White spirit production method and system based on deep learning

Abstract

The invention belongs to the technical field of white spirit production, and discloses a white spirit production method and system based on deep learning. Comprising the following steps: acquiring various planned physicochemical indexes of an upper-layer fermentation product, a middle-layer fermentation product and a lower-layer fermentation product in a target fermentation tank as original characteristics; respectively obtaining interlayer difference between any two layers of each original feature, a mean value and a standard deviation between three layers of each original feature, and interaction values of different planned physicochemical indexes in the same layer, so as to obtain a plurality of corresponding enhanced features; inputting the original features and the enhanced features into a first prediction model to obtain a predicted top temperature, and inputting the original features, the enhanced features and the predicted top temperature into a second prediction model to obtain a predicted alcoholic strength; adjusting the planned physicochemical index based on the predicted top temperature and the predicted alcohol content to obtain an iterative agency index, and repeating the steps until the optimal value is reached; and performing formula adjustment on the fermentation product in the target fermentation tank, and putting into production. The monitoring efficiency is high, and the quality of the white spirit can be effectively optimized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of liquor production, and in particular to a liquor production method and system based on deep learning. BACKGROUND

[0002] The production process of liquor includes, in sequence, material selection, koji making, fermentation, distillation, aging, blending and filling. Among them, fermentation is a key step in the production process of liquor. Specifically, under suitable temperature, the sugar in the fermentation material in the fermentation tank reacts with the enzyme to generate alcohol.

[0003] In the actual fermentation process, the fermentation material in the fermentation tank needs to be periodically sampled to monitor the fermentation process and determine the fermentation quality. However, this kind of monitoring method based on manual sampling has the following obvious defects. On the one hand, the monitoring process is tedious and cannot match the current modern and automated liquor production process, thereby affecting the entire production rhythm and reducing the production efficiency.

[0004] On the other hand, since the entire fermentation process needs to be kept in a sealed state, the sampling process will cause oxygen to enter the fermentation tank, resulting in excessive production of acid and affecting the quality of the final liquor. More than that, since the ingredients cannot be adjusted during the fermentation process to avoid affecting the stability of the entire fermentation process, even if the actual data of manual sampling reflects defects in the fermentation process, it can only be based on the optimization of the next fermentation, but cannot adjust the current fermentation process; thereby causing a significant lag between the fermentation strategy adjustment and this periodic sampling. SUMMARY

[0005] The present application aims to provide a liquor production method and system based on deep learning to solve the technical problems of low efficiency and lagging relationship between fermentation strategy adjustment caused by manual sampling in the prior art.

[0006] To achieve the above-mentioned purpose, the present application proposes the following technical solutions:

[0007] In a first aspect, the present application provides a liquor production method based on deep learning, comprising:

[0008] Obtaining each planned physicochemical index of the upper fermentation material, the middle fermentation material and the lower fermentation material in the target fermentation tank as each original feature;

[0009] Based on the fermentation material level division, obtaining the interlayer difference between any two layers of each original feature, the mean and standard deviation between three layers, and the interaction value of different planned physicochemical indexes in the same layer, and then obtaining a plurality of corresponding enhanced features;

[0010] The first prediction model is based on a neural network, and a double-path convolution kernel directional decomposition is used, and an attention mechanism is introduced to extract information within each feature.

[0011] The planned physicochemical indexes are adjusted based on the predicted top temperature and the predicted alcohol content to obtain iterative physicochemical indexes.

[0012] The above steps are repeated in sequence until the predicted top temperature and the predicted alcohol content reach optimal values, and the corresponding iterative physicochemical indexes are used as final physicochemical indexes.

[0013] Further, the original features include the water punching time, the amount of powder, the amount of rice husk, the amount of fermented grains, the moisture, the acidity, the starch amount, the caproic acid amount, the saccharifying power of Daqu, the fermenting power of Daqu, the daily maximum temperature, the daily minimum temperature, the pool temperature, and the temperature change rate within a preset time period of the upper layer, the middle layer, and the lower layer of the fermented grains.

[0014] Further, the first prediction model is based on a neural network, and a double-path convolution kernel directional decomposition is used, and an attention mechanism is introduced to extract information within each feature.

[0015] In the first prediction model and the second prediction model, the gradient difference between the same original features or enhanced features between different fermented grain levels is calculated based on a longitudinal convolution path to obtain the correlation between different layers of the same feature.

