Method for analyzing sensory effect of aroma components of euphausia superba seasoning

Abstract

The application provides a method for analyzing the sensory effect of Euphausia superba flavoring aroma components, and relates to the technical field of food processing, comprising the following steps: preparing Euphausia superba flavoring samples and detecting volatile components; evaluating the sensory attributes of the Euphausia superba flavoring samples; performing OAV screening of the volatile components based on a dynamic threshold adjustment mechanism; performing standardization processing, enhancement and division on the volatile component data and the sensory attribute data; constructing an AMSO integrated model based on a random forest model and an XGBoost model and performing training; performing SHAP-based feature importance analysis and feature aroma component screening on each volatile component; constructing an aroma component-sensory attribute mapping relationship by using the feature aroma components screened based on OAV, the feature aroma components screened based on SHAP and the SHAP values corresponding to the feature aroma components; and applying the trained AMSO integrated model and the aroma component-sensory attribute mapping relationship to the production process of Euphausia superba flavoring.

Description

Technical Field

[0001] This invention relates to the field of food processing technology, and in particular to a method for analyzing the sensory effects of aroma components in Antarctic krill seasoning. Background Technology

[0002] Antarctic krill (Euphausia superba) is the most abundant single biological resource on Earth, with an annual catch of 60-100 million tons. Rich in high-quality protein, unsaturated fatty acids, flavor peptides, and free amino acids, it is an ideal raw material for preparing natural seasonings. However, the processing of Antarctic krill generates a complex system of volatile components, typically containing hundreds of volatile compounds across eight major categories: aldehydes, ketones, alcohols, esters, pyrazines, furans, sulfur-containing compounds, and nitrogen-containing compounds. These components work together through complex synergistic and antagonistic interactions to determine the characteristic flavor of the product. Accurately identifying key aroma markers from this vast library of volatile components and quantitatively analyzing the contribution of each component to specific sensory attributes has become a critical scientific problem that urgently needs to be solved in the industrial production of Antarctic krill seasonings.

[0003] Currently, the detection of volatile components in food mainly relies on instrumental analysis techniques such as headspace solid-phase microextraction-gas chromatography-mass spectrometry (HS-SPME-GC-MS) and gas chromatography-olfactometry-mass spectrometry (GC-O-MS), which can identify 50-200 volatile compounds in a single experiment. Traditional methods for screening characteristic aroma components are mainly based on the odor activity value (OAV) theory, which is the ratio of the concentration of volatile components to their sensory threshold (OAV=C / T). It is generally believed that components with OAV>1 directly contribute to the overall aroma, and the higher the OAV value, the more significant the contribution. However, this method has the following five limitations: (1) The problem of missing threshold data is serious. Currently, the volatile components whose sensory thresholds have been determined account for only 50-70% of the components detected in food. A large number of novel compounds or components in complex matrices lack threshold data, making it impossible to calculate their OAV values. As a result, they are ignored in the screening process and may miss important potential aroma markers. (2) Difficulty in identifying synergistic and antagonistic effects. Traditional OAV theory is based on the independent contribution of a single component and does not consider the interaction between volatile components. Studies have shown that some components with low OAV values ​​(OAV<1) may significantly affect the overall flavor through synergistic effects with other components. For example, trace amounts of sulfur-containing compounds (dimethyl sulfide, dimethyl disulfide, etc.) can significantly enhance meat and umami characteristics even though their OAV values ​​are close to 1. In addition, some components with high OAV values ​​may be masked by other components due to antagonistic effects, and their actual contribution is lower than theoretically expected. Traditional methods are unable to capture these complex nonlinear interactions; (3) The fixed threshold standard is not adaptable enough. Traditional OAV screening usually uses a fixed threshold (such as OAV≥1 or OAV≥10) as the screening standard, without taking into account the differences in the overall concentration distribution of volatile components caused by changes in different batches of raw materials and processing conditions. When the freshness of the raw materials is high and the overall concentration of volatile components is low, the fixed threshold may be too strict, resulting in the omission of important components; conversely, when the processing intensity is high and the concentration of volatile components generally increases, the fixed threshold may be too lenient, introducing a large number of noise components and reducing the specificity and accuracy of screening. (4) Traditional multivariate statistical methods have limited interpretability. Although multivariate statistical methods such as principal component analysis (PCA) and orthogonal partial least squares discriminant analysis (OPLS-DA) can identify the differential components between samples, their output results (such as principal component loadings and VIP values) mainly reflect the contribution of components to sample classification. They cannot directly quantify the degree and direction (promotion or inhibition) of each component's contribution to specific sensory attributes (such as baking aroma, fishy smell, sweet aroma, etc.), resulting in an ambiguous correspondence between screening results and sensory evaluation. (5) Insufficient robustness and generalization ability of the model. Food flavor data are usually characterized by high dimensionality, small sample size, and high noise. Single statistical or machine learning models are easily affected by data fluctuations, outliers, and overfitting, resulting in poor reproducibility of screening results and limited applicability to different batches of samples. In addition, existing methods lack interpretability analysis of model prediction results. The model is regarded as a "black box" and it is difficult to reveal the molecular mechanism of flavor formation, which limits its application and promotion in industrial production. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides an analytical method for the sensory effects of aroma components in Antarctic krill seasonings. By integrating odor activity value (OAV) theory, dynamic threshold adaptive adjustment mechanism, multi-model AMSO ensemble model, and SHAP (SHapley Additive exPlanations) interpretability analysis, the method aims to accurately screen characteristic aroma substances from a large number of volatile components and quantitatively analyze their contribution to specific sensory attributes.

[0005] To achieve the above objectives, the present invention provides a method for analyzing the sensory effects of aroma components in Antarctic krill seasoning, comprising the following steps: S1. Prepare Antarctic krill seasoning samples and use headspace solid-phase microextraction-gas chromatography-mass spectrometry to detect the volatile components of each sample, obtaining a volatile component data matrix X∈ containing compound name, CAS number, chemical classification, retention time, retention index RI, mass spectrometry matching degree, and relative content. , where n is the number of samples and m is the number of volatile components; S2. Establish a sensory evaluation team and use quantitative descriptive analysis to evaluate the sensory attributes of Antarctic krill seasoning samples, including sensory attribute data matrix Y∈, which includes roasted aroma, fishy smell, sweet aroma, umami, and amine flavor. , where p is the number of sensory attributes; S3. Perform OAV screening based on dynamic threshold adjustment mechanism on volatile components to obtain a preliminary set of characteristic aroma components A; S4. Standardize, enhance, and divide the volatile component data matrix X and the sensory attribute data matrix Y to obtain the training set, validation set, and test set; S5. Construct and train an AMSO ensemble model based on a random forest model and an XGBoost model. S6. Perform SHAP-based feature importance analysis and feature aroma component screening for each volatile component to obtain the SHAP-screened feature aroma component set B; S7. Using the characteristic aroma component set A based on OAV screening and the characteristic aroma component set B based on SHAP screening, as well as the SHAP values ​​corresponding to the characteristic aroma components, construct the aroma component-sensory attribute mapping relationship. S8. The trained AMSO ensemble model and aroma component-sensory attribute mapping relationship are applied to the production process of Antarctic krill seasoning.

[0006] Optionally, step S2 includes: S2.1, Establishment of the Sensory Evaluation Team

[0007] Select 10-15 evaluators aged 20-50, without olfactory impairment, and sensitive to seafood flavors, with a balanced male-to-female ratio; The basic olfactory sensitivity test and the three-point test method are used to screen evaluators to ensure that they have good olfactory discrimination ability; Evaluators will undergo 2-4 weeks of training in sensory descriptive terms, including the definition of sensory attributes, smelling of reference standard samples, and the use of intensity scales. Consistency tests will be conducted to ensure the reliability of the evaluation results.

