Construction method of self-adaptive transmission content selection system for Internet of Vehicles

Abstract

The invention relates to the technical field of Internet of Vehicles and distributed machine learning, and discloses an Internet of Vehicles-oriented self-adaptive transmission content selection system construction method, which comprises the steps of initialization setting, local model training and interaction content preparation, information interaction and content transmission, model aggregation updating and updating termination. According to the method, an adaptive transmission content selection algorithm is adopted, so that each working node can dynamically sense a real-time channel state and node resources, transmission model gradient or lightweight data prototype characteristics are intelligently selected, and the communication load is reduced while the model performance is guaranteed; meanwhile, an information caching mechanism is introduced, a least recently used caching updating mechanism is designed, and a spatio-temporal information transmission network is formed by means of high-speed movement of vehicles, so that multi-hop propagation of information in the network is realized, the information interaction coverage range is effectively expanded, and the model performance and training efficiency in a mobile scene are improved.

Description

Technical Field

[0001] The present invention relates to the field of Internet of Vehicles and distributed machine learning technology, and more specifically, to a method for constructing an adaptive transmission content selection system for the Internet of Vehicles. Background Art

[0002] With the rapid development of mobile IoT and edge computing, training data sources have been shifting from cloud data centers to end devices. Federated learning has made an irreplaceable contribution to enabling training using end-device data while protecting privacy. However, the single-level tree topology of centralized federated learning places enormous communication pressure on parameter servers. Furthermore, with the massive growth of end devices, parameter servers have an increasing impact on the entire system, becoming a significant communication bottleneck. For example, using 100 end devices to collaboratively train a VGG16 model (approximately 138MB), a single global interaction consumes over 13GB of bandwidth from the federated learning parameter server, which can easily cause network congestion when wirelessly communicating between mobile devices. Although some research has proposed hierarchical federated learning, introducing edge servers to alleviate the pressure on parameter servers in centralized federated learning, both parameter servers and edge servers can be considered aggregation centers for federated learning, making them highly important and easily targeted by attacks, posing serious security risks.

[0003] Decentralized federated learning (DFL) has emerged to alleviate the communication pressure on parameter servers in centralized federated learning architectures and avoid single points of failure. The concept of DFL can be traced back to the decentralized parallel stochastic gradient descent algorithm D-PSGD, which is considered the first work to demonstrate that D-PSGD has the same convergence rate as centralized algorithms. Its underlying theory is consensus-based distributed optimization. In DFL, worker nodes exchange data through a communication topology. Each node only needs to communicate with a limited number of neighbors, rather than relying on parameter servers. Generally, the performance of decentralized distributed learning is determined by the communication topology. Higher connectivity results in less information loss in model aggregation and faster convergence, but this can lead to network congestion, long communication times, and high synchronization costs. However, by limiting the number of neighbors a node can have and compressing transmitted content, the bandwidth requirements for model aggregation can be reduced, accelerating the training process. Due to these advantages, DFL has been widely used in various fields, including military, healthcare, recommendation systems, and the Internet of Vehicles. Especially in the field of Internet of Vehicles, the performance of mobile devices (such as mobile phones, tablets, laptops, self-driving cars, drones and autonomously navigated robots) has been rapidly improving, enabling tasks combining deep learning and distributed decision-making (such as traffic light and signal recognition, road condition perception, intelligent obstacle avoidance and intelligent road planning, etc.) to be transferred to mobile terminals. This not only utilizes the computing power of mobile devices, but also protects user privacy.

[0004] However, decentralized federated learning still has at least the following problems: 1) Device heterogeneity mainly includes device computing power heterogeneity, energy consumption heterogeneity, and data distribution imbalance. Among them, computing power heterogeneity refers to the differences in computing power among devices, which will affect the speed and efficiency of model training. Energy consumption heterogeneity means that the energy of devices is limited, and frequent communications, local training, aggregation and other operations will consume a lot of battery resources. Unbalanced data distribution is a classic problem of federated learning. Especially in the decentralized training process of the Internet of Vehicles scenario, the computing power, energy and data set distribution of the devices participating in the training vary greatly. It is necessary to select appropriate neighbors to complete the training and achieve a trade-off between communication cost and model performance; 2) Time-varying networks refer to networks that change in real time and have limited and heterogeneous communication bandwidth. Given limited communication channels and network conditions, the size of the transmitted content will severely impact system training efficiency. It is crucial to strike a balance between data exchange volume and model performance within limited resources.

