A joint routing and resource management method and system based on a space-time attention mechanism

Abstract

The application discloses a kind of joint routing and resource management method and system based on space-time attention mechanism, specifically is: from source node, current node collects the observation information of other nodes in its perception range, and carries out data extraction;Graph attention network aggregates the information of neighborhood node to current node by multi-head attention mechanism;Time series neural network utilizes the encoder module in the framework of Transformer to process the time series information of historical nodes on the path;Using deep reinforcement learning method, according to the information after aggregation, the node selection and resource allocation of next hop are carried out;After completing node selection, relevant data and agent model are transferred to next hop node, and the node selection and resource management of next hop are carried out, until connected to destination node.The application has low resource consumption, high quality of established communication path, high resource management efficiency, can provide high-quality, high-reliability communication service for network users, and has broad application prospect.

Description

Technical Field

[0001] This invention relates to the field of communication network systems, and in particular to a joint routing and resource management method based on a spatiotemporal attention mechanism. Background Technology

[0002] With the rapid development of IoT devices, wireless ad hoc networks are widely used due to their flexible network topology. Invention patent CN202210460232.X discloses a distributed model training method and apparatus for joint routing decision-making and intra-network aggregation. This method determines resource constraints for joint routing and intra-network aggregation problems based on network resources in the distributed training network; it partitions the distributed training model according to the resource constraints to determine the allocation ratio of the distributed training model on each parameter server; it simplifies the resource constraints according to the allocation ratio, and determines the routing decision-making and intra-network aggregation scheme for the distributed training model based on the simplified resource constraints. This invention, through the formulation of joint routing and intra-network aggregation schemes, fully considers gradient routing selection and intra-network aggregation location, making full use of network resources and further improving the speed of distributed training. Invention patent CN201510069482.0 discloses a joint cooperative routing and channel allocation method in a wireless mesh network. It obtains the original topology and link state information of the entire network based on the broadcast information of network nodes; and constructs a new network topology based on virtual nodes and virtual links using a cooperative module virtualization method. The new network topology establishes an objective function model for the sum of satisfaction levels of each service flow in a multi-concurrent wireless mesh network, initially determining the cooperative transmission paths and the working channels on the uplinks of these paths. Local channel adjustments are then performed to ensure that the number of channels allocated to each node does not exceed its configured radio frequency count, and the final cooperative routing and channel allocation scheme are determined for each service flow. This invention effectively reduces co-channel interference while fully utilizing the spatial diversity gain brought by cooperative communication, thereby maximizing the overall network throughput.

[0003] However, in wireless ad hoc networks, path quality is typically closely related to routing strategies and transmission power. Traditional routing and resource management methods mostly treat these two issues separately, ignoring their coupling relationship, which leads to an inability to fully utilize network resources to establish high-quality communication paths. Furthermore, due to the limited awareness capabilities of nodes and the complex and varied network topology, how to achieve efficient routing and resource allocation within a limited awareness range is a pressing problem to be solved.

[0004] Graph Neural Networks (GNNs) have garnered significant attention in communication networks due to their powerful ability to perceive graph information and generalize to diverse scenarios. This development fills a significant gap in the routing algorithm literature, where much research has focused primarily on fixed topologies from training to testing, neglecting the flexibility of connections between nodes. GNNs are used to efficiently extract features from network topology, thereby assisting routing decisions at the network layer. As a parameterized strategy, GNNs capture network topology information for routing purposes. The computational framework or network structure is modeled as a graph, and Graph Attention Networks (GATs) are used to integrate topology configuration, vertex features, dynamic dependencies between nodes, and complex features to achieve feature extraction. Furthermore, electromagnetic interference between physical layer links can be described using graphs, enabling GNN-based resource allocation methods. GATs are designed to learn channel gain magnitudes between links and manage cumulative interference patterns by leveraging attention-driven neighbor aggregation mechanisms, while Graph Convolutional Networks (GCNs) can also be strategically deployed to address congestion interference modeling problems.

