Quality evaluation and grading method of preserved vegetable and pork with black bean

Abstract

The application discloses a quality evaluation and grading method for preserved plum and pork, and belongs to the technical field of food quality analysis and intelligent grading, and aims to solve the technical problems that traditional sensory evaluation of preserved plum and pork is highly subjective and poor in consistency, and multi-dimensional physicochemical index data is complex and redundant, and it is difficult to directly correlate comprehensive quality and realize automatic grading. The method determines the texture profile analysis parameters, basic nutritional ingredients and volatile flavor substances of the sample, determines the key quality indicators through two rounds of screening of odor activity value and variable importance projection value, extracts principal components by using principal component analysis for dimension reduction, finally inputs the principal component scores into a pre-trained partial least squares regression model to obtain quality prediction scores, and grades according to the quality prediction scores. The method is mainly used for objective and efficient quality evaluation and automatic grading of preserved plum and pork products, and provides a reliable technical means for quality control in industrial production.

Description

Technical Field

[0001] This invention belongs to the field of food quality analysis and intelligent grading technology, specifically relating to a method for quality evaluation and grading of braised pork with preserved mustard greens. Background Technology

[0002] Currently, in the industrial production and quality control of traditional meat products such as braised pork belly with preserved mustard greens, establishing an objective, accurate, and efficient quality evaluation and grading system remains a pressing challenge. Existing quality assessment methods primarily rely on sensory evaluation, where trained evaluators score products based on color, aroma, taste, and texture. While sensory evaluation directly reflects consumer preferences, this method has significant limitations and instability. First, sensory evaluation results are easily influenced by individual evaluator differences, fatigue, psychological expectations, and external environmental factors, leading to poor repeatability, strong subjectivity, and difficulty in establishing unified and quantifiable standards. Second, the organization and implementation of sensory evaluation are costly, time-consuming, and labor-intensive, failing to meet the demands of modern food production lines for rapid, real-time testing and grading of large quantities of products.

[0003] From a physicochemical perspective, researchers and producers have attempted to use instrumental methods such as texture profile analysis (TPA), nutritional component detection, and volatile flavor compound identification to objectively measure braised pork with preserved mustard greens, hoping to supplement or replace sensory evaluation. However, these methods also face many difficulties in practical application. The quality of braised pork with preserved mustard greens is a complex system composed of multiple dimensions and indicators, including texture, nutrition, and flavor, with complex interactions and redundancies among the indicators. Directly using all the original indicator data will not only lead to the "curse of dimensionality" due to excessive dimensionality, increasing the computational burden of model construction and the risk of overfitting, but it is also difficult to effectively discover and condense the key core indicators that truly determine the final sensory quality. For example, among the many detected volatile flavor compounds, not every one contributes significantly to the overall flavor; including them all indiscriminately in the model will result in noise masking the effective signal. At the same time, different quality indicators have different units and dimensions. How to effectively integrate and comprehensively analyze these heterogeneous data to form a mathematical model that can stably and accurately predict the final comprehensive sensory score is a technical challenge.

[0004] Therefore, how to overcome the subjectivity of sensory evaluation, the limitations of a single model, and the variability caused by different storage conditions, and construct an objective method that is robust, adaptable, and can accurately reflect the overall quality of braised pork with preserved mustard greens and achieve automated grading, based on fully integrating multi-dimensional objective indicator information, is a long-standing technical problem in actual production in this field that has not yet been satisfactorily solved. Summary of the Invention

[0005] The quality evaluation of traditional dishes such as braised pork belly with preserved mustard greens has long relied on sensory assessments, which are highly subjective and inconsistent. However, relying solely on instrument-measured physicochemical indicators is problematic because the data is too dimensional and the relationships between indicators are complex, making it difficult to directly correlate them with sensory quality and thus hindering the formation of a stable, efficient, and quantifiable automated grading system. This invention aims to solve the technical problem of how to screen out the key indicators that truly affect quality from multidimensional and heterogeneous raw quality data, and to construct a system through scientific modeling methods that can accurately predict sensory scores, thereby achieving objective grading.

[0006] In the classification process based on model-predicted scores, samples near the classification boundary may be misclassified due to slight fluctuations in their predicted values, leading to unstable classification results. This invention aims to address how to utilize an objective data dimension other than the predictive model to assist in the adjudication of boundary samples, thereby improving the robustness and accuracy of classification decisions. This is a problem that requires further investigation.

[0007] The evolution of quality indicators such as texture and flavor of braised pork belly with preserved mustard greens differs significantly under two main storage conditions: frozen and room temperature. Using a uniform model to evaluate all samples would lead to decreased prediction accuracy because the model fails to adequately learn the quality characteristics under different conditions. Therefore, this invention aims to solve the problem of how to make the evaluation system adaptable to different storage conditions while maintaining high accuracy.

[0008] The performance of partial least squares regression (PLSR) models is highly dependent on the settings of their key parameters (such as the number of latent variables). This invention aims to address how to construct a high-performance, reliable PLSR model through a standardized and verifiable training process, avoiding overfitting or underfitting, and ensuring stable predictive ability on unknown samples. This is a core element in guaranteeing the effectiveness of the entire grading method.

[0009] The principal component score F is a crucial objective basis for boundary verification, and its calculation depends on the determination of the weights of each principal component. This invention aims to address how to provide experimentally validated, specific, and clear weighting coefficients for sample groups under different storage conditions, making the calculation of the overall score F standardized, operable, and able to truly reflect the quality information structure of each sample group. This is a problem that needs to be specifically solved.

[0010] When faced with complex data, a single PLSR model may exhibit instability (large variance) in its prediction results. This invention aims to address how to effectively reduce the prediction variance of a single model and improve the overall prediction stability and generalization ability through model ensemble techniques, thereby obtaining a more robust quality prediction model, which is necessary to further improve the reliability of grading.

[0011] While ensemble models are stable, significant discrepancies (large standard deviations) can arise in the predictions of individual sub-models when forecasting certain specific samples. In such cases, relying solely on the weighted average of the models reduces reliability. This invention aims to address how to design a dynamic decision-making mechanism that trusts the model when internal consensus is high, and introduces an independent objective evaluation system (F-score) for correction when consensus is low. This ensures a more prudent and reliable final prediction in all situations, which is crucial for improving the method's intelligence and fault tolerance.

[0012] When using principal component analysis (PCA) scores (F) for boundary verification, a set of specific, objective, and operable judgment rules is needed. This invention aims to solve the problem of how to quantify the similarity between the F-value of boundary samples and the distribution of F-values ​​of typical samples of known grades, and based on this, to formulate clear conditions for grade adjustment or downgrading. This makes the verification process systematic and avoids secondary subjective judgments, which is a rule refinement problem that must be solved to achieve automated boundary verification.

[0013] The grading boundary should not be an absolute, rigid point, but rather a reasonable interval to accommodate minor errors in the prediction model. This invention aims to address how to scientifically determine the width of this boundary interval so that it can effectively identify true critical samples without negatively impacting overall grading efficiency due to an excessively wide interval. This is a parameter that needs to be quantified to optimize the grading strategy.

[0014] The final grading requires clear grading standards. This invention aims to address how to determine specific and reasonable score thresholds based on the actual distribution of model-predicted scores and the industry's general understanding of product grading (such as excellent, medium, and qualified). Mapping continuous predicted scores to discrete quality levels is the final step in completing the entire grading process and requires a clear and reasonable setting method.

[0015] To achieve these objectives and other advantages of the present invention, a method for quality evaluation and grading of braised pork with preserved mustard greens is provided, comprising the following steps: S1. Determine multiple quality indicators of the preserved mustard greens and braised pork samples to obtain raw data; the multiple quality indicators include texture profile analysis parameters, basic nutritional components and volatile flavor substances; S2. Select key quality indicators from the raw data, specifically including the following sub-steps: S21. For volatile flavor substances, calculate the odor activity value (OAV) of each substance, and select substances with OAV ≥ 1 to form the first flavor substance set; S22. Using the comprehensive sensory score as the dependent variable, and all substances in the first flavor substance set, as well as all texture profile analysis parameters and basic nutritional components measured in step S1, as independent variables, perform a first partial least squares regression analysis to calculate the variable importance projection value of each independent variable; S23. Select volatile flavor substances with variable importance projection values ​​greater than 1 from the first flavor substance set, and record them as key volatile flavor substances; The key quality indicators include: shear strength, hardness, adhesiveness, elasticity, cohesiveness, tackiness, chewiness, resilience, total amino acids, moisture, protein, fat, and the key volatile flavor substances; S3. Based on principal component analysis, perform dimensionality reduction on the data of the key quality indicators and extract N principal components with a cumulative variance contribution rate of more than 80% and an eigenvalue greater than 1. S4. Input the scores of the N principal components into the pre-trained partial least squares regression model for quality prediction to obtain the quality prediction score of the sample; the partial least squares regression model for quality prediction uses the principal component scores as input variables and the comprehensive sensory score as output variables, and is obtained through sample training. S5. Classify the quality grade of the preserved mustard greens and braised pork samples according to the quality prediction score.

[0016] Preferably, step S3 further includes: obtaining the principal component comprehensive score F for each sample based on principal component analysis; wherein the principal component comprehensive score F is obtained by weighting and summing the scores of each principal component using the proportion of the eigenvalue of each principal component to the sum of the extracted principal component eigenvalues ​​as the weight. In step S5, when the quality prediction score is within the preset grade division boundary range, the grade of the sample is reviewed or adjusted with reference to the principal component comprehensive score F.

