Prediction method for daqu fermentation environment based on multi-modal data fusion and deep learning

Abstract

The present application belongs to the technical field of intelligent control of brewing process, and discloses a Daqu fermentation environment prediction method based on multi-modal data fusion and deep learning, comprising: 1 collecting historical environment data in one or more complete cycles of the Daqu fermentation process, and obtaining a data set after pre-processing the temperature and humidity data of the Daqu fermentation room; 2 performing feature screening on the multi-source data collected by the sensor by using a triple guarantee strategy, focusing on feature construction, time sequence law mining and three-stage optimization for feature screening; 3 predicting the future temperature and humidity changes in the Daqu fermentation room by using an ElasticNet algorithm model. The present application combines environment parameters and Daqu block state information, has the advantages of strong time sequence modeling capability, good non-linear fitting, fast regulation and response, etc., can effectively buffer the influence of equipment response delay, significantly reduce environmental fluctuations and quality risks, and thus effectively guarantees the stability of environment control in the whole process of Daqu block cultivation.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control technology for brewing processes, specifically relating to a method for predicting the fermentation environment of Daqu (a type of starter culture) based on multimodal data fusion and deep learning, and its application. Background Technology

[0002] As the core fermenting agent in baijiu brewing, the quality of daqu directly determines the yield and flavor characteristics of baijiu. Parameters such as temperature and humidity of the fermentation environment and the temperature of the daqu itself are key factors affecting the microbial community structure and enzyme activity of the daqu. In traditional daqu fermentation, environmental control relies heavily on manual experience, resulting in problems such as delayed response and low control precision, leading to poor uniformity of daqu fermentation and significant fluctuations in product quality.

[0003] Existing methods for predicting the fermentation environment of Daqu (a type of starter culture) mainly include traditional mathematical models such as linear regression, support vector machines, and grey prediction models, as well as some prediction schemes based on single deep learning networks. However, these methods have significant limitations: First, traditional models struggle to handle nonlinear relationships and long-term time-series dependencies during fermentation, resulting in large prediction errors and high RMSE values. Second, single deep learning models often ignore the impact of spatial heterogeneity, failing to fully explore the correlation features between different monitoring points. Third, existing technologies often separate prediction and control, employing fixed threshold control strategies that cannot adapt to the dynamic needs of each stage of fermentation. Fourth, the data preprocessing process is simple and fails to effectively address the redundancy and outliers of multi-source parameters, affecting the model's generalization ability.

[0004] Furthermore, existing prediction methods mostly focus only on the temperature parameter of the fermentation starter culture (Daqu), lacking a comprehensive consideration of key environmental factors such as humidity and temperature within the fermentation room, and failing to form a closed-loop system of "data collection and feature screening - accurate prediction - dynamic control". These shortcomings lead to insufficient timeliness and accuracy in the control of the fermentation environment, easily causing problems such as the inactivation of beneficial bacteria and abnormal enzymatic reactions, thus hindering the intelligent upgrading of Baijiu brewing. Therefore, it is urgent to develop an intelligent prediction method that can take into account spatial heterogeneity, temporal dependence, and nonlinear characteristics, and combine it with dynamic control strategies to achieve precise control of the fermentation environment of Daqu. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting the fermentation environment of Daqu (a type of starter culture) based on multimodal data fusion and deep learning. By fusing environmental parameters and Daqu block state information, the accuracy of environmental prediction can be improved, thereby effectively ensuring the stability of environmental control throughout the entire Daqu block cultivation process.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: The present invention provides a method for predicting the fermentation environment of Daqu (a type of starter culture) based on multimodal data fusion and deep learning, characterized by the following steps: Step 1: Collect historical environmental time series data that affect the fermentation of Daqu in the fermentation room; the historical observation time series data includes: environmental temperature value, environmental humidity value and Daqu production process parameters; Step 2: The historical observation environment time series is preprocessed using a triple guarantee strategy based on spline interpolation and local outlier factors to obtain the preprocessed historical observation environment time series; Step 3: Process the preprocessed historical observation environment time series to obtain the normalized historical observation environment feature sequence X; Step 4: Based on X, construct the objective function of the ElasticNet linear regression model using equation (1) and solve it to obtain the optimal regression coefficient vector. : (1) In equation (1), For historical duration, Let t be the actual value of the observed environment. Let X be the environmental feature vector observed at time t. The regression coefficients to be learned. For regularization strength, It is a weighting balancing factor, and ∈(0,1); Indicates transpose; Step 5: Output the future using equation (2) Predicted values ​​of the observation environment at any time : (2) Step 6: Construct the mean squared error loss function L of the model using equation (3): (3) In equation (3), For t Continuously monitor the actual values ​​of the environment; Predicting the duration for the future; Step 7: Train the ElasticNet linear regression model using gradient descent and calculate the mean squared error loss function L to update the model parameters until the loss function L converges, thereby obtaining the optimal prediction model for the historical observation environment of Daqu fermentation, which is used to predict the future trend of the observation environment of Daqu fermentation in real time.

