Vehicle clustering method and collaborative caching method for vehicle-mounted named data network

Abstract

The invention provides a vehicle clustering method and a collaborative caching method for a vehicle-mounted named data network, and belongs to the field of named data networks. The vehicle clustering method comprises the steps that vehicle features of vehicle nodes in a communication range are acquired, and relative features between the vehicle nodes are calculated after filling; obtaining connectivity between the vehicle nodes by using a pre-constructed connectivity prediction model; obtaining a connectivity confidence weighted graph of the vehicle nodes according to the connectivity between the vehicle nodes; and then cluster head vehicles and corresponding vehicle clusters are obtained. According to the vehicle clustering method provided by the invention, the dynamic characteristics of the vehicles are considered, and the connectivity between the vehicle nodes is predicted and updated in real time, so that dynamic clustering and real-time adjustment are realized, and the clustering accuracy and the communication efficiency are improved.

Description

Technical Field

[0001] The present invention relates to a vehicle clustering method and a collaborative caching method for a vehicle named data network, and belongs to the field of named data networks. Background Art

[0002] In modern transportation systems, the vehicle named data network (VNDN), as an emerging technology, is gradually becoming a key component of intelligent transportation systems and smart cities. VNDN combines the mobility of vehicle networks and the content-centric characteristics of named data networks, aiming to provide efficient and reliable information exchange services. With the rapid development of intelligent applications such as autonomous driving, in-vehicle information services, driving safety applications, and in-vehicle augmented reality, the demand for low latency and high bandwidth in the network is increasing day by day. These applications not only require fast network responses but also a large amount of computing and storage resources to perform complex tasks.

[0003] Traditional cloud computing models face challenges in handling these requirements. Due to the long distance between the cloud server and the vehicle and limited network bandwidth, high transmission latency may occur when the vehicle accesses cloud data. In addition, the high mobility of the vehicle causes the network topology to change continuously, further increasing the instability of communication. To solve these problems, the concept of vehicle edge computing (VEC) emerged, which deploys computing and storage resources at the vehicle edge to reduce network bandwidth pressure and task response time.

[0004] A key component of VEC is vehicle edge caching, which realizes low-latency data caching by deploying servers near the vehicle. This caching strategy can be deployed in roadside infrastructure, such as macro base stations, cellular towers, or roadside units, or directly in the vehicle. By performing vehicle-to-vehicle or vehicle-to-infrastructure communication between vehicles or between vehicles and roadside units, popular content can be stored and transmitted, effectively increasing the distribution range of the content.

[0005] In the vehicle networking environment, although VEC technology can significantly improve the quality of service, there are obvious deficiencies in existing research on vehicle clustering methods, which directly affect the performance of clustering-based caching strategies. Existing clustering methods mostly adopt static strategies, fail to adapt to the dynamic movement of vehicles and changes in the network environment, and ignore the social attributes and content popularity between vehicles, resulting in inaccurate clustering results and low communication efficiency. In addition, these methods usually lack the ability to dynamically adjust, cannot respond to the joining, leaving, or movement of vehicles according to real-time data, and adopt a single fixed replacement strategy in the caching strategy, failing to dynamically adjust according to content popularity and vehicle mobility, resulting in low caching efficiency. Summary of the Invention

[0006] The object of the present invention is to overcome the deficiencies in the prior art, and provide a vehicle clustering method and a cooperative caching method for a vehicular named data network, which solve the problems that the vehicle clustering fails to adapt to the dynamic movement of vehicles and the changes in the network environment, and the content caching ignores the social attributes and content popularity among vehicles.

[0007] To achieve the above object, the present invention is implemented by the following technical solutions:

[0008] In the first aspect, the present invention provides a vehicle clustering method for a vehicular named data network, with a roadside unit as the execution entity, including:

[0009] S1: Obtain the vehicle characteristics of vehicle nodes within the communication range, fill them, and calculate the relative characteristics between vehicle nodes;

[0010] S2: Based on the vehicle characteristics of vehicle nodes and the relative characteristics between vehicle nodes, use a pre-constructed connectivity prediction model to obtain the connectivity between vehicle nodes;

[0011] S3: According to the connectivity between each vehicle node and other vehicle nodes, obtain a connectivity confidence weighted graph of vehicle nodes;

[0012] S4: Based on the connectivity confidence weighted graph of vehicle nodes, obtain cluster head vehicles and corresponding vehicle clusters.

