An optimized processing method of plateau fresh rose

Abstract

The application relates to an optimized processing method of plateau fresh food roses. The optimized processing method of the plateau fresh food roses comprises the following steps: collecting multi-source data of the plateau fresh food roses, and preprocessing the multi-source data; analyzing fresh food rose quality influence factors, and screening out core influence factors; constructing a relational plateau fresh food rose database; respectively constructing prediction models for active ingredients and sensory quality indexes; fusing the active ingredient prediction model and the sensory quality index prediction model to construct a prediction formula of the active ingredients and the sensory indexes; and according to target requirements of users, solving an optimized whole processing scheme through the prediction model and the prediction formula. The scheme provided by the application can change a traditional processing process into data-driven accurate prediction and decision, realize accurate regulation and control and quality controllability of a plateau fresh food rose processing process, and improve stability and predictability of product quality.

Description

Technical Field

[0001] This application relates to the field of rose processing technology, and in particular to an optimized processing method for fresh highland roses. Background Technology

[0002] Highland fresh-eating roses thrive in a unique environment characterized by high altitude, strong ultraviolet radiation, and large diurnal temperature variations. Their petals accumulate active ingredients such as total phenols, total flavonoids, and anthocyanins, with levels far exceeding those of rose varieties grown in plains areas. These active ingredients endow them with extremely high economic value in food processing, cosmetics research and development, and health product production. Ultimately, the efficacy and market competitiveness of highland fresh-eating rose products depend on the retention efficiency of active ingredients during processing, as well as the stability of sensory qualities such as color and aroma. Therefore, the scientific and precise nature of processing techniques is crucial to the industry's development.

[0003] Currently, the quality of fresh-eating roses in the entire supply chain, from planting to processing, is influenced by multiple factors, including variety, maturity, drying temperature, and extraction liquid-to-material ratio. These factors exhibit significant synergistic or antagonistic interactions. Existing technologies often employ single-factor optimization methods, focusing only on the independent effects of individual variables and failing to explore the complex interactions between multiple factors. This results in optimization schemes that struggle to balance the retention of active ingredients with the improvement of sensory quality, leading to significant batch-to-batch variations and poor quality stability. Furthermore, processing companies cannot predict processing parameters based on target product requirements. The production process necessitates repeated trial and error adjustments, which not only wastes raw materials and increases production costs but also prolongs the production cycle, significantly reducing production efficiency and hindering the industry's large-scale and standardized development.

[0004] Therefore, it is urgent to build a data-driven intelligent processing technology system to achieve precise control and quality control in the processing of fresh highland roses, and provide core support for industrial upgrading. Summary of the Invention

[0005] To address or partially address the problems existing in related technologies, this application provides an optimized processing method for fresh highland roses, aiming to solve the technical problem of predicting the processing quality of fresh highland roses based on the influence of multiple factors.

[0006] This application provides an optimized processing method for fresh highland roses, including: Multi-source data on fresh highland roses were collected and preprocessed. The preprocessing included Z-score standardization of continuous variables and one-hot encoding of discrete variables, and K-nearest neighbor interpolation or targeted experiments to supplement missing data. Based on the processed multi-source data, we analyzed the factors affecting the quality of fresh roses and screened out the core influencing factors. Among them, the analysis of factors affecting the quality of fresh roses included, with other variables fixed, analyzing the quantitative impact of single variable factors on the quality of fresh roses; and using an orthogonal experimental design, combining multiple linear regression and XGBoost regression models to explore the interaction patterns of multiple factors. A five-level data relational database of plateau fresh roses was constructed, consisting of a variety information table, a planting condition table, a processing parameter table, a quality index table, and a correlation table. Based on a relational database of fresh highland roses, prediction models for active ingredients and sensory quality indicators were constructed. The active ingredient prediction model for fresh highland roses includes a multiple linear regression model and a support vector regression model for predicting total phenols and total flavonoids, and a random forest model for predicting anthocyanins. The sensory quality indicator prediction model for fresh highland roses includes a BP neural network for predicting color indicators and an XGBoost model for predicting aroma indicators. A Stacking ensemble learning strategy was adopted to fuse an active ingredient prediction model and a sensory quality index prediction model. Based on the fusion model, key influencing factors and their weights were extracted, and prediction formulas for active ingredients and sensory indicators were constructed. Based on the user's target needs, a complete processing solution is obtained by using prediction models and formulas to determine the optimal variety, planting altitude, maturity, processing method, and drying process.

[0007] Optionally, the quantitative impact of single-variable factors on the quality of fresh roses can be analyzed, including: With other variables fixed, we explored the quantitative effects of a single variable on the color, active ingredients, and aroma of fresh roses. Pearson correlation coefficient was used to perform correlation analysis on the target independent variable and quality index data of highland fresh roses to test normality. If the data conformed to a normal distribution, Pearson correlation analysis was performed directly. If the data did not conform to a normal distribution, the data was first transformed to conform to a normal distribution before analysis was performed. For discrete independent variables, first confirm the compatibility between the binary data after one-hot encoding of the discrete independent variables and the quality indicators, and then conduct correlation analysis. The correlation coefficient between the target independent variable and each quality indicator was calculated using the Pearson correlation coefficient formula. Based on the calculated correlation coefficient values, the direction of linear association between the target independent variable and each quality indicator was determined, and the level of association strength was judged. At the same time, the specific values ​​of the correlation coefficients of each association were recorded to form correlation analysis data.

[0008] Optionally, explore the patterns of interaction among multiple factors, including: Using key variables as influencing factors, and combining the optimal values ​​of single influencing factors with actual production feasibility, an influencing factor-numerical table is determined, and an orthogonal experimental scheme is constructed. Conduct experiments according to the plan, with each group set up two or more repeated experiments to ensure data reliability. During the experiment, extraneous variables are strictly controlled and standardized operating procedures are followed. After the experiment was completed, the core quality indicators of each combination were measured, the raw data were recorded, and the data were preprocessed: outlier data exceeding 1.5 times the interquartile range were removed by box plot method; the mean and standard deviation of each group of repeated data were calculated, and if the coefficient of variation was >5%, the experiment was repeated to verify the stability of the data; the quality indicator data were standardized by Z-score to eliminate the influence of dimensions. Orthogonal experimental data were coded into variables. Discrete variables were converted into numerical values ​​using one-hot coding, while continuous variables retained the standardized data. Based on the coded variables, pairwise interaction terms are constructed, and the main effect terms of each variable are combined to form a set of independent variables. With a single quality index as the dependent variable, a multiple linear regression model is constructed using the stepwise regression method. The model is kept simple and effective by screening independent variables. After constructing the multiple linear regression model, a significance test is performed. The F-test is used to verify the overall significance of the model and ensure that the set of independent variables has significant explanatory power for the dependent variable. Significant interaction terms are screened by t-test, and interaction terms with no significant effect are removed. The model determination coefficient R² is calculated. If R² ≥ 0.6, the model fit is good. Otherwise, variable coding needs to be optimized, data supplemented, or the model needs to be reconstructed. Based on the preprocessed data, we constructed an XGBoost regression model with significant interaction terms and main effect terms as input features and quality indicators as output targets. After the model is trained, the importance score of each input feature is calculated based on the number of splits and split gain of the decision tree. The features are sorted from high to low and the top-ranked interaction items are selected as high-contribution combinations. The intensity of interaction effects is quantified by marginal effect analysis, interaction effect values ​​are calculated, synergistic or antagonistic relationships between variables are determined, and the interaction effect levels of each combination are clarified by classifying the absolute value of the effect value to the mean of the quality index.

