Vehicle edge caching method and device based on de-noising diffusion probability model
By generating HI pairs with hash codes and numbers and uploading them to roadside units, integrated knowledge of similar vehicle users is obtained for local distillation training. This solves the privacy protection and communication overhead problems in vehicle edge caching, and achieves efficient model training and improved cache hit rate.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-09
- Publication Date
- 2026-03-13
AI Technical Summary
Model training in vehicle edge caching is difficult to protect user privacy, has excessive communication overhead, and is not robust.
By generating HI pairs with hash codes and numbers and uploading them to roadside units, integrated knowledge of similar vehicle users is obtained for local distillation training, reducing communication overhead and improving model training efficiency. The knowledge is updated using roadside units and cached content is requested.
It effectively protects user privacy, significantly reduces communication overhead, improves model training efficiency and cache hit rate, and has good robustness.
Smart Images
Figure CN121665289A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of vehicle edge caching technology, and in particular to a vehicle edge caching method and apparatus based on a denoising diffusion probability model. Background Technology
[0002] With the rapid growth in demand for autonomous driving and connected vehicle applications, vehicles need to acquire map navigation data and multimedia content in real time to improve driving efficiency and user experience. Vehicle edge caching technology can significantly reduce the latency of content requests and the pressure on the core network by pre-caching content of interest to vehicle users (VUs) in Roadside Units (RSUs). However, the caching capacity of RSUs is limited; therefore, accurately predicting the content of interest to vehicle users is crucial.
[0003] Denoising Diffusion Probabilistic Models (DDPMs) have garnered significant attention due to their superior generative performance. DDPMs can generate high-quality and diverse content samples through an inverse denoising process. Applying DDPMs to edge caching can, to some extent, achieve more accurate content prediction, enabling vehicles to access content more suited to their needs at the RSU (Remote Subsystem Unit). However, the training and deployment of DDPMs require substantial computational resources, making them difficult to apply directly to computationally constrained in-vehicle environments. The lightweight U-Net architecture of DDPMs has proven effective in reducing the model's computational resource requirements. Therefore, adopting a lightweight DDPM (LDPM) is an effective solution. However, DDPM training still requires users' personal data, which is inconsistent with the growing demand for user privacy protection.
[0004] While Federated Learning (FL) can protect user privacy by sharing local models instead of raw data, each iteration of FL requires the transfer of complete model parameters between the vehicle and the server, resulting in significant communication overhead.
[0005] Furthermore, the high mobility of vehicles poses a significant challenge to the robustness of vehicle edge caching. Vehicles may leave the coverage area of the RSU before model training is complete, leading to training failure. Previous methods typically performed vehicle selection before training to ensure successful completion. However, this approach is not suitable for scenarios with slow model convergence, large amounts of local data, and large model sizes. Summary of the Invention
[0006] Therefore, the technical problem to be solved by the present invention is to overcome the problems of existing vehicle edge caching model training being difficult to protect user privacy, having excessive communication overhead, and poor robustness.
[0007] To address the aforementioned technical problems, this invention provides a vehicle edge caching method based on a denoising diffusion probability model, comprising: When a vehicle user enters the coverage area of a roadside unit, a hash code is generated based on local data, and an HI pair consisting of the hash code and the vehicle user's ID is uploaded to the roadside unit. The roadside unit stores the HI pairs of the current vehicle user and obtains multiple similar vehicle users with similar preferences based on the HI pairs; it then obtains integrated knowledge based on the knowledge of the multiple similar vehicle users; and finally sends the integrated knowledge to the current vehicle user. The current vehicle user uses local data and integrated knowledge to perform local distillation training on a lightweight denoising diffusion probability model; based on the lightweight denoising diffusion probability model that has completed local distillation training, the current vehicle user's model output knowledge is obtained; based on the model output knowledge, the current vehicle user's content recommendation list is obtained; the current vehicle user sends the KI pair consisting of the model output knowledge and the vehicle user's ID, as well as the content recommendation list, to the roadside unit. The roadside unit updates its knowledge of the current vehicle user based on the KI and requests cached content from the macro base station based on the content recommendation list of vehicle users within its coverage area.