[0005] Furthermore, the temporal information from the first few steps in a Markov Decision Process (MDP) can be used to compensate for the partial perception capabilities of nodes, thereby improving the performance of local decision-making. Long Short-Term Memory (LSTM) networks excel at learning long-term dependencies in sequences and are used to capture the periodic features of communication channels, thus improving the accuracy of channel state prediction. The proposed Transformer architecture can effectively capture long-term dependencies in sequence data. When processing each time step, the model can utilize an attention mechanism to focus on any position in the sequence, rather than simply performing sequential computation. The Transformer is applied to represent the Q-function in Deep Reinforcement Learning (DRL), thereby providing more accurate value estimates for various state-action pairs. Due to the presence of the multi-head attention mechanism, different parts of the input sequence can be focused on simultaneously, capturing complex spatiotemporal dependencies. The Transformer computes the relationship between actions and observations through a causal attention network, enabling the agent to optimize decisions within its corresponding action space using a policy gradient algorithm. In addition, the spatial feature extraction capabilities of GAT are combined with the temporal series modeling of Temporal Convolutional Networks (TCNs) to extract highly correlated spatiotemporal data patterns, effectively capturing the dynamic evolution patterns of channel features. Summary of the Invention

[0006] The purpose of this invention is to provide a joint routing and resource management method for communication networks that features low resource consumption, high-quality established communication paths, and high efficiency in resource management.

[0007] The technical solution to achieve the purpose of this invention is: a joint routing and resource management method based on a spatiotemporal attention mechanism, comprising the following steps:

[0008] Step 1: Starting from the source node, the current node collects observation information from other nodes within its sensing range and extracts the data.

[0009] Step 2: The graph attention network aggregates information from neighboring nodes to the current node through a multi-head attention mechanism;

[0010] Step 3: The temporal neural network uses the encoder module in the Transformer framework to process the temporal information of historical nodes on the path;

[0011] Step 4: Employ deep reinforcement learning methods to select the next hop node and allocate resources based on the aggregated information;

[0012] Step 5: After completing node selection, transfer the observation information to the selected node, perform the next hop node selection and resource management, until the destination node is connected.

[0013] A joint routing and resource management system based on a spatiotemporal attention mechanism is provided. The system is used to implement the joint routing and resource management method based on the spatiotemporal attention mechanism. The system includes a data collection module, an aggregation module, a time series information processing module, a decision module, and a model transfer module.

[0014] The data collection module starts from the source node, collects observation information from other nodes within its sensing range, and extracts the data.

[0015] The aggregation module, based on a graph attention network, aggregates information from neighboring nodes to the current node through a multi-head attention mechanism;

[0016] The temporal information processing module uses a temporal neural network and the encoder module in the Transformer framework to process the temporal information of historical nodes on the path.

[0017] The decision-making module employs deep reinforcement learning to select the next hop node and allocate resources based on the aggregated information.

[0018] The model transfer module, after completing node selection, transfers the observation information to the selected node, performs the next hop node selection and resource management, until it connects to the destination node.

[0019] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned joint routing and resource management method based on a spatiotemporal attention mechanism.

[0020] Compared with the prior art, the significant advantages of this invention are:

[0021] (1) Using spatial attention mechanism, the local perception information of nodes is used to learn the relationship between nodes and aggregate the information of neighboring nodes. Through multi-head attention mechanism, the information of neighboring nodes is aggregated to the current node, thereby expanding the spatial perception range of the node, enabling the node to obtain more spatial information within a limited perception range, and improving the accuracy of routing selection.

[0022] (2) The temporal attention mechanism is adopted, and the encoder module in the Transformer framework is used to process the temporal information of historical nodes on the path, so that the current node can pay attention to the key information on the historical path, thereby enhancing the temporal awareness of the node, effectively overcoming the limitation of the node's perception range, and improving the efficiency of resource management.

[0023] (3) Deep reinforcement learning (DRL) is used to solve the partially observable Markov decision process (POMDP) ​​problem. The outputs of GAT and time attention mechanism are used as state inputs. The optimal routing and transmission power adjustment strategies are learned through the DRL model. It can adapt to different network topologies, including random distribution, multi-flow coexistence and clustered distribution, and has good robustness.

