Campus refrigerator sharing method based on federal learning and corollary equipment and application program thereof
Through federated learning model pool technology, local model training and global knowledge gathering are carried out in campus shared refrigerators, which solves user privacy and data security issues, realizes intelligent management of diversified and personal needs of the refrigerator system, and improves system efficiency and users Experience.
Patent Information
- Application Number
- CN202410916998.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-09
- Publication Date
- 2024-11-01
AI Technical Summary
The existing campus shared refrigerator technology has user privacy and data security problems, the risk of personal information leakage is high, and the ability to expand and use data resources is insufficient, and the smart management and service prediction functions of diversified needs and personal needs have not been realized.
The federated learning model pool technology is introduced to protect user privacy through local model training, and only transmit model parameters to reduce communication overhead, realize global knowledge aggregation and model collaborative learning, and build a smart management and service prediction system.
Effectively protect user privacy, reduce data leakage risks, improve system efficiency and robustness, achieve the satisfaction of diverse and personal needs, and improve the intelligent management and service prediction capabilities of refrigerator systems.
Smart Images

Figure CN118886992A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the fields of electronic technology, communication technology and large model technology, and relates to the configuration, installation, debugging and optimization of control, communication, security, signal detection and large model equipment and software, and specifically relates to a campus shared refrigerator method based on federated learning and its supporting equipment and application programs. Background Art
[0002] In today's era, refrigerators have become a very popular household appliance. With the diversification and personalization of contemporary university campus life, storing raw vegetables and fruits, as well as foods with a short shelf life such as bread and fresh milk in school has become a need for many college students. Correspondingly, college students are increasingly dependent on refrigerators, so shared refrigerators have emerged in contemporary campuses.
[0003] The existing campus shared refrigerator is an innovative model that uses the Internet of Things technology and cloud computing platform to provide students with convenient, safe, and energy-saving food storage and exchange services. Students can use the mobile phone application to scan the QR code on the refrigerator to realize the operation of the refrigerator, temperature adjustment, food storage and retrieval, etc. At the same time, the refrigerator will automatically record the usage of each user and upload it to the cloud to form a big data analysis system for optimizing the operation management and resource allocation of the refrigerator. Driven by the Internet of Things technology and cloud computing platform, campus shared refrigerators have become an innovative model that provides students with convenient, safe, and energy-saving food storage and exchange services. However, with the development of technology, campus shared refrigerators are also facing a series of challenges involving the expansion of refrigerator data value applications, user privacy, and data security.
[0004] Through the application, students can remotely control the refrigerator at any time, which improves the convenience of use while also raising concerns about user privacy and data security. The cameras and sensors on the refrigerator may leak sensitive personal information, and remote storage and transmission are at risk of being attacked by hackers. For example: the cameras and sensors on the refrigerator may leak the user's personal information, such as facial features, ID number, mobile phone number, etc., causing unnecessary trouble and risks to the user. Remote data storage and transmission may be attacked or stolen by hackers, resulting in the leakage or abuse of sensitive data such as user usage records, food preferences, and health status. The refrigerator's data analysis system may infringe on the user's privacy rights, such as making personalized recommendations, behavioral analysis, and consumption guidance for users, affecting the user's independent choice and judgment.
[0005] Although the shared refrigerator system is recorded in the Chinese invention patent "A shared refrigerator system and its interactive method" (publication number CN110009823A), including a mobile terminal, an enterprise cloud platform, a PC management terminal, an operator IOT platform, a NB-IOT device, an electric control board and a refrigerator; the enterprise cloud platform is respectively connected to the mobile terminal, the PC management terminal and the operator IOT platform, and the operator IOT platform is connected to the electric control board through the NB-NOT device. The present invention uploads the information of the refrigerator to the enterprise cloud platform through the PC management terminal, and the user obtains the deposit and rental information of the refrigerator uploaded to the enterprise cloud platform through the mobile terminal, and pays the deposit and rental. After the payment is successful, the operator IOT platform sends a control instruction to the electric control board through the BN-IOT device to control the refrigerator to work, and the enterprise cloud platform counts whether the rental time has expired. Once it expires and the fee has not been paid, the operation of the refrigerator is stopped, which greatly facilitates the user to start the refrigerator at any time, meets the temporary user's demand for the refrigerator, realizes the refrigerator sharing mode, is simple to operate, and improves the user's interactive experience.
