Distilled liquor mixed drinking scheme design method and system based on sensory data

Abstract

The invention relates to the technical field of distilled liquor, discloses a distilled liquor mixed drinking scheme design method and system based on sensory data, and aims at solving the problem that an existing distilled liquor mixed drinking scheme design mode is poor in accuracy. The scheme mainly comprises the following steps: collecting sensory evaluation data of a user on each base liquor of the distilled liquor and a plurality of mixed drinking schemes under different drinking conditions, and obtaining mixed drinking data of the plurality of mixed drinking schemes; constructing and training a mixed drinking flavor prediction model according to the mixed drinking data; and constructing a mixed drinking scheme design model based on an evolutionary algorithm and the mixed drinking flavor prediction model, obtaining target sensory data, target base wine types and target drinking conditions of the user, inputting the target sensory data, the target base wine types and the target drinking conditions into the mixed drinking scheme design model, and obtaining a mixed drinking scheme design result through iterative optimization. The method improves the accuracy, individuation and innovativeness of mixed drinking scheme design, and is particularly suitable for mixed drinking development of distilled liquor.

Description

Technical Field

[0001] This invention relates to the field of distilled spirits technology, and specifically to a method and system for designing distilled spirits blending schemes based on sensory data. Background Technology

[0002] Distilled spirits, such as baijiu, whiskey, and brandy, can no longer be enjoyed neat to meet increasingly diverse consumer demands. Mixing has become an important direction for product innovation and enhancing the consumer experience. Currently, the design of mixed spirit drinking solutions mainly relies on two methods:

[0003] The first category is the rational design method based on chemical composition analysis. This method uses instrumental analysis techniques such as gas chromatography-mass spectrometry (GC-MS) to determine the composition and content of volatile flavor compounds in the wine, and then infers the flavor characteristics after blending. However, this method has significant limitations: First, the relationship between the superposition of chemical components and sensory perception is not a simple linear one; complex synergistic or antagonistic effects make it difficult to accurately predict flavor. Second, instrumental testing is expensive and time-consuming, making it difficult to respond quickly to market changes. Finally, this method completely fails to capture and reflect consumers' subjective preferences and emotional experiences, leading to designs that may deviate from market demands.

[0004] The second category is traditional optimization methods based on consumer sensory evaluations or expert experience. This method collects sensory scores (e.g., 1-5 points) from sommeliers or consumers on different blended samples under specific conditions, and then uses statistical analysis or simple optimization algorithms to generate recommended solutions. While this method incorporates subjective perception to some extent, it still faces several bottlenecks: First, sensory data collection typically relies on fixed scoring forms, requiring users to input according to pre-defined structured indicators. This rigid format fails to effectively capture the diverse and ambiguous feedback generated by users in real-world scenarios, such as natural language or speech (e.g., "smooth on the palate but a bit spicy aftertaste," "want more fruitiness, not too stimulating," etc.). Second, traditional predictive models (such as linear regression) struggle to accurately characterize the complex mapping relationships between base spirit types, blending ratios, drinking conditions, and complex sensory results, making it impossible to accurately predict blended flavors. Ultimately, this results in poor accuracy of the generated personalized distilled spirit blending solutions. Summary of the Invention

[0005] This invention aims to address the problem of poor accuracy in existing distilled spirits blending scheme design methods, and proposes a distilled spirits blending scheme design method and system based on sensory data.

[0006] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0007] In a first aspect, the present invention provides a method for designing a distilled spirits blending scheme based on sensory data, the method comprising:

[0008] Sensory evaluation data of users on various base spirits of distilled spirits and multiple blending schemes under different drinking conditions are collected to obtain blending data of multiple blending schemes. The sensory evaluation data is obtained by structuring and standardizing the user's natural language text or speech text through a language big data model. The blending data includes the types of base spirits in the blending scheme, the blending ratio of each base spirit in the blending scheme, the sensory evaluation data of each base spirit under different drinking conditions, and the sensory evaluation data of the blending scheme under different drinking conditions.

[0009] Based on the mixed drinking data, a mixed drinking flavor prediction model is constructed and trained. The mixed drinking flavor prediction model is used to predict the sensory data of the corresponding mixed drinking scheme under the drinking conditions based on the types of base wines, the mixing ratio of each base wine and the drinking conditions in the mixed drinking scheme.

[0010] A blending scheme design model is constructed based on evolutionary algorithms and blending flavor prediction models. The target sensory data, target base wine types, and target drinking conditions of users are obtained and input into the blending scheme design model. Through iterative optimization, the blending scheme design results are obtained.