[0005] Based on this, the present invention provides a method for constructing an adaptive transmission content selection system for the Internet of Vehicles to solve the above-mentioned technical problems. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a method for constructing an adaptive transmission content selection system for the Internet of Vehicles.

[0007] To achieve the above object, the present invention provides the following technical solutions: A first aspect of the present invention provides a method for constructing an adaptive transmission content selection system for an Internet of Vehicles, the method comprising: S1. Initialization setup: Determine the working nodes and their characteristics in the decentralized federated learning system DFL, initialize the global local model, and set the training parameters; Among them, the characteristics include the network bandwidth, energy consumption, and computing power of the worker nodes; S2. Local model training and interactive content preparation: Monitor the characteristics of each worker node in real time, perform local model training on the local data of each worker node that meets the requirements, and synchronously generate two types of interactive resources after training. Among them, the two types of interactive resources are local model weight update information and lightweight local data category prototype information; S3, Information Interaction and Content Transmission: An adaptive transmission content selection algorithm and information caching mechanism are used to complete information exchange between each working node and its neighboring working nodes within the instantaneous communication range; S4, model aggregation update: Each working node shares the information generated by local training and the content received from neighboring working nodes to complete the update and iteration of the local model; S5. Update termination: Repeat steps S2 to S4 until the preset convergence condition is met or the maximum communication rounds are reached.

[0008] The present invention is further configured as follows: in step S1, the network bandwidth is calculated as follows: The inbound bandwidth and outbound bandwidth of worker node k in round t are expressed as and ; Assuming that the bandwidth is evenly distributed, the link The network bandwidth is , where 、 They represent the number of neighbor nodes of worker node k and worker node j in round t respectively.

[0009] The present invention is further configured as follows: in step S1, the energy consumption is calculated as follows: During the training phase, the training energy consumption of the working node k in the tth round is , where represents the number of processor cycles required for worker node k to execute a data sample, is the local dataset of device k, Representation dataset The number of samples, represents the processor cycle frequency of worker node k, is the number of local iterations, Indicates the effective capacitance coefficient of the device k calculation chip; In the model interaction stage, the training energy consumption of the working node k in the tth round is , where ,in, is the transmission rate of device k, is the bandwidth, is the background noise, is the channel gain of working node k, is the communication time of the device, The size of the data to be transferred; The total energy consumption of working node k is .

[0010] The present invention is further configured such that, in step S1, the computing capability satisfies the following constraints: , where is the link between working nodes k and j bandwidth, For the Wheel Chain bandwidth resources, is the energy budget of worker node k in round t.

[0011] The present invention is further configured as follows: in step S2, the model training process is as follows: Each worker node participating in the training uses the local training dataset to calculate its stochastic gradient; The worker nodes are then leveraged to update their local models.

[0012] The present invention is further configured as follows: in step S3, the adaptive interactive content selection algorithm specifically analyzes the following process: Each working node in the Internet of Vehicles adaptively selects the interactive content based on the current network bandwidth and the energy of the working node; Among them, the interactive content includes two categories: local model weight update information and lightweight local data category prototype information.

[0013] The present invention is further configured as follows: in step S3, during information interaction and content transmission, in the interaction phase, priority is given to sharing of local data category prototype information, and when network bandwidth permits, sharing of local model weight update information is added.

[0014] The present invention is further configured such that, in step S3, the information caching mechanism includes: Allow each worker node to store and transmit local model updates received from the worker nodes it encounters in its historical interactions; The least recently used principle (LRU) is used to update information cache, and timestamps are used for filtering.

[0015] The present invention is further configured as follows: in step S4, the update iteration formula of the local model is: , where represents the subgraph containing worker node k and its neighbors, Indicates the aggregation rule used, is the local model weight of worker node k after aggregation in round t, is the set of neighboring working nodes of working node k, represents the candidate local model to be aggregated by worker node i in round t

[0016] The second aspect of the present invention: An electronic device is also provided, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, the method for constructing the above-mentioned adaptive transmission content selection system for the Internet of Vehicles is implemented.

[0017] The third invention of the present invention: also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed, it implements the method for constructing the above-mentioned adaptive transmission content selection system for the Internet of Vehicles.