[0008] S2.2 Determination of Sensory Attribute Vocabulary

[0009] Through preliminary experiments and group discussions, the sensory attribute vocabulary of Antarctic krill seasoning was determined using the free choice descriptive method and the consensus vocabulary method. The sensory attribute vocabulary includes roasted aroma, fishy smell, sweet aroma, umami, and amine flavor. For each sensory attribute, provide reference standard samples with 3-5 concentration gradients to establish an intensity scale;

[0010] S2.3 Implementation of Sensory Evaluation

[0011] Sample preparation: Antarctic krill seasoning samples were removed from refrigeration 24 hours before evaluation and brought to room temperature 2 hours before evaluation. 5-10 mL of sample was weighed into a 50 mL brown sample cup with a lid and labeled with a three-digit random code to avoid psychological suggestion from the evaluator. Evaluation environment: The evaluation was conducted in a sensory evaluation room that meets the ISO 8589 standard. The room temperature was controlled at 20-25℃, the relative humidity was 50-70%, there were no odor interferences, and the lighting was natural light or fluorescent light. The evaluators evaluated independently without communicating with each other. Evaluation process: Each evaluation provides 4-6 samples, which are presented in a randomized or balanced design. The evaluator scores the intensity of each sensory attribute by smell, using a continuous linear scale of 0-10. Each sample is evaluated 3 times, with an interval of at least 10 minutes between each evaluation. Water and unsalted biscuits are provided during the evaluation process to eliminate odor interference from the previous sample. Each sample is rested for 2-3 minutes after evaluation. Data Processing: Analysis of variance was performed on the scores from all evaluators to test the significance of differences between samples, between evaluators, and between replicates. Outliers were removed, and the mean score and standard error for each sample on different sensory attributes were calculated. A sensory attribute data matrix Y∈ , where p is the number of sensory attributes;

[0012] S2.4 Data Quality Control: Two-way ANOVA is used to assess the evaluators' discrimination ability and repeatability, calculate the consistency among evaluators, and eliminate the data of evaluators with poor discrimination ability or poor repeatability to ensure the reliability of sensory evaluation results.

[0013] Optionally, step 3 includes: S3.1 Data Collection and Threshold Query: Sensory thresholds Ti of known volatile components are obtained by consulting aroma compound databases and relevant literature, and a threshold database is established; S3.2 Concentration Quantification or Semi-Quantification: The volatile components are quantitatively or semi-quantitatively analyzed using the external standard method, internal standard method, or peak area normalization method to obtain the concentration Ci of each volatile component in the sample; S3.3 Calculate the Odor Activity (OAV) value for each known volatile component using the following formula:

[0014] S3.4 Dynamic Threshold Setting: Statistically analyze the OAV value distribution of known volatile components in all samples and calculate the mean. and standard deviation Set dynamic thresholds:

[0015] Where k is the adjustment coefficient;

[0016] S3.5, Preliminary Screening: For OAV ≥ The volatile components were screened to form a preliminary set of characteristic aroma components, A.

[0017] Optionally, step S4 includes: S4.1 Data Standardization: Z-score standardization is performed on the volatile component data matrix X and the sensory attribute data matrix Y to eliminate the influence of differences in concentration magnitudes between different components and differences in scoring scales between different sensory attributes. The standardization formula is:

[0018]

[0019] S4.2 Data Augmentation: The original samples are resampled and augmented with replacement using the bootstrap method to generate augmented datasets, thereby increasing the amount of data for model training and improving generalization ability; S4.3 Dataset partitioning: After merging the original samples and enhanced samples, they are randomly divided into training set, validation set and test set according to a preset ratio.

[0020] Optionally, step S5 includes: S5.1 Random Forest Model Construction

[0021] Hyperparameter settings: Number of decision trees ∈[100, 500], maximum tree depth ∈[5, 15], minimum number of split samples ∈[2, 10], minimum number of leaf node samples ∈[1, 5], the largest eigenvalue ∈['sqrt', 'log2', None], the optimal parameter combination is determined by grid search combined with k-fold cross-validation, and the evaluation index is the cross-validation determination coefficient CVR²; Model training: using the standardized relative content of volatile components As input features, the intensity of each sensory attribute As the target variable, p independent random forest models are constructed, and the models are trained on the training set. The training time and computational complexity are recorded. Model validation: Evaluate the predictive performance of the random forest model on the validation set and calculate metrics such as the coefficient of determination R², root mean square error RMSE, and mean absolute error MAE.

[0022] S5.2, XGBoost Model Construction

[0023] Hyperparameter settings: Number of trees ∈[100, 500], learning rate ∈[0.01, 0.3], maximum tree depth Subsample ∈ [3, 10], subsample ∈ [0.6, 1.0], column sampling ratio ∈[0.6, 1.0], with regularization parameters lambda∈[0, 10] and alpha∈[0, 10]. The Bayesian optimization algorithm is used to optimize the hyperparameters, with the optimization objective being to maximize the validation set R². Model training: using the standardized relative content of volatile components As input features, the intensity of each sensory attribute As the target variable, p independent XGBoost models are constructed, and the early_stopping_rounds parameter is set. Training is stopped early when the performance on the validation set does not improve within a certain number of consecutive rounds to prevent overfitting. Model validation: Evaluate the predictive performance of the XGBoost model on the validation set and calculate metrics such as the coefficient of determination R², root mean square error RMSE, and mean absolute error MAE.

[0024] S5.3, Multi-model AMSO integrated model: Weighted average ensemble: A dynamic weighted average method based on validation set performance is used to integrate the prediction results of the random forest and XGBoost models. The prediction values ​​of the AMSO ensemble model are:

[0025]

[0026]

[0027] Among them, w RF and w XGBoost Let w be the weight coefficients of the random forest model and the XGBoost model, respectively, satisfying w RF +w XGBoost =1, and These are the outputs of the Random Forest model and the XGBoost model, respectively. and These are the coefficients of determination for the Random Forest model and the XGBoost model on the validation set, respectively.

[0028] Model performance evaluation: The performance of the AMSO ensemble model was evaluated on an independent test set. The determination coefficient R², root mean square error RMSE, mean absolute error MAE, and mean absolute percentage error MAPE of the AMSO ensemble model were calculated. The superiority of the AMSO ensemble model was verified by comparing it with the random forest model, the XGBoost model, and traditional multivariate statistical methods.

[0029] Optionally, step 6 includes: S6.1 SHAP Value Calculation

[0030] The KernelSHAP algorithm is employed, and weighted linear regression is used to approximate the SHAP value: First, multiple subsets of characteristic aroma components are randomly generated in the characteristic aroma component space; then, the model prediction value is calculated for each subset of characteristic aroma components, and different weights are assigned according to the size of the subset; finally, the SHAP value of each characteristic aroma component is solved by weighted linear regression. The formula for calculating the SHAP value φi is as follows:

[0031] Where N is the set of all characteristic aroma components, S is the subset of characteristic aroma components that does not contain characteristic aroma component i, f(S) is the model prediction based on the subset S of characteristic aroma components, |S| is the size of the subset S, and |N| is the total number of characteristic aroma components.

