Hypoglycemic solid beverage formula generation method and device based on deep learning model

Abstract

The application provides a blood glucose-lowering solid beverage formula generation method and device based on a deep learning model. The method obtains user input data, which includes raw material information; searches for blood glucose-lowering solid beverage formula data corresponding to the raw material information from a blood glucose-lowering solid beverage formula database, and generates blood glucose-lowering solid beverage formula data corresponding to the raw material information through a pre-trained formula design model; the formula design model has been verified for blood glucose-lowering effect after training is completed; and in the case of finding, the found blood glucose-lowering solid beverage formula data and the blood glucose-lowering solid beverage formula data generated by the model are displayed. The technical solution provided by the application can meet specific needs of different users.

Description

TECHNICAL FIELD

[0001] The embodiment of the present application relates to the technical field of food science and artificial intelligence, and particularly relates to a blood glucose lowering solid beverage formula generation method and device based on a deep learning model. BACKGROUND

[0002] With the improvement of people's living standards and the enhancement of health consciousness, the demand for functional food and beverage is increasing. Especially for the special needs of chronic disease patients, such as diabetic patients, developing food and beverage with specific health efficacy has become a research hotspot. Diabetes is a common metabolic disease, and patients need to strictly control their diet to maintain stable blood glucose levels. Therefore, it is of great significance to develop a solid beverage formula that can effectively lower blood glucose.

[0003] Currently, the design method of blood glucose lowering solid beverage formula mainly relies on traditional trial-and-error method and experience accumulation. This method not only consumes time and effort, but also is difficult to ensure the effectiveness and safety of the blood glucose lowering solid beverage formula. In recent years, with the development of artificial intelligence and machine learning technology, some researches have begun to try to use these technologies to assist in the design of solid beverage formula. For example, rule-based expert systems and machine learning algorithms (such as decision trees, support vector machines, etc.) have improved the efficiency and accuracy of solid beverage formula design to some extent.

[0004] However, these methods still have some limitations. For example, diabetic patients may have other complications, and the physical conditions of different diabetic patients vary greatly, so their needs have highly personalized characteristics, and existing methods are difficult to meet the specific needs of different users. SUMMARY

[0005] The embodiment of the present application provides a blood glucose lowering solid beverage formula generation method based on a deep learning model, to solve the problem that existing methods are difficult to meet the specific needs of different users.

[0006] In a first aspect, the embodiment of the present application provides a blood glucose lowering solid beverage formula generation method based on a deep learning model, comprising:

[0007] Obtaining user input data, the user input data including raw material information;

[0008] Finding blood glucose lowering solid beverage formula data corresponding to the raw material information from a blood glucose lowering solid beverage formula database, and generating blood glucose lowering solid beverage formula data corresponding to the raw material information through a pre-trained formula design model; the formula design model has passed blood glucose lowering effect verification after training is completed;

[0009] In the case of finding, the found hypoglycemic solid beverage formula data and the model generated hypoglycemic solid beverage formula data are displayed.

[0010] Optionally, the training process of the formula design model comprises:

[0011] Obtaining plant active ingredient data determined to have hypoglycemic function and known hypoglycemic solid beverage formula data;

[0012] The plant active ingredient data and the solid beverage formula data are respectively pre-processed to form a functional raw material database and a hypoglycemic solid beverage formula database;

[0013] A formula design model is constructed using a deep learning technology, and the formula design model is trained based on the functional raw material database and the hypoglycemic solid beverage formula database until a preset stop training condition is met to obtain a pre-trained formula design model.

[0014] Optionally, the formula design model is trained based on the functional raw material database and the hypoglycemic solid beverage formula database until a preset stop training condition is met to obtain a pre-trained formula design model, comprising:

[0015] Determine the main task and the sub-task; the main task is to generate known hypoglycemic solid beverage formula data corresponding to the plant active ingredient data; the sub-task includes at least one of the following: determining the formula composition structure and the raw material proportion relationship in different known hypoglycemic solid beverage formula data, determining the formula mode corresponding to the raw material information in the plant active ingredient data, and determining the associated raw material information corresponding to the raw material information in the plant active ingredient data;

[0016] Based on the functional raw material database and the hypoglycemic solid beverage formula database, the formula design model is trained in a multi-task learning manner until a preset stop training condition is met to obtain a pre-trained formula design model.

[0017] Optionally, during the training process of the formula design model, the method further comprises:

[0018] Regularization techniques, including L1 regularization and L2 regularization, are introduced.

[0019] Optionally, during the training process of the formula design model, the method further comprises:

[0020] The hyperparameters such as learning rate, training rounds and batch size are initialized, the parameters of the formula design model are updated using a stochastic gradient descent optimizer, and the mean square error loss is taken as the loss function of the formula design model.

[0021] Optionally, during the training process of the formula design model, the method further comprises:

[0022] A data balancing technique is introduced to cope with the difference in data distribution among different tasks, which includes a data enhancement strategy and a specific weight adjustment strategy.

[0023] Optionally, the structure of the deep learning model comprises an input layer, a hidden layer and an output layer, wherein the input layer receives user input data and generates an input feature vector based on the user input data; the user input data comprises at least one input parameter; the hidden layer gradually processes the input feature vector through two fully connected layers to obtain an intermediate feature vector; and the output layer generates the hypoglycemic solid beverage formula data corresponding to the raw material information based on the intermediate feature vector.

[0024] Optionally, the two fully connected layers comprise a first fully connected layer and a second fully connected layer; the hidden layer comprises the first fully connected layer and the second fully connected layer connected in sequence; the first fully connected layer comprises a first linear transformation module and a first ReLU activation function connected in sequence, and the second fully connected layer comprises a second linear transformation module and a second ReLU activation function connected in sequence.

[0025] The hidden layer gradually processes the input feature vector through two fully connected layers, comprising:

[0026] The input feature vector is linearly transformed by the first linear transformation module to obtain a first linear feature vector;

[0027] The first linear feature vector is converted into a nonlinear feature vector by the first ReLU activation function;

[0028] The nonlinear feature vector is linearly transformed by the second linear transformation module to obtain a second linear feature vector;

[0029] The second linear feature vector is converted into an intermediate feature vector by the second ReLU activation function.

