Artificial intelligence-based baijiu base liquor grading method and system

Abstract

The application discloses a liquor base liquor grading method and system based on artificial intelligence. The grading method comprises the following steps: collecting VCD spectrum images of base liquor samples as a feature set; classifying and grading the collected base liquor samples according to artificial product evaluation, corresponding the obtained grading categories to the feature set to form a label data set; and training a feedforward neural network model through the label data set to obtain a grading model. The application can accurately grade base liquor of complex liquor with a complex flavor type such as a rich and mellow flavor type under the condition of low cost and fewer samples, and significantly improves the base liquor grading efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to the technical field of a feedforward neural network grading method. BACKGROUND

[0002] The preparation of Baijiu (liquor) generally includes material selection, koji making, fermentation, distillation, aging, blending, and filling, wherein the liquor just distilled is base liquor, which has a high degree of alcohol and can cause certain harm to the human body if directly consumed, and thus needs to be blended in a certain proportion. Different types of base liquor can be blended to obtain drinking liquor with different qualities, tastes, and styles, and thus the grading, storage, and combination of base liquor of the same grade are of great importance and can lay a solid foundation for the next step of Baijiu blending.

[0003] At present, the traditional base liquor classification or grading method mainly includes sensory evaluation and instrumental analysis. The sensory evaluation mainly relies on artificial liquor tasters, consumes a large amount of manpower and time cost, and is prone to human error. The instrumental analysis includes chromatography, chromatography-mass spectrometry, spectroscopy, electronic tongue and electronic nose, and array sensor, etc.

[0004] Some existing technologies further combine instrumental analysis with computational algorithms to conduct in-depth research on the flavor indicators of base liquor. For example, in the research of "Analysis of the Flavor Quality of Maotai Flavor Baijiu in the Large Return Process Based on Multivariate Chromatography and Principal Component Analysis", multivariate chromatography was used to analyze the changes in the content of flavor substances in seven batches of base liquor stored for 1-3 years, including esters, alcohols, aldehydes, acids, ketones, and pyrazines, and principal component analysis was used to obtain that lactic acid and ethyl lactate are positively correlated with the sensory quality of base liquor in the large return process of Maotai flavor Baijiu. However, this method has some obvious defects, such as being mainly applicable to one type of base liquor and being difficult to apply to other types, especially Baijiu with complex flavors, and being mainly aimed at the content of flavor substances, while the actual Baijiu grade division is based on more complex comprehensive indicators, resulting in difficulty in achieving ideal accuracy in grading.

[0005] On the other hand, Fuyu flavor Baijiu has unique taste characteristics of "clear color and transparency, complex flavors, soft and sweet entry, mellow and full-bodied, harmonious aroma, and long-lasting aftertaste", and "front thick, middle clear, and back sauce", and its flavor is complex and there is less relevant research in the early stage, and it is also difficult to achieve accurate judgment by artificial evaluation. SUMMARY

[0006] In view of the defects of the prior art, the purpose of the present application is to provide a grading method and system that can accurately grade base liquor of complex flavors such as Fuyu flavor Baijiu at low cost and with fewer samples, which can quickly, simply, and accurately judge the grade of base liquor and improve the efficiency and quality of Baijiu preparation.

[0007] The technical scheme of the present application is as follows:

[0008] The liquor base grading method based on artificial intelligence comprises the following steps:

[0009] S1, collecting a vibration circular dichroism spectrum image (VCD image) of a base liquor sample, and forming a feature set from characteristic peaks and / or fingerprint peaks of the vibration circular dichroism spectrum image;

[0010] S2, classifying and grading the collected base liquor sample according to artificial product evaluation, and corresponding the obtained grading level to the feature set of the base liquor sample from which it comes, to form a label data set;

[0011] S3, performing a feedforward neural network classification training through the label data set, and the classification model obtained after the training or further testing is a liquor base grading model, through which the grading of liquor base is realized.

[0012] According to some preferred embodiments of the present application, the feature set further comprises production information of the collected base liquor sample, and the production information comprises a base liquor fermentation category and a base liquor production time, and the base liquor fermentation category comprises bottom dregs, middle dregs or cap dregs.

[0013] According to some preferred embodiments of the present application, the characteristic peaks and / or fingerprint peaks comprise characteristic peaks and / or fingerprint peaks of one or more of the following flavor substances: acetaldehyde, ethyl acetate, n-propanol, sec-butyl alcohol, acetal, isobutyl alcohol, n-butyl alcohol, ethyl butyrate, isoamyl alcohol, ethyl valerate, ethyl lactate, ethyl hexanoate and ethyl oleate.