[0013] Further, the S1 includes: using a null value filling method based on spatio-temporal similarity to fill the null values in vehicle characteristics; wherein, the null value filling method based on spatio-temporal similarity includes:

[0014] Calculate the spatial similarity between two vehicle nodes: ; Wherein, is the spatial similarity between two vehicle nodes, is the dimension of vehicle characteristics, is the th feature in is the th feature in and are the vehicle characteristics of vehicle node and vehicle node at time;

[0015] Calculate the temporal similarity of vehicle nodes: ; Wherein, is the vehicle node The time similarity, is the time factor, is the time interval;

[0016] Based on the spatial similarity and time similarity, fill in the zero values in the vehicle features: ;

[0017] Calculate the relative features between vehicle nodes, and the calculation formula is: ; Among them, are the speed, position and acceleration of vehicle node respectively, are the speed, position and acceleration of vehicle node respectively.

[0018] Furthermore, the connectivity prediction model adopts a recurrent neural network model; the connectivity prediction model adopts a binary cross-entropy loss function; the activation function of the output layer of the connectivity prediction model is the Softmax function; the output label of the connectivity prediction model is the connectivity between vehicle nodes ; ; Among them, is the communication range of the vehicle node.

[0019] Furthermore, a denoising autoencoder is used to perform feature dimensionality reduction on the input of the connectivity prediction model; noise perturbations are added to the input layer of the denoising autoencoder; including:

[0020] Normalize and combine the vehicle characteristics , and the relative features to generate the first feature vector of two vehicle nodes;

[0021] Input the first feature vector into the denoising autoencoder to obtain the feature vectors of two vehicle nodes for input into the connectivity prediction model;

[0022] The training process of the denoising autoencoder includes: calculating the error between the feature vector and the first feature vector, and updating the weights of the denoising autoencoder through the backpropagation algorithm until the error is within the preset range.

[0023] Furthermore, the connectivity confidence weighted graph is , among which, is the set of vehicle nodes; is the edge set of; is the weight function for calculating The weights of each side in Among them, ; is the eigenvector distance, and the calculation formula is: ; In the formula, and are respectively the and th embedding vectors in the eigenvector, and is the dimension of the embedding vector;

[0024] The above S3 includes:

[0025] For any vehicle node , obtain the connectivity confidence interval of the vehicle node : ;

[0026] According to the connectivity confidence interval of the vehicle node and the connectivity between vehicle nodes, obtain edge set;

[0027] In response to , then add the edge between and to , and the corresponding weight is 1;

[0028] In response to , then add the edge between and to , and calculate the corresponding weight through the weight function .

[0029] Furthermore, the above S4 includes: For the connectivity confidence weighted graph of the vehicle node ,

[0030] In response to , then is the cluster head vehicle; among them, is neighbor node set of, is the distance threshold, is the neighbor threshold;

[0031] The vehicle cluster is composed of a cluster head vehicle and member vehicles, and the member vehicles are vehicle nodes with connectivity to the cluster head vehicle.

[0032] In a second aspect, the present invention provides a collaborative caching method based on vehicle clustering, with the cluster head vehicle as the execution entity, including:

[0033] Obtaining important nodes of the vehicle cluster according to the node betweenness of all vehicle nodes in the vehicle cluster;

[0034] Receiving the requested content and the corresponding content popularity sent by the member vehicles;

[0035] Caching the requested content into the important nodes of the vehicle cluster according to the content popularity.

[0036] Further, the node betweenness is stored in the router in the vehicular named data network, and the calculation formula of the node betweenness of the vehicle node is: ; ; where , is the set of all vehicle nodes in the vehicle cluster, is the number of the shortest paths from node to node through node , is the total number of all shortest paths from node to node ;

[0037] The important nodes of the vehicle cluster include: primary important nodes and secondary important nodes;

[0038] The obtaining of the important nodes of the vehicle cluster includes: sorting all vehicle nodes in the vehicle cluster in descending order according to the node betweenness; first marking the top vehicle nodes as primary important nodes, and then marking the top vehicle nodes among the remaining vehicle nodes as secondary important nodes; where and are preset values.