[0008] Preferably, the roadside unit periodically synchronizes knowledge cache from the macro base station, the steps of which include: The roadside unit periodically uploads all vehicle user information it stores to the macro base station, including HI pairs and KI pairs; The macro base station receives vehicle user information uploaded by all roadside units within its coverage area, and saves the latest information for each vehicle user based on the upload time, thus obtaining the updated vehicle user information. Macro base stations send the updated vehicle user information to each roadside unit within their coverage area.
[0009] Preferably, the method by which vehicle users generate hash codes based on local data is as follows: Vehicle users encode local data using a pre-trained encoder to obtain the current vehicle user's hash code.
[0010] Preferably, the method for obtaining multiple similar vehicle users with similar preferences to the current vehicle user based on HI is as follows: Calculate the cosine similarity between the HI pair of the current vehicle user and the HI pairs of other vehicle users stored in the roadside unit, and select the top C vehicle users with the largest cosine similarity as similar vehicle users.
[0011] Preferably, the integrated knowledge is acquired based on the knowledge of multiple similar vehicle users, and the method is as follows: Acquire knowledge from multiple users of similar vehicles, calculate their average value, and use it as integrated knowledge.
[0012] Preferably, the current vehicle user utilizes local data and integrated knowledge to perform local distillation training on the lightweight denoising diffusion probability model, as follows: By encoding local data using a pre-trained encoder, a low-dimensional latent space can be obtained. ; For low-dimensional potential space Each data in By utilizing ensemble knowledge, a lightweight denoising diffusion probability model is trained locally using an optimization function.
[0013] Preferably, the optimization function is expressed as follows: ; in, Let i be the optimization function for vehicle user i, where i is the vehicle user's ID. These are the parameters of a lightweight denoising diffusion probability model for vehicle user i. For the low-dimensional potential space of vehicle user i, For data in low-dimensional latent space, To train the loss function, As a weighting factor, For divergence term, For the softmax function, For temperature parameters, To integrate knowledge.
[0014] Preferably, based on the model output knowledge, the content recommendation list for the current vehicle user is obtained, including the following steps: The model output knowledge is processed by a pre-trained decoder to obtain reconstructed data; each dimension of the reconstructed data includes a score for each content item. The reconstructed data is averaged along each dimension to obtain the target score for each content item; Select the top M content items with the highest target ratings as the content recommendation list for the current vehicle user.
[0015] Preferably, the roadside unit requests cached content from the macro base station based on a content recommendation list for all vehicle users within its coverage area, using the following method: The roadside unit calculates a score for each content item based on the vehicle user's location and speed, according to the content recommendation list of the vehicle user; The roadside unit selects the top N content items with the highest scores and requests cached content from the macro base station.
[0016] The present invention also provides a vehicle edge buffer device based on a denoising diffusion probability model, comprising: The hash code upload module is used to generate a hash code based on local data when a vehicle user enters the coverage area of the roadside unit, and upload a HI pair consisting of the hash code and the vehicle user's number to the roadside unit; An integrated knowledge acquisition module is used by the roadside unit to store the HI pairs of the current vehicle users and to acquire multiple similar vehicle users with similar preferences based on the HI pairs; to acquire integrated knowledge based on the knowledge of multiple similar vehicle users; and to send the integrated knowledge to the current vehicle user. The content recommendation list acquisition module is used by the current vehicle user to perform local distillation training on a lightweight denoising diffusion probability model using local data and integrated knowledge; based on the lightweight denoising diffusion probability model that has completed local distillation training, the module obtains the model output knowledge of the current vehicle user; based on the model output knowledge, the module obtains the content recommendation list of the current vehicle user; and the current vehicle user sends the KI pair consisting of the model output knowledge and the vehicle user's ID, as well as the content recommendation list, to the roadside unit. The content update module is used by roadside units to update the knowledge of the current vehicle users based on KI pairs, and to request cached content from macro base stations based on the content recommendation list of vehicle users within their coverage area.
