Intelligent resource management method and system for wireless power supply mobile edge computing network

Abstract

The invention discloses an intelligent resource management method and system for a wireless power supply mobile edge computing network. According to the intelligent resource management method and system, an attention mechanism enhanced multi-task neural model is constructed, the multi-task and multi-task combined learning training is adopted to optimize the multi-task neural model, wireless channel gain information in the wireless power supply mobile edge computing network is input, a prediction result of a computing decision and resource allocation is obtained, and the resource allocation efficiency is improved. And therefore, the maximum network performance can be obtained. According to the method, calculation unloading decision making and resource scheduling can be efficiently achieved, it is ensured that the system keeps high performance and low delay in a rapidly-changing environment, the calculation complexity of online decision making is remarkably reduced, complex calculation is transferred to an offline training stage, and the method is suitable for edge devices with limited resources; and meanwhile, the wireless channel gain can be dynamically monitored, so that task unloading decision and resource allocation can be carried out, and the dynamic change of the equipment state can be effectively processed by utilizing a mask mechanism.

Description

Technical Field

[0001] The present invention relates to wireless communication and edge computing technologies, and provides an intelligent resource management method and system for a wireless powered mobile edge computing network. Background Art

[0002] At present, the resource management of wireless powered mobile edge computing (WPMEC) networks mainly adopts two types of methods: traditional optimization algorithms and methods based on deep reinforcement learning (DRL). Traditional optimization algorithms obtain resource allocation strategies through mathematical modeling and solving. This type of method can obtain stable management and scheduling decisions, but it is difficult to obtain the optimal solution in real time in a rapidly changing wireless channel environment. Especially in large-scale WPMEC networks, due to the NP-Hard characteristics of the problem, the computational complexity of traditional methods is too high to meet the needs of real-time system optimization. The DRL-based method is mainly used to optimize computational offloading decisions. Although the decision-making performance has been improved to a certain extent, resource allocation optimization still relies on traditional methods, and it is difficult to optimize computational offloading and resource allocation at the same time. At the same time, the DRL method requires a lot of online learning and environmental interaction to obtain effective strategies, which is difficult to achieve in resource-constrained and delay-sensitive networks. In addition, the DRL method usually requires additional storage resources to save learning data, which increases the complexity of the system.

[0003] Therefore, there is an urgent need for a new intelligent resource management method that can achieve near real-time joint optimization in WPMEC networks with low computational complexity, low resource consumption and high scalability. Summary of the invention

[0004] The purpose of the present invention is to provide an intelligent resource management method for a wireless powered mobile edge computing network, so that mobile devices can continuously perform task offloading and data processing in a low-power environment, thereby improving network performance, especially in delay-sensitive and resource-constrained scenarios.

[0005] A second object of the present invention is to provide an intelligent resource management system for a wireless powered mobile edge computing network.

[0006] The first object of the present invention is achieved by the following technical solution: The intelligent resource management method for a wireless powered mobile edge computing network comprises the following steps:

[0007] S1. Configure a data interface on a remote controller of a wireless powered mobile edge computing network that needs to make computing decisions and allocate resources, and clone the collected system status information to a server of a running data collection module to form a data set;

[0008] S2. Construct a multi-task neural network model enhanced by an attention mechanism, wherein the multi-task neural network model includes an input layer, a shared hidden layer, an attention mechanism layer, a task offloading decision subtask module, and a resource allocation strategy subtask module, wherein the input layer, the shared hidden layer, and the attention mechanism layer are connected in sequence, the task offloading decision subtask module and the resource allocation strategy subtask module are connected in parallel and are both connected to the attention mechanism layer, and the task offloading decision subtask module and the resource allocation strategy subtask module are connected to each other;

[0009] S3, using multi-task joint learning to train a multi-task neural network model to optimize the multi-task neural network model;

[0010] S4. Deploy the optimized multi-task neural network model to input the wireless channel gain information in the wireless powered mobile edge computing network to obtain the prediction results of computing decisions and resource allocation to maximize the network performance.

[0011] Preferably, the system status includes information on the number of wireless devices in the wireless powered mobile edge computing network and information on wireless channel gain between the edge server and the wireless devices.

[0012] Preferably, the data set includes a training set, a validation set and a test set.

