Digital twinning-based multi-dimensional parameter monitoring system for liquor fermentation process

Abstract

The invention discloses a Baijiu fermentation process multi-dimensional parameter monitoring system based on digital twinning. The Baijiu fermentation process multi-dimensional parameter monitoring system comprises a system supporting layer, a model building layer, a data interaction layer and a twinning application layer. The system support layer is used for providing hardware and software bases for the monitoring system; the model construction layer is used for realizing the virtualization capability of the monitoring system; the data interaction layer is used for data flow and interaction between a physical world and a virtual model; and the twin application layer is used for providing an interface for interacting with a user, carrying out visual display on dynamic data, and supporting the user to carry out operation and intervention in the fermentation process. According to the invention, accurate, real-time, numerical and intelligent monitoring effects can be realized, and reliable technical support is provided for dynamic environmental factor control and product quality stability.

Description

Technical Field

[0001] The present invention relates to the field of liquor fermentation monitoring, and in particular to a liquor fermentation process multidimensional parameter monitoring system based on digital twin. Background Art

[0002] Liquor fermentation is the core of traditional winemaking process. It not only determines the flavor, aroma and taste of liquor, but also is the key to inheriting the millennium wine culture. Through the delicate fermentation process, the sugar in the raw materials can be converted into alcohol, and rich flavor substances are generated at the same time, giving liquor a unique quality and charm.

[0003] At present, the monitoring technology of liquor fermentation process is relatively backward and cannot meet the needs of high-precision and real-time monitoring. This affects the accuracy and reliability of monitoring data, making it difficult to accurately control the fermentation process. Although a large amount of monitoring data has been collected, the ability to process and analyze the data is limited, and it is impossible to fully utilize the monitoring data to guide the optimization and adjustment of the fermentation process, which in turn affects the improvement of liquor quality.

[0004] Therefore, in order to solve the above problems, a multi-dimensional parameter monitoring system for the liquor fermentation process based on digital twin is needed, which can achieve accurate, real-time, numerical and intelligent monitoring effects, and provide reliable technical support for dynamic environmental factor control and product quality stability. Summary of the invention

[0005] In view of this, the purpose of the present invention is to overcome the defects in the prior art and provide a multi-dimensional parameter monitoring system for the liquor fermentation process based on digital twins, which can achieve accurate, real-time, numerical and intelligent monitoring effects, and provide reliable technical support for dynamic environmental factor control and product quality stability.

[0006] The multi-dimensional parameter monitoring system of liquor fermentation process based on digital twin of the present invention comprises a system support layer, a model building layer, a data interaction layer and a twin application layer;

[0007] The system support layer is used to provide hardware and software foundation for the monitoring system;

[0008] The model building layer is used to realize the virtualization capability of the monitoring system;

[0009] The data interaction layer is used for data flow and interaction between the physical world and the virtual model;

[0010] The twin application layer is used to provide an interface for interacting with users, visualize dynamic data, and support users to operate and intervene in the fermentation process.

[0011] Furthermore, the system support layer:

[0012] The fermentation tank, main control board, sensor and intelligent terminal are used as physical infrastructure, and the physical infrastructure collects environmental parameters in the fermentation process in real time; the environmental parameters include temperature, humidity, pH value and carbon dioxide concentration; the sensor data is transmitted to the background by using WIFI and MQTT message transmission protocol; based on VMware 16 virtualization platform, the system runs on CentOS 7 operating system; in the virtual environment, Unreal Engine 5 is used to drive the 3D visualization engine, and users monitor the fermentation process through the human-computer interaction interface.

[0013] Furthermore, the model building layer:

[0014] SolidWorks, Unreal Engine, and Blender were used to perform 3D modeling of the physical fermentation tank and its environment. The fermentation process in the physical world was reproduced through simulation modeling of actual equipment and scenes. Optimized rendering was used to improve the realism and rendering efficiency of the 3D model, so that the fermentation process remained consistent with the actual physical process.

[0015] Furthermore, the data interaction layer:

[0016] Collect, transmit, store and convert multi-source data to form a data management core; use MySQL database to store environmental parameters; and standardize heterogeneous data from different sources through the ODBC interface.