[0017] Preferably, in step S3, the analysis based on principal component analysis specifically involves: dividing the samples into a frozen group and a room temperature group according to the storage conditions; performing principal component analysis independently on the key quality index data of the two groups of samples, and calculating the principal component comprehensive score F for each group of samples. In steps S4 and S5, for the samples in the frozen group and the room temperature group, the principal component scores extracted from the corresponding group and the independently trained partial least squares regression model are used for prediction and classification.

[0018] Preferably, in step S4, the training process of the pre-trained partial least squares regression model includes: The sample dataset is divided into training and test sets according to a certain ratio; Cross-validation is used on the training set to select the optimal number of latent variables for the model; A partial least squares regression model was established and validated on the test set. The partial least squares regression model was successfully trained when the prediction determination coefficient was greater than 0.90 and the prediction root mean square error was less than 1.0.

[0019] Preferably, for frozen samples, the top three principal components with a cumulative variance contribution rate of 100% are extracted, and the formula for calculating their principal component composite score F is: F = 0.554 × F1 + 0.303 × F2 + 0.143 × F3; For the samples in the room temperature group, the top three principal components with a cumulative variance contribution rate of 100% are extracted, and the formula for calculating their principal component composite score F is: F = 0.528 × F1 + 0.246 × F2 + 0.226 × F3; F1, F2, and F3 represent the scores of the first, second, and third principal components of the corresponding groups, respectively.

[0020] Preferably, in step S4, the pre-trained partial least squares regression model is a Bagging-based model. The ensemble model is obtained by training using the following method: M1. Set the integration size K, K≥3; Perform K random samplings with replacement from the original training sample set that determines the key quality indicators. Each sampling has the same size as the original set, generating K subsets of training sets. Samples that are not selected form the corresponding K out-of-bag sample sets. M2. For the k-th subset of training, using the principal component scores extracted in step S3 as input and the comprehensive sensory score as output, independently train a partial least squares regression sub-model M. k Where k=1,...,K; M3, for the kth sub-model M k Using its out-of-bag sample set for prediction, the out-of-bag prediction determination coefficient R is calculated. 2 oob,k According to formula ω k = R 2 oob,k / Σ(R 2 oob,i ), calculate its weight coefficients, where i = 1 to K; M4. Save all K sub-models {M1, …, M} K} and its weight coefficients {ω1, …,ω K Together, they constitute the Bagging ensemble model.

[0021] Preferably, when applying to a specific storage group, the final quality prediction is performed according to the following procedure: A. For the sample to be tested, call the trained Bagging ensemble model of its corresponding group to obtain the predicted values ​​{y1, …, y} of the K sub-models. K}; B. Calculate the initial model prediction score Y for this sample. model =Σ(ω k ×y k ), where ω k These are the weight coefficients for the corresponding sub-models; simultaneously, the standard deviation σ of these K predicted values ​​is calculated. y ; C. Call the principal component comprehensive score F of the sample; the principal component comprehensive score F is calculated using the predetermined principal component weighting coefficients of the group, wherein the weighting coefficients for the frozen group are (0.554, 0.303, 0.143) and for the room temperature group are (0.528, 0.246, 0.226); D. Perform dynamic fusion decision-making and calculate the final quality prediction score Y: Y = λ × Y model + (1 - λ)× F; where the fusion weight λ is based on the standard deviation σ. y Dynamically determined, the range of λ is (0, 1]: if σ y If ≤ δ, then λ = 1.0; if σ y If δ > 0, then λ = γ / σ y The discreteness threshold δ and the adjustment parameter γ are determined by analyzing the σ values ​​of the samples in the training set. y The correlation between the predicted residuals and the predicted residuals is optimized and determined.

[0022] Preferably, in step S5, the sample is graded or adjusted with reference to the principal component comprehensive score F, which specifically includes the following sub-steps: S51. When the predicted quality score Y of a sample is within the boundary interval between adjacent high-grade A and low-grade B, obtain the principal component composite score F of the sample, and at the same time obtain the average principal component composite score Favg of the predetermined grade A reference sample. A With standard deviation σ A And the average principal component composite score (Favg) of the B-level reference sample. B With standard deviation σ B ; S52. Calculate the standardized deviation D of the overall principal component score F of the sample relative to grade A. A The calculation formula is: D A =|F - Favg A | / σ A , where σA is the standard deviation of the F value of the A-level reference sample; at the same time, calculate its standardized deviation D relative to grade B B , and the calculation formula is: D B = |F - Favg B | / σ B , where σ B is the standard deviation of the F value of the B-level reference sample; S53. Perform grade determination: If D A is less than or equal to D B , and at the same time satisfies that F is greater than or equal to Favg A , then adjust the grade of the sample upward or confirm it as grade A; if D B is less than D A , and at the same time satisfies that F is less than or equal to Favg B , then adjust the grade of the sample downward or confirm it as grade B; if none of the above conditions are met, then maintain the initial grading result based on the quality prediction score Y.

[0023] Preferably, the preset grade division boundary interval described in step S5 is: Let the score boundary for grade division be T, then the boundary interval is [T - β, T + β], where β is determined according to the standard deviation of the sample prediction scores in the training set, and the value range is 0.3×ρ ≤ β ≤ 0.7×ρ, and ρ is the standard deviation of the prediction scores.

[0024] Preferably, the quality grade division described in step S5 is three levels, specifically: Sort the quality prediction scores Y from high to low, and determine two segmentation thresholds T1 and T2, where T1 > T2, corresponding to the boundaries between the excellent level and the intermediate level, and between the intermediate level and the qualified level respectively; if Y ≥ T1, then determine it as the excellent level; if T2 ≤ Y < T1, then determine it as the intermediate level; if Y < T2, then determine it as the qualified level; where the values of T1 and T2 are jointly determined by analyzing the distribution of the prediction scores of the samples in the training set, combining the actual distribution of the sensory comprehensive scores and the industry standards.

[0025] The present invention has at least the following beneficial effects: First, this invention provides a complete and objective quality evaluation process for braised pork belly with preserved mustard greens, from data to decision-making. By combining initial screening using sensory contribution (OAV) with fine screening using model contribution (VIP) values, it can efficiently and accurately identify substances with key influences on the overall flavor from a large number of flavor compounds, avoiding interference from irrelevant variables. Subsequently, PCA is used to reduce the dimensionality of all screened key indicators, effectively eliminating collinearity and information redundancy among indicators, providing concise and information-rich input for subsequent modeling. Finally, the PLSR model accurately correlates objective indicators with sensory scores, realizing sensory quality prediction based on instrument data. This frees quality grading from absolute dependence on manual sensory evaluation, improving the objectivity, consistency, and efficiency of the evaluation, and providing a reliable technical foundation for industrial grading.

[0026] Secondly, by simultaneously calculating the principal component composite score F during principal component analysis, an objective reference dimension based on the statistical information of all key indicators is provided for the grading decision-making, independent of the prediction model. When the model prediction score of a sample falls within the grade boundary range where misjudgment is likely, the F score is introduced for verification or adjustment, essentially adding a "verification procedure" to the grading process. This effectively improves the grading system's ability to handle critical samples, reduces grade misjudgments caused by minor fluctuations in model predictions, and makes the final grading results more robust and reliable.

[0027] Third, by grouping samples according to storage conditions (frozen / room temperature) and conducting independent principal component analysis and model training, the entire evaluation system can adapt to the changing characteristics of braised pork belly with preserved mustard greens under different storage environments. This grouped modeling strategy allows the models constructed for the frozen and room temperature groups to focus on learning the mapping relationship between quality indicators and sensory scores under their specific conditions, thereby improving the model's prediction accuracy for samples in their respective groups. This solves the problem of poor adaptability of general models and ensures that this method maintains high accuracy in evaluation under different actual storage and distribution scenarios.

[0028] Fourth, a clearly defined PLSR model training and validation process is established, including dataset partitioning, cross-validation optimization, and a dual acceptance criterion based on the prediction determination coefficient and root mean square error (RMSE). 2 With a RMSE >0.90 and RMSE <1.0, this process provides a standardized and actionable guide for building high-performance, highly reliable prediction models. This workflow effectively prevents overfitting and underfitting during model training, ensuring that the obtained model not only performs well on the training set but, more importantly, exhibits stable and excellent predictive performance on the unknown test set. This guarantees the accuracy and generalization ability of the entire quality prediction process from the outset, laying a solid model foundation for subsequent grading.

[0029] Fifth, by employing a Bagging ensemble strategy to construct the final quality prediction model, the stability and generalization ability of the prediction are effectively improved. This method generates multiple training subsets through sampling with replacement, trains multiple PLSR sub-models, and then assigns weights to them based on out-of-bag sample performance evaluation before ensemble integration. This design effectively reduces the prediction variance (instability) caused by the randomness of training data in a single model, making the prediction results of the ensemble model smoother and more reliable. The weighted ensemble approach further strengthens the voice of the better-performing sub-model in the final decision, thus obtaining a more robust and interference-resistant quality predictor as a whole.

[0030] Sixth, by designing a dynamic fusion decision-making mechanism, the prediction results (Y) of the integrated model are creatively integrated. model The model intelligently weights and fuses the objective statistical score (F) to generate the final quality prediction score Y. Its core innovation lies in the fact that the weight λ is based on the model's predicted internal consensus (standard deviation σ). y Dynamic adjustment: When consensus is high, the model is fully trusted; when consensus is low, the model weight is reduced, and the weight of the objective score F is increased accordingly. This gives the method self-evaluation and correction capabilities. When faced with "difficult" samples where the model struggles to reach a consensus, it can automatically introduce another independent evaluation system for correction, thereby greatly improving the reliability and intelligence of the final prediction results, especially when dealing with abnormal or complex samples.

