Obese children recipe generation system based on deep learning neural network

Abstract

This application relates to the field of information processing technology, specifically to a recipe generation system for obese children based on deep learning neural networks. It includes a server and terminal devices connected to the server. The server includes: a cloud platform connected to the retail interfaces of various fresh food markets to store information on ingredients currently being sold in each market; and a cloud processor that receives location information from the terminal devices, extracts information on ingredients currently being sold in fresh food markets near the terminal devices, and sends this information to the terminal devices. This invention uses the cloud platform to acquire and integrate information on ingredients for sale in nearby fresh food markets in real time, combines this with precise positioning to build a localized ingredient database, and utilizes a deep learning model to comprehensively analyze individual physical characteristics data (such as BIM) and user recipe preferences, dynamically generating highly personalized recommended recipes.

Description

Technical Field

[0001] This application relates to the field of information processing technology, and more specifically, to a recipe generation system for obese children based on deep learning neural networks. Background Technology

[0002] Childhood obesity rates continue to rise, becoming a serious public health problem. Obesity poses multiple threats to children's physical and mental health, and can induce problems such as insulin resistance, fatty liver, hypertension, abnormal bone development, and low self-esteem.

[0003] To address childhood obesity, the core of existing intervention strategies lies in dietary management through the generation of specific meal plans. This approach aims to scientifically design meal structures, control total calorie intake, optimize the proportions of macronutrients (carbohydrates, protein, and fat), and ensure an adequate supply of micronutrients such as vitamins and minerals. This, in turn, creates a negative energy balance while meeting the basic needs for children's growth and development, thereby achieving the goals of weight control and body fat reduction. Nutritionists or related health management software typically use this principle to customize personalized or standardized dietary plans for obese children.

[0004] Current mainstream recipe generation methods are overly formulaic. These methods primarily (or solely) rely on children's BMI (Body Mass Index) or basal metabolic rate for calculations, mechanically applying nutritional models to output calorie and nutrient ratio requirements. For parents, this model often results in recipes containing specific, expensive, or hard-to-obtain ingredients, with frequent changes and complex cooking processes, leading to high ingredient procurement costs, time-consuming and labor-intensive processes, and extreme difficulty in implementation. For children, such recipes are not designed according to their own tastes, making it difficult to stimulate their willingness to eat and even causing resistance.

[0005] In conclusion, there is currently a lack of a method for generating children's recipes that can be applied to family life. Summary of the Invention

[0006] The summary section of this application is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description section below. This summary section is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0007] Some embodiments of this application propose a recipe generation system for obese children based on deep learning neural networks to address the technical problems mentioned in the background section above.

[0008] As a first aspect of this application, some embodiments of this application provide a recipe generation system for obese children based on deep learning neural networks, comprising: a server and a terminal device connected to the server, wherein:

[0009] The servers include:

[0010] The cloud platform connects to the retail interfaces of various fresh food markets to store information on the ingredients being sold in each market.

[0011] The cloud processor receives the location information of the terminal device, extracts information on the ingredients being sold in the fresh food market in the vicinity of the terminal device, and sends it to the terminal device.

[0012] Terminal equipment includes:

[0013] A locator is used to obtain current location information;

[0014] Data storage is used to store food information sent by the cloud processor and generate a food database;

[0015] The data collector is used to collect users' recipe information, which includes the number of dishes and the vital signs data of each child.

[0016] Instruction memory, which stores one or more instructions;

[0017] The terminal processor is configured to execute one or more instructions stored in the instruction memory, input recipe information and vital sign data into the recipe generation model to select the required ingredients from the ingredient database, and generate a recommended recipe.

[0018] The recipe generation model generates recommended recipes based on the target children's preferred tastes and target BMI.

[0019] This invention provides a deep learning neural network-based recipe generation system for obese children. It acquires and integrates real-time information on available ingredients from nearby fresh food markets via a cloud platform, combines this with precise location data to build a localized ingredient database, and utilizes a deep learning model to comprehensively analyze individual physical characteristics (such as BMI) and user dietary preferences to dynamically generate highly personalized recommended recipes. This solution effectively overcomes the core shortcomings of existing technologies, such as overly formulaic recipe generation, reliance on BMI leading to difficulties in ingredient procurement (high difficulty for parents to implement), and insufficient appeal of dishes (low compliance among children). It achieves localized recipe adaptation (reducing the difficulty and cost of procurement for parents), personalized customization (enhancing the appeal of dishes to children's taste and visual appeal), and intelligent generation (beyond simple formula calculations), thereby significantly improving the practical feasibility of dietary intervention programs for obese children and children's active acceptance.

[0020] Furthermore, the terminal equipment also includes:

[0021] The input device allows users to modify the ingredient database, adding or deleting ingredients.

