A method for controlling the fermentation of Daqu (a type of starter culture) and a Daqu fermentation system
By constructing a fermentation controller for Daqu (a type of starter culture) based on deep learning and fuzzy logic control, the problem of intelligent temperature and humidity control during the fermentation process was solved, achieving precise environmental parameter control, improving fermentation quality and stability, and reducing energy consumption.
Patent Information
- Application Number
- CN202411479564.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-23
- Publication Date
- 2026-01-30
- Estimated Expiration
- 2044-10-23
AI Technical Summary
In the current fermentation process of Daqu (a type of starter culture), temperature and humidity control is not intelligent, resulting in large differences in fermentation quality, making it impossible to guarantee uniformity and posing a health hazard to operators.
An adaptive weighted fusion algorithm based on deep learning and a fuzzy logic control algorithm, combined with neural networks and swarm intelligence optimization strategies, are used to construct a fermentation controller for koji (a type of starter culture). The controller precisely controls temperature and humidity through actuators, thereby achieving accurate adjustment of environmental parameters.
It improved fermentation quality, reduced energy consumption, ensured the stability of fermentation quality and the environment, and reduced labor intensity.
Smart Images
Figure CN119414900B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application belongs to the technical field of brewing and automatic control, and particularly relates to a Daqu fermentation control method and a Daqu fermentation system. BACKGROUND
[0002] As a solid-state brewing fermentation agent, Daqu plays a very key role in the brewing process. In the production of liquor, there is a saying that "the aroma is relied on distillation, and the aroma is relied on fermentation". The so-called aroma is that the active microorganisms and enzymes in the Daqu are mixed into the grain dregs containing starch after high-temperature cooking and gelatinization, participate in the fermentation process, and produce complex aroma and flavor substances. A large amount of heat is generated during the fermentation of Daqu, and the evaporation of water in Daqu is accompanied by the rise of temperature, and the concentrations of O2 and CO2 also change. Daqu plays a soul role in the brewing fermentation process, and the environment of the Daqu room, as the fermentation place of Daqu, greatly affects the quality of Daqu.
[0003] At present, the Daqu fermentation in some domestic distilleries still adopts the traditional Daqu room fermentation method, that is, the temperature and humidity and environmental parameters such as O2 and CO2 in the Daqu room during the fermentation process are controlled by artificial timed humidification, opening and closing of doors and windows, and turning of Daqu, so that the Daqu and the environment in the Daqu room need to be observed by the operating personnel, and then the doors and windows are opened and closed (to adjust the temperature, humidity, and moisture removal), the Daqu is turned multiple times, the Daqu is burned, the Daqu is collected, and the Daqu is covered with a straw curtain, and other operations are performed to control the environmental parameters during the fermentation of Daqu to achieve the best environmental state required for the fermentation of Daqu. The traditional Daqu fermentation production process not only relies on the subjective experience of the Daqu workshop operating personnel to make decisions, but also is prone to poor Daqu fermentation process control quality, and the fermentation quality of Daqu in different batches or different Daqu rooms of the same batch and different positions in the same Daqu room has great difference, so that the uniformity is difficult to guarantee. Moreover, the production environment during the fermentation of Daqu is poor, and the high-temperature and high-humidity environment is not conducive to the health of the workshop operating personnel.
[0004] In view of the above problems, at present, many researchers build intelligent koji room, such as koji fermentation workshop temperature and humidity monitoring system of a liquor factory, which uses DS18B20 and DHIH10N as temperature and humidity detection devices to form a temperature and humidity detection and regulation system based on CAN bus network and adopts PID control algorithm for system control, thereby improving the automation level and reducing the labor intensity. For example, a company proposes digitalization in the production process of Daqu, applies a digital control system to its own liquor factory, collects, analyzes and processes data of environmental parameters such as temperature, humidity and carbon dioxide, and performs digital and automatic opening and closing operation on the windows of the koji room according to the results, thereby changing the traditional management mode of koji room koji making and realizing the leap to digital management. For example, a group successfully applies automatic opening and closing of doors and windows to Daqu fermentation koji room for the first time, detects parameters such as temperature and humidity of the koji core and the environment in the koji room, and automatically opens the doors and windows for ventilation when the data exceeds the limited value, thereby cooling the koji room, maintaining stable fermentation of Daqu and ensuring the fermentation quality of Daqu.
[0005] The above research results promote the industrial intelligentization of Daqu fermentation through modern scientific technology, but all have the problem of not combining with the actual characteristics of the koji room and the temperature and humidity characteristics of the Daqu fermentation environment, and the intelligentization of temperature and humidity regulation and control in the Daqu fermentation process has not been realized, and the quality and uniformity of Daqu fermentation cannot be guaranteed. SUMMARY
[0006] The present application aims to provide a Daqu fermentation control method and a Daqu fermentation system, which adopts a self-adaptive weighted fusion algorithm based on deep learning to fuse temperature detection data and detection data, and then proposes a Daqu fermentation controller based on a group intelligence optimization strategy through a fuzzy logic control algorithm and a neural network, so that the optimized Daqu fermentation controller can meet the control requirements of the Daqu fermentation system on temperature and humidity, realize more accurate start and stop time of the actuator, ensure the accuracy of the koji room environment temperature and humidity control and good following of the environment temperature and humidity, improve the fermentation quality, reduce the energy consumption and ensure the fermentation quality.
[0007] The technical scheme adopted by the present application is:
[0008] A Daqu fermentation system, comprising:
[0009] a koji room;
[0010] an actuator, which is dispersedly arranged in the koji room; wherein the actuator comprises a heater, a humidifier, an electric window, an internal circulating fan and an external circulating fan;
[0011] a temperature sensor, which is dispersedly arranged in the koji room and collects temperature detection data of different monitoring points in the koji room;
[0012] A humidity sensor is arranged in the koji room in a scattered manner to collect humidity detection data of different monitoring points in the koji room.