[0016] The matrix product between a preset number of original features and / or enhanced features and a transverse convolution kernel within the same fermented grain level is calculated based on a transverse convolution path to obtain the correlation between different features in the same layer.

[0017] The feature region weight of each original feature and enhanced feature of each layer is calculated based on the attention mechanism to highlight the core region.

[0018] Further, the first prediction model is based on a neural network, and a double-path convolution kernel directional decomposition is used, and an attention mechanism is introduced to extract information within each feature.

[0019] judging that a preset update cycle is reached, obtaining each verification original feature corresponding to an actual production and a verification actual top temperature in a historical database, and obtaining each verification enhanced feature based on each verification original feature;

[0020] inputting each verification original feature and verification enhanced feature into the first prediction model to obtain a verification predicted top temperature, and inputting each verification original feature and verification enhanced feature and the verification actual top temperature into the second prediction model to obtain a verification predicted alcohol content;

[0021] when an error between the verification predicted top temperature and the verification actual top temperature is greater than a preset top temperature error threshold, re-calling the historical database to train and optimize the first prediction model; and when an error between the verification predicted alcohol content and the verification actual alcohol content is greater than a preset alcohol content error threshold, re-calling the historical database to train and optimize the second prediction model.

[0022] In a second aspect, the technical solution 1 provides a liquor production system based on deep learning, comprising:

[0023] a first obtaining module, configured to obtain each planned physicochemical index of upper, middle and lower fermentation materials in a target fermentation tank as each original feature;

[0024] a second obtaining module, configured to obtain, based on fermentation material level division, a difference between any two layers of each original feature, a mean and a standard deviation among three layers, and an interaction value of different planned physicochemical indexes in the same layer, and then obtain a plurality of corresponding enhanced features;

[0025] a prediction module, configured to input all original features and all enhanced features into a first prediction model to obtain a predicted top temperature, and then input all original features, all enhanced features and the predicted top temperature into a second prediction model to obtain a predicted alcohol content; wherein the first prediction model and the second prediction model are both built based on a neural network, adopt double-path convolution kernel directional decomposition, and introduce an attention mechanism to extract information in each feature;

[0026] an index adjusting module, configured to adjust each planned physicochemical index based on the predicted top temperature and the predicted alcohol content to obtain each iterative physicochemical index;

[0027] an optimized production module, configured to repeatedly execute the above steps until the predicted top temperature and the predicted alcohol content both reach optimal values, and use each iterative physicochemical index as each final physicochemical index; and adjust a formula according to the final physicochemical index and put the fermentation materials into production.

[0028] Further, the prediction module comprises:

[0029] a judging unit configured to judge a preset update cycle, acquire each verification original feature corresponding to actual production and a verification actual top temperature in a historical database, and acquire each verification enhanced feature based on each verification original feature;

[0030] a verification unit configured to input each verification original feature and verification enhanced feature into the first prediction model to obtain a verification predicted top temperature, and input each verification original feature, verification enhanced feature and verification actual top temperature into the second prediction model to obtain a verification predicted alcohol content;

[0031] an optimization unit configured to, when an error between the verification predicted top temperature and the verification actual top temperature is greater than a preset top temperature error threshold, re-call the historical database to train and optimize the first prediction model; and when an error between the verification predicted alcohol content and the verification actual alcohol content is greater than a preset alcohol content error threshold, re-call the historical database to train and optimize the second prediction model.

[0032] Further, the prediction module comprises:

[0033] a longitudinal convolution unit configured to, in the first prediction model and the second prediction model, calculate gradient differences between the same original feature or enhanced feature between different fermentation material levels based on a longitudinal convolution path to acquire the correlation between different layers of the same feature;

[0034] a transverse convolution unit configured to, based on a transverse convolution path, calculate a matrix product between a preset number of original features and / or enhanced features sequentially adjacent in the same fermentation material level and a transverse convolution kernel to acquire the correlation between different features in the same layer;

[0035] an attention mechanism unit configured to, based on the attention mechanism, perform feature region weight calculation on each original feature and enhanced feature of each layer to highlight a core region.

[0036] In a third aspect, the technical solution provides an electronic device, comprising at least one processor, the processor being coupled with a memory, the memory storing a computer program, the computer program being configured to be executed by the processor to perform the method.