[0017] Compared with the prior art, the above-described technical solution of the present invention has the following advantages: This invention discloses a vehicle edge caching method based on a denoising diffusion probability model. First, it encodes the vehicle user's local data into hash codes and sends them to roadside units. The hash codes have a much smaller dimensionality than the original local data, ensuring that sharing hash codes between the vehicle user and the roadside unit does not leak user privacy. Second, the invention obtains similar users with similar preferences to the current vehicle user from the roadside unit, aggregates their knowledge into integrated knowledge, and sends it to the vehicle user, reducing communication overhead. Third, the vehicle user trains a local denoising diffusion probability model based on local data and integrated knowledge. Training in a low-dimensional space significantly reduces the computational resources required for training and improves training efficiency. Furthermore, introducing integrated knowledge during training effectively improves the model's prediction accuracy for content of interest to the user. Finally, the content recommendation list generated by the trained model is sent to the roadside unit. The roadside unit requests cached content from the macro base station based on the content recommendation lists of all vehicle users within its coverage area. This allows content of interest to the user to be pre-cached at the roadside unit, reducing the latency for vehicle users accessing content and maximizing the cache hit rate. This invention exhibits good robustness to vehicle speed changes and significantly reduces communication overhead while improving cache hit rate. Attached Figure Description
[0018] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is a flowchart of a vehicle edge caching method based on a denoising diffusion probability model according to the present invention; Figure 2 This is a schematic diagram of a highway scene according to the present invention; Figure 3 Convergence curve of local LDPM training for vehicle users; Figure 4 A graph showing the change in latency for requesting content as cache capacity changes; Figure 5 This is a comparison chart showing how cache hit rate changes with vehicle speed. Detailed Implementation
[0019] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.
[0020] Reference Figure 1 As shown, this invention provides a vehicle edge caching method based on a denoising diffusion probability model, comprising: S1: When a vehicle user enters the coverage area of a roadside unit, a hash code is generated based on local data, and an HI pair consisting of the hash code and the vehicle user's number is uploaded to the roadside unit; S2: The roadside unit stores the HI pairs of the current vehicle user and obtains multiple similar vehicle users with similar preferences based on the HI pairs; it obtains integrated knowledge based on the knowledge of multiple similar vehicle users; and the roadside unit sends the integrated knowledge to the current vehicle user. S3: The current vehicle user uses local data and integrated knowledge to perform local distillation training on the lightweight denoising diffusion probability model; based on the lightweight denoising diffusion probability model that has completed local distillation training, the current vehicle user's model output knowledge is obtained; based on the model output knowledge, the current vehicle user's content recommendation list is obtained; the current vehicle user sends the KI pair consisting of the model output knowledge and the vehicle user's ID, as well as the content recommendation list, to the roadside unit. S4: The roadside unit updates the knowledge of the current vehicle user based on the KI and requests cached content from the macro base station based on the content recommendation list of vehicle users within its coverage area.
[0021] Specifically, the present invention considers as follows Figure 2The diagram illustrates a highway system scenario. In this scenario, the vehicle edge computing network consists of three layers: macro base stations (MBS), roadside units, and vehicle users. Each roadside unit... Connected to a macro base station via a reliable wired link, while vehicle users... It is located within the coverage area of the roadside unit.
[0022] Each vehicle user has their own local data. It is equipped with a communication module, a distillation module, and a locally pre-trained encoder and decoder network.