[0006] However, the existing shared refrigerator technology has the following shortcomings: there is no centralized model analysis application pool, the ability to expand and apply refrigerator usage data resources is insufficient, and personal information is not well protected when multiple people share the refrigerator. The personal information of people who rent and use the refrigerator will be stored on remote and local servers, and will be transmitted on the shared refrigerator network, which can easily lead to the exposure and leakage of personal privacy. It mainly solves the real-time management and service of refrigerator storage and consumption, and does not truly realize the intelligent management and service prediction functions of the shared refrigerator system with multiple demands, personalized demands, and multiple transactions under the large model.
[0007] It is necessary to protect the privacy and data security of users and expand the use value of refrigerator data while ensuring the normal operation of refrigerator functions. This is a very challenging difficulty. Traditional encryption or anonymization methods often increase computational complexity and communication overhead in the IoT scenario, reducing system efficiency and performance. Federated Learning is a new paradigm of encrypted distributed machine learning that has emerged in recent years. It allows all participants to collaborate on AI without leaving the local data, achieving "knowledge sharing without data sharing" and improving the effectiveness of their respective AI models. It is the most effective solution to the two "bottleneck" problems of "data islands" and "data privacy protection" in the current AI industry. Therefore, how to introduce federated learning model pool technology to build a campus shared refrigerator is one of the research and development directions in this field. Summary of the invention
[0008] In response to the above problems, the present invention provides a campus shared refrigerator method based on federated learning.
[0009] The introduction of federated learning model pool technology can solve the above problems.
[0010] First, federated learning technology protects user privacy to a great extent by training local models on local nodes, because personal information and sensitive data do not need to be transmitted to the cloud, reducing the risk of data leakage. Secondly, the original data is not transmitted on the network, and federated learning significantly reduces the possibility of being hacked or stolen, improving data security. Then, by transmitting only model parameters instead of huge raw data, the technology reduces communication overhead and latency, thereby improving the overall efficiency of the system, making IoT applications such as campus shared refrigerators more real-time and responsive.
[0011] Second, in addition, local nodes can dynamically adjust model parameters according to their own data characteristics and environmental changes, and update synchronously with the cloud, which enhances the robustness and scalability of the system. Most importantly, federated learning achieves the convergence of global knowledge through the collaborative training of multiple local nodes, promotes the collaborative learning of models, and provides a more comprehensive data foundation for the intelligent analysis and optimization of the system's model pool.
[0012] Third, with the help of current big data technology, learning models for multi-level demands such as multiple demands, individual demands, and different transaction behaviors are introduced on the federated learning side. After continuous iteration and updating, a federated learning model pool is built to realize the intelligent management of the shared refrigerator system and the prediction of different service demands.
[0013] The specific technical solution is: a campus shared refrigerator method based on federated learning, including:
[0014] 1) Model initialization:
[0015] a. Set up various appropriate models according to the needs of shared refrigerator management scenarios;
[0016] b. Set appropriate conditions and requirements for model training, interaction, and updating;
[0017] c. Initialize various models in the cloud model pool;
[0018] 2) Federated learning training steps:
[0019] Step 1: Data segmentation and local training: Each campus shared refrigerator is regarded as a local node. The local node device obtains the latest comprehensive global model parameters from the remote server and uses local data for model training, including user's switch, temperature adjustment, food storage and retrieval operation records, usage habits, etc. Then, distributed data segmentation technology is used to ensure that each local node has enough samples to train the model while complying with the privacy protection principle.
[0020] Step 2: Upload model parameters to the cloud: After training on the local node, only upload model parameters instead of original data to the cloud, using secure transmission protocols and end-to-end encryption technology to ensure secure transmission of model parameters and data security;
[0021] Step 3: Cloud model parameter aggregation: Federated learning weighted algorithm technology is used in the cloud to aggregate the model parameters uploaded by all local nodes into a global model;
[0022] Step 4: Send the global model to each local node: Send the global model parameters generated in the cloud to each local node, and use security protocols to ensure the secure transmission of model parameters and data security;
[0023] Step 5: Regular updates and collaborative learning: Set up a regular federated learning execution plan to ensure that the system can learn new information from data on all nodes.