[0017] Furthermore, the twin application layer:

[0018] The 3D visualization interface supports users to monitor environmental parameters during the fermentation process in real time, and also supports users to operate and intervene in the fermentation process;

[0019] Through the analysis and modeling of historical data, anomaly detection and prediction functions are formed. When certain environmental parameters are abnormal, an early warning will be automatically issued to remind operators to take timely measures;

[0020] Based on the data short-term prediction model, trend analysis of key variables in the fermentation process is carried out;

[0021] Combined with the data management module, it integrates multi-source data and provides in-depth analysis to support intelligent control and optimization of the fermentation process.

[0022] Further, the data short-term prediction model includes a temperature and humidity prediction model;

[0023] The temperature and humidity prediction model is a CNN-BiLSTM-Attention model based on sand cat group optimization; wherein, by simulating the global and local search mechanism of the predation behavior of the sand cat group, the hyperparameters of the temperature and humidity prediction model are globally optimized, so that the temperature and humidity prediction model escapes from the local optimum, optimizes the high-dimensional parameter space, and automatically adjusts the network structure.

[0024] Furthermore, the convolution module of the temperature and humidity prediction model adopts a one-dimensional convolution with a step size of 1, the number of convolution kernels is set to 32, and the convolution kernel size is 3; the convolution layer uses the SAME filling strategy, and the pooling layer is configured with a pooling kernel size of 2; the BiLSTM layer is configured with 50 neurons;

[0025] The Adam optimizer is used to train the prediction model, with the initial learning rate set to 0.01, the number of iterations to 100, the L2 regularization coefficient to 0.001, and the Dropout ratio to 0.1.

[0026] The initial population size of the sand cat swarm optimization algorithm is set to 20, the maximum number of iterations is 20, the sensitivity range is 2, and the fitness function uses the mean absolute percentage error as the evaluation indicator to optimize the number of convolution kernels, the number of BiLSTM neurons, the learning rate, and the L2 regularization coefficient in the prediction model.

[0027] Furthermore, outlier processing and mean processing are performed on the collected data:

[0028] Outlier processing, including:

[0029] If the difference between the temperature or humidity data monitored in a short period of time and the historical average value exceeds 3 times the standard deviation, the data is determined to be an abnormal value, and the abnormal value is replaced by the mean of the valid data before and after it;

[0030] Mean processing, including:

[0031] The average of the data collected every half hour is used as the temperature and humidity monitoring data.

[0032] The beneficial effects of the present invention are as follows: a multi-dimensional parameter monitoring system for the liquor fermentation process based on digital twins disclosed in the present invention forms four levels by constructing a system support layer, a model building layer, a data interaction layer and a twin application layer. The levels complement each other and constitute an organic whole, covering the complete mapping process from physical to virtual twins. The levels are closely connected through the real-time flow of data, which provides a theoretical basis and technical path for the digital and intelligent monitoring of the fermentation process, improves the monitoring, management and optimization capabilities of the fermentation process, and ensures that the virtual model can reflect the physical state in real time and accurately, thereby improving the stability and accuracy of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0034] Figure 1 It is a schematic diagram of the monitoring system framework of the present invention;

[0035] Figure 2 It is a data transmission structure diagram of the present invention;

[0036] Figure 3 It is a flow chart of the data interaction structure between MYSQL, UE and MATLAB of the present invention;

[0037] Figure 4 It is a schematic diagram of the mapping scheme of the twin model of the present invention to the physical entity;

[0038] Figure 5 It is a schematic diagram of the initial stage of the sand cat colony optimization algorithm of the present invention;

[0039] Figure 6 It is a schematic diagram of the structure of the CNN-BiLSTM-Attention model of the present invention;

[0040] Figure 7 Flowchart of CNN-BiLSTM-Attention model optimized for SCSO of the present invention;

[0041] Figure 8 It is a schematic diagram of the evolution curve of the present invention. DETAILED DESCRIPTION

[0042] The present invention is further described below in conjunction with the accompanying drawings, as shown in the drawings:

[0043] This embodiment discloses a multi-dimensional parameter monitoring system for liquor fermentation process based on digital twin, including a system support layer, a model building layer, a data interaction layer and a twin application layer;

[0044] The system support layer is used to provide hardware and software foundation for the monitoring system;

[0045] The model building layer is used to realize the virtualization capability of the monitoring system;

[0046] The data interaction layer is used for data flow and interaction between the physical world and the virtual model;

[0047] The twin application layer is used to provide an interface for interacting with users, visualize dynamic data, and support users to operate and intervene in the fermentation process.

