An intelligent blending system and method for Chinese liquor based on machine learning

Abstract

The present application provides a liquor intelligent blending system and method based on machine learning. An initial prediction of the blending ratio between liquor and base liquor is made through the content of flavor components in the liquor and base liquor to obtain a first blending parameter. Projection clustering is performed on the near-infrared spectral data of the liquor and base liquor to construct a spectral feature space, and then the baseline offset of each spectral feature wavelength in the spectral feature space is determined. Further, the characteristic wavelengths of the liquor and base liquor are screened out from the spectral feature space through the baseline offset, and then a secondary prediction of the blending ratio between the liquor and base liquor is made based on the characteristic wavelengths to obtain a second blending parameter. Based on the first blending parameter and the second blending parameter, the intelligent blending model is coupled and trained, and then the blending ratio between the liquor and base liquor is predicted through the trained intelligent blending model. By adopting the solution of the present application, intelligent coupling estimation of the liquor blending ratio can be realized, thereby improving the blending accuracy in liquor blending.

Description

Technical Field

[0001] This application relates to the field of machine learning technology. More specifically, this application relates to a liquor intelligent blending system and method based on machine learning. Background Art

[0002] Machine learning plays a key role in optimizing formulas and improving product consistency during the liquor blending process. By analyzing historical data and various parameters during the brewing process, machine learning models can identify key factors affecting liquor quality, thereby predicting the best combinations of different raw materials. This technology not only improves blending efficiency but also can quickly respond to changes in consumer taste preferences and flexibly adjust production strategies to ensure the high quality and market competitiveness of the final product.

[0003] In the prior art, traditional liquor blending techniques usually rely on a single data source, such as chemical components (flavor substances) or sensory evaluation. The single data source method often cannot comprehensively characterize the complexity of liquor, ignoring the interaction between flavor components and physical properties, resulting in insufficient prediction accuracy of the blending ratio. By using machine learning models to perform coupled analysis on multi-dimensional data features of liquor, effectively integrating flavor components and near-infrared spectroscopy data, and capturing the deep relationships between them, chemical and physical information can be integrated to make up for the one-sidedness caused by a single data source. Therefore, how to achieve intelligent coupled estimation of the liquor blending ratio to improve the blending accuracy in liquor blending has become a difficult problem faced by the industry. Summary of the Invention

[0004] This application provides a liquor intelligent blending system and method based on machine learning, which can achieve intelligent coupled estimation of the liquor blending ratio, thereby improving the blending accuracy in liquor blending.

[0005] In the first aspect, this application provides a method for training a liquor intelligent blending model based on machine learning, which is used for a liquor intelligent blending system to perform liquor intelligent blending, and includes the following steps:

[0006] Initialize and construct an intelligent blending model for liquor blending based on a siamese neural network;

[0007] Collect the content of flavor components of liquor and base liquor, and perform an initial prediction on the blending ratio between the liquor and the base liquor according to the content of flavor components of the liquor and the base liquor to obtain the first blending parameter of the intelligent blending model;

[0008] Collect the near-infrared spectroscopy data of the liquor and the base liquor, perform projection clustering on the near-infrared spectroscopy data to construct a spectral feature space, and then determine the baseline offset of each spectral feature wavelength according to the difference gradient between the spectral feature wavelengths in the spectral feature space;

[0009] Based on the baseline offset of each spectral characteristic wavelength, the characteristic wavelengths of the baijiu and the base liquor are screened out from the spectral characteristic space, and then the blending ratio between the baijiu and the base liquor is secondarily predicted according to the screened characteristic wavelengths to obtain the second blending parameter of the intelligent blending model;

[0010] Based on the first blending parameter and the second blending parameter, the intelligent blending model is coupled and trained, and then the blending ratio between the baijiu and the base liquor is predicted by the trained intelligent blending model.

[0011] Preferably, the initial prediction of the blending ratio between the baijiu and the base liquor according to the flavor component contents of the baijiu and the base liquor to obtain the first blending parameter of the intelligent blending model specifically includes:

[0012] Standardize the flavor component contents of the baijiu and the base liquor to obtain the standard flavor component content data;

[0013] Predict the initial blending ratio of the baijiu and the base liquor according to the standard flavor component content data, and use the initial blending ratio as the first blending parameter of the intelligent blending model.

[0014] Preferably, the construction of the spectral characteristic space by projection clustering of the near-infrared spectral data specifically includes:

[0015] Use principal component analysis to extract the main characteristics of the near-infrared spectral data to obtain a spectral characteristic set;

[0016] Perform band-level clustering on the spectral characteristics in the spectral characteristic set to obtain spectral characteristic clusters in multiple bands;

[0017] Project and map all the spectral characteristic clusters into a unified characteristic space to obtain the spectral characteristic space.

[0018] Preferably, determining the baseline offset of each spectral characteristic wavelength according to the difference gradient between the spectral characteristic wavelengths in the spectral characteristic space specifically includes:

[0019] Use the difference method to determine the difference gradient between the spectral characteristic wavelengths in the spectral characteristic space;

[0020] Based on the normal distribution hypothesis, determine the baseline wavelength of the spectral characteristics in the spectral characteristic space according to all the difference gradients;

[0021] Determine the baseline offset of the corresponding spectral characteristic wavelength through the baseline wavelength and each spectral characteristic wavelength.