[0024] Step 6: Monitoring and maintenance: Establish a monitoring system to regularly check the operating status and model performance of each local node, handle any abnormal situations that may arise in a timely manner, and ensure the stable operation of the entire federated learning system.
[0025] 3) Global model aggregation, update and use of remote model pool:
[0026] a. Each time a new round of model parameter aggregation is performed, it is necessary to evaluate whether the local node model is a malicious model or an unreliable model. Malicious models and unreliable models need to be distinguished;
[0027] b. Malicious and unreliable models are eliminated, and the remaining local node models will be aggregated to form a new global model;
[0028] c. The new global model is added to the model pool, and authorized users can view and use the corresponding model through the mobile application;
[0029] d. For user use and transfer models, facial recognition registration and contract signing are carried out online through mobile phone applications to ensure the compliance and legal transfer of ownership;
[0030] 4) Maintenance and optimization of model pool:
[0031] a. Based on actual application conditions, regularly supplement and optimize the corresponding models in the model pool to improve model accuracy and enhance user stickiness and experience of using refrigerators;
[0032] b. Continuously maintain and upgrade the shared refrigerator system to meet technological development and users’ personalized and multi-level needs.
[0033] Furthermore, differential privacy technology is introduced in step three of the federated learning training procedure to enhance user privacy protection and ensure that individual data is not leaked during the aggregation process.
[0034] Furthermore, in step 4 of the federated learning training step, the scheduling system of federated learning is utilized to accurately manage the periodic update process to maximize the system performance.
[0035] Furthermore, in the global model aggregation, update and use of the remote model pool, the evaluation of malicious models and unreliable models and the global model aggregation operations include;
[0036] First, perform an average operation on all local models in the previous round to obtain an average model, and calculate and evaluate the deviation difference between each local model uploaded to the remote server in this round and the average model of the previous round;
[0037] Second, if the deviation difference between the uploaded local model and the average model is lower than the specified first threshold, it is identified as a suspected normal model to be estimated; otherwise, it is identified as a malicious model; the deviation between the local model and the average model is evaluated in the following way:
[0038]
[0039] It represents the deviation between the nth local model in the Kth round and the average model in the K-1th round, that is, the Minkowski shift distance. represents the average model of the K-1th round, represents the nth local model of the Kth round of model federation learning;
[0040] Third, according to the deviation Size, assign the normalized contribution weight λ of the nth local suspected normal model node to the global model aggregation n , the deviation degree and contribution weight show an inverse relationship;
[0041] Fourth, the normalized contribution weight set of the Kth round is (λ1,λ2....λ n ), if the contribution weight dispersion is higher than the specified second threshold, the local suspected normal model to be estimated in this round is identified as an unreliable model, and the K+1 round of model training and learning is required again until the contribution weight dispersion is lower than the specified second threshold and is identified as the final normal model. The model dispersion degree is evaluated by the following methods:
[0042]
[0043] δ K represents the standard deviation of the normalized contribution weight of the local suspected normal model node to the global model aggregation in the Kth round, λ irepresents the normalized contribution weight of the i-th model in the K-th round, Represents the average value of normalized contribution weights in round K;
[0044] Fifth, all normal models are aggregated, that is, all normal models are weighted according to their contribution weights to obtain a global model.
[0045] Furthermore, in the global model aggregation, update and use of the remote model pool, the global model forms a large and comprehensive shared refrigerator model pool, including models such as resource scheduling, consumption recommendations, and user preferences, providing model testing functions for authorized users. Models that have not been trained through federated learning are regarded as initial global models.
[0046] Furthermore, the local model training layer of the edge refrigerator device is used for refrigerator and user data segmentation and edge node training. The local node device trained by federated learning obtains the latest comprehensive global model parameters from the remote server, uses local data to independently perform learning and training tasks, and submits the trained local node model parameters to the remote server. The server authenticates the local training node model and confirms that it is not a malicious or unreliable node model before it can join the global model aggregation calculation of federated learning training.
