Intelligent refrigerator food material management system and method based on AI large model

Abstract

The application discloses an intelligent refrigerator sticker food material management system based on an AI large model, which comprises an intelligent refrigerator sticker, a mobile phone APP and a cloud end; the intelligent refrigerator sticker comprises an SRAM storage and computing integrated chip, a Bluetooth communication module, an LED indicator lamp and a buzzer; the mobile phone APP comprises a food material recognition model and a recipe recommendation engine, the mobile phone APP communicates with the intelligent refrigerator sticker through Bluetooth, sends control instructions, receives recognition results, and is responsible for the collection of food material images and the management of user preference data; the cloud end comprises an AI large model and a food material knowledge base, the AI large model provides intelligent services including food material attribute analysis and recipe generation; the food material knowledge base is used for storing massive food material information; the cloud end and the mobile phone APP perform data interaction through the Internet; the application utilizes an end-cloud fusion architecture, embeds an embedded AI chip in the intelligent refrigerator sticker, realizes the rapid recognition, attribute extraction and state monitoring of food materials, and effectively solves the calculation time delay problem in a low-power consumption scene.

Description

Technical Field

[0001] This invention relates to the fields of smart home and artificial intelligence, specifically to a smart refrigerator magnet food management system and method based on an AI big data model. Background Technology

[0002] With the development of artificial intelligence technology, introducing AI capabilities into home appliances has become an important development direction in the smart home field. Most existing smart refrigerators manage food through barcodes, QR codes, etc., but the recognition accuracy is not high and they cannot obtain detailed attributes of the food.

[0003] Meanwhile, existing technologies lack a personalized understanding of user preferences and cannot provide intelligent recipe recommendation services. Therefore, there is an urgent need for an intelligent refrigerator magnet system based on AI big data models to improve the level of intelligence in food management. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides an intelligent refrigerator magnet food management system and method based on an AI large model. The system uses intelligent refrigerator magnets as physical carriers and utilizes built-in embedded AI models and Bluetooth communication modules to work in collaboration with a mobile APP to achieve intelligent services such as accurate identification of food, attribute extraction, shelf-life warning, and recipe recommendation, thereby significantly improving the user experience.

[0005] The technical solution of the present invention is as follows:

[0006] The AI-based smart refrigerator magnet food management system adopts an integrated edge-cloud architecture, including smart refrigerator magnets, mobile APP and cloud.

[0007] The smart refrigerator magnet includes an SRAM in-memory computing chip, a Bluetooth communication module, an LED indicator, and a buzzer. The SRAM in-memory computing chip integrates an embedded AI model and a local food knowledge graph, enabling the extraction and preliminary identification of food image features locally. The Bluetooth module is used to communicate with a mobile app, receive control commands, and upload identification results. The LED indicator displays the freshness of the food, and the buzzer emits a notification sound.

[0008] The mobile app communicates with the smart refrigerator magnet via Bluetooth, sends control commands, receives recognition results, and is responsible for collecting food images and managing user preference data.

[0009] The cloud platform includes a large AI model and a food knowledge base. The large AI model provides intelligent services including food attribute analysis and recipe generation; the food knowledge base stores massive amounts of food information; and the cloud platform and the mobile app interact via the internet.

[0010] Furthermore, in the aforementioned edge-cloud integrated architecture, "edge" refers to the terminal, including mobile apps and smart refrigerator magnets; "edge" refers to edge computing, including the SRAM in-memory computing chip within the smart refrigerator magnet; and "cloud" refers to cloud computing, including cloud servers.

[0011] Furthermore, the input to the embedded AI model is an image of the food ingredients; the food ingredient image is converted into a low-dimensional semantic vector v = [v1, v2, ..., v...]. 64 ];in, i = 1, 2, ..., 64;

[0012] The low-dimensional semantic vector v is divided into different sub-vectors, each representing a different feature category:

[0013] v_texture=[v1, v2, ..., v 28 / / Texture and shape information

[0014] v_color_size=[v 29 v 30 , ..., v 52 / / Color and size information

[0015] v_other = [v 53 v 54 , ..., v 64 / / Other visual features;

[0016] Extract high-level feature information from the image and match it with the food attribute vectors in the food knowledge base; output the weight values ​​for food classification to achieve food recognition.