[0037] In a fourth aspect, the technical solution provides a computer readable storage medium, which stores a computer program, the computer program being used to be executed by a processor to implement the method.

[0038] Advantages:

[0039] According to the technical scheme, the liquor production method based on deep learning is provided to solve the problems of low efficiency, influence on the quality of liquor and inability to adjust the current fermentation process in the artificial sampling and detection process.

[0040] The technical scheme introduces a prediction model based on deep learning into the production process of liquor. Considering that the pits are divided into three layers in actual production, the data of only one layer cannot accurately describe the whole pit condition and will affect the prediction accuracy. The planned physicochemical indexes of the upper layer, middle layer and lower layer of the target fermentation pit are obtained as the original features. Considering the data island characteristics and limitation caused by the secrecy of production data of each manufacturer in liquor production, the technical scheme obtains enhanced features by performing interlayer difference processing, statistical processing and interaction processing based on the original features. The technical scheme improves the data volume and is also beneficial to deeper feature association, thereby realizing accurate prediction under limited data. In the specific prediction process, considering that in the solid-state fermentation process of liquor, if the top temperature is too high or too low, it will affect the final alcohol content, and is closely related to the physicochemical indexes when entering the pit. Therefore, the top temperature is first predicted by the first prediction model, and the alcohol content is predicted by the second prediction model combined with the features and the prediction result of the first prediction model. In the specific prediction process, the double-path convolution kernel directional decomposition is adopted, and the attention mechanism is introduced to extract information in each feature, which reduces the model calculation parameter amount, improves the convergence speed, and also effectively improves the prediction accuracy. At this time, the above process is iterated until the predicted top temperature and the predicted alcohol content reach the optimal values, and the corresponding iteration physicochemical index is the best production process parameter. Then, it is put into production, and any detection in the fermentation process is not needed, and the fermentation quality can be ensured.

[0041] It should be understood that all combinations of the foregoing concepts and additional concepts described in greater detail below can be seen as being part of the inventive subject matter of the present disclosure provided such concepts are not mutually inconsistent.

[0042] The foregoing and other aspects, embodiments and features of the present teachings can be better understood from the following description of the present teachings taken in conjunction with the accompanying drawings. Other features of the present teachings, such as exemplary embodiments thereof, will be apparent from the following description of the present teachings and associated drawings. BRIEF DESCRIPTION OF DRAWINGS

[0043] The accompanying drawings are not intended to be drawn to scale. In the drawings, like or similar components throughout the figures can be denoted with the same reference numerals. In the interest of clarity, not all components of the embodiments are shown in each figure. Now, embodiments of various aspects of the present application will be described by example with reference to the accompanying drawings, in which:

[0044] Figure 1 Flow chart of the deep learning-based liquor production method described in the present embodiment;

[0045] Figure 2 Flow chart of the original feature and enhanced feature information extraction performed in the present embodiment;

[0046] Figure 3 Flow chart of the periodic iteration optimization of the model;

[0047] Figure 4 Structural block diagram of the deep learning-based liquor production system described in the present embodiment;

[0048] Figure 5 Structural block diagram of the electronic device described in the present embodiment. DETAILED DESCRIPTION

[0049] To make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions of the embodiments of the present application will be described clearly and completely below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the described embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without any creative effort fall within the scope of protection of the present application. Unless otherwise defined, the technical terms or scientific terms used herein should have the usual meanings understood by those of ordinary skill in the art.

[0050] The terms "first", "second", and similar terms used herein do not denote any order, quantity, or importance, but are used to distinguish different components. Similarly, the singular forms "a", "an", and "the" do not denote the quantity limitation, but denote the existence of at least one, unless the context clearly indicates otherwise. The terms "comprise", "include", and similar terms mean that the components or objects appearing before the terms "comprise" or "include" cover the features, integers, steps, operations, elements, and / or assemblies listed after the terms "comprise" or "include", and do not exclude the existence or addition of one or more other features, integers, steps, operations, elements, assemblies, and / or sets thereof. "Up", "down", "left", "right", and the like are used to represent relative positional relationships, which may change accordingly when the absolute positions of the described objects change.