[0024] (4) The priority experience replay mechanism is adopted. Different replay probabilities are assigned according to the estimation error of the experience. Priority is given to replaying the experience with larger errors, thereby improving the training efficiency of the model, reducing training time, and improving the efficiency of resource management. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the joint routing and resource management method based on the spatiotemporal attention mechanism of this invention.

[0026] Figure 2 This is the network model diagram of the present invention.

[0027] Figure 3 This is a comparison chart of the training results of the present invention and the fully connected model.

[0028] Figure 4 This is an example diagram of the routing results of the present invention.

[0029] Figure 5 This is an example diagram of the routing results for a fully connected model. Detailed Implementation

[0030] This invention discloses a joint routing and resource management method based on a spatiotemporal attention mechanism, aiming to establish high-quality communication paths in multi-hop networks through joint routing and resource management. This method employs hop-by-hop optimized deep reinforcement learning (DRL) to solve the inherent partially observable Markov decision problem (POMDP) ​​in path establishment, and integrates graph neural networks, temporal neural networks, and deep reinforcement learning techniques.

[0031] The principle of this invention is as follows: This invention aggregates the observation information of each node within the observation area of ​​a node based on a graph neural network, and then uses a temporal neural network to aggregate historical observation information on the path to expand the observation area of ​​the node. Then, it uses deep reinforcement learning (DRL) to select the next hop node and allocate resources based on the above aggregated information.

[0032] This invention provides a joint routing and resource management method based on a spatiotemporal attention mechanism, comprising the following steps:

[0033] Step 1: Starting from the source node, the current node collects observation information from other nodes within its sensing range and extracts the data.

[0034] Step 2: The graph attention network aggregates information from neighboring nodes to the current node through a multi-head attention mechanism;

[0035] Step 3: The temporal neural network uses the encoder module in the Transformer framework to process the temporal information of historical nodes on the path;

[0036] Step 4: Employ deep reinforcement learning methods to select the next hop node and allocate resources based on the aggregated information;

[0037] Step 5: After completing node selection, transfer the observation information to the selected node, perform the next hop node selection and resource management, until the destination node is connected.

[0038] As a specific example, in step 1, the current node collects observation information from other nodes within its sensing range and performs data extraction. The extracted data includes geographical location information, signal-to-noise ratio (SINR), remaining node energy, data transmission rate, and processing load, wherein:

[0039] Geographic location information: The coordinates of neighboring nodes, used to calculate the distance and direction between nodes;

[0040] Signal-to-noise ratio (SINR): The ratio of the signal strength of neighboring nodes to the interference plus noise, used to evaluate link quality;

[0041] Node remaining energy: The battery power of neighboring nodes, used to assess the node's sustainable communication capability;

[0042] Data transmission rate: The current data transmission rate of neighboring nodes, used to evaluate the throughput capacity of the link;

[0043] Processing load: Metrics of processing tasks or resource consumption, including CPU, memory, and bandwidth resource utilization, or the number of currently running tasks.

[0044] As a specific example, the graph attention network described in step 2 aggregates information from neighboring nodes to the current node through a multi-head attention mechanism, as follows:

[0045] Step 2.1: Construct a feature vector for each node, including geographical location, signal-to-noise ratio, remaining battery power, data transmission rate, and processing load;

[0046] Step 2.2: Calculate the attention weights between the current node and its neighboring nodes, and dynamically allocate the weights based on the feature similarity and topological relationship between the nodes;

[0047] Step 2.3: Based on the attention weight, aggregate the information of the neighboring nodes to the current node to provide graph feature information for path selection.

[0048] As a specific example, the calculation process for steps 2.2 and 2.3 is as follows:

[0049] Attention coefficient The calculation is as follows:

[0050] (1)

[0051] in, Indicates observation information, , express Time Node , Observed values; and These are the weight vector and weight matrix of the attention mechanism, respectively;

[0052] Under the multi-head attention mechanism, the observation information is updated as follows:

[0053] (2)

[0054] in, The number of heads indicating multi-head attention; Indicates the first Attention coefficient of each attention head; Indicates matrix concatenation;

[0055] The output of graph attention is an aggregation of the updated observations:

[0056] (3)

[0057] in, This represents the aggregated observations.