[0017] Furthermore, the local food knowledge graph organizes data using nodes and edges, facilitating rapid querying and reasoning;

[0018] Each node represents an ingredient, including ID, name, shelf life, nutritional information, and recommended recipes; represented by a vector as: N_i = [name_i, shelf_life_i, nutrition_i, recipes_i];

[0019] Edges are used to represent relationships between nodes, including cooking compatibility and substitution relationships: represented by an adjacency matrix A[i][j] = {1 if ingredients i and j are compatible, 0 otherwise}.

[0020] The completed knowledge graph is stored in a custom binary format for easy access and application later.

[0021] Furthermore, the AI ​​big model includes a large-scale language model (LLM) and a food knowledge graph. The LLM is used to extract attribute information of food ingredients, including taste, flavor, and nutritional value. The food knowledge graph is used to construct the matching rules and correlations between food ingredients.

[0022] Furthermore, the computational method of the intelligent refrigerator magnet food management system based on AI large model utilizes edge-cloud collaborative computing, that is, distributing computing tasks to edge devices and cloud servers, and is characterized by including the following steps:

[0023] Users place food items in front of the smart refrigerator magnet, triggering a mobile app to capture images via the camera. The edge device first preprocesses the images, extracts feature vectors, and calls an embedded AI model for preliminary classification to identify the types of food items. Subsequently, the edge device queries the local food knowledge graph to obtain basic attribute information of the food items.

[0024] Edge devices synchronize recognition results and related data to a mobile app via Bluetooth communication modules. The mobile app then uploads the data to the cloud, triggering the retrieval enhancement and RAG generation process. After receiving the data, the cloud system performs vectorization processing and conducts similarity retrieval in the knowledge graph to find relevant recipes and preservation suggestions.

[0025] Ultimately, personalized recommendations are generated in the cloud and returned to the mobile app for users to view.

[0026] Furthermore, the RAG retrieval mechanism includes the following steps:

[0027] 1) Data Upload: Users upload relevant data for food identification to the cloud via a mobile app;

[0028] 2) Vectorization: The uploaded data is processed and converted into a high-dimensional vector representation;

[0029] 3) Retrieval process: Use a vector database to retrieve the stored food knowledge graph data; find the most relevant nodes by calculating the similarity between the uploaded data vector and the vectors of each node in the food knowledge graph;

[0030] 4) Generate personalized suggestions: Generate personalized recipe suggestions and freshness reminders based on the information of the most relevant nodes.

[0031] The "large-scale AI model" in this invention refers to a large-scale language model, which is a neural network model trained on massive amounts of text data and possesses powerful natural language understanding and generation capabilities. Representative models include GPT-4, Wenxin Yiyan, and Tongyi Qianwen. These models, through self-supervised learning and transfer learning, can be widely applied to various natural language processing tasks. This invention utilizes the large-scale AI model to achieve intelligent services such as extracting ingredient attributes, generating recipes, and making recommendations.

[0032] This invention employs an embedding model to learn representations of food ingredient attributes. Embedding models are a type of representation learning model that maps discrete variables to continuous dense vectors, highlighting semantic similarity and finding wide application in recommender systems, semantic retrieval, and other fields. In this invention, the embedding model maps food ingredient attributes (such as name, shelf life, taste, and nutritional components) to low-dimensional dense vectors, facilitating subsequent matching, retrieval, and recommendation tasks.

[0033] The beneficial effects of this invention are as follows:

[0034] 1) By utilizing an edge-cloud converged architecture, an embedded AI chip is built into the smart refrigerator magnet to achieve rapid identification, attribute extraction, and status monitoring of food ingredients, effectively solving the problem of computing latency in low-power scenarios.

[0035] 2) By building an entry point for food image acquisition and user interaction through a mobile APP, and utilizing cloud-based AI big data models to achieve in-depth understanding of food information and intelligent personalized recipe recommendations, a "one-stop" food management solution is provided.

