High-temperature yeast fermentation process temperature prediction method based on deep learning

Abstract

The invention discloses a high-temperature yeast fermentation process temperature prediction method based on deep learning, and belongs to the technical field of yeast making technologies, and the method comprises the steps: collecting the core temperature of a yeast block in a fermentation bin and variable parameters related to the core temperature, taking the variable parameters collected at the same time point as a sample, storing the sample in an original data table, and processing the original data; determining a historical window and a prediction view of the model, using historical and current time series data of the processed data to construct a feature vector used for predicting the core temperature of the block, using the feature vector as a training sample of the model, and adopting a variable screening method based on an information theory to perform data dimension reduction on an input sample; modeling based on a long short-term memory neural network architecture, and inputting a training sample for training; and predicting the core temperature of the curved block in real time according to the collected variable parameters by using the trained prediction model. The whole fermentation condition in the fermentation bin can be quickly and accurately controlled, and key control points in the fermentation process are determined, so that the yeast turning process is optimized.

Description

Technical Field

[0001] The present invention relates to the field of fermentation engineering, and particularly to a method for predicting the temperature of the high-temperature Daqu fermentation process based on deep learning. Background Art

[0002] High-temperature Daqu, also known as soy sauce-flavored Daqu, is named because the highest temperature reached during the fermentation of the Qu cake is 60-65°C. When producing soy sauce-flavored liquor, the amount of Qu used is the largest among all liquors, and the ratio of grains to Qu medicine reaches 1:1. During the liquor-making process, Daqu is not only an important saccharifying, fermenting, and flavor-producing agent, but also a raw material for the fermentation process, providing a basis for the formation of the soy sauce flavor characteristics of soy sauce-flavored Baijiu. During the Daqu fermentation process, heat and CO 2 are generated due to the metabolic activities of microorganisms, causing the temperature of the Qu cake to rise and moisture to be lost, thereby causing changes in environmental temperature and humidity, air components and other parameters, and the environmental parameters in turn affect the Daqu fermentation. Therefore, analyzing the variation law of environmental parameters during the cultivation process of high-temperature Daqu is of great significance for controlling the Daqu fermentation state and precisely controlling the Qu-making process. The regulation of the Daqu fermentation process lacks objective data guidance. Traditional Qu-making mainly relies on personal experience, regularly checking the fermentation degree and appearance of the Qu cake, and then adjusting specific fermentation environment parameters through operations such as turning the Qu and opening the window to exhaust moisture. The existing Daqu turning time is summarized by Qu-making workers based on experience. Since the temperature change trends of Daqu in the fermentation warehouse are not consistent under different seasons and different operations, workers lack a scientific and accurate judgment method for the best turning time. So far, no method for predicting the temperature of the Qu cake during the Daqu fermentation process has been reported. Therefore, it is urgent to establish the relationship between the changes in environmental variables such as the temperature, humidity, CO 2 , O 2 in the fermentation warehouse and the center temperature of the Qu cake for scientific guidance of the high-temperature Daqu fermentation process.

[0003] In the early stage in China, the temperature, humidity and other parameters at individual points in the fermentation warehouse were monitored, and discrete sampling was carried out according to time periods to fit the relationship between the environmental temperature in the fermentation warehouse and the center temperature of the Qu cake, but the gas concentration in the fermentation warehouse and external environmental variables were not considered. The research on the parameter simulation of the Qu house environment is still in the laboratory stage. Researchers used Fluent fluid simulation software to model and simulate the spatial structure layout of the experimental Qu house, and calculated and solved the Qu house environment with the Qu cake as the heat source to obtain the temperature and velocity distribution cloud maps under different air inlet and outlet methods and different Qu cake placement forms; by studying the physical phenomena involved in the related changes in temperature and humidity in the Qu house, and combining theories such as computational fluid dynamics and heat and mass transfer, a natural convection heat and moisture transfer simulation model of the Qu house and a heat and moisture transfer simulation model of the Qu house under the influence of the flow field were established. The models established under laboratory conditions in the above research results have a large difference in the number of samples and spatial structure from the actual production environment, and the guiding significance of the models is not strong.