[0030] Optionally, the hidden layer comprises a residual connection module between the first fully connected layer and the second fully connected layer.

[0031] After the first linear feature vector is converted into a nonlinear feature vector by the first ReLU activation function, the method further comprises:

[0032] The input feature vector and the nonlinear feature vector are added to obtain a feature vector after residual connection.

[0033] In a second aspect, the embodiments of the present application provide a hypoglycemic solid beverage formula generation device based on a deep learning model, comprising:

[0034] The acquisition module is configured to acquire user input data, wherein the user input data comprises raw material information.

[0035] The search generation module is configured to search for hypoglycemic solid beverage formula data corresponding to the raw material information from a hypoglycemic solid beverage formula database, and generate hypoglycemic solid beverage formula data corresponding to the raw material information through a pre-trained formula design model.

[0036] The display module is configured to display the searched hypoglycemic solid beverage formula data and the model-generated hypoglycemic solid beverage formula data in the case of searching.

[0037] The embodiment of the present application provides a hypoglycemic solid beverage formula generation method based on a deep learning model, which comprises the following steps: acquiring user input data, wherein the user input data comprises raw material information; searching for hypoglycemic solid beverage formula data corresponding to the raw material information from a hypoglycemic solid beverage formula database, and generating hypoglycemic solid beverage formula data corresponding to the raw material information through a pre-trained formula design model; the formula design model has been verified for hypoglycemic effect at the end of training; and in the case of searching, the searched hypoglycemic solid beverage formula data and the model-generated hypoglycemic solid beverage formula data are displayed. In the embodiment of the present application, the user can input different raw material information according to his own needs, and in the case of searching, the searched hypoglycemic solid beverage formula data and the model-generated hypoglycemic solid beverage formula data are provided to the user, and in the case of not searching, the model-generated hypoglycemic solid beverage formula data is provided to the user. Therefore, the embodiment of the present application can generate hypoglycemic solid beverage formula data meeting the individual needs of different users through the formula design model.

[0038] These and other aspects of the present application will become more apparent from the following description. BRIEF DESCRIPTION OF DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without any creative effort.

[0040] Figure 1 A flowchart of a hypoglycemic solid beverage formula generation method based on a deep learning model provided by the embodiment of the present application;

[0041] Figure 2Another flowchart of a method for generating a hypoglycemic solid beverage formula based on a deep learning model is provided for the embodiments of the present application.

[0042] Figure 3 A flowchart of training a deep learning model is provided for the embodiments of the present application.

[0043] Figure 4 Another flowchart of a method for generating a hypoglycemic solid beverage formula based on a deep learning model is provided for the embodiments of the present application.

[0044] Figure 5 A schematic diagram of experimental results for verifying the hypoglycemic function is provided for the embodiments of the present application.

[0045] Figure 6 A structural schematic diagram of a device for generating a hypoglycemic solid beverage formula based on a deep learning model is provided for the embodiments of the present application.

[0046] Figure 7 A structural schematic diagram of a computing device is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0047] To better enable persons skilled in the art to understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings.

[0048] Diabetes is a metabolic disease characterized by high blood sugar. Long-term high blood sugar can lead to various complications, seriously affecting the quality of life of diabetic patients. Diabetes treatment and management require a combination of dietary control, drug treatment and lifestyle adjustment. Currently, there are some hypoglycemic drugs on the market, but long-term use may cause some side effects. Therefore, developing a natural food with hypoglycemic function has become an urgent need for people. In the management of diabetes, dietary control is one of the key links, and hypoglycemic solid beverage as a convenient food form has great market potential and practical application value.

[0049] However, the traditional design of hypoglycemic solid beverage formula mainly relies on experience and trial-and-error method, lacks systematicness and scientificity, and is difficult to meet the needs of diabetic patients for individualization and scientific verification.

[0050] In order to solve the problem that the traditional hypoglycemic solid beverage formula design cannot meet the personalized and scientific verification needs of diabetic patients, the embodiment of the present application provides a hypoglycemic solid beverage formula generation method based on a deep learning model. Different user input data correspond to different pre-trained formula design models. The embodiment of the present application can design a personalized hypoglycemic solid beverage formula according to the user input data (for example, raw material information). At the same time, the embodiment of the present application can also verify the hypoglycemic effect of the hypoglycemic solid beverage formula data designed by the formula design model, and then prove the practicability and effectiveness of the formula design model.

[0051] In some of the processes described in the description, claims, and schematic diagrams of the present application, a plurality of operations are included in a specific order, but it should be clear that these operations can be executed or in parallel without the order in which they appear in this text. The serial numbers of the operations, such as 11, 12, etc., are only used to distinguish different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes can include more or fewer operations, and these operations can be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this text are used to distinguish different messages, devices, modules, etc., and do not represent the order of precedence. Also, "first" and "second" are different types.

[0052] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0053] Figure 1 A flowchart of a hypoglycemic solid beverage formula generation method based on a deep learning model provided by the embodiment of the present application is shown in Figure 1 The method comprises:

[0054] S11, obtaining user input data, the user input data including raw material information.

[0055] The user input data can include raw material information. The above-mentioned raw material information can be raw material name, raw material number, etc. Raw material identification. The raw material information is used to reflect the user's demand.

[0056] S12, find the hypoglycemic solid beverage formula data corresponding to the raw material information from the hypoglycemic solid beverage formula database, and generate the hypoglycemic solid beverage formula data corresponding to the raw material information through the pre-trained formula design model. The formula design model has been verified by the hypoglycemic effect after training.

[0057] It should be understood that the hypoglycemic solid beverage formula database contains pre-processed hypoglycemic solid beverage formula data, and different hypoglycemic solid beverage formula data can correspond to the same or different raw material information. The hypoglycemic solid beverage formula can be referred to as a formula, which is a formula containing corresponding raw materials and having a hypoglycemic function. The hypoglycemic solid beverage formula data generated by the model is the specific data of the generated hypoglycemic solid beverage formula. The generated hypoglycemic solid beverage formula can be understood as a formula generated by the model, a new formula, etc.

