Yeast frozen dough fermentation process control method based on fermentation kinetics model

Abstract

The present invention relates to the field of data processing technology, and in particular to a yeast frozen dough fermentation process control method based on a fermentation kinetics model. The method comprises the following steps: obtaining frozen dough data and yeast activity data; constructing a yeast dynamics basic model according to the yeast activity data to obtain a yeast dynamics basic model; extracting dough component characteristics according to the frozen dough data to obtain frozen dough component characteristic data; performing yeast fermentation dynamics evolution according to the frozen dough component characteristic data and the yeast dynamics basic model to obtain frozen dough fermentation process data; performing fermentation process evaluation according to the frozen dough fermentation process data to obtain fermentation process evaluation data, so as to perform yeast frozen dough fermentation process control auxiliary operations. The present invention can effectively simulate the control method of the fermentation process to improve the uniformity and stability of frozen dough fermentation, thereby improving the quality of baked products.

Description

Technical Field

[0001] The invention relates to the technical field of data processing, and in particular to a yeast frozen dough fermentation process control method based on a fermentation kinetics model. Background Art

[0002] In the food industry, yeast frozen dough is widely used in the production of various baked products. Frozen dough plays an important role in modern food processing because it can be stored for a long time and quickly thawed and fermented when needed. However, during the thawing and fermentation process of frozen dough, its fermentation effect will be affected by many factors such as yeast activity, dough composition, temperature gradient and external environment, resulting in problems such as uneven fermentation and unstable product quality. Therefore, how to accurately control the fermentation process of frozen dough and ensure the uniformity and stability of dough fermentation has become an important research topic in the food industry.

[0003] Traditional frozen dough fermentation control methods mainly rely on experience and experiments to improve the fermentation effect by adjusting the fermentation time, temperature and humidity. However, this method lacks in-depth understanding and prediction of the fermentation process, and is difficult to adapt to the individual differences of different batches of dough during the fermentation process, resulting in fluctuations in product quality. In addition, traditional methods cannot effectively simulate the differences in yeast fermentation dynamics inside and on the surface of the dough, especially after the frozen dough is thawed, the inconsistency of the fermentation speed between the surface and the inside is more prominent. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes a yeast frozen dough fermentation process control method based on a fermentation kinetics model to solve at least one of the above technical problems.

[0005] The present application provides a yeast frozen dough fermentation process control method based on a fermentation kinetics model, the method comprising:

[0006] S1, obtaining frozen dough data and yeast activity data;

[0007] S2. constructing a yeast dynamics basic model based on the yeast activity data to obtain a yeast dynamics basic model; extracting dough component features based on the frozen dough data to obtain frozen dough component feature data;

[0008] S3, performing yeast fermentation dynamics evolution according to the frozen dough component characteristic data and the yeast dynamics basic model to obtain the frozen dough fermentation process data;

[0009] S4. Perform fermentation process evaluation based on the frozen dough fermentation process data to obtain fermentation process evaluation data for performing yeast frozen dough fermentation process control auxiliary operations.

[0010] In the present invention, by acquiring frozen dough data and yeast activity data, and combining with the constructed yeast dynamic basic model, the behavior of frozen dough under different fermentation conditions can be accurately simulated and predicted, so that the fermentation process can be more accurately controlled, ensuring that the dough is always in the best state during the fermentation process, and finally improving the quality and stability of bread. The method not only takes into account the basic component characteristics of frozen dough, but also combines the dynamic characteristics of yeast activity, so that the fermentation process can be dynamically adjusted according to the changes of multidimensional characteristics, ensuring a comprehensive understanding and control of the fermentation process, and avoiding the errors caused by single-dimensional analysis. Through real-time evaluation of fermentation process data, the system can quickly identify anomalies or deviations in the fermentation process, and provide control auxiliary operations in time, make dynamic adjustments during the fermentation process, reduce variability in production, and improve production efficiency. Through strict fermentation process control, the consistency and quality of bread products can be significantly improved. Accurate fermentation control means that the dough can achieve the expected effect every time it is fermented, avoiding product quality problems caused by unstable fermentation process. Since the method is based on the fermentation dynamics model, it can adapt to the characteristics of different batches of frozen dough. Whether it is changes in dough ingredients, yeast activity, or environmental conditions, the system can dynamically adjust the fermentation process to ensure the consistency and high quality of different batches of production.

[0011] Optionally, S1 includes:

[0012] S11, obtaining frozen dough ingredient data through a preset frozen dough ingredient input interface;

[0013] S12, collecting physical properties of frozen dough through intelligent sensors to obtain physical property data of frozen dough;

[0014] S13, integrating the frozen dough ingredient data and the frozen dough physical property data to obtain frozen dough data;

[0015] S14. Obtain yeast activity data.

[0016] The present invention can fully obtain the composition data and physical property data of the frozen dough through the combination of the preset frozen dough component input interface and the intelligent sensor, ensure the comprehensiveness and accuracy of the data, and provide a solid foundation for the subsequent fermentation process control. By integrating the frozen dough component data and the physical property data, the complexity of data processing can be effectively reduced, and a comprehensive data set can be generated, which improves the data processing efficiency and ensures the accuracy of model construction and fermentation process prediction in the subsequent steps. The obtained frozen dough data and yeast activity data together provide high-quality input for the construction of the fermentation kinetics model, so that the model can more accurately simulate and predict the fermentation process of the frozen dough, thereby improving the accuracy of fermentation control and ensuring product quality. By collecting the physical property data of the frozen dough through the intelligent sensor, real-time monitoring of the fermentation process can be achieved. Once an abnormality occurs during the fermentation process, the system can dynamically adjust the fermentation conditions according to the real-time data to avoid fermentation out of control and ensure the consistency and stability of the product.

[0017] Optionally, S2 includes:

[0018] S21, generating basic parameters of yeast fermentation according to the yeast activity data to obtain basic parameter data of yeast fermentation;

[0019] S22, constructing a yeast dynamics basic model according to the yeast fermentation basic parameter data and the frozen dough physical property data to obtain a yeast dynamics basic model;

[0020] S23. Extract dough component features according to the frozen dough data to obtain frozen dough component feature data.

[0021] In the present invention, by generating yeast fermentation basic parameter data according to yeast activity data, scientific and accurate input parameters can be provided for the yeast dynamic basic model, ensuring that the construction process of the model is based on the actual yeast activity situation, so that the model can more accurately reflect the real dynamics in the fermentation process. The basic parameter data of yeast fermentation is combined with the physical property data of frozen dough to construct a dynamic model, so that the model can fully consider the interaction between yeast and dough in the fermentation process, improve the accuracy and reliability of the model, and better predict and control the fermentation process. By extracting dough component characteristics from frozen dough data, it is ensured that the control of the fermentation process is based on accurate dough component information. Accurate component feature extraction can improve the responsiveness of the fermentation process model and ensure that the model can effectively guide the fermentation process under different conditions. The dynamic model constructed by combining the basic parameter data of yeast fermentation and the physical property data of frozen dough can more accurately simulate the fermentation process under different conditions, improve the prediction ability of the fermentation process, provide more reliable reference information in the fermentation process, and reduce the risk of fermentation failure. By obtaining and utilizing the physical property data and yeast activity data of a specific batch of frozen dough, personalized control of the fermentation process can be achieved. Each batch of frozen dough can be adjusted according to its specific physical and chemical characteristics, ensuring consistent fermentation results and high-quality output.

[0022] Optionally, the dough component feature extraction is specifically as follows:

[0023] Performing multimodal data alignment according to frozen dough data to obtain multimodal alignment data;

[0024] Performing multimodal feature extraction according to the multimodal alignment data to obtain multimodal feature data;

[0025] Calculating feature weights according to the multimodal feature data to obtain multimodal feature weighted data;

[0026] Multimodal feature fusion is performed based on multimodal feature weighted data to obtain frozen dough ingredient feature data.

[0027] In the present invention, through multimodal data alignment and feature extraction, the method can extract feature information of multiple dimensions from frozen dough data, improve the comprehensiveness of data processing, ensure the key features of dough are captured from different angles and levels, and thus enhance the understanding and control of the fermentation process. The multimodal data alignment step ensures the consistency of frozen dough data from different modes in time and space dimensions. This alignment process reduces the deviation between different data sources, ensures the accuracy of subsequent feature extraction and fusion, and lays a solid foundation for generating reliable component feature data. Through feature weight calculation, the system can identify and amplify the most critical multimodal features, which ensures that in the subsequent fermentation process control, important feature data can better influence the decision of the model, optimize the utilization of feature data, and enable the model to guide the fermentation process more accurately. The multimodal feature fusion step integrates data of different modes to generate unified frozen dough component feature data, which not only simplifies the complexity of subsequent analysis, but also improves the integrity and consistency of the data, so that the fermentation kinetics model can make more accurate predictions based on comprehensive feature information. By fusing weighted feature data from different modalities, the generated dough component feature data is more comprehensive and accurate, enabling the yeast dynamics model to better cope with changes under various production conditions and ensure the consistency and reliability of the fermentation process in different batches and under different conditions. The precise extraction and fusion of multimodal features enables the model to respond to subtle changes during the fermentation process, thereby achieving more sophisticated process control, improving product quality, and ensuring that each fermentation process is in optimal condition.