[0013] A koji fermentation controller is constructed based on a fuzzy logic control algorithm and a neural network fusion, and is optimized by using a swarm intelligence optimization strategy; the optimized koji fermentation controller uses a deep learning-based adaptive weighted fusion algorithm to perform data fusion processing on the temperature detection data and the humidity detection data, and then outputs a temperature deviation and a temperature deviation change rate of an actual koji room temperature value and a target temperature value, and a humidity deviation and a humidity deviation change rate of an actual koji room humidity value and a target humidity value as input signals, and outputs an opening time of each execution mechanism as an output signal; wherein the temperature deviation change rate is a ratio of the temperature deviation to the target temperature value, and the humidity deviation change rate is a ratio of the humidity deviation to the target humidity value.
[0014] A host computer receives all the output signals, and converts each output signal into a corresponding Boolean signal, and then controls the execution of the corresponding execution mechanism according to the Boolean signal to start or stop operation.
[0015] Further, the koji fermentation controller includes an antecedent network and a consequent network.
[0016] The first layer of the antecedent network is composed of an input layer, and each node in the input layer is connected to each component of an input vector and transmitted to the next layer; the input vector is the input signal.
[0017] The second layer of the antecedent network is composed of a fuzzification layer, and the input layer is data-fuzzed by the fuzzification layer through a Gaussian membership function.
[0018] The third layer of the antecedent network is composed of a rule layer, and the rule layer is composed of 49*49 neuron nodes, each neuron node of the rule layer representing a fuzzy rule.
[0019] The fourth layer of the antecedent network is composed of a normalization layer, which is used for normalization calculation and serves as a connection weight of the third layer of the consequent network; the neuron node number of the normalization layer is 49*49.
[0020] The first layer of the consequent network is composed of the input layer, and the input vector is transmitted to the next layer.
[0021] The second layer of the consequent network consists of a rule calculation layer, which is composed of 49*49 neuron nodes. Each neuron node in the rule calculation layer represents a calculation rule. The rule calculation layer is used to calculate the consequent of the fuzzy rule.
[0022] The third layer of the consequent network consists of an output layer, which performs a weighted sum of the output of the normalization layer and the output of the rule calculation layer.
[0023] Furthermore, the host computer also has a visualization module; the visualization module is used to visualize and display chart data.
[0024] Based on the same inventive concept, the present invention also provides a method for controlling the fermentation of Daqu (a type of starter culture), comprising the following steps:
[0025] Step S1: Construct a fermentation controller for Daqu (a type of starter culture) based on the fusion of fuzzy logic control algorithm and neural network, and optimize the Daqu fermentation controller using a swarm intelligence optimization strategy;
[0026] Step S2: Collect temperature and humidity data from different monitoring points within the cursive room, and use a deep learning-based adaptive weighted fusion algorithm to fuse the temperature and humidity data to obtain the fused actual temperature and humidity values of the cursive room. Then, output the temperature deviation and rate of change between the actual temperature and the target temperature, and output the humidity deviation and rate of change between the actual humidity and the target humidity. The rate of change of temperature deviation is the ratio of the temperature deviation to the target temperature value, and the rate of change of humidity deviation is the ratio of the humidity deviation to the target humidity value.
[0027] Step S3: The temperature deviation and the rate of change of the temperature deviation, as well as the humidity deviation and the rate of change of the humidity deviation, are used as input signals to be imported into the optimized Daqu fermentation controller for learning and training, and the opening time of each actuator is used as the output signal; wherein, the actuator includes a heater, a humidifier, an electric window, an internal circulation fan and an external circulation fan;
[0028] Step S4: The host computer receives all the output signals and converts each output signal into a corresponding Boolean signal. Then, the host computer controls the corresponding actuator to start or stop based on the Boolean signal.
[0029] Furthermore, the specific process of optimizing the Daqu fermentation controller using a swarm intelligence optimization strategy in step S1 is as follows:
[0030] Step S11: Initialize the Daqu fermentation controller; wherein, the Daqu fermentation controller includes a preprocessor network and a postprocessor network;
[0031] Suppose there are m individuals in a D-dimensional search space, where each individual represents a parameter combination of the antecedent network and the consequent network. In the swarm intelligence optimization strategy, the fitness of each individual is the difference between the actual start-stop time of the actuator and the expected start-stop time. The position and velocity of each individual are updated based on its fitness. The individual's velocity and position are updated using the following formula based on its individual extreme value and the swarm extreme value:
[0032] ;
[0033] ;
[0034] In the formula, This represents the speed of the i-th individual in the j-th dimension of the D-dimensional search space during the (k+1)-th iteration of the swarm intelligence optimization strategy;
[0035] This represents the speed of the i-th individual in the k-th iteration of the swarm intelligence optimization strategy in the j-th dimension of the D-dimensional search space;
[0036] Indicates inertia weight;
[0037] c1 and c2 represent acceleration constants;
[0038] r1 and r2 represent random numbers in the interval [0, 1].
[0039] This represents the optimal position of the i-th individual in the j-th dimension of the D-dimensional search space during the k-th iteration of the swarm intelligence optimization strategy.
[0040] This represents the position of the i-th individual in the j-th dimension of the D-dimensional search space during the k-th iteration of the swarm intelligence optimization strategy;
[0041] This represents the position of the i-th individual in the j-th dimension of the D-dimensional search space during the (k+1)-th iteration of the swarm intelligence optimization strategy.
[0042] This represents the global optimal position of the i-th individual in the k-th iteration of the swarm intelligence optimization strategy in the D-dimensional search space;
[0043] Step S12: Preset the maximum number of iterations for the swarm intelligence optimization strategy, and evaluate the fitness of each individual after each iteration; after reaching the preset maximum number of iterations, terminate the iteration and execute the next step;
[0044] Step S13: Preset the maximum number of iterations for the Daqu fermentation controller, select the optimal value of the individual as the combined parameter weight of the antecedent network and the consequent network and import it into the Daqu fermentation controller; after reaching the preset maximum number of iterations, terminate the iteration and obtain the optimized Daqu fermentation controller.