[0023] The speed of each vehicle follows an independent and identically distributed truncated Gaussian distribution. Let... Let represent the speed of vehicle user i within the coverage of the r-th roadside unit. probability density function It can be represented as: ; in, and These are the maximum and minimum speed thresholds for the vehicle, respectively. It is a Gaussian error function. The mean, Let be the variance, i be the vehicle user's ID, and r be the roadside unit ID.
[0024] Each RSU is equipped with a communication module, a content caching module, and a knowledge caching (KC) module. The KC module includes a hash-to-index (HI) module and a knowledge-to-index (KI) module. The HI and KI modules are used to store HI pairs (composed of the vehicle user's hash code and ID) and KI pairs (composed of the vehicle user's knowledge and ID), respectively, uploaded by the vehicle. The content caching module has a limited capacity, caching a maximum of N content fragments.
[0025] This invention assumes that the MBS stores all available content. When content requested by a vehicle user is already cached in the RSU, the RSU directly transmits the content to the vehicle. If it is not cached, the RSU must first request the content from the MBS before transmitting it to the vehicle, which leads to higher content request latency. The goal of this invention is to accurately predict the content that the VU is interested in and pre-cache it at the RSU, thereby reducing the latency of VU accessing content and maximizing the cache hit rate of cached content.
[0026] In the vehicle edge caching algorithm, during the execution phase between vehicles and RSUs, each vehicle i uploads its content recommendation list when entering the r-th RSU. This content recommendation list It is generated by vehicle users using local models. The process of interaction between vehicle users and roadside units to generate the content recommendation list is divided into S1 to S3, which are described in detail below.
[0027] In S1, the hash code is generated locally by the vehicle user. The generation method is as follows: Vehicle user i utilizes pre-trained encoders For local data Encode the hash code of the current vehicle user i. The formula is expressed as: .
[0028] The encoder is a deeply trained neural network whose output dimension is much smaller than the dimension of the original data, so sharing the hash code between vehicle users and RSUs does not compromise privacy. The pre-trained encoder and decoder networks are first trained using publicly available datasets and then fine-tuned using local data.
[0029] The vehicle user will combine the hash code and the current vehicle user's ID into a HI pair and upload it to the RSU.
[0030] Upon receiving the HI pair uploaded by vehicle user i, the RSU will immediately save the uploaded HI pair and update the corresponding content in the HI module, that is: .
[0031] The roadside unit uses HI (Highly Important Person) pairs to obtain multiple similar vehicle users with similar preferences to the current vehicle user. The method is as follows: RSU calculates the cosine similarity between the HI pair of the current vehicle user and the HI pairs of other vehicle users stored in the roadside unit, and selects the top C vehicle users with the highest cosine similarity as similar vehicle users. The formula is expressed as: ; ; in, For the HI pair of vehicle user i in roadside unit r. For vehicle users in roadside unit r HI is correct. For users of similar vehicles, This is the minimum threshold for cosine similarity.
[0032] When acquiring similar vehicle users, this invention calculates the similarity between vehicle users, rather than the similarity between individual samples. The main reason is that vehicle users only rate a limited number of items, resulting in a large number of zero values in their data. Directly calculating the similarity between each sample would treat these unrated items as content that VUs are not interested in, leading to biased predictions.
[0033] RSU extracts knowledge about similar vehicle users from the KC module for C similar vehicle users to the current vehicle user i.
[0034] Preferably, the integrated knowledge is acquired based on the knowledge of multiple similar vehicle users, and the method is as follows: RSU retrieves knowledge about C similar vehicle users to the current vehicle user i from the KC module; then, it calculates the average of the knowledge from all similar vehicle users as the integrated knowledge, expressed by the formula: ; ; in, For knowledge of users of similar vehicles, This refers to the knowledge of the c-th similar vehicle user of vehicle user i in roadside unit r. C represents the total number of users with similar vehicles. The integrated knowledge generated for roadside unit r for vehicle user i.
[0035] The roadside unit r will integrate knowledge Send to vehicle user i.
