A method for predicting the temperature and humidity of koji turning during the fermentation process of Daqu

Abstract

This invention belongs to the field of temperature and humidity prediction technology for brewing koji (a type of starter culture), and relates to a method for predicting the temperature and humidity of the koji blocks during fermentation. The method includes: 1. Collecting temperature and humidity data for the first three days before the koji is placed in the fermentation chamber; 2. Performing missing data imputation on the temperature and humidity data; 3. Using the imputed temperature data / or temperature and humidity data as input features, and the temperature of the koji during the first turning as the output feature, training a neural network model to obtain a temperature prediction model; or, using the imputed temperature and humidity data as input features, and the humidity of the koji during the first turning as the output feature, training a neural network model to obtain a humidity prediction model; 4. Predicting the temperature of the koji during the first turning based on the temperature prediction model; or, predicting the humidity of the koji during the first turning based on the humidity prediction model.

Description

Technical Field

[0001] This invention belongs to the field of temperature and humidity prediction technology for brewing koji, and relates to a method for predicting the temperature and humidity of koji blocks during the fermentation process. Background Technology

[0002] The temperature and humidity changes of the koji blocks during the fermentation process are related to variables such as ambient temperature, humidity, and process parameters. The temperature and humidity of the first turning of the koji block have a particularly significant impact on the fermentation quality. Extensive experiments have shown a correlation between the temperature and humidity changes of the koji blocks in the first three days and those in the subsequent 7-9 days. Regarding the prediction of the temperature of the koji blocks after the first turning, traditional methods rely on workers' experience or a limited number of temperature measurement points, resulting in poor accuracy. Existing technologies have also disclosed methods for predicting koji block temperature using curve fitting, employing one type of curve model for winter koji, spring and autumn koji, and another type for summer koji. While this method addresses the problem of prediction through fitting to some extent, the parameters of the fitted curve are difficult to obtain, and in some special cases, the parameters are difficult to modify. Especially in different seasons, environments, and sensor placement, the parameters of the fitted curve can be affected to some extent, making it difficult to practically apply to temperature prediction in different koji storage areas and seasons. Currently, there is no method for predicting the humidity of the koji blocks after the first turning.

[0003] Based on this, the present invention provides a method for predicting the temperature and humidity of koji blocks during the fermentation process in different seasons, which has important guiding significance for timely adjustment of koji production measures. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting the temperature and humidity of the fermentation blocks during the fermentation process of Daqu (a type of starter culture).

[0005] According to a first aspect of the present invention, the present invention provides a method for predicting the temperature and humidity of koji blocks during the fermentation process of koji, the method comprising:

[0006] 1. Collect temperature and humidity data for the three days prior to the fermentation of Daqu (a type of starter culture);

[0007] 2. Perform missing data filling processing on the collected temperature and humidity data respectively;

[0008] 3. Using the temperature data after missing data filling and / or the temperature and humidity data as input features, and the temperature of the koji block corresponding to one turning of the koji as the output feature, an SVR neural network model is trained to construct a temperature prediction model; or, using the temperature data and humidity data after missing data filling as input features, and the humidity of the koji block corresponding to one turning of the koji as the output feature, an SVR neural network model is trained to construct a humidity prediction model.

[0009] Fourth, predict the temperature of the Daqu (fermented liquor) to be predicted during one turning of the fermentation process based on the temperature prediction model; or, predict the humidity of the Daqu to be predicted during one turning of the fermentation process based on the humidity prediction model.

[0010] In some embodiments of the present invention, in step one, the sampling time interval for collecting temperature and humidity data for the three days prior to the fermentation of Daqu (a type of starter culture) is 1-2 hours per time.

[0011] In some embodiments of the present invention, in step three, the temperature data is the temperature data for the three days before the Daqu (a type of starter culture) is put into the fermentation chamber; the humidity data is the humidity data for the three days before the Daqu is put into the fermentation chamber.

[0012] In some embodiments of the present invention, the temperature data for the three days prior to fermentation of the Daqu (a type of starter culture) refers to the temperature data at the same time from the first day of fermentation to the third day of fermentation; the humidity data for the three days prior to fermentation of the Daqu refers to the humidity data at the same time from the first day of fermentation to the third day of fermentation.

[0013] In some embodiments of the present invention, the missing data filling process includes: using the KNN method to fill in the missing temperature and / or humidity data in the time series acquired in step one.

[0014] In some embodiments of the present invention, the missing filling includes the following steps:

[0015] (1) Calculate the Euclidean distance between each data sample and the current missing value;

[0016] (2) Based on the results calculated in step (1), select the 5 data samples with the shortest Euclidean distance, take the weighted average value as the imputation value;

[0017] (3) Based on the filling value obtained in step (2), if the filling is not completed, return to step (1) and continue the missing filling process; if the filling is completed, then end and proceed to step (4);

[0018] (4) Based on the filling results of step (3) and the collected data, new time series data samples are generated. The newly generated time series data samples are expanded according to a sampling time frequency that is shorter than the time interval of the collected data samples to obtain the missing data.

[0019] In some embodiments of the present invention, the shorter sampling time frequency includes a sampling frequency of 5-15 min / time.

[0020] In some embodiments of the present invention, the Euclidean distance between each data sample and the current missing value is obtained by...

[0021] The calculation is performed using the formula shown in Equation 1:

[0022]

[0023] In Equation 1, d i Representing point (x) i y i ) and point (x) j y j The Euclidean distance between x and y, where i represents the index of the i-th point, j represents the index of the j-th point, and x is the distance between x and y. i Let y represent the x-coordinate of the i-th point. i Let x represent the ordinate value of the i-th point. j y represents the x-coordinate of the j-th point. j This represents the ordinate value of the j-th point.

[0024] In some embodiments of the present invention, the fill value is calculated using the formula shown in Equation 2:

[0025]

[0026] In Equation 2, i represents the 5 data samples with the shortest Euclidean distance to the missing values, i.e., i = 1, 2, 3, 4, 5; d i x represents the Euclidean distance between the 5 data samples with the shortest Euclidean distance to the missing value and the missing value. i x represents the x-coordinate of the i-th sample point with the shortest Euclidean distance to the missing value, and x is the imputation value for the missing value.

[0027] In some embodiments of the present invention, the method further includes: optimizing the constructed temperature prediction model / or humidity prediction model using the Sparrow Search Algorithm (SSA), and predicting the temperature / or humidity of the Daqu (a type of liquor) to be predicted during a single turning of the Daqu based on the optimized temperature prediction model / or humidity prediction model.