[0013] Preferably, the task offloading subtask module comprises a TOD-fully connected layer, a Sigmoid layer, a TOD-mask layer, a thresholding layer and a task offloading decision output layer connected in sequence;

[0014] TOD-fully connected layer, used to convert the output of the attention mechanism layer into the input of the Sigmoid layer;

[0015] The Sigmoid layer is used to map the input transformed by the TOD-fully connected layer to obtain an output between 0 and 1. s ;

[0016] TOD-mask layer is used to calculate the mask operator M1 from the input data and the output O of the Sigmoid layer s Perform mask operation: O S ⊙M1;

[0017] The thresholding layer is used to map the output of the TOD-mask layer into a binary value of 0 or 1, indicating that the wireless device task is calculated locally or offloaded to the edge server for calculation;

[0018] The task offloading decision output layer is used to output the task offloading decision predicted by the model.

[0019] Preferably, the resource allocation subtask module comprises a Concat layer, a RAD-fully connected layer, a RAD-mask layer, a SoftMax layer and a decision-making output layer connected in sequence;

[0020] The Concat layer is used to concatenate the output of the attention mechanism layer and the output of the TOD-mask layer of the task offloading subtask module into the input of the resource allocation subtask module;

[0021] RAD-fully connected layer, used to convert the output of the Concat layer into the input of the RAD-mask layer;

[0022] The RAD-mask layer calculates the mask operator M2 from the task offloading decision output layer and applies it to the output O of the RAD-fully connected layer. R Perform mask operation: O R (i)·M2(i)+(-∞)·(1-M2(i)), where i is used to identify different wireless devices;

[0023] SoftMax layer, which maps the output of the RAD-mask layer to a probability distribution in the interval (0,1);

[0024] The resource allocation decision output layer outputs the resource allocation decision predicted by the model.

[0025] Preferably, the optimization process of step S3 includes the following steps:

[0026] S31. In the initial stage of training, the multi-task neural network model is initialized according to the number of wireless devices in the target wireless power edge computing network and the custom hyperparameters. θ , where θ represents the set of trainable parameters of the model;

[0027] S32. In each round of training, the multi-task neural network model trains and learns the input sample Input respectively, and the output obtained by the task offloading decision prediction module of the corresponding multi-task neural network model is O t , the loss function Loss of the task offloading subtask t By output O t The loss value of the task offloading label Label1 is:

[0028] Loss t =Γ1(O t ,Label1);

[0029] S33, the output obtained by the resource allocation decision prediction module of the corresponding multi-task neural network model is r , the loss function Loss of the resource allocation subtask r By output O rThe loss value of the resource allocation label Label2 is:

[0030] Loss r =Γ2(O s ,Label2);

[0031] S34, according to the loss of the task offloading subtask and the resource allocation subtask, the mixed loss Loss is calculated by the mixed parameter γ j :

[0032] Loss j =γ×Loss t +(1-γ)×Loss r ;

[0033] S34, according to the loss value Loss j For multi-task neural network model θ After adjusting the parameters, we can get the optimized multi-task neural network model.

[0034] The second object of the present invention is achieved through the following technical solution: an intelligent resource management system for a wireless power supply mobile edge computing network, used to achieve the intelligent resource management for a wireless power supply mobile edge computing network of the first object, comprising:

[0035] A data collection module, used to collect system information of the wireless power supply mobile edge computing network to form a data set;

[0036] A data processing module is used to process the data set to obtain training samples;

[0037] The model construction module is used to build a multi-task neural network model enhanced by the attention mechanism, and use training samples to train and optimize the multi-task neural network model;

[0038] A decision prediction module is used to perform model reasoning on data in the deployed wireless power mobile edge computing network environment and output a resource management decision plan, which includes specific task offloading decisions and resource allocation decisions;

[0039] The decision execution module selects local computing or edge offloading for wireless devices according to the task offloading decision, and allocates resources to the wireless devices according to the resource allocation decision.

[0040] The present invention has the following advantages over the prior art:

[0041] 1. The present invention adopts a multi-task neural network learning method, which can simultaneously optimize computational offloading decisions and resource allocation, capture and utilize relevant feature patterns between tasks, avoid repeated calculations, and achieve collaborative optimization, thereby improving the generalization ability of the model and achieving an effect close to the global optimum. At the same time, it can achieve near-real-time optimization decisions in a rapidly changing wireless environment, greatly reducing decision-making delays, thereby significantly improving real-time performance. In addition, the training cost and time can be reduced, and the results of two tasks can be obtained in one reasoning, thereby improving the reasoning speed.

[0042] 2. Through the innovative mask mechanism, the present invention can effectively handle the dynamic changes in the number of user devices in the wireless network, ensure the reasonable allocation of resources, maintain the high performance and real-time performance of the model, and enhance the scalability.

