensionless), ice crystal fractal dimension (dimensionless), ice crystal growth rate (μm / s), freezing temperature (℃), freezing rate (℃ / min), and freezing time (min). Then, the system performs Z-score standardization processing on each parameter, that is, subtracting the average value of the parameter in all historical samples and dividing by the standard deviation, so that all parameters are converted into dimensionless values with a mean of 0 and a standard deviation of 1. After standardization, the system inputs the ice crystal-related multi-scale information (such as equivalent diameter, roundness, etc.) into the established “ice crystal multi-scale information-texture damage relationship model”, which outputs the correlation strength of these features with thawing loss rate and muscle fiber fracture index. The system automatically selects the ice crystal features significantly related to texture damage as the core input variables for subsequent scoring modeling based on the correlation strength and statistical analysis (such as p value less than 0.05).

[0101] A special decision analysis software module is run on the server, or the analytic hierarchy process is implemented using tools such as the `pyDecision` library of Python. The working process is as follows: for the parameter system determined for scoring after screening (which may be less than 11 items), the domain expert compares the importance of these parameters two by two through the interactive interface provided by the software, and scores according to the 1-9 scale method to form a judgment matrix. The software automatically calculates the maximum eigenvalue of the matrix and the corresponding normalized eigenvector, which is the weight coefficient (w i ) of each parameter, and performs consistency check (CR value needs to be less than 0.1). After obtaining the weight, the system establishes the final scoring model using the multiple linear regression algorithm. This algorithm uses the normalized parameter values (x i) is the independent variable, and the comprehensive quality score in the expert comprehensive evaluation or historical data is the dependent variable. The regression coefficient (i.e. the weight w i The fine-tuning value of the micro-adjustment value) and the bias term (b) are solved. Finally, the scoring formula is determined, in which S is linearly mapped to the range of 0-100 points.

[0102] The scoring calculation engine integrated in the server can output the calculation results to the monitoring terminal or display screen through the API interface. When new sample data is input, the scoring calculation engine automatically calls the processed standardized parameter value and fixed weight coefficient, substitutes them into the formula, and performs real-time calculation. The calculated S value is a specific value between 0 and 100. The system has built-in grade determination logic: when 80≤S≤100, it is determined to be high quality; when 60≤S<80, it is determined to be good; when 40≤S<60, it is determined to be general; and when 0≤S<40, it is determined to be poor. The score and grade results are stored as a record in the database and can be displayed on the human-machine interface of the upper computer in real time for the operator to view.

[0103] By standardizing and integrating up to 11 parameters such as raw material characteristics, process conditions, and micro-ice crystal characteristics from different sources, and screening key features, the problem of multi-source heterogeneous data being unable to be unified for comprehensive evaluation is solved. The weights are determined using the analytic hierarchy process combined with expert knowledge, so that the scoring model not only reflects the objective data rules, but also takes into account the experience and cognition in actual production, improving the rationality and acceptability of the scoring system. The final generated 0-100 point quantitative score and four clear grade thresholds (when 80≤S≤100, it is determined to be high quality; when 60≤S<80, it is determined to be good; when 40≤S<60, it is determined to be general; and when 0≤S<40, it is determined to be poor) provide an intuitive, unified, and repeatable objective evaluation standard for meat frozen quality. This standardized scoring output provides clear and direct decision-making basis for subsequent automatic optimization and control of process parameters and shelf life prediction.

[0104] In another technical solution, the method for establishing the ice crystal growth kinetics prediction model in step four is as follows:

[0105] Real-time acquisition of meat sample freezing temperature, freezing rate, temperature-time curve, etc., combined with freezing time and original component data of frozen meat samples including moisture content, protein content, fat content, and initial ice crystal multi-scale information obtained through microscopic imaging technology including ice crystal equivalent diameter, ice crystal growth rate, ice crystal roundness, ice crystal stretchability, and ice crystal fractal dimension, etc., a compound kinetics model is constructed by combining classical nucleation theory, KGT model, and Langer model; and the LSTM-XGBoost hybrid machine learning algorithm is used to realize dynamic prediction of ice crystal growth characteristics.

[0106] In the technical solution, the device for real-time data acquisition can include a set of T-type thermocouple temperature sensors, a data acquisition module, and an industrial control computer... Its working process is, after the freezing program starts, the data acquisition module reads the voltage signal of each thermocouple at a fixed frequency (for example, once every second) and converts it into a temperature value. The monitoring software in the control computer receives these temperature data and calculates the temperature change rate, i.e. the freezing rate (unit: ℃ / min) in real time. At the same time, the software identifies the stable stage (i.e. the phase transition platform period) where the temperature change rate significantly decreases by analyzing the real-time collected temperature-time curve, records the duration of this stage as Δt, and takes the average value of the starting and ending temperatures of this stage as the characteristic parameter representing the phase transition process. These real-time collected temperature, rate, phase transition characteristic parameters and freezing time data are temporarily stored in the computer memory and are ready to be sent to the downstream prediction model.

[0107] Its working process is, the program first calls the collected real-time supercooling degree (ΔT), the known freezing point (Tm) of the meat sample, the latent heat of phase transition (ΔH, which can be obtained from the table, for example, the latent heat of phase transition of water is about 334 kJ / m 3 ) and interface energy (σ) and other parameters, and substitutes them into the classical nucleation theory formula r*=2σTm / (ΔHΔT) to calculate the critical radius of ice crystals (r*) under the current conditions. Then, in the growth stage, the program uses the KGT model to substitute the calculated driving force (Δμ) and real-time supercooling degree (ΔT) into the formula, where β and γ are constant parameters obtained by fitting the experimental data in the early stage, to estimate the theoretical growth rate (v) of ice crystals. For the competitive growth of ice crystals, the program introduces the Langer model, uses diffusion coefficient (D) and characteristic supercooling degree (ΔT0) and other parameters, and estimates the growth speed of the tip of the ice crystal through the formula. These intermediate variables calculated by different physical models together constitute a composite model framework describing the nucleation and growth kinetics of ice crystals, and the output provides physically meaningful feature input for the subsequent machine learning model.