[0004] In the prior art, Patent CN114840807A discloses "A Method for Predicting the Temperature of Qu Blocks during Daqu Fermentation Process". This patent classifies Daqu of different rounds according to the temperature in the fermentation warehouse, specifically divided into spring and autumn season Qu, summer Qu, and winter Qu. Then, it establishes a fitting model between the fermentation days and the temperature of the Qu block, and further predicts the fermentation temperature of the Qu block to prejudge in advance whether the temperature of the Qu block meets the requirements for turning the Qu. This patent does not understand the Qu block fermentation process. During the Qu block fermentation process, heat and CO 2 are generated due to the microbial metabolic activities, causing the temperature of the Qu block to rise and the moisture to be lost, thus causing changes in parameters such as the environmental temperature and humidity and the air components. The environmental parameters in turn continue to affect the Daqu fermentation. Therefore, relying solely on the temperature in the fermentation warehouse for classification is not perfect.

[0005] Patent CN116822355A discloses "A Method for Real-time Online Monitoring and Controlling the Quality of Daqu Fermentation". This patent takes the physical and chemical indexes of the Qu block quality as the goal, establishes a connection between the changes in environmental parameters and the physical and chemical indexes of the Qu block quality, and then establishes a BP neural network model to realize the online detection of the Qu block quality (physical and chemical indexes). However, the fermentation process is a continuous process. The environmental parameters and the Qu block quality at the previous several moments will have an impact on the subsequent fermentation quality of the Qu block. The BP neural network model is a static model, and this patent does not consider the influence of the environmental parameter state and the Qu block fermentation state at the previous several moments on the Qu block fermentation state at the current moment during the continuous fermentation process of the Qu block. Summary of the Invention

[0006] The present invention aims to solve the problems existing in the prior art in the prediction research method for the core temperature of Qu blocks during the high-temperature Daqu fermentation process, and proposes a temperature prediction method for high-temperature Daqu fermentation process based on deep learning. This method comprehensively collects the core temperature of Qu blocks and the related variable parameters, and constructs a time series prediction model with variable parameters as the input and the core temperature of Qu blocks as the output for prediction, which can quickly and accurately grasp the overall fermentation situation in the fermentation warehouse, clarify the key control points in the fermentation process, and thus optimize the turning Qu process.

[0007] In order to achieve the above-mentioned invention purpose, the technical solution of the present invention is as follows: The present invention provides a temperature prediction method for high-temperature Daqu fermentation process based on deep learning, including the following steps: Data collection step, collecting the core temperature of Qu blocks in the fermentation warehouse and the related variable parameters at a set frequency, including the CO 2 concentration in the warehouse, the O 2 concentration in the warehouse, the temperature in the warehouse, the humidity in the warehouse, the window wind speed, the outdoor temperature, and the outdoor humidity; Data processing step: The core temperature of the curved block collected at the same time point and the corresponding variable parameters are stored as a sample in the original data table, and the error, invalid, abnormal, and missing data in the original data table are processed; Feature construction step: Determine the historical window and prediction horizon of the model, use the historical and current time series data of the processed data to construct a feature vector for predicting the core temperature of the curved block as the training sample of the model, and use a variable screening method based on information theory to reduce the dimension of the initial input sample in the training sample; Model training step: Build a sequence-to-sequence prediction model based on the long short-term memory neural network architecture, and use the sample set after data dimension reduction to train and test the model; Online prediction step: Use the trained prediction model to predict the core temperature of the curved block in real time according to the currently collected variable parameters.

[0008] Furthermore, the processing of the error, invalid, abnormal, and missing data in the original data table includes: correcting and deleting the obvious errors and invalid values in the original data table, identifying and deleting the potential abnormal values for the abnormal values by combining the empirical values with the descriptive statistical features of each variable field, and then using the multiple missing data filling method based on random forest to fill the missing data. The filling method is: first, extract the samples with complete field information from the original data table after abnormal identification processing to construct a new data set, then use the new data set to establish a random forest prediction model for each parameter, and predict and fill the missing values by combining the random forest with other parameter data.

[0009] Furthermore, in the feature construction step, at each moment t , intercept the corresponding historical change sequence according to the historical window of each variable, and splice all variable sequences as the initial input sample of the model; then obtain the change sequence of the target variable in the future for a period of time according to the prediction horizon as the output sample of the model.

[0010] Furthermore, the mathematical expression between the input sample and the output sample is: ; where represents the continuous change of the target variable y in the future for a period of horizon , is the current and historical window sequence of the target variable y, indicating the vector of the change of y from time to ; ~ is the current and historical window sequence of the related variable X; the parameters , ...... denotes the length of the historical window sequence, denotes the time instant.