[0043] 3. The attention mechanism introduced in the present invention enables the model to better handle large-scale wireless powered mobile edge computing networks while still maintaining near-optimal performance, reflecting its adaptability to large-scale networks.

[0044] 4. The present invention greatly reduces the computational complexity of online decision-making by transferring complex calculations to the offline training stage, enabling it to run efficiently on resource-constrained edge devices, thereby achieving a significant improvement in computing efficiency.

[0045] 5. The model design of the present invention takes into account the resource limitations of edge computing and achieves low memory usage, enabling it to be deployed and applied on a variety of edge devices, thereby reducing memory requirements. Compared with deep reinforcement learning methods that require a large amount of online learning and memory resources, the model of the present invention significantly reduces resource consumption during the deployment phase, is more suitable for application on edge devices, and has better resource utilization. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a structural diagram of the multi-task neural network model enhanced by the attention mechanism for intelligent resource management of wireless powered mobile edge computing networks of the present invention.

[0047] Figure 2 It is a flowchart of the multi-task joint learning training of the present invention. DETAILED DESCRIPTION

[0048] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0049] Example 1

[0050] like Figure 1 and Figure 2 As shown, the intelligent resource management method for a wireless powered mobile edge computing network comprises the following steps:

[0051] S1. Configure a data interface on the remote controller of a wireless powered mobile edge computing network that needs to make computing decisions and allocate resources, and clone the collected system status information to the edge server of the running data collection module to form a data set; specifically, the system status includes the number of wireless devices in the wireless powered mobile edge computing network and the wireless channel gain information between the edge server and the wireless device; the data set includes a training set, a verification set and a test set.

[0052] The system status information of this embodiment uses public data on the Internet. Specifically, there are 8 different sizes of wireless power mobile edge computing network data of 5 (10,000 samples), 6 (10,000 samples), 7 (10,000 samples), 8 (10,000 samples), 9 (10,000 samples), 10 (30,000 samples), 20 (30,000 samples), and 30 wireless devices (30,000 samples). Each sample includes: the wireless channel gain between all devices and the wireless power edge server, the best binary offloading decision, and the best resource allocation decision (including the best time for the wireless power edge server to broadcast RF energy for charging wireless devices and the best time allocated to wireless devices for task offloading).

[0053] S2. Construct a multi-task neural network model enhanced by an attention mechanism, wherein the multi-task neural network model enhanced by an attention mechanism includes a task offloading decision subtask module and a resource allocation strategy subtask module composed of an input layer, a shared hidden layer, an attention mechanism layer and a composite network connected in sequence;

[0054] The input layer is used to receive the wireless channel gain h between all devices and the wireless power edge server, and its size is consistent with the number of wireless devices n in the wireless power mobile edge computing network. At the same time, the shared hidden layer is connected to the input layer, including 3 fully connected layers and a ReLU layer, which is used to convert the input features into an intermediate representation shared between the two subtasks.

[0055] The attention mechanism layer converts the intermediate representation into feature representations in multiple subspaces and performs weighted aggregation on these features to generate a richer and more accurate feature representation. Specifically, first, the intermediate representation generates a query vector Q, a key vector K, and a value vector V through three sets of linear transformations: Q = W Q g,K=W K g,V=W V g, where W Q , W K , W V is a learnable weight matrix. The attention weight is calculated by the dot product between Q and K, and the output of the attention mechanism layer is weighted by this attention weight V: d k Indicates the length of the key vector K.

[0056] The task offloading subtask module includes a TOD-fully connected layer, a Sigmoid layer, a TOD-mask layer, a thresholding layer, and a task offloading decision output layer connected in sequence. The TOD-fully connected layer is connected to the attention mechanism layer to convert the output of the attention mechanism layer into the input of the Sigmoid layer. The Sigmoid layer maps the input to obtain an output between the interval (0, 1). s The TOD-mask layer is used to handle the case of dynamic input dimension changes caused by equipment downtime or failure. First, the TOD-mask layer calculates the mask operator M1 from the input data:

[0057]

[0058] in, is an indicator function, which takes the value 1 when the condition in the brackets is true, otherwise it takes the value 0; i is used to identify different wireless devices. Then the TOD-mask layer outputs the Sigmoid layer s Perform mask operation: O s ⊙M1; where ⊙ represents element-by-element multiplication. Thereafter, the thresholding layer maps the output of the TOD-layer to a binary value of 0 or 1, indicating that the wireless device task is calculated locally or offloaded to the edge server for calculation. Finally, the task offloading decision output layer outputs the task offloading decision predicted by the model.