[0018] Compared with the prior art, the present invention has the following beneficial effects: The present invention provides a method for constructing an adaptive transmission content selection system for the Internet of Vehicles. The method adopts an adaptive transmission content selection algorithm, so that each working node can dynamically perceive the real-time channel status and node resources, and intelligently select the transmission model gradient or lightweight data prototype features, so as to reduce the communication load while ensuring the model performance. At the same time, an information caching mechanism is introduced, and a least recently used cache update mechanism is designed. With the help of the spatiotemporal information transmission network formed by the high-speed movement of vehicles, multi-hop propagation of information in the network is realized, the coverage of information interaction is effectively expanded, and the model performance and training efficiency in mobile scenarios are improved. The method can effectively overcome the influence of intermittent communication of each working node in the high-speed mobile scenario on the convergence and accuracy of the model, and can be well applied to scenarios such as the Internet of Vehicles and smart transportation. Without sharing private data, the method can improve the prediction speed of vehicles and traffic conditions, enhance the intelligence level of real-time traffic decision-making, enhance the intelligence level of traffic management, public safety and environmental monitoring in smart cities, achieve optimal allocation of resources, and have broad application value and promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A flow chart of the method provided by the present invention; Figure 2 A schematic diagram of the overall process of the method provided by the present invention; Figure 3 Generate and interact with category prototypes for the present invention; Figure 4 This is a schematic diagram showing the effect of the model test accuracy of the present invention; in, Figure 4 (a) is a schematic diagram of the model test accuracy of the experimental group VGG9-MNIST-NonIIDp=0.8; Figure 4 (b) is a schematic diagram of the model test accuracy of the experimental group VGG9-CIFAR10-NonIIDp=0.8; Figure 4(c) is a schematic diagram of the model test accuracy of the experimental group ResNet9-CIFAR10-NonIIDp=0.8; Figure 4 (d) is a schematic diagram of the model test accuracy of the experimental group ResNet9-CIFAR100-NonIIDp=0.8; Figure 5 This is a schematic diagram showing the effect of the training time used to achieve uniform accuracy in the present invention; in, Figure 5 (a) is a schematic diagram showing the effect of training time required to achieve uniform accuracy for VGG9-MNIST in the experimental group; Figure 5 (b) is a schematic diagram showing the effect of training time used by the experimental group VGG9-CIFAR10 to achieve uniform accuracy; Figure 5 (c) is a schematic diagram showing the effect of training time required for ResNet9-CIFAR10 in the experimental group to achieve uniform accuracy; Figure 5 (d) is a schematic diagram showing the effect of training time required for ResNet9-CIFAR100 in the experimental group to achieve uniform accuracy; Figure 6 This is a schematic diagram showing the effect of the communication overhead consumed in achieving uniform accuracy in the present invention; in, Figure 6 (a) is a schematic diagram showing the effect of the communication overhead consumed by the experimental group VGG9-MNIST to achieve uniform accuracy; Figure 6 (b) is a schematic diagram showing the effect of the communication overhead consumed by the experimental group VGG9-CIFAR10 to achieve uniform accuracy; Figure 6 (c) is a schematic diagram showing the effect of the communication overhead consumed by the experimental group ResNet9-CIFAR10 to achieve uniform accuracy; Figure 6 (d) is a schematic diagram showing the effect of the communication overhead consumed by the experimental group ResNet9-CIFAR100 to achieve uniform accuracy; Figure 7 A schematic structural diagram of an electronic device provided by the present invention; Figure 8 A schematic structural diagram of a computer-readable storage medium provided by the present invention. DETAILED DESCRIPTION

[0020] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0021] Example 1 See also Figure 1 、 Figure 2 and Figure 3 As shown, this embodiment provides a method for constructing an adaptive transmission content selection system for an Internet of Vehicles, the method comprising: S1. Initialization setup: Determine the working nodes and their characteristics in the decentralized federated learning system DFL, initialize the global local model, and set the training parameters; Among them, the characteristics include the network bandwidth, energy consumption and computing power of the working nodes.

[0022] As an example, assume that in a decentralized federated learning system DFL consisting of N worker nodes, each worker node k uses a local dataset Train local model weights ,The goal of DFL is to find the model weight that minimizes the weighted average loss of N clients : (1) ; in, is the local loss function of the k-th worker node, For local model weights and a single data sample losses, represents the objective function of DFL.