[0022] This invention allows users to make real-time adjustments to the localized food database (adding new available ingredients or deleting disliked or overpriced ingredients), enabling recipes to dynamically adapt to users' actual purchasing power (adding readily available and affordable ingredients), accurately avoid ingredients that children dislike (deleting disliked options), and proactively control food costs (deleting expensive ingredients).

[0023] Furthermore, the food information includes the food name and the unit price of the food;

[0024] Recipe generation models include:

[0025] The target generator generates target parameters and parameter change information based on the input vital sign data. The target parameters include target BMI, target weight, and target body fat, and the parameter change information includes the rate of change of BMI, the rate of change of weight, and the rate of change of body fat.

[0026] An energy training network generates nutritional information based on input vital sign data and a pre-set target BMI.

[0027] The recipe generation network generates recommended recipes based on nutritional information, children's preferred tastes, and ingredient information.

[0028] In the technical solution provided in this application, nutritional information and dish selection are processed separately when generating recommended recipes. This ensures accurate control over the nutritional content ingested by the user while also allowing for fine-grained control over the introduced taste variables. It avoids the gradient explosion problem caused by conflating taste and nutrition in a single model.

[0029] Furthermore, the energy training network includes:

[0030] The encoding network is used to transform the input vital sign data into feature labels X, and obtain all feature labels X to get the input data sequence TX;

[0031] A positive feedback neural network outputs nutrient content information Y based on the input feature label X, so as to minimize the error between the feature label X and the parameter change information;

[0032] X = {x1, x2, x3}

[0033] x1 = current BMI - previous BMI, x2 = current weight - previous weight, x3 represents current body fat - previous body fat;

[0034] Y = {y1, y2, y3, y4}, where y1 represents the carbohydrate intake required for the day, y2 represents the protein intake required for the day, and y3 represents the fat intake required for the day.

[0035] In this approach, by making the feature labels as close as possible to the parameter change information, the weight loss process of children can be controlled, avoiding rapid weight loss through malnutrition, which could negatively impact children's normal development.

[0036] Furthermore, positive feedback neural networks include:

[0037] Feedback separator, used to separate parameter variation signals from a data sequence TX. ;

[0038] ;

[0039] in, This represents the parameter change signal in the parameter change sequence. Let represent the compensation signal for the t-th feature label in the data sequence TX, where t represents the index of the feature label. Indicates the feedback intensity coefficient. Indicates the lag step size;

[0040] The parameter changer, with a built-in neural network structure, generates nutrient content information Y based on the input parameter change signal. t ;

[0041] ;

[0042] Represents the neural network structure. Y represents the model parameters of the neural network structure. t This represents the nutrient content information generated at the t-th time step;

[0043] The hysteresis feedback calculator calculates the feedback signal based on the output nutrient content information Yt. ;

[0044] ; This represents the feedback gain coefficient.

[0045] A feedback compensator generates a compensation signal for the next time step based on the feedback signal.

[0046] ;in, This represents the exponential moving average function. This represents the feedback strength coefficient, which is shared by the feedback compensator and the feedback separator.

[0047] The technical solution provided in this application adds feedback calculation to the input and output parts of the neural network model. This feedback calculation is primarily based on feedback signals generated by a hysteresis feedback calculator, which are then added to the original signal. This allows the hysteresis relationship between ingested substances and a child's weight to be considered, resulting in a more accurate training effect.

[0048] Furthermore, the loss function of a positive feedback neural network for:

[0049] ;

[0050] This represents all learnable parameters of a positive feedback neural network, where t represents the time step index. This represents the actual change in feature labels. Indicates time interval, Used to represent regularization parameters , Indicates information about parameter changes;

[0051] This represents the tracking error term, which guides the feedback compensator to generate a compensation signal that can follow the actual data sequence TX. Represents the rate of change;

[0052] This indicates a safe option for weight loss.

[0053] In the solution provided in this application, the tracking error term ensures that the feedback adjustment mechanism can accurately respond to real-time data changes during training, improving the immediacy and accuracy of compensation. The rate of change term can improve error accumulation in the system feedback loop. The weight loss safety feature guides the weight loss process to move as close as possible to the preset parameter changes, increasing the safety of weight loss.

[0054] Furthermore, the recipe generation network includes:

[0055] The encoder is used to compile the food information in the food database into food data W, W={W1, W2, W3, W4}; W1, W2, W3 represent the carbohydrate content, fat content and protein content per unit weight, respectively, and W4 represents the retail price of the food per unit weight.

[0056] The ingredient information filter filters ingredient data based on the input recipe budget and quantity, resulting in an ingredient selection library R; it then selects ingredients with a retail price per unit weight less than... Food data is entered into the food selection database;

[0057] The recipe generation module generates recommended recipes from the ingredient selection library R so that the nutritional content of the recommended recipes for children is close to the nutritional information.

[0058] By introducing consumption budget data in this solution, many irrelevant dishes can be eliminated from the menu selection, thereby further reducing the workload of subsequent model training.

[0059] Furthermore, .