[0009] Optionally, core impact factors can be selected, including: A two-dimensional evaluation index system is constructed. The contribution of a single factor is calculated based on the mean absolute value of the correlation coefficient between the variable and the quality index Pearson, reflecting the strength of the independent influence. The contribution of multiple factors in interaction is calculated based on the mean importance score of the high-contribution interaction combination involving the variable, reflecting the strength of the interaction contribution. Objective weights are determined using the entropy weight method, and preliminary subjective weights are replaced by index standardization, information entropy calculation, and entropy weight solution. Technical personnel in relevant fields conducted multiple rounds of scoring using the Delphi method until opinions converged, and the average score was taken as the subjective weight. The overall weight is calculated by combining objective and subjective weights in a 6:4 ratio. The top 4 variables with the highest overall weights are then selected to determine the final core influencing factors.

[0010] Optional predictive models for the active ingredients of highland fresh roses include: For the three core active ingredients of fresh highland roses—total phenols, total flavonoids, and anthocyanins—exclusive predictive models were constructed to form a benchmark-complementary-precision system. The prediction model for total phenols and total flavonoids uses multiple linear regression as a benchmark, incorporating key factors such as variety, altitude and maturity. Variables are screened through stepwise regression to eliminate collinearity. A support vector regression model is constructed simultaneously, and a radial basis function kernel function is used to capture nonlinear correlations. The penalty parameter and kernel function parameter are optimized through grid search. The anthocyanin prediction model uses a random forest model, based on an ensemble of 100 decision trees. The training set is generated by bootstrap resampling, and 1 / 3 of the features of each tree are randomly selected for splitting. Parameters such as setting the maximum tree depth to 10-30 layers are used to improve the robustness of the model.

[0011] Optional sensory quality index prediction models for highland fresh roses include: A backpropagation neural network for color indicators was constructed based on the characteristics of color indicators, employing a 6-input, 2-hidden-3-output architecture. The input layers correspond to six key factors: variety, altitude, maturity, pretreatment method, drying process, and extraction process. The hidden layers consist of 12→8 gradient nodes to extract non-linear correlation features, using the Sigmoid activation function. The output layer corresponds to the three core coordinate values ​​L in the color space of the color indicator. a b ; A gradient boosting decision tree XGBoost model for aroma indicators was constructed based on the characteristics of aroma indicators. The output is the total number of volatile substances plus the relative content of characteristic components. The input includes 6 key factors and gas chromatography-mass spectrometry auxiliary features. A step-by-step iteration with a learning rate of 0.05 was adopted, and the maximum tree depth was set to 8 layers, the minimum number of split samples was 4, and the column sampling rate was 0.8 to enhance the prediction sensitivity of low-content components. The training data covers multiple varieties, altitudes, and process scenarios. The parameters of the two models are optimized through grid search, with a focus on reducing the prediction error of the content components.

[0012] Optionally, a model that integrates active ingredient prediction and sensory quality indicator prediction may be included, including: The Stacking ensemble learning strategy is adopted, using five basic models—multivariate linear regression, support vector regression, gradient boosting decision tree, random forest, and BP neural network—as the basic models and logistic regression as the meta-model. The prediction results of each basic model are weighted and fused to improve prediction accuracy. Based on the quality index characteristics of fresh highland roses, five basic models were adapted and preprocessed in a targeted manner. For the active ingredients total phenols and total flavonoids, which exhibit significant linear correlation, multiple linear regression and support vector regression models were retained. For the active ingredient anthocyanins, which are significantly affected by multi-factor nonlinearity, a random forest model was incorporated. For the sensory indicator color, which is easily affected by interference, a backpropagation neural network model was retained. For the complex indicator aroma, which contains trace components, a gradient boosting decision tree model was adopted. The five models were divided into multiple subsets for separate training. The active ingredient prediction subset included multiple linear regression, support vector regression, and random forest, while the sensory quality prediction subset included backpropagation neural network and gradient boosting decision tree. A 5-fold cross-validation strategy was used to train each basic model. In each round of validation, 4 / 5 of the training set data was used for modeling, and 1 / 5 of the data was used to generate prediction results, which served as input features for the meta-model. This process was iteratively repeated to generate prediction features for the entire training set. The test set directly generated prediction features using the trained basic models. The output results of each model are standardized to eliminate differences in the output dimensions of different models; correlation analysis is used to screen model combinations with a correlation of <0.7 in the prediction results to preserve model diversity and improve the integration effect. The Stacking ensemble learning framework is constructed by building a two-level ensemble framework consisting of a base model layer and a meta-model layer. This framework enhances the predictive capabilities of each model. Logistic regression is selected as the meta-model, and the predicted features generated by the base model layer are used as the input to the meta-model. The measured values ​​of the quality index are used as the output target. The training objective function of the meta-model is constructed, and weights are assigned based on the coefficient of determination R² of the validation sets of each base model. By introducing weighting coefficients, the prediction results of each base model are weighted and fused, thereby strengthening the contribution of high-quality models and weakening the influence of models with large errors. A hierarchical hyperparameter optimization system is constructed, employing a two-step strategy of coarse optimization via grid search and fine optimization via Bayesian optimization to accurately optimize the hyperparameters of the basic and meta-models. The basic model sets the optimization range according to the characteristics of the model, first determining the optimal range of parameters, and then using Bayesian optimization to fit the mapping relationship between parameters and performance to locate the optimal combination. The meta-model optimizes the regularization strength and number of iterations for logistic regression, and determines the parameters by combining 5-fold cross-validation.

[0013] Optionally, predictive formulas for active ingredients and sensory indicators can be constructed, including: Feature importance analysis was conducted based on the optimized fusion model to screen for key active ingredients and color a. The core factors that significantly affect the value are the variety, planting altitude, maturity, drying temperature, and pretreatment method. The linear and nonlinear fitting results of basic models such as MLR and SVR in the active ingredient fusion model are finally extracted into a linear superposition formula structure after weighted fusion by the meta-model. Color a The nonlinear fitting characteristics of the BP neural network in the value fusion model are extracted into a formula structure in the form of logistic regression to match the changing pattern of the redness index. The training parameters of the fusion model provide the initial data source for the formula weight coefficients, enabling the transfer of the model's predictive ability to the formula, thus obtaining the prediction formula for key active ingredients in fresh roses: (1) In the formula, This is a key active ingredient in fresh highland roses. For constant terms; ~ These are the weight coefficients for each variable and interaction term; Variety codes are assigned, with different values ​​selected based on testing for different varieties; For planting altitude; For maturity; This refers to the drying temperature. Key interaction items; This is random error; Color index a The formula for predicting the rose redness of fresh roses from high-altitude areas: (2) In the formula, For color value; These are the logistic regression coefficients; Encode the preprocessing method; coefficient , The acquisition relies on the Stacking fusion model as its core, combining data calibration and statistical testing to achieve accurate solutions; utilizing the feature weight analysis function of the fusion model, the initial weights of each core factor and interaction term are extracted as... The initial value reflects the contribution intensity of each factor to the content of active ingredients; Color a Value formula The initial values ​​are derived from the connection weights between the output and hidden layers of the BP neural network in the fusion model, and are obtained after logistic regression transformation. constant term , These correspond to the predicted baseline values ​​of the fusion model when each factor takes a baseline value; The extracted initial weights are substituted into the preprocessed full dataset, and the least squares method is used for calibration. The weight coefficients are fine-tuned with the goal of minimizing the mean square error between the predicted and measured values ​​to eliminate the error caused by model simplification. For the encoding and assignment of discrete variables, the assignment rules are optimized by combining the prediction bias of the fusion model for different varieties and preprocessing methods to ensure the matching of the encoded value and the weight coefficient. The significance of each weight coefficient was verified by t-test, and insignificant coefficient terms were removed. The consistency between the formula for constructing the weights after calibration and the prediction results of the fusion model was verified by the coefficient of determination R², thus ensuring the reliability of the weight coefficients.