[0011] Furthermore, the historical window parameter of the relevant variable X is determined by the autocorrelation function based on mutual information, and its calculation formula is: ; In the formula, denotes the i-th relevant variable, is the historical window parameter, MI denotes mutual information; denotes probability.

[0012] Furthermore, when determining the historical window parameter of the relevant variable X, to avoid the interference of accidental factors in the data samples on the autocorrelation function and the window estimation results, it is also necessary to obtain a stable and reliable statistical test of the autocorrelation function and the historical window based on Bootstrap sampling, and obtain the final historical window parameter of the relevant variable X, which specifically includes the following steps: Suppose to perform rounds of Bootstrap sampling. In each round of sampling, sample with replacement from the time series samples at the current time instant and the time series samples corresponding to the time delay to extract samples of the same scale, denoted as and and ; Calculate the autocorrelation function for each round, that is ; Obtain the mean, variance, and confidence interval of all rounds of autocorrelation functions; Suppose the significance level is , and use the one-sided test method to perform a statistical test on the results. If the autocorrelation function obtained by Bootstrap is significantly greater than 0, it is considered that the historical window value is significant. At this time, the new autocorrelation function is: ; Use the above new autocorrelation function to calculate and obtain the historical window parameters of each variable, and obtain the final historical window parameter of the relevant variable X.

[0013] Furthermore, the prediction horizon of the target variable y is jointly determined by the target variable y itself and the time series effect of other relevant variables X on the target variable y. The longest horizon that all variables can provide is selected as the theoretical prediction horizon of the target variable y.

[0014] Furthermore, the length of the prediction horizon determined by the target variable y itself is calculated using the autocorrelation function ACF based on mutual information. The prediction horizon of the relevant variable X for the target variable y is calculated according to the autocorrelation function as: , Even the minimum delay value that reduces the ACF to near 0.

[0015] Furthermore, the data dimensionality reduction of the training samples using the variable screening method based on information theory includes: Let all candidate variable sets be , first determine a part of the modeling variable set based on business experience and expert knowledge; then, adopt the forward incremental variable selection strategy to select from the remaining variable set the variables that can improve the predictive information amount about the target variable , until all remaining variables cannot provide additional improvement in predictive information, and then terminate. The selected variables and the business experience variables together are used as the modeling variable set to participate in the subsequent predictive modeling of the target variable .

[0016] Furthermore, in the model training step, the sample set after data dimensionality reduction is divided into a training set and a test set according to a ratio of 7:3. The training set is used for forward calculation, and the obtained prediction results are used to calculate the loss function determined by the error with the actual target values in the training set. Then, an optimizer is used to update the model parameters through backpropagation. When the loss function decreases to the lowest with the parameter update iteration, the training ends; finally, the test set data is substituted into the model to obtain the prediction output, and it is compared with the actual target results in the test set to calculate indexes such as prediction accuracy and error, and the model evaluation is completed.

[0017]

[0017] In summary, the present invention has the following advantages: 1. The present invention can quickly and accurately grasp the overall fermentation situation in the fermentation tank, clarify the key control points in the fermentation process, and thus optimize the turning process by comprehensively collecting the core temperature of the koji block and the variable parameters related to the core temperature of the koji block, and constructing a time series prediction model with the variable parameters as the input and the core temperature of the koji block as the output for prediction; 2. The present invention collects the characteristic data related to modeling at different spatial positions, different koji pile areas, and typical koji blocks in the fermentation tank, including the core temperature of the koji block and the in-tank CO 2 concentration, the in-tank O 2 concentration, the in-tank temperature, the in-tank humidity, the window wind speed, the outdoor temperature, and the outdoor humidity. The collected data is more comprehensive and has a stronger correlation with the process; 3. When processing the collected original data, the present invention identifies, corrects, deletes, etc. possible errors, invalidities, missing values, and abnormal data in the original data table, and based on the multiple missing data filling method of random forest, explores and utilizes the linear and non - linear correlation information between the missing data and other field data to predict and fill the missing values, thereby improving the data quality and ensuring the accuracy, integrity, and consistency of the data; 4. The present invention constructs a neural network model of deep learning with long short - term memory (LSTM). When constructing the training samples, the environmental parameters and the state of the koji blocks at the previous several moments are incorporated into the model, making the prediction more scientific and accurate; 5. When determining the historical window parameters of each variable, the present invention uses mutual information combined with the autocorrelation function of Bootstrap to measure any linear and non - linear correlations between the fermentation process data, and can obtain more stable and reliable results, thereby improving the prediction performance of the training model; 6. The present invention uses a variable screening method based on information theory to perform data dimensionality reduction on the training samples without transforming the data, thus avoiding the loss of prediction information during the transformation process and having better prediction accuracy and interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a schematic diagram of data collection for the present invention; Figure 2 is a technical implementation roadmap of the present invention; Figure 3 shows the time - series changes of each variable signal; Figure 4 shows the composition of the input samples and output samples; Figure 5 shows the main structure of the neural network; Figure 6 shows the calculation logic of the memory unit module cell pointed to by box A in the neural network structure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] To more clearly illustrate the present invention, the following further describes the present invention in conjunction with the preferred embodiments and the drawings. Those skilled in the art should understand that the content specifically described below is illustrative rather than restrictive, and should not be used to limit the protection scope of the present invention.

