Personalized model based on edge device environment difference and training method thereof
By conducting personalized model training and aggregation on edge devices and combining generalization training of the whole data, the model adaptability and robustness problems caused by environmental differences are solved, and the model is efficiently adapted and widely used.
Patent Information
- Application Number
- CN202510253473.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-05
- Publication Date
- 2025-06-24
AI Technical Summary
The prior art fails to fully utilize the computing power of the device when training models on edge devices, and due to environmental differences, the feedback of the same set of models after being deployed on different devices is large, which affects the adaptability and robustness of the model.
A personalized model training method based on the environmental differences of edge devices is adopted, and edge devices are divided into groups by clustering and multiple model training is carried out on each group, models within the group are aggregated, and models are generalized using the whole data to improve the adaptability and robustness of the model.
This method has achieved remarkable results in balancing personalization and generalization, so that the model can not only accurately meet the needs of individual users, but also have stronger robustness and adaptability, improving the performance and flexibility of the overall system.
Smart Images
Figure CN120196945A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method for training an edge device model, and more particularly to a personalized model training method based on the environmental differences of edge devices. Background Art
[0002] With the widespread application of renewable energy and the popularization of Internet of Things (IoT) devices globally, the distribution network is facing unprecedented challenges. The traditional centralized computing model used to be the mainstream in power system monitoring and management, but it has many deficiencies. First, data collection and analysis mainly rely on remote data centers, resulting in limited real-time data processing capabilities. This limitation makes the response time of the distribution system to emergencies longer, affecting the stability and reliability of power supply.
[0003] Entering the 21st century, the rapid development of IoT technology has prompted more and more sensors, smart meters, and devices to be connected to the power grid, generating a large amount of data in real time. This change not only improves the monitoring ability of the distribution network status but also promotes in-depth analysis of its operation. However, with the explosion of data volume, the traditional centralized computing model faces huge challenges, especially in terms of data transmission delay and bandwidth limitation, and cannot meet the requirements of real-time monitoring and response.
[0004] To solve these problems, edge computing technology has emerged. Edge computing emphasizes data processing near the data generation source, significantly reducing data transmission delay and enhancing the real-time response ability of the system. This concept provides a new solution for the power industry, prompting the industry to explore the application potential of edge computing in real-time monitoring, fault detection, and data analysis.
[0005] In the context of the increasingly in-depth construction of smart grids, edge computing has gradually become an important part of them, supporting the access and management of distributed energy and effectively coping with the volatility of renewable energy. Through real-time data processing by edge devices, the distribution system can quickly identify and respond to potential faults or instability factors, thereby improving the reliability of power supply.
[0006] The performance of edge devices is constantly enhancing and can now deploy complex models on these devices for real-time monitoring and fault detection. These devices not only have stronger computing and storage capabilities but also support the application of deep learning and machine learning algorithms, enabling them to efficiently analyze and process a large amount of data. By running intelligent models on edge devices, the system can quickly identify potential problems and respond in a timely manner, thus significantly improving the reliability and stability of the distribution network.
[0007] Despite the continuous improvement in the performance of edge devices, due to significant environmental differences, the feedback obtained after deploying the same set of models on different edge devices may vary significantly. These differences may stem from various factors such as the geographical location of the device, operating conditions, and load characteristics. Current technologies rely on uploading the data of each edge device to a central server, where the central server trains the model with all the data. This training method does not fully utilize the computing power of edge devices. Summary of the Invention
[0008] The purpose of the present invention is to solve the above-mentioned deficiencies of the prior art, and thus provide a personalized model training method based on the environmental differences of edge devices.
[0009] A personalized model training method based on the environmental differences of edge devices includes the following steps:
[0010] Cluster and divide edge devices into groups according to the data of each edge device;
[0011] Train the group model multiple times on each group;
[0012] Aggregate the models trained by the edge devices within the group;
[0013] Train the group model at least once using all the data to enhance the generalization and robustness of the group model;
[0014] Determine whether the model converges. If it converges, obtain the final model. If it does not converge, return to re-divide the groups.
[0015] Cluster and divide edge devices into groups according to the data of each edge device, specifically:
[0016] Based on the environment where each edge device is located, obtain the data feature point D symbolizing its own characteristics;
[0017] Based on the data feature point D, calculate the cosine similarity between the data of each edge device and the zero point d σ ;
[0018] Cluster the edge devices according to the cosine similarity, and divide the edge nodes with the same data characteristics into the same node.