[0051] In the actual fermentation process, the fermentation in the fermentation tank needs to be regularly sampled to monitor the fermentation process and determine the fermentation quality. However, such a monitoring method based on manual sampling is not only cumbersome and inefficient, but also introduces oxygen during the opening and sampling process, which affects the quality of the liquor. Even more so, even if the actual data of manual sampling reflects defects in the fermentation process, it can only be based on the optimization of the next fermentation, but cannot adjust the current fermentation process; further leading to a significant lag between fermentation strategy adjustment and such regular sampling. Based on this, the present embodiment aims to provide a liquor production method based on deep learning to solve the above technical defects.

[0052] The liquor production method based on deep learning described in the present embodiment will be described in detail below with reference to the accompanying drawings.

[0053] In combination with Figure 1 As shown in the figure, the method comprises the following steps:

[0054] Step S202, obtaining each planned physicochemical index of the upper layer fermentation, the middle layer fermentation and the lower layer fermentation in the target fermentation tank as each original feature.

[0055] Considering that the parameters of different depths of the pit in actual production are different, therefore, the present embodiment divides the pit into three layers of upper, middle and lower, and correspondingly obtains each planned physicochemical index of the upper layer fermentation, the middle layer fermentation and the lower layer fermentation in the target fermentation tank as each original feature, thereby accurately describing the whole pit condition and improving the accuracy of subsequent prediction.

[0056] The liquor described in the embodiment mainly refers to Luzhou-flavor liquor. Considering that the main factors affecting the fermentation condition of the pit-entered fermented grains of Luzhou-flavor liquor include pit-entered temperature, moisture, acidity, starch, dosage of Daqu, dosage of bran, and fermented grains, and the factors interact with each other. For example, starch in the pit is first converted into glucose under the action of amylase, and then converted into alcohol or other trace components under the action of alcohol-forming enzyme. When the amount of starch input is less than the saccharification and alcohol-forming capacity, that is, less than 17%, increasing the starch content will increase the heat. The high-in and high-out mode is adopted for the starch of Wuliangye, and the amount of high-in starch is always greater than the saccharification and alcohol-forming capacity, so increasing the concentration of pit-entered starch will not increase the heat. When the concentration is low, increasing the starch can increase the heat, and when the starch concentration increases by a certain amount, further increasing the starch will not increase the heat. For another example, too low pit-entered moisture will cause the fermented grains to dry, high acidity, and the fermented grains to be not soft; abnormal saccharification and fermentation, high residual starch, and low liquor yield. If the moisture is too high, it will cause slow saccharification and fermentation, long fermentation period, fast microbial reproduction, large amount of acid production, incomplete fermentation, sticky fermented grains, and difficult steaming of grains. Increasing the amount of water input, compressing the gap between the fermented grains, and expelling air will reduce the amount of air in the fermented grains, thereby reducing the amount of microbial reproduction, slowing down the saccharification and fermentation speed, reducing the heat production speed, and slowing down the fermentation temperature rising speed. For another example, acid substances are important flavor components of Luzhou-flavor liquor, and appropriate organic acids in the fermented grains are precursors for forming flavor components of Luzhou-flavor liquor, which can be esterified to form various ester substances. Therefore, when the acidity in the fermented grains is not enough, the produced liquor is not strong-flavor and monotonous; but high acidity will inhibit the growth and reproduction of beneficial microorganisms (mainly yeast), thereby producing little or no liquor. At the same time, the fermentation power of Daqu is high, and the alcohol production capacity is strong, which can convert more Tang into alcohol and improve the liquor yield.

[0057] Based on this, the embodiment considers introducing the following physicochemical indexes as original features. Specifically, they include the water hitting time, the amount of powder grains, the amount of rice husks, the amount of fermented grains, the moisture, the acidity, the starch amount, the caproic acid amount, the saccharifying power of Daqu, the fermentation power of Daqu, the daily maximum air temperature, the daily minimum air temperature, the pit-entered temperature, and the temperature change rate in a preset time period of the upper fermented substance, the middle fermented substance, and the lower fermented substance. In specific implementation, the temperature change rate in the preset time period is the temperature change rate in 7 days after entering the pit. The temperature change rate can be obtained through empirical fitting.

[0058] Step S204, based on the fermented substance level division, the layer difference between any two layers of each original feature, the mean and standard deviation between three layers, and the interaction value of different planned physicochemical indexes in the same layer are respectively obtained, and then a plurality of corresponding enhanced features are obtained.

