Intelligent classification method and device for distilled liquor
By obtaining the multimodal characteristics of distilled liquor samples and establishing a high-dimensional abstract neural network model, the problem of low efficiency of existing distilled liquor classification methods is solved, and a more efficient and intelligent distilled liquor classification is achieved.
Patent Information
- Application Number
- CN202411733421.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-29
- Publication Date
- 2025-03-04
AI Technical Summary
The existing distilled liquor classification methods rely on experience, are inefficient and lack intelligent auxiliary means.
The intelligent classification method of distilled liquor is adopted to obtain the multimodal characteristics of distilled liquor samples, including sample photos and compound characteristics, and a high-dimensional abstract neural network model is established to obtain the color characteristics. After splicing the color characteristics with the compound characteristics, the intelligent classification model of distilled liquor is used to classify it.
It improves the efficiency of distilled liquor classification, reduces the dependence on traditional artificial experience, and makes the classification more intelligent and standardized.
Smart Images
Figure CN119557789A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of distilled liquor classification, and in particular to an intelligent distilled liquor classification method and device. Background Art
[0002] In the production of distilled spirits, the classification of distilled spirits is an important means of evaluating distilled spirits, and is also crucial to understanding their flavor, quality, and applicable occasions. However, the classification of distilled spirits is currently mainly carried out by experienced experts, which relies on the experience of experts and is inefficient. There is currently a lack of intelligent classification methods to assist experts in efficient distilled spirit classification. Summary of the invention
[0003] In order to improve the efficiency of distilled liquor classification while reducing reliance on expert experience, the present application provides a method and device for intelligent classification of distilled liquor.
[0004] The technical solution adopted by the present invention to solve the above problems is:
[0005] Intelligent classification method for distilled spirits, including:
[0006] Step 1: obtaining a distilled liquor sample database, wherein the distilled liquor sample database contains multimodal features of distilled liquor and corresponding category features, wherein the multimodal features include sample photos and compound features;
[0007] Step 2, establishing a high-dimensional abstract neural network model for obtaining color features based on sample photos;
[0008] Step 3, obtaining color features based on a high-dimensional abstract neural network model;
[0009] Step 4: reconstruct the color feature into a one-dimensional tensor by flattening, and concatenate the one-dimensional color feature with the compound feature;
[0010] Step 5: Using the concatenated data as input and the corresponding category features as output, an intelligent classification model for distilled liquor is established and trained;
[0011] Step 6. Obtain sample photos and compound features of the distilled liquor to be classified, and obtain the color features of the distilled liquor to be classified based on a high-dimensional abstract neural network model; reconstruct the color features into a one-dimensional tensor by flattening, and splice the one-dimensional color features with the compound features; input the spliced data into the trained distilled liquor intelligent classification model to obtain the distilled liquor classification.
[0012] Further, compound characteristics include esters, alcohols, acids, aromatic compounds, aldehydes, ketones, furans and pyrazines.
[0013] Furthermore, the compounds were characterized by mass spectrometry and flame ionization detection.
[0014] Furthermore, the category characteristics include at least one of: flavor, wine type, brand, region, year, characteristics and country.
[0015] Furthermore, the step 2 is specifically as follows: characterizing the sample photo as a three-dimensional photo tensor, using RGB encoding color in the channel dimension of the three-dimensional photo tensor, and processing the other two dimensions of the three-dimensional photo tensor to have the same resolution; using the processed three-dimensional photo tensor to train an image autoencoder neural network model, the image autoencoder neural network model includes an image encoder and an image decoder, and when the difference between the three-dimensional photo tensor generated by the image decoder and the input three-dimensional photo tensor meets the requirements, the image encoder of the image autoencoder neural network model at this time is used as a high-dimensional abstract neural network model.
[0016] Furthermore, the step 1 specifically includes: obtaining a dynamically updated distilled liquor sample database.
[0017] Furthermore, in step 1, if the updated part includes the multimodal features of distilled liquor and the corresponding category features, the distilled liquor intelligent classification model is trained based on the updated part;
[0018] If the updated part only contains the multimodal features of distilled liquor, the predicted probability of the classification result is obtained based on the high-dimensional abstract neural network model and the distilled liquor intelligent classification model. If the predicted probability of the classification result is greater than or equal to the preset probability threshold, the classification is used as the corresponding category feature. If the predicted probability of the classification result is less than the preset probability threshold, the classification is obtained based on the distilled liquor sample database using a clustering algorithm, and the classification is used as the corresponding category feature. The distilled liquor intelligent classification model is trained based on the updated part.