[0048] The overall framework of the liquor fermentation multidimensional parameter monitoring system constructed by the present invention is as follows: Figure 1As shown in the figure, the system architecture can be divided into four layers: system support layer, model building layer, data interaction layer and twin application layer. Each layer complements each other to form an organic whole, which comprehensively improves the monitoring, management and optimization capabilities of the fermentation process.

[0049] In this embodiment, the system support layer provides a solid hardware and software foundation for the entire multidimensional parameter monitoring system (or digital twin system). In terms of hardware, equipment such as fermentation tanks, main control boards, intelligent sensors and intelligent terminals constitute the physical infrastructure of the system, and these devices collect key environmental parameters such as temperature, humidity, pH value and carbon dioxide concentration in real time during the fermentation process. By using lightweight message transmission protocols such as WIFI and MQTT, sensor data can be efficiently transmitted to the background. In order to ensure the stable operation of the system, virtualization technology plays a key role in the support layer. Based on VMware16 virtualization platform, the system runs on CentOS 7 operating system. In the virtual environment, Unreal Engine 5 (UE5) is used to drive the three-dimensional visualization engine, and users can intuitively monitor the fermentation process through the human-computer interaction interface. This architecture not only improves the flexibility and scalability of the system, but also provides a strong basic support for the upper module.

[0050] In this embodiment, based on the system support layer, the model building layer further promotes the virtualization capability of the multi-dimensional parameter monitoring system. The core task of this layer is to use tools such as SolidWorks, Unreal Engine and Blender to perform high-precision three-dimensional modeling of the physical fermentation tank and its environment. Through simulation modeling of actual equipment and scenes, the system effectively reproduces the fermentation process in the physical world, providing a basis for virtual-real synchronization. In addition, the application of optimized rendering technology greatly improves the realism and rendering efficiency of the three-dimensional model, ensuring a high degree of consistency between the fermentation process presented by the system in the virtual environment and the actual physical process. The accurate reproduction of the scene provides a basic platform for real-time monitoring and predictive analysis. Subsequent data collection and monitoring are intuitively displayed through the virtual scene, enhancing the user experience and the effectiveness of decision-making.

[0051] In this embodiment, the model building layer builds a bridge between the virtual and real environments, while the data interaction layer ensures the data flow and interaction between the physical world and the virtual model. This layer is responsible for the collection, transmission, storage and conversion of multi-source data, forming the data management core of the system. The sensor collects physical data such as temperature, humidity, pH value and carbon dioxide concentration in real time through WIFI and MQTT protocol, and transmits these data to the MySQL database for storage.

[0052] The data interaction layer not only needs to deal with data storage issues, but also ensure the integration and consistency of different data sources. Through the ODBC interface, the system can standardize heterogeneous data from different sources to ensure that data can be smoothly converted and integrated, thereby providing reliable input for subsequent data analysis. The design of this layer ensures efficient interaction and seamless flow of multi-source data in the system, providing a solid data foundation for real-time monitoring and analysis.

[0053] In this embodiment, the twin application layer is the direct interactive interface between the system and the user, and is also a concrete manifestation of the system's analysis and decision-making functions. On the three-dimensional visualization interface, users can monitor key parameters of the fermentation process in real time, such as temperature, humidity, pH value, and carbon dioxide concentration. This interface not only provides a visual display of dynamic data, but also supports users to operate and intervene in the fermentation process.

[0054] Through the analysis and modeling of historical data, the system has the function of anomaly detection and prediction. When certain key parameters are abnormal (such as excessive temperature or abnormal pH value), the system will automatically issue an early warning to remind operators to take timely measures. Based on the short-term prediction function of the data, the system can also perform trend analysis on key variables in the fermentation process to optimize the fermentation control strategy. Combined with the data management module, the twin application layer can integrate multi-source data and provide in-depth analysis, thereby effectively supporting the intelligent control and optimization of the fermentation process.

[0055] Among them, the present invention uses the EMQX Internet of Things platform based on the MQTT protocol as a bridge for information transmission between the physical system and the twin system to ensure the efficiency and reliability of information transmission. MQTT is a lightweight, publish-subscribe-based message transmission protocol suitable for resource-constrained devices and low-bandwidth, high-latency or unstable network environments. The MQTT architecture has three components: publisher, server, and subscriber. The publisher publishes messages with specific topics to the server, subscribers subscribe to these topics, and then the server sends messages for each topic to the subscribers of the topic.