[0060] Where g1 is the maximum consumption amount entered by the user, a1 is the number of dishes entered by the user, a2 is the pre-set weight of ingredients consumed for each dish, and a2 is a pre-set fixed value. min The retail price of the cheapest ingredient among all ingredient data.

[0061] Furthermore, the recipe generation module includes:

[0062] The generator produces preliminary recommended recipes based on nutritional information and the number of dishes.

[0063] The first discriminator determines whether the preliminary recipe fits the budget. If it does, the preliminary recommended recipe is sent to the second discriminator.

[0064] The second discriminator determines whether the nutritional content of the amount of food consumed by the child in the preliminary recommended diet is close to the nutritional content information. If it is close to the nutritional content information, the preliminary recommended diet is output as the recommended diet.

[0065] Furthermore, the generator and the second discriminator form an adversarial network, with the first discriminator acting as a random activator for the generator to increase the richness of the recommended recipes.

[0066] In the technical solution provided in this application, the first discriminator, acting as a standalone discriminator, increases the randomness of recommended recipe generation, guiding the generator to produce more recipes. Because there is no adversarial relationship between the first discriminator and the generator, information unrelated to weight loss, such as monetary value, will not affect the generator. Correspondingly, the adversarial network formed by the second discriminator and the generator allows the generator to increasingly learn the user's actual taste preferences in the recommended recipes.

[0067] The beneficial effects of the technical solution in this application are as follows: The technical solution provided in this application, when generating recommended recipes, differs from the traditional method of crawling standard recipes from a database. Instead, it divides the cooking process into ingredient selection and cooking method. This ingredient selection is more in line with the actual situation of family meals. For example, Dongpo pork obtained using traditional methods has a complex preparation process, making it unsuitable for daily family use. The recipe obtained using this solution, however, consists of braised pork belly, green beans, and potatoes. Although cooking pork belly, green beans, and potatoes together as a braised dish is not a standard recipe, it is very practical for families; simply follow the steps for braising the green beans, potatoes, and pork belly together. Thus, the solution provided in this application can more accurately focus on how to reduce users' obesity rates. Attached Figure Description

[0068] Figure 1 This is a schematic diagram of a recipe generation system for obese children based on deep learning neural networks.

[0069] Figure 2 This is a schematic diagram of the terminal device.

[0070] Figure 3 A schematic diagram of the structure of the recipe generation model.

[0071] Figure 4 This is a schematic diagram of a positive feedback neural network.

[0072] Figure 5 This is a structural diagram of the recipe generation module. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments. The same reference numerals in the accompanying drawings represent the same components. It should be noted that the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the described embodiments of this application without creative effort are within the scope of protection of this application.

[0074] Compared to the embodiments shown in the accompanying drawings, feasible embodiments within the scope of this application may have fewer components, other components not shown in the drawings, different components, differently arranged components, or components with different connections, etc. Furthermore, two or more components in the drawings may be implemented in a single component, or a single component shown in the drawings may be implemented as multiple separate components.

[0075] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” and similar terms used in this specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not necessarily indicate a quantity limitation. Terms such as “upper” and “lower” are used only to indicate relative positional relationships, and these relative positional relationships may change accordingly when the absolute position of the described object changes.

[0076] refer to Figure 1 Example 1: A recipe generation system for obese children based on deep learning neural networks includes a server and terminal devices. The server provides data storage, large-scale data processing, and external information access functions. User terminal devices include mobile phones, desktop computers, laptops, etc., supporting multiple users to interact with the server through their respective terminals.

[0077] The server comprises a cloud platform and a cloud processor. The cloud platform provides information access interfaces and stores data, connecting to the retail interfaces of various fresh food markets. Retailers upload their location and real-time information on available ingredients through this interface. The cloud platform aggregates and stores real-time data on available ingredients from each market.

[0078] The cloud processor receives the location information of the terminal device, extracts real-time information on available ingredients from nearby fresh food markets, and sends it to the terminal device. For example, the terminal device sends its location coordinates (point A) to the cloud processor, which then extracts information on available ingredients from various fresh food markets within a 500-meter radius of point A (nearby area) and sends it back to the terminal device. Thus, the terminal device can obtain information on nearby available ingredients based on its location.

[0079] refer to Figure 2 The terminal equipment includes: a locator, a data storage device, an input device, a data acquisition device, an instruction storage device, and a terminal processor. The locator is used to obtain current location information. The locator is a positioning module that can determine the current geographical location through satellite positioning or the operator's communication network.

[0080] The data storage device stores the ingredient information issued by the server and saves the ingredient information to generate an ingredient database. The ingredient database records the names of the ingredients that can currently be purchased, as well as their corresponding unit prices.

[0081] The input device allows users to modify the ingredient database, adding or deleting new ingredients. For example, if a user can buy carrots or fish at a nearby market, they can add them to the ingredient database using the input device, or delete unwanted ingredients.