[0047] The transmission service layer is used for the transmission and interaction of model parameters between the local node model and the global model of the cloud server during federated learning training;
[0048] The cloud service layer is used for remote servers to aggregate model parameters and maintain and manage each model in the model pool. The integrated monitoring system monitors the operating status and model performance of local nodes in real time and handles possible abnormal situations in a timely manner. The remote server uses federated learning weighted algorithm technology to aggregate the model parameters uploaded by all local nodes into a global model. Mobile applications are developed on the remote server to realize remote management of shared refrigerator model resources and visualization and query management of related data.
[0049] Another object of the present invention is to provide a local microcomputer device, the computing device includes a memory, a processor and a signal transmission device, the memory stores the local model algorithm program, the processor executes the algorithm program, the signal transmission device transmits the model parameter results of the processor executing the algorithm program to a cloud server, and the microcomputer executes the federated learning training steps as described in claim 2.
[0050] Another object of the present invention is to provide a remote server device, which includes a memory and a processor. The memory stores an algorithm program for calculating global model parameters, data and information sent from a mobile phone application, and the processor executes the algorithm program, performs functional transformation, maintenance, testing and optimization of the mobile phone application, etc. The processor executes parameter aggregation, updating, and collaborative learning of the federated learning training steps as described in claim 2 and application maintenance and optimization as described in claim 1, thereby improving the user experience of using a shared refrigerator.
[0051] Another object of the present invention is to provide a mobile application, which is a shared refrigerator management and data display query interface, providing query and management of the federated learning training time of each model, the latest global model update and issuance time, the existing model status of the model pool, and the user status of authorized login to view the model pool. The mini program software queries and manages the model resources as described in claim 1.
[0052] The beneficial effects of the present invention are as follows: while ensuring the normal operation of campus shared refrigerators, federated learning provides a forward-looking solution to the problem of privacy and data security. At the same time, it provides a centralized model analysis application pool platform through mobile phone applications, so that the application of data resources generated by shared refrigerators can be expanded, which has the commercial potential and economic benefits of the industry. This innovative method combines solving information security with creating economic benefits, which not only points out the direction for the further development of campus shared refrigerator technology, but also opens up a new solution to the privacy challenges in IoT applications.
[0053] The local nodes of federated learning can dynamically adjust model parameters according to their own data characteristics and environmental changes, and update them synchronously with the cloud, which enhances the robustness and scalability of the system. Personal information and sensitive data do not need to be transmitted to the cloud, reducing the risk of data leakage. By transmitting only model parameters instead of huge raw data, this technology reduces communication overhead and latency, thereby improving the overall efficiency of the system. Federated learning achieves the convergence of global knowledge through the collaborative training of multiple local nodes, promotes the collaborative learning of models, and provides a more comprehensive data foundation for the intelligent analysis and optimization of the system. Related models such as resource scheduling and sharing, consumer behavior prediction, character portrait analysis, face recognition, and queue waiting analysis are introduced to form a large model pool to achieve intelligent management and services for items, customers, and resources, and meet multi-faceted and multi-level needs such as diversified needs, personalized needs, and special needs. BRIEF DESCRIPTION OF THE DRAWINGS
[0054] Figure 1 This is a schematic diagram of the functional architecture of the campus shared refrigerator system based on federated learning of the present invention;
[0055] Figure 2 This is a diagram of the training steps of federated learning for the campus shared refrigerator system model;
[0056] Figure 3 This is a closed-loop flow chart of the learning, training and operation of the campus shared refrigerator system model based on federated learning;
[0057] Figure 4 This is a schematic diagram of independent learning and training of the local model based on data reorganization in the local refrigerator;
[0058] Figure 5 This is a schematic diagram of the main interface of the WeChat applet management of the shared refrigerator on the mobile phone;
[0059] Figure 6 This is a schematic diagram of the order interface called by the cloud model after the authorized user signs the usage agreement. DETAILED DESCRIPTION
[0060] In order to make the technical problems and technical solutions solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0061] like Figure 1 As shown in the figure, the system realizes data interaction and transmission between local refrigerators, user mobile phones, and remote Alibaba Cloud servers and micro servers through technologies such as Alibaba Cloud servers, WeChat servers, mobile wireless networks, the Internet, and WeChat mini-program terminals. The system is divided into three layers: the local model training layer of the edge refrigerator device, the transmission service layer, and the cloud service layer, which specifically include:
[0062] The local model training layer of the edge refrigerator device is used for refrigerator and user data segmentation and edge node training. The local node device of the federated learning training obtains the latest comprehensive global model parameters from the remote server, uses local data to independently perform learning and training tasks, and submits the trained local node model parameters to the remote server. The server authenticates the local training node model and confirms that it is not a malicious or unreliable node model before it can join the global model aggregation calculation of the federated learning training.