[0108] The server is deployed with deep learning frameworks such as TensorFlow, PyTorch and XGBoost library. The server is connected with the control computer through network. The working process is that firstly, the physical model variables (ice crystal critical radius, growth rate, etc.), the collected real-time process data (freezing temperature, rate, time) and the fixed original component data (moisture, protein, fat content) and initial ice crystal multi-scale information of the batch meat sample obtained from the database are integrated to form a multi-scale information data set changing with time sequence. The data set is input into a hybrid model architecture. The time sequence part is processed by LSTM neural network, which can contain 2 hidden layers and 64 neurons in each layer to capture the dynamic evolution trend of ice crystal characteristics. At the same time, all features are also input into an XGBoost regression model to learn the complex static relationship between features and target (ice crystal equivalent diameter at next moment). The two models are trained and predicted in parallel. In the inference (prediction) stage, the output values of the LSTM model and the XGBoost model are weighted and fused by the optimal weight ratio determined in advance by the Bayesian optimization algorithm, so as to obtain more accurate prediction results of the dynamic evolution of ice crystal multi-scale information.

[0109] By collecting real-time production line data and combining with physical model, the prediction model can perceive and respond to the dynamic changes of the current actual freezing conditions, improving the on-site applicability of the model. The classical nucleation and growth kinetics theory model is combined with the advanced LSTM-XGBoost machine learning algorithm, which has the explainability of physical process and the high precision ability of data-driven model to handle complex nonlinear relationship. This hybrid modeling strategy can more stably predict the growth behavior of ice crystals under the coupling action of multiple factors, providing real-time and reliable theoretical basis for dynamically adjusting the freezing process (such as time) based on the prediction results, so as to push the control of freezing process from experience judgment to a new stage of model prediction and real-time optimization.

[0110] In another technical solution, the composite kinetics model comprises:

[0111] Nucleation stage: calculate the critical radius of ice crystal r*=2σT m / (ΔHΔT), wherein σ is the interface energy, the value range is 0.02~0.05J / m 2 ; ΔH is the latent heat of phase transition, J / m 3 ; ΔT is the supercooling degree; T m is the freezing point of the meat sample;

[0112] Growth stage: use KGT model to describe the growth rate of ice crystal , wherein Δμ is the driving force, the value range is 10 3 ~10 5J / mol; β and γ are fitting parameters;

[0113] Competitive growth: Langer model is introduced to analyze the growth rate of ice crystal tips. Where D is the diffusion coefficient, and its value ranges from 10. -9 ~10 -8 m 2 / s; ΔT0 is the characteristic undercooling.

[0114] In this technical solution, the "equipment" relied upon for this stage of calculation mainly consists of a computer or server running the calculation program. The program requires specific input parameters, among which the supercooling ΔT is provided by a real-time temperature sensor monitoring system, i.e., the sample's current temperature and its freezing point T. m The difference. The interfacial energy σ is a system-related physical property parameter; for the ice-water system in meat tissue, its value ranges from 0.02 J / m³. 2 Up to 0.05 J / m 2 The specific value can be determined by consulting relevant literature or previous calibration experiments. The latent heat of phase change ΔH mainly depends on the moisture content of the sample, and is approximately 334 MJ / m for water. 3 For meat, the moisture content can be estimated proportionally. These parameters are input into the program code, which executes the formula r*=2σT. m The formula / (ΔHΔT) is used for iterative calculation. This formula is written as a separate function that iteratively calculates each time new temperature data (T) is received. m When ΔT is reached, the function is invoked. The calculated critical ice crystal radius r* (usually in meters) is output as an intermediate variable, indicating the minimum size of the ice crystal nucleus that can stably exist and begin to grow under the current supercooled conditions. This calculation result is stored in memory and passed to subsequent growth models as one of the input conditions.

[0115] The program receives calculation results from the nucleation stage and real-time supercooling ΔT data. The driving force for ice crystal growth, Δμ, is physically represented by the change in free energy per unit volume of material as it transitions from the liquid phase to the solid phase. For an ice-water system, its typical value ranges from 10... 3 J / mol to 10 5 The values ​​are between J / mol, and specific values ​​can be obtained from thermodynamic data tables or through experimental calibration. β and γ in the model are fitting parameters; they do not have universally fixed values ​​and require prior experimental data on ice crystal growth rates at different degrees of supercooling. Then, a nonlinear regression algorithm (such as the least squares method) is used to refine the formula. The fitting is performed to determine. After obtaining the specific values of β and γ, the program substitutes the real-time ΔT, Δμ, and the fitting parameters into the KGT model formula, i.e. the theoretical ice crystal growth rate v can be calculated, which is usually in meters per second. The rate v represents the average advancing speed of the ice crystal interface in the normal direction, and is one of the key dynamic indicators for describing the speed of ice crystal volume growth.

[0116] The program needs to call two parameters of diffusion coefficient D and characteristic supercooling ΔT0. The diffusion coefficient D reflects the ease of water molecule migration in muscle tissue matrix, and the specific value is affected by the temperature, tissue structure and composition of the meat sample, which can be obtained by experimental measurement or literature value. The characteristic supercooling ΔT0 is a material constant related to interface dynamics, which needs to be calibrated by experiment. The program substitutes the real-time supercooling ΔT, D and ΔT0 into the Langer model formula for calculation. The ice crystal tip growth rate v tip calculated by this formula describes the rate at which the tips of those crystals with favorable growth direction extend rapidly during the competition of ice crystal growth. This rate is usually higher than the average growth rate v calculated by the KGT model, and is of great significance for explaining the ice crystal morphology (such as the formation of dendrites) and the non-uniformity of spatial distribution. The v tip value output by the program, together with variables such as r* and v, constitutes a set of complex dynamic parameters for describing the evolution of ice crystal population.

[0117] By clearly dividing the formation and development of ice crystals during freezing into three physical stages of nucleation, growth and competition growth, and respectively using the verified classical theoretical models for mathematical description, a clear theoretical framework is provided for understanding the micro mechanism of ice crystal formation. Each model provides key calculation parameters (such as critical radius, average growth rate, tip speed) according to the physical nature of different stages, so that the description of ice crystal behavior is extended from a single static size to a dynamic multi-scale level containing nucleation difficulty, volume growth and morphology competition. Taking the output of these theoretical models as features, input into the subsequent data-driven machine learning model, can provide prior knowledge with clear physical meaning for machine learning, thereby enhancing the interpretability and extrapolation prediction ability of the hybrid model, so that the prediction of ice crystal growth in complex meat systems is based on a more solid physical foundation.