[0023] In this implementation, it should be noted that DFL generally includes Round local iteration and In each round of global aggregation, local model training, model exchange, and model aggregation are performed sequentially.

[0024] In addition, in this step, due to the limitation of communication capacity, the bandwidth available to each working node in the DFL is limited. Assuming that the bandwidth is evenly distributed to neighbors, the bandwidth of a link is determined by the minimum of its inbound and outbound bandwidths. Therefore, the calculation process of the network bandwidth is provided as follows: The inbound bandwidth and outbound bandwidth of worker node k in round t are expressed as and ; Assuming that the bandwidth is evenly distributed, the link The network bandwidth is: (2); Where, 、 They represent the number of neighbor nodes of worker node k and worker node j in round t respectively.

[0025] At the same time, it should be noted that due to the asymmetry of the link, the link and The bandwidth may be different.

[0026] In addition, in this step, energy consumption includes the training phase and the model interaction phase. The specific energy consumption calculation process is as follows: During the training phase, the training energy consumption of the working node k in the tth round is: (3); Where, represents the number of processor cycles required for worker node k to execute a data sample, is the local dataset of device k, Representation dataset The number of samples, represents the processor cycle frequency of worker node k, is the number of local iterations, Indicates the effective capacitance coefficient of the device k calculation chip; In the model interaction stage, the training energy consumption of the working node k in the tth round is: (4); Where, ,in, is the transmission rate of device k, is the bandwidth, is the background noise, is the channel gain of working node k, is the communication time of the device, The size of the data to be transferred; The total energy consumption of working node k is: (5); In a mobile decentralized federated learning scenario, due to communication resource constraints, each worker node may not be able to request models from all neighboring worker nodes. Therefore, each worker node needs to select an appropriate subset of neighbors. More neighbors will help converge faster, but will consume more bandwidth. Assume that the bandwidth resource of each link in round t is In practice, since bandwidth is time-varying and measurable, the bandwidth constraint for each link is : (6); On the other hand, a worker node with less energy may not be able to complete the entire process of local training, model interaction, and model aggregation, resulting in information loss. To mitigate the impact, the total energy consumption of worker node k in round t should not exceed the energy budget of that round. ,Right now: (7); Therefore, the computing power satisfies the following constraints: (8); Where, is the link between working nodes k and j bandwidth, For the Wheel Chain bandwidth resources, is the energy budget of worker node k in round t.

[0027] In addition, in order to effectively balance the requirements of low overhead and high performance, and to achieve rapid model convergence by disseminating neighboring nodes and its own model information as much as possible when the network is not stable, our goal can be expressed as optimizing the global loss function in formula (1) under the condition of limited communication bandwidth and energy of working nodes: (9).

[0028] S2. Local model training and interactive content preparation: Monitor the characteristics of each worker node in real time, perform local model training on the local data of each worker node that meets the requirements, and synchronously generate two types of interactive resources after training. Among them, the two types of interactive resources are local model weight update information and lightweight local data category prototype information.

[0029] The model training process is as follows: Each worker node participating in the training uses the local training dataset to calculate its stochastic gradient ; Then, use worker node k to update its local model: (10); Where, is the learning rate, represents the local model of worker node k at the beginning of the tth round of global training (usually the latest global model), represents the candidate local model to be aggregated for worker node k in round t.

[0030] S3. Information interaction and content transmission: An adaptive transmission content selection algorithm and information caching mechanism are used to complete information interaction between each working node and its neighboring working nodes within the instantaneous communication range.

[0031] It should be noted that after each working node completes the local model training, the local model It sends the data to its connected neighbors through the set communication network and receives the neighbors' local models.

[0032] The communication network in DFL can be modeled as a connected undirected graph , recorded as .in, Represents the vertex, which is the collection of working nodes. Represents the set of links connecting working nodes. For example, in round t, Represents the set of links connecting working nodes, using the adjacency matrix Represents the network topology of P2P communication in DFL, where if , then, otherwise it is 0. The neighbor set of working node k is expressed as The degree matrix is ​​defined as a diagonal matrix where .

[0033] In addition, during information interaction and content transmission, in the interaction phase, priority is given to sharing local data category prototype information, and when network bandwidth permits, sharing of local model weight update information is increased.