[0028] Optionally, the multimodal data alignment is specifically:

[0029] Perform three-dimensional modeling according to the frozen dough data to obtain a three-dimensional model of the frozen dough;

[0030] Finite element calculation is performed according to the frozen dough three-dimensional model and the frozen dough data to obtain frozen dough finite element data;

[0031] Extract the finite element data of frozen dough and obtain the multimodal data of spatial distribution of frozen dough;

[0032] The spatial matrix of the multimodal data of spatial distribution of frozen dough is constructed to obtain multimodal alignment data;

[0033] The data extraction is as follows:

[0034] Perform spatial similarity screening according to the frozen dough finite element data to obtain the frozen dough spatial similarity screening data;

[0035] Extract spatial distribution multimodal data from the frozen dough spatial similarity screening data to obtain the frozen dough spatial distribution multimodal data;

[0036] The spatial similarity screening is as follows:

[0037] Perform clustering processing on frozen dough finite element data to obtain finite element clustering data;

[0038] According to the finite element clustering data, the frozen dough finite element data are divided into adjacent spaces to obtain finite element space division data;

[0039] The similarity of the frozen dough finite element data is calculated and screened according to the finite element space partitioning data to obtain the frozen dough space similarity screening data.

[0040] In the present invention, by three-dimensional modeling and finite element calculation of frozen dough data, the mechanical behavior and internal structure of dough under different conditions can be accurately simulated, which provides a reliable basis for subsequent data extraction and fermentation process control, so that the model can better reflect the actual physical state of the dough. The method can capture fine spatial distribution characteristics from the finite element data of frozen dough through spatial similarity screening and multimodal data extraction, including not only the overall spatial structure of the dough, but also local similarities and differences, ensuring that the control of the fermentation process can fully consider spatial factors. Through clustering processing and adjacent space division, the method can effectively classify and organize the finite element data, thereby reducing the complexity of calculation and ensuring the accuracy of data processing. In particular, in the process of spatial similarity screening, regions with similar mechanical and physical properties can be accurately identified, which improves the accuracy and robustness of the model. Through spatial matrix construction, spatially distributed multimodal data from different data sources are aligned. This alignment ensures the consistency of multimodal data in the spatial dimension, avoids possible spatial deviations between different data sources, and makes subsequent data fusion more reliable. Through spatial partitioning and similarity calculation of frozen dough finite element data, this method can dynamically adjust the parameters and structure of the model to adapt it to the characteristics of different batches of dough, so that the model can still maintain efficient and accurate control capabilities when facing different production conditions. Spatial similarity screening and clustering processing effectively reduce the amount of data to be processed, optimize the calculation process, and reduce resource consumption. Through screening and clustering, the model can focus on processing the most important and representative feature data, thereby improving calculation efficiency and reducing unnecessary calculation overhead.

[0041] Optionally, the feature weight calculation is specifically as follows:

[0042] Calculating a first feature weight according to the multimodal feature data to obtain first multimodal feature weighted data;

[0043] Performing a second feature weight calculation based on the multimodal feature data to obtain second multimodal feature weighted data, wherein the first feature weight calculation and the second feature weight calculation are different feature weight calculation methods;

[0044] Feature fitting is performed on the first multimodal feature weighted data and the second multimodal feature weighted data to obtain multimodal feature weighted data.

[0045] In the present invention, by using two different feature weight calculation methods for the same set of multimodal feature data, the importance of features can be analyzed from multiple perspectives. The impact of each feature in the fermentation process can be evaluated from multiple dimensions, thereby improving the accuracy of fermentation process control. Since the first and second feature weight calculation methods are different, this reduces the deviation or error caused by a single weight calculation method. The feature weighted data calculated by the two different methods are fused and fitted, which can enhance the reliability of the final weighted data and ensure that the fermentation process control is more stable and accurate. Through the feature fitting process, the first multimodal feature weighted data is fused with the second multimodal feature weighted data, which can integrate the advantages of the two calculation methods, generate higher quality multimodal feature weighted data, provide more accurate input for the subsequent fermentation process model, and optimize the entire fermentation control process. Using multiple feature weight calculation methods and fitting effectively reduces the calculation deviation that may be caused by a single calculation method, so that the model shows stronger robustness and stability in practical applications, especially in complex and changeable production environments. Through the combination and fitting of different weight calculation methods, the model can better adapt to the feature changes of different batches of frozen dough. Regardless of changes in the physical properties of the dough or environmental conditions, the model can dynamically adjust feature weights to ensure consistency and controllability of the fermentation process.

[0046] Optionally, the first feature weight calculation is specifically as follows:

[0047]

[0048] W i is the weighted data of the first multimodal feature, X i is the i-th multimodal feature data, J i is the signal-to-noise ratio data corresponding to the i-th multimodal feature data, K i is the spatial correlation data corresponding to the i-th multimodal feature data, L i is the temporal stability data corresponding to the i-th multimodal feature data, e is the natural exponential term, m i is the confidence data corresponding to the i-th multimodal feature data, n i is the frequency domain feature density data corresponding to the i-th multimodal feature data, P i is the complexity metric data corresponding to the i-th multimodal feature data, Qi is the geometric distribution coefficient data corresponding to the i-th multimodal feature data.

[0049] The present invention can evaluate the importance of each multimodal feature data from multiple dimensions, which helps to more accurately reflect the actual impact of the data and improve the prediction accuracy of the model. The main function of this formula is to calculate the weights for the multimodal feature data so that in the subsequent feature fusion or decision-making process, the influence of each feature can be fully considered to ensure that the model's evaluation of the features is accurate and comprehensive. Used to deal with the nonlinear effects of the signal-to-noise ratio, and sinK i It is used to deal with the impact of spatial correlation on weights and ensure the consistency of data under different spatial conditions. The natural index term e converts the confidence data m i With the frequency domain feature density data n i Combine and amplify or reduce the influence on the weight at an exponential level to reflect the changing trend of the data in the time dimension and its credibility. The exponential function can sensitively capture the influence of time changes on the weight. Processing complexity measure P i and the geometric distribution coefficient Q i The relationship between the two parameters is balanced. Through the inverse tangent function, the influence of the ratio of the two parameters on the weight is linearized, avoiding the excessive influence of extreme values ​​on the weight calculation. The combination of derivative and exponential function enables the formula to dynamically respond to changes in feature data, showing greater flexibility and adaptability in practical applications.

[0050] Optionally, the second feature weight calculation is specifically as follows:

[0051] Perform convolution calculation according to the multimodal feature data to obtain multimodal feature convolution layer data;

[0052] Perform clustering calculation on the multi-modal feature convolution layer data to obtain multi-modal feature clustering data;

[0053] Performing pooling calculation on the multimodal feature clustering data to obtain multimodal feature pooling data;

[0054] Perform masked multi-head self-attention calculation on the multimodal feature data to obtain multimodal masked multi-head self-attention data;

[0055] Feature weighted calculation is performed according to the multimodal masked multi-head self-attention data and the multimodal feature pooling data to obtain second multimodal feature weighted data.

[0056] In the present invention, through convolution calculation, deeper features can be extracted from multimodal feature data, local patterns and complex feature relationships in the data can be captured, and the expressiveness and prediction accuracy of the model can be improved. Clustering calculation is performed on the convolution layer data so that similar features can be effectively grouped, the accuracy of feature classification is improved, the influence of redundant information is reduced, and the model is helpful to better distinguish important features from minor features in feature weighted calculation. Through pooling calculation, the feature data can be reduced in dimension, the amount of data can be reduced while retaining key feature information, the computational efficiency of the model is improved, and the risk of overfitting is reduced, so that the model can still maintain efficient performance with fewer computing resources. Masked multi-head self-attention calculation can capture global relationships in feature data, especially the long-range dependencies between features, so that the model can comprehensively consider the complex interactions between features, thereby more accurately weighting the influence of each feature and improving the global perception ability of the model. Masking technology allows selective attention to specific features or feature combinations in self-attention calculations, ignoring irrelevant or noise features, making the calculation process more targeted, and can further improve the accuracy of feature weighted calculations. By combining multimodal masked multi-head self-attention data and multimodal feature pooling data for weighted calculation, the model can utilize deep features, local patterns and global relationships to achieve more accurate feature weighting, which can better reflect the actual role of each feature in the fermentation process and improve the final weighted data quality.

[0057] Optionally, S3 includes:

[0058] S31, obtaining yeast distribution data, wherein the yeast distribution data includes dough surface distribution data and dough internal distribution data;

[0059] S32, performing surface yeast fermentation dynamics evolution according to dough surface distribution data, frozen dough component characteristic data and yeast dynamics basic model to obtain first frozen dough fermentation process data;

[0060] S33, performing internal yeast fermentation dynamics evolution according to the dough internal distribution data, the frozen dough component characteristic data and the yeast dynamics basic model to obtain second frozen dough fermentation process data;

[0061] S34, generating fermentation process data from the first frozen dough fermentation process data and the second frozen dough fermentation process data to obtain the frozen dough fermentation process data.

