Intelligent meal system and method based on metabolic management

Abstract

This invention discloses an intelligent dietary system and method based on metabolic management. The system includes a data acquisition unit for collecting basic user information and meal records; a cloud service unit for dynamically assessing the user's current blood glucose status and dietary restriction needs based on dietary guidelines and individual user data; a user interaction unit for providing an intuitive and user-friendly interface, allowing users to view recommended plans, provide feedback on implementation status, and adjust dietary preferences; and an external ecosystem connection unit for acquiring real-time or historical blood glucose data from the cloud for dynamically adjusting dietary recommendations. The method of the intelligent dietary system based on metabolic management has the following advantages: lowered usage threshold; real-time closed-loop operation; improved execution accuracy; and seamless integration.

Description

Technical Field

[0001] This invention relates to an intelligent dietary system based on metabolic management, and more particularly to an intelligent dietary system and method based on metabolic management. Background Technology

[0002] In recent years, with the continuous rise in diabetes prevalence, lifestyle interventions centered on diet, exercise, blood glucose monitoring, and behavioral habits have become a key aspect of diabetes prevention and long-term management. Currently, diabetes lifestyle intervention services on the market mainly rely on digital platforms such as mobile applications (Apps) and WeChat mini-programs, or employ one-on-one / group guidance service models. Simultaneously, various smart monitoring devices, wearable terminals, and intelligent dietary analysis tools are gradually becoming more widespread, attempting to improve intervention efficiency through data-driven and intelligent methods. However, existing technologies and products still have significant limitations in practical applications, specifically as follows: 1) Smart food scales: such as the "Mint Health" smart scale, can only weigh and roughly estimate calories, cannot identify food types, cannot provide dynamic ratio suggestions based on the user's personalized metabolic data, and has no voice interaction function.

[0003] 2) Smart Dietary System: There are some plates with weighing functions on the market (such as "Hapi"), but they are limited in function, only recording the total weight, without zone recognition, algorithm interaction, or blood glucose data linkage, and cannot be used for chronic disease management.

[0004] 3) Diet tracking apps: such as "Sugar Nurse" and "Livongo", rely on users to manually input or upload photos, resulting in data lag and low accuracy, making real-time intervention impossible.

[0005] In summary, existing products mainly focus on the software level (such as various diabetes management apps) or single-function smart hardware (such as smart scales and blood glucose meters), and still have the following shortcomings: High learning threshold: Users need to master nutritional knowledge and manually calculate, which is difficult to persist with; Passive data collection: Relying on users to actively enter data, there are subjective resistance, delays, and errors; Untimely intervention: Apps cannot perceive the eating process in real time, and guidance is delayed; Low execution accuracy: Lacking quantitative tools, the theoretical ratio deviates greatly from the actual intake; Detached from life scenarios: Software and hardware are separated and cannot be integrated into daily cooking and dining processes. Summary of the Invention

[0006] To address the shortcomings of the aforementioned technologies, this invention provides an intelligent dietary system and method based on metabolic management.

[0007] To solve the above technical problems, the technical solution adopted by the present invention is: an intelligent dietary system based on metabolic management, including a data acquisition unit, a cloud service unit, a user interaction unit, and an external ecological connection unit; Data acquisition unit: used to collect users' basic information and dining records; Cloud service unit: used to dynamically assess a user's current blood glucose status and dietary restriction needs based on the Dietary Guidelines for Residents and individual user data; User interaction unit: Used to provide an intuitive and user-friendly interface, supporting users to view recommended plans, provide feedback on implementation status, and adjust dietary preferences; External ecosystem connection unit: used to obtain users' real-time or historical blood glucose data through the cloud, and to dynamically adjust dietary recommendations.

[0008] Preferably, the data acquisition unit includes a processing module, a weighing sensor, a communication module, a voice interaction module, and a camera module; The processing module is responsible for receiving multimodal data collected by the weighing sensor and camera, calling the carbon-egg ratio algorithm for processing, and feeding the results back to the user through the voice interaction module and the display screen, while synchronizing with the cloud through the communication module. The camera module uses dual cameras: the top camera shoots vertically downwards at the plate for food recognition, volume estimation, and plate area segmentation; the front camera faces the user and is used for face recognition, user proximity detection, and voice interaction assistance. The voice interaction module is used to provide real-time voice guidance to users, relay the guidance from the processing module to users, and interact with users.

[0009] Preferably, the cloud service unit adopts initial suggestions based on population statistics during the initial user phase. As the user accumulates diet-glucose response data during the usage process, it dynamically updates metabolic classification and personalized parameters, and achieves accurate suggestions through convergence judgment and anomaly detection. The cloud service unit is equipped with a carbon-egg ratio algorithm and a reinforcement learning model. By using users' dietary records, a complete user profile is first generated, and then the target ratio for each meal is adjusted in real time.

[0010] The cloud service unit pushes food purchasing suggestion reports to users through the user interaction unit, and users provide feedback on their actual purchases in the user interaction unit.

[0011] Preferably, the cloud service unit automatically adjusts the reminder level based on user usage frequency and abnormal status, and dynamically controls the reminder frequency through user response behavior, while leveraging social support networks to improve user compliance; the reminder levels include in-device reminders, app push notifications, third-party assistance reminders, and manual follow-ups.

[0012] A method for a smart dietary system based on metabolic management includes the following steps: Step S1, User Identification: Upon first use, the user completes the binding process by scanning a QR code; Step S2, Food Recognition and Weighing: After the user places the food into the corresponding area, the weighing sensor measures the weight increment of each area in real time. Through camera image recognition and weighing sensor data fusion, the food type, weight and volume are automatically determined, and voice interaction is used to supplement the recognition. Step S3, Carbon-to-Egg Ratio Calculation: The processing module calls the carbon-to-egg ratio algorithm, and calculates the target carbon-to-egg volume ratio for the current meal by combining the user's metabolic classification, kidney function, and BMI correction factor. Step S4, Real-time Feedback and Guidance: Compare the actual weight of each area with the target ratio and provide real-time guidance through voice broadcast; Step S5, Intervention Mode Switching and Blood Glucose Monitoring Reminder: The system provides differentiated blood glucose monitoring guidance based on the user's current intervention mode.

[0013] Preferably, in step S1, the front-facing camera automatically confirms the user's identity through facial recognition or by connecting to a mobile phone via Bluetooth, and downloads the user's metabolic classification, current intervention mode, and five-point physiological rhythm threshold from the cloud; it can automatically switch between multiple users.

[0014] Preferably, in step S2, the top camera automatically captures images of the food, identifies the type of food through a local lightweight AI model or by uploading to the cloud, estimates its volume and weight, and verifies it by fusing it with the weighing data.

[0015] If the recognition confidence is low, the user will be asked via voice, and the user can then confirm via voice or touchscreen.

[0016] In step S2, food identification includes the following steps: Step S21, Region Segmentation: Use a lightweight semantic segmentation model to separate the plate region in the image from the background and identify the boundaries of each region; Step S22, Food Area Detection: Within each partition, independent food blocks are identified through edge detection and connected component analysis; if two food items are stacked in the same partition, the user is prompted to separate the food items. Step S23, Food Classification: For each food piece, call the image classification model; the model pre-training includes a self-built dataset of common Chinese ingredients; Step S24, Volume and Weight Estimation: Combining the pixel area of ​​the food block in the image, the camera focal length, and the known size of the plate, estimate the volume of the food. Simultaneously, receive the actual weight of the area from the weighing sensor and call the multimodal fusion algorithm. Based on the food category, query the preset density library and calculate the estimated image weight = estimated image volume × density; The weighted average is calculated with the actual weight measured by the weighing sensor, and the weights are dynamically adjusted according to lighting conditions and the complexity of the food's shape; if the difference between the two is >30%, a voice inquiry is triggered. Step S25, Confidence Assessment and Interactive Confirmation: If the classification confidence is <80% or the image conflicts with the weighing sensor data, the user will be asked via voice; the user can answer via voice or manually select via touchscreen. The system uses the user's confirmation as a new sample and uploads it to the cloud for subsequent model optimization.