[0059] In this step, considering the data island characteristics and limitations caused by the confidentiality of production data in liquor production, the original features are used as the basis to obtain enhanced features through inter-layer difference processing, statistical processing and interaction processing; while improving the data volume, it is also conducive to deeper feature association, thereby realizing accurate prediction under limited data.

[0060] In specific implementation, the enhanced features obtained based on inter-layer difference include: the difference value of the middle layer fermented grains and the lower layer fermented grains, the difference value of the middle layer fermented grains and the upper layer fermented grains, and the difference value of the upper layer fermented grains and the lower layer fermented grains. The enhanced features obtained based on statistical processing include: the mean value of the three-layer water punching time, the mean value of the three-layer rice hull amount, the mean value of the three-layer powder grain amount, the standard deviation of the three-layer Daqu saccharifying power, the standard deviation of the three-layer Daqu fermenting power, and the standard deviation of the three-layer temperature change rate. The enhanced features obtained based on interaction processing include: the product of the same layer powder grain amount and rice hull amount, and the ratio of the same layer daily maximum temperature and daily minimum temperature.

[0061] Step S206, input all original features and all enhanced features into the first prediction model to obtain a predicted top temperature, and then input all original features, all enhanced features and the predicted top temperature into the second prediction model to obtain a predicted alcohol content.

[0062] In specific implementation, corresponding denoising and standardization processing are performed before inputting the model.

[0063] Specifically, for the temperature change of solid-state fermentation of liquor, slow warming in the early stage (12-15 days, warming by about 15°C), maintaining the temperature for a period of time (the top temperature can be maintained for 7-10 days), and slow reduction in the later stage (to the end of fermentation, the top temperature is reduced by 3-5°C), that is, slow in the early stage, steady in the middle stage, and slow in the later stage. If the intermediate temperature of the current fermentation cannot be maintained, it will affect the final quality of the liquor. Based on this, the top temperature is predicted.

[0064] In this embodiment, the first prediction model and the second prediction model are both built based on neural networks. Specifically, in order to improve the convergence speed and improve the prediction accuracy, the model is set as follows: directional decomposition of convolution kernel is performed by double path, and attention mechanism is introduced to extract information within each feature. Specifically, combined with Figure 2 As shown in FIG. 6, the method comprises the following steps:

[0065] Step S20602, in the first prediction model and the second prediction model, the gradient difference between the same original feature or enhanced feature between different fermentation material levels is calculated based on the longitudinal convolution path to obtain the correlation between the same features between different layers.

[0066] Step S20604, based on the transverse convolution path, the matrix product between a preset number of original features and / or enhanced features sequentially adjacent in the same fermentation level and the transverse convolution kernel is calculated to obtain the correlation between different features in the same layer.

[0067] Step S20606, based on the attention mechanism, the feature region weight of each original feature and enhanced feature of each layer is calculated to highlight the core region.

[0068] Specifically, the first prediction model structure of the first prediction model is shown in Table 1 below.

[0069] Table 1 First prediction model structure

[0070]

[0071] The second prediction model structure of the second prediction model is shown in Table 2 below.

[0072] Table 2 Second prediction model structure

[0073]

[0074]

[0075] Based on the above model design, the first prediction model has a validation set loss value of 1.13 after iterative training, and the fitting degree reaches 95%; the second prediction model has a validation set loss value of 0.5 after iterative training, and the fitting degree reaches 85%; both show good prediction ability of the model.

[0076] The parameter amount of the two prediction models is shown in Table 3 below, and the total parameter amount at this time is only 6,625kb; which is much smaller than the total parameter amount (specifically 1.14Mb) of the ordinary CNN model prediction.

[0077] Table 3 Parameter amount calculation

[0078] Layer Parameter calculation formula Parameter amount Design optimization point PathA longitudinal convolution (3×1×1)×32+32 128 Narrow kernel (3, 1) focus on inter-layer relationship PathB transverse convolution (1×5×1)×32+32 192 Wide kernel (1, 5) capture feature interaction Fusion 1x1 convolution (1×1×64)×64+64 4,160 Dimension reduction while fusing double paths Attention mechanism (1×1×64)×1+1 65 Lightweight dynamic weighting Output layer 64×32+32,32×1+1 2,080 Global pooling instead of Flatten

[0079] As a preferred embodiment, in order to further improve the prediction accuracy of the model, the model is combined with the attention mechanism as shown in the following formula (2) to further improve the prediction accuracy of the model. Figure 3 The model is periodically optimized as follows:

[0080] Step S20642, it is judged that the preset update period is reached, the verification original features, verification actual top temperature corresponding to one actual production in the historical database are obtained, and the corresponding verification enhanced features are obtained based on the verification original features.