[0020] The present invention provides a temperature prediction method for the high - temperature Daqu fermentation process based on deep learning, as Figure 2 shown, including the following steps: Step 1. Data collection Deploy a digital monitoring system for the fermentation silos, collect measurable characteristic data at key positions, doors, windows, and the external environment of the fermentation silos, and send the collected characteristic data to the data acquisition system. Then, send the real-time data to the server through a redundant switching network. The architecture of the digital monitoring system is as shown in Figure 1 . The digital monitoring system for the fermentation silos integrates functions such as data acquisition, storage, and visualization. It can monitor and record parameters such as the temperature of the koji blocks, and the temperature, humidity, etc. inside the silos in real time during the koji-making fermentation process. The data can be adjusted according to minutes, hours, and days.

[0021] During specific implementation, the air volume and heat exchange volume of the window can be collected and calculated by installing an anemometer at the window; the external environment parameters can be detected by arranging temperature and humidity sensors on the outer side of the fermentation silo building; the temperature changes at three points, namely the upper, middle, and lower parts of the koji blocks, can be detected by inserting from the convex side of the koji blocks to a depth of about 7 cm; temperature and humidity sensors, as well as oxygen and carbon dioxide gas concentration detection devices, can be installed on the top of the fermentation silo to monitor and analyze the distribution of the environmental parameters inside the silo. Finally, the characteristic data of the fermentation silos collected are transmitted to a server that uses a high-performance parallel computing form of multiple GPUs + CPUs. In the present invention, the server has functions such as data storage, historical data query, large-screen display, deep learning model training and deployment, and visual three-dimensional interface.

[0022] During the implementation of this step, the sensors deployed on-site are combined with simulation to sample and record the characteristic data related to modeling at certain time intervals (30 minutes in this embodiment). The important collected variable information is shown in Table 1 below.

[0023] Table 1 Modeling data information

[0024] As can be seen from Table 1, all the variables used for prediction and the target variables during the process have actual physical meanings and are continuous variables. Therefore, corresponding methods are adopted for this continuous variable in the subsequent data cleaning, dimensionality reduction, modeling, and evaluation processes.

[0025] Step 2: Data processing The purpose of this step is to improve the data quality by processing errors, invalid values, anomalies, and missing values in the original data to ensure the accuracy, integrity, and consistency of the data, including two sub-steps: abnormal data identification and missing data filling.

[0026] The original data table as shown in Table 2 is obtained by collecting data according to the variables in Table 1. Each row in the table corresponds to a sampling sample, each column corresponds to a variable field, and the field t in the first column is the sampling time.

[0027] Table 2 Original data table format

[0028]

[0029] Identify possible errors, invalidities, missing values, and abnormal data in the original data table. According to business knowledge and experience, correct and delete obvious error records in the table. In addition, potential outliers can be identified by combining the descriptive statistical characteristics of each variable field in the table. For example, for variables with a normal distribution, outliers can be detected through the principle to check for outliers in the data; for variables with a non-normal distribution, box plots and other methods can be used to check for outliers.

[0030] After the above checks, mark and delete all positions in the original data table that are determined to be outliers, and regard them as missing values. Then, calculate the missing rate for each field in the processed data table. If the field missing rate is greater than 20%, the field is directly deleted; otherwise, the field is retained and the missing values are filled.