[0058] Optionally, the hypoglycemic effect verification can be a zebrafish experiment verification. After the model training is completed, the embodiment of the present application can obtain experimental data, and then analyze the experimental data to obtain experimental results, and then obtain the result of whether the formula design model verification is passed.

[0059] S13, in the case of finding, display the found hypoglycemic solid beverage formula data and the hypoglycemic solid beverage formula data generated by the model; or, in the case of not finding, display the hypoglycemic solid beverage formula data generated by the model.

[0060] The embodiment of the present application provides a hypoglycemic solid beverage formula generation method based on a deep learning model. The method can improve the efficiency and accuracy of hypoglycemic solid beverage formula design through a deep learning model. The core is to use a deep learning algorithm, combine big data and food science knowledge, and design an effective hypoglycemic solid beverage formula for diabetic patients or people who want to control blood sugar. In the embodiment of the present application, the user can input different raw material information according to his own needs. The embodiment of the present application can provide the user with the found hypoglycemic solid beverage formula data and the hypoglycemic solid beverage formula data generated by the model in the case of finding, and provide the user with the hypoglycemic solid beverage formula data generated by the model in the case of not finding. Therefore, the embodiment can generate hypoglycemic solid beverage formula data that meets the individual needs of different users through the formula design model. At the same time, the embodiment of the present application can also verify the formula design model through a zebrafish experiment based on the hypoglycemic solid beverage formula data designed by the formula design model after the model training is completed, and then prove the hypoglycemic effect of the hypoglycemic solid beverage formula data, so as to prove the practicability and effectiveness of the formula design model.

[0061] The above process is the application process of the formula design model. For the formula design model, reasonable parameter design can improve the accuracy of the formula design model. Therefore, before executing the application process of the formula design model, the embodiment can design the training process of the formula design model. It should be understood that the training process of the formula design model mainly includes data collection, data processing and model construction. The embodiment takes an exemplary training process of the formula design model as an example for analysis as follows:

[0062] In some optional embodiments, the training process of the formula design model comprises:

[0063] Step 21, obtaining plant active ingredient data determined to have hypoglycemic function and known hypoglycemic solid beverage formula data.

[0064] It should be understood that obtaining can be understood as collecting, and function can be understood as efficacy. The plant active ingredient data determined to have hypoglycemic function can refer to the plant active ingredient data with hypoglycemic function that can be collected by consulting literature, network or patent documents; the known hypoglycemic solid beverage formula data can refer to the hypoglycemic solid beverage formula data that can be collected by consulting literature, network or patent documents.

[0065] Among them, the plant active ingredient data includes raw material information, active ingredient information in the raw material that plays a hypoglycemic function, literature, website or patent document providing raw material information and active ingredient information, and other data sources (including: title, author, publication time, patent number, etc.), other function information of the raw material (such as antioxidant, antitumor, immune regulation, etc.), etc. The hypoglycemic solid beverage formula data includes the raw materials in the formula, the addition amount of each raw material in the formula, other function information of the formula, literature, website or patent document providing the formula, and other data sources (including: title, author, publication time, patent number, etc.), etc.

[0066] Step 22, pre-processing the plant active ingredient data and the solid beverage formula data respectively to form a functional raw material database and a hypoglycemic solid beverage formula database.

[0067] It should be understood that pre-processing can include cleaning, integration, etc. Specifically, the present embodiment can clean the collected various data to remove abnormal values and noise data, and convert different types of data into a unified data format and range, so as to facilitate the training and application of the formula design model. Among them, data cleaning involves checking errors, inconsistencies and missing values in the data, and the purpose is to ensure that the database is accurate and complete, so as to facilitate subsequent analysis. For example, deleting duplicate records, correcting incorrect entries or filling in missing data. Abnormal values refer to data that are significantly different from other data in the database, which are removed in the present embodiment. Noise data refers to data containing random errors or inaccuracies, and removing these data helps to reduce interference in the model training process and improve the performance of the formula design model. Unified data format is a key step to ensure that the formula design model can correctly understand and process data. Through data preprocessing, the model training process can be effectively carried out, and finally a high-quality prediction result can be obtained.

[0068] It should also be understood that the integration of pre-processed data can form a complete database. Therefore, the present embodiment can establish a functional raw material database and a hypoglycemic solid beverage formula database through the integration step.

[0069] For example, the functional raw material database includes plant active ingredients with hypoglycemic function, including polysaccharides, polyphenols, alkaloids, flavonoids, saponins and triterpenes. These plant active ingredients are commonly found in various plant foods, and data sources can include scientific literature, patent libraries and experimental research, etc. The hypoglycemic solid beverage formula database includes the name, formula composition structure, raw material ratio relationship and other functions of the hypoglycemic solid beverage formula. The establishment of the functional raw material database and the hypoglycemic solid beverage formula database not only helps to organize and store a large amount of information, but also improves the efficiency of data retrieval and analysis.

[0070] Step 23, using deep learning technology to construct a formula design model, and training the formula design model based on the functional raw material database and the hypoglycemic solid beverage formula database until the preset stop training condition is met, to obtain a pre-trained formula design model. Wherein, the functional raw material database can be referred to as the raw material library, and the hypoglycemic solid beverage formula database can be understood as the formula library.

[0071] It should be understood that in step 23, the present embodiment trains the formula design model for the development target, which includes: formula composition structure and raw material ratio relationship, formula mode, synergy between raw materials in the formula, etc. In other words, in order to better consider the formula composition structure and raw material ratio relationship, formula mode, synergy between raw materials in the formula, the present embodiment can use deep learning technology to obtain the formula design model.

[0072] In the present embodiment, the formula design model is also called a deep learning model. The deep learning model uses the powerful data processing capability of neural networks to analyze and learn a large amount of plant active ingredient data and hypoglycemic solid beverage formula data, so as to reflect the complex relationship between the data and improve the accuracy of the formula design model.