[0118] In another technical solution, the machine learning algorithm process is:

[0119] The following parameters are taken as input features, including: ice crystal critical radius, ice crystal growth rate, freezing temperature, freezing rate, characteristic parameters representing the phase change process, freezing time, and original component data of frozen meat samples including moisture content, protein content, fat content, initial ice crystal multi-scale information obtained by micro-imaging technology including ice crystal equivalent diameter, ice crystal growth rate, ice crystal roundness, ice crystal stretchability and ice crystal fractal dimension, etc.

[0120] The time series data is processed using an LSTM neural network to output ice crystal multiscale information prediction values;

[0121] An XGBoost algorithm is used to establish a multi-factor correlation model to output ice crystal multiscale information prediction values;

[0122] The weight proportions w1 and w2 of the LSTM and XGBoost models are determined by a Bayesian optimization algorithm;

[0123] In the real-time running phase of the system, to adapt to online data streams and maintain prediction accuracy, a periodic weight fine-tuning strategy is adopted: after processing N (for example, N = 100) new sample data, the error between the actual observed value and the model prediction value of this part of the sample data is used as feedback, and small-batch gradient descent is used to fine-tune w1 and w2; or after each complete production batch ends, a round of fast Bayesian optimization is performed using the accumulated data of the batch to update the weight proportions. The final prediction value of the ice crystal multiscale information = w1 x LSTM output value + w2 x XGBoost output value.

[0124] In the technical solution, the server receives data from different sources through the network: time series data of ice crystal critical radius (r*, unit: m) and ice crystal growth rate (v, unit: m / s) from a composite dynamics model calculation program; real-time data of freezing temperature (unit: °C), freezing rate (unit: °C / min), and freezing time (unit: min) from a sensor system installed on a freezing device; fixed original component data of the batch meat sample, including moisture content (unit: %), protein content (unit: %), and fat content (unit: %), from a laboratory information management system (LIMS) or a database; ice crystal initial characteristics obtained by microscopic imaging at the initial moment of freezing, including ice crystal equivalent diameter (unit: pm), ice crystal growth rate (unit: pm / s), ice crystal roundness (dimensionless), ice crystal elongation (dimensionless), and ice crystal fractal dimension (dimensionless), from an ice crystal multiscale information database. The data server aligns, cleans, and standardizes the preprocessing of these data with different sources and dimensions to form a feature matrix with fixed dimensions, where each row of the matrix represents a sample at a time step, and each column represents an input feature. The feature matrix is stored in the memory or cache of the server, waiting to be called for model inference.

[0125] TensorFlow or PyTorch deep learning framework is installed on the server. LSTM model is deployed as a submodule in the framework. The architecture of the model can be configured as follows: the number of input layer neurons is equal to the dimension of the input features; it can contain two LSTM hidden layers, each of which can be set to 64 neurons, and use tanh as the activation function; the output layer is a fully connected layer, and the number of neurons is equal to the dimension of the ice crystal multi-scale information (such as equivalent diameter, roundness, etc.) to be predicted. The working process is that the generated feature matrix is cut into fixed length (such as 10 time steps) sequence fragments in chronological order, and input into the trained LSTM model. The LSTM network handles the time-dependent relationship according to its internal gating mechanism (input gate, forget gate, output gate), and finally outputs a vector at the last time step, which is the prediction value of the ice crystal multi-scale information at the next time (or several times). The prediction value is temporarily stored.

[0126] XGBoost machine learning library is installed on the server. An XGBoost regression model trained for the current prediction task is loaded into memory. The parameters of the model can be set in advance, such as the number of trees is set to 100, the learning rate is set to 0.1, and the maximum depth of the tree is set to 6. Its working process is different from LSTM, it does not explicitly handle the time sequence of data. The program flattens all feature values in the feature matrix at the current time and previous time steps as a high-dimensional static feature vector, and inputs it into the XGBoost model. XGBoost model based on its internal integrated large number of decision trees, forward propagation on the static feature vector, each tree gives a prediction value, finally the prediction results of all trees are weighted average, output another ice crystal multi-scale information prediction value vector. The prediction value is also temporarily stored.

[0127] A model fusion and optimization program is run on the server. The program calls a Bayesian optimization algorithm library (such as `bayes_opt` or `Optuna`). During the model training phase, the optimization algorithm takes the prediction error (such as root mean square error) on the pre-reserved validation set as the objective function, and searches for the optimal combination of output weights w1 of the LSTM model and w2 of the XGBoost model as hyperparameters in the range of 0 to 1. After the optimization process is completed, a set of optimal weight combinations that minimize the validation set error is determined, for example, w1 = 0.6, w2 = 0.4. In the model prediction (inference) phase, the program directly uses this set of fixed optimal weights. Its working process is that the program reads and outputs the prediction value vector from the cache, and then performs element-wise weighted summation on the two vectors according to the formula: final prediction value = w1 x LSTM output value + w2 x XGBoost output value. The final vector obtained after summation is the final prediction result of the hybrid model for the ice crystal multi-scale information at the future time, which will be output for subsequent quality score calculation or process decision-making.

[0128] By explicitly defining the diversified input feature system from the physical model, real-time process to the initial state of the sample, a comprehensive information base for the machine learning model is provided to describe the state of the freezing system. The LSTM neural network is used to specifically process the dynamics features strongly related to time evolution, effectively capturing the sequence dependence and historical trends of ice crystal growth. The XGBoost algorithm is used in parallel to model, which can fully learn the complex and nonlinear static relationship between input features and output targets, especially the influence of non-time series factors (such as raw material composition). Finally, the optimal fusion weights of the two heterogeneous models are automatically determined by Bayesian optimization, rather than subjective setting, so that the final prediction value can adaptively combine the advantages of time series model and static model, thereby improving the accuracy of dynamic evolution prediction of ice crystal multi-scale information and the robustness of the model as a whole. This hybrid architecture provides a feasible technical path for dealing with similar complex industrial process prediction problems that have both time series and multi-factor coupling.

