A liquor flavor analysis method and system based on artificial intelligence

Abstract

The application discloses a liquor flavor analysis method and system based on artificial intelligence, relates to the technical field of artificial intelligence data analysis, and comprises the following steps: collecting the spectrum and electrochemical parameters of a liquor sample, performing constant-temperature control and dark current correction on the collected signals, and generating a standardized characteristic vector; inputting the standardized characteristic vector into a liquor flavor vertical large model for analysis, generating a model fusion characteristic vector by a convolutional neural network and in combination with random forest mapping physicochemical parameters, and outputting the sweetness grade, process type, fragrance type classification and flavor structure of the liquor sample by using an attention mechanism; in the model training stage, the standardized characteristic vector of a historical liquor sample and corresponding artificial product evaluation data are used as a training sample set to construct a mapping relationship between the spectrum, physicochemical parameters and artificial product evaluation results, and the liquor flavor vertical large model is iteratively updated. The method realizes intelligent replacement of the function of an artificial liquor maker, and improves practical value.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence data analysis technology, specifically to a method and system for analyzing the flavor of baijiu (Chinese liquor) based on artificial intelligence. Background Technology

[0002] With the rapid development of artificial intelligence technology and multimodal data analysis methods, spectral analysis and electrochemical detection are widely used in food and beverage quality assessment. Especially in the liquor industry, spectral technology can quickly capture flavor-related components such as total acid and total esters in the liquor, while electrochemical sensors can achieve real-time measurement of indicators such as alcohol content and methanol. In recent years, machine learning algorithms such as convolutional neural networks (CNN) and random forests have been increasingly applied to deep feature extraction and pattern recognition of chemical and spectral data. Multimodal fusion models enable the efficient integration and analysis of complex sample information, providing new technical means for flavor determination and process optimization.

[0003] Existing methods for analyzing the flavor of baijiu (Chinese liquor) primarily rely on manual sensory evaluation or single physicochemical tests, resulting in issues such as strong subjectivity, difficulty in data standardization, low efficiency, and poor batch-to-batch consistency. Some automated methods based on spectroscopy and electrochemistry lack systematic correction for environmental interference; for example, temperature fluctuations or dark current drift can affect the accuracy of spectral measurements, leading to significant fluctuations in absorbance signals and an inability to stably generate standardized feature vectors. Traditional data analysis models typically use only a single data source, making it difficult to simultaneously integrate spectral and physicochemical parameters. They also lack the ability to extract deep nonlinear features from complex, multidimensional samples, thus failing to reach the level of human sommeliers in areas such as sweetness grading, process type determination, and refined flavor analysis. Furthermore, existing methods fail to achieve dynamic feature weight allocation and multidimensional judgment fusion, resulting in poor stability of judgment results across different samples or batches. This makes it difficult to generate quantifiable standardized outputs and provides a reliable data foundation for subsequent intelligent process optimization or model iteration. In summary, existing technologies cannot simultaneously eliminate environmental interference, unify standardized features, integrate spectral and physicochemical parameters, perform deep nonlinear feature extraction, and achieve dynamic judgment in baijiu flavor analysis. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is that existing methods for analyzing the flavor of baijiu rely on subjective judgment by human senses, cannot effectively integrate and standardize spectral and physicochemical parameters, and lack deep feature extraction and dynamic judgment mechanisms. The invention also addresses how to achieve artificial intelligence-based spectral and electrochemical multimodal fusion, standardized feature generation, deep nonlinear feature extraction, and dynamic judgment of baijiu sweetness and processing type.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: Spectroscopic and electrochemical parameters of baijiu samples are collected; the collected signals are subjected to isothermal control and dark current correction to generate standardized feature vectors; the standardized feature vectors are input into a large-scale baijiu flavor vertical model for analysis; a model fusion feature vector is generated by mapping physicochemical parameters using a convolutional neural network and a random forest; and an attention mechanism is used to output the sweetness level, processing type, aroma classification, and flavor structure of the baijiu samples. During the model training phase, the standardized feature vectors of historical baijiu samples and corresponding human tasting data are used as the training sample set to construct the mapping relationship between spectra, physicochemical parameters, and human tasting results, enabling iterative updates of the large-scale baijiu flavor vertical model.