[0011] Furthermore, the sensory evaluation data and predicted sensory data include eleven categories of sensory indicators, including blending, smoothness, likability, fruity aroma, grain aroma, floral aroma, sweetness, sourness, bitterness, smoky or oak barrel aroma, and irritation or alcoholic aroma.

[0012] The target sensory data includes ten categories of sensory indicators: blending, smoothness, fruity aroma, grain aroma, floral aroma, sweetness, sourness, bitterness, smokiness or oak barrel aroma, and irritation or alcoholic aroma.

[0013] Furthermore, the drinking conditions and target drinking conditions include room temperature, adding ice for 15 seconds, and the ice melting by half after being added;

[0014] The base spirits include baijiu, whiskey, rum, and brandy.

[0015] Furthermore, the design model for the mixed beverage scheme employs a fitness function for iterative optimization, specifically including:

[0016] Multiple candidate blending schemes are randomly generated based on the user's target base wine. Each candidate blending scheme includes the type of base wine, drinking conditions, and the blending ratio of each base wine.

[0017] For each candidate mixed beverage scheme, the mixed beverage flavor prediction model is invoked to make a prediction, and the predicted sensory data of each candidate mixed beverage scheme is obtained. The target sensory data and the predicted sensory data are substituted into the fitness function to calculate the fitness value of each candidate mixed beverage scheme.

[0018] Evolutionary iteration is performed based on the fitness values ​​of each candidate mixing scheme. Selection, crossover, and mutation are used to generate a new generation of candidate mixing schemes, and the fitness values ​​of the new generation of candidate mixing schemes are recalculated.

[0019] Once the number of iterations reaches the threshold, one or more candidate mixed drinking schemes with fitness values ​​higher than the fitness threshold are taken as the design results of the mixed drinking scheme.

[0020] Furthermore, the formula for the fitness function is:

[0021] ;

[0022] in, Represents the fitness value. This represents the weighting coefficient, with a value range of [0, 1]. L This indicates the liking index in the predicted sensory data. This represents the sensory error, which is the sum of weighted errors of corresponding sensory indicators between the target sensory data and the predicted sensory data. Indicates to The scaling constant is used for normalization. This indicates a constraint or penalty.

[0023] Furthermore, the constraint penalty term For matching and constraint penalties Alcohol content range restriction penalties Sparsity penalty The sum of these terms is calculated using the following formulas:

[0024] ;

[0025] ;

[0026] ;

[0027] ;

[0028] in, Indicates the first candidate mixed beverage regimen The mixing ratio of base spirits, This indicates the number of different types of base spirits in the candidate blending recipes, and ABV represents the alcohol content. Indicates the drinking condition correction factor. and Indicates the upper and lower limits of alcohol content. , and These represent the corresponding weighting coefficients.

[0029] Furthermore, the evolutionary algorithm employs one or more of the following: differential genetic algorithm, evolutionary algorithm, particle swarm optimization algorithm, simulated annealing algorithm, or multi-objective optimization algorithm.

[0030] Furthermore, the flavor prediction model for mixed drinks is implemented using a deep neural network model, including one or more combinations of one-dimensional convolutional neural networks, recurrent neural networks, or fully connected neural networks.

[0031] Furthermore, after obtaining the design results of the mixed beverage scheme, the method further includes:

[0032] Receive users' personal preference ratings and natural language feedback on the design results of the mixed drink solution, and use a large language model to transform them into structured indicators to form user historical preference data;

[0033] When generating the design results of the mixed drink scheme, personalized mixed drink schemes are selected based on the user's historical preference data, and corresponding natural language explanations are generated.

[0034] Secondly, the present invention provides a system for designing distilled spirits blending schemes based on sensory data, used to implement the method for designing distilled spirits blending schemes based on sensory data described in the first aspect, the system comprising:

[0035] The mixed drinking data acquisition unit is used to collect sensory evaluation data of users on various base spirits of distilled spirits and multiple mixed drinking schemes under different drinking conditions, and to obtain mixed drinking data of multiple mixed drinking schemes. The sensory evaluation data is obtained by structuring and standardizing the user's natural language text or speech text through a language big data model. The mixed drinking data includes the types of base spirits in the mixed drinking scheme, the mixing ratio of each base spirit in the mixed drinking scheme, the sensory evaluation data of each base spirit under different drinking conditions, and the sensory evaluation data of the mixed drinking scheme under different drinking conditions.