[0014] According to some preferred embodiments of the present application, the classification and grading according to artificial product evaluation comprises: according to the fermentation category, the base liquor is classified into five large categories of cap dregs fermentation, one-time middle dregs fermentation, two-time middle dregs fermentation, one-time bottom dregs fermentation and two-time bottom dregs fermentation, and thereafter, according to the base liquor quality, each large category is further classified into the first to fifth grades, and the base liquor quality decreases in turn from the first grade to the fifth grade.

[0015] According to some preferred embodiments of the present application, the feedforward neural network comprises: an input layer for inputting feature set data, an output layer for outputting base liquor grade categories, and a hidden layer; according to a weight setting, the hidden layer forms: a first classifier and a second classifier connected with the input layer, a first sub-classifier and a second sub-classifier connected with the first classifier, wherein the first classifier is a three-classifier, the second classifier is a five-classifier, the first sub-classifier is a two-classifier, and the second sub-classifier is a three-classifier.

[0016] According to some preferred embodiments of the present application, the feedforward neural network comprises, sequentially connected, a first convolutional layer, a first max-pooling layer, a second convolutional layer, a second max-pooling layer, a first Dropout layer, a Flatten layer, a first and a second fully connected layer, a second Dropout layer, and a third fully connected layer,

[0017] According to some preferred embodiments of the present application, the first convolutional layer contains 64 convolutional kernels with a size of 10, the scanning step length is 1, the pool size of the first max-pooling layer is 10, the second convolutional layer is set the same as the first convolutional layer, the second max-pooling layer is set the same as the first max-pooling layer, the Dropout rate of the first and second Dropout layers is 0.25, the number of neurons of the first and second fully connected layers is 128 and 64 respectively, the activation function is relu, and the number of neurons of the third fully connected layer is 5, and the activation function is softmax.

[0018] According to some preferred embodiments of the present application, the Adam optimizer is used in the classification training of the feedforward neural network, the Batch size is 64, and the learning rate is 0.00001.

[0019] According to the above grading method, an artificial intelligence-based base liquor grading system can be further obtained, which comprises a storage medium storing programs and / or models and / or structural data for implementing the above grading method.

[0020] The VCD (Vibrational Circular Dichroism) spectrum image is used in the grading method of the present application, which can measure the absorption difference of chiral substances at a vibration frequency range for left and right circularly polarized light, and the determination of chiral enantiomers can determine the geographical origin of the product matrix and other information, which can be used for quality control and detection of sample adulteration. The brewing of pure grain liquor belongs to natural fermentation, and there are also chiral molecules, especially two enantiomers of ethyl lactate (D-type and L-type), and three optical isomers of 2,3-butanediol, namely S-(+)-2,3-butanediol, R-(-)-2,3-butanediol, and meso-(R,S)-butanediol (meso), which can be tested by VCD spectrum.

[0021] The grading method of the present application constructs the VCD spectrum image of the base liquor sample and the base liquor production information, and then combines the feedforward neural network to realize the precise grading of the base liquor through the method of machine learning. Through the precise grading of the base liquor, the reasonable regulation of the base liquor can be realized, such as combining base liquors of different seasons, different fermentation periods and different factories, increasing the types and contents of different trace components between base liquors, and improving the product quality and taste.

[0022] The application can intelligently grade base liquor into storage, store and merge the same grade, and make a solid foundation for the next liquor blending, which can save 60%-80% of the workload of manual evaluation, effectively improve the efficiency of base liquor classification, and improve the scientificity and accuracy of base liquor grading.

[0023] The selected feedforward neural network classification model has high performance, does not need to do feature selection when processing high-dimensional data, has fast training speed and high efficiency, can detect the influence between features, has the effect of error balance on unbalanced feature data, and can still maintain high accuracy in the case of feature loss.

[0024] The feedforward neural network grading model of the application can achieve 99.5% accuracy in the base liquor classification task, which is much higher than other non-artificial classification methods in the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 It is a specific grading process schematic diagram of the grading system of the application.

[0026] Figure 2 It is a specific grading model structure schematic diagram of the application. DETAILED DESCRIPTION

[0027] The application will be described in detail below in conjunction with the embodiments and drawings, but it should be understood that the embodiments and drawings are only used to exemplarily describe the application, and cannot constitute any limitation on the protection scope of the application. All reasonable modifications and combinations within the scope of the inventive concept of the application fall within the protection scope of the application.