[0058] As a specific example, the temporal neural network described in step 3 utilizes the encoder module in the Transformer framework to process the temporal information of historical nodes on the path, as follows:

[0059] Step 3.1: Extract observation information of historical nodes along the path, including geographical location, signal-to-noise ratio, and remaining battery power, as a time-series feature sequence;

[0060] Step 3.2: Through the self-attention mechanism of Transformer, aggregate the temporal information of historical nodes on the path to enhance the temporal awareness of nodes;

[0061] Step 3.3: Combine the aggregated temporal information with the information aggregated by the graph attention network to form a comprehensive state representation.

[0062] As a specific example, in step 3, the attention aggregation result of historical step observation information is calculated as follows:

[0063] (4)

[0064] in, This represents the result of attention calculation for the perceived data sequence; , These represent the query vector, key vector, and value vector, respectively. This represents the dimensions of the query vector and key vector, i.e., the dimensions of each attention head. It is a normalized exponential function;

[0065] After multi-head attention merging, the final output result for:

[0066] (5)

[0067] in, The number of heads processed for historical observation information. For the index of attention heads; This indicates matrix concatenation.

[0068] As a specific example, step 4 employs a deep reinforcement learning method to select the next hop node and allocate resources based on the aggregated information, as detailed below:

[0069] Step 4.1: Input the observation information processed by the spatiotemporal attention mechanism as the state into the deep reinforcement learning model;

[0070] Step 4.2: The deep reinforcement learning model transforms the observed data into the decision space and outputs the Q-value.

[0071] (6)

[0072] Where W is the dimension transformation matrix, which transforms the dimension of the Q function to the action dimension; , , These represent the object, action, and model parameters, respectively.

[0073] Through Q value The argmax process obtains node selection and resource allocation decisions, including the selection of the next-hop node and the power settings of that node;

[0074] Step 4.3: Design a reward function that comprehensively considers factors such as link quality, energy consumption, latency, and throughput to evaluate the performance of path selection and resource allocation;

[0075] Step 4.4: Through interaction with the environment, dynamically learn the optimal path selection and resource allocation strategies to accelerate model convergence.

[0076] As a specific example, after completing node selection as described in step 5, the relevant data and agent model are transferred to the next-hop node for next-hop node selection and resource management, until the connection to the destination node is established, as follows:

[0077] Step 5.1: After completing node selection, record the observation information of the current node. Aggregate information and transition the state of the DRL model to the next hop node;

[0078] Step 5.2: The next hop node becomes the new current node. Repeat steps 1 to 4 to continue path selection and resource allocation.

[0079] Step 5.3: Perform path selection and resource allocation hop by hop until a successful connection to the destination node is established, thus completing the establishment of the entire communication path.

[0080] The present invention also provides a joint routing and resource management system based on a spatiotemporal attention mechanism. The system is used to implement the joint routing and resource management method based on the spatiotemporal attention mechanism. The system includes a data collection module, an aggregation module, a time series information processing module, a decision module, and a model transfer module.

[0081] The data collection module starts from the source node, collects observation information from other nodes within its sensing range, and extracts the data.

[0082] The aggregation module, based on a graph attention network, aggregates information from neighboring nodes to the current node through a multi-head attention mechanism;

[0083] The temporal information processing module uses a temporal neural network and the encoder module in the Transformer framework to process the temporal information of historical nodes on the path.

[0084] The decision-making module employs deep reinforcement learning to select the next hop node and allocate resources based on the aggregated information.

[0085] The model transfer module, after completing node selection, transfers the observation information to the selected node, performs the next hop node selection and resource management, until it connects to the destination node.

[0086] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned joint routing and resource management method based on a spatiotemporal attention mechanism.

