CNN-LSTM hybrid model-based yeast fermentation environment prediction method and application thereof
Through the combination of CNN-LSTM mixed model and fuzzy PID controller, the traditional method is solved to capture the insufficient timing and nonlinear relationships in the Daqu fermentation environment, and the precise regulation of the Daqu fermentation environment is achieved, and the quality and wine production rate of liquor are improved.
Patent Information
- Application Number
- CN202510406538.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-02
- Publication Date
- 2025-08-01
AI Technical Summary
Traditional prediction methods cannot effectively capture the timing and nonlinear relationship in the Daqu fermentation environment, resulting in the disconnection of the temperature and humidity prediction value from the actual dynamic evolution process, the real-time and computational efficiency are inefficient, and the regulation strategies lack dynamic adaptability, which affects the quality of liquor and the yield rate.
The CNN-LSTM hybrid model is used to combine random forest screening features and fuzzy PID controller for real-time regulation. The spatial features and long and short-term memory network capture timing dependence are extracted through convolutional neural networks, and the key points of random forest screening are combined to optimize data processing and prediction accuracy, and dynamically adjust temperature and humidity.
The prediction accuracy and real-time regulation of the Daqu fermentation environment are improved, the uniformity and stability of the fermentation process are ensured, and the quality and wine production rate of liquor are improved.
Smart Images
Figure CN120408397A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of brewing, and specifically relates to a prediction method for the Daqu fermentation environment based on a CNN-LSTM hybrid model and its application. Background Art
[0002] As the core fermenting agent for Chinese liquor brewing, the quality of Daqu is closely related to the liquor yield of Chinese liquor. The temperature and humidity environment in the koji-making room is one of the key factors affecting the quality of Daqu and the liquor-making effect. By optimizing the selection of Daqu raw materials, improving the koji-making process, and precisely controlling the temperature and humidity in the koji-making room, the quality of Daqu can be significantly improved, and the liquor yield and quality of Chinese liquor can be increased. With the continuous development and technological innovation of the liquor industry, it is necessary to further strengthen the refined research and intelligent management and control of Daqu and the koji-making room environment, deeply explore the interaction mechanism between microorganisms and environmental factors, and provide strong support for the efficient, high-quality, and sustainable development of Chinese liquor brewing.
[0003] In the field of predicting the Daqu fermentation environment, some current distilleries still rely on prediction methods based on traditional mathematical models, such as Linear Regression, Partial Least Squares Regression (PLSR), Grey Prediction Model, and Support Vector Machine (SVM). These methods construct static prediction models through historical temperature and humidity data, attempting to reveal the potential relationship between environmental parameters and fermentation effects. However, the fermentation environment in the Qu room has significant characteristics of time series, non-linearity, and spatial variability. Traditional models expose multiple limitations when dealing with these complex characteristics: (1) Insufficient time series modeling: During the fermentation process, the dynamic changes of temperature and humidity have strong time-dependent characteristics. For example, the heat accumulation caused by microbial metabolic activities has a lag effect. The current environmental state is not only affected by immediate parameters but is also closely related to historical data. Traditional models (such as linear regression and SVM) are mostly based on static assumptions and cannot effectively capture the long-term dependence relationships in time series (such as periodic fluctuations and trend accumulation), resulting in the disconnection between the predicted values of temperature and humidity and the actual dynamic evolution process. (2) Weak ability to fit non-linear relationships: Daqu fermentation involves the coupled effects of multiple factors such as microbial community metabolism and enzyme activity changes. There is a complex non-linear interaction relationship between temperature, humidity, and fermentation efficiency. For example, an increase in temperature may promote the growth of microorganisms in a certain stage, but it will inhibit the activity after exceeding the threshold. Traditional linear regression and PLSR models are limited by linear assumptions and are difficult to characterize such non-linear mechanisms. Although SVM can handle some non-linear problems through kernel functions, its computational complexity is high, parameter tuning is difficult, and its ability to extract features from high-dimensional time series data is limited. (3) Low real-time performance and computational efficiency: Although the grey prediction model is suitable for small sample scenarios, its prediction accuracy drops sharply as the data dimension increases, and it is difficult to adapt to the high dynamic changes of the fermentation environment. When facing large-scale time series data, the model training and inference speeds of SVM and PLSR are slow, unable to meet the requirements of real-time monitoring. In addition, traditional methods mostly rely on offline batch processing, and there is a lag in data feedback, making it difficult to trigger control instructions in a timely manner, resulting in the adjustment of environmental parameters lagging behind the changes in the actual fermentation state. (4) Lack of dynamic adaptability of control strategies: Existing methods usually separate the prediction and control links and rely on fixed thresholds or manual experience for environmental intervention. For example, when the predicted temperature deviates from the set range, only the heating equipment is simply started or stopped for adjustment, without considering the dynamic requirements of the fermentation stage (such as rapid temperature increase in the early stage and slow temperature decrease in the later stage). This static control strategy is difficult to cope with the non-linear and time-varying control requirements during the fermentation process, easily causing over-regulation or insufficient response, and exacerbating product quality fluctuations.
