Internet of vehicles cache prediction method based on cloud edge collaboration

Abstract

The invention discloses an Internet of Vehicles cache prediction method based on cloud edge collaboration, and belongs to the field of computer networks. The problems that the number of requests is large and the performance of edge nodes is limited exist in the Internet of Vehicles under an edge computing architecture, and the cache hit rate at the edge nodes is improved through collaborative prediction of the edge nodes and a cloud server. And a decision maker is used at the edge node to determine whether the prediction result of the edge node needs to be verified by the cloud server, so that the accuracy and the efficiency can be better balanced. Experiments are carried out on an Amazon data set, and compared with an existing cache prediction method, the method can effectively improve the cache hit rate of edge nodes and reduce data transmission delay.

Description

Technical Field

[0001] The present invention belongs to the field of computer networks, and in particular relates to the design of a collaborative caching mechanism, and specifically to a vehicle network cache prediction method based on cloud-edge collaboration. Background Art

[0002] As the number of vehicles increases exponentially, data requests in traditional connected vehicles (IoVs) are also increasing. However, the cloud computing model, which emphasizes centralized data processing in the cloud, results in high latency during data transmission, failing to meet user needs. Consequently, edge computing has become a common network architecture in IoVs. Edge computing pushes data processing to edge devices closer to the terminal, enabling lower latency and more efficient data processing. Furthermore, using NDN networks to connect terminals and edge devices is a mainstream solution. Because NDN networks are content-centric, requests from vehicle terminals to edge devices do not require end-to-end connectivity, as is required with IP networks. This effectively overcomes the rapid changes in topology state within IoVs. Summary of the Invention

[0003] The present invention provides a vehicle network cache prediction method based on cloud-edge collaboration in edge computing, which solves the problems of excessive vehicle requests and limited edge node performance in the prior art.

[0004] The present invention is implemented through the following technical solution: a method for predicting Internet of Vehicles cache based on cloud-edge collaboration, comprising the following steps:

[0005] Step 1: When the edge node receives the request sent by the terminal, it generates a request sequence in chronological order and uses the approximate model for prediction;

[0006] The edge node receives interest packets from different terminals and collects the user's request sequence in chronological order, and then combines the names of their interest packets into a request sequence. The request sequence is word-embedded and position-embedded. The word embedding is performed by maintaining a trainable matrix To complete the vocabulary mapping, where V is the size of the vocabulary and d is the dimension of the word embedding. Each word in the request sequence is indexed in the embedding matrix to find the corresponding vector, so that the entire request sequence is combined into a matrix. Then, position embedding is performed on this matrix. Position embedding generates the code by alternating sine and cosine functions and obtains the matrix X. The specific formulas are shown in Formula 1 and Formula 2. Where PE represents position encoding, pos is the position in the sequence, i is the dimension, and d is the position of the word. model Represents the dimension of word embedding.

[0007]

[0008]

[0009] Convert X into query matrix Q, key-value matrix K and value matrix V through linear transformation. Obtain query matrix Q through formula 3, obtain key-value matrix K through formula 4, and obtain value matrix V through formula 5.

[0010] Q=XW Q (Formula 3)

[0011] K=XW K (Formula 4)

[0012] V=XW V (Formula 5)

[0013] Among them, W Q ,W K , These are all learnable parameter matrices that allow the model to learn to adjust its overall focus based on different contexts. After obtaining the Q, K, and V matrices, we need to calculate the attention score matrix. This is done by scaling the dot product attention matrix R. The calculation process is shown in Equation 6.

[0014]

[0015] Among them This is to prevent data instability and gradient disappearance problems. If the value of d is large, the overall value may be very large if the dot product is not scaled, thus affecting the subsequent Softmax function.

[0016] After obtaining the attention score matrix R, Softmax normalization is applied to it. The Softmax function is applied to each row of S, and the attention weight matrix A is obtained, as shown in Formula 7.

[0017]

[0018] Softmax normalization converts the attention score into a probability distribution, ensuring that the sum of the elements in each row of the matrix is 1. This approach allows the model to focus more on the most important part of the whole and reduce attention to unimportant information.