[0129] In another technical solution, the determination method of the optimal freezing process and the optimal frozen shelf life in step four is:

[0130] Step a, obtain the original component data of the meat sample and the initial ice crystal multi-scale information from the ice crystal multi-scale information database, and set the initial freezing temperature, rate and time;

[0131] Step b, normalize the initial freezing parameters, original component data and initial ice crystal multi-scale information according to the method of step three, screen significant ice crystal feature information, determine weight coefficients, and calculate the initial freezing quality score S0;

[0132] Step c, predicting the change of ice crystal characteristic information under the current freezing time length by using the real-time ice crystal growth kinetics prediction model:

[0133] If S0≥80, the current freezing condition and freezing time length are maintained in high quality; if 60≤S0<80, the freezing rate is fine-tuned and the freezing time length is maintained in good quality; if 40≤S0<60, the freezing parameters are adjusted according to the preset rules in general quality: the freezing time length is shortened preferentially, and the freezing temperature and rate are adjusted in a stepwise manner within a set range based on the feedback of the ice crystal growth prediction model, and the score S is recalculated until S≥60; if S0<40, the process reset program is started in poor quality: in the database or the preset process library, the freezing process parameter combination matched with the current raw material components and having a historical score≥60 is selected as a new initial process, and the optimization process of steps b to c is repeated until the score is improved to the general quality and above; the shelf life data of the meat under different freezing quality scores are collected, and a correlation model of the freezing quality score and the shelf life is established by using the Cox proportional hazards model, and the optimal freezing shelf life is determined by substituting the dynamically optimized score into the model.

[0134] In the technical solution, the device for implementing the feature can include a data server storing ice crystal multi-scale information, an industrial control computer (PLC or industrial computer) running process control software, and a main controller of the freezing equipment itself. The data server is usually deployed in a laboratory or central computer room. The industrial control computer is located in the control room of the freezing workshop and is connected with the data server and the freezing equipment controller through industrial Ethernet. The working process is that when a new batch of meat samples enters the freezing process, the operator or the automatic code scanning system inputs the sample identification code of the batch on the control computer. The control computer sends a query request to the data server, and the data server retrieves and returns the original component data (moisture, protein, and fat content) of the batch of meat samples and the initial ice crystal multi-scale information (such as initial equivalent diameter, roundness, etc.) obtained through previous detection from the ice crystal multi-scale information database according to the identification code. At the same time, the operator sets the initial parameters of this freezing according to the product specifications on the human-machine interface of the control computer, including the freezing temperature (such as -35℃), the freezing rate (such as -1℃ / min), and the planned freezing time length (such as 240 minutes). The control computer packages these initial parameters, component data, and initial ice crystal information and prepares to send them to the score calculation engine.

[0135] The server receives data packets from the control computer via the network. A frozen quality score calculation engine is running on the server. The working process is that the score calculation engine first pre-processes the received data, including 11 parameters (components, initial ice crystal information, initial freezing parameters) are standardized (Z-score standardization) consistent with the training. Subsequently, the engine calls the stored weight coefficients and multiple linear regression model, and substitutes the standardized parameter values into the formula for calculation. The score calculated is the initial frozen quality score S0. According to the preset grade threshold, S0 will be immediately classified: if 80≤S0≤100, marked as "high quality"; if 60≤S0<80, marked as "good"; if 40≤S0<60, marked as "general"; if 0≤S0<40, marked as "poor". The score and grade results are fed back to the control computer and displayed.

[0136] The control computer triggers the corresponding adjustment logic according to the received initial score S0 and its grade. If 80≤S0≤100 (high quality), the control computer sends an instruction to the freezing equipment controller to maintain the current set freezing temperature, rate and duration. If 60≤S0<80 (good), the control computer sends a fine-tuning freezing rate instruction to the freezing equipment while keeping the freezing duration unchanged, for example, adjusting the rate from -1℃ / min to -0.9℃ / min, in order to obtain better ice crystal morphology. If 40≤S0<60 (general), the control computer will first shorten the planned freezing duration (for example, from 240 minutes to 180 minutes), and at the same time adjust the freezing temperature and rate (for example, the temperature is raised to -30℃, and the rate is adjusted to -2℃ / min), and then based on the new adjusted parameters, the score calculation engine is called again to calculate a new score S. This process can be iterated until the newly calculated S value is improved to the "good" or "high quality" range. If 0≤S0<40 (poor), the control computer determines that the initial process is not feasible, and needs to completely redesign a set of freezing process parameter combination, and repeat the above score and adjustment process until the score meets the standard.

[0137] In the long-term production or experimental process, the system will continuously collect the actual shelf life data of the meat corresponding to different final frozen quality scores (S) (i.e. the time of storage to spoilage, unit: day). These data are stored in the data server. Subsequently, on the statistical analysis computer, using R language, SPSS or Python `lifelines` library and other tools, the shelf life time is used as the dependent variable, and the corresponding frozen quality score S is used as the covariate to fit the Cox proportional risk model, the model form is h(t|S) = h0(t)·exp(βS). Where h(t|S) is the risk function under the score S, h0(t) is the baseline risk function, and β is the regression coefficient of the score S. After the model fitting is completed, for the final frozen quality score S obtained after dynamic optimization final , which is substituted into the Cox model. The model outputs the risk ratio relative to the baseline state under this score, and the expected optimal frozen shelf life with controllable quality degradation risk can be calculated, providing a basis for logistics and sales planning.

[0138] By connecting the initial setting, real-time score, model prediction and process adjustment into an automated decision loop, the frozen process is transformed from static preset to dynamic optimization, which can be personalized for different batches of raw materials. Based on the grading control strategy formulated based on the clear score threshold (80≤S≤100 for high quality, 60≤S<80 for good, 40≤S<60 for general, and 0≤S<40 for poor), the process adjustment has a rule to follow, which not only avoids excessive intervention in high-quality cases, but also ensures timely and effective correction in general or poor cases. Finally, the Cox proportional risk model is introduced to link the microscopic ice crystal score with the macroscopic product shelf life, so that the goal of process optimization is not limited to obtaining a high immediate score, but is extended to achieving a predictable and economic optimal storage length, thereby completing the whole-chain quality control closed loop from process control to warehouse logistics management.