[0063] The transmission service layer is used for the transmission and interaction of model parameters between the local node model and the global model of the cloud server during federated learning training;
[0064] The cloud service layer is used by Alibaba Cloud Server to realize model parameter aggregation and model maintenance and management in the model pool. The integrated monitoring system monitors the operating status and model performance of local nodes in real time and handles possible abnormal situations in a timely manner. Alibaba Cloud Server uses federated learning weighted algorithm technology to aggregate the model parameters uploaded by all local nodes into a global model. The WeChat applet for mobile shared refrigerators is developed on the WeChat server to realize remote management of shared refrigerator model resources and visualization and query management of related data.
[0065] The local shared refrigerator of the local model training layer of the edge refrigerator device is composed of a collection of smart refrigerators distributed in different universities, buildings, and areas. Each refrigerator is composed of independent storage compartments with refrigeration and freezing functions. It is connected to the Internet through Raspberry Pi and WiFi communication technology to achieve communication and interaction with Alibaba Cloud servers, which can not only meet students' needs for independent partitioned refrigerators and realize full utilization of resources, but also ensure the safety of electricity use in student dormitories.
[0066] The shared refrigerator data transmission in the transmission service layer is connected to the Internet through the Raspberry Pi microcontroller and Wifi communication technology of the local shared refrigerator and transmitted to the server of the cloud service layer. The mobile terminal can access the Internet through the 5G wireless network and transmit it to the server of the cloud service layer.
[0067] The WeChat applet on the cloud service layer allows users to rent a refrigerator by scanning the QR code on the refrigerator or logging into the WeChat applet on the mobile phone. This allows users to easily “go around campus with their mobile phone” and quickly transmit, store and update data. Students can also use their recorded facial information to take items, which further facilitates students’ lives. In addition, the applet also has menu recipe recommendations, expiration reminders and other functions. At the same time, the local shared refrigerator can learn and update local model parameters on a regular or manual basis.
[0068] like Figure 2 , Figure 3 and Figure 4 As shown in the figure, the privacy protection for campus shared refrigerators is achieved by combining the cloud-based centralized model aggregation pool of federated learning and the local distributed model learning and training. The specific operation process, steps and technical support are as follows: 1) Model initialization:
[0069] a. Build a campus shared refrigerator system and provide a model federated learning scenario;
[0070] b. Set up various appropriate models according to the needs of shared refrigerator management scenarios;
[0071] c. Set appropriate conditions and requirements for model training, interaction, and updating;
[0072] d. Initialize various models in the cloud model pool;
[0073] 2) Federated learning training steps:
[0074] Step 1: Data segmentation and local training: Each campus shared refrigerator is regarded as a local node. The local node device obtains the latest comprehensive global model parameters from the remote server and uses local data for model training, including user's switch, temperature adjustment, food storage and retrieval operation records, usage habits, etc. Then, distributed data segmentation technology is used to ensure that each local node has enough samples to train the model while complying with the privacy protection principle.
[0075] Step 2: Upload model parameters to the cloud: After training on the local node, only upload model parameters instead of original data to the cloud, using secure transmission protocols and end-to-end encryption technology to ensure secure transmission of model parameters and data security;
[0076] Step 3: Cloud model parameter aggregation: Federated learning weighted algorithm technology is used in the cloud to aggregate the model parameters uploaded by all local nodes into a global model;
[0077] Step 4: Send the global model to each local node: Send the global model parameters generated in the cloud to each local node, and use security protocols to ensure the secure transmission of model parameters and data security;
[0078] Step 5: Regular updates and collaborative learning: Set up a regular federated learning execution plan to ensure that the system can learn new information from data on all nodes.
