A location-aware federated learning training method and apparatus

Abstract

This application relates to the fields of intelligent connected vehicles and edge computing technology, specifically, to a location-aware federated learning (VEFL) training method and apparatus, which can, to some extent, solve the problem that existing federated learning methods cannot be well adapted to VEFL application scenarios. The method includes: acquiring the historical trajectory and resource information of the client, the resource information including the client's computing and communication capabilities; in each round of training, performing trajectory prediction based on the historical trajectory to obtain the trajectory prediction result for each client, and estimating the response latency of each client based on the resource information and the trajectory prediction result; based on the response latency and trajectory prediction result of each client, selecting a corresponding sample client set and a test client set for each cluster; assigning corresponding cluster models to the sample client set and the test client set for local training and testing tasks, repeating the above process until the training termination condition is met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent networked vehicles and edge computing, in particular to a federated learning training method and device based on location awareness. BACKGROUND

[0002] Federated Learning (FL) is a new machine learning paradigm that has great potential in data privacy protection and resource utilization. Through Client-Edge Federated Learning (VEFL), intelligent clients can train models locally and only share the update results of the models without transmitting sensitive raw data. This not only protects user privacy, but also reduces the delay and overhead of data transmission, improving the accuracy and real-time performance of the model.

[0003] In the application scenario of VEFL, the data collected by clients from different geographical areas has different statistical characteristics, and the high mobility of clients further exacerbates the statistical heterogeneity, that is, the data distribution collected by clients may change with their location changes. Therefore, existing solutions propose to bind the FL model with geographical areas and train the same FL model within an area where the data follows a highly similar distribution to alleviate the adverse effects of statistical heterogeneity.

[0004] However, in VEFL, clients are both data collectors and model users. Due to the change of client location, if a cluster model is assigned to a client only according to its current location, the data collected by the client in the past may not match the data distribution suitable for the model, resulting in bias in model training. If a cluster model is assigned to a client only according to the data it has collected, the model may not be suitable for the data collected by the client in the present and future, resulting in unreasonable test results and ideal application accuracy, affecting the real-time performance of the model. Moreover, different clients have different computing and communication capabilities, i.e., system heterogeneity, and resource-poor clients are likely to become "stragglers", greatly reducing the training efficiency. Therefore, the existing federated learning method cannot adapt well to the VEFL application scenario. SUMMARY

[0005] To solve the problem that the existing federated learning method cannot adapt well to the VEFL application scenario, the present application provides a federated learning training method and device based on location awareness to adapt to mobile scenarios and improve training efficiency.

[0006] Embodiments of the present application are implemented as follows:

[0007] In a first aspect, the present application provides a federated learning training method based on location awareness, comprising:

[0008] acquire a historical trajectory and a resource condition of the client, the resource condition including a computing capability and a communication capability of the client;

[0009] In each round of training, trajectory prediction is performed according to the historical trajectory to obtain a trajectory prediction result of each client, and a response delay of each client is estimated according to the resource condition and the trajectory prediction result;

[0010] Based on the response delay and the trajectory prediction result of each client, a corresponding set of sampling clients and a corresponding set of test clients are selected for each cluster;

[0011] The set of sampling clients and the set of test clients are assigned to a corresponding cluster model for local training and testing tasks, and the above process is repeated until an end condition of training is met.

[0012] In a possible implementation, in each round of training, trajectory prediction is performed according to the historical trajectory to obtain a trajectory prediction result of each client, and a response delay of each client is estimated according to the resource condition and the trajectory prediction result, and the method further includes:

[0013] Trajectory prediction is performed based on the historical trajectory to obtain a trajectory prediction result;

[0014] The size of a training data set of the client is estimated according to the historical trajectory and the trajectory prediction result;

[0015] Based on the size of the training data set, the computing capability, the model size and the communication capability, a response delay of each client is estimated.