[0056] EMQX is an open source distributed MQTT message server with strong stability, scalability and clustering capabilities. It can establish connections with various devices and achieve efficient data transmission. In order to realize data transmission between the physical entity of the fermentation system and the twin system through the EMQX server, the hardware system uses ESP32-WROOM-32 as the core processor, uses C language to write programs in Arduino, connects to the Internet through its own integrated WIFI module, and matches the server's IP address to establish a connection with the EMQX Internet of Things platform and topic subscription, thereby realizing data transmission between the hardware system and the EMQX Internet of Things platform. In the twin system, write the control blueprint through UE5, install the Mqtt Client plug-in, call the CreateMqtt Client node and the Connect node to establish a communication connection with the EMQX server, call the Subscribe node to complete the corresponding topic subscription, and realize the connection and message communication between the twin system and the EMQX platform, thereby finally realizing data communication between the physical system and the twin system. The data transmission structure is as follows: Figure 2 shown.

[0057] MySQL is an open source and free relational database management system. By storing data in different tables, it effectively improves the speed and flexibility of data calls. In the digital twin system for liquor fermentation monitoring designed by the present invention, data interfaces with UE5 and MATLAB are mainly developed to realize data interaction and processing. The database is used to store real-time monitoring data such as temperature, humidity, pH value, carbon dioxide concentration, etc. during the fermentation process and the corresponding prediction data. In addition, information such as abnormal warnings and operation logs generated during the operation of the system will also be stored in the database for further analysis and management. The data interaction structure flow chart between MYSQL, UE and MATLAB is shown in the figure. Figure 3 shown.

[0058] The data interaction between MYSQL and UE twin system is realized through MySQL and MariaDB Integration plug-in. First, configure the MYSQL database operating environment, use the database management tool Navicat to create the database and the required data tables; secondly, configure the MySQL and MariaDB Integration file in Unreal Engine, call the relevant nodes in the blueprint program to create a database connection module, and connect to the MYSQL database by matching the corresponding database name, user name, password, port and other information; finally, store the system operation data and the real-time data collected by the monitoring system in the corresponding data table.

[0059] MATLAB is connected to the MYSQL database to feed real-time monitoring data into the prediction model and to feed back the prediction value of the next stage to the digital twin system in a timely manner to ensure the stable operation of the fermentation process. The implementation of data communication between MATLAB and MYSQL depends on the ODBC (Open Database Connectivity) driver and the Database Toolbox toolbox that comes with MATLAB. The ODBC driver is a standard interface for exchanging data between databases and applications. Combined with the Database Toolbox that comes with MATLAB, it can provide a powerful and flexible way to process and analyze large amounts of data stored in the database.

[0060] In order to achieve accurate mapping of the virtual model of the liquor fermentation twin system to the physical system model, it is first necessary to perform high-precision three-dimensional modeling from multiple dimensions such as geometric structure, physical properties, behavioral characteristics and rule constraints to characterize the physical characteristics of the fermentation system. This includes accurate descriptions of geometric and physical parameters such as size, shape, and material properties, as well as comprehensive capture of dynamic data changes during the liquor fermentation process. Subsequently, by associating, merging and integrating models of various dimensions, it is ensured that the twin system can respond to state changes of the physical system in real time, dynamically update and reflect its latest physical state and performance data, thereby achieving a high degree of consistency between the virtual system and the physical system. The mapping scheme of the twin model to the physical entity is as follows: Figure 4 shown.

[0061] The present invention takes the physical entity of the liquor fermentation tank as the modeling object, first uses SolidWorks 3D modeling software to complete the construction and assembly of the 3D model, then imports the model into Blender software, performs rendering, mapping and material addition, and finally forms a virtual model with high fidelity to the physical entity, and finally imports the model into the virtual scene of UE5, and uses the level blueprint to debug the relevant functions to ensure the accuracy and functionality of the model in details. In the interactive design, combined with the Create Mqtt Client plug-in in the control blueprint, the data collected by the physical system in real time is synchronously obtained, and the data is assigned to each level blueprint after format conversion and filtering analysis, and the virtual model is driven by real-time data, so as to achieve the effect of "reflecting the real with the virtual"; at the same time, based on the MySQL and MariaDB Integration plug-in, the system can store monitoring data and running data in the MYSQL database, and support database access, call historical data and predictive data, so as to achieve the goal of "predicting the real with the virtual". In addition, the UMG (Unreal Dynamic Graphics UI Designer) in UE5 can synchronously apply data streams to UI elements, use UI panel components to build a visual interface, and combine Echarts to realize the visual display of data charts.

