KR20200072754A - Alcohol fermentation system with complex neural network algorithm - Google Patents

Abstract

The present invention relates to a liquor fermentation system applied with a complex neural network algorithm which accurately controls a traditional liquor fermentation process of a parallel double fermentation method in which saccharification and liquor fermentation processes are performed at the same time, thereby producing liquor in accordance with a desired purpose without abnormal fermentation. The liquor fermentation system analyzes various state information of a fermenter through a complex neural network algorithm for brewing the traditional liquor in which liquefaction, saccharification and liquor fermentation are simultaneously performed while yeast is working, and provides information on temperature control timing and temperature control by analyzing a current fermentation state in response to entire fermentation processes and goals, thereby having an effect of increasing possibilities of successful fermentation of the traditional liquor by even users with low skill levels.

Description

Alcohol fermentation system with complex neural network algorithm

The present invention relates to a mainstream fermentation system to which a complex neural network algorithm is applied, in particular, to precisely control the fermentation process of the traditionalism of the parallel fermentation method in which saccharification and alcohol fermentation are simultaneously performed, so that alcoholic beverages according to a desired purpose can be produced without abnormal fermentation. The present invention relates to a mainstream fermentation system to which a complex neural network algorithm is applied.

Beer, wine and whiskey are all over the world. On the other hand, makgeolli and soju (traditional distilled soju), despite their great advantages, do not occupy the active or mainstream flow even in Korea. There are two reasons for this, the first is the absence of story and content about the sake, and the second is the lack of systematization of the brewing method.

Among them, various stories and contents about alcohol are being developed, and interest in makgeolli or traditional liquor is rising with the Korean Wave. However, since the brewing method is not systematized, the difference in taste occurs even if the same ingredients are used in the case of traditionalism.

In the case of makgeolli, which is currently highly interested, the manufacturing method is simplified by using a single yeast strain, saccharomyces, which is cultivated using immigration in godubap. The method of fermentation by adjusting the titer for saccharification and injecting a single yeast therein minimizes the probability of failure, but it is manufactured in a way that the taste is simple and the taste is matched through the process of fermentation and seasoning to produce mass rice wine. In the case, the taste is mostly similar.

On the other hand, in the case of beer and wine, systematization that determines taste and quality is made by analyzing data accumulated for a long time. For example, in the case of beer, it was dataized that the combination of ingredients, such as color and writing, would result in a certain taste with a certain turbidity. In addition, it is in a state where it has been automated by systemizing the conditions for saccharification and fermentation.

In fact, in the case of wine, depending on the material and the aging method varies, the fermentation process is relatively simple, so the fermentation itself can be easily automated.

In the case of beer, the fermentation process is more complicated than that of wine, but the saccharification process and the alcohol fermentation process are separated, so data is relatively easy to automate.

However, in the case of traditionalism, since the saccharification process and the alcohol fermentation process are performed at the same time, there is a high probability of failure due to the explosive growth of microorganisms due to abnormal fermentation even with a slight temperature control failure, and the complex conditions need to be managed simultaneously. The taste is also different.

Therefore, even if data is collected for brewing traditional liquor, it is difficult to grasp the numerous effects on saccharification and fermentation, and it is difficult to properly manage it because the conditions for saccharification and fermentation are different. In particular, it is difficult to construct a control method by statistically simple analysis of various measurement data and glycation and fermentation states because various conditions and current glycation and fermentation states are complicated and correlated with each other.

Furthermore, considering these complex conditions, saccharification can be maximized and alcohol fermentation can be precisely controlled, but in this case, alcohol yield and frequency are high, but various flavors of liquor disappear, thus maintaining a balance of fragrance and taste, not increasing alcohol fermentation yield. In terms of controlling fermentation, it is difficult to systemize traditional brewing by simply collecting the data and making it statistical.

Therefore, by gathering and analyzing information on the process of saccharification and fermentation according to the type of traditionalism to be produced, but by learning the collected data through a neural network rather than a simple statistical method, we want to complexly analyze traditional fermentation that requires various analysis by situation. It is expected that there will be attempts. However, since it is difficult to determine the fermenter state in which microorganisms multiply in real time by simply changing the information or data, even if you prepare and learn the data that succeeded in actual brewing, it is only a degree to predict what the current state is. There is a limit to the application, and it is not easy to respond to cases in which the taste of alcohol is changed, environmental conditions are changed, or the quality of materials is changed.

It is not easy to collect enough labeled learning data that has been successful in brewing more realistically, and even if some data is prepared, a simple neural network is used to learn what information to learn and how to use it for actual fermentation control. It is also difficult to do.

Korean Registered Patent No. 10-1903528, [Optimized Control System for Landfill Gas and Biogas Power Plant Using Deep Learning Technique] Korean Registered Patent No. 10-1765652, [ICT based digital traditional fermented food manufacturing container]

The objective of the embodiments of the present invention is to detect various state information of the fermenter through sensors for brewing a traditional liquor in which liquefaction, saccharification, and alcohol fermentation are simultaneously performed while yeast is applied, and confirm the actual fermentation state through a microscope and a camera. Analyzes through the first neural network to determine the current fermentation status, and analyzes and controls the current fermentation status in response to the entire fermentation process and target through the second neural network using the determined image information and other sensor information as input. It is to provide a mainstream fermentation system to which a complex neural network algorithm is applied that can generate basic information for a.

Another object of the embodiment of the present invention is to set the current fermentation status information obtained through the complex neural network algorithm as a reward value in comparison with the target status in the entire fermentation process, and perform temperature control to perform various sensing information and image information of the fermenter. It is to provide a mainstream fermentation system to which a complex neural network algorithm is applied to provide a targeted liquor based on a complex state by enabling automatic temperature control for optimal reward generation through reinforcement learning.