[0032] S6.2 Global SHAP Feature Importance Analysis For each volatile component, the average absolute value of its SHAP values ​​across all samples is calculated as the global importance score for that volatile component:

[0033] Where φi,j is the SHAP value of the i-th component in the j-th sample; For all volatile components according to Sort the features in descending order to construct a global ranking of SHAP feature importance;

[0034] S6.3 Determination of SHAP Threshold and Screening of Feature Components

[0035] Calculate the SHAP threshold, and then perform SHAP filtering on each sensory attribute according to the SHAP threshold to obtain a set of p feature components. By merging the characteristic component sets of all sensory attributes, we obtain the characteristic aroma component set screened by SHAP. ;

[0036] S6.4 Integration of Comprehensive Screening Results

[0037] The final set of characteristic aroma components C = A∪B is obtained by taking the union of the characteristic aroma component set A based on OAV screening and the characteristic aroma component set B based on SHAP screening.

[0038] Optionally, step 7 includes: (1) For each sensory attribute, extract the corresponding SHAP value, sort them in descending order of absolute value, and obtain the Top N contribution components of the attribute; (2) Generate an aroma component mapping spectrum that includes component name, CAS number, chemical classification, relative content, OAV value, SHAP value, and aroma description; (3) The positive or negative value of SHAP indicates whether the volatile components have a promoting or inhibiting effect on sensory properties; (4) The nonlinear relationship and interaction between the concentration of characteristic components and the intensity of sensory attributes were analyzed by using SHAP dependency graph.

[0039] Optionally, step 8 includes: (1) Quality monitoring: By detecting the volatile components of the product, the intensity of its sensory attributes is predicted using the AMSO integrated model, thereby achieving rapid evaluation of the product's flavor quality and replacing time-consuming and labor-intensive sensory evaluation. (2) Flavor control: Based on the target flavor requirements, the key aroma components that need to be controlled and their target concentration ranges are identified by SHAP contribution ranking; (3) Process optimization: Combining the generation pathways and influencing factors of key aroma components, reverse calculations are made to determine the optimization direction of processing parameters, guiding the precise control of the production process; (4) Formula design: Based on the AMSO integrated model, the sensory properties of different ingredient combinations are predicted to realize the intelligent design and optimization of Antarctic krill seasoning formula.

[0040] After adopting the above technical solution, the beneficial effects of the present invention are as follows: 1. This invention proposes for the first time an intelligent mapping method for food aroma components based on the AMSO (Adaptive Multi-model Synergistic Optimization) machine learning framework. It integrates the prior knowledge of OAV theory, dynamic threshold adaptive adjustment mechanism, multi-model AMSO ensemble model and SHAP interpretability analysis, and overcomes the five major limitations of traditional methods (threshold missing, difficulty in identifying synergistic effects, insufficient adaptability of fixed threshold, limited interpretability, and poor robustness). It achieves the goal of accurately screening characteristic aroma substances from more than 100 volatile components and quantitatively analyzing their contribution to specific sensory attributes. 2. The dynamic threshold adjustment mechanism introduced in this invention can adaptively adjust the OAV screening criteria according to the changes in different batches of Antarctic krill raw materials and processing conditions, which significantly improves the robustness and adaptability of the method in actual production. 3. The multi-model AMSO ensemble model based on dynamic weighting of the validation set R² adopted in this invention fully leverages the complementary advantages of the random forest and XGBoost models, significantly improving the accuracy and stability of predictions. 4. Based on the SHAP framework, this invention can not only screen out directly contributing components with high OAV values, but also identify low-threshold components and undetermined threshold components that affect flavor through synergistic effects, thus expanding the coverage of aroma component screening. The positive or negative SHAP value can also reveal the promoting or inhibiting effect of components on sensory attributes. 5. The AMSO integrated model established in this invention can be directly applied to the production line of Antarctic krill seasoning. By detecting the volatile component composition of the product, the model can predict the intensity of sensory attributes, thereby achieving rapid evaluation of flavor quality. Based on the target flavor requirements, the key aroma components that need to be regulated can be identified through SHAP contribution ranking, guiding the optimization of processing parameters. 6. The method of this invention is not only applicable to the aroma analysis of Antarctic krill seasonings, but can also be extended to the flavor analysis of other marine biological resources (such as salmon, tuna, scallops, etc.), meat products (such as ham, sausage, dried meat, etc.), fermented foods (such as soy sauce, broad bean paste, shrimp paste, etc.), and other complex food systems. By adjusting the model parameters and sensory attribute definitions, it can adapt to the characteristic flavor evaluation needs of different foods. In addition, the AMSO framework can be extended to other sensory dimensions (such as taste, texture, etc.) and quality indicators (such as nutritional components, functional activities, etc.) to construct a multi-dimensional intelligent food quality evaluation system. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 Heatmap of sensory properties of Antarctic krill seasoning samples; Figure 2 A comparison chart of the prediction performance of the Random Forest model, XGBoost model, and AMSO ensemble model on the test set; Figure 3 Heatmap of SHAP contribution of 20 characteristic volatile components to 5 sensory attributes. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] In recent years, machine learning techniques have gradually emerged in the field of food science, with ensemble learning algorithms such as Random Forest (RF) and Extreme Gradient Boosting (XGBoost) demonstrating excellent performance in classification and regression tasks. Random Forest constructs multiple decision trees through bootstrap sampling and random feature selection strategies, exhibiting advantages such as strong resistance to overfitting, insensitivity to missing values, and the ability to assess feature importance. XGBoost, based on the Gradient Boosting Decision Tree (GBDT) framework, iteratively builds decision trees to gradually reduce prediction errors, introduces regularization terms to control model complexity, and employs second-order Taylor expansion to improve optimization accuracy, demonstrating superior performance in handling high-dimensional sparse data and nonlinear relationships. However, RF and XGBoost each have their advantages and disadvantages: RF tends to provide stable prediction results but has relatively weak fine-grained fitting capabilities; XGBoost can finely fit complex nonlinear relationships in data but is prone to overfitting under small sample conditions. A single model cannot simultaneously guarantee both prediction accuracy and robustness.

[0045] SHAP (SHapley Additive exPlanations), an interpretability framework based on Shapley values ​​in game theory, quantifies the marginal contribution of each feature to the model's prediction results, providing a transparent explanation mechanism for "black box" models. The core idea of ​​SHAP is to decompose the model's prediction results into the sum of the marginal contributions of each input feature. The SHAP value of each feature represents the degree and direction of its influence on the prediction result. The calculation of SHAP values ​​is based on the fair allocation principle of Shapley values. By traversing all possible feature combinations, the change in the model's prediction value before and after adding a certain feature is calculated, and a weighted average is performed according to the size of the combination. SHAP not only provides a global ranking of feature importance but also provides explanations of feature contributions at the local sample level, with theoretical guarantees such as consistency, local accuracy, and handling of missing features. Although SHAP has been widely used in computer vision, natural language processing, and other fields, its application in food flavor analysis is still in its early stages. There are no research reports on integrating SHAP with OAV theory and multi-model AMSO ensemble model systems for the mapping of aroma components in complex food systems.

[0046] Furthermore, existing research lacks an adaptive optimization mechanism for the screening process of volatile components. In actual production, factors such as the harvesting season, freshness, and storage conditions of Antarctic krill feedstock can lead to changes in the baseline concentration levels and distribution characteristics of volatile components. Adjustments to processing parameters (enzymatic hydrolysis time, temperature, pH, heat treatment intensity, etc.) can also significantly alter the formation and transformation pathways of volatile components. Traditional methods, employing fixed screening criteria and model parameters, cannot adapt to such dynamic changes, resulting in significant fluctuations in screening results across different batches, making it difficult to meet the robustness and consistency requirements of industrial production.

[0047] To this end, an intelligent aroma component mapping technology integrating OAV theory prior knowledge, multi-model collaborative decision-making, dynamic threshold adaptive adjustment, and interpretability analysis was developed to achieve precise analysis and control of the flavor quality of Antarctic krill seasonings, and to promote the standardization and intelligent manufacturing process of the products.