[0073] It should also be understood that, as shown in Figure 2 In the training process of the formula design model, step 21 realizes data collection, step 22 realizes data processing and the establishment of the database (i.e. the raw material library and the formula library), and step 23 realizes model construction. In the formula design model (i.e. Figure 2After the deep learning model in the S10 is trained, the hypoglycemic effect of the solid beverage formula generated by the formula design model is verified through a zebrafish experiment, and then the practicability and effectiveness of the formula design model are verified. After verification, the formula design model is deployed on a server, and the application process of the formula design model is performed by using the server. In the application process of the formula design model, in the S11, the user demand is sent to the server where the deep learning model is deployed through the network, so that the server generates a formula based on the user demand through the deep learning model in the S13, that is, a formula containing corresponding raw materials and having a hypoglycemic function is generated by using the formula design model according to the user input data (including raw material information). Therefore, compared with the prior art, the embodiments of the present application have the following positive effects:

[0074] The embodiments of the present application use deep learning technology in machine learning to learn and analyze a large amount of plant active ingredient data and hypoglycemic solid beverage formula data, improve the effectiveness and accuracy of formula design, verify the practicability and effectiveness of the formula design model through a zebrafish experiment, and provide personalized hypoglycemic solid beverage formulas for users according to user demand, in other words, the formula design model can provide reliable hypoglycemic solid beverages for diabetic patients or people who want to control blood sugar.

[0075] It should be noted that the development target can be one or more, and the embodiments of the present application do not specifically limit this. Based on different development targets, the embodiments of the present application can train the formula design model in a multi-task learning manner to improve the adaptability of the formula design model in different scenarios.

[0076] In some optional embodiments, in the step 23, the formula design model is trained based on the functional raw material database and the hypoglycemic solid beverage formula database until a preset stop training condition is met, to obtain a pre-trained formula design model, including:

[0077] In the step 231, a main task and a sub-task are determined; the main task is to generate known hypoglycemic solid beverage formula data corresponding to the plant active ingredient data; the sub-task includes at least one of the following: determining the formula composition structure and the raw material proportion relationship in different known hypoglycemic solid beverage formula data, determining the formula mode corresponding to the raw material information in the plant active ingredient data, and determining the associated raw material information corresponding to the raw material information in the plant active ingredient data.

[0078] In the step 232, the formula design model is trained in a multi-task learning manner based on the functional raw material database and the hypoglycemic solid beverage formula database until a preset stop training condition is met, to obtain a pre-trained formula design model.

[0079] In the present embodiment, a recipe design model is trained in a multi-task learning manner. Compared with the prior art, in the model training process, the present embodiment can simultaneously process multiple related tasks through multi-task learning, specifically including: task one, analyzing the database of hypoglycemic solid beverage recipes to understand the recipe composition structure and raw material proportion relationship of different hypoglycemic solid beverage recipes; task two: determining common recipe patterns, for example, taking a certain main function as the core of the hypoglycemic raw material, and matching other raw materials (or auxiliary raw materials) to enhance the hypoglycemic effect or improve the taste; task three: studying the synergistic effect between raw materials in the hypoglycemic solid beverage recipe, for example, some raw material combinations can produce better hypoglycemic effect; task four: based on the analysis of raw materials and recipes, generate a hypoglycemic solid beverage recipe containing user input raw materials. Through sharing the underlying fully connected neural network structure, knowledge sharing can be achieved between different tasks, improving the learning efficiency and generalization ability of the recipe design model. In contrast, the traditional single-task model cannot fully utilize the correlation between tasks. In recipe design, plant active ingredient data and hypoglycemic solid beverage recipe data can be represented as fixed-length vectors, and the fully connected neural network structure can effectively learn the complex relationship between these vectors.

[0080] In some optional embodiments, during the training process of the recipe design model, there is a limitation of easy fitting, and the method further includes: introducing regularization techniques, including L1 regularization and L2 regularization. The present embodiment adopts the regularization technique to overcome the limitation of easy overfitting in the training process.

[0081] In some optional embodiments, during the training process of the recipe design model, the method further includes: initializing hyperparameters such as learning rate, training rounds and batch size, updating the parameters of the recipe design model using a stochastic gradient descent optimizer, and taking the mean square error loss as the loss function of the recipe design model.

[0082] The parameters of the formula design model can be referred to as model parameters. The hyperparameters such as learning rate lr, training rounds epochs, and batch size batch size are defined. The learning rate determines the step size of the model parameter update, the training rounds determine the number of times the deep learning model learns the data, and the batch size affects the amount of data processed each time. After creating a model instance model = SimpleModel(), the loss function criterion is defined as the mean square error loss, which is a loss function used for regression problems, which measures the average squared difference between the model output and the true label. In this embodiment, a stochastic gradient descent optimizer can be selected to update the model parameters. During the training process, the model parameters and the learning rate are passed as inputs to the stochastic gradient descent optimizer described above. The DataLoader class provided by the PyTorch framework is used to create a data loader train_loader, which divides the dataset into batches of size batch_size and sets the random data order to improve the training effect and generalization ability of the formula design model. The dataset is composed of a functional material database and a hypoglycemic solid beverage formula database. Each data in the dataset includes a preprocessed plant active ingredient data in the functional material database and a preprocessed hypoglycemic solid beverage formula data corresponding to the preprocessed plant active ingredient data in the hypoglycemic solid beverage formula database. Model evaluation needs to consider multiple task evaluation indicators, and more hyperparameter optimization is needed during adjustment, including hyperparameters specific to multi-task learning.

[0083] In some optional embodiments, in terms of training process, the embodiment can optimize the loss functions of multiple tasks at the same time, regularize using task correlation, and consider the data distribution difference between different tasks using data balancing techniques. Therefore, in the training process of the formula design model, the method further includes: introducing a data balancing technique to cope with the difference in data distribution between different tasks. The data balancing technique includes a data enhancement strategy and a specific weight adjustment strategy. The present application uses a more complex data enhancement strategy to cope with data quality and diversity requirements, and a specific weight adjustment strategy to balance the learning progress of different tasks.