[0019] Then, in order to evaluate the global importance of each token in the sequence, it is necessary to sum the columns of the attention weight matrix A and generate a row vector s, as shown in Formula 8.

[0020]

[0021] The essence of obtaining the row vector s is to accumulate the attention values of other positions in the self-attention mechanism on the corresponding position. The higher the accumulated score, the more important the token at that position is in the entire sequence. Finally, the sorting output is performed. The original tokens are sorted based on the vector s to obtain the approximate model prediction result S and the probability value P = s corresponding to each token.

[0022] Step 2: The edge node uses the decision maker to evaluate the prediction result and choose to generate a content request or a verification request;

[0023] Map the probability value P to the latent space The specific process is shown in Formula 9.

[0024]

[0025] Where μ is the mean vector, which represents the center position in the latent space. σ is the standard deviation vector, which controls the width of the latent space distribution. ∈ is the standard normal distribution. The noise vector sampled in is used to introduce randomness and ensure the diversity of the generated latent variables.

[0026] The decoder generates a decision action a from the latent space t ∈{NoVerify,Verify}, as shown in Formula 10.

[0027] a t =Decoder(z) (Formula 10)

[0028] Here the decision action a t It can be one of two actions: NoVerify or Verify. NoVerify means that the prediction result of the approximate model is used without verification by the cloud server, while Verify means that the prediction result of the approximate model needs to be verified by the original model in the cloud server.

[0029] In generating decision action a t After that, it will not be executed directly, but will be executed based on the decision of e-greedy, that is, the decision action a will be adopted with probability e t , randomly selecting one of the two actions, Verify or NoVerify, with a probability of 1-e. This strategy helps balance exploration and exploitation in the multi-armed bandit model and prevents the model from converging prematurely. The specific process is shown in Equation 11.

[0030]

[0031] The performance evaluation of the decision maker depends on the change of cache hit rate. If the cache hit rate at the next moment is higher than the current moment, it means that the decision maker's decision has received positive feedback. The probability e of the final action of the decision will be updated according to the cache hit rate brought by the decision result. When the cumulative number of user requests received by the edge node reaches the threshold N, the edge node will trigger the cache prediction process. The decision maker will record the cache hit rate at the current moment t as h t After completing the entire cache prediction process, the next departure time is t+1, and the hit rate at this time is h t+1 The specific update rule is shown in Formula 12.

[0032] e k+1 =e k +ηsign(h t+1 -h t ) (Formula 12)

[0033] where e k is the current probability value, η is the learning rate, which determines the step size of each update. sign(h t+1 -h t ) is the sign function, which will be calculated based on the hit rate h at the next moment. t+1 and the current hit rate h t This update rule allows the decision maker to dynamically adjust the value of e based on the rewards brought by the actual decision action, thereby gradually optimizing the subsequent decision process.

[0034] Step 3: The cloud server uses the original model to verify the approximate model results;

[0035] The cloud server will receive a content request or verification request from the edge node. If the request received is a content request, it will simply return the content requested by the edge node; if the request received is a verification request, it is necessary to verify the prediction result of the edge node and return the data content corresponding to the verification result of the cloud server.

[0036] In the previous process, the original input sequence X and the prediction sequence S of the approximate model have been obtained, where X = [x1, x2, x3, ..., x m ],S=[s1,s2,s3,...,s t For each step t, it is necessary to verify the token s generated by the approximate model. t Is it reasonable? At each step, the original model M is used l To calculate the probability distribution of the same position, and select the word with the highest probability, as shown in Formula 13.

[0037] P l (yt |X,S ＜t )=softmax(M l X,S ＜t ) (Formula 13)

[0038] Among them, P l represents the probability distribution of the lth strategy or model branch, y t represents the output label or action at time step t, X represents the input data, and S ＜t Represents the historical state sequence up to time step t-1, that is, the past context information. This gives a probability distribution calculated by the original model, and then the word y with the highest probability can be selected based on the entire probability distribution. t , as shown in Formula 14.