Another object of the embodiments of the present invention is through a GAN (Generative Adversarial Network)-based intelligent network that generates saccharification and fermentation processes for obtaining sake of desired taste and alcoholicity for the reinforcement learning based on the brewing information provided by the user. By virtually generating and automatically determining rewards based on this goal, practical control can be performed while learning automatically in the reinforcement learning unit, so that changes in ingredients or changes in taste are instructed so that automatic fermentation control can be expected accordingly. It is to provide a mainstream fermentation system using a complex neural network algorithm.

Another object of the embodiments of the present invention is to provide a mainstream fermentation system to which a complex neural network algorithm is applied to solve a biased learning problem of reinforcement learning by synchronizing an intelligent network in an asynchronous manner through several physically distributed fermenters.

In order to achieve the above object, the mainstream fermentation system to which the composite neural network algorithm according to an embodiment of the present invention is applied is a fermentation tank composed of a plurality of sensors, a camera and a temperature control unit, and an intelligent network fermentation management unit analyzing the fermentation status of the fermentation tank A mainstream fermentation system comprising: an intelligent network fermentation management unit, comprising: a sensor data collection unit collecting numerical data through a plurality of information collection sensors configured in a fermentation tank; An image collection unit for collecting a fermentation status image of a fermentation target of a fermentation tank; A drive control unit for controlling the temperature of the fermenter; A user interface unit providing a current fermentation status to a user and collecting brewing information including a fermentation target and a fermentation target from the user; A first neural network that determines the fermentation status based on the collected image of the image collection unit, and outputs the current fermentation status by inputting the output of the first neural network and the brewing information collected through the sensor data collection unit and the user interface unit. It includes an intelligent network learning unit.

As an example, a reward determination unit for calculating a reward by comparing current fermentation status information output by the multi-learning network learning unit with a target brewing step and a step-by-step target; Using the output of the reward determination unit as a reward value, and the information collected through the sensor data collection unit and the information obtained through the first neural network through the image collection unit as status information, execution information for optimally controlling the temperature of the fermenter is generated, and It may include a reinforcement learning unit for controlling the driving control unit.

As an example, the reward determination unit includes a brewing type determination unit that determines a master species to be brewed based on the brewing information; A brewing step determination unit for determining a position among preset brewing steps through a current cumulative process; A step-by-step target derivation unit for calculating target fermentation status information in a current brewing step; A comparison with the target fermentation status of the step-by-step target derivation unit and the current fermentation status provided through the fermentation status analysis unit may include a comparison unit for calculating a reward.

As another example, the reward determination unit is a GAN (Generative Adversarial Network) based intelligent network model that generates saccharification and fermentation processes and fermentation targets for each process through learning to obtain a desired taste and alcoholic liquor based on brewing information provided by a user. Including, it is possible to calculate the reward based on the achievement level by comparing the current fermentation status information of the fermentation status analysis unit for each process and each process-generated goal generated through the GAN-based intelligent network model.

As an example, the reinforcement learning unit learns the collected state information and outputs an appropriateness for each temperature control as a probability value to provide a control signal of the driving control unit, an actor applied with an optimal control policy neural network model, and the same state as the execution unit It may include a criterion applied to the value calculation neural network model to output the expected compensation information when the driving control unit controls the fermentation environment according to the state information as input as the information to a predetermined temperature. Furthermore, the main reinforcement learning unit includes a main reinforcement learning unit that collects and reflects the learning results of the reinforcement learning unit through a communication network, and the main reinforcement learning unit comprises a reinforcement learning unit corresponding to a plurality of fermentation tanks and a communication network, and each learning result from a plurality of reinforcement learning units It can be collected asynchronously and synchronized with the reinforcement learning unit that provided the corresponding learning result.

The mainstream fermentation system using the complex neural network algorithm according to an embodiment of the present invention analyzes various state information of the fermentation tank through a complex neural network algorithm for brewing a traditional liquor where liquefaction, saccharification, and alcohol fermentation are simultaneously performed while yeast is applied. The state can be analyzed in response to the entire fermentation process and targets to provide information on temperature control timing and temperature control, so that even a user with low skill can increase the likelihood of success in targeting traditional fermentation.

In addition, the embodiment of the present invention repeats learning by applying a reinforcement learning unit that learns temperature control in a direction to achieve the best brewing based on current fermentation status information obtained through a complex neural network algorithm and collected sensing and image information. This has the effect of making it possible for anyone to succeed in brewing to provide the desired sake taste.

Furthermore, an embodiment of the present invention applies a fermentation process and a generative adversarial network (GAN) that automatically generates a target for each process based on brewing information provided by a user, and thus a process for target brewing and a target for each process Only by learning to automatically generate close to the actual, the actual control aimed at this has the effect of enabling the automatic fermentation control accordingly by instructing a change in material or a change in taste by learning through the reinforcement learning unit.

1 is a conceptual diagram of a fermentation process for explaining the mainstream fermentation method.
2 is a flow chart showing an exemplary manufacturing process of traditionalism.
3 is a photomicrograph showing a normal yeast fermentation state and an abnormal fermentation state appearing in a traditional fermentation process.
4 is a block diagram showing a mainstream fermentation system to which an algorithm is applied to a complex neural network according to an embodiment of the present invention.
5 is a block diagram showing the configuration of the intelligent network fermentation management unit according to an embodiment of the present invention.
6 is a detailed partial configuration diagram for explaining a practical operation method of the intelligent network fermentation management unit according to an embodiment of the present invention.
7 shows an example of a microscopic shape analysis unit which is a part of a multi-intellectual network learning unit according to an embodiment of the present invention.
8 is an exemplary view for explaining a fermentation state analysis unit according to an embodiment of the present invention.
9 is a conceptual diagram illustrating an expanded configuration of a reinforcement learning unit according to an embodiment of the present invention.

The present invention as described above will be described in detail through the accompanying drawings and embodiments.

It should be noted that the technical terms used in the present invention are only used to describe specific embodiments, and are not intended to limit the present invention. In addition, technical terms used in the present invention should be interpreted as meanings generally understood by a person having ordinary knowledge in the technical field to which the present invention belongs, unless otherwise defined in the present invention. It should not be interpreted as a meaning or an excessively reduced meaning. In addition, when the technical term used in the present invention is a wrong technical term that does not accurately represent the spirit of the present invention, it should be understood as being replaced by a technical term that can be correctly understood by those skilled in the art. In addition, the general terms used in the present invention should be interpreted as defined in the dictionary or in context before and after, and should not be interpreted as an excessively reduced meaning.