[0036] 3) A distributed food knowledge base management mechanism is adopted to achieve knowledge sharing and incremental updates between smart refrigerator magnets, mobile APP and cloud, which saves storage space and improves the intelligence level of the system.

[0037] 4) By comprehensively utilizing IoT, big data and AI technologies, using smart refrigerator magnets as the physical carrier and large-scale language models and knowledge graphs as the intelligence engine, an integrated "device-edge-cloud" food management system has been built, setting a new benchmark for the smart home field.

[0038] 5) The system has a high recognition accuracy, provides comprehensive and detailed descriptions of the attributes of ingredients, and offers highly personalized recipe recommendations, which helps improve users' healthy eating habits; in addition, the shelf-life warning function can significantly reduce food waste. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the system structure of the present invention;

[0040] Figure 2 This is a schematic diagram of the system flow of the present invention. Detailed Implementation

[0041] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0042] like Figure 1-2 As shown, this is an intelligent refrigerator magnet food management system based on an AI big data model.

[0043] System Architecture:

[0044] The intelligent refrigerator magnet food management system of this invention adopts an integrated "edge-cloud" architecture, mainly composed of three parts: intelligent refrigerator magnets, a mobile APP, and cloud services. "Edge" refers to the mobile APP and intelligent refrigerator magnets, "edge" refers to the embedded AI chip (SRAM in-memory computing chip) within the refrigerator magnet, and "cloud" refers to the cloud server deploying large-scale models. In this architecture, edge computing is mainly reflected in the extraction and matching of food image features by the intelligent refrigerator magnets. Compared with traditional centralized cloud computing, edge computing can significantly reduce data transmission bandwidth and latency, and alleviate the computational pressure on the cloud. Simultaneously, the collaboration between edge devices and the cloud enables an optimal combination of real-time processing on the edge and intelligent processing in the cloud.

[0045] The smart refrigerator magnet ("edge") is the core physical carrier of the system, with a built-in SRAM in-memory computing chip ("edge"), a Bluetooth communication module, and peripheral devices such as LED indicators and buzzers. The SRAM in-memory computing chip integrates an embedded AI model and a local food knowledge base, which can extract and preliminarily identify food image features locally. The Bluetooth module is used to communicate with a mobile APP, receive control commands, and upload recognition results. The LED indicator is used to display the freshness of the food, and the buzzer is used to emit a prompt sound.

[0046] The embedded AI model transforms food images captured by a mobile phone into low-dimensional semantic vectors, extracting high-level feature information from the images. This information can then be matched with food attribute vectors in a built-in knowledge base to achieve food recognition on the device. In this way, refrigerator magnets do not need to store the original image data, significantly reducing data transmission volume and storage overhead. Simultaneously, the extraction of semantic features provides a more refined and effective information representation for subsequent attribute matching and recognition tasks.

[0047] The steps for deploying an embedded AI model are as follows:

[0048] 1) Model Selection and Training: Select a base model (such as MobileNetV2) and fine-tune it on a specific seafood image dataset. Improve the model's accuracy and robustness in seafood classification through transfer learning and data augmentation techniques.

[0049] 2) Model quantization: The fine-tuned model is quantized and converted into a low-bit representation (such as int8 format). This reduces the model's storage space and computational resource requirements, adapting to the limitations of embedded devices.

[0050] 3) Embedded environment preparation: Install the necessary runtime environment and libraries (such as TensorFlowLite or ONNX Runtime) on the target device to support model loading and execution.

[0051] 4) Model Integration: Embed the quantized model file into the refrigerator magnet's storage system. Ensure the model can effectively interact with the device's processor and other hardware components.

[0052] 5) Real-time inference: Develop the client-side application to process user input (such as images captured by the camera) and call the fine-tuned model for real-time inference. This involves extracting feature vectors and performing classification through image preprocessing steps (such as scaling and normalization).

[0053] 6) Performance Optimization: Monitor the model's inference performance and further optimize it as needed. Techniques such as model pruning and knowledge distillation can be used to improve inference speed and reduce resource consumption. By fine-tuning the model deployment on the device side, smart refrigerator magnets can quickly and accurately identify food types locally, improving user experience and reducing reliance on the cloud, thus achieving more efficient food management.