[0017] Preferably, in step S5, the intervention mode includes: In the heavy intervention mode, the system automatically obtains real-time CGM data through the cloud, dynamically adjusts the carbon-to-protein ratio recommendation based on the user's current blood glucose level and blood glucose trend during the meal, and automatically compares the CGM data 2 hours after the meal to assess the blood glucose impact of the meal for algorithm optimization. The light intervention mode includes the following scenarios: When a user has not linked any blood glucose monitoring device, the system only provides dietary guidance based on metabolic classification and does not provide blood glucose reminders, but still reminds the user to monitor blood glucose as a health reminder through the five-point rhythm. When a user binds a CGM device but is not currently wearing it, the system reminds the user to take measurements according to the five-point rhythm and records the values ​​through photo recognition or voice input. When a user wears a CGM device but the system determines that the current period is a low-intensity intervention period, the system prioritizes using real-time CGM data for mild intervention guidance and only triggers voice reminders when postprandial blood glucose levels rise abnormally.

[0018] Preferably, in step S5, the system automatically detects the status of the user's blood glucose monitoring device and automatically switches between heavy intervention mode and light intervention mode; After each meal, the processing module automatically uploads the food data, including food type, weight, carbohydrate-to-protein ratio, and time, to the cloud to update the user's food records. The cloud-based system combines user CGM / BGM data to assess postprandial blood glucose response and dynamically optimize subsequent recommendations. If an anomaly is detected, the intervention mode is switched via the cloud and the user is alerted via voice prompts on the dining tray.

[0019] Compared with the prior art, the present invention has the following beneficial effects: 1) Lowering the barrier to entry: Through partitioned design and automation functions, users can accurately execute diet plans without any learning.

[0020] 2) Achieve real-time closed-loop: The intelligent diet system automatically collects dietary data, interacts with cloud algorithms in real time, and provides instant feedback and guidance through voice interaction.

[0021] 3) Improve execution accuracy: Integrate weighing, image recognition, and user historical data to ensure that the carbon-egg ratio is accurate to the gram.

[0022] 4) Bridging the last mile: Physicalizing digital therapy, making the smart diet system a personal smart assistant in the home, and truly integrating it into the user's daily life.

[0023] 5) Achieving a leap from passive recording to proactive planning: Existing technologies mostly focus on real-time recording and feedback during user meals. This invention introduces a weekly food procurement suggestion function for the first time. Based on retrospective analysis of users' long-term dietary data, it proactively plans future dietary structure, helping users optimize their diet from the source, demonstrating stronger foresight and proactivity. Attached Figure Description

[0024] Figure 1 This is a system architecture diagram of the present invention.

[0025] Figure 2 This is an architecture diagram of the cloud service unit of the present invention.

[0026] Figure 3 for Figure 2 The corresponding product structure diagram. Detailed Implementation

[0027] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0028] like Figure 1 The system shown is a smart dietary system based on metabolic management. The system consists of four parts: a data acquisition unit (hardware layer and software layer), a cloud service platform, a user interaction terminal (App / mini-program, real-time voice guidance, and abnormal warning notification), and external ecosystem connection. The four parts communicate in real time via wireless network (Wi-Fi / Bluetooth / 5G).

[0029] The cloud service platform architecture is built on mainstream cloud platforms (Alibaba Cloud / Tencent Cloud / Huawei Cloud), and includes the following steps: 1. Equipment registration and certification: Each intelligent food preparation system is programmed with a unique device ID and key during production; Upon first use, the device connects to the internet via Wi-Fi and sends its device ID and key to the cloud for authentication; the cloud returns an access token (JWT), which is carried in all subsequent communications. 2. Data communication protocol: MQTT protocol: used for real-time bidirectional communication (device status reporting, cloud command delivery), maintaining a long connection; HTTPS protocol: used for file uploads (such as user-confirmed image samples) and batch data synchronization; WebSocket protocol: used for real-time data synchronization with the app; 3. Cloud-based algorithm deployment: The carbon-egg ratio algorithm and reinforcement learning model are deployed in the algorithm service container; every day at midnight, the cloud performs incremental training of the model based on the previous day's data from all users; new model parameters are pushed to the local device via MQTT, supporting OTA remote upgrades. 4. Dynamic User Profile Evolution Mechanism: The system adopts a progressive personalized convergence strategy to solve the cold start problem for new users, and continuously optimizes the accuracy of suggestions as data accumulates. 4.1 Initial Stage: When a user uses the system for the first time, the system only obtains basic information (age, gender, height, weight, type of diabetes, chief complaint of comorbidities and complications, medication status). The cloud generates an initial metabolic classification estimate based on a population statistical model (mean data stratified by age, BMI, and diabetes type), and marks it as "confidence: low". The voice guidance adopts a conservative style, such as "Based on your initial information, we recommend a carbohydrate intake of about 150 grams, which will be further optimized as your data improves." The system clearly informs users: "The first 6-9 meals are a learning period. As you increase the number of meals, the suggestions will become more accurate." 4.2 Data accumulation phase: Each time a user completes a meal record (including food type, weight, and post-meal blood glucose response), the system records the "diet-blood glucose" response data pair for that meal; The cloud-based algorithm performs weighted learning on the following dimensions: Food level: User's glycemic response characteristics to different foods (e.g., rice with the same carbohydrate content vs. mixed grain rice). Time-related factors: Differences in insulin sensitivity among users at different times of day (morning / noon / evening); Combination level: User's blood glucose response pattern to a specific "carbohydrate-to-protein ratio"; After accumulating 3-9 meals of valid data, the system reassesses the user's metabolic classification and updates the confidence label. 4.3 Convergence Judgment and Active Verification: The system sets a convergence threshold: when the deviation between the suggested blood glucose response and the actual blood glucose response for 5 consecutive meals is less than the preset error range (e.g., the difference between the predicted blood glucose value and the measured value 2 hours after a meal is <1.5 mmol / L), it is determined that the user profile has converged to a stable state. After the convergence, the system will inform the user via voice: "I understand your eating habits, and subsequent suggestions will be more precise." If no new data is input for an extended period (e.g., 30 days), the system will automatically lower the confidence level and revert to a conservative recommendation mode. 4.4 Anomaly Detection and User Profile Correction: If a user's blood glucose response after a meal deviates significantly from the prediction (e.g., predicted increase of 2 mmol / L, actual increase of 5 mmol / L), the system marks it as an "abnormal event"; The following actions are triggered in the cloud: Voice prompt: "Your post-meal blood sugar is higher than expected. Did you eat any special foods or have any changes in your physical condition?" If a user confirms this (e.g., "I ate fried food today"), the system will mark the food as "high-risk food" and add it to the user's personalized food blacklist. If deviations occur continuously, the system will automatically reduce the confidence level of the current profile and re-enter the convergence process; In the initial stage of user experience, the system adopts initial suggestions based on population statistics. As users accumulate "diet-blood glucose" response data during use, the system dynamically updates metabolic classification and personalized parameters. Through convergence judgment and anomaly detection, the system proactively evolves from "fuzzy suggestions" to "precise suggestions," thereby improving the early user experience.