[0082] The data collector is used to collect users' recipe information, which includes the number of dishes and children's vital signs data.

[0083] In this solution, the user refers to the operator of the terminal device, corresponding to the parents within the family. The child is the individual needing to lose weight and is the target for generating the recommended meal plan. The core of this solution is that it doesn't generate a separate meal plan for the child, but rather allows parents and children to share the same plan. By adjusting the meal plan to suit the child's taste preferences, it guides the child to eat more or less of specific foods.

[0084] For example, if a user plans to prepare four dishes, they can enter the four dishes into the recipe information. Children's physical characteristics data include weight, body fat percentage, height, age, and BMI. These characteristics need to be measured by the user using tools such as a body fat scale and ruler. In practice, the terminal device, body fat scale, and electronic height measurement device can be connected within a local area network, with the data collector directly collecting the required physical characteristics data from the body fat scale and electronic height measurement device.

[0085] The instruction storage unit stores one or more instructions; the terminal processor is configured to execute these instructions to implement the following process: inputting recipe information and vital sign data collected by the data acquisition unit into the recipe generation model. This model generates recommended recipes based on user-defined taste preferences and target BMI, combined with the ingredient database in the data storage unit.

[0086] Example 2: Example 1 provided a system capable of generating recipes based on nearby ingredients. This system can combine the user's surrounding market conditions to provide recommended recipes. Example 2, based on Example 1, provides a specific method for generating recommended recipes. The recommended recipes are mainly provided by a recipe generation model. Specifically:

[0087] refer to Figure 3 The recipe generation model includes a target generator, an energy training network, and a recipe generation network.

[0088] The target generator generates target parameters and parameter change information based on the input vital sign data. Target parameters include target BMI, target weight, and target body fat percentage, while parameter change information includes the rate of change in BMI, the rate of change in weight, and the rate of change in body fat percentage.

[0089] The target parameters are standard BMI, weight, and body fat for children. Adjusting children's vital signs data to these target parameters is the core objective of this system.

[0090] The parameter change information represents the rate of decrease for each vital sign data point each day. The parameter changes should be as gradual as possible to avoid adverse health effects caused by dieting or reducing nutrient intake.

[0091] The target generator, based on current medical standards, sets a standard target BMI, target weight, and target body fat for each child. The generator can retrieve these targets using a built-in data table, based on the child's height, age, and gender. Parameter changes can be configured according to the child's age, gender, height, and other information.

[0092] The energy training network generates nutrient content information based on input vital signs data and target BMI, target weight, and target body fat.

[0093] Nutritional information includes carbohydrates, protein, and fat for the day. Vitamins, minerals, and water, when adequately supplied, are largely unrelated to weight changes. These micronutrients can be easily obtained through supplements, and adjustments to these micronutrient levels are not correlated with weight correction; therefore, they are not included in the generation of recommended meal plans.

[0094] Nutritional content information Y, Y={y1, y2, y3}, where y1 represents the carbohydrate intake required for the day, y2 represents the protein intake required for the day, and y3 represents the fat intake required for the day.

[0095] The energy training network includes:

[0096] The encoding network is used to transform the input vital sign data into feature labels X, and obtain all feature labels X to get the input data sequence TX;

[0097] X = {x1, x2, x3}, where x1 = current BMI - previous BMI, x2 = current weight - previous weight, and x3 represents current body fat - previous body fat.

[0098] The feature label represents the weight loss progress of the object. The nutritional content information is continuously corrected through the feature label X until the feature label X is as close as possible to the parameter change information.

[0099] Encoding networks are used to transform redundant information sequences into input data sequences. By comparing the input data sequences, the dimensionality of the input data sequences can be effectively reduced, thus reducing the complexity of the model.

[0100] refer to Figure 4 A positive feedback neural network outputs nutrient content information Y based on the input feature label X, so as to minimize the error between the feature label X and the parameter change information.

[0101] Positive feedback neural networks include:

[0102] The feedback separator is used to separate the parameter change signal from the data sequence TX. The feature label represents the weight loss progress of the object. The nutritional content information is continuously corrected through the feature label X until the feature label X is as close as possible to the parameter change information.

[0103] Encoding networks are used to transform redundant information sequences into input data sequences. By comparing the input data sequences, the dimensionality of the input data sequences can be effectively reduced, thus reducing the complexity of the model.

[0104] refer to Figure 4 A positive feedback neural network outputs nutrient content information Y based on the input feature label X, so that the feature label X is as close as possible to the parameter change information.

[0105] Positive feedback neural networks include:

[0106] Feedback separator, used to separate parameter variation signals from a data sequence TX.

[0107] The feature label represents the weight loss progress of the object. The nutritional content information is continuously corrected through the feature label X until the feature label X is as close as possible to the parameter change information.

[0108] Encoding networks are used to transform redundant information sequences into input data sequences. By comparing the input data sequences, the dimensionality of the input data sequences can be effectively reduced, thus reducing the complexity of the model.