[0062] In this embodiment, the data short-term prediction model includes a temperature and humidity prediction model;

[0063] The temperature and humidity prediction model is a CNN-BiLSTM-Attention model based on sand cat group optimization; wherein, by simulating the global and local search mechanism of the predation behavior of the sand cat group, the hyperparameters of the temperature and humidity prediction model are globally optimized, so that the temperature and humidity prediction model escapes from the local optimum, optimizes the high-dimensional parameter space, and automatically adjusts the network structure.

[0064] Among them, the Sand Cat Swarm Optimization (SCSO) algorithm simulates the survival behavior of sand cats in the natural environment, especially their unique mechanism in the process of prey search and hunting. With their excellent low-frequency noise perception ability, sand cats are able to detect low-frequency sounds in the range of 0 to 2kHz. This allows sand cats to sense the location of prey through extremely subtle sounds and accurately capture prey. This perception and hunting process provides a theoretical basis for the core mechanism of SCSO, which is to optimize the problem-solving process by simulating the search and hunting behavior of sand cats.

[0065] (1) Population initialization

[0066] In the SCSO optimization algorithm, each sand cat represents a solution to a problem. In the optimization of d-dimensional problems, the sand cat is represented as a 1×d-dimensional array, such as Figure 5 shown.

[0067] The set of all sand cat groups is represented as:

[0068] SC i ={SC1,SC2,....,SC n}; i∈[1,n] (1)

[0069] Among them, SC i represents the i-th sand cat. The sand cat individual represents a solution to the optimization problem, which is (x1, x2, ..., x d ), xi are all floating point numbers and must be between the upper and lower bounds.

[0070] In the initialization phase of SCSO, we first need to calculate the problem size and dimension (N pop ×N d ), (pop=1,2,......,d) Use the sand cat group to establish a candidate matrix. Then, define the fitness function, use the fitness function to evaluate the fitness of each sand cat, and finally select the best individual.

[0071] (2) Search Behavior

[0072] In nature, sand cats rely on low-frequency noise to search for prey. They have the ability to perceive low-frequency sound waves below 2kHz. Therefore, in the SCSO algorithm, the parameter r representing the hearing ability is G It decreases linearly from 2 to 0 to gradually approach the prey and avoid missing or skipping. The update method is as follows:

[0073]

[0074] Where r G is the normal sensitivity range; S M N is the auditory characteristic of the sand cat, usually set to 2; C is the current iteration number; N MAX is the maximum number of iterations.

[0075] To avoid local optimality, the position of each sand cat is random, and each sand cat has a different sensitivity to prey. Therefore, the sensitivity range of each sand cat is defined as:

[0076] r=r G ×rand(0,1) (3)

[0077] Where r is the sensitivity range of each sand cat; rand(0,1) is a random number between 0 and 1.

[0078] The hunting behavior of sand cats mainly includes search behavior and attack behavior. The balance coefficient R is used to control the conversion between search and attack behavior. The calculation formula of the balance coefficient is:

[0079] R=2r G ×rand(0,1)-r G (4)

[0080] When |R|>1, the sand cat starts to search for prey. Each sand cat will update its position based on its own sensitivity range, current position and optimal solution. The position update calculation formula is:

[0081] P(t+1)=r×(P b (t)-rand(0,1)×P c (t)) (5)

[0082] Where P(t+1) is the updated position; P b (t) is the optimal solution; P c (t) is the current position.

[0083] (3) Aggressive behavior

[0084] When |R|≤1, the sand cat starts to attack the prey and randomly generates a position near the prey based on its current position and the global optimal solution, so that it gradually reaches the best hunting position. The position update calculation formula is:

[0085] P rnd =|P b (t)×rand(0,1)-P c (t)| (6)

[0086] Where P rnd is a randomly generated position near the prey; P b (t) is the optimal solution; P c (t) is the current position.