[0084] In some optional embodiments, the structure of the deep learning model includes an input layer, a hidden layer, and an output layer. In the model application process, the input layer receives user input data and generates an input feature vector based on the user input data. The user input data includes at least one input parameter. The hidden layer processes the input feature vector step by step through two fully connected layers to obtain an intermediate feature vector. The output layer generates hypoglycemic solid beverage formula data corresponding to the raw material information based on the intermediate feature vector.

[0085] In some optional embodiments, the hidden layer includes a fully connected layer fc1, a ReLU activation function, and a fully connected layer fc2. During the model training process, the deep learning model can be trained under the PyTorch framework. First, define simpleModel as the underlying architecture for subsequent training and application, aiming to make regression predictions on given input data. During the processing of the deep learning model, it gradually processes the respective input data through two fully connected layers fc1 and fc2. First, the fully connected layer fc1 maps the specific attributes (i.e., input features) in the user input data to a first linear feature vector. The input features are selected and extracted to better represent the relationship between the input data and the model output. Second, the fully connected layer fc2 further maps the first linear feature vector to an output, thereby completing the entire data processing and prediction process. This hierarchical structure allows the deep learning model to gradually learn the complex mapping relationship from the input data to the output. Specifically, in the forward propagation function forward of the deep learning model, first, the user input data is linearly transformed through the fully connected layer fc1, then the ReLU activation function is applied to increase the non-linear representation ability of the deep learning model, and finally the final output result is obtained through the fully connected layer fc2.

[0086] During the model training process, the outer loop corresponds to the number of training rounds (epochs). The initial training rounds or epochs are initialized. At the beginning of each round, the cumulative loss is initialized to zero. The inner loop iterates through each batch of data in the data loader, as shown in Figure 3 For each batch, first, perform gradient zeroing on the stochastic gradient descent optimizer to prevent the gradient from the previous round from adversely affecting the current model parameter update. Then, perform forward propagation to pass the input data to the deep learning model to obtain the output result. Subsequently, the loss is calculated by backpropagation (calculating the loss value between the output of the deep learning model and the true label) to determine the gradient of the loss with respect to the model parameters. Then, the model parameters are updated using the stochastic gradient descent optimizer based on the calculated gradient. Finally, the loss of the current batch is accumulated, and training information including the current round number, average loss, and perplexity is periodically printed to observe the progress and stability of the training process.

[0087] In other words, throughout the model training process, the deep learning model strives to minimize the mean square error loss function. By continuously adjusting its parameters, the output of the deep learning model is as close as possible to the true label value. In each training round, after batch data is obtained from the data loader, gradient zero is performed to prevent gradient accumulation, then forward propagation is performed to calculate the loss by forward propagation, so that the input data passes through two fully connected layers and the ReLU activation function in the middle to obtain the output. Then calculate the loss, determine the gradient of the loss function with respect to the model parameters by backpropagation, and finally update the model parameters according to the gradient using the stochastic gradient descent optimizer. This process is repeated continuously, so that the deep learning model gradually learns the mapping relationship between the input data and the output label, and can accurately predict new user input data.

[0088] In some other optional embodiments, the two fully connected layers include a first fully connected layer and a second fully connected layer; the hidden layer includes the first fully connected layer and the second fully connected layer connected in sequence; the first fully connected layer includes a first linear transformation module and a first ReLU activation function connected in sequence, and the second fully connected layer includes a second linear transformation module and a second ReLU activation function connected in sequence. In combination with the structure of the hidden layer, the function of the hidden layer is refined. Specifically, in the model application process, the hidden layer gradually processes the input feature vector through the two fully connected layers, including:

[0089] Step a1, linearly transforming the input feature vector through the first linear transformation module to obtain a first linear feature vector. Step a2, converting the first linear feature vector into a nonlinear feature vector using the first ReLU activation function. Step a3, linearly transforming the nonlinear feature vector through the second linear transformation module to obtain a second linear feature vector. Step a4, converting the second linear feature vector into an intermediate feature vector using the second ReLU activation function.

[0090] Wherein, the first linear transformation module and the second linear transformation module are linear transformation modules with the same structure or different structures, and are located in different fully connected layers. Similarly, the first ReLU activation function and the second ReLU activation function correspond to ReLU activation functions with the same formula or different formulas, and are located in different fully connected layers.

[0091] This embodiment can automatically select and extract the most useful features for the task through step-by-step linear transformation and nonlinear conversion, reducing the risk of overfitting. At the same time, each fully connected layer is divided into a linear transformation module and a ReLU activation function, making the design of the deep learning model more modular. This modular design facilitates the expansion and optimization of the deep learning model, and more layers or adjustments to the structure of the layers can be added according to specific task requirements.

[0092] In some optional embodiments, the hidden layer can further include a residual connection module located between the first fully connected layer and the second fully connected layer; after step a2, the first linear feature vector is converted into a nonlinear feature vector using the first ReLU activation function, the method further includes:

[0093] Step a5, adding the input feature vector and the nonlinear feature vector to obtain a feature vector after residual connection.

[0094] Residual connection enables the deep learning model to utilize both the input feature vector and the nonlinear feature vector simultaneously. This feature fusion mechanism enhances the expressive power of the model, enabling the deep learning model to better capture complex patterns and relationships in the data. Moreover, the introduction of the residual connection module makes the design of the deep learning model more modular, allowing for the easy addition or removal of residual connections between different layers without significantly affecting the overall performance of the deep learning model. This modular design improves the scalability and flexibility of the deep learning model.

[0095] After the model training process is completed, the deep learning model can be deployed to a server. In the deployment code of the server, a model instance is first created, then the model parameters are saved based on the PyTorch framework, and the deep learning model is set to evaluation mode to ensure that no training operations such as gradient calculation are performed during inference. Further, the deep learning model can be deployed using Web service or containerization. Web service deployment creates a Web service using a Web framework such as Flask or Django. The deep learning model is loaded into the Web service, and user input data is received through a Hypertext Transfer Protocol (HTTP) request based on the network, and the result is returned after model inference. For example, when using Flask, the embodiment can create a route to receive POST requests, obtain input data from the POST requests, and convert the preprocessed user input data into a tensor format using functions provided by the PyTorch framework. Then the tensor format user input data is transmitted into the deep learning model to obtain the output, and finally the result is returned in JSON format. Containerization deployment refers to packaging the deep learning model and its dependencies into a container image using Docker. By creating a Docker file, determining the base image, installing dependencies, copying model files, and setting startup commands, the container image is built and the container is run to ensure the consistency and portability of the deep learning model in different environments. The embodiment can also use tools such as Postman or curl to send requests to the deployed service to verify whether the deep learning model can correctly respond and return accurate results. At the same time, the resource usage of the server is monitored to ensure the performance and stability of the deep learning model.