[0081] Step S20644, inputting each item of verification original feature and verification enhanced feature into the first prediction model to obtain verification predicted top temperature, and inputting each item of verification original feature and verification enhanced feature and verification actual top temperature into the second prediction model to obtain verification predicted alcohol content.

[0082] Step S20646, when the error between the verification predicted top temperature and the verification actual top temperature is greater than the preset top temperature error threshold, re-calling the historical database to train and optimize the first prediction model; when the error between the verification predicted alcohol content and the verification actual alcohol content is greater than the preset alcohol content error threshold, re-calling the historical database to train and optimize the second prediction model.

[0083] At this time, through steps S20642-S20646, the models can be periodically optimized according to the actual prediction performance of the models at a preset period.

[0084] Step S208, adjusting each item of planned physicochemical index based on the predicted top temperature and the predicted alcohol content to obtain each item of iterative physicochemical index.

[0085] Step S210, repeating the above steps in sequence until the predicted top temperature and the predicted alcohol content reach optimal values, taking each item of iterative physicochemical index as each item of final physicochemical index; and adjusting the formula of the fermentation material in the target fermentation tank according to the physicochemical indexes and putting into production.

[0086] As can be seen from the above, the embodiment innovatively provides a liquor production method based on a machine learning algorithm, and optimizes the input features, the model itself, and the continuous prediction process in sequence, so that the corresponding prediction model can accurately predict the top temperature and the alcohol content. At this time, it is no longer necessary to manually sample the fermentation tank, which affects the quality of the liquor, and the fermentation process can be directly adjusted, thereby ensuring the fermentation quality of each fermentation process.

[0087] The above program can run in a processor, or can also be stored in a memory (or called computer readable storage medium), the computer readable medium includes permanent and non-permanent, movable and non-movable medium can be realized by any method or technology information storage. Information can be computer readable instructions, data structure, program module or other data. Examples of computer storage medium include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), electrically erasable programmable read only memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic disk storage or other magnetic storage device or any other non-transmission medium, which can be used to store information that can be accessed by a computing device. According to the definition in this paper, computer readable medium does not include temporary computer readable medium, such as modulated data signal and carrier wave.

[0088] These computer programs can also be loaded into a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate computer implemented processing, so that the instructions executed on the computer or other programmable device provide a process for implementing the steps specified in the flow Figure 1 One flow or multiple flows and / or blocks Figure 1 One block or multiple blocks, the steps of the function specified in the block or blocks can be realized by different modules.

[0089] The embodiment also provides a liquor production system based on deep learning, which combines Figure 4 As shown in the figure, the system comprises:

[0090] The first acquisition module is used for acquiring each planned physicochemical index of the upper layer fermentation, the middle layer fermentation and the lower layer fermentation in the target fermentation tank as each original feature.

[0091] The second acquisition module is used for acquiring the interlayer difference between any two layers of each original feature, the mean and standard deviation among three layers, and the interaction value of different planned physicochemical indexes in the same layer based on the fermentation level division, and then obtaining a plurality of corresponding enhanced features.

[0092] The prediction module is configured to input all original features and all enhanced features into a first prediction model to obtain a predicted top temperature, and then input all original features, all enhanced features and the predicted top temperature into a second prediction model to obtain a predicted alcohol content, wherein the first prediction model and the second prediction model are both built based on a neural network, adopt double-path convolution kernel directional decomposition, and introduce an attention mechanism to extract information in each feature.

[0093] The index adjustment module is configured to adjust each planned physicochemical index based on the predicted top temperature and the predicted alcohol content to obtain each iteration physicochemical index.

[0094] The optimization production module is configured to repeatedly execute the above steps until the predicted top temperature and the predicted alcohol content both reach optimal values, so as to take each iteration physicochemical index as a final physicochemical index, and adjust a formula of a fermentation product in a target fermentation tank according to the final physicochemical index and then put the fermentation product into production.

[0095] Since the system is built based on the method, the above-described details are not repeated here.

[0096] For example, the prediction module includes:

[0097] The judgment unit is configured to judge whether a preset update cycle is reached, obtain each verification original feature and verification actual top temperature corresponding to actual production in a historical database, and obtain each verification enhanced feature based on each verification original feature.