[0087] The attack range of the sand cat can be regarded as a circle. The sand cat group can move in all directions along the circumference. The attack direction of each sand cat is determined by the random angle θ. In this way, the sand cat can reach the hunting position without falling into the local optimum. The specific update method is:

[0088] P(t+1)=P b (t)-r×P rnd ×cos(θ),0°≤θ≤360°(7)

[0089] The Sand Cat Swarm Optimization Algorithm can switch between search and attack behaviors based on its own adaptive capabilities according to the changes in the R value, and dynamically adjust the search strategy in the multi-dimensional parameter space. It can perform fine search and optimization for key hyperparameters in the model, such as convolution kernel size, number of LSTM units, weights of the Attention mechanism, etc., and adjust network parameters such as weights, biases, and learning rates to reduce training errors and improve the model's fitting performance when processing complex time series data. Through this adaptive optimization strategy, the model training process is more stable and efficient, and can better capture key features in the data.

[0090] In addition, CNN-BiLSTM-Attention is a hybrid neural network model that combines the advantages of CNN, BiLSTM and Attention mechanism, and is suitable for processing complex time series data, especially multidimensional data with spatial and temporal dependencies. CNN and BiLSTM use different methods for feature extraction. CNN uses convolution kernels to extract spatial features from input data and identify short-term fluctuations and local features of data such as temperature and humidity; BiLSTM relies on its internal bidirectional LSTM structure to handle global time dependencies; and the Attention mechanism selectively focuses on important feature points, dynamically assigns weights, enhances the decision-making ability of the model, and improves the utilization efficiency of key information. Therefore, this combined model makes full use of the advantages of different modules and can handle time series features, long-term and short-term dependencies, and dynamic selection of important features. Specific model features such as Figure 6 As shown:

[0091] First, a dataset containing the historical data of temperature and humidity and the time series of their key influencing factors is created, and the data is used as the model input after data preprocessing; in the CNN layer, the convolution layer extracts local features through convolution operations, and the pooling layer downsamples the features extracted by the convolution layer to reduce the data dimension and retain key information. Next, in the BiLSTM layer, the data output by the CNN is trained and learned from both the front and back directions through a bidirectional LSTM network to capture the temporal dependency of the predicted data; then, the data processed by the BiLSTM layer is input into the attention mechanism layer, and different weights are assigned to the data at different time points, so that the model focuses on the key parts; finally, the feature vector processed by the attention mechanism is passed to the fully connected layer to output the prediction result.

[0092] In addition, to prevent overfitting, the Dropout layer is introduced in the CNN layer and attention mechanism layer of the model. This mechanism randomly discards some neurons during the training process to reduce excessive dependence between neurons and prevent co-adaptation between neurons. It can avoid overfitting and improve the generalization ability and training efficiency of the model.

[0093] Although the CNN-BiLSTM-Attention model has demonstrated strong spatiotemporal feature extraction capabilities in capturing temperature and humidity changes during liquor fermentation, its inherent non-convex optimization problem makes the model prone to local optimal solutions. In addition, due to the complexity of the model, especially when facing long time series data, a large number of trainable parameters are often required, which increases the risk of overfitting, thereby limiting its generalization ability and performance on new data. Manual parameter adjustment is not only time-consuming and inefficient, but also difficult to effectively explore the huge parameter space.

[0094] Therefore, the present invention introduces the SCSO optimization algorithm, which globally optimizes the hyperparameters of the model by simulating the global and local search mechanism of the predation behavior of sand cat groups, so that the model can effectively escape the local optimum, optimize the high-dimensional parameter space, and automatically adjust the network structure, thereby improving the training efficiency, enhancing the generalization ability of the model, and significantly reducing the risk of overfitting. The specific optimization process is as follows Figure 7 As shown:

[0095] The SCSO algorithm has the characteristics of strong global search capability, strong adaptability, fast convergence speed, and simple and easy implementation. It has good performance and competitive advantages when dealing with complex, multi-peak, and nonlinear optimization problems. The fermentation environment prediction model of the fermented mash in the present invention is a time series prediction model with multiple feature inputs and a single output. Therefore, using SCSO to optimize the CNN-BiLSTM-Attention model will effectively improve the performance of the model.