[0007] As a preferred embodiment of the artificial intelligence-based baijiu flavor analysis method described in this invention, the method for collecting spectral and electrochemical parameters of baijiu samples includes: sequentially irradiating a quartz cuvette with a four-wavelength LED light source; collecting absorbance signals at each wavelength using a silicon photodiode; inputting the absorbance signals into a signal amplification module and processing them through a low-pass filter; performing continuous sampling on the main control board and smoothing the signals using a moving average method; acquiring a dark current baseline with the light source off and automatically subtracting the dark current when measuring light intensity; and collecting samples from each baijiu sample multiple times and calculating the average value to form a standardized spectral feature vector.

[0008] As a preferred embodiment of the artificial intelligence-based baijiu flavor analysis method described in this invention, the generation of the standardized feature vector includes: injecting a baijiu sample into an electrochemical sensor detection cell; acquiring electrochemical signals through an alcohol content sensor, a methanol sensor, and an auxiliary sensor; inputting the electrochemical signals into a signal amplification module and processing them through a low-pass filter; performing moving average filtering on the main control board and combining it with temperature compensation logic to eliminate environmental interference on the detection of the chemical properties of the baijiu; and synchronously encapsulating the electrochemical signals and the standardized spectral feature vector to form a standardized feature vector for the baijiu sample.

[0009] As a preferred embodiment of the artificial intelligence-based baijiu flavor analysis method described in this invention, the generation of the model fusion feature vector includes: inputting a standardized spectral feature vector; performing multi-layer convolution processing on the total acid and total ester absorbance signals of the baijiu sample collected under four-wavelength LED illumination; extracting high-dimensional nonlinear spectral features by combining activation functions and pooling operations; outputting a high-dimensional spectral feature vector; inputting the electrochemical parameters in the standardized feature vector into a random forest; calculating the probability distribution of physicochemical parameters in the baijiu sample by constructing multiple decision trees; outputting a physicochemical parameter probability vector; concatenating the physicochemical parameter probability vector output by the random forest with the high-dimensional spectral feature vector in a fusion layer to generate the model fusion feature vector of the baijiu sample.

[0010] As a preferred embodiment of the artificial intelligence-based baijiu flavor analysis method described in this invention, the step of using an attention mechanism to output the sweetness level, processing type, aroma classification, and flavor structure of a baijiu sample includes: inputting the model fusion feature vector into the attention mechanism; calculating attention weights based on the contribution of each input feature to the baijiu flavor, aroma, and quality; dynamically allocating the proportion of features in the judgment; generating a weighted representation of each feature through matrix operations and normalization, and updating the baijiu sample representation corresponding to the fusion feature vector; inputting the baijiu sample representation corresponding to the fusion feature vector into a fully connected layer; forming a probability distribution vector by linearly mapping the fusion feature vector and combining it with normalization; determining the sweetness level, fermentation process type, and aroma category of the baijiu sample through the probability distribution vector; and constructing a flavor structure vector through the fusion feature vector.

[0011] As a preferred embodiment of the artificial intelligence-based baijiu flavor analysis method described in this invention, the training sample set includes: archiving the spectral feature vectors of historical baijiu samples after uniform dimension and standardization processing, and establishing an index; performing convolutional feature extraction and decision tree mapping processing on the spectral features and physicochemical features of each baijiu sample respectively to generate a high-dimensional spectral feature representation and a probability vector of physicochemical parameters; and forming a preliminary fusion feature vector in the fusion layer.

[0012] As a preferred embodiment of the artificial intelligence-based baijiu flavor analysis method described in this invention, the iterative update of the baijiu flavor vertical category model includes: pairing the fused feature vector of each training sample with the corresponding human evaluation result; calculating the error between the model output and the human evaluation result using a supervised learning method; adopting a batch training strategy and iterative optimization processing during the training process; iteratively adjusting the weights of the convolutional neural network kernel, the parameters of the random forest decision tree, and the weights of the fusion layer, and performing gradient updates in conjunction with the loss function and optimization algorithm.

[0013] As a preferred embodiment of the artificial intelligence-based baijiu flavor analysis system of the present invention, it includes: a data acquisition module, a model output module, and a model training module; the data acquisition module is used to acquire the spectral and electrochemical parameters of baijiu samples, perform isothermal control and dark current correction on the acquired signals, and generate standardized feature vectors; the model output module is used to input the standardized feature vectors into a large-scale baijiu flavor vertical model for analysis, generate model fusion feature vectors by mapping physicochemical parameters through a convolutional neural network and a random forest, and output the sweetness level, processing type, aroma classification, and flavor structure of the baijiu samples using an attention mechanism; the model training module is used to construct the mapping relationship between the spectral and physicochemical parameters and the results of the human evaluation of baijiu samples as a training sample set during the model training stage, and perform iterative updates of the large-scale baijiu flavor vertical model.

