Edge computing multi-quality of service prediction method, system, device and storage medium

Abstract

The present application relates to the field of electric communication, specifically to edge computing multi-service quality prediction method, system, device and storage medium, the method of the present application, first, the user request information, network environment information and service running information which influence the service quality of edge server are fused to obtain edge server vector, and based on multi-head attention processing, residual connection and normalization processing, the edge server vector is extracted to obtain the edge server feature vector; then, the edge server feature vector is processed by multi-layer time sequence feature analysis, the relationship between different scene factor features is obtained, and the nonlinear expression ability of different scene factor features is enhanced, to obtain the edge server feature correlation vector; finally, the edge server feature correlation vector is processed by multi-task quality prediction to obtain several service quality prediction values, realizing high accuracy and high reliability of edge server prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of electric communication technology, in particular to an edge computing multi-service quality prediction method, system, device and storage medium. BACKGROUND

[0002] With the rapid development of mobile edge computing technology, a large number of edge computing services from different fields and with diversified functions have emerged on the network, greatly improving the experience and convenience of users. In actual application, users often have specific quality requirements for the services they use, which are usually reflected in different dimensions of service quality (QoS), such as expecting the response time to be no more than 60 minutes, the service availability to be higher than 85%, and the service reliability to be more than 90%. In order to ensure that the user's demand is met, the service system needs to select or integrate the optimal edge computing service that meets the user's multi-dimensional service quality requirements from a large number of service options, so in the increasingly complex network and service environment, accurate prediction of service quality has become a crucial task to improve user experience.

[0003] The existing technologies include linear regression method, logistic regression method, exponential smoothing method, support vector machine, random forest, Bayesian network and artificial neural network, which often analyze past service behaviors to predict future service quality. However, these methods are difficult to effectively capture complex time series patterns and nonlinear relationships, and have low nonlinear expression ability when dealing with service quality problems in the current mobile edge computing environment where the change of each situational factor will cause fluctuations and influences of service quality, which cannot fully explore the complex relationship between service quality factors and parameters, resulting in a decline in service quality prediction performance and a reduction in user experience. SUMMARY

[0004] The purpose of the present application is to provide an edge computing multi-service quality prediction method, system, device and storage medium.

[0005] The technical solution of the present application is as follows:

[0006] An edge computing multi-service quality prediction method, comprising:

[0007] S1, obtaining user request information, network environment information and service running information of an edge server, respectively embedding and processing, splicing to obtain an edge server vector; the edge server vector is processed by multi-head attention, residual connection and normalization to obtain an edge server feature vector; the user request information includes user connection quantity and random task data size; the network environment information includes network bandwidth, network delay and operation speed; the service running information includes total amount of tasks to be processed, task load and queue task quantity;

[0008] S2, the edge server feature vector is processed by multi-layer time sequence feature analysis to obtain an edge server feature correlation vector; in the process of multi-layer time sequence feature analysis: the output of the current layer time sequence feature analysis processing is fused with the edge server feature vector to obtain a current layer initial fusion vector; the current layer initial fusion vector is processed by nonlinear expression ability enhancement to obtain a current layer fusion feature vector as the input of the next layer time sequence feature analysis processing; the nonlinear expression ability enhancement processing is specifically: the current layer initial fusion vector is multiplied by the current layer weight to obtain a current layer nonlinear processing vector; the current layer nonlinear processing vector is mapped to obtain a current layer initial fusion mapping value; it is judged whether the current layer initial fusion mapping value is greater than the mapping value threshold; if greater, the current layer nonlinear processing vector is processed by first nonlinear processing based on the edge server feature vector mean and the current layer initial fusion vector mean to obtain a first nonlinear enhancement vector as the current layer fusion feature vector; if not greater, the current layer nonlinear processing vector is processed by second nonlinear processing based on the edge server feature vector variance and the current layer initial fusion vector variance to obtain a second nonlinear enhancement vector as the current layer fusion feature vector;

[0009] S3, the edge server feature correlation vector is processed by multi-task quality prediction to obtain a plurality of service quality prediction values.

[0010] In S2, the first nonlinear processing is realized by the following formula: y i,1 =(ω1(μ-μ0)+γ)X i , and the second nonlinear processing is realized by the following formula: y i,1 , y i,2 are the i-th layer first nonlinear enhancement vector and the i-th layer second nonlinear enhancement vector respectively, ω1 and ω2 are the first learning rate and the second learning rate respectively, μ and μ0 are the edge server feature vector mean and the current layer initial fusion vector mean respectively, σ and σ0 are the edge server feature vector variance and the current layer initial fusion vector variance respectively, γ and ε are the first slope value and the second slope value respectively, γ>ε, X i is the i-th layer initial fusion vector.