[0109] refer to Figure 4 A positive feedback neural network outputs nutrient content information Y based on the input feature label X, so that the feature label X is as close as possible to the parameter change information.

[0110] Positive feedback neural networks include:

[0111] The feedback separator is used to separate the parameter change signal from the data sequence TX. The feature label represents the weight loss progress of the object. The nutritional content information is continuously corrected through the feature label X until the feature label X is as close as possible to the parameter change information.

[0112] Encoding networks are used to transform redundant information sequences into input data sequences. By comparing the input data sequences, the dimensionality of the input data sequences can be effectively reduced, thus reducing the complexity of the model.

[0113] refer to Figure 4 A positive feedback neural network outputs nutrient content information Y based on the input feature label X, so that the feature label X is as close as possible to the parameter change information.

[0114] Positive feedback neural networks include:

[0115] Feedback separator, used to separate parameter variation signals from a data sequence TX. ;

[0116] ;

[0117] in, This represents the parameter change signal in the parameter change sequence. Let represent the compensation signal for the t-th feature label in the data sequence TX, where t represents the index of the feature label. Indicates the feedback intensity coefficient. Indicates the lag step size;

[0118] The parameter changer, with a built-in neural network structure, generates nutrient content information Y based on the input parameter change signal. t ;

[0119] ;

[0120] Represents the neural network structure. Y represents the model parameters of the neural network structure. t This represents the nutrient content information generated at the t-th time step;

[0121] The parameter changer is a neural network model that can generate nutrient content information based on the input parameter change signal.

[0122] Hysteresis feedback calculator, based on the output nutrient content information Y t Calculate feedback signal ;

[0123] ; This represents the feedback gain coefficient.

[0124] A feedback compensator generates a compensation signal for the next time step based on the feedback signal.

[0125] ;in, This represents the exponential moving average function. This represents the feedback strength coefficient, which is shared by the feedback compensator and the feedback separator.

[0126] This scheme constructs a neural network structure with feedback performance. This design enables the compensation signal to offset the lag effect in real time through precise time step control and adaptive parameter adjustment, so as to achieve accurate tracking of parameter changes by feature labels. It can accurately adjust and monitor features with obvious lag phenomena such as weight, body fat and BMI.

[0127] Loss function of positive feedback neural network for:

[0128] ;

[0129] This represents all learnable parameters of a positive feedback neural network, where t represents the time step index. This represents the actual change in feature labels. Indicates time interval, Used to represent regularization parameters , This indicates information about parameter changes.

[0130] This represents the tracking error term, which guides the feedback compensator to generate a compensation signal that can follow the actual data sequence TX. This represents the rate of change term, used to reduce tracking latency and improve the dynamic response performance of the system.

[0131] This indicates the safety aspects of weight loss, guiding children to make their weight loss efficiency as close as possible to or equal to the parameter change information.

[0132] This solution divides the generation of recommended recipes into two parts: an energy training network to adjust the required nutrients and a recipe generation network to learn children's taste preferences so that the substances children ingest match the nutritional information as closely as possible.

[0133] The recipe generation network includes an encoder, an ingredient information filter, and a recipe generation module.

[0134] The encoder is used to compile the food information in the food database into food data W, W={w1, w2, w3, w4}; w1, w2, w3 represent the carbohydrate content, fat content and protein content per unit weight, respectively, and w4 represents the retail price per unit weight of food.

[0135] The ingredient data W includes the nutritional value of the market and the required amount.

[0136] The ingredient information filter filters ingredient data based on the input recipe budget and quantity, resulting in an ingredient selection library R; specifically, it selects ingredients with a retail price per unit weight less than... Food data is entered into the food selection database; .

[0137] Where g1 is the maximum consumption amount entered by the user, a1 is the number of dishes entered by the user, a2 is the pre-set weight of ingredients consumed for each dish, and a2 is a pre-set fixed value. minThis represents the retail price of the cheapest ingredient among all ingredient data, with m0 representing the unit weight, which is the default minimum purchase quantity. The ingredient information filter can remove expensive ingredients that the user cannot afford, minimizing the complexity when generating recipes later.

[0138] After obtaining all the ingredient data W, distribute all the ingredient data into a 4-dimensional dictionary according to w1, w2, w3, w4, to obtain the ingredient dictionary WD.

[0139] The recipe generation module generates recommended recipes from the food dictionary WD to ensure that the nutritional content of the recommended recipes for children is close to the nutritional information.

[0140] refer to Figure 5 The recipe generation module includes: a generator, a first discriminator, and a second discriminator;

[0141] The generator produces preliminary recommended recipes based on nutritional information and the number of dishes.