[0031] In this embodiment, considering that there may be mutual coupling between different variables, filling only the information of a single variable field may result in the loss of the associated information between fields in the filled data, which may further lead to problems such as reduced accuracy in analysis and modeling. To address this, this solution uses a multiple missing data filling method based on random forest (RF): First, extract the samples with complete field information from the data table after anomaly identification and processing to construct a new data set , and then use to establish a random forest prediction model for each field. If a sample field has missing values, the missing values are predicted and filled by combining the random forest with the data of other fields. This filling method can maximize the exploration and utilization of the linear and non-linear association information between the missing data and the data of other fields.

[0032] Step 3: Feature construction In this step, the processed data is converted into a feature data set that can be used by machine learning and deep learning models, including a feature construction sub-step and a data dimensionality reduction sub-step.

[0033] The effect of feature construction is strongly correlated with the scenario it faces and the selection of the prediction model. For the koji-making process, there may be complex non-linear time-series coupling relationships between the predicted target core temperature of the koji block and other factors such as temperature, humidity, gas composition, and wind speed during the process. For example, the temperature, humidity, O 2 and CO 2Concentrations, etc. will affect the fermentation process inside the koji blocks and the changes in the core temperature. The heat generated during fermentation at the core temperature of the koji blocks will in turn affect the changes in other variables. Further, considering that the fermentation chamber is a relatively large physical space, the transmission of the influence between variables may also have non-instantaneity, resulting in a certain time lag and inertia. In order to achieve as accurate a prediction as possible for the temperature changes of the koji blocks, it is necessary to include the information with characteristics such as time lag and inertia in the modeling feature set. Therefore, this embodiment uses the historical and current time series data in Table 1 and to predict the changes within a future period of time .

[0034] Figure 3 shows the relationship between the prediction modeling data and the target from the perspective of the time series changes of each variable signal. Each row in the figure corresponds to the time series change of a variable. Among them, the core temperature of the koji block at the current moment is denoted as ; its historical window sequence for prediction is denoted as , indicating the vector of changes from time to . The parameter represents the length of the historical window sequence for prediction modeling. The model needs to predict the continuous changes of the target variable within a future prediction horizon , corresponding to the vector . For other modeling variables , their current values are ~ , and the historical window sequences are ~ .

[0035] As shown in Figure 4 , the prediction model will use the current values and historical window sequence values of the modeling variables , , …, to predict the continuous changes within a future period . The mathematical expression is: ; where the undetermined parameters are the historical window lengths , , …, of each variable and the prediction horizon , with the unit being the number of sample points

[0036] Different from the characteristics of discrete production in general manufacturing, the Daqu fermentation process has a long cycle, many influencing factors, and exhibits complex long-term coupling and dependence relationships. Therefore, a modeling method that can adapt to the above process characteristics is required. Among the relevant methods, time series prediction based on deep learning has received extensive attention and application in recent years due to its excellent modeling effect. In the construction of deep learning models such as recursive neural network (RNN) and long short-term memory (LSTM), the historical window and the prediction horizon as external hyperparameters independent of the model have an important impact on the prediction performance. On the one hand, a longer historical window can incorporate sufficient time series information for prediction, but it will also increase the complexity and training difficulty of the prediction model and reduce the computational efficiency; on the other hand, the prediction horizon can help people anticipate and control interventions in the future changes of the process in advance; but as the value increases, the uncertainty of the model prediction will continue to increase, and the complexity, training difficulty, and computational burden of the model will also continue to increase. Therefore, for the Daqu fermentation process, appropriate settings of the historical window and the prediction horizon are very important for the model effect.

[0037] From the perspective of information theory, and respectively reflect the propagation ability of the variable state information at the current moment along the time in the reverse and forward directions. The parameters of each variable can be determined by methods such as the auto-correlation function (ACF). If the information of a certain variable is propagated linearly in time series, the ACF based on the Pearson correlation coefficient can be used to calculate its parameters. However, considering that there may be linear and non-linear correlations between data at the same time, the present invention uses mutual information (MI) to replace the linear Pearson coefficient and proposes the following auto-correlation function calculation method: — Equation (1).