[0030] 5. User data synchronization: The device locally stores the dietary records of the past 30 days; each time it connects to the network, the new data is compressed and uploaded to the cloud; the cloud merges data from multiple terminals (device, App, CGM) to generate a complete user profile; 6. Reminder Service and Third-Party Monitoring Mechanism: The system provides a multi-layered reminder mechanism designed to encourage frequent device use while avoiding excessive interruptions that could cause user annoyance. 6.1 The reminder service includes the following four levels of reminders:

[0031] Level 1 Reminder (In-device): If the device is not used beyond the preset meal time, the tray will announce "It's time to eat" via voice. This can only happen once per meal. Level 2 Reminder (App Push): If the app remains inactive for 30 minutes after the Level 1 reminder, an App notification will be sent, up to twice per day. Level 3 reminders (third-party supervision): If no food is consumed for 2 consecutive meals or no data is available for 24 consecutive hours, a text message / WeChat message will be sent to the authorized contact person (family member / caregiver), up to 3 times per week; Level 4 Reminder (Manual Intervention): If the user has not used the service for 3 consecutive days, the system will mark the user as a "lost user" and a service person will make a manual follow-up call, which can be done a maximum of once per month. 6.2 Third-party supervision mechanism: When users use the app for the first time, they can authorize the addition of 1-3 "health assistants" (such as spouse, children, caregivers). The scope of authorization includes: receiving alerts about abnormal user device usage and viewing user dietary record summaries (de-sensitized); Example of a third-party notification: "The [user name] you are following has not used the smart meal system for two consecutive meals. Please help to check on them." After receiving the reminder, the third party can send an encouraging message (such as "Remember to eat on time") with one click through the mini-program, and the system will forward the message to the user's app; Users can turn off third-party notifications in the app, or set a "do not disturb" period (e.g., do not send notifications from 9:00 PM to 7:00 AM the next day). 6.3 Reminder Frequency Control and Anti-Harassment Mechanism: The system records the user's response to reminders (whether the device is used after the reminder). If a user responds to the reminder five times consecutively, the system will maintain the current reminder frequency. If a user fails to respond to an alert three times in a row, the system will automatically downgrade the alert frequency (e.g., skipping from level 2 alerts and going directly to level 3 alerts). If a user actively turns off app notifications, the system will automatically pause second-level and higher-level notifications, retaining only first-level in-device notifications. All notification sending time, type, and user response data are recorded in the cloud for subsequent optimization of notification strategies; The system automatically adjusts the reminder level (in-device reminders, app push notifications, third-party assistance reminders, and manual follow-ups) based on user usage frequency and abnormal status, and dynamically controls the reminder frequency through user response behavior to avoid excessive disturbance, while leveraging social support networks to improve user compliance.

[0032] 7. Multi-tenant isolation: Dietary data and health records of different users are strictly isolated at the database level; all sensitive data (facial features, medical records) are encrypted and stored, and the keys are managed independently by the users; 8. Weekly Food Procurement Suggestion Generation Mechanism: The system adds a food analysis engine and recommendation service in the cloud, aiming to extend dietary management from "meal-by-meal guidance" to "cycle planning," helping users optimize their diet from the source. The mechanism operates as follows: 8.1 Data Collection and Review: The system automatically summarizes the user's complete dietary records for the past 14 days (configurable), including: The intake and proportion of macronutrients (carbohydrates, protein, and fat) at each meal; The actual intake frequency and weight of various foods (staple foods, protein, vegetables, fruits, and oils); Estimated intake of key micronutrients (such as dietary fiber, vitamin C, potassium, etc., based on food identification results and correlation with nutrition databases). Postprandial blood glucose response data (if CGM / BGM linkage is present); 8.2 Nutritional Deficiency Analysis: The cloud-based analytics engine compares the user's actual intake over the past two weeks with the following benchmarks: 1. General Dietary Guidelines Benchmark: Refer to the recommended intake of various foods for the corresponding age and gender in the "Chinese Dietary Guidelines (2022)".

[0033] 2. Personalized chronic disease management benchmarks: Based on the user's metabolic classification, renal function (eGFR), blood glucose control targets, etc., personalized targets are set for the ratio of carbohydrates to protein, dietary fiber, sodium intake, etc. (for example, the recommended upper limit of protein is appropriately reduced for those with renal insufficiency).

[0034] 3. User historical baseline: Compare the user's own data from the past few weeks to identify significant fluctuations or persistent under- or over-intake.

[0035] Output a "Nutritional Deficiency Report", for example:

[0036] "Over the past two weeks, your average daily dietary fiber intake was 15 grams, which is below the recommended 25 grams. We suggest you increase your intake of leafy green vegetables and whole grains."

[0037] "Your total protein intake is up to standard, but the proportion of protein in your dinner is too high. We suggest shifting some of it to lunch."

[0038] "You consume fruit infrequently, so your vitamin C intake may be insufficient."

[0039] 8.3 Seasonal Ingredient Recommendations and Gap Filling: The system has access to or has a built-in seasonal ingredient database (which can be updated regularly), prioritizing the recommendation of high-quality ingredients for the current season (e.g., asparagus and peas in spring; lotus root and yam in autumn).

[0040] Targeted recommendations based on the "Nutritional Deficiency Report": If you are deficient in dietary fiber, it is recommended to eat seasonal high-fiber vegetables (such as broccoli, celery, and mushrooms).

[0041] If you are deficient in vitamin C, seasonal fruits (such as oranges, kiwis, and strawberries) are recommended, along with suggested serving sizes (such as "one medium-sized kiwi per day").

[0042] If users have been consuming refined staple foods frequently recently, it is recommended to replace them with seasonal whole grains (such as fresh corn, sweet potatoes, and yams).

[0043] For users with chronic diseases, the system will provide additional precautions. For example, for those with renal insufficiency, when recommending high-fiber vegetables, it will remind them to "avoid high-potassium vegetables (such as spinach and potatoes) and blanch them before eating"; for diabetic patients, when recommending fruits, it will explicitly suggest that "they should be eaten between meals, and the portion should not exceed 200 grams".

[0044] 8.4 Procurement List Generation and Output: Based on nutritional gap analysis and seasonal recommendations, the system automatically generates a "Procurement Recommendation List for Next Week," with an example of the list structure: This week's recommended shopping list (Date: [Month] [Day] - [Month] [Day], 202X): Based on your dietary analysis over the past two weeks, we recommend the following seasonal ingredients to optimize your diet: 1. Leafy green vegetables (to supplement dietary fiber and vitamins): Seasonal recommendations: Spinach (500g), romaine lettuce (500g); Reason: Your vegetable intake was low last week. Spinach is rich in dietary fiber and iron, and is suitable for the current season.

[0045] 2. Protein (Maintain high-quality protein intake): Recommended ingredients: skinless chicken breast (600g), firm tofu (400g); Reason: Your protein intake is up to standard, but the proportion of red meat is too high. It is recommended to increase white meat and plant protein intake.

[0046] 3. Staple foods (increase the proportion of whole grains): Seasonal recommendations: Fresh corn (3-4 ears), sweet potato (500g); Reason: You have been mainly consuming refined rice and white flour for the past two weeks. It is recommended to partially replace these with fresh corn and sweet potatoes to increase dietary fiber and stabilize blood sugar.

