Baijiu blending big data analysis decision-making system and method

Abstract

The invention discloses a Baijiu blending big data analysis decision system and method, and belongs to the technical field of intelligent food manufacturing and flavor modeling. Physicochemical indexes, gas chromatography-mass spectrometry, sensory evaluation and historical proportioning schemes of multiple batches of raw wine are collected, and a heterogeneous flavor original data set is constructed; performing unified feature extraction and vectorization processing on the data to generate a flavor multi-modal feature matrix; constructing a flavor mapping map based on an attention mechanism, and extracting key flavor contribution features; using the target flavor label to guide a neural generation network to generate a wine base combination candidate set, and outputting a matching suggestion; performing subjective consistency scoring through a flavor simulation model and an expert scoring discrimination model, and screening an optimal matching scheme; the method has the advantages of being high in modeling precision, high in sensory consistency, high in blending efficiency and the like, and is suitable for intelligent matching design and personalized seasoning optimization of complex flavor products.

Description

Technical Field

[0001] This invention relates to the field of intelligent food manufacturing and flavor modeling technology, specifically to a big data analysis and decision-making system and method for blending baijiu (Chinese liquor). Background Technology

[0002] Traditional baijiu blending relies on experienced distillers' subjective perception of the liquor's flavor and their long-accumulated technological knowledge. Although some companies have introduced physicochemical testing methods and flavor wheel analysis systems, existing technologies struggle to meet the demands for high-precision, high-consistency intelligent blending due to significant heterogeneity in the composition of multiple batches of raw liquor, low-frequency and fragmented sensory evaluation data, and increasingly differentiated blending requirements. Several technical challenges currently exist: 1. The concentrations of many key flavor compounds are at trace levels, and their changes have a high-dimensional nonlinear mapping effect on sensory evaluation, which is difficult to characterize using traditional regression modeling methods. 2. The experimental sample size is limited, and the sensory expert evaluation has strong subjective bias, resulting in a sparse distribution of training samples and a low label signal-to-noise ratio. 3. Data types such as the physicochemical indicators of raw wine, GC-MS spectra of volatile components, sensory flavor evaluation, and historical blending scheme texts vary greatly, and there is a lack of a unified feature extraction and fusion analysis framework. 4. The language used to describe the target flavor is vague and cannot be quantified: words such as "mellow", "sweet", and "long aftertaste" appear frequently in expert corpora, but lack standardized definitions and quantifiable explanations.

[0003] The aforementioned problems have led to a continued heavy reliance on manual labor in the current baijiu blending process, which has hampered batch consistency, product innovation, and the inheritance of technological knowledge. Summary of the Invention

[0004] The purpose of this invention is to provide a big data analysis and decision-making system and method for blending baijiu (Chinese liquor) to address the shortcomings in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a big data analysis and decision-making method for blending baijiu (Chinese liquor), comprising: Acquire physicochemical index data, GC-MS volatile component spectra, historical sensory evaluation records, and blending ratio scheme texts of multiple batches of raw wine to construct a heterogeneous flavor raw dataset; A unified feature transformation was performed on the heterogeneous flavor original dataset. A spectral autoencoder was used to extract the latent component concentration feature vectors of volatile components. Word vector encoding technology was used to vectorize the text ratio scheme and construct a flavor multimodal feature matrix at a unified scale. Based on the flavor multimodal feature matrix, a micro-macro flavor mapping map is constructed. An attention mechanism is used to learn the contribution weight of each flavor feature to the target flavor label, and the flavor contribution distribution map of each sample is output. The target blending flavor requirement is input as a label vector. A back-guided neural network is used to generate a candidate set of base spirits that are closest to the target flavor, and the corresponding blending ratio suggestions are output. The corresponding ratio recommendations are reconstructed through perceptual simulation. Combined with the discrimination model of historical expert evaluation, the subjective consistency score of each candidate scheme is calculated, and the one with the highest score is selected as the preferred mixing scheme. Based on the preferred blending scheme, check whether its sensory semantic distance with the target flavor vector meets the threshold requirement. If it does not meet the threshold, update the parameters of the reverse generation network until the flavor fitting accuracy is met.

[0006] Preferably, the step of constructing the flavor multimodal feature matrix at a uniform scale includes: Peak normalization was performed on the gas chromatography-mass spectra, and a self-encoding neural network consisting of a convolutional encoder and a loop decoder was used to extract the concentration feature vector of potential volatile components in the spectra. The text of the blending scheme is structurally analyzed to extract the relationship between the types of base wines and their corresponding proportions. The blending scheme is then converted into a numerical vector representation using a word embedding model-based encoding method. Standardize the mapping between physicochemical index data and sensory evaluation labels to construct a physicochemical feature submatrix and a sensory feature submatrix under a unified dimension. By dimensional alignment and feature concatenation of spectral vectors, text proportion vectors, and physicochemical and sensory feature sub-matrices, a flavor multimodal feature matrix at a unified scale is constructed, which serves as the input feature benchmark for flavor mapping modeling.

[0007] Preferably, the step of constructing a micro-macro flavor mapping map based on the flavor multimodal feature matrix includes: The flavor multimodal feature matrix is ​​input into a multilayer perceptron to extract high-order nonlinear flavor feature representations; Attention matching is performed on the higher-order nonlinear flavor features and the target flavor label vector to generate the attention weight distribution of each feature dimension relative to the target label; Calculate the weighted flavor response vector by summing the features according to their corresponding attention weights to obtain the flavor response result for each sample; Based on the similarity between the flavor response vector and the target label, a flavor contribution distribution map is generated, representing the degree of contribution of each flavor feature to the achievement of the target flavor.