[0054] Under the edge-cloud collaborative architecture, the food knowledge base adopts a distributed storage and management approach. The smart refrigerator magnet has built-in basic attributes of commonly used foods, the mobile APP stores user preference data, and the cloud has a massive amount of food encyclopedia information.

[0055] The local ingredient knowledge graph is a structured dataset designed to provide detailed information and relationships about ingredients; it organizes data through nodes and edges for easy querying and reasoning.

[0056] Steps for building a food knowledge base:

[0057] 1) Node definition:

[0058] Each node represents a type of ingredient and includes the following attributes:

[0059] ID: Unique identifier;

[0060] Name: The name of the seafood;

[0061] Shelf life: The storage time for this seafood;

[0062] Nutritional composition: Key nutritional information (such as protein, fat, vitamins, etc.);

[0063] Recommended recipes: Cooking recipes related to this seafood;

[0064] 2) Definition of edge relation:

[0065] Edges are used to represent the relationships between nodes, and mainly include:

[0066] Culinary compatibility: Indicates which ingredients can be cooked together to enhance the flavor of the dish.

[0067] Substitution Relationship: Provides alternative options for ingredients, helping users choose other suitable substitutes when a certain type of seafood is unavailable.

[0068] 3) Data storage:

[0069] The completed knowledge graph is stored in a custom binary format for easy access and application later.

[0070] The mobile app ("the app") is installed on the user's mobile phone, serving as the direct entry point for human-computer interaction. The app has a built-in food recognition model and recipe recommendation engine, and can access cloud-based AI model services. The app communicates with the refrigerator magnets via Bluetooth, sending control commands, receiving recognition results, and managing food image acquisition and user preference data.

[0071] In this embodiment, the food identification model adopts the Transformer architecture LLM large language model, which includes: multi-head attention mechanism to process food text description, positional encoding to maintain food feature sequence information, residual connection and layer normalization;

[0072] Specific implementation steps:

[0073] 1) Input food images are processed by a Convolutional Neural Network (CNN) to extract features and then converted into text descriptions; 2) The text is converted into vector representations through an embedding layer; 3) Deep semantic features are extracted through a multi-layer Transformer encoder; 4) Finally, food attribute labels are output through a classification head.

[0074] Recipe recommendation engine: Based on a hybrid recommendation AI algorithm, including: content-based feature extraction module, diversity re-ranking algorithm, and deep learning ranking module.

[0075] The cloud deploys large-scale AI models and a big data knowledge base of ingredients. The AI ​​models provide intelligent services such as ingredient attribute analysis and recipe generation. The knowledge base stores massive amounts of ingredient information, such as nutritional components, optimal storage methods, and cooking techniques. Data interaction between the cloud and the mobile app occurs via the internet. The cloud deploys large-scale AI models trained on massive amounts of ingredient data, including large-scale language models and ingredient knowledge graphs. The large-scale language models enable deep understanding and extraction of ingredient attribute information, such as taste, flavor, and nutritional value. The ingredient knowledge graph depicts the pairing patterns and relationships between ingredients. Based on the inference results of these models, the cloud provides personalized recipe recommendation services. The RAG (Retrieval-Augmented Generation) retrieval mechanism combines the advantages of information retrieval and generative models, aiming to provide users with personalized suggestions and information.

[0076] The system of this invention adopts edge-cloud collaborative computing:

[0077] Edge-cloud collaborative computing is a technology that distributes computing tasks between edge devices (smart refrigerator magnets) and cloud servers to optimize resource utilization and improve response speed. In practice, the user places seafood in front of the smart refrigerator magnet, triggering the phone's camera to capture an image. The edge device first preprocesses the image, extracting feature vectors, and then uses a locally fine-tuned model for preliminary classification, identifying the type of seafood. Subsequently, the edge device queries a local knowledge graph to obtain detailed attribute information about the ingredients, including shelf life and recommended recipes.