[0087] In summary, this invention has the following characteristics: First, the Graph Attention Network (GAT) can extract graph information related to neighboring nodes, accurately capturing the topological structure and interrelationships between nodes, providing key graph feature information for path selection. Second, by processing the observation information of historical nodes on the path through the autoencoder module in the Transformer framework, the historical state change patterns of nodes on the path are effectively extracted, providing a time-dimensional reference for path optimization. Finally, based on the graph information extracted by the GAT and the historical observation information processed by the autoencoder module, the Deep Reinforcement Learning (DRL) module dynamically selects the optimal path using a hop-by-hop optimization approach to maximize the overall performance of the network. Furthermore, this invention utilizes local graph information and historical time data on the path to expand the observable range of nodes, thereby significantly improving the network's situational awareness and enabling the network to more accurately predict and optimize path selection, reducing path selection errors caused by incomplete information. During model training, a priority experience replay mechanism is introduced, prioritizing experiences according to their importance, and replaying those experiences more valuable to model training first, thereby accelerating the model's convergence speed and improving overall training efficiency.

[0088] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0089] Example

[0090] like Figure 1 , Figure 2As shown, the present invention provides a joint routing and resource management method based on a spatiotemporal attention mechanism, comprising the following steps:

[0091] Step 1: Starting from the source node, the current node collects observation information from other nodes within its sensing range and extracts data. The extracted data includes geographical location information, signal-to-noise ratio (SINR), node remaining energy, data transmission rate, and processing load.

[0092] Geographic location information: The coordinates of neighboring nodes, used to calculate the distance and direction between nodes;

[0093] Signal-to-noise ratio (SINR): The ratio of the signal strength of neighboring nodes to the interference plus noise, used to evaluate link quality;

[0094] Node remaining energy: The battery power of neighboring nodes, used to assess the node's sustainable communication capability;

[0095] Data transmission rate: The current data transmission rate of neighboring nodes, used to evaluate the throughput capacity of the link;

[0096] Processing load: Metrics of processing tasks or resource consumption, including CPU, memory, and bandwidth resource utilization, or the number of currently running tasks.

[0097] This data will serve as input for subsequent path selection and resource allocation, providing fundamental data support for the network's intelligent decision-making.

[0098] Step 2: The graph attention network aggregates information from neighboring nodes to the current node through a multi-head attention mechanism, as follows:

[0099] Step 2.1: Construct a feature vector for each node, including geographical location, signal-to-noise ratio, remaining battery power, data transmission rate, and processing load;

[0100] Step 2.2: Calculate the attention weights between the current node and its neighboring nodes, and dynamically allocate the weights based on the feature similarity and topological relationship between the nodes. This mechanism can automatically allocate different weights based on the feature similarity and topological relationship between the nodes, so that important neighboring nodes have a greater influence on the current node.

[0101] Step 2.3: Based on attention weights, aggregate information from neighboring nodes to the current node, providing graph feature information for path selection. This method effectively integrates information from neighboring nodes, providing more comprehensive graph feature information for subsequent path selection.

[0102] Through the above steps, GAT can dynamically allocate attention weights, thereby more accurately capturing the interaction relationships between nodes and providing more accurate graph feature information for path selection.

[0103] Step 3: The temporal neural network uses the encoder module in the Transformer framework to process the temporal information of historical nodes on the path, as follows:

[0104] Step 3.1: Extract observation information of historical nodes along the path, including geographical location, signal-to-noise ratio, and remaining battery power, as a time-series feature sequence;

[0105] Step 3.2: Through the self-attention mechanism of Transformer, the temporal information of historical nodes on the path is aggregated to enhance the temporal awareness of the nodes. This mechanism can automatically identify and focus on important nodes and events in the historical path, thereby enhancing the temporal awareness of the nodes.

[0106] Step 3.3: Combine the aggregated temporal information with the information aggregated by the graph attention network to form a comprehensive state representation. The aggregated temporal information will serve as input for subsequent path selection and resource allocation.

[0107] Through the steps described above, the Transformer framework can effectively handle dynamic changes on the path, enabling nodes to make more rational decisions based on historical state trends.