In addition, the singular expression used in the present invention includes a plural expression unless the context clearly indicates otherwise. In the present invention, terms such as “consisting of” or “comprising” should not be construed to include all of the various components, or various stages, described in the invention, including some of the components or some stages It may or may not be construed as further comprising additional components or steps.

Further, terms including ordinal numbers such as first and second used in the present invention may be used to describe elements, but the elements should not be limited by terms. The terms are used only to distinguish one component from another component. For example, the first component may be referred to as a second component without departing from the scope of the present invention, and similarly, the second component may be referred to as a first component.

Hereinafter, preferred embodiments according to the present invention will be described in detail with reference to the accompanying drawings, but the same or similar elements will be given the same reference numbers regardless of the reference numerals, and repeated descriptions thereof will be omitted.

In addition, in the description of the present invention, when it is determined that detailed descriptions of related known technologies may obscure the subject matter of the present invention, detailed descriptions thereof will be omitted. In addition, it should be noted that the accompanying drawings are only for easy understanding of the spirit of the present invention and should not be interpreted as limiting the spirit of the present invention by the accompanying drawings.

Hereinafter, embodiments of the present invention will be described with reference to FIGS. 1 to 9.

1 is a conceptual diagram of a fermentation process for explaining the mainstream fermentation method.

Brewing can be largely divided into saccharification step, alcohol fermentation step and aging step. First, the step of saccharification is to break down polysaccharides into monosaccharides, which prepare the alcoholic fermentation by breaking the chains of polysaccharides in which monosaccharides are woven into a plurality of chains. Glycosylase performs this task.

When the saccharification process is completed, the desired liquor can be obtained through the process of fermenting monosaccharides into alcohol through yeast, and these saccharification and alcohol fermentation processes are performed respectively or in combination.

The single fermentation process shown in FIG. 1A is applied to fruit wines such as wine, and since it is already directly used for fruit, which is a monosaccharide or disaccharide, it does not require a saccharification process in the first place, and when only yeast is injected, alcohol fermentation occurs to complete fruit wine. Of course, these fruit wines have different flavors in the characteristics of the fruit itself or in the later ripening process than in this fermentation process, but the alcohol fermentation process itself can be performed without major failure, and for this reason, the fruit naturally falls off and becomes liquor.

FIG. 1B shows a single-line fermentation process in which sugar is first produced, for example, beer, and yeast is added to generate alcohol. As shown, the barley is germinated to form a malt, and after crushing it, hops and water are added to make the malt saccharified with oligosaccharides or glucose. Then, yeast is injected into the saccharified wort to ferment alcohol to produce beer.

As shown in the figure, since the process of saccharifying malt and the process of alcohol fermentation are divided, it is possible to produce beer through an automated process by collecting and analyzing data on the fermentation process.

For example, after saccharification at an appropriate temperature (55 to 75 degrees) in a saccharification tank, the temperature is lowered to 20 degrees or less within a short period of time, transferred to a prepared fermentation tank, and then the yeast is added to fermentation temperature (ale: 15 to 25 degrees/ Lager: 5~15 degrees), alcohol is naturally produced.

However, in the case of traditional liquor (eg, Takju or Yakju), the saccharification process and the alcohol fermentation process are simultaneously performed as shown in FIG. 1C.

As shown in the figure, yeast is sprayed on the starch raw material to generate yeast enzyme. Then, yeast is added to the yeast to cultivate the yeast until a certain amount is obtained to make a main yeast. This is called undersake or undersake and corresponds to a kind of starter in which yeast is purely cultured. This is applied to the starch raw material to make sake and make liquor. In this process, liquefaction, saccharification, and alcohol fermentation occur simultaneously to make yakju or takju. This process is called concurrent rejuvenation, and if the starch is not appropriately saccharified before alcohol fermentation, alcohol fermentation does not occur, and if the temperature is not properly adjusted in the alcohol fermentation process, abnormal fermentation occurs and microorganisms explode explosively by bacteria. The alcohol will be rested.

In the case of saccharification, it is made at a relatively high temperature. When saccharification is heated at a temperature between 65 and 68 degrees, saccharification is achieved by enzymatic action such as amylase. However, since the yeast responsible for alcohol fermentation dies at a high temperature optimized for such saccharification, temperature control becomes extremely important when saccharification and alcohol fermentation occur simultaneously.

In particular, when the enzyme changes the starch of the grain to sugar, the enzyme has an optimum temperature and acidity for each enzyme, and if it exceeds a certain temperature range, the enzyme is inactivated.A protein with a large size among the protein components of the grain makes the body of the liquor cloudy and sticky. And medium-sized proteins contribute to the formation and maintenance of bubbles, and small-sized proteins become nutrients for yeast. Therefore, it is necessary to cut a large sized protein into a medium or smaller size, which is called protein adjustment, and for this, a process of maintaining the conducting grain for 20 to 30 minutes at a temperature of 50 to 55 degrees is necessary. That is, it is also important to control and maintain the temperature for adjusting the body feeling.

In the end, in the case of fermentation, in the case of yeast, the treatment by glycation enzyme and the appropriate temperature for alcohol fermentation are different, so that these two things occur smoothly at a single temperature, and also alcohol that is fermented when alcohol fermentation is performed after saccharification. Since there is an organic linkage in which abnormal fermentation is suppressed, temperature control in the initial fermentation process is extremely important.

In the case of such a combination fermentation, there is also a purpose to increase the taste and flavor, because the balance of saccharification and alcohol fermentation needs to be well controlled. When the saccharification process and the alcohol fermentation process are approached from the viewpoint of the alcohol yield, the alcohol yield and alcohol frequency are high, but the flavor disappears from the alcohol when the alcohol is fermented rapidly after initial saccharification at a maximum through the purification enzyme.