[0139] The application also provides a system for determining the frozen quality of meat based on ice crystal multi-scale information, comprising:

[0140] An ice crystal multi-scale information acquisition module: integrating an image processing unit and a dynamics analysis unit, for receiving ice crystal images and dynamic data obtained by external microscopic imaging equipment and sensors, extracting ice crystal spatial scale, time scale and structure scale features;

[0141] A three-dimensional database module: storing raw material component data, frozen process parameters and ice crystal multi-scale information, establishing an ice crystal multi-scale information database with multi-scale index;

[0142] Texture damage analysis module: connect the thawing loss rate detection unit and the muscle fiber breakage index detection unit, perform Pearson / Spearman correlation analysis of ice crystal multi-scale information and meat quality indicators, construct and output the ice crystal multi-scale information-texture damage relationship model;

[0143] Quality score calculation engine: based on the analytic hierarchy process to determine the weight coefficient, through the multiple linear regression model to input the three-dimensional database data and the texture damage relationship model for dimensionless processing, real-time calculation of 0-100 frozen quality score and quality grade division;

[0144] Freezing process and shelf life optimization system:

[0145] Ice crystal growth prediction sub-module: fusion of nucleation theory, KGT model and Langer model complex kinetics algorithm, combined with LSTM-XGBoost machine learning to predict the evolution of ice crystal multi-scale information in real time;

[0146] Process parameter control sub-module: dynamically adjust the freezing temperature, rate and time according to the quality score;

[0147] Shelf life prediction sub-module: through the Cox proportional hazards model to map the frozen quality score S to the optimal frozen shelf life;

[0148] Among them, the Cox proportional hazards model is defined as: h(t|S)=h0(t)⋅exp(βS); wherein h(t|S) is the risk function under the damage score S, h0(t) is the baseline risk function, and β is the regression coefficient of the score S.

[0149] In the technical solution, the ice crystal multi-scale information acquisition module can be a combination of software and hardware integrated in an industrial computer. The hardware part can include an image acquisition card for receiving digital image signals from an external scanning electron microscope or a low-temperature confocal microscope, and a multi-channel data acquisition card for connecting a temperature sensor (such as a thermocouple) installed in a freezing device to obtain dynamic temperature data. The industrial computer is placed in a freezing laboratory or a workshop control room. Its working process is that the image acquisition card transmits the received ice crystal microscopic images to the computer memory, and the image processing unit (software) running in the computer calls library functions such as OpenCV to perform image analysis and automatically extract the spatial scale and structural scale of ice crystals. At the same time, the data acquisition card reads the voltage signal of the temperature sensor in real time, converts it into temperature and freezing rate through calculation, and records the phase transition time and other time scale characteristics by the dynamics analysis unit (software). The three-dimensional database module can be a relational database management system (such as MySQL) deployed on another server in the same machine room. The server is connected to the above industrial computer through a local area network. Its working process is that it receives and sorts the ice crystal multi-scale information data from the acquisition module, imports the raw material component data from the laboratory information management system, and reads the set process parameters from the freezing equipment controller, writes them into the database table according to the unified data structure, and establishes a composite index based on the process type, raw material batch and time for fast query.

[0150] The texture damage analysis module can rely on a computer installed in a texture analysis laboratory. The computer is connected to a texture analyzer (for measuring muscle fiber breakage index) and a centrifuge (for assisting in measuring thawing loss rate) through a data interface. Its working process is that the operator detects the thawed meat samples, and the detection results (thawing loss rate percentage and muscle fiber breakage index value) are automatically or manually entered into the computer. The texture damage analysis software running in the computer then retrieves the ice crystal multi-scale information corresponding to these meat samples from the database server, and uses the built-in statistical analysis toolkit (such as SciPy) to perform Pearson correlation analysis, calculate the correlation coefficient and p value of each ice crystal feature and damage index, and finally train and output a random forest regression model as the ice crystal multi-scale information-texture damage relationship model according to the analysis results. The quality score calculation engine can be deployed on a dedicated application server. The server obtains the relationship model and raw data through the network, its internal program first standardizes the 11 input parameters, then calls the pre-stored weight coefficients determined by the analytic hierarchy process, performs linear combination calculation according to the formula, and outputs a score S between 0 and 100 in real time, and determines the grade according to the threshold value, and returns the result to the system main interface or downstream control unit.

[0151] The frozen process dynamic optimization system is a software system deployed on the core control server of the workshop. Its ice crystal growth prediction submodule is a background service program that periodically obtains real-time process data and raw material data from the database, and calls the composite kinetics model (integration of nucleation, KGT, Langer model) for calculation, while driving the trained LSTM-XGBoost hybrid machine learning model to predict the multi-scale information evolution trend of ice crystals in the future period of time. The parameter control submodule runs on a programmable logic controller or an industrial control computer, which receives real-time scores S from the quality score calculation engine. If S is lower than the current process target level (for example, lower than the "good" threshold of 60), the submodule will generate adjustment instructions according to the preset optimization rule library, such as "adjust the freezing rate from -1 ℃ / min to -1.5 ℃ / min", and send it to the actuator (such as compressor, throttle valve) of the freezing equipment through the industrial network. The freezing duration mapping submodule runs on the data analysis server, which continuously collects the final score S and the actual shelf life (days) of the corresponding product in the historical production data, and uses statistical software to fit the Cox proportional hazards model h(t|S)=h0(t)⋅exp(βS). When a batch of products is completed and obtains the final score S final After that, the submodule substitutes this score into the model to calculate the recommended optimal frozen shelf life under the acceptable quality decay risk, and provides this period to the warehouse management system.

[0152] By integrating the dispersed ice crystal image acquisition, kinetics analysis, quality detection, score calculation and process control functions into a software and hardware cooperative system, the whole process automation from data perception to decision execution is realized. The system not only can statically evaluate the frozen quality, but more importantly, through the internal closed-loop dynamic optimization mechanism, it can adaptively adjust the temperature, rate and other key parameters in the freezing process according to the real-time quality score and ice crystal growth prediction. Finally, by correlating the microscopic ice crystal score with the macroscopic product shelf life through a statistical model, the output of the system is no longer limited to process parameter recommendations, but can directly guide the subsequent logistics and warehouse planning, thereby forming a complete technical solution covering frozen processing, quality control and supply chain management, and improving the overall intelligent control level of the meat freezing industry.