[0038] Furthermore, on the basis of mutual information, in order to avoid the interference of accidental factors in the data samples on the auto-correlation function and the window estimation results, the present invention also designs a method for obtaining a stable and reliable historical window for the statistical test of the auto-correlation function based on Bootstrap sampling. The implementation steps are as follows: First, assume that rounds of Bootstrap sampling are performed. In each sampling, samples of the same scale are drawn with replacement from the time-series samples at the current moment and the time-series samples corresponding to the time delay . Denote them as and ; ; Next, calculate the autocorrelation function for each round, that is, ; Then, obtain statistical indicators such as the mean, variance, and confidence interval of the autocorrelation functions for all rounds. Assume the significance level is (usually taken as 0.05, corresponding to a 95% confidence interval). Use a one-sided test to perform a statistical test on the results. If the autocorrelation function obtained by Bootstrap is significantly greater than 0, it is considered that the historical window value is significant. At this time, the autocorrelation function is: .

[0039] The autocorrelation function combining mutual information and Bootstrap described above can measure any linear and non-linear associations between fermentation process data and obtain more statistically stable and reliable results.

[0040] On the other hand, the time-series action information of the target variable itself and other variables on jointly determines the prediction horizon . Among them, the prediction horizon length that itself can determine is calculated based on the autocorrelation function of mutual information (i.e., Equation (1)), while the prediction horizon of other variables on can be calculated according to the following formula as: , that is, the minimum time delay value that reduces the ACF to near 0. Finally, the theoretical prediction horizon of should be the longest horizon that all variables can provide:

[0041] When the parameters such as the historical window and prediction horizon of each variable are determined, the prediction input samples can be obtained in the Figure 4 way: that is, at each moment t, intercept the corresponding historical change sequence according to the historical window parameters of each prediction variable, and then splice all variable sequences as the initial input sample of the model; then obtain the change sequence of the target in the future for a period of time according to the prediction horizon as the output sample of the model.

[0042] Furthermore, since there are temporal correlations among the variable signals, there may be redundant information among the variable features in the obtained input samples. This not only fails to further increase the amount of information for model prediction but also increases the complexity of the input data and the model structure parameters, which is not conducive to modeling applications. Therefore, it is necessary to select or transform the features in the initial input samples through data dimensionality reduction techniques to reduce the variable dimension while retaining the effective information therein, thereby achieving refined prediction.

[0043] The present invention combines business experience with data and proposes a variable screening method based on information theory for process modeling, which reduces its complexity while ensuring the prediction accuracy of the model, and improves interpretability, generalization ability, and robustness. The implementation steps are as follows: Let all candidate variable sets be , and first determine a part of the modeling variable set based on business experience and expert knowledge; Next, further select variables from the remaining variables that can increase the amount of information for predicting to improve the prediction accuracy. The present invention adopts a forward incremental variable selection (IVS) strategy to select the remaining variables: assume that at the -th step, the selected variable set is , which includes the business experience variable and the variables screened out through data in the previous steps. Then, the variable selected in this round should achieve the maximum additional information increase regarding the target on the basis of the already selected variables, that is: ; where the mutual information MI has been defined above. Next, can be added to the selected variable set : ; Continuously repeat the above variable selection process until the iteration terminates when all remaining variables cannot provide additional improvement in prediction information: ; At this point, the selected features and the business experience variable together serve as the modeling variable set to participate in the subsequent prediction modeling of .

[0044] The variable selection method proposed by the present invention can be applied to industrial process data with complex non - linear correlation effects, and is different from traditional data - dimensionality - reduction algorithms such as independent principal component analysis. It does not require data transformation, thus avoiding the loss of prediction information during the transformation process and having better prediction accuracy and interpretability.

[0045] Step Four: Deep - learning modeling Since the present invention involves the problem of multi - variable multi - step time - series prediction, a sequence - to - sequence (Seq2Seq) prediction model based on the long short - term memory (LSTM) neural network architecture is selected for modeling.

[0046] As Figure 5 shown, it presents the main structure of the Seq2Seq neural network. Each moment corresponds to a memory - cell module. The whole network is formed by connecting memory - cell modules at different moments in series along the time order, corresponding to Figure 3 correspondingly. Figure 6 It shows Figure 5 the calculation logic of the memory - cell module cell pointed by box A in , which contains three gates: the forget gate, the input gate, and the output gate. The memory - cell module at each moment receives the input tensor at this time, the weight matrix , the bias , the hidden state at the previous time step, and the memory - cell state , and then returns the hidden state at the current time step and the updated memory - cell state . In this example, the input of the memory - cell module at moment corresponds to the core temperature of the block at moment, the signal obtained after data dimensionality reduction, ,...