[0039] Further, the content popularity is obtained by a popularity prediction model deployed on the member vehicles;

[0040] The popularity prediction model is used to predict the future popularity value according to the historical feature data of the requested content; where the historical feature data includes: historical popularity features and geographical location features;

[0041] The popularity prediction model adopts an encoder-decoder structure with a bidirectional long short-term memory (BiLSTM) network and a multi-head attention mechanism, including: an encoder based on the BiLSTM network and a decoder centered on multiple attentions.

[0042] Further, the requested content is cached in the important nodes of the vehicle cluster according to the content popularity, including: dividing the content popularity into X levels; obtaining a probability matrix based on the number of important nodes in the vehicle cluster and the content popularity levels; sending the probability matrix to all important nodes in the vehicle cluster so that all important nodes perform caching according to the probability matrix.

[0043] Among them, the probability matrix is used to pair important nodes and content popularity levels; the number of its rows and columns is equal to the number of important nodes in the vehicle cluster and the number of content popularity levels respectively; the elements of the probability matrix are the probabilities that the th important node is used to cache the requested content with a content popularity of ; ; then it means that the th important node is not used to cache the requested content with a content popularity of ; the calculation formula is: ; Among them, is the number of nodes in the vehicle cluster, is the proportion of popular content among content popularities, the number of nodes in the vehicle cluster, ; then it means that the th important node is not used to cache the requested content with a content popularity of ; is the first weight factor, is the second weight factor, .

[0044] Compared with the prior art, the beneficial effects achieved by the present invention:

[0045] (1) The vehicle clustering method provided by the present invention considers the dynamic characteristics of vehicles, predicts and updates the connectivity between vehicle nodes in real time, realizes dynamic clustering and real-time adjustment, so as to improve the accuracy of clustering and communication efficiency.

[0046] (2) The collaborative caching strategy based on vehicle clustering provided by the present invention can optimize the selection of caching nodes according to content popularity and vehicle mobility prediction, and comprehensively considers the dynamic characteristics, social attributes and content popularity of vehicles. Improve the utilization rate of caching resources, reduce content access latency, and significantly improve the performance and service quality of vehicular ad hoc networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is the flowchart of the vehicle clustering method in Embodiment 1 of the present invention;

[0048] Figure 2 It is the flowchart of the collaborative caching method based on vehicle clustering in Embodiment 2 of the present invention;

[0049] Figure 3 It is the comparison chart of cache hit rates of different caching strategies in Embodiment 3 of the present invention;

[0050] Figure 4 It is the comparison chart of latencies of different caching strategies in Embodiment 3 of the present invention;

[0051] Figure 5 It is the comparison chart of the influence of cache capacity parameters on the link load performance of different caching strategies in Embodiment 3 of the present invention. Detailed implementation manners

[0052] The terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or are inherent to these processes, methods, products or devices. The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the drawings in the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all the embodiments.

[0053] Embodiment 1

[0054] According to the first aspect of the present invention, this embodiment provides a vehicle clustering method for an on-vehicle named data network, with a roadside unit as the execution entity, as Figure 1 shown, including:

[0055] Step 1: Obtain the vehicle characteristics of vehicle nodes within the communication range, perform filling, and calculate the relative characteristics between vehicle nodes.

[0056] Specifically, the vehicle characteristics of vehicle node at time are , where are respectively the speed, position, and acceleration of vehicle node .

[0057] Due to environmental interference, there may be null values in the vehicle characteristics, and null value filling is required.

[0058] In some specific embodiments, a null value filling method based on spatio-temporal similarity is used to fill the null values in vehicle characteristics. Spatial similarity measures the similarity of vehicle characteristics between vehicle nodes with similar traffic environments, and temporal similarity measures the change of vehicle characteristics over time. The shorter the time interval, the higher the probability of similarity of vehicle characteristics.