[0014] Optionally, preprocessing of multi-source data may be performed, including: We collected literature data on fresh highland roses, production logs from processing enterprises, monitoring data from planting bases, and quality inspection reports from the market. The data dimensions covered rose raw materials, processing technology, and finished product quality. Rose raw materials included variety type, planting altitude, and maturity. Processing technology included pretreatment methods, drying processes, and extraction processes. Finished product quality included color indicators, active ingredients, and aroma components. Among these, the solvent type for variety type, pretreatment methods, drying processes, and extraction processes were discrete variables, while the spatial values ​​for planting altitude, processing parameters, active ingredient content, and color indicators were continuous variables. Maturity was a graded variable. The Z-score standardization method was used to normalize continuous variables and eliminate the influence of dimensions; the categories and values ​​of each discrete variable were sorted out and uniformly named. To count the number of categories of discrete variables, one-hot encoding is used to convert the n categories of variables into n binary variables. When a value is taken from a certain category, the corresponding variable is set to 1, and the rest are set to 0. Integrate the encoded data with standardized continuous variables to ensure a consistent format; In terms of missing value identification, the standardized dataset is scanned to record the location of missing values ​​and related background information; the missing rate of each variable is calculated, and the variables are divided into low missing rate and high missing rate according to the missing rate, and a missing rate statistical report is generated. Using the missing variable as the target variable and other variables as features, a random forest classification model is constructed. The missing pattern is determined by the feature importance score. If there is no significant correlation with other variables, it is considered random missing; if there is a strong correlation, it is considered non-random missing. For data with low missing values, K-nearest neighbor interpolation was used for imputation, and the optimal K value was determined through cross-validation. Euclidean distance was used as a measure of similarity between samples to calculate the distance between the sample containing the missing value and other non-missing samples. Based on the determined K value and the distance measurement results, the K nearest non-missing samples to the sample containing the missing value were selected as reference samples for interpolation. For continuous missing variables, the mean of the corresponding variable values ​​of the K nearest neighbor samples was used as the imputation value for the missing value. For discrete missing variables, the mode of the corresponding variable values ​​of the K nearest neighbor samples was used as the imputation value for the missing value. For key data with high missing rates, targeted experiments were designed to supplement the data, and matching materials were selected according to the missing background; independent variables such as process parameters and dependent variables such as finished product quality were set to cover the missing dimensions; each group was repeated 3-5 times; and the testing process was standardized to ensure that the results are reproducible.

[0015] Optionally, a complete processing scheme is derived, including the optimal variety, planting altitude, maturity, processing method, and drying process, including: Receive customized requests from users, which should include at least the target product type, the expected target value or range of core quality indicators, and acceptable cost constraints. Perform compliance verification on the received user request data, including checking whether the indicators are within the range supported by the model and whether the values ​​are within a reasonable range; After successful validation, the user-inputted raw target value is converted into a standardized data format that the model can process, and the following two computation paths are started in parallel: Path 1: Input the standardized demand data into the fusion prediction model to obtain an output containing the prediction results and their confidence levels; Path 2 involves substituting the standardized demand data into the prediction formula to obtain the recommendation results; The decision is made based on the confidence level of the model output in Path 1. If the confidence level is ≥80%, the processing path corresponding to the model prediction result is adopted as the preferred recommended solution; if the confidence level is <80%, the processing path corresponding to the formula calculation result is adopted as the preferred recommended solution. The preferred selection scheme should include at least the recommended rose variety, planting altitude range, raw material maturity, pretreatment method, drying process parameters and extraction process parameters; The generated optimization scheme, its corresponding theoretical quality index predictions, and key operational points are converted into graphic and textual formats and presented to users intuitively.

[0016] The technical solution provided in this application may include the following beneficial effects: By conducting single and multi-factor impact analysis using multi-source data, a database of highland fresh roses was established. Predictive models for active ingredients and sensory quality indicators were constructed and fused using a Stacking ensemble learning strategy to extract key factor weights and construct quantifiable prediction formulas. Finally, based on customized user needs, the processing plan was solved in reverse using the prediction models and formulas, transforming the traditional processing process into data-driven precision prediction and decision-making. This enabled precise control and quality control of the highland fresh rose processing process, improving the stability and predictability of product quality.

[0017] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0018] The above and other objects, features and advantages of this application will become more apparent from the more detailed description of exemplary embodiments thereof in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components in the exemplary embodiments thereof.

[0019] Figure 1 This is a schematic flowchart illustrating the optimized processing method for fresh highland roses as shown in the embodiments of this application; Figure 2 This is a schematic diagram of the technical route for the optimized processing method of fresh highland roses as shown in the embodiments of this application. Detailed Implementation

[0020] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to make this application more thorough and complete, and to fully convey the scope of this application to those skilled in the art.

[0021] The technical solutions of the embodiments of this application are described in detail below with reference to the accompanying drawings.

[0022] Figure 1 This is a schematic flowchart illustrating the optimized processing method for fresh highland roses as shown in the embodiments of this application.

[0023] In some implementations, see Figure 1 An optimized processing method for fresh highland roses includes: S101. Collect multi-source data on fresh highland roses and preprocess the multi-source data; Specifically, the study systematically collected literature data on highland fresh-eating roses, production logs from processing enterprises, monitoring data from planting bases, and market circulation quality inspection reports. The data dimensions cover the entire processing chain of highland fresh-eating roses, from raw materials to processing technology to finished product quality. Raw material data includes variety type, planting altitude, and maturity level, with maturity graded into five levels based on petal unfolding and color. Processing technology includes pretreatment methods, drying processes, and extraction processes. Pretreatment methods include washing methods, types and concentrations of color-protecting agents; drying processes include drying methods, temperature, time, and air speed; extraction processes include solvent type, material-to-liquid ratio, extraction temperature, time, and pressure. Finished product quality includes color indicators, active ingredients, and aroma components. Color indicators include the three core coordinate values ​​L in the color space. a b The active ingredients include the content of total phenols, total flavonoids, and anthocyanins, while the aroma components include the types and relative contents of volatile aromatic substances. Among these, variety type, pretreatment method, drying process, and solvent type in extraction process are discrete variables; planting altitude, processing parameters, active ingredient content, and color index are continuous variables; and maturity is a graded variable.

[0024] Z-score standardization was used to normalize continuous variables and eliminate the influence of dimensions. The categories and values ​​of discrete variables were identified and standardized, such as distinguishing between "hot air drying" and "hot air drying". The number of categories of discrete variables was counted, and one-thermal coding was used to convert the n categories of variables into n binary variables. When a value belongs to a certain category, the corresponding variable is set to 1, and the rest are set to 0. The coded data was then integrated with the standardized continuous variables to ensure a consistent format.