[0016] In a possible implementation, the method of selecting, based on the response delay and the trajectory prediction result of each client, a corresponding set of sampling clients and a corresponding set of test clients for each cluster further includes:

[0017] The entire training area is divided into a plurality of non-overlapping sub-areas, and each sub-area is bound to a dedicated cluster model;

[0018] According to the historical trajectory and the trajectory prediction result of the client, a set of training clients is determined for each cluster;

[0019] According to the response delay and a sampling ratio of the client, a set of sampling clients of each cluster is determined in the set of training clients.

[0020] In a possible implementation, the training method further includes a cluster model allocation strategy, and the cluster model allocation strategy includes:

[0021] determining a sub-area where the client will stay in future according to the trajectory prediction result of the client;

[0022] assigning the client to a test client set of a cluster of sub-areas where the client will stay in future.

[0023] In a possible implementation, the cluster model assignment strategy comprises:

[0024] determining whether the client is a client that has just entered the training area according to the historical trajectory of the client, and the client will not participate in the training task immediately if the client is a client that has just entered the training area.

[0025] In a possible implementation, the cluster model assignment strategy comprises:

[0026] determining whether the client will leave the training area soon according to the trajectory prediction result of the client, and the client will not participate in the test task if the client will leave the training area soon.

[0027] In a possible implementation, the cluster model assignment strategy comprises:

[0028] when each client in the sampling client set receives a corresponding cluster model to perform a training task, the client filters data collected in the range of the cluster as a training data set;

[0029] when each client in the test client set receives a corresponding cluster model to perform a test task, the client filters data collected recently as a test data set.

[0030] In a possible implementation, the cluster model assignment strategy comprises:

[0031] in each round of training, the training client set is divided into multiple layers by a clustering algorithm according to the response delay of each client, and the response delay between clients in each layer is close;

[0032] after selecting a layer of the training client set, a group of clients are randomly selected as a sampling client set to perform a training task according to a sampling ratio.

[0033] In a possible implementation, the cluster model assignment strategy comprises:

[0034] determining a stay time of the client in each cluster according to the historical trajectory;

[0035] determining the cluster with the most client stay time as the current participating training cluster of the client.

[0036] In a second aspect, the present application provides a position-aware federated learning training device, comprising:

[0037] a data collection module configured to acquire historical trajectories and resource conditions of the clients, wherein the resource conditions include computing capabilities and communication capabilities of the clients;

[0038] a data processing module configured to, in each round of training, perform trajectory prediction according to the historical trajectories to obtain trajectory prediction results of each client, and estimate response delays of each client according to the resource conditions and the trajectory prediction results;

[0039] a data division module configured to select, based on the response delays and the trajectory prediction results of each client, a corresponding set of sampling clients and a corresponding set of test clients for each cluster;

[0040] a training and testing module configured to assign corresponding cluster models to the set of sampling clients and the set of test clients for local training and testing tasks, and repeat the above process until an end condition of training is met.

[0041] The technical solutions provided by the present application can achieve at least the following beneficial effects:

[0042] The position-aware federated learning training method and device provided by the present application can predict the future position trajectories of the clients based on their historical trajectories by taking advantage of the predictability of the movement of the clients, thereby determining the cluster models most suitable for the distribution of the data sets of each client and estimating the response delays of each client, alleviating the adverse effects of device heterogeneity, further improving the training efficiency, and taking into account the dual identity of the clients in VEFL, i.e., the clients are both collectors of data and users of models, and assigning appropriate cluster models for testing and application to the clients according to the historical trajectories and the trajectory prediction results, to ensure the effectiveness of the testing and the accuracy of the application. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without any creative labor.

[0044] Figure 1 is a flowchart of a position-aware federated learning training method according to an exemplary embodiment of the present application.

[0045] Figure 2 is a schematic diagram of an implementation procedure of a location-aware based federated learning training algorithm according to an example embodiment of the present application;

[0046] Figure 3 is a schematic diagram of federated learning training area and cluster division according to an example embodiment of the present application;

[0047] Figure 4 is a schematic diagram of the first three rounds of client location and sampling state according to an example embodiment of the present application;

[0048] Figure 5 is a schematic diagram of the structure of a location-aware based federated learning training device according to an example embodiment of the present application.