[0014] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement a method for analyzing the flavor of baijiu (Chinese liquor) based on artificial intelligence.

[0015] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of an artificial intelligence-based method for analyzing the flavor of baijiu (Chinese liquor).

[0016] The beneficial effects of this invention are as follows: This invention achieves high-precision and repeatable determination of Baijiu flavor through a systematic combination of multimodal feature acquisition and intelligent analysis. Specifically, by combining multi-wavelength LED spectral acquisition with electrochemical sensor measurement, the total acid, total ester, alcohol content, and methanol content of Baijiu are obtained simultaneously and with high precision. Dark current correction, filtering, and smoothing are used to form standardized feature vectors, thereby ensuring data stability and multimodal integrity. Furthermore, convolutional neural networks are used to extract high-dimensional nonlinear features of the spectrum, and random forests are used to perform probabilistic mapping of physicochemical parameters to generate fused feature vectors. This achieves complementary expression of spectral and physicochemical information, improving the sensitivity and interpretability of the flavor determination model. Finally, an attention mechanism is used to dynamically allocate feature weights, inputting weighted features into a fully connected layer to output sweetness level, processing type, aroma classification, and flavor structure, achieving intelligent and multi-dimensional determination of Baijiu flavor. This not only solves the problems of low accuracy in traditional single-modal detection and strong subjectivity in human tasting, but also forms a repeatable, interpretable, and intelligent flavor evaluation system, realizing intelligent replacement of the function of human sommeliers and improving the reliability, scientificity, and practical value of flavor analysis. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 The above is an overall flowchart of an artificial intelligence-based method for analyzing the flavor of baijiu (Chinese liquor) provided in Embodiment 1 of the present invention.

[0019] Figure 2 The flowchart shows a large-scale model of baijiu flavor vertical category, which is a baijiu flavor analysis method based on artificial intelligence provided in Embodiment 1 of the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0021] Example 1, referring to Figures 1-2 As an embodiment of the present invention, an artificial intelligence-based method for analyzing the flavor of baijiu (Chinese liquor) is provided, comprising:

[0022] S1: Collect the spectral and electrochemical parameters of the liquor sample. Collect the absorbance signals of total acid and total ester by illuminating the quartz cuvette with a four-wavelength LED light source. At the same time, use silicon photocell and electrochemical sensor to collect the alcohol content and methanol content. Perform isothermal control and dark current correction on the collected signals and generate standardized feature vectors.

[0023] Furthermore, the spectral and electrochemical parameters of the liquor samples were collected by sequentially illuminating a quartz cuvette with four-wavelength LED light sources; collecting absorbance signals at each wavelength using a silicon photodiode; inputting the absorbance signals into a signal amplification module and processing them through a low-pass filter; performing continuous sampling on the main control board and smoothing the signals using a moving average method; acquiring the dark current baseline with the light source off and automatically subtracting the dark current when measuring light intensity; and collecting samples from each liquor sample multiple times and calculating the average value to form a standardized spectral feature vector.

[0024] It should also be noted that a preferred scheme for collecting spectral and electrochemical parameters of baijiu samples specifically includes, for each baijiu sample , Indicates the first One sample of baijiu (Chinese liquor) Indicates the number of samples. , This indicates the total number of samples. The samples are injected into a quartz cuvette. The excellent optical transmittance of quartz creates a stable optical path environment, ensuring stable transmission of the optical signal during the detection process. Four-wavelength LED light sources are used for sequential illumination. , Indicates the first Each LED light source wavelength, Indicates the number index of the four LED light sources, Light intensity , This indicates the light energy density provided by a specific wavelength LED light source on the surface of a liquor sample, with a pulse frequency of 1 kHz and a duty cycle of 50%. By using a fixed-wavelength LED light source array to sequentially illuminate the sample, multi-wavelength spectral information can be acquired without the need for complex grating spectrometers or mechanical scanning structures, simplifying the structure of spectral detection equipment and improving the integration and portability of the detection system.

[0025] Silicon photovoltaic cells collect transmitted light signals Dark current was collected when the light source was off. The net transmitted light intensity was obtained. , is represented as:

[0026]

[0027] in, Indicates the first The first sample of baijiu was in the first... Net transmitted light intensity signal after dark current correction under each LED wavelength channel. Indicates the first The first sample of baijiu was in the first... The original transmitted light signal under each LED wavelength channel express The first sample of baijiu was in the first... By acquiring the dark current baseline under each LED wavelength channel and subtracting it in real time during the detection process, the influence of ambient light interference and photoelectric device drift on the detection results can be effectively eliminated, thereby improving the stability of the detection signal and reducing manual calibration operations. The transmitted light intensity of the blank reference liquid is utilized. Calculate absorbance , is represented as:

[0028]

[0029] in, Indicates the first The first sample of baijiu was in the first... The absorbance of each LED wavelength channel reflects the total acid or total ester content; Indicates the first The transmitted light intensity of the blank reference liquid under each LED wavelength channel.