[0078] Next, the edge device synchronizes the recognition results and related data to the user's mobile app via Bluetooth. The mobile app then uploads this data to the cloud, triggering the Retrieval Augmentation (RAG) process. Upon receiving the data, the cloud system performs vectorization processing and conducts similarity searches within a knowledge graph to find relevant recipes and preservation suggestions. Finally, the cloud generates personalized recommendations and returns them to the mobile app for the user to view. Through this collaborative work between the edge and cloud, the system achieves efficient data processing and personalized services, enhancing the user experience. The smart refrigerator magnet synchronizes data with the user's mobile app via Bluetooth, utilizing Bluetooth Low Energy (BLE) as a key technology. Throughout this process, data security and privacy protection are paramount. Key technologies include data encryption and anonymization; during data transmission, encryption technology is used to protect user data and ensure privacy; and edge computing privacy protection is crucial.

[0079] The RAG retrieval mechanism described above is as follows:

[0080] 1) Data Upload: Users upload relevant data for seafood identification (such as category, shelf life, etc.) to the cloud via a mobile app.

[0081] 2) Vectorization: The uploaded data is processed and transformed into a high-dimensional vector representation. This process typically uses Natural Language Processing (NLP) technology to convert text information into a format that can be processed by computers.

[0082] 3) Retrieval Process: The stored knowledge graph data is retrieved using a vector database (such as FAISS or Annoy). The most relevant nodes are found by calculating the similarity between the uploaded data vector and the vectors of each node in the knowledge graph.

[0083] 4) Generate personalized suggestions: Once relevant nodes are found, the system will generate personalized recipe suggestions and preservation reminders based on the information from these nodes. This process incorporates Natural Language Generation (NLG) technology to transform the search results into easy-to-understand suggestions.

[0084] The "edge-cloud" architecture of this invention fully leverages the advantages of different computing layers: the edge device handles data collection and user interaction, the edge device performs local real-time computing, and the cloud provides powerful intelligent service support. Compared with traditional centralized cloud computing, the distributed computing paradigm of this system can significantly reduce data transmission bandwidth and latency, alleviating the computing pressure on the cloud. Simultaneously, the collaboration between edge and cloud enables a perfect combination of real-time interaction and intelligent decision-making, resulting in a smoother and more intelligent user experience. This architectural model represents a new trend in intelligent system design in the Internet of Things era.

[0085] The workflow of this invention is as follows:

[0086] 1) Adding Ingredients: The user places the ingredients in the refrigerator and presses and holds the switch button on the corresponding refrigerator magnet. The refrigerator magnet sends a photo request to the mobile app. The user takes a picture of the ingredients and uploads it through the app. The refrigerator magnet uses an embedded AI model to extract image features, matches them with local ingredient attributes, and returns the recognition results to the app via Bluetooth. At the same time, an LED indicator light illuminates, showing the freshness of the ingredients.

[0087] 2) Food Management: Users can view the types, quantities, and expiration dates of food in the refrigerator at any time via the app; when a food item is nearing its expiration date, a yellow light on the refrigerator magnet will illuminate and a reminder sound will be emitted at regular intervals. If the food item has expired, a red light will illuminate, and the app will send a notification message; after the user removes the food item, double-clicking the corresponding refrigerator magnet button will automatically update the food list.

[0088] 3) Recipe Recommendations: Users can set their personal dietary preferences in the app and record their meals over a period of time by taking photos. The phone periodically uploads the preference data to the cloud, where it is analyzed by an AI model to create a user taste profile. When a user needs to cook, the cloud, combined with real-time ingredients in the refrigerator, intelligently recommends delicious recipes that suit the user's taste.

[0089] 4) Knowledge Base Updates: The latest food encyclopedia information, such as new varieties and cooking methods, is regularly updated and synced to the mobile app via incremental updates from the cloud. Users can also add new food entries to refrigerator magnets through the app according to their needs. In addition, user-reported food recognition results are also uploaded to the cloud to optimize the AI ​​model.