[0049] Reference signs:

[0050] 1, data acquisition module; 2, data processing module; 3, data division module; 4, training and testing module. DETAILED DESCRIPTION

[0051] In order to make the purposes, embodiments and advantages of the present application clearer, the following will combine the drawings in the example embodiments of the present application to clearly and completely describe the example embodiments of the present application. Obviously, the described example embodiments are only some of the embodiments of the present application, not all the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not intended to limit the present application.

[0052] It should be noted that the brief description of the terms in the present application is only for the convenience of understanding the following described embodiments, and is not intended to limit the embodiments of the present application. Unless otherwise specified, these terms should be understood according to their ordinary and general meanings.

[0053] The terms "first", "second", "third" and the like in the specification and claims of the present application and the above drawings are used to distinguish similar or similar objects or entities, and do not necessarily mean to limit the specific order or sequence, unless otherwise specified. It should be understood that the terms used in this way can be interchanged under appropriate circumstances.

[0054] The terms "include" and "have" and any variations thereof are intended to cover but not exclusive inclusion, for example, a product or device including a series of components does not necessarily limit to all the components clearly listed, but can include other components not clearly listed or inherent to these products or devices.

[0055] In order to facilitate the technical solutions of the application, the following first describes some concepts related to the present application.

[0056] Before the position-aware federated learning training method provided by the embodiments of the present application is explained and described, the application scenarios and implementation environments of the embodiments of the present application are introduced.

[0057] Federated learning (FL) is a new machine learning paradigm that has great potential in data privacy protection and resource utilization. Through client edge federated learning (VEFL), intelligent clients can train models locally and only share the update results of the models without transmitting sensitive raw data. This not only protects user privacy, but also reduces the delay and overhead of data transmission, improves the accuracy and real-time performance of the model.

[0058] In the application scenario of VEFL, the data collected by clients from different geographical areas has different statistical characteristics, and the high mobility of clients further exacerbates the statistical heterogeneity, that is, the data distribution collected by the client may change with its location, therefore, the existing scheme proposes to bind the FL model with the geographical area, and train the same FL model within the area where the data follows a highly similar distribution to alleviate the adverse effects of statistical heterogeneity.

[0059] However, in VEFL, the client is both a data collector and a model user. Due to the change of the client's location, if the client is only assigned a cluster model according to the current location, the data collected by the client in the past may not match the data distribution suitable for the model, resulting in introducing bias in model training; if the client is only assigned a corresponding cluster model according to the data the client has collected, the model may not be suitable for the data the client collects currently and in the future, and cannot obtain reasonable test results and ideal application accuracy, affecting the real-time performance of the model; and different clients have different computing and communication capabilities, that is, system heterogeneity, and resource-poor clients are easy to become "stragglers", greatly reducing the training efficiency; making the existing federated learning method not well adapted to the VEFL application scenario.

[0060] Based on this, the present application provides a position-aware federated learning training method and device, which guides the training and application process according to the resource situation and movement prediction result of the client to adapt to the actual mobile scenario and improve the training efficiency.

[0061] Next, the technical solutions of the present application and how the technical solutions of the present application solve the above technical problems will be described in detail through embodiments and in combination with the drawings. Each embodiment can be combined with each other, and the same or similar concepts or processes can not be described again in some embodiments. Obviously, the described embodiments are part of the embodiments of the present application, not all.

[0062] Figure 1 FIG. 1 is a flowchart of a position-aware federated learning training method according to an example embodiment of the present application.

[0063] In an example embodiment, as shown in FIG. 2, a position-aware federated learning training method is provided, which can include the following steps: Figure 1

[0064] Step 100: Obtain the historical trajectory and resource situation of the client, wherein the resource situation includes the computing and communication capabilities of the client.