[0036] The blending flavor prediction unit is constructed and trained based on the blending data. It is used to predict the sensory data of the corresponding blending scheme under the drinking conditions based on the types of base wines, the blending ratio of each base wine and the drinking conditions in the blending scheme.

[0037] The blending scheme design unit is used to construct a blending scheme design model based on evolutionary algorithms and blending flavor prediction models. It obtains the user's target sensory data, target base wine type, and target drinking conditions, inputs them into the blending scheme design model, and obtains the blending scheme design result through iterative optimization.

[0038] The beneficial effects of this invention are as follows: The method and system for designing distilled spirit blending schemes based on sensory data provided by this invention solves the problems of traditional methods that rely on fixed scoring tables for sensory data collection and cannot capture personalized user descriptions by introducing a large language model to parse the user's natural language or speech, thereby reducing the user input threshold and improving the breadth and authenticity of data collection; by constructing a blending flavor prediction model, it can accurately learn and predict the complex nonlinear relationship between the type of base spirit, the blending ratio, the drinking conditions, and the final sensory result of the blending, overcoming the problem of inaccurate prediction caused by relying solely on chemical components or simple linear extrapolation in traditional methods, and improving the accuracy of blending scheme design; by combining evolutionary algorithms with the blending flavor prediction model to construct a blending scheme design model, after receiving the user's target flavor, target type of base spirit, and target drinking conditions, it can actively search, evaluate, and iteratively optimize in a huge formula space, and finally directly output an executable blending scheme that meets the requirements, thereby improving the personalization and innovation of the blending scheme. Attached Figure Description

[0039] Figure 1 A flowchart illustrating the method for designing a distilled spirits blending scheme based on sensory data, provided for an embodiment;

[0040] Figure 2 A schematic diagram of the structure of the distilled spirits mixing scheme design system based on sensory data provided in the embodiment. Detailed Implementation

[0041] Existing technologies for designing mixed spirits mainly rely on chemical composition analysis or sensory scoring using fixed forms. The former is costly and fails to reflect subjective preferences, while the latter, with its rigid format, struggles to capture users' genuine natural language descriptions and cannot accurately predict the flavor of the mixed drink. This invention uses a large-scale language model to parse natural language, achieving the structuring and standardization of sensory data. It then constructs a deep learning model to accurately predict the flavor of the mixed drink, improving the accuracy of mixed drink design. Simultaneously, it utilizes evolutionary algorithms to automatically generate optimized solutions that meet multiple constraints, thereby achieving automatic generation of mixed drink solutions and significantly improving the efficiency and personalization of mixed drink innovation.

[0042] Specifically, this invention first involves data collection and structuring. By introducing a pre-trained large-scale language model, the fuzzy, subjective descriptions of user sensory evaluations—such as natural language or speech—are parsed and mapped onto pre-defined structured sensory indicators, which are then standardized. Next, a blending flavor prediction model is established capable of accurately predicting the sensory data of blended drinks under any base wine type, blending ratio, and drinking conditions. This model learns the complex mapping relationship between base wine type, blending ratio, drinking conditions, and the sensory evaluation data of the base wine and the sensory evaluation data of the blending scheme through training. After training, no actual mixing is required; only a new base wine type, blending recipe, and drinking conditions are input, and the model can forward propagate to calculate the predicted sensory data of the new blending recipe under the corresponding drinking conditions. Finally, an evolutionary algorithm is combined with the blending flavor prediction model to construct a blending scheme design model. After receiving the user's target flavor, target base wine type, and target drinking conditions, the model actively searches, evaluates, and iteratively optimizes within a vast recipe space to find the optimal recipe that best meets the user's flavor goals and is practically feasible, thus obtaining the blending scheme design result.

[0043] The technical solutions in this embodiment 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.

[0044] Figure 1 A flowchart illustrating a method for designing distilled spirit blending schemes based on sensory data is shown. Please refer to [link / reference]. Figure 1 The method includes the following steps:

[0045] Step 1: Collect sensory evaluation data from users on various base spirits of distilled spirits and multiple blending schemes under different drinking conditions to obtain blending data for multiple blending schemes.

[0046] The sensory evaluation data is obtained by structuring and standardizing the user's natural language text or speech text through a language big data model. The mixed drinking data includes the types of base wines in the mixed drinking scheme, the mixing ratio of each base wine in the mixed drinking scheme, the sensory evaluation data of each base wine under different drinking conditions, and the sensory evaluation data of the mixed drinking scheme under different drinking conditions.