That is, traditional brewing does not simply maximize the level of alcohol fermentation, but because it prioritizes the quality of brewing obtained from alcohol fermentation, it is necessary to perform fermentation control while maintaining a balance of aroma and taste, not from an efficiency point of view. For example, rather than saccharifying all starch into monosaccharides, non-fermented sugars (some polysaccharides) that remain in the fermentation tank without being partially fermented can be made deliberately to enhance sweetness and increase body feel. To this end, it may be possible to adjust the temperature of saccharification or use glutinous rice with a lot of amylopectin instead of non-glutinous rice or reduce the amount of water. This can cause a precise temperature control is required.

2 is a flow chart showing an exemplary manufacturing process of traditionalism.

As shown in the figure, first, the grain containing starch is called water so that the saccharifying enzyme can decompose well, and it is usually managed by heating for easy gelatinization. The main method is to add suitable water to make porridge, to make snow white or steam with godubap, and mix the starch that has undergone this treatment with yeast in an appropriate ratio. Typically, for the saccharification power of 300SP, mix from 9% to 13% to 15%. When used a lot, more than 20% to 25% may be used.

After this process, saccharification and fermentation process is started by adding appropriate water in proportion to the weight of rice and yeast to the sterilized container or fermenter.

In the case of the saccharification and fermentation process, it is necessary to maintain a constant temperature for saccharification in the initial stage, and when saccharification starts, yeast cells are activated by heat generated during the saccharification process thereafter, and the yeast bacteria perform alcoholic fermentation and for other germs As for the control, fermentation is maintained stably after the initial fermentation occurs due to the alcohol generated, thereby completing yakju or takju.

In other words, since both the fermentation agent and the fermentation agent, yeast are used, and concurrent fermentation with simultaneous saccharification and fermentation is carried out through one fermentation tank, the optimum temperature for saccharification of yeasts (55~75 degrees) and alcohol fermentation temperature (25) The difference between the provinces and provinces is usually around 25 to 35 degrees. If the fermentation starts at about 35 degrees, the intermediate temperature of saccharification and fermentation temperature, the kinetic energy of the yeast is converted into thermal energy, so that it quickly becomes more than 38 degrees Celsius. At this temperature, the yeast is exhausted and further growth is suppressed, and the fermentation temperature When it goes over 40 degrees, the yeast begins to die. Therefore, the fermentation temperature of traditional liquor is fermented according to the temperature favored by yeast, which is a temperature lower than the saccharification temperature.

Since fermentation occurs at such a low temperature, the grain must be crushed to effectively saccharify or sufficiently cooled in hot water and then cooled before being placed in the fermentation tank. In the following procedure, the activation temperature of the yeast is about 25 degrees (temperature inside the fermentation tank) through a stirring process. Fermentation.

FIG. 3 shows photographs of yeast bacteria and lactobacillus or acetic acid bacteria that are normally activated during saccharification and alcohol fermentation. When the temperature control is normally performed as shown in FIG. 3A, yeast bacteria are activated to generate alcohol and the produced alcohol controls the bacteria, and when the temperature control fails, bacteria are explosively generated as shown in FIG. 3B. During abnormal fermentation, acetic acid bacteria that cause the sour to appear, such as rods or sausage-type abnormal bacteria (lactic acid bacteria) or black spots, appear.

After all, it is necessary to calculate the optimum temperature control timing and control temperature required for saccharification and alcohol fermentation, the type and condition of the material for saccharification, the characteristics of the yeast used and various temperature conditions, and the state of the fermentation vessel that varies differently according to these various conditions. Since it is difficult to specify (temperature, acidity, carbon dioxide, specific gravity, degree of yeast activation, size or frequency of bubbles inside the fermenter, etc.), until now, after considering these various situations, the fermentation state is visually observed and the smell and taste The fermentation temperature was adjusted by judging empirically while confirming through.

The present invention integrates these various conditions to accurately grasp the current fermentation state, and then analyzes the threshold value rather than the control method after grasping the changed state beyond the actual threshold to control the temperature immediately before that. In order to prevent and obtain the desired flavor, we apply a neural network algorithm that takes into account the actual experience judgment method.

4 is a block diagram showing a mainstream fermentation system to which an algorithm is applied to a complex neural network according to an embodiment of the present invention.

As shown, the mainstream fermentation system to which the algorithm is applied to the complex neural network includes a fermentation tank 10 composed of a plurality of sensors, a camera, and a temperature control unit, and an intelligent network fermentation management unit 100 that analyzes the fermentation state of the fermentation tank 10.

The fermentation tank 10 basically includes a stirring unit 12 for stirring the fermentation target 20 and a motor unit 11 for rotating the stirring unit 12 to increase or decrease the temperature inside the fermentation tank. A temperature control unit 13 composed of a heater or a chiller, an extraction pipe 14 configured to check the internal state through a quartz tube 15 while selectively extracting a fermentation object inside the fermentation tank through a valve, and a fermentation tank A temperature sensor unit 16 for measuring the inside and outside temperature and the outside temperature, a hydrometer 17 configured in the extraction pipe 14, an acidity measuring unit 18 for measuring acidity, and measuring carbon dioxide It has a carbon dioxide measuring unit 19 for. In addition, it includes an internal camera unit 30 for photographing an image of the fermentation object inside the fermentation tank, and a microscope camera unit 40 for photographing a microscope image of the fermentation object extracted through the extraction pipe 14. Of course, in addition to the various sensors shown, an additional sensor including a sugar sensor for measuring sugar content, an acoustic sensor for measuring fermentation sound, and a camera unit for observing the quartz tube 15 may be configured, and the camera units 30 and 40 For this, a configuration for replacing a light source or a sample may be further included.