[0022] Preferably, screening out the characteristic wavelengths of the baijiu and the base liquor from the spectral characteristic space through the baseline offset of each spectral characteristic wavelength specifically includes:

[0023] Compare the baseline offset of each spectral characteristic wavelength with the offset threshold;

[0024] When the baseline offset is less than or equal to the preset offset threshold, the spectral characteristic wavelength corresponding to the baseline offset is used as the effective spectral characteristic wavelength, and then all the effective spectral characteristic wavelengths are obtained;

[0025] All the effective spectral characteristic wavelengths are used as the characteristic wavelengths of the baijiu and the base liquor.

[0026] Preferably, the coupled training of the intelligent blending model based on the first blending parameter and the second blending parameter specifically includes:

[0027] Determine the coupling correlation between the first blending parameter and the second blending parameter;

[0028] Determine the shared weights between the two branch networks in the intelligent blending model according to the coupling correlation;

[0029] Use the first blending parameter and the second blending parameter as the prior parameters of the two branch networks in the intelligent blending model respectively, and then train the intelligent blending model through a preset training data set.

[0030] Preferably, the flavor component content refers to the specific concentrations of various volatile compounds in the baijiu and the base liquor.

[0031] In a second aspect, the present application provides a machine learning-based intelligent baijiu blending system, including an intelligent baijiu blending model training unit, and the intelligent baijiu blending model training unit includes:

[0032] A construction module for initializing and constructing an intelligent blending model for baijiu blending based on a siamese neural network;

[0033] A processing module for collecting the flavor component contents of the baijiu and the base liquor, and making an initial prediction of the blending ratio between the baijiu and the base liquor according to the flavor component contents of the baijiu and the base liquor to obtain the first blending parameter of the intelligent blending model;

[0034] The processing module is further configured to collect the near-infrared spectral data of the baijiu and the base liquor, perform projection clustering on the near-infrared spectral data to construct a spectral feature space, and then determine the baseline offset of each spectral characteristic wavelength according to the difference gradient between the spectral characteristic wavelengths in the spectral feature space;

[0035] The processing module is further configured to screen out the characteristic wavelengths of the white liquor and the base liquor from the spectral feature space according to the baseline offset of each spectral characteristic wavelength, and then perform a secondary prediction on the blending ratio between the white liquor and the base liquor based on the screened characteristic wavelengths to obtain a second blending parameter of the intelligent blending model;

[0036] An execution module is configured to perform coupled training on the intelligent blending model based on the first blending parameter and the second blending parameter, and then predict the blending ratio between the white liquor and the base liquor through the trained intelligent blending model.

[0037] In a third aspect, the present application provides a computer device, which includes a memory and a processor. The memory stores code, and the processor is configured to obtain the code and execute the above-mentioned method for training an intelligent blending model of white liquor based on machine learning.

[0038] In a fourth aspect, the present application provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the above-mentioned method for training an intelligent blending model of white liquor based on machine learning is implemented.

[0039] The technical solution provided by the disclosed embodiments of the present application has the following beneficial effects:

[0040] In the embodiments of the present application, first, an intelligent blending model for blending white liquor is initially constructed based on a Siamese neural network; the content of flavor components of white liquor and base liquor is collected, and an initial prediction is made on the blending ratio between the white liquor and the base liquor according to the content of flavor components of the white liquor and the base liquor to obtain a first blending parameter of the intelligent blending model; the near-infrared spectral data of the white liquor and the base liquor is collected, and projection clustering is performed on the near-infrared spectral data to construct a spectral feature space, and then the baseline offset of each spectral characteristic wavelength is determined according to the difference gradient between the spectral characteristic wavelengths in the spectral feature space; the characteristic wavelengths of the white liquor and the base liquor are screened out from the spectral feature space according to the baseline offset of each spectral characteristic wavelength, and then a secondary prediction is made on the blending ratio between the white liquor and the base liquor based on the screened characteristic wavelengths to obtain a second blending parameter of the intelligent blending model; coupled training is performed on the intelligent blending model based on the first blending parameter and the second blending parameter, and then the blending ratio between the white liquor and the base liquor is predicted through the trained intelligent blending model.

[0041] It can be seen that in this application, based on the first blending parameter and the second blending parameter, the intelligent blending model is coupled and trained, and then the blending ratio between white liquor and base liquor is predicted through the trained intelligent blending model. Among them, first, the blending ratio between white liquor and base liquor is initially predicted according to the flavor component content to obtain the first blending ratio, and this first blending ratio is used as the first blending parameter of the intelligent blending model. The first blending parameter acts as prior knowledge in the intelligent blending model, enabling the intelligent blending model to make reasonable predictions of the blending ratio in the initial stage, laying a foundation for subsequent coupling optimization based on spectral data, enhancing the processing ability of the intelligent blending model for multi-dimensional features, and thus improving the preliminary coupling effect between flavor components and near-infrared spectra. Secondly, the spectral characteristic wavelengths that can best reflect the differences between white liquor and base liquor are selected based on the baseline offset amounts of different wavelengths in the spectral characteristic space. The spectral characteristic wavelengths reflect the subtle changes in the physical properties of white liquor, which can better supplement the deficiencies of flavor components. On this basis, the second blending ratio is predicted using the selected spectral characteristic wavelengths, and this second blending ratio is used as the second blending parameter of the intelligent blending model. Further, through the correlation coupling analysis of the first blending parameter and the second blending parameter, the complementarity between different data sources (i.e., flavor components and near-infrared spectra) is realized, making up for the one-sidedness caused by a single data source, and then facilitating the subsequent coupling training of the intelligent blending model to improve the ability to grasp the complex chemical (i.e., flavor components) and physical (i.e., near-infrared spectra) relationships in white liquor blending, thereby improving the blending accuracy in white liquor blending. In summary, the implementation scheme of this application can realize the intelligent coupling estimation of the white liquor blending ratio, thereby improving the blending accuracy in white liquor blending. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 is an exemplary flowchart of a method for training an intelligent white liquor blending model based on machine learning in some embodiments of this application;