[0025] Specifically, in terms of missing value identification, a standardized dataset is scanned to record the location of missing values ​​and related background information. The missing value rate for each variable is calculated as: Missing value rate = Number of missing values ​​ / Total number of missing values ​​× 100%. Variables are categorized into low missing value rate (<10%) and high missing value rate (≥10%), and a missing value statistical report is generated. Using the missing variable as the target variable and other variables as features, a random forest classification model is constructed, with the dependent variable set to: Missing = 1, Non-missing = 0. The missing value pattern is determined by feature importance scores; if there is no significant correlation with other variables, it is considered random missing; if a strong correlation exists, it is considered non-random missing. For data with a missing value rate <10%, K-nearest neighbor interpolation was used for data supplementation. The specific steps were as follows: First, the optimal K value was determined using cross-validation. Second, Euclidean distance was used as a measure of similarity between samples to calculate the distance between the sample containing the missing value and other non-missing samples. Third, based on the determined K value and distance measurement results, the K nearest non-missing samples were selected as reference samples for interpolation. For continuous missing variables, the mean of the corresponding variable values ​​of the K nearest neighbor samples was used as the supplementary value for the missing value. For discrete missing variables, the mode of the corresponding variable values ​​of the K nearest neighbor samples was used as the supplementary value for the missing value. For critical data with a missing value rate ≥10%, targeted experiments were designed for data supplementation. The specific steps were as follows: First, matching materials were selected according to the missing background, such as variety, altitude, and maturity. Independent variables (process parameters) and dependent variables (components or color, aroma, and flavor) were set to cover the missing dimensions. Each group was repeated 3-5 times. Standardized detection procedures were implemented, such as HPLC for component analysis and GC-MS for aroma analysis, to ensure reproducible results.

[0026] S102. Based on the processed multi-source data, analyze the factors affecting the quality of fresh roses and screen out the core influencing factors. Specifically, single-factor influence analysis was conducted, fixing other variables to explore the quantitative impact of individual variables on the color, active ingredients, and aroma of fresh roses. Correlation analysis was performed based on Pearson correlation coefficients to test the normality of the target independent variable and quality index data for highland fresh roses. If the data conformed to a normal distribution, Pearson correlation analysis was performed directly; if not, the data underwent normalization transformations, such as logarithmic or square root transformations, until the data conformed to a normal distribution before analysis. For discrete independent variables, the compatibility between their one-hot encoded binary data and the quality indicators was confirmed before conducting correlation analysis. The Pearson correlation coefficient formula was used to calculate the correlation coefficient *r* between the target independent variable and each quality indicator. The correlation coefficient *r* ranged from -1 to 1, where *r* > 0 indicated a positive correlation (increasing the independent variable leads to an increase in the quality indicator), *r* < 0 indicated a negative correlation (increasing the independent variable leads to a decrease in the quality indicator), and *r* = 0 indicated no linear correlation. The closer *|r|* was to 1, the stronger the linear association. *|r|* ≥ 0.8 indicated a very strong correlation, 0.6 ≤ *r|* < 0.8 indicated a strong correlation, 0.4 ≤ *r|* < 0.6 indicated a moderate correlation, 0.2 ≤ *r|* < 0.4 indicated a weak correlation, and *r|* < 0.2 indicated a very weak correlation or no correlation. Based on the calculated correlation coefficient *r* values, the direction of the linear association between the target independent variable and each quality indicator was determined, and the level of association strength was assessed. Simultaneously, the specific values ​​of the correlation coefficients for each association were recorded to form a correlation analysis table, providing a basis for subsequent trend and pattern summarization.

[0027] Specifically, the study explored the interaction patterns of multiple factors, using key variables as influencing factors. A numerical table of influencing factors was determined by combining the optimal level range of single factors with actual production feasibility, and an orthogonal experimental design was constructed. The experiment was conducted according to the design, with three replicates per group to ensure data reliability. Irrelevant variables were strictly controlled during the experiment, and standardized operating procedures were followed. After the experiment, the core quality indicators of each combination were measured and the raw data were recorded. Data preprocessing was then performed, using box plots to remove outliers exceeding 1.5 times the interquartile range. The mean and standard deviation of the replicated data for each group were calculated; if the coefficient of variation was greater than 5%, the experiment was repeated to verify data stability. Z-score standardization was applied to the quality indicator data to eliminate the influence of dimensions, laying the foundation for subsequent modeling. Variable encoding was performed on the orthogonal experimental data. Discrete variables were converted to numerical values ​​using one-hot encoding, while continuous variables retained the standardized data. Pairwise interaction terms were constructed based on the coded variables, and independent variable sets were formed by combining the main effect terms of each variable. A multiple linear regression model was constructed using stepwise regression with a single quality indicator as the dependent variable. Automatic selection of independent variables ensured the model's simplicity and effectiveness. After model construction, significance tests were performed: an F-test (α=0.05, P<0.05) verified the overall significance of the model, ensuring that the independent variable set had significant explanatory power for the dependent variable; a t-test (α=0.05, P<0.05) screened significant interaction terms and removed those with no significant effect. The model's coefficient of determination (R²) was calculated. If R² ≥ 0.6, the model fit was considered good; otherwise, variable coding needed optimization, data supplementation, or model reconstruction. Based on the preprocessed data, significant interaction terms and main effect terms were used as input features, and quality indicators were used as output targets to construct an XGBoost regression model.

[0028] A grid search algorithm combined with 5-fold cross-validation was used to optimize model parameters, including learning rate, number of decision trees, and maximum tree depth, with the goal of minimizing the mean squared error of cross-validation to determine the optimal parameter combination. After model training, the importance score of each input feature was calculated based on the number of decision tree splits and split gain. The features were then sorted from highest to lowest score, and the top 20% of interaction terms were selected as high-contribution combinations. Marginal effect analysis was used to quantify the strength of the interaction effect, calculating the interaction effect value (the difference between the combined effect value and the sum of the individual effects) to determine whether there was a synergistic or antagonistic relationship between variables. The interaction effect was then graded according to the ratio of the absolute value of the effect value to the mean of the quality index: strong >10%, medium 5%-10%, and weak <5%, thus clarifying the level of interaction effect for each combination.

[0029] A dual-dimensional evaluation index system was constructed. The contribution of a single factor was calculated based on the mean absolute value of the Pearson correlation coefficient between the variable and the quality index, reflecting the strength of independent influence. The contribution of multiple factors in interaction was calculated based on the mean importance score of high-contribution interaction combinations involving the variable, reflecting the strength of interaction contribution. The entropy weight method was used to determine objective weights, replacing preliminary subjective weights through index standardization, information entropy calculation, and entropy weight solution. Five to seven experts in relevant fields were invited to score multiple times using the Delphi method until consensus was reached (i.e., the coefficient of variation <10%), and the mean was taken as the subjective weight. The objective and subjective weights were combined in a 6:4 ratio to calculate the comprehensive weight. The top four variables with the highest comprehensive weights were selected, ultimately identifying variety, maturity, drying temperature, and extraction-to-liquid ratio as the core influencing factors, focusing on key variables for subsequent prediction model optimization.

[0030] S103. Construct a five-level data relationship database of plateau fresh roses, consisting of variety information table, planting condition table, processing parameter table, quality index table, and association table. Specifically, a relational database for processing and quality of plateau fresh roses was constructed based on MySQL. A five-level data table structure was designed, consisting of a variety information table, a planting condition table, a processing parameter table, a quality index table, and a correlation table. The correlation table includes a single-factor influence correlation analysis table and a multi-factor interaction influence table. This enables many-to-many correlation queries of plateau fresh rose processing data and supports data addition, deletion, updating, and multi-condition filtering functions, providing data support for subsequent model construction.

[0031] S104. Based on the relational highland fresh rose database, predictive models were constructed for active ingredients and sensory quality indicators, respectively. A basic predictive model for the active ingredients of highland fresh roses was constructed, with dedicated predictive models built for three core active ingredients: total phenols, total flavonoids, and anthocyanins, forming a "benchmark-complementary-precision" system. The total phenol and total flavonoid models used a multiple linear regression (MLR) model as a benchmark, incorporating key factors such as variety, altitude, and maturity, as well as linear and quadratic terms. Variables were screened through stepwise regression, and collinearity was eliminated when VIF < 10. Simultaneously, a support vector regression (SVR) model was constructed, employing a radial basis function (RBF) kernel to capture nonlinear associations, and optimizing the penalty parameters C and C' through grid search. (0.1~100), and kernel function parameters g, g (0.01~10).