[0096] In this embodiment, in the CNN-BiLSTM-Attention fermentation environment parameter prediction model based on sand cat group optimization proposed in the present invention, in order to effectively capture local and global features in time series data while avoiding overfitting and waste of resources, the convolution module of the temperature and humidity prediction model adopts a one-dimensional convolution with a step size of 1, the number of convolution kernels is set to 32, and the convolution kernel size is 3; the convolution layer uses the SAME filling strategy, and the pooling layer is configured with a pooling kernel of size 2; the BiLSTM layer is configured with 50 neurons; wherein, the SAME filling strategy is a flexible filling method, which allows the convolutional neural network to extract the features of the input data while keeping the output data size the same as the input data size in a certain dimension.

[0097] The Adam optimizer is used to train the prediction model, with the initial learning rate set to 0.01, the number of iterations to 100, the L2 regularization coefficient to 0.001, and the Dropout ratio to 0.1.

[0098] The initial population size of the sand cat swarm optimization algorithm is set to 20, the maximum number of iterations is 20, the sensitivity range is 2, and the fitness function uses the mean absolute percentage error as the evaluation indicator to optimize the number of convolution kernels, the number of BiLSTM neurons, the learning rate and the L2 regularization coefficient in the prediction model, so as to obtain the evolution curve as shown in the figure. Figure 8 As shown in Table 1, the hyperparameter optimization range and results of the CNN-BiLSTM-Attention network based on sand cat group optimization are shown in Table 1.

[0099] Table 1

[0100]

[0101] In this embodiment, in view of the nonlinear changes in environmental parameters during liquor fermentation, the present invention performs systematic outlier identification and mean smoothing processing on the collected data:

[0102] 1) Outlier processing. Liquor fermentation is a slow-changing dynamic process. Therefore, if the difference between the temperature or humidity data monitored in a short period of time and the historical average value exceeds 3 times the standard deviation, the data is judged as an outlier. In order to reduce the impact of abnormal data on the overall analysis, the identified outlier will be replaced by the mean of the adjacent valid data before and after to ensure the continuity and stability of the data.

[0103]

[0104] Where X n is the data collection value at the abnormal moment; is the historical mean of each data; X n ′ is the value after abnormal data processing; σ is the standard deviation.

[0105] 2) Mean processing: The data collection frequency of the fermentation monitoring system of the present invention is once every 5 minutes. After long-term observation of the collected data, it is found that the data has no obvious fluctuations. Therefore, in order to reduce the amount of data processing, the temperature and humidity monitoring data will be processed every half hour using a mean algorithm.

[0106] In order to verify the performance of the SCSO-CNN-BiLSTM-Attention prediction model proposed in the present invention, the unoptimized CNN-BiLSTM-Attention model was used as a control. The temperature, humidity, pH and CO2 data of the whole process of the fermentation test were input into the trained model to verify the prediction performance of the model for the whole cycle of the fermentation process.

[0107] By observing the temperature prediction results, we can find that the CNN-BiLSTM-Attention model can roughly capture the trend of temperature changes. However, there is a significant deviation from the actual value in the stage of rapid temperature rise and fall, especially in the stage of temperature rise and peak. The model's prediction curve has a certain lag near the peak, reflecting that the model is not responsive enough in the fast-changing area. The CNN-BiLSTM-Attention model optimized by the sand cat swarm optimization algorithm shows better fitting accuracy. The sand cat swarm optimization algorithm helps the model better capture the subtle features of drastic temperature fluctuations, making the prediction results more consistent with the actual change trend.

[0108] From the humidity prediction results, it can be found that the humidity is relatively stable during the fermentation process. Both models can capture the overall change trend of humidity. The CNN-BiLSTM-Attention model optimized by the sand cat swarm optimization algorithm shows higher accuracy. In some small fluctuation areas, the optimized model can more accurately capture subtle changes in humidity, showing higher sensitivity and further reducing the prediction error.

[0109] By comparing the evaluation indicators in Table 2, it can be found that the sand cat swarm optimization algorithm significantly improves the overall performance of the CNN-BiLSTM-Attention model in temperature and humidity prediction, especially in the temperature prediction task. The optimized model performs well in all evaluation indicators, significantly reduces the error (MAE, MAPE, RMSE), and significantly improves the model fitting effect (R2). The results show that the CNN-BiLSTM-Attention model based on sand cat swarm optimization proposed in the present invention has better prediction accuracy and generalization ability.

[0110] Table 2

[0111]

[0112] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. A multi-dimensional parameter monitoring system for liquor fermentation process based on digital twin, characterized by: It includes system support layer, model building layer, data interaction layer and twin application layer; The system support layer is used to provide hardware and software foundation for the monitoring system; The model building layer is used to realize the virtualization capability of the monitoring system; The data interaction layer is used for data flow and interaction between the physical world and the virtual model; The twin application layer is used to provide an interface for interacting with users, visualize dynamic data, and support users to operate and intervene in the fermentation process.