[0019] Intelligent classification device for distilled liquor, comprising:
[0020] A database module, used to store multimodal features of distilled liquor and their corresponding category features, wherein the multimodal features include sample photos and compound features;
[0021] Color feature acquisition module, which acquires the color features in sample photos based on a high-dimensional abstract neural network model;
[0022] A feature splicing module is used to reconstruct the color feature into a one-dimensional tensor by flattening, and to splice the one-dimensional color feature with the compound feature;
[0023] The intelligent classification module obtains the classification of distilled liquor based on the spliced data and the intelligent classification model of distilled liquor.
[0024] Furthermore, the database module is also provided with an updating unit for dynamically updating the data in the database.
[0025] Furthermore, a category feature acquisition unit is also provided in the database module, and the category feature acquisition unit acquires the classification of distilled liquor based on a clustering algorithm.
[0026] Compared with the prior art, the present invention has the following beneficial effects: processing photos of distilled liquor to obtain color characteristics, splicing the color characteristics with the characteristics of compounds that can be quantified in the distilled liquor as the input of an intelligent classification model for distilled liquor, and completing classification based on the intelligent classification model for distilled liquor. The entire process is more efficient and reduces excessive reliance on traditional manual experience, making classification more intelligent and standardized. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 It is a flow chart of the intelligent classification method for distilled liquor;
[0028] Figure 2 Build a schematic diagram for a high-dimensional abstract neural network model. DETAILED DESCRIPTION
[0029] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0030] like Figure 1 As shown, the intelligent classification method for distilled liquor includes:
[0031] Step 1: Obtain a distilled liquor sample database, wherein the distilled liquor sample database contains multimodal features of distilled liquor and corresponding category features, wherein the multimodal features include sample photos and compound features.
[0032] The compound characteristics must include esters, alcohols, acids, aromatic compounds, aldehydes, ketones, furans and pyrazines, and may also include other compounds that experts or companies believe may affect the flavor and taste of distilled spirits. The compound characteristics can be measured by compound detection techniques including mass spectrometry and hydrogen flame ionization detector.
[0033] In this embodiment, comprehensive two-dimensional gas chromatography is used for compound determination. From the corresponding peaks, a total of 23 peaks including one-dimensional times of 556, 628, 7040, 708, 748, 796, 800, 896, 1212, 1288, 1372, 1436, 1444, 1472, 1532, 1540, 1616, 1676, 1748, 1756, 2108, 2148, 2304, etc. are selected as compounds, and their peak areas are used as compound characteristics; the peak areas of the 23 peaks corresponding to the wine sample are [811604, 413948, 941288, 1191063, ...], which are standardized. In this embodiment, the standardization method used is:
[0034]
[0035] Among them, x min Take 708.329, x max The standardized compound characteristics of distilled liquor samples are [0.086, 0.044, 0.100, 0.127, …].
[0036] Category characteristics include: at least one of flavor, liquor, brand, region, year, characteristics and country. Among them, flavor may include: strong aroma, sauce aroma, rice aroma, light aroma, soy aroma, Laobaigan aroma, mixed aroma, Dong aroma, Xiaoqu aroma, special aroma, sesame aroma, Fengxiang, etc., and flavor may also include other flavors that experts or enterprises need to classify. Liquor may include: brandy, whiskey, vodka, rum, gin, tequila, Chinese liquor, etc., and liquor may also include other liquor that experts or enterprises need to classify. Brands may include: WuXX, MaoX, LuzhouXX, JianXX, LangX, XiX, YangX, YuXX, BaiXX, BaiXX, RoyalXX, XuanXX, MaXX, etc., and brands may also include other brands that experts or enterprises need to classify; regions may include: Sichuan Yibin, Sichuan Luzhou, Guizhou Renhuai, Shanxi Fenyang, French Cognac, etc., and regions may also include other regions that experts or enterprises need to classify. Years may include: 10 years, 20 years, 30 years, 50 years, etc., and years may also include other years that experts or companies need to classify. Characteristics may include: fragrant, sweet, thick, strong, refreshing, muddy, rotten, astringent, etc., and characteristics may also include other characteristics that experts or companies need to classify. Countries may include: China, France, Cuba, the Netherlands, the United Kingdom, Russia, Poland, Sweden, the United States, Canada, Japan, Mexico, etc., and countries may also include other countries that experts or companies need to classify. Other categories that experts or companies are concerned about may also be included, for example, the East-West category, including: East, West, etc.
[0037] In this embodiment, the distilled liquor category feature is an Eastern and Western category, including Eastern and Western categories, and is represented by a one-hot vector, that is, the distilled liquor category feature corresponding to the East is [1,0], and the distilled liquor category feature corresponding to the West is [0,1].
[0038] Step 2: Establish a high-dimensional abstract neural network model for obtaining color features based on sample photos. Figure 2 As shown, specifically: the sample photo is represented as a three-dimensional photo tensor, RGB encoding color is used in the channel dimension of the three-dimensional photo tensor, and the other two dimensions of the three-dimensional photo tensor are processed to have the same resolution; the processed three-dimensional photo tensor is used to train the image autoencoder neural network model, the image autoencoder neural network model includes an image encoder and an image decoder, when the difference between the three-dimensional photo tensor generated by the image decoder and the input three-dimensional photo tensor meets the requirements, the image encoder of the image autoencoder neural network model at this time is used as a high-dimensional abstract neural network model.