[0028] In some embodiments of the present invention, the method further includes: optimizing the penalty parameter and the variance of the Gaussian kernel function in the constructed temperature prediction model / or humidity prediction model using the Sparrow Search Algorithm (SSA); optimizing and training the temperature prediction model / or humidity prediction model based on the optimal values ​​of the penalty parameter and the variance of the Gaussian kernel function obtained after optimization to obtain an optimized temperature prediction model / or humidity prediction model; and predicting the temperature / or humidity of the Daqu to be predicted during a single turning of the Daqu based on the optimized temperature prediction model / or humidity prediction model.

[0029] In some embodiments of the present invention, the optimization of the penalty parameter and the variance of the Gaussian kernel function in the constructed temperature prediction model / or humidity prediction model using the Sparrow Search Algorithm (SSA) includes the following steps:

[0030] 1) Preprocess the data samples and divide the preprocessed data into training set and test set;

[0031] 2) Set the population size N, the maximum number of generations T, initialize the parameters of the prediction model, and set the penalty parameter and the variance of the Gaussian kernel function in the prediction model as the optimization values ​​of the sparrow search algorithm.

[0032] 3) Treat the penalty parameter and the variance of the Gaussian kernel function as sparrows, calculate the fitness of each individual, rank the fitness of each individual, select the sparrows with the highest fitness from the population as discoverers, and the remaining sparrows become followers;

[0033] 4) Update the positions of the discoverer, follower, and watcher;

[0034] 5) Compare the fitness of the new individual location with that of the original individual location, and retain the better individuals;

[0035] 6) Determine whether the set number of iterations has been reached or whether the optimal solution that meets the accuracy requirements has been found. If the set number of iterations has been reached or the optimal solution that meets the accuracy requirements has been found, stop the search; otherwise, return to step 4) and continue the search.

[0036] 7) Output the optimal values ​​of the penalty parameter and the variance of the Gaussian kernel function, and retrain the training data to obtain the optimized temperature prediction model / humidity prediction model.

[0037] In some embodiments of the present invention, step 1) of the method for preprocessing the data sample includes: mapping the original data to the range [0,1] through Min-Max normalization, and performing normalization processing using the formula shown in Equation 12.

[0038]

[0039] In Equation 12, x new Here are the normalized data, and x is the original sample data. max x is the maximum value of the sample data. min This represents the minimum value of the sample data.

[0040] In some embodiments of the present invention, the optimization of the penalty parameter and the variance of the Gaussian kernel function in the constructed temperature prediction model / or humidity prediction model using the Sparrow Search Algorithm (SSA) further includes:

[0041] 8) Input the validation set data into the optimized temperature prediction model / humidity prediction model for prediction, evaluate and analyze the performance of the prediction model.

[0042] In some embodiments of the present invention, in step 4), the positions of the discoverer, follower, and vigilant are updated and calculated using the following formula, which includes:

[0043] The location of the discoverer is updated using the formula shown in Equation 13:

[0044]

[0045] In Equation 13, This indicates the position of the discoverer at the t-th iteration. Let j represent the position of the discoverer in the (t+1)th iteration, j represent the discoverer's ID, i represent the dimension of the position vector, and α be a random number in the range (0,1]. iter,max R2 is the ratio of the current iteration count to the maximum iteration count. R2 is the warning value, which is a random number uniformly distributed in the interval [0,1). ST is the safety value, which ranges from [0.5,1]. L is the vector difference between the discoverer's position and the global optimal solution's position, which is a 1×m dimensional matrix where each element is 1. Q is a parameter that controls the step size.

[0046] In some embodiments of the present invention, the position of the follower is updated using the formula shown in Equation 14:

[0047]

[0048] In Equation 14, Q is a random vector; X w This indicates the position of the current worst-case global solution. This represents the position of the current global optimal solution. A represents the Euclidean distance between the current sparrow and the global optimal solution; + Given a 1×m matrix, where each element is assigned the value 1 or -1, and A + =A T (AA T ) -1 L represents the side length of the current search space.

[0049] In some embodiments of the present invention, the position of the vigilant is updated using the formula shown in Equation 15:

[0050]

[0051] In Equation 15, X b f represents the location of the global optimal solution. j f represents the fitness value of the j-th sparrow; g f is the fitness value of the global optimal solution. bε is the worst fitness value globally; β represents the control distance coefficient, which follows a normal distribution with a mean of 0 and a variance of 1; K represents the sensitivity coefficient of the vigilant, which is the direction of the sparrow's movement and takes a random number in the range [-1, 1]; ε is a minimum constant.

[0052] In some embodiments of the present invention, step four, which involves predicting the temperature of the Daqu (a type of starter culture) during its first turning based on the constructed temperature prediction model, includes: obtaining temperature data or temperature and humidity data for the three days prior to the fermentation of the Daqu as input features, inputting them into the temperature prediction model, and predicting the temperature of the Daqu during its first turning based on the output results.

[0053] In some embodiments of the present invention, the prediction of the humidity of the Daqu to be predicted during the first turning of the Daqu based on the constructed humidity prediction model includes: obtaining the temperature and humidity data of the Daqu to be predicted for three days before entering the fermentation chamber as input features, inputting them into the humidity prediction model, and predicting the humidity of the Daqu to be predicted during the first turning of the Daqu based on the output results.

[0054] In some embodiments of the present invention, the prediction model is as shown in equation (3):

[0055]

[0056] In Equation 3, i is the index of the i-th training sample, j is the index of the j-th training sample, k is the index of the k-th predicted sample, and x is the feature vector of the input sample. j Let α be the feature vector of the j-th training sample. i Let α be the weight of the j-th training sample. j * Let b be another set of weights for the j-th training sample, and y be the bias term. k This is the predicted output for the input sample x.

[0057] In some embodiments of the present invention, the solution process of the prediction model includes:

[0058] Step 1: Calculate the error;

[0059] Step 2: Based on the error-free result calculated in Step 1, assuming the tolerable deviation is ε, calculate the insensitive loss function.

[0060] Step 3: Solve the optimization problem to obtain the parameters.

[0061] Step 4: Calculate the solution to the original optimization problem;

[0062] Step 5: Calculate the output of the SVR model to obtain the predicted value.

[0063] In some embodiments of the present invention, in the first step, the error is calculated using the formula shown in Equation 4:

[0064] E i =f(x) i )-y i (Equation 4);

[0065] In Equation 4, f(x) i ) represents the predicted value, y i Let i be the index of the i-th training sample, which is the training data.

[0066] In some embodiments of the present invention, in the second step, the insensitive loss function is calculated using the formula shown in Equation 5.

[0067]

[0068] In some embodiments of the present invention, in the third step, the parameters are calculated using the formula shown in Equation 6.

[0069]

[0070] In the formula, C is the penalty parameter, and α i For the Lagrange multiplier associated with the i-th training sample, For another set of Lagrange multipliers associated with the i-th training sample, α j For the Lagrange multiplier associated with the j-th training sample, For another set of Lagrange multipliers related to the j-th training sample, x i For the data of the i-th sample, x j Let i be the index of the j-th sample, where i is the index of the i-th sample and j is the index of the j-th sample.