[0079] Step 6: Monitoring and maintenance: Establish a monitoring system to regularly check the operating status and model performance of each local node, handle any abnormal situations that may arise in a timely manner, and ensure the stable operation of the entire federated learning system.
[0080] 3) Global model aggregation, update and use of remote model pool:
[0081] a. Each time a new round of model parameter aggregation is performed, it is necessary to evaluate whether the local node model is a malicious model or an unreliable model. Malicious models and unreliable models need to be distinguished;
[0082] b. Malicious and unreliable models are eliminated, and the remaining local node models will be aggregated to form a new global model;
[0083] c. The new global model is added to the model pool, and authorized users can view and use the corresponding model through the shared refrigerator WeChat applet;
[0084] d. For the user use and transfer model, facial recognition registration and contract signing are carried out online through the shared refrigerator WeChat applet to ensure the compliance and legal transfer of ownership;
[0085] 4) Maintenance and optimization of model pool:
[0086] a. Based on actual application conditions, regularly supplement and optimize the corresponding models in the model pool to improve model accuracy and enhance user stickiness and experience of using refrigerators;
[0087] b. Continuously maintain and upgrade the shared refrigerator system to meet technological development and users’ personalized and multi-level needs.
[0088] Evaluation of malicious and unreliable models and global model aggregation operations include;
[0089] First, perform an average operation on all local models in the previous round to obtain an average model, and calculate and evaluate the deviation difference between each local model uploaded to the remote server in this round and the average model of the previous round;
[0090] Second, if the deviation difference between the uploaded local model and the average model is lower than the specified first threshold, it is identified as a suspected normal model to be estimated; otherwise, it is identified as a malicious model; the deviation between the local model and the average model is evaluated in the following way:
[0091]
[0092] It represents the deviation between the nth local model in the Kth round and the average model in the K-1th round, that is, the Minkowski shift distance. represents the average model of the K-1th round, represents the nth local model of the Kth round of model federation learning;
[0093] Third, according to the deviation Size, assign the normalized contribution weight λ of the nth local suspected normal model node to the global model aggregation n , the deviation degree and contribution weight show an inverse relationship;
[0094] Fourth, the normalized contribution weight set of the Kth round is (λ1,λ2....λ n ), if the contribution weight dispersion is higher than the specified second threshold, the local suspected normal model to be estimated in this round is identified as an unreliable model, and the K+1 round of model training and learning is required again until the contribution weight dispersion is lower than the specified second threshold and is identified as the final normal model. The model dispersion degree is evaluated by the following methods:
[0095]
[0096] δ K represents the standard deviation of the normalized contribution weight of the local suspected normal model node to the global model aggregation in the Kth round, λ i represents the normalized contribution weight of the i-th model in the K-th round, Represents the average value of normalized contribution weights in round K;
[0097] Fifth, all normal models are aggregated, that is, all normal models are weighted according to their contribution weights to obtain a global model.
[0098] An application embodiment of the present invention provides a local Raspberry Pi microcomputer device, wherein the computing device includes a memory, a processor and a signal transmission device, wherein the memory stores a local model algorithm program, the processor executes the algorithm program, and the signal transmission device transmits the model parameter results of the processor executing the algorithm program to a cloud server, so that the processor executes the steps of a campus shared refrigerator method based on federated learning.
[0099] The application embodiment of the present invention provides a remote Alibaba Cloud and WeChat server device, the server device includes a memory and a processor, the memory stores the algorithm program for calculating the global model parameters, the data and information sent by the mobile phone application, the processor executes the algorithm program, the mobile phone application function transformation maintenance, testing and optimization, etc., so that the processor executes the steps of the campus shared refrigerator method based on federated learning.
[0100] The application embodiment of the present invention provides a mobile phone WeChat applet, which is a shared refrigerator management and data display query tool, providing the federated learning training time of each model, the latest global model update and issuance time, the existing model status of the model pool, and the authorized login to call the model of the model pool, etc. The WeChat applet is used to realize a campus shared refrigerator based on federated learning.