[0011] The operation of the multi-layer time sequence feature analysis in S2 can be realized by a time sequence analysis model composed of a plurality of time sequence feature analysis modules; in the time sequence analysis model, the plurality of time sequence feature analysis modules are distributed in a rectangular array, the data propagation direction of the same column time sequence feature analysis module is from bottom to top, and the data propagation direction of the same row time sequence feature analysis module is from left to right.

[0012] In the time sequence feature analysis module, the operation of the time sequence feature analysis processing is specifically: after being processed by the first activation function, the hyperbolic tangent function and the forgetting processing respectively, the input is fused to obtain a memory feature; the input is processed by the second activation function to obtain a second nonlinear feature; the memory feature is processed by the hyperbolic tangent function and fused with the second nonlinear feature to obtain an output.

[0013] The total amount of tasks to be processed in S1 is the cumulative amount of all task data of the edge server in the first neighborhood time period; the task load is the sum of the number of tasks being run and waiting to run of the edge server in the second neighborhood time period, and the second neighborhood time period is within the first neighborhood time period.

[0014] In S1, the average geographic location and the average use time period of all users connected to the edge server are also embedded and spliced with the edge server vector to obtain an edge server optimization vector, which is used for multi-head attention processing, residual connection and normalization processing.

[0015] In S3, each task quality prediction processing is realized by a multi-layer fully connected neural network layer.

[0016] An edge computing multi-service quality prediction system for implementing the above-mentioned edge computing multi-service quality prediction method, comprising:

[0017] An edge server feature vector generation module is used to obtain user request information, network environment information and service running information of the edge server, which are respectively embedded and spliced to obtain an edge server vector; the edge server vector is processed by multi-head attention, residual connection and normalization to obtain an edge server feature vector; the user request information includes the number of user connections and the size of random task data; the network environment information includes network bandwidth, network delay and operation speed; the service running information includes the total amount of tasks to be processed, the task load and the queue task amount;

[0018] The edge server feature correlation vector generation module is configured to perform multi-layer time sequence feature analysis processing on the edge server feature vector to obtain an edge server feature correlation vector.

[0019] The service quality prediction value generation module is configured to perform multi-task quality prediction processing on the edge server feature correlation vector to obtain a plurality of service quality prediction values.

[0020] An edge computing multi-service quality prediction device includes a processor and a memory, wherein the processor implements the edge computing multi-service quality prediction method described above when executing a computer program stored in the memory.

[0021] A computer-readable storage medium is configured to store a computer program, wherein the computer program is executed by a processor to implement the edge computing multi-service quality prediction method described above.

[0022] The edge computing multi-service quality prediction method provided by the application has the following beneficial effects:

[0023] The edge computing multi-service quality prediction method provided by the application has the following beneficial effects: The edge computing multi-service quality prediction method provided by the application has the following beneficial effects: Attached Figure Description

[0024] The solutions and advantages of this application will become clear to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention.

[0025] In the attached diagram:

[0026] Figure 1 This is a flowchart illustrating the prediction method in this embodiment.

[0027] Figure 2 As shown in the example, this is a comparison chart of the service reliability accuracy of the five prediction methods;

[0028] Figure 3 This is a comparison chart of the service time accuracy of five prediction methods in the embodiment;

[0029] Figure 4 As shown in the example, this is a grouped comparison chart of the service reliability accuracy of the five prediction methods;

[0030] Figure 5 The following is a comparison chart of the service time accuracy of the five prediction methods in the example. Detailed Implementation

[0031] Exemplary embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings.

[0032] This embodiment provides a method for predicting the quality of service (QS) of multiple services in edge computing. (See also...) Figure 1 ,include:

[0033] S1. Obtain user request information, network environment information, and service operation information from the edge server. After embedding processing, these information are concatenated to obtain the edge server vector. The edge server vector is then processed through multi-head attention, residual connection, and normalization to obtain the edge server feature vector. User request information includes the number of user connections and the size of random task data. Network environment information includes network bandwidth, network latency, and processing speed. Service operation information includes the total number of tasks to be processed, task load, and queue task volume.