[0047] 4. Fruits (to supplement vitamins, but in moderation): Seasonal recommendation: Strawberries (300g, about 3-4 per day); Reason: You consumed less fruit last week. Strawberries have a low glycemic index and are rich in vitamin C, so they can be eaten in moderation between meals.

[0048] Output channels: A "This Week's Food Procurement Recommendations" report will be pushed to users every Monday via their App / Mini Program.

[0049] Users can easily copy or share the list with family members, or add it to their cart with a single click on supported fresh food e-commerce platforms (authorization required). The system allows users to mark recommended ingredients as "disliked" or "allergic," and subsequent recommendations will automatically avoid these categories.

[0050] 8.5 Closed-loop feedback and optimization: Users can rate the weekly recommendation list in the App (e.g., "very useful", "some are not applicable"), or provide feedback on actual purchase / consumption.

[0051] The system links user feedback with their subsequent dietary records to continuously optimize the recommendation algorithm. For example, if a user does not purchase recommended ingredients for several consecutive weeks, the system will adjust its recommendation strategy and try other similar ingredients.

[0052] The system aggregates users' dietary records from the past two weeks and identifies nutritional gaps by comparing them with general dietary guidelines, personalized chronic disease management benchmarks, and users' historical baselines. Combined with a seasonal food database, it automatically generates a "gap-filling" shopping list, extending from precise guidance for single meals to periodic dietary planning.

[0053] like Figure 3 As shown, the corresponding product of the intelligent meal system mainly consists of a base and a partitioned plate. The plate is placed on the base, with a pressure sensor at the bottom (for weighing); a camera at the top covers the food in the plate to identify the type, properties, and volume of the food. A processing chip performs personalized carbon-to-protein ratio calculations and provides real-time voice feedback. Figure 2 As shown, it mainly includes the following modules: Processing chip: Core computing unit, running RT-OS, handling multimodal data fusion and algorithm calculation, selected model ARM Cortex-M4 / M7 (such as Rockchip RV1126, with NPU). The system incorporates a high-performance chip (such as the ARM Cortex series) and runs a lightweight embedded real-time operating system (RTOS), preferably RT-Thread (a domestically developed open-source RTOS) or FreeRTOS (the RTOS with the highest market share). The system has the following characteristics: Multi-task scheduling: It can handle multiple concurrent tasks such as weighing sensor data acquisition, camera image recognition, voice interaction, and cloud communication simultaneously, ensuring real-time response.

[0054] Low resource consumption: The kernel is small (the smallest kernel can be reduced to 3KB RAM), compatible with ARM Cortex-M series chips, and meets low power consumption requirements.

[0055] Device driver framework: Provides standardized sensor drivers, camera drivers, and Wi-Fi / Bluetooth protocol stacks to simplify hardware adaptation.

[0056] Multi-user management: Supports local storage of profile data such as user identifiers, metabolic classifications, and carbon-to-protein ratio thresholds for multiple users, and enables fast loading when switching users through task scheduling.

[0057] Secure boot and encrypted storage: Supports encrypted storage of sensitive user information (such as health data) to prevent data leakage.

[0058] The system is responsible for receiving multimodal data collected by weighing sensors and cameras, processing it using a carbon-egg ratio algorithm, and feeding the results back to the user through a voice broadcast unit and a display screen. It also synchronizes with the cloud through a communication module. Top camera: Shoots the plate vertically downwards for food recognition and volume estimation. Model OV5640, 5 megapixels, 120° wide-angle. Front-facing camera: Pointed towards the user for facial recognition and user proximity detection. Model OV2640, 2 megapixels, with infrared illumination. To achieve both accurate user identification and food recognition, this system preferably uses a dual-camera configuration: The top camera (Camera A) is mounted on the top bracket of the base. Its height is adjustable to ensure complete coverage of the entire plate area. It can be tilted vertically downwards or slightly tilted to ensure that the food image is distortion-free. It works continuously while the user places the food, and can collect multiple frames of images per meal for multimodal fusion analysis. It is mainly used to identify food types (rice, chicken breast, broccoli, etc.); estimate food volume (combining image pixel area and plate size); and identify the position of food in different sections of the plate.

[0059] The front-facing camera (Camera B) is mounted on the front of the base (it can be hidden in the logo area), facing the user; it can be level or slightly tilted upwards to ensure complete capture of the user's facial features; it has a built-in infrared fill light to ensure recognition performance in low-light environments; it is in low-power standby mode by default and is woken up when the infrared sensor detects the user approaching; its main functions include: User identification: When a user approaches, their face is automatically photographed, and a facial recognition algorithm is run to identify the user's identity; User Status Monitoring: Assists in monitoring user status during the dining process (optional function). Voice interaction assistance: When a user answers a question by voice, lip reading can be captured (advanced feature).

[0060] The system adopts a dual-camera collaborative mechanism: when a user approaches, the infrared sensor wakes up, the front camera starts, and face recognition is successful; after the front camera starts, it loads the user's personalized parameters (metabolic capacity level, carbon dioxide threshold), the front camera enters sleep mode, and the top camera is activated. The user places food, and the top camera continuously captures images to identify the food. If the recognition confidence is low, the system asks the user a question via voice, and the front camera helps capture the user's response. Once the meal preparation process is complete, the system enters standby mode.

[0061] The two cameras share the same image processing unit and work collaboratively through multi-task scheduling of the processing module. After preprocessing, the image data is sent to the face recognition algorithm module and the food recognition algorithm module, respectively. The dual-camera mechanism has the following advantages: Clear division of labor: Each camera focuses on the task best suited to it, resulting in higher recognition accuracy; Smooth user experience: Users can be recognized simply by walking naturally, without having to deliberately bend down to get closer to the plate; All-weather operation: The front-facing camera has infrared fill light, so it can still recognize objects in low-light environments; System power consumption optimization: Dual cameras work in shifts to avoid increased power consumption caused by simultaneous operation.

[0062] Embedded real-time operating system: multi-task scheduling, driver management, multi-user data storage, selected model RT-Thread / FreeRTOS, minimum kernel 3KB RAM;

[0063] Food placement area: removable partitioned trays with independent pressure transmission structure to transfer weight to the weighing sensor, stainless steel support column + silicone isolation layer, compatible with TAL220 sensor.

[0064] High-precision weighing sensors are located at the bottom of the base, enabling real-time measurement of the weight of the partitioned plates and food in each area, with an accuracy of ±1g. These sensors collect the food weight and transmit it to the processing module.

[0065] A resistance strain gauge pressure sensor (such as the domestically produced TAL220 or HX711 sensor) is preferred, and its technical parameters are as follows: Measurement principle: The strain gauge inside the sensor deforms with pressure changes, causing a change in resistance value. This change is converted into a voltage signal by a Wheatstone bridge and then into a digital quantity by a 24-bit high-precision ADC (such as the HX711 chip).

[0066] Accuracy Guarantee: The single sensor has a range of 0-3kg, a nonlinear error of <0.02%FS, and an overall accuracy of ±1g, meeting the precise requirements of carbon-egg ratio calculation for food weight.

[0067] Independent measurement in multiple zones: Each zone of the plate is independently equipped with a pressure sensor (at least 2 or 4, depending on the device model), which is connected to the ADC interface of the processing module to achieve independent weight acquisition of the carbohydrate zone, protein zone, vegetable zone, and extended zone, avoiding measurement interference caused by food stacking or crossing zones.