[0071] In the third step, α is solved using the SMO algorithm. i and α j Each time, two variables α are chosen. i and α j And fix other parameters. Thus, after parameter initialization, SMO continuously executes the following two steps until convergence:

[0072] (1) Select a pair of variables α that need to be updated. i and α j ;

[0073] (2) Fix α i and α j Solving for the parameters other than those specified yields the updated α.i and α j .

[0074] In some embodiments of the present invention, in the fourth step, the solution to the original optimization problem is calculated using the formulas shown in Equations 7 and 8:

[0075]

[0076]

[0077] In equations 7 and 8, Let x be the difference between two distinct Lagrange multipliers related to the i-th training sample. i For the data of the i-th sample, x j Let i be the index of the j-th sample, and j be the index of the j-th sample.

[0078] In some embodiments of the present invention, in the fifth step, the output of the SVR model is calculated using the formula shown in Equation 9:

[0079]

[0080] In Equation 9, κ(x, x) i Let be the Gaussian kernel function, which is calculated using the formula shown in Equation 10:

[0081]

[0082] In the formula, σ i Let be the variance of the Gaussian function.

[0083] According to a second aspect of the present invention, the present invention also provides an application of the method described in any of the first aspects of the present invention in the field of baijiu brewing. In some embodiments of the present invention, the baijiu brewing field includes the field of daqu fermentation preparation technology; the baijiu brewing field also includes the prediction of temperature and humidity during daqu fermentation.

[0084] Compared with the prior art, the present invention has the following beneficial effects:

[0085] (1) In order to accurately predict the temperature and humidity during the fermentation process of Daqu and better guide the production of Daqu, this invention is based on a large amount of data collected from different Daqu blocks and different time periods. The temperature and humidity prediction model of Daqu blocks is trained and learned by classifying and dividing the data into blocks. Finally, the temperature and humidity prediction model of Daqu is constructed. The temperature and humidity prediction model provided by this invention effectively solves the problem of temperature and humidity prediction during the fermentation process of Daqu.

[0086] (2) The prediction method provided by the present invention can predict the temperature and humidity during the fermentation process of Daqu with high accuracy and small error between the predicted value and the actual value. At the same time, the prediction method provided by the present invention can simultaneously achieve accurate prediction of the temperature and humidity of Daqu blocks in different seasons. Compared with the existing technology that requires fitting different models for Daqu in different seasons, the prediction method provided by the present invention is more practical and can be better applied to the actual process production of Daqu fermentation / making. Attached Figure Description

[0087] Figure 1 This is the result of data missing data imputation in Embodiment 1 of this invention; wherein, Figure 1 (A) is the result of missing data imputation for temperature data. Figure 1 (B) is the result of missing data imputation for humidity data;

[0088] Figure 2 This is a structural illustration of the method for predicting the temperature and humidity of a curved block using different methods in Embodiment 1 of the present invention;

[0089] Figure 3 This is a schematic diagram illustrating the construction of the temperature prediction model in Embodiment 1 of the present invention;

[0090] Figure 4 This is a schematic diagram illustrating the construction of the humidity prediction model in Embodiment 1 of the present invention;

[0091] Figure 5 This is a flowchart illustrating the optimization of the SVR prediction model using the Sparrow Search Algorithm (SSA) in Embodiment 1 of the present invention.

[0092] Figure 6 The curves showing the mean square error versus the number of optimization iterations obtained by using different optimization models to predict temperature and humidity in Embodiment 1 of the present invention are shown. Figure 6 (A) shows the adaptation curves of mean square error versus number of optimization iterations obtained by using different optimization models to predict temperature; Figure 6 (B) shows the adaptation curves of mean square error versus number of optimization iterations obtained by using different optimization models to predict humidity;

[0093] Figure 7 This is a comparison between the predicted values ​​and the actual values ​​obtained by using different optimization models to predict temperature and humidity in Embodiment 1 of the present invention; wherein, Figure 7 (A) is a comparison between the predicted temperature value and the actual temperature value obtained by using different optimization models in Embodiment 1 of the present invention; Figure 7(B) is a comparison between the predicted humidity value and the actual humidity value obtained by using different optimization models in Embodiment 1 of the present invention;

[0094] Figure 8 This refers to the result obtained by linearly fitting the predicted values ​​of temperature and humidity obtained by using different optimization models in Embodiment 1 of the present invention with the actual values; wherein, Figure 8 (A) is the linear fitting result between the predicted temperature value and the actual temperature value obtained by using different optimization models in Embodiment 1 of the present invention; Figure 8 (B) is the linear fitting result between the predicted humidity value and the actual humidity value obtained by using different optimization models in Embodiment 1 of the present invention;

[0095] Figure 9 This presents a comparison of the prediction results and prediction errors for the temperature and humidity of curved block A using different optimization models in Embodiment 2 and Comparative Example 1 of the present invention; wherein, Figure 9 (a) Comparison results of temperature prediction for block A using different optimization models in Embodiment 2 and Comparative Example 1 of the present invention. Figure 9 (b) Comparison of prediction errors for predicting the temperature of block A using different optimization models in Embodiment 2 and Comparative Example 1 of the present invention; Figure 9 (c) Comparison results of predicting the humidity of curved block A using different optimization models in Embodiment 2 and Comparative Example 1 of the present invention. Figure 9 (d) is a comparison of the prediction errors of the humidity prediction of block A using different optimization models in Embodiment 2 and Comparative Example 1 of the present invention;

[0096] Figure 10 This presents a comparison of the prediction results and prediction errors for the temperature and humidity of curved block B using different optimization models in Embodiment 2 and Comparative Example 1 of the present invention; wherein, Figure 10 (a) Comparison results of temperature prediction for block B using different optimization models in Embodiment 2 and Comparative Example 1 of the present invention. Figure 10 (b) Comparison of prediction errors for predicting the temperature of block B using different optimization models in Embodiment 2 and Comparative Example 1 of the present invention; Figure 10 (c) Comparison results of predicting the humidity of curved block B using different optimization models in Embodiment 2 and Comparative Example 1 of the present invention. Figure 10 (d) is a comparison of the prediction errors of the humidity prediction of block B using different optimization models in Embodiment 2 and Comparative Example 1 of the present invention;