[0059] Specifically, the null value filling method based on spatio-temporal similarity includes:

[0060] Step 11: According to the Pearson correlation coefficient, the spatial similarity between vehicle node and vehicle node at time is: ; where is the spatial similarity of the two vehicle nodes, is the dimension of vehicle characteristics, is the th feature in is the th feature in and are the vehicle characteristics of vehicle node and vehicle node at time respectively.

[0061] It should be noted that during the calculation process, the null values in vehicle characteristics are taken as 0 for calculation.

[0062] Step 12: Calculate the temporal similarity of vehicle nodes: ; where is the temporal similarity of vehicle node , is the time factor, is the time interval.

[0063] This formula is used to measure the influence of time dynamics on vehicle characteristics. As the time interval increases, the similarity decreases exponentially.

[0064] Step 13: Based on spatial similarity and temporal similarity, calculate the zero filling value of vehicle characteristics: .

[0065] After completing the filling of null values in vehicle characteristics, calculate the relative characteristics between pairwise vehicle nodes. The relative characteristics between node and are calculated by the following formula: ; Among them, are the speed, position, and acceleration of the vehicle node respectively.

[0066] Step 2: Based on the vehicle characteristics of the vehicle nodes and the relative characteristics between the vehicle nodes, use the pre-constructed connectivity prediction model to obtain the connectivity between the vehicle nodes.

[0067] Specifically, vehicle motion is time-related, and the recurrent neural network model RNN can exhibit time-dynamic behavior. Therefore, the connectivity prediction model adopts the recurrent neural network model.

[0068] The output label of the connectivity prediction model is the connectivity between the vehicle nodes , and the expression is: ; Among them, is the communication range of the vehicle node.

[0069] Since connectivity prediction is a binary classification problem, the connectivity prediction model uses the Softmax function as the activation function of the output layer, which normalizes the RNN output into a probability distribution consisting of two probabilities.

[0070] Cross-entropy is an error metric that is useful in problems with targets of 0 or 1 when the output of the model can be regarded as representing independent hypotheses. Therefore, the connectivity prediction model adopts the binary cross-entropy loss function.

[0071] In addition, during the training process of the connectivity prediction model, the backpropagation algorithm is used to calculate the gradient of the loss function for each weight and bias through the chain rule, and the weights and biases of one layer are updated at a time.

[0072] In some specific embodiments, in order to reduce the complexity of the connection prediction method, a denoising autoencoder is used to perform feature dimensionality reduction on the input of the connectivity prediction model, and noise perturbations are added to the input layer of the denoising autoencoder.

[0073] The specific process includes: in order to unify the dimensions of the vehicle characteristics, before inputting the data into the denoising autoencoder, the vehicle characteristics , and the relative characteristics are normalized and merged to generate the first feature vector of the two vehicle nodes.

[0074] Normalization is to scale the data into a standard range, such as 0 to 1, which can ensure that different features are compared on the same scale. Merging feature vectors is to combine multiple features into one vector as the input of the denoising autoencoder, and the denoising autoencoder outputs the feature vector.

[0075] The denoising autoencoder is used to learn and extract useful information from data containing noise (interference). Its training process includes: calculating the error between the feature vector and the first feature vector, and updating the weights of the denoising autoencoder through the backpropagation algorithm until the error is within the preset range.

[0076] Step 3: According to the connectivity between each vehicle node and other vehicle nodes, obtain the connectivity confidence weighted graph of the vehicle nodes.

[0077] It should be noted that the connectivity confidence weighted graph reflects the connectivity between vehicle nodes. The connectivity confidence weighted graph is , where is the set of vehicle nodes; is the edge set of is the weight function, which is used to calculate the weights of each edge in Among them, ; is the feature vector distance, and the calculation formula is: ; In the formula, and are respectively the and th embedding vectors in the feature vector, and is the dimension of the embedding vector.

[0078] Specifically, Step 3 includes:

[0079] Step 31: Obtain the connectivity confidence interval of the vehicle node : .