[0004] In summary, limitations make it difficult to guarantee the accuracy and reliability of traditional prediction methods. Specifically, there are large deviations between the predicted values and the actual values of temperature and humidity (such as an increase in RMSE), and the control of spatial uniformity fails, which in turn leads to problems such as microbial metabolic disorders and abnormal enzyme activities. For example, prediction errors may cause the temperature in a local area to be too high, resulting in the inactivation of beneficial bacteria, or the humidity to be too low, causing the enzymatic reaction to stagnate, ultimately reducing the liquor yield and the consistency of liquor flavor. At the same time, the lag of data feedback and control measures further amplifies the uncertainty of the fermentation process, increases raw material waste and the defective rate, and restricts the intelligent upgrade of the brewing process. Summary of the Invention
[0005] The purpose of the present invention is to provide a method for predicting the Daqu fermentation environment based on a CNN-LSTM hybrid model and its application. This method combines a convolutional neural network (CNN) to extract spatial features, a long short-term memory network (LSTM) to capture temporal dependencies, and introduces a random forest (RF) for key point screening, so as to take into account non-linearity, temporality, and spatial heterogeneity, and provide theoretical support for real-time and accurate control.
[0006] To achieve the above purpose, the present invention adopts the following technical solutions:
[0007] A method for predicting the Daqu fermentation environment based on a CNN-LSTM hybrid model includes the following steps:
[0008] Step 1: Collect the temperature and humidity data of the upper, middle, and lower layers in the Daqu fermentation room, and preprocess the temperature and humidity data of the Daqu fermentation room to obtain data sets for the upper, middle, and lower layers;
[0009] Step 2: Use the random forest model for the upper, middle, and lower layer data respectively, calculate the feature importance scores, and select the data of the top three important points in each layer to form a feature subset;
[0010] Step 3: Use the CNN-LSTM hybrid model to predict the future temperature and humidity changes in the Daqu fermentation room;
[0011] The CNN-LSTM hybrid model includes an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, a transition layer, a first LSMT layer, a second LSMT layer, and an output layer connected in sequence; the convolutional layer slides along the time axis to capture local features of adjacent time steps; then the transition layer flattens the multi-dimensional feature tensor into a one-dimensional vector; then the LSMT layer learns the long-term dependencies of the time series and predicts the temperature and humidity, and finally the output layer outputs the predicted values of the temperature and humidity.
[0012] Further, the temperature and humidity data is collected by temperature and humidity sensors, and the temperature and humidity sensors are arranged on the upper, middle, and lower layers of the Daqu fermentation chamber. At least 5 sensors are set on each layer, located in the middle and four corners of the Daqu fermentation chamber.
[0013] Further, the preprocessing includes: First, the moving average method is used to process the noise of the temperature and humidity data during the fermentation process in the fermentation chamber; then, the mean interpolation method is used to supplement the missing values; then, the Pearson correlation coefficient of the data at different points is calculated to screen the features with strong correlation; finally, the min-max normalization is used to map the data to the range of [-1, 1].