[0007] The method for predicting the fermentation environment of Daqu (a type of starter culture) based on multimodal data fusion and deep learning described in this invention is characterized in that step 2 includes the following steps: Step 2.1: Use spline interpolation to smooth and fill the missing data in the historical observation environment time series to obtain the smoothed historical observation environment time series; Step 2.2: Use the local outlier factor algorithm to identify and remove outliers from the smoothed and filled historical environmental time series to obtain the historical environmental time series after anomaly removal; Step 2.3: Resample and smooth the historical observation environment time series after removing anomalies to obtain the preprocessed historical observation environment time series.

[0008] Furthermore, step 3 includes the following steps: Step 3.1: Process the preprocessed historical observation environment time series using trigonometric function mapping to obtain the sine and cosine fluctuation characteristic series of the historical observation environment; Step 3.2: Process the sine and cosine fluctuation characteristic sequences of the historical observation environment using first-order difference operations to obtain the short-term fluctuation characteristic sequences of the historical observation environment; Step 3.3: Normalize the short-term fluctuation characteristic sequence of the historical observation environment to obtain the normalized historical observation environment characteristic sequence X.

[0009] The present invention provides an electronic device, including a memory and a processor, characterized in that the memory is used to store a program supporting the processor in performing the method described therein, and the processor is configured to execute the program stored in the memory.

[0010] The present invention discloses a computer-readable storage medium storing a computer program, characterized in that the computer program is executed by a processor to perform the steps of the method described thereon.

[0011] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention performs smooth filling, missing value processing, and normalization on the environmental data of the fermentation room of Daqu (a type of starter culture), which reduces the impact of sensor data redundancy and missing anomalies, reduces data dimensionality, ensures the consistency of the dataset, improves training efficiency, and enhances the predictive performance and robustness of the model.

[0012] 2. This invention employs a deep learning method based on the ElasticNet algorithm model, which can more fully explore the potential relationships between key features, significantly improve prediction accuracy, and also significantly enhance model stability and adaptability.

[0013] 3. This invention is based on a preset standard range of temperature and humidity for Daqu fermentation, and dynamically generates environmental control commands according to the degree of deviation between the predicted temperature and humidity values ​​and the standard range. When the predicted temperature and humidity will deviate from the target range, the system automatically triggers a door and window opening and closing control strategy. By adjusting the environmental ventilation volume and heat and humidity exchange rate, it achieves active regulation of the temperature and humidity of the fermentation environment, maintaining it within a suitable range for Daqu fermentation. This provides a theoretical basis and practical guidance for the optimization and intelligentization of the Daqu fermentation process, achieving the goal of improving the quality of the finished Daqu and ensuring the key to a high yield of excellent products. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the process of Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the ElasticNet algorithm model of the present invention; Figure 3a Temperature curve of the upper layer of Daqu (a type of starter culture) in an embodiment of the present invention within a Daqu fermentation room; Figure 3b Temperature curve of the lower layer of Daqu (a type of starter culture) in an embodiment of the present invention within a Daqu fermentation room; Figure 3c A temperature curve of the present invention in a fermentation room of a certain type of koji (a type of Chinese liquor) as an example; Figure 3d The humidity curve of the present invention is shown in an example of a fermentation room for a certain type of koji (a type of Chinese liquor). Detailed Implementation

[0015] In this embodiment, a method for predicting the fermentation environment of Daqu (a type of Chinese koji) based on multimodal data fusion and deep learning combines data preprocessing with a triple-protection strategy, feature construction focusing on temporal pattern mining, and feature screening to implement third-order optimization. This approach balances nonlinearity, temporal sequence, and spatial heterogeneity, providing theoretical support for real-time and precise control. It effectively buffers the impact of equipment response delays, significantly reducing environmental fluctuations and quality risks. This solves the problem of traditional Daqu management relying on "post-event control" after temperature deviations, which suffers from control lag and the potential for large-scale control to trigger environmental fluctuations. The method possesses advantages such as strong temporal modeling capabilities, excellent nonlinear fitting, and fast control response. Figure 1 As shown, specifically, the method includes the following steps: Step 1: Temperature and humidity sensors are installed on both the upper and lower layers of the fermentation chamber, with at least 5 sensors on each layer. These sensors are located in the middle and four corners of the fermentation chamber. Depending on the size of the fermentation chamber, more sensors can be installed in the central area to collect historical environmental time series data on the upper and lower layers that affect the fermentation of Daqu. The historical time series data includes: ambient temperature values, ambient humidity values, and koji-making process parameters of the upper and lower layers.