[0034] S2, the edge server feature vector is processed by multi-layer time sequence feature analysis to obtain an edge server feature correlation vector; in the process of multi-layer time sequence feature analysis: the output of the current layer time sequence feature analysis is fused with the edge server feature vector to obtain a current layer initial fusion vector; the current layer initial fusion vector is processed by nonlinear expression ability enhancement to obtain a current layer fusion feature vector as the input of the next layer time sequence feature analysis; the nonlinear expression ability enhancement processing is specifically: the current layer initial fusion vector is multiplied by the current layer weight to obtain a current layer nonlinear processing vector; the current layer nonlinear processing vector is mapped to obtain a current layer initial fusion mapping value; it is judged whether the current layer initial fusion mapping value is greater than the mapping value threshold; if greater, the current layer nonlinear processing vector is processed by first nonlinear processing based on the edge server feature vector mean and the current layer initial fusion vector mean to obtain a first nonlinear enhancement vector as the current layer fusion feature vector; if not greater, the current layer nonlinear processing vector is processed by second nonlinear processing based on the edge server feature vector variance and the current layer initial fusion vector variance to obtain a second nonlinear enhancement vector as the current layer fusion feature vector;

[0035] S3, the edge server feature correlation vector is processed by multi-task quality prediction to obtain several service quality prediction values.

[0036] S1, the user request information, network environment information and service running information of the edge server are obtained, embedded and spliced to obtain an edge server vector; the edge server vector is processed by multi-head attention, residual connection and normalization to obtain an edge server feature vector.

[0037] The user request information, network environment information and service running information affecting the service quality of the edge server are fused to obtain an edge server vector, and the edge server vector is extracted based on multi-head attention, residual connection and normalization to obtain an edge server feature vector.

[0038] First, the scenario factors affecting the edge server computing effect and the edge server service quality are obtained, including user request information reflecting the edge server to be processed tasks, network environment information reflecting the interaction efficiency between users and edge servers and service running information reflecting the current edge server usage.

[0039] The user request information includes the number of user connections and the size of random task data. The number of user connections is the number of devices using the edge server at the same time, and the size of random task data is the size of any task data in all task data of all users connected to the edge server, indicating the amount of data required for task transmission.

[0040] The network environment information includes network bandwidth, network delay and computing speed. The network bandwidth is the transmission rate of a wireless local area network (WLAN), the network delay is the time required for data to be transmitted from a sending end to a receiving end, and the computing speed is a performance measurement index of the edge server, usually expressed in millions of instructions per second (MIPS).

[0041] The service running information includes total amount of tasks to be processed, task load and queue task amount. The total amount of tasks to be processed is the total amount of tasks to be processed, indicating the cumulative amount of all task data of the edge server in the first neighborhood time period; the task load is the sum of the number of tasks being run and waiting to be run by the edge server in the second neighborhood time period, which is within the first neighborhood time period; and the queue task amount is the amount of tasks to be processed in a queuing state, indicating the amount of tasks that have not yet started processing but are already in the queue.

[0042] Then, the user request information, the network environment information and the service running information, i.e., the number of user connections, the random task data size, the network bandwidth, the network delay, the computing speed, the total amount of tasks to be processed, the task load and the queue task amount, are respectively embedded and processed, and then spliced to obtain an edge server vector that can reflect both the processing performance characteristics of the edge server and the environment characteristics of the edge server.

[0043] Finally, in order to focus on the characteristics of different situational factors, the edge server vector is first subjected to multi-head attention processing to focus on different aspects of the input, then subjected to residual connection processing to focus on the differences in characteristics before and after multi-head attention processing, and finally subjected to normalization processing to obtain an edge server feature vector.

[0044] In order to further improve the richness of the edge server vector and improve the accuracy of subsequent prediction results, the average geographic location and the average use time period of all users connected to the edge server are also embedded and processed, and then spliced with the edge server vector to obtain an edge server optimization vector, which is used for multi-head attention processing, residual connection and normalization processing.

[0045] S2, the edge server feature vector is subjected to multi-layer time series feature analysis processing to obtain an edge server feature association vector.

[0046] The edge server feature vector is subjected to multi-layer time series feature analysis processing to obtain an edge server feature association vector.

[0047] The operation of the multi-layer time sequence feature analysis processing can be implemented by a time sequence analysis model composed of a plurality of time sequence feature analysis modules, so that the long-term time dependence of the data in the edge server feature vector can be effectively captured, the correlation of the data at different time steps can be understood, and the running state of the edge server can be better analyzed and predicted.