[0068] Temperature compensation: The sensor has a built-in temperature compensation circuit, which can maintain stable accuracy in an operating environment of -10℃ to 40℃, making it suitable for home kitchen environments.

[0069] Calibration mechanism: The system supports zero-point calibration (zeroing when the disk is empty) and standard weight calibration (periodic verification with a 100g standard weight) to ensure that the accuracy does not decrease with long-term use.

[0070] Voice broadcaster: Triggers voice guidance based on state machine, outputs through speaker, uses I2S interface, and has a pre-stored MP3 voice library; built-in speaker, used to provide real-time voice guidance to users, feedback the guidance from the processing module to users, and interact with users.

[0071] Voice-triggered scenarios: The processing module triggers voice broadcasting in the following typical scenarios, as shown in the table below:

[0072] The following table shows examples of real-time dietary guidance broadcast content:

[0073] The following table shows examples of abnormal warning and guidance broadcast content:

[0074] User voice command: The user says "What should I eat today?" → The system replies "Recommended lunch: 150 grams of carbohydrates and 120 grams of protein"; Meal Completion and Summary: When the user presses the "Meal Configuration Complete" button, the "Meal Record for This Session Has Been Saved, Carbon to Egg Ratio 1.2:1, Meeting Target Requirements" message appears.

[0075] The tray detects that the user has left and enters a sleep state.

[0076] The processing module runs a state machine, which determines the current state in real time based on sensor inputs (weighing, camera, user voice) and data pushed from the cloud. It then calls the corresponding MP3 file from the pre-stored voice library and outputs it to the speaker amplifier through the I2S interface.

[0077] Communication module: Supports Wi-Fi, Bluetooth, and 4G / 5G to ensure real-time communication with the cloud and mobile phones. Selected model: Wi-Fi: ESP32-C3; Bluetooth: Built-in 4G Quectel EC200A; Wi-Fi / Bluetooth modules connect to the processing module via SPI or UART interface; 4G / 5G modules connect via USB or PCIe interface, requiring an additional SIM card slot; All communication modules share the same power management unit and support sleep mode to reduce power consumption.

[0078] Optional small LED screen or indicator lights can be added to display current status, battery level, network connection, etc., as shown in the table below:

[0079] The status and meaning of the indicator lights are shown in the table below:

[0080] The main control chip (processing chip) controls the LED brightness through PWM (pulse width modulation) to achieve a breathing effect; the state switching is uniformly managed by the state machine, which makes a comprehensive judgment based on the input of various sensors.

[0081] The display screen is driven by the processing module through a GUI graphics library, preferably LVGL (a lightweight graphics library) or Shibing GUI (accompanying RT-Thread), as shown in the table below:

[0082] The system has a built-in rechargeable battery and supports wireless charging or USB-C wired charging. The power management hardware includes: Charging Management: When USB-C or wireless charging is connected, the PMIC automatically detects the input voltage and adopts a constant current and constant voltage charging mode (first charging with a constant current of 1A to 4.2V, then switching to constant voltage charging until the current drops to the cutoff value). Battery monitoring: The battery meter calculates the remaining battery power in real time using the coulomb counting method with an accuracy of ±3%. The main controller reads the battery percentage every 10 seconds for display and low battery warning. Power path management: When an external power source is plugged in, the PMIC automatically switches to external power supply + simultaneous charging mode; after the power source is unplugged, it seamlessly switches to battery power.

[0083] Low-power modes include the following three forms: Standby mode: After 5 minutes of inactivity, the display turns off, network connection is maintained, and power consumption is 50mA. Deep sleep: After 30 minutes of inactivity, Wi-Fi is disconnected, only the RTC remains active, and power consumption is <1mA; Wireless charging: The receiving coil conforms to the Qi standard and is compatible with most wireless chargers on the market.

[0084] The partitioned plates are detachable, with the plates independent of the base, making them easy to remove, replace, and clean. The bottom of the base is mainly divided into carbohydrate and protein zones (with additional zones that can be added as needed, such as vegetable, fruit, and fat zones), facilitating the placement of different plates.

[0085] Schematic icons (such as rice icons or meat chunk icons) can be printed on the edges of the zones to guide users to place the food correctly. The pressure transmission structure refers to the mechanical device located at the bottom of each zone, used to independently transfer the weight of the food to the base weighing sensor. This system preferably adopts an independent support column structure, specifically implemented as follows: Each section (carbohydrate section, protein section, vegetable section, and extended section) has a stainless steel support column at the bottom center, with a diameter of 8mm, a height of 5mm, and a hemispherical top.

[0086] The support column and the plate are integrally injection molded (embedded), ensuring that they will not loosen or shift when placed in the embedded position. A 1mm thick silicone insulating layer surrounds the support column to absorb any slight tilting of the plate and prevent interference between the different sections.

[0087] A load cell (such as TAL220) is fixed in the base for each zone, with the force-bearing surface of the cell facing upwards and in contact with the top of the support column.

[0088] When a user places food in a specific compartment, the support column at the bottom of that compartment transmits the weight of the food vertically downwards to the corresponding sensor, which generates an electrical signal. The support columns in other compartments are unaffected or only minimally affected by the insulating layer, thus enabling independent measurement of the weight in each compartment.

[0089] Pressure transmission structures have the following advantages: 1. The mechanical path is clear, and the weight of each partition directly affects the corresponding sensor without cross-interference; 2. The support column and the sensor are in point contact, reducing the impact of friction; 3. The silicone insulating layer adapts to minor deformations of the plate, ensuring reliable contact; 4. The plates can be easily removed for washing, and the exposed support columns are easy to clean.

[0090] Image acquisition and preprocessing adopt a dual-camera acquisition mechanism: the top camera continuously acquires images of the plate at a rate of 5-10 frames per second, and triggers shooting when it detects movement of an object above the plate; the front camera is in standby mode in low power mode, and is awakened when the infrared sensor detects the user approaching to capture the user's facial image.

[0091] The acquired images are first preprocessed in the processing module: 1. Automatic white balance and exposure correction: Eliminates color cast caused by different light sources (daylight, LED, fluorescent lamp); 2. Distortion Correction: Geometric correction is performed to address barrel distortion caused by wide-angle lenses; 3. Image Enhancement: Adaptive histogram equalization improves detail visibility in low-light environments; 4. ROI Extraction: The top camera crops out the plate area based on the preset plate outline; the front camera crops out the face area based on the face detection results.

[0092] The preprocessed images are processed in two parallel paths: the image from the front camera enters the face recognition path, and the image from the top camera enters the food recognition path.

[0093] The facial recognition process (user identification) includes the following steps: Step 1, Face Detection: Use the lightweight face detection model MTCNN (Multi-task Cascaded Convolutional Networks) or the mobile version of RetinaFace; Locate the bounding box of the face and 5 key points (eyes, nose tip, corners of mouth) in the image; If no face is detected, the alternative solution (Bluetooth RSSI near-field recognition or voice inquiry) is triggered. Step 2, Face Alignment: Based on the detected key points, the face is rotated and corrected to a standard pose through affine transformation; Step 3, Feature Extraction: Input the aligned face image into a lightweight version of MobileFaceNet or ArcFace; extract 128-dimensional or 256-dimensional face feature vectors (Face Embedding). Step 4, Feature Comparison: Calculate the cosine similarity between the extracted feature vectors and the locally stored user feature library; If the similarity is greater than the preset threshold (e.g., 0.85), it is determined to be the corresponding registered user; if it is less than the threshold, it is determined to be a new user, triggering the registration process or visitor mode. Step 5, Identity Loading: After successful identification, the system loads the user's metabolic classification, carbon protein threshold, and five physiological rhythm parameters from local flash memory or the cloud, in preparation for providing personalized guidance.