[0032] The anthocyanin model uses a random forest (RF) model, based on an ensemble of 100 decision trees. The training set is generated by bootstrap resampling, and 1 / 3 of the features of each tree are randomly selected for splitting. Parameters such as setting the maximum tree depth to 10-30 layers are used to improve the robustness of the model.

[0033] A comprehensive validation system was established to ensure accuracy and stability. Parameter optimization involved adjusting variable terms and L2 regularization parameters based on MLR. The Support Vector Regression (SVR) model employed 5-fold cross-validation combined with grid search to optimize parameters, aiming for minimum RMSE. The Random Forest (RF) model adjusted parameters such as the number and depth of trees, using out-of-bag (OOB) error to select the optimal configuration. Internal validation used RMSE, MAE, and R² on the validation set, requiring R² ≥ 0.85. Prediction errors showed no significant differences across different varieties and altitude ranges. Overfitting was addressed through enhanced regularization. External testing used the test set to verify generalization ability, requiring the evaluation metrics to differ from the validation set by ≤ 10%. Residual normality and lack of trend tests validated the reasonableness of the fit.

[0034] Establish a multi-dimensional mechanism to ensure practicality and stability: Data quality control, including data traceability and labeling, clearly defining the source and testing methods, and eliminating low-reliability data. Robustness testing, fine-tuning input variables, requiring prediction result fluctuations ≤5%. Interpretability analysis, analyzing feature importance through regression coefficients, kernel function contribution, and node splitting gain, clarifying factor influence weights.

[0035] A sensory quality prediction model for highland fresh roses was constructed, with dedicated models built for color and aroma indicators to improve prediction accuracy. A backpropagation (BP) neural network for color indicators was established, employing a "6-input - 2-hidden-3-output" architecture. The input layers correspond to six key factors: variety, altitude, maturity, pretreatment method, drying process, and extraction process. The hidden layers consist of 12→8 gradient nodes to extract non-linear correlation features, using the Sigmoid activation function. The output layer corresponds to L... a b The Adam optimizer was used with an initial learning rate of 0.001, and the algorithm was iterated for 10 rounds. L2 regularization was introduced to suppress overfitting.

[0036] An XGBoost model for aroma indicators was established, with the output being "total types of volatile substances + relative content of characteristic components". The input includes 6 core factors and gas chromatography-mass spectrometry (GC-MS) auxiliary features. A step-by-step iterative learning rate of 0.05 was adopted, with a maximum tree depth of 8 layers, a minimum number of split samples of 4, and a column sampling rate of 0.8 to enhance the sensitivity of low-content component prediction. The training data covers multiple varieties, altitudes, and processing scenarios.

[0037] The parameters of the BP neural network for color indicators and the XGBoost model for aroma indicators were optimized using grid search, with a focus on reducing the prediction error of content components. Internal validation requirements were: color model R² ≥ 0.85, RMSE ≤ 0.3 (L... ) / 0.2(a / b The aroma model's predictive accuracy for characteristic components is ≤8%, and the overall accuracy for all categories is ≥90%. External testing ensures that the difference from the validation set metrics is ≤10%, using 50% and 100% cross-validation to guarantee stability. A comparative analysis of the predicted trends for low-content aroma components and a special evaluation of process interference simulation for the color model are conducted. A data traceability and quality rating system is established, eliminating data with a reliability <85%. Factor contribution is clarified through weight analysis and feature importance ranking. An incremental learning interface is reserved to incorporate new data for parameter optimization. Error warning thresholds are set: color RMSE > 0.5 and aroma RE > 10%, triggering a re-examination and re-optimization process if these thresholds are exceeded.

[0038] S105. The Stacking ensemble learning strategy is used to fuse the active ingredient prediction model and the sensory quality index prediction model. Based on the fusion model, key influencing factors and their weights are extracted, and prediction formulas for active ingredients and sensory indicators are constructed. Specifically, a stacking ensemble learning strategy is adopted, using multiple linear regression, support vector regression, gradient boosting decision tree, random forest and BP neural network as the basic models, and logistic regression as the meta-model. The prediction results of each basic model are weighted and fused to improve prediction accuracy.

[0039] The basic model selection, adaptation, and preprocessing were conducted based on the quality indicators of fresh highland roses, including active ingredients, color, and aroma. Five basic models were specifically adapted and uniformly preprocessed. For active ingredients with significant linear correlations, such as total phenols and total flavonoids, the Multiple Linear Regression (MLR) and Support Vector Regression (SVR) models were retained. For components significantly affected by multi-factor nonlinearity, such as anthocyanins, the Random Forest (RF) model was incorporated. For sensory indicators easily affected by interference, such as color, the Backpropagation (BP) neural network model was retained. For complex indicators containing trace components, such as aroma, the Gradient Boosting Decision Tree (XGBoost) model was adopted. The output results of each model were standardized and normalized to the [0,1] interval to eliminate differences in the dimensions of different model outputs, laying the foundation for subsequent fusion. Simultaneously, to avoid redundancy in the basic models, correlation analysis was used to screen model combinations with a correlation of <0.7 in the prediction results, preserving model diversity to improve the integration effect.

[0040] The Stacking ensemble learning framework is constructed, creating a two-tier ensemble framework of "base model layer - meta-model layer" to enhance the predictive capabilities of each model. In the base model layer design, five types of models are divided into multiple subsets for separate training. The active ingredient prediction subset includes multiple linear regression (MLR), support vector regression (SVR), and random forest (RF). The sensory quality prediction subset includes backpropagation (BP) neural network and gradient boosting decision tree (XGBoost) models. A 5-fold cross-validation strategy is used to train each base model. In each round of validation, 4 / 5 of the training set data is used for modeling, and 1 / 5 of the data is used to generate prediction results, which serve as the input features for the meta-model. This process iterates to generate prediction features for the entire training set. The test set directly generates prediction features from the trained base models. In the meta-model layer design, logistic regression is chosen as the meta-model due to its strong fitting ability to multi-dimensional input features and stable generalization performance. The prediction features generated by the base model layer are used as the input to the meta-model, with the measured quality index as the output target, to construct the meta-model training objective function. By introducing weighting coefficients, weights are assigned based on the coefficient of determination R² of each base model's validation set. The higher the R², the greater the weight. This achieves weighted fusion of the prediction results of each base model, strengthens the contribution of high-quality models, and weakens the influence of models with larger errors.

[0041] A hierarchical hyperparameter optimization system is constructed, employing a two-step strategy of coarse optimization via grid search and fine optimization via Bayesian optimization to accurately optimize the hyperparameters of the basic and meta-models. The basic model has its optimization range set according to its characteristics, first determining the approximate optimal range of parameters, and then efficiently locating the optimal combination by fitting the parameter-performance mapping relationship through Bayesian optimization. The meta-model optimizes the regularization strength and iteration count for logistic regression, and determines the parameters using 5-fold cross-validation. A multi-dimensional validation system is established, using hierarchical validation with validation and test sets. The requirements are: R² ≥ 0.90 for the fused model on the test set, RMSE reduced by more than 15% compared to the single optimal model, and prediction bias ≤ 8%; robustness testing is performed on simulated production parameter fluctuations, requiring an accuracy decrease of ≤ 5%. An optimization log tracking system, model update interface, and performance early warning mechanism are established to ensure model reliability.