[0047] The entire Seq2Seq network is formed by connecting an encoder and a decoder in series, which contain and memory cells respectively. The theoretical values of the number of module units in the encoder - decoder are: . The actual setting also needs to be determined in combination with factors such as application accuracy and calculation efficiency. For the prediction task at moment , first, all variables and variables from moment to is organized into tensors and sent one by one to the corresponding memory unit modules of Seq2Seq. Then, in the order shown by the red part on the left, the hidden states of each memory unit module Cell in the encoder are calculated sequentially from left to right Figure 5 and the memory states as shown in the left red part, calculate the hidden states of each memory unit module Cell in the encoder sequentially from left to right . For example, the memory unit corresponding to time will receive the current , , , …, as inputs, as well as the and from the previous time step, and calculate the and for this time step

[0048] Next, perform the decoder calculation. Use the hidden state output at time as the input to the first memory unit module of the decoder. Combine and to obtain the output and of this unit module, and extract the corresponding variable value from as the first prediction result . And so on, complete the calculation of the remaining decoder Cells, and finally obtain the prediction results for the change of from to

[0049] The offline modeling of Seq2Seq is divided into two stages: training and testing. First, divide all the and samples after dimensionality reduction into a training set and a test set according to a ratio of 7:3. Use the training set to perform the above-mentioned predictive forward calculation, and calculate the loss function determined by the error between the obtained prediction results and the actual target values in the training set. Use optimizers such as stochastic gradient descent (SGD) and Adam to perform backpropagation updates on the model parameters. When the loss function decreases to the lowest with the parameter updates and iterations, end the training. Next, substitute the test set data into the model to obtain the predicted output, and compare it with the actual target results in the test set to calculate indicators such as prediction accuracy and error, and complete the model evaluation

[0050] Since the targets in this example are all continuous variables, the mean absolute error MAE, mean square error MSE, root mean square error RMSE, root mean square logarithmic error RMSLE, mean absolute percentage error MAPE, coefficient of determination R​2 Evaluate the training effect of the model using accuracy and error evaluation metrics.

[0051] Step Five: Online Prediction After Step Four, the trained Seq2Seq model can be obtained and stored as a file. It is loaded in the actual environment, deployed on a high-performance workstation, and used to predict the temperature of the curved block in real time according to the collected data.

[0052] Specifically, at each prediction moment , obtain the historical and current window sequences of each and variables with a length of . Then, perform anomaly identification and missing value filling on the data of each window according to the method in Step Two. Next, perform dimensionality reduction on , , …, and according to the dimensionality reduction method in Step Three. Suppose the window samples of , , …, after dimensionality reduction are . Input them together with , , …, into the Seq2Seq model, and finally obtain the prediction result of the core temperature of the curved block .

[0053] The above is only a preferred embodiment of the present invention, and does not impose any form of limitation on the present invention. Any simple modification or equivalent change made to the above embodiments based on the technical essence of the present invention falls within the protection scope of the present invention.

Claims

1. A temperature prediction method for high-temperature Daqu fermentation process based on deep learning, characterized in that: The steps include: Data collection step, collecting the core temperature of the fermentation block and related variable parameters at a set frequency, the related variables include CO2 concentration in the warehouse, O2 concentration in the warehouse, temperature in the warehouse, humidity in the warehouse, window wind speed, outdoor temperature and outdoor humidity; Data processing step: storing the core temperature of the block and the corresponding variable parameters collected at the same time point as a sample in the original data table, and processing the erroneous, invalid, abnormal and missing data in the original data table; The feature construction step determines the historical window and prediction horizon of the model, uses the historical and current time series data of the processed data to construct the feature vector for predicting the core temperature of the block as the training sample of the model, and uses the variable screening method based on information theory to reduce the data dimension of the initial input samples in the training samples; Model training step: build a sequence-to-sequence prediction model based on the long short-term memory neural network architecture, and use the sample set after data dimension reduction to train and test the model; In the online prediction step, the trained prediction model is used to predict the core temperature of the block in real time based on the currently collected variable parameters.

2. The method for predicting temperature of a high-temperature Daqu fermentation process based on deep learning according to claim 1, characterized in that: The processing of erroneous, invalid, abnormal and missing data in the original data table includes: correcting and deleting obvious errors and invalid values ​​in the original data table, identifying and deleting potential abnormal values ​​by using empirical values ​​combined with the descriptive statistical characteristics of each variable field, and then filling the missing data with a multiple missing data filling method based on random forest. The filling method is: first, extracting samples with complete field information from the original data table after abnormal identification processing to construct a new data set, and then using the new data set to establish a random forest prediction model for each parameter, and predicting and filling the missing values ​​through random forest combined with other parameter data.