[0094] From shooting to identity verification, the recognition time is less than 500ms; the number of users supported is up to 10 people stored locally, with no limit on the number of users in the cloud.

[0095] The food recognition process (AI image recognition) includes the following steps: Step 1, Plate Region Segmentation: Use a lightweight semantic segmentation model (such as the mobile version of UNet or DeepLabV3) to separate the plate region from the background in the image and identify the boundaries of each segment; Step 2, Food Area Detection: Within each partition, independent food blocks are identified through edge detection and connected component analysis; if two food items are stacked in the same partition, the system will prompt the user to "please separate the food items"; Step 3, Food Classification: For each food piece, call the image classification model MobileNetV3 or EfficientNet-Lite; the model is pre-trained on a self-built dataset containing 120+ common Chinese ingredients, including the following categories: Staple foods: white rice, brown rice, steamed buns, noodles, sweet potatoes, pumpkin, yams, potatoes, corn, etc. Protein sources: chicken breast, fish, shrimp, eggs, tofu, beef, pork, etc. Vegetables: broccoli, bok choy, spinach, tomatoes, cucumbers, etc. Fruits: apples, bananas, blueberries, oranges, etc.; Oils and fats: nuts, avocados, cooking oils (must be manually confirmed); Output the category probability distribution for each food piece, and take the category with the highest probability as the recognition result; Step 4, Volume / Weight Estimation: Combining the pixel area of ​​the food piece in the image, the camera focal length, and the known size of the plate, estimate the food volume. Simultaneously, receive the actual weight of the area from the weighing sensor and call the multimodal fusion algorithm. Query the preset density library based on food category (e.g., rice 0.6g / ml, chicken breast 1.0g / ml); calculate "image estimated weight" = image estimated volume × density; The weighted average is calculated with the "sensor-measured weight," with the weight dynamically adjusted based on lighting conditions and the complexity of the food's shape. If the difference between the two is greater than 30%, a voice question is triggered: "I recognized it as refined rice, and the weight is 150 grams, is that correct?" Step 5, Confidence Assessment and Interactive Confirmation: If the classification confidence is <80% or the image and sensor data conflict, the system will ask the user via voice. Users can select via voice (e.g., "This is sticky rice") or manually via touchscreen; The system will use the user's confirmation results as new samples and upload them to the cloud for subsequent model optimization (including adding high-risk food warnings and adding new food ingredient models). From shooting to output, the recognition time is less than 1 second, the classification accuracy is greater than 90% (common ingredients), and the weight estimation error is less than 10% (after fusion).

[0096] Data processing and storage adopt a local priority principle: the facial feature library and commonly used food classification models are stored on the local chip (such as Rockchip RV1126's 8GB eMMC) to ensure that it can still work when there is no network. De-identified upload: Only the de-identified feature vectors and recognition results are uploaded to the cloud (the original images are not uploaded) for model iteration; Encrypted storage: All biometric data is stored using AES-256 encryption to prevent leakage; The aforementioned AI models can all run on ARM Cortex-M4 / M7+NPU architecture (such as the Rockchip RV1126 with its built-in NPU, achieving a computing power of 2 TOPS), or be optimized for operation on a pure MCU via CMSIS-NN. If an edge-cloud collaborative solution is adopted, only preliminary detection is performed locally, with complex recognition data uploaded to the cloud for processing. Dual cameras share the same image processing unit, achieving collaborative work through multi-task scheduling of the processing module.

[0097] The intelligent dietary system has a built-in chip pre-loaded with a core algorithm module. The system can optionally connect with external blood glucose monitoring devices (including continuous glucose monitoring (CGM) and fingertip blood glucose meters with BGM) to acquire real-time or historical blood glucose data from the user via the cloud, enabling dynamic adjustments to dietary recommendations. The system's usage is as follows: 1. User Identification: Upon first use, users scan the QR code on the plate using the app to complete the binding process. Afterward, the plate automatically verifies the user's identity via facial recognition using the front-facing camera (or via Bluetooth connection to a mobile phone) and downloads personalized data from the cloud, including the user's metabolic classification (level 1-5), current intervention mode (severe / mild), and five-point physiological rhythm thresholds. Multiple users can be automatically switched.

[0098] 2. Food recognition and weighing: After the user places the food into the corresponding area, the weighing sensor measures the weight increment of each area in real time.

[0099] The top camera automatically captures images of food, identifies the type of food (such as rice, chicken breast, broccoli) through a local lightweight AI model (or uploads to the cloud for recognition), and estimates its volume / weight (which is then fused and verified with weighing data).

[0100] If the recognition confidence level is low, the user will be asked via voice, and the user can provide further confirmation via voice or touchscreen.

[0101] By fusing camera image recognition and weighing sensor data, the system automatically determines the type, weight, and volume of food, and uses voice interaction to supplement the recognition, thereby improving accuracy.

[0102] 3. Carbon-to-Egg Ratio Calculation: The chip calls the carbon-to-egg ratio algorithm, and combines correction factors such as user metabolic classification, kidney function (eGFR), and BMI to calculate the target carbon-to-egg volume ratio for the current meal.

[0103] At the same time, refer to the five physiological rhythms: if it is breakfast time, consider the risk of dawn phenomenon; if it is dinner time, consider the risk of nocturnal hypoglycemia and automatically adjust the proportion (such as increasing the proportion of protein in dinner to delay absorption).

[0104] The dynamic carbon-to-protein ratio calculation method based on user metabolic classification and physiological rhythm, combined with cloud-synchronized user profiles (metabolic classification, kidney function, BMI, current intervention mode), adjusts the target ratio for each meal in real time.

[0105] 4. Real-time feedback and guidance: Compare the actual weight of each area with the target ratio and provide immediate guidance via voice broadcast, for example:

[0106] "The protein content is sufficient, but the carbohydrate content needs to be increased by 30 grams."

[0107] "The refined rice you chose raises blood sugar quickly. We recommend replacing it with mixed grain rice or increasing the protein content."

[0108] If the user is currently in heavy intervention mode, the guidance tone will be more stringent; in light intervention mode, it will be more gentle.

[0109] As the user places food, the system proactively broadcasts guidance and suggestions based on real-time weighing data compared to the target ratio, achieving an interactive experience of "adjusting as you place".

[0110] 5. Intervention mode switching and blood glucose monitoring reminders: The system provides differentiated blood glucose monitoring guidance based on the user's current intervention mode (heavy intervention / light intervention).

[0111] The system automatically detects the status of the user's blood glucose monitoring device (CGM online / offline, BGM availability) and automatically switches between a heavy intervention mode (providing intensive guidance by linking real-time CGM data) and a light intervention mode (providing appropriate reminders based on five physiological rhythms, compatible with CGM and BGM data sources) to achieve continuous intervention in different scenarios.

[0112] The system can optionally be linked with a continuous glucose monitoring device (CGM) or a fingertip blood glucose meter (BGM) to obtain real-time or historical blood glucose data through the cloud and dynamically adjust dietary recommendations; it also supports recognizing BGM and CGM values ​​by taking pictures with a camera and automatically recording the user's blood glucose data.