[0014] Further, the first convolutional layer, the second convolutional layer, and the third convolutional layer are all composed of convolution and max pooling. The first convolutional layer has 32 convolutional kernels, the kernel size is 3, the activation function is ReLU, the stride is 1, and the padding is Same; the second convolutional layer has 32 convolutional kernels, the kernel size is 3; the third convolutional layer has 64 convolutional kernels, the kernel size is 3; the pooling window of max pooling is 2.
[0015] Further, the first LSTM layer has 32 neurons, and the second LSTM layer has 64 neurons.
[0016] The present invention also provides a method for dynamically regulating the temperature and humidity of the Daqu fermentation environment based on fuzzy PID, including using a fuzzy PID controller to adjust the temperature and humidity of the Daqu fermentation environment in real time according to the temperature and humidity prediction values; the temperature and humidity prediction values are obtained according to the above-mentioned prediction method for the Daqu fermentation environment;
[0017] The model of the fuzzy PID controller is:
[0018]
[0019] In the formula, u(t) is the high-level time of the device for controlling the temperature and humidity of the fermentation environment, and e(t) is the temperature and humidity error; K p 、K i 、K d are the proportional coefficient, integral coefficient, and differential coefficient of the fuzzy PID controller respectively, and ΔK p 、ΔK i 、ΔK d are the PID parameter adjustment amounts output by the fuzzy PID controller respectively;
[0020] The temperature and humidity error e(t) is the deviation between the temperature and humidity prediction value and the target value; the target value is formulated based on expert experience.
[0021] Compared with the prior art, the present invention has the following beneficial effects:
[0022] (1) Denoise, perform correlation analysis, handle missing values, and normalize the environmental data of the Daqu fermentation chamber, reducing the impact of redundant and missing abnormal sensor data, ensuring the consistency of the dataset, and improving the robustness of the model. Then, use the random forest model to select features from the sensor point data distributed in three layers of the Daqu fermentation chamber space, obtain the most critical point data for each layer to construct a feature subset, screen out the key points that have the most influence on temperature and humidity prediction, effectively reduce the data dimension, and improve the model efficiency. Combining the feature extraction ability of the CNN-LSTM model for spatio-temporal features further optimizes the prediction accuracy, providing high-quality input data for subsequent fuzzy PID control.
[0023] (2) Adopt the deep learning method of the CNN-LSTM hybrid model, use a 3-layer 1D-CNN structure, with convolution kernel sizes of 3×1×1, 5×1×1, and 3×1×1 respectively, and the number of channels increasing to 64, effectively capturing the spatial correlation of multi-point temperature and humidity in the fermentation chamber. Use a bidirectional LSTM structure, where the forward layer captures historical trends and the reverse layer learns future potential patterns; enable the cell state to directly participate in the gating calculation to reduce the modeling error of temporal features; compared with traditional prediction methods, the prediction accuracy is greatly improved.
[0024] (3) Apply the predicted temperature and humidity values to the control of the Daqu fermentation environment, and dynamically adjust the environmental temperature and humidity of the Daqu fermentation environment based on the predicted temperature and humidity values using a fuzzy PID controller based on the prediction model, precisely improving the temperature and humidity during the Daqu fermentation process and enhancing the environmental uniformity; providing a theoretical basis and practical guidance for the optimization and intelligentization of the Daqu fermentation process, achieving the purpose of improving the quality of the finished Daqu, which is the key to ensuring the excellent rate. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] Figure 1 It is a schematic flowchart of Embodiment 1 of the present invention.
[0026] Figure 2 It is a schematic structural diagram of the CNN-LSTM hybrid model of the present invention.
[0027] Figure 3 It is a schematic flowchart of Embodiment 2 of the present invention.