[0096] As Figure 4 shown, the embodiment provides still another method for generating a hypoglycemic solid beverage formula based on a deep learning model, which comprises obtaining user requirements, using a deep learning model to design a formula, obtaining a formula generated by the model, verifying the formula generated by the model, and taking the verified formula as a hypoglycemic solid beverage formula. Specifically, as Figure 4 shown, the model application process is exemplarily illustrated:

[0097] Step 1: Obtain user input data reflecting user requirements. The user input data can be the name of the raw material input by the user for formula design. In order to efficiently obtain the name of the raw material, the embodiment can develop a user-friendly interface or tool, which can integrate automatic completion and error checking functions.

[0098] Step 2: Send the obtained user input data to a target server deployed with a deep learning model. First, the user input data (such as the name of the raw material) is encapsulated, at which time the user input data is in the network interface layer. The user input data usually exists in binary form, but in order to be transmitted on the network, the user input data can be encapsulated into specific data packets (or data frames). Each data packet contains the data itself, the source address, the destination address, and some necessary control information. Subsequently, the data packet can enter the transport layer, which is responsible for establishing a communication connection between the source and the destination. The Transmission Control Protocol (TCP) ensures the integrity and sequential transmission of data packets to the target server through confirmation mechanism, retransmission mechanism and flow control. When the data packet enters the network layer, the network layer is responsible for allocating an Internet Protocol Address (IP) for the data packet, and according to the information in the routing table, the data packet is sent from the source network to the target network. In the network layer, the data packet can be split and recombined multiple times to adapt to the bandwidth and transmission capacity of different network links. The data link layer encapsulates the data packet into a frame and adds necessary control information such as the source Media Access Control (MAC) address and the target MAC address for transmission in a local area network or a wide area network. At the same time, the data link layer is also responsible for error detection and correction to ensure the accuracy of the data frame during transmission. When the data packet reaches the target server, the Network Interface Card (NIC) of the target server will perform a receiving operation and pass the data packet to the operating system of the target server for processing. In this embodiment, the source server and the target server can refer to a computer.

[0099] Step 3, generating a formula. Specifically, the present embodiment can process the raw material information contained in the received user input data through the target server, first search the existing information in the hypoglycemic solid beverage formula database, if the raw material information exists, call all the existing formula data of the formula in the hypoglycemic solid beverage formula database(including: the formula composition structure containing the corresponding raw material and the raw material proportion relationship, the literature or patent file of the formula and other information); secondly, through the deep learning model, taking the raw material information as the input, taking the combination mode of other raw materials and the proportion relationship of raw materials as the output, providing one or more model generated formulas. On the contrary, if there is no formula containing the raw material information in the hypoglycemic solid beverage formula database, the model generated formula is directly used. Finally, all the retrieved existing formulas(no display if there is none) and the model generated formula are recommended to the user.

[0100] For example, when the user inputs "corn silk powder", the model generated formula is made of the following components by mass fraction: corn silk powder 8 parts, astragalus powder 2 parts, barley sprout juice powder 10 parts, and burdock seed powder 10 parts.

[0101] Step 4, verification function. In order to evaluate the hypoglycemic function of the model generated formula involved in the present embodiment, and further determine the practicability of the deep learning model, the present embodiment carries out zebra fish experiment on the hypoglycemic function of the model generated formula, and the experimental conditions are as follows:

[0102] The corn silk powder, astragalus powder, barley sprout juice powder and burdock seed powder are weighed according to the raw material proportion relationship: 4:1:5:5, and then mixed uniformly to obtain a solid beverage.

[0103] The present example groups and doses zebrafish animals. For example, 100 adult zebrafish (4 months old, weight 0.5±0.1 g, AB strain wild type, male to female ratio 1:1) are from the same growing environment. The zebrafish are raised in a table water system, and the feeding environment is 14 h light and 10 h dark alternately, and the temperature is 27±2℃. These zebrafish are adapted in the feeding environment for 7 days, and the filtration and oxygenation are maintained, and the ordinary feed is fed once a day. In order to induce a type 2 diabetes mellitus (T2DM) model, the zebrafish are randomly divided into 5 groups, 20 in each group. Normal control group: blank control group (Normal Control, NC). Diabetes model group: model control group (Model Control, MC), low concentration solid beverage treatment group (Low Solid Beverage, LSB), medium concentration solid beverage treatment group (Medium Solid Beverage, MSB), and high concentration solid beverage treatment group (High Solid Beverage, HSB). The zebrafish in the diabetes model group are placed in a 111 mM glucose solution, soaked for 10 h a day, and fed with high-fat feed containing 10% cholesterol for 28 days. The blank control group is raised in water without glucose.

[0104] During the experiment, the zebrafish are fed once a day. Each fish is fed 20 mg of feed per day. After 28 days, the glucose soaking is stopped. Subsequently, all zebrafish are placed in a clean water environment for feeding, the NC group is fed with ordinary feed, and the MC group is fed with the same feed. The LSB, MSB, and HSB groups are added with 10%, 30%, and 50% solid beverage in the ordinary feed, respectively, to prepare mixed feed for feeding, and the process lasts for 28 days.

[0105] At the end of the experiment (e.g., on day 56), the zebrafish are fasted for 12 h, and then soaked in 2-4℃ cold water for 15 min. At the same time, when the samples are collected, the surface moisture of the zebrafish is wiped off with a paper towel, and the body length and weight of the zebrafish are measured, and the body mass index (BMI) of the zebrafish is calculated according to the formula: body weight (g) divided by the square of body length (cm 2 ).