[0101] It should be noted that the embodiments of the present invention can be implemented through a combination of software and hardware. The hardware part can be implemented using a dedicated server, a dedicated logic control chip and a microcomputer, and the software part can be stored in a related memory and called and executed by a corresponding instruction execution system, such as a third-party server or microprocessor. We tested this technology in a university campus scenario, using foam boxes, single-core refrigeration kits and relays to assemble and simulate local shared refrigerator equipment, and developed a mobile phone shared refrigerator WeChat applet as a dedicated software tool for refrigerator management through WeChat servers and computers. Figure 5 , Figure 6 shown.
[0102] The present invention is described in detail above through specific and preferred embodiments, but those skilled in the art should understand that the present invention is not limited to the embodiments described above, and any modifications, equivalent substitutions, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A campus shared refrigerator method based on federated learning, characterized in that: A cloud-based centralized model pool for shared refrigerator management serving the campus was built using federated learning technology. Through the automatic training, evaluation, interaction and update process of local node models and remote global models, the identity verification of training nodes, regular aggregation and update of global models, and the composition of model pools by different types of global models are realized. Authorized users can query and use related models in the model pool through mobile applications according to their own needs, and other functions are realized to achieve secure sharing of shared refrigerator data. At the same time, the models and applications of the model pool are continuously maintained and optimized to meet the needs of different users for refrigerators and enhance the user experience of shared refrigerators.
2. According to the method of a campus shared refrigerator based on federated learning in claim 1, it is characterized in that: include: 1) Model initialization: a. Set up various appropriate models according to the needs of shared refrigerator management scenarios; b. Set appropriate conditions and requirements for model training, interaction, and updating; c. Initialize various models in the cloud model pool; 2) Federated learning training steps: Step 1: Data segmentation and local training: Each campus shared refrigerator is regarded as a local node. The local node device obtains the latest comprehensive global model parameters from the remote server and uses local data for model training, including user's switch, temperature adjustment, food storage and retrieval operation records, usage habits, etc. Then, distributed data segmentation technology is used to ensure that each local node has enough samples to train the model while complying with the privacy protection principle. Step 2: Upload model parameters to the cloud: After training on the local node, only upload model parameters instead of original data to the cloud, using secure transmission protocols and end-to-end encryption technology to ensure secure transmission of model parameters and data security; Step 3: Cloud model parameter aggregation: Federated learning weighted algorithm technology is used in the cloud to aggregate the model parameters uploaded by all local nodes into a global model; Step 4: Send the global model to each local node: Send the global model parameters generated in the cloud to each local node, and use security protocols to ensure the secure transmission of model parameters and data security; Step 5: Regular updates and collaborative learning: Set up a regular federated learning execution plan to ensure that the system can learn new information from data on all nodes. Step 6: Monitoring and maintenance: Establish a monitoring system to regularly check the operating status and model performance of each local node, handle any abnormal situations that may occur in a timely manner, and ensure the stable operation of the entire federated learning system. 3) Global model aggregation, update and use of remote model pool: a. Each time a new round of model parameter aggregation is performed, it is necessary to evaluate whether the local node model is a malicious model or an unreliable model. Malicious models and unreliable models need to be distinguished; b. Malicious and unreliable models are eliminated, and the remaining local node models will be aggregated to form a new global model; c. The new global model is added to the model pool, and authorized users can view and use the corresponding model through the mobile application; d. For user use and transfer models, facial recognition registration and contract signing are carried out online through mobile phone applications to ensure the compliance and legal transfer of ownership; 4) Maintenance and optimization of model pool: a. Based on actual application conditions, regularly supplement and optimize the corresponding models in the model pool to improve model accuracy and enhance user stickiness and experience of using refrigerators; b. Continuously maintain and upgrade the shared refrigerator system to meet technological development and users’ personalized and multi-level needs.
3. According to the method of sharing refrigerators on campus based on federated learning in claim 2, it is characterized in that: Differential privacy technology is introduced in step three of the federated learning training process to enhance user privacy protection and ensure that individual data is not leaked during the aggregation process.
4. According to the method of sharing refrigerators on campus based on federated learning in claim 2, it is characterized in that: In step 4 of the federated learning training procedure, the scheduling system of federated learning is used to accurately manage the regular update process to maximize system performance.