[0008] Preferably, the step of generating a flavor contribution distribution map based on the similarity between the flavor response vector and the target label includes: The flavor response vector and the target flavor label vector of each sample are input into the similarity calculation module, and the cosine similarity function is used to calculate the similarity score between them. The original attention weight matrix is ​​normalized and adjusted based on the similarity score. The normalized feature weights are mapped onto a two-dimensional plane, and the degree of matching of each flavor feature with the target label is visualized in the form of a heatmap, forming a flavor contribution distribution map. The flavor contribution distribution map is segmented by thresholding to extract the set of high-contribution features, which are then used as priority flavor factors for subsequent candidate base wine screening and blending optimization.

[0009] Preferably, the step of generating a candidate set of base wine combinations that most closely resemble the target flavor using a reverse-guided neural network includes: The target blending flavor requirement is encoded into a fixed-dimensional label vector through a flavor label embedding model; a neural generative network is constructed with the original wine ratio vector as the generation target, and the label vector is input into the generator as a guiding condition; the set of high-contribution flavor features extracted from the flavor contribution distribution map is used as an additional input guiding vector; the generator and discriminator are optimized through adversarial training so that the generator output can simulate the feature distribution of real historical blending schemes and is as close as possible to the target flavor vector in the flavor response space, thereby outputting a candidate set of original wine combinations and their suggested ratios.

[0010] Preferably, the step of calculating the subjective consistency score of each candidate solution and selecting the one with the highest score as the preferred matching solution includes: Each candidate ratio vector is input into the trained flavor response prediction model, and the corresponding flavor response vector is generated through simulation. Construct a subjective consistency discrimination model based on historical sensory evaluation data, input the flavor response vector, and output a subjective consistency score; The discrimination model was trained under supervision using historical evaluation corpus from sensory experts, and the cross-entropy loss function was aligned with the evaluation labels. All candidate formulation schemes are sorted according to subjective consistency scores, and the candidate scheme with the highest score is selected as the final preferred blending formulation scheme.

[0011] Preferably, based on the preferred blending scheme, the step of verifying whether its sensory semantic distance with the target flavor vector meets the threshold requirement includes: The optimal blending ratio is input into the flavor response prediction model, and the corresponding flavor response vector is output by simulation. The flavor response vector and the target flavor label vector are calculated using a sensory semantic distance function, which is a normalized cosine distance function. A sensory semantic distance threshold is set. When the distance does not meet the threshold requirement, the feedback generation network adjusts the input noise and label vector, guiding the generator to re-output the original wine combination scheme.

[0012] This invention also provides a big data analysis and decision-making system for blending baijiu (Chinese liquor), comprising: Data acquisition module: acquires physicochemical index data, GC-MS volatile component spectra, historical sensory evaluation records and blending ratio scheme text of multiple batches of raw wine, and constructs a heterogeneous flavor raw dataset; Feature extraction module: Performs unified feature transformation on the heterogeneous flavor raw dataset, uses spectral autoencoder to extract latent component concentration feature vectors of volatile components, uses word vector encoding technology to vectorize the text ratio scheme, and constructs a flavor multimodal feature matrix at a unified scale; Flavor mapping modeling module: Based on the flavor multimodal feature matrix, construct a micro-macro flavor mapping map, use an attention mechanism to learn the contribution weight of each flavor feature to the target flavor label, and output the flavor contribution distribution map of each sample; Flavor Reverse Generation Module: Input the target blending flavor requirement as a label vector, use a reverse-guided neural network to generate a candidate set of base wine combinations that are closest to the target flavor, and output corresponding blending suggestions; Perceptibility simulation evaluation module: Performs perceptibility simulation reconstruction on the corresponding ratio recommendations, combines the discrimination model of historical expert evaluation, calculates the subjective consistency score of each candidate scheme, and selects the one with the highest score as the preferred mixing scheme; Flavor Fitting Verification Module: Based on the preferred blending scheme, verify whether its sensory semantic distance with the target flavor vector meets the threshold requirement. If it does not meet the threshold, update the parameters of the reverse generation network until the flavor fitting accuracy is met.

[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention, by constructing a heterogeneous flavor data fusion framework, achieves a unified expression of the physicochemical indicators, GC-MS spectra, sensory evaluations, and historical blending schemes of baijiu raw materials. It is the first to propose a micro-macro flavor mapping method driven by a multimodal feature matrix, solving the problems of "scattered data, incomparable indicators, and difficulty in quantifying subjective factors" in the traditional blending process. By utilizing an attention mechanism to establish a flavor contribution distribution map, it effectively quantifies the influence path of each component on the target flavor, providing interpretable evidence for data-driven blending optimization and significantly improving the accuracy and transparency of baijiu flavor modeling.

[0014] 2. This invention introduces a reverse-guided neural network to transform sensory flavor targets into base spirit blending schemes, constructs a closed-loop optimization mechanism, and integrates expert sensory evaluation and flavor response prediction models. Through dual constraints of subjective consistency scoring and sensory semantic distance, it achieves personalized and automated intelligent blending decisions. Experimental verification shows that the method of this invention outperforms existing methods in terms of flavor matching accuracy, expert satisfaction, and scheme innovation, and possesses good industrial applicability. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0016] Figure 1 This is a flowchart of the method of the present invention.

[0017] Figure 2 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0019] Example 1, please refer to Figure 1 As shown in this embodiment, the big data analysis and decision-making method for blending baijiu includes: We acquired physicochemical index data, GC-MS volatile component spectra, historical sensory evaluation records, and blending ratio scheme texts from multiple batches of raw wine to construct a heterogeneous flavor raw dataset.

[0020] In this invention, a complete heterogeneous flavor dataset must first be constructed to support subsequent multimodal feature extraction and flavor mapping modeling. This step specifically includes the following: Data collection on physicochemical properties of raw spirits: Samples of raw spirits from multiple different production batches were collected. These physicochemical properties include, but are not limited to, basic physicochemical parameters such as alcohol content, total acidity, total esters, ethyl acetate, ethyl lactate, acetal, methanol, fusel oil, pH value, and conductivity. This type of data is generally presented in structured tabular form, with standardized field definitions and dimensions, facilitating subsequent standardization.