[0031] Seventh, this invention provides a set of specific, quantitative, and logically rigorous boundary verification operation rules. This is achieved by calculating the "standardized deviation" (D0) of the F-value of the boundary sample relative to the distribution of F-values ​​of two adjacent level references. A and D B ), and set clear judgment conditions (such as D) A ≤D B And F≥Favg A (If the score is adjusted to Level A), the review process, which could have been subjective, is transformed into a completely objective data calculation and comparison. This ensures the fairness, consistency, and automation of the boundary review, effectively utilizes the additional information provided by the F score, makes the determination of the level of critical samples more accurate and evidence-based, and further refines and improves the final stage of the hierarchical decision-making process.

[0032] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the method for evaluating and grading the quality of braised pork with preserved mustard greens according to the present invention. Figure 2Basic information on 8 commercially available pre-prepared dishes with preserved mustard greens and braised pork belly; Figure 3 The results of standardized treatment of physicochemical indicators of frozen preserved mustard greens with braised pork; Figure 4 The results of standardized treatment of physicochemical indicators of preserved mustard greens with braised pork at room temperature; Figure 5 A graph showing the eigenvalues ​​and cumulative contribution rate of the correlation matrix for the frozen group; Figure 6 The eigenvectors of the principal components of the frozen group's quality indicators; Figure 7 These are the eigenvalues ​​and cumulative contribution rates of the correlation matrix for the room temperature group. Figure 8 This represents the eigenvectors of the principal components of the quality indicators for the room temperature group. Detailed Implementation

[0034] The present invention will now be described in further detail so that those skilled in the art can implement it based on the description.

[0035] It should be understood that terms such as “having,” “comprising,” and “including” as used herein do not exclude the presence or addition of one or more other elements or combinations thereof.

[0036] like Figure 1 As shown, this invention provides a method for quality evaluation and grading of braised pork with preserved mustard greens, comprising the following steps: S1. Determine multiple quality indicators of the preserved mustard greens and braised pork samples to obtain raw data; the multiple quality indicators include texture profile analysis parameters, basic nutritional components and volatile flavor substances; S2. Select key quality indicators from the raw data, specifically including the following sub-steps: S21. For volatile flavor substances, calculate the odor activity value (OAV) of each substance, and select substances with OAV ≥ 1 to form the first flavor substance set; S22. Using the comprehensive sensory score as the dependent variable, and all substances in the first flavor substance set, as well as all texture profile analysis parameters and basic nutritional components measured in step S1, as independent variables, perform a first partial least squares regression analysis to calculate the variable importance projection value of each independent variable; S23. Select volatile flavor substances with variable importance projection values ​​greater than 1 from the first flavor substance set, and record them as key volatile flavor substances; The key quality indicators include: shear strength, hardness, adhesiveness, elasticity, cohesiveness, tackiness, chewiness, resilience, total amino acids, moisture, protein, fat, and the key volatile flavor substances; S3. Based on principal component analysis, perform dimensionality reduction on the data of the key quality indicators and extract N principal components with a cumulative variance contribution rate of more than 80% and an eigenvalue greater than 1. S4. Input the scores of the N principal components into the pre-trained partial least squares regression model for quality prediction to obtain the quality prediction score of the sample; the partial least squares regression model for quality prediction uses the principal component scores as input variables and the comprehensive sensory score as output variables, and is obtained through sample training. S5. Classify the quality grade of the preserved mustard greens and braised pork samples according to the quality prediction score.

[0037] In the above embodiments, this method integrates multi-source instrument detection data to establish an objective and quantifiable quality prediction and grading system, aiming to solve the problems of strong subjectivity, poor consistency, and difficulty in directly linking a single physicochemical indicator to comprehensive quality in traditional sensory evaluation.

[0038] First, various quality indicators of the preserved mustard greens and braised pork samples were measured to obtain raw data. These indicators mainly cover three categories: texture profile analysis parameters, basic nutritional components, and volatile flavor compounds. Texture profile analysis parameters are typically measured using a texture analyzer, reflecting the mechanical properties of the sample, such as hardness, elasticity, and cohesion. Basic nutritional components include moisture, protein, fat, and total amino acids, reflecting the basic nutritional composition of the sample. Volatile flavor compounds are identified and quantified using techniques such as gas chromatography-mass spectrometry (GC-MS), reflecting the flavor characteristics of the sample. This step is the data foundation construction stage, aiming to comprehensively collect multi-dimensional objective information reflecting the overall quality of the preserved mustard greens and braised pork, providing raw materials for subsequent analysis. In practice, to ensure data comparability, all samples must be subjected to standardized sampling sites, pretreatment methods, and instrument detection conditions. For example, texture profile analysis can use a cylindrical probe to simulate chewing twice at a fixed speed to obtain a series of TPA parameters; before flavor compound detection, pretreatment such as headspace solid-phase microextraction is required to enrich volatile components.

[0039] Next, key quality indicators (KPIs) are selected from the raw data to identify those with a critical impact on the final sensory quality. This selection process involves three sub-steps. First, for volatile flavor compounds, the odor activity value (OVA) is calculated for each compound. The OVA is the ratio of a compound's concentration to its sensory threshold, quantifying its actual contribution to the overall flavor. Typically, compounds with an OVA greater than or equal to 1 are initially identified as contributing to the flavor, forming the first set of flavor compounds. For example, dozens of volatile compounds may be detected in braised pork belly with preserved mustard greens, but through OVA screening, only about ten key flavor active compounds may be retained, significantly reducing noise in subsequent analyses. Second, using the overall sensory score assessed by a professional sensory evaluation team as the dependent variable, and all compounds in the first flavor compound set, along with all texture profile analysis parameters and basic nutritional components measured in the first step, as independent variables, a partial least squares (PLSR) regression analysis is performed. The PLSR model is a multivariate statistical method capable of handling multicollinearity data and building predictive models. The primary objective of this analysis is not to build a final predictive model, but rather to calculate the Variable Importance Projection (VIP) values ​​for each independent variable. The VIP value measures the importance of each independent variable in explaining changes in the dependent variable; a higher VIP value indicates that the variable is more critical to predicting sensory scores. The third step involves screening volatile flavor compounds with VIP values ​​greater than 1 from the first set of flavor compounds, designating them as key volatile flavor compounds as defined in this method. Thus, after two rounds of screening (OAV initial screening and VIP value fine screening), the final set of key quality indicators includes: texture parameters such as shear strength, hardness, adhesiveness, elasticity, cohesion, stickiness, chewiness, and resilience; basic nutritional indicators such as total amino acids, moisture, protein, and fat; and the screened key volatile flavor compounds. The core of this step lies in data dimensionality reduction and feature selection, which removes redundant and irrelevant variables, focusing on the core indicator set that truly affects quality, laying a clear and efficient data foundation for subsequent modeling.

[0040] Then, the data matrix of the selected key quality indicators is dimensionality reduced using principal component analysis (PCA). Since there may still be some correlation between key indicators, directly using them for modeling could lead to information overlap and excessive model complexity. PCA, through orthogonal transformation, converts multiple potentially correlated variables into a few independent principal components, which can retain the information of the original data to the greatest extent. In this step, N principal components with a cumulative variance contribution rate exceeding 80% and an eigenvalue greater than 1 are extracted. A cumulative variance contribution rate exceeding 80% means that these N principal components can explain more than 80% of the variation information in the original key indicator data, while an eigenvalue greater than 1 is a commonly used criterion for principal component retention. For example, in practical applications, extracting 3 or 4 principal components may suffice. After PCA transformation, each sample receives a score on these N principal components, which serves as the input to the subsequent prediction model. This step further simplifies the data structure, eliminates collinearity among indicators, and condenses multidimensional information into a few comprehensive variables, improving the robustness and computational efficiency of the subsequent model.

[0041] Subsequently, the scores of the N principal components are input into a pre-trained partial least squares regression model for quality prediction to obtain the predicted quality score of the sample. The quality prediction partial least squares regression model uses the principal component scores as input variables and the overall sensory score as the output variable, and is trained using a sufficient number of samples with known sensory scores. Model training is an independent and crucial step, as it establishes the mathematical mapping relationship between objective principal component scores and subjective sensory scores. A well-trained model should be able to accurately predict the sensory score of unknown samples. In practical deployment, for a new sample to be tested, only the aforementioned steps to obtain its principal component scores are needed, which can then be input into the model to instantly obtain its predicted quality score, achieving a rapid and objective conversion from instrument data to sensory evaluation.

[0042] Finally, the quality grades of the preserved mustard greens and braised pork samples are determined based on the calculated quality prediction scores. This requires pre-setting thresholds for grading. For example, the quality prediction scores can be sorted from high to low, and the grading thresholds can be determined by combining industry conventions (such as excellent, medium, and qualified grades) and the actual distribution of sample scores. If the predicted score is higher than the first threshold, it is classified as excellent; if it is between the first and second thresholds, it is classified as medium; and if it is lower than the second threshold, it is classified as qualified. In this way, continuous prediction scores are mapped to discrete quality grades, thereby achieving automated and standardized grading of large-volume products.

[0043] In summary, this embodiment constructs a systematic objective evaluation system for braised pork belly with preserved mustard greens by organically combining multiple technical steps such as OAV screening, VIP value screening, PCA dimensionality reduction, PLSR modeling, and threshold grading. Compared with existing technologies, this method achieves significant technical effects. First, it completely changes the excessive reliance of traditional quality control on human sensory evaluation. By using instrument detection and mathematical models, it achieves objectivity and standardization of the evaluation process, significantly improving the consistency and repeatability of evaluation results and avoiding biases caused by subjective differences among evaluators. Second, through scientific data screening and dimensionality reduction methods, this method accurately extracts core quality driving factors from massive amounts of raw detection indicators and establishes a robust predictive model between them and the final sensory experience, solving the technical problem that multidimensional heterogeneous data cannot be directly used for quality judgment. Finally, the entire process can be automated or semi-automated, greatly improving the efficiency of rapid quality detection and grading of braised pork belly with preserved mustard greens on the production line. It provides a feasible technical path for large-scale, intelligent food quality control, and has substantial improvements in efficiency, objectivity, and consistency compared to traditional methods.