[0090] Example 1: Smart Refrigerator Magnet for Seafood and Fish

[0091] 1. Hardware configuration:

[0092] SDRAM storage: High-speed CMOS synchronous DRAM 128MB RAM;

[0093] Communication: Bluetooth 5.1 BLE, 32-bit ARM Cortex-M0 core, maximum operating frequency of 64MHz; it integrates 48KB SRAM and 512KB Flash.

[0094] Display: None;

[0095] 2. Software Architecture:

[0096] a) Fine-tuning the model:

[0097] Base model: MobileNetV2 (quantized version);

[0098] Input dimensions: 224x224x3 (for reference only);

[0099] Output: Weight values ​​(1-5 points) for 100 common seafood species.

[0100] 64-dimensional feature vectors;

[0101] v = [v1, v2, ..., v] 64 ]

[0102] in, i = 1, 2, ..., 64

[0103] The low-dimensional semantic vector v is divided into different sub-vectors, each representing a different feature category:

[0104] v_texture=[v1, v2, ..., v 28 / / Texture and shape information;

[0105] v_color_size=[v 29 v 30 , ..., v 52 / / Color and size information; v_other = [v 53v 54 , ..., v 64 / / Other visual features (surface properties, state, comprehensive information);

[0106] b) Local knowledge graph:

[0107] Number of nodes: 100 common types of seafood and fish;

[0108] Each node has the following attributes: name, shelf life, nutritional information, and recommended recipes.

[0109] Edge relationships: culinary compatibility, ingredient substitution relationships;

[0110] Node representation: Each food item node can be represented as a vector:

[0111] N_i=[name_i, shelf_life_i, nutrition_i, recipes_i]

[0112] N_apple = ["Apple", 14, [52, 0.2, 13.8, 0.3], ["Apple Pie", "Fruit Salad"]]

[0113] Edge relationships can be represented using an adjacency matrix or an adjacency list.

[0114] Adjacency matrix example:

[0115] A[i][j] = {1, if ingredients i and j are compatible with 0, otherwise}

[0116] 3. Workflow:

[0117] Step 1: The user places the seafood in front of the refrigerator magnet, triggering image capture;

[0118] Step 2: Preprocess the image on the edge and extract feature vectors;

[0119] Step 3: Use the local fine-tuning model for initial classification;

[0120] Step 4: Query the local knowledge graph to obtain detailed attributes;

[0121] Step 5: Sync data with the mobile app via Bluetooth;

[0122] Step 6: The mobile app uploads the data to the cloud, triggering RAG vector retrieval;

[0123] Step 7: Personalized recipe suggestions and preservation reminders are returned via the cloud. The cloud system uses user-uploaded seafood data, combined with seafood characteristics and cooking knowledge from the database, to generate recommendations that meet user needs, enhancing user experience and improving the efficiency of ingredient utilization.

[0124] 4. Core Algorithm:

[0125] a) Fish freshness assessment:

[0126] Freshness is evaluated using an SVM classifier based on image color and texture features.

[0127] 5. Data Structures:

[0128] The model and knowledge graph are stored using a custom binary format.

[0129] 6. Example of cloud-edge-device collaborative computing:

[0130] Refrigerator magnet: Initial classification identified as "sea bass"

[0131] Mobile App: Performs high-precision image segmentation and extracts fish gill color features.

[0132] Refrigerator magnets: Based on color characteristics, their freshness is assessed as "excellent".

[0133] Cloud-based: Based on users' dietary habits, recommend a "Steamed Sea Bass" recipe.

[0134] 7. Edge computing and privacy protection:

[0135] Only the processed feature vectors and classification results are transmitted; the original images are not uploaded.

[0136] User dietary preferences are stored locally, while only anonymized data is stored in the cloud.

[0137] Through the implementation of the above technologies, this smart refrigerator magnet for seafood can quickly and accurately identify and manage seafood ingredients while protecting user privacy, and provide personalized preservation and cooking suggestions, thereby improving users' food management efficiency and dietary quality.