[0080] In the vehicle edge network, the connection confidence interval represents the stability of the communication link between vehicle nodes. When the distance between two vehicles is within this confidence interval, it means the possibility that the communication link between these two vehicles remains connected. If the distance exceeds this range, the connection may decrease and the communication link may be disconnected.

[0081] Step 32: According to the connectivity confidence interval of the vehicle node and the connectivity between vehicle nodes, obtain the edge set of

[0082] Specifically, based on the eigenvector distance decide whether to add the edge and between to If , then add and the edge between to and the corresponding weight is 1; If , that is, the eigenvector distance is within the connectivity confidence interval, then add and the edge between to At this time, calculate the corresponding weight through the weight function .

[0083] Step 4: Based on the connectivity confidence weighted graph of vehicle nodes, obtain the cluster head vehicle and the corresponding vehicle clusters.

[0084] Specifically, for the connectivity confidence weighted graph of the vehicle node , if , then is the cluster head vehicle; where is the neighbor node set of is the distance threshold is the neighbor threshold.

[0085] At this time, the remaining vehicles with connectivity to the cluster head are member vehicles, and the cluster head vehicle and the member vehicles together form a vehicle cluster.

[0086] The vehicle clustering method in this embodiment predicts vehicle connectivity through a recurrent neural network (RNN), and based on this prediction result, selects the vehicle with the highest connectivity as the cluster head vehicle and forms a vehicle cluster. This method considers the dynamic characteristics of vehicles, predicts and updates the connectivity between vehicle nodes in real time, and realizes dynamic clustering and real-time adjustment to improve the accuracy of clustering and communication efficiency.

[0087] Embodiment 2

[0088] This embodiment provides a collaborative caching method based on the vehicle clustering described in Embodiment 1, with the cluster head vehicle as the execution subject, as shown in Figure 2 , including:

[0089] Step A: Obtain the important nodes of the vehicle cluster according to the node betweenness of all vehicle nodes in the vehicle cluster.

[0090] It should be noted that the routers in the VNDN network do not have an information table for recording node betweenness. Therefore, in the present invention, an information record table for nodes is added in the routers of the NDN network to record node betweenness information. The information record table is shown in Table 1 as follows:

[0091] Table Information Record Table

[0092] Indexes Betweenness centrality Value 1.5

[0093] In Table 1, the first column Indexes is the index column, and betweenness centrality under this column is the node betweenness. The second column Value in the table is the value corresponding to the index.

[0094] As is well known, for vehicle nodes the calculation formula for node betweenness is: ; wherein , is the set composed of all vehicle nodes in the vehicle cluster, is the number of the shortest paths from node to node through node , is the total number of all the shortest paths from node to node .

[0095] All member vehicles and cluster head vehicles are rated according to node betweenness, including: first-level important nodes and second-level important nodes.

[0096] In some specific embodiments, obtaining the important nodes of the vehicle cluster includes: sorting all vehicle nodes in the vehicle cluster in descending order according to node betweenness; first marking the first vehicle nodes as first-level important nodes, and then marking the first vehicle nodes among the remaining vehicle nodes as second-level important nodes; marking the remaining nodes as unimportant nodes, wherein and are preset values.

[0097] The cluster head vehicle maintains a list of important nodes of this vehicle cluster and the betweenness rating of each node in the list. If is the number of nodes in this vehicle cluster, the list contains first-level important nodes and second-level important nodes.

[0098] Step B: Receive the request content and the corresponding content popularity sent by the member vehicle.

[0099] It should be noted that the content popularity is obtained by the popularity prediction model deployed on the member vehicles. The popularity prediction model is used to predict the future popularity value according to the historical feature data of the requested content. The historical feature data includes historical popularity features, social relationships, and geographical location features.

[0100] In some specific embodiments, the popularity prediction model adopts an encoder-decoder structure with a bidirectional long short-term memory (BiLSTM) network and a multi-head attention mechanism, including an encoder based on the BiLSTM network and a decoder centered on multiple attentions.

[0101] Step C: Cache the requested content into the important nodes of the vehicle cluster according to the content popularity.

[0102] Specifically, Step C includes: first dividing the content popularity into levels; then, obtaining a probability matrix according to the number of important nodes in the vehicle cluster and the content popularity levels; and finally sending the probability matrix to all the important nodes in the vehicle cluster so that all the important nodes perform caching according to the probability matrix.