[0034] It's important to note that in mobile scenarios, worker nodes need to quickly select neighbors for information exchange based on current bandwidth and energy consumption. Dynamic bandwidth fluctuations mean that some neighbors may not be able to meet transmission requirements within a specific time period, while devices with insufficient energy cannot participate in model training and updates. To effectively meet these constraints, a flexible and efficient information exchange method is required to ensure that each worker node can exchange the maximum amount of information within limited communication resources and energy budgets.

[0035] The adaptive interactive content selection algorithm mainly selects interactive content adaptively based on the current network bandwidth and the energy of the working node.

[0036] Among them, exchanging local model weight update information is a common practice in federated learning. Each working node uses a weighted average aggregation strategy to complete the local model update: (11); Where, is the model weight of worker node k in round t, represents the set of neighbor nodes of working node k in round t, Indicates the number of neighbor nodes of working node k in round t.

[0037] Exchanging local data category prototype information can correct for the lack of concentrated distribution of category representations under high mobility. Furthermore, because prototype information is much smaller than weight information, communication overhead can be reduced. The calculation of local data category prototype information also includes two stages: local category prototype preparation and global category prototype aggregation. It is worth noting that, similar to the aggregation of global model weights, in decentralized federated learning, the aggregation of global category prototypes is also the aggregation of the respective global category prototypes by each worker node. The training goal is still to achieve consistency in the global category prototypes of each worker node.

[0038] As an example, assume that worker nodes i and k have their local datasets and Completed training on and can have different label statistical distributions. Our goal is to make the global category prototypes of each working node consistent and to identify all categories of data. For the complete category set of the dataset, in the initial stage, each working node may only have global categories There can be overlap between the working nodes.

[0039] In the local category prototype preparation phase, each worker node completes the calculation of the category prototype based on the local data set. The worker node aggregates a prototype for each category of the local data, such as Figure 3 As shown, worker node 1 will get 4 category prototypes, worker node 3 will get 5 category prototypes, and worker node 5 will get 3 category prototypes. Assume that the geometric center of the z-th category data is defined as the prototype , then in worker node k, the prototype of category z is the local dataset of worker node k The mean of the embedding vector of the z-th class sample in can be expressed as: (12); Where, Is a local dataset A subset of , consisting of training samples belonging to the zth class, is the forward calculation of the model, It is a calculation method from the input to the representation layer of the intermediate representation. The representation layer refers to the layers other than the last layer.

[0040] We then introduce a method based on category prototypes to calibrate the problem of insufficient distribution of category representations under high mobility. At this point, the goals of decentralized federated learning can be improved as follows: (13); in, , For local model weights and a single data sample losses, is the number of neighbor nodes, is the number of samples belonging to category z in all working nodes, Is a measure of local prototype and the corresponding global prototype The regularization term for the distance between .

[0041] A certain type of prototype globally for each working node By aggregating the prototypes shared by its neighbor nodes, we get: (14); Where, represents the prototype of category z from worker node k, represents the number of worker nodes with class z, Represents a collection of working nodes The total number of samples of category z in .

[0042] It is worth noting that in decentralized federated learning, since the local global model of each node will eventually converge, the local global model obtained by aggregating neighbor prototypes of each node is the global prototype of the system. The approximate value of is: (15); Therefore, the label of sample x is Predict the representation vector of the available measurement samples and prototype The L2 distance between them is expressed as: (16); As an example, now Figure 2 From the perspective of worker node 1, the process of exchanging category prototype information is explained, such as Figure 3 shown.

[0043] After each worker node completes local training and prepares its own category prototype, it exchanges information. Worker node 1 will receive information shared by its neighbors, worker nodes 2, 3, and 5. In the aggregation phase, in addition to completing weight aggregation according to formula (11), prototype aggregation is also completed by category according to formula (14), ultimately forming a global set of all-category prototypes, which is broadcast to its neighbors in the next round of information sharing.

[0044] Considering the high-speed, decentralized nature of interaction, communication between worker nodes can only occur within each other's transmission range. If worker nodes encounter each other infrequently, the information available for aggregation will not be sufficient to support rapid model convergence. To expand the scope of model information sharing, we introduce an information caching mechanism to store data received by worker nodes from previously encountered nodes. When a worker node encounters a new node, it not only transmits its own data but also the cached data of other nodes.

[0045] The information caching mechanism includes: Allow each worker node to store and transmit local model updates received from the worker nodes it encounters in its historical interactions; The least recently used principle (LRU) is used to update information cache, and timestamps are used for filtering.