[0042] Specifically, a personalized prediction formula for highland fresh roses was developed. Based on an optimized fusion model, feature importance analysis was conducted to screen out key active ingredients and color a. The core factors significantly influencing the value, with active ingredients focusing on variety, planting altitude, maturity, drying temperature, and pretreatment methods; the linear and nonlinear fitting results of basic models such as multiple linear regression (MLR) and support vector regression (SVR) in the active ingredient fusion model, after weighted fusion by the meta-model, are finally refined into a linear superposition formula structure, containing main effect terms and interaction terms; color a Due to the nonlinear fitting characteristics of the BP neural network in the fusion model, a formula structure in the form of logistic regression was extracted to match the changing patterns of the redness index. Finally, the training parameters of the fusion model provided the initial data source for the formula weight coefficients, realizing the transfer of the model's predictive ability to the formula. Therefore, the prediction formula for the key active ingredients of fresh roses was obtained: (1) In the formula, This is a key active ingredient in fresh highland roses. For constant terms; ~ The weight coefficients for each variable and interaction term are obtained from training the ensemble model; For variety coding, different discrete variables are assigned different values ​​based on the test, such as double-petaled red rose = 1, bitter water rose = 2, others = 3; For planting altitude; The maturity level is 1-5, which is normalized to 0-1. This is the drying temperature, which is normalized to 0-1; Key interaction items; This is random error.

[0043] Color index a The formula for predicting the rose redness of fresh roses from high-altitude areas is as follows: (2) In the formula, For color value; ~ These are the logistic regression coefficients; Encode the preprocessing method: color protection = 1, no color protection = 0; Weighting coefficient , The acquisition relies on the Stacking fusion model as its core, combining data calibration and statistical testing to achieve accurate solutions. First, the feature weight analysis function of the fusion model is used to extract the initial weights of each core factor and interaction term as [the initial weights]. The initial values, derived from the regression coefficients of the corresponding factors in the meta-model, reflect the contribution intensity of each factor to the content of active ingredients; color a Value formula The initial values ​​are derived from the connection weights between the output and hidden layers of the BP neural network in the fusion model, obtained after logistic regression transformation; constant term , These correspond to the predicted baseline values ​​of the fusion model when each factor is set to a baseline value. Then, the extracted initial weights are substituted into the preprocessed full dataset, and calibration is performed using the least squares method. The weight coefficients are fine-tuned to minimize the mean square error between the predicted and measured values, eliminating errors caused by model simplification. Simultaneously, for the coding of discrete variables, the assignment rules are optimized based on the prediction bias of the fusion model for different varieties and preprocessing methods, ensuring the matching between the coded values ​​and the weight coefficients. Finally, a t-test is used to verify the significance of each weight coefficient, eliminating insignificant coefficient terms. The consistency between the formula for constructing the calibrated weights and the prediction results of the fusion model is verified using the coefficient of determination R², requiring R² ≥ 0.88 to ensure the reliability of the weight coefficients.

[0044] The predictive model was validated and its reliability assessed through experiments. Three typical plateau rose varieties—Yunnan Double-Petaled Red Rose, Gansu Kushui Rose, and Qinghai Menyuan Rose—were selected. Twelve combined experiments were designed, involving different altitudes (2000m, 2800m, 3500m), maturity levels (3 and 5), and drying processes (vacuum freeze-drying and hot air drying). The deviations between actual quality indicators and model predictions were measured. Paired t-tests were used to verify the absence of significant differences between predicted and actual values ​​(P>0.05). The relative error (RE) was calculated, requiring that experimental groups with RE≤5% account for ≥90% of the total, thus validating the model's reliability in practical applications. For scenarios with significant errors during validation, 10 additional targeted experimental data sets were added to each scenario. The model was retrained, and weighting coefficients were adjusted to reduce systematic errors and ensure its applicability in the unique plateau environment.

[0045] S106. Based on the user's target needs, use prediction models and formulas to solve for the complete processing scheme of the optimal variety, planting altitude, maturity, processing method and drying process.

[0046] The system receives customized user requests, which must include at least the target product type, expected target values ​​or ranges for core quality indicators, and acceptable cost constraints. It then performs compliance checks on the received user request data, including verifying whether the indicators are within the model's supported range and whether the values ​​are within a reasonable range. After successful verification, the system converts the user's original target values ​​into a standardized data format that the model can process, and initiates the following two computational paths in parallel: Path 1: Input the standardized demand data into the fusion prediction model to obtain an output containing the prediction results and their confidence levels; Path 2 involves substituting the standardized demand data into the prediction formula to obtain the recommendation results; The decision is made based on the confidence level of the model output in Path 1. If the confidence level is ≥80%, the processing path corresponding to the model prediction result is adopted as the preferred recommended solution. If the confidence level is <80%, the processing path corresponding to the formula calculation result is adopted as the preferred recommended solution. The preferred recommended solution should include at least the recommended rose variety, planting altitude range, raw material maturity, pretreatment method, drying process parameters, and extraction process parameters. The generated optimized solution and its corresponding theoretical quality index prediction values ​​and key operation points are converted into graphic and textual forms and presented to the user intuitively.

[0047] In some implementations, WeChat mini-program functions are designed based on predictive models: with "simple Q&A + precise recommendation" as the core, three main functional modules are designed: The requirement input module allows users to specify the target ingredients, application scenarios, and cost constraints through drop-down selections or numerical input.

[0048] The recommendation module is optimized by calling the prediction model and prediction formula. With the goal of maximizing the content of the target component, the optimal combination of processing parameters is solved, and specific recommended values ​​for variety, planting altitude, maturity, pretreatment and drying process are output. The confidence of the prediction model recommendation results is evaluated. When the confidence of the model prediction result is ≥80%, the prediction model result is adopted first, and the formula is used as an auxiliary. When the confidence of the model prediction result is <80%, the output result of the prediction formula is taken as the standard.

[0049] Results Display Module: Displays the theoretical quality indicators, expected effects, and key operational points of the recommended solution in a graphic and textual format, and supports solution export and sharing.

[0050] The technical architecture adopts a three-tier architecture of "front-end-back-database". The front-end is developed based on the WeChat native framework to implement the user interface; the back-end uses the Spring Boot framework to build the server, integrates the prediction model algorithm and optimization solver, processes front-end requests and returns recommendation results; the database uses a combination of MySQL and Redis, with MySQL storing basic data and Redis caching frequently queried results to improve response speed.

[0051] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. An optimized processing method of highland fresh rose, characterized in that, include: Multi-source data on fresh highland roses were collected, and the multi-source data were preprocessed. The preprocessing included Z-score standardization of continuous variables, one-hot encoding of discrete variables, and K-nearest neighbor interpolation or targeted experiments to supplement missing data. Based on the processed multi-source data, the factors influencing the quality of fresh roses were analyzed, and core influencing factors were screened out. The analysis of factors influencing the quality of fresh roses included, with other variables fixed, analyzing the quantitative impact of a single variable on the quality of fresh roses; and using an orthogonal experimental design, combining multiple linear regression and XGBoost regression models to explore the interaction patterns of multiple factors. A five-level data relational database of plateau fresh roses was constructed, consisting of a variety information table, a planting condition table, a processing parameter table, a quality index table, and a correlation table. Based on the relational highland fresh rose database, prediction models for active ingredients and sensory quality indicators were constructed respectively. Among them, the prediction model for active ingredients of highland fresh rose includes a multiple linear regression model and a support vector regression model for predicting total phenols and total flavonoids, and a random forest model for predicting anthocyanins. The prediction model for sensory quality indicators of highland fresh rose includes a BP neural network for predicting color indicators and an XGBoost model for predicting aroma indicators. A Stacking ensemble learning strategy was adopted to fuse an active ingredient prediction model and a sensory quality index prediction model. Based on the fusion model, key influencing factors and their weights were extracted, and prediction formulas for active ingredients and sensory indicators were constructed. Based on the user's target needs, a complete processing solution is obtained by using prediction models and formulas to determine the optimal variety, planting altitude, maturity, processing method, and drying process.