[0142] The recommended recipe is R;

[0143] R = {r1, r2, ... r} i ...r m}, where m represents the number of dishes, i represents the index of the dish, and r i This indicates the i-th dish in the preliminary recommended recipe;

[0144] , Indicates r i Cooking methods include boiling, stir-frying, pan-frying, braising, and deep-frying; Indicates r i The first ingredient in the dish. Indicates r i The k-th ingredient in the list, where k represents the ingredient index and K represents the total number of ingredients.

[0145] To reduce the dimensionality of the model, s1 is used as the main ingredient, accounting for 50% of the total weight of the dish. The weight of the other ingredients is the same. When there are only two ingredients, each ingredient accounts for 50% of the weight.

[0146] The weight of ingredients consumed in each dish is a2, and the unit weight is m0, where the unit weight represents the minimum purchase quantity of each ingredient. Therefore, the maximum quantity of ingredients in a given dish is K. .

[0147] Thus, the generator uses the number of dishes to randomly generate m dish codes from the ingredient dictionary (WD), while the cooking method for each ingredient is randomly generated. This method can generate a random recommended recipe. (Currently, most common ingredients support the four cooking methods of frying, stir-frying, braising, and stewing.)

[0148] The generator includes:

[0149] The initial setup layer sets the range of total food intake for the child in the initial state, as well as the range of the proportion of each dish consumed. The nutritional content information is used as a constraint to randomly generate an initial recipe z. The initial recipe z conforms to the nutritional content information within the set range of total food intake and the range of the proportion of each dish consumed.

[0150] For example, if a child's total daily food intake is 500g, and there are 5 dishes, each with a consumption range of 10% to 30%, then given the known nutritional information, we can determine each child's daily requirements for carbohydrates, protein, and fat. The generated initial diet plan z uses the total food intake of 500g and the 10% to 30% consumption range for each dish, combined with the nutrients provided by each dish, to determine the daily range for carbohydrates, protein, and fat. If the nutritional information falls within the predicted ranges for carbohydrates, protein, and fat for the day, then the initial diet plan z is effective.

[0151] Multiple hidden layers, each employing a fully connected transformation of the initial recipe z using an activation function; the activation function is ReLU; this application has three hidden layers, and the gradient propagation formula for the three hidden layers is as follows:

[0152] ;

[0153] in, This represents the gradient of the loss function LG with respect to the weights of the first layer, where LG represents the value of the loss function. This represents the weight matrix of the k-th hidden layer. Represents the norm, This represents the linear output of the k-th hidden layer, where k represents the index of the hidden layer.

[0154] Output layer, the output to the hidden layer Perform a linear transformation to obtain the output vector ;

[0155] K represents the total number of hidden layers, a K This represents the output of the last hidden layer, b o W represents the bias vector of the output layer. o This represents the weight matrix of the output layer; The output vector of the output layer;

[0156] The reshaping layer transforms the two-dimensional output vector. Reshape into a vector matrix ;

[0157] , ;r i Representing a vector matrix The i-th row vector in the middle;

[0158] Range constraint layer, for vector matrix Apply range constraints to obtain the recommended recipe R;

[0159] The constraint methods are as follows:

[0160] For vector matrix Feature extraction was performed to obtain the compressed value of carbohydrate intake. Fat intake compression value Protein intake compression value :

[0161] ; ; ;

[0162] ;

[0163] ;

[0164] ;

[0165] Where Car is the carbohydrate intake value corresponding to vector matrix JR, Fat is the fat intake value corresponding to vector matrix JR, and Pro is the protein intake value corresponding to vector matrix JR; v i These are weighting coefficients used to describe the proportion of children consuming the i-th dish. This represents the carbohydrate content of the k-th ingredient in the i-th dish. This represents the fat content of the k-th ingredient in the i-th dish. This represents the fat content of the k-th ingredient in the i-th dish;

[0166] Compressibility of carbohydrate intake Fat intake compression value Protein intake compression value Linear scaling is used to obtain the corrected value for carbohydrate intake. Corrected value for fat intake Protein intake correction value ;

[0167] ;

[0168] ;

[0169] ;

[0170] , ;

[0171] , ;

[0172] , ;

[0173] y1, y2, and y3 represent the carbohydrate intake, protein intake, and fat intake from the nutritional information, respectively. This indicates the lower limit of carbohydrate intake derived from nutritional information. This indicates the upper limit of carbohydrate intake based on nutritional information. This indicates the lower limit of fat intake based on nutritional information. This indicates the upper limit of fat intake based on nutritional information. This represents the lower limit parameter for carbohydrates. This indicates the upper limit parameter for carbohydrates. This represents the lower limit parameter for fat. This represents the upper limit parameter for fat intake. This represents the lower limit parameter for protein levels. This represents the upper limit parameter for protein levels;

[0174] Based on carbohydrate intake correction value Corrected value for fat intake Protein intake correction value Adjust the names of ingredients in each dish to select the closest combination from the ingredient database W as the recommended recipe R.