[0047] In practical applications, recruit no fewer than 20 trained users and collect their sensory evaluation data on each base spirit of the distilled spirit, as well as sensory evaluation data on 100 mixed drinking schemes under different drinking conditions.

[0048] In this embodiment, the sensory evaluation data and predicted sensory data include eleven categories of sensory indicators: blending, smoothness, likability, fruitiness, graininess, floral aroma, sweetness, acidity, bitterness, smokiness or oak aging, and irritation or alcoholic taste. In practical applications, users can rate the sensory indicators of each base spirit and blending scheme, with a rating range of 1–5 points. A score of 1 indicates a low perceived level of the sensory indicator in the sample, while a score of 5 indicates a high perceived level. Users can also input natural language or speech descriptions during evaluation, such as "smooth on the palate, with a noticeable fruity aroma, but somewhat irritating." The pre-trained language model can automatically parse this description and map it into a structured sensory indicator score.

[0049] The scores for the eleven sensory indicators in the sensory evaluation data are as follows:

[0050] Integration (1-5 points): Indicates the harmony of the various components of the wine; Smoothness (1-5 points): Indicates the smoothness of the taste; Likability (1-5 points): Indicates the overall liking; Fruitiness (1-5 points): Indicates the intensity of fruit aromas; Graininess (1-5 points): Indicates the intensity of grain aromas; Floral Aroma (1-5 points): Indicates the intensity of floral aromas; Sweetness (1-5 points): Indicates the intensity of perceived sweetness; Acidity (1-5 points): Indicates the intensity of perceived acidity; Bitterness (1-5 points): Indicates the intensity of perceived bitterness; Smoky or Oak Barrel (1-5 points): Indicates the intensity of smoky or oak barrel flavors; Stimulation or Alcoholic Taste (1-5 points): Indicates the intensity of alcoholic stimulation.

[0051] After obtaining the scores for each sensory index, the maximum value of 5 was used for standardization. That is, 1 point is standardized to 0.2, and 5 points are standardized to 1.0. For example, at room temperature, a 1:1 mixture of 52-degree strong-aroma baijiu and Martell brandy was prepared. The scores for the eleven sensory indexes, namely blending, smoothness, likability, fruit aroma, grain aroma, floral aroma, sweetness, acidity, bitterness, smokiness or oak barrel aroma, and irritation or alcohol aroma, were as follows: 4.25, 3.25, 4.25, 2.25, 3.75, 3, 3.25, 2.25, 2.5, 3, 3.25. After standardization, the scores were as follows: 0.85, 0.65, 0.85, 0.45, 0.75, 0.60, 0.65, 0.45, 0.50, 0.60, 0.65.

[0052] Simultaneously, the drinking conditions of the mixed drink scheme are recorded and represented in the form of a one-hot vector. In this embodiment, the drinking conditions include room temperature, adding ice for 15 seconds, and ice melting by half after adding ice. For example, room temperature is denoted as [1, 0, 0], adding ice for 15 seconds is denoted as [0, 1, 0], and ice melting by half is denoted as [0, 0, 1].

[0053] Regarding the types of base spirits, this embodiment considers four types of distilled spirits: baijiu, whiskey, rum, and brandy. Base spirits are labeled using one-hot vectors. For example, if a mixed drink consists of baijiu and brandy, the base spirit type vector is [1, 0, 0, 1]. If it is only a single type of alcohol, for example, strong-aroma baijiu is denoted as [1, 0, 0, 0], and brandy as [0, 0, 0, 1]. Different brands of the same type of base spirit are not differentiated; instead, differences are directly represented by sensory evaluation data.

[0054] The blending ratio refers to the volume percentage of each type of base spirit (such as pure baijiu, pure whiskey, etc.) in the blending scheme. In this embodiment, the blending ratio is represented by a real number vector to ensure that the sum of all ratios is 1. For example, [0.6, 0.4] represents 60% baijiu and 40% brandy.

[0055] Step 2: Construct and train a mixed drinking flavor prediction model based on the mixed drinking data. The mixed drinking flavor prediction model is used to predict the sensory data of the corresponding mixed drinking scheme under the drinking conditions based on the types of base wines, the mixing ratio of each base wine and the drinking conditions in the mixed drinking scheme.

[0056] It is understandable that the blending flavor prediction model learns the sensory evaluation data of each base wine under different drinking conditions through training, and learns the complex mapping relationship between the base wine type, blending ratio, drinking conditions and sensory evaluation data of the blending scheme. After training, there is no need for actual blending. Just input the new base wine type, blending recipe and drinking conditions, and the model can forward propagate to calculate the predicted sensory data of the new blending recipe under the corresponding drinking conditions.