[0062] In the present invention, by respectively obtaining the yeast distribution data of the dough surface and the interior, and performing targeted fermentation dynamics evolution simulation, the method can comprehensively cover the fermentation process of the dough, accurately predict the fermentation behavior of the dough at different levels, thereby optimizing the fermentation conditions and improving the uniformity and quality of the product. The method processes the fermentation process of the dough surface and the interior separately, and performs independent evolution calculations based on the respective characteristic data and yeast dynamics model. The refined hierarchical control can better cope with the fermentation needs of different regions and prevent product defects caused by uneven fermentation, such as excessive expansion of the surface or insufficient internal fermentation. By using yeast distribution data and frozen dough component characteristic data to perform differentiated fermentation dynamics evolution on the surface and the interior, the fermentation parameters can be flexibly adjusted to adapt to different types of dough and production environments, ensuring that the fermentation process can be customized according to specific needs, thereby improving the flexibility and adaptability of production.

[0063] Optionally, the fermentation process data generation is specifically as follows:

[0064] Performing spatial matching processing on the first frozen dough fermentation process data and the second frozen dough fermentation process data to obtain fermentation process matching data;

[0065] The first frozen dough fermentation process data and the second frozen dough fermentation process data are screened according to the fermentation process matching data to obtain fermentation process spatial overlap data and fermentation process spatial non-overlap data;

[0066] The secondary fermentation data of the fermentation process are obtained by calculating the secondary fermentation according to the spatial overlap number of the fermentation process;

[0067] According to the secondary fermentation data of the fermentation process, the spatial non-overlapping data of the fermentation process are linearly weighted and screened for overlap to obtain the frozen dough fermentation process data.

[0068] In the present invention, spatial matching processing is performed on the first and second frozen dough fermentation process data to ensure the spatial consistency of fermentation data from different levels, which helps to accurately reflect the overall changes of the entire dough during the fermentation process and improves the simulation accuracy of the model for the fermentation process. By screening the fermentation process data for spatial overlap and non-overlap, the fermentation behavior of different regions in the dough can be distinguished, the difference between the surface and internal fermentation processes can be revealed, and more accurate data support can be provided for subsequent fermentation control and optimization. Secondary fermentation calculation is performed on spatial overlap data, which can further optimize the fermentation process of these regions, help to make up for the deficiencies that may exist in the first fermentation process, ensure the uniformity and sufficiency of fermentation, and ultimately improve the fermentation quality of the dough. Linear weighting and overlap screening are performed on spatial non-overlap data, which can be used to perform more detailed control on these regions, optimize the details that are easily overlooked in the fermentation process, and ensure that the fermentation process of the entire dough is more balanced and consistent. Through secondary fermentation calculation and linear weighted processing, the fermentation process data finally generated is more accurate, the error caused by the unevenness in the fermentation process is reduced, the prediction accuracy and reliability of the model are enhanced, and the fermentation quality in the production process is ensured to reach the optimal level. By screening and weighting the overlap of different fermentation areas, producers can better control the key variables in the fermentation process, which helps to adjust the fermentation conditions in time in actual production, reduce fermentation unevenness, and ensure product consistency and high quality.

[0069] The objects of the present invention are:

[0070] 1. The present invention firstly uses a preset interface and intelligent sensors to accurately collect the composition and physical property data of frozen dough, and combines it with yeast activity data to form a multi-dimensional and comprehensive data foundation.

[0071] 2. Through in-depth analysis of yeast activity data, accurate basic parameters of fermentation dynamics are generated, and combined with the physical property data of frozen dough, a basic yeast dynamics model is constructed to accurately simulate the behavior of yeast during the fermentation process, thereby providing accurate dynamic predictions and ensuring the efficiency and stability of the fermentation process. The multimodal characteristics of the data are utilized to ensure that the model can capture the key characteristics of the dough in different dimensions, improving the prediction accuracy and adaptability of the model.

[0072] 3. The present invention simulates the evolution of yeast fermentation dynamics and generates detailed data on the frozen dough fermentation process. In this process, the fermentation behaviors on the surface and inside of the dough are processed and simulated separately, ensuring comprehensive coverage and fine control of the dough fermentation process. By separately processing overlapping and non-overlapping data, the present invention not only improves the accuracy and consistency of the fermentation process data, but also effectively reduces the deviations and uncertainties that may occur during the fermentation process, ensuring the uniformity and quality stability of the dough during the fermentation process.

[0073] 4. The present invention can timely discover and adjust possible abnormal conditions in the fermentation process through real-time evaluation of the data of the frozen dough fermentation process. Through the data-driven feedback mechanism, the system can dynamically adjust the fermentation parameters to ensure that the fermentation process is always carried out in the best state. Real-time monitoring and feedback control greatly improve the controllability and production efficiency of the fermentation process and reduce product defects caused by unstable fermentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Other features, objects and advantages of the present application will become more apparent by reading the detailed description of non-limiting implementations made with reference to the following drawings:

[0075] Figure 1 A flowchart showing a method for controlling a yeast frozen dough fermentation process based on a fermentation kinetics model according to an embodiment of the present invention is provided;

[0076] Figure 2 A flowchart showing a method for collecting data of yeast frozen dough according to an embodiment is shown;

[0077] Figure 3 A flowchart showing a method for constructing a yeast dynamics basic model and extracting frozen dough component characteristics according to an embodiment of the present invention is shown;

[0078] Figure 4 A flowchart showing a method for evolving yeast fermentation dynamics according to an embodiment of the present invention is provided;

[0079] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0080] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0081] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0082] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0083] See also Figures 1 to 4 The present application provides a yeast frozen dough fermentation process control method based on a fermentation kinetics model, the method comprising:

[0084] S1, obtaining frozen dough data and yeast activity data;

[0085] In one embodiment provided by the present invention, the physical properties of the frozen dough (such as temperature, humidity, and water content) are monitored in real time using instruments such as temperature and humidity sensors, and the dough is photographed using a camera to record the original frozen dough data.

[0086] The sugar, protein and fat contents in the frozen dough were detected by biochemical analysis methods (such as high performance liquid chromatography) to obtain dough composition data.

[0087] The yeast activity tester was used to measure the fermentation activity of yeast, and the fermentation efficiency of yeast under different temperature and time conditions was recorded to obtain yeast activity data.

[0088] S2. constructing a yeast dynamics basic model based on the yeast activity data to obtain a yeast dynamics basic model; extracting dough component features based on the frozen dough data to obtain frozen dough component feature data;

[0089] In one embodiment provided by the present invention, a nonlinear regression method is used to fit yeast activity data and a yeast kinetic model based on the Michaelis-Menten equation is established.

[0090] The parameters of the fitting model are optimized (such as using a genetic algorithm or a particle swarm algorithm) to determine the optimal model parameters and obtain a basic yeast dynamics model.

[0091] The model is validated through cross-validation method to ensure the generalization ability of the model.

[0092] Principal component analysis (PCA) was used to reduce the dimension of frozen dough ingredient data, extract the main component features, and obtain dough ingredient characteristic data.

[0093] Clustering algorithms (such as K-means clustering) are used to perform cluster analysis on dough ingredient data, identify dough samples with different characteristic categories, and obtain frozen dough ingredient characteristic data.

[0094] S3, performing yeast fermentation dynamics evolution according to the frozen dough component characteristic data and the yeast dynamics basic model to obtain the frozen dough fermentation process data;

[0095] The present invention provides an embodiment in which the yeast dynamics basic model constructed in S2 is used to simulate the fermentation dynamics changes of yeast under different dough composition conditions according to dough composition characteristic data.

[0096] The dynamic evolution of yeast fermentation process is simulated by numerical simulation methods (such as Runge-Kutta method) to obtain fermentation kinetic data at different times.

[0097] The simulation results are fused to generate the fermentation process data of frozen dough during the entire fermentation process.

[0098] Perform time series analysis on the fermentation process data to extract key time points and characteristic values ​​(such as maximum yeast activity time and sugar consumption rate) in the fermentation process.

[0099] Use machine learning algorithms (such as random forests) to classify fermentation process data and identify the main factors affecting the fermentation process.

[0100] S4. Perform fermentation process evaluation based on the frozen dough fermentation process data to obtain fermentation process evaluation data for performing yeast frozen dough fermentation process control auxiliary operations.

[0101] In one embodiment provided by the present invention, a mapping model between the fermentation process and the dough quality is constructed through multiple regression analysis to evaluate the influence of the fermentation process on the final dough quality.

[0102] The decision tree algorithm is used to establish the fermentation process control rules, and the fermentation conditions (such as temperature and time) are adjusted in real time according to the fermentation process data to optimize the dough quality.

[0103] The evaluation data is input into the fermentation process control system, and the PID controller is used to dynamically adjust the fermentation conditions.

[0104] Based on the evaluation results, the fermentation process is optimized to ensure that the dough reaches the optimal fermentation state and improve the final quality of the frozen dough.

[0105] In the present invention, by acquiring frozen dough data and yeast activity data, and combining with the constructed yeast dynamic basic model, the behavior of frozen dough under different fermentation conditions can be accurately simulated and predicted, so that the fermentation process can be more accurately controlled, ensuring that the dough is always in the best state during the fermentation process, and finally improving the quality and stability of bread. The method not only takes into account the basic component characteristics of frozen dough, but also combines the dynamic characteristics of yeast activity, so that the fermentation process can be dynamically adjusted according to the changes of multidimensional characteristics, ensuring a comprehensive understanding and control of the fermentation process, and avoiding the errors caused by single-dimensional analysis. Through real-time evaluation of fermentation process data, the system can quickly identify anomalies or deviations in the fermentation process, and provide control auxiliary operations in time, make dynamic adjustments during the fermentation process, reduce variability in production, and improve production efficiency. Through strict fermentation process control, the consistency and quality of bread products can be significantly improved. Accurate fermentation control means that the dough can achieve the expected effect every time it is fermented, avoiding product quality problems caused by unstable fermentation process. Since the method is based on the fermentation dynamics model, it can adapt to the characteristics of different batches of frozen dough. Whether it is changes in dough ingredients, yeast activity, or environmental conditions, the system can dynamically adjust the fermentation process to ensure the consistency and high quality of different batches of production.