[0036] Vehicle users utilize local data and received integrated knowledge to perform local distillation training.
[0037] Vehicle user i utilizes local data and integrated knowledge The lightweight denoising diffusion probability model is trained by local distillation, as follows: Because LDPM's low-dimensional space training can significantly reduce the required computational resources and improve training efficiency, a pre-trained encoder is used to encode the local data before training, transforming the original local data... Mapping to a lower-dimensional latent space yields the lower-dimensional latent space. Then, a local LDPM is trained in this low-dimensional latent space; For low-dimensional potential space Each data in By utilizing ensemble knowledge, a lightweight denoising diffusion probability model is trained locally using an optimization function.
[0038] Specifically, the formula for the optimization function is as follows: ; in, Let i be the optimization function for vehicle user i, where i is the vehicle user's ID. These are the parameters of a lightweight denoising diffusion probability model for vehicle user i. For the low-dimensional potential space of vehicle user i, For data in low-dimensional latent space, To train the loss function, As a weighting factor, For divergence term, For the softmax function, For temperature parameters, To integrate knowledge.
[0039] After completing local distillation training, vehicle user i uses the local LDPM that has completed local distillation training to perform a backdiffusion process, outputting knowledge based on the Gaussian noise generation model. The model outputs knowledge. Including F data points, there are F represents the number of sample data points.
[0040] Based on model output knowledge Get the content recommendation list for vehicle user i The steps include: Output knowledge from the model The pre-trained decoder processes each data point. Mapping back to the original data dimensions yields the reconstructed data; the individual data points of the reconstructed data are... The formula is expressed as: , For decoder networks; The reconstructed data is used for content popularity prediction. Compared with the original data, it better reflects VU's preferences for content because it contains fewer zero values. Specifically, each dimension of the reconstructed data includes a score for each content item, with a total of F dimensions. The reconstructed data is averaged along each dimension to obtain the target score for each content item; Assuming there are K content items in total, the reconstructed data after averaging by dimension is... The length of is K, which can be represented as The formula is: ; Target score It can reflect the preferences of vehicle user i; the higher the target score, the more interested vehicle user i is in that content. The top M content items with the highest target ratings will be selected as the content recommendation list for the current vehicle user. .
[0041] Vehicle user i sends the model output knowledge and the KI pair consisting of the vehicle user's ID, along with a content recommendation list, to the roadside unit.
[0042] Among them, KI outputs knowledge from the model of vehicle users. It consists of a number, that is: .
[0043] The roadside unit updates its knowledge of the current vehicle user based on the KI and requests cached content from the macro base station based on the content recommendation list of vehicle users within its coverage area.
[0044] In the execution phase between RSUs and MBS in the vehicle edge caching algorithm, when a vehicle user within the coverage area of a certain RSU uploads a content recommendation list, that RSU sends a request to the macro base station to update its cached content. Furthermore, to prevent the information stored in the KC module from becoming outdated after a vehicle leaves the RSU coverage area, the RSUs periodically update and synchronize the content of the KC module, i.e., updating the knowledge of the current vehicle user based on the KI (Knowledge Base Station).
[0045] Specifically, the roadside unit requests cached content from the macro base station based on a content recommendation list for all vehicle users within its coverage area, using the following method: The roadside unit calculates a score for each content item in the content recommendation list for each vehicle user based on the vehicle user's location and speed; since vehicle users who stay in the roadside unit for a longer time are more likely to request more content, their content recommendation list should be given greater weight. For each content item k, the formula for calculating its score is: ; in, For the roadside unit r, the score for content item k is given. For the geometry of vehicle users within the coverage area of roadside unit r, Location weighting factor The coverage area of the roadside unit. For indicator functions, The distance from vehicle user i to the entrance of the roadside unit. The speed of vehicle user i;
[0046] The roadside unit selects the top N content items with the highest scores and requests cached content from the macro base station.