[0021] GC-MS volatile component spectral acquisition: The above-mentioned raw wine samples were qualitatively and quantitatively analyzed using gas chromatography-mass spectrometry (GC-MS) to obtain the volatile organic compound response peak diagrams at specific retention times. The spectral data are presented in a two-dimensional matrix or sequence format, containing information such as peak area, retention time, and peak intensity, to reveal the flavor molecular characteristics at the microscopic level.

[0022] Historical sensory evaluation data integration: This involves collecting historical sensory evaluation records from baijiu tasters on the corresponding raw liquor samples. Evaluation content includes, but is not limited to, dimensions such as aroma, taste, harmony, finish, and typicality. Sensory evaluation data may be in the form of scoring tables (e.g., 1-10 point scales), grading labels (e.g., "mellow," "rich"), or unstructured natural language descriptions (e.g., "smooth on the palate, with a lingering sweetness"). This type of data is highly subjective and sparsely distributed, requiring normalization and structured text parsing.

[0023] Text extraction of blending ratio schemes: Extracting the ratio records formed in past actual blending experiments, including the types of base spirits used, the corresponding volume ratios (or mass ratios), and the name or code of the target product. This type of data usually comes from process archives or electronic record systems and needs to be standardized in format and linked to sensory evaluation data at the sample level.

[0024] For the four types of data mentioned above, a unified sample ID indexing rule is set to achieve a one-to-one correspondence between physicochemical data, spectral data, sensory evaluation, and blending formula. By binding sample IDs, a raw wine data entity with complete flavor attribute tags can be formed, constituting a multi-source heterogeneous flavor raw dataset.

[0025] Through the above steps, this invention establishes a basic framework for flavor big data, encompassing physicochemical properties, microscopic chemical spectra, subjective sensory evaluation, and actual process pathways. This framework provides high-quality data support for subsequent flavor feature fusion, sensory prediction modeling, and intelligent blending scheme generation. Compared to traditional single-data source analysis methods, the dataset structure of this invention is more comprehensive and its information dimensions are richer, which helps improve the model's ability to understand flavor complexity and proportioning strategies.

[0026] A unified feature transformation is performed on the heterogeneous flavor original dataset. A spectral autoencoder is used to extract the latent component concentration feature vectors of volatile components. Word vector encoding technology is used to vectorize the text ratio scheme and construct a flavor multimodal feature matrix at a unified scale.

[0027] After constructing the original heterogeneous flavor dataset, this invention achieves a unified quantitative expression of baijiu flavor through a multimodal feature extraction and fusion mechanism, providing a foundation for subsequent flavor mapping modeling. This step includes the following sub-steps: Gas chromatography-mass spectrometry (GC-MS) data for each raw wine sample were entered in CSV format, containing two fields: retention time (in minutes) and response intensity (in count values). Each sample contained 1000 to 2000 data points.

[0028] First, the spectral data is interpolated and resampled to a fixed length, such as 2048 equally spaced points, to form a fixed-length time series. Linearly normalize the response intensity, calculate the maximum peak value in the spectrum, and divide all response values ​​by this maximum value to make them fall into the [0,1] interval; Background peaks with a normalization intensity below 0.01 are filtered out to remove interference noise.

[0029] The encoder uses a three-layer one-dimensional convolutional neural network with kernel sizes of 5, 3, and 3, and channel numbers of 32, 64, and 128 respectively. The activation function uses a rectified linear unit (ReLU); The pooling operation uses max pooling, with a pooling window size of 2 for each layer. The encoded output is flattened and then connected to a fully connected layer, which outputs a 128-dimensional latent component concentration feature vector. The decoder is a bidirectional gated recurrent unit network (Bi-GRU) with the encoded vector as the initial state and a sequence length of 2048. The loss function is the mean squared error (MSE) between the input sequence and the reconstructed sequence.

[0030] Each original wine sample outputs a 128-dimensional vector of potential component concentrations, which is saved as a NumPy array for subsequent multimodal feature concatenation.

[0031] Each formulation is described in natural language, for example: "Mix and blend 40% strong aroma type A, 30% light aroma type B, and 30% soy sauce aroma type C."

[0032] Input the historical blending and proportioning scheme in text form, and first extract the original wine type, batch number and corresponding volume ratio value using regular expression methods.

[0033] Organize the results into a structured dictionary format, such as: {"Strong Aroma Type A":0.40,"Light Aroma Type B":0.30,"Soy Sauce Aroma Type C":0.30}.

[0034] Construct a corpus containing at least 50,000 historical collation schemes; The Word2Vec Skip-Gram model was used for training, with word vector dimension set to 64, window size to 5, and minimum word frequency to 2. Each wine type name is embedded as a 64-dimensional vector, multiplied by its proportion, and then all vectors are weighted, summed, and L2 normalized to obtain the final proportion text vector.

[0035] Each matching scheme corresponds to a 64-dimensional matching vector, with a unified format of a floating-point vector array.

[0036] The physicochemical index data are presented in a quantitative structured table, with fields including 10 conventional indicators such as alcohol content, total acid, and total esters. Sensory evaluation is presented as a scoring sheet or text label, and converted into eight dimensions, including aroma, harmony, aftertaste, and mouthfeel.

[0037] For physicochemical indicators, the maximum-minimum normalization method is used: each indicator value is subtracted from the minimum value of the entire sample and then divided by the range; Sensory scores were standardized using Z-score. If the words are tags, such as "rich", "moderate", or "light", they are mapped to scores by constructing a dictionary (e.g., rich is 0.9, light is 0.3). Finally, a unified physical and chemical feature sub-matrix (10-dimensional) and sensory feature sub-matrix (8-dimensional) are generated, which are represented as a vector group.