[0097] Figure 11This presents a comparison of the prediction results and prediction errors for the temperature and humidity of curved block C using different optimization models in Embodiment 2 and Comparative Example 1 of the present invention; wherein, Figure 11 (a) Comparison results of temperature prediction for curved block C using different optimization models in Embodiment 2 and Comparative Example 1 of the present invention. Figure 11 (b) Comparison of prediction errors for predicting the temperature of block C using different optimization models in Embodiment 2 and Comparative Example 1 of the present invention; Figure 11 (c) Comparison results of predicting the humidity of curved block C using different optimization models in Embodiment 2 and Comparative Example 1 of the present invention. Figure 11 (d) is a comparison of the prediction errors of the humidity prediction of the curved block C using different optimization models in Embodiment 2 and Comparative Example 1 of the present invention;

[0098] Figure 12 This presents a comparison of the prediction results and prediction errors of the temperature and humidity of curved block D using different optimization models in Embodiment 2 and Comparative Example 1 of the present invention; wherein, Figure 12 (a) Comparison results of temperature prediction for curved block D using different optimization models in Embodiment 2 and Comparative Example 1 of the present invention. Figure 12 (b) Comparison of prediction errors for predicting the temperature of block D using different optimization models in Embodiment 2 and Comparative Example 1 of the present invention; Figure 12 (c) Comparison results of predicting the humidity of curved block D using different optimization models in Embodiment 2 and Comparative Example 1 of the present invention. Figure 12 (d) shows the comparison results of prediction errors for predicting the humidity of curved block D using different optimization models in Embodiment 2 and Comparative Example 1 of the present invention. Detailed Implementation

[0099] The following specific embodiments further illustrate the technical solution of the present invention. These specific embodiments do not represent a limitation on the scope of protection of the present invention. Non-essential modifications and adjustments made by others based on the concept of the present invention still fall within the scope of protection of the present invention.

[0100] Example 1

[0101] In the fermentation process of brewing koji, three stages of operation are carried out according to the fermentation status: first turning, second turning, and koji removal. Among them, the turning operation is a key step to ensure normal and uniform fermentation of the koji blocks, and has a significant impact on the fermentation quality of koji. In particular, the temperature of the koji blocks at the time of the first turning directly affects the overall fermentation effect. In order to accurately predict the temperature of koji during the fermentation process and guide the koji production work, this invention aims to monitor the temperature and humidity data of the koji blocks for the first three days of fermentation, train an SSA-SVR prediction model based on SVR, and use the trained SSA-SVR prediction model to accurately predict the temperature and humidity on the 9th day after the first turning of the koji. Based on this, this embodiment provides a method for predicting the temperature and humidity of koji blocks during the fermentation process of koji based on the SSA-SVR model, including the following steps:

[0102] I. Data Collection:

[0103] Temperature and humidity data were collected by automatic temperature and humidity sensors during the fermentation process of Daqu. Spatially, different parts of the Daqu were selected for monitoring, including the door, center, and window of the fermentation chamber. In terms of sampling time, the first measurement was started after the Daqu was placed in the chamber and timed, and then measured every hour thereafter, for the entire fermentation cycle of Daqu.

[0104] II. Data missing information handling:

[0105] The temperature and humidity data collected in Step 1 are time series data, meaning they include temperature and humidity data for different locations of the large curved blocks at each consecutive time point (e.g., every hour). However, the temperature and humidity time series data obtained in Step 1 has temporal discontinuities. Missing values ​​in the data can lead to biases and incorrect conclusions when training the prediction model. To improve the accuracy of the subsequent prediction model, this step further performs missing data imputation on the temperature and humidity data obtained in Step 1. Specifically, the KNN method is used to impute missing temperature or humidity data in the collected time series. The specific steps are as follows:

[0106] (1) Calculate the Euclidean distance between each data sample and the current missing value:

[0107]

[0108] In Equation 1, d i Representing point (x) i y i ) and point (x) j y jThe Euclidean distance between x and y, where i represents the index of the i-th point, j represents the index of the j-th point, and x is the distance between x and y. i Let y represent the x-coordinate of the i-th point. i Let x represent the ordinate value of the i-th point. j y represents the x-coordinate of the j-th point. j This represents the ordinate value of the j-th point;

[0109] (2) Based on the results calculated in step (1), select the 5 data samples with the shortest Euclidean distance, and take the weighted average as the imputation value x:

[0110]

[0111] In Equation 2, i represents the 5 data samples with the shortest Euclidean distance to the missing values, i.e., i = 1, 2, 3, 4, 5; d i x represents the Euclidean distance between the 5 data samples with the shortest Euclidean distance to the missing value and the missing value. i x represents the x-coordinate of the i-th sample point with the shortest Euclidean distance to the missing value, and x is the imputation value for the missing value.

[0112] (3) If the filling is not completed, return to step (1) and continue the missing filling process. If the filling is completed, then end and proceed to step (4).

[0113] (4) Based on the filling results of step (3) and the data obtained in step one, new time series data samples are generated. The newly generated time series data samples are then expanded at a frequency of once every 10 minutes to obtain expanded data.

[0114] III. Construction of the Prediction Model:

[0115] Based on the data imputation process in step two above, temperature data and humidity data for each moment (in 10-minute intervals) of the fermentation process from the start of fermentation to the end of the entire fermentation cycle were obtained. Based on the obtained data, a prediction model was constructed:

[0116] (1) First, the input parameters are screened and compared. The specific steps are as follows: By studying the fermentation process, this invention provides the following four methods to predict the temperature and humidity corresponding to the 9th day of fermentation of the koji blocks. The four methods are as follows:

[0117] Group A: Predict the temperature on the 9th day of fermentation of koji blocks by using the temperature three days before the koji blocks are put into the fermentation chamber;

[0118] Group B: Predict the temperature on the 9th day of fermentation by measuring the temperature and humidity of the koji blocks in the fermentation chamber during the first three days of fermentation.

[0119] Group C: Predict the humidity on the 9th day of fermentation of koji blocks by measuring the humidity of the koji blocks three days before fermentation.

[0120] Group D: Predict the humidity on the 9th day of fermentation of koji blocks by using the temperature and humidity of the koji blocks in the fermentation chamber three days before fermentation;

[0121] The four prediction methods described above were tested. Using the SVR regression neural network trained according to these methods, all parameters, training samples, and prediction samples were identical except for the prediction method (input parameters). The structure of the prediction method is as follows: Figure 2 As shown, the results are as follows: Group A: The average absolute error of temperature prediction using temperature is 0.3201; Group B: The average absolute error of temperature prediction using both temperature and humidity is 0.3147; Group C: The average absolute error of humidity prediction using humidity is 0.7470; Group D: The average absolute error of humidity prediction using both temperature and humidity is 0.4024. In the comparison of the four groups, Groups A and B are for temperature prediction. The average absolute errors of the two methods are close, differing by only 0.0054. However, due to the larger number of input parameters and more complex calculations in method B, the final temperature prediction method is determined to be using the temperature of the koji blocks three days before fermentation as the input parameter to predict the temperature on the 9th day of fermentation. Groups C and D are for humidity prediction. Based on the above prediction results, method D is significantly better than method C. Therefore, the final humidity prediction method is determined to be using the temperature and humidity of the koji blocks three days before fermentation to predict the humidity on the 9th day of fermentation.