[0153] Example 1

[0154] This example takes chicken breast as the experimental object to demonstrate the specific implementation process of the method and system.

[0155] 1. Sample preparation and initial data acquisition Fresh chicken breast meat (within 24 hours postmortem) was purchased from a local market as experimental samples. First, the original composition data of the meat samples were determined by using a near-infrared spectroscopy rapid analyzer, and the average moisture content of a sample was 74.2%, the protein content was 22.1%, and the fat content was 2.5%. Subsequently, the meat samples were cut into small pieces with a size of 2 cm x 1 cm x 1 cm for later use.

[0156] 2. Establishment of ice crystal multiscale information database

[0157] The prepared chicken sample pieces were placed in a programmed freezer for freezing treatment. Three different freezing processes were set: Group A was -20°C, freezing rate -1°C / min; Group B was -40°C, freezing rate -5°C / min; and Group C was -60°C, freezing rate -30°C / min (as shown in FIG. 1, which compares the freezing rate and freezing time characteristics at different temperatures). During the initial nucleation period (about 1-2 minutes after the temperature passed the freezing point), the rapid growth period (temperature rapidly decreased), and the stable period (close to the target temperature) of each freezing process, the sample pieces were quickly removed. Figure 10

[0158] The removed samples were immediately sliced using a freezing microtome, and the slice thickness was controlled at 15 microns. Before slicing, the samples were embedded with a specific concentration of sodium carboxymethyl cellulose solution. The prepared frozen sections were quickly transferred to the low-temperature sample stage of a scanning electron microscope for observation. As shown in FIG. 2, it is a comparison of ice crystal microscopic imaging collected at different freezing temperatures, which intuitively shows the significant effect of temperature on ice crystal size and morphology. After gold spraying in the electron microscope vacuum chamber, the microscopic digital images of ice crystals were collected at an acceleration voltage of 5 kV and a magnification of 2000 times. Figure 4

[0159] The collected images were analyzed by image processing software. As shown in FIGS. 3 and 4, they are the comparison of ice crystal equivalent diameter statistics and the comparison of ice crystal relative area analysis obtained by image analysis, which quantitatively reveals the differences in ice crystal growth at different freezing temperatures. For each image, the following ice crystal multiscale information was extracted: spatial scale characteristics (ice crystal equivalent diameter, distribution density); structural scale characteristics (roundness, stretch, fractal dimension); for a series of images at different time points under the same process, further calculate the time scale characteristics (ice crystal growth rate, nucleation rate estimated based on the number of new ice crystals). Figures 5-8 Figure 9

[0160] ​​​​Finally, the corresponding frozen process parameters (temperature, rate, time point) of each group of samples, raw material component data (moisture, protein, fat content), and extracted ice crystal multi-scale information data are stored in a database constructed by a MySQL database management system in a structured manner, forming an ice crystal multi-scale information database.

[0161] 3. Construction of texture damage relationship model

[0162] After completing image acquisition, the frozen chicken samples are thawed and the quality indicators are detected. As shown in FIGS. 3 and 4, the appearance state and microstructure of chicken breast after different freezing temperature treatments are compared, clearly showing the apparent and microscopic differences of freezing damage. After thawing, the surface moisture is absorbed with filter paper and weighed, then centrifuged at 2000 rpm for 10 minutes, weighed again, and the thawing loss rate is calculated. At the same time, another thawed meat sample is used to perform shear force testing using a shear probe equipped with a texture analyzer, and the maximum shear force value is taken as the representation value of the muscle fiber fracture index. Figure 2 Figure 3 The ice crystal multi-scale information (such as average equivalent diameter, average roundness, ice crystal density, etc.) corresponding to these measured samples is retrieved from the database and arranged with the measured thawing loss rate and shear force data. Using the SciPy library of Python, the Pearson correlation coefficient between each ice crystal feature and the two quality indicators is calculated. Analysis shows that the ice crystal equivalent diameter, ice crystal roundness, and ice crystal density have the most significant correlation with the thawing loss rate and shear force (p value is less than 0.05).

[0163] Based on these significantly correlated features, a random forest regression algorithm in the scikit-learn library is used to construct a prediction model. With ice crystal equivalent diameter, roundness, and density as input features, and thawing loss rate and shear force as output targets, 70% of the data is used for model training, and the remaining 30% is used for testing. The trained model can be used as an "ice crystal multi-scale information-texture damage relationship model" to predict the theoretical texture damage degree of new samples based on their ice crystal features.

[0164] 4. Frozen quality score and grade division

[0165] When a new batch of frozen chicken needs to be evaluated for quality, the system performs the following operations: First, obtain the component data of the batch of chicken from the database, and the ice crystal multi-scale information obtained through initial imaging (or provided by the prediction model). At the same time, input its frozen process parameters (temperature, rate, planned duration). The system will standardize these 11 parameters.

[0166] When a new batch of frozen chicken needs to be evaluated for quality, the system performs the following operations: First, obtain the component data of the batch of chicken from the database, and the ice crystal multi-scale information obtained through initial imaging (or provided by the prediction model). At the same time, input its frozen process parameters (temperature, rate, planned duration). The system will standardize these 11 parameters.

[0167] ​The weight coefficients have been pre-determined by the analytic hierarchy process. For example, the weight of moisture content, ice crystal equivalent diameter and freezing rate can be higher. The system calls the multiple linear regression scoring model, multiplies the normalized parameter values by the respective weights and sums them up, plus the bias term, to calculate a comprehensive score (S) between 0 and 100. The scoring criteria are: 80≤S≤100 is excellent, 60≤S<80 is good, 40≤S<60 is fair, and S<40 is poor. For example, a batch of chicken frozen at -35℃ with a process of -2℃ / min, the calculated S value is 72, then the system determines its quality grade as "good".