[0048] In the time sequence analysis model, the plurality of time sequence feature analysis modules are arranged in a rectangular array, the data propagation direction of the time sequence feature analysis modules in the same column is from bottom to top, and the data propagation direction of the time sequence feature analysis modules in the same row is from left to right. For example, if the time sequence analysis model is composed of 9 time sequence feature analysis modules, the 9 time sequence feature analysis modules are arranged in a 3*3 array, the time sequence feature analysis module at the lower left corner is a module for performing first-layer time sequence feature analysis processing, and is used for processing the edge server feature vector, and the time sequence feature analysis module at the upper right corner is a module for performing last-layer time sequence feature analysis processing, and is used for outputting the obtained edge server feature correlation vector.

[0049] In addition, in the time sequence analysis model, the plurality of time sequence feature analysis modules can also be connected in series, so that the feature analysis efficiency can be improved.

[0050] In the time sequence feature analysis processing, the operation is specifically as follows: after being processed by the first activation function, the hyperbolic tangent function and the forgetting processing respectively, the inputs are fused to obtain a memory feature; the input is processed by the second activation function to obtain a second nonlinear feature; and the memory feature is processed by the hyperbolic tangent function and fused with the second nonlinear feature (which can be realized by vector multiplication) to obtain an output.

[0051] The operation of the current-layer time sequence feature analysis processing can be implemented by the following formula:

[0052] f t =σ(W f ·[h t-1 ,x t ]+a f ),

[0053] i t =σ(W i ·[h t-1 ,x t ]+a i ),

[0054] c t =tanh(W c ·[h t-1 ,x t ]+a c ),

[0055] C t =f t ·Ct-1 +i t ·c t ,

[0056] O t =σ(W O ·[h t-1 ,x t ]+a O ),

[0057] h t =O t ·tanh(C t ),

[0058] h t-1 The output at time t-1, x t Let W be the initial input at time t, and let W be the fused feature vector from the previous layer. f For forgetting weights, a f For the forgetting compensation parameter, f t Let be the forgotten output at time t, σ be the sigmoid function, and i t As the first nonlinear characteristic, W i Let a be the weight of the first activation function. i c is the compensation parameter for the first activation function. t The first tangent feature (obtained by hyperbolic tangent processing of the input), W c As the first tangent weight, a c C is the first tangent compensation parameter. t For the memory features at time t, C t-1 For the memory features at time t-1, O t As the second nonlinear characteristic, W O a is the weight of the second activation function. O h is the compensation parameter for the second activation function. t This is the output at time t.

[0059] The operation of time series feature analysis can also be as follows: the input data is standardized to obtain a standard feature vector; the covariance matrix of the standard feature vector is obtained, which reflects the degree of linear correlation between features in the standard feature vector; the covariance matrix is ​​decomposed to obtain eigenvalues; based on the eigenvalues, the principal component vector of the input is obtained, which contains the main information of the original data, while reducing the dimensionality of the data, and each principal component is independent of each other, eliminating the linear correlation between the original features; the transpose of the standard feature vector is multiplied with the principal component vector, and then added to the input to obtain the output.

[0060] In the process of the multi-layer time sequence feature analysis processing, the output of the current layer time sequence feature analysis processing and the edge server feature vector are fused (which can be achieved by transverse splicing on a one-dimensional vector), not only the key information of the original input is retained, but also the loss in the layer-by-layer information transmission is effectively made up, ensuring that important features are always retained and strengthened, obtaining the current layer initial fusion vector; the current layer initial fusion vector is subjected to nonlinear expression capability enhancement processing, improving the capture ability of complex features and enhancing the nonlinear expression capability of features, obtaining the current layer fusion feature vector as the input of the next layer time sequence feature analysis processing.

[0061] The nonlinear expression capability enhancement processing is specifically as follows: the current layer initial fusion vector is multiplied by the current layer weight to obtain a current layer to-be-nonlinearly-processed vector; the current layer to-be-nonlinearly-processed vector is mapped to a specific numerical value to obtain a current layer initial fusion mapping value; it is determined whether the current layer initial fusion mapping value is greater than a mapping value threshold; if yes, the current layer to-be-nonlinearly-processed vector is subjected to first nonlinear processing based on the edge server feature vector mean and the current layer initial fusion vector mean to obtain a first nonlinear enhancement vector as the current layer fusion feature vector; if no, the current layer to-be-nonlinearly-processed vector is subjected to second nonlinear processing based on the edge server feature vector variance and the current layer initial fusion vector variance to obtain a second nonlinear enhancement vector as the current layer fusion feature vector.