[0046] As an example, assume that the cache data of worker node k in round t is The working nodes not only exchange their own local information, but also share each other's cache information. Then, in the aggregation phase, the improved formula (11) is as follows: (17); Similarly, the prototype aggregation introduces Buffer and further improves formula (14) as follows: (18); in, is the weight of local data, and Respectively represent the weight information and prototype information in the Buffer.

[0047] Furthermore, to ensure the final convergence of the model, the joint graph of the peer-to-peer network across multiple rounds is assumed to be strongly connected, and peer-to-peer communication is assumed to be non-blocking. Furthermore, to mitigate the impact of stale information on the system, we propose a least recently used (LRU) cache update method that uses timestamps to assist in filtering and maximize cache benefits.

[0048] In this way, each working node can not only directly obtain the information of its neighbors, but also indirectly obtain the information of its neighbors’ neighbors, which helps promote the dissemination of data information and improve the convergence speed of DFL. It is worth noting that the peer-to-peer network topology is dynamic and sparse. Considering the impact of outdated information, a threshold is set. Control the impact of stale information on the system and remove stale information greater than the threshold in the cache of each working node. When all cached information is within an acceptable staleness range, fresher information is selected to replace the outdated information within the maximum capacity of the cache pool. After information exchange is completed, each worker node completes the aggregation of local model information in parallel.

[0049] S4, model aggregation update: Each working node shares the information generated by local training and the content received from neighboring working nodes to complete the update and iteration of the local model; Among them, the update iteration formula of the local model is: , where represents the subgraph containing worker node k and its neighbors, Indicates the aggregation rule used, is the local model weight of worker node k after aggregation in round t, is the set of neighboring working nodes of working node k, represents the candidate local model to be aggregated by worker node i in round t.

[0050] S5. Update termination: Repeat steps S2 to S4 until the preset convergence condition is met or the maximum communication rounds are reached.

[0051] In this embodiment, the maximum communication rounds of the experimental group of the training data set MNIST are set to 100 rounds, and the maximum communication rounds of the experimental group of the training data sets CIFAR10 and CIFAR100 are set to 200 rounds.

[0052] Performance testing: 1) The datasets used in the experiment include the MNIST handwritten digit dataset, and the CIFAR10 and CIFAR100 datasets created by the Canadian Institute for Advanced Research. These datasets are classified into two types of data distribution: independent and identically distributed (IID) and non-independent and identically distributed (Non-IID). In the IID scenario, worker nodes were randomly assigned datasets of equal size. In the Non-IID scenario, the p value was used to adjust the proportion of the client's actual data type to simulate the Non-IID distribution of the data. In the experiment, p was set to 0.2, 0.5, and 0.8. 2) Model: The experiment used the visual geometry group model VGG9 and the residual neural network ResNet9 as training models. The VGG9 model was trained on the MNIST and CIFAR10 datasets, and the ResNet9 model was trained on the CIFAR10 and CIFAR100 datasets. 3) Key parameter settings: In the experiment, the VGG9 model group was set to 10 worker nodes, and the ResNet9 model group was set to 20 worker nodes. The maximum training rounds for the MNIST dataset group were 100, and the maximum training rounds for the CIFAR10 and CIFAR100 groups were set to 200; the early stopping setting was 20 for both. In addition, the learning rate was set to 0.01 for the MNIST and CIFAR10 datasets, and the learning rate was set to 0.005 for the CIFAR100 dataset. The SGD optimizer was used for both, with a momentum of 0.9, a number of local updates of 2, and a batch size of 32. 4) Communication Configuration: We assume that nodes communicate with their neighbors via a local area network (LAN) or wide area network (WAN). We simulate the heterogeneity and dynamics of peer-to-peer networks by randomly assigning bandwidth to each node. We also simulate different latency for each node.

[0053] Four sets of experiments were conducted to evaluate the overall performance of our method: VGG9-MNIST, VGG9-CIFAR10, ResNet9-CIFAR10, and ResNet9-CIFAR100. The best accuracy of our method and the baseline methods in these four sets of experiments are shown in Table 1.