[0108] Step 4: Employ deep reinforcement learning methods to select the next hop node and allocate resources based on the aggregated information, as detailed below:

[0109] Step 4.1: Input the observation information processed by the spatiotemporal attention mechanism into the deep reinforcement learning model as the state information; the state information includes the feature vector of the current node, the aggregation information of the neighboring nodes, and the temporal information of the historical nodes on the path.

[0110] Step 4.2: The deep reinforcement learning model outputs actions, including the selection of the next hop node and the power setting of that node; the action space includes selecting a neighboring node as the next hop node and setting the corresponding power.

[0111] Step 4.3: Design a reward function that comprehensively considers factors such as link quality, energy consumption, latency, and throughput to evaluate the performance of path selection and resource allocation. The reward function comprehensively considers link quality, latency, such as SINR, path length, and node processing capacity, and balances the contributions of different indicators through weighting coefficients.

[0112] Step 4.4: Through interaction with the environment, the system dynamically learns the optimal path selection and resource allocation strategies to accelerate model convergence. During training, a priority experience replay mechanism is introduced, prioritizing experiences based on their importance and replaying those experiences that are more valuable for model training, thereby accelerating model convergence and improving overall training efficiency.

[0113] Through the above steps, the DRL model can dynamically optimize path selection and resource allocation based on the real-time status of the network, thereby achieving high-quality communication path establishment in multi-hop networks.

[0114] Step 5: After completing node selection, transfer the relevant data and agent model to the next-hop node, perform next-hop node selection and resource management, until connecting to the destination node, as follows:

[0115] Step 5.1: After completing the node selection, transfer the observation data, aggregation information, and DRL model state of the current node to the next hop node;

[0116] Step 5.2: The next hop node becomes the new current node. Repeat steps 1 to 4 to continue path selection and resource allocation.

[0117] Step 5.3: Perform path selection and resource allocation hop by hop until a successful connection to the destination node is established, thus completing the establishment of the entire communication path.

[0118] By using this hop-by-hop optimization method, the present invention can ensure that each decision is based on the best available information, thereby achieving globally optimal path selection and resource allocation.

[0119] A joint routing and resource management device based on a spatiotemporal attention mechanism includes a data collection module, a data processing module, a decision-making module, and a model transfer module;

[0120] The data collection module is used to collect observation information of other nodes within the current node's sensing range, starting from the source node, and to extract the data.

[0121] The data processing module includes a graph attention network and a temporal neural network, which are used to aggregate the observation information of each node within the node's observation area and the historical observation information on the path, thereby expanding the node's observation area.

[0122] The decision-making module employs a deep reinforcement learning method to select the next hop node and allocate resources based on the aggregated information, thus solving some observable Markov decision process problems.

[0123] The model transfer module is used to transfer relevant data and agent models to the next hop node after completing node selection, and continue to select the next hop and manage resources until the target node is connected.

[0124] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned joint routing and resource management method based on a spatiotemporal attention mechanism.

[0125] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the aforementioned joint routing and resource management method based on a spatiotemporal attention mechanism.

[0126] Figure 3 The graph shows a comparison of the training results of the method of this invention and the fully connected model. It can be seen that the method of this invention can learn and utilize valuable experience more efficiently, accelerate the convergence process, and significantly improve the performance and efficiency of the algorithm. This not only improves the training speed of the model, but also enhances its adaptability and robustness in complex network environments. Figure 4 This is an example diagram of the routing results using the method in this invention. Figure 5 The diagram shows an example of routing results using a fully connected model. It can be seen that the method of this invention can more accurately predict the optimal action selection, establish higher quality network links, and improve the efficiency of IoT joint routing and resource management.

[0127] This invention can be implemented in software and deployed on nodes in a wireless network. By running this software, the nodes can automatically learn optimal routing and resource allocation strategies, thereby establishing high-quality communication paths. In practical applications, the model's parameters and hyperparameters can be adjusted according to different network scenarios and requirements to achieve optimal performance.

[0128] This invention is applicable to various complex multi-hop network environments, such as wireless sensor networks, Internet of Things (IoT) networks, industrial internet, and distributed computing networks. By combining routing and resource management, it optimizes the allocation of network resources, ensuring efficient utilization and further improving overall network performance. This method not only effectively addresses challenges such as node energy consumption, signal strength, network topology changes, and dynamic traffic variations, but also provides network users with high-quality, high-reliability communication services, demonstrating broad application prospects and significant practical importance.