[0062] The first nonlinear processing is achieved by the following formula: y i,1 =(ω1(μ-μ0)+γ)X i The second nonlinear processing is achieved by the following formula: y i,1 , y i,2 are the i-th layer first nonlinear enhancement vector and the i-th layer second nonlinear enhancement vector, respectively, ω1 and ω2 are the first learning rate and the second learning rate, respectively, μ and μ0 are the edge server feature vector mean and the current layer initial fusion vector mean, respectively, σ and σ0 are the edge server feature vector variance and the current layer initial fusion vector variance, respectively, γ and ε are the first slope value and the second slope value (greater than 0 but less than 1), respectively, and γ>ε, X i is the i-th layer to-be-nonlinearly-processed vector.

[0063] S3, the edge server feature association vector is subjected to multi-task quality prediction processing to obtain a plurality of service quality prediction values.

[0064] In the multi-task quality prediction processing, the weight and bias value of different task quality predictions are different, and each task quality prediction processing is implemented through a plurality of full connection neural networks, and each full connection neural network is responsible for fine feature extraction of input features to meet the requirements of a specific task.

[0065] The operation of each full connection neural network processing can be implemented through the following formula:

[0066] Z p =σ p (β p Z p-1 +c p ),

[0067] Z p is the output of the pth full connection neural network, σ p is an activation function of the pth full connection neural network, β p is a weight matrix of the pth full connection neural network, Z p-1 is the output of the pth full connection neural network, c p is a bias term of the pth full connection neural network.

[0068] The multi-task quality prediction processing includes running time prediction and service recommendation reliability prediction.

[0069] The embodiment also provides an edge computing multi-service quality prediction system for implementing the edge computing multi-service quality prediction method.

[0070] The edge server feature vector generation module is configured to obtain user request information, network environment information and service running information of the edge server, perform embedding processing on the user request information, the network environment information and the service running information respectively, splice the embedding processing results to obtain an edge server vector, perform multi-head attention processing, residual connection and normalization processing on the edge server vector to obtain an edge server feature vector, and the user request information includes user connection quantity and random task data size, the network environment information includes network bandwidth, network delay and operation speed, and the service running information includes total quantity of tasks to be processed, task load and queue task quantity.

[0071] The edge server feature correlation vector generation module is configured to obtain an edge server feature correlation vector by performing multi-layer time sequence feature analysis processing on the edge server feature vector. During the multi-layer time sequence feature analysis processing, the output of the current layer time sequence feature analysis processing is fused with the edge server feature vector to obtain a current layer initial fusion vector. The current layer initial fusion vector is subjected to nonlinear expression capability enhancement processing to obtain a current layer fusion feature vector as the input of the next layer time sequence feature analysis processing. The nonlinear expression capability enhancement processing is specifically performed as follows: the current layer initial fusion vector is multiplied by the current layer weight to obtain a current layer nonlinear processing vector after weighting processing. The current layer nonlinear processing vector is subjected to mapping processing to obtain a current layer initial fusion mapping value. It is determined whether the current layer initial fusion mapping value is greater than a mapping value threshold. If yes, the first nonlinear processing is performed on the current layer nonlinear processing vector based on the edge server feature vector mean and the current layer initial fusion vector mean to obtain a first nonlinear enhancement vector as the current layer fusion feature vector. If no, the second nonlinear processing is performed on the current layer nonlinear processing vector based on the edge server feature vector variance and the current layer initial fusion vector variance to obtain a second nonlinear enhancement vector as the current layer fusion feature vector.

[0072] The service quality prediction value generation module is configured to obtain a plurality of service quality prediction values by performing multi-task quality prediction processing on the edge server feature correlation vector.

[0073] When the system is implemented based on the above method, the training sample set (formed by a plurality of scenario factors of a plurality of edge servers and corresponding multi-task service quality label values) is obtained, and the system is trained by using the training data set. When the loss value obtained by performing weighted summation on the difference between the prediction value and the label value of each sample is less than the loss threshold, the training is ended.

[0074] In addition, to improve the stability of system calculation, the first-order moment deviation and the second-order moment deviation of the edge server feature correlation vector are corrected in each training process, and the parameter gradient reverse processing is performed to obtain the corrected training parameter, which is used to update the bias term of the full connection neural network in the multi-layer time sequence feature analysis processing.

[0075] The parameter gradient reverse processing is implemented by the following formula:

[0076]

[0077] β t is the corrected training parameter at time t, β t-1 is the corrected training parameter at time t-1, η is the gradient reverse processing coefficient, s t is the corrected first-order moment deviation, r t is the corrected second-order moment deviation, and ε is the compensation value.