3. The method for predicting temperature of high-temperature Daqu fermentation process based on deep learning as claimed in claim 1, characterized in that: In the feature construction step, at each time , according to the historical window of each variable, the corresponding historical change sequence is intercepted, and all variable sequences are spliced ​​as the initial input sample of the model; then the change sequence of the target variable in a period of time in the future is obtained according to the prediction horizon as the output sample of the model.

4. The method for predicting temperature of a high-temperature Daqu fermentation process based on deep learning as claimed in claim 3, characterized in that: The mathematical expression between input samples and output samples is: ； in, Indicates the target variable y in the future horizon Continuous changes within is the current and historical window sequence of the target variable y, indicating that y has been to The vector of change; ~ is the current and historical window sequence of the relevant variable X; parameter , ...... Indicates the length of the historical window sequence, Indicates time.

5. The method for predicting temperature of a high-temperature Daqu fermentation process based on deep learning according to claim 1 or 4, characterized in that: History window parameters of the relevant variable X It is determined by the autocorrelation function based on mutual information, and its calculation formula is: ； In the formula, represents the i-th related variable, is the historical window parameter, MI represents mutual information; Represents probability.

6. The method for predicting temperature of a high-temperature Daqu fermentation process based on deep learning according to claim 5, characterized in that: When determining the historical window parameters of the relevant variable X, in order to avoid the interference of accidental factors in the data sample on the autocorrelation function and the window estimation results, a stable and reliable autocorrelation function statistical test historical window is obtained based on Bootstrap sampling to obtain the final historical window parameters of the relevant variable X, which specifically includes the following steps: Set up Bootstrap sampling round, in each round In sampling, the time series samples from the current moment are replaced and the corresponding delay A time series sample We draw samples of the same size from and ; Calculate the autocorrelation function of each round, that is ; Get All The mean, variance, and confidence interval of the round autocorrelation function; Set the significance level to , a one-sided test is used to perform a statistical test on the results. If the autocorrelation function obtained by Bootstrap is significantly greater than 0, the historical window value is considered significant. At this time, the new autocorrelation function is: ； The new autocorrelation function is used to calculate the historical window parameters of each variable, and the final historical window parameters of the related variable X are obtained.

7. The method for predicting temperature of high-temperature Daqu fermentation process based on deep learning as claimed in claim 5, characterized in that: The prediction horizon of the target variable y is determined by the target variable y itself and the temporal effects of other related variables X on the target variable y. The longest horizon that all variables can provide is selected as the theoretical prediction horizon of the target variable y.

8. The method for predicting temperature of high-temperature Daqu fermentation process based on deep learning according to claim 7, characterized in that: The predicted horizon length determined by the target variable y itself The autocorrelation function ACF based on mutual information is used for calculation. The prediction horizon of the relevant variable X for the target variable y is calculated according to the autocorrelation function as follows: , That is, the ACF is reduced to the minimum delay value near 0.

9. The method for predicting temperature of a high-temperature Daqu fermentation process based on deep learning according to claim 1, characterized in that: The variable screening method based on information theory is used to reduce the dimension of training samples, including: Suppose the set of all candidate variables is First, determine a set of modeling variables based on business experience and expert knowledge ; Then, the forward incremental variable selection strategy is used to select the remaining variables from the set Select the ones that can improve the target variable Variables that predict information , until all remaining variables can no longer provide additional prediction information, and the selected variables and business experience variables are combined. Together as a set of modeling variables to participate in the subsequent target variable predictive modeling.

10. The method for predicting temperature of high-temperature Daqu fermentation process based on deep learning according to claim 1, characterized in that: In the model training step, the sample set after data dimensionality reduction is divided into training set and test set in a ratio of 7:

3. The training set is used for forward calculation, and the obtained prediction result is compared with the actual target value in the training set to calculate the loss function determined by the error. Then the optimizer is used to back-propagate and update the model parameters. The training is terminated when the loss function drops to the minimum with the iteration of parameter update. Finally, the test set data is substituted into the model to obtain the predicted output, which is compared with the actual target result in the test set, and the prediction accuracy and error index are calculated to complete the model evaluation.

Images