[0045] Furthermore, in step S2, before performing the fusion process, the collected temperature detection data or humidity detection data is first subjected to a consistency check.
[0046] Furthermore, the specific process of performing a consistency check on the temperature detection data in step S2 is as follows:
[0047] Determine one by one whether the difference between the temperature detection data currently collected by the same temperature sensor and the temperature detection data collected at the previous moment is less than or equal to the corresponding temperature monitoring threshold.
[0048] If the temperature is less than or equal to the corresponding temperature monitoring threshold, then the temperature detection data currently collected by the same temperature sensor will undergo subsequent fusion processing.
[0049] If the temperature exceeds the corresponding temperature monitoring threshold, the average temperature value of all temperature detection data collected by the same temperature sensor is used for subsequent fusion processing.
[0050] The specific process of performing a consistency check on the humidity detection data in step S2 is as follows:
[0051] Determine one by one whether the difference between the humidity detection data currently collected by the same humidity sensor and the humidity detection data collected at the previous moment is less than or equal to the corresponding humidity monitoring threshold.
[0052] If the humidity is less than or equal to the corresponding humidity monitoring threshold, then the humidity detection data currently collected by the same humidity sensor will undergo subsequent fusion processing.
[0053] If the humidity exceeds the corresponding humidity monitoring threshold, the average humidity value of all humidity detection data collected by the same humidity sensor is used for subsequent fusion processing.
[0054] Furthermore, in step S2, the temperature detection data undergoes data fusion processing to obtain the fused actual temperature value of the flute. It is obtained through the following formula:
[0055] ;
[0056] ;
[0057] ;
[0058] In the formula, This indicates the actual temperature value of the fused cursive chamber;
[0059] This represents the latest temperature detection data of the p-th temperature sensor out of n temperature sensors, or the average temperature detection data of all the p-th temperature sensors.
[0060] This represents the variance of the temperature detection data of the p-th temperature sensor out of n temperature sensors;
[0061] This represents the average value of temperature data detected by n temperature sensors.
[0062] Similarly, the process of fusing the humidity detection data in step S2 is the same as the process of fusing the temperature detection data.
[0063] Furthermore, during the learning and training process in the Daqu fermentation controller in step S3, the backpropagation algorithm is used to change the weights of the second layer of the consequent network in the Daqu fermentation controller, and the center value and width of the Gaussian membership function used in the second layer of the antecedent network in the Daqu fermentation controller are adjusted, so that the Daqu fermentation controller can stably track and optimize the target value.
[0064] Furthermore, the specific process of learning and training in the Daqu fermentation controller in step S3 is as follows:
[0065] The learning algorithm for the weights of the second layer of the consequent network is as follows:
[0066] ;
[0067] ;
[0068] In the formula, The loss function;
[0069] Let be the weight change of the k-th output and the ith input in the p-th computation rule of the consequent network;
[0070] This is the output of the p-th computation rule in the consequent network;
[0071] This is the output of the k-th neuron node in the p-th computation rule of the consequent network;
[0072] Let p be the expected output of the p-th computation rule in the consequent network;
[0073] Let be the membership degree of the j-th fuzzy set;
[0074] Let i be the component of the i-th input vector;
[0075] r is the learning rate;
[0076] l represents the number of iterations of the Daqu fermentation controller;
[0077] In parameter value Once determined, the impact of the error on each layer of the predecessor network is then verified, and the derivation formula is as follows;
[0078] ;
[0079] ;
[0080] ;
[0081] In the formula, This represents the error term of the i-th neuron node in the fifth layer of the preceding network; the fifth layer of the preceding network consists of an output layer.
[0082] This represents the error term of the k-th neuron node in the fourth layer of the antecedent network;
[0083] This represents the error term of the k-th neuron node in the third layer of the antecedent network;
[0084] t i This represents the expected output of the i-th output node;
[0085] y i This represents the actual output of the i-th output node;
[0086] y ik This represents the connection weight between the k-th neuron in the fourth layer of the preceding network and the i-th neuron in the fifth layer.
[0087] a i This represents the membership degree of the i-th input variable;
[0088] The final result is:
[0089] ;
[0090] ;
[0091] In the formula, E is the loss function;
[0092] This represents the error term of the i-th neuron node in the k-th layer;
[0093] x i This represents the value of the i-th input variable;
[0094] c ik This represents the center value of the k-th Gaussian membership function corresponding to the i-th input used in the second layer of the preceding network;
[0095] c jk This represents the center value of the k-th Gaussian membership function corresponding to the j-th input used in the second layer of the preceding network;
[0096] q ik This represents the width of the k-th Gaussian membership function corresponding to the i-th input used in the second layer of the preamble network;
[0097] The second layer of the preamble network uses the center value c of the k-th Gaussian membership function corresponding to the i-th input. ik and width value q ik The learning rules are:
[0098] ;
[0099] ;
[0100] In the formula, This represents the center value of the k-th Gaussian membership function corresponding to the i-th input used in the second layer of the antecedent network during the (l+1)-th iteration of the fermentation controller.
[0101] This represents the center value of the k-th Gaussian membership function corresponding to the i-th input used in the second layer of the antecedent network during the l-th iteration of the fermentation controller.
[0102] This represents the width of the k-th Gaussian membership function corresponding to the i-th input used in the second layer of the preamble network during the (l+1)-th iteration of the fermentation controller.
[0103] This represents the width of the k-th Gaussian membership function corresponding to the i-th input used in the second layer of the antecedent network during the l-th iteration of the fermentation controller.
[0104] r is the learning rate;
[0105] E is the loss function.