[0103] It should be noted that the probability matrix is used to pair the important nodes and the content popularity levels; the number of its rows and columns is equal to the number of important nodes and the number of content popularity levels in the vehicle cluster respectively; the elements of the probability matrix represent the probability that the th important node is used to cache the requested content with a content popularity of , ; represents that the th important node is not used to cache the requested content with a content popularity of .

[0104] In some specific embodiments, the probability matrix is a matrix with rows and columns. The calculation formula for the element in the probability matrix is: ; In the formula, and are weight factors, where . Each element of the probability matrix is calculated through the importance level and the popularity level therein. The reciprocal function is used here to assign higher probabilities to more important nodes and more popular content categories.

[0105] The principle of designing the probability matrix is that the first-level important nodes are used to cache the most popular content, the second-level important nodes are used to cache the second-most popular content, and the less popular content is cached by the unimportant routers. Therefore, for non-zero , it must satisfy > and > .

[0106] The cluster head vehicle collects information on the importance level of nodes and the content popularity. After the collection is completed, a probability matrix is calculated, and the result and the corresponding content popularity level are sent to all important nodes in this vehicle cluster. After receiving the result from the cluster head vehicle, all important nodes perform caching using the probability matrix.

[0107] After receiving the probability matrix and the list of popularity levels sent by the cluster head vehicle, each important node first obtains the caching probability for each content popularity level and caches the content according to the specified caching probability in the node.

[0108] Embodiment 3

[0109] To evaluate the effectiveness of the caching method described in Embodiment 2, this embodiment provides a simulation experiment using the collaborative caching method described in Embodiment 2 for caching.

[0110] This simulation experiment simulates an urban information network environment containing roadside units RSU, which is composed of intelligent vehicles. The MovieLens dataset is used for the experiment. The MovieLens dataset contains approximately 1 million rating records of 4000 movies from 6000 users. The MovieLens dataset is used to simulate the content request and access patterns in the automotive edge caching environment, where movie ratings are considered as content popularity metrics, and the user geographical distribution and rating habits are used to simulate the distribution and movement characteristics of vehicle nodes.

[0111] In addition, the cache sizes of each caching vehicle are set to 300 MB and 2 GB respectively. The following channel model is used: path loss (dB) 36.8 + 36.7log(d), where d is the distance in meters, the logarithmic positive mask parameter is 7 dB, the antenna gain is 5 dBi, the small-scale fading follows a Rayleigh distribution with unit variance, and the channel bandwidth is 20 MHz.

[0112] Finally, the cache hit rate, user latency, and link load, these 3 evaluation criteria of different methods are respectively counted. The caching strategies for comparison are LCE, LCD, ProCache, CL4M, ECSMADRL, and PBC.

[0113] (1) Cache hit rate

[0114] Figure 3 Shows the comparison graph of the cache hit rate change trends between the present invention and other caching strategies. The preposedscheme is the experimental data of the present invention.

[0115] It can be seen from Figure 3 that the present invention is always superior to the other six caching strategies. In all simulation experiments where the Zipf parameter and the cache capacity parameter vary within the range of 0.05 to 0.25, the present invention is superior to the six caching strategies. When the cache capacity parameter varies within the range of 0.05 to 0.25, the average cache hit rate of the present invention is 47%, which is 2.1% higher than that of ECSMADRL (44.9%), 3.8% higher than that of PBC (43.2%), 5.6% higher than that of LCD (41.4%), 8.6% higher than that of CL4M (38.4%), 15.6% higher than that of LCE (31.4%), and 12.5% higher than that of ProbCache (34.5%). In summary, the present invention can improve the cache hit rate in various situations. The present invention takes into account the importance of nodes and the placement location of popular content, thereby improving the utilization rate of the cache.

[0116] (2) User delay

[0117] Delay is an important indicator to measure network performance. Figure 4 shows the delays of the seven caching strategies, where the Zipf parameter , and the cache capacity parameter is from 0.05 to 0.25. Generally, when the cache capacity parameter increases, the delay will naturally decrease because more content objects can be cached in the network, so the time delay is lower.