[0038] Each sample outputs 18-dimensional standardized numerical features, saved as JSON or a matrix format.

[0039] Input vector dimensions summary: Spectral feature vector: 128 dimensions; Text matching vector: 64 dimensions; Physicochemical index vector: 10 dimensions; Sensory evaluation vector: 8 dimensions.

[0040] All feature vectors are first normalized to the unit mean; for those with insufficient dimensions, zero padding is not performed, but a 2-layer perceptron upscaling network is used to upscale the 10-dimensional and 8-dimensional vectors to 32 dimensions respectively; after upscaling, all vectors are concatenated into a single flavor feature vector with a total of 256 dimensions.

[0041] The 256-dimensional vectors of each sample are arranged sequentially to form a flavor multimodal feature matrix with N1 samples. This matrix will serve as the standard input for subsequent flavor mapping networks (such as flavor contribution graph models), possessing a unified structure, unified semantics, and high interpretability.

[0042] Based on the flavor multimodal feature matrix, a micro-macro flavor mapping map is constructed. An attention mechanism is used to learn the contribution weight of each flavor feature to the target flavor label, and the flavor contribution distribution map of each sample is output.

[0043] After constructing the multimodal flavor feature matrix, this invention further establishes a micro-to-macro flavor mapping relationship through a multilayer perceptron and attention mechanism, and combines this with target flavor labels to analyze the contribution of each flavor feature to the formation of the target taste. This step includes the following sub-steps: The multimodal flavor feature vector corresponding to each original wine sample is input into a multilayer perceptron network structure. The multilayer perceptron consists of three fully connected neural networks: the first layer maps the original 256-dimensional feature input to 128 dimensions; the second layer maps the 128 dimensions to 64 dimensions; and the third layer outputs the final 32-dimensional higher-order flavor feature vector. Each layer uses a rectified linear unit function as its activation function; the loss function is the minimization of the target flavor prediction error. This multilayer perceptron captures the complex interaction structure between flavor components through nonlinear mapping relationships, achieving deep fusion of multi-source information.

[0044] The 32-dimensional higher-order flavor feature vector of each sample is matched with the target flavor label vector using attention matching. The target flavor label vector is generated by a flavor label embedding model using an experienced expert's language description, and its dimension is uniformly set to 32.

[0045] A dot-product attention-based computational method is employed. The calculation involves: performing linear transformations on both the sample flavor feature vector and the target label vector; performing a dot product operation on the transformed vectors; scaling the dot product result and applying a softmax function to obtain the attention weight distribution of each flavor feature dimension relative to the target label; and normalizing the weight values ​​to between 0 and 1. This weight distribution measures the attention and sensitivity of each flavor feature dimension to achieving the target flavor.

[0046] Based on the obtained attention weight vector, it is multiplied element-wise with the original high-order flavor feature vector, and the results are summed to obtain the weighted flavor response vector.

[0047] Subsequently, the flavor response vector and the target flavor label vector are input into the similarity calculation module. The similarity calculation uses a cosine similarity function, where the value is the cosine of the angle between the two vectors, ranging from -1 to 1; the closer to 1, the closer the similarity. The specific calculation method is as follows: calculate the magnitude of each vector; perform a dot product on the two vectors; divide the dot product by the magnitude product to obtain the similarity score. This score serves as the core indicator for evaluating the degree of match between the current flavor profile and the target flavor label.

[0048] The obtained attention weight vector is used as input and normalized based on similarity weights to ensure that the sum of the weights is 1, while enhancing the ability to distinguish high-matching dimensions.

[0049] The normalized feature weights are mapped onto a two-dimensional visualization plane to construct a flavor contribution distribution map. The distribution map is displayed in the form of a heatmap, with each flavor dimension serving as one of the coordinate axes on the map, and the intensity of the color representing its contribution.

[0050] To identify key flavor factors, a threshold (e.g., a weight value greater than 0.05) is applied to the flavor contribution distribution map, and feature indices with weights higher than the threshold and their corresponding values ​​are extracted to form a set of high-contribution features.

[0051] This set of flavor factors is used as priority retention or enhancement factors in subsequent raw wine screening and blending ratio optimization processes, which helps to improve the accuracy of target flavor matching and the efficiency of scheme design.

[0052] To address the issue that high-contribution flavor feature sets are sparse feature sets with variable dimensions during the extraction stage, a fixed-length mapping mechanism based on a unified flavor feature index space is introduced to ensure the consistency of input dimensions and the integrity of information representation in subsequent generation processes. Specifically, a unified high-order flavor feature index system is pre-constructed during model training, defining all high-order flavor features in a fixed order as 32 standard feature dimensions, which serve as the global flavor representation benchmark. After thresholding the flavor contribution distribution map, feature indices with weights higher than the threshold and their corresponding weights are extracted to form a high-contribution feature set. Based on their positions in the unified index space, the corresponding weight values ​​are mapped to the corresponding positions in the 32-dimensional vector. Unselected dimensions are initially assigned a value of 0, thus forming a structured sparse vector with a clear semantic correspondence. This mapping process is not simply zero-padding; rather, index alignment ensures that each dimension corresponds to a specific flavor semantic, enabling the resulting vector to maintain dimensional consistency while possessing clear physical meaning and interpretability, thereby providing a stable input foundation for subsequent generation processes.