[0122] (2) Based on the research results of step (1) above, the SVR model is further constructed:

[0123] First, the temperature prediction model is constructed: the input layer contains the daily temperatures at the same time each day for the three days prior to the fermentation of the koji blocks, and the output layer contains the temperature on the 9th day of fermentation. The structure is as follows: Figure 3 As shown; Humidity prediction model construction: The input layer consists of the humidity and temperature of the koji blocks at the same time each day for the three days before fermentation, and the output layer is the humidity on the 9th day of fermentation. The structure is as follows. Figure 4 As shown; the specific construction process is as follows:

[0124] Construction of the SVR Neural Network Temperature Prediction Model: The temperature data obtained above for the first to third day of fermentation of Daqu koji blocks (with a 10-minute time interval) and the ninth day of fermentation (with a 10-minute time interval) are divided into a training set and a test set, with a ratio of 7:3 between the training set and the test set. The SVR neural network prediction model is trained and constructed using the data in the training set. Specifically, the temperature data of Daqu koji blocks from the first to the third day of fermentation in the training set is used as the input feature, and the temperature of the corresponding Daqu koji block at one turning time (i.e., the ninth day of fermentation) is used as the output feature. The SVR neural network model (a model combining support vector machine regression (SVR) and neural network) is used for training to construct the SVR neural network temperature prediction model.

[0125] Construction of the SVR Neural Network Humidity Prediction Model: The temperature and humidity data of the Daqu koji blocks obtained above from the first day to the third day of fermentation (with a time interval of 10 minutes) and the temperature data corresponding to the ninth day of fermentation (with a time interval of 10 minutes) are divided into a training set and a test set, with a ratio of 7:3 between the training set and the test set. The SVR neural network prediction model is trained and constructed using the data in the training set. Specifically, the temperature and humidity data of the Daqu koji blocks from the first day to the third day of fermentation are used as input features, and the humidity of the Daqu koji blocks at the time of one turning (i.e., the ninth day of fermentation) is used as output features. The SVR neural network model (a model that combines support vector machine regression (SVR) and neural network) is used for training to construct the SVR neural network humidity prediction model.

[0126] The SVR neural network temperature / humidity prediction model constructed in the above manner is shown in Equation 3:

[0127]

[0128] In Equation 3, i is the index of the i-th training sample, j is the index of the j-th training sample, k is the index of the k-th predicted sample, and x is the feature vector of the input sample. j Let α be the feature vector of the j-th training sample. i Let α be the weight of the j-th training sample. j * Let b be another set of weights for the j-th training sample, and y be the bias term. k This is the predicted output for the input sample x.

[0129] The calculation process of the SVR neural network prediction model is as follows:

[0130] Step 1: Calculate the error:

[0131] E i =f(x) i )-y i (Equation 4);

[0132] In Equation 4, f(x) i ) represents the predicted value, y i For training data, i is the index of the i-th training sample;

[0133] Step 2: Assuming the tolerable deviation is ε, calculate the insensitive loss function ε. i * :

[0134]

[0135] Step 3: Solve the optimization problem to obtain the parameters.

[0136]

[0137] In the formula, C is the penalty parameter, and α i For the Lagrange multiplier associated with the i-th training sample, For another set of Lagrange multipliers associated with the i-th training sample, α j For the Lagrange multiplier associated with the j-th training sample, For another set of Lagrange multipliers related to the j-th training sample, x i For the data of the i-th sample, x j Let i be the index of the i-th sample, and j be the index of the j-th sample.

[0138] Solving α using the SMO algorithm i and α j Each time, two variables α are chosen. i and α j And fix other parameters. Thus, after parameter initialization, SMO continuously executes the following two steps until convergence:

[0139] (1) Select a pair of variables α that need to be updated. i and α j ;

[0140] (2) Fix α i and α j Solving for the parameters other than those specified yields the updated α. i and α j (The parameters are solved based on equation (6);

[0141] Step 4: Calculate the solution to the original optimization problem:

[0142]

[0143]

[0144] In equations 7 and 8, Let x be the difference between two distinct Lagrange multipliers related to the i-th training sample. i For the data of the i-th sample, x j Let i be the index of the j-th sample, and j be the index of the j-th sample.

[0145] Step 5: Calculate the output of the SVR model:

[0146]

[0147] In Equation 3, κ(x, x) i Let ) be the Gaussian kernel function, and its formula is as follows:

[0148]

[0149] In the formula, σ i Let be the variance of the Gaussian function.

[0150] Based on the above calculation process, the corresponding predicted value is obtained.

[0151] IV. Optimization of the SVR Neural Network Temperature / Humidity Prediction Model

[0152] In step three of the calculation process, the values ​​of σ and C affect α. * , w * b * The solution has a significant impact on the problem. This invention further employs the Sparrow Search Algorithm (SSA) to optimize the penalty parameter (C) and the variance (σ) of the Gaussian kernel function to obtain suitable parameters. The optimal values ​​of the penalty parameter and the variance of the Gaussian kernel function for the prediction model are found by simulating the behavior of sparrows during foraging. The core of this algorithm is to introduce a group of "sparrows" into the solution space, utilizing their information sharing and cooperation to gradually approach the optimal solution. Parameter settings typically include the initial population size, maximum number of iterations, and local search radius. The initial population should cover the entire solution space to ensure diversity, while the local search radius can be adjusted according to the problem characteristics to enhance the balance between exploration and development. The specific steps are as follows:

[0153] Step 1: Preprocess the data samples to facilitate model training, and divide the preprocessed data into training and testing sets. The specific preprocessing method is as follows: Min-Max normalization is used to map the original data to the range [0,1]. The normalization formula is as follows:

[0154] Where, x new Here are the normalized data, and x is the original sample data. max x is the maximum value of the sample data. min This represents the minimum value of the sample data.

[0155] Step 2: Set the population size N, the maximum number of generations T, initialize the parameters of the prediction model, and set the penalty parameter and the variance of the Gaussian kernel function in the prediction model as the optimization values ​​of the sparrow search algorithm.

[0156] Step 3: Treating the penalty parameter and the variance of the Gaussian kernel function as sparrows, calculate the fitness of each individual and rank them. Select the sparrows with the highest fitness from the population as discoverers, and the remaining sparrows as followers. A two-dimensional population of n sparrows can be represented as shown in Equation 12:

[0157]

[0158] Step 4: Discoverer Location Update:

[0159]

[0160] In Equation 12, This indicates the position of the discoverer at the t-th iteration. Let j represent the position of the discoverer in the (t+1)th iteration, j represent the discoverer's ID, i represent the dimension of the position vector, and α be a random number in the range (0,1]. iter,max R2 is the ratio of the current iteration count to the maximum iteration count. R2 is the warning value, which is a random number uniformly distributed in the interval [0,1). ST is the safety value, which ranges from [0.5,1]. L is the vector difference between the discoverer's position and the global optimal solution's position, which is a 1×m dimensional matrix where each element is 1. Q is a parameter that controls the step size.