[0106] The sample weight (g) was weighed, 9 times the volume of normal saline was added, the fish body tissue was treated with a high-speed handheld homogenizer for 1 min to ensure complete lysis of the tissue. The supernatant was collected after centrifugation at 2500 r / min for 10 min at 4°C for biochemical analysis, and the determination operation was performed according to the procedures and steps specified in the reagent kit accompanying instructions. According to the manufacturer's instructions, the biochemical kit was used to detect the glucose (Glucose, GLU), triglyceride (Triglycerides, TG), total protein (Total Protein, TP), total cholesterol (Total Cholesterol, T-CHO), high-density lipoprotein cholesterol (High-Density Lipoprotein Cholesterol, HDL-C) and low-density lipoprotein cholesterol (Low-Density Lipoprotein Cholesterol, LDL-C) levels in the whole body sample. In all experiments, P < 0.05 was considered significant. Among them, * represents P < 0.05, ** represents P < 0.01, *** represents P < 0.001, and **** represents P < 0.0001.

[0107] Figure 5 As shown, the BMI, GLU, TG, T-CHO, and LDL-C of the diabetes model group were significantly higher than those of the normal group, and the HDL-C was significantly lower than that of the normal group, which indicated that feeding with a 111 mM glucose solution combined with cholesterol feed could induce glucose and lipid metabolism disorders in adult zebrafish, causing them to develop diabetes. This phenomenon is similar to the detection results of blood lipids in clinical diabetes. After stopping the high-sugar and high-fat diet, BMI, GLU, TG, T-CHO, and LDL-C remained at a high level, and HDL-C remained at a low level, indicating that the zebrafish type 2 diabetes model constructed in the application was stable and effective. Based on this, the model formula intervention experiment was performed on the diseased zebrafish. Solid beverage intervention can reduce the degree of obesity, glucose level, triglyceride level, cholesterol level, and low-density lipoprotein level in zebrafish, and increase the high-density lipoprotein level, which indicates that solid beverage intervention can improve the symptoms of type 2 diabetes in zebrafish.

[0108] The above experimental data can indicate that the solid beverage formula obtained by the method described in the application has a significant hypoglycemic function.

[0109] The above hypoglycemic effect verification method can improve the reliability of the formula generated by the model.

[0110] Figure 6 A structure diagram of a hypoglycemic solid beverage formula generation device based on a deep learning model provided by the application is shown in Figure 6 As shown, the device comprises:

[0111] The acquisition module 61 is configured to acquire user input data, and the user input data includes raw material information.

[0112] The search and generation module 62 is configured to search for the hypoglycemic solid beverage formula data corresponding to the raw material information from a hypoglycemic solid beverage formula database, and generate the hypoglycemic solid beverage formula data corresponding to the raw material information through a pre-trained formula design model; the formula design model has been verified for a hypoglycemic effect after training.

[0113] The display module 63 is configured to display the searched hypoglycemic solid beverage formula data and the model-generated hypoglycemic solid beverage formula data in the case of searching.

[0114] Figure 6 The hypoglycemic solid beverage formula generation device based on the deep learning model can perform Figure 1 The implementation principle and technical effects of the hypoglycemic solid beverage formula generation method based on the deep learning model in the embodiments are not repeated. The specific operation manner of each module and unit of the hypoglycemic solid beverage formula generation device based on the deep learning model in the embodiments has been described in detail in the embodiments related to the method, and will not be described in detail here.

[0115] In one possible design, Figure 6 The hypoglycemic solid beverage formula generation device based on the deep learning model in the embodiments can be implemented as a computing device, such as a server. Figure 7 As shown, the computing device can include a storage component 71 and a processing component 72.

[0116] The storage component 71 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 72.

[0117] The processing component 72 is configured to acquire user input data, and the user input data includes raw material information; search for hypoglycemic solid beverage formula data corresponding to the raw material information from a hypoglycemic solid beverage formula database, and generate the hypoglycemic solid beverage formula data corresponding to the raw material information through a pre-trained formula design model; the formula design model has been verified for a hypoglycemic effect after training; and in the case of searching, display the searched hypoglycemic solid beverage formula data and the model-generated hypoglycemic solid beverage formula data.

[0118] The processing component 72 can include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component can also be one or more Application-Specific Integrated Circuits (ASICs), Digital Signal Processors (DSPs), Digital Signal Process Device (DSPD), Programmable Logic Devices (PLDs), Field Programmable Gate Arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components, for executing the above method.

[0119] The storage component 71 is configured to store various types of data to support the operation of the terminal. The storage component can be implemented by any type of volatile or non-volatile storage devices or their combinations, such as Random Access Memory (RAM), Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0120] Of course, the computing device can also include other components, such as an input / output interface, a display component, a communication component, etc.

[0121] The input / output interface provides an interface between the processing component and peripheral interface modules, which can be output devices, input devices, etc.

[0122] The communication component is configured to facilitate wired or wireless communication between the computing device and other devices, etc.

[0123] The computing device can be a physical device or an elastic computing host provided by a cloud computing platform, and the computing device can refer to a cloud server, and the processing component, the storage component, etc. can be basic server resources rented or purchased from the cloud computing platform.

[0124] The embodiment of the application further provides a computer storage medium, which stores a computer program, and the computer program can realize the above method when being executed by a computer. Figure 1 The embodiment of the application further provides a computer storage medium, which stores a computer program, and the computer program can realize the above method when being executed by a computer.

[0125] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described devices and modules can refer to the corresponding processes in the foregoing method embodiments, which will not be described here.

[0126] The device embodiments described above are only schematic, and the units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place, or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment scheme according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0127] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software and the necessary general hardware platform, and of course, can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes a plurality of instructions to make a computer device (which can be a personal computer, server, or network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

[0128] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the application, and not to limit them; although the application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the application.