[0118] It can be seen from Figure 4 that when the cache capacity parameter is 0.25, the delays of all caching strategies are the lowest. Among them, the average delay of the solution proposed by the present invention is 61.74 ms, which is 3.3% lower than that of ECSMADRL (63.9 ms), 5.2% lower than that of PBC (65.16 ms), 6.9% lower than that of LCD (66.32 ms), 8.1% lower than that of CL4M (67.21 ms), 9.6% lower than that of ProbCache (68.31 ms), and 10.3% lower than that of LCE (68.9 ms).

[0119] (3) Link load

[0120] Figure 5 shows the influence of different cache capacity parameters on the link load performance when the Zipf parameter .

[0121] It can be seen from Figure 5It can be seen that the average link load of the present invention is 245.8 bytes, which is 2.4% lower than that of LCD (251.9 bytes), 3.15% lower than that of CL4M (253.8 bytes), 3.7% lower than that of ProbCache (255.5 bytes), 4.4% lower than that of LCE (257.3 bytes), 1.1% lower than that of PBC (248.6 bytes), and 1.3% higher than that of ECSMADRL (242.6 bytes).

[0122] The vehicle optimizes caching and data transmission through hierarchical caching that clusters and predicts content popularity. This requires continuous updating of node and content information to ensure the accuracy of caching decisions. These information updates and caching decision processes increase the traffic in the network because they involve frequent data transmissions between vehicles and between vehicles and RSU. This frequent data transmission may lead to a slightly higher link load. Generally speaking, the link load performance of the present invention is superior to that of LCD, LCE, CL4M, ProbCache, and PBC.

[0123] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them; those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present application.

Claims

1. A vehicle clustering method for a vehicle-mounted named data network, with a roadside unit as the execution subject, characterized in that: include: S1: Obtain vehicle features of vehicle nodes within the communication range, fill them in, and calculate relative features between vehicle nodes; S2: Based on the vehicle characteristics of the vehicle nodes and the relative characteristics between the vehicle nodes, the connectivity between the vehicle nodes is obtained using a pre-built connectivity prediction model; S3: Obtain a connectivity confidence weighted graph of the vehicle nodes according to the connectivity between each vehicle node and other vehicle nodes; S4: Based on the connectivity confidence weighted graph of vehicle nodes, the cluster head vehicle and the corresponding vehicle cluster are obtained.

2. The vehicle clustering method of the vehicle named data network according to claim 1, characterized in that: The S1 includes: The empty value filling method based on spatiotemporal similarity is used to fill the empty values ​​in vehicle characteristics; Wherein, the method for filling in empty values ​​based on spatiotemporal similarity includes: Calculate the spatial similarity of two vehicle nodes: ； in, is the spatial similarity of the two vehicle nodes, is the dimension of vehicle characteristics, yes Middle Features, yes Middle Features and The vehicle nodes are and vehicle nodes exist Time vehicle characteristics; Calculate the time similarity of vehicle nodes: ； in, It is a vehicle node The temporal similarity of is the time factor, is the time interval; Based on spatial and temporal similarity, fill the zero values ​​in vehicle features: ； Calculate the relative characteristics between vehicle nodes, the calculation formula is: ； in, The vehicle nodes are velocity, position and acceleration, The vehicle nodes are velocity, position and acceleration.

3. The vehicle clustering method of the vehicle named data network as claimed in claim 2, characterized in that: The connectivity prediction model adopts a recurrent neural network model; The connectivity prediction model adopts a binary cross entropy loss function; The activation function of the output layer of the connectivity prediction model is a Softmax function; The output label of the connectivity prediction model is the connectivity between vehicle nodes. ; ； in, is the communication range of the vehicle node.