[0161] Step 5: Follow the location update:

[0162]

[0163] In Equation 13, Q is a random vector; X w This indicates the position of the current worst-case global solution. This represents the position of the current global optimal solution. A represents the Euclidean distance between the current sparrow and the global optimal solution; + Given a 1×m matrix, where each element is assigned the value 1 or -1, and A + =A T (AA T ) -1 L represents the side length of the current search space.

[0164] Step 6: Guardian Position Update:

[0165]

[0166] In Equation 14, X b f represents the location of the global optimal solution. j f represents the fitness value of the j-th sparrow; g f is the fitness value of the global optimal solution. b ε is the worst fitness value globally; β represents the control distance coefficient, which follows a normal distribution with a mean of 0 and a variance of 1; K represents the sensitivity coefficient of the vigilant, which is the direction of the sparrow's movement and takes a random number in the range [-1, 1]; ε is a minimum constant.

[0167] Step 6: Compare the fitness of the new individual positions with that of the original individual positions, and retain the better individuals;

[0168] Step 7: Determine whether the set number of iterations has been reached or whether the optimal solution that meets the accuracy requirements has been found. If the set number of iterations has been reached or the optimal solution that meets the accuracy requirements has been found, stop the search; otherwise, return to step 4 and continue the search.

[0169] Step 8: Output the optimal values ​​of the penalty parameter and the variance of the Gaussian kernel function, and retrain the training data to obtain the optimized prediction model;

[0170] Step 9: Input the validation set data into the optimized prediction model for prediction, evaluate and analyze the performance of the prediction model.

[0171] The process of optimizing the SVR prediction model using SSA described above is as follows: Figure 5 As shown.

[0172] Meanwhile, this embodiment further compares and optimizes the penalty parameter (C) and the variance (σ) of the Gaussian kernel function in the calculation process of step three based on particle swarm optimization (PSO) and genetic algorithm (GA). The specific comparison results are as follows:

[0173] In this embodiment, 1564 samples were obtained after preprocessing the data as the initial dataset, including 872 temperature data and 692 humidity data. The initial dataset was shuffled, and the normalized data from the training and test sets were input into the network for training. The optimization algorithm population size was set to 30, the maximum number of generations was set to 50, the discoverer ratio of the sparrow search algorithm was 0.2, the guard ratio was 0.1, and the safety value was 0.8; the learning rate of the particle swarm optimization algorithm was 1.5; and the generation gap of the genetic algorithm was 0.95, the crossover probability was 0.7, and the mutation probability was 0.01.

[0174] The mean squared error is further used to evaluate the prediction models obtained through different optimization methods, and finally the optimized parameters are obtained. The iterative results are as follows: Figure 6 As shown. According to Figure 6 As shown in the iterative results, the Sparrow Search algorithm has a faster optimization speed than the Genetic Algorithm and the Particle Swarm Optimization (PSO) algorithm. Under the same iterative parameters, its optimization results are also better than those of the Genetic Algorithm and the PSO algorithm. Therefore, the Sparrow Search algorithm is more suitable for training and optimizing the prediction model of this invention.

[0175] V. Training and Optimization of the SSA-SVR Prediction Model

[0176] Based on the optimization results in step four, and using the optimal values ​​of σ and C obtained by optimizing the penalty parameter (C) and the variance (σ) of the Gaussian kernel function using the Sparrow Search Algorithm (SSA), the SVR prediction model is further optimized and trained to obtain the SSA-SVR prediction model.

[0177] Meanwhile, this embodiment also further compares the prediction capabilities of the corresponding PSO-SVR and GA-SVR prediction models obtained by optimizing σ and C based on particle swarm optimization and genetic algorithm training. The specific comparison results are as follows:

[0178] In this embodiment, the optimized temperature and humidity prediction model was tested using 893 temperature samples and 670 humidity samples. The resulting errors and fitting results are shown below. Figure 7 , 8 As shown. According to Figure 7 , Figure 8 The results shown, after extensive data training and verification, demonstrate that the predicted values ​​of temperature and humidity on the 9th day of fermentation of koji blocks using the SSA-SVR prediction model provided by this invention are approximately on a straight line with the actual values, exhibiting a high degree of overlap. Specifically, R... 2The values ​​are 0.896 and 0.913, respectively. Meanwhile, the SSA-SVR prediction model provided by this invention was used to predict the temperature and humidity on the 9th day of fermentation of koji blocks. The average absolute error of the temperature prediction was 0.063564, the mean square error was 0.0051966, and the average absolute percentage error was 0.1392%. The average absolute error of the humidity prediction was 0.055888, the mean square error was 0.004275, and the average absolute percentage error was 0.17593%. It can be seen that the prediction results obtained by the SSA-SVR prediction model provided by this invention for the temperature and humidity on the 9th day of fermentation of koji blocks have small errors compared with the actual values, and can be used in actual industrial production to predict the temperature and humidity on the 9th day of fermentation of koji blocks. Furthermore, predictions using the PSO-SVR and GA-SVR models yielded mean absolute errors of 0.0914 and 0.0997 for temperature, mean square errors of 0.00872 and 0.00925, and mean absolute percentage errors of 0.283% and 0.391%, respectively. Humidity predictions showed mean absolute errors of 0.0871 and 0.0914, mean square errors of 0.00834 and 0.00897, and mean absolute percentage errors of 0.278% and 0.324%, respectively.

[0179] Example 2

[0180] This embodiment, based on the SSA-SVR prediction model constructed in Embodiment 1, further predicts the temperature and humidity of Daqu (a type of liquor) in different seasons, as detailed below:

[0181] 1. Select different large-sized pieces of liquor (A, B, C, D) from the four seasons of spring, summer, autumn, and winter respectively;

[0182] 2. Obtain temperature and humidity data for Daqu blocks in different seasons three days before they are put into the fermentation chamber;

[0183] Third, the temperature and humidity data of Daqu koji blocks obtained in step two three days before entering the fermentation chamber are used as input features and input into the SSA-SVR prediction model constructed in Example 1 to predict the humidity data corresponding to Daqu koji blocks in different seasons; and the temperature data of Daqu koji blocks obtained in step two three days before entering the fermentation chamber are used as input features and input into the SSA-SVR prediction model constructed in Example 1 to predict the temperature data corresponding to Daqu koji blocks in different seasons.