[0078] The embodiment also provides an edge computing multi-service quality prediction device, comprising a processor and a memory, wherein the processor implements the edge computing multi-service quality prediction method described above when executing the computer program stored in the memory.

[0079] The embodiment also provides a computer readable storage medium for storing a computer program, wherein the computer program is executed by a processor to implement the edge computing multi-service quality prediction method described above.

[0080] The edge computing multi-service quality prediction method provided by the embodiment first fuses user request information, network environment information and service running information that affect the service quality of an edge server to obtain an edge server vector, and performs feature extraction on the edge server vector based on multi-head attention processing, residual connection and normalization processing to obtain an edge server feature vector; then performs multi-layer time sequence feature analysis processing on the edge server feature vector to obtain the relationship between different scenario factor features and enhance the nonlinear expression ability of different scenario factor features, and obtains an edge server feature correlation vector; finally, the edge server feature correlation vector is subjected to multi-task quality prediction processing to obtain a plurality of service quality prediction values, thereby achieving high accuracy and high reliability of edge server prediction.

[0081] To verify the effect of the method of the embodiment, the following experiment is performed.

[0082] Experimental data. In the experiment, an edge computing simulation platform EdgeCloudSim is used to construct a scenario in which elderly people in a simulated nursing home environment select a specific service S. During the experiment, the scenario information when the elderly people select the service S each time and the multiple service quality (hereinafter referred to as QoS) indicators of the service S corresponding thereto are recorded, and the service quality indicators include running time and accuracy. A total of 200,000 service selection records are collected in the experiment, and each record describes the specific situation of the elderly people when selecting the service S and the service quality performance embodied by the selection. These data not only include the selection behavior of the user, but also cover various QoS parameters in the service execution process, providing a rich information basis for subsequent analysis.

[0083] Evaluation index. In order to evaluate the effectiveness of the method of the embodiment, the experiment uses the mean absolute error (MAE), the mean square error (MSE) and the accuracy (Acc) to measure the deviation between the predicted value and the true value of QoS. The smaller the values of MAE and MSE, the higher the prediction accuracy, and the larger the value of Acc, the higher the prediction accuracy.

[0084] Experimental parameters. The experimental parameters are set as shown in Table 1. The operating system is Windows 11, the graphics card is RTX3070Ti, the programming language is python3.6, and the deep learning framework is Tensorflow.

[0085] Table 1 Experimental parameter settings

[0086] Parameter Value epoch 200 batch_size 256 optimizer Adam Loss function mean_squared_error learning rate 0.001

[0087] Comparison with prior art. In the experiment, the Cross-stitch method, the U-MTL method, the MTDNN method and the IMTDNN method were selected to compare the performance with the method of the embodiment, and to further verify the effectiveness of the method of the embodiment.

[0088] MTDNN method and IMTDNN method, and the method of the embodiment, and to further verify the effectiveness of the method of the embodiment.

[0089] Experimental results 1. In order to fully verify the effectiveness of the method of the embodiment and avoid the influence of data difference on prediction fluctuation, the experiment sets different proportions of test sets (hereinafter referred to as TDS, the number is 10%, 20% and 30%) to more comprehensively evaluate the performance of the method under different data scales, so as to ensure the stability and reliability of the results. In each experiment, three key indicators of MAE, MSE and Acc are calculated. Finally, by taking the average of 10 experimental results, more stable and reliable evaluation results are obtained, as shown in Tables 2 and 3.

[0090] Table 2 Performance summary of different methods on reliable prediction under different number of test sets

[0091]

[0092] Table 3 Performance summary of different methods on running time under different number of test sets

[0093]

[0094] From the data results of Tables 2 and 3, it can be seen that the method of the embodiment is superior to the other four existing methods in predicting two QoSs (reliability and running time) in evaluation indicators Acc, MAE and MSE. In particular, on the test data set with TDS=30%, the method of the embodiment is superior to the other existing methods in predicting two QoSs in evaluation indicator Acc, with a gain of 0.9%; it is also superior to the other existing methods in indicator MAE, with a gain of 4%; and it is also superior to the other existing methods in MSE, with a gain of 13.3%. From the above experimental results, it can be concluded that the method of the embodiment not only can efficiently process shared features and task-specific features, but also can better reflect the QoS changes under different scenarios, enhance the robustness of feature representation, reduce information loss, and finally achieve more accurate prediction effect.