[0019] Based on the data feature point D, calculate the cosine similarity between the data of each edge device and the zero point d σ , and the formula is:
[0020]
[0021] Where: D is the data feature of the edge device, d iDenote the data feature points of the $i$-th edge device, and $S$ represents the relative distance calculated based on the cosine similarity between each edge device and the zero point.
[0022] Cluster the edge devices according to the cosine similarity. The specific clustering formula is:
[0023]
[0024] where: $S$ i represents the cosine similarity between the data of the $i$-th edge device and the zero point $d$ σ , that is, the relative distance from the zero point. $S$ j represents the cosine similarity between the data of the $j$-th edge device and the zero point $d$ σ , and $\sigma$ represents the radius during clustering.
[0025] The model optimization formula for training the group model on each group is:
[0026]
[0027] where: $m$ represents the model parameters, is the loss function, minimize is the optimization method for finding the minimum value, and $u$ represents the data used for training. Here, all the data comes from the edge devices within the group, i.e., GroupData.
[0028] The model optimization formula for training the group model using all the data is:
[0029]
[0030] $u$ represents the data used for training. At this time, the data used for training is the data of all edge devices, i.e., AllData.
[0031] On the other hand, the present invention provides a personalized model based on the environmental differences of edge devices trained by the method described in the above solution.
[0032] Compared with the prior art, the advantages of the present invention are:
[0033] 1. This method first trains the model using similar data to maximize the adaptability of the specific edge device's environment, and then trains it with all the data to improve the generalization ability of the model. This method has achieved remarkable results in balancing personalization and generalization, enabling the model to not only accurately meet the needs of individual users, but also have stronger robustness and adaptability in a wider data distribution, thereby improving the overall performance and flexibility of the system.
[0034] 2. This method: First, perform personalized pre-training using similar data to enable the model to quickly converge to a specific task, and then use all data for generalization training. This two-stage training strategy effectively reduces the training time: The pre-training with similar data reduces the complexity of the initial parameters, thereby accelerating the speed of model optimization, while the joint training with all data avoids the time consumption required for multiple iterations while retaining personalized performance, overall improving the training efficiency.
[0035] 3. Through the combination of personalization and generalization, this method achieves compatibility with device differences and large-scale applications, and is suitable for expansion to more edge devices and user groups, supporting large-scale deployment in edge computing environments. Especially in edge computing networks, the personalization stage ensures the training efficiency of each device, while the generalization stage improves cross-device consistency, thereby enhancing the scalability of the model and the stability of the overall system. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 It is a flowchart of the method of the present invention. DETAILED DESCRIPTION OF THE INVENTION
[0037] The present invention performs personalization on the models deployed on each edge device according to the specific environment and usage scenario of the edge device. This personalized adjustment can not only improve the adaptability and accuracy of the model, but also optimize its performance in actual applications. By analyzing the historical data and real-time monitoring information of each edge device, we can adjust the parameters and structure of the model to better meet the requirements of the specific environment, thereby achieving more efficient fault detection and performance monitoring. This customized method provides a more flexible and intelligent solution for the application of edge computing in the distribution network.
[0038] Such as Figure 1 , a personalized model training method based on the environmental differences of edge devices of the present invention includes the following steps:
[0039] Cluster and divide edge devices into groups according to the data of each edge device;
[0040] Based on the environment where each edge device is located, obtain the data feature point D symbolizing its own characteristics; specifically: count the number of each landform in the surrounding environment of each edge device, and combine them to form a high-dimensional data feature point of the data characteristics symbolizing its own characteristics;
[0041] For example: collect the deployment environment of edge devices, such as how many poles are around, and whether each pole is located in farmland, the city or the forest. Then record, for example, that there are 5 farmlands, 3 cities, and 0 forests to be inspected near edge device A, then record it as (5, 3, 0). Edge device B has 0 farmlands, 0 cities, and 9 forests, then record B as (0, 0, 9). Among them, (5, 3, 0) and (0, 0, 9) are the data feature points of A and B respectively.
[0042] Based on the data feature point D, calculate the cosine similarity between the data of each edge device and the zero point d σ The formula is:
[0043]
[0044] where: D is the data feature of the edge device, d i represents the data feature point of the i-th edge device.
[0045] Cluster the edge devices according to the cosine similarity, and divide the edge nodes with the same data features into the same node.
[0046] The specific clustering formula is:
[0047]
[0048] where: S i represents the cosine similarity between the data of the i-th edge device and the zero point d σ S j represents the cosine similarity between the data of the j-th edge device and the zero point d σ σ represents the radius during clustering.
[0049] Train the group model on each group, and the model optimization formula is:
[0050]
[0051] where: m represents the model parameters, is the loss function, minimize is the optimization method for finding the minimum value, and u represents the data used for training. Here, all the data comes from the edge devices within the group GroupData.