[0053] Building upon the aforementioned index alignment, to avoid information fragmentation caused by inactive dimensions remaining zero for extended periods, this invention further introduces a weight diffusion and reconstruction mechanism based on flavor semantic association. Specifically, utilizing the semantic distance or statistical correlation between high-order flavor features, the weights of high-contribution features are diffused to their adjacent feature dimensions according to a preset decay function. This allows previously inactive dimensions to acquire non-zero response values ​​related to their semantic proximity, thus achieving a smooth transition from local sparse expression to global continuous expression. Subsequently, the diffused vector is normalized to eliminate the impact of weight scale differences between different samples and ensure overall numerical stability. Through the continuous processing of index alignment—weight diffusion—normalization, the original sparse high-contribution feature set can be stably converted into a fixed 32-dimensional dense flavor feature vector. This not only avoids the invalid noise problem introduced by direct zero padding but also enhances the ability of the generation process to express key flavor information, enabling those skilled in the art to directly implement this technical solution based on the description.

[0054] The target blending flavor requirement is input as a label vector. A back-guided neural network is used to generate a candidate set of base spirits that are closest to the target flavor, and the corresponding blending ratio suggestions are output.

[0055] After obtaining the flavor contribution distribution map and the set of high-contribution flavor features, this invention further constructs a reverse-guided neural network based on the target flavor label vector to generate a candidate set of base wine combinations that closely approximate the target flavor, and outputs corresponding blending suggestions. The specific steps are as follows: The target flavor profile is determined by sensory experts using a flavor descriptive lexicon, such as "smooth, well-balanced, and long-lasting." First, these descriptive terms are standardized, keywords are extracted, and then vectorized using a trained flavor tag embedding model.

[0056] The flavor tag embedding model uses the Skip-Gram method and is trained on a corpus containing at least 100,000 historical sensory evaluations, with an embedding dimension of 32. Each flavor keyword is mapped to a 32-dimensional vector, and multiple keywords are weighted and averaged or attention-weighted summed to obtain a unified 32-dimensional target tag vector.

[0057] A neural generative network is constructed with the goal of generating outputs using the proportion vector of the original wine combination. The generator employs a conditional fully connected neural network, with the following structure: Input layer: Accepts a 128-dimensional vector by concatenating a random noise vector (length set to 64), a flavor label vector (32-dimensional), and a high-contribution flavor feature vector (length uniformly set to 32-dimensional); Hidden layers: There are 3 layers, each with 256, 128, and 64 neurons respectively, and the activation function for all of them is rectified linear unit; Output layer: Outputs a proportion vector of length N (where N is the number of types of base wine), with each dimension representing the proportion of the corresponding base wine.

[0058] The output blending vector is normalized by Softmax so that the sum of all original wine blending ratios equals 1, ensuring practical compatibility.

[0059] The set of high-contribution flavor features extracted from the flavor contribution distribution map in the previous step is input into the generator as an additional flavor guidance vector to enhance the generator's ability to capture key flavor factors. This guidance vector is structurally connected in the input layer through feature masking, allowing it to participate in the generator's forward propagation.

[0060] The guiding vector sets the gating weights to control the degree of their influence on the final output. The trainable gating factor can be dynamically adjusted during model training, thereby achieving controlled output of personalized flavor features.

[0061] A discriminator is constructed to determine whether the generated base wine combination is consistent with the distribution of historical blending schemes, and at the same time, whether its flavor response vector matches the target flavor label.

[0062] The discriminator input is: the proportion vector of the generated sample and its flavor response vector (obtained from the trained flavor response model); The discriminator outputs the degree of match between the true and false labels and the flavor profile. The loss function employs a hybrid loss mechanism, which includes adversarial loss (used to approximate the true data distribution) and flavor similarity loss (using the cosine similarity between the target label vector and the flavor response vector as an evaluation metric).

[0063] By alternately optimizing the parameters of the generator and discriminator, the generator continuously improves the realism of the generated samples and the accuracy of flavor matching. The training process is iterative until the similarity between the generated samples and the target flavor stably exceeds a set threshold (e.g., 0.95).

[0064] After training, multiple noise inputs and target labels are continuously sampled from the generator to generate no fewer than 100 candidate combinations of base wines. Each combination is then input into the flavor response model to calculate its flavor vector, which is then ranked based on similarity to the target label vector.

[0065] Select the top K (e.g., top 5) similarity scores for the blending schemes, and output the final candidate set of original wine combinations and their corresponding blending suggestions to provide data support for subsequent sensory verification and experimental blending.

[0066] This step innovatively achieves reverse reasoning of the original liquor combination from the target flavor label by constructing a guided neural network, overcoming the bottleneck problem of "difficulty in reverse expression of flavor perception" in traditional baijiu blending, and has a high degree of automation and personalized customization capabilities.

[0067] To verify whether the reverse-guided neural network proposed in this invention can achieve higher flavor fitting accuracy and sensory expert satisfaction when generating original wine combination ratio schemes compared to traditional similar flavor retrieval methods, the following experiment was designed: The training data sources include: 4,232 blending schemes of raw spirits collected from a winery over the past five years; each scheme is accompanied by physicochemical indicators, GC-MS spectra, sensory evaluation scores and flavor label text.

[0068] The data processing methods include: all flavor label texts are preprocessed and converted into 32-dimensional label vectors; the original wine combination scheme is converted into an N-dimensional ratio vector (N=9, representing 9 main types of original wine); and the flavor response vector is obtained from the flavor feature mapping through a trained attention model.

[0069] Model training settings include: The generator takes 128 dimensions as input and outputs a 9-dimensional ratio vector. The discriminator is a three-layer fully connected network, and its output is whether it is a real sample and a flavor fit score. Training epochs: 100 epochs, batch size: 64, optimizer: Adam, learning rate: 0.0002; The baseline comparison method was K-nearest neighbor historical sample flavor matching (K=3).

[0070] Evaluation indicators include: Flavor Similarity Score (FSS): The cosine similarity between the flavor response vector generated using the scheme and the target label vector. The closer the value is to 1, the better the fit.

[0071] Expert Satisfaction Rate (ESR): The generated proposal was blind-tasted by five national first-class wine tasters to evaluate whether it "fits the target description" using a Yes / No scoring method.