[0039] y t =argmax y P l (y t =y|X,S ＜t ) (Formula 14)

[0040] This gives the validation output y of the original model t , then we need to compare the prediction results y of the original model t and the prediction results of the edge node approximation model s t Are they the same? If they are, the approximate model's prediction results are accepted. If they are different, the approximate model's prediction results are discarded and the original model's prediction results are used for update. The cloud server then transmits the data corresponding to the final verification result back to the edge node that issued the request. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 CECCP structure diagram;

[0042] Figure 2 Experimental topology diagram;

[0043] Figure 3 Hit rate results of algorithms with different cache sizes;

[0044] Figure 4 Hit rate results of the algorithm for different user request sequence sizes;

[0045] Figure 5 Recall rate and normalized discounted cumulative gain experimental results;

[0046] Figure 6 Latency results for different buffer sizes;

[0047] Figure 7 Latency results for different user request sequence sizes. DETAILED DESCRIPTION

[0048] A method for predicting cache in an Internet of Vehicles (IoV) based on cloud-edge collaboration includes the following steps:

[0049] Step 1: When the edge node receives the request sent by the terminal, it generates a request sequence in chronological order and uses the approximate model for prediction;

[0050] The edge node receives interest packets from different terminals and collects the user's request sequence in chronological order, and then combines the names of their interest packets into a request sequence. The request sequence is word-embedded and position-embedded. The word embedding is performed by maintaining a trainable matrix To complete the vocabulary mapping, where V is the size of the vocabulary and d is the dimension of the word embedding. Each word in the request sequence is indexed in the embedding matrix to find the corresponding vector, so that the entire request sequence is combined into a matrix. Then, position embedding is performed on this matrix. Position embedding generates the code by alternating sine and cosine functions and obtains the matrix X. The specific formulas are shown in Formula 1 and Formula 2. Where PE represents position encoding, pos is the position in the sequence, i is the dimension, and d is the position of the word. model Represents the dimension of word embedding.

[0051]

[0052] Convert X into query matrix Q, key-value matrix K and value matrix V through linear transformation. Obtain query matrix Q through formula 3, obtain key-value matrix K through formula 4, and obtain value matrix V through formula 5.

[0053] Q=XW Q (Formula 3)

[0054] K=XW K (Formula 4)

[0055] V=XW V (Formula 5)

[0056] Among them, W Q ,W K , These are all learnable parameter matrices that allow the model to learn to adjust its overall focus based on different contexts. After obtaining the Q, K, and V matrices, we need to calculate the attention score matrix. This is done by scaling the dot product attention matrix R. The calculation process is shown in Equation 6.

[0057]

[0058] Among them This is to prevent data instability and gradient disappearance problems. If the value of d is large, the overall value may be very large if the dot product is not scaled, thus affecting the subsequent Softmax function.

[0059] After obtaining the attention score matrix R, Softmax normalization is applied to it. The Softmax function is applied to each row of S, and the attention weight matrix A is obtained, as shown in Formula 7.

[0060]

[0061] Softmax normalization converts the attention score into a probability distribution, ensuring that the sum of the elements in each row of the matrix is 1. This approach allows the model to focus more on the most important part of the whole and reduce attention to unimportant information.

[0062] Then, in order to evaluate the global importance of each token in the sequence, it is necessary to sum the columns of the attention weight matrix A and generate a row vector s, as shown in Formula 8.

[0063]

[0064] The essence of obtaining the row vector s is to accumulate the attention values of other positions in the self-attention mechanism on the corresponding position. The higher the accumulated score, the more important the token at that position is in the entire sequence. Finally, the sorting output is performed. The original tokens are sorted based on the vector s to obtain the approximate model prediction result S and the probability value P = s corresponding to each token.

[0065] Step 2: The edge node uses the decision maker to evaluate the prediction result and choose to generate a content request or a verification request;

[0066] Map the probability value P to the latent space The specific process is shown in Formula 9.

[0067]

[0068] Where μ is the mean vector, which represents the center position in the latent space. σ is the standard deviation vector, which controls the width of the latent space distribution. ∈ is the standard normal distribution. The noise vector sampled in is used to introduce randomness and ensure the diversity of the generated latent variables.