[0057] In this embodiment, the mixed beverage flavor prediction model is implemented using a deep neural network model, including one or more combinations of one-dimensional convolutional neural networks, recurrent neural networks, or fully connected neural networks.

[0058] In practical applications, for a new blended beverage recipe, a corresponding blending ratio vector, drinking condition vector, and base wine type vector are generated, and these are combined into a comprehensive input vector and input into the blended beverage flavor prediction model. The blended beverage flavor prediction model extracts features through a multi-layer fully connected network and a one-dimensional convolutional structure, and finally outputs predicted sensory data, namely the predicted values ​​of eleven types of sensory indicators.

[0059] Step 3: Construct a mixed drinking scheme design model based on evolutionary algorithms and mixed drinking flavor prediction models. Obtain the user's target sensory data, target base wine type, and target drinking conditions, and input them into the mixed drinking scheme design model. Through iterative optimization, obtain the mixed drinking scheme design results.

[0060] In this embodiment, the design model for the mixed beverage scheme is constructed based on an evolutionary algorithm and a mixed beverage flavor prediction model, wherein the evolutionary algorithm adopts one or more of the following: genetic algorithm, differential evolution algorithm, particle swarm optimization algorithm, simulated annealing algorithm, or multi-objective optimization algorithm.

[0061] The mixed beverage scheme design model uses a fitness function for iterative optimization, specifically including the following steps:

[0062] Step 31: Randomly generate multiple candidate blending schemes based on the user's target base wine. Each candidate blending scheme includes the type of base wine, drinking conditions, and the blending ratio of each base wine.

[0063] This step is based on the population initialization principle of evolutionary algorithms. By uniformly sampling within the solution space, the population is initialized, generating candidate solutions. This provides diverse initial solutions for subsequent evolutionary iterations, ensuring the algorithm can explore a broad search space and avoid prematurely getting trapped in local optima. Each candidate blending solution obtained has a corresponding base wine type, blending ratio, and drinking conditions.

[0064] Step 32: For each candidate mixed beverage scheme, call the mixed beverage flavor prediction model to make a prediction, obtain the predicted sensory data of each candidate mixed beverage scheme, substitute the target sensory data and the predicted sensory data into the fitness function to calculate the fitness value of each candidate mixed beverage scheme.

[0065] This embodiment transforms abstract sensory objectives into computable fitness values. A mapping relationship between formulation and flavor is established through a mixed beverage flavor prediction model, and then multi-objective optimization is achieved through a fitness function.

[0066] In practical applications, based on the mixed-drinking flavor prediction model constructed in step 2, the base wine type vector, mixing ratio vector, and drinking condition vector corresponding to each candidate mixed-drinking scheme are formed into a comprehensive input vector and input into the mixed-drinking flavor prediction model. The mixed-drinking flavor prediction model extracts features through a multi-layer fully connected network and a one-dimensional convolutional structure, and finally outputs predicted sensory data, namely the predicted values ​​of eleven types of sensory indicators. Then, the fitness value of the candidate mixed-drinking scheme is calculated using the target sensory data and the predicted sensory data.

[0067] In this embodiment, the formula for the fitness function is:

[0068] ;

[0069] in, Represents the fitness value. This represents the weighting coefficient, with a value range of [0, 1]. L This indicates the liking index in the predicted sensory data. This represents the sensory error, which is the sum of weighted errors of corresponding sensory indicators between the target sensory data and the predicted sensory data. Indicates to The scaling constant is used for normalization. This indicates a constraint or penalty.

[0070] In this embodiment, the target sensory data includes ten categories of sensory indicators: blending degree, smoothness, fruity aroma, grainy aroma, floral aroma, sweetness, sourness, bitterness, smoky or oaky aroma, and pungent or alcoholic aroma. That is, in the above fitness function, sensory error... It represents the sum of weighted errors of ten sensory indicators (excluding liking indicators) between the target sensory data and the predicted sensory data.

[0071] The above scheme achieves objective and quantitative evaluation of the mixed drinking scheme through fitness, and the fitness function uses weight coefficients... By balancing the dual optimization objectives of meeting the target and liking, and by using constraint penalty terms to ensure the practical feasibility of the generated solutions, the accuracy and personalization of mixed beverage solution design are improved.