[0098] The verification unit is configured to input each verification original feature and verification enhanced feature into the first prediction model to obtain a verification predicted top temperature, and input each verification original feature, verification enhanced feature and verification actual top temperature into the second prediction model to obtain a verification predicted alcohol content.

[0099] The optimization unit is configured to, when an error between the verification predicted top temperature and the verification actual top temperature is greater than a preset top temperature error threshold, re-call the historical database to train and optimize the first prediction model, and when an error between the verification predicted alcohol content and the verification actual alcohol content is greater than a preset alcohol content error threshold, re-call the historical database to train and optimize the second prediction model.

[0100] For another example, the prediction module includes:

[0101] The first processing unit is configured to perform standardization processing on each original feature and enhanced feature to obtain a first single-channel image.

[0102] The first prediction unit is configured to perform denoising processing on the first single-channel image to obtain a first target image, and input the first target image into the first prediction model to obtain a predicted top temperature.

[0103] The second processing unit is configured to perform standardization processing on the original features, the enhanced features, and the predicted top temperature to obtain a second single-channel image.

[0104] The second prediction unit is configured to perform denoising processing on the second single-channel image to obtain a second target image, and input the second prediction model to obtain a predicted alcohol content.

[0105] In combination Figure 5 As shown in the embodiments, the electronic device comprises at least one processor coupled with a memory, and the memory stores a computer program configured to be executed by the processor to perform the method.

[0106] Meanwhile, the embodiments also provide a computer-readable storage medium having a computer program stored thereon, and the computer program is used to be executed by a processor to implement the method.

[0107] Since the system, the electronic device, and the storage medium are all used to implement or execute the method, they have the technical advantages of not needing manual sampling to optimize the current fermentation process and to guide the subsequent fermentation process.

[0108] Although the present application has been disclosed with reference to the preferred embodiments, it is not intended to limit the present application. Those skilled in the art who have ordinary knowledge can make various modifications and improvements without departing from the spirit and scope of the present application. Therefore, the protection scope of the present application should be defined by the claims.

Claims

1. A method for producing baijiu (Chinese liquor) based on deep learning, characterized in that, include: The planned physicochemical indicators of the upper, middle and lower fermentation products in the target fermentation tank are obtained as the original characteristics. Based on the hierarchical division of fermentation products, the interlayer difference between any two layers of each original feature, the mean and standard deviation between the three layers, and the interaction values ​​of different planned physicochemical indicators within the same layer are obtained, thereby obtaining several corresponding enhancement features. All original features and all enhanced features are simultaneously input into the first prediction model to obtain the predicted top temperature. Then, all original features, all enhanced features, and the predicted top temperature are input into the second prediction model to obtain the predicted alcohol content. Both the first and second prediction models are built based on neural networks, using dual-path convolution kernel directional decomposition and introducing an attention mechanism to extract information from each feature. Based on the predicted top temperature and the predicted alcohol content, the various planned physicochemical indicators are adjusted to obtain various iterative physicochemical indicators; Repeat the above steps until the predicted top temperature and the predicted alcohol content reach their optimal values. Use the corresponding iterative physicochemical indicators as the final physicochemical indicators. Adjust the formula of the fermentation material in the target fermentation tank according to these indicators and put it into production.

2. The deep learning-based liquor production method according to claim 1, characterized in that, Each original characteristic includes: watering time of the upper fermentation material, middle fermentation material and lower fermentation material, amount of powdered grain, amount of rice husk, amount of mash, moisture, acidity, starch content, hexanoic acid content, saccharification power of Daqu, fermentation power of Daqu, daily maximum temperature, daily minimum temperature, temperature of entering the pool, and temperature change rate within the preset time.

3. The deep learning-based liquor production method according to claim 1, characterized in that, The process involves simultaneously inputting all original features and all enhanced features into a first prediction model to obtain the predicted top temperature, and then inputting all original features, all enhanced features, and the predicted top temperature into a second prediction model to obtain the predicted alcohol content; including: In the first and second prediction models, the gradient difference between the same original feature or enhanced feature between different fermentation layers is calculated based on the vertical convolution path to obtain the correlation between the same feature in different layers. The lateral convolution path is used to calculate the matrix product between a predetermined number of sequentially adjacent original features and / or enhanced features within the same fermentation layer and the lateral convolution kernel to obtain the correlation between different features in the same layer. Based on the attention mechanism, feature region weights are calculated for each original feature and enhanced feature in each layer to highlight the core region.