[0106] The beneficial effects of this invention are:
[0107] This invention provides a method and system for controlling the fermentation of Daqu (a type of starter culture). It employs a deep learning-based adaptive weighted fusion algorithm to fuse temperature and humidity data. Then, by fusing fuzzy logic control and neural networks, a Daqu fermentation controller based on a swarm intelligence optimization strategy is proposed. This controller defines the correlation parameters between the input conditions and output results of the fermentation process as individuals in the search space, and performs global optimization on the antecedent and consequent networks of the Daqu fermentation network model. It combines the actual characteristics of the fermentation chamber with the temperature and humidity characteristics of the Daqu fermentation environment, applying this approach to the fermentation process. The optimized Daqu fermentation controller meets the temperature and humidity control requirements of the Daqu fermentation system, enabling more precise start-stop times for actuators such as internal and external circulation fans, electric windows, heaters, and humidifiers. This ensures accurate temperature and humidity control in the fermentation chamber and good temperature and humidity tracking, improving fermentation quality, reducing energy consumption, and guaranteeing fermentation quality. Attached Figure Description
[0108] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0109] Figure 1 This is a front view of the curved room in Example 1.
[0110] Figure 2 This is a top view of the curved room in Example 1.
[0111] Figure 3 This is a schematic diagram of the fermentation control method of Daqu in Example 2.
[0112] Figure 4 This is a schematic diagram of the process of optimizing the fermentation controller of Daqu using a swarm intelligence optimization strategy in Example 2.
[0113] Figure 5 This is a schematic diagram of the processing flow of temperature detection data or humidity detection data in Example 2.
[0114] The attached figures are labeled as follows:
[0115] 100 - Perforated plate, 200 - Electric window, 300 - Humidifier, 400 - External circulation fan, 500 - Internal circulation fan, 600 - Heater, 700 - Bend, 800 - Temperature sensor, 900 - Humidity sensor. Detailed Implementation
[0116] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.
[0117] The following disclosure provides many different embodiments or examples for implementing different structures of the present invention. To simplify the disclosure of the present invention, the components and arrangements of specific examples are described below. Of course, these are merely examples and are not intended to limit the present invention.
[0118] The embodiments of the invention will now be described in detail with reference to the accompanying drawings.
[0119] Example 1
[0120] A fermentation system for traditional Chinese koji (a type of starter culture) includes: a fermentation room, an actuator, a temperature sensor, a humidity sensor, and a host computer.
[0121] The actuators are distributed throughout the room; the actuators include a heater, a humidifier, an electric window, an internal circulation fan, and an external circulation fan.
[0122] The temperature sensors are distributed throughout the curing chamber to collect temperature data at different monitoring points. The temperature accuracy and range are: ±0.3℃, 0℃~80℃.
[0123] The humidity sensors are distributed throughout the cursive room to collect humidity data at different monitoring points. The humidity accuracy and range are: ±3% RH, 10%RH to 100%RH.
[0124] The fermentation controller for Daqu (a type of starter culture) is constructed based on a fuzzy logic control algorithm and a neural network, and optimized using a swarm intelligence optimization strategy. The optimized controller uses a deep learning-based adaptive weighted fusion algorithm to fuse the temperature and humidity detection data. The output signals are the temperature deviation and its rate of change between the actual and target temperatures in the fermentation chamber, and the humidity deviation and its rate of change between the actual and target humidity values. The output signal is the start time of each actuator. The temperature deviation rate is the ratio of the temperature deviation to the target temperature value, and the humidity deviation rate is the ratio of the humidity deviation to the target humidity value.
[0125] The host computer receives all the output signals and converts each output signal into a corresponding Boolean signal. The host computer then controls the corresponding actuator to start or stop operations based on the Boolean signals. Simultaneously, the host computer also has a visualization module; this module is used to visualize chart data, such as temperature and humidity curves for each monitoring point.
[0126] Figure 1 This is a front view of the curved room in Example 1. Figure 2 This is a top view of the curved room in Example 1. For example... Figure 1 and Figure 2 The layout of the cursive room shown in the diagram includes four perforated plates 100 on the left and right side walls, an electric window 200 on each side of the rear wall and in the middle of the front wall, a door in the middle of the rear wall, a humidifier 300 between the electric window 200 and the door, an external circulation fan 400 and an internal circulation fan 500 on each side of the front wall, and a heater 600 between the external circulation fan 400 and the electric window 200 on the front wall. The cursive room contains four cursive racks 700; each rack 700 has three layers, with 16 large cursive pieces per layer, totaling 192 pieces. Temperature sensors 800 and humidity sensors 900 are evenly distributed across the upper, middle, and lower layers of the cursive room, with five monitoring points on each layer; each temperature sensor 800 and humidity sensor 900 records temperature and humidity data every 6 minutes.
[0127] During the fermentation control of Daqu, the theoretical temperature and humidity curve of Daqu fermentation is used as the control target. The temperature and humidity data of the fermentation room environment after adaptive weighted fusion algorithm is used as the input data of Daqu fermentation controller. The output realizes the start and stop time of actuators such as internal circulation fan, external circulation fan, electric window, heater, and humidifier. The host computer controls the corresponding actuators to start or stop.
[0128] The fermentation controller for Daqu (a type of starter culture) includes a preprocessor network and a postprocessor network.
[0129] The first layer of the preamble network consists of an input layer, where each node is connected to a component of the input vector and transmitted to the next layer. The input vector is the input signal, namely the temperature deviation and its rate of change, as well as the humidity deviation and its rate of change. The second layer of the preamble network consists of a fuzzification layer, which fuzzifies the input data using a Gaussian membership function. The third layer of the preamble network consists of a rule layer, which has 49*49 neurons, each representing a fuzzy rule. The fourth layer of the preamble network consists of a normalization layer, which performs normalization calculations and serves as the connection weights for the third layer of the consequent network. The normalization layer has 49*49 neurons.
[0130] The first layer of the consequent network consists of the input layer, which transmits the input vector to the next layer. The second layer of the consequent network consists of a rule calculation layer, which is composed of 49*49 neuron nodes, each neuron node of the rule calculation layer representing a calculation rule. The rule calculation layer is used to calculate the consequent of the fuzzy rule. The third layer of the consequent network consists of an output layer, which performs a weighted summation of the output of the normalization layer and the output of the rule calculation layer.