[0072] Furthermore, in this embodiment, the constraint penalty term For matching and constraint penalties Alcohol content range restriction penalties Sparsity penalty The sum of these terms is calculated using the following formulas:

[0073] ;

[0074] ;

[0075] ;

[0076] ;

[0077] in, Indicates the first candidate mixed beverage regimen The mixing ratio of base spirits, This indicates the number of different types of base spirits in the candidate blending recipes, and ABV represents the alcohol content. Indicates the drinking condition correction factor. and Indicates the upper and lower limits of alcohol content. , and These represent the corresponding weighting coefficients.

[0078] For example, , and Let the values ​​be 100, 30, and 0.05 respectively. Assume a candidate mixing scheme with a mixing ratio of [0.6, 0.4]. and If we set the values ​​to 0.35 and 0.40 respectively, then by calculating the above penalty terms, we can obtain the matching and constraint penalties. If the alcohol content is set to 0, and the combined alcohol content is set to 0.38 (38 degrees), then the alcohol content range constraint penalty is... A value of 0 indicates a sparsity penalty. It is 0.07.

[0079] Specifically, this embodiment transforms complex physical constraints and process requirements into calculable penalty terms, which are embedded in the fitness function of the evolutionary algorithm. By introducing a measure of constraint violation into the objective function, the constrained optimization problem is transformed into an unconstrained optimization problem, guiding the search process to automatically converge towards the feasible region.

[0080] The aforementioned constraint penalty terms, by simultaneously considering proportion and constraint penalties, alcohol content range constraints, and sparsity penalties, ensure that the output blending scheme satisfies the proportion and constraints, keeps the alcohol content as close to the target range as possible, and minimizes the number of base spirit types, thus reducing the number of subsequent evolutionary iterations. This ensures that the generated blending scheme not only has excellent flavor but also possesses the feasibility and economic viability for industrial production, achieving a perfect transition from theoretical optimization to practical application.

[0081] Step 33: Perform evolutionary iteration based on the fitness values ​​of each candidate mixing scheme, using selection, crossover, and mutation to generate a new generation of candidate mixing schemes, and recalculate the fitness values ​​of the new generation of candidate mixing schemes.

[0082] This embodiment is based on an evolutionary algorithm, which gradually improves the quality of the population by simulating the natural evolutionary process. The selection operation retains superior individuals, the crossover operation integrates excellent traits, and the mutation operation introduces new exploration directions, thereby generating a new generation of candidate mixed drinking schemes. In this embodiment, the mutation probability is 0.05, and the crossover probability is 0.8. For the generated new generation of candidate mixed drinking schemes, steps 32 and 33 are repeated for evolutionary iteration.

[0083] Step 34: When the number of iterations reaches the threshold, one or more candidate mixed drinking schemes with fitness values ​​higher than the fitness threshold are taken as the mixed drinking scheme design results.

[0084] In this embodiment, the iteration terminates when the population evolution reaches a stable state or meets preset conditions, significantly improving computational efficiency while ensuring solution quality. After the number of iterations reaches a threshold (e.g., 200 times), or when the fitness value stabilizes (the fitness value increases only slightly after multiple iterations), one or more candidate mixed drinking schemes with fitness values ​​higher than the fitness threshold are used as the mixed drinking scheme design results. When there are multiple candidate mixed drinking schemes, alternative schemes are provided to enhance practicality.

[0085] In this embodiment, after obtaining the design results of the mixed beverage scheme, the method further includes:

[0086] The system receives user ratings and natural language feedback on the design results of mixed beverage schemes, and uses a large language model to transform these ratings into structured indicators, forming user historical preference data. When generating mixed beverage scheme design results in the future, the system selects personalized mixed beverage schemes based on the user historical preference data and generates corresponding natural language explanations.

[0087] In practical applications, a large language model can be used to analyze users' natural language feedback, transforming it into a structured sensory indicator adjustment vector to accurately capture users' implicit preference information. Secondly, an exponentially weighted moving average algorithm is employed to fuse new feedback with historical preference data, forming a continuously evolving personalized preference vector. Finally, in the subsequent solution generation process, this preference vector is used to personalize and re-rank candidate mixed drinks, generating natural language explanations to justify the recommendations. By continuously collecting explicit user feedback (personal preference scores) and implicit feedback (natural language), a dynamically updated user preference model is established, enabling the system to become increasingly accurate with each interaction, gaining a more precise understanding of user taste preferences and ultimately achieving adaptive personalized recommendations.