[0048] This invention provides a method for analyzing the sensory effects of aroma components in Antarctic krill seasoning, comprising the following steps:

[0049] S1. Prepare Antarctic krill seasoning samples and detect volatile components.

[0050] S1.1 Preparation of Antarctic krill seasoning samples

[0051] Fresh Antarctic krill (Euphausia superba) caught in the Antarctic waters between January and February 2024 were selected as raw materials. They were immediately flash-frozen at -40°C on board the ship and then frozen and stored at -18°C during transport to the laboratory. After thawing (and refrigerated overnight at 4°C), the krill shells and internal organs were removed, and they were washed clean with deionized water, drained, and homogenized to make krill paste (material-to-liquid ratio 1:0, i.e., pure krill meat).

[0052] Krill paste was divided into four groups according to different processing methods, and three biological parallel samples were prepared for each group: (1) Trypsin hydrolysis group for 4h (numbered T4-1, T4-2, T4-3): Take 500g of krill paste, add an equal volume of phosphate buffer (50mM, pH7.5), add 0.2% (w / w, based on protein) of trypsin (enzyme activity ≥250U / mg), and hydrolyze in a 50℃ water bath for 4h, stirring every 30min during the process to ensure uniform hydrolysis; after the hydrolysis is completed, inactivate the enzyme in a 100℃ water bath for 10min. (2) Alkaline protease hydrolysis group for 4h (numbered A4-1, A4-2, A4-3): Take 500g of krill paste, add an equal volume of phosphate buffer (50mM, pH9.0), add 0.2% (w / w) of alkaline protease (enzyme activity ≥200U / mg), and hydrolyze in a 50℃ water bath for 4h, stirring every 30min during the process; after the hydrolysis is completed, inactivate the enzyme in a 100℃ water bath for 10min. (3) Baking at 120℃ (R120-1, R120-2, R120-3): Take 500g of krill paste, place it on a baking tray with a thickness of about 1.0-1.5cm, and bake in a 120℃ hot air oven for 20 minutes, turning it over every 5 minutes to ensure even heating; during the baking process, the moisture gradually evaporates, and the final moisture content drops to about 50-60%; (4) Lactic acid bacteria fermentation 24h group (numbered LF24-1, LF24-2, LF24-3): Take 500g of krill paste, inoculate with Streptococcus thermophilus (inoculation amount 10^6 CFU / g), mix evenly and ferment in a constant temperature incubator at 37℃ for 24h. During this period, the pH value is measured every 6h (from the initial pH 6.8 to the final pH 4.5-5.0).

[0053] After all samples were processed, 15% (w / w) sodium chloride (NaCl, analytical grade) was added, and the mixture was concentrated in a 60°C water bath to a solid content of 40-45% (measured using a handheld refractometer). The mixture was then homogenized (10,000 rpm, 2 min), filled into sterile glass vials, and stored at 4°C for later use. Strict aseptic techniques were maintained throughout the sample preparation process to avoid contamination by exogenous microorganisms.

[0054] S1.2, Detection of volatile components

[0055] (1) Sample pretreatment: Accurately weigh 5.0g of Antarctic krill seasoning sample into a 20mL headspace vial, add internal standard 2-methyl-3-heptanone (concentration 20μg / kg), and seal the vial.

[0056] (2) SPME extraction: A 50 / 30μm DVB / CAR / PDMS three-phase extraction head was used. Before extraction, the headspace vial was aged at 250℃ for 30 min to remove residues. The headspace vial was placed on a 60℃ constant temperature magnetic stirrer at 500 rpm for equilibration for 15 min. Then the extraction head was inserted into the headspace and extracted for 40 min. After extraction, the extraction head was immediately inserted into the GC inlet and desorbed at 260℃ for 4 min. The desorption mode was splitless injection.

[0057] (3) GC-MS conditions: Instruments: Agilent 7890B gas chromatograph-5977A mass spectrometer; Chromatographic column: HP-5MS capillary column (30m×0.25mm×0.25μm); Carrier gas: high-purity helium (99.999% purity), constant flow mode, flow rate 1.2 mL / min; Inlet temperature: 260℃, splitless injection, septum purge flow rate: 3 mL / min; Temperature rise program: Initial temperature 40℃ held for 3 min, then rise to 140℃ at 4℃ / min (held for 0 min), then rise to 240℃ at 8℃ / min (held for 8 min), total run time 49 min; Transmission line temperature: 280℃; MS conditions: Ionization mode: EI; ionization energy: 70 eV; ion source temperature: 230℃; quadrupole temperature: 150℃; scan range: m / z 35-450; scan mode: Full Scan; solvent delay: 5 min.

[0058] (4) Data processing: Data acquisition and processing were performed using Agilent MassHunter software; volatile components were identified by matching with the NIST 17 mass spectrometry library (matching degree >80%), and confirmed by retention index (RI) (RI deviation <20); retention index was calculated using C7-C30 n-alkane standards; the relative content (%) of each component was calculated using the peak area normalization method.

[0059] (5) Detection results: A total of 120 volatile components were identified, including 26 aldehydes, 18 ketones, 24 alcohols, 15 esters, 12 acids, 10 pyrazines, 10 furans, and 5 sulfur-containing compounds. A volatile component data matrix X∈ was obtained, containing compound name, CAS number, chemical classification, retention time, retention index (RI), mass spectrometry matching degree, and relative content. 12 represents the number of samples, and 120 represents the number of volatile components.

[0060] S2. Sensory evaluation and quantitative description of properties of Antarctic krill seasoning samples.

[0061] S2.1, Establishment of the Sensory Evaluation Team

[0062] Twelve evaluators (6 men and 6 women, aged 25-45) with no olfactory impairment and sensitivity to seafood flavors will be selected. All evaluators passed the basic olfactory sensitivity test (isovaleric acid threshold <1.0mg / L) and the three-point test consistency test (accuracy >75%) to ensure that the evaluators have good olfactory discrimination ability; Evaluators will undergo 2-4 weeks of training in sensory descriptive terms, including the definition of sensory attributes, smelling of reference standard samples, and the use of intensity scales. Consistency tests will be conducted to ensure the reliability of the evaluation results.

[0063] S2.2 Determination of Sensory Attribute Vocabulary

[0064] Through preliminary experiments and group discussions, using a free-choice descriptive method and a consensus-based vocabulary method, sensory attribute vocabulary for Antarctic krill seasoning was determined. The sensory attribute vocabulary includes: (1) Roasted / toasted aroma: similar to the aroma of toasted bread, nuts, caramel, and cocoa, mainly contributed by Maillard reaction products (such as pyrazines, furans, and aldehydes). The reference standard is the aroma of roasted malt or roasted sesame seeds. (2) Fishy smell: The typical fishy and metallic smell of marine fish, mainly contributed by trimethylamine, lipid oxidation products (such as aldehydes and ketones), and sulfur-containing compounds. The reference standard is the smell of fresh sea fish or seaweed. (3) Sweet: The pleasant sweet aroma of flowers and fruits, mainly contributed by esters, aromatic aldehydes and alcohols. The reference standard is the aroma of apple, banana and rose. (4) Umami: A rich and delicious taste (the contribution of aroma to umami is evaluated by smell), mainly contributed by sulfur-containing compounds (dimethyl sulfide, methylthiopropionaldehyde, etc.), nitrogen-containing compounds, nucleotides and amino acid-related aromas. The reference standard is the aroma of chicken soup and mushrooms. (5) Amine odor: A pungent odor similar to ammonia or putrid smell, mainly contributed by trimethylamine, indole, and lower amines. The reference standard is a diluted ammonia solution. For each sensory attribute, provide reference standard samples with 3-5 concentration gradients to establish an intensity scale;