[0095] Experimental results 2. To verify the convergence of the method of the embodiment, 70% of the experimental data is taken as the training set, and the remaining 30% as the test set, to evaluate the performance of the five different methods. In this way, not only can the effectiveness of model training be ensured, but the generalization ability on unseen data can also be accurately measured. The experimental results are shown in Figure 2 and Figure 3 , where Figure 2 shows the prediction results of service reliability, Figure 3 shows the prediction results of service running time. In Figure 2 and Figure 3 , the vertical coordinate represents the prediction accuracy, and the horizontal coordinate represents the number of iterations of the algorithm. From Figure 2 and Figure 3 , it can be seen that with the increase of the number of iterations, the service accuracy and service time prediction accuracy of the five methods all show an upward trend. However, the method of the embodiment not only performs well in prediction accuracy, but also has a significantly faster convergence speed than the other four methods. The method of the embodiment can quickly improve the prediction accuracy at the early stage of iteration and maintain stable growth in subsequent iterations, which indicates that the method of the embodiment can find the optimal solution in a relatively short time, thereby improving the prediction efficiency.

[0096] Experimental results 3. To comprehensively evaluate the prediction ability stability of the method of the embodiment, 50% of the data in the test set (TDS=30%) is randomly selected as a sub-test set, and a total of 10 independent samplings are performed, thereby forming 10 different sub-test sets. Subsequently, the method of the embodiment and the other four methods are used to test the 10 sub-test sets. The experimental results are shown in Figure 4 and Figure 5 , where Figure 4 shows the prediction accuracy of the five different methods for service reliability, and Figure 5 focuses on the prediction accuracy of service running time. In the two figures, the vertical coordinate represents the prediction accuracy, and the horizontal coordinate is the serial number of the test data set. From Figure 4 and Figure 5 , it can be clearly seen that the method of the embodiment performs well in predicting QoS reliability and service time, with prediction accuracy stably around 0.89 and 0.96, respectively. This result shows that no matter how the test data changes, the method of the embodiment can maintain a high prediction accuracy, showing its excellent stability and reliability. In contrast, the prediction accuracy of the other four algorithms fluctuates greatly, especially when dealing with different sub-test sets, and the performance of these algorithms differs significantly, further highlighting the advantages of the method of the embodiment.

Claims

1. An edge computing multi-quality of service prediction method, characterized in that, The application relates to a method for predicting service quality of an edge server. The method comprises the following steps: S1, obtaining user request information, network environment information and service running information of the edge server, respectively carrying out embedded processing, splicing and obtaining an edge server vector; The edge server vector is subjected to multi-head attention processing, residual connection and normalization processing to obtain an edge server feature vector; the user request information comprises user connection quantity and random task data size; the network environment information comprises network bandwidth, network delay and operation speed; the service running information comprises total quantity of tasks to be processed, task load and queue task quantity; S2, the edge server feature vector is subjected to multi-layer time sequence feature analysis processing to obtain an edge server feature correlation vector; In the process of multi-layer time sequence feature analysis processing: the output of the current layer time sequence feature analysis processing is subjected to fusion processing with the edge server feature vector to obtain a current layer initial fusion vector; the current layer initial fusion vector is subjected to nonlinear expression capacity enhancement processing to obtain a current layer fusion feature vector as the input of the next layer time sequence feature analysis processing; The operation of the nonlinear expression capacity enhancement processing is specifically as follows: the current layer initial fusion vector is multiplied by the current layer weight to obtain a current layer nonlinear processing vector; the current layer nonlinear processing vector is subjected to mapping processing to obtain a current layer initial fusion mapping value; whether the current layer initial fusion mapping value is greater than a mapping value threshold is judged; if yes, the current layer nonlinear processing vector is subjected to first nonlinear processing based on the edge server feature vector mean value and the current layer initial fusion vector mean value to obtain a first nonlinear enhancement vector as the current layer fusion feature vector; The first nonlinear processing is achieved through the following formula: The second nonlinear processing is achieved through the following formula: , , The first i The first nonlinear enhancement vector of the layer and the second i The second nonlinear enhancement vector of the layer, , These are the first learning rate and the second learning rate, respectively. , These are the mean of the feature vectors of the edge servers and the mean of the initial fusion vector of the current layer, respectively. , These are the variances of the feature vectors of the edge servers and the variance of the initial fusion vector of the current layer, respectively. , These are the first slope value and the second slope value, respectively. , For the first i Initial fusion vector of the layer; If not, the current layer nonlinear processing vector is subjected to second nonlinear processing based on the edge server feature vector variance and the current layer initial fusion vector variance to obtain a second nonlinear enhancement vector as the current layer fusion feature vector; 2.The edge computing multi-quality of service prediction method of claim 1, wherein, S3, the edge server feature correlation vector is subjected to multi-task quality prediction processing to obtain a plurality of service quality prediction values. In the S2, the operation of the multi-layer time sequence feature analysis processing can be realized by a time sequence analysis model composed of a plurality of time sequence feature analysis modules; 3.The edge computing multi-quality of service prediction method of claim 2, wherein, In the time sequence analysis model, the plurality of time sequence feature analysis modules are arranged in a rectangular array, the data propagation direction of the time sequence feature analysis modules in the same column is from bottom to top, and the data propagation direction of the time sequence feature analysis modules in the same row is from left to right. In the time sequence feature analysis module, the operation of the time sequence feature analysis processing is specifically as follows: the input is subjected to fusion processing after being subjected to first activation function processing, hyperbolic tangent processing and forgetting processing to obtain a memory feature; the input is subjected to second activation function processing to obtain a second nonlinear feature; 4.The edge computing multi-quality of service prediction method of claim 1, wherein, the memory feature is subjected to hyperbolic tangent processing and then fused with the second nonlinear feature to obtain the output. In the S1, the total quantity of tasks to be processed is the cumulative quantity of all task data of the edge server in a first neighborhood time period; the task load is the sum of the number of tasks being run and the number of tasks waiting to be run of the edge server in a second neighborhood time period, and the second neighborhood time period is in the first neighborhood time period.