[0088] In summary, this embodiment achieves intelligent collection and standardization of sensory data by introducing a large language model, constructing a precise mixed-drink flavor prediction model that accurately depicts the complex nonlinear relationship between the mixing ratio, drinking conditions, and sensory characteristics. Combining evolutionary algorithms and multi-constraint fitness functions, it enables the rapid generation of optimized solutions that meet the user's target flavor while satisfying practical constraints such as mixing ratios, alcohol content ranges, and formula simplicity. Furthermore, it establishes a personalized preference model through user feedback, forming a closed-loop learning system of recommendation, feedback, and optimization. Ultimately, it constructs a complete technical system from data-driven to intelligent generation, addressing the industry pain points of traditional mixed-drink design—relying on human experience, being inefficient, and failing to meet personalized needs—and improving the accuracy, efficiency, and user experience of distilled spirits product development.

[0089] Based on the above technical solutions, this embodiment also proposes a distilled spirits blending scheme design system based on sensory data, used to implement the distilled spirits blending scheme design method based on sensory data described in the embodiment. Please refer to [link to relevant documentation]. Figure 2 The system includes:

[0090] The mixed drinking data acquisition unit is used to collect sensory evaluation data of users on various base spirits of distilled spirits and multiple mixed drinking schemes under different drinking conditions, and to obtain mixed drinking data of multiple mixed drinking schemes. The sensory evaluation data is obtained by structuring and standardizing the user's natural language text or speech text through a language big data model. The mixed drinking data includes the types of base spirits in the mixed drinking scheme, the mixing ratio of each base spirit in the mixed drinking scheme, the sensory evaluation data of each base spirit under different drinking conditions, and the sensory evaluation data of the mixed drinking scheme under different drinking conditions.

[0091] The blending flavor prediction unit is constructed and trained based on the blending data. It is used to predict the sensory data of the corresponding blending scheme under the drinking conditions based on the types of base wines, the blending ratio of each base wine and the drinking conditions in the blending scheme.

[0092] The blending scheme design unit is used to construct a blending scheme design model based on evolutionary algorithms and blending flavor prediction models. It obtains the user's target sensory data, target base wine type, and target drinking conditions, inputs them into the blending scheme design model, and obtains the blending scheme design result through iterative optimization.

[0093] It is understood that since the distilled spirits mixing scheme design system based on sensory data described in this embodiment is a system for implementing the distilled spirits mixing scheme design method based on sensory data described in the embodiment, the system disclosed in the embodiment is relatively simple to describe because it corresponds to the method disclosed in the embodiment. For relevant parts, please refer to the description of the method, and it will not be repeated here.

Claims

1. A method for designing mixed drinking schemes of distilled spirits based on sensory data, characterized in that, The method includes: Sensory evaluation data of users on various base spirits of distilled spirits and multiple blending schemes under different drinking conditions are collected to obtain blending data of multiple blending schemes. The sensory evaluation data is obtained by structuring and standardizing the user's natural language text or speech text through a language big data model. The blending data includes the types of base spirits in the blending scheme, the blending ratio of each base spirit in the blending scheme, the sensory evaluation data of each base spirit under different drinking conditions, and the sensory evaluation data of the blending scheme under different drinking conditions. Based on the mixed drinking data, a mixed drinking flavor prediction model is constructed and trained. The mixed drinking flavor prediction model is used to predict the sensory data of the corresponding mixed drinking scheme under the drinking conditions based on the types of base wines, the mixing ratio of each base wine and the drinking conditions in the mixed drinking scheme. A blending scheme design model is constructed based on evolutionary algorithms and blending flavor prediction models. The target sensory data, target base wine types, and target drinking conditions of users are obtained and input into the blending scheme design model. Through iterative optimization, the blending scheme design results are obtained.

2. The method for designing a distilled spirits blending scheme based on sensory data according to claim 1, characterized in that, The sensory evaluation data and predicted sensory data include eleven categories of sensory indicators, including blending, smoothness, likability, fruity aroma, grain aroma, floral aroma, sweetness, sourness, bitterness, smoky or oak barrel aroma, and irritation or alcoholic aroma. The target sensory data includes ten categories of sensory indicators: blending, smoothness, fruity aroma, grain aroma, floral aroma, sweetness, sourness, bitterness, smokiness or oak barrel aroma, and irritation or alcoholic aroma.

3. The method for designing a distilled spirits blending scheme based on sensory data according to claim 1, characterized in that, The drinking conditions and target drinking conditions include room temperature, adding ice for 15 seconds, and ice melting by half after adding ice. The base spirits include baijiu, whiskey, rum, and brandy.