[0106] Optionally, S1 includes:

[0107] S11, obtaining frozen dough ingredient data through a preset frozen dough ingredient input interface;

[0108] The present invention provides an embodiment in which a frozen dough ingredient input interface is developed using a graphical user interface (GUI) design tool (such as Qt, JavaFX), through which a user can input the proportion of various dough ingredients (such as flour, sugar, fat, additives, etc.).

[0109] The interface includes preset ingredient drop-down menus or input boxes, as well as unit conversion functions to ensure standardization and uniformity of data input.

[0110] The validation function of input data is added through programming logic to ensure that the input data meets the preset range and format requirements. For example, the flour ratio must not exceed 100%, and the content of additives should be within the specified percentage range.

[0111] Use regular expressions to verify the input data format in real time to avoid data entry errors.

[0112] S12, collecting physical properties of frozen dough through intelligent sensors to obtain physical property data of frozen dough;

[0113] In one embodiment provided by the present invention, intelligent sensors (such as temperature and humidity sensors, spectrum analyzers) are deployed in a frozen dough production line to monitor the physical properties of the frozen dough (such as temperature, humidity, texture) in real time.

[0114] The sensor is connected to the central control system through the Internet of Things (IoT) platform to transmit real-time data to ensure the continuity and accuracy of data collection. Use a collection program (such as a Python or C++-based program) to read data from the sensor regularly and format the data (such as converting temperature data into Celsius format).

[0115] Using the MQTT protocol or HTTP API, the collected data on the physical properties of frozen dough are transmitted to the central database for storage and subsequent processing.

[0116] S13, integrating the frozen dough ingredient data and the frozen dough physical property data to obtain frozen dough data;

[0117] In one embodiment provided by the present invention, an integration script (such as a Python or R language script) is written to match and integrate the frozen dough ingredient data and the physical property data in the database, and the data is associated based on the batch number or timestamp.

[0118] The two types of data tables are inner joined (INNER JOIN) through the SQL query language to generate a data table containing composition and physical properties.

[0119] S14. Obtain yeast activity data.

[0120] In one embodiment provided by the present invention, a yeast activity measuring instrument (such as a carbon dioxide emission measuring instrument, a photometer) is used to measure the fermentation activity of yeast under different temperature and time conditions, and the carbon dioxide emission or absorbance value is recorded as an activity indicator.

[0121] Through a series of experiments, the activity data of yeast under different conditions are obtained to form a yeast kinetic characteristic curve.

[0122] Use experimental data recording software (such as Excel or laboratory information management system LIMS) to record the yeast activity test results and store them by experimental conditions (such as temperature, time).

[0123] The activity data is smoothed (eg, using a moving average method) to eliminate random fluctuations in the measurement and obtain a more stable yeast activity data curve.

[0124] The present invention can fully obtain the composition data and physical property data of the frozen dough through the combination of the preset frozen dough component input interface and the intelligent sensor, ensure the comprehensiveness and accuracy of the data, and provide a solid foundation for the subsequent fermentation process control. By integrating the frozen dough component data and the physical property data, the complexity of data processing can be effectively reduced, and a comprehensive data set can be generated, which improves the data processing efficiency and ensures the accuracy of model construction and fermentation process prediction in the subsequent steps. The obtained frozen dough data and yeast activity data together provide high-quality input for the construction of the fermentation kinetics model, so that the model can more accurately simulate and predict the fermentation process of the frozen dough, thereby improving the accuracy of fermentation control and ensuring product quality. By collecting the physical property data of the frozen dough through the intelligent sensor, real-time monitoring of the fermentation process can be achieved. Once an abnormality occurs during the fermentation process, the system can dynamically adjust the fermentation conditions according to the real-time data to avoid fermentation out of control and ensure the consistency and stability of the product.

[0125] Optionally, S2 includes:

[0126] S21, generating basic parameters of yeast fermentation according to the yeast activity data to obtain basic parameter data of yeast fermentation;

[0127] In one embodiment of the present invention, the raw data obtained from the yeast activity tester generally includes yeast fermentation activity curves under multiple experiments. First, data cleaning technology is used to remove outliers and noise data (such as filtering outliers by IQR method).

[0128] Interpolate time series data to fill in possible missing values. Linear interpolation or spline interpolation can be used.

[0129] The yeast activity curve was fitted using a nonlinear fitting algorithm (such as the Levenberg-Marquardt algorithm) to extract key fermentation parameters, including the maximum specific growth rate (μ_max), half-saturation constant (K_s), and generation coefficient (Y_x / s).

[0130] The yeast activity curve was segmented by curve fitting, and the growth rate and decay rate of the yeast were calculated at different time points, and these values ​​were recorded as basic fermentation parameters.

[0131] The extracted fermentation parameters were statistically analyzed (e.g., by multiple comparisons or analysis of variance) to verify the correlation and consistency between the parameters.

[0132] The fermentation parameters were further optimized using genetic algorithm or particle swarm optimization algorithm to improve the accuracy of parameters and model fit.

[0133] S22, constructing a yeast dynamics basic model according to the yeast fermentation basic parameter data and the frozen dough physical property data to obtain a yeast dynamics basic model;

[0134] The present invention provides an embodiment in which a suitable model structure is selected according to the biodynamic principle of yeast fermentation, such as a Monod model, a Logistic model or a Baranyi-Roberts model. The model should be able to reflect the fermentation behavior of yeast under different physical conditions. The yeast kinetic process is constructed using differential equations, in which the change of yeast concentration over time is the core of the model. The model equation can be expressed as: where X is the yeast concentration, S is the substrate concentration, k8 is the death rate constant, and u ；<＝ K is the maximum specific growth rate, which indicates the maximum growth rate of yeast cells under optimal conditions. C is the half-saturation constant or Michaelis constant. It means that when the matrix concentration S reaches K C When , the growth rate of yeast reaches half of the maximum growth rate. The basic parameter data of yeast fermentation obtained in S1 are used as model parameters, and the differential equations are solved by numerical solution techniques (such as Euler method or Runge-Kutta method) to simulate the growth and fermentation process of yeast under the physical properties of frozen dough. The model parameters are fitted by least squares method or Bayesian estimation to maximize the fit between the model and the experimental data. The fitting process can use optimization tools such as the optimize module in the SciPy library. The model is verified by cross-validation method, and the data set is divided into training set and test set to evaluate the prediction accuracy and generalization ability of the model. According to the verification results, the model structure and parameters are adjusted, and correction items (such as temperature compensation factor) are introduced when necessary to improve the accuracy of the model.

[0135] S23. Extract dough component features according to the frozen dough data to obtain frozen dough component feature data.

[0136] The present invention provides an embodiment of selecting and reducing the dimension:

[0137] Key component features related to yeast fermentation are extracted from frozen dough data, such as sugar content, moisture content, pH value, etc. The principal component analysis (PCA) method is used to reduce the dimensionality of high-dimensional component data to reduce the feature dimension and retain the main information.

[0138] If the number of features is large, a selection algorithm based on feature importance (such as random forest-based feature selection) can be used to determine the most important component features.

[0139] Use multivariate linear regression analysis or polynomial regression analysis to determine the relationship between each component characteristic of frozen dough and yeast fermentation kinetic parameters. Construct a regression model to describe the effect of component characteristics on the fermentation process.

[0140] The dough component characteristics are clustered through cluster analysis (such as K-means clustering), the dough samples are divided into different categories, and the characteristic differences between different categories are analyzed.

[0141] The extracted component feature data are stored in a structured database and a unique identifier is associated with each sample.

[0142] In the present invention, by generating yeast fermentation basic parameter data according to yeast activity data, scientific and accurate input parameters can be provided for the yeast dynamic basic model, ensuring that the construction process of the model is based on the actual yeast activity situation, so that the model can more accurately reflect the real dynamics in the fermentation process. The basic parameter data of yeast fermentation is combined with the physical property data of frozen dough to construct a dynamic model, so that the model can fully consider the interaction between yeast and dough in the fermentation process, improve the accuracy and reliability of the model, and better predict and control the fermentation process. By extracting dough component characteristics from frozen dough data, it is ensured that the control of the fermentation process is based on accurate dough component information. Accurate component feature extraction can improve the responsiveness of the fermentation process model and ensure that the model can effectively guide the fermentation process under different conditions. The dynamic model constructed by combining the basic parameter data of yeast fermentation and the physical property data of frozen dough can more accurately simulate the fermentation process under different conditions, improve the prediction ability of the fermentation process, provide more reliable reference information in the fermentation process, and reduce the risk of fermentation failure. By obtaining and utilizing the physical property data and yeast activity data of a specific batch of frozen dough, personalized control of the fermentation process can be achieved. Each batch of frozen dough can be adjusted according to its specific physical and chemical characteristics, ensuring consistent fermentation results and high-quality output.