[0043] Figure 2 is a schematic structural diagram of a siamese neural network in some embodiments of this application;

[0044] Figure 3 is a schematic flowchart of a process for determining a spectral characteristic space in some embodiments of this application;

[0045] Figure 4 is a schematic structural diagram of an intelligent white liquor blending model training unit in some embodiments of this application;

[0046] Figure 5 is a schematic structural diagram of a computer device for implementing a method for training an intelligent white liquor blending model based on machine learning in some embodiments of this application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0047] To better understand the technical solution of this application, the technical solution of the application will be described in detail below in conjunction with the accompanying drawings of the specification and specific embodiments.

[0048] Reference Figure 1 , this figure is an exemplary flowchart of a training method for a liquor intelligent blending model based on machine learning in some embodiments of this application. The training method 100 for the liquor intelligent blending model based on machine learning mainly includes the following steps:

[0049] In step 101, an intelligent blending model for liquor blending is initialized and constructed based on a siamese neural network.

[0050] It should be noted that a siamese neural network (SNN) usually consists of two or more branch networks sharing weights. Each branch network is responsible for processing different input features. In this application, an intelligent blending model for liquor blending is initialized and constructed based on a siamese neural network. The core is to use the structure of the siamese neural network to perform contrastive learning on different input features to predict the optimal blending ratio between liquor and base liquor.

[0051] Specifically, when implemented, reference Figure 2 , this figure is a schematic structural diagram of a siamese neural network in some embodiments of this application. First, a double-branch network structure can be designed. Each branch network receives different feature information of liquor and base liquor, such as the content of flavor components and near-infrared spectrum data. Then, the content of flavor components is used as the initialization parameter of the first branch network, and the near-infrared spectrum data is used as the initialization parameter of the second branch network. Finally, a loss function of the siamese neural network is designed, and similarity measurement is performed through the loss function. In this application, a contrastive loss function can be used as the loss function of the siamese neural network to complete the initialization and construction of the intelligent blending model.

[0052] In step 102, the content of flavor components of liquor and base liquor is collected, and the blending ratio between the liquor and the base liquor is initially predicted according to the content of flavor components of the liquor and the base liquor to obtain the first blending parameter of the intelligent blending model.

[0053] It should be noted that the content of flavor components in the solution of this application refers to the specific concentrations of various volatile compounds in liquor and base liquor, including alcohols, esters, acids, aldehydes, ketones, and phenolic compounds. These components are quantitatively determined in gas chromatography-mass spectrometry analysis and can provide basic data for predicting the blending ratio. Since the content of flavor components of liquor and base liquor in different batches will vary, the liquor and base liquor used in this application during training are the same batch as those used for liquor blending after training, and a new round of training and blending is performed for each batch.

[0054] In specific implementation, the collection of the flavor component contents of the baijiu and the base liquor can be achieved by the following method, that is: an appropriate amount of baijiu and base liquor can be taken, first diluted and filtered to remove solid impurities to ensure that the sample is suitable for gas chromatography-mass spectrometry analysis, then a suitable gas chromatography column (such as a capillary column) is selected, appropriate temperature programs, carrier gas flow rates and injection volumes are set for compound separation, and then the separated compounds are introduced into a mass spectrometer, the analysis conditions of the ion source and the mass spectrometry are set, a suitable ionization method (such as electron bombardment or chemical ionization) is selected for mass spectrometry detection, and then the mass spectrometry diagrams of each component are obtained through mass spectrometry analysis, professional software is used for data processing to quantitatively analyze the flavor components in the baijiu and the base liquor and generate a flavor component spectrum, and finally, based on the generated flavor component spectrum, the main flavor components are identified and quantified, and the contents of the main flavor components are used as the flavor component contents for subsequent blending ratio prediction.

[0055] In some embodiments, the initial prediction of the blending ratio between the baijiu and the base liquor according to the flavor component contents of the baijiu and the base liquor to obtain the first blending parameter of the intelligent blending model can be achieved by the following steps:

[0056] Perform standardization processing on the flavor component contents of the baijiu and the base liquor to obtain standard flavor component content data;

[0057] Predict the initial blending ratio of the baijiu and the base liquor according to the standard flavor component content data, and use the initial blending ratio as the first blending parameter of the intelligent blending model.

[0058] In specific implementation, first, quantile standardization in the prior art can be used to perform standardization processing on the flavor component contents of the baijiu and the base liquor, and then the data set obtained after the standardization processing is used as the standard flavor component content data; then, the standard flavor component content data is used as the input parameter of a pre-trained baijiu blending model, and the baijiu blending model will analyze based on the features and relationships learned during training, and predict the blending ratio of the baijiu and the base liquor through the baijiu blending model to obtain the initial blending ratio of the baijiu and the base liquor, and then use this initial blending ratio as the first blending parameter of the intelligent blending model; it should be noted that the first blending parameter in this application refers to the blending ratio predicted based on the flavor component contents of the baijiu and the base liquor, and can be used as the prior knowledge of the intelligent blending model.