[0044] In one specific embodiment, step S3 further includes: obtaining the principal component comprehensive score F for each sample based on principal component analysis; wherein the principal component comprehensive score F is obtained by weighting and summing the scores of each principal component using the proportion of the eigenvalue of each principal component to the sum of the extracted principal component eigenvalues ​​as the weight. In step S5, when the quality prediction score is within the preset grade division boundary range, the grade of the sample is reviewed or adjusted with reference to the principal component comprehensive score F.

[0045] In the above implementation, after completing Principal Component Analysis (PCA) and extracting N principal components, not only are the scores of each sample on the principal components obtained, but also a comprehensive score is calculated using the eigenvalues ​​of these principal components (representing the ability of each principal component to explain the variation in the original data). Specifically, the proportion of each principal component's eigenvalue to the sum of the extracted principal component eigenvalues ​​is used as the weight of that principal component's score. Then, the scores of the sample on each principal component are weighted and summed to obtain the sample's comprehensive principal component score F. This F value is essentially a single comprehensive statistic that compresses and integrates information from multiple key quality indicators. It is independent of the subsequent prediction model and purely reflects the sample's relative position in the multivariate space composed of all key indicators. For example, the first principal component with the largest eigenvalue is usually assigned the largest weight because it carries the most variation in the original information.

[0046] For the quality prediction score, a "preset grade division boundary interval" is also set. This interval is not a rigid score line, but a small range extending to both sides of the theoretical grade boundary score (such as the score dividing line between excellent and intermediate grades). For example, if the boundary score is 80 points, then this boundary interval might be [78, 82]. The width of this interval can be determined based on the performance volatility of the prediction model on historical data, such as taking a multiple of the standard deviation of the predicted score. A reasonable reference range could be 0.4 to 0.6 times the standard deviation, with an optimal value of 0.5 times. When the quality prediction score of a sample falls exactly within this boundary interval, it means that it is in a "potentially up or down" critical state, and relying solely on the model's predicted score for grading may carry the risk of misjudgment.

[0047] When the predicted quality score falls within the preset grade division boundary range, the system will no longer rely solely on the prediction model's results, but will instead "refer" to the sample's principal component composite score F for grade review or adjustment. The review logic is as follows: the F score provides an objective perspective directly synthesized from the original key indicators, without being "translated" by the model. If the F score of a critical sample significantly skews towards the typical distribution area of ​​adjacent higher-grade samples' F scores, its grade may be upgraded; conversely, it may be downgraded. This adds a verification procedure based on the essential characteristics of the original data to the grading decision, making the judgment of samples near the boundary more prudent and reliable.

[0048] In one specific embodiment, step S3, the analysis based on principal component analysis, specifically involves: dividing the samples into a frozen group and a room temperature group according to storage conditions; independently performing principal component analysis on the key quality index data of the two groups of samples, and calculating the principal component comprehensive score F of each group of samples. In steps S4 and S5, for the samples in the frozen group and the room temperature group, the principal component scores extracted from the corresponding group and the independently trained partial least squares regression model are used for prediction and classification.

[0049] The above implementation further refines how to address the impact of different storage conditions on the quality of braised pork with preserved mustard greens. The core of this approach is to employ a "group modeling" strategy to adapt to data heterogeneity. Before analysis, the braised pork with preserved mustard greens samples need to be clearly divided into different groups based on their storage conditions. A typical example is dividing them into a "frozen group" and a "room temperature group." This is because freezing and room temperature storage significantly affect the texture of meat products (e.g., changes in texture due to juice loss after thawing), the volatility of flavor compounds, and the stability of certain nutrients. Mixing these samples with vastly different conditions for analysis would obscure their unique quality evolution patterns, leading to impure extracted comprehensive characteristics and consequently affecting the prediction accuracy of subsequent models.

[0050] For the pre-defined frozen and room temperature groups, principal component analysis was performed independently on the sample data within each group. This means that a dedicated set of principal components (including eigenvalues, scores, and overall score F) was calculated for the frozen group data, and a different set of principal components was calculated for the room temperature group data. Similarly, the overall principal component score F was calculated for each group of samples. The resulting F value is a comprehensive indicator that is comparable within the sample population under its specific storage conditions, and better reflects the quality under those conditions.

[0051] In step S4 for quality prediction and step S5 for grading, a single general model is no longer used; instead, a strict grouping principle is followed. For the frozen group samples to be tested, the principal component scores extracted from the frozen group data, along with a partial least squares regression model specifically trained on frozen group samples, are used for prediction and grading. For the room temperature group samples, the corresponding room temperature group principal component scores and model are used. This "matching" approach ensures that the patterns learned by the model perfectly match the conditions of the samples being tested.

[0052] This implementation significantly improves the adaptability and predictive accuracy of the evaluation system across various practical application scenarios. By identifying and isolating storage conditions—a key source of variation—and establishing targeted analytical models for different conditions, it effectively overcomes the performance degradation problem caused by the inherent data heterogeneity of "one-size-fits-all" general models. This allows the method to maintain high-precision quality evaluation capabilities for both cold chain products and products stored at ambient temperature, thereby expanding the method's applicability and practical value.

[0053] In one specific implementation, step S4, the training process of the pre-trained partial least squares regression model includes: The sample dataset is divided into training and test sets according to a certain ratio; Cross-validation is used on the training set to select the optimal number of latent variables for the model; A partial least squares regression model was established and validated on the test set. The partial least squares regression model was successfully trained when the prediction determination coefficient was greater than 0.90 and the prediction root mean square error was less than 1.0.

[0054] In the above implementation, the focus is on the specific and standardized training process for constructing the quality prediction partial least squares regression model, which aims to ensure that the final model has excellent performance and reliability and avoids overfitting or underfitting.

[0055] For the data preparation stage of model training, in order to objectively evaluate the model's generalization ability (i.e., its predictive ability on new samples), not all existing samples can be used for training. Therefore, the complete dataset of collected samples with sensory comprehensive rating labels needs to be divided into two parts according to a certain ratio: a training set and a test set. The training set is used for "learning" and building the model, while the test set is used for the final "exam" to validate the model. Common division ratios are 7:3 or 8:2, for example, randomly allocating 70% of the samples to the training set and the remaining 30% to the test set is a commonly used and effective practice.

[0056] Partial least squares regression models have a key hyperparameter—the number of latent variables. This parameter controls the model's complexity; too few latent variables may lead to the model failing to fully learn the patterns in the data (underfitting), while too many may even learn the random noise in the data (overfitting). This implementation uses cross-validation to automatically select the optimal number of latent variables within the training set. Cross-validation is a robust parameter optimization technique, such as the commonly used five-fold cross-validation. This process automatically finds the best balance between model complexity and predictive power.

[0057] After building a PLSR model on the full training set with the selected optimal parameters, it must be applied to a test set that has never been used for training or parameter optimization to verify its performance. This implementation sets two explicit performance thresholds: a prediction determination coefficient (R²) greater than 0.90 and a prediction root mean square error (RMSE) less than 1.0. R² measures the goodness of fit between the model's predicted values ​​and the true values; the closer to 1, the better. RMSE measures the average error magnitude of the predictions. Only when the model simultaneously meets both of these stringent conditions on an independent test set is it considered successfully trained and ready for practical use. These two thresholds together ensure that the model has extremely high prediction accuracy and stability.

[0058] In one specific implementation, for the frozen sample, the top three principal components with a cumulative variance contribution rate of 100% are extracted, and the formula for calculating the principal component comprehensive score F is: F = 0.554 × F1 + 0.303 × F2 + 0.143 × F3; For the samples in the room temperature group, the top three principal components with a cumulative variance contribution rate of 100% are extracted, and the formula for calculating their principal component composite score F is: F = 0.528 × F1 + 0.246 × F2 + 0.226 × F3; F1, F2, and F3 represent the scores of the first, second, and third principal components of the corresponding groups, respectively.

[0059] In the above embodiments, the appropriate number of principal components to extract during principal component analysis is crucial for samples from specific storage groups. The claims specify that for both frozen and room-temperature samples, the "top three principal components with a cumulative variance contribution rate of 100%" should be extracted. This means that only three principal components are needed to fully explain all the variation (variance) of all key quality indicator data within each group. This is a highly efficient and thorough dimensionality reduction result, demonstrating that within the framework of grouped modeling, the inherent structure of each group of data is very clear and can be perfectly summarized by a very small number of comprehensive variables. This provides an extremely concise and complete input for subsequent modeling.

[0060] The specific weighting coefficients for calculating the overall principal component score F are as follows: For frozen samples, the F-value is obtained by weighted summation of the scores of the first, second, and third principal components (F1, F2, F3), with weights of 0.554, 0.303, and 0.143, respectively. For samples stored at room temperature, the weights are 0.528, 0.246, and 0.226, respectively. These weights are precisely the ratios of the eigenvalues ​​of each principal component to the sum of the eigenvalues ​​of the three principal components. For example, the first principal component in the frozen group explains the largest portion of the variation (55.4%), and therefore has the greatest weight in the overall score. These specific coefficients are determined values ​​calculated through principal component analysis of actual experimental data, rather than theoretical assumptions. They make the calculation of the F-score completely standardized and operational, and these weights themselves reflect the differences in the relative importance of various quality indicators (such as texture, flavor, and nutrition) in constituting the overall quality under different storage conditions.