[0047] Preferably, the roadside unit periodically synchronizes knowledge cache from the macro base station, the steps of which include: The roadside unit periodically uploads information on all vehicle users stored in its KC module to the macro base station, including HI pairs and KI pairs; The macro base station receives vehicle user information uploaded by all roadside units within its coverage area, and saves the latest information for each vehicle user based on the upload time, thus obtaining the updated vehicle user information. ; The formula for obtaining updated vehicle user information is expressed as follows: ; ; in, The time when vehicle user i uploads the HI pair to roadside unit r. A roadside unit for storing the latest information of vehicle user i. The total number of vehicle users; Macro base stations will update vehicle user information Send to each roadside unit within the coverage area.
[0048] To verify the effectiveness of the method of the present invention, relevant simulation experiments were conducted in this embodiment, and the experimental results are as follows.
[0049] Table 1. Cache hit rate and communication overhead under different schemes
[0050] As shown in Table 1, the cache hit rate of various schemes gradually increases with the increase of cache capacity. This is because a higher cache capacity allows RSU to cache more content, making it more likely that VUs will request the required content from RSU. Oracle achieves the best performance and represents the theoretical optimal value because it can predict future requests from VUs. It can be seen that, except for Oracle, the method of this invention performs best because this scheme trains a personalized local LDPM for each VU, which can accurately predict the content of interest for each VU. AsyFed's performance is lower than CAFR because, for high-speed moving vehicles, AsyFed has difficulty completing training before they leave the RSU coverage area. FedAvg is more severely affected by vehicle speed, thus its performance is inferior to AsyFed. The N-τ-greedy algorithm performs the worst because it is a non-learning algorithm. Furthermore, it can be seen that the method of this invention has significantly lower communication overhead than other schemes. Compared with traditional FedAvg and AsyFed, communication overhead is reduced by more than 98%. This is because the method of this invention only transmits a small number of HI and KI pairs, rather than the complete model parameter set.
[0051] Figure 3The convergence curve of the local LDPM training in this invention is shown. It can be observed that as the number of training rounds increases, the model loss gradually decreases and converges at approximately 300 rounds.
[0052] Figure 4 The diagram illustrates the changes in content request latency for various schemes under different cache capacities. It can be seen that as the cache capacity increases, the content request latency of all schemes gradually decreases. This is because a larger cache capacity makes it more likely that VUs will retrieve the requested content from nearby RSUs, thereby reducing content request latency. Furthermore, it can be observed that the method of this invention achieves the lowest content request latency, second only to Oracle. This is because the method of this invention has a higher cache hit rate, making it more likely that VUs will request the content they are interested in, thus reducing content request latency.
[0053] Figure 5 The maximum cache hit rate of various schemes is shown at different vehicle speeds. It can be seen that the method of this invention is almost unaffected by vehicle speed. This is because the method of this invention performs model training locally, without frequent interaction with the RSU; even if a vehicle enters a new RSU, a content recommendation list matching the VU's preferences can be uploaded immediately. Furthermore, this invention also considers the vehicle's movement characteristics and dynamically updates the cached content. CAFR is also almost unaffected by changes in vehicle speed, but the method of this invention performs better because it trains a personalized local LDPM for each VU. As vehicle speed increases, the cache hit rate of FedAvg and AsyFed gradually decreases, because the vehicle leaves the RSU too quickly, causing the model training to fail.