[0059] It should be noted that the pre-trained baijiu blending model in this application refers to a machine learning model trained on historical flavor component content data, aiming to predict the blending ratio between baijiu and base liquor. This baijiu blending model uses known flavor component content data and the corresponding blending ratios to learn the relationships and patterns between them, so as to achieve the prediction of the blending ratio. It should also be noted that the algorithm structure of the baijiu blending model in this application can adopt a support vector machine. In other embodiments, the algorithm structure of the baijiu blending model can also adopt other algorithm structures, which are not limited here.

[0060] In step 103, the near-infrared spectral data of the baijiu and the base liquor are collected, and projection clustering is performed on the near-infrared spectral data to construct a spectral feature space. Then, the baseline offset of each spectral feature wavelength is determined according to the difference gradient between the spectral feature wavelengths in the spectral feature space.

[0061] It should be noted that the solution of this application is based on the different responses of near-infrared spectra to the molecular vibration and rotation characteristics of chemical components (i.e., the chemical components of baijiu and base liquor), so as to quantitatively analyze the component content in the sample. The implementation principle of using near-infrared spectra to measure the component content in baijiu and base liquor is as follows: Near-infrared spectra are mainly analyzed through the absorption characteristics of molecular vibration and rotation. Different compounds (such as alcohols, esters, acids, etc.) have unique absorption peaks in specific wavelength ranges. When near-infrared light irradiates the sample, the molecules in the sample will absorb light of specific wavelengths, generating an absorption spectrum. Each chemical component corresponds to a specific absorption band. Therefore, the components in the sample can be identified and quantified spectroscopically. Further, by establishing a standard curve or regression model of known components and comparing the measured spectral data with the component concentrations of known samples, the component content in the sample (baijiu and base liquor) can be quantitatively analyzed.

[0062] Specifically, the collection of the near-infrared spectral data of the baijiu and the base liquor can be implemented in the following way: First, take appropriate samples of baijiu and base liquor, and ensure that the samples are clean and free of bubbles to avoid affecting the spectral measurement results. Then, set an appropriate wavelength range (usually 800 - 2500 nm) according to the instrument instructions, select an appropriate scanning speed and number of times, and then place the sample in the sample chamber of the spectrometer and start the instrument to collect spectra. Each sample can be measured multiple times to improve the reliability of the data. Finally, the collected near-infrared spectral data are recorded in real time, usually stored in the form of a relationship diagram between spectral intensity and wavelength, so as to obtain the near-infrared spectral data of the baijiu and the base liquor.

[0063] In some embodiments, refer to Figure 3 As shown, this figure is a schematic flowchart of the process for determining the spectral feature space in some embodiments of this application. In this embodiment, the projection clustering of the near-infrared spectral data to construct the spectral feature space can be implemented by the following steps:

[0064] In step 1031, the main features of the near-infrared spectral data are extracted by using principal component analysis to obtain a spectral feature set.

[0065] In step 1032, the spectral features in the spectral feature set are subjected to band-level clustering to obtain spectral feature clusters in multiple bands.

[0066] In step 1033, all the spectral feature clusters are projected and mapped into a unified feature space to obtain a spectral feature space.

[0067] In specific implementation, first, all the near-infrared spectral data are standardized, and then the standardized data are integrated into a matrix, where the rows represent the compound components in the sample and the columns represent the wavelengths. Further, the covariance matrix is calculated using the principal component analysis algorithm, and the first few principal components (usually the components with a cumulative contribution rate of 95%) are extracted, and the spectral features of all the extracted components are combined into a set as the spectral feature set. Then, using the hierarchical clustering algorithm, a distance matrix is calculated based on the similarity (such as Euclidean distance or cosine similarity) between the spectral features in the spectral feature set, and then according to the set distance threshold, similar spectral features are gradually merged to form spectral feature clusters in multiple bands. Finally, the principal components can be used as coordinate axes to construct a multi-dimensional feature space, and then the center points of each spectral feature cluster are mapped into the multi-dimensional feature space to form a spectral feature space.

[0068] In some embodiments, determining the baseline offset of each spectral feature wavelength according to the difference gradient between the spectral feature wavelengths in the spectral feature space can be implemented by the following steps:

[0069] Use the difference method to determine the difference gradient between the spectral feature wavelengths in the spectral feature space.

[0070] Based on the normal distribution hypothesis, determine the baseline wavelength of the spectral features in the spectral feature space according to all the difference gradients.

[0071] Determine the baseline offset of the corresponding spectral feature wavelength through the baseline wavelength and each spectral feature wavelength.

[0072] In specific implementation, first, the average difference in wavelength between the selected spectral feature and other spectral features in the spectral feature space is calculated using the difference method in the prior art, and this average difference is used as the wavelength difference gradient between the selected spectral feature and other spectral features, so as to obtain the wavelength difference gradient between each spectral feature and other spectral features; assuming that each difference gradient conforms to a normal distribution, the average value of all difference gradients is calculated, and this average value is used as the gradient mean value, and then the wavelength corresponding to or closest to the gradient mean value is selected as the baseline wavelength. Assuming that the wavelength corresponding to or closest to the gradient mean value is 600 nm, then the selected wavelength 600 nm is used as the baseline wavelength; the wavelengths of each spectral feature are obtained from the acquired spectral feature space, and the Euclidean distance between the baseline wavelength and the wavelengths of each spectral feature can be used as the baseline offset of the corresponding spectral feature wavelength.