[0061] With a clearly defined number of principal components (3) and precise weighting coefficients, in practical applications, for any test sample belonging to a specific group, as long as its scores on the three principal components of the corresponding group are calculated, its principal component composite score F can be quickly and unambiguously calculated according to the given formula. This eliminates arbitrariness in the calculation process and ensures the consistency of calculation results across different batches and by different operators.

[0062] In one specific implementation, in step S4, the pre-trained partial least squares regression model is a Bagging ensemble-based model, which is trained using the following method: M1. Set the integration size K, K≥3; Perform K random samplings with replacement from the original training sample set that determines the key quality indicators. Each sampling has the same size as the original set, generating K subsets of training sets. Samples that are not selected form the corresponding K out-of-bag sample sets. M2. For the k-th subset of training, using the principal component scores extracted in step S3 as input and the comprehensive sensory score as output, independently train a partial least squares regression sub-model M. k Where k=1,...,K; M3, for the kth sub-model M k Using its out-of-bag sample set for prediction, the out-of-bag prediction determination coefficient R is calculated. 2 oob,k According to formula ω k = R 2 oob,k / Σ(R 2 oob,i ), calculate its weight coefficients, where i = 1 to K; M4. Save all K sub-models {M1, …, M} K} and its weight coefficients {ω1, …,ω K Together, they constitute the Bagging ensemble model.

[0063] In the above implementation, the pre-trained partial least squares regression model is defined as a Bagging ensemble model. Bagging ensemble models aim to improve the stability and accuracy of predictions through group decision-making. The core idea of ​​Bagging is "bootshoe resampling." An ensemble size K is set, with a minimum of 3. Then, from the original, complete training sample set with determined key quality indicators, K random samplings with replacement are performed. Each sampling draws the same number of samples as the original set (due to replacement, some samples may be drawn repeatedly, while others may not be drawn). This generates K sub-training sets that differ slightly in content. Simultaneously, the samples not drawn in each sampling naturally form a corresponding "out-of-bag sample set," which will be used to subsequently evaluate the performance of its corresponding sub-model.

[0064] For each of the generated sub-training sets, a complete partial least squares regression sub-model is trained independently. The training process includes calculating principal component scores using the sub-training set (based on the parameters of its storage group), cross-validating to select the optimal number of latent variables, and building the model. Ultimately, this results in K functionally identical PLSR sub-models with slightly different internal parameters due to minor differences in the training data.

[0065] For the k-th trained sub-model, instead of using its training data for evaluation, its corresponding out-of-bag sample set is used for prediction. The coefficient of determination of this prediction result is calculated and denoted as the out-of-bag prediction determination coefficient R. 2 oob,k This coefficient objectively reflects the predictive ability of the sub-model for "unseen" data. Then, the R-squared values ​​of all K sub-models are calculated. 2 oob,k Sum the values, using the R-values ​​of each sub-model. 2 oob,k Dividing the value by this sum yields the weight coefficient ω of the sub-model.k The better the performance of the sub-model, the higher its weight. Finally, these K weighted sub-models are saved, and together they constitute the final Bagging ensemble model.

[0066] This implementation method effectively reduces the prediction variance (i.e., instability) caused by the randomness of training data in a single model by constructing multiple models and performing weighted ensemble. Even if a sub-model is biased due to interference from specific samples, other sub-models and the weighting mechanism can correct it. The ensemble model exhibits stronger anti-interference ability and more stable prediction output when facing complex and variable data, making the final classification basis more reliable.

[0067] In one specific implementation, when applied to a particular storage group, the final quality prediction is performed according to the following process: A. For the sample to be tested, call the trained Bagging ensemble model of its corresponding group to obtain the predicted values ​​{y1, …, y} of the K sub-models. K}; B. Calculate the initial model prediction score Y for this sample. model =Σ(ω k ×y k ), where ω k These are the weight coefficients for the corresponding sub-models; simultaneously, the standard deviation σ of these K predicted values ​​is calculated. y ; C. Call the principal component comprehensive score F of the sample; the principal component comprehensive score F is calculated using the predetermined principal component weighting coefficients of the group, wherein the weighting coefficients for the frozen group are (0.554, 0.303, 0.143) and for the room temperature group are (0.528, 0.246, 0.226); D. Perform dynamic fusion decision-making and calculate the final quality prediction score Y: Y = λ × Y model + (1 - λ)× F; where the fusion weight λ is based on the standard deviation σ. y Dynamically determined, the range of λ is (0, 1]: if σ y If ≤ δ, then λ = 1.0; if σ y If δ > 0, then λ = γ / σ y The discreteness threshold δ and the adjustment parameter γ are determined by analyzing the σ values ​​of the samples in the training set. y The correlation between the predicted residuals and the predicted residuals is optimized and determined.

[0068] In the above implementation, based on the Bagging ensemble model and the principal component comprehensive score F, an intelligent dynamic fusion decision mechanism is designed to generate the final quality prediction score Y. Its core is to dynamically adjust the degree of trust in different information sources according to the consensus within the model.

[0069] For a new test sample, the pre-trained Bagging ensemble model corresponding to its storage group is first invoked. The principal component score of this sample is then input into all K sub-models, resulting in K slightly different predicted values ​​y1 to y2. K Next, based on the weights ω of each sub-model... k Calculate the weighted average of these predictions to obtain the "initial prediction score" Y. model At the same time, calculate the standard deviation σ of these K predicted values. y This standard deviation σ y σ is a key metric that quantifies the dispersion of predictions made by all sub-models within an ensemble model for a given sample; it represents the "internal consensus." y The smaller the value, the higher the consensus among all sub-models; σ y The larger the value, the more significant the discrepancy between the sub-models.

[0070] The system will simultaneously call the principal component composite score F calculated independently for that sample. The final quality prediction score Y is not simply a matter of taking Y. model Instead, Y model A weighted fusion result with F: Y = λ × Y model + (1 - λ) × F. Where the fusion weight λ is not a fixed value, but rather depends on the consensus degree σ. y Dynamically determined. The rule is as follows: Set a dispersion threshold δ. If σ y If ≤ δ, it indicates a high degree of consensus within the model. In this case, the model can be fully trusted. Let λ = 1.0, i.e., Y = Y model If σ y A value greater than δ indicates significant internal discrepancies within the model. In this case, the confidence level in the model should be reduced, and the value of λ becomes γ / σ. y (Where γ is an adjustment parameter). This means σ y The larger the dispersion threshold (the greater the discrepancy), the smaller λ becomes, and the greater the weight (1-λ) of the objective statistical score F in the final score Y. The dispersion threshold δ and the adjustment parameter γ are determined by optimizing the prediction results of all samples in the training set. The specific method is as follows: First, the trained Bagging ensemble model is used to predict the training set samples, and the standard deviation σ of the sub-model prediction for each sample is calculated. y and its model prediction residuals; subsequently, by analyzing σ yThe correlation with the predicted residuals (e.g., plotting a scatter plot or calculating different σ) y The average residual of the interval), δ is set as the critical value that can distinguish between the "low residual-low dispersion" and "high residual-high dispersion" groups. It can usually be taken as the training set sample σ. y The 10%–30% quantile of the distribution, for example, δ = 0.15 in one implementation. Then, at σ... y Within the sample range of >δ, a grid search method is used to find the optimal value of γ within a reasonable interval (e.g., γ ∈ [0.1, 1.0]), such that according to the formula λ = γ / σ y and Y = λ×Y model The root mean square error of the final prediction score Y calculated by + (1-λ)×F is minimized, resulting in γ=0.18 in one optimization. Finally, the effect of the selected parameters on the prediction performance needs to be verified on the test set to ensure the effectiveness and robustness of the dynamic fusion mechanism.

[0071] When the ensemble model is highly confident in its prediction of a particular sample (high consensus), the model's judgment is adopted. When the model itself is hesitant or the results of the sub-models differ significantly, the system automatically recognizes that relying solely on the model's results is risky. Therefore, it switches to referencing another completely independent, objective comprehensive index F calculated directly from the original data for correction. This allows the system to output a more prudent and reliable prediction when facing easily confused or anomalous "difficult" samples.

[0072] In one specific embodiment, step S5 involves reviewing or adjusting the sample's grade based on the principal component comprehensive score F, specifically including the following sub-steps: S51. When the predicted quality score Y of a sample is within the boundary interval between adjacent high-grade A and low-grade B, obtain the principal component composite score F of the sample, and at the same time obtain the average principal component composite score Favg of the predetermined grade A reference sample. A With standard deviation σ A And the average principal component composite score (Favg) of the B-level reference sample. B With standard deviation σ B ; S52. Calculate the standardized deviation D of the overall principal component score F of the sample relative to grade A. A The calculation formula is: D A =|F - Favg A | / σ A , where σ A The standard deviation of the F-values ​​for the reference sample at level A is given; simultaneously, the standardized deviation D relative to level B is calculated. B The calculation formula is: D B = |F - FavgB | / σ B , where σ B The standard deviation of the F-values ​​for the B-level reference sample; S53. Determine the level: If D A Less than or equal to D B And simultaneously satisfy F greater than or equal to Favg A If the sample grade is D, then the grade should be adjusted upwards or confirmed as grade A; if D B Less than D A And simultaneously satisfy F less than or equal to Favg B If the condition is not met, the sample grade will be adjusted down or confirmed as grade B; if none of the above conditions are met, the initial grading result based on the quality prediction score Y will be maintained.