[0065] S2.3, Implementation of Sensory Evaluation

[0066] Sample preparation: Antarctic krill seasoning samples were taken out of the refrigeration condition (4℃) 24 hours before evaluation and brought to room temperature (20-25℃) 2 hours before evaluation. 8 mL of the sample was weighed into a 50 mL brown sample cup with a lid and labeled with a three-digit random code to avoid psychological suggestion from the evaluator. Evaluation environment: The evaluation was conducted in a sensory evaluation room that meets the ISO 8589 standard. The room temperature was controlled at 20-25℃, the relative humidity was 50-70%, there were no odor interferences, and the lighting was natural light or fluorescent lamps (illuminance >1000 lux). The evaluators evaluated independently without communicating with each other. Evaluation process: Each evaluation provides 4-6 samples, presented in a randomized or balanced design. Evaluators score the intensity of each sensory attribute by smell, using a continuous linear scale of 0-10 (0 represents no perception or very weak, 10 represents very strong). Each sample is evaluated 3 times (over 3 days), with a 10-minute interval between each evaluation. During the evaluation process, clean water and unsalted biscuits are provided to eliminate odor interference from the previous sample. Each sample is rested for 2-3 minutes after evaluation. Data Processing: Analysis of variance was performed on the scores from all evaluators to test the significance of differences between samples, between evaluators, and between replicates. Outliers were removed, and the mean score and standard error for each sample on different sensory attributes were calculated. A sensory attribute data matrix Y∈ ;

[0067] S2.4 Data Quality Control: The ability and repeatability of the evaluators are assessed by two-way ANOVA, the consistency among evaluators is calculated, and the data of evaluators with poor discrimination ability (P<0.05) or poor repeatability are eliminated (Cronbach's α=0.86) to ensure the reliability of the sensory evaluation results.

[0068] The sensory evaluation results of the 12 samples are as follows (mean ± standard error): Baking aroma: T4 group 5.3±0.3^c, A4 group 6.7±0.2^b, R120 group 9.1±0.2^a, LF24 group 6.1±0.3^bc (different letters indicate significant differences, P<0.05, the same below); Fishy odor: T4 group 5.1±0.2^a, A4 group 4.3±0.2^ab, R120 group 2.9±0.2^c, LF24 group 3.9±0.2^b; Sweet aroma: T4 group 4.9±0.2^c, A4 group 5.7±0.2^b, R120 group 6.6±0.2^a, LF24 group 7.7±0.2^a; Umami: T4 group 8.1±0.2^a, A4 group 8.1±0.2^a, R120 group 7.4±0.2^b, LF24 group 7.9±0.2^ab; Amine odor: T4 group 3.9±0.2^a, A4 group 3.0±0.2^b, R120 group 1.9±0.1^c, LF24 group 2.0±0.1^c; Sensory evaluation results showed that the baking group (R120) had the highest roasted aroma and the lowest fishy and amine aroma; the fermentation group (LF24) had the highest sweet aroma; and the enzymatic hydrolysis groups (T4 and A4) had higher umami flavor. Detailed data can be found in [link to data]. Figure 1 .

[0069] S3. Perform OAV screening on volatile components based on a dynamic threshold adjustment mechanism.

[0070] S3.1 Data Collection and Threshold Query: Sensory threshold data of 87 out of 120 volatile components were obtained by consulting Flavornet, Pherobase databases and related literature (using water or oil as the medium).

[0071] S3.2 Concentration Quantification or Semi-Quantification: A standard curve is established using the internal standard method (2-methyl-3-heptanone) to perform absolute quantification of volatile components and obtain the concentration (μg / kg) of each component in the sample; for volatile components for which a standard curve has not been established, the relative content is calculated using the peak area normalization method, and semi-quantification is performed assuming the detector response factor is 1.

[0072] S3.3 Calculate the Odor Activity (OAV) values ​​of 87 volatile components with known sensory thresholds, using the following formula:

[0073] A total of 87 × 12 = 1044 OAV values ​​were obtained from 12 samples;

[0074] S3.4 Dynamic Threshold Setting: Statistically analyze the distribution of 1044 OAV values ​​and calculate the mean. =6.8, standard deviation =11.3, calculate the dynamic threshold:

[0075] Where k is an adjustment coefficient, ranging from 0.5 to 2.0, preferably from 0.8 to 1.2, and most preferably 1.0; when the OAV distribution varies significantly between sample batches, the dynamic threshold can adaptively adjust the screening criteria, enhancing the robustness of the method; specifically, when Higher (e.g.) When the threshold is >10), the dynamic threshold automatically increases, focusing on the most significant contributing components; when Lower (e.g.) When <5), the dynamic threshold automatically decreases to avoid missing potentially important components;

[0076] Set the adjustment coefficient k=1.0, and calculate the dynamic threshold. ;

[0077] S3.5 OAV Screening Results: A total of 42 volatile components with OAV≥1 were screened out, forming the initial set of characteristic aroma components A;

[0078] S4. Standardize, enhance, and segment the volatile component data matrix X and the sensory attribute data matrix Y.

[0079] S4.1 Data Standardization: Z-score standardization is performed on the volatile component data matrix X and the sensory attribute data matrix Y to eliminate the influence of differences in concentration magnitudes between different components and differences in scoring scales between different sensory attributes. The standardization formula is:

[0080]

[0081] S4.2 Data Augmentation: The bootstrap method was used to resample and augment the 12 original samples with replacement. The resampling was repeated 100 times to generate 100 augmented samples. The original samples and augmented samples were merged to obtain a dataset of 112 samples. The bootstrap method can increase the number of training samples without changing the data distribution by random sampling with replacement, while preserving the statistical characteristics of the original data.

[0082] S4.3 Dataset partitioning: The dataset of 112 samples is randomly divided into a training set (for model parameter training), a validation set (for hyperparameter tuning and model selection), and a test set (for independent performance evaluation) in a ratio of 7:2:1.

[0083] S5. Construct and train an AMSO ensemble model based on a random forest model and an XGBoost model.

[0084] S5.1 Random Forest Model Construction and Training

[0085] Random forest model construction: Using the standardized relative content of 120 volatile components as input features and the intensity of 5 sensory attributes as target variables, 5 independent random forest models were constructed using Python 3.8 and the scikit-learn library (version 1.0.2). Hyperparameter settings and optimization: Number of decision trees ∈[100, 500], maximum tree depth ∈[5, 15], minimum number of split samples ∈[2, 10], minimum number of leaf node samples ∈[1, 5], the largest eigenvalue ∈['sqrt', 'log2', None]; The optimal parameter combination is determined using grid search combined with 5-fold cross-validation (k=3-5): number of decision trees. =300, maximum tree depth =10, minimum number of split samples =5, minimum number of leaf node samples =2; Model validation: The predictive performance of the random forest model was evaluated on the validation set. The calculated values ​​were: roasted aroma R² = 0.81, fishy aroma R² = 0.78, sweet aroma R² = 0.75, umami aroma R² = 0.80, amine aroma R² = 0.77, and average R² = 0.78.