5. According to the method of sharing refrigerators on campus based on federated learning in claim 2, it is characterized in that: In the global model aggregation, update and use of the remote model pool, the evaluation of malicious models and unreliable models and the global model aggregation operations include: First, perform an average operation on all local models in the previous round to obtain an average model, and calculate and evaluate the deviation between each local model uploaded to the remote server in this round and the average model of the previous round; Second, if the deviation between the uploaded local model and the average model is lower than the first threshold, it is identified as a suspected normal model to be estimated; otherwise, it is identified as a malicious model; the deviation between the local model and the average model is evaluated in the following way: It represents the deviation between the nth local model in the Kth round and the average model in the K-1th round, that is, the Minkowski shift distance. represents the average model of the K-1th round, represents the nth local model of the Kth round of model federation learning; Third, according to the deviation Size, assign the normalized contribution weight λ of the nth local suspected normal model node to the global model aggregation n , the deviation degree and contribution weight show an inverse relationship; Fourth, the normalized contribution weight set of the Kth round is (λ1, λ2....λ n ), if the contribution weight dispersion is higher than the specified second threshold, the local suspected normal model to be estimated in this round is identified as an unreliable model, and the K+1th round of model training and learning is required again until the contribution weight dispersion is lower than the specified second threshold and is identified as the final normal model; the model dispersion degree is evaluated by the following method: δ K represents the standard deviation of the normalized contribution weight of the local suspected normal model node to the global model aggregation in the Kth round, λ i represents the normalized contribution weight of the i-th model in the K-th round, Represents the average value of normalized contribution weights in round K; Fifth, all normal models are aggregated, that is, all normal models are weighted according to their contribution weights to obtain a global model.
6. A campus shared refrigerator method based on federated learning according to claim 2, characterized in that: In the global model aggregation, update and use of the remote model pool, the global model forms a large and comprehensive shared refrigerator model pool, including models such as resource scheduling, consumption recommendations, and user preferences, providing model testing and usage functions for authorized users. Models that have not been trained through federated learning are regarded as initial global models.
7. A campus shared refrigerator method based on federated learning according to claim 1, characterized in that: Also includes: The local model training layer of the edge refrigerator device is used for refrigerator and user data segmentation and edge node training. The local node device of the federated learning training obtains the latest comprehensive global model parameters from the remote server, uses local data to independently perform learning and training tasks, and submits the trained local node model parameters to the remote server. The server authenticates the local training node model and confirms that it is not a malicious or unreliable node model before it can be added to the global model aggregation calculation of the federated learning training; the transmission service layer is used for the transmission and interaction of model parameters between the local node model of the federated learning training and the global model of the cloud server; The cloud service layer is used for remote servers to aggregate model parameters and maintain and manage models in the model pool. The integrated monitoring system monitors the operating status and model performance of local nodes in real time and handles possible abnormal situations in a timely manner. The remote server uses federated learning weighted algorithm technology to aggregate the model parameters uploaded by all local nodes into a global model. Mobile applications are developed on the remote server to realize remote management of shared refrigerator model resources and visualization and query management of related data.
8. A local microcomputer device, the computing device includes a memory, a processor and a signal transmission device, the memory stores and calculates a local model algorithm program, the processor executes the algorithm program, the signal transmission device transmits the model parameter results of the processor executing the algorithm program to a cloud server, and the microcomputer executes the federated learning training steps as described in claim 2.
9. A remote server device, the server device includes a memory and a processor, the memory stores the algorithm program for calculating global model parameters, data and information sent from the mobile application, the processor executes the algorithm program, the mobile application function modification maintenance, testing and optimization, etc. The processor executes the parameter aggregation, update, collaborative learning of the federated learning training steps as described in claim 2 and the application maintenance and optimization as described in claim 1, thereby improving the user experience of using a shared refrigerator.
10. A mobile application, which is a shared refrigerator management and data display query interface, provides query and management of the federated learning training time of each model, the latest global model update and release time, the existing model status of the model pool, and the user status of authorized login to view the model pool. The application software is the model resource query and management as described in claim 1.
Citation Information
Cited By
Vaccine cold chain collaborative control system and method based on edge collaborative learning
CN122694334A