[0069] The decoder generates a decision action a from the latent space t ∈{NoVerify,Verify}, as shown in Formula 10.

[0070] at =Decoder(z) (Formula 10)

[0071] Here the decision action a t It can be one of two actions: NoVerify or Verify. NoVerify means that the prediction result of the approximate model is used without verification by the cloud server, while Verify means that the prediction result of the approximate model needs to be verified by the original model in the cloud server.

[0072] In generating decision action a t After that, it will not be executed directly, but will be executed based on the decision of e-greedy, that is, the decision action a will be adopted with probability e t , randomly selecting one of the two actions, Verify or NoVerify, with a probability of 1-e. This strategy helps balance exploration and exploitation in the multi-armed bandit model and prevents the model from converging prematurely. The specific process is shown in Equation 11.

[0073]

[0074] The performance evaluation of the decision maker depends on the change of cache hit rate. If the cache hit rate at the next moment is higher than the current moment, it means that the decision maker's decision has received positive feedback. The probability e of the final action of the decision will be updated according to the cache hit rate brought by the decision result. When the cumulative number of user requests received by the edge node reaches the threshold N, the edge node will trigger the cache prediction process. The decision maker will record the cache hit rate at the current moment t as h t After completing the entire cache prediction process, the next departure time is t+1, and the hit rate at this time is h t+1 The specific update rule is shown in Formula 12.

[0075] e k+1 =e k +ηsign(h t+1 -h t ) (Formula 12)

[0076] where e k is the current probability value, η is the learning rate, which determines the step size of each update. t+1 -h t ) is the sign function, which will be calculated based on the hit rate h at the next moment. t+1 and the current hit rate h t This update rule allows the decision maker to dynamically adjust the value of e based on the rewards brought by the actual decision action, thereby gradually optimizing the subsequent decision process.

[0077] Step 3: The cloud server uses the original model to verify the approximate model results;

[0078] The cloud server will receive a content request or verification request from the edge node. If the request received is a content request, it will simply return the content requested by the edge node; if the request received is a verification request, it is necessary to verify the prediction result of the edge node and return the data content corresponding to the verification result of the cloud server.

[0079] In the previous process, the original input sequence X and the prediction sequence S of the approximate model have been obtained, where X = [x1, x2, x3, ..., x n ],S=[s1,s2,s3,...,s t For each step t, it is necessary to verify the token s generated by the approximate model. t Is it reasonable? At each step, the original model M is used l To calculate the probability distribution of the same position, and select the word with the highest probability, as shown in Formula 13.

[0080] P l (y t |X,S ＜t )=softmax(M l X,S ＜t ) (Formula 13)

[0081] Among them, P l represents the probability distribution of the lth strategy or model branch, y t represents the output label or action at time step t, X represents the input data, and S ＜t Represents the historical state sequence up to time step t-1, that is, the past context information. This gives a probability distribution calculated by the original model, and then the word y with the highest probability can be selected based on the entire probability distribution. t , as shown in Formula 14.

[0082] y t =argmax y P l (y t =y|X,S ＜t ) (Formula 14) This gives the validation output y of the original model t , then we need to compare the prediction results y of the original model t and the prediction results of the edge node approximation model s tAre they the same? If they are, the approximate model's prediction results are accepted. If they are different, the approximate model's prediction results are discarded and the original model's prediction results are used for update. The cloud server then transmits the data corresponding to the final verification result back to the edge node that issued the request.

[0083] Example 1:

[0084] In order to test the performance of the method of the present invention, the UPCS algorithm and the RecFormer algorithm were used. Figure 2 The topology shown is compared with the CECCP, a vehicle network cache prediction method based on cloud-edge collaboration proposed in this invention. The CECCP structure is as follows: Figure 1 As shown in the figure, relevant data such as cache hit rate, Recall, NDCG and latency are also counted.