[0065] Step 200: In each round of training, perform trajectory prediction according to the historical trajectory to obtain the trajectory prediction result of each client, and estimate the response delay of each client according to the resource situation and the trajectory prediction result.

[0066] Step 300: Based on the response delay and the trajectory prediction result of each client, select a corresponding set of sampling clients and a set of test clients for each cluster.

[0067] Step 400: Assign the corresponding cluster model to the set of sampling clients and the set of test clients for local training and testing tasks, and repeat the above process until the end condition of training is met.

[0068] As can be seen, some embodiments of the present application take into account that the geographical dispersion of data and the mobility of clients further aggravate statistical heterogeneity, and use the predictability introduced by mobility to issue appropriate cluster models to clients to perform training or testing tasks, thereby alleviating the adverse effects of statistical heterogeneity, and for the system heterogeneity between clients in the VEFL application scenario, a position-aware hierarchical sampling strategy is designed, which selects a set of sampling clients with similar response delays for each cluster according to the resource situation and location trajectory of the client, to improve training efficiency.

[0069] In a possible implementation, considering that the client is both a data collector and a model user, and due to its high mobility, it can cross different clusters covered by the area in a short time, therefore, the cluster model it participates in training can be different from the cluster model used for testing and application.

[0070] To this end, some embodiments of the present application make full use of the predictability of mobility by using a position prediction method.

[0071] ​Assuming that the entire training area is divided into K clusters, for each federated learning training round r, if the server Θ determines that the client has previously stayed more frequently within the coverage range of the kth cluster according to the historical and predicted location trajectory of the client, it will be allocated to the client set N k (r) participating in the training process of the kth cluster in the rth round

[0072] If it is most likely to move within the coverage range of the kth cluster in the future, the server Θ will divide it into the client set N k (r) participating in the testing or applying the model w of the kth cluster in the rth round.

[0073] When the server Θ finds that the client i has just entered the entire FL training area for a short time, that is, i has collected little data within the area, it will not let i participate in the training task immediately.

[0074] In addition, when the server Θ predicts that the client i is about to leave the entire FL training area, it will not let i participate in the subsequent testing task to avoid introducing errors.

[0075] On this basis, in order to further reduce the influence of statistical heterogeneity brought by mobility, when each client i∈N k (r) receives the kth cluster model w k (r-1) is ready for local training, it will filter the data collected within the coverage area of the kth cluster as the training data set.

[0076] While the client in N k (r) is more reasonable and more in line with the actual application scenario.

[0077] Figure 2 is the implementation flowchart of the location-aware federated learning training algorithm according to an example embodiment of the present application.

[0078] In a possible implementation manner, as shown in Figure 2 , in the training process of each cluster model, some embodiments of the present application adopt a hierarchical sampling strategy,

[0079] that is, a group of clients S k (r) with similar response delays T i (r) are selected in N k (r) to perform the actual training task, so as to alleviate the adverse effects of system heterogeneity.

[0080] The specific implementation flow of the hierarchical sampling strategy is as follows: in each round r, the server Θ first obtains the historical trajectory Ji (t) and current computing capability C i (t) and communication capability θ i (t), and determines the training data set size D of the estimated client i according to the historical and predicted position trajectory i (t);

[0081] Then, according to the data set size D i (t), the computing capability C i (t), the model size a and the communication capability θ i (t), the delay size T i (r) of each client i is estimated

[0082] Then, the clustering algorithm is used to divide N i (r) into K' layers according to T k (r), wherein the delay difference between the clients in the same layer is relatively small;

[0083] Finally, Θ selects one layer according to the sampling ratio σ to randomly select a group of clients S k (r) to perform the training task. The parameter K' and how to select one layer in the K' layers for random sampling can be adjusted according to different levels of device heterogeneity.