[0131] Example 2
[0132] Figure 3 This is a schematic diagram of the fermentation control method for Daqu (a type of starter culture) in Example 2. Figure 3 As shown, the fermentation control method for Daqu (a type of starter culture) includes the following steps:
[0133] Step S1: Construct a fermentation controller for Daqu (a type of starter culture) based on the fusion of fuzzy logic control algorithm and neural network, and optimize the Daqu fermentation controller using a swarm intelligence optimization strategy;
[0134] Step S2: Collect temperature and humidity data from different monitoring points within the cursive room, and use a deep learning-based adaptive weighted fusion algorithm to fuse the temperature and humidity data to obtain the fused actual temperature and humidity values of the cursive room. Then, output the temperature deviation and rate of change between the actual temperature and the target temperature, and output the humidity deviation and rate of change between the actual humidity and the target humidity. The rate of change of temperature deviation is the ratio of the temperature deviation to the target temperature value, and the rate of change of humidity deviation is the ratio of the humidity deviation to the target humidity value.
[0135] Step S3: The temperature deviation and the rate of change of the temperature deviation, as well as the humidity deviation and the rate of change of the humidity deviation, are used as input signals to be imported into the optimized Daqu fermentation controller for learning and training, and the opening time of each actuator is used as the output signal; wherein, the actuator includes a heater, a humidifier, an electric window, an internal circulation fan and an external circulation fan;
[0136] Step S4: The host computer receives all the output signals and converts each output signal into a corresponding Boolean signal. Then, the host computer controls the corresponding actuator to start or stop based on the Boolean signal.
[0137] The fermentation control method in this embodiment employs a deep learning-based adaptive weighted fusion algorithm to fuse temperature and humidity data. Then, by fusing fuzzy logic control and neural networks, a swarm intelligence optimization strategy-based fermentation controller is proposed. This controller defines the correlation parameters between the input conditions and output results of the fermentation process as individuals in the search space, performing global optimization on the antecedent and consequent networks of the fermentation network model. It combines the actual characteristics of the fermentation chamber with the temperature and humidity characteristics of the fermentation environment, applying this approach to the fermentation process. The optimized fermentation controller meets the temperature and humidity control requirements of the fermentation system, enabling more precise start-stop times for actuators such as internal and external circulation fans, electric windows, heaters, and humidifiers. This ensures accurate temperature and humidity control in the fermentation chamber and good temperature and humidity tracking, improving fermentation quality, reducing energy consumption, and guaranteeing fermentation quality.
[0138] It should be noted that the order of steps S1 and S2 in this embodiment can be reversed or the two steps can be executed simultaneously.
[0139] In this embodiment, the fermentation controller based on the fusion of fuzzy logic control algorithm and neural network includes an antecedent network and a consequent network. The first layer of the antecedent network consists of an input layer, where each node is connected to a component of the input vector and transmitted to the next layer. The input vector is the input signal, namely temperature deviation and its rate of change, as well as humidity deviation and its rate of change. The second layer of the antecedent network consists of a fuzzification layer, which fuzzifies the input layer data using a Gaussian membership function. The third layer of the antecedent network consists of a rule layer, which comprises 49*49 neurons, each representing a fuzzy rule. The fourth layer of the antecedent network consists of a normalization layer, which performs normalization calculations and serves as the connection weight for the third layer of the consequent network. The normalization layer has 49*49 neurons.
[0140] The first layer of the consequent network consists of the input layer, which transmits the input vector to the next layer. The second layer of the consequent network consists of a rule calculation layer, which is composed of 49*49 neuron nodes, each neuron node of the rule calculation layer representing a calculation rule. The rule calculation layer is used to calculate the consequent of the fuzzy rule. The third layer of the consequent network consists of an output layer, which performs a weighted summation of the output of the normalization layer and the output of the rule calculation layer.
[0141] Figure 4 This is a schematic diagram of the process for optimizing the fermentation controller of Daqu (a type of starter culture) using a swarm intelligence optimization strategy in Example 2. Figure 4 As shown in the figure, the specific process of optimizing the Daqu fermentation controller using a swarm intelligence optimization strategy in step S1 is as follows:
[0142] Step S11: Initialize the Daqu fermentation controller; wherein, the Daqu fermentation controller includes a preprocessor network and a postprocessor network;
[0143] Suppose there are m individuals in a D-dimensional search space, where each individual represents a parameter combination of the antecedent network and the consequent network. In the swarm intelligence optimization strategy, the fitness of each individual is the difference between the actual start-stop time of the actuator and the expected start-stop time. The position and velocity of each individual are updated based on its fitness. The individual's velocity and position are updated using the following formula based on its individual extreme value and the swarm extreme value:
[0144] ;
[0145] ;
[0146] In the formula, This represents the speed of the i-th individual in the j-th dimension of the D-dimensional search space during the (k+1)-th iteration of the swarm intelligence optimization strategy;
[0147] This represents the speed of the i-th individual in the k-th iteration of the swarm intelligence optimization strategy in the j-th dimension of the D-dimensional search space;
[0148] Indicates inertia weight;
[0149] c1 and c2 represent acceleration constants;
[0150] r1 and r2 represent random numbers in the interval [0, 1].
[0151] This represents the optimal position of the i-th individual in the j-th dimension of the D-dimensional search space during the k-th iteration of the swarm intelligence optimization strategy.
[0152] This represents the position of the i-th individual in the j-th dimension of the D-dimensional search space during the k-th iteration of the swarm intelligence optimization strategy;
[0153] This represents the position of the i-th individual in the j-th dimension of the D-dimensional search space during the (k+1)-th iteration of the swarm intelligence optimization strategy.
[0154] This represents the global optimal position of the i-th individual in the k-th iteration of the swarm intelligence optimization strategy in the D-dimensional search space;
[0155] Step S12: Preset the maximum number of iterations for the swarm intelligence optimization strategy, and evaluate the fitness of each individual after each iteration; after reaching the preset maximum number of iterations, terminate the iteration and execute the next step; wherein, the smaller the fitness, the closer it is to the optimal solution;
[0156] Step S13: Preset the maximum number of iterations for the Daqu fermentation controller, select the optimal value of the individual as the combined parameter weight of the antecedent network and the consequent network and import it into the Daqu fermentation controller; after reaching the preset maximum number of iterations, terminate the iteration and obtain the optimized Daqu fermentation controller.