[0016] Step 2: The historical observation environment time series is preprocessed using a triple protection strategy based on spline interpolation and local outlier factors to obtain the preprocessed historical observation environment time series. In this embodiment, since the sensor may lose data due to power outages, communication interference or physical damage, or generate outliers due to extreme fermentation environments (such as local high temperature and high humidity), and high frequency noise is common in the original data, three-step preprocessing is required to ensure data quality.

[0017] Step 2.1: Use spline interpolation to smooth and fill the missing data in the historical observation environment time series to obtain the smoothed historical observation environment time series; in this embodiment, the spline interpolation method is specifically cubic spline interpolation to fit the nonlinear fluctuation law of the historical observation environment data.

[0018] Step 2.2: The Local Outlier Factor (LOF) algorithm is used to identify and remove outliers from the smoothed historical environmental time series, resulting in a historical environmental time series after anomaly removal. In this embodiment, the LOF algorithm is employed to accurately identify outliers based on local density differences, effectively removing or adjusting outlier samples in the temperature and humidity data, improving data quality, and preventing the negative impact of abnormal disturbances on model training. The reason for choosing LOF instead of a simple global threshold is that temperature and humidity changes in the fermentation room have local nonlinear characteristics: for example, during the turning stage, the temperature at a certain point may rise or fall in a short period of time. At this time, this point is an outlier relative to its neighborhood. LOF can capture such local anomalies by comparing the local reachability density of each point with its k nearest neighbors.

[0019] Step 2.3: Resample and smooth the historical observation environment time series after removing anomalies to obtain the preprocessed historical observation environment time series; Step 3: Process the preprocessed historical observation environment time series using trigonometric function mapping to obtain the sine and cosine fluctuation characteristic series of the historical observation environment; The first-order difference operation shown in Equation (4) is used to process the sine and cosine fluctuation characteristic sequence of the historical observation environment to obtain the short-term fluctuation characteristic sequence of the historical observation environment. (4) In equation (4), and The sine and cosine wave characteristics of the observation environment at time t and time t-1 are respectively. This represents the difference between the two.

[0020] The short-term fluctuation characteristic sequence of historical observation environment is normalized to obtain the normalized historical observation environment characteristic sequence X.

[0021] Step 4: Traditional linear regression models often focus on fitting all features, neglecting the complex collinear relationships and redundant information that may exist among variables in the fermentation environment of Daqu (a type of koji). Furthermore, considering that each environmental variable sequence has its own scale of change and fluctuation characteristics, it is difficult to achieve accurate and stable prediction results using ordinary least squares methods. However, simply using the L1 regularized lasso algorithm can only achieve sparse feature selection, but it is prone to randomly discarding one feature from a strongly correlated feature group; while using only L2 regularized ridge regression can suppress collinearity, it cannot effectively remove irrelevant features. By introducing ElasticNet linear regression that combines L1 sparse regularization and L2 smoothing regularization, a prediction method that emphasizes both feature selection and coefficient shrinkage is proposed. This method more effectively handles the redundancy and collinearity problems in multi-source environmental data, thereby reducing the risk of overfitting, reducing computational overhead, and improving the model's generalization ability. The objective function of the ElasticNet linear regression model is then constructed and solved to obtain the optimal regression coefficient vector. ,like Figure 2 As shown; (1) In equation (1), For historical duration, Let t be the actual value of the observed environment. Let X be the environmental feature vector observed at time t. The regression coefficients to be learned. For regularization strength, It is a weighting balancing factor, and ∈(0,1); This indicates transpose.

[0022] Step 4.1: This embodiment uses a recursive approach, using the features at the current time t. Predict the value at time t+1 Then Continue prediction as part of the input features (lag term) And so on, to obtain the predicted value for the next t+k hours. Using equation (2), the future value is output. Predicted values ​​of the observation environment at any time : (2) Step 5: This loss function directly measures the cumulative error of multi-step prediction, and reflects the actual control requirements better than the single-step loss. The mean square error loss function L of the model is constructed using equation (3): (3) In equation (3), For t Continuously monitor the actual values ​​of the environment; Predict the duration for the future.