[0086] S5.2, XGBoost model construction: XGBoost Model Construction: Using the standardized relative content of 120 volatile components as input features and the intensity of 5 sensory attributes as target variables, 5 independent XGBoost models were constructed using Python 3.8 and the xgboost library (version 1.5.1). Hyperparameter settings and optimization: Number of trees ∈[100, 500], learning rate ∈ [0.01, 0.3] (also known as shrinkage or eta, controlling the contribution weight of each tree), maximum tree depth ∈[3, 10], subsample ∈[0.6, 1.0] (row sampling ratio), column sampling ratio Given the parameters ∈ [0.6, 1.0], regularization parameters lambda ∈ [0, 10] (L2 regularization) and alpha ∈ [0, 10] (L1 regularization); Bayesian optimization is used to optimize hyperparameters, with the objective function being to maximize the validation set R², to determine the optimal parameter combination: the number of trees. =350, learning rate =0.08, maximum tree depth =7, subsample ratio = 0.85, column sampling ratio =0.80; Model validation: The predictive performance of the XGBoost model was evaluated on the validation set. The calculated R² values ​​were: roasted aroma R² = 0.84, fishy aroma R² = 0.80, sweet aroma R² = 0.78, umami aroma R² = 0.83, amine aroma R² = 0.79, and average R² = 0.81.

[0087] S5.3, Multi-model AMSO integrated model: Weighted average ensemble: The prediction results of the random forest and XGBoost models are integrated using a dynamic weighted average method based on the validation set R². The prediction values ​​of the AMSO ensemble model are:

[0088]

[0089]

[0090] Among them, w RF and w XGBoost The weights of the Random Forest and XGBoost models are respectively, satisfying w RF +w XGBoost =1, and These are the outputs of the Random Forest model and the XGBoost model, respectively. and These are the coefficients of determination for the Random Forest model and the XGBoost model on the validation set, respectively. Taking baking aroma as an example, w RF =0.81 / (0.81+0.84)=0.491, w XGBoost =0.84 / (0.81+0.84)=0.509, and the weight coefficients of the other four sensory attributes are calculated similarly; Model validation: The predictive performance of the AMSO ensemble model was evaluated on the validation set. The calculated values ​​were: roasted aroma R² = 0.87, fishy aroma R² = 0.83, sweet aroma R² = 0.81, umami aroma R² = 0.86, amine aroma R² = 0.82, and average R² = 0.84. The performance of the AMSO ensemble model was evaluated on an independent test set. Its superiority was verified by comparing it with random forest and XGBoost models. Figure 2 As shown; Advantages of the AMSO ensemble model: Random Forest provides stable prediction results through a voting mechanism of multiple independent decision trees and has strong robustness to outliers and noisy data, but its fine-fitting ability is relatively weak; XGBoost can finely fit the complex nonlinear relationships of data through a gradient boosting serial learning mechanism and has a strong ability to capture the interaction between features, but it is prone to overfitting under small sample conditions; Through dynamic weighted ensemble, the accuracy of model prediction is guaranteed (by utilizing the fine-fitting ability of XGBoost) while reducing the risk of overfitting of a single model (by utilizing the robustness of random forest), so that the AMSO ensemble model can still maintain excellent performance under small sample conditions.

[0091] S6. Perform SHAP-based eigenvalue importance analysis and eigenvalue screening for each volatile component.

[0092] S6.1 Theoretical basis of SHAP value:

[0093] SHAP, based on the Shapley value concept in game theory, decomposes the model prediction result into the sum of the marginal contributions of each input feature. For the i-th feature, aroma component, its SHAP value φi is calculated as follows:

[0094] Where N is the set of all characteristic aroma components, S is the subset of characteristic aroma components that does not contain characteristic aroma component i, f(S) is the model prediction based on the subset S of characteristic aroma components, |S| is the size of the subset S, and |N| is the total number of characteristic aroma components. This formula calculates the change in the model's predicted value before and after adding the characteristic aroma component i by iterating through all possible combinations of characteristic aroma components, and performs a weighted average according to the size of the combination, thereby fairly distributing the contribution of each feature.

[0095] S6.2 SHAP Value Calculation

[0096] The SHAP analysis of the AMSO ensemble model was performed using the shap library (version 0.41.0) in Python. Due to the large amount of computation, 50 representative samples from the training set were selected for SHAP analysis. For each sensory attribute of the AMSO ensemble model, the SHAP value matrix (50×120) of 120 volatile components on the 50 samples was calculated.

[0097] S6.3, Global SHAP Feature Importance Analysis

[0098] For each volatile component, the mean absolute SHAP of its values ​​across 50 samples is calculated as the global importance score for that volatile component.

[0099] Where φi,j is the SHAP value of the i-th component in the j-th sample. The larger the Global_SHAP_i, the more significant the contribution of the component to the sensory attributes.

[0100] For all ingredients according to Sort the features in descending order to construct a global ranking of SHAP feature importance;

[0101] Taking baking aromas as an example, the top 10 volatile components and their SAP values ​​are as follows: 1,2,3,5-Trimethylpyrazine (SHAP=0.842) 2. 2-Methylbutyraldehyde (SHAP=0.798) 3. 3-Methylbutyraldehyde (SHAP=0.765) 4. 2,5-Dimethylpyrazine (SHAP=0.731) 5. 2-Pentylfuran (SHAP=0.682) 6. Furfural (SHAP=0.651) 7. 2,3-Dimethylpyrazine (SHAP=0.598) 8. Phenylacetaldehyde (SHAP=0.567) 9. 2-Ethyl-3,5-dimethylpyrazine (SHAP=0.542) 10. Octal (SHAP=0.515)

[0102] S6.4 Determination of SHAP Threshold and Screening of Feature Components

[0103] The SHAP threshold is calculated using the cumulative contribution rate method: the sum of Global_SHAP_i for all components is calculated. The components are arranged in descending order according to Global_SHAP, and then added one by one until the cumulative contribution rate reaches a preset value (usually 80%-90%, preferably 85%). The corresponding component set is the characteristic aroma component set B based on SHAP screening. Another method for calculating the SHAP threshold is to calculate the Global_SHAP_i mean for all components. and standard deviation ,set up Wherein, α is an adjustment coefficient, with a value ranging from 0.5 to 2.0, preferably 1.0; The SHAP threshold was calculated using the top 85% cumulative contribution rate method. Based on the SHAP threshold, the number of characteristic components screened out for each of the five sensory attributes was calculated as follows: 28 roasted aromas, 26 fishy aromas, 30 sweet aromas, 25 umami aromas, and 22 amine aromas. After merging and removing duplicates, a total of 65 components were screened out by at least one sensory attribute (set B).

[0104] S6.5 Integration of Comprehensive Screening Results

[0105] The union of the characteristic aroma component set A (42 types) screened based on OAV and the characteristic aroma component set B (65 types) screened based on SHAP is taken to obtain the final characteristic aroma component set C = A∪B, which contains 71 components. This set includes three important aroma components: Category 1 (22 types): Components with high OAV value and high SHAP contribution (A∩B). These components directly determine the main flavor of Antarctic krill seasoning and are core aroma markers that can be identified by both traditional methods and machine learning methods. Category 2 (20 types): Components with low OAV values ​​(or OAV values ​​that do not reach the dynamic threshold) but high SHAP contributions (B\A). Although the concentration of these components is close to or below the sensory threshold, they have a significant impact on the overall flavor through synergistic effects with other components. They are key components that are easily overlooked by traditional OAV methods. The third category (29 types): components with high SHAP contribution but whose threshold was not determined. These components have often been overlooked in previous studies, but the AMSO ensemble model has revealed their potential importance through machine learning and interpretability analysis, providing clues for the discovery of novel aroma biomarkers.

[0106] S7. Construct a mapping relationship between aroma components and sensory attributes.