[0157] Figure 5 This is a schematic diagram of the processing flow for temperature or humidity detection data in Example 2. Figure 5 As shown in the figure, in step S2 of this embodiment, before performing the fusion process, the collected temperature detection data or humidity detection data is first subjected to consistency detection to improve the accuracy of data fusion.
[0158] More specifically, the process of performing consistency detection on the temperature detection data in step S2 is as follows: It is determined whether the difference between the temperature detection data currently collected by the same temperature sensor and the temperature detection data collected at the previous moment is less than or equal to the corresponding temperature monitoring threshold; if it is less than or equal to the corresponding temperature monitoring threshold, subsequent fusion processing is performed on the temperature detection data currently collected by the same temperature sensor; if it is greater than the corresponding temperature monitoring threshold, the average temperature value of all temperature detection data collected by the same temperature sensor is taken and subsequent fusion processing is performed.
[0159] The specific process of performing consistency detection on the humidity detection data in step S2 is as follows: determine whether the difference between the humidity detection data currently collected by the same humidity sensor and the humidity detection data collected at the previous moment is less than or equal to the corresponding humidity monitoring threshold; if it is less than or equal to the corresponding humidity monitoring threshold, perform subsequent fusion processing on the humidity detection data currently collected by the same humidity sensor; if it is greater than the corresponding humidity monitoring threshold, take the average humidity detection value of all humidity detection data collected by the same humidity sensor and perform subsequent fusion processing.
[0160] For example, in the fermentation room of Example 1, a total of 15 temperature sensors are set up. Each temperature sensor detects the environment every 10 seconds during the fermentation process, forming a continuous temperature detection data. The method to determine the consistency of this data is to determine whether the difference between the latest detection value and the previous detection value of the same temperature sensor is less than or equal to the temperature monitoring threshold. Similarly, the detection data of the 15 humidity sensors set up in the fermentation room are processed in a similar way.
[0161] In this embodiment, the temperature detection data is fused in step S2 to obtain the fused actual temperature value of the flute room using the following formula:
[0162] ;
[0163] ;
[0164] ;
[0165] In the formula, This indicates the actual temperature value of the fused cursive chamber;
[0166] This represents the latest temperature detection data of the p-th temperature sensor out of n temperature sensors, or the average temperature detection data of all the p-th temperature sensors.
[0167] This represents the variance of the temperature detection data of the p-th temperature sensor out of n temperature sensors;
[0168] This represents the average value of temperature data detected by n temperature sensors.
[0169] Similarly, the process of fusing the humidity detection data in step S2 is the same as the process of fusing the temperature detection data.
[0170] In this embodiment, during the learning and training process of the Daqu fermentation controller in step S3, the backpropagation algorithm is used to change the weights of the second layer of the consequent network in the Daqu fermentation controller, and to adjust the center value and width of the Gaussian membership function used in the second layer of the antecedent network in the Daqu fermentation controller, so that the Daqu fermentation controller can stably track and optimize the target value.
[0171] Furthermore, the specific process of learning and training in the Daqu fermentation controller in step S3 is as follows:
[0172] The learning algorithm for the weights of the second layer of the consequent network is as follows:
[0173] ;
[0174] ;
[0175] In the formula, The loss function;
[0176] Let be the weight change of the k-th output and the ith input in the p-th computation rule of the consequent network;
[0177] This is the output of the p-th computation rule in the consequent network;
[0178] This is the output of the k-th neuron node in the p-th computation rule of the consequent network;
[0179] Let p be the expected output of the p-th computation rule in the consequent network;
[0180] Let be the membership degree of the j-th fuzzy set;
[0181] Let i be the component of the i-th input vector;
[0182] r is the learning rate;
[0183] l represents the number of iterations of the Daqu fermentation controller;
[0184] In parameter value Once determined, the impact of the error on each layer of the predecessor network is then verified, and the derivation formula is as follows;
[0185] ;
[0186] ;
[0187] ;
[0188] In the formula, This represents the error term of the i-th neuron node in the fifth layer of the preceding network; the fifth layer of the preceding network consists of an output layer.
[0189] This represents the error term of the k-th neuron node in the fourth layer of the antecedent network;
[0190] This represents the error term of the k-th neuron node in the third layer of the antecedent network;
[0191] t i This represents the expected output of the i-th output node;
[0192] y i This represents the actual output of the i-th output node;
[0193] y ik This represents the connection weight between the k-th neuron in the fourth layer of the preceding network and the i-th neuron in the fifth layer.
[0194] a i This represents the membership degree of the i-th input variable;
[0195] The final result is:
[0196] ;
[0197] ;
[0198] In the formula, E is the loss function;
[0199] This represents the error term of the i-th neuron node in the k-th layer;
[0200] x i This represents the value of the i-th input variable;
[0201] c ik This represents the center value of the k-th Gaussian membership function corresponding to the i-th input used in the second layer of the preceding network;
[0202] c jk This represents the center value of the k-th Gaussian membership function corresponding to the j-th input used in the second layer of the preceding network;
[0203] q ik This represents the width of the k-th Gaussian membership function corresponding to the i-th input used in the second layer of the preamble network;
[0204] The second layer of the preamble network uses the center value c of the k-th Gaussian membership function corresponding to the i-th input. ikand width value q ik The learning rules are:
[0205] ;
[0206] ;
[0207] In the formula, This represents the center value of the k-th Gaussian membership function corresponding to the i-th input used in the second layer of the antecedent network during the (l+1)-th iteration of the fermentation controller.
[0208] This represents the center value of the k-th Gaussian membership function corresponding to the i-th input used in the second layer of the antecedent network during the l-th iteration of the fermentation controller.