[0028] Figure 4 It is a temperature and humidity curve graph of the present invention in a certain Daqu fermentation chamber. DETAILED DESCRIPTION OF THE INVENTION
[0029] Embodiment 1
[0030] As Figure 1 shown, a method for predicting the Daqu fermentation environment based on a CNN-LSTM hybrid model provided in this embodiment includes the following steps:
[0031] Step 1: Collect the temperature and humidity data of the upper, middle, and lower layers in the Daqu fermentation chamber, and preprocess the temperature and humidity data of the Daqu fermentation chamber to obtain the data sets of the upper, middle, and lower layers.
[0032] In this embodiment, temperature and humidity sensors are used to collect the temperature and humidity data of the Daqu fermentation chamber. Specifically, the temperature and humidity sensors are arranged in the upper, middle, and lower layers of the Daqu fermentation chamber, with at least 5 sensors set in each layer, all located in the middle and four corners of the Daqu fermentation chamber. According to the area of the Daqu fermentation chamber, several more sensors can be set in the middle area. The temperature data and humidity data during the Daqu fermentation process are collected through the set temperature and humidity sensors.
[0033] The preprocessing includes: (1) Using the moving average method to process the noise of the temperature and humidity data during the fermentation process in the koji room; smoothing the data, filtering out short-term fluctuations and random noise; enhancing the stability of the subsequent model by smoothing the temperature and humidity data.
[0034] During the sensor collection process, it may be affected by environmental interference or other factors, resulting in noise in the data; the moving average method is used to smooth the temperature and humidity data, calculate the average value within a certain time window, effectively filter out short-term fluctuations and random noise, and make the data smoother and closer to the true value. For the temperature and humidity data at each time point, take the average value of several time points before and after it as the corrected value at this time point, thereby reducing the impact of noise on the data.
[0035] (2) Sensor failures, data transmission interruptions, or human operation errors may cause data loss. To ensure the integrity of the data, the mean interpolation method is used to supplement the missing values; the missing values x i-1 and x i+1 are the time data at the moment before and after the missing data.
[0036] (3) During the Daqu fermentation process, the temperature and humidity at different points in the koji room may have a high degree of correlation. Calculate the Pearson correlation coefficient of the data at different points, and screen out the features with strong correlation; avoid redundant features interfering with model training.
[0037]
[0038] Among them, R is the Pearson correlation coefficient, X i is the i-th observation value of variable X, Y i is the i-th observation value of variable Y, and are the mean values of the two variables respectively; R ∈ [-1, 1]; when R > 0.8, there is a strong positive correlation, and select the point with the highest importance score. When R < 0.3, there is a weak correlation, which may be noise or independent features, and may be an abnormal sensor, requiring manual inspection.
[0039] By quantifying the spatial correlation of sensor data, it provides a scientific basis for feature selection and model input optimization, and ultimately improves the accuracy of temperature and humidity prediction and the real-time performance of regulation.
[0040] (4) Use min-max normalization to map the data to the range of [-1, 1], retaining the independent distribution characteristics of each layer; eliminate the dimension difference and accelerate the model convergence.
[0041]
[0042] Among them, is the normalized value, x min is the minimum value, x max is the maximum value;
[0043] In order to make the temperature and humidity data at different points comparable and improve the training efficiency of the subsequent model, the data is normalized. Avoiding the training deviation of the model caused by the dimension difference of the data, and at the same time accelerating the convergence speed of the model. The normalized data will be used as the basic data set for the input of the subsequent model. Combining hierarchical independent processing and inverse normalization operations to ensure the reliability and practicality of the algorithm in actual industrial scenarios.
[0044] Step 2: Use the random forest model for the upper, middle, and lower layer data respectively, calculate the feature importance scores, and select the data of the top three important points in each layer to form a feature subset to reduce redundant information.