[0072] Composition Novelty Index (CNI): Calculates the average distance between the generated composition scheme and the historical scheme in Euclidean space to evaluate whether it has combinatorial innovation.

[0073] The experimental results include: Table 1: Comparison of Flavor Matching Effects and Sensory Evaluations of Different Blending Methods

[0074] The experimental data show that the reverse-guided neural network method proposed in this invention improves the flavor fitting accuracy by about 8.1% and the sensory expert satisfaction rate by 24% compared with the traditional K-nearest neighbor retrieval method. It also has a significantly higher innovation in the combination of base spirits and can be used to generate high-fit blending schemes that are not already in history, thereby significantly improving the flavor consistency and personalized development capability of baijiu products.

[0075] The corresponding ratio recommendations are reconstructed through perceptual simulation. Combined with the discrimination model based on historical expert evaluations, the subjective consistency score of each candidate scheme is calculated, and the one with the highest score is selected as the preferred blending scheme.

[0076] Each candidate blend of base spirits is input as a blending vector into the trained flavor response prediction model. The model takes an N-dimensional blending vector (where N is the number of base spirit types, set to 9) as input and outputs a M-dimensional flavor response vector (where M is the number of flavor labels, set to 32) to simulate the combined effect of blending on sensory characteristics.

[0077] The flavor response prediction model employs a multilayer perceptron architecture, and the network includes: Input layer: 9-dimensional balancing vector; Hidden layers: 2 layers, each containing 64 and 32 neurons respectively, with rectified linear units as the activation function; Output layer: Outputs a 32-dimensional flavor response vector; The training loss function uses the mean squared error function, and the training data comes from a paired dataset of historical proportions and flavor labels.

[0078] After training, the model can predict the corresponding sensory flavor characteristics based on any combination of proportions, realizing the simulation reconstruction "from proportion to flavor".

[0079] The flavor response vector obtained above is input into the subjective consistency discrimination model. This model aims to determine whether the flavor response meets the historical experts' perception of consistency in high-quality blending schemes.

[0080] The model adopts a binary classification structure. The input is a 32-dimensional flavor response vector, and the output is a single scalar with a value range of 0 to 1, which is used to represent the subjective consistency score of the flavor scheme. The higher the value, the closer it is to the historical high-quality flavor.

[0081] The model structure is as follows: Input layer: 32-dimensional flavor response vector; Hidden layers: 2 layers, with 64 and 32 nodes respectively, and the activation function is the hyperbolic tangent function; Output layer: The Sigmoid activation function outputs a score between 0 and 1; The label data comes from the sensory evaluation levels corresponding to the historical matching schemes, which are divided into three levels: "excellent", "good" and "average". The "excellent" samples are marked as 1, and the rest are marked as 0, which constitutes a binary classification task.

[0082] To improve the accuracy of subjective consistency ratings and their consistency with human senses, this invention employs labeled supervised learning to train the subjective consistency discrimination model.

[0083] Training data sources include: historical blending sample ratio schemes; corresponding expert scores or flavor description texts; and evaluation level labels formed by scores or text mapping (e.g., "excellent" is labeled as 1, and "average" is labeled as 0).

[0084] Cross-entropy is used as the loss function to minimize the error between the model's output probability and the true label. The Adam optimizer is used during training, with a learning rate of 0.0005, 50 training epochs, and a batch size of 32. The F1 score and accuracy are monitored using a validation set to ensure the model can accurately distinguish between "high-quality sensory solutions."

[0085] The subjective consistency discrimination model does not directly use the original expert evaluations as the training basis. Instead, it systematically processes historical sensory evaluation data before training. Specifically, for multiple expert evaluation records corresponding to the same original wine sample, unified identification binding and aggregation are first performed to form a multi-evaluation dataset under the same flavor dimension. Then, the score distribution in this dataset is discretized to remove abnormal evaluations that deviate from the overall distribution center. Weights are then assigned based on the stability and consistency of each expert's historical evaluations, and the remaining evaluations are weighted and fused to obtain a bias-free fused evaluation result. On this basis, the original sensory description text is transformed into a numerical vector expression through flavor tag embedding and jointly mapped with the fused score result. This establishes a correspondence between subjective descriptions and numerical evaluations in a unified semantic space, achieving a structured reconstruction of sensory evaluation data.

[0086] After reconstruction, the reconstructed flavor label vectors are further aligned with the physicochemical indicators and GC-MS volatile component characteristics of the original wine through multimodal consistency processing. Only evaluation data consistent with the changing trends of objective characteristics are retained, thereby further reducing noise interference caused by subjective bias. Finally, the evaluation results after bias removal, semantic unification, and consistency screening are used as training labels to input the subjective consistency discrimination model. This allows the model to learn the mapping relationship between flavor response and multi-expert consensus, rather than the subjective judgment of a single expert. Therefore, through multi-stage processing and constraints on subjective evaluation data, it is transformed into a stable and reliable training basis.

[0087] All candidate base spirit blends were evaluated using flavor response prediction and subjective consistency scoring to obtain corresponding scores. The candidate blends were then ranked in descending order of their scores, and the blend with the highest score was selected as the final preferred blending ratio recommendation.

[0088] This scheme will be highly consistent with the target label in the flavor semantic space and achieve the highest matching degree in subjective sensory judgment. It has the ability to meet the standards in both flavor rationality and sensory perceptibility at the same time, and can be used for subsequent experimental verification and actual production formulation.

[0089] Based on the preferred blending scheme, check whether its sensory semantic distance with the target flavor vector meets the threshold requirement. If it does not meet the threshold, update the parameters of the reverse generation network until the flavor fitting accuracy is met.