[0209] This represents the width of the k-th Gaussian membership function corresponding to the i-th input used in the second layer of the preamble network during the (l+1)-th iteration of the fermentation controller.
[0210] This represents the width of the k-th Gaussian membership function corresponding to the i-th input used in the second layer of the antecedent network during the l-th iteration of the fermentation controller.
[0211] r is the learning rate;
[0212] E is the loss function.
Claims
1. A koji fermentation control method characterized by, The method comprises the following steps: Step S1, a Daqu fermentation controller based on fuzzy logic control algorithm and neural network fusion is constructed, and a colony intelligence optimization strategy is used to optimize the Daqu fermentation controller. The correlation parameters of input conditions and output results of the fermentation process are defined as individuals in the search space. The global optimization of the antecedent network and the consequent network of the Daqu fermentation network model is performed. The actual characteristics of the Daqu room and the temperature and humidity characteristics of the Daqu fermentation environment are combined and applied in the Daqu fermentation process. Step S2, temperature detection data and humidity detection data of different monitoring points in the Daqu room are collected. The temperature detection data and humidity detection data are processed by using a self-adaptive weighted fusion algorithm based on deep learning to obtain the actual temperature value and the actual humidity value of the Daqu room. Then, the temperature deviation and the temperature deviation change rate of the actual temperature value of the Daqu room and the target temperature value are outputted, and the humidity deviation and the humidity deviation change rate of the actual humidity value of the Daqu room and the target humidity value are outputted. The temperature deviation change rate is the ratio of the temperature deviation to the target temperature value, and the humidity deviation change rate is the ratio of the humidity deviation to the target humidity value. Step S3, the temperature deviation and the temperature deviation change rate, and the humidity deviation and the humidity deviation change rate are taken as input signals and introduced into the optimized Daqu fermentation controller for learning and training, and the opening time of each execution mechanism is taken as an output signal. The execution mechanism includes a heater, a humidifier, an electric window, an internal circulating fan and an external circulating fan. Step S4, all the output signals are received by the upper computer, and the output signals are converted into corresponding Boolean signals one by one. Then, the upper computer controls the corresponding execution mechanism to start or stop according to the Boolean signals. The Daqu fermentation controller comprises an antecedent network and a consequent network. In the learning and training of the Daqu fermentation controller in step S3, the back propagation algorithm is used to change the weight value of the second layer of the consequent network in the Daqu fermentation controller, and the center value and the width of the Gaussian membership function used in the second layer of the antecedent network in the Daqu fermentation controller are adjusted, so that the Daqu fermentation controller can stably track the optimized target value.
2. The Daqu fermentation control method according to claim 1, characterized in that, The specific process of optimizing the Daqu fermentation controller by using the colony intelligence optimization strategy in step S1 is as follows: Step S11, the Daqu fermentation controller is initialized. There are m individuals in the D-dimensional search space, and the specific meaning of each individual is the parameter combination of the antecedent network and the consequent network. The fitness of each individual in the colony intelligence optimization strategy is the difference between the actual output of the execution mechanism and the expected start-stop time. The position and speed of each individual are updated according to the fitness of each individual. The speed and position of each individual are updated by using the following formula: ； ； In the formula, represents the speed of the ith individual in the jth dimension of the D-dimensional search space at the k+1th iteration of the swarm intelligence optimization strategy. vi,j,k represents the velocity of the ith individual in the jth dimension of the D-dimensional search space at the kth iteration of the swarm intelligence optimization strategy; denotes the inertial weight; c1 and c2 represent acceleration constants; r1 and r2 represent random numbers in the interval [0, 1]. Xi,j,k represents the individual optimal position of the ith individual in the jth dimension of the D-dimensional search space at the kth iteration of the swarm intelligence optimization strategy; Xi,j,k represents the position of the ith individual in the jth dimension of the D-dimensional search space at the kth iteration of the swarm intelligence optimization strategy; Xi k+1j represents the position of the ith individual in the jth dimension of the D-dimensional search space at the k+1th iteration of the swarm intelligence optimization strategy; Xi k represents the global optimal position of the ith individual at the kth iteration of the swarm intelligence optimization strategy in the D-dimensional search space; Step S12, preset the maximum iteration number of the swarm intelligence optimization strategy, and evaluate the fitness of each individual after each iteration; after reaching the preset maximum iteration number, terminate the iteration and perform the next step; Step S13, preset the maximum iteration number of the Daqu fermentation controller, select the optimal value of the individual as the combined parameter weight of the antecedent network and the consequent network, and import it into the Daqu fermentation controller; after reaching the preset maximum iteration number, terminate the iteration, and obtain the optimized Daqu fermentation controller.
3. The Daqu fermentation control method according to claim 2, characterized in that, Before performing the fusion processing in step S2, the consistency of the collected temperature detection data or humidity detection data is detected.
4. The Daqu fermentation control method according to claim 3, characterized in that, The specific process of performing consistency detection on the temperature detection data in step S2 is: Determine whether the difference between the temperature detection data currently collected by the same temperature sensor and the temperature detection data collected at the last time is less than or equal to the corresponding temperature monitoring threshold value; If it is less than or equal to the corresponding temperature monitoring threshold value, perform subsequent fusion processing on the temperature detection data currently collected by the same temperature sensor; If it is greater than the corresponding temperature monitoring threshold value, take the temperature detection average value of all the temperature detection data collected by the same temperature sensor to perform subsequent fusion processing; The specific process of performing consistency detection on the humidity detection data in step S2 is: Determine whether the difference between the humidity detection data currently collected by the same humidity sensor and the humidity detection data collected at the last time is less than or equal to the corresponding humidity monitoring threshold value; If it is less than or equal to the corresponding humidity monitoring threshold value, perform subsequent fusion processing on the humidity detection data currently collected by the same humidity sensor; If it is greater than the corresponding humidity monitoring threshold value, take the humidity detection average value of all the humidity detection data collected by the same humidity sensor to perform subsequent fusion processing.