On the other hand, the intelligent network fermentation management unit 100 is the basic information about the object to be fermented from the user, the type of the material and the mixing ratio (whether rice, glutinous rice, whether they are mixed, porridge, snow white, or godubap, Receives brewing information, including fermentation targets, such as how much water is included, information about strains such as yeast and yeast, and what kind of liquor you want to ferment with (type, sugar content, body feeling, flavor, alcohol content, etc.) And can guide the current status. Based on this guide information, the user can control the fermenter. If the user's operation is successful, the user's operation can be learned with a high weight and stored as a learning information labeled in the database.

5 is a block diagram showing the configuration of the intelligent network fermentation management unit 100 according to an embodiment of the present invention.

As shown, the sensor data collection unit 110 for collecting numerical data through a plurality of information collection sensors 16 to 19 configured in the fermentation tank 10 and a fermentation state image of the fermentation target of the fermentation tank 10 ( 30, 40), the image collection unit 120, the driving control unit 130 for controlling the temperature of the fermentation tank 10, and providing the current fermentation status to the user and brewing including the fermentation target and the fermentation target from the user The user interface unit 180 for collecting information, the first neural network for determining the fermentation status based on the collected image of the image collection unit 120, the output of the first neural network, and the collected data of the sensor data collection unit 110 And a multi-intelligent network learning unit 140 that outputs the current fermentation status (estimation) by using the brewing information collected through the user interface 180 as an input. This will be described later in detail with reference to FIG. 6.

The multi-intelligent network learning unit 140 may output the current fermentation state at a certain stage in the fermentation process and the state for the saccharification and fermentation state as a probability value. For example, in the fermentation process defined as a plurality of steps, the probability of the A-3 step is 87%, the current saccharification is 35%, the probability is 92%, and the fermentation is 12%, the probability is 88%. Can. These outputs can be changed to the required format by adjusting the output stage as necessary (for example, outputs such as 35% saccharification, 12% fermentation, or 67% saccharification compared to the target reaching level, and 45% fermentation). May be provided).

Even if only such a configuration is present, the user can predict the current fermentation state based on the output of the multi-intelligent network learning unit 140 and manually control the appropriate temperature, thereby increasing the probability of success in brewing by the manager.

On the other hand, in the embodiment of the present invention, since a specific control method may vary depending on the main species to be brewed (takju, yakju, etc.) or basic materials for brewing a large stem, after categorizing them according to the needs of these broad categories of liquor, The neural network analysis may be performed by performing separate learning for each category, and the output of the learning and multi-intelligent network learning unit 140 may be provided by classifying categories based on brewing information input by a user.

In this way, it is possible to increase the success rate of brewing by providing information on timing that requires temperature control by estimating the current fermentation status based on various information, but the user must manually control the temperature by referring to the estimated current status or Since the temperature is simply adjusted according to the control algorithm, it is difficult to reliably secure the mainstream of desired quality.

Therefore, in the present invention, a reinforcement learning method of managing fermentation at an appropriate temperature at an appropriate timing is further applied.

In the illustrated example, A3C (Asynchronous Advantage Actor-Critic) based in-depth reinforcement learning method is applied. A3C in-depth reinforcement learning method is used by Deep Mind to perform video game play with artificial intelligence, and in reality, these scores are compensated by compensating the scores obtained through state information and behavior of the current video game environment. By learning the process of performing control for game play in the direction of getting high, the game can be gradually played like a human player, and furthermore, it is possible to play games that overtake people. Currently, games such as simple brick breaking can be applied relatively effectively to the control of the present invention, which controls only the temperature, since it has been a long time since it overtakes human play.

Practically, this A3C-based in-depth reinforcement learning method requires a configuration to simply copy the learning environment and experience the same game in various different situations through various settings (otherwise, learning is biased due to limited experience and performance decreases). There must be an accurate reward (score of the game), but it is not easy to apply it to an ambiguous situation for a real fermenter.

Normally, in A3C deep reinforcement learning, state information on a control target for grasping the current situation, control information (actor) to increase the reward by changing the reward information (reward) and state information obtained according to the result of the control is used. .

In the present invention, the output of the reward determination unit 150 and the reward determination unit 150 for calculating a reward by comparing the current fermentation status information output by the multi-intelligence network learning unit 140 with a target brewing step and a step-by-step target is a reward value. Optimally controlling the temperature of the fermenter 10 by using information collected through the sensor data collection unit 110 and information classified through the first neural network as images obtained through the image collection unit 120 as state information. It includes a reinforcement learning unit 160 for generating the execution information and controls the driving control unit 130 through this.

Meanwhile, a database storing a plurality of neural networks constituting the multi-intelligence network learning unit 140 and cumulative information on a fermentation process and a fermentation state that the reward determination unit 150 and the reinforcement learning unit 160 can refer to for learning. It may include (170).

Then, the method for estimating the current fermentation state through these configurations and the method for automatically performing the fermentation control based on the configuration will be described in more detail with reference to FIGS. 6 to 8.

6 is a detailed partial configuration diagram for explaining a practical operation method of the intelligent network fermentation management unit 100 according to an embodiment of the present invention, a more specific configuration of the multiple intelligent network learning unit 140 and the reward determination unit 150 It is for explaining the exemplary configuration and interworking with the reinforcement learning unit 160.

The illustrated multi-intelligent network learning unit 140 analyzes the fermentation state through a multiple neural network model using various current fermenter state information and brewing information provided by a user through input.

In particular, the first neural network model is applied to distinguish the fermentation status based on image information in order to process the internal image of the fermentation tank or the microscopic image of the fermentation object, which is difficult to distinguish numerically, in addition to the numerical data obtained through various sensors.

The illustrated multi-intelligent network learning unit 140 can determine from the microscope image whether the current fermentation state is a yeast activation state, a state in which other microorganisms have occurred, or an abnormal fermentation state by another microorganism, etc. It includes a microscope image analysis unit 141 to which the image classification neural network model is applied, and a camera image analysis unit 142 to which the image classification neural network model is applied to check the current fermentation state from the image inside the fermenter (foam state). As described above, the outputs of the microscope image analysis unit 141 and the camera image analysis unit 142, which learn image information that cannot be quantified numerically and provide the output to quantify the state, together with other sensor data and brewing information. And a fermentation state analysis unit 143 that analyzes and outputs current saccharification and fermentation states as inputs.