[0023] Step 6: Train the ElasticNet linear regression model using gradient descent and calculate the mean squared error loss function L to update the model parameters until the loss function L converges, thereby obtaining the optimal prediction model for the historical observation environment of Daqu fermentation, which is used to predict the future trend of the observation environment of Daqu fermentation in real time.

[0024] The example was applied to a Daqu fermentation room. The predicted maximum temperature difference between the upper and lower layers was 1.9°C, the maximum ambient humidity difference was 4.2%Rh, and the maximum ambient temperature difference was 3.3°C. Throughout the fermentation cycle, in terms of the accuracy and uniformity of the control system, the temperature difference between the actual and predicted values ​​was reduced to within 2°C, and the coverage rate of the prediction error within ±2°C exceeded 99.27%. The room temperature difference between the actual and predicted values ​​was reduced to within 2°C, and the coverage rate of the prediction error within ±2°C exceeded 98.98%. The humidity difference between the actual and predicted values ​​was reduced to within 2%Rh, and the coverage rate of the prediction error within ±2%Rh exceeded 97.38%. Using this method, the ambient temperature and humidity of the fermentation room and the Daqu temperature curve changes over the next 8 hours during the Daqu fermentation process were accurately predicted. The overall fluctuations between the actual and predicted values ​​were small, with minimal discrepancies. Figure 3a , Figure 3b , Figure 3c , Figure 3d As shown.

[0025] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications and substitutions based on the technical solutions and inventive concepts provided by the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the fermentation environment of Daqu (a type of Chinese koji) based on multimodal data fusion and deep learning, characterized in that, Includes the following steps: Step 1: Collect historical environmental time series data that affect the fermentation of Daqu in the fermentation room; the historical observation time series data includes: environmental temperature value, environmental humidity value and Daqu production process parameters; Step 2: The historical observation environment time series is preprocessed using a triple guarantee strategy based on spline interpolation and local outlier factors to obtain the preprocessed historical observation environment time series; Step 3: Process the preprocessed historical observation environment time series to obtain the normalized historical observation environment feature sequence X; Step 4: Based on X, construct the objective function of the ElasticNet linear regression model using equation (1) and solve it to obtain the optimal regression coefficient vector. : (1) In equation (1), For historical duration, Let t be the actual value of the observed environment. Let X be the environmental feature vector observed at time t. The regression coefficients to be learned. For regularization strength, It is a weighting balancing factor, and ∈(0,1); Indicates transpose; Step 5: Output the future using equation (2) Predicted values ​​of the observation environment at any time : (2) Step 6: Construct the mean squared error loss function L of the model using equation (3): (3) In equation (3), For t Continuously monitor the actual values ​​of the environment; Predicting the duration for the future; Step 7: Train the ElasticNet linear regression model using gradient descent and calculate the mean squared error loss function L to update the model parameters until the loss function L converges, thereby obtaining the optimal prediction model for the historical observation environment of Daqu fermentation, which is used to predict the future trend of the observation environment of Daqu fermentation in real time.

2. The method for predicting the fermentation environment of Daqu (a type of Chinese koji) based on multimodal data fusion and deep learning according to claim 1, characterized in that, Step 2 includes the following steps: Step 2.1: Use spline interpolation to smooth and fill the missing data in the historical observation environment time series to obtain the smoothed historical observation environment time series; Step 2.2: Use the local outlier factor algorithm to identify and remove outliers from the smoothed and filled historical environmental time series to obtain the historical environmental time series after anomaly removal; Step 2.3: Resample and smooth the historical observation environment time series after removing anomalies to obtain the preprocessed historical observation environment time series.

3. The method for predicting the fermentation environment of Daqu (a type of Chinese koji) based on multimodal data fusion and deep learning according to claim 1, characterized in that, Step 3 includes the following steps: Step 3.1: Process the preprocessed historical observation environment time series using trigonometric function mapping to obtain the sine and cosine fluctuation characteristic series of the historical observation environment; Step 3.2: Process the sine and cosine fluctuation characteristic sequences of the historical observation environment using first-order difference operations to obtain the short-term fluctuation characteristic sequences of the historical observation environment; Step 3.3: Normalize the short-term fluctuation characteristic sequence of the historical observation environment to obtain the normalized historical observation environment characteristic sequence X.

4. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports a processor in executing the method of any one of claims 1-3, the processor being configured to execute the program stored in the memory.

5. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program is executed by the processor to perform the steps of the method according to any one of claims 1-3.

Images