5. The Daqu fermentation control method according to claim 4, characterized in that, The step S2 is to perform data fusion processing on the temperature detection data, and the actual temperature value of the room after fusion is This is obtained by the following formula: ； ； ； In the formula, represents the actual temperature value of the fusion room. temperature detection data of the pth temperature sensor or an average value of all temperature detection data of the pth temperature sensor; a variance of temperature detection data of a pth temperature sensor among n temperature sensors; an average of temperature detection data of n temperature sensors; Similarly, the fusion processing of the humidity detection data in step S2 refers to the fusion processing process of the temperature detection data.
6. The Daqu fermentation control method according to claim 5, characterized in that, The specific process of learning and training in the Daqu fermentation controller in step S3 is: The learning algorithm of the weights of the second layer of the consequent network is: ； ； In the formula, is a loss function; update the weight of the kth output and the ith input in the pth rule in the back-propagation network; output of the pth rule in the consequent network; outputp k = outputp k-1 + w p k x p k compute the expected output for the pth rule in the consequent network; Membership of the jth fuzzy set; is the component of the i-th input vector; r is the learning rate; l is the iteration number of the Daqu fermentation controller; In the parameter values After the determination, the influence of the check error on each layer in the antecedent network is verified, and the formula is as follows: ； ； ； wherein represents the error term for the i-th neuron node in the fifth layer in the front-end network; The fifth layer of the antecedent network is composed of an output layer; represents an error term for the kth neuron node in the fourth layer in the antecedent network; represents the error term for the kth neuron node in the third layer in the antecedent network; t i represents the desired output of the i-th output node; y i represents the actual output of the i-th output node; y ik represents the connection weight between the kth neuron node in the fourth layer and the ith neuron node in the fifth layer of the antecedent network; a i μi represents the membership of the ith input variable; Finally, we get: ； ； In the formula, E is the loss function; represents the error term of the i-th neuron node at the k-th layer; x i Xi represents the value of the i-th input variable; c ik Ck(i) represents the center value of the kth Gaussian membership function corresponding to the ith input employed in the second layer of the antecedent network; c jk Cjk represents the center value of the kth Gaussian membership function corresponding to the jth input employed in the second layer of the antecedent network; q ik width of the kth Gaussian membership function corresponding to the ith input employed in the second layer of the antecedent network; The center value c of the kth Gaussian membership function corresponding to the ith input employed in the second layer of the antecedent network ik and the width value q ik The learning rule is: ； ； wherein represents the center value of the kth Gaussian membership function employed in the second layer of the antecedent network corresponding to the ith input pair at the (l+1)th iteration of the Daqu fermentation controller. Ck(i) represents the center value of the kth Gaussian membership function employed in the second layer of the antecedent network corresponding to the ith input pair for the lth Daqu fermentation controller iteration; represents the width of the kth Gaussian membership function employed in the second layer of the antecedent network corresponding to the ith input pair for the (l+1)th iteration of the large koji fermentation controller; represents the width of the kth Gaussian membership function employed in the second layer of the antecedent network corresponding to the ith input pair for the lth Daqu fermentation controller iteration; r is the learning rate; E is the loss function.
7. A Daqu fermentation system for implementing the Daqu fermentation control method according to any one of claims 1-6, characterized in that, It includes: A curve room; An actuator, which is dispersedly arranged in the curve room; wherein the actuator includes a heater, a humidifier, an electric window, an internal circulation fan and an external circulation fan; A temperature sensor, which is dispersedly arranged in the curve room and collects temperature detection data of different monitoring points in the curve room; A humidity sensor, which is dispersedly arranged in the curve room and collects humidity detection data of different monitoring points in the curve room; The Daqu fermentation controller is constructed based on a fuzzy logic control algorithm and a neural network fusion, and is optimized by using a swarm intelligence optimization strategy; the optimized Daqu fermentation controller uses an adaptive weighted fusion algorithm based on deep learning to perform data fusion processing on the temperature detection data and the humidity detection data, and then outputs a temperature deviation and a temperature deviation change rate of an actual temperature value of a Daqu room from a target temperature value, and a humidity deviation and a humidity deviation change rate of an actual humidity value of the Daqu room from a target humidity value as input signals, and outputs an opening time of each execution mechanism as an output signal; wherein the temperature deviation change rate is a ratio of the temperature deviation to the target temperature value, and the humidity deviation change rate is a ratio of the humidity deviation to the target humidity value; The host computer receives all the output signals, and converts the output signals into corresponding Boolean signals one by one, and then controls the execution opening or stopping operation of the corresponding execution mechanism according to the Boolean signals.
8. The Daqu fermentation system according to claim 7, characterized in that, The Daqu fermentation controller comprises an antecedent network and a consequent network; The first layer of the antecedent network is composed of an input layer, and each node in the input layer is connected to each component of an input vector and is transmitted to the next layer; the input vector is the input signal; The second layer of the antecedent network is composed of a fuzzification layer, and the input layer is data-fuzzed by the fuzzification layer through a Gaussian membership function; The third layer of the antecedent network is composed of a rule layer, and the rule layer is composed of 49*49 neuron nodes, and each neuron node of the rule layer represents a fuzzy rule; The fourth layer of the antecedent network is composed of a normalization layer, which is used for normalization calculation and serves as a connection weight of the third layer of the consequent network; the neuron node number of the normalization layer is 49*49; The first layer of the consequent network is composed of the input layer, and the input vector is transmitted to the next layer; The second layer of the consequent network is composed of a rule calculation layer, and the rule calculation layer is composed of 49*49 neuron nodes, and each neuron node of the rule calculation layer represents a calculation rule; the rule calculation layer is used to calculate the consequent of the fuzzy rule; The third layer of the consequent network is composed of an output layer, and the output layer performs weighted summation on the output of the normalization layer and the output of the rule calculation layer.
9. The Daqu fermentation system according to claim 7, characterized in that, The host computer also has a visualization module; the visualization module is used for chart data visualization display.