2. The optimized processing method of high altitude fresh rose according to claim 1, wherein, The analysis of the quantitative impact of single-variable factors on the quality of fresh roses includes: With other variables fixed, we explored the quantitative effects of a single variable on the color, active ingredients, and aroma of fresh roses. Pearson correlation coefficient was used to perform correlation analysis on the target independent variable and quality index data of highland fresh roses to test normality. If the data conformed to a normal distribution, Pearson correlation analysis was performed directly. If the data did not conform to a normal distribution, the data was first transformed to conform to a normal distribution before analysis was performed. For discrete independent variables, first confirm the compatibility between the binary data after one-hot encoding of the discrete independent variables and the quality indicators, and then conduct correlation analysis. The correlation coefficient between the target independent variable and each quality indicator was calculated using the Pearson correlation coefficient formula. Based on the calculated correlation coefficient values, the direction of linear association between the target independent variable and each quality indicator was determined, and the level of association strength was judged. At the same time, the specific values ​​of the correlation coefficients of each association were recorded to form correlation analysis data.

3. The optimized processing method of high altitude fresh rose as claimed in claim 1, wherein, The process of uncovering the patterns of interaction among multiple factors includes: Using key variables as influencing factors, and combining the optimal value of a single influencing factor with the actual production feasibility to determine 0, an orthogonal experimental scheme is constructed. Conduct experiments according to the plan, with each group set up two or more repeated experiments to ensure data reliability. During the experiment, extraneous variables are strictly controlled and standardized operating procedures are followed. After the experiment was completed, the core quality indicators of each combination were measured, the raw data were recorded, and the data were preprocessed: outlier data exceeding 1.5 times the interquartile range were removed by box plot method; the mean and standard deviation of each group of repeated data were calculated, and if the coefficient of variation was >5%, the experiment was repeated to verify the stability of the data; the quality indicator data were standardized by Z-score to eliminate the influence of dimensions. Orthogonal experimental data were coded into variables. Discrete variables were converted into numerical values ​​using one-hot coding, while continuous variables retained the standardized data. Based on the coded variables, pairwise interaction terms are constructed, and the main effect terms of each variable are combined to form a set of independent variables. With a single quality index as the dependent variable, a multiple linear regression model is constructed using the stepwise regression method. The model is kept simple and effective by screening independent variables. After constructing the multiple linear regression model, a significance test is performed. The F-test is used to verify the overall significance of the model and ensure that the set of independent variables has significant explanatory power for the dependent variable. Significant interaction terms are screened by t-test, and interaction terms with no significant effect are removed. The model determination coefficient R² is calculated. If R² ≥ 0.6, the model fit is good. Otherwise, variable coding needs to be optimized, data supplemented, or the model needs to be reconstructed. Based on the preprocessed data, we constructed an XGBoost regression model with significant interaction terms and main effect terms as input features and quality indicators as output targets. After the model is trained, the importance score of each input feature is calculated based on the number of splits and split gain of the decision tree. The features are sorted from high to low and the top-ranked interaction items are selected as high-contribution combinations. The intensity of interaction effects is quantified by marginal effect analysis, interaction effect values ​​are calculated, synergistic or antagonistic relationships between variables are determined, and the interaction effect levels of each combination are clarified by classifying the absolute value of the effect value to the mean of the quality index.

4. The optimized processing method of high altitude fresh rose according to claim 1, wherein, The core influencing factors identified include: A two-dimensional evaluation index system is constructed. The contribution of a single factor is calculated based on the mean absolute value of the correlation coefficient between the variable and the quality index Pearson, reflecting the strength of the independent influence. The contribution of multiple factors in interaction is calculated based on the mean importance score of the high-contribution interaction combination involving the variable, reflecting the strength of the interaction contribution. Objective weights are determined using the entropy weight method, and preliminary subjective weights are replaced by index standardization, information entropy calculation, and entropy weight solution. Technical personnel in relevant fields conducted multiple rounds of scoring using the Delphi method until opinions converged, and the average score was taken as the subjective weight. The overall weight is calculated by combining objective and subjective weights in a 6:4 ratio. The top 4 variables with the highest overall weights are then selected to determine the final core influencing factors.

5. The optimized processing method for fresh highland roses according to claim 1, characterized in that, The predictive model for the active ingredients of fresh highland roses includes: For the three core active ingredients of fresh highland roses—total phenols, total flavonoids, and anthocyanins—exclusive predictive models were constructed to form a benchmark-complementary-precision system. The prediction model for total phenols and total flavonoids uses multiple linear regression as a benchmark, incorporating key factors such as variety, altitude and maturity. Variables are screened through stepwise regression to eliminate collinearity. A support vector regression model is constructed simultaneously, and a radial basis function kernel function is used to capture nonlinear correlations. The penalty parameter and kernel function parameter are optimized through grid search. The anthocyanin prediction model uses a random forest model, based on an ensemble of 100 decision trees. The training set is generated by bootstrap resampling, and 1 / 3 of the features of each tree are randomly selected for splitting. The maximum tree depth is set to 10-30 layers to improve the robustness of the model.

6. The optimized processing method for fresh highland roses according to claim 1, characterized in that, The sensory quality index prediction model for plateau fresh-eating roses includes: A backpropagation neural network for color indicators was constructed based on the characteristics of color indicators, employing a 6-input, 2-hidden-3-output architecture. The input layers correspond to six key factors: variety, altitude, maturity, pretreatment method, drying process, and extraction process. The hidden layers consist of 12→8 gradient nodes to extract non-linear correlation features, using the Sigmoid activation function. The output layer corresponds to the three core coordinate values ​​L in the color space of the color indicator. a b ; A gradient boosting decision tree XGBoost model for aroma indicators was constructed based on the characteristics of aroma indicators. The output is the total number of volatile substances plus the relative content of characteristic components. The input includes 6 key factors and gas chromatography-mass spectrometry auxiliary features. A step-by-step iteration with a learning rate of 0.05 was adopted, and the maximum tree depth was set to 8 layers, the minimum number of split samples was 4, and the column sampling rate was 0.8 to enhance the prediction sensitivity of low-content components. The training data covers multiple varieties, altitudes, and process scenarios. The parameters of the two models are optimized through grid search, with a focus on reducing the prediction error of the content components.