5. The edge computing multi-quality of service prediction method of claim 1, wherein, In the S1, the average geographic position and the average use time period of all users of the acquired connection edge server are also included, and after embedding processing is respectively performed, the edge server vector is spliced to obtain an edge server optimization vector, which is used to perform the operations of the multi-head attention processing, residual connection, and normalization processing.

6. The edge computing multi-quality of service prediction method of claim 1, wherein, In the S3, each task quality prediction processing is implemented through a multi-layer fully connected neural network layer.

7. An edge computing multi-quality of service prediction system for implementing the edge computing multi-quality of service prediction method of claim 1, characterized in that, The method comprises: An edge server feature vector generation module is configured to acquire user request information, network environment information, and service running information of an edge server, perform embedding processing on the information respectively, splice the information, and obtain an edge server vector; perform multi-head attention processing, residual connection, and normalization processing on the edge server vector, and obtain an edge server feature vector; the user request information comprises the number of user connections and random task data size; the network environment information comprises network bandwidth, network delay, and operation speed; and the service running information comprises the total amount of tasks to be processed, task load, and queue task amount. An edge server feature correlation vector generation module is configured to perform multi-layer time sequence feature analysis processing on the edge server feature vector, and obtain an edge server feature correlation vector; in the multi-layer time sequence feature analysis processing, the output of current layer time sequence feature analysis processing is fused with the edge server feature vector to obtain a current layer initial fusion vector; the current layer initial fusion vector is subjected to nonlinear expression capacity enhancement processing to obtain a current layer fusion feature vector as the input of next layer time sequence feature analysis processing; the nonlinear expression capacity enhancement processing is specifically performed as follows: the current layer initial fusion vector is multiplied by a current layer weight to obtain a current layer to-be-nonlinearly-processed vector through weighted processing; the current layer to-be-nonlinearly-processed vector is subjected to mapping processing to obtain a current layer initial fusion mapping value; it is determined whether the current layer initial fusion mapping value is greater than a mapping value threshold; if yes, the current layer to-be-nonlinearly-processed vector is subjected to first nonlinear processing based on the edge server feature vector mean value and the current layer initial fusion vector mean value to obtain a first nonlinear enhancement vector as the current layer fusion feature vector; if no, the current layer to-be-nonlinearly-processed vector is subjected to second nonlinear processing based on the edge server feature vector variance and the current layer initial fusion vector variance to obtain a second nonlinear enhancement vector as the current layer fusion feature vector; A service quality prediction value generation module is configured to perform multi-task quality prediction processing on the edge server feature correlation vector to obtain a plurality of service quality prediction values.

8. An edge computing multi-quality of service prediction device, characterized by, A device comprises a processor and a memory, wherein the processor implements the edge computing multi-service quality prediction method of any one of claims 1-6 when executing a computer program saved in the memory.

9. A computer-readable storage medium, characterized in that, A computer program is stored, wherein the computer program is executed by a processor to implement the edge computing multi-service quality prediction method of any one of claims 1-6.

Images