[0084] It should be understood that although each step in the flowchart involved in the above embodiment is displayed in sequence as indicated, these steps are not necessarily executed in the order as indicated. Unless explicitly stated herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other orders. Moreover, at least part of the steps in the flowchart involved in the above embodiment can include multiple steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution sequence of these steps or stages is not necessarily sequential, but can be executed in rotation or alternation with at least part of other steps or steps or stages in other steps.

[0085] Corresponding to the above-mentioned embodiments of the location-aware federated learning training method, using the same technical concept, the present application also provides embodiments of a location-aware federated learning training device.

[0086] Figure 5 is a structural schematic diagram of a location-aware federated learning training device according to an example embodiment of the present application.

[0087] In an example embodiment, as Figure 5 shown, the location-aware federated learning training device comprises:

[0088] Data collection module 1: used to obtain the historical trajectory and resource situation of the client, the resource situation including the computing and communication capabilities of the client;

[0089] Data processing module 2: used to perform trajectory prediction according to the historical trajectory in each round of training to obtain the trajectory prediction result of each client, and to estimate the response delay of each client according to the resource situation and the trajectory prediction result;

[0090] Data division module 3: used to select the corresponding sample client set and test client set for each cluster based on the response delay and the trajectory prediction result of each client;

[0091] Training and testing module 4: used to assign the sample client set and the test client set to the corresponding cluster model for local training and testing tasks, and to repeat the above process until the end condition of training is met.

[0092] The specific limitations of the location-aware federated learning training device can be referred to the limitations of the location-aware federated learning training method described above, which will not be repeated here. The various modules in the above location-aware federated learning training device can be realized by software, hardware and their combinations in whole or in part. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to the above modules.

[0093] Working process:

[0094] In order to make the technical solutions and advantages of the embodiments of the present application more clear, the technical solutions in the embodiments of the present application will be described clearly and completely below in combination with an embodiment of the present application. Obviously, the described embodiments are a part of the embodiments of the present application, not all the embodiments.

[0095] Figure 3 is a schematic diagram of federated learning training area and cluster division shown by an exemplary embodiment of the present application, Figure 4 is a schematic diagram of the first three rounds of client position and sampling state shown by an exemplary embodiment of the present application.

[0096] As Figure 3As shown, the entire training area is divided into two parts: the coverage area M1 of the first cluster and the coverage area M2 of the second cluster. Initially, there are 8 cars in the entire training area, numbered 1, 4, 5, 8, 3, 6, 7, and 9. In each round, the client first uploads its resource status and historical trajectory to the server. After the server predicts the location of the client, it estimates the response latency of each client. Some embodiments of this application assume that the actual latency of each car is as shown in Table 1. The car number also represents the sorting order of its latency. The smaller the number, the lower the latency of the corresponding car. It is also assumed that the latency of the car remains unchanged in each round.

[0097] Table 1. Examples of client latency times

[0098] Client 1 2 3 4 5 6 7 8 9 Latency time 1 2 3 4 5 16 17 18 19

[0099] Next, some embodiments of this application will be used as a simple example to illustrate the specific implementation process of each round of cluster model allocation and hierarchical sampling strategy.

[0100] like Figure 3 As shown, in the first round, the server learned from historical trajectories that cars numbered 1, 4, 6, and 8 had been staying within the coverage area M1 of the first cluster, while cars numbered 3, 5, 7, and 9 had been staying within the coverage area M2 of the second cluster. Based on position prediction, the server knew that car number 3 was about to enter M1 and car number 4 was about to enter M2, thus determining N. 1 (1) = {1, 4, 6, 8} N 2 (1) = {3, 5, 7, 9} Next, the server uses the K-Means clustering algorithm to cluster N 1 (1) and N 2 (1) The system is divided into two layers, with the layers having lower latency ({1,4} and {3,5}) and the layers having higher latency ({6,8} and {7,9}). In each subsequent round, some embodiments of this application will alternately select the layer with higher or lower latency for sampling. Therefore, in the first round, some embodiments of this application select the layer with lower latency for random sampling. Since the number of clients is small, the sampling ratio σ is set to 100%, i.e., S 1 (1) = {1, 4}, S 2 (1) = {3, 5}. Afterwards, the server will use the initial cluster model w 1 The range of (0) and M1, w 2 The ranges of (0) and M2 are respectively sent to S 1 (1) and S 2 (1), sampling client i∈S1 (1)∪S 2 (1) Select data collected within the coverage area of ​​the cluster as the training dataset to train the corresponding cluster model. After training, transfer the new local model w i (1) Send to the server, the server has a local model for each cluster k. Aggregation is performed to obtain a new cluster model w. 1 (1) and w 2 (1) Then send them to each and Conduct the test, and then upload the results to the server after the test is completed.