[0054] Table 1 Best model accuracy

[0055] From Table 1, we can see that our method can achieve comparable accuracy to the baseline in different models and datasets. In particular, it outperforms all baselines in the NonIIDp=0.8 scenario: in the VGG9-MNIST, VGG9-CIFAR10, ResNet9-CIFAR10, and ResNet9-CIFAR10 scenarios, the model accuracy of our method is improved by an average of 4.71%, 9.41%, 10.88%, and 5.93% respectively compared with the baseline. In addition, Figure 4 The model test accuracy of each set of experiments using our method in the NonIIDp=0.8 scenario is shown. The results show that our method achieves more stable accuracy and minimizes oscillations during training. This method achieves higher accuracy than the baseline in highly heterogeneous environments due to the introduction of an information caching mechanism and the design of a least recently used cache update mechanism. This method leverages the spatiotemporal information transmission network formed by high-speed vehicle movement to achieve multi-hop information propagation within the network, effectively expanding the coverage of information interaction and improving model performance in mobile scenarios.

[0056] In addition, we counted the training time and communication volume used by each group of methods to achieve the same accuracy. Figure 5 and Figure 6 For example, Figure 5 (a) shows the time required for our method and comparison methods to achieve the same accuracy under varying dataset heterogeneity in the VGG9-MNIST scenario. The horizontal axis represents the dataset distribution and the accuracy level achieved by the shuffled methods, while the vertical axis represents training time. Experimental results show that our method, DFLTrans, outperforms all baselines in both training time and communication overhead, achieving consistent accuracy across different models and datasets.

[0057] Specifically, if Figure 5 As shown in the data, in terms of training time, our method reduces training time by an average of 47.96%, 43.21%, 51.13%, and 76.50% compared to the suboptimal method in the IID, NonIID (p=0.2), NonIID (p=0.5), and NonIID (p=0.8) scenarios, respectively. The advantage of our method becomes more pronounced as dataset heterogeneity increases. This is primarily due to the design of the caching mechanism, which trades storage space for training efficiency.

[0058] In terms of communication traffic overhead, such as Figure 6 As shown in the results, in the IID, NonIID (p=0.2), NonIID (p=0.5), and NonIID (p=0.8) scenarios, our method reduces communication overhead by an average of 43.55%, 33.47%, 38.48%, and 69.51%, respectively, compared to the suboptimal method. Similarly, as dataset heterogeneity increases, the advantage of our method in terms of communication overhead becomes more pronounced. This is due to the design of the adaptive content selection module, which selects smaller interactive content for transmission when communication conditions are limited, reducing communication overhead and improving training efficiency. Furthermore, the design of the caching mechanism ensures the richness of training data, thereby ensuring model accuracy.

[0059] In summary, the method provided by this application focuses on limited and time-varying communication networks, as well as edge devices with varying computing power and energy consumption, and proposes a decentralized federated learning method for heterogeneous devices with adaptive transmission content selection. The method provided by this application effectively reduces communication overhead and ensures model performance through an adaptive interactive content selection algorithm; at the same time, with the help of a caching mechanism to expand the amount of transmitted information, the model performance and training efficiency in mobile scenarios are improved. The method provided by this application is quite effective in reducing training time and communication overhead.

[0060] Example 2 See also Figure 7As shown, this embodiment discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements any one of the methods provided by the above methods to construct an adaptive transmission content selection system for the Internet of Vehicles.

[0061] Since the electronic device described in this embodiment is an electronic device used to implement the method for constructing an adaptive transmission content selection system for the Internet of Vehicles in the embodiment of this application, based on the method for constructing an adaptive transmission content selection system for the Internet of Vehicles described in the embodiment of this application, those skilled in the art will be able to understand the specific implementation methods and various variations of the electronic device of this embodiment, so how the electronic device implements the method in the embodiment of this application will not be described in detail here. As long as those skilled in the art implement the electronic device used in the method for constructing an adaptive transmission content selection system for the Internet of Vehicles in the embodiment of this application, it falls within the scope of protection to be provided by this application.

[0062] Example 3 See also Figure 8 As shown, this embodiment discloses a computer-readable storage medium, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements any one of the methods provided by the above methods to construct an adaptive transmission content selection system for the Internet of Vehicles.

[0063] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters, weights and thresholds in the formulas are set by technicians in this field according to actual conditions.

