Clustering algorithm-based new edge device matching high-generalization model and construction method thereof

Abstract

The invention discloses a clustering algorithm-based new edge device matching high-generalization model and a construction method thereof, and the construction method comprises the following steps: generating a data feature point D representing the data feature of each edge device based on the environment where each edge device is located; classifying the data feature points D by using Euclidean distance; extracting different types of edge devices to form a group containing all types of data; training the initial model on the group until the model converges to obtain a group model; and aggregating the small group models to obtain a high generalization model. The data of each category of equipment is effectively represented in the group, and meanwhile, proportional balance is performed to ensure that the model has high generalization on the basis of seeing different categories of data. For a newly added edge device, good performance can be obtained by using the generalization model based on multi-class data training even under the condition that the environment and task characteristics of the edge device are unknown.

Description

Technical Field

[0001] The present invention relates to the technical field of edge device model construction, and specifically to a new edge device supporting highly generalized model based on a clustering algorithm and its construction method. Background Art

[0002] In an edge computing network, the addition of new edge devices often faces many challenges. Especially when these devices have no historical data, it is impossible to have a full understanding of their environment, which makes the deployment of models more complex and difficult. Since new devices have not participated in previous computing processes, they cannot effectively utilize existing historical data and models, so the effect of directly migrating historical training models by existing methods is not satisfactory. The reason is that the data used to train these historical models may not be suitable for the environment faced by new devices, resulting in a significant decline in model performance and inability to effectively meet the actual needs of new devices.

[0003] In an edge computing network, the environments and conditions of devices may vary greatly. For example, data distribution changes caused by different geographical locations, uneven hardware performance, different user usage scenarios, etc. Therefore, when new devices first join an edge computing network, the lack of historical data and environmental information makes it difficult for them to directly adapt to tasks in the network. Traditional models are usually trained in specific environments and lack generalization ability, making it difficult to obtain ideal performance in new environments. In addition, the computing power, storage resources, and network bandwidth of edge devices are often strictly limited, making it difficult to effectively deploy complex existing models directly, thus affecting the rapid response and actual application effect of new devices. Therefore, more effective strategies need to be developed to help these newly added edge devices achieve rapid model adaptation and stable operation. Summary of the Invention

[0004] The purpose of the present invention is to solve the above-mentioned deficiencies of the prior art, and thus provide a new edge device supporting highly generalized model based on a clustering algorithm and its construction method.

[0005] A construction method for a new edge device supporting highly generalized model based on a clustering algorithm includes the following steps:

[0006] Generate data feature points D representing its own data characteristics based on the environment where each edge device is located;

[0007] Classify the data feature points D using the Euclidean distance;

[0008] Extract different types of edge devices to form a group containing all types of data;

[0009] Train an initial model on the group until the model converges to obtain a group model;

[0010] Aggregate the models of each group to obtain a highly generalized model.

[0011] Classify the data feature points D using the Euclidean distance, specifically:

[0012] Calculate the Euclidean distance between each edge device and the data origin D of the central server x The formula is:

[0013]

[0014] Where: S i Represents the Euclidean distance between the i-th edge device and the data origin D of the central server x ; D i Represents the data feature point of the i-th edge device;

[0015] According to the Euclidean distance between each edge device and the data origin D of the central server x Use the K-means clustering algorithm to classify each edge device to obtain different categories:

[0016] C i ={d(S i , S j ) ≤ ∈}.

[0017] Extract edge devices of different categories to form groups containing data of all categories, specifically:

[0018] Statistically calculate the total amount of data carried by each edge device within each category;

[0019] Select edge devices from each category according to the total amount of data carried to form a group.

[0020] When selecting edge devices, the sum of the total amount of data carried by all edge devices within each group needs to be balanced.

[0021] Train the initial model on the groups until the model converges to obtain the group models, specifically:

[0022] Use the data within each group to train the initial model as:

[0023]

[0024] Where: w t+1 Is the model after the (t + 1)-th round of training, w t Is the model after the t-th round of training, Represents the loss function; u i Represents the data carried by the i-th edge device within the group;

[0025] Repeat the training multiple times until the loss function converges to obtain the models for each group.

[0026] The present invention also provides a highly generalized model for new edge devices based on a clustering algorithm constructed by the method described in the above solution.