[0073] It should be noted that the baseline offset in this application measures the offset degree of the spectral feature wavelength relative to its reference baseline. The larger the baseline offset, the greater the offset degree of the spectral feature wavelength relative to its reference baseline; the smaller the baseline offset, the smaller the offset degree of the spectral feature wavelength relative to its reference baseline. It should also be noted that in liquor blending, spectral data may be affected by background light, impurities in the sample or environmental factors, resulting in the up and down offset of spectral features. By calculating the baseline offset, these interferences can be eliminated, so as to obtain more accurate spectral information.

[0074] In step 104, the characteristic wavelengths of the liquor and the base liquor are screened out from the spectral feature space through the baseline offset of each spectral feature wavelength, and then the blending ratio between the liquor and the base liquor is predicted again based on the screened characteristic wavelengths to obtain the second blending parameter of the intelligent blending model.

[0075] In some embodiments, screening out the characteristic wavelengths of the liquor and the base liquor from the spectral feature space through the baseline offset of each spectral feature wavelength can be implemented by the following steps:

[0076] Compare the baseline offset of each spectral feature wavelength with the offset threshold.

[0077] When the baseline offset is less than or equal to the preset offset threshold, the spectral feature wavelength corresponding to this baseline offset is used as the effective spectral feature wavelength, and then all effective spectral feature wavelengths are obtained.

[0078] All the effective spectral feature wavelengths are used as the characteristic wavelengths of the liquor and the base liquor.

[0079] In specific implementation, first, an offset threshold can be preset in advance. The offset threshold can be set based on a large number of baseline offsets. In some embodiments, the average value of all baseline offsets can be used as the offset threshold. In other embodiments, other methods can also be used for setting, which is not limited here. Secondly, the baseline offset of each spectral characteristic wavelength is compared with the offset threshold. The spectral characteristic wavelength corresponding to the baseline offset less than or equal to the offset threshold is used as the effective spectral characteristic wavelength, and the spectral characteristic wavelength corresponding to the baseline offset greater than the offset threshold is used as the invalid spectral characteristic wavelength, and the invalid spectral characteristic wavelengths are removed, so as to obtain all the effective spectral characteristic wavelengths. Then, all the effective spectral characteristic wavelengths are used as the characteristic wavelengths of the baijiu and the base liquor.

[0080] In some embodiments, the second blending parameter of the intelligent blending model can be obtained by performing a secondary prediction on the blending ratio between the baijiu and the base liquor according to the selected characteristic wavelengths, which can be realized by the following steps:

[0081] Perform normalization processing on the selected characteristic wavelengths to obtain the normalized characteristic wavelengths;

[0082] Construct a baijiu blending prediction model according to the normalized characteristic wavelengths;

[0083] Predict the second blending parameter of the intelligent blending model through the baijiu blending prediction model.

[0084] In specific implementation, first, the selected characteristic wavelengths can be normalized by using the min-max normalization technique in the prior art, and the data obtained after the normalization processing is used as the normalized characteristic wavelengths. Then, a machine learning algorithm (such as a convolutional neural network) can be used to learn the correlation between the characteristic wavelengths and the blending ratio, so as to pre-train the baijiu blending model in advance. Further, the normalized characteristic wavelengths are input into the pre-trained baijiu blending model for initialization, and then the baijiu blending model is used to identify the characteristic wavelengths and map out the blending ratio between the baijiu and the base liquor, so as to complete the training of the baijiu blending prediction model. Finally, the blending ratio output by the baijiu blending prediction model is used as the second blending parameter of the intelligent blending model. It should be noted that the second blending parameter in this application refers to the blending ratio predicted based on the near-infrared spectral characteristic wavelengths of the baijiu and the base liquor.

[0085] In step 105, based on the first blending parameter and the second blending parameter, the intelligent blending model is coupled and trained, and then the blending ratio between the baijiu and the base liquor is predicted through the trained intelligent blending model.

[0086] In some embodiments, based on the first blending parameter and the second blending parameter, the coupled training of the intelligent blending model can be implemented by the following steps:

[0087] Determine the coupling correlation degree between the first blending parameter and the second blending parameter;

[0088] Determine the shared weight between two branch networks in the intelligent blending model according to the coupling correlation degree;

[0089] Respectively use the first blending parameter and the second blending parameter as the prior parameters of two branch networks in the intelligent blending model, and then train the intelligent blending model through a preset training data set.

[0090] It should be noted that in this application, through the coupled training of the first blending parameter (based on the flavor component content) and the second blending parameter (based on the near-infrared spectrum), the chemical and physical characteristics can be integrated, the coupling correlation degree between the two can be quantified, and then the shared weight of the dual-branch network in the intelligent blending model can be optimized. This process utilizes the complementary advantages of different data sources. By using the two parameters as prior knowledge, the learning ability and prediction accuracy of the model are further improved, the balance between flavor and molecular structure in the liquor blending process is ensured, and a more accurate blending ratio prediction is achieved.