[0073] In the above implementation, specific, quantitative, and automatically executable judgment rules are provided for "reviewing or adjusting the grade based on the principal component comprehensive score F", which transforms the review of boundary samples from subjective judgment to objective calculation.

[0074] The review process begins when a sample's final predicted quality score Y falls within a pre-defined boundary between two adjacent grades (high grade A and low grade B). The system first obtains the sample's principal component composite score F. Simultaneously, it retrieves two key reference pieces of information from historical data: the average F-value (Favg) of all reference samples at grade A (i.e., typical samples explicitly classified as grade A). A and standard deviation σ A ; and the average F-value of the reference sample for grade B. B and standard deviation σ B These statistics can be calculated and saved from the training set during the model training phase, representing the typical distribution center (mean) and dispersion (standard deviation) of each level on the F-score dimension.

[0075] To determine which grade the boundary sample is more biased towards, calculate the degree of deviation (D) of its F-value relative to the typical distribution of grade A. A The formula is D A = |F - Favg A | / σ A This value indicates how many standard deviations of Grade A itself are far from the center of Grade A for the sample's F-value. Similarly, calculate the standardized deviation D relative to Grade B. B = |F - Favg B | / σ B The advantage of using "standardization" is that it eliminates the influence of different F-value fluctuation ranges among different levels, making the comparison more equitable. For example, a value 1.5 times σ away from the center of level A... AThe sample, and a sample 0.8 times σ away from the center of class B. B Although the absolute values ​​of the samples may be far apart, the latter are actually "closer" after standardization.

[0076] The judgment is based on two conditions: first, comparing D A and D B The relative magnitude of the F-value is checked, and secondly, the absolute position of the F-value relative to the rank center is examined. The specific rule is: if D... A ≤ D B (That is, the F value of this sample is closer to the typical distribution of grade A), and at the same time F ≥ Favg A (That is, if the sample's F value is not lower than the average level of Grade A), then the sample's grade will be adjusted upwards or confirmed as Grade A. If D B <D A (i.e., closer to grade B), and at the same time F ≤ Favg B (i.e., not higher than the average level of grade B), then the sample grade is adjusted down or confirmed as grade B. If neither of the above two conditions is met (e.g., although D...), the sample grade is adjusted down or confirmed as grade B. A Smaller, but F is lower than Favg A If the F-value is not clear, it indicates that the "attribution" intention of the sample is unclear, and the initial classification result based on the predicted score Y should be maintained. These conditions together ensure that the adjustment is prudent and well-founded.

[0077] This implementation thoroughly standardizes and objectifies the crucial but subjective step of boundary verification. By introducing the concept of standardized deviation and setting rigorous dual judgment conditions, this rule can automatically and consistently process all boundary samples, completely avoiding inconsistencies that may arise from human intervention. It cleverly utilizes the additional information dimension provided by the F-score, making the classification of critical samples no longer an ambiguous guess, but a rational decision based on explicit mathematical calculations, thereby further improving the accuracy and fairness of the entire classification system in the smallest details.

[0078] In one specific implementation, the preset grade division boundary interval in step S5 is: let the score boundary of the grade division be T, then the boundary interval is [T-β, T+β], where β is determined according to the distribution standard deviation of the predicted scores of the samples in the training set, and the value range is 0.3×ρ≤β≤0.7×ρ, where ρ is the standard deviation of the predicted scores.

[0079] In the above implementation, the grading is typically based on one or more defined score thresholds T. For example, the threshold for excellent and intermediate grades is T1, and the threshold for intermediate and acceptable grades is T2. The boundary interval is not these isolated points, but rather an interval formed by extending a width β to both sides of the threshold T, i.e., [T - β, T + β]. Any sample whose predicted score falls within this interval will be considered a boundary sample and trigger a possible review procedure. This design acknowledges that any prediction model has an inherent, small prediction error.

[0080] The setting of β is not arbitrary, but based on the predictive stability exhibited by the prediction model on the training set (or an independent validation set). Specifically, it is necessary to calculate the standard deviation ρ of the final quality prediction scores of all samples in the training set. This ρ value reflects the overall fluctuation range of the model's predicted scores. The value of β should be proportional to ρ, and the claims suggest a reasonable range: 0.3ρ ≤ β ≤ 0.7ρ. For example, if ρ = 2.0 points, then β can be chosen between 0.6 and 1.4 points. If the value of β is too small (e.g., less than 0.3ρ), the boundary interval is too narrow, which may not effectively capture samples that fall near the threshold due to normal model fluctuations, resulting in insufficient activation of the review mechanism; if the value of β is too large (e.g., greater than 0.7ρ), the interval is too wide, which will cause too many samples that should be clearly graded to be sent to the review process, reducing grading efficiency. A preferred balance value is 0.5ρ.

[0081] By linking β to the intrinsic volatility (ρ) of the model's predictions, the width of the boundary interval can adapt to the performance of different prediction models. For a model with very accurate predictions and low volatility (small ρ), its boundary interval automatically narrows; for a model with slightly higher volatility (large ρ), its boundary interval widens accordingly to accommodate more uncertainty. This makes the fault tolerance mechanism of the entire hierarchical system adaptive and scientific, ensuring secondary verification when needed while avoiding unnecessary verification overhead.

[0082] This implementation provides a data-driven method to rationally set the width of the "fuzzy zone," enabling the resource-intensive boundary verification step to be precisely located on the samples that truly need it. This avoids two extremes caused by improper interval settings: either the interval is too narrow, missing critical samples that need verification and affecting the accuracy of classification; or the interval is too wide, allowing a large number of samples to enter the verification process, slowing down the overall classification speed. Through scientific quantification, this method maintains high processing efficiency while ensuring the robustness of the classification results.

[0083] In one specific embodiment, the quality level in step S5 is divided into three levels, specifically: sort the quality prediction scores Y from high to low, and determine two segmentation thresholds T1 and T2, where T1>T2, corresponding to the boundaries between the excellent level and the intermediate level, and between the intermediate level and the qualified level respectively; if Y ≥ T1, it is determined as the excellent level; if T2 ≤ Y<T1, it is determined as the intermediate level; if Y<T2, it is determined as the qualified level; where the values of T1 and T2 are jointly determined by analyzing the distribution of the prediction scores of the samples in the training set, combining the actual distribution of the sensory comprehensive scores and the industry standards.

[0084] In the above embodiment, the three-level classification system is defined and how to determine the score thresholds between levels is clarified, completing the final mapping from continuous prediction scores to discrete quality levels. This method finally divides the quality of the braised pork with preserved vegetables into three clear levels, which can generally be named "excellent level", "intermediate level" and "qualified level" from high to low. This three-level classification conforms to the common quality level practices in the food industry and is also easy for production management and consumers to understand. The classification basis is the quality prediction score Y finally calculated for the samples. To achieve the three-level classification, two segmentation points, that is, the thresholds T1 and T2, need to be determined on the numerical axis of the prediction scores Y of all samples, and T1>T2. T1 is the boundary between the excellent level and the intermediate level, and T2 is the boundary between the intermediate level and the qualified level. For any sample to be classified, its level determination rule is very simple: if Y ≥ T1, it is determined as the excellent level; if T2 ≤ Y<T1, it is determined as the intermediate level; if Y<T2, it is determined as the qualified level. This set of rules is clear and non-overlapping, ensuring that each sample has and only has one level. The determination of the thresholds is not subjectively set, but based on the analysis of historical data and combined with actual requirements. The specific process is: after the model training is completed, use the prediction scores Y of all samples in the training set (or a larger dataset containing representative samples) to observe their distribution. At the same time, refer to the distribution of the true sensory comprehensive scores of these samples (for example, the approximate score ranges corresponding to "excellent", "medium" and "qualified" considered by sensory experts), as well as the industry general standards or enterprise internal control standards in this product field. By analyzing the distribution histogram and quantiles of the prediction score Y and comparing with the sensory scores and industry requirements, the positions of T1 and T2 can be reasonably determined. For example, T1 can be set near the critical point of the top 30% of the samples in the prediction score ranking, and T2 can be set near the critical point of the bottom 30% of the samples, but this needs to be fine-tuned according to the actual data distribution and standard requirements. The core goal is to make the classification results based on Y have the highest possible consistency with the classification results based on sensory scores and industry consensus. Specific embodiments: 1. Select 8 commercially available prefabricated dishes of braised pork with preserved vegetables, transport them back to the laboratory and store them temporarily according to the product storage conditions for later use. The basic information is shown in Figure 2Eight commercially available braised pork belly with preserved mustard greens dishes can be used as eight sample data sources or eight treatment groups. Each sample data source can contain 24 samples (three parallel samples, samples from different time points, and samples under different storage conditions).

[0086] II. Quality Inspection.

[0087] 2.1 Determination of moisture, fat, and protein: Moisture content was determined using the direct drying method as specified in GB 5009.3-2016 "Determination of Moisture in Food". Fat content was determined using the Soxhlet extraction method as specified in GB 5009.6-2016 "Determination of Fat in Food". Protein content was determined using the Kjeldahl method as specified in GB 5009.5-2016 "Determination of Protein in Food".

[0088] 2.2 Texture Properties Determination: The thawed sample was cut into 1 cm × 1 cm × 0.5 cm pieces along the muscle fiber direction. The texture of the sample was determined using a texture analyzer. A cylindrical probe (P / 50) was selected, and the initial speed was 1.00 mm•s. -1 The speed during the test was 5.00 mm•s. -1 The speed after testing was 5.00 mm•s. -1 The strain is 75%, and the automatic (force) is 0.049 N.