[0113] 5.1 When the user wears CGM, it is in heavy intervention mode: The system automatically obtains real-time CGM data through the cloud and dynamically adjusts the carbohydrate-to-protein ratio recommendation based on the user's current blood glucose level and blood glucose trend (rate of increase / decrease) during meals. For example, if pre-meal blood sugar is already high (>7.8mmol / L), the system will prompt with a voice message: "Current blood sugar is high. It is recommended to reduce the total carbohydrate intake for lunch today by 20% and prioritize protein intake."

[0114] Two hours after a meal, CGM data is automatically compared to assess the glycemic impact of the meal, which is then used for algorithm optimization.

[0115] 5.2 The light intervention mode is applicable to the following scenarios: the user has not bound any blood glucose monitoring device; the user has bound a CGM device but is not currently wearing it (the device is offline); the user is wearing a CGM device but the system determines that the current situation is "low-intensity intervention" (if the user's blood glucose is stable for a long period of time, the system will actively downgrade to reduce disturbance).

[0116] If the user has a usable CGM device (worn and online): the system prioritizes using real-time CGM data for light intervention guidance, but the frequency of reminders is reduced (to avoid intensive guidance in heavy intervention mode), and voice reminders are only triggered when postprandial blood glucose rises abnormally: "Your postprandial blood glucose is high, it is recommended to reduce carbohydrates in the next meal"; If the user has a CGM device but does not wear it, or only has a BGM device: the system will remind the user to take measurements according to the five-point rhythm and record the values ​​through photo recognition or voice input; If the user does not have any blood glucose monitoring device: the system only provides dietary guidance based on metabolic classification and does not provide blood glucose-related reminders, but still reminds the user of "blood glucose monitoring is recommended" as a health reminder through five-point rhythm; The strategy for controlling the frequency of blood glucose monitoring adopts the principle of "the worse the condition, the more frequent the monitoring," as detailed below: (1) Weekly monitoring frequency setting: The system divides users into three levels based on their historical blood glucose control level (based on CGM data or BGM records from the past 7-14 days):

[0117] User tiers are dynamically updated weekly, and the system reassesses and adjusts the target number of monitoring sessions for the week based on data from the past 7 days every Monday.

[0118] (2) The dynamic allocation strategy adopts the principle of "the worse the quality, the denser the allocation", as follows: Within the set weekly limit, the system dynamically allocates monitoring frequency to the "most important" time points based on the user's historical blood glucose performance at various time points: Node Risk Scoring: The system calculates a risk score for each of the five circadian rhythm nodes (morning fasting, after breakfast, after lunch, after dinner, after exercise, and at night). Risk score = Historical frequency of blood glucose exceedances × Exceedance magnitude coefficient

[0119] For example: If a user has experienced 3 instances of elevated blood glucose levels (>10 mmol / L) after breakfast in the past 7 days, with an average elevation of 2.5 mmol / L, then the risk score for this event is 3 / 7 × (2.5 / 10) = 0.107. Allocation by score: Allocate the available monitoring times (3-5 times) this week to the corresponding nodes in descending order of risk score. Allocation example: If a user has the highest risk after breakfast, followed by after dinner, and then on an empty stomach, and there are 4 monitoring sessions scheduled for this week, the system will allocate the following times: 2 times after breakfast, 1 time after dinner, and 1 time on an empty stomach. The distribution of reminder times actually received by users will reflect the principle of "testing more where the problem is".

[0120] (3) Node convergence and dynamic adjustment: If a high-risk node shows that blood glucose levels are within the target range for two consecutive weeks (all monitoring values ​​of the node are within the target range), the system will automatically reduce the risk score of the node and redistribute the monitoring times to other nodes that still have risks. If blood glucose control continues to improve at a certain point, the frequency of reminders for that point will gradually decrease until it is classified as "maintenance monitoring" (1-2 times per month). The system regularly provides feedback to users: "You completed 4 blood glucose monitoring sessions this week, focusing on post-breakfast and post-dinner blood glucose levels. Compared to last week, the post-breakfast blood glucose target achievement rate has increased from 40% to 75%. Next week, we will reduce post-breakfast monitoring and increase fasting monitoring." (4) User-defined and exemption mechanism: Users can manually adjust the weekly monitoring target through the App (within the range of ±1 times of the system's recommended value), or specify the nodes that they wish to focus on monitoring; If a user is unable to complete the number of monitoring sessions recommended by the system due to special circumstances (such as illness or travel), they can manually set a "temporary exemption period" (maximum 7 days). During this period, the system will suspend monitoring reminders and automatically resume them after the exemption period ends. If a user's completion rate is less than 50% of the recommended number of times for two consecutive weeks, the system will automatically upgrade the user to the next monitoring level (e.g., from moderate control to poor control) and increase the intensity of reminders. The system sets the total number of weekly monitoring sessions (3-5 times) based on the user's historical blood glucose control level, and dynamically allocates the number of monitoring sessions according to the historical risk scores of each of the five rhythm nodes. It prioritizes the allocation of limited monitoring resources to the period with the worst blood glucose control, realizing the precise monitoring strategy of "testing more where the blood glucose is worse". It also supports dynamic adjustment after node convergence and user-defined exemption mechanisms.

[0121] 5.3 Automatic switching of intervention mode: The system detects whether the user has bound a valid CGM device via the cloud (data uploaded in the last 24 hours): If the CGM device is detected to be online, it will automatically switch to heavy intervention mode and use CGM data to provide refined guidance.

[0122] If the CGM device is offline for more than 24 hours, it will automatically switch to the light intervention mode. When switching, the user will be informed by voice: "The CGM device is offline. Switching to light intervention mode. Please press the reminder to test your blood glucose."

[0123] Users can also manually specify the intervention mode through the App.

[0124] 6. Data Upload and Linkage: After each meal, the plate automatically uploads the food data (food type, weight, carbohydrate-to-protein ratio, and time) to the cloud to update the user's food records.

[0125] The cloud-based system combines user CGM / BGM data to assess postprandial blood glucose response and dynamically optimize subsequent recommendations (reinforcement learning).

[0126] If an anomaly is detected (such as exceeding the limit for multiple consecutive meals), the cloud can trigger an intervention mode switch (such as switching from mild intervention to severe intervention) and remind the user via voice prompts on the plate.

[0127] This invention aims to provide an intelligent dietary system that embodies the management logic of five-point monitoring, dual-module intervention, and a carbon-to-protein ratio algorithm, and can be directly deployed in the user's home, offering the following advantages: 1) Integrated design: Weighing, recognition, calculation and feedback are integrated into the plate itself, eliminating the need for users to operate their mobile phones. The process is natural and smooth, which is more in line with the dining table scenario.

[0128] 2) Easy to operate: Users only need to put the food into the corresponding compartment, and the dietary system will automatically complete the weighing, identification, calculation and feedback, without any learning cost.

[0129] 3) Real-time dynamic guidance: Provides immediate feedback during the user's food placement process, rather than recording it afterward, and can correct deviations on the spot to ensure that every meal is executed accurately. 4) Real-time interactive closed loop: The dietary system's built-in chip communicates with the cloud-based App in real time, automatically records dietary data, and dynamically adjusts the recommended ratio based on the user's metabolic capacity classification and real-time blood glucose (through CGM / BGM linkage), providing instant guidance through voice dialogue.

[0130] 5) Precise execution: Through zoned plates, high-precision weighing, and image recognition, the carbon-to-protein ratio is ensured to be executed precisely, transforming the complex management of macronutrients into a simple action of "just put it in the right place".