[0028] According to the technical scheme of the application, a specific base liquor grading method of liquor based on artificial intelligence comprises:

[0029] S1 collects VCD images of base liquor samples, and the feature set is composed of characteristic peaks and / or fingerprint peaks of the images;

[0030] Further, the feature set can further include generation information of the base liquor sample, such as base liquor fermentation category and base liquor production time, and the base liquor fermentation category includes bottom dregs, middle dregs or cover dregs.

[0031] The characteristic peaks and / or fingerprint peaks are preferably characteristic peaks and / or fingerprint peaks of some flavoring substances, such as one or more of acetaldehyde, ethyl acetate, n-propanol, sec-butyl alcohol, ethyl acetal, isobutyl alcohol, n-butyl alcohol, ethyl butyrate, isoamyl alcohol, ethyl valerate, ethyl lactate, ethyl hexanoate and ethyl oleate.

[0032] S2 classifies and grades the collected base wine samples according to human evaluation, and matches the obtained classification categories and grades with the feature sets of the base wine samples from which they come to form a labeled dataset. In subsequent model training, this dataset can be further divided into training set and test set to train and test the model respectively.

[0033] Because the types, quantities, and proportions of microorganisms in the kiln mud are unevenly distributed in the upper, middle, and lower layers of the kiln, the fermentation status and characteristics of the mash at different levels within a single kiln vary. Therefore, the base liquor can be divided into five major categories: top mash fermentation, primary middle mash fermentation, secondary middle mash fermentation, primary bottom mash fermentation, and secondary bottom mash fermentation. Furthermore, based on the corresponding quality of the base liquor, such as the content of alcohol, total esters, total acids, total aldehydes, fusel oils, and various trace components like acids, esters, alcohols, and aldehydes in the mash, each fermentation category is further divided into grades one through five.

[0034] Specifically, the base liquor is divided into five types: G, Z1, Z2, D1, and D2, corresponding to fermentation in the top layer, primary fermentation in the middle layer, secondary fermentation in the middle layer, primary fermentation in the bottom layer, and secondary fermentation in the bottom layer, respectively. The G, Z1, Z2, D1, and D2 types are further divided into five grades: A1, A2, A2-, A3+, and A3.

[0035] Each level corresponds to:

[0036] A1: Overall quality is good;

[0037] A2: Cellar aroma with a hint of lees, relatively mellow, sweet, and clean;

[0038] A2-: Slightly sour and muddy smell, with a fermented aroma, relatively mellow and sweet, and relatively clean;

[0039] A3+: Cellar aroma, relatively mellow and sweet, slightly sour, with a slightly astringent aftertaste.

[0040] A3: Fragrant, musty, astringent, bran-like, lees-like, fragrant and musty, mixed flavors

[0041] In the above classification, A2 and A2- base wines lack certain aroma compounds, while A3+ and A3 base wines contain foreign substances.

[0042] S3 trains a grading model using a labeled dataset. The resulting model, after training or further testing, is the grading model for baijiu base liquor. This model enables the grading of baijiu base liquor.

[0043] Based on the above classification method, referring to Figure 1A specific hierarchical system includes an input layer for inputting feature set data, an output layer for outputting base liquor grade categories, and a hidden layer; according to a weight setting, the hidden layer forms a first classifier (illustrated classifier 1) and a second classifier (illustrated classifier 2) connected to the input layer, a first sub-classifier (illustrated classifier 1) and a second sub-classifier (illustrated classifier 2) connected to the first classifier, wherein the first classifier is a three-classifier, the second classifier is a five-classifier, the first sub-classifier is a two-classifier, and the second sub-classifier is a three-classifier.

[0044] In the above system, abnormal data in the input model can be directly classified into five levels by the second classifier; non-abnormal data can be divided into three categories of A1 level, A2 and A3+ mixed level, and A2- and A3 mixed level by the first feedforward neural classifier; then the A2 and A3+ mixed level is divided into A2 level and A3+ level by the first sub-classifier, and the A2- and A3 mixed level is divided into A2- level and A3 level by the second sub-classifier, to realize the differentiation and identification of A, A2, A2-, A3, and A3+ five levels.

[0045] In some specific embodiments, referring to the accompanying drawings Figure 2 The feedforward neural network used to build the hierarchical model includes, in sequence, a first convolutional layer, a first max-pooling layer, a second convolutional layer, a second max-pooling layer, a first Dropout layer, a Flatten layer, a first and a second fully connected layer, a second Dropout layer, and a third fully connected layer,

[0046] Preferably, the first convolutional layer contains 64 convolutional kernels with a size of 10 and a scanning step of 1, the pool size of the first max-pooling layer is 10, the second convolutional layer is set the same as the first convolutional layer, the second max-pooling layer is set the same as the first max-pooling layer, the Dropout rate of the first and second Dropout layers is 0.25, the number of neurons of the first and second fully connected layers is 128 and 64 respectively, the activation function is relu, the number of neurons of the third fully connected layer is 5, and the activation function is softmax. The first convolutional layer serves as the input layer, and the third fully connected layer serves as the output layer.