[0107] For each sensory attribute, the corresponding SHAP value is extracted, and the values ​​are sorted in descending order of absolute value to obtain the top 10 contributing components of that attribute. The top 10 contributing factors for each sensory attribute are listed below: Top 10 Baking Aromas: 1,2,3,5-Trimethylpyrazine (SHAP=+0.842) → cocoa aroma, roasted aroma 2. 2-Methylbutyraldehyde (SHAP=+0.798) → Malt aroma, roasted aroma 3. 3-Methylbutyraldehyde (SHAP=+0.765) → Chocolate aroma, toasty aroma 4. 2,5-Dimethylpyrazine (SHAP=+0.731) → Nutty and roasted aroma 5. 2-Pentylfuran (SHAP=+0.682) → Bean and nutty aromas 6. Furfural (SHAP=+0.651) → Caramel aroma, bread aroma 7. 2,3-Dimethylpyrazine (SHAP=+0.598) → Nutty and roasted aroma 8. Phenylacetaldehyde (SHAP=+0.567) → Rose scent, sweet scent 9. 2-Ethyl-3,5-dimethylpyrazine (SHAP=+0.542) → roasted, potato-like aroma 10. Octaldehyde (SHAP=+0.515) → Citrus and fatty aromas

[0108] Top 10 fishy smells: 1. Trimethylamine (SHAP=+0.891) → fishy smell, ammonia smell 2. 1-Octen-3-ol (SHAP=+0.762) → Mushroom aroma, metallic taste 3. 2,4-Heptadienal (SHAP=+0.683) → Fatty oxidized odor, fishy odor 4. 2,4-Decadienal (SHAP=+0.621) → Oily odor, fishy odor 5. Indole (SHAP=+0.584) → Fecal odor, floral scent 6. 1-Octen-3-one (SHAP=+0.547) → Metallic taste, mushroom taste 7. Heptanal (SHAP=+0.512) → fatty aroma, green aroma 8. Octaldehyde (SHAP=+0.489) → Citrus and fatty aromas 9. Nonanal (SHAP=+0.465) → Citrus aroma, fatty aroma 10. Hexanal (SHAP=+0.442) → Grassy and fatty aroma

[0109] The list of the top 10 contributing ingredients for sweetness, umami, and amine flavors is not shown; please see below for details. Figure 3 .

[0110] (2) Generate an aroma component mapping spectrum that includes component name, CAS number, chemical classification, relative content, OAV value, SHAP value, and aroma description; (3) The positive or negative SHAP value indicates whether the component promotes or inhibits the sensory attribute: a positive SHAP value indicates that the component enhances the intensity of the sensory attribute, and a negative SHAP value indicates that the component reduces the intensity of the sensory attribute. (4) Analyze the nonlinear relationship and interaction between the concentration of characteristic components and the intensity of sensory attributes using the SHAP dependence plot;

[0111] S8. The trained AMSO ensemble model and aroma component-sensory attribute mapping relationship are applied to the production process of Antarctic krill seasoning. (1) Quality monitoring: By detecting the volatile components of the product, the intensity of its sensory attributes is predicted using the AMSO integrated model, thereby achieving rapid evaluation of the product's flavor quality and replacing time-consuming and labor-intensive sensory evaluation. (2) Flavor control: Based on the target flavor requirements (such as enhancing the roasted aroma and reducing the fishy smell), the key aroma components that need to be controlled and their target concentration ranges are identified by SHAP contribution ranking. (3) Process optimization: Based on the generation pathways and influencing factors of key aroma components, reverse the calculation of the optimization direction of processing parameters (such as enzymatic hydrolysis time, temperature, pH, heat treatment temperature and time, fermentation conditions, etc.) to guide the precise control of the production process. (4) Formula design: Based on the AMSO integrated model, the sensory properties of different ingredient combinations are predicted to realize the intelligent design and optimization of Antarctic krill seasoning formula.

[0112] Although the present invention has been disclosed above with reference to embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications and refinements without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A method for analyzing the sensory effects of aroma components in Antarctic krill seasoning, characterized in that, Includes the following steps: S1. Prepare Antarctic krill seasoning samples and use headspace solid-phase microextraction-gas chromatography-mass spectrometry to detect the volatile components of each sample, obtaining a volatile component data matrix X∈ containing compound name, CAS number, chemical classification, retention time, retention index RI, mass spectrometry matching degree, and relative content. , where n is the number of samples and m is the number of volatile components; S2. Establish a sensory evaluation team and use quantitative descriptive analysis to evaluate the sensory attributes of Antarctic krill seasoning samples, including sensory attribute data matrix Y∈, which includes roasted aroma, fishy smell, sweet aroma, umami, and amine flavor. , where p is the number of sensory attributes; S3. Perform OAV screening based on dynamic threshold adjustment mechanism on volatile components to obtain a preliminary set of characteristic aroma components A; S4. Standardize, enhance, and divide the volatile component data matrix X and the sensory attribute data matrix Y to obtain the training set, validation set, and test set; S5. Construct and train an AMSO ensemble model based on a random forest model and an XGBoost model. S6. Perform SHAP-based eigenvalue importance analysis and characteristic aroma component screening for each volatile component to obtain the SHAP-screened characteristic aroma component set. ; S7. Using the characteristic aroma component set A based on OAV screening and the characteristic aroma component set B based on SHAP screening, as well as the SHAP values ​​corresponding to the characteristic aroma components, construct the aroma component-sensory attribute mapping relationship. S8. The trained AMSO ensemble model and aroma component-sensory attribute mapping relationship are applied to the production process of Antarctic krill seasoning.

2. The method for analyzing the sensory effects of aroma components in Antarctic krill seasoning according to claim 1, characterized in that, Step S2 includes: S2.1, Establishment of the Sensory Evaluation Team Select 10-15 evaluators aged 20-50, without olfactory impairment, and sensitive to seafood flavors, with a balanced male-to-female ratio; The basic olfactory sensitivity test and the three-point test method are used to screen evaluators to ensure that they have good olfactory discrimination ability; Evaluators will undergo 2-4 weeks of training in sensory descriptive terms, including the definition of sensory attributes, smelling of reference standard samples, and the use of intensity scales. Consistency tests will be conducted to ensure the reliability of the evaluation results. S2.2 Determination of Sensory Attribute Vocabulary Through preliminary experiments and group discussions, the sensory attribute vocabulary of Antarctic krill seasoning was determined using the free choice descriptive method and the consensus vocabulary method. The sensory attribute vocabulary includes roasted aroma, fishy smell, sweet aroma, umami, and amine flavor. For each sensory attribute, provide reference standard samples with 3-5 concentration gradients to establish an intensity scale; S2.3 Implementation of Sensory Evaluation Sample preparation: Antarctic krill seasoning samples were removed from refrigeration 24 hours before evaluation and brought to room temperature 2 hours before evaluation. 5-10 mL of sample was weighed into a 50 mL brown sample cup with a lid and labeled with a three-digit random code to avoid psychological suggestion from the evaluator. Evaluation environment: The evaluation was conducted in a sensory evaluation room that meets the ISO 8589 standard. The room temperature was controlled at 20-25℃, the relative humidity was 50-70%, there were no odor interferences, and the lighting was natural light or fluorescent light. The evaluators evaluated independently without communicating with each other. Evaluation process: Each evaluation provides 4-6 samples, which are presented in a randomized or balanced design. The evaluator scores the intensity of each sensory attribute by smell, using a continuous linear scale of 0-10. Each sample is evaluated 3 times, with an interval of at least 10 minutes between each evaluation. Water and unsalted biscuits are provided during the evaluation process to eliminate odor interference from the previous sample. Each sample is rested for 2-3 minutes after evaluation. Data Processing: Analysis of variance was performed on the scores from all evaluators to test the significance of differences between samples, between evaluators, and between replicates. Outliers were removed, and the mean score and standard error for each sample on different sensory attributes were calculated. A sensory attribute data matrix Y∈ , where p is the number of sensory attributes; S2.4 Data Quality Control: Two-way ANOVA is used to assess the evaluators' discrimination ability and repeatability, calculate the consistency among evaluators, and eliminate the data of evaluators with poor discrimination ability or poor repeatability to ensure the reliability of sensory evaluation results.