[0090] To ensure that the selected optimal blending scheme not only scores the highest in the model output but also exhibits a high degree of consistency with the preset target flavor in the sensory semantic space, this invention proposes a semantic verification method based on flavor vector distance. This method is used to construct a closed-loop control mechanism for flavor generation, further improving the sensory accuracy of the model's generated results. The specific steps are as follows: The preferred blending ratio is input as a vector into a trained flavor response prediction model to generate the corresponding flavor response vector. The flavor response prediction model employs a multilayer perceptron structure, consisting of an input layer, two hidden layers, and an output layer. The output is a 32-dimensional flavor vector, representing the predicted intensity of 32 standard flavor labels. This simulation process modeled the sensory performance of the blending ratio in actual tasting, forming the basis for subsequent semantic distance calculations.

[0091] The flavor response vector obtained above, along with the target flavor label vector, is input into the distance calculation module. This module uses the normalized cosine distance function as the method for calculating sensory semantic distance. The specific calculation method is as follows: First, perform L2 normalization on the two vectors, which means dividing each vector by its own magnitude. Then calculate the dot product between the two normalized vectors; Finally, the following function is used to calculate the distance value: sensory semantic distance = 1 minus the dot product of the two vectors.

[0092] The obtained sensory semantic distance ranges from 0 to 2. If the two are completely consistent, the distance is 0; if the directions are completely opposite, the distance is 2. To improve the sensitivity of semantic discrimination, this invention linearly compresses the distance to between 0 and 1 and defines it as a semantic distance index.

[0093] To ensure that the sensory difference between the generated solution and the target flavor is within an acceptable range, this invention sets a fixed sensory semantic distance threshold. The threshold is set to 0.1, indicating that the angle between the flavor response vector and the target label should be close to 0, meaning that the sensory expression is almost completely identical.

[0094] If the calculated sensory semantic distance is less than or equal to 0.1, it is considered to meet the target flavor requirement; if it is greater than the threshold, it is considered that the scheme deviates from the target flavor description in actual sensory perception and feedback correction is required.

[0095] When the optimal matching scheme fails the semantic distance verification, the system uses this result as a feedback signal to input into the reverse-guided neural network. The specific operation is as follows: The original target label vector is kept unchanged; the random noise input or high-contribution flavor feature guide vector is adjusted to enable the generator to explore different potential combination spaces; the generator re-outputs the original wine combination ratio scheme; the simulation and verification process of the new scheme is repeated; the process is repeated until the generated scheme meets the standard requirement that the sensory semantic distance is less than or equal to 0.1.

[0096] In this way, the present invention forms a closed-loop control chain of "target flavor → generated scheme → flavor simulation → semantic verification → feedback optimization", which enables the output scheme to have higher sensory interpretability and matching degree while having a high model score.

[0097] Example 2, please refer to Figure 2 As shown in this embodiment, the big data analysis and decision-making system for baijiu blending includes: Data acquisition module: acquires physicochemical index data, GC-MS volatile component spectra, historical sensory evaluation records and blending ratio scheme text of multiple batches of raw wine, and constructs a heterogeneous flavor raw dataset; Feature extraction module: Performs unified feature transformation on the heterogeneous flavor raw dataset, uses spectral autoencoder to extract latent component concentration feature vectors of volatile components, uses word vector encoding technology to vectorize the text ratio scheme, and constructs a flavor multimodal feature matrix at a unified scale; Flavor mapping modeling module: Based on the flavor multimodal feature matrix, construct a micro-macro flavor mapping map, use an attention mechanism to learn the contribution weight of each flavor feature to the target flavor label, and output the flavor contribution distribution map of each sample; Flavor Reverse Generation Module: Input the target blending flavor requirement as a label vector, use a reverse-guided neural network to generate a candidate set of base wine combinations that are closest to the target flavor, and output corresponding blending suggestions; Perceptibility simulation evaluation module: Performs perceptibility simulation reconstruction on the corresponding ratio recommendations, combines the discrimination model of historical expert evaluation, calculates the subjective consistency score of each candidate scheme, and selects the one with the highest score as the preferred mixing scheme; Flavor Fitting Verification Module: Based on the preferred blending scheme, verify whether its sensory semantic distance with the target flavor vector meets the threshold requirement. If it does not meet the threshold, update the parameters of the reverse generation network until the flavor fitting accuracy is met.

[0098] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A big data analysis and decision-making method for blending baijiu (Chinese liquor), characterized by: include: Acquire physicochemical index data, GC-MS volatile component spectra, historical sensory evaluation records, and blending ratio scheme texts of multiple batches of raw wine to construct a heterogeneous flavor raw dataset; A unified feature transformation was performed on the heterogeneous flavor original dataset. A spectral autoencoder was used to extract the latent component concentration feature vectors of volatile components. Word vector encoding technology was used to vectorize the text ratio scheme and construct a flavor multimodal feature matrix at a unified scale. Based on the flavor multimodal feature matrix, a micro-macro flavor mapping map is constructed. An attention mechanism is used to learn the contribution weight of each flavor feature to the target flavor label, and the flavor contribution distribution map of each sample is output. The target blending flavor requirement is input as a label vector. A back-guided neural network is used to generate a candidate set of base spirits that are closest to the target flavor, and the corresponding blending ratio suggestions are output. The corresponding ratio recommendations are reconstructed through perceptual simulation. Combined with the discrimination model of historical expert evaluation, the subjective consistency score of each candidate scheme is calculated, and the one with the highest score is selected as the preferred mixing scheme. Based on the preferred blending scheme, check whether its sensory semantic distance with the target flavor vector meets the threshold requirement. If it does not meet the threshold, update the parameters of the reverse generation network until the flavor fitting accuracy is met.