[0059] Among them, the resource allocation subtask module includes a Concat layer, a RAD-fully connected layer, a RAD-mask layer, a SoftMax layer, and a resource allocation decision output layer connected in sequence. The Concat layer splices the output of the attention mechanism layer and the TOD-mask layer output of the task offloading subtask module into the input of the resource allocation subtask module; the RAD-fully connected layer connects the Concat layer and converts the output of the Concat layer into the input of the RAD-mask layer. The RAD-mask layer is used to ensure that resources are not allocated to unnecessary wireless devices. The RAD-mask layer first calculates the mask operator M2 from the output x of the task offloading decision output layer:

[0060]

[0061] Then the output O of the RAD-fully connected layer R Perform mask operation: O R(i)·M2(i)+(-∞)·(1-M2(i)); the SoftMax layer maps the output of the RAD-mask layer to a probability distribution in the interval (0,1). The resource allocation decision output layer outputs the resource allocation decision predicted by the model.

[0062] The multi-task neural network model enhanced by the attention mechanism designed in this way can simultaneously optimize computational offloading decisions and resource allocation, capture and utilize relevant feature patterns between tasks, avoid repeated calculations, and achieve collaborative optimization, thereby improving the generalization ability of the model and achieving an effect close to the global optimal. At the same time, it can achieve near-real-time optimization decisions in a rapidly changing wireless environment, greatly reducing decision delays, thereby significantly improving real-time performance. In addition, the training cost and time can be reduced, and the results of two tasks can be obtained in one reasoning, thereby improving the reasoning speed. At the same time, through the innovative masking mechanism, the present invention can effectively handle the situation where the number of user devices in the wireless network changes dynamically, ensure reasonable allocation of resources, maintain the high performance and real-time performance of the model, and enhance scalability. In addition, the attention mechanism introduced by the model enables the model to better handle large-scale wireless powered mobile edge computing networks, and still maintain near-optimal performance, reflecting the adaptability to large-scale networks.

[0063] S3. Use multi-task joint learning to train a multi-task neural network model enhanced by an attention mechanism to optimize the multi-task neural network model enhanced by an attention mechanism;

[0064] The optimization process of step S3 includes the following steps:

[0065] S31. In the initial stage of training, the multi-task neural network model enhanced by the attention mechanism is initialized according to the number of wireless devices in the target wireless power edge computing network and the custom hyperparameters. θ , where θ represents the set of trainable parameters of the model; set the initial parameters, the training batch size is set to 16, the training rounds are set to 50 times, the learning rate λ is initially 0.001, and every 10 training rounds, λ is reduced to 0.1 times the original. The optimizer uses the Adam optimizer, the mixing parameter γ is set to 0.01, the loss function of the task offloading subtask uses the mean square error loss function, and the loss function of the resource allocation subtask uses the cross entropy loss function.

[0066] S32. In each round of training, the multi-task neural network model enhanced by the attention mechanism trains and learns the input sample Input respectively, and the output obtained by the task offloading decision prediction module of the corresponding multi-task neural network model enhanced by the attention mechanism is O t , the loss function Loss of the task offloading subtask t By output O tThe loss value of the task offloading label Label1 is:

[0067] Loss t =Γ1(O t ,Label1);

[0068] S33, the output obtained by the resource allocation decision prediction module of the corresponding attention mechanism enhanced multi-task neural network model is r , the loss function Loss of the resource allocation subtask r By output O r The loss value of the resource allocation label Label2 is:

[0069] Loss r =Γ2(O s ,Label2);

[0070] S34, according to the loss of the task offloading subtask and the resource allocation subtask, the mixed loss Loss is calculated by the mixed parameter γ j :

[0071] Loss j =γ×Loss t +(1-γ)×Loss r ;

[0072] S34, according to the mixed loss value Loss j Multi-task neural network model π enhanced with attention mechanism using back-propagation algorithm θ After adjusting the parameters, the multi-task neural network model with enhanced attention mechanism obtained by training is the optimized multi-task neural network model with enhanced attention mechanism obtained through multi-task joint learning training.