[0129] This embodiment also provides a computer-readable storage medium storing a computer program. When the computer program is run by a processor, it executes the steps of the joint routing and resource management method based on the spatiotemporal attention mechanism described in the above method embodiments. The storage medium can be a volatile or non-volatile computer-readable storage medium.

[0130] In addition, this embodiment also provides a computer program product, on which a computer program is stored. When the computer program is run by a processor, it executes the steps of the joint routing and resource management method based on the spatiotemporal attention mechanism provided in the above embodiment. For details, please refer to the above method embodiment, which will not be repeated here.

[0131] The aforementioned computer program product can be implemented through hardware, software, or a combination thereof. In one optional embodiment, the computer program product is specifically embodied in a computer storage medium, which can be a volatile or non-volatile computer-readable storage medium. In another optional embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0132] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices and apparatuses described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. In this embodiment, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division; in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection may be through some communication interfaces; the indirect coupling or communication connection of devices or units may be electrical, mechanical, or other forms.

[0133] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0134] In addition, in this embodiment, each functional unit can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0135] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the essence of this technical solution, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0136] Finally, it should be noted that the above-described embodiments are merely specific implementations of this disclosure, used to illustrate the technical solutions of this disclosure, and not to limit it. The protection scope of this disclosure is not limited thereto. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this disclosure; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be covered within the protection scope of this disclosure. Therefore, the protection scope of this disclosure should be determined by the protection scope of the claims.

Claims

1. A joint routing and resource management method based on a spatiotemporal attention mechanism, characterized in that, Includes the following steps: Step 1: Starting from the source node, the current node collects observation information from other nodes within its sensing range and extracts the data. Step 2: The graph attention network aggregates information from neighboring nodes to the current node through a multi-head attention mechanism; Step 3: The temporal neural network uses the encoder module in the Transformer framework to process the temporal information of historical nodes on the path; Step 4: Employ deep reinforcement learning methods to select the next hop node and allocate resources based on the aggregated information; Step 5: After completing node selection, transfer the observation information to the selected node, perform the next hop node selection and resource management, until the destination node is connected.

2. The joint routing and resource management method based on spatiotemporal attention mechanism according to claim 1, characterized in that, In step 1, the current node collects observation information from other nodes within its sensing range and extracts data. The extracted data includes geographical location information, signal-to-noise ratio (SINR), remaining node energy, data transmission rate, and processing load. Geographic location information: The coordinates of neighboring nodes, used to calculate the distance and direction between nodes; Signal-to-noise ratio (SINR): The ratio of the signal strength of neighboring nodes to the interference plus noise, used to evaluate link quality; Node remaining energy: The battery power of neighboring nodes, used to assess the node's sustainable communication capability; Data transmission rate: The current data transmission rate of neighboring nodes, used to evaluate the throughput capacity of the link; Processing load: Metrics of processing tasks or resource consumption, including CPU, memory, and bandwidth resource utilization, or the number of currently running tasks.

3. The joint routing and resource management method based on spatiotemporal attention mechanism according to claim 2, characterized in that, The graph attention network described in step 2 aggregates information from neighboring nodes to the current node through a multi-head attention mechanism, as detailed below: Step 2.1: Construct a feature vector for each node, including geographical location, signal-to-noise ratio, remaining battery power, data transmission rate, and processing load; Step 2.2: Calculate the attention weights between the current node and its neighboring nodes, and dynamically allocate the weights based on the feature similarity and topological relationship between the nodes; Step 2.3: Based on the attention weight, aggregate the information of the neighboring nodes to the current node to provide graph feature information for path selection.