[0089] 2.3 Determination of Free Amino Acids: Sample Pretreatment: Weigh 1 g (accurate to 0.001 g) of homogenized preserved mustard greens with braised pork sample and place it in a 50 mL centrifuge tube. Add 50 mL of pre-cooled 0.01 mol•L⁻¹ ammonium chloride solution. -1 Hydrochloric acid solution. The mixture was placed in an ultrasonic extractor (300 W, 40 kHz) and extracted at 25°C for 30 min. After extraction, it was allowed to stand for 10 min to allow solid-liquid separation. 2 mL of the supernatant was transferred to a 5 mL centrifuge tube, and an equal volume (2 mL) of 8% (w / v) sulfosalicylic acid solution was added. The mixture was vortexed for 1 min to mix thoroughly. The mixture was then incubated at 4°C and 10000 r•min. -1 Centrifuge for 10 min under the specified conditions. Filter the supernatant through a 0.22 μm aqueous microporous membrane (polyethersulfone) and transfer it to a 2 mL vial for analysis using liquid chromatography-mass spectrometry (LC-MS). LC conditions: SeQuant® ZIC®-HILIC column, column temperature 40℃, injection volume 2 μL, flow rate 0.4 mL / min. -1 Mobile phase A: 0.1% formic acid aqueous solution, mobile phase B: 0.1% formic acid acetonitrile solution.

[0090] 2.4 Determination of volatile flavor compounds: Weigh 3 g (accurate to 0.001 g) of sample into a headspace vial, add 1.5 μL of 2-methyl-3-heptanone (1.68 μg·mL⁻¹). -1 After mixing thoroughly, the sample is ready for testing.

[0091] HS-SPME conditions: Sample preheating temperature 60 ℃, preheating time 20 min, extraction temperature 60 ℃, extraction time 40 min.

[0092] GC conditions: DB-WAX (30 m × 0.25 mm × 0.25 μm, Agilent Technologies) capillary column; helium (He) flow rate 1.0 mL / min, injection volume 1.0 μL. Splitless injection. Temperature program: initial column temperature 40 ℃, hold for 3 min, then increase at 2 ℃·min. -1 Raise to 70 °C, then increase by 4 °C·min -1 Rise to 130 °C, and finally reduce to 10 °C·min. -1 Raise the temperature to 230 °C and maintain for 10 minutes.

[0093] MS conditions: ion source temperature 230℃, quadrupole temperature 150℃, EI ionization 70 eV, acquisition mode: scanning mode, scan quality range: 50–550 m·z -1 .

[0094] Qualitative analysis of volatile compounds: The retention index (RI) calculated from the retention times of the molecular ion peak and n-alkanes (C7-C40) in the mass spectrometry database (NIST 20.0) of the compound to be tested is compared with that of the volatile compounds to be determined.

[0095] Quantitative analysis of volatile compounds: The semi-quantitative concentration (mg·kg) is calculated based on the peak area ratio of the volatile compound to an internal standard of known concentration. -1 ).

[0096] Odor activity value (OAV) calculation: OAV characterizes the contribution of each aroma component to the total aroma of food, and is calculated as follows: OAV = C / T. OAV ≥ 1 indicates that the flavor substance contributes significantly to the overall flavor, where C is the absolute concentration of the odor substance and T is the sensory threshold.

[0097] 2.5 To eliminate differences in units and orders of magnitude among the various indicators, data standardization is required before principal component analysis. A certain correlation between the indicators is a prerequisite for using principal component analysis. Therefore, X1-X16 represent shear force, hardness, adhesiveness, elasticity, cohesion, viscousity, chewiness, resilience, total amino acids, moisture, protein, fat, styrene, eugenol, ethyl maltol, and 2-pentyl-furan (eight volatile flavor compounds common to commercially available braised pork belly with preserved mustard greens and OAV and VIP > 1).

[0098] Figure 3 The results of standardized treatment of physicochemical indicators of frozen preserved mustard greens with braised pork; Figure 4 The results of standardized treatment of physicochemical indicators of preserved mustard greens with braised pork at room temperature.

[0099] Sixteen indicators measured in the frozen and room temperature groups of braised pork with preserved mustard greens were used to form 4 × 16 matrices. Principal component analysis was performed using SPSS 27.0 software to obtain the eigenvalues ​​and variance contribution rates of each component. Figure 5 The eigenvalues ​​and cumulative contribution rates of the correlation matrix of the frozen group are used to determine the number of principal components based on the principle that the eigenvalues ​​are greater than 1 and the cumulative variance contribution rate exceeds 80%. Figure 6 The eigenvectors are the principal components of the frozen group's quality indicators.

[0100] The eigenvector equation of principal component 1 is: F1=-0.107X1+0.076X2-0.065X3+0.062X4+0.070X5+0.101X6+0.081X7+0.100X8-0.107 X9-0.071X 10 -0.108 X 11 +0.074 X 12 -0.112 X 13 +0.056X 14 +0.071X 15 -0.036X 16 The eigenvector equation of principal component 2 is: F2=0.054X1-0.139X2+0.166X3+0.168X4+0.161X5+0.088X6+0.134X7+0.065X8-0.019X9-0.027X 10 +0.057X 11 -0.014X 12 +0.003X 13 -0.178X 14 -0.155X 15 -0.110X16 The eigenvector equation of principal component 3 is: F3=-0.069X1+0.129X2+0.062X3+0.086X4-0.015X5+0.065X6+0.104X7+0.152X8-0.129X9+0.335X 10 +0.029X 11 -0.328X 12 -0.060X 13 +0.027X 14 -0.085X 15 +0.343X 16 .

[0101] Since the first three components already provide 100% of the original data information for the frozen group, these three principal components can be used to replace the original multiple complex indicators for model building. The weight of each principal component is determined by the proportion of its eigenvalue to the total extracted principal component eigenvalues. Based on this, a comprehensive evaluation model F can be established, with the following equation: F = 0.554F1 + 0.303F2 + 0.143F3.

[0102] Figure 7 These are the eigenvalues ​​and cumulative contribution rates of the correlation matrix for the room temperature group. Figure 8 This represents the eigenvectors of the principal components of the quality indicators for the room temperature group.

[0103] The eigenvector equation of principal component 1 is: F1=0.113X1-0.098X2-0.024X3+0.043X4+0.098X5+0.110X6+0.104X7+0.000X8-0.074X9+0.017X 10 +0.108 X 11 -0.091 X 12 +0.115 X 13 +0.113X 14 +0.092X 15 +0.031X 16 The eigenvector equation of principal component 2 is: F2=-0.074X1+0.141X2+0.05X3-0.118X4+0.004X5+0.066X6+0.009X7+0.247X8-0.162X9-0.137X 10 -0.103X 11 +0.134X 12 -0.059X 13-0.072X 14 +0.156X 15 +0.201X 16 The eigenvector equation of principal component 3 is: F3=-0.16X1-0.017X2+0.266X3+0.225X4+0.155X5+0.068X6+0.133X7+0.065X8+0.126X9-0.230X 10 -0.023X 11 +0.100X 12 +0.000X 13 -0.028X 14 -0.038X 15 -0.153X 16 Since the first three components already provide 100% of the original data information for the room temperature group, these three principal components can be used to replace the original multiple complex indicators for model building. The weight of each principal component is determined by the proportion of its eigenvalue to the sum of all extracted principal component eigenvalues. Based on this, a comprehensive evaluation model F can be established, with the following equation: F = 0.528F1 + 0.246F2 + 0.226F3.

[0104] 2.6 PLSR analysis was performed on the frozen group and the room temperature group respectively.

[0105] For the frozen group (or room temperature group), there are 96 samples in each group.

[0106] Sample splitting: In the model training part, the 96 samples in this group were randomly divided into a training set (about 67 samples) and a test set (about 29 samples) at a ratio of about 70%.

[0107] Model Validation: Based on this partition, perform cross-validation, modeling, and independent test set validation, and report R² and RMSEP. The sample size at this point is sufficient to support the conclusion that "R² > 0.90, RMSE < 1.0".

[0108] Bagging ensemble: It can successfully perform sampling with replacement on a training set of 67 samples to generate a meaningful sub-training set and out-of-bag sample set.

[0109] Threshold Determination: The final grading thresholds T1 and T2 need to be determined collaboratively through the following steps, taking into account the statistical distribution of the model's predicted scores (Y values), the actual grade assignment of the comprehensive sensory evaluation, and relevant industry standards or enterprise internal control standards: Establish a grade reference label for the training set samples: A trained sensory evaluation team conducts blind evaluations of all samples in the training set and assigns a clear sensory grade label (excellent, average, or qualified) to each sample according to a predetermined scoring standard (e.g., 9-point system, ≥7 points is excellent, 5-7 points is average, and <5 points is acceptable).

[0110] Obtain the final prediction score Y for the training set samples: Use the trained model (including dynamic fusion decision) to perform back-substitution prediction on all training set samples to obtain the final quality prediction score Y for each sample.

[0111] Analyze the distribution of Y values ​​across different sensory levels. Based on this distribution analysis, find the Y value split point that maximizes the distinction between adjacent levels. For example, it can be set as follows: T1 (Excellent / Intermediate Boundary): Set in the boundary region between the Y-value distributions of "Excellent" and "Intermediate" samples, such that the Y-value of the vast majority of "Excellent" samples is ≥ T1.

[0112] T2 (Medium / Qualified Level Boundary): Set in the boundary region between the Y-value distribution of "Medium" and "Qualified" samples, such that the Y-value of the vast majority of "Medium" samples is ≥ T2.

[0113] Validation and Fine-tuning: The initially set T1 and T2 values ​​were applied to independent test sets to check the consistency between the grading results based on Y values ​​and the sensory evaluation results (e.g., calculation accuracy, Kappa coefficient). Simultaneously, referring to the descriptions of product grades in industry standards such as GB / T 22210-2008 Sensory Evaluation Specification for Meat and Meat Products, the thresholds were rationally assessed and fine-tuned as necessary to ensure that the grading results conformed to both data patterns and industry consensus.