[0091] It should also be noted that the coupling correlation degree in this application is an index to measure the correlation degree between the first blending parameter and the second blending parameter. The smaller the coupling correlation degree, the greater the correlation degree between the first blending parameter and the second blending parameter; the larger the coupling correlation degree, the smaller the correlation degree between the first blending parameter and the second blending parameter. In addition, the value range of the coupling correlation degree in this application is between 0 and 1.

[0092] Preferably, in the above embodiment, the coupling correlation degree between the first blending parameter and the second blending parameter can be determined by the following steps:

[0093] Obtain the flavor component contents of the liquor and the base liquor, and then determine the content variance of the flavor component contents;

[0094] Obtain the near-infrared spectrum data of the liquor and the base liquor, and then determine the spectrum variance of the near-infrared spectrum data;

[0095] Determine the fluctuation correlation degree of the determination of the compound component content in the sample through the content variance and the spectrum variance;

[0096] Determine the parameter difference between the first blending parameter and the second blending parameter;

[0097] Compensate the fluctuation correlation degree through the parameter difference value, and use the compensated fluctuation correlation degree as the coupling correlation degree between the first blending parameter and the second blending parameter.

[0098] In specific implementation, first, obtain the flavor component contents of the white liquor and the base liquor through the method of the above embodiment, then calculate the variance of the flavor component contents corresponding to the white liquor, and use this variance as the white liquor variance. Further, calculate the variance of the flavor component contents corresponding to the base liquor, and use this variance as the base liquor variance. Then, use the average value of the white liquor variance and the base liquor variance as the content variance of the flavor component contents. Secondly, obtain the near-infrared spectrum data of the white liquor and the base liquor through the method of the above embodiment. Calculate the spectrum variance of the white liquor through the near-infrared spectrum data corresponding to the white liquor, and use the spectrum variance of the white liquor as the white liquor spectrum variance. Calculate the spectrum variance of the base liquor through the near-infrared spectrum data corresponding to the base liquor, and use the spectrum variance of the base liquor as the base liquor spectrum variance. Further, the average value of the white liquor spectrum variance and the base liquor spectrum variance can be used as the spectrum variance of the near-infrared spectrum data. Thirdly, the cosine similarity between the content variance and the spectrum variance can be used as the fluctuation correlation degree for measuring the content of the compound components in the sample. Then, the absolute difference value between the first blending parameter and the second blending parameter can be used as the parameter difference value. Finally, the natural exponent of the opposite number of the parameter difference value can be used as the compensation factor. Furthermore, compensate the fluctuation correlation degree through the compensation factor, which is beneficial to reducing the systematic error introduced by measurement, so as to improve the accuracy of the coupling correlation degree. As a preferred embodiment, compensating the fluctuation correlation degree through the compensation factor can be implemented in the following manner, that is: multiply the compensation factor and the fluctuation correlation degree, and use the obtained product as the coupling correlation degree between the first blending parameter and the second blending parameter.

[0099] In specific implementation, determining the shared weight between two branch networks in the intelligent blending model according to the coupling correlation can be achieved in the following manner, that is: the coupling correlation can be used to evaluate the correlation between features, that is, the magnitude of the coupling correlation can be used as the shared weight between two branch networks in the intelligent blending model. For example, when the coupling correlation is 0.85, the shared weight of the intelligent blending model can be initialized to 0.85, and the shared weight vector W is [0.85, 0.85]. There are two branch networks in the intelligent blending model (i.e., Branch A and Branch B), where each branch network receives different input features. For example, the first branch network receives the flavor component content data of white liquor and base liquor, and the second branch network receives the near-infrared spectrum data of white liquor and base liquor; taking the first blending parameter and the second blending parameter as the prior parameters of the two branch networks in the intelligent blending model respectively, and then training the intelligent blending model through a preset training data set can be achieved in the following manner, that is: First, take the first blending parameter as the prior parameter of the first branch network and the second blending parameter as the prior parameter of the second branch network. It should be noted that the prior parameters in this application can be used as the initial constraint conditions of the network to guide the training direction of the model; then, take the flavor component content data (i.e., the flavor component content of white liquor and base liquor) in the preset training data set as the input of the first branch network in the intelligent blending model, take the near-infrared spectrum data (i.e., the near-infrared spectrum data of white liquor and base liquor) in the preset training data set as the input of the second branch network in the intelligent blending model, and then let the two branch networks learn and optimize their respective blending features respectively, and perform coupling through the shared layer, and use an appropriate loss function (such as mean square error) to optimize the shared parameters and minimize the prediction error. When further evaluating the model performance, use the validation set for testing to ensure the generalization ability of the model, so as to complete the training of the intelligent blending model based on the siamese neural network.

[0100] In specific implementation, predicting the blending ratio between the white liquor and the base liquor through the trained intelligent blending model can be achieved in the following manner, that is: in the intelligent blending model, the first branch network receives the flavor component content data corresponding to the first blending parameter, and the second branch network receives the near-infrared spectrum data corresponding to the second blending parameter. The intelligent blending model processes and analyzes the input features through its shared weight, calculates the output of each branch, and then shares and couples the outputs of the two branches to generate the blending ratio between the white liquor and the base liquor.