[0168] 5. Ice crystal growth kinetics prediction and process dynamic optimization

[0169] For the ongoing freezing process, the system starts the dynamic optimization module. The temperature sensor collects the temperature data of the chicken sample in real time, calculates the supercooling degree (ΔT) and the instantaneous freezing rate. The ice crystal growth prediction sub-module combines physical models and machine learning models: first, according to the classical nucleation theory, KGT model and Langer model (using pre-set physical parameters, such as interface energy σ=0.03 J / m 2 , diffusion coefficient D=5×10 -9 m 2 / s), calculate the critical radius of ice crystal and the theoretical growth rate. At the same time, the real-time process data, raw material components and initial ice crystal characteristics are input into a trained LSTM-XGBoost hybrid machine learning model to predict the evolution trend of ice crystal equivalent diameter and roundness in the future period of time.

[0170] Assuming that the final ice crystal state predicted under the initial process corresponds to a quality score S0 calculated as 65 (good). The parameter adjustment sub-module decides to fine-tune the process parameters to seek improvement according to the "good" level strategy. It sends instructions to the freezing equipment controller to fine-tune the freezing rate from -2℃ / min to -1.8℃ / min. The system re-performs ice crystal growth prediction and quality score calculation based on the new rate. This iterative process can be repeated until the score tends to be stable or reaches the target level.

[0171] 6. Optimal frozen shelf life prediction

[0172] The actual shelf life data (days until sensory or microbiological indicators fail) of batches of frozen chicken meat with different final scores (S) in history were collected and stored at -18℃. Using statistical analysis software, the Cox proportional hazards model was fitted with shelf life days as the time variable and quality score S as the covariate: h(t|S) = h0(t)⋅exp(βS). After the model fitting, for a batch of frozen chicken meat with final optimized score of 85 (high quality), its score was substituted into the model, and the expected optimal shelf life significantly extended relative to the baseline sample could be calculated, thus providing quantitative basis for cold chain logistics and sales planning.

[0173] 7. System integration operation

[0174] All the above steps constitute a complete prediction and optimization closed loop. As shown in the figure, it is the system workflow diagram of the present application for determining the frozen quality of meat based on ice crystal multi-scale information, which clearly illustrates the whole process logic from data acquisition, database construction, modeling, scoring to dynamic optimization and shelf life prediction. Each module exchanges data and transmits instructions through a local area network, forming an automated operation system from micro-ice crystal monitoring to macro-quality control. Figure 1

[0175] Although the embodiments of the present application have been disclosed as above, it is not limited to the application listed in the specification and embodiments, and can be fully applied to various fields suitable for the present application. Those skilled in the art can easily realize additional modifications, and therefore the present application is not limited to specific details and examples shown and described herein, without departing from the general concept defined by the claims and equivalent scope.​

Claims

1. A method for determining the freezing quality of meat based on multi-scale information of ice crystals, characterized in that, Comprise the following steps: Step one, obtain the ice crystal multi-scale information of frozen livestock and poultry meat samples at different stages and under different conditions through microscopic imaging technology, integrate the freezing process parameters, meat raw material component data and ice crystal multi-scale information, and establish an ice crystal multi-scale information database; Step two, detect the quality indicators of the frozen meat samples after thawing, correlate them with the corresponding ice crystal multi-scale information, and construct an ice crystal multi-scale information-texture damage relationship model; Step three, according to the data integrated in step one and the ice crystal multi-scale information-texture damage relationship model constructed in step two, score and grade the frozen meat quality; Step four, based on the ice crystal multi-scale information obtained in step one and the ice crystal multi-scale information-texture damage relationship model obtained in step two, establish an ice crystal growth kinetics prediction model, optimize the freezing time and establish a frozen optimal quality prediction model, and the frozen optimal quality prediction model determines the optimal freezing process and optimal frozen shelf life corresponding to the optimal frozen quality through the frozen quality score of step three; Wherein, the specific method of frozen meat quality scoring in step three is: The original component data of meat including moisture content, protein content, fat content, ice crystal multi-scale information and freezing temperature, freezing rate, freezing time of freezing process are dimensionless processed, based on the ice crystal multi-scale information-texture damage relationship model established in step two, the ice crystal multi-scale information characteristics significantly related to thawing loss rate and muscle fiber fracture index are selected by correlation analysis; The weight coefficients of each parameter are determined by analytic hierarchy process, and the evaluation model is established by multiple linear regression method, and the frozen quality score is calculated according to the following formula: ； where w i is a weight coefficient for each parameter, x i is a normalized parameter value, b is a bias term, and S is a frozen product quality score, with S ranging from 0 to 100, wherein 80≤S≤100 is excellent, 60≤S<80 is good, 40≤S<60 is fair, and 0≤S<40 is poor.

2. A method of determining the freezing quality of meat based on ice crystal multiscale information according to claim 1, characterized in that, The ice crystal multi-scale information includes spatial scale characteristics, time scale characteristics and structure scale characteristics; wherein, the spatial scale characteristics include ice crystal equivalent diameter, distribution density and regional aggregation degree; the time scale characteristics include ice crystal growth rate and nucleation rate; the structure scale characteristics include ice crystal roundness, stretchability and fractal dimension.

3. The method of determining the freezing quality of meat based on ice crystal multiscale information according to claim 1, characterized in that, The establishment process of the ice crystal multi-scale information database in step one is: Select different types of meat samples, measure their original component data, including moisture content, protein content and fat content; Use scanning electron microscope to image the frozen meat samples, collect the ice crystal images of the samples at the initial nucleation stage, rapid growth stage and stable stage during freezing, cover the temperature gradient of -5℃ to -80℃ and the freezing rate of 0.1℃ / min to 100℃ / min, and control the thickness of the frozen section at 10-20μm; Extract the ice crystal spatial scale, time scale and structure scale information through image processing algorithm, including equivalent diameter, ice crystal growth rate, ice crystal roundness, ice crystal stretchability and ice crystal fractal dimension, and establish a three-dimensional ice crystal multi-scale information database according to the freezing process parameters-raw material components-freezing time.

4. The method of determining the freezing quality of meat based on ice crystal multi-scale information according to claim 1, characterized in that, The specific method for constructing the ice crystal multi-scale information-texture damage relationship model in step two is: determining the thawing loss rate and muscle fiber breakage index of the thawed frozen meat; analyzing the correlation between the ice crystal multi-scale information and the texture damage index through Pearson or Spearman correlation coefficient; selecting a multiple linear regression, random forest or support vector machine algorithm to construct the ice crystal multi-scale information-texture damage relationship model according to the correlation analysis results and data characteristics.