[0045] Random Forest (RF) is an ensemble learning algorithm based on decision trees. By constructing multiple decision trees and aggregating their results, it improves the generalization ability and stability of the model. The preprocessed upper, middle, and lower layer temperature and humidity data are respectively input into the random forest model, and each layer of data is processed independently. Each layer contains 5 points (10-dimensional features of temperature + humidity), and the model needs to evaluate the contribution of each point to the temperature and humidity prediction. Based on Gini importance or permutation importance, calculate the importance scores of the data of each point. For example, if the temperature data of a certain point frequently participates in node splitting in most decision trees and significantly reduces the impurity, then its importance score is higher. Sort the feature importance scores of each layer of data in descending order, and select the data of the top three important points in each layer (a total of 3 points × temperature / humidity = 6-dimensional features) to form the optimal feature subset. For example, in the upper layer data, the importance of Sensor 1 (temperature), Sensor 3 (humidity), and Sensor 5 (temperature) is the highest, so the data of these 3 points are selected.
[0046] The random forest quantifies the importance of each sensor point, screens out the key points that have the greatest influence on temperature and humidity prediction, effectively reduces the data dimension and improves the model efficiency. Combining with the ability of the CNN-LSTM model to extract spatio-temporal features, the prediction accuracy is further optimized, providing high-quality input data for subsequent fuzzy PID control.
[0047] Step 3: Use the CNN-LSTM hybrid model to predict the future temperature and humidity changes in the Daqu fermentation room.
[0048] The CNN-LSTM hybrid model includes an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, a transition layer, a first LSMT layer, a second LSMT layer, and an output layer connected in sequence.
[0049] Among them, the structures of the first convolutional layer, the second convolutional layer, and the third convolutional layer are similar, all composed of convolution and max pooling. The first convolutional layer has 32 convolutional kernels, the kernel size is 3, the activation function is ReLU, the stride is 1, and the padding is Same; the second convolutional layer has 32 convolutional kernels, the kernel size is 3; the third convolutional layer has 64 convolutional kernels, the kernel size is 3; the pooling window of max pooling is 2. By sliding the convolutional layer along the time axis, local patterns in adjacent time steps (such as sudden temperature rise and humidity fluctuations) are captured; by max pooling, the data dimension is reduced, key features are retained, and the robustness of the model to small fluctuations is enhanced.
[0050] The transition layer flattens the multi-dimensional feature tensor output by the convolutional layer into a one-dimensional vector, realizing data integration and shape unification, and serving as the input of the LSTM layer.
[0051] The structures of the first LSTM layer and the second LSTM layer are the same. The first LSTM layer has 32 neurons and returns the complete sequence; the second LSTM layer has 64 neurons and returns the final output; learning the long-term dependencies of the time series, such as the correlation between the temperature rise trend in the initial stage of fermentation (0-24h) and the stable stage in the middle stage (24-72h). Gating mechanism: Dynamically adjust the information flow through the forget gate, input gate, and output gate to avoid gradient disappearance / explosion.
[0052] The output layer uses a fully connected layer to output the predicted values of temperature and humidity for the next N hours.
[0053] The model training of the CNN-LSTM hybrid model includes: collecting a large amount of temperature data and humidity data collected by the temperature and humidity sensors arranged in Step 1, preprocessing the data, constructing a feature subset using a random forest, and dividing the feature subset into a training set, a validation set, and a test set at a ratio of 6:2:2. The CNN-LSTM hybrid model is trained using the training set, the trained CNN-LSTM hybrid model is verified using the validation set, overfitting is monitored, and hyperparameters are adjusted. Finally, the CNN-LSTM hybrid model is tested using the test set, and the generalization performance of the model is finally evaluated.