[0047] In some specific embodiments, the Adam optimizer is used in the classification training of the feedforward neural network, the Batch size is 64, and the learning rate is 0.00001.

[0048] In specific embodiments, the above system can be based on the Flask framework of Python to build a front-end API interface with HTML and CSS, and deployed on a web server.

[0049] And on the server by uploading data files in compliance with the format requirements such as csv, the prediction classification results of data are generated. The obtained prediction results can include base liquor number, production time, base liquor category, grading results, and can be used for generation and download.

[0050] The above embodiments are only preferred embodiments of the present application, and the protection scope of the present application is not limited to the above embodiments. Any technical solutions falling within the concept of the present application shall fall within the protection scope of the present application. It should be pointed out that, for ordinary skilled persons in the art, improvements and refinements without departing from the principles of the present application shall also be considered as falling within the protection scope of the present application.

Claims

1. A method for grading baijiu base liquor based on artificial intelligence, characterized in that, It includes: S1 Collect vibrational circular dichroism spectral images of base wine samples, and form a feature set from the characteristic peaks and / or fingerprint peaks of the vibrational circular dichroism spectral images; S2 classifies and grades the collected base wine samples according to human evaluation, and matches the obtained grading levels with the feature sets of the base wine samples from which they came, forming a label dataset; S3 uses the labeled dataset to perform feedforward neural network classification training. The resulting classification model after training or further testing is the grading model for baijiu base liquor. The grading of baijiu base liquor is achieved through this model. The feedforward neural network includes: an input layer for inputting feature set data, an output layer for outputting base wine grade categories, and a hidden layer; according to the weight settings, the hidden layer forms: a first classifier and a second classifier connected to the input layer, and a first sub-classifier and a second sub-classifier connected to the first classifier, wherein the first classifier is a three-classifier, the second classifier is a five-classifier, the first sub-classifier is a two-classifier, and the second sub-classifier is a three-classifier; The feedforward neural network comprises: a first convolutional layer, a first max-pooling layer, a second convolutional layer, a second max-pooling layer, a first dropout layer, a flattened layer, a first and second fully connected layers, a second dropout layer, and a third fully connected layer, all connected in sequence. The first convolutional layer contains 64 convolutional kernels of size 10 and a stride of 1. The pool size of the first max-pooling layer is 10. The second convolutional layer and the second max-pooling layer have the same settings as the first convolutional layer and the same settings as the first max-pooling layer. The dropout rate of the first and second dropout layers is 0.

25. The number of neurons in the first and second fully connected layers are 128 and 64, respectively, and the activation function is ReLU. The number of neurons in the third fully connected layer is 5, and the activation function is softmax.

2. The method for grading baijiu base liquor according to claim 1, characterized in that, The feature set also includes the production information of the collected base liquor samples, which includes the fermentation category and production time of the base liquor. The fermentation category of the base liquor includes bottom fermentation, middle fermentation, or top fermentation.

3. The method for grading baijiu base liquor according to claim 1, characterized in that, The characteristic peaks and / or fingerprint peaks include one or more of the following flavor compounds: acetaldehyde, ethyl acetate, n-propanol, sec-butanol, acetal, isobutanol, n-butanol, ethyl butyrate, isoamyl alcohol, ethyl valerate, ethyl lactate, ethyl hexanoate, and ethyl oleate.

4. The method for grading base liquor of baijiu according to claim 1, characterized in that, The classification and grading based on human evaluation includes: according to the fermentation type, the base liquor is divided into five major categories: top fermentation, primary middle fermentation, secondary middle fermentation, primary bottom fermentation, and secondary bottom fermentation. According to the quality of the base liquor, each major category is further divided into five grades, from the first grade to the fifth grade, with the quality of the base liquor decreasing sequentially from the first grade to the fifth grade.

5. The method for grading base liquor of baijiu according to claim 1, characterized in that, The Adam optimizer was used for classification training of the feedforward neural network, with a batch size of 64 and a learning rate of 0.00001.

6. An artificial intelligence-based grading system for baijiu base liquor, comprising a storage medium storing program and / or model and / or structural data for implementing the grading method according to any one of claims 1-5.

Images