In an embodiment of the present invention, the microscope image analysis unit 141 learns the input image through a convolutional neural network (CNN)-based neural network model, as shown in FIG. 7, and displays the current yeast or other fungi from the input microscope image. It can be configured to output. The microscope image analysis unit 141 may provide, for example, the presence or activity of yeast as a probability value or a normalized numerical value, and the presence or activity of a bacterium may also be provided as a probability value or a normalized numerical value.

The camera image analysis unit 142 may also have the same or similar configuration to the microscope image analysis unit 141.

As described above, by quantitatively using the real-time fermenter internal image and the microscopic image of the fermentation target that were not previously used for data analysis, it is possible to greatly increase the reliability of the fermentation state analyzer 143 that estimates the current saccharification and fermentation process.

FIG. 8 shows a configuration example of the fermentation state analysis unit 143, and as shown, results of image analysis through a first neural network model configured in the microscope image analysis unit 141 or the camera image analysis unit 142 are different numerical values. An input management unit that collects as input as red sensor data, and a Recurrent Neural Network (RNN) that estimates the current state of glycation and fermentation based on time-series information of the existing glycation and fermentation based on these various inputs (or an improvement thereof) RNN analysis unit to estimate the current state of glycosylation and fermentation using LSTM (Long Short-Term Memory), or GRU (Gated Recurrent Unit) based circulatory neural network model, and for example, the current glycosylation and fermentation process It may include an output management unit configured to provide a step estimate and an estimate of the saccharification and fermentation status at the step.

In this way, in order to cyclically learn several sensor measurement parts for visual judgment parts and other sensory judgment parts, such as taste and aroma, which are important criteria in the existing empirical fermentation management, in an embodiment of the present invention, images are analyzed and applied. Based on the primary neural network model that classifies the glycosylation or fermentation status to which the image belongs (such as the degree of activation of yeast or the presence of other bacteria), the output of the primary neural network model, and sensor data and brewing information collected from other sensors By using the multi-intelligent network learning unit 140 configured with multiple secondary neural network models for estimating the saccharification and fermentation state, the experienced person can accurately distinguish the saccharification and fermentation state through the current situation.

When the current saccharification and fermentation states are accurately classified, it is empirically judged the temperature control point and the degree of temperature control for the desired flavor and alcohol frequency while suppressing abnormal fermentation, and in the embodiment of the present invention, the temperature control depending on this experience The timing and temperature control can be automatically performed through reinforcement learning.

As illustrated, the reinforcement learning unit 160 is, for example, an A3C deep reinforcement learning unit, and learns how to perform control in consideration of the current state in the direction of increasing compensation. For this, the most important part provides accurate compensation Is to do.

Since the brewing according to the embodiment of the present invention is not simply to increase the efficiency of alcohol fermentation, but to brew alcohol with proper alcohol content that provides a desired taste, aroma, and body feel, it is necessary to properly provide compensation for the current state. Through the reward determination unit 150, a reward for the current (glycosylated) fermentation state, which is the output of the multiple intelligent network learning unit 140, is provided.

The reward determining unit 150 does not determine whether to reach it based on a specific brewing control method, but sets only a brewing step to obtain a desired flavor of liquor and a target to be reached in the step, and based on this, the actual fermentation state Rewards are calculated based on the deviation from.

As illustrated, the reward determination unit 150 determines the location of the brewing type determination unit 151 that determines the master species to be brewed based on the brewing information received from the user, and the position of the preset brewing step through the current cumulative process. The brewing step determination unit 152, the target derivation unit 153 for calculating the target fermentation status information in the current brewing step, and the target fermentation status and fermentation status analysis unit 143 of the step derivation unit 153 It may include a target comparison determining unit 154 for calculating the reward by comparing the current fermentation status provided through.

Here, the brewing type determination unit 151 determines what kind of alcohol you want to obtain based on the brewing information (material and ratio, material state, yeast type, desired taste, body feeling, alcohol content, etc.) received from the user. This is for categorizing and classifying alcoholic beverages at a time rather than managing all kinds of alcoholic beverages at once. Can.

The brewing step determination unit 152 has a process process standard for a general brewing step for fermenting alcoholic beverages of a selected category and judges the current brewing step based on the analysis contents of the past fermentation status analysis unit 143 that has been continuously collected. . For example, it determines whether it is the initial stage of saccharification, the stage where saccharification and alcohol fermentation are started at the same time, the stage at which alcohol is produced through alcohol fermentation, or the stage of controlling the frequency through alcohol fermentation, and determines sugar content or flavor. Steps for adjustment may also be included.

The step-by-step goal derivation unit 153 derives information about a situation to be careful in, and a target saccharification and fermentation state to be reached in the current fermentation step.

The brewing step determination unit 152 and the information about the brewing step and the step-by-step target, which are the basis of the step-by-step goal derivation unit 153, are information previously stored in the database 170 based on the similar similar brewing process and the step-by-step target value. It may be information generated by an experienced user and stored in the database 170. In other words, in order to brew such liquor, these processes are required in the saccharification and fermentation process, and only the steps and targets of the approximate process that each process needs to reach a certain level of saccharification and fermentation are set. By collecting a large number of such data, it is possible to select and use the information closest to the brewing information input by the user.

However, in this case, it is difficult to cope with various variables such as the specific type and combination ratio of the material, the provided state, the mixing ratio with water, the type of yeast used, the desired aroma and flavor, body feeling, and the frequency of alcohol. Criteria mean that fermentation could potentially be controlled with the wrong goal.

Therefore, in the present invention, in order to accurately determine the reward, the corresponding brewing step determination unit 152 and the step-by-step target derivation unit 153 apply GAN (Generative Adversarial Network)-based intelligent network model 155 to correspond to various brewing information. It is possible to generate information about the fermentation process and glycation and fermentation targets to be reached in each process in close proximity to reality. In other words, this GAN-based intelligent network model is the actual brewing data for various variables such as the specific type and combination ratio of the material, the provided state, the mixing ratio with water, the type of yeast used, the desired aroma and flavor, body feeling, and alcohol content. B. By learning based on the information on the saccharification and fermentation process and fermentation target for each process presented by the skilled person, it is possible to generate new data that is closest to the actual data or expert data when the variables are changed.