[0143] Optionally, the dough component feature extraction is specifically as follows:

[0144] Performing multimodal data alignment according to frozen dough data to obtain multimodal alignment data;

[0145] In one embodiment provided by the present invention, multimodal data of frozen dough is collected from different data sources, including image data (such as microscopic images), physical property data (such as temperature and humidity), and component data (such as chemical component ratio).

[0146] Preliminary preprocessing of each modality data, such as denoising of image data (using Gaussian filter) and standardization of physical data and composition data (such as Z-score standardization), is performed to ensure the quality and consistency of the data.

[0147] The dynamic time warping (DTW) algorithm is used to align multimodal data on different time axes to ensure the synchronization of each modality in the time dimension. DTW can achieve alignment by minimizing the distance between time series.

[0148] For the alignment of image data with other modal data, feature matching algorithms (such as SIFT or ORB) can be used to extract key points in the image and match them with features in other modal data to achieve spatial alignment.

[0149] The Pearson correlation coefficient or mutual information was used to quantify the correlation between the aligned data to ensure the accuracy of the alignment.

[0150] If the alignment result is not satisfactory, adjust the alignment parameters or choose a more suitable data alignment method and repeat the alignment process.

[0151] Performing multimodal feature extraction according to the multimodal alignment data to obtain multimodal feature data;

[0152] In one embodiment of the present invention, a feature extraction method is selected based on the aligned data. For example, a convolutional neural network (CNN) is used to extract deep features for image data, and a statistical feature extraction method (such as mean, variance, kurtosis) is used to extract features for physical and component data.

[0153] For image feature extraction, a pre-trained ResNet or VGG network can be used for transfer learning to extract representative image feature vectors.

[0154] The multimodal alignment data is input into the respective feature extraction modules to extract the corresponding modal feature data. For example, the convolution and pooling operations of CNN on image data can obtain image feature vectors; physical and compositional data can extract features through linear regression or principal component analysis (PCA).

[0155] After extracting features from the data of each modality, a corresponding feature matrix is ​​formed and saved in a structured database to ensure that the feature data of each sample can correspond to its original data.

[0156] The feature data of each modality is standardized so that the feature data can be compared and fused on the same scale. The standardization method can include Z-score standardization or Min-Max scaling.

[0157] Calculating feature weights according to the multimodal feature data to obtain multimodal feature weighted data;

[0158] In one embodiment of the present invention, a feature weight calculation method based on information entropy or mutual information is used to quantify the importance of each modal feature in the overall feature. For example, the information entropy method evaluates the weight of a feature by calculating the entropy value of the feature distribution.

[0159] For classification problems, the weights can be calculated using feature importance evaluation methods based on classification performance (such as random forest feature importance).

[0160] For each modal feature, the information gain or mutual information relative to the target variable (such as dough fermentation effect) is calculated as the basis for feature weighting.

[0161] The feature weights of each mode are normalized so that the sum of the feature weights of all modes is 1, ensuring the rationality of the feature weighting process.

[0162] The calculated weights are applied to the feature data of each modality to generate multimodal feature weighted data.

[0163] Multimodal feature fusion is performed based on multimodal feature weighted data to obtain frozen dough ingredient feature data.

[0164] In one embodiment provided by the present invention, a multimodal feature fusion method is selected, such as a simple weighted average method, a principal component analysis (PCA) fusion method, or a feature fusion method based on deep learning (such as a multimodal neural network).

[0165] If deep learning is used, a multi-layer perceptron (MLP) can be constructed to fuse multimodal features and ultimately generate a unified feature representation.

[0166] The weighted multimodal feature data is fused. When using the weighted average method, the feature values ​​of different modes are weighted and summed according to their weights. The PCA method maps multimodal features to a common low-dimensional space by reducing dimensionality.

[0167] If a deep learning method is used, the feature data of each modality is used as input to train the fusion model and obtain the final fusion feature representation.

[0168] In the present invention, through multimodal data alignment and feature extraction, the method can extract feature information of multiple dimensions from frozen dough data, improve the comprehensiveness of data processing, ensure the key features of dough are captured from different angles and levels, and thus enhance the understanding and control of the fermentation process. The multimodal data alignment step ensures the consistency of frozen dough data from different modes in time and space dimensions. This alignment process reduces the deviation between different data sources, ensures the accuracy of subsequent feature extraction and fusion, and lays a solid foundation for generating reliable component feature data. Through feature weight calculation, the system can identify and amplify the most critical multimodal features, which ensures that in the subsequent fermentation process control, important feature data can better influence the decision of the model, optimize the utilization of feature data, and enable the model to guide the fermentation process more accurately. The multimodal feature fusion step integrates data of different modes to generate unified frozen dough component feature data, which not only simplifies the complexity of subsequent analysis, but also improves the integrity and consistency of the data, so that the fermentation kinetics model can make more accurate predictions based on comprehensive feature information. By fusing weighted feature data from different modalities, the generated dough component feature data is more comprehensive and accurate, enabling the yeast dynamics model to better cope with changes under various production conditions and ensure the consistency and reliability of the fermentation process in different batches and under different conditions. The precise extraction and fusion of multimodal features enables the model to respond to subtle changes during the fermentation process, thereby achieving more sophisticated process control, improving product quality, and ensuring that each fermentation process is in optimal condition.

[0169] Optionally, the multimodal data alignment is specifically:

[0170] Perform three-dimensional modeling according to the frozen dough data to obtain a three-dimensional model of the frozen dough;

[0171] In one embodiment of the present invention, two-dimensional cross-sectional image data or X-ray / CT scan data of frozen dough are collected and used to construct a three-dimensional model.

[0172] The voxel-based 3D reconstruction algorithm is used to reconstruct the 2D cross-sectional image data into a 3D model. The algorithm steps include: image segmentation, feature extraction, voxel generation, and voxel splicing.

[0173] Import the voxel data into a 3D reconstruction tool (such as MATLAB or Python's PyVista library) to generate a 3D model of the frozen dough.

[0174] The generated 3D model is smoothed and meshed to reduce the number of triangle faces and improve the computational efficiency of the model.

[0175] Finite element calculation is performed according to the frozen dough three-dimensional model and the frozen dough data to obtain frozen dough finite element data;

[0176] The present invention provides an embodiment in which a finite element mesh is formed on a three-dimensional model of frozen dough using a Delaunay triangulation or an octree partitioning method. The goal of the meshing is to decompose the three-dimensional model into a finite number of small units.

[0177] The material properties of the 3D model are assigned to the finite element mesh according to the physical properties of the dough (e.g. elastic modulus, density).

[0178] Select finite element analysis software (such as ANSYS, Abaqus) to perform mechanical or thermal simulation calculations on the frozen dough three-dimensional model.

[0179] Set boundary conditions and initial conditions, such as applied external forces or temperature fields, calculate the stress and strain distribution of the dough under these conditions, and obtain the finite element data of the frozen dough.

[0180] Post-process the finite element calculation results, such as visualization of stress and strain fields, extraction and analysis of result data. Use contour lines or contour surfaces to display the stress and deformation inside the frozen dough.

[0181] Extract the finite element data of frozen dough and obtain the multimodal data of spatial distribution of frozen dough;

[0182] In one embodiment, the present invention performs unsupervised clustering processing on finite element data, and uses algorithms such as K-means and DBSCAN to divide the finite element data into different spatial clustering regions. Each cluster represents an area with similar mechanical properties or temperature fields inside the dough. Finite element clustering data is generated based on the clustering results, and the cluster category to which each grid unit belongs is identified.

[0183] Based on the clustering results, the finite element model is divided into adjacent spaces. The model is divided into multiple adjacent spatial regions using a Voronoi diagram or a Delaunay triangulation. The adjacency relationship within each cluster is determined, and finite element space partitioning data is generated to describe the adjacency information between each spatial region. The similarity between adjacent spatial regions is calculated using a spatial similarity calculation method (such as cosine similarity or Euclidean distance), and a spatial region with high similarity is selected to form the frozen dough spatial similarity screening data. The similarity screening data is stored in the form of an adjacency matrix.

[0184] The spatial matrix of the multimodal data of spatial distribution of frozen dough is constructed to obtain multimodal alignment data;

[0185] In one embodiment, the present invention provides a method for associating spatial similarity screening data with physical and component data of the original frozen dough to extract multimodal data corresponding to spatially similar regions. For each similar spatial region, the corresponding image, physical properties and component data are extracted to form a spatially distributed multimodal data set.

[0186] The data extraction is as follows:

[0187] Perform spatial similarity screening according to the frozen dough finite element data to obtain the frozen dough spatial similarity screening data;

[0188] In one embodiment of the present invention, the extracted multimodal data is subjected to dimensionality reduction processing (such as PCA or t-SNE) to reduce the data dimension and retain the main features. The multimodal data is standardized to ensure that different modal data are analyzed on the same scale.

[0189] Extract spatial distribution multimodal data from the frozen dough spatial similarity screening data to obtain the frozen dough spatial distribution multimodal data;

[0190] In one embodiment, the present invention provides a method for associating spatial similarity screening data with physical and component data of the original frozen dough to extract multimodal data corresponding to spatially similar regions. For each similar spatial region, the corresponding image, physical properties and component data are extracted to form a spatially distributed multimodal data set.