[0054] The model optimization of the CNN-LSTM hybrid model includes using an Adam optimizer to adaptively adjust the learning rate; the initial value of the learning rate is 0.001, the batch size is 32, balancing the training speed and memory limitations. Regularization adds Dropout (ratio = 0.2) between LSTM layers to prevent overfitting. L2 regularization: applies a penalty term (λ = 0.001) to the weights of the fully connected layers. The evaluation metrics for optimization are the root mean square error RMSE, the mean absolute error MAE, and the coefficient of determination R 2 。
[0055] The convolutional layer CNN extracts local features from the input temperature and humidity time series data, including short-term fluctuations, mutations, etc., which can help the model better understand the local laws of temperature and humidity changes. After each convolution layer, a max pooling operation is used to reduce the data dimension while retaining key features. This design can reduce the computational complexity without losing important information; then, a transition layer Flatten operation is used to convert the multi-dimensional tensor output into a one-dimensional vector, flatten the multi-dimensional feature tensor into a one-dimensional vector, extract the local features, and use them as the input of the subsequent LSTM layer, providing rich spatio-temporal feature information for the model; the LSTM layer helps the model learn the time series characteristics of temperature and humidity data, so as to better predict future temperature and humidity changes. A two-layer LSTM structure is constructed, and the non-linear characteristics can effectively accelerate the convergence speed of the model and improve the fitting ability of the model to complex data. The CNN-LSTM hybrid model realizes the accurate prediction of the temperature and humidity in the Daqu fermentation environment by extracting spatial local features through CNN and modeling time dependencies through LSTM. Its hierarchical design (3 layers of CNN + 2 layers of LSTM) and dynamic hyperparameter tuning strategy significantly improve the prediction accuracy and real-time control ability, providing core technical support for the intelligent upgrade of the brewing process.
[0056] Example 2
[0057] This embodiment provides a method for dynamically regulating the temperature and humidity of the Daqu fermentation environment based on fuzzy PID, including: using a fuzzy PID controller to adjust the temperature and humidity of the Daqu fermentation environment in real time according to the temperature and humidity prediction values in Embodiment 1.
[0058] The model of the fuzzy PID controller is as follows:
[0059]
[0060] In the formula, u(t) is the high-level time of the device for controlling the temperature and humidity of the fermentation environment, and e(t) is the temperature and humidity error; K p 、K i 、K d are the proportional coefficient, integral coefficient, and differential coefficient of the fuzzy PID controller respectively, and ΔK p 、ΔK i 、ΔK d are the PID parameter adjustment amounts output by the fuzzy PID controller respectively.
[0061] According to the deviation e(t) between the predicted temperature and humidity value and the target value provided in Embodiment 1, the high-level time of the PWM signal is dynamically adjusted, thereby controlling the operating states of actuators such as heaters, humidifiers, and ventilation equipment; the target value is obtained based on manual experience and historical fermentation data and can provide ideal environmental parameters for the fermentation process.
[0062] Taking the temperature and humidity deviation e(t) and the deviation rate Δe(t) as the input quantities of the fuzzy controller, ΔK p 、ΔK i 、ΔK d parameters are obtained; the initial PID controller K p 、K i 、K d parameter values are obtained using the Ziegler-Nichols method, and the initial temperature and humidity data of the Daqu fermentation room are set; within each fuzzy PID control period, the difference e(t) between the actual temperature and humidity of the Daqu fermentation room and the set temperature and humidity is calculated; the control output value u(t) is obtained using the fuzzy PID controller model, and according to the output value, the high-level time of the device for controlling the temperature and humidity in the fermentation room is adjusted.
[0063] The fuzzy PID controller realizes the high-precision control of the Daqu fermentation environment through the coordination of dynamic parameter adjustment and prediction model; it provides a reliable solution for the intelligence and high efficiency of Daqu fermentation, and significantly improves the quality and stability of liquor production.
[0064] Apply Example 2 to a certain Daqu fermentation chamber, and the maximum environmental temperature difference between its upper, middle, and lower layers is reduced to 1.53 °C, and the maximum environmental humidity difference is reduced to 6.8% Rh. During the entire fermentation cycle, in terms of the accuracy and control uniformity of the control system, the temperature difference between the three-layer environments is reduced to 2 °C, and the humidity difference is reduced to 7% Rh. In terms of the followability of the control system, the temperature difference control with the target curve is reduced from 2.4 °C to 1.4 °C, and the humidity difference control with the target curve is reduced to 5.2% Rh. Using this method, accurately control the environmental temperature and humidity in the Daqu fermentation chamber to follow the target temperature and humidity curve, and the temperature and humidity in each area of the Daqu fermentation chamber are kept uniform, as Figure 4 shown.