[0039] The image encoder and image decoder are constructed using convolutional neural networks, the image encoder is a downsampling convolutional neural network composed of a convolutional layer and a pooling layer, and the image decoder is an upsampling convolutional neural network composed of a deconvolutional layer or an upsampling layer. This embodiment uses a loss function to characterize the difference and performs model optimization through back propagation of the neural network.
[0040] Step 3: Obtain color features based on a high-dimensional abstract neural network model.
[0041] Step 4: Reconstruct the color feature into a one-dimensional tensor by flattening, and concatenate the one-dimensional color feature with the compound feature. If the key color feature obtained by the image encoder is a 3×3 three-dimensional tensor with 3 channels, then after flattening, it becomes a one-dimensional vector with a length of 27, which is the one-dimensional key color feature; the flattening method uses the tensor.view(-1) function of the PyTorch library; then the torch.cat function of the PyTorch library is used to concatenate the one-dimensional key color feature with the standardized key compound feature of the distilled liquor sample as the input feature of the distilled liquor intelligent classification model.
[0042] Step 5, using the concatenated data as input and the corresponding category features as output, establish and train a distilled liquor intelligent classification model; the distilled liquor intelligent classification model is constructed based on a neural network model that can be used for one-dimensional data classification, including MLP, RNN, and LSTM. During training, the data in the database is divided into a training set and a test set, and the training set is used to train the model. The model corresponding to the minimum loss function of the test set and the accuracy of the test set model reaching a preset threshold is selected as the trained distilled liquor intelligent classification model.
[0043] In this embodiment, the distilled liquor sample feature data set has a total of 20 distilled liquor sample key feature data, and a test set is taken at 20%, with 16 data as training sets and 4 data as test sets.
[0044] The loss function used in the training process of the distilled liquor intelligent classification method is the cross entropy loss function:
[0045]
[0046] Where loss is the loss function, N is the number of samples, C is the total number of distilled liquor categories, and y is the ic is a binary indicator (0 or 1), which is 1 if the true category of sample i is c and 0 otherwise; p ic is the probability that the model predicts that sample i belongs to category c.
[0047] The category with the highest classification probability of the distilled liquor sample is taken as the category of the distilled liquor sample. In this embodiment, since two categories, Eastern and Western, are considered, it can be considered that a sample belongs to either the Eastern category or the Western category, and the sum of the probabilities of being predicted as the Eastern category and the Western category is 1. Therefore, the category with the highest classification probability of the distilled liquor sample is the category with a classification probability of the distilled liquor sample greater than 0.5.
[0048] After training, the model of this embodiment can achieve 100% accuracy in the test set. Therefore, the model can effectively realize the intelligent classification of distilled liquor. The results of the test set predicted by the distilled liquor intelligent classification model are shown in the following table:
[0049] Table 1 Prediction results of distilled liquor intelligent classification model
[0050]
[0051] Step 6. Obtain sample photos and compound features of the distilled liquor to be classified, and obtain the color features of the distilled liquor to be classified based on a high-dimensional abstract neural network model; reconstruct the color features into a one-dimensional tensor by flattening, and splice the one-dimensional color features with the compound features; input the spliced data into the trained distilled liquor intelligent classification model to obtain the distilled liquor classification.
[0052] In order to improve the accuracy of the model, a dynamically updated database of distilled liquor samples can also be obtained to train the model. The updated data can be divided into two categories according to the content:
[0053] The first category is the updated part, which includes the multimodal features of distilled liquor and the corresponding category features. The high-dimensional abstract neural network model and the distilled liquor intelligent classification model are trained based on the updated part.
[0054] The second category is the updated part that only contains the multimodal features of distilled liquor. At this time, it is necessary to first obtain the category features corresponding to the multimodal features. The acquisition method is to obtain the predicted probability of the classification result based on the high-dimensional abstract neural network model and the distilled liquor intelligent classification model. If the predicted probability of the classification result is greater than or equal to the preset probability threshold, the classification is used as the corresponding category feature. If the predicted probability of the classification result is less than the preset probability threshold, the classification is obtained based on the distilled liquor sample database using a clustering algorithm, and the classification is used as the corresponding category feature. Then, the high-dimensional abstract neural network model and the distilled liquor intelligent classification model are trained based on the multimodal features of distilled liquor and the corresponding category features. The clustering algorithm can be K-means clustering (K-Means clustering), hierarchical clustering, density-based spatial clustering of noisy applications (DBSCAN), Gaussian mixture model (GMM), t-SNE (t-distributed random neighbor embedding), etc.