[0131] 6) Last-mile delivery: Make the smart diet system a digital nutritionist in the user's home, providing 24 / 7 personalized service, and completely break down the physical barrier from software to the user's home.

[0132] The above embodiments are not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the technical solution of the present invention are also within the protection scope of the present invention.

Claims

1. A metabolic management based intelligent meal system, characterized in that: It includes a data acquisition unit, a cloud service unit, a user interaction unit, and an external ecosystem connection unit; The data acquisition unit is used to collect users' basic information and dining records. The cloud service unit is used to dynamically assess a user's current blood glucose status and dietary restriction needs based on the Dietary Guidelines for Residents and individual user data. The user interaction unit is used to provide an intuitive and user-friendly interface, supporting users to view recommended plans, provide feedback on implementation status, and adjust dietary preferences. The external ecological connection unit is used to obtain real-time or historical blood glucose data from users via the cloud, and to dynamically adjust dietary recommendations.

2. The metabolic management based smart meal system as claimed in claim 1, wherein: The data acquisition unit includes a processing module, a weighing sensor, a communication module, a voice interaction module, and a camera module; The processing module is responsible for receiving multimodal data collected by the weighing sensor and camera, calling the carbon-egg ratio algorithm for processing, and feeding the results back to the user through the voice interaction module and the display screen, while synchronizing with the cloud through the communication module. The camera module uses dual cameras: the top camera shoots vertically downwards at the plate for food recognition, volume estimation, and plate area segmentation; the front camera faces the user and is used for face recognition, user proximity detection, and voice interaction assistance. The voice interaction module is used to provide real-time voice guidance to the user, provide feedback on the guidance from the processing module to the user, and interact with the user.

3. The intelligent dietary system based on metabolic management according to claim 1, characterized in that: The cloud service unit adopts initial suggestions based on population statistics during the initial user phase. As the user accumulates diet-glucose response data during the usage process, it dynamically updates metabolic classification and personalized parameters, and achieves accurate suggestions through convergence judgment and anomaly detection. The cloud service unit is equipped with a carbon-egg ratio algorithm and a reinforcement learning model. By analyzing the dietary records provided by users, a complete user profile is first generated, and then the target ratio for each meal is adjusted in real time. The cloud service unit pushes food purchasing suggestion reports to users through the user interaction unit, and users provide feedback on their actual purchases in the user interaction unit.

4. The intelligent dietary system based on metabolic management according to claim 1, characterized in that: The cloud service unit automatically adjusts the reminder level based on user usage frequency and abnormal status, and dynamically controls the reminder frequency through user response behavior. At the same time, it utilizes social support networks to improve user compliance. The reminder levels include in-device reminders, app push notifications, third-party assistance reminders, and manual follow-ups.

5. A method for a smart dietary system based on metabolic management as described in claim 1, characterized in that: Includes the following steps: Step S1, User Identification: Upon first use, the user completes the binding process by scanning a QR code; Step S2, Food Recognition and Weighing: After the user places the food into the corresponding area, the weighing sensor measures the weight increment of each area in real time. Through camera image recognition and weighing sensor data fusion, the food type, weight and volume are automatically determined, and voice interaction is used to supplement the recognition. Step S3, Carbon-to-Egg Ratio Calculation: The processing module calls the carbon-to-egg ratio algorithm, and calculates the target carbon-to-egg volume ratio for the current meal by combining the user's metabolic classification, kidney function, and BMI correction factor. Step S4, Real-time Feedback and Guidance: Compare the actual weight of each area with the target ratio and provide real-time guidance through voice broadcast; Step S5, Intervention Mode Switching and Blood Glucose Monitoring Reminder: The system provides differentiated blood glucose monitoring guidance based on the user's current intervention mode.

6. The method of the intelligent dietary system based on metabolic management according to claim 5, characterized in that: In step S1, the front-facing camera automatically confirms the user's identity through facial recognition or by connecting to a mobile phone via Bluetooth, and downloads the user's metabolic classification, current intervention mode, and five-point physiological rhythm threshold from the cloud; it can automatically switch between multiple users.

7. The method of the intelligent dietary system based on metabolic management according to claim 5, characterized in that: In step S2, the top camera automatically captures images of the food, identifies the type of food through a local lightweight AI model or by uploading to the cloud, estimates its volume and weight, and verifies it by fusing it with the weighing data. If the recognition confidence is low, the user will be asked via voice, and the user can then confirm via voice or touchscreen.

8. The method of the intelligent dietary system based on metabolic management according to claim 7, characterized in that: In step S2, food identification includes the following steps: Step S21, Region Segmentation: Use a lightweight semantic segmentation model to separate the plate region in the image from the background and identify the boundaries of each region; Step S22, Food Area Detection: Within each partition, independent food blocks are identified through edge detection and connected component analysis; if two food items are stacked in the same partition, the user is prompted to separate the food items. Step S23, Food Classification: For each food piece, call the image classification model; the model pre-training includes a self-built dataset of common Chinese ingredients; Step S24, Volume and Weight Estimation: Combining the pixel area of ​​the food block in the image, the camera focal length, and the known size of the plate, estimate the volume of the food. Simultaneously, receive the actual weight of the area from the weighing sensor and call the multimodal fusion algorithm. Based on the food category, query the preset density library and calculate the estimated image weight = estimated image volume × density; The weighted average is calculated with the actual weight measured by the weighing sensor, and the weights are dynamically adjusted according to lighting conditions and the complexity of the food's shape; if the difference between the two is >30%, a voice inquiry is triggered. Step S25, Confidence Assessment and Interactive Confirmation: If the classification confidence is <80% or the image conflicts with the weighing sensor data, the user will be asked via voice; the user can answer via voice or manually select via touchscreen. The system uses the user's confirmation as a new sample and uploads it to the cloud for subsequent model optimization.

9. The method of the intelligent dietary system based on metabolic management according to claim 5, characterized in that: In step S5, the intervention modes include: In the heavy intervention mode, the system automatically obtains real-time CGM data through the cloud, dynamically adjusts the carbon-to-protein ratio recommendation based on the user's current blood glucose level and blood glucose trend during the meal, and automatically compares the CGM data 2 hours after the meal to assess the blood glucose impact of the meal for algorithm optimization. The light intervention mode includes the following scenarios: When a user has not linked any blood glucose monitoring device, the system only provides dietary guidance based on metabolic classification and does not provide blood glucose reminders, but still reminds the user to monitor blood glucose as a health reminder through the five-point rhythm. When a user binds a CGM device but is not currently wearing it, the system reminds the user to take measurements according to the five-point rhythm and records the values ​​through photo recognition or voice input. When a user wears a CGM device but the system determines that the current period is a low-intensity intervention period, the system prioritizes using real-time CGM data for mild intervention guidance and only triggers voice reminders when postprandial blood glucose levels rise abnormally.

10. The method of the intelligent dietary system based on metabolic management according to claim 5, characterized in that: In step S5, the system automatically detects the status of the user's blood glucose monitoring device and automatically switches between heavy intervention mode and light intervention mode. After each meal, the processing module automatically uploads the food data, including food type, weight, carbohydrate-to-protein ratio, and time, to the cloud to update the user's food records. The cloud-based system combines user CGM / BGM data to assess postprandial blood glucose response and dynamically optimize subsequent recommendations. If an anomaly is detected, the intervention mode will be switched via the cloud and the user will be notified via voice prompts on the dining tray.

Images