[0064] The above embodiments can be implemented in whole or in part via software, hardware, firmware, or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product comprises one or more computer instructions or computer programs. When loaded or executed on a computer, the processes or functions described in accordance with the embodiments of the present invention are fully or partially performed. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another. For example, the computer instructions can be transferred from one website, computer, server, or data center to another website, computer, server, or data center via a wired or wireless network. The computer-readable storage medium can be any available medium accessible by a computer, or a data storage device such as a server or data center that contains a collection of one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, or magnetic tapes), optical media (e.g., DVDs), or semiconductor media. The semiconductor media can be a solid-state drive.

[0065] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0066] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0067] In the several embodiments provided by the present invention, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only one type. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0068] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0069] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0070] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

[0071] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for constructing an adaptive transmission content selection system for the Internet of Vehicles, characterized in that: The method comprises: S1. Initialization setup: Determine the working nodes and their characteristics in the decentralized federated learning system DFL, initialize the global local model, and set the training parameters; Among them, the characteristics include the network bandwidth, energy consumption, and computing power of the worker nodes; S2. Local model training and interactive content preparation: Monitor the characteristics of each worker node in real time, perform local model training on the local data of each worker node that meets the requirements, and synchronously generate two types of interactive resources after training. Among them, the two types of interactive resources are local model weight update information and lightweight local data category prototype information; S3, Information Interaction and Content Transmission: An adaptive transmission content selection algorithm and information caching mechanism are used to complete information exchange between each working node and its neighboring working nodes within the instantaneous communication range; S4, model aggregation update: Each working node shares the information generated by local training and the content received from neighboring working nodes to complete the update and iteration of the local model; S5. Update termination: Repeat steps S2 to S4 until the preset convergence condition is met or the maximum communication rounds are reached.

2. The method according to claim 1, characterized in that In step S1, the network bandwidth is calculated as follows: The inbound bandwidth and outbound bandwidth of worker node k in round t are expressed as and ; Assuming that the bandwidth is evenly distributed, the link The network bandwidth is , where 、 They represent the number of neighbor nodes of worker node k and worker node j in round t respectively.

3. The method according to claim 1, characterized in that In step S1, the energy consumption is calculated as follows: During the training phase, the training energy consumption of the working node k in the tth round is , where represents the number of processor cycles required for worker node k to execute a data sample, is the local dataset of device k, Representation dataset The number of samples, represents the processor cycle frequency of worker node k, is the number of local iterations, Indicates the effective capacitance coefficient of the device k calculation chip; In the model interaction stage, the training energy consumption of the working node k in the tth round is , where ,in, is the transmission rate of device k, is the bandwidth, is the background noise, is the channel gain of working node k, is the communication time of the device, The size of the data to be transferred; The total energy consumption of working node k is .

4. The method for constructing an adaptive transmission content selection system for the Internet of Vehicles according to claim 3, characterized in that: In step S1, the computing capacity satisfies the following constraints: , where is the link between working nodes k and j bandwidth, For the Wheel Chain bandwidth resources, is the energy budget of worker node k in round t.

5. The method for constructing an adaptive transmission content selection system for the Internet of Vehicles according to claim 1, characterized in that: In step S2, the model training process is as follows: Each worker node participating in the training uses the local training dataset to calculate its stochastic gradient; The worker nodes are then leveraged to update their local models.

6. The method for constructing an adaptive transmission content selection system for the Internet of Vehicles according to claim 1, characterized in that: In step S3, during information interaction and content transmission, in the interaction phase, priority is given to sharing local data category prototype information, and if network bandwidth permits, sharing of local model weight update information is added.

7. The method for constructing an adaptive transmission content selection system for the Internet of Vehicles according to claim 1, characterized in that: In step S3, the information caching mechanism includes: Allow each worker node to store and transmit local model updates received from the worker nodes it encounters in its historical interactions; The least recently used principle (LRU) is used to update information cache, and timestamps are used for filtering.

8. The method for constructing an adaptive transmission content selection system for the Internet of Vehicles according to claim 1, characterized in that: In step S4, the update iteration formula of the local model is: , where represents the subgraph containing worker node k and its neighbors, Indicates the aggregation rule used, is the local model weight of worker node k after aggregation in round t, is the set of neighboring working nodes of working node k, represents the candidate local model to be aggregated by worker node i in round t.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, it implements the method for constructing the adaptive transmission content selection system for the Internet of Vehicles as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed, implements the method for constructing the adaptive transmission content selection system for the Internet of Vehicles as described in any one of claims 1 to 8.

Images