[0175] The specific plan is as follows:

[0176] S1: Calculate the correction value The difference with Car, according to The difference between Car and the vector matrix is ​​then processed sequentially. Replace the ingredients with the highest carbohydrate content until the Car value calculated by the vector matrix JR is reached and corrected. The difference between ` is less than the preset threshold;

[0177] S2: Calculate the correction value `and The difference, according to `and The difference is successively applied to the vector matrix Replace the ingredients with the highest fat content until the vector matrix JR is calculated. With correction value The difference between ` is less than the preset threshold;

[0178] S3: Calculate the correction value `and The difference, according to `and The difference is successively applied to the vector matrix Replace the ingredients with the highest protein content until the vector matrix JR is calculated. With correction value The difference between ` is less than the preset threshold;

[0179] S4: Adjust the weight of all ingredients in the vector matrix JR until the weight calculated by the vector matrix JR is obtained. The difference with Car, `and The difference and `and Only when the difference is minimized can the resulting vector matrix JR correspond to the recommended recipe R.

[0180] When replacing ingredients, the process actually involves replacing them based on their coordinates in the ingredient dictionary, thereby selecting the optimal vector matrix JR as the initial recommended recipe generated by the generator.

[0181] The first discriminator determines whether the preliminary recipe meets the budget amount based on the recipe budget. If it does, the preliminary recommended recipe is sent to the second discriminator.

[0182] The first discriminator primarily serves as a filter, with relatively little constraint on recipe selection based on price. Furthermore, there's no clear correlation between price and children's liking for certain recipes. Therefore, introducing this mechanism into the generator or discriminator could easily lead to overfitting, preventing accurate selection based on user preferences and impacting model accuracy. To address this, it's used as a separate discriminator. With constraints on the ingredient dictionary, over 50% of randomly generated recommended recipes pass the first discriminator. Thus, the first discriminator acts as a random on / off switch for the generator, restarting it without warning.

[0183] The second discriminator determines whether the nutritional content of the amount of food consumed by the child in the preliminary recommended diet is close to the nutritional content information. If it is close to the nutritional content information, the preliminary recommended diet is output as the recommended diet.

[0184] The second discriminator includes: a PointNet layer with a built-in fully connected network that shares weights, used to process the recommended recipe R into a high-dimensional matrix H;

[0185] The global max pooling layer performs global max pooling on the high-dimensional matrix H to generate feature vectors.

[0186] The multi-layer perceptron extracts multi-layer perceptron features from the feature vectors to generate the conformity degree D(R) of the recommended recipes. If the conformity degree is higher than the preset threshold, it means that the recommended recipe R is qualified and is finally output. If it is not qualified, it is regenerated.

[0187] ;

[0188] , , These represent the activation outputs of the first, second, and third layers in a multi-layered perceptron. This represents the weight matrix of the first layer of the multi-layer sensing layer. This represents the weight matrix of the second layer of the multi-layer sensing layer. This represents the weight matrix of the third layer of the multi-sensor layer. This represents the output layer weight matrix of the multilayer sensing layer. This represents the bias vector of the first layer of the multi-layer perceptron. This represents the bias vector of the second layer of the multi-layer perceptron. This represents the bias vector of the third layer of the multi-layer perceptron. This represents the output layer bias vector of the multi-layer sensing layer.

[0189] The discriminator learns to distinguish the relationship between recommended recipes and children's consumption amounts. The generator, on the other hand, learns to deceive the discriminator. The specific process is as follows: Discriminator training: The generator is fixed, and the discriminator is trained using real data and generated trajectories, causing it to give high scores to real trajectories and low scores to generated trajectories, maximizing the score difference between the two.

[0190] Generator training: Fix the discriminator and adjust the generator parameters to make the generated data obtain higher discriminator scores and minimize the distribution difference with the real trajectory.

[0191] Alternating optimization: The above two steps are executed cyclically, forming a dynamic game—the discriminator continuously improves its discrimination ability, and the generator continuously improves its generation quality, until the generated trajectory is indistinguishable from the real trajectory. The real data here refers to children's consumption of the recommended recipes, i.e., their taste preferences.

[0192] In practice, the generator and discriminator models are pre-trained on a server using a large amount of taste preference data from children of the same age, gender, and BMI. The terminal device then accesses the server to retrieve and load these model parameters. For children's weight loss, there is no clearly defined timeframe; it's a long-term process. In practice, children's weight, body fat percentage, BMI change rate, and daily consumption of different dishes can be used as labeled data to train the generator and the second discriminator. Correspondingly, the energy training network can also be trained.