[0085] exist Figure 3 and Figure 4 As the cache capacity increases, the cache hit rates of all algorithms increase. This is because the cache has sufficient capacity to ensure hits for incoming requests. However, as the cache capacity increases, the hit rates of CECCP and UPCS become increasingly similar. This is because the impact of the algorithms on the cache hit rate decreases when the cache size reaches a certain level. As the size of the user request sequence increases, the hit rates of all algorithms decrease. This is because, given a fixed cache size, a larger number of user requests results in fewer cache hits and a lower hit rate.

[0086] exist Figure 5 In the relevant evaluation metrics, CECCP slightly outperformed the other two models and was on par with the RecFormer model in ndcg@10. After the Top-k metric was increased from 10 to 50, the performance of all models improved. This is because this change increases the size of the candidate set, giving the model more opportunities to predict content of interest to users and reducing the likelihood of omissions.

[0087] exist Figure 6 and Figure 7In the study, user-side latency was compared between an IoV system that deployed the CECCP model and an IoV system that did not deploy an edge computing architecture. The edge computing architecture using the CECCP model significantly improved data transmission speed compared to the architecture without the model. As the cache size increases, the cache hit rate increases, so edge nodes no longer need to send requests to the cloud. The latency is now only the transmission delay between the terminal device and the edge node, resulting in a continuous decrease in average latency. As the number of user requests increases, the cache hit rate decreases, forcing the edge node to request content from the cloud server. The latency is the sum of the latency from the terminal device to the edge node and the latency from the edge node to the cloud server, resulting in a continuous increase in average latency.

Claims

1. A method for predicting Internet of Vehicles cache based on cloud-edge collaboration, characterized in that: Here are the steps: 1) The edge node collects the request sequence sent by the terminal and uses the approximate model to make predictions; 2) The edge node uses the decision maker to evaluate its own prediction results. If the decision maker gives a non-verification decision, the edge node will send a content request to the cloud server. If the decision maker gives a verification request, the edge node will send a verification request to the cloud server. 3) If the cloud server receives a content request, it will send the corresponding content in the request back to the edge node. If it receives a verification request, it will use the original model to verify the approximate model result. After obtaining the verification result, it will send the content corresponding to the verification result back to the edge node.

2. The method for predicting Internet of Vehicles cache based on cloud-edge collaboration according to claim 1, characterized in that: In the above 1), the specific method is: The edge node receives interest packets from different terminals and collects the user's request sequence in chronological order, and continuously combines the names of the interest packets into a request sequence; The request sequence is word-embedded and position-embedded. The word embedding is achieved by maintaining a trainable matrix To complete the vocabulary mapping, where V is the size of the vocabulary and d is the dimension of the word embedding; each word in the request sequence is indexed in the embedding matrix to find the corresponding vector, and the entire request sequence is combined into a matrix; Position embedding is performed on this matrix. Position embedding generates the code by alternating sine and cosine functions and obtains the matrix X. The formulas are shown in Formula 1 and Formula 2, where PE represents position encoding, pos is the position in the sequence, i is the dimension, and d model Represents the dimension of word embedding; Convert X into query matrix Q, key-value matrix K and value matrix V through linear transformation. Obtain query matrix Q through formula 3, key-value matrix K through formula 4, and value matrix V through formula 5. Q=XW Q (Formula 3)K=XW K (Formula 4)V=XW V (Formula 5) Where, And they are all learnable parameter matrices. These matrices enable the model to learn to adjust the overall focus according to different contexts. After obtaining the Q, K, and V matrices, the attention score matrix is calculated. The attention score matrix R is calculated by scaling the dot product attention. The calculation process is shown in Formula 6. Among them To prevent data instability and gradient disappearance, the value of d is large, and the overall value is not scaled and the dot product is performed, which will result in a larger value, affecting the subsequent Softmax function; After obtaining the attention score matrix R, apply Softmax normalization, apply the Softmax function to each row of S, and obtain the attention weight matrix A, as shown in Formula 7; Softmax normalization can convert the attention score into a probability distribution, ensuring that the sum of the elements in each row of the matrix is 1, allowing the model to focus on the important parts of the whole and reduce attention to unimportant information; In order to evaluate the global importance of each token in the sequence, it is necessary to sum the columns of the attention weight matrix A and generate a row vector s, as shown in Formula 8; The attention values of other positions in the self-attention mechanism to the corresponding position are accumulated to obtain the row vector s. The higher the accumulated score, the more important the token at the corresponding position is in the entire sequence. Finally, the sorting output is performed. The original tokens are sorted based on the vector s to obtain the prediction result S of the approximate model, as well as the probability value P = s corresponding to each token.