[0184] The prediction performance was evaluated using MAE, MSE, and MAPE, and the results are shown in Tables 1 and 2.

[0185] Table 1 Temperature Prediction MAE, MSE, and MAPE Values

[0186]

[0187] Table 2. Predicted humidity values: MAE, MSE, and MAPE

[0188]

[0189] Among them, MAE (Mean Absolute Error) is a statistical indicator that measures the difference between predicted and actual values; MSE (Mean Squared Error) is an abbreviation for mean squared error, used to measure the difference between predicted and actual values, and is a measure of the degree of difference between the estimator and the estimated quantity; MAPE (Mean Absolute Percentage Error) is a commonly used indicator to evaluate the prediction accuracy of regression models, used to measure the difference between predicted and actual values, and expressed as a percentage. The smaller the MAPE value, the better the prediction performance of the model. According to the results presented in Tables 1 and 2, the SSA-SVR prediction model provided by this invention can simultaneously achieve accurate prediction of temperature and humidity of Daqu (a type of Chinese liquor) in different seasons.

[0190] Comparative Example 1

[0191] Based on Example 2, this comparative example further compares the effects of using different prediction models to predict the temperature and humidity of Daqu (a type of Chinese liquor) in different seasons, as follows:

[0192] Based on Example 2, this comparative example further predicts the temperature and humidity of Daqu (a type of Chinese liquor) in different seasons using the PSO-SVR prediction model and the GA-SVR prediction model, respectively. The results are shown in Tables 3 and 4.

[0193] Table 3 Temperature Prediction MAE, MSE, and MAPE Values

[0194]

[0195] Table 4. Predicted humidity values ​​for MAE, MSE, and MAPE

[0196]

[0197] Combining Table 1, Table 2, Table 3, Table 4, and Figure 9 , Figure 10 , Figure 11 , Figure 12The results show that, when predicting the temperature of block A, compared with PSO-SVR, the MAE, MSE, and MAPE predicted using the SSA-SVR model were reduced by 19.41%, 38.4%, and 19.37%, respectively; compared with GA-SVR, the MAE, MSE, and MAPE predicted using the SSA-SVR model were reduced by 19.41%, 38.4%, and 19.37%, respectively. When predicting humidity, compared with PSO-SVR, the MAE, MSE, and MAPE predicted using the SSA-SVR model were reduced by 46.05%, 75%, and 46.1%, respectively; compared with GA-SVR, the MAE, MSE, and MAPE predicted using the SSA-SVR model were reduced by 73.12%, 93.67%, and 73.15%, respectively.

[0198] When predicting the temperature of block B, compared with PSO-SVR, the MAE, MSE, and MAPE predicted using the SSA-SVR model were reduced by 26.64%, 39.76%, and 26.67%, respectively; compared with GA-SVR, the MAE, MSE, and MAPE predicted using the SSA-SVR model were reduced by 42.84%, 69.7%, and 42.77%, respectively. When predicting humidity, compared with PSO-SVR, the MAE, MSE, and MAPE predicted using the SSA-SVR model were reduced by 48.16%, 78.99%, and 48.13%, respectively; compared with GA-SVR, the MAE, MSE, and MAPE predicted using the SSA-SVR model were reduced by 75.81%, 95.26%, and 75.82%, respectively.

[0199] When predicting the temperature of block C, compared with PSO-SVR, the MAE, MSE, and MAPE predicted using the SSA-SVR model were reduced by 43.04%, 58.16%, and 42.99%, respectively; compared with GA-SVR, the MAE, MSE, and MAPE predicted using the SSA-SVR model were reduced by 43.27%, 58.59%, and 43.20%, respectively. When predicting humidity, compared with PSO-SVR, the MAE, MSE, and MAPE predicted using the SSA-SVR model were reduced by 24.68%, 31.82%, and 24.74%, respectively; compared with GA-SVR, the MAE, MSE, and MAPE predicted using the SSA-SVR model were reduced by 32.95%, 43.04%, and 32.98%, respectively.

[0200] When predicting the temperature of block D, compared with PSO-SVR, the MAE, MSE, and MAPE predicted using the SSA-SVR model were reduced by 31.01%, 43.48%, and 30.98%, respectively; compared with GA-SVR, the MAE, MSE, and MAPE predicted using the SSA-SVR model were reduced by 35.27%, 49.35%, and 35.23%, respectively. When predicting humidity, compared with PSO-SVR, the MAE, MSE, and MAPE predicted using the SSA-SVR model were reduced by 39.13%, 48.65%, and 39.19%, respectively; compared with GA-SVR, the MAE, MSE, and MAPE predicted using the SSA-SVR model were reduced by 41.80%, 52.50%, and 41.83%, respectively.

[0201] Based on the above analysis of the temperature and humidity prediction of koji blocks in four seasons using different prediction models, SSA-SVR shows improvements in MAE, MSE, and MAPE compared to PSO-SVR and GA-SVR. It is evident that the SSA-SVR prediction model provided by this invention is significantly better at predicting the temperature and humidity of koji blocks and the temperature and humidity of koji blocks in different seasons.

[0202] It is understood that this invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of this invention. Furthermore, under the teachings of this invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of this invention. Therefore, this invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this invention are within the protection scope of this invention.

Claims

1. A method for predicting the temperature and humidity of koji blocks during fermentation, characterized in that, The method includes:

1. Collect temperature and humidity data for the three days prior to the fermentation of Daqu (a type of starter culture); 2. Perform missing data filling processing on the collected temperature and humidity data respectively; 3. Using the temperature data after missing data filling and / or the temperature and humidity data as input features, and the temperature of the koji block corresponding to one turning of the koji as the output feature, an SVR neural network model is trained to construct a temperature prediction model; or, using the temperature data and humidity data after missing data filling as input features, and the humidity of the koji block corresponding to one turning of the koji as the output feature, an SVR neural network model is trained to construct a humidity prediction model. Fourth, predict the temperature of the Daqu (fermented liquor) to be predicted during one turning of the fermentation process based on the temperature prediction model; or, predict the humidity of the Daqu to be predicted during one turning of the fermentation process based on the humidity prediction model.

2. The method according to claim 1, characterized in that, In step one, the sampling interval for collecting temperature and humidity data for the three days prior to the fermentation of the starter culture is 1-2 hours per sampling.

3. The method according to claim 1, characterized in that, The missing data filling process includes: using the KNN method to fill in the missing temperature and / or humidity data in the time series obtained in step one.