[0030] The absorbance signal is amplified 10 times by a high input impedance amplifier module, and a low-pass filter removes high-frequency noise. The main control board samples at 100Hz, and the signal is smoothed and filtered using a moving average method. Outliers. Each wavelength was sampled five times, and the average value was used to form the spectral feature vector. , is represented as:

[0031]

[0032] in, Indicates a sample of baijiu (Chinese liquor) spectral eigenvectors, This represents the average absorbance of the sample at four wavelengths. Multiple automatic samplings and averaging yield stable and reliable spectral characteristic data, ensuring detection accuracy while improving the repeatability of results.

[0033] It should be noted that generating the standardized feature vector includes: injecting the liquor sample into the detection cell of the electrochemical sensor, acquiring electrochemical signals through the alcohol content sensor, methanol sensor and auxiliary sensors; inputting the electrochemical signals into the signal amplification module and processing them through low-pass filtering; performing moving average filtering on the main control board, and combining temperature compensation logic to eliminate the interference of the environment on the detection of the chemical properties of the liquor; and synchronously encapsulating the electrochemical signals and standardized spectral feature vectors to form the standardized feature vector of the liquor sample.

[0034] It should also be noted that a preferred scheme for generating standardized feature vectors specifically includes processing the liquor samples... The current signal is injected into an electrochemical detection cell and obtained through alcohol content, methanol, and auxiliary sensors. , Indicates a sample of baijiu (Chinese liquor) The original current signal of the electrochemical sensor, This corresponds to alcohol content, methanol, total acid, and total esters. The signal is amplified 20 times by a high input impedance amplification module, and a low-pass filter removes high-frequency noise above 50Hz. The main control board uses a sliding average window of 5 times, and temperature compensation is performed in conjunction with real-time temperature to obtain the corrected electrochemical signal. To form electrochemical characteristic vectors , is represented as:

[0035]

[0036] in, Indicates the first Electrochemical feature vectors of individual baijiu samples This indicates the alcohol content measurement after signal amplification, filtering, and temperature compensation. This represents the measured methanol content after signal amplification, filtering, and temperature compensation. The corrected electrochemical signal representing the total acid content. The corrected electrochemical signal representing the total ester content. Spectral eigenvector. With electrochemical eigenvectors Synchronously encapsulate and normalize to the [0, 1] interval to form a standardized feature vector. , is represented as:

[0037]

[0038] By fusing and normalizing the spectral and electrochemical detection results, optical and physicochemical characteristics of baijiu samples can be obtained simultaneously in a single detection process, forming more complete and stable multimodal characteristic data. This provides reliable input data for baijiu flavor analysis models, improves the automation level of the detection process, and enhances the practical operability of the equipment.

[0039] It should also be noted that the absorbance signals of total acid and total ester are collected by illuminating a quartz cuvette with a four-wavelength LED light source, while the alcohol content and methanol content are collected simultaneously using a silicon photocell and an electrochemical sensor, achieving synchronous acquisition of multi-dimensional physicochemical indicators of the liquor sample. Combined with constant temperature control and dark current correction, interference from ambient temperature fluctuations and photoelectric drift on the measurement data can be eliminated, ensuring the stability and repeatability of the acquired signals. Furthermore, a high input impedance amplification module and a low-pass filter are used to enhance and suppress noise in the absorbance and electrochemical signals. Standardized spectral and electrochemical feature vectors are generated in the main control board using moving average and multiple repeated sampling to obtain the average value, enabling the original signals to be transformed into numerical features of a uniform scale, facilitating subsequent AI modeling and processing. The technical challenge lies in ensuring the consistency of measurement accuracy and numerically quantifiable features under conditions of multi-dimensional signal acquisition and environmental interference. Compared with existing technologies, this invention optimizes and integrates different types of physicochemical parameters through dual-channel spectroscopy and electrochemistry, isothermal closed-loop, and dark current correction mechanisms, providing high-fidelity and quantifiable basic data input for the analysis of baijiu flavor, and achieving the goal of simultaneous, continuous, and accurate measurement that is difficult for traditional human wine tasters to achieve.