Claims

1. An intelligent refrigerator magnet food management system based on an AI big data model, employing an integrated edge-cloud architecture, characterized by: This includes smart refrigerator magnets, mobile apps, and cloud services; The smart refrigerator magnet includes an SRAM in-memory computing chip, a Bluetooth communication module, an LED indicator, and a buzzer. The SRAM in-memory computing chip integrates an embedded AI model and a local food knowledge graph, enabling the extraction and preliminary identification of food image features locally. The Bluetooth module is used to communicate with a mobile app, receive control commands, and upload identification results. The LED indicator displays the freshness of the food, and the buzzer emits a notification sound. The mobile app communicates with the smart refrigerator magnet via Bluetooth, sends control commands, receives recognition results, and is responsible for collecting food images and managing user preference data. The cloud and the mobile app interact via the internet. The cloud includes an AI big model and a food knowledge base. The AI ​​big model provides intelligent services including food attribute analysis and recipe generation. The food knowledge base is used to store massive amounts of food information. In the aforementioned edge-cloud integrated architecture, "edge" refers to the terminal, including mobile apps and smart refrigerator magnets; "edge" refers to edge computing, including the SRAM in-memory computing chip within the smart refrigerator magnet; and "cloud" refers to cloud computing, including cloud servers. Utilizing edge-cloud collaborative computing, i.e., distributing computing tasks to edge devices and cloud servers, includes the following steps: Users place food items in front of the smart refrigerator magnet, triggering a mobile app to capture images via the camera; the edge device first preprocesses the images, extracts feature vectors, and calls an embedded AI model for preliminary classification to identify the types of food items. Subsequently, the edge device queries the local food knowledge graph to obtain basic food attribute information; Edge devices synchronize recognition results and related data to a mobile app via Bluetooth communication modules. The mobile app then uploads the data to the cloud, triggering the retrieval enhancement and RAG generation process. After receiving the data, the cloud system performs vectorization processing and conducts similarity retrieval in the knowledge graph to find relevant recipes and preservation suggestions. Ultimately, personalized recommendations are generated in the cloud and returned to the mobile app for users to view.

2. The intelligent refrigerator magnet food management system based on an AI large model as described in claim 1, characterized in that, The embedded AI model takes food images as input and converts them into low-dimensional semantic vectors v = [v1, v2, ..., v...]. 64 ];in, The low-dimensional semantic vector v is divided into different sub-vectors, each representing a different feature category: v_texture=[v1, v2, ..., v 28 Texture and shape information; v_color_size=[v 29 v 30 , ..., v 52 Color and size information; v_other = [v 53 v 54 , ..., v 64 Other visual features; Extract high-level feature information from the image and match it with the food attribute vectors in the food knowledge base; output the weight values ​​for food classification to achieve food recognition.

3. The intelligent refrigerator magnet food management system based on an AI large model according to claim 1, characterized in that, The local food knowledge graph organizes data using nodes and edges, facilitating rapid querying and reasoning. Each node represents an ingredient, including ID, name, shelf life, nutritional information, and recommended recipes; represented by a vector as: N_i = [name_i, shelf_life_i, nutrition_i, recipes_i]; Edges are used to represent relationships between nodes, including cooking compatibility and substitution relationships: represented by an adjacency matrix A[i][j] = {1 if ingredients i and j are compatible, 0 otherwise}. The completed knowledge graph is stored in a custom binary format for easy access and application later.

4. The intelligent refrigerator magnet food management system based on an AI large model according to claim 1, characterized in that, The AI ​​big model includes a large-scale language model (LLM) and a food knowledge graph. The LLM is used to extract the attribute information of food ingredients, and the food knowledge graph is used to construct the matching rules and correlations between food ingredients.

5. The calculation method for the intelligent refrigerator magnet food management system based on an AI large model according to claim 1, characterized in that, The RAG retrieval mechanism includes the following steps: 1) Data Upload: Users upload relevant data for food identification to the cloud via a mobile app; 2) Vectorization: The uploaded data is processed and converted into a high-dimensional vector representation; 3) Retrieval process: Use a vector database to retrieve the stored food knowledge graph data; find the most relevant nodes by calculating the similarity between the uploaded data vector and the vectors of each node in the food knowledge graph; 4) Generate personalized suggestions: Generate personalized recipe suggestions and freshness reminders based on the information of the most relevant nodes.

Images