4. The joint routing and resource management method based on spatiotemporal attention mechanism according to claim 3, characterized in that, The calculation process for steps 2.2 and 2.3 is as follows: Attention coefficient The calculation is as follows: (1) in, Indicates observation information, , express Time Node , Observed values; and These are the weight vector and weight matrix of the attention mechanism, respectively; Under the multi-head attention mechanism, the observation information is updated as follows: (2) in, The number of heads indicating multi-head attention; Indicates the first Attention coefficient of each attention head; Indicates matrix concatenation; The output of graph attention is an aggregation of the updated observations: (3) in, This represents the aggregated observations.

5. The joint routing and resource management method based on spatiotemporal attention mechanism according to claim 4, characterized in that, The temporal neural network described in step 3 utilizes the encoder module in the Transformer framework to process the temporal information of historical nodes along the path, as detailed below: Step 3.1: Extract observation information of historical nodes along the path, including geographical location, signal-to-noise ratio, and remaining battery power, as a time-series feature sequence; Step 3.2: Through the self-attention mechanism of Transformer, aggregate the temporal information of historical nodes on the path to enhance the temporal awareness of nodes; Step 3.3: Combine the aggregated temporal information with the information aggregated by the graph attention network to form a comprehensive state representation.

6. The joint routing and resource management method based on spatiotemporal attention mechanism according to claim 5, characterized in that, In step 3, the attention aggregation result of historical step observation information is calculated as follows: (4) in, This represents the result of attention calculation for the perceived data sequence; , These represent the query vector, key vector, and value vector, respectively. This represents the dimensions of the query vector and key vector, i.e., the dimensions of each attention head. It is a normalized exponential function; After multi-head attention merging, the final output result for: (5) in, The number of heads processed for historical observation information. For the index of attention heads; This indicates matrix concatenation.

7. The joint routing and resource management method based on spatiotemporal attention mechanism according to claim 6, characterized in that, Step 4 describes the use of deep reinforcement learning to select the next hop node and allocate resources based on the aggregated information, as detailed below: Step 4.1: Input the observation information processed by the spatiotemporal attention mechanism as the state into the deep reinforcement learning model; Step 4.2: The deep reinforcement learning model transforms the observed data into the decision space and outputs the Q-value. (6) Where W is the dimension transformation matrix, which transforms the dimension of the Q function to the action dimension; , , These represent the object, action, and model parameters, respectively. Through Q value The argmax process obtains node selection and resource allocation decisions, including the selection of the next-hop node and the power settings of that node; Step 4.3: Design a reward function that comprehensively considers factors such as link quality, energy consumption, latency, and throughput to evaluate the performance of path selection and resource allocation; Step 4.4: Through interaction with the environment, dynamically learn the optimal path selection and resource allocation strategies to accelerate model convergence.

8. The joint routing and resource management method based on spatiotemporal attention mechanism according to claim 7, characterized in that, After completing node selection as described in step 5, the relevant data and agent model are transferred to the next-hop node for next-hop node selection and resource management until the destination node is connected, as detailed below: Step 5.1: After completing node selection, record the observation information of the current node. Aggregate information and transition the state of the DRL model to the next hop node; Step 5.2: The next hop node becomes the new current node. Repeat steps 1 to 4 to continue path selection and resource allocation. Step 5.3: Perform path selection and resource allocation hop by hop until a successful connection to the destination node is established, thus completing the establishment of the entire communication path.

9. A joint routing and resource management system based on a spatiotemporal attention mechanism, characterized in that, The system is used to implement the joint routing and resource management method based on spatiotemporal attention mechanism as described in any one of claims 1 to 8. The system includes a data collection module, an aggregation module, a time series information processing module, a decision module, and a model transfer module. The data collection module starts from the source node, collects observation information from other nodes within its sensing range, and extracts the data. The aggregation module, based on a graph attention network, aggregates information from neighboring nodes to the current node through a multi-head attention mechanism; The temporal information processing module uses a temporal neural network and the encoder module in the Transformer framework to process the temporal information of historical nodes on the path. The decision-making module employs deep reinforcement learning to select the next hop node and allocate resources based on the aggregated information. The model transfer module, after completing node selection, transfers the observation information to the selected node, performs the next hop node selection and resource management, until it connects to the destination node.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the joint routing and resource management method based on the spatiotemporal attention mechanism as described in any one of claims 1 to 8.

Images