[0040] S2: Input the standardized feature vector into the large-scale model of baijiu flavor vertical category for analysis. Through convolutional neural network and combined with random forest mapping of physicochemical parameters, generate model fusion feature vector. Use attention mechanism to output the sweetness level, process type, aroma classification and flavor structure of baijiu samples.

[0041] Furthermore, refer to Figure 2 The generated model fusion feature vector includes: inputting a standardized spectral feature vector; performing multi-layer convolution processing on the total acid and total ester absorbance signals of the liquor sample collected under four-wavelength LED illumination; extracting high-dimensional nonlinear spectral features by combining activation functions and pooling operations; outputting a high-dimensional spectral feature vector; inputting the electrochemical parameters in the standardized feature vector into a random forest; calculating the probability distribution of physicochemical parameters in the liquor sample by constructing multiple decision trees; outputting a physicochemical parameter probability vector; concatenating the physicochemical parameter probability vector output by the random forest with the high-dimensional spectral feature vector in a fusion layer to generate the model fusion feature vector of the liquor sample.

[0042] It should also be noted that a preferred approach to fusing feature vectors specifically includes normalizing the spectral vectors extracted via spectral feature extraction (CNN). The input is a convolutional neural network (CNN), which processes the data through multiple convolutions, activation functions, and pooling to output high-dimensional nonlinear spectral features. Represented as:

[0043]

[0044] in, Indicates the first The standardized spectral feature vectors of a baijiu sample are used to extract high-dimensional nonlinear features through a convolutional neural network. This represents the feature extraction function of a convolutional neural network. It standardizes the physicochemical feature vectors. Input a random forest model, construct multiple decision trees, calculate the probability contribution of each physicochemical parameter in flavor determination, and output the probability vector of the physicochemical parameters. , is represented as:

[0045]

[0046] in, This represents the prediction function of a random forest, which incorporates high-dimensional linear spectral features. With physicochemical parameter probability vector Concatenate to generate a fused feature vector , is represented as:

[0047]

[0048] It should be noted that, referring to Figure 2 This paper utilizes an attention mechanism to output the sweetness level, fermentation process type, aroma type, and flavor structure of baijiu samples. The process involves: inputting the model's fused feature vector into the attention mechanism; calculating attention weights based on the contribution of each input feature to the baijiu's flavor, aroma, and quality; dynamically allocating the proportion of each feature in the decision-making process; generating a weighted representation of each feature through matrix operations and normalization, and updating the baijiu sample representation corresponding to the fused feature vector; inputting the baijiu sample representation corresponding to the fused feature vector into a fully connected layer; forming a probability distribution vector by linearly mapping the fused feature vector and combining it with normalization; determining the sweetness level, fermentation process type, and aroma type of the baijiu sample using the probability distribution vector; and constructing a flavor structure vector using the fused feature vector.

[0049] It should also be noted that a preferred scheme for using the attention mechanism to output the sweetness level, processing type, aroma classification, and flavor structure of baijiu samples specifically includes fusing feature vectors. Input attention mechanism, through a trainable scoring function Calculate the contribution score for each feature , is represented as:

[0050]

[0051] in, Indicates the first The th baijiu sample in the fused feature vector One characteristic, This represents the feature index in the fused feature vector, which scores the contribution of each feature. Attention weights are obtained through softmax normalization. , is represented as:

[0052]

[0053] in, This represents the total dimension of the fused feature vector. Indicates the first The relative importance of each feature in the final flavor determination. Weighted feature vector. , is represented as:

[0054]

[0055] Weighted eigenvectors Input a fully connected layer, output a sweetness probability vector , is represented as:

[0056]

[0057] The highest probability corresponds to the category of low sweetness / medium sweetness / high sweetness level, among which, Indicates the first The probability vector of sweetness level for each sample of baijiu (Chinese liquor). Indicates the first The probability that a sample of baijiu belongs to the low-sweetness type. Show the first The probability that a sample of baijiu belongs to the medium-sweet type. Show the first The probability that a given baijiu sample belongs to the high-sweetness type; the weighted feature vector Input a fully connected layer, output a process probability vector , is represented as:

[0058]

[0059] The most probable category corresponds to solid / semi-solid / liquid fermentation processes, among which, Indicates the first The process probability vector of a sample of baijiu (Chinese liquor). Indicates the first The probability that a sample of baijiu (Chinese liquor) uses solid-state fermentation. No. The probability that a sample of baijiu (Chinese liquor) uses a semi-solid fermentation process. No. The probability that a given baijiu sample belongs to the liquid fermentation process; the weighted feature vector Input a fully connected layer, output a fragrance probability vector , is represented as:

[0060]

[0061] The most probable corresponding categories are soy sauce aroma / strong aroma / light aroma, among which, Indicates the first The aroma probability vector of a baijiu sample. Indicates the first The probability vector of a given baijiu sample belonging to the sauce-aroma type. Indicates the first The probability vector of a sample of baijiu belonging to the strong-aroma type. Indicates the first The probability vector of each baijiu sample belonging to the light aroma type; the weighted feature vector Input a fully connected layer to generate a flavor structure vector. , is represented as:

[0062]

[0063] The evaluation of body, smoothness, finish, aroma intensity, and flavor harmony by simulated human sommeliers was conducted. Indicates the first Flavor structure representation vectors for each baijiu sample Indicates the first The body and flavor score of each baijiu sample Indicates the first The smoothness score of each sample of baijiu. Indicates the first The aftertaste length score of each baijiu sample Indicates the first Aroma intensity score for each sample of baijiu. Indicates the first Flavor harmony score for each sample of baijiu.

[0064] It should also be noted that by processing spectral signals through multi-layer convolution, the model can automatically capture the high-dimensional nonlinear features of total acid and total esters in baijiu samples, improving the sensitivity to sweetness and aroma. By combining random forest mapping of physicochemical parameters, a quantitative probability mapping is established between alcohol content, methanol content, and other physicochemical indicators and flavor categories, allowing low-dimensional structured data to play a full role in the judgment. The fusion layer unifies the high-dimensional spectral features and physicochemical parameter probability vectors to generate a fused feature vector, realizing the integration of multi-source information and avoiding judgment bias caused by a single data source. Furthermore, an attention mechanism is introduced to dynamically allocate weights according to the contribution of each feature to the flavor, aroma, and quality of baijiu, enhancing the accuracy and robustness of the judgment. Through a fully connected layer, the weighted fused feature vector is mapped to a probability distribution vector, outputting sweetness level, process type, aroma category, and flavor structure, realizing the quantitative and interpretable judgment of baijiu samples. This enables the model to simulate the comprehensive evaluation ability of human wine tasters and supports sweetness grading, process identification, and aroma judgment.

[0065] S3: During the model training phase, standardized feature vectors of historical baijiu samples and corresponding human evaluation data are used as training sample sets to construct the mapping relationship between spectral and physicochemical parameters and human evaluation results, and to iteratively update the large-scale baijiu flavor vertical model.

[0066] Furthermore, the training sample set includes archiving the spectral feature vectors of historical liquor samples after unifying the dimensions and standardizing them, and establishing an index; performing convolutional feature extraction and decision tree mapping on the spectral and physicochemical features of each liquor sample to generate high-dimensional spectral feature representations and physicochemical parameter probability vectors; and forming a preliminary fusion feature vector in the fusion layer.

[0067] It should also be noted that a preferred scheme for the training sample set specifically includes standardizing the spectral feature vectors of historical baijiu samples. With electrochemical eigenvectors Archiving and indexing management are implemented to ensure that training data is traceable and accessible in sequence. Indicates the training sample index , This represents the total number of training samples, with each sample including the collection time. And a unique identifier. Spectral eigenvectors After being processed for dimensionality unification and standardization, the data is input into a convolutional neural network, where high-dimensional nonlinear spectral features are extracted through multi-layer convolution operations, nonlinear activation functions, and pooling operations. , is represented as:

[0068]

[0069] in, Indicates the first High-dimensional spectral feature representations of training samples are used to capture the complex spectral patterns of total acids and total esters under LED irradiation at different wavelengths. Electrochemical feature vectors are then used... Input a random forest model, construct multiple decision trees to calculate the probability contribution of each physicochemical parameter in flavor determination, and output a probability vector of the physicochemical parameters. , is represented as:

[0070]

[0071] in, Indicates the first The probability distribution vector of the physicochemical parameters of each training sample in flavor determination. Then, in the fusion layer... and Concatenate to generate a preliminary fused feature vector. :

[0072]

[0073] in, Indicates the first Preliminary fused feature vectors of training samples.

[0074] It should be noted that the iterative update of the large-scale model for the flavor vertical of baijiu includes: pairing the fused feature vector of each training sample with the corresponding human evaluation results; using supervised learning methods to calculate the error between the model output and the human evaluation results; adopting a batch training strategy and iterative optimization during training; iteratively adjusting the weights of the convolutional neural network kernels, the parameters of the random forest decision tree, and the weights of the fusion layer, and performing gradient updates in combination with loss functions and optimization algorithms.