4. The method for designing a distilled spirits blending scheme based on sensory data according to claim 1, characterized in that, The design model for the mixed beverage scheme uses a fitness function for iterative optimization, specifically including: Multiple candidate blending schemes are randomly generated based on the user's target base wine. Each candidate blending scheme includes the type of base wine, drinking conditions, and the blending ratio of each base wine. For each candidate mixed beverage scheme, the mixed beverage flavor prediction model is invoked to make a prediction, and the predicted sensory data of each candidate mixed beverage scheme is obtained. The target sensory data and the predicted sensory data are substituted into the fitness function to calculate the fitness value of each candidate mixed beverage scheme. Evolutionary iteration is performed based on the fitness values ​​of each candidate mixing scheme. Selection, crossover, and mutation are used to generate a new generation of candidate mixing schemes, and the fitness values ​​of the new generation of candidate mixing schemes are recalculated. Once the number of iterations reaches the threshold, one or more candidate mixed drinking schemes with fitness values ​​higher than the fitness threshold are taken as the design results of the mixed drinking scheme.

5. The method for designing a distilled spirits blending scheme based on sensory data according to claim 4, characterized in that, The formula for the fitness function is: ； in, Represents the fitness value. This represents the weighting coefficient, with a value range of [0, 1]. L This indicates the liking index in the predicted sensory data. This represents the sensory error, which is the sum of weighted errors of corresponding sensory indicators between the target sensory data and the predicted sensory data. Indicates to The scaling constant is used for normalization. This indicates a constraint or penalty.

6. The method for designing a distilled spirits blending scheme based on sensory data according to claim 5, characterized in that, The constraint penalty item For matching and constraint penalties Alcohol content range restriction penalties Sparsity penalty The sum of these terms is calculated using the following formulas: ； ； ； ； in, Indicates the first candidate mixed beverage regimen The mixing ratio of different base spirits. This indicates the number of different types of base spirits in the candidate blending recipes, and ABV represents the alcohol content. Indicates the drinking condition correction factor. and Indicates the upper and lower limits of alcohol content. , and These represent the corresponding weighting coefficients.

7. The method for designing a distilled spirits blending scheme based on sensory data according to claim 1, characterized in that, The evolutionary algorithm employs one or more of the following: genetic algorithm, differential evolution algorithm, particle swarm optimization algorithm, simulated annealing algorithm, or multi-objective optimization algorithm.

8. The method for designing a distilled spirits blending scheme based on sensory data according to claim 1, characterized in that, The flavor prediction model for mixed drinks is implemented using a deep neural network model, including one or more combinations of one-dimensional convolutional neural networks, recurrent neural networks, or fully connected neural networks.

9. The method for designing a distilled spirits blending scheme based on sensory data according to claim 1, characterized in that, After obtaining the design results of the mixed beverage scheme, the method further includes: Receive users' personal preference ratings and natural language feedback on the design results of the mixed drink solution, and use a large language model to transform them into structured indicators to form user historical preference data; When generating the design results of the mixed drink scheme, personalized mixed drink schemes are selected based on the user's historical preference data, and corresponding natural language explanations are generated.

10. A system for designing distilled spirits blending schemes based on sensory data, characterized in that: The system is used to implement the method for designing distilled spirit blending schemes based on sensory data as described in any one of claims 1 to 9, the system comprising: The mixed drinking data acquisition unit is used to collect sensory evaluation data of users on various base spirits of distilled spirits and multiple mixed drinking schemes under different drinking conditions, and to obtain mixed drinking data of multiple mixed drinking schemes. The sensory evaluation data is obtained by structuring and standardizing the user's natural language text or speech text through a language big data model. The mixed drinking data includes the types of base spirits in the mixed drinking scheme, the mixing ratio of each base spirit in the mixed drinking scheme, the sensory evaluation data of each base spirit under different drinking conditions, and the sensory evaluation data of the mixed drinking scheme under different drinking conditions. The blending flavor prediction unit is constructed and trained based on the blending data. It is used to predict the sensory data of the corresponding blending scheme under the drinking conditions based on the types of base wines, the blending ratio of each base wine and the drinking conditions in the blending scheme. The blending scheme design unit is used to construct a blending scheme design model based on evolutionary algorithms and blending flavor prediction models. It obtains the user's target sensory data, target base wine type, and target drinking conditions, inputs them into the blending scheme design model, and obtains the blending scheme design result through iterative optimization.

Images