[0065] The above is only the preferred implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modification and replacement based on the technical solutions and inventive concepts provided by the present invention should be covered within the protection scope of the present invention.
Claims
1. A prediction method for the Daqu fermentation environment based on a CNN-LSTM hybrid model, characterized in that, It includes the following steps: Step 1: Collect the temperature and humidity data of the upper, middle, and lower layers in the Daqu fermentation chamber, and preprocess the temperature and humidity data of the Daqu fermentation chamber to obtain the data sets of the upper, middle, and lower layers; Step 2: Respectively use the random forest model for the upper, middle, and lower layer data, calculate the feature importance scores, and select the top three point data in terms of importance for each layer to form a feature subset; Step 3: Use the CNN-LSTM hybrid model to predict the future temperature and humidity changes in the Daqu fermentation chamber; The CNN-LSTM hybrid model includes an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, a transition layer, a first LSMT layer, a second LSMT layer, and an output layer connected in sequence; the convolutional layer slides along the time axis to capture the local features of adjacent time steps; Then, the multi-dimensional feature tensor is flattened into a one-dimensional vector through the transition layer; then, the long-term dependence relationship of the time series is learned through the LSMT layer, and the temperature and humidity are predicted, and finally the temperature and humidity prediction values are output through the output layer.
2. The method for predicting the Daqu fermentation environment based on the CNN-LSTM hybrid model according to claim 1, wherein The temperature and humidity data are collected by temperature and humidity sensors, and the temperature and humidity sensors are arranged on the upper, middle, and lower layers of the Daqu fermentation chamber, with at least 5 sensors set on each layer, located in the middle and four corners of the Daqu fermentation chamber.
3. The prediction method for the Daqu fermentation environment based on the CNN-LSTM hybrid model according to claim 1, characterized in that: The preprocessing includes: First, use the moving average method to process the noise of the temperature and humidity data during the fermentation process of the Qu chamber; then, use the mean interpolation method to supplement the missing values; then, calculate the Pearson correlation coefficient of the data at different points and screen the features with strong correlation; finally, use the minimum-maximum normalization to map the data to the range of [-1,1].
4. The method for predicting the Daqu fermentation environment based on the CNN-LSTM hybrid model according to claim 1, wherein: The first convolutional layer, the second convolutional layer, and the third convolutional layer are all composed of convolution and max pooling. The first convolutional layer has 32 convolutional kernels, the kernel size is 3, the activation function is ReLU, the stride is 1, and the padding is Same; the second convolutional layer has 32 convolutional kernels, and the kernel size is 3; the third convolutional layer has 64 convolutional kernels, and the kernel size is 3; the pooling window of max pooling is 2.
5. The method for predicting the Daqu fermentation environment based on the CNN-LSTM hybrid model according to claim 1, wherein: The first LSTM layer has 32 neurons, and the second LSTM layer has 64 neurons.
6. A method for dynamically regulating the temperature and humidity of the Daqu fermentation environment based on fuzzy PID, characterized in that: It includes using a fuzzy PID controller to adjust the temperature and humidity of the Daqu fermentation environment in real time according to the temperature and humidity prediction values; the temperature and humidity prediction values are obtained according to the Daqu fermentation environment prediction method described in any one of claims 1-5; The model of the fuzzy PID controller is: where \(u(t)\) is the high-level time of the device for controlling the temperature and humidity of the fermentation environment, and \(e(t)\) is the temperature and humidity error; \(K\) p , \(K\) i , \(K\) d are the proportional coefficient, integral coefficient, and differential coefficient of the fuzzy PID controller respectively, and \(\Delta K\) p , \(\Delta K\) i , \(\Delta K\) d are the PID parameter adjustment amounts output by the fuzzy PID controller respectively; The temperature and humidity error e(t) is the deviation between the temperature and humidity prediction value and the target value; The target value is formulated based on expert experience.