[0114] Example: In a specific implementation, for a batch of frozen preserved mustard greens with braised pork training samples (n=67), the Y-value distribution was as follows: the Y-value range for excellent samples (sensory evaluation) was [8.2, 9.5], for medium-grade samples it was [6.5, 8.3], and for qualified samples it was [4.0, 6.6]. After analysis, T1 = 8.0 and T2 = 6.5 were set. At these thresholds, the classification accuracy within the training set reached 94.0%, and the classification accuracy for the test set (n=29) reached 91.4%, with good consistency between the classification results and the industry expert review results.

[0115] The number of devices and processing scale described herein are for the purpose of simplifying the description of the invention. Applications, modifications, and variations of the invention will be readily apparent to those skilled in the art.

[0116] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details.

Claims

1. A method for evaluating and grading the quality of braised pork with preserved mustard greens, characterized in that, Includes the following steps: S1. Determine multiple quality indicators of the preserved mustard greens and braised pork samples to obtain raw data; the multiple quality indicators include texture profile analysis parameters, basic nutritional components and volatile flavor substances; S2. Select key quality indicators from the raw data, specifically including the following sub-steps: S21. For volatile flavor substances, calculate the odor activity value (OAV) of each substance, and select substances with OAV ≥ 1 to form the first flavor substance set; S22. Using the comprehensive sensory score as the dependent variable, and all substances in the first flavor substance set, as well as all texture profile analysis parameters and basic nutritional components measured in step S1, perform a partial least squares regression analysis to calculate the variable importance projection value of each variable; S23. From the first flavor substance set, select volatile flavor substances with variable importance projection values ​​greater than 1, and record them as key volatile flavor substances; The key quality indicators include: shear force, hardness, adhesiveness, elasticity, cohesiveness, tackiness, chewiness, resilience, total amino acids, moisture, protein, fat, and the key volatile flavor compounds. S3. Based on principal component analysis, perform dimensionality reduction on the data of the key quality indicators and extract N principal components with a cumulative variance contribution rate of more than 80% and an eigenvalue greater than 1. S4. Input the scores of the N principal components into the pre-trained partial least squares regression model for quality prediction to obtain the quality prediction score of the sample; the partial least squares regression model for quality prediction uses the principal component scores as input variables and the comprehensive sensory score as output variables, and is obtained through sample training. S5. Classify the quality grade of the preserved mustard greens and braised pork samples according to the quality prediction score.

2. The method for quality evaluation and grading of braised pork with preserved mustard greens as described in claim 1, characterized in that, Step S3 further includes: obtaining the principal component comprehensive score F for each sample based on principal component analysis; wherein, the principal component comprehensive score F is obtained by weighting and summing the scores of each principal component using the proportion of the eigenvalue of each principal component to the total sum of the extracted principal component eigenvalues ​​as the weight. In step S5, when the quality prediction score is within the preset grade division boundary range, the grade of the sample is reviewed or adjusted with reference to the principal component comprehensive score F.

3. The method for quality evaluation and grading of braised pork with preserved mustard greens as described in claim 2, characterized in that, In step S3, the analysis based on principal component analysis specifically involves: dividing the samples into a frozen group and a room temperature group according to the storage conditions; performing principal component analysis independently on the key quality index data of the two groups of samples, and calculating the principal component comprehensive score F for each group of samples. In steps S4 and S5, for the samples in the frozen group and the room temperature group, the principal component scores extracted from the corresponding group and the independently trained partial least squares regression model are used for prediction and classification.

4. The method for quality evaluation and grading of braised pork with preserved mustard greens as described in claim 3, characterized in that, In step S4, the training process of the pre-trained partial least squares regression model includes: The sample dataset is divided into training and test sets according to a certain ratio; Cross-validation is used on the training set to select the optimal number of latent variables for the model; A partial least squares regression model was established and validated on the test set. The partial least squares regression model was successfully trained when the prediction determination coefficient was greater than 0.90 and the prediction root mean square error was less than 1.

0.

5. The method for quality evaluation and grading of braised pork with preserved mustard greens as described in claim 4, characterized in that, For frozen samples, the top three principal components with a cumulative variance contribution rate of 100% are extracted, and the formula for calculating their principal component composite score F is: F = 0.554 × F1 + 0.303 × F2 + 0.143 × F3; For the samples in the room temperature group, the top three principal components with a cumulative variance contribution rate of 100% are extracted, and the formula for calculating their principal component composite score F is: F = 0.528 × F1 + 0.246 × F2 + 0.226 × F3; F1, F2, and F3 represent the scores of the first, second, and third principal components of the corresponding groups, respectively.

6. The method for quality evaluation and grading of braised pork with preserved mustard greens as described in claim 5, characterized in that, In step S4, the The pre-trained partial least squares regression model is a Bagging ensemble model, which is trained using the following method: M1. Set the integration size K, K≥3; Perform K random samplings with replacement from the original training sample set that determines the key quality indicators. Each sampling has the same size as the original set, generating K subsets of training sets. Samples that are not selected form the corresponding K out-of-bag sample sets. M2. For the k-th subset of training, using the principal component scores extracted in step S3 as input and the comprehensive sensory score as output, independently train a partial least squares regression sub-model M. k Where k=1,...,K; M3, for the kth sub-model M k Using its out-of-bag sample set for prediction, the out-of-bag prediction determination coefficient R is calculated. 2 oob,k According to formula ω k = R 2 oob,k / Σ(R 2 oob,i ), calculate its weight coefficients, where i = 1 to K; M4. Save all K sub-models {M1, …, M} K } and its weight coefficients {ω1, …,ω K Together, they constitute the Bagging ensemble model.

7. The method for quality evaluation and grading of braised pork with preserved mustard greens as described in claim 6, characterized in that, When applying to a specific storage group, the final quality prediction should be performed according to the following procedure: A. For the sample to be tested, call the trained Bagging ensemble model of its corresponding group to obtain the predicted values ​​{y1, …, y} of the K sub-models. K }; B. Calculate the initial model prediction score Y for this sample. model =Σ(ω k ×y k ), where ω k These are the weight coefficients for the corresponding sub-models; simultaneously, the standard deviation σ of these K predicted values ​​is calculated. y ; C. Call the principal component comprehensive score F of the sample; the principal component comprehensive score F is calculated using the predetermined principal component weighting coefficients of the group, wherein the weighting coefficients for the frozen group are (0.554, 0.303, 0.143) and for the room temperature group are (0.528, 0.246, 0.226); D. Perform dynamic fusion decision-making and calculate the final quality prediction score Y: Y = λ × Y model + (1 - λ)× F; where the fusion weight λ is based on the standard deviation σ y Dynamically determined: if σ y If ≤ δ, then λ = 1.0; if σ y If δ > σ, then λ = γ / σ y The discreteness threshold δ and the adjustment parameter γ are determined by analyzing the σ values ​​of the samples in the training set. y The correlation between the predicted residuals and the predicted residuals is optimized and determined.

8. The method for quality evaluation and grading of braised pork with preserved mustard greens as described in claim 7, characterized in that, In step S5, the sample is graded or adjusted with reference to the principal component comprehensive score F, which specifically includes the following sub-steps: S51. When the predicted quality score Y of a sample is within the boundary interval between adjacent high-grade A and low-grade B, obtain the principal component composite score F of the sample, and at the same time obtain the average principal component composite score Favg of the predetermined grade A reference sample. A With standard deviation σ A And the average principal component composite score (Favg) of the B-level reference sample. B With standard deviation σ B ; S52. Calculate the standardized deviation D of the overall principal component score F of the sample relative to grade A. A The calculation formula is: D A =|F- Favg A | / σ A , where σ A The standard deviation of the F-values ​​for the reference sample at level A is given; simultaneously, the standardized deviation D relative to level B is calculated. B The calculation formula is: D B = |F - Favg B | / σ B , where σ B The standard deviation of the F-values ​​for the B-level reference sample; S53. Determine the level: If D A Less than or equal to D B And simultaneously satisfy F greater than or equal to Favg A If the sample grade is D, then the grade should be adjusted upwards or confirmed as grade A; if D B Less than D A And simultaneously satisfy F less than or equal to Favg B If the condition is not met, the sample grade will be adjusted down or confirmed as grade B; if none of the above conditions are met, the initial grading result based on the quality prediction score Y will be maintained.

9. The method for quality evaluation and grading of braised pork with preserved mustard greens as described in claim 3, characterized in that, The preset grade division boundary interval mentioned in step S5 is as follows: Let the score boundary of the grade division be T, then the boundary interval is [T-β, T+β], where β is determined according to the standard deviation of the distribution of the predicted scores of the samples in the training set, and the value range is 0.3×ρ≤β≤0.7×ρ, where ρ is the standard deviation of the predicted scores.

10. The method for quality evaluation and grading of braised pork with preserved mustard greens as described in claim 3, characterized in that, The quality grades described in step S5 are divided into three levels, specifically: the quality prediction scores Y are sorted from high to low, and two segmentation thresholds T1 and T2 are determined, where T1 > T2, corresponding to the boundaries between excellent and intermediate, and intermediate and qualified grades, respectively; if Y ≥ T1, it is judged as excellent; if T2 ≤ Y < T1, it is judged as intermediate; if Y < T2, it is judged as qualified; the values ​​of T1 and T2 are determined by analyzing the distribution of prediction scores of samples in the training set, combined with the actual distribution of sensory comprehensive scores and industry standards.

Images