2. The big data analysis and decision-making method for blending baijiu according to claim 1, characterized in that: The steps for constructing a flavor multimodal feature matrix at a uniform scale include: Peak normalization was performed on the gas chromatography-mass spectra, and a self-encoding neural network consisting of a convolutional encoder and a cyclic decoder was used to extract the concentration feature vector of potential volatile components in the spectra. The text of the blending scheme is structurally analyzed to extract the relationship between the types of base wines and their corresponding proportions. The blending scheme is then converted into a numerical vector representation using a word embedding model-based encoding method. Standardize the mapping between physicochemical index data and sensory evaluation labels to construct a physicochemical feature submatrix and a sensory feature submatrix under a unified dimension. By dimensional alignment and feature concatenation of spectral vectors, text proportion vectors, and physicochemical and sensory feature sub-matrices, a flavor multimodal feature matrix at a unified scale is constructed, which serves as the input feature benchmark for flavor mapping modeling.

3. The big data analysis and decision-making method for blending baijiu according to claim 1, characterized in that: The steps for constructing a micro-macro flavor mapping map based on the flavor multimodal feature matrix include: The flavor multimodal feature matrix is ​​input into a multilayer perceptron to extract high-order nonlinear flavor feature representations; Attention matching is performed on the higher-order nonlinear flavor features and the target flavor label vector to generate the attention weight distribution of each feature dimension relative to the target label; Calculate the weighted flavor response vector by summing the features according to their corresponding attention weights to obtain the flavor response result for each sample; Based on the similarity between the flavor response vector and the target label, a flavor contribution distribution map is generated, representing the degree of contribution of each flavor feature to the achievement of the target flavor.

4. The big data analysis and decision-making method for blending baijiu according to claim 3, characterized in that: The step of generating a flavor contribution distribution map based on the similarity between the flavor response vector and the target label includes: The flavor response vector and the target flavor label vector of each sample are input into the similarity calculation module, and the cosine similarity function is used to calculate the similarity score between them. The original attention weight matrix is ​​normalized and adjusted based on the similarity score. The normalized feature weights are mapped onto a two-dimensional plane, and the degree of matching of each flavor feature with the target label is visualized in the form of a heatmap, forming a flavor contribution distribution map. The flavor contribution distribution map is segmented by thresholding to extract the set of high-contribution features, which are then used as priority flavor factors for subsequent candidate base wine screening and blending optimization.

5. The big data analysis and decision-making method for blending baijiu according to claim 1, characterized in that: The step of generating a candidate set of base wine combinations that most closely resemble the target flavor using a reverse-guided neural network includes: The target blending flavor requirement is encoded into a fixed-dimensional label vector through a flavor label embedding model; a neural generative network is constructed with the original wine ratio vector as the generation target, and the label vector is input into the generator as a guiding condition; the set of high-contribution flavor features extracted from the flavor contribution distribution map is used as an additional input guiding vector; the generator and discriminator are optimized through adversarial training so that the generator output can simulate the feature distribution of real historical blending schemes and is as close as possible to the target flavor vector in the flavor response space, thereby outputting a candidate set of original wine combinations and their suggested ratios.

6. The big data analysis and decision-making method for blending baijiu according to claim 1, characterized in that: The step of calculating the subjective consistency score for each candidate solution and selecting the one with the highest score as the preferred matching solution includes: Each candidate ratio vector is input into the trained flavor response prediction model, and the corresponding flavor response vector is generated through simulation. Construct a subjective consistency discrimination model based on historical sensory evaluation data, input the flavor response vector, and output a subjective consistency score; The discrimination model was trained under supervision using historical evaluation corpus from sensory experts, and the cross-entropy loss function was aligned with the evaluation labels. All candidate blending schemes are sorted according to subjective consistency scores, and the candidate scheme with the highest score is selected as the final preferred blending scheme.

7. The big data analysis and decision-making method for blending baijiu according to claim 1, characterized in that: Based on the preferred blending scheme, the step of verifying whether its sensory semantic distance with the target flavor vector meets the threshold requirement includes: The optimal blending ratio is input into the flavor response prediction model, and the corresponding flavor response vector is output by simulation. The flavor response vector and the target flavor label vector are calculated using a sensory semantic distance function, which is a normalized cosine distance function. A sensory semantic distance threshold is set. When the distance does not meet the threshold requirement, the feedback generation network adjusts the input noise and label vector, guiding the generator to re-output the original wine combination scheme.

8. A big data analysis and decision-making system for blending baijiu (Chinese liquor), used to implement the big data analysis and decision-making method for blending baijiu as described in any one of claims 1-7, characterized in that: include: Data acquisition module: acquires physicochemical index data, GC-MS volatile component spectra, historical sensory evaluation records and blending ratio scheme text of multiple batches of raw wine, and constructs a heterogeneous flavor raw dataset; Feature extraction module: Performs unified feature transformation on the heterogeneous flavor raw dataset, uses spectral autoencoder to extract latent component concentration feature vectors of volatile components, uses word vector encoding technology to vectorize the text ratio scheme, and constructs a flavor multimodal feature matrix at a unified scale; Flavor mapping modeling module: Based on the flavor multimodal feature matrix, construct a micro-macro flavor mapping map, use an attention mechanism to learn the contribution weight of each flavor feature to the target flavor label, and output the flavor contribution distribution map of each sample; Flavor Reverse Generation Module: Input the target blending flavor requirement as a label vector, use a reverse-guided neural network to generate a candidate set of base wine combinations that are closest to the target flavor, and output corresponding blending suggestions; Perceptibility simulation evaluation module: Performs perceptibility simulation reconstruction on the corresponding ratio recommendations, combines the discrimination model of historical expert evaluation, calculates the subjective consistency score of each candidate scheme, and selects the one with the highest score as the preferred mixing scheme. Flavor Fitting Verification Module: Based on the preferred blending scheme, verify whether its sensory semantic distance with the target flavor vector meets the threshold requirement. If it does not meet the threshold, update the parameters of the reverse generation network until the flavor fitting accuracy is met.

Images