7. The optimized processing method for fresh highland roses according to claim 1, characterized in that, The fusion of the active ingredient prediction model and the sensory quality index prediction model includes: The Stacking ensemble learning strategy is adopted, using five basic models—multivariate linear regression, support vector regression, gradient boosting decision tree, random forest, and BP neural network—as the basic models and logistic regression as the meta-model. The prediction results of each basic model are weighted and fused to improve prediction accuracy. Based on the quality index characteristics of fresh highland roses, five basic models were adapted and preprocessed in a targeted manner. For the active ingredients total phenols and total flavonoids, which exhibit significant linear correlation, multiple linear regression and support vector regression models were retained. For the active ingredient anthocyanins, which are significantly affected by multi-factor nonlinearity, a random forest model was incorporated. For the sensory indicator color, which is easily affected by interference, a backpropagation neural network model was retained. For the complex indicator aroma, which contains trace components, a gradient boosting decision tree model was adopted. The five models were divided into multiple subsets for separate training. The active ingredient prediction subset included multiple linear regression, support vector regression, and random forest, while the sensory quality prediction subset included backpropagation neural network and gradient boosting decision tree. A 5-fold cross-validation strategy was used to train each basic model. In each round of validation, 4 / 5 of the training set data was used for modeling, and 1 / 5 of the data was used to generate prediction results, which served as input features for the meta-model. This process was iteratively repeated to generate prediction features for the entire training set. The test set directly generated prediction features using the trained basic models. The output results of each model are standardized to eliminate differences in the output dimensions of different models; correlation analysis is used to screen model combinations with a correlation of <0.7 in the prediction results to preserve model diversity and improve the integration effect. The Stacking ensemble learning framework is constructed by building a two-level ensemble framework consisting of a base model layer and a meta-model layer. This framework enhances the predictive capabilities of each model. Logistic regression is selected as the meta-model, and the predicted features generated by the base model layer are used as the input to the meta-model. The measured values ​​of the quality index are used as the output target. The training objective function of the meta-model is constructed, and weights are assigned based on the coefficient of determination R² of the validation sets of each base model. By introducing weighting coefficients, the prediction results of each base model are weighted and fused, thereby strengthening the contribution of high-quality models and weakening the influence of models with large errors. A hierarchical hyperparameter optimization system is constructed, employing a two-step strategy of coarse optimization via grid search and fine optimization via Bayesian optimization to accurately optimize the hyperparameters of the basic and meta-models. The basic model sets the optimization range according to the characteristics of the model, first determining the optimal range of parameters, and then using Bayesian optimization to fit the mapping relationship between parameters and performance to locate the optimal combination. The meta-model optimizes the regularization strength and number of iterations for logistic regression, and determines the parameters by combining 5-fold cross-validation.

8. The optimized processing method for fresh highland roses according to claim 1, characterized in that, The predictive formula for constructing active ingredients and sensory indicators includes: Feature importance analysis was conducted based on the optimized fusion model to screen for key active ingredients and color a. The core factors that significantly affect the value are the variety, planting altitude, maturity, drying temperature, and pretreatment method. The linear and nonlinear fitting results of the basic models of the multiple linear regression model and the support vector regression model in the active ingredient fusion model are finally extracted into a linear superposition formula structure after weighted fusion of the meta-model. Color a The nonlinear fitting characteristics of the BP neural network in the value fusion model are extracted into a formula structure in the form of logistic regression to match the changing pattern of the redness index. The training parameters of the fusion model provide the initial data source for the formula weight coefficients, enabling the transfer of the model's predictive ability to the formula, thus obtaining the prediction formula for key active ingredients in fresh roses: （1） In the formula, This is a key active ingredient in fresh highland roses. For constant terms; ~ These are the weight coefficients for each variable and interaction term; Variety codes are assigned, with different values ​​selected based on testing for different varieties; For planting altitude; For maturity; This refers to the drying temperature. Key interaction items; This is random error; Color index a The formula for predicting the rose redness of fresh roses from high-altitude areas: （2） In the formula, For color value; These are the logistic regression coefficients; Encode the preprocessing method; coefficient , The acquisition relies on the Stacking fusion model as its core, combining data calibration and statistical testing to achieve accurate solutions; utilizing the feature weight analysis function of the fusion model, the initial weights of each core factor and interaction term are extracted as... The initial value reflects the contribution intensity of each factor to the content of active ingredients; Color a Value formula The initial values ​​are derived from the connection weights between the output and hidden layers of the BP neural network in the fusion model, and are obtained after logistic regression transformation. constant term , These correspond to the predicted baseline values ​​of the fusion model when each factor takes a baseline value; The extracted initial weights are substituted into the preprocessed full dataset, and the least squares method is used for calibration. The weight coefficients are fine-tuned with the goal of minimizing the mean square error between the predicted and measured values ​​to eliminate the error caused by model simplification. For the encoding and assignment of discrete variables, the assignment rules are optimized by combining the prediction bias of the fusion model for different varieties and preprocessing methods to ensure the matching of the encoded value and the weight coefficient. The significance of each weight coefficient was verified by t-test, and insignificant coefficient terms were removed. The consistency between the formula for constructing the weights after calibration and the prediction results of the fusion model was verified by the coefficient of determination R², thus ensuring the reliability of the weight coefficients.

9. The optimized processing method for fresh highland roses according to claim 1, characterized in that, The preprocessing of the multi-source data includes: We collected literature data on fresh highland roses, production logs from processing enterprises, monitoring data from planting bases, and quality inspection reports from the market. The data dimensions covered rose raw materials, processing technology, and finished product quality. Rose raw materials included variety type, planting altitude, and maturity. Processing technology included pretreatment methods, drying processes, and extraction processes. Finished product quality included color indicators, active ingredients, and aroma components. Among these, the solvent type for variety type, pretreatment methods, drying processes, and extraction processes were discrete variables, while the spatial values ​​for planting altitude, processing parameters, active ingredient content, and color indicators were continuous variables. Maturity was a graded variable. The Z-score standardization method was used to normalize continuous variables and eliminate the influence of dimensions; the categories and values ​​of each discrete variable were sorted out and uniformly named. To count the number of categories of discrete variables, one-hot encoding is used to convert the n categories of variables into n binary variables. When a value is taken from a certain category, the corresponding variable is set to 1, and the rest are set to 0. Integrate the encoded data with standardized continuous variables to ensure a consistent format; In terms of missing value identification, the standardized dataset is scanned to record the location of missing values ​​and related background information; the missing rate of each variable is calculated, and the variables are divided into low missing rate and high missing rate according to the missing rate, and a missing rate statistical report is generated. Using the missing variable as the target variable and other variables as features, a random forest classification model is constructed. The missing pattern is determined by the feature importance score. If there is no significant correlation with other variables, it is considered random missing; if there is a strong correlation, it is considered non-random missing. For data with low missing values, K-nearest neighbor interpolation was used for imputation, and the optimal K value was determined through cross-validation. Euclidean distance was used as a measure of similarity between samples to calculate the distance between the sample containing the missing value and other non-missing samples. Based on the determined K value and the distance measurement results, the K nearest non-missing samples to the sample containing the missing value were selected as reference samples for interpolation. For continuous missing variables, the mean of the corresponding variable values ​​of the K nearest neighbor samples was used as the imputation value for the missing value. For discrete missing variables, the mode of the corresponding variable values ​​of the K nearest neighbor samples was used as the imputation value for the missing value. For key data with high missing rates, targeted experiments were designed to supplement the data, and matching materials were selected according to the missing background; independent variables such as process parameters and dependent variables such as finished product quality were set to cover the missing dimensions; each group was repeated 3-5 times; and the testing process was standardized to ensure that the results are reproducible.

10. The optimized processing method for fresh highland roses according to claim 1, characterized in that, The complete processing scheme for optimizing the variety, planting altitude, maturity, processing method, and drying process includes: Receive customized requirements from users, which include at least the target product type, the expected target value or range of core quality indicators, and acceptable cost constraints; Perform compliance verification on the received user request data, including checking whether the indicators are within the range supported by the model and whether the values ​​are within a reasonable range; After successful validation, the user-inputted raw target value is converted into a standardized data format that the model can process, and the following two computation paths are started in parallel: Path 1: Input the standardized demand data into the fusion prediction model to obtain an output containing the prediction results and their confidence levels; Path 2 involves substituting the standardized demand data into the prediction formula to obtain the recommendation result; The decision is made based on the confidence level of the model output in Path 1. If the confidence level is ≥80%, the processing path corresponding to the model prediction result is adopted as the preferred recommended solution; if the confidence level is <80%, the processing path corresponding to the formula calculation result is adopted as the preferred recommended solution. The preferred selection scheme should include at least the recommended rose variety, planting altitude range, raw material maturity, pretreatment method, drying process parameters and extraction process parameters; The generated optimization scheme, its corresponding theoretical quality index predictions, and key operational points are converted into graphic and textual formats and presented to users intuitively.

Images