[0027] The method of the present invention classifies edge devices according to their environmental and data characteristics through clustering analysis to train a general model with good generalization ability on different devices. In this way, newly added devices can quickly benefit from the models suitable for their environments. Even when lacking historical data at the first time of joining, they can achieve efficient application and adaptation in a short time. Such a highly generalized model can not only effectively improve the performance of new devices in the edge computing network, but also adapt to diverse edge environments by enhancing the robustness of the model, thereby improving the scalability and reliability of the entire network.

[0028] For devices newly added to the edge computing network, these new devices face adaptation problems due to unknown factors such as environment and tasks. Since it is impossible to accurately understand the environmental and task characteristics of new devices, directly using existing trained models often fails to ensure good results. Therefore, the method of the present invention first clusters existing edge devices, groups them according to data characteristics and task scenarios, and then extracts representative devices from each group to construct a group containing different categories of data for training.

[0029] This training method ensures that the data of each category of device is effectively represented within the group and balances the proportions at the same time to ensure that the model can have high generalization ability based on seeing different categories of data. For newly added edge devices, using this generalized model trained based on multi-category data can obtain good performance even when their environmental and task characteristics are unknown.

[0030] The model obtained by the present invention can quickly adapt to new devices, reduce the cold start time, and provide strong robustness and adaptability, ensuring that new devices can achieve effective application at the initial stage. This is crucial for the continuous scalability in the edge computing network, can cope with the challenges of large differences in device environments and unclear task requirements, and ensure that each newly added device can smoothly integrate into the overall network and provide stable services. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flowchart of the method for constructing a highly generalized model for new edge devices based on a clustering algorithm of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0032] The following will elaborate on a highly generalized model for new edge devices based on a clustering algorithm provided according to a preferred embodiment of the present invention. Through this embodiment, we can fully understand and practice the technical solution of the present invention.

[0033] A method for constructing a highly generalized model for new edge devices based on a clustering algorithm of the present invention includes the following steps:

[0034] Based on the environment where each edge device is located, generate data feature points D symbolizing its own data characteristics; specifically: count the quantity of each type of landform in the surrounding environment of each edge device, and combine them to form high-dimensional data feature points of data characteristics symbolizing its own features; for example: collect the deployment environment of the edge device, such as how many utility poles are around, and whether each utility pole is in farmland, the city or the forest. Then record, for example, that there are 5 farmlands, 3 cities, and 0 forests among the utility poles that need to be inspected near Edge Device A, and record A as (5, 3, 0). Edge Device B has 0 farmlands, 0 cities, and 9 forests, so 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.

[0035] Classify the data feature points D using the Euclidean distance; specifically:

[0036] Calculate the Euclidean distance between each edge device and the data origin D of the central server x The formula is:

[0037]

[0038] Where: S i represents the Euclidean distance between the i-th edge device and the data origin D of the central server x ; D i represents the data feature point of the i-th edge device;

[0039] According to the Euclidean distance between each edge device and the data origin D of the central server x , use the K-means clustering algorithm to classify each edge device to obtain different categories:

[0040] C i ={d(S i , S j ) ≤ ∈}.

[0041] Extract edge devices of different categories to form a group containing data of all categories; specifically:

[0042] Statistically count the total amount of data carried by each edge device within each category;

[0043] Select edge devices from each category according to the total amount of data carried to form a group.

[0044] Train the initial model on each group until the model converges to obtain the group model; specifically:

[0045] Use the data within each group to train the initial model as:

[0046]

[0047] where: w t+1 is the model after the (t + 1)-th round of training, w t is the model after the t-th round of training, u i represents the data carried by the i-th edge device within the group, represents the loss function;

[0048] Repeat the training multiple times until the loss function converges to obtain the group models.

[0049] Aggregate the group models to obtain a highly generalized model.

[0050] Example:

[0051] 1. Initialize parameters and data dimensionality reduction: Generate data feature points D based on the environment of each edge device. D x is the data origin of the central server. Using the Euclidean distance S as a symbol of the distance of data feature points, cluster the data features of edge devices into C, group the edge devices into G, and determine the number of communication rounds as T.