2. The multi-dimensional parameter monitoring system for liquor fermentation process based on digital twin according to claim 1 is characterized in that: The system support layer: The fermentation tank, main control board, sensor and intelligent terminal are used as physical infrastructure, and the physical infrastructure collects environmental parameters in the fermentation process in real time; the environmental parameters include temperature, humidity, pH value and carbon dioxide concentration; the sensor data is transmitted to the background by using WIFI and MQTT message transmission protocol; based on VMware 16 virtualization platform, the system runs on CentOS 7 operating system; in the virtual environment, Unreal Engine 5 is used to drive the 3D visualization engine, and users monitor the fermentation process through the human-computer interaction interface.

3. The multi-dimensional parameter monitoring system for liquor fermentation process based on digital twin according to claim 1 is characterized in that: The model builds layers: SolidWorks, UnrealEngine and Blender are used to perform 3D modeling of the physical fermentation tank and its environment. The fermentation process in the physical world is reproduced through simulation modeling of actual equipment and scenes. Optimized rendering is used to improve the realism and rendering efficiency of the 3D model, so that the fermentation process remains consistent with the actual physical process.

4. The multi-dimensional parameter monitoring system for liquor fermentation process based on digital twin according to claim 1 is characterized in that: The data interaction layer: Collect, transmit, store and convert multi-source data to form a data management core; use MySQL database to store environmental parameters; and standardize heterogeneous data from different sources through the ODBC interface.

5. The multi-dimensional parameter monitoring system for liquor fermentation process based on digital twin according to claim 1 is characterized in that: The twin application layer: The 3D visualization interface supports users to monitor environmental parameters during the fermentation process in real time, and also supports users to operate and intervene in the fermentation process; Through the analysis and modeling of historical data, anomaly detection and prediction functions are formed. When certain environmental parameters are abnormal, an early warning will be automatically issued to remind operators to take timely measures; Based on the data short-term prediction model, trend analysis of key variables in the fermentation process is carried out; Combined with the data management module, it integrates multi-source data and provides in-depth analysis to support intelligent control and optimization of the fermentation process.

6. The multi-dimensional parameter monitoring system for liquor fermentation process based on digital twin according to claim 5 is characterized in that: The data short-term prediction model includes a temperature and humidity prediction model; The temperature and humidity prediction model is a CNN-BiLSTM-Attention model based on sand cat group optimization; wherein, by simulating the global and local search mechanism of the predation behavior of the sand cat group, the hyperparameters of the temperature and humidity prediction model are globally optimized, so that the temperature and humidity prediction model escapes from the local optimum, optimizes the high-dimensional parameter space, and automatically adjusts the network structure.

7. The multi-dimensional parameter monitoring system for liquor fermentation process based on digital twin according to claim 6 is characterized in that: The convolution module of the temperature and humidity prediction model adopts a one-dimensional convolution with a step size of 1, the number of convolution kernels is set to 32, and the convolution kernel size is 3; the convolution layer uses the SAME filling strategy, and the pooling layer is configured with a pooling kernel size of 2; the BiLSTM layer is configured with 50 neurons; The Adam optimizer is used to train the prediction model, with the initial learning rate set to 0.01, the number of iterations to 100, the L2 regularization coefficient to 0.001, and the Dropout ratio to 0.

1. The initial population size of the sand cat swarm optimization algorithm is set to 20, the maximum number of iterations is 20, the sensitivity range is 2, and the fitness function uses the mean absolute percentage error as the evaluation indicator to optimize the number of convolution kernels, the number of BiLSTM neurons, the learning rate, and the L2 regularization coefficient in the prediction model.

8. The multi-dimensional parameter monitoring system for liquor fermentation process based on digital twin according to claim 4 is characterized in that: Perform outlier and mean processing on the collected data: Outlier processing, including: If the difference between the temperature or humidity data monitored in a short period of time and the historical average value exceeds 3 times the standard deviation, the data is determined to be an abnormal value, and the abnormal value is replaced by the mean of the valid data before and after it; Mean processing, including: The average of the data collected every half hour is used as the temperature and humidity monitoring data.

Images