[0073] The present invention uses the above-mentioned multi-task joint learning training to train the multi-task neural network model enhanced by the attention mechanism, which enables a lightweight model to learn the rules of the data set and the parameter rules of the heavy model at the same time. The multi-task joint learning proposed in the present invention can ensure that the model is forward calculated once, and at the same time obtain the decision-making scheme for the two tasks of task offloading and resource allocation, while maintaining a high degree of accuracy. The time and memory consumption of model reasoning are greatly reduced, and near-real-time optimization decisions are achieved in a rapidly changing wireless environment. The algorithm flow pseudo code of multi-task joint learning training is as follows:

[0074]

[0075] S4. The optimized attention mechanism enhanced multi-task neural network model is deployed to the target wireless power mobile edge computing network through the decision prediction module. The data processing module processes the system status data to obtain the input of this attention mechanism enhanced multi-task neural network model. This model is inferred to form a specific resource management plan for computing offloading decisions and resource allocation decisions; this specific resource management decision plan executes the corresponding decision in the corresponding wireless power mobile edge computing network through the decision execution module.

[0076] In summary, the method of the present invention adopts a multi-task neural network learning method, which can simultaneously optimize computational offloading decisions and resource allocation, capture and utilize relevant feature patterns between tasks, avoid repeated calculations, and achieve collaborative optimization, thereby improving the generalization ability of the model and achieving a near-global optimal effect. At the same time, it can achieve near-real-time optimization decisions in a rapidly changing wireless environment, greatly reducing decision delays, thereby significantly improving real-time performance. In addition, the method of the present invention greatly reduces the computational complexity of online decisions by transferring complex calculations to the offline training stage, enabling it to run efficiently on resource-constrained edge devices, thereby achieving a significant improvement in computing efficiency. The multi-task joint training method can reduce training costs and time, and the results of two tasks can be obtained in one reasoning, thereby improving the reasoning speed.

[0077] The attention mechanism introduced in the multi-task model design of the method of the present invention enables the model to better handle large-scale wireless powered mobile edge computing networks and still maintain near-optimal performance, reflecting the adaptability to large-scale networks. In addition, through the innovative masking mechanism, it can effectively handle the dynamic changes in the number of user devices in the wireless network, ensure the rational allocation of resources, maintain the high performance and real-time performance of the model, and enhance scalability. At the same time, the design of the model takes into account the resource limitations of edge computing and achieves low memory usage, so that it can be deployed and applied on a variety of edge devices, thereby reducing memory requirements. Compared with deep reinforcement learning methods that require a large amount of online learning and memory resources, the model of the present invention significantly reduces resource consumption during the deployment phase, is more suitable for application on edge devices, and has better resource utilization.

[0078] Embodiment 2:

[0079] This embodiment discloses an intelligent resource management system for a wireless power supply mobile edge computing network, which is used to achieve the intelligent resource management for a wireless power supply mobile edge computing network of the first objective, including:

[0080] A data collection module, used to collect system information of the wireless power supply mobile edge computing network to form a data set;

[0081] A data processing module is used to process the data set to obtain training samples;

[0082] The model construction module is used to build a multi-task neural network model enhanced by the attention mechanism, and use training samples to train and optimize the multi-task neural network model;

[0083] The decision prediction module is used to perform model reasoning on the data in the deployed network environment to output resource management decision prediction results, and use the optimized multi-task neural network model to convert the original decision information into a specific resource management decision plan to obtain the resource allocation decision;

[0084] The decision execution module selects local computing or edge offloading for the wireless device according to the task offloading decision information, and allocates resources to the wireless device according to the resource allocation decision.

[0085] The above specific implementation modes are preferred embodiments of the present invention and cannot be used to limit the present invention. Any other changes or other equivalent replacement methods that do not deviate from the technical solution of the present invention are included in the protection scope of the present invention.

Claims

1. An intelligent resource management method for a wireless powered mobile edge computing network, characterized in that: The following steps are involved: S1. Configure a data interface on a remote controller of a wireless powered mobile edge computing network that needs to make computing decisions and allocate resources, and clone the collected system status information to a server of a running data collection module to form a data set; S2. Construct a multi-task neural network model enhanced by an attention mechanism, wherein the multi-task neural network model includes an input layer, a shared hidden layer, an attention mechanism layer, a task offloading decision subtask module, and a resource allocation strategy subtask module, wherein the input layer, the shared hidden layer, and the attention mechanism layer are connected in sequence, the task offloading decision subtask module and the resource allocation strategy subtask module are connected in parallel and are both connected to the attention mechanism layer, and the task offloading decision subtask module and the resource allocation strategy subtask module are connected to each other; S3, using multi-task joint learning to train a multi-task neural network model to optimize the multi-task neural network model; S4. Deploy the optimized multi-task neural network model to input the wireless channel gain information in the wireless powered mobile edge computing network to obtain the prediction results of computing decisions and resource allocation to maximize the network performance.