[0191] The spatial similarity screening is as follows:

[0192] Perform clustering processing on frozen dough finite element data to obtain finite element clustering data;

[0193] According to the finite element clustering data, the frozen dough finite element data are divided into adjacent spaces to obtain finite element space division data;

[0194] The similarity of the frozen dough finite element data is calculated and screened according to the finite element space partitioning data to obtain the frozen dough space similarity screening data.

[0195] In one embodiment, the present invention performs unsupervised clustering processing on finite element data, and uses algorithms such as K-means and DBSCAN to divide the finite element data into different spatial clustering regions. Each cluster represents an area with similar mechanical properties or temperature fields inside the dough. Finite element clustering data is generated based on the clustering results, and the cluster category to which each grid unit belongs is identified.

[0196] Based on the clustering results, the finite element model is divided into adjacent spaces. The model is divided into multiple adjacent spatial regions using a Voronoi diagram or a Delaunay triangulation. The adjacency relationship within each cluster is determined, and finite element space partitioning data is generated to describe the adjacency information between each spatial region. The similarity between adjacent spatial regions is calculated using a spatial similarity calculation method (such as cosine similarity or Euclidean distance), and a spatial region with high similarity is selected to form the frozen dough spatial similarity screening data. The similarity screening data is stored in the form of an adjacency matrix.

[0197] In the present invention, by three-dimensional modeling and finite element calculation of frozen dough data, the mechanical behavior and internal structure of dough under different conditions can be accurately simulated, which provides a reliable basis for subsequent data extraction and fermentation process control, so that the model can better reflect the actual physical state of the dough. The method can capture fine spatial distribution characteristics from the finite element data of frozen dough through spatial similarity screening and multimodal data extraction, including not only the overall spatial structure of the dough, but also local similarities and differences, ensuring that the control of the fermentation process can fully consider spatial factors. Through clustering processing and adjacent space division, the method can effectively classify and organize the finite element data, thereby reducing the complexity of calculation and ensuring the accuracy of data processing. In particular, in the process of spatial similarity screening, regions with similar mechanical and physical properties can be accurately identified, which improves the accuracy and robustness of the model. Through spatial matrix construction, spatially distributed multimodal data from different data sources are aligned. This alignment ensures the consistency of multimodal data in the spatial dimension, avoids possible spatial deviations between different data sources, and makes subsequent data fusion more reliable. Through spatial partitioning and similarity calculation of frozen dough finite element data, this method can dynamically adjust the parameters and structure of the model to adapt it to the characteristics of different batches of dough, so that the model can still maintain efficient and accurate control capabilities when facing different production conditions. Spatial similarity screening and clustering processing effectively reduce the amount of data to be processed, optimize the calculation process, and reduce resource consumption. Through screening and clustering, the model can focus on processing the most important and representative feature data, thereby improving calculation efficiency and reducing unnecessary calculation overhead.

[0198] Optionally, the feature weight calculation is specifically as follows:

[0199] Calculating a first feature weight according to the multimodal feature data to obtain first multimodal feature weighted data;

[0200] In one embodiment provided by the present invention, information gain is selected as the first feature weight calculation method. Information gain calculates the degree to which the uncertainty of a feature to a target variable is reduced, and is used to measure the importance of the feature. For each modal feature, the entropy value of the feature and the conditional entropy between the feature and the target variable are calculated. Information gain is calculated. The information gain value of each feature calculated is used as a weight to generate first multimodal feature weighted data.

[0201] Performing a second feature weight calculation based on the multimodal feature data to obtain second multimodal feature weighted data, wherein the first feature weight calculation and the second feature weight calculation are different feature weight calculation methods;

[0202] In one embodiment of the present invention, mutual information is selected as the second feature weight calculation method. Mutual information measures the dependency between two variables and can be used to evaluate the strength of association between a feature and a target variable. The mutual information between each modal feature and the target variable is calculated. The mutual information value of each feature is used as a weight to generate second multimodal feature weighted data.

[0203] Feature fitting is performed on the first multimodal feature weighted data and the second multimodal feature weighted data to obtain multimodal feature weighted data.

[0204] The present invention provides an embodiment in which linear regression or weighted average method is selected as a feature fitting method. Through fitting, the results of different feature weight calculation methods can be combined to obtain more stable feature weights. The first multimodal feature weighted data and the second multimodal feature weighted data are used as inputs and fitted using a linear regression model. Alternatively, if the weighted average method is selected, the two weighted data are directly weighted averaged.

[0205] In the present invention, by using two different feature weight calculation methods for the same set of multimodal feature data, the importance of features can be analyzed from multiple perspectives. The impact of each feature in the fermentation process can be evaluated from multiple dimensions, thereby improving the accuracy of fermentation process control. Since the first and second feature weight calculation methods are different, this reduces the deviation or error caused by a single weight calculation method. The feature weighted data calculated by the two different methods are fused and fitted, which can enhance the reliability of the final weighted data and ensure that the fermentation process control is more stable and accurate. Through the feature fitting process, the first multimodal feature weighted data is fused with the second multimodal feature weighted data, which can integrate the advantages of the two calculation methods, generate higher quality multimodal feature weighted data, provide more accurate input for the subsequent fermentation process model, and optimize the entire fermentation control process. Using multiple feature weight calculation methods and fitting effectively reduces the calculation deviation that may be caused by a single calculation method, so that the model shows stronger robustness and stability in practical applications, especially in complex and changeable production environments. Through the combination and fitting of different weight calculation methods, the model can better adapt to the feature changes of different batches of frozen dough. Regardless of changes in the physical properties of the dough or environmental conditions, the model can dynamically adjust feature weights to ensure consistency and controllability of the fermentation process.

[0206] Optionally, the first feature weight calculation is specifically as follows:

[0207]

[0208] W i is the weighted data of the first multimodal feature, X i is the i-th multimodal feature data, J i is the signal-to-noise ratio data corresponding to the i-th multimodal feature data, K i is the spatial correlation data corresponding to the i-th multimodal feature data, L i is the temporal stability data corresponding to the i-th multimodal feature data, e is the natural exponential term, m i is the confidence data corresponding to the i-th multimodal feature data, n i is the frequency domain feature density data corresponding to the i-th multimodal feature data, P i is the complexity metric data corresponding to the i-th multimodal feature data, Q i is the geometric distribution coefficient data corresponding to the i-th multimodal feature data.

[0209] The present invention can evaluate the importance of each multimodal feature data from multiple dimensions, which helps to more accurately reflect the actual impact of the data and improve the prediction accuracy of the model. The main function of this formula is to calculate the weights for the multimodal feature data so that in the subsequent feature fusion or decision-making process, the influence of each feature can be fully considered to ensure that the model's evaluation of the features is accurate and comprehensive. Used to deal with the nonlinear effects of the signal-to-noise ratio, and sinK i It is used to deal with the impact of spatial correlation on weights and ensure the consistency of data under different spatial conditions. The natural index term e converts the confidence data m i With the frequency domain feature density data n i Combine and amplify or reduce the influence on the weight at an exponential level to reflect the changing trend of the data in the time dimension and its credibility. The exponential function can sensitively capture the influence of time changes on the weight. Processing complexity measure P i and the geometric distribution coefficient Q i The relationship between the two parameters is balanced. Through the inverse tangent function, the influence of the ratio of the two parameters on the weight is linearized, avoiding the excessive influence of extreme values ​​on the weight calculation. The combination of derivative and exponential function enables the formula to dynamically respond to changes in feature data, showing greater flexibility and adaptability in practical applications.

[0210] Optionally, the second feature weight calculation is specifically as follows:

[0211] Perform convolution calculation according to the multimodal feature data to obtain multimodal feature convolution layer data;

[0212] In one embodiment of the present invention, a one-dimensional or two-dimensional convolution operation is used to perform convolution calculation on multimodal feature data. The purpose of the convolution operation is to extract local patterns or structural features in the feature data. A convolution kernel (such as 3x3 or 5x5) and a step size are selected.

[0213] The convolutional layer in the convolutional neural network (CNN) is used to calculate the input multimodal feature data. The output of the convolutional layer, that is, the multimodal feature convolutional layer data, reflects the local pattern of each feature. Applying a nonlinear activation function, such as ReLU (Rectified Linear Unit), to the convolutional layer output data increases the nonlinear expression ability of the model.

[0214] Perform clustering calculation on the multi-modal feature convolution layer data to obtain multi-modal feature clustering data;

[0215] In one embodiment of the present invention, clustering analysis is performed on the convolutional layer data using a clustering algorithm such as K-means or DBSCAN. The purpose of clustering is to group similar feature data for subsequent processing.

[0216] Perform dimensionality reduction on the convolutional layer data (such as using PCA or t-SNE) to reduce the dimension of the data and retain the main features.

[0217] Using the selected clustering algorithm, the convolutional layer data is divided into several clusters to generate multimodal feature clustering data. The clustering result can be represented as the cluster identifier to which each data point belongs.

[0218] Performing pooling calculation on the multimodal feature clustering data to obtain multimodal feature pooling data;

[0219] In one embodiment of the present invention, a maximum pooling or average pooling operation is used to perform pooling processing on the clustered data. The purpose of pooling is to reduce the dimension of the data while retaining the most important information.

[0220] Apply pooling operations to the data of each cluster, for example, max pooling selects the maximum value within each pooling window. Generate multimodal feature pooling data, reduce the dimension of the data, and highlight the main information of the feature.