[0052] Aggregate the models trained by the edge devices within the group;
[0053] Train the group model with all the data at least once to enhance the generalization and robustness of the group model; the model optimization formula for training the group model with all the data is:
[0054]
[0055] u represents the data used for training. At this time, the data used for training is the data of all edge devices, that is, AllData.
[0056] Determine whether the model converges. If it converges, obtain the final model. If it does not converge, return to re-divide the groups.
[0057] On the other hand, the present invention provides a personalized model based on the environmental differences of edge devices trained using the solution described in the above technical solution.
[0058] Embodiment
[0059] 1. Initialize parameters and clustering: Set the edge device set E, the total number of communication rounds as T, and the data feature of the edge device as D.
[0060] For the communication round t = 0, the central server sends down the reference system data zero point d σ ;
[0061] Collect the data feature points D of the environment where each edge device is located;
[0062] This method of clustering pays more attention to the tightness of the data. Therefore, the cosine similarity is used as a measure of distance. The edge node calculates the cosine similarity between its own data feature D and d σ as the distance S and uploads it to the central server;
[0063] 2. Division of groups of edge devices of the same type:
[0064] According to the clustering of cosine similarity, edge nodes with the same data features are divided into the same node. Pay more attention to the tightness between data of the same category, so the cosine similarity is used for measurement;
[0065] The central server sends down the initial model to the edge devices;
[0066] 3. Training of personalized models within the group:
[0067] For t = x + i to y + i, each edge device uses its own data to train the model and aggregates within the group;
[0068] For t = x + y + i, each group uploads the group model to the central server for training using the central data;
[0069] Send down the group model after training with the central data to the edge devices within the corresponding group;
[0070] Among them, y - x is equal to the number of rounds of local personalized training, and the specific number of rounds is determined according to the data size, with all data participating in training as the standard.
[0071] 4. Personalized edge model:
[0072] At t = T, the edge device downloads the corresponding model from the central server;
[0073] The method proposed above measures the relative distance using cosine similarity for the data features of edge nodes and clusters them to obtain multiple groups with similar data types. After the edge nodes within the group use local data to train the model 4 times to increase the sensitivity of the model to the actual environment, they then use the central server data to train once to increase its generalization and robustness, enabling the model to have higher sensitivity to the corresponding edge node data while having a certain degree of robustness to avoid overfitting of the model, thereby achieving that each edge node has a personalized edge model to better adapt to its work.
Claims
1. A personalized model training method based on edge device environment differences, characterized in that: The following steps are involved: Cluster edge devices into groups based on the data of each edge device; Train the group model multiple times on each group; Aggregate models trained on edge devices within a group; Use the entire data to train the group model at least once to enhance the generalization and robustness of the group model; Determine whether the model converges. If so, obtain the final model. If not, return to re-divide the groups.
2. According to claim 1, a personalized model training method based on edge device environment differences is characterized in that: The edge devices are clustered into groups based on the data of each edge device, specifically: Based on the environment in which each edge device is located, obtain the data feature points D that symbolize its own characteristics; Based on the data feature point D, calculate the data of each edge device and the zero point d σ The cosine similarity of According to the cosine similarity, the edge devices are clustered and the edge nodes with the same data features are divided into the same node.
3. According to claim 2, a personalized model training method based on edge device environment differences is characterized in that: Based on the data feature point D, calculate the data of each edge device and the zero point d σ The cosine similarity of is: Where: D is the edge device data feature, d i represents the data feature point of the i-th edge device, and S represents the relative distance between each edge device and the zero point based on the cosine similarity.
4. The personalized model training method based on edge device environment differences according to claim 2 is characterized in that: According to the cosine similarity, the edge devices are clustered. The specific clustering formula is: Where: S i Represents the data of the i-th edge device and the zero point d σ The cosine similarity is the relative distance from zero, S j Represents the data of the jth edge device and the zero point d σ The cosine similarity of , σ represents the radius of clustering.
5. The personalized model training method based on edge device environment differences according to claim 1 is characterized in that: The model optimization formula for training the group model on each group is: Where: m represents the model parameter, is the loss function, minimize is the optimization method for finding the minimum value, and u represents the data used for training. Here, all data comes from the edge device GroupData in the group.
6. The personalized model training method based on edge device environment differences according to claim 1 is characterized in that: The model optimization formula for training the group model using all data is: u represents the data used for training. In this case, the data used for training is the data of all edge devices, that is, AllData.
7. A personalized model based on the differences in edge device environments obtained by training using the method described in 1-6.