3. The method for analyzing the sensory effects of aroma components in Antarctic krill seasoning according to claim 1, characterized in that, Step 3 includes: S3.1 Data Collection and Threshold Query: Sensory thresholds Ti of known volatile components are obtained by consulting aroma compound databases and relevant literature, and a threshold database is established; S3.2 Concentration Quantification or Semi-Quantification: The volatile components are quantitatively or semi-quantitatively analyzed using the external standard method, internal standard method, or peak area normalization method to obtain the concentration Ci of each volatile component in the sample; S3.3 Calculate the Odor Activity (OAV) value for each known volatile component using the following formula: S3.4 Dynamic Threshold Setting: Statistically analyze the OAV value distribution of known volatile components in all samples and calculate the mean. and standard deviation Set dynamic thresholds: Where k is the adjustment coefficient; S3.5, Preliminary Screening: For OAV ≥ The volatile components were screened to form a preliminary set of characteristic aroma components, A.

4. The method for analyzing the sensory effects of aroma components in Antarctic krill seasoning according to claim 1, characterized in that, Step S4 includes: S4.1 Data Standardization: Z-score standardization is performed on the volatile component data matrix X and the sensory attribute data matrix Y to eliminate the influence of differences in concentration magnitudes between different components and differences in scoring scales between different sensory attributes. The standardization formula is: S4.2 Data Augmentation: The original samples are resampled and augmented with replacement using the bootstrap method to generate augmented datasets, thereby increasing the amount of data for model training and improving generalization ability; S4.3 Dataset partitioning: After merging the original samples and enhanced samples, they are randomly divided into training set, validation set and test set according to a preset ratio.

5. The method for analyzing the sensory effects of aroma components in Antarctic krill seasoning according to claim 1, characterized in that, Step S5 includes: S5.1 Random Forest Model Construction Hyperparameter settings: Number of decision trees ∈[100, 500], maximum tree depth ∈[5, 15], minimum number of split samples ∈[2, 10], minimum number of leaf node samples ∈[1, 5], the largest eigenvalue ∈['sqrt', 'log2', None], the optimal parameter combination is determined by grid search combined with k-fold cross-validation, and the evaluation index is the cross-validation determination coefficient CVR²; Model training: using the standardized relative content of volatile components As input features, the intensity of each sensory attribute As the target variable, p independent random forest models are constructed, and the models are trained on the training set. The training time and computational complexity are recorded. Model validation: Evaluate the predictive performance of the random forest model on the validation set and calculate metrics such as the coefficient of determination R², root mean square error RMSE, and mean absolute error MAE. S5.2, XGBoost Model Construction Hyperparameter settings: Number of trees ∈[100, 500], learning rate ∈[0.01, 0.3], maximum tree depth Subsample ∈ [3, 10], subsample ∈ [0.6, 1.0], column sampling ratio ∈[0.6, 1.0], with regularization parameters lambda∈[0, 10] and alpha∈[0, 10]. The Bayesian optimization algorithm is used to optimize the hyperparameters, with the optimization objective being to maximize the validation set R². Model training: using the standardized relative content of volatile components As input features, the intensity of each sensory attribute As the target variable, p independent XGBoost models are constructed, and the early_stopping_rounds parameter is set. Training is stopped early when the performance on the validation set does not improve within a certain number of consecutive rounds to prevent overfitting. Model validation: Evaluate the predictive performance of the XGBoost model on the validation set and calculate metrics such as the coefficient of determination R², root mean square error RMSE, and mean absolute error MAE. S5.3, Multi-model AMSO integrated model: Weighted average ensemble: A dynamic weighted average method based on validation set performance is used to integrate the prediction results of the random forest and XGBoost models. The prediction values ​​of the AMSO ensemble model are: Among them, w RF and w XGBoost Let w be the weight coefficients of the random forest model and the XGBoost model, respectively, satisfying w RF +w XGBoost =1, and These are the outputs of the Random Forest model and the XGBoost model, respectively. and These are the coefficients of determination for the Random Forest model and the XGBoost model on the validation set, respectively. Model performance evaluation: The performance of the AMSO ensemble model was evaluated on an independent test set. The determination coefficient R², root mean square error RMSE, mean absolute error MAE, and mean absolute percentage error MAPE of the AMSO ensemble model were calculated. The superiority of the AMSO ensemble model was verified by comparing it with the random forest model, the XGBoost model, and traditional multivariate statistical methods.

6. The method for analyzing the sensory effects of aroma components in Antarctic krill seasoning according to claim 1, characterized in that, Step 6 includes: S6.1 SHAP Value Calculation The KernelSHAP algorithm is employed, and weighted linear regression is used to approximate the SHAP value: First, multiple subsets of characteristic aroma components are randomly generated in the characteristic aroma component space; then, the model prediction value is calculated for each subset of characteristic aroma components, and different weights are assigned according to the size of the subset; finally, the SHAP value of each characteristic aroma component is solved by weighted linear regression. The formula for calculating the SHAP value φi is as follows: Where N is the set of all characteristic aroma components, S is the subset of characteristic aroma components that does not contain characteristic aroma component i, f(S) is the model prediction based on the subset S of characteristic aroma components, |S| is the size of the subset S, and |N| is the total number of characteristic aroma components. S6.2 Global SHAP Feature Importance Analysis For each volatile component, the average absolute value of its SHAP values ​​across all samples is calculated as the global importance score for that volatile component: Where φi,j is the SHAP value of the i-th component in the j-th sample; For all volatile components according to Sort the features in descending order to construct a global ranking of SHAP feature importance; S6.3 Determination of SHAP Threshold and Screening of Feature Components Calculate the SHAP threshold, and then perform SHAP filtering on each sensory attribute according to the SHAP threshold to obtain a set of p feature components. By merging the characteristic component sets of all sensory attributes, we obtain the characteristic aroma component set screened by SHAP. ; S6.4 Integration of Comprehensive Screening Results The final set of characteristic aroma components C = A∪B is obtained by taking the union of the characteristic aroma component set A based on OAV screening and the characteristic aroma component set B based on SHAP screening.

7. The method for analyzing the sensory effects of aroma components in Antarctic krill seasoning according to claim 1, characterized in that, Step 7 includes: (1) For each sensory attribute, extract the corresponding SHAP value, sort them in descending order of absolute value, and obtain the Top N contribution components of the attribute; (2) Generate an aroma component mapping spectrum that includes component name, CAS number, chemical classification, relative content, OAV value, SHAP value, and aroma description; (3) The positive or negative value of SHAP indicates whether the volatile components have a promoting or inhibiting effect on sensory properties; (4) The nonlinear relationship and interaction between the concentration of characteristic components and the intensity of sensory attributes were analyzed by using SHAP dependency graph.

8. The method for analyzing the sensory effects of aroma components in Antarctic krill seasoning according to claim 1, characterized in that, Step 8 includes: (1) Quality monitoring: By detecting the volatile components of the product, the intensity of its sensory attributes is predicted using the AMSO integrated model, thereby achieving rapid evaluation of the product's flavor quality and replacing time-consuming and labor-intensive sensory evaluation. (2) Flavor control: Based on the target flavor requirements, the key aroma components that need to be controlled and their target concentration ranges are identified by SHAP contribution ranking; (3) Process optimization: Combining the generation pathways and influencing factors of key aroma components, reverse calculations are made to determine the optimization direction of processing parameters, guiding the precise control of the production process; (4) Formula design: Based on the AMSO integrated model, the sensory properties of different ingredient combinations are predicted to realize the intelligent design and optimization of Antarctic krill seasoning formula.

Images