[0221] Perform masked multi-head self-attention calculation on the multimodal feature data to obtain multimodal masked multi-head self-attention data;

[0222] The present invention provides an embodiment in which a masked multi-head self-attention layer is designed using the multi-head self-attention mechanism in the Transformer model. Multi-head self-attention can capture the relationship and dependency between different parts of the data. The multimodal feature data is linearly transformed to generate query (Q), key (K) and value (V) matrices. The attention weights are calculated and combined with the mask matrix. The outputs of multiple attention heads are spliced ​​and linearly transformed to generate the final multimodal masked multi-head self-attention data.

[0223] Feature weighted calculation is performed according to the multimodal masked multi-head self-attention data and the multimodal feature pooling data to obtain second multimodal feature weighted data.

[0224] In one embodiment of the present invention, the multimodal masked multi-head self-attention data and the pooled data are combined, and weighted calculation is performed using a weighted average or linear combination method. The masked multi-head self-attention data and the pooled data are linearly combined to obtain second multimodal feature weighted data.

[0225] In the present invention, through convolution calculation, deeper features can be extracted from multimodal feature data, local patterns and complex feature relationships in the data can be captured, and the expressiveness and prediction accuracy of the model can be improved. Clustering calculation is performed on the convolution layer data so that similar features can be effectively grouped, the accuracy of feature classification is improved, the influence of redundant information is reduced, and the model is helpful to better distinguish important features from minor features in feature weighted calculation. Through pooling calculation, the feature data can be reduced in dimension, the amount of data can be reduced while retaining key feature information, the computational efficiency of the model is improved, and the risk of overfitting is reduced, so that the model can still maintain efficient performance with fewer computing resources. Masked multi-head self-attention calculation can capture global relationships in feature data, especially the long-range dependencies between features, so that the model can comprehensively consider the complex interactions between features, thereby more accurately weighting the influence of each feature and improving the global perception ability of the model. Masking technology allows selective attention to specific features or feature combinations in self-attention calculations, ignoring irrelevant or noise features, making the calculation process more targeted, and can further improve the accuracy of feature weighted calculations. By combining multimodal masked multi-head self-attention data and multimodal feature pooling data for weighted calculation, the model can utilize deep features, local patterns and global relationships to achieve more accurate feature weighting, which can better reflect the actual role of each feature in the fermentation process and improve the final weighted data quality.

[0226] Optionally, S3 includes:

[0227] S31, obtaining yeast distribution data, wherein the yeast distribution data includes dough surface distribution data and dough internal distribution data;

[0228] In one embodiment of the present invention, imaging technology (such as high-resolution microscopy or CT scanning) is used to obtain yeast distribution data on the surface and inside of the dough. For surface data, the density and distribution uniformity of surface yeast are mainly concerned; for internal data, the focus is on collecting yeast concentration distribution at different depths.

[0229] The imaging data is processed, such as using edge detection and segmentation algorithms (such as Otsu threshold segmentation) to extract the yeast distribution from the image to form surface and internal yeast distribution data.

[0230] S32, performing surface yeast fermentation dynamics evolution according to dough surface distribution data, frozen dough component characteristic data and yeast dynamics basic model to obtain first frozen dough fermentation process data;

[0231] In one embodiment provided by the present invention, the yeast dynamics basic model constructed in S2 is used in combination with dough surface distribution data and frozen dough component characteristic data to simulate the fermentation dynamics evolution process of surface yeast.

[0232] The simulation process can use partial differential equations (PDE) to describe the growth, diffusion and consumption of yeast. Numerical solutions use numerical calculation techniques such as finite difference method (FDM) or finite element method (FEM).

[0233] Set the initial and boundary conditions such as surface temperature, humidity, and initial yeast concentration according to the kinetic model.

[0234] The time-stepping simulation is performed to calculate the concentration change of surface yeast and the amount of fermentation products generated, and the data of the first frozen dough fermentation process are obtained. The simulation output data includes the yeast concentration distribution at each time point and the cumulative amount of fermentation products.

[0235] S33, performing internal yeast fermentation dynamics evolution according to the dough internal distribution data, the frozen dough component characteristic data and the yeast dynamics basic model to obtain second frozen dough fermentation process data;

[0236] In one embodiment of the present invention, a similar dynamic model is used to simulate the yeast distribution inside the dough. Due to the complexity of the internal environment, factors such as the internal temperature gradient, diffusion of oxygen and nutrients need to be considered.

[0237] The internal dynamics model needs to combine the diffusion equation and the reaction equation and be numerically solved using the finite element method (FEM).

[0238] The initial conditions and boundary conditions inside the dough are set, and simulation calculations are performed to obtain the changes in yeast concentration and the generation of fermentation products at different locations inside the dough.

[0239] The obtained fermentation process data of the second frozen dough reflects the fermentation progress of each layer inside the dough and can be used to analyze the fermentation uniformity and quality inside the dough.

[0240] S34, generating fermentation process data from the first frozen dough fermentation process data and the second frozen dough fermentation process data to obtain the frozen dough fermentation process data.

[0241] The present invention provides an embodiment in which the surface fermentation process data and the internal fermentation process data are fused, and a weighted average method or other data fusion technology can be used. The weight can be allocated according to the fermentation contribution of the surface layer and the internal layer. Through the fusion algorithm, the surface and internal fermentation data are integrated to generate complete frozen dough fermentation process data. The data generated includes time series data, cumulative data of fermentation products, and the final yeast concentration distribution.

[0242] In the present invention, by respectively obtaining the yeast distribution data of the dough surface and the interior, and performing targeted fermentation dynamics evolution simulation, the method can comprehensively cover the fermentation process of the dough, accurately predict the fermentation behavior of the dough at different levels, thereby optimizing the fermentation conditions and improving the uniformity and quality of the product. The method processes the fermentation process of the dough surface and the interior separately, and performs independent evolution calculations based on the respective characteristic data and yeast dynamics model. The refined hierarchical control can better cope with the fermentation needs of different regions and prevent product defects caused by uneven fermentation, such as excessive expansion of the surface or insufficient internal fermentation. By using yeast distribution data and frozen dough component characteristic data to perform differentiated fermentation dynamics evolution on the surface and the interior, the fermentation parameters can be flexibly adjusted to adapt to different types of dough and production environments, ensuring that the fermentation process can be customized according to specific needs, thereby improving the flexibility and adaptability of production.

[0243] Optionally, the fermentation process data generation is specifically as follows:

[0244] Performing spatial matching processing on the first frozen dough fermentation process data and the second frozen dough fermentation process data to obtain fermentation process matching data;

[0245] In one embodiment of the present invention, a distance-based nearest neighbor matching algorithm (Nearest Neighbor Matching) or an ICP (Iterative Closest Point) algorithm is used to perform spatial matching processing on the first frozen dough fermentation process data and the second frozen dough fermentation process data.

[0246] The spatial coordinates of the two data sets were aligned to ensure that the surface and internal fermentation process data were compared in the same spatial coordinate system.

[0247] The distance between each point pair is calculated, and the point pair with the smallest distance is selected as the matching pair to form the fermentation process matching data. The matching data includes the coordinates of the matching point pair and the corresponding fermentation state.

[0248] The first frozen dough fermentation process data and the second frozen dough fermentation process data are screened according to the fermentation process matching data to obtain fermentation process spatial overlap data and fermentation process spatial non-overlap data;

[0249] In one embodiment provided by the present invention, the overlap degree of the matching data is calculated, and a similarity metric (such as cosine similarity or Euclidean distance) is used to determine the spatial consistency of the surface and internal fermentation data.

[0250] According to the overlap threshold, the matching data is divided into spatial overlap data and spatial non-overlap data. The overlap data are those point pairs whose similarity exceeds the threshold, and the non-overlap data are those point pairs whose similarity is lower than the threshold.

[0251] The screening results are stored as fermentation process spatial overlapping data and fermentation process spatial non-overlapping data for subsequent processing.

[0252] The secondary fermentation data of the fermentation process are obtained by calculating the secondary fermentation according to the spatial overlap number of the fermentation process;

[0253] In one embodiment of the present invention, a secondary fermentation kinetic model is constructed based on spatial overlap data. The model needs to consider the mutual influence between the surface and the interior, such as the secondary fermentation reaction triggered by the fermentation products on the surface penetrating into the interior.

[0254] The secondary fermentation process can be described using the reaction-diffusion equation or modified kinetic model. The reaction term and the diffusion term can be coupled to form a nonlinear PDE.

[0255] The fermentation status of the overlapping area is simulated again to simulate the possible fermentation deepening or product accumulation process.

[0256] The secondary fermentation process is solved by numerical methods (such as finite element method) to generate secondary fermentation data, which includes the fermentation state changes of each spatial point in the secondary fermentation process.

[0257] According to the secondary fermentation data of the fermentation process, the spatial non-overlapping data of the fermentation process are linearly weighted and screened for overlap to obtain the frozen dough fermentation process data.

[0258] In one embodiment provided by the present invention, weighted processing is performed on spatial non-overlapping data, and linear interpolation and weighted averaging are performed on the non-overlapping data according to the fermentation change trend in the secondary fermentation data. The results after linear weighting are screened for overlap to ensure that the final fermentation process data is consistent with the fermentation trend of the overlapped data. The screened data are integrated into the final frozen dough fermentation process data, which represents the global state of the entire dough during the fermentation process.