4. The method according to claim 3, characterized in that, The missing information filling includes the following steps: (1) Calculate the Euclidean distance between each data sample and the current missing value; (2) Based on the results calculated in step (1), select the 5 data samples with the shortest Euclidean distance, take the weighted average value as the imputation value; (3) Based on the filling value obtained in step (2), if the filling is not completed, return to step (1) and continue the missing filling process; If the filling is completed, the process ends and proceeds to step (4); (4) Based on the filling results of step (3) and the collected data, new time series data samples are generated. The newly generated time series data samples are expanded according to a sampling time frequency that is shorter than the time interval of the collected data samples to obtain the missing data. Preferably, the shorter sampling time frequency includes a sampling frequency of 5-15 min / time.

5. The method according to claim 4, characterized in that, The Euclidean distance between each data sample and the current missing value is calculated using the formula shown in Equation 1: In Equation 1, d i Representing point (x) i y i ) and point (x) j y j The Euclidean distance between x and y, where i represents the index of the i-th point, j represents the index of the j-th point, and x is the distance between x and y. i Let y represent the x-coordinate of the i-th point. i Let x represent the ordinate value of the i-th point. j y represents the x-coordinate of the j-th point. j This represents the ordinate value of the j-th point; Preferably, the fill value is calculated using the formula shown in Equation 2: In Equation 2, i represents the 5 data samples with the shortest Euclidean distance to the missing values, i.e., i = 1, 2, 3, 4, 5; d i x represents the Euclidean distance between the 5 data samples with the shortest Euclidean distance to the missing value and the missing value. i x represents the x-coordinate of the i-th sample point with the shortest Euclidean distance to the missing value, and x is the imputation value for the missing value.

6. The method according to claim 1, characterized in that, The method further includes: optimizing the constructed temperature prediction model / or humidity prediction model using the Sparrow Search Algorithm (SSA), and predicting the temperature / or humidity of the Daqu (a type of liquor) to be predicted during a single turning of the liquor based on the optimized temperature prediction model / or humidity prediction model.

7. The method according to claim 1, characterized in that, The method further includes: using the Sparrow Search Algorithm (SSA) to optimize the penalty parameter and the variance of the Gaussian kernel function in the constructed temperature prediction model / or humidity prediction model; based on the optimal values ​​of the penalty parameter and the variance of the Gaussian kernel function obtained after optimization, optimizing and training the temperature prediction model / or humidity prediction model to obtain the optimized temperature prediction model / or humidity prediction model; and predicting the temperature / or humidity of the Daqu to be predicted during a single turning of the Daqu based on the optimized temperature prediction model / or humidity prediction model. Preferably, the optimization of the penalty parameter and variance of the Gaussian kernel function in the constructed temperature prediction model / or humidity prediction model using the Sparrow Search Algorithm (SSA) includes the following steps: 1) Preprocess the data samples and divide the preprocessed data into training set and test set; 2) Set the population size N, the maximum number of generations T, initialize the parameters of the prediction model, and set the penalty parameter and the variance of the Gaussian kernel function in the prediction model as the optimization values ​​of the sparrow search algorithm. 3) Treat the penalty parameter and the variance of the Gaussian kernel function as sparrows, calculate the fitness of each individual, rank the fitness of each individual, select the sparrows with the highest fitness from the population as discoverers, and the remaining sparrows become followers; 4) Update the positions of the discoverer, follower, and watcher; 5) Compare the fitness of the new individual location with that of the original individual location, and retain the better individuals; 6) Determine whether the set number of iterations has been reached or whether the optimal solution that meets the accuracy requirements has been found. If the set number of iterations has been reached or the optimal solution that meets the accuracy requirements has been found, stop the search; otherwise, return to step 4) and continue the search. 7) Output the optimal values ​​of the penalty parameter and the variance of the Gaussian kernel function, and retrain the training data to obtain the optimized temperature prediction model / or humidity prediction model. Preferably, the optimization of the penalty parameter and the variance of the Gaussian kernel function in the constructed temperature prediction model / or humidity prediction model using the Sparrow Search Algorithm (SSA) further includes: 8) Input the validation set data into the optimized temperature prediction model / humidity prediction model for prediction, evaluate and analyze the performance of the prediction model.

8. The method according to claim 7, characterized in that, In step 4), the positions of the discoverer, follower, and vigilant are updated using the following formula, which includes: The location of the discoverer is updated using the formula shown in Equation 13: In Equation 13, This indicates the position of the discoverer at the t-th iteration. Let j represent the position of the discoverer in the (t+1)th iteration, j represent the discoverer's number, i represent the dimension of the position vector, and α be a random number in the range (0,1]. I iter,max R2 is the ratio of the current iteration count to the maximum iteration count. R2 is the warning value, which is a random number uniformly distributed in the interval [0,1). ST is the safety value, which ranges from [0.5,1]. L is the vector difference between the discoverer's position and the global optimal solution's position, which is a 1×m dimensional matrix where each element is 1. Q is a parameter that controls the step size. The position of the follower is updated using the formula shown in Equation 14: In Equation 14, Q is a random vector; X w This indicates the position of the current worst-case global solution. This represents the position of the current global optimal solution. A represents the Euclidean distance between the current sparrow and the global optimal solution; + Given a 1×m matrix, where each element is assigned the value 1 or -1, and A + =A T (AA T ) -1 L represents the side length of the current search space. The position of the watchman is updated using the formula shown in Equation 15: In Equation 15, X b f represents the location of the global optimal solution. j f represents the fitness value of the j-th sparrow; g f is the fitness value of the global optimal solution. b ε is the worst fitness value globally; β represents the control distance coefficient, which follows a normal distribution with a mean of 0 and a variance of 1; K represents the sensitivity coefficient of the vigilant, which is the direction of the sparrow's movement and takes a random number in the range [-1, 1]; ε is a minimum constant.

9. The method according to claim 1, characterized in that, In step four, the prediction of the temperature of the Daqu to be predicted during the first turning of the Daqu based on the constructed temperature prediction model includes: obtaining the temperature data or temperature and humidity data of the Daqu to be predicted for three days before entering the fermentation chamber as input features, inputting them into the temperature prediction model, and predicting the temperature of the Daqu to be predicted during the first turning of the Daqu based on the output results. Preferably, the method of predicting the humidity of the Daqu (a type of starter culture) during a first turning of the starter culture based on the constructed humidity prediction model includes: obtaining the temperature and humidity data of the Daqu before it is put into the fermentation chamber for three days as input features, inputting them into the humidity prediction model, and predicting the humidity of the Daqu during a first turning of the starter culture based on the output results; Preferably, the prediction model is as shown in equation (3): In Equation 3, i is the index of the i-th training sample, j is the index of the j-th training sample, k is the index of the k-th predicted sample, and x is the feature vector of the input sample. j Let α be the feature vector of the j-th training sample. i Let α be the weight of the j-th training sample. j * Let b be another set of weights for the j-th training sample, and y be the bias term. k This is the predicted output for the input sample x.

10. The application of the method as described in any one of claims 1-9 in the field of liquor brewing.

Images