3. The method for predicting Internet of Vehicles cache based on cloud-edge collaboration according to claim 1, characterized in that: In the above 2), the specific method is: Map the probability value P to the latent space The specific process is shown in Formula 9. Where μ is the mean vector, representing the center position in the latent space, σ is the standard deviation vector, controlling the width of the latent space distribution, and ∈ is the vector from the standard normal distribution. The noise vector sampled in is used to introduce randomness and ensure the diversity of the generated latent variables; The decoder generates a decision action a from the latent space t ∈{NoVerify,Verify}, as shown in Formula 10, a t =Decoder(z) (Formula 10) Decision action a t It is one of the two actions NoVerify or Verify. NoVerify means that the prediction result of the approximate model is used without verification by the cloud server, and Verify means that the prediction result of the approximate model needs to be verified by the original model in the cloud server. In generating decision action a t It will not be executed directly afterwards, but will be executed based on the decision of e-greedy, and the decision action a will be adopted with probability e t , randomly selects one of the two actions, Verify and NoVerify, with a probability of 1-e, thereby balancing the exploration and exploitation in the multi-armed bandit model and preventing the model from converging prematurely, as shown in Formula 11; The performance evaluation of the decision maker depends on the change in the cache hit rate. If the cache hit rate at the next moment is higher than the current moment, it means that the decision maker's decision has received positive feedback. The probability e of the final action will be updated based on the cache hit rate brought by the decision result. When the cumulative number of user requests received by the edge node reaches the threshold N, the edge node will trigger the cache prediction process, and the decision maker will record the cache hit rate at the current time t as h t After completing the entire cache prediction process, the next departure time is t+1, and the hit rate at this time is h t+1 , the update rule is shown in formula 12; e k+1 =e k +ηsign(h t+1 -h t ) (Formula 12) where e k is the current probability value, η is the learning rate, which determines the step size of each update, sign(h t+1 -h t ) is the sign function, according to the hit rate h at the next moment t+1 and the current hit rate h t The size of determines whether a positive or negative value is returned; this update rule enables the decision maker to dynamically adjust the value of e based on the rewards brought by the actual decision action, and gradually optimize the subsequent decision process.

4. The method for predicting Internet of Vehicles cache based on cloud-edge collaboration according to claim 1, characterized in that: In the above 3), the specific method is: The cloud server will receive a content request or verification request from the edge node. If the request is a content request, it will return the content requested by the edge node. If the request is a verification request, it will verify the prediction result of the edge node and return the data content corresponding to the verification result of the cloud server. In the previous process, the original input sequence X and the prediction sequence S of the approximate model have been obtained, where X = [x1, x2, x3, ..., x n ],S=[s1,s2,s3,...,s t ], for each step t, verify the token generated by the approximate model, i.e. s t Is it reasonable to use the original model M at each step? l To calculate the probability distribution of the same position, and select the word with the highest probability, as shown in Formula 13; P l (y t |X,S ＜t )=softmax(M l X,S ＜t ) (Formula 13) Among them, P l represents the probability distribution of the lth strategy or model branch, y t represents the output label or action at time step t, X represents the input data, and S ＜t Represents the historical state sequence up to time step t-1, that is, the past context information; thus, a probability distribution calculated by the original model is obtained, and the word y with the highest probability is selected according to the entire probability distribution t , as shown in formula 14; y t = argmax y P l (y t = y|X,S ＜t ) (Formula 14) Finally, the verification output y of the original model is obtained t , and then compare the prediction results y of the original model t and the prediction results of the edge node approximation model s t Are they the same? If they are the same, the prediction result of the approximate model is received. If they are different, the prediction result of the approximate model is abandoned and the prediction result of the original model is used for update. The cloud server returns the data content corresponding to the final verification result to the edge node that issued the request.

Images