5. The method of determining the freezing quality of meat based on ice crystal multiscale information according to claim 1, characterized in that, The method for establishing the ice crystal growth kinetics prediction model in step four is: Real-time acquisition of the freezing temperature, freezing rate and temperature-time curve of the meat sample, extraction of the characteristic parameters representing the phase change process based on the temperature-time curve, combination of the freezing time and the original component data of the frozen meat sample and the initial ice crystal multi-scale information, fusion of the classical nucleation theory, KGT model and Langer model, and construction of a composite kinetics model; then, combining the LSTM-XGBoost hybrid machine learning algorithm to construct an ice crystal growth prediction model, and outputting the dynamic evolution results of the ice crystal multi-scale information.

6. A method of determining the freezing quality of meat based on ice crystal multiscale information according to claim 5, characterized in that, The composite kinetics model includes: Nucleation stage: critical radius of ice crystals is calculated based on classical nucleation theory where σ is the interfacial energy, ranging from 0.02 to 0.05 J / m 2 ; ΔH is the latent heat of phase transition, J / m 3 ; ΔT is the supercooling degree; T m is the meat-like ice point; Growth stage: KGT model was used to describe the ice crystal growth rate where Δμ is the driving force, ranging from 10 3 ~10 5 J / mol; β, γ are fitting parameters; Competitive growth: Langer model is introduced to analyze the growth rate of ice crystal tip where D is the diffusion coefficient, ranging from 10 -9 ~10 -8 m 2 / s; AT0 is the characteristic supercooling.

7. A method of determining the freezing quality of meat based on ice crystal multiscale information according to claim 6, characterized in that, The machine learning algorithm process is: The following parameters are used as input features, including: ice crystal critical radius, ice crystal growth rate, freezing temperature, freezing rate, characteristic parameters representing the phase change process, freezing time, and original component data of the frozen meat including moisture content, protein content and fat content; the initial ice crystal multi-scale information obtained through microscopic imaging technology includes ice crystal equivalent diameter, ice crystal growth rate, ice crystal roundness, ice crystal stretchability and ice crystal fractal dimension; The LSTM neural network is used to process time series data, and the ice crystal multi-scale information prediction value is outputted; The XGBoost algorithm is used to establish a multi-factor correlation model, and the ice crystal multi-scale information prediction value is outputted; In the model training stage, the weight proportions w1 and w2 of the LSTM and XGBoost models are determined through the Bayesian optimization algorithm; in the real-time prediction stage, based on the feedback of the prediction error in the sliding window, the incremental learning or periodic weight fine-tuning strategy is adopted to dynamically update w1 and w2; The final prediction value of the ice crystal multi-scale information = w1*LSTM output value + w2*XGBoost output value.

8. The method of determining the freezing quality of meat based on ice crystal multiscale information according to claim 1, wherein, The method for determining the optimal freezing process and the optimal frozen shelf life in step four is: Step a, obtain the original component data and the initial ice crystal multi-scale information of the meat sample from the ice crystal multi-scale information database, and set the initial freezing temperature, rate and time; Step b, dimensionless processing of the initial freezing parameters, original component data and initial ice crystal multi-scale information according to the method of step three, screening of significant ice crystal multi-scale information, determination of the weight coefficient, and calculation of the initial freezing quality score S0; Step c, use the ice crystal growth kinetics prediction model to predict the change of the ice crystal multi-scale information under the current freezing time: If S0≥80, the current freezing condition and freezing duration are maintained if the quality is excellent; if 60≤S0<80, the freezing rate is fine-tuned and the freezing duration is maintained if the quality is good; if 40≤S0<60, the freezing parameters are adjusted according to the preset rules if the quality is general: the freezing duration is shortened preferentially, and the freezing temperature and rate are adjusted in a stepwise manner within a set range based on the feedback of the ice crystal growth prediction model, and the score S is recalculated until S≥60; if S0<40, the process reset program is started if the quality is poor: in the database or the preset process library, a combination of freezing process parameters that matches the current raw material components and has a historical score≥60 is selected as a new initial process, and the optimization process of steps b to c is repeated until the score is improved to the general or above range; the shelf life data of the meat under different freezing quality scores are collected, and a correlation model of the freezing quality score and the shelf life is established by using the Cox proportional hazards model, and the optimal frozen shelf life is determined by substituting the dynamically optimized score into the model.

9. A system for determining the freezing quality of meat based on multi-scale information of ice crystals, characterized in that, It comprises: An ice crystal multi-scale information acquisition module: integrating an image processing unit and a dynamics analysis unit, for receiving ice crystal images and dynamic data acquired by external microscopic imaging equipment and sensors, extracting ice crystal spatial scale, time scale and structure scale information; A three-dimensional database module: storing raw material component data, freezing process parameters and ice crystal multi-scale information, and establishing an ice crystal multi-scale information database with multi-scale index; A texture damage analysis module: connecting a thawing loss rate detection unit and a muscle fiber fracture index detection unit, performing Pearson / Spearman correlation analysis of ice crystal multi-scale information and meat quality indicators, and constructing and outputting an ice crystal multi-scale information-texture damage relationship model; A quality score calculation engine: determining weight coefficients based on the analytic hierarchy process, and performing dimensionless processing on the input three-dimensional database data and texture damage relationship model by using a multiple linear regression model, to calculate a frozen quality score of 0-100 points in real time and divide the quality grade; A freezing process and shelf life optimization system: An ice crystal growth prediction sub-module: combining a nucleation theory, a KGT model and a Langer model, and using a composite dynamics algorithm to predict the evolution of ice crystal multi-scale information in real time by using LSTM-XGBoost machine learning; A process parameter control sub-module: dynamically adjusting the freezing temperature, rate and duration according to the quality score; A shelf life prediction sub-module: mapping the frozen quality score S to the optimal frozen shelf life by using the Cox proportional hazards model; The Cox proportional hazards model is defined as: h(t|S)=h0(t)·exp(βS); wherein h(t|S) is the risk function under the frozen quality score S, h0(t) is the baseline risk function, and β is the regression coefficient of the frozen quality score S.

Images