Claims

1. A method for generating a hypoglycemic solid beverage formula based on a deep learning model, characterized in that, The method comprises the following steps: Obtaining user input data, which includes raw material information, such as raw material name or raw material number; Searching for hypoglycemic solid beverage formula data corresponding to the raw material information from a hypoglycemic solid beverage formula database, and generating hypoglycemic solid beverage formula data corresponding to the raw material information through a pre-trained formula design model; The formula design model has been verified for hypoglycemic effect after training, and the verification is realized through zebrafish experiments. The hypoglycemic solid beverage formula database contains pre-processed hypoglycemic solid beverage formula data, and different hypoglycemic solid beverage formula data correspond to the same or different raw material information; In the case of finding, display the found hypoglycemic solid beverage formula data and the model generated hypoglycemic solid beverage formula data; The training process of the formula design model comprises the following steps: Obtaining plant active ingredient data that has been determined to have hypoglycemic function and known hypoglycemic solid beverage formula data, wherein the plant active ingredient data includes raw material information, active ingredient information in the raw material that plays a hypoglycemic function, and data sources providing raw material information and active ingredient information; Preprocessing the plant active ingredient data and the solid beverage formula data respectively to form a functional raw material database and a hypoglycemic solid beverage formula database; Using deep learning technology to construct a formula design model; Determining the main task and the sub-task; the main task is to generate known hypoglycemic solid beverage formula data corresponding to the plant active ingredient data; the sub-task includes at least one of the following: determining the formula composition structure and the raw material proportion relationship in different known hypoglycemic solid beverage formula data, determining the formula mode corresponding to the raw material information in the plant active ingredient data, and determining the associated raw material information corresponding to the raw material information in the plant active ingredient data; Based on the functional raw material database and the hypoglycemic solid beverage formula database, the formula design model is trained in a multi-task learning manner. In the model training process, the knowledge sharing between different tasks is realized by sharing the bottom full connection neural network structure, until the preset stopping training condition is met, to obtain the pre-trained formula design model.

2. The method of claim 1, wherein, During the training process of the formula design model, the method further comprises the following steps: Introducing regularization techniques, including L1 regularization and L2 regularization.

3. The method of claim 1, wherein, During the training process of the formula design model, the method further comprises the following steps: Initializing hyperparameters such as learning rate, training rounds and batch size, using a stochastic gradient descent optimizer to update the parameters of the formula design model, and taking mean square error loss as the loss function of the formula design model.

4. The method of claim 1, wherein, During the training process of the formula design model, the method further comprises the following steps: Introducing data balancing techniques to cope with the difference in data distribution between different tasks, and the data balancing techniques include data enhancement strategies.

5. The method of claim 1, wherein, The structure of the deep learning model comprises an input layer, a hidden layer and an output layer, wherein the input layer receives user input data and generates an input feature vector based on the user input data; the user input data comprises at least one input parameter; the hidden layer gradually processes the input feature vector through two fully connected layers to obtain an intermediate feature vector; and the output layer generates hypoglycemic solid beverage formula data corresponding to the raw material information based on the intermediate feature vector.

6. The method of claim 5, wherein, The two fully connected layers comprise a first fully connected layer and a second fully connected layer; the hidden layer comprises the first fully connected layer and the second fully connected layer connected in sequence; the first fully connected layer comprises a first linear transformation module and a first ReLU activation function connected in sequence, and the second fully connected layer comprises a second linear transformation module and a second ReLU activation function connected in sequence; The hidden layer gradually processes the input feature vector through the two fully connected layers, comprising: linearly transforming the input feature vector through the first linear transformation module to obtain a first linear feature vector; converting the first linear feature vector into a nonlinear feature vector by using the first ReLU activation function; linearly transforming the nonlinear feature vector through the second linear transformation module to obtain a second linear feature vector; converting the second linear feature vector into an intermediate feature vector by using the second ReLU activation function.

7. The method of claim 6, wherein, The hidden layer comprises a residual connection module located between the first fully connected layer and the second fully connected layer; After converting the first linear feature vector into a nonlinear feature vector by using the first ReLU activation function, the method further comprises: adding the input feature vector and the nonlinear feature vector to obtain a feature vector after residual connection.

8. A device for generating a hypoglycemic solid beverage formula based on a deep learning model, characterized in that, comprising: an acquisition module configured to acquire user input data, wherein the user input data comprises raw material information, and the raw material information is a raw material name or a raw material number; a searching and generating module configured to search for hypoglycemic solid beverage formula data corresponding to the raw material information from a hypoglycemic solid beverage formula database, and generate hypoglycemic solid beverage formula data corresponding to the raw material information by using a pre-trained formula design model; the formula design model has passed a hypoglycemic effect verification at the end of training, the verification is realized through a zebrafish experiment, the hypoglycemic solid beverage formula database contains preprocessed hypoglycemic solid beverage formula data, and different hypoglycemic solid beverage formula data correspond to the same or different raw material information; a display module configured to display the searched hypoglycemic solid beverage formula data and the hypoglycemic solid beverage formula data generated by the model in the case of searching; the training process of the formula design model comprises: acquiring plant active ingredient data determined to have a hypoglycemic function and known hypoglycemic solid beverage formula data, wherein the plant active ingredient data comprises raw material information, active ingredient information in the raw material that plays a hypoglycemic function, and a data source providing the raw material information and the active ingredient information; preprocessing the plant active ingredient data and the solid beverage formula data respectively to form a functional raw material database and a hypoglycemic solid beverage formula database; A recipe design model is constructed using a deep learning technology; A main task and a sub-task are determined; the main task is to generate known hypoglycemic solid beverage recipe data corresponding to plant active ingredient data; the sub-task includes at least one of the following: determining the recipe composition structure and the proportion relationship of raw materials in different known hypoglycemic solid beverage recipe data, determining a recipe mode corresponding to raw material information in the plant active ingredient data, and determining associated raw material information corresponding to the raw material information in the plant active ingredient data; Based on the functional raw material database and the hypoglycemic solid beverage recipe database, a recipe design model is trained in a multi-task learning manner; in the model training process, knowledge sharing between different tasks is achieved by sharing a bottom-layer fully connected neural network structure, until a preset stop training condition is met, to obtain a pre-trained recipe design model.

Images