In general, a GAN-based intelligent network model uses a pair of a generation model and a classification model to generate a realistic fake. The classification model sets the standard by learning the brewing success data or expert-provided data for each brewing information, and the generation model generates similar fake data and provides it to the classification model so that it can be classified as real data. Competitors develop with each other and provide real results that cannot be faked in the classification model. In particular, the GAN-based generated data operates in a direction to understand the concept, not simply a result of random combination (in the case of the currently implemented GAN image generating system, an image of the left face is provided and an image of the right face is generated). If you do, you can see the process of gradually turning to the right face while rotating the face in the step-by-step process). That is, if the variable is changed while learning the basic fermentation process and the fermentation state for each process, the values that reflect the set variable values are actually fermented instead of randomly generating similar values containing noise and making the closest one. It has the characteristics of the process and the target fermentation state, but the process and state information is operated in a form that gradually adjusts to the actual value in a slightly excessive state. In other words, even if new brewing information is provided and there is no actual brewing step or step-by-step target for it, applying the brewing step or step-by-step fermentation goal provided through the GAN-based intelligent network model enables professional-level brewing step setting and step-by-step fermentation goal setting do.

On the other hand, the target comparison determination unit 154 compares the target fermentation state with the current estimated fermentation state provided through the fermentation state analysis unit 143 to determine compensation for the deviation, but not simply using the size of the deviation In the case of the deviation reduction direction in the normal direction, more rewards are provided according to the reduction of the deviation, and if the deviation is abnormal, a penalty for minus the reward is imposed, and the reward is given more or less depending on the type of deviation occurrence. You can decide whether to give or not. Furthermore, it can be divided into long-term rewards (rewards depending on how much progress is made to reach the goal) and short-term rewards (rewards based on reaching short-term goals in the current stage). For example, if the long-term reward can be lowered even if the short-term reward is large, it is possible to provide a criterion for the reinforcement learning unit to operate so that the short-term reward is largest in an environment in which the long-term reward is increased.

The reinforcement learning unit 160 learns the collected state information (which may include all sensor data, image data, and brewing information, or may use only sensor data and image data when classifying to operate only for a specific category) In order to provide a control signal to the driving control unit 130, the optimum control policy neural network model for outputting the fitness for each temperature control as a probability value is applied, and the same state information as the execution unit 161 and the execution unit 161 are applied. A criterion 162 to which a value-calculated neural network model is applied to output expected compensation information (rewards) when the driving control unit 130 controls the fermentation environment according to the state information as input as a preset temperature. It may include.

Through this configuration, if the user provides the desired brewing information, and the fermentation target corresponding to the information is put into the fermentation tank and the fermentation is started, the desired flavor and alcoholic beverage can be automatically made, which is a success. As the amount of learning is repeatedly increased, full automation may be gradually possible.

However, in the case of the reinforcement learning unit, when the operation is performed only for a single control target, the learning is biased, and in a slightly different situation, a quality problem in which the control becomes abnormal may occur.

Therefore, it is necessary to be able to learn through various experiences, and since the original A3C deep reinforcement learning method targets computer games, copy these computer games to build various sub-environments with slightly different settings, and then collect different experiences. You can use the integration method. However, since this method cannot be used in an embodiment of the present invention targeting an actual fermentation tank, learning data for various control environments that each experience through interworking with reinforcement learning units configured in correspondence with fermentation tanks configured in various environments through a network. Apply configuration to share. However, if the main reinforcement learning unit accepts the learning data for each environment in a simple synchronous manner and updates the intelligent network and provides it to all the reinforcement learning units, it is difficult to obtain different experiences because the control operation is performed according to the same intelligent network. It works by mixing asynchronous intelligent network update and synchronized intelligent network update.

9 shows a communication network configuration capable of rapidly increasing the quality of artificial intelligence control while reducing the learning bias by expanding the reinforcement learning unit.

As shown, the main reinforcement learning unit 160_A which collects and reflects the learning results of the reinforcement learning unit through a communication network is constituted, and the main reinforcement learning unit 160_A is configured to correspond to a plurality of fermentation tanks distributed in various environments. Constructs a communication network with the learning units 160_1 to 160_n.

Each of the reinforcement learning units 160_1 to 160_n becomes a local reinforcement learning unit and learns while performing fermentation control for a fermentation tank that is linked to the reinforcement learning method to update its neural network, and updates the learning data to the main reinforcement learning unit 160_A To asynchronously. The main reinforcement learning unit 160_A receiving the learning data of a specific local reinforcement learning unit (for example, 160_1) in an asynchronous manner reflects the learning data and updates its neural network, and then displays the result of the specific local reinforcement learning unit. (Example: 160_1) and perform synchronization between the two. In this case, the corresponding local reinforcement learning unit (for example, 160_1) performs subsequent learning through the latest integrated neural network at the present time, and the other local reinforcement learning units 160_2 to 160_n each use synchronized neural networks at different times. You will have a different experience as you control the fermenter.

Subsequently, when another local reinforcement learning unit (for example, 160_4) provides its own learning data to the main reinforcement learning unit 160_A, the main reinforcement learning unit 160_A updates the neural network after receiving it and updates it again. It is synchronized with the local reinforcement learning unit (eg 160_4), and another local reinforcement learning unit (eg 160_4) controls the fermenter through the latest updated neural network.

As described above, the main reinforcement learning unit 160_A always maintains a neural network reflecting the latest learning information, but since each collected learning information is an experience obtained based on different neural networks, it is possible to reduce the problem of learning bias and increase control quality. There will be.