[0101] like Figure 4 As shown, in the second round, the server, based on historical trajectories, knows that cars numbered 1, 3, 6, and 8 were in M1 over a period of time, while cars numbered 4, 5, 7, and 9 were in M2. Car number 2 had just entered M1, and based on position prediction, the server knows that car number 1 is about to leave M1. Therefore, N is determined. 1 (2) = {1, 3, 6, 8} N 2 (2) = {4, 5, 7, 9} Next, the server uses the K-Means clustering algorithm to cluster N 1 (2) and N 2 (2) Divide into two layers, with the lower latency layers being {1, 3} and {4, 5}, and the higher latency layers being {6, 8} and {7, 9}. In the second round, some embodiments of this application select the higher latency layer for random sampling with σ = 100%, i.e., S 1 (2) = {6, 8}, S 2 (2) = {7, 9}. Afterwards, the server will send model w... 1 (1) and the range of M1, w 2 (1) and the range of M2 are respectively sent to S 1 (2) and S 2 (2), sampling client i∈S 1 (2)∪S 2 (2) Use the filtered data as the training dataset to train the corresponding cluster model. After training, transfer the new local model w. i (2) Send to the server, the server has a local model for each cluster k. Aggregation is performed to obtain a new cluster model w. 1 (2) and w 2 (2) Then send them to each and Conduct the test, and then upload the results to the server after the test is completed.

[0102] As Figure 4 shown, in the third round, the server learns from the historical trajectories that in the past time, the trolleys numbered 2, 3, 6 and 8 stay in M1, the trolleys numbered 4, 5, 7 and 9 stay in M2, and the trolley numbered 1 has left M1, and according to the position prediction result, the eight trolleys numbered 2-9 will still move in the current area in the future, thereby determining N 1 (3) = {2, 3, 6, 8}, N 2 (3) = {4, 5, 7, 9}, Then, the server also uses the clustering algorithm to divide N 1 (3) and N 2 (3) into 2 layers respectively, where the lower delay layer is {2, 3} and {4, 5} respectively, and the higher delay layer is {6, 8} and {7, 9} respectively. In the third round, some embodiments of the application select the lower delay layer to randomly sample with σ = 100%, that is, S 1 (3) = {2, 3}, S 2 (3) = {4, 5}.

[0103] The subsequent training and testing steps are similar to the first two rounds, and the above steps are repeated every round until the training cutoff time is reached, the test result reaches the expected accuracy, or the cluster model converges.

[0104] The technical features of the above embodiments can be combined in any way. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, but as long as the combination of the technical features does not exist contradictory, it should be considered as the scope of the present disclosure.

[0105] The above embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be noted that for those skilled in the art, without departing from the concept of the present application, some modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of the patent of the present application should be subject to the appended claims.