[0055] This embodiment also provides a distilled liquor intelligent classification device, including:
[0056] A database module, used to store multimodal features of distilled liquor and their corresponding category features, wherein the multimodal features include sample photos and compound features;
[0057] Color feature acquisition module, which acquires the color features in sample photos based on a high-dimensional abstract neural network model;
[0058] A feature splicing module is used to reconstruct the color feature into a one-dimensional tensor by flattening, and to splice the one-dimensional color feature with the compound feature;
[0059] The intelligent classification module obtains the classification of distilled liquor based on the spliced data and the intelligent classification model of distilled liquor.
[0060] Furthermore, the database module is also provided with an updating unit for dynamically updating the data in the database; and a category feature acquisition unit, wherein the category feature acquisition unit acquires the classification of distilled liquor based on a clustering algorithm.
Claims
1. An intelligent classification method for distilled liquor, characterized in that: include: Step 1: obtaining a distilled liquor sample database, wherein the distilled liquor sample database contains multimodal features of distilled liquor and corresponding category features, wherein the multimodal features include sample photos and compound features; Step 2, establishing a high-dimensional abstract neural network model for obtaining color features based on sample photos; Step 3, obtaining color features based on a high-dimensional abstract neural network model; Step 4: reconstruct the color feature into a one-dimensional tensor by flattening, and concatenate the one-dimensional color feature with the compound feature; Step 5: Using the concatenated data as input and the corresponding category features as output, an intelligent classification model for distilled liquor is established and trained; Step 6. Obtain sample photos and compound features of the distilled liquor to be classified, and obtain the color features of the distilled liquor to be classified based on a high-dimensional abstract neural network model; reconstruct the color features into a one-dimensional tensor by flattening, and splice the one-dimensional color features with the compound features; input the spliced data into the trained distilled liquor intelligent classification model to obtain the distilled liquor classification.
2. The intelligent classification method for distilled liquor according to claim 1, characterized in that: Compound features include esters, alcohols, acids, aromatic compounds, aldehydes, ketones, furans and pyrazines.
3. The intelligent classification method for distilled liquor according to claim 2, characterized in that: The compounds were characterized by mass spectrometry and flame ionization detection.
4. The intelligent classification method for distilled liquor according to claim 1, characterized in that: Category characteristics include: at least one of flavor, wine type, brand, region, year, characteristics and country.
5. The intelligent classification method for distilled liquor according to claim 1, characterized in that: The step 2 specifically includes: characterizing the sample photo as a three-dimensional photo tensor, using RGB coding color in the channel dimension of the three-dimensional photo tensor, and processing the other two dimensions of the three-dimensional photo tensor to have the same resolution; using the processed three-dimensional photo tensor to train an image autoencoder neural network model, the image autoencoder neural network model includes an image encoder and an image decoder, and when the difference between the three-dimensional photo tensor generated by the image decoder and the input three-dimensional photo tensor meets the requirements, the image encoder of the image autoencoder neural network model at this time is used as a high-dimensional abstract neural network model.
6. The intelligent classification method for distilled liquor according to claim 1, characterized in that: The step 1 specifically includes: obtaining a dynamically updated distilled liquor sample database.
7. The intelligent classification method for distilled liquor according to claim 6, characterized in that: In step 1, if the updated part includes the multimodal features of distilled liquor and the corresponding category features, the distilled liquor intelligent classification model is trained based on the updated part; If the updated part only contains the multimodal features of distilled liquor, the predicted probability of the classification result is obtained based on the high-dimensional abstract neural network model and the distilled liquor intelligent classification model. If the predicted probability of the classification result is greater than or equal to the preset probability threshold, the classification is used as the corresponding category feature. If the predicted probability of the classification result is less than the preset probability threshold, the classification is obtained based on the distilled liquor sample database using a clustering algorithm, and the classification is used as the corresponding category feature. The distilled liquor intelligent classification model is trained based on the updated part.
8. Intelligent classification device for distilled liquor, characterized in that: include: A database module, used to store multimodal features of distilled liquor and their corresponding category features, wherein the multimodal features include sample photos and compound features; Color feature acquisition module, which acquires the color features in sample photos based on a high-dimensional abstract neural network model; A feature splicing module is used to reconstruct the color feature into a one-dimensional tensor by flattening, and to splice the one-dimensional color feature with the compound feature; The intelligent classification module obtains the classification of distilled liquor based on the spliced data and the intelligent classification model of distilled liquor.
9. The intelligent classification device for distilled liquor according to claim 8, characterized in that: The database module is also provided with an updating unit for dynamically updating the data in the database.
10. The intelligent classification device for distilled liquor according to claim 9, characterized in that: The database module is also provided with a category feature acquisition unit, and the category feature acquisition unit acquires the classification of distilled liquor based on a clustering algorithm.