2. The intelligent resource management method for a wireless powered mobile edge computing network according to claim 1, characterized in that: The system status includes information about the number of wireless devices in the wireless powered mobile edge computing network and information about the gain of wireless channels between the edge server and the wireless devices.

3. The intelligent resource management method for a wireless powered mobile edge computing network according to claim 1, characterized in that: The data set includes a training set, a validation set and a test set.

4. The intelligent resource management method for a wireless powered mobile edge computing network according to claim 1, characterized in that: The task offloading subtask module comprises a TOD-fully connected layer, a Sigmoid layer, a TOD-mask layer, a thresholding layer and a task offloading decision output layer connected in sequence; TOD-fully connected layer, used to convert the output of the attention mechanism layer into the input of the Sigmoid layer; The Sigmoid layer is used to map the input transformed by the TOD-fully connected layer to obtain an output between 0 and 1. s ; TOD-mask layer is used to calculate the mask operator M1 from the input data and the output O of the Sigmoid layer s Perform mask operation: O S ⊙M1; The thresholding layer is used to map the output of the TOD-mask layer into a binary value of 0 or 1, indicating that the wireless device task is calculated locally or offloaded to the edge server for calculation; The task offloading decision output layer is used to output the task offloading decision predicted by the model.

5. The intelligent resource management method for a wireless powered mobile edge computing network according to claim 4, characterized in that: The resource allocation subtask module comprises a Concat layer, a RAD-fully connected layer, a RAD-mask layer, a SoftMax layer and a decision-making output layer connected in sequence; The Concat layer is used to concatenate the output of the attention mechanism layer and the output of the TOD-mask layer of the task offloading subtask module into the input of the resource allocation subtask module; RAD-fully connected layer, used to convert the output of the Concat layer into the input of the RAD-mask layer; The RAD-mask layer calculates the mask operator M2 from the task offloading decision output layer and applies it to the output O of the RAD-fully connected layer. R Perform mask operation: O R (i)·M2(i)+(-∞)·(1-M2(i)), where i is used to identify different wireless devices; SoftMax layer, maps the output of RAD--mask layer to a probability distribution in the interval (0, 1); The resource allocation decision output layer outputs the resource allocation decision predicted by the model.

6. The intelligent resource management method for a wireless powered mobile edge computing network according to claim 1, characterized in that: The optimization process of step S3 includes the following steps: S31. In the initial stage of training, the multi-task neural network model is initialized according to the number of wireless devices in the target wireless power edge computing network and the custom hyperparameters. θ , where θ represents the set of trainable parameters of the model; S32. In each round of training, the multi-task neural network model trains and learns the input sample Input respectively, and the output obtained by the task offloading decision prediction module of the corresponding multi-task neural network model is O t , the loss function Loss of the task offloading subtask t By output O t The loss value of the task offloading label Label1 is: Loss t =Γ1(O t ,Label1); S33, the output obtained by the resource allocation decision prediction module of the corresponding multi-task neural network model is r , the loss function Loss of the resource allocation subtask r By output O r The loss value of the resource allocation label Label2 is: Loss r =Γ2(O s ,Label2); S34, according to the loss of the task offloading subtask and the resource allocation subtask, the mixed loss Loss is calculated by the mixed parameter γ j : Loss j =γ×Loss t +(1-c)×Loss r ; S34, according to the loss value Loss j For multi-task neural network model θ After adjusting the parameters, we can get the optimized multi-task neural network model.

7. An intelligent resource management system for a wireless powered mobile edge computing network, characterized in that: The intelligent resource management for a wireless power supply mobile edge computing network according to any one of claims 1 to 6 comprises: A data collection module, used to collect system information of the wireless power supply mobile edge computing network to form a data set; A data processing module is used to process the data set to obtain training samples; The model construction module is used to build a multi-task neural network model enhanced by the attention mechanism, and use training samples to train and optimize the multi-task neural network model; A decision prediction module is used to perform model reasoning on data in the deployed wireless power mobile edge computing network environment and output a resource management decision plan, which includes specific task offloading decisions and resource allocation decisions; The decision execution module selects local computing or edge offloading for wireless devices according to the task offloading decision, and allocates resources to the wireless devices according to the resource allocation decision.

Images