[0075] It should also be noted that a preferred approach for iteratively updating the large-scale model of baijiu flavor verticals specifically includes using the initial fused vector as input for subsequent supervised learning, integrating multi-source features from spectral and physicochemical information. During training, a mini-batch training strategy is employed, with a set number of iterations. and learning rate For each batch of samples, the model outputs... Compared with human evaluation results The error between them is expressed through the loss function. The calculation is performed and expressed as follows:

[0076]

[0077] in, The model predicts the first The sweetness level, processing type, aroma classification, and flavor structure vector of each training sample. Indicates the corresponding number The vector of human evaluation results for each training sample. This indicates the batch size. The gradient descent method is used to update the convolutional neural network kernel weights, random forest decision tree parameters, and fusion layer weights. , is represented as:

[0078]

[0079] Through iterative training, the model can accurately map and fuse feature vectors. Compared with human evaluation results The relationship between these parameters enables precise determination of the sweetness level, processing type, aroma type, and flavor structure of baijiu samples. Simultaneously, to enhance generalization ability, regularization constraints are introduced during the training process. This forms the total loss function. , is represented as:

[0080]

[0081] in, This represents the regularization coefficient. Through the total loss function, the large-scale model for baijiu flavor categories can learn the mapping rules between spectral and physicochemical parameters and human evaluation results, forming a stable and reliable fusion feature representation.

[0082] It should also be noted that standardized management of training data is achieved by uniformly processing and indexing historical baijiu samples; the fused features are paired with human tasting results, and the prediction error is calculated through a loss function, realizing a supervised learning mechanism aimed at the experience of human baijiu tasters; gradient updates are combined to iteratively optimize the parameters of the convolutional network, random forest, and fusion layer, enabling the model to gradually approach the results of human tasting; at the same time, regularization constraints are introduced to suppress overfitting and improve the model's adaptability to unknown samples; a stable mapping relationship is established between spectra and physicochemical parameters and human tasting results, enabling the model to simulate the comprehensive judgment of human baijiu tasters on sweetness, processing type, aroma type, and flavor structure.

[0083] Example 2 is an embodiment of the present invention, which provides an artificial intelligence-based baijiu flavor analysis system, including a data acquisition module, a model output module, and a model training module.

[0084] The data acquisition module is used to collect the spectral and electrochemical parameters of baijiu samples, perform isothermal control and dark current correction on the acquired signals, and generate standardized feature vectors. The model output module is used to input the standardized feature vectors into the baijiu flavor vertical category model for analysis. Through convolutional neural networks and random forest mapping of physicochemical parameters, it generates model fusion feature vectors and uses an attention mechanism to output the sweetness level, processing type, aroma classification, and flavor structure of baijiu samples. The model training module is used to use the standardized feature vectors of historical baijiu samples and the corresponding human evaluation data as training sample sets during the model training stage to construct the mapping relationship between the spectrum, physicochemical parameters, and human evaluation results, and to iteratively update the baijiu flavor vertical category model.

Claims

1. A method for analyzing the flavor of baijiu (Chinese liquor) based on artificial intelligence, characterized in that, include: The spectral and electrochemical parameters of the liquor samples were collected, and the collected signals were subjected to isothermal control and dark current correction to generate standardized feature vectors. The standardized feature vectors are input into the large-scale model of Baijiu flavor vertical category for analysis. The model fusion feature vector is generated by using a convolutional neural network and combining it with random forest to map physicochemical parameters. The attention mechanism is used to output the sweetness level, process type, aroma type classification and flavor structure of Baijiu samples. During the model training phase, standardized feature vectors of historical baijiu samples and corresponding human evaluation data are used as training sample sets to construct the mapping relationship between spectral and physicochemical parameters and human evaluation results, and to iteratively update the large-scale baijiu flavor vertical model.

2. The method for analyzing the flavor of baijiu based on artificial intelligence as described in claim 1, characterized in that: The spectral and electrochemical parameters of the collected liquor samples include, The quartz cuvette was illuminated sequentially using a four-wavelength LED light source. Silicon photovoltaic cells collect absorbance signals at various wavelengths; The absorbance signal is input to the signal amplification module and processed by low-pass filtering; The main control board performs continuous sampling and uses a moving average method to smooth the signal; The dark current baseline is collected when the light source is off, and the dark current is automatically subtracted when measuring light intensity; Each baijiu sample was collected multiple times, and the average value was calculated to form a standardized spectral feature vector.