[0054] In summary, the vehicle edge caching method based on a denoising diffusion probability model described in this invention encodes the local data of vehicle users into hash codes and sends them to roadside units. The hash code's dimensionality is much smaller than the original local data, ensuring that sharing the hash code between vehicle users and roadside units does not leak user privacy. Secondly, this invention obtains similar users with similar preferences to the current vehicle user from the roadside unit, aggregates their knowledge into integrated knowledge, and sends it to the vehicle user, reducing communication overhead. Furthermore, the vehicle user trains a local denoising diffusion probability model based on local data and integrated knowledge. Training in a low-dimensional space significantly reduces the computational resources required for training and improves training efficiency. Introducing integrated knowledge during training effectively improves the model's prediction accuracy for content of interest to the user. Finally, the content recommendation list generated by the trained model is sent to the roadside unit. The roadside unit requests cached content from the macro base station based on the content recommendation lists of all vehicle users within its coverage area, allowing content of interest to the user to be pre-cached at the roadside unit, reducing the latency for vehicle users accessing content and maximizing the cache hit rate. This invention exhibits good robustness to vehicle speed changes and significantly reduces communication overhead while improving cache hit rate.
[0055] Based on the above-mentioned vehicle edge caching method based on a denoising diffusion probability model, the present invention also provides a vehicle edge caching device based on a denoising diffusion probability model, comprising: The hash code upload module is used to generate a hash code based on local data when a vehicle user enters the coverage area of the roadside unit, and upload a HI pair consisting of the hash code and the vehicle user's number to the roadside unit; An integrated knowledge acquisition module is used by the roadside unit to store the HI pairs of the current vehicle users and to acquire multiple similar vehicle users with similar preferences based on the HI pairs; to acquire integrated knowledge based on the knowledge of multiple similar vehicle users; and to send the integrated knowledge to the current vehicle user. The content recommendation list acquisition module is used by the current vehicle user to perform local distillation training on a lightweight denoising diffusion probability model using local data and integrated knowledge; based on the lightweight denoising diffusion probability model that has completed local distillation training, the module obtains the model output knowledge of the current vehicle user; based on the model output knowledge, the module obtains the content recommendation list of the current vehicle user; and the current vehicle user sends the KI pair consisting of the model output knowledge and the vehicle user's ID, as well as the content recommendation list, to the roadside unit. The content update module is used by roadside units to update the knowledge of the current vehicle users based on KI pairs, and to request cached content from macro base stations based on the content recommendation list of vehicle users within their coverage area.
[0056] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0057] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0058] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0059] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0060] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.
Claims
1. A vehicle edge caching method based on a denoising diffusion probability model, characterized in that, include: When a vehicle user enters the coverage area of a roadside unit, a hash code is generated based on local data, and an HI pair consisting of the hash code and the vehicle user's ID is uploaded to the roadside unit. The roadside unit stores the HI pairs of the current vehicle user and obtains multiple similar vehicle users with similar preferences based on the HI pairs; it also obtains integrated knowledge based on the knowledge of multiple similar vehicle users. The roadside unit will integrate the knowledge and send it to the current vehicle user; Currently, vehicle users are using local data and integrated knowledge to perform local distillation training on a lightweight denoising diffusion probability model; Based on the lightweight denoised diffusion probability model that has completed local distillation training, the model output knowledge of the current vehicle user is obtained; based on the model output knowledge, the content recommendation list of the current vehicle user is obtained. The current vehicle user sends a KI pair consisting of the model output knowledge and the vehicle user's ID, along with a content recommendation list, to the roadside unit; The roadside unit updates its knowledge of the current vehicle user based on the KI and requests cached content from the macro base station based on the content recommendation list of vehicle users within its coverage area.
2. The vehicle edge caching method based on a denoising diffusion probability model according to claim 1, characterized in that, Roadside units periodically synchronize knowledge cache from macro base stations. The steps include: The roadside unit periodically uploads all vehicle user information it stores to the macro base station, including HI pairs and KI pairs; The macro base station receives vehicle user information uploaded by all roadside units within its coverage area, and saves the latest information for each vehicle user based on the upload time, thus obtaining the updated vehicle user information. Macro base stations send the updated vehicle user information to each roadside unit within their coverage area.
3. The vehicle edge caching method based on a denoising diffusion probability model according to claim 1, characterized in that, The method used by vehicle users to generate hash codes based on local data is as follows: Vehicle users encode local data using a pre-trained encoder to obtain the current vehicle user's hash code.