[0052] When t = 0 communication rounds, perform the following steps. The central server sends down the reference system data feature points D x ;

[0053] Collect the data features D of the environment where each edge device is located;

[0054] Here, clustering pays more attention to data differences, so Euclidean distance is used for clustering. The edge nodes calculate the Euclidean distance S between their own data features D and D x (the data origin of the central server) and upload it to the central server;

[0055] 2. Cluster according to Euclidean distance:

[0056] The central server classifies these devices using the K-means clustering algorithm according to the distance S between each edge device and the data origin of the central server, obtaining different categories C. First, the central server calculates the distances between each edge device and other devices, and selects appropriate clustering centers as the initial central points based on these distances. Then, using the iterative process of the K-means algorithm, the edge devices are gradually classified into the closest clustering centers, and the positions of each central point are updated until convergence. Finally, the central server obtains a set of categories C, each category containing edge devices with relatively close distances;

[0057] 3. Divide into groups:

[0058] After completing the K-means clustering, the central server selects edge devices from each category and counts the total amount of data carried by the devices within each category. Through these statistical information, the central server can ensure that the total amount of data carried by the devices in each category is basically the same, maintaining balance as much as possible to avoid too much or too little data in a certain category. Then, according to the balance requirement of the total data volume, the central server selects an appropriate number of devices from each category and divides them into different groups G. Each group G contains devices from all categories and can ensure that the total data volume of each group is close to equilibrium

[0059] 4. Train a highly generalized model:

[0060] When t = 1, the central server distributes the initial model to each group;

[0061] When t = 2 to t = T - 1, each edge device within each group trains the initial model using its own data;

[0062] After training is completed, all local models within the group are uploaded to the central server;

[0063] The central server aggregates all local models within the group into a new group model and distributes it to the internal nodes of group G;

[0064] The central server collects all the final group models of all groups and then aggregates them to obtain a highly generalized model. New edge nodes can use this model to greatly enhance their generalization ability.

[0065] The method proposed above reduces the dimensionality of the data features of the edge nodes and then clusters. To ensure the data balance of each group, it is more inclined to the differences of edge devices, so the Euclidean distance is used. After classification, group G is obtained. The data within group G is balanced in terms of data volume and rich in data categories. The entire process is parallel computed to maximize the computing power of the edge devices, making the group model have extremely strong generalization ability to cope with various situations that new nodes may encounter.

Claims

1. A method for constructing a high-generalization model for supporting new edge devices based on a clustering algorithm, characterized in that: The following steps are involved: Based on the environment in which each edge device is located, generate data feature points D that represent its own data characteristics; Classify the data feature points D using Euclidean distance; Extract different categories of edge devices to form a group containing all categories of data; Train the initial model on the group until the model converges to obtain the group model; The models of each group are aggregated to obtain a highly generalized model.

2. According to the method for constructing a new edge device supporting high generalization model based on clustering algorithm according to claim 1, it is characterized in that: The data feature points D are classified using Euclidean distance, specifically: Calculate the data origin D of each edge device and the central server x The Euclidean distance between them is: Where: S i Represents the data origin D between the i-th edge device and the central server x The Euclidean distance between i Represents the data feature point of the i-th edge device; According to the data origin D of each edge device and the central server x The Euclidean distance is used to classify each edge device using the K-means clustering algorithm to obtain different categories: C i ={d(S i ,S j )≤∈}。 3. According to the method for constructing a new edge device supporting high generalization model based on clustering algorithm according to claim 1, it is characterized in that: Different categories of edge devices are extracted to form a group containing all categories of data, specifically: Count the total amount of data carried by each edge device in each category; Edge devices are selected from each category to form a group based on the total amount of data they carry.

4. According to claim 3, a method for constructing a new edge device supporting high generalization model based on a clustering algorithm is characterized in that: When selecting edge devices, the total amount of data carried by all edge devices in each group needs to be balanced.

5. The method for constructing a new edge device supporting high generalization model based on a clustering algorithm according to claim 1, characterized in that: Train the initial model on the group until the model converges to obtain the group model, specifically: The initial model trained using the data from each group is: Where: w t+1 is the model after t+1 rounds of training, w t is the model after the tth round of training, represents the loss function; u i Represents the data carried by the i-th edge device in the group; Repeat the training multiple times until the loss function converges and obtain the models of each group.

6. A highly generalized model for supporting new edge devices based on a clustering algorithm constructed using any of the methods described in items 1-5.

Images