In the end, the more the mainstream fermentation system that applied the algorithm to the complex neural network including the intelligent network fermentation management unit corresponding to the fermentation tank is used, the higher the control performance becomes, and various types of success data are labeled and collected. The estimated quality is also increased, so that it is possible to increase the success rate of brewing by performing appropriate control before abnormal fermentation occurs while making sake of desired flavor. Therefore, ultimately, it is possible to successfully and automatically produce the desired liquor even if the type of fermentation target to be input is diverse, the yeast is diverse, and the environment is diverse, and in the same brewing conditions, even if there is material variation or environmental variation, a similar taste is obtained. It is possible to provide uniform and personalized alcohol production.

Meanwhile, the apparatus referred to in the present invention may be implemented with a hardware component, a software component, and/or a combination of a hardware component and a software component. The systems, devices and components described in embodiments include, for example, processors, controllers, arithmetic logic units (ALUs), digital signal processors (micro signal processors), microcomputers, field programmable arrays (FPAs), PLUs ( It may be implemented using one or more general purpose computers or special purpose computers, such as a programmable logic unit, microprocessor, or any other device capable of executing and responding to instructions. Further, such a server or device may run an operating system (OS) and one or more software applications running on the operating system. In addition, the device may access, store, manipulate, process, and generate data in response to the execution of the software. For convenience of understanding, a server or a device may be described as one being used, but a person having ordinary skill in the art, the processing device may include a plurality of processing elements and/or a plurality of types of processing. It can be seen that it may contain elements. For example, a server or device may include a plurality of processors or a processor and a controller. In addition, other processing configurations, such as parallel processors, are possible. In addition, the device can control various sensors directly through the interface or through a separate hardware/software board or module for controlling a driver or sensor, and can cover all various logical, electrical, and optical modifications for this. have.

Furthermore, in the embodiments of the present invention, devices for collecting data and components thereof may be devices including hardware or software for collecting information of physical and electrical sensors, and receiving and processing electrical signals. Interface, drive driver, communication device, power supply, microcontroller for control, storage, operating system or firmware, and may include sensors as needed.

In the above, preferred embodiments according to the present invention have been illustrated and described. However, the present invention is not limited to the above-described embodiments, and various modifications can be made to any person skilled in the art to which the present invention pertains without departing from the gist of the present invention appended in the claims. .

10: fermentation tank 11: motor unit
12: stirring unit 13: temperature control unit
14: extraction pipe 15: quartz pipe
16: temperature sensor unit 17: hydrometer
18: acidity measurement unit 19: carbon dioxide measurement unit
30: internal camera unit 40: microscope camera unit
100: intelligent network fermentation management unit 110: sensor data collection unit
120: image collection unit 130: drive control unit
140: multiple intelligence network learning unit 150: reward determination unit
160: reinforcement learning unit 170: database
180: user interface

Claims (6)

A mainstream fermentation system comprising a fermentation tank comprising a plurality of sensors, a camera, and a temperature control unit and an intelligent network fermentation management unit analyzing the fermentation status of the fermentation tank,
The intelligent network fermentation management unit,
A sensor data collection unit collecting numerical data through a plurality of information collection sensors configured in the fermenter;
An image collection unit for collecting a fermentation status image of a fermentation target of the fermenter;
A driving control unit for controlling the temperature of the fermenter;
A user interface unit providing a current fermentation status to a user and collecting brewing information including a fermentation target and a fermentation target from the user;
The current fermentation status is output by inputting the output of the first neural network and the first neural network that determines the fermentation status based on the collected image of the image collection unit, and the brewing information collected through the sensor data collection unit and the user interface unit. A mainstream fermentation system that applies an algorithm to a complex neural network including a multi-intelligent network learning unit.
The method according to claim 1, Reward determining unit for calculating a reward by comparing the current fermentation status information output by the multi-intelligent network learning unit with a target brewing step and a step-by-step target;
The execution information for optimally controlling the temperature of the fermenter using the output of the reward determination unit as a reward value and the information collected through the sensor data collection unit and the information obtained through the first neural network through the image collection unit as state information. Mainstream fermentation system using a complex neural network algorithm, characterized in that it comprises a reinforcement learning unit for generating and controlling the driving control unit.
The method according to claim 1, The reward determination unit
A brewing type determination unit determining a master species to be brewed based on the brewing information;
A brewing step determination unit for determining a position among preset brewing steps through a current cumulative process;
A step-by-step target derivation unit for calculating target fermentation status information in a current brewing step;
Mainstream fermentation system using a complex neural network algorithm, characterized in that it comprises a target comparison determining unit for calculating a reward by comparing the target fermentation status of the step-by-step target derivation unit and the current fermentation status provided through the fermentation status analysis unit.
The method according to claim 2, The reward determination unit is based on the brewing information provided by the user, based on the GAN (Generative Adversarial Network) that generates saccharification and fermentation processes and fermentation targets for each process through learning to obtain the desired taste and alcoholic alcohol. It includes an intelligent network model, and a complex neural network characterized by comparing the current fermentation status information of the fermentation status analyzer with respect to each process and targets generated through the GAN-based intelligent network model and calculating rewards based on the achievement level Alcoholic fermentation system using the algorithm.
The method according to claim 2, The reinforcement learning unit is an execution unit (actor) to which the optimal control policy neural network model is applied, which learns the collected state information and outputs a fitness value for each temperature control as a probability value for providing a control signal of the driving control unit. A criterion applied with a value calculation neural network model that outputs expected compensation information when the driving control unit controls the fermentation environment according to the state information by inputting the same state information as the negative. Mainstream fermentation system applying an algorithm to a complex neural network, characterized in that.
The method according to claim 5, It includes a main reinforcement learning unit for collecting and reflecting the learning results of the reinforcement learning unit through a communication network, the main reinforcement learning unit comprises a reinforcement learning unit corresponding to a plurality of fermentation tanks and a communication network, and a plurality of reinforcement learning Mainstream fermentation system applying a complex neural network algorithm, characterized in that each learning result is collected asynchronously from the department and synchronized with the reinforcement learning unit that provided the learning result.

Images