[0259] In the present invention, spatial matching processing is performed on the first and second frozen dough fermentation process data to ensure the spatial consistency of fermentation data from different levels, which helps to accurately reflect the overall changes of the entire dough during the fermentation process and improves the simulation accuracy of the model for the fermentation process. By screening the fermentation process data for spatial overlap and non-overlap, the fermentation behavior of different regions in the dough can be distinguished, the difference between the surface and internal fermentation processes can be revealed, and more accurate data support can be provided for subsequent fermentation control and optimization. Secondary fermentation calculation is performed on spatial overlap data, which can further optimize the fermentation process of these regions, help to make up for the deficiencies that may exist in the first fermentation process, ensure the uniformity and sufficiency of fermentation, and ultimately improve the fermentation quality of the dough. Linear weighting and overlap screening are performed on spatial non-overlap data, which can be used to perform more detailed control on these regions, optimize the details that are easily overlooked in the fermentation process, and ensure that the fermentation process of the entire dough is more balanced and consistent. Through secondary fermentation calculation and linear weighted processing, the fermentation process data finally generated is more accurate, the error caused by the unevenness in the fermentation process is reduced, the prediction accuracy and reliability of the model are enhanced, and the fermentation quality in the production process is ensured to reach the optimal level. By screening and weighting the overlap of different fermentation areas, producers can better control the key variables in the fermentation process, which helps to adjust the fermentation conditions in time in actual production, reduce fermentation unevenness, and ensure product consistency and high quality.

[0260] The objects of the present invention are:

[0261] 1. The present invention firstly uses a preset interface and intelligent sensors to accurately collect the composition and physical property data of frozen dough, and combines it with yeast activity data to form a multi-dimensional and comprehensive data foundation.

[0262] 2. Through in-depth analysis of yeast activity data, accurate basic parameters of fermentation dynamics are generated, and combined with the physical property data of frozen dough, a basic yeast dynamics model is constructed to accurately simulate the behavior of yeast during the fermentation process, thereby providing accurate dynamic predictions and ensuring the efficiency and stability of the fermentation process. The multimodal characteristics of the data are utilized to ensure that the model can capture the key characteristics of the dough in different dimensions, improving the prediction accuracy and adaptability of the model.

[0263] 3. The present invention simulates the evolution of yeast fermentation dynamics and generates detailed data on the frozen dough fermentation process. In this process, the fermentation behaviors on the surface and inside of the dough are processed and simulated separately, ensuring comprehensive coverage and fine control of the dough fermentation process. By separately processing overlapping and non-overlapping data, the present invention not only improves the accuracy and consistency of the fermentation process data, but also effectively reduces the deviations and uncertainties that may occur during the fermentation process, ensuring the uniformity and quality stability of the dough during the fermentation process.

[0264] 4. The present invention can timely discover and adjust possible abnormal conditions in the fermentation process through real-time evaluation of the data of the frozen dough fermentation process. Through the data-driven feedback mechanism, the system can dynamically adjust the fermentation parameters to ensure that the fermentation process is always carried out in the best state. Real-time monitoring and feedback control greatly improve the controllability and production efficiency of the fermentation process and reduce product defects caused by unstable fermentation.

[0265] Therefore, from any point of view, the embodiments should be regarded as illustrative and non-restrictive, and the scope of the present invention is limited by the attached application documents rather than the above description, and it is intended that all changes falling within the meaning and scope of equivalent elements of the application documents are included in the present invention.

[0266] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A yeast frozen dough fermentation process control method based on a fermentation kinetics model, characterized in that: The method comprises: S1, obtaining frozen dough data and yeast activity data; S2. constructing a yeast dynamics basic model according to the yeast activity data to obtain a yeast dynamics basic model; extracting dough component features according to the frozen dough data to obtain frozen dough component feature data; S3, performing yeast fermentation dynamics evolution according to the frozen dough component characteristic data and the yeast dynamics basic model to obtain the frozen dough fermentation process data; S4, performing fermentation process evaluation according to the frozen dough fermentation process data to obtain fermentation process evaluation data for performing yeast frozen dough fermentation process control auxiliary operations; S3 includes: Acquire yeast distribution data, wherein the yeast distribution data includes dough surface distribution data and dough internal distribution data; According to dough surface distribution data, frozen dough component characteristic data and yeast dynamics basic model, surface yeast fermentation dynamics evolution is performed to obtain the first frozen dough fermentation process data; According to the dough internal distribution data, the frozen dough component characteristic data and the yeast dynamics basic model, the internal yeast fermentation dynamics evolution is performed to obtain the second frozen dough fermentation process data; The first frozen dough fermentation process data and the second frozen dough fermentation process data are used to generate fermentation process data to obtain the frozen dough fermentation process data; The fermentation process data generation is specifically as follows: Performing spatial matching processing on the first frozen dough fermentation process data and the second frozen dough fermentation process data to obtain fermentation process matching data; The first frozen dough fermentation process data and the second frozen dough fermentation process data are screened according to the fermentation process matching data to obtain fermentation process spatial overlap data and fermentation process spatial non-overlap data; The secondary fermentation data of the fermentation process are obtained by calculating the secondary fermentation according to the spatial overlap number of the fermentation process; According to the secondary fermentation data of the fermentation process, the spatial non-overlapping data of the fermentation process are linearly weighted and screened for overlap to obtain the frozen dough fermentation process data.

2. The method according to claim 1, characterized in that S1 includes: Obtaining frozen dough ingredient data through a preset frozen dough ingredient input interface; The physical properties of frozen dough are collected by intelligent sensors to obtain the physical properties data of frozen dough; Integrate frozen dough ingredient data and frozen dough physical property data to obtain frozen dough data; Obtain yeast activity data.

3. The method according to claim 2, characterized in that S2 include: Generate basic parameters of yeast fermentation according to yeast activity data to obtain basic parameter data of yeast fermentation; The yeast dynamics basic model is constructed according to the yeast fermentation basic parameter data and the frozen dough physical property data to obtain the yeast dynamics basic model; Dough component features are extracted based on the frozen dough data to obtain frozen dough component feature data.

4. The method according to claim 3, characterized in that The dough component feature extraction is as follows: Performing multimodal data alignment according to frozen dough data to obtain multimodal alignment data; Performing multimodal feature extraction according to the multimodal alignment data to obtain multimodal feature data; Calculating feature weights according to the multimodal feature data to obtain multimodal feature weighted data; Multimodal feature fusion is performed based on multimodal feature weighted data to obtain frozen dough ingredient feature data.

5. The method according to claim 4, characterized in that The multimodal data alignment is specifically as follows: Perform three-dimensional modeling according to the frozen dough data to obtain a three-dimensional model of the frozen dough; Finite element calculation is performed according to the frozen dough three-dimensional model and the frozen dough data to obtain frozen dough finite element data; Extract the finite element data of frozen dough and obtain the multimodal data of spatial distribution of frozen dough; The spatial matrix of the multimodal data of spatial distribution of frozen dough is constructed to obtain multimodal alignment data; The data extraction is as follows: Performing spatial similarity screening according to the frozen dough finite element data to obtain the frozen dough spatial similarity screening data; Extract spatial distribution multimodal data from the frozen dough spatial similarity screening data to obtain the frozen dough spatial distribution multimodal data; The spatial similarity screening is as follows: Perform clustering processing on frozen dough finite element data to obtain finite element clustering data; According to the finite element clustering data, the frozen dough finite element data are divided into adjacent spaces to obtain finite element space division data; The similarity of the frozen dough finite element data is calculated and screened according to the finite element space partitioning data to obtain the frozen dough space similarity screening data.

6. The method according to claim 4, characterized in that Features The weight calculation is as follows: Calculating a first feature weight according to the multimodal feature data to obtain first multimodal feature weighted data; Performing a second feature weight calculation based on the multimodal feature data to obtain second multimodal feature weighted data, wherein the first feature weight calculation and the second feature weight calculation are different feature weight calculation methods; Feature fitting is performed on the first multimodal feature weighted data and the second multimodal feature weighted data to obtain multimodal feature weighted data.

7. The method according to claim 6, characterized in that The calculation of the first feature weight is as follows: ； is the first multimodal feature weighted data, For the Multimodal feature data, For the The signal-to-noise ratio data corresponding to the multimodal feature data, For the The spatial correlation data corresponding to the multimodal feature data, For the The temporal stability data corresponding to the multimodal feature data, is the natural exponential term, For the The confidence data corresponding to the multimodal feature data, For the The frequency domain feature density data corresponding to the multimodal feature data, For the The complexity measurement data corresponding to the multimodal feature data, For the The geometric distribution coefficient data corresponding to the multimodal feature data.

8. The method according to claim 6, characterized in that The second feature weight calculation is specifically as follows: Perform convolution calculation according to the multimodal feature data to obtain multimodal feature convolution layer data; Perform clustering calculation on the multi-modal feature convolution layer data to obtain multi-modal feature clustering data; Performing pooling calculation on the multimodal feature clustering data to obtain multimodal feature pooling data; Perform masked multi-head self-attention calculation on the multimodal feature data to obtain multimodal masked multi-head self-attention data; Feature weighted calculation is performed according to the multimodal masked multi-head self-attention data and the multimodal feature pooling data to obtain second multimodal feature weighted data.

Images