Claims

1. A location-aware based federated learning training method, characterized in that, The method comprises the following steps: acquiring historical trajectories and resource conditions of the clients, wherein the resource conditions comprise computing capabilities and communication capabilities of the clients; in each round of training, trajectory prediction is performed according to the historical trajectories to obtain trajectory prediction results of each client, and response delays of each client are estimated according to the resource conditions and the trajectory prediction results, which comprises the following steps: trajectory prediction is performed according to the historical trajectories to obtain trajectory prediction results; the training data set size of the client is estimated according to the historical trajectories and the trajectory prediction results; and the response delay of each client is estimated based on the training data set size, the computing capability, the model size and the communication capability; based on the response delay and the trajectory prediction result of each client, a corresponding set of sampling clients and a corresponding set of test clients are selected for each cluster, which comprises the following steps: the entire training area is divided into a plurality of non-overlapping sub-areas, and each sub-area is bound to a dedicated cluster model; a set of training clients is determined for each cluster according to the historical trajectories and the trajectory prediction results of the clients; and a set of sampling clients of each cluster is determined in the set of training clients according to the response delay and a sampling ratio of the clients; the set of sampling clients and the set of test clients are allocated corresponding cluster models to perform local training and testing tasks, and the above process is repeated until the end condition of training is met. 2.The location-aware based federated learning training method of claim 1, wherein, The training method further comprises a cluster model allocation strategy, and the cluster model allocation strategy comprises the following steps: determining the sub-area where the client will stay in the future according to the trajectory prediction result of the client; allocating the client to the set of test clients of the cluster in the sub-area where the client will stay in the future. 3.The location-aware based federated learning training method of claim 2, wherein, The cluster model allocation strategy comprises the following steps: determining whether the client is a client that has just entered the training area according to the historical trajectory of the client, and the client that has just entered the training area will not immediately participate in the training task. 4.The location-aware based federated learning training method of claim 2, wherein, The cluster model allocation strategy comprises the following steps: determining whether the client is a client that will leave the training area according to the trajectory prediction result of the client, and the client that will leave the training area will not participate in the testing task. 5.The location-aware based federated learning training method of claim 2, wherein, The cluster model allocation strategy comprises the following steps: when each client in the set of sampling clients receives a corresponding cluster model to perform a training task, the data collected within the coverage range of the cluster is screened as a training data set; when each client in the set of test clients receives a corresponding cluster model to perform a testing task, the data recently collected by the client is screened as a testing data set. 6.The location-aware based federated learning training method of claim 1, wherein, The method of allocating corresponding cluster models to the set of sampling clients and the set of test clients to perform local training and testing tasks, and repeating the above process until the end condition of training is met, further comprises the following steps: in each round of training, the set of training clients is divided into multiple layers by a clustering algorithm according to the response delay of each client, and the response delays between the clients in each layer are close to each other; after selecting a layer of the set of training clients, a group of clients are randomly selected as the set of sampling clients to perform a training task according to a sampling ratio.

7. The location-aware based federated learning training method of claim 1, wherein, The method further comprises determining, according to the historical trajectory of the client, a cluster in which the client is currently participating in training, and further comprising: determining, according to the historical trajectory, a stay time of the client in each cluster; determining a cluster in which the client stays the longest time as the cluster in which the client is currently participating in training.

8. A location-aware based federated learning training apparatus, comprising: Comprise: a data acquisition module for acquiring the historical trajectory and resource situation of the client, the resource situation comprising the computing ability and communication ability of the client; a data processing module for, in each round of training, performing trajectory prediction according to the historical trajectory to obtain a trajectory prediction result of each client, and estimating a response delay of each client according to the resource situation and the trajectory prediction result, comprising: performing trajectory prediction based on the historical trajectory to obtain a trajectory prediction result; estimating the size of the training dataset of the client according to the historical trajectory and the trajectory prediction result; and estimating the response delay of each client based on the training dataset size, computing ability, model size and communication ability; a data division module for selecting a corresponding sample client set and test client set for each cluster based on the response delay and the trajectory prediction result of each client, comprising: dividing the entire training area into a plurality of non-overlapping sub-areas, each sub-area being bound to a dedicated cluster model; determining a training client set for each cluster according to the historical trajectory and the trajectory prediction result of the client; and determining a sample client set for each cluster from the training client set according to the response delay and a sampling ratio of the client; a training and testing module for assigning the sample client set and the test client set to a corresponding cluster model for local training and testing tasks, and repeating the above process until a termination condition of training is met.

Images