[0101] It should be noted that in the solution of this application, an intelligent blending model is constructed based on a siamese neural network, and the flavor components and near-infrared spectral data are comprehensively utilized to generate the first blending parameter and the second blending parameter respectively, so as to realize the fusion analysis of multi-dimensional information. Through projection clustering and characteristic wavelength screening, the model not only optimizes the understanding of chemical and physical characteristics, but also improves the prediction accuracy in the coupled training. Finally, this solution effectively improves the prediction accuracy of the blending ratio in the liquor blending process, ensures the consistency of the finished liquor in flavor and quality, and meets the market demand for high-quality liquor.

[0102] On the other hand, in some embodiments, the present application provides a liquor intelligent blending system based on machine learning. The liquor intelligent blending system based on machine learning includes a training unit for the liquor intelligent blending model. Refer to Figure 4 , which is a schematic structural diagram of the training unit for the liquor intelligent blending model in some embodiments of the present application. The training unit 400 for the liquor intelligent blending model specifically includes: a construction module 401, a processing module 402, and an execution module 403, which are described as follows:

[0103] The construction module 401 is mainly used in the present application to initialize and construct an intelligent blending model for liquor blending based on a siamese neural network;

[0104] The processing module 402 is used in the present application to collect the flavor component contents of liquor and base liquor, and perform an initial prediction on the blending ratio between the liquor and the base liquor according to the flavor component contents of the liquor and the base liquor, so as to obtain the first blending parameter of the intelligent blending model;

[0105] The processing module 402 is further used in the present application to collect the near-infrared spectral data of the liquor and the base liquor, perform projection clustering on the near-infrared spectral data to construct a spectral feature space, and then determine the baseline offset of each spectral characteristic wavelength according to the difference gradient between the spectral characteristic wavelengths in the spectral feature space;

[0106] The processing module 402 is further used in the present application to screen out the characteristic wavelengths of the liquor and the base liquor from the spectral feature space through the baseline offset of each spectral characteristic wavelength, and then perform a secondary prediction on the blending ratio between the liquor and the base liquor according to the screened characteristic wavelengths, so as to obtain the second blending parameter of the intelligent blending model;

[0107] The execution module 403 is mainly used in the present application to perform coupled training on the intelligent blending model based on the first blending parameter and the second blending parameter, and then predict the blending ratio between the liquor and the base liquor through the trained intelligent blending model.

[0108] In addition, the present application also provides a computer device, which includes a memory and a processor. The memory stores code, and the processor is configured to obtain the code and execute the above-mentioned method for training a Baijiu intelligent blending model based on machine learning.

[0109] In some embodiments, referring to Figure 5 , this figure is a schematic structural diagram of a computer device for implementing the method for training a Baijiu intelligent blending model based on machine learning in some embodiments of the present application. The method for training a Baijiu intelligent blending model based on machine learning in the above embodiments can be implemented by Figure 5 the computer device shown. The computer device 500 includes at least one processor 501, a communication bus 502, a memory 503, and at least one communication interface 504.

[0110] The processor 501 can be a general-purpose central processing unit (CPU) or an application-specific integrated circuit (ASIC).

[0111] The communication bus 502 can be used to transfer information between the above components.

[0112] The memory 503 can be a read-only memory (ROM) or other types of static storage devices that can store static information and instructions, a random access memory (RAM), or other types of dynamic storage devices that can store information and instructions. It can also be an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM), or other optical disc storage, optical disc storage (including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), magnetic disks, or other magnetic storage devices, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory 503 can exist independently and be connected to the processor 501 through the communication bus 502. The memory 503 can also be integrated with the processor 501.

[0113] Among them, the memory 503 is used to store the program code for executing the solution of this application, and is controlled by the processor 501 to execute. The processor 501 is used to execute the program code stored in the memory 503. The program code may include one or more software modules. The above-mentioned method for training the intelligent blending model of Baijiu based on machine learning in the embodiment can be implemented by one or more software modules in the processor 501 and the program code in the memory 503.

[0114] The communication interface 504, using any device such as a transceiver, is used to communicate with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0115] In a specific implementation, as an embodiment, the computer device may include multiple processors, and each of these processors may be a single-CPU processor or a multi-CPU processor. Here, the processor may refer to one or more devices, circuits, and / or processing cores for processing data (such as computer program instructions).

[0116] The above-mentioned computer device may be a general-purpose computer device or a special-purpose computer device. In a specific implementation, the computer device may be a desktop computer, a laptop computer, a network server, a personal digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. The embodiments of this application do not limit the type of the computer device.

[0117] In addition, this application also provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, it implements the above-mentioned method for training the intelligent blending model of Baijiu based on machine learning.

[0118] Although the preferred embodiments of this application have been described, those skilled in the art can make additional changes and modifications once they know the basic creative concept. Therefore, the appended claims are intended to be interpreted to include the preferred embodiments and all changes and modifications falling within the scope of this application.

[0119] Obviously, those skilled in the art can make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalent technologies, this application is also intended to include these changes and modifications.