4. The vehicle clustering method of the vehicle named data network as claimed in claim 3, characterized in that: A denoising autoencoder is used to perform feature dimension reduction on the input of the connectivity prediction model; and noise disturbance is added to the input layer of the denoising autoencoder; comprising: The vehicle characteristics , and relative characteristics Normalize and merge to generate the first feature vectors of the two vehicle nodes; Inputting the first feature vector into a denoising autoencoder to obtain feature vectors of two vehicle nodes for input into a connectivity prediction model; The training process of the denoising autoencoder includes: calculating the error between the feature vector and the first feature vector, and updating the weight of the denoising autoencoder through a back propagation algorithm until the error is within a preset range.

5. The vehicle clustering method of the vehicle-mounted named data network as claimed in claim 4, characterized in that: The connectivity confidence weighted graph is ,in, is a collection of vehicle nodes; yes The edge set of is the weight function used to calculate The weight of each edge in ; in, ; in, is the feature vector distance, calculated as: ； In the formula, and They are the eigenvectors and No. embedding vectors, is the dimension of the embedding vector; The S3 includes: For any vehicle node , Get vehicle node Connectivity confidence interval for : ; According to the vehicle node The connectivity confidence interval and the connectivity between vehicle nodes are obtained The edge set of In response to , then and Add the edges between , and the corresponding weight is 1; response , then and Add the edges between and through the weight function Calculate the corresponding weights.

6. The vehicle clustering method of the vehicle named data network as claimed in claim 5, characterized in that: The S4 includes: For vehicle nodes Connectivity confidence weighted graph of , In response to ,but is the cluster head vehicle; in, for The set of neighbor nodes of is the distance threshold, is the neighbor threshold; The vehicle cluster consists of a cluster head vehicle and member vehicles, and the member vehicles are vehicle nodes that have connectivity with the cluster head vehicle.

7. A collaborative caching method based on vehicle clustering as claimed in any one of claims 1 to 6, characterized in that: The cluster head vehicle is the execution subject, including: Obtain the important nodes of the vehicle cluster according to the node betweenness of all vehicle nodes in the vehicle cluster; Receive request content and corresponding content popularity sent by member vehicles; The requested content is cached in important nodes of the vehicle cluster according to the content popularity.

8. The collaborative caching method based on vehicle clustering as claimed in claim 7, characterized in that: The node betweenness is stored in the router in the vehicle named data network. The calculation formula of node betweenness is: ； in, , is the set of all vehicle nodes in the vehicle cluster, Through the node Slave Node To Node The number of shortest paths, It is a slave node To Node The total number of all shortest paths; The important nodes of the vehicle cluster include: first-level important nodes and second-level important nodes; The important nodes for obtaining the vehicle cluster include: Sort all vehicle nodes in the vehicle cluster in descending order according to the node betweenness; First put the front The vehicle nodes are marked as first-level important nodes, and the remaining vehicle nodes are marked as The vehicle nodes are marked as secondary important nodes; among them, and is the default value.

9. The collaborative caching method based on vehicle clustering as claimed in claim 8, characterized in that: The content popularity is obtained by a popularity prediction model deployed on a member vehicle; The popularity prediction model is used to predict future popularity values ​​based on historical feature data of the request content; wherein the historical feature data includes: historical popularity features and geographic location features; The popularity prediction model adopts an encoder-decoder structure with a bidirectional long short-term memory BiLSTM network and a multi-head attention mechanism, including: an encoder based on a BiLSTM network and a decoder centered on multiple attentions.

10. The collaborative caching method based on vehicle clustering as claimed in claim 9, characterized in that: Cache the requested content to important nodes in the vehicle cluster based on content popularity, including: Divide content popularity into level; Obtain a probability matrix based on the number of important nodes and content popularity level of the vehicle cluster; The probability matrix is ​​sent to all important nodes of the vehicle cluster so that all important nodes perform caching according to the probability matrix; The probability matrix is ​​used to pair important nodes and content popularity levels; the number of its rows and columns are equal to the number of important nodes and content popularity levels of the vehicle cluster, respectively; Elements of the probability matrix For the Important nodes are used to cache content with a popularity of The probability of the request content is calculated as follows: ； in, is the number of nodes in the vehicle cluster, for The proportion of popular content in the content popularity, the number of nodes in the vehicle cluster, ; It means the Important nodes are not used to cache content with a popularity of The content of the request; is the first weight factor, is the second weight factor, .

Images