3. The method for analyzing the flavor of baijiu based on artificial intelligence as described in claim 2, characterized in that: The generation of standardized feature vectors includes, A sample of baijiu (Chinese liquor) was injected into the detection cell of an electrochemical sensor, and electrochemical signals were obtained through an alcohol content sensor, a methanol sensor, and an auxiliary sensor. The electrochemical signal is input to the signal amplification module and processed by a low-pass filter. The main control board performs moving average filtering and combines it with temperature compensation logic to eliminate environmental interference with the detection of the chemical properties of the wine. Electrochemical signals and standardized spectral feature vectors are encapsulated simultaneously to form standardized feature vectors for baijiu samples.

4. The method for analyzing the flavor of baijiu based on artificial intelligence as described in claim 3, characterized in that: The generated model integrates feature vectors including, Input a standardized spectral feature vector, perform multi-layer convolution processing on the total acid and total ester absorbance signals of the liquor sample collected under four-wavelength LED illumination, combine activation function and pooling operation to extract high-dimensional nonlinear spectral features, and output a high-dimensional spectral feature vector. The electrochemical parameters in the standardized feature vector are input into a random forest. Multiple decision trees are constructed to calculate the probability distribution of physicochemical parameters in the liquor samples, and the probability vector of physicochemical parameters is output. The probability vector of physicochemical parameters output by the random forest and the high-dimensional spectral feature vector are concatenated in the fusion layer to generate the model fusion feature vector of the liquor sample.

5. The method for analyzing the flavor of baijiu based on artificial intelligence as described in claim 4, characterized in that: The method of using attention mechanisms to output the sweetness level, processing type, aroma classification, and flavor structure of baijiu samples includes, The model integrates feature vector input attention mechanism, calculates attention weights based on the contribution of each input feature to the flavor, aroma and quality of baijiu, and dynamically allocates the proportion of features in the judgment. Attention weights generate a weighted representation of each feature through matrix operations and normalization, and update the baijiu sample representation corresponding to the fused feature vector; Input the liquor sample representation corresponding to the fused feature vector into the fully connected layer; A probability distribution vector is formed by linearly mapping the fused feature vector and combining it with normalization. The sweetness level, fermentation process type, and aroma category of the baijiu sample were determined by the probability distribution vector. Flavor structure vectors are constructed by fusing feature vectors.

6. The method for analyzing the flavor of baijiu based on artificial intelligence as described in claim 5, characterized in that: The training sample set includes, The spectral feature vectors of historical liquor samples are archived and indexed after being processed with unified dimensions and standardization. For each baijiu sample, convolutional feature extraction and decision tree mapping are performed on the spectral and physicochemical features to generate a high-dimensional spectral feature representation and a probability vector of physicochemical parameters. A preliminary fusion feature vector is formed in the fusion layer.

7. The method for analyzing the flavor of baijiu based on artificial intelligence as described in claim 6, characterized in that: The iterative update of the large-scale model for the flavor vertical category of baijiu includes, Pair the fused feature vector of each training sample with the corresponding human evaluation result; Supervised learning methods are used to calculate the error between the model output and the human evaluation results; During the training process, a batch training strategy and iterative optimization are adopted; The weights of the convolutional neural network kernels, the parameters of the random forest decision tree, and the weights of the fusion layer are iteratively adjusted, and gradient updates are performed in conjunction with the loss function and optimization algorithm.

8. A baijiu flavor analysis system based on artificial intelligence, employing the baijiu flavor analysis method based on artificial intelligence as described in any one of claims 1 to 7, characterized in that: It includes a data acquisition module, a model output module, and a model training module; The data acquisition module is used to acquire the spectral and electrochemical parameters of the liquor sample, perform isothermal control and dark current correction on the acquired signal, and generate a standardized feature vector. The model output module is used to input the standardized feature vector into the large-scale model of baijiu flavor for analysis. Through convolutional neural network and combined with random forest mapping of physicochemical parameters, it generates model fusion feature vector and uses attention mechanism to output the sweetness level, process type, aroma classification and flavor structure of baijiu samples. The model training module is used to construct the mapping relationship between the spectral and physicochemical parameters and the human evaluation results of historical baijiu samples as training sample sets during the model training stage, and to iteratively update the large baijiu flavor vertical model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the artificial intelligence-based baijiu flavor analysis method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the artificial intelligence-based baijiu flavor analysis method according to any one of claims 1 to 7.

Images