Claims

1. A method for training a liquor intelligent blending model based on machine learning, which is used in a liquor intelligent blending system for liquor intelligent blending, characterized in that: The steps include: Construct an intelligent blending model for liquor blending based on twin neural network initialization; Collecting the flavor component contents of the liquor and the base wine, and making an initial prediction of the blending ratio between the liquor and the base wine according to the flavor component contents of the liquor and the base wine, to obtain a first blending parameter of the intelligent blending model; Collecting near-infrared spectral data of the liquor and base wine, performing projection clustering on the near-infrared spectral data to construct a spectral feature space, and then determining the baseline offset of each spectral feature wavelength according to the difference gradient between each spectral feature wavelength in the spectral feature space; The characteristic wavelengths of the liquor and the base wine are screened out from the spectral feature space by the baseline offset of each spectral characteristic wavelength, and then the blending ratio between the liquor and the base wine is secondary predicted based on the screened characteristic wavelengths to obtain the second blending parameter of the intelligent blending model; Based on the first blending parameter and the second blending parameter, coupling training is performed on the intelligent blending model, and then the blending ratio between the liquor and the base liquor is predicted by the trained intelligent blending model; The coupling training of the intelligent blending model based on the first blending parameter and the second blending parameter specifically includes: Determining a coupling correlation between the first blending parameter and the second blending parameter; Determining a sharing weight between two branch networks in the intelligent blending model according to the coupling correlation; The first blending parameter and the second blending parameter are respectively used as prior parameters of two branch networks in the intelligent blending model, and then the intelligent blending model is trained using a preset training data set; Wherein, determining the coupling correlation between the first blending parameter and the second blending parameter specifically includes: Obtaining the flavor component contents of liquor and base liquor, and then determining the content variance of the flavor component contents; Acquire near infrared spectral data of liquor and base liquor, and then determine the spectral variance of the near infrared spectral data; Determine the fluctuation correlation of the content determination of the compound component in the sample through the content variance and the spectral variance; Determining a parameter difference between the first blending parameter and the second blending parameter; The fluctuation correlation is compensated by the parameter difference, and the compensated fluctuation correlation is used as the coupling correlation between the first blending parameter and the second blending parameter.

2. The method according to claim 1, characterized in that The blending ratio between the liquor and the base wine is initially predicted according to the flavor component content of the liquor and the base wine, and the first blending parameter of the intelligent blending model is obtained, which specifically includes: Standardizing the flavor component contents of the liquor and base liquor to obtain standard flavor component content data; The initial blending ratio of the liquor and the base liquor is predicted according to the standard flavor component content data, and the initial blending ratio is used as the first blending parameter of the intelligent blending model.

3. The method according to claim 1, characterized in that The projecting and clustering of the near infrared spectral data to construct a spectral feature space specifically includes: Extracting the main features of the near-infrared spectral data using principal component analysis to obtain a spectral feature set; Performing band hierarchical clustering on the spectral features in the spectral feature set to obtain spectral feature clusters of multiple bands; All spectral feature clusters are projected and mapped into a unified feature space to obtain a spectral feature space.

4. The method according to claim 1, characterized in that Determining the baseline offset of each spectral feature wavelength according to the difference gradient between each spectral feature wavelength in the spectral feature space specifically includes: Determine the difference gradient between the wavelengths of each spectral feature in the spectral feature space using a difference method; Based on a normal distribution assumption, determining a baseline wavelength of the spectral feature in the spectral feature space according to all difference gradients; The baseline offset of the corresponding spectral characteristic wavelength is determined by the baseline wavelength and each spectral characteristic wavelength.

5. The method according to claim 1, characterized in that Screening out the characteristic wavelengths of the liquor and base wine from the spectral feature space by using the baseline offset of each spectral characteristic wavelength specifically includes: Performing offset threshold comparison on the baseline offset of each spectral feature wavelength; When the baseline offset is less than or equal to a preset offset threshold, the spectral characteristic wavelength corresponding to the baseline offset is taken as the effective spectral characteristic wavelength, and then all effective spectral characteristic wavelengths are obtained; All effective spectral characteristic wavelengths are used as characteristic wavelengths of the liquor and base wine.

6. The method according to claim 1, characterized in that The flavor component content refers to the specific concentration of various volatile compounds in the liquor and base wine.

7. A liquor intelligent blending system based on machine learning, which uses the method described in any one of claims 1 to 6 to train a liquor intelligent blending model, the system comprising a liquor intelligent blending model training unit, characterized in that: The liquor intelligent blending model training unit comprises: A construction module is used to construct an intelligent blending model for liquor blending based on the twin neural network initialization; A processing module, for collecting the flavor component contents of the liquor and the base wine, and making an initial prediction of the blending ratio between the liquor and the base wine according to the flavor component contents of the liquor and the base wine, so as to obtain a first blending parameter of the intelligent blending model; The processing module is further used to collect near-infrared spectral data of the liquor and base wine, perform projection clustering on the near-infrared spectral data to construct a spectral feature space, and then determine the baseline offset of each spectral feature wavelength according to the difference gradient between each spectral feature wavelength in the spectral feature space; The processing module is further used to filter out the characteristic wavelengths of the liquor and the base wine from the spectral feature space by using the baseline offset of each spectral characteristic wavelength, and then perform a secondary prediction on the blending ratio between the liquor and the base wine based on the filtered characteristic wavelengths to obtain a second blending parameter of the intelligent blending model; An execution module is used to perform coupling training on the intelligent blending model based on the first blending parameter and the second blending parameter, and then predict the blending ratio between the liquor and the base liquor through the trained intelligent blending model.

8. A computer device, comprising a memory and a processor, wherein the memory stores a code, characterized in that: The processor is configured to obtain the code and execute the white wine intelligent blending model training method based on machine learning as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by the processor, the machine learning-based intelligent liquor blending model training method as described in any one of claims 1 to 6 is implemented.

Images