[0193] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A recipe generation system for obese children based on deep learning neural networks, characterized in that, include: The server and the terminal devices connected to the server, wherein: The servers include: The cloud platform connects to the retail interfaces of various fresh food markets to store information on the ingredients being sold in each market. The cloud processor receives the location information of the terminal device, extracts information on the ingredients being sold in the fresh food market in the vicinity of the terminal device, and sends it to the terminal device. Terminal equipment includes: A locator is used to obtain current location information; Data storage is used to store food information sent by the cloud processor and generate a food database; The data collector is used to collect users' recipe information, which includes the number of dishes and the vital signs data of each child. Instruction memory, which stores one or more instructions; The terminal processor is configured to execute one or more instructions stored in the instruction memory, input recipe information and vital sign data into the recipe generation model to select the required ingredients from the ingredient database, and generate a recommended recipe. The recipe generation model generates recommended recipes based on the target children's preferred tastes and target BMI; Ingredient information includes the ingredient name and unit price; Recipe generation models include: The target generator generates target parameters and parameter change information based on the input vital sign data. The target parameters include target BMI, target weight, and target body fat, and the parameter change information includes the rate of change of BMI, the rate of change of weight, and the rate of change of body fat. An energy training network generates nutritional information based on input vital sign data and a pre-set target BMI. A recipe generation network generates recommended recipes based on nutritional information, children's preferred tastes, and ingredient information. The energy training network includes: The encoding network is used to transform the input vital sign data into feature labels X, and obtain all feature labels X to get the input data sequence TX; A positive feedback neural network outputs nutrient content information Y based on the input feature label X, so as to minimize the error between the feature label X and the parameter change information; X = {x1, x2, x3} x1 = current BMI - previous BMI, x2 = current weight - previous weight, x3 represents current body fat - previous body fat; Y = {y1, y2, y3}, where y1 represents the carbohydrate intake required for the day, y2 represents the protein intake required for the day, and y3 represents the fat intake required for the day; Positive feedback neural networks include: Feedback separator, used to separate parameter variation signals from a data sequence TX. ; ; in, This represents the parameter change signal in the parameter change sequence. Let represent the compensation signal for the t-th feature label in the data sequence TX, where t represents the index of the feature label. Indicates the feedback intensity coefficient. Indicates the lag step size; The parameter changer, with a built-in neural network structure, generates nutrient content information Y based on the input parameter change signal. t ; ； Represents the neural network structure. Y represents the model parameters of the neural network structure. t This represents the nutrient content information generated at the t-th time step; Hysteresis feedback calculator, based on the output nutrient content information Y t Calculate feedback signal ; ; Indicates the feedback gain coefficient; A feedback compensator generates a compensation signal for the next time step based on the feedback signal. ;in, This represents the exponential moving average function. This represents the feedback strength coefficient, which is shared by the feedback compensator and the feedback separator.

2. The diet plan generation system for obese children based on deep learning neural networks according to claim 1, characterized in that: Terminal equipment also includes: The input device allows users to modify the ingredient database, adding or deleting ingredients.

3. The diet plan generation system for obese children based on deep learning neural networks according to claim 1, characterized in that: Loss function of positive feedback neural network for: ； This represents all learnable parameters of a positive feedback neural network, where t represents the time step index. This represents the actual change in feature labels. Indicates time interval, Used to represent regularization parameters , Indicates information about parameter changes; This represents the tracking error term, which guides the feedback compensator to generate a compensation signal that can follow the actual data sequence TX. Represents the rate of change; This indicates a safe option for weight loss.

4. The recipe generation system for obese children based on deep learning neural networks according to claim 1, characterized in that: Recipe generation networks include: The encoder is used to compile the food information in the food database into food data W, W={W1, W2, W3, W4}; W1, W2, W3 represent the carbohydrate content, fat content and protein content per unit weight, respectively, and W4 represents the retail price of the food per unit weight. The ingredient information filter filters ingredient data based on the input recipe budget and quantity, resulting in an ingredient selection library R; it then selects ingredients with a retail price per unit weight less than... Food data is entered into the food selection database; The recipe generation module generates recommended recipes from the ingredient selection library R so that the nutritional content of the recommended recipes for children is close to the nutritional information.

5. The recipe generation system for obese children based on deep learning neural networks according to claim 4, characterized in that: ； Where g1 is the maximum spending amount entered by the user, a1 is the number of dishes entered by the user, a2 is the pre-set weight of ingredients consumed for each dish, m0 is the minimum purchase quantity of ingredients, and g min The retail price of the cheapest ingredient among all ingredient data.

6. The recipe generation system for obese children based on deep learning neural networks according to claim 4, characterized in that: The recipe generation module includes: The generator produces preliminary recommended recipes based on nutritional information and the number of dishes. The first discriminator determines whether the preliminary recipe meets the budget amount based on the recipe budget. If it does, the preliminary recommended recipe is sent to the second discriminator. The second discriminator determines whether the nutritional content of the amount of food consumed by the child in the preliminary recommended diet is close to the nutritional content information. If it is close to the nutritional content information, the preliminary recommended diet is output as the recommended diet.

7. The recipe generation system for obese children based on deep learning neural networks according to claim 6, characterized in that: The generator and the second discriminator form an adversarial network, with the first discriminator acting as a random activator for the generator to increase the richness of the recommended recipes.

Images