4. The deep learning-based liquor production method according to claim 1, characterized in that, The process involves simultaneously inputting all original features and all enhanced features into a first prediction model to obtain the predicted top temperature, and then inputting all original features, all enhanced features, and the predicted top temperature into a second prediction model to obtain the predicted alcohol content; including: Determine when the preset update cycle has been reached, obtain the original verification features and the actual top temperature of each actual production in the historical database, and obtain the corresponding enhanced verification features based on the original verification features. The original verification features and enhanced verification features are input into the first prediction model to obtain the verification predicted top temperature, and the original verification features, enhanced verification features, and actual verification top temperature are input into the second prediction model to obtain the verification predicted alcohol content. If the error between the predicted top temperature and the actual top temperature is greater than a preset top temperature error threshold, the first prediction model is retrained and optimized using the historical database; if the error between the predicted alcohol content and the actual alcohol content is greater than a preset alcohol content error threshold, the second prediction model is retrained and optimized using the historical database.

5. A deep learning-based liquor production system, characterized in that, include: The first acquisition module is used to acquire various planned physicochemical indicators of the upper, middle and lower fermentation products in the target fermentation tank as each original feature. The second acquisition module is used to acquire, based on the fermentation product hierarchical division, the interlayer difference between any two layers of each original feature, the mean and standard deviation between three layers, and the interaction values ​​of different planned physicochemical indicators within the same layer, thereby obtaining several corresponding enhancement features. The prediction module is used to simultaneously input all original features and all enhanced features into the first prediction model to obtain the predicted top temperature, and then input all original features, all enhanced features and the predicted top temperature into the second prediction model to obtain the predicted alcohol content; wherein, both the first prediction model and the second prediction model are built based on neural networks, using dual-path convolution kernel directional decomposition, and introducing an attention mechanism to extract information within each feature; The indicator adjustment module is used to adjust various planned physicochemical indicators based on the predicted top temperature and the predicted alcohol content to obtain various iterative physicochemical indicators; The production optimization module is used to repeat the above steps sequentially until the predicted top temperature and the predicted alcohol content both reach their optimal values, and to use the corresponding iterative physicochemical indicators as the final physicochemical indicators; the fermentation material in the target fermentation tank is then adjusted according to these indicators and put into production.

6. The deep learning-based liquor production system according to claim 5, characterized in that, The prediction module includes: The judgment unit is used to determine when the preset update cycle has been reached, obtain the original verification features and the actual top temperature of each actual production in the historical database, and obtain the corresponding enhanced verification features based on the original verification features. The verification unit is used to input the original verification features and the enhanced verification features into the first prediction model to obtain the predicted verification top temperature, and to input the original verification features, the enhanced verification features, and the actual verification top temperature into the second prediction model to obtain the predicted verification alcohol content. The optimization unit is used to, when determining that the error between the verified predicted top temperature and the verified actual top temperature is greater than a preset top temperature error threshold, retrain and optimize the first prediction model by calling the historical database; and to, when determining that the error between the verified predicted alcohol content and the verified actual alcohol content is greater than a preset alcohol content error threshold, retrain and optimize the second prediction model by calling the historical database.

7. The deep learning-based liquor production system according to claim 5, characterized in that, The prediction module includes: The longitudinal convolutional unit is used in the first and second prediction models to calculate the gradient difference between the same original feature or enhanced feature between different fermentation layers based on the longitudinal convolutional path to obtain the correlation between the same feature in different layers. A lateral convolutional unit is used to calculate the matrix product between a predetermined number of sequentially adjacent original features and / or enhanced features within the same fermentation layer and the lateral convolutional kernel based on the lateral convolutional path to obtain the correlation between different features in the same layer. The attention mechanism unit is used to calculate the feature region weights of each original feature and enhanced feature in each layer based on the attention mechanism in order to highlight the core region.

8. An electronic device, characterized in that, It includes at least one processor coupled to a memory storing a computer program configured to be executed by the processor to perform the method of any one of claims 1-4.

9. A computer-readable storage medium, characterized in that, It stores a computer program thereon, which is executed by a processor to implement the method of any one of claims 1-4.

Images