4. The vehicle edge caching method based on a denoising diffusion probability model according to claim 1, characterized in that, Based on HI (Internet Information Services), multiple similar vehicle users with similar preferences to the current vehicle user are obtained using the following method: Calculate the cosine similarity between the HI pair of the current vehicle user and the HI pairs of other vehicle users stored in the roadside unit, and select the top C vehicle users with the largest cosine similarity as similar vehicle users.
5. The vehicle edge caching method based on a denoising diffusion probability model according to claim 1, characterized in that, The method for integrating knowledge based on knowledge from multiple similar vehicle users is as follows: Acquire knowledge from multiple users of similar vehicles, calculate their average value, and use it as integrated knowledge.
6. The vehicle edge caching method based on a denoising diffusion probability model according to claim 1, characterized in that, Currently, vehicle users utilize local data and integrated knowledge to perform local distillation training on a lightweight denoising diffusion probability model. The method is as follows: By encoding local data using a pre-trained encoder, a low-dimensional latent space can be obtained. ; For low-dimensional potential space Each data in By utilizing ensemble knowledge, a lightweight denoising diffusion probability model is trained locally using an optimization function.
7. The vehicle edge caching method based on a denoising diffusion probability model according to claim 1, characterized in that, The formula for the optimization function is as follows: in, Let i be the optimization function for vehicle user i, where i is the vehicle user's ID. These are the parameters of a lightweight denoising diffusion probability model for vehicle user i. For the low-dimensional potential space of vehicle user i, For data in low-dimensional latent space, To train the loss function, As a weighting factor, For divergence term, For the softmax function, For temperature parameters, To integrate knowledge.
8. The vehicle edge caching method based on a denoising diffusion probability model according to claim 1, characterized in that, Based on the model's output knowledge, the following steps are taken to obtain the content recommendation list for the current vehicle user: The model output knowledge is processed by a pre-trained decoder to obtain reconstructed data; each dimension of the reconstructed data includes a score for each content item. The reconstructed data is averaged along each dimension to obtain the target score for each content item; Select the top M content items with the highest target ratings as the content recommendation list for the current vehicle user.
9. A vehicle edge caching method based on a denoising diffusion probability model according to claim 1, characterized in that, The roadside unit requests cached content from the macro base station based on a content recommendation list for all vehicle users within its coverage area, using the following method: The roadside unit calculates a score for each content item based on the vehicle user's location and speed, according to the content recommendation list of the vehicle user; The roadside unit selects the top N content items with the highest scores and requests cached content from the macro base station.
10. A vehicle edge buffer device based on a denoising diffusion probability model, characterized in that, include: The hash code upload module is used to generate a hash code based on local data when a vehicle user enters the coverage area of the roadside unit, and upload a HI pair consisting of the hash code and the vehicle user's number to the roadside unit; An integrated knowledge acquisition module is used to store the HI pairs of the current vehicle users in the roadside unit, and to acquire multiple similar vehicle users with similar preferences based on the HI pairs; and to acquire integrated knowledge based on the knowledge of multiple similar vehicle users. The roadside unit will integrate the knowledge and send it to the current vehicle user; The content recommendation list acquisition module is used by the current vehicle user to perform local distillation training on a lightweight denoising diffusion probability model using local data and integrated knowledge. Based on the lightweight denoised diffusion probability model that has completed local distillation training, the model output knowledge of the current vehicle user is obtained; based on the model output knowledge, the content recommendation list of the current vehicle user is obtained. The current vehicle user sends a KI pair consisting of the model output knowledge and the vehicle user's ID, along with a content recommendation list, to the roadside unit; The content update module is used by roadside units to update the knowledge of the current vehicle users based on KI pairs, and to request cached content from macro base stations based on the content recommendation list of vehicle users within their coverage area.