A wireless communication multi-link load balancing and intelligent optimization method

Abstract

The application discloses a wireless communication multi-link load balancing and intelligent optimization method, and belongs to the technical field of wireless communication.The method comprises the following steps: collecting multi-modal link state data of a target communication terminal in a multi-link concurrent transmission environment; performing space-time correlation analysis and processing on the multi-modal link state data to generate a multi-link joint feature tensor containing a link interference coupling coefficient, a load dynamic offset vector and a service flow space-time distribution mode; calling a pre-trained deep separable attention network to perform load balancing decision generation processing on the multi-link joint feature tensor, and outputting a target load distribution coefficient of each link and a flow scheduling path matrix; and performing intelligent decomposition and mapping processing on original service flow according to the target load distribution coefficient to generate a multi-link distribution strategy based on service flow granularity.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, specifically to a method for multi-link load balancing and intelligent optimization in wireless communication. Background Technology

[0002] In multi-link concurrent transmission scenarios, wireless communication terminals need to manage traffic distribution across various heterogeneous links such as Wi-Fi and cellular networks simultaneously. Existing load balancing technologies generally rely on single-dimensional link quality indicators, such as static threshold-based traffic distribution based solely on real-time throughput or signal strength. This ignores the complex electromagnetic interference coupling effects between multiple links and the dynamic distribution characteristics of service flows in the spatiotemporal dimension, causing load allocation decisions to lag behind the rapid changes in channels and services.

[0003] Existing machine learning-based load balancing technologies mostly treat each link as an independent entity, extracting features separately and then simply concatenating or weighting them together. These methods fail to effectively model the topological relationships between links and the spatial propagation patterns of signal interference, and also lack the ability to finely analyze causal and temporal dependencies in historical sequence data. When faced with sudden link congestion or interference fluctuations, the resulting traffic splitting decisions are difficult to accurately match the service quality requirements of the business flow, easily leading to some links being overloaded while others are underutilized.

[0004] How to jointly extract spatial interference coupling relationships and temporal load evolution patterns from multimodal link data, and construct a unified feature space that can deeply characterize the state of multi-link systems, is an urgent problem to be solved. Another direction that needs breakthrough is to design an efficient attention computing architecture that can accurately decode the aforementioned joint features and map multi-granularity business paths while avoiding an explosion in model parameters. Summary of the Invention

[0005] The purpose of this invention is to provide a method for multi-link load balancing and intelligent optimization in wireless communication, addressing the decision-making lag caused by the failure of existing technologies to deeply integrate inter-link interference coupling, dynamic load shift, and the spatiotemporal distribution characteristics of service flows. This invention constructs a multi-link joint feature tensor to uniformly express the spatial topology and temporal evolution information of the links; it utilizes a deep separable attention network to extract differentiated features through separate spatial and temporal coding and cross-fusion them, achieving real-time load allocation and path scheduling at the service flow granularity.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a method for multi-link load balancing and intelligent optimization in wireless communication, comprising: collecting multi-modal link status data of a target communication terminal in a multi-link concurrent transmission environment; performing spatiotemporal correlation analysis on the multi-modal link status data to generate a multi-link joint feature tensor, wherein the multi-link joint feature tensor includes inter-link interference coupling coefficients, load dynamic offset vectors, and spatiotemporal distribution patterns of service flows, thereby achieving in-depth characterization of heterogeneous link status information and effective fusion of spatiotemporal correlation features; calling a pre-trained deep separable attention network to perform load balancing decision generation processing on the multi-link joint feature tensor, outputting the target load allocation coefficient and traffic scheduling path matrix for each link, utilizing the structural characteristics of the network to reduce computational complexity while ensuring decision accuracy, and improving the load balancing response speed in dynamic environments; and performing intelligent decomposition and mapping processing on the original service traffic according to the target load allocation coefficient to generate a multi-link allocation strategy based on service flow granularity, thereby refining traffic allocation from coarse-grained link level to service flow level, improving resource scheduling accuracy and service quality assurance capabilities.

[0007] As a technical solution of the present invention, the step of performing spatiotemporal correlation analysis on multimodal link state data to generate a multi-link joint feature tensor includes: the multimodal link state data containing a channel state information sequence, a real-time load rate sequence, a service flow type identifier, and geospatial location coordinates for each link; performing temporal differential coding on the channel state information sequence to obtain a channel quality fluctuation feature vector for each link, thereby highlighting the changing trend of the channel state; and performing sliding window statistical feature extraction on the real-time load rate sequence to generate a multidimensional load feature vector containing the load rate mean, variance, and peak distribution, thereby reflecting... The dynamic range and statistical patterns of the load are reflected; the service flow type identifier is mapped to a one-hot encoded vector, and a spatial proximity matrix is ​​constructed by combining the geographic location coordinates to introduce service type preferences and spatial correlation between links; the channel quality fluctuation feature vector, multi-dimensional load feature vector, one-hot encoded vector and spatial proximity matrix are input into a preset spatiotemporal graph convolutional network for node feature aggregation and edge weight update processing, and the output multi-link joint feature tensor with spatiotemporal correlation information is output, so that the generated tensor can synchronously express the individual state of the link, the mutual influence between links and the spatial pattern of service distribution, providing a feature basis for accurate decision making.

[0008] Preferably, the step of inputting the channel quality fluctuation feature vector, multidimensional load feature vector, one-hot encoded vector, and spatial proximity matrix into a preset spatiotemporal graph convolutional network for node feature aggregation and edge weight update processing includes: using each link as a graph node, and using the geographical distance and signal-to-interference ratio between links as initial edge weights to construct an initial link topology graph, thereby introducing the physical spatial relationship and radio frequency interference relationship between links into the graph structure; for each node in the initial link topology graph, aggregating the channel quality fluctuation feature vector and multidimensional load feature vector of its first-order neighboring nodes to generate aggregated node neighborhood features, enabling the node to perceive the load and channel conditions of surrounding links; updating the node's own features based on the one-hot encoded vector, and dynamically adjusting the edge weights between adjacent nodes in combination with the spatial proximity matrix to generate an optimized graph structure containing service type preferences, enabling the graph structure to autonomously adjust the correlation strength between links according to service type differences; inputting the optimized graph structure and corresponding node features into multiple layers of graph convolutional layers for iterative message passing and feature updating, outputting a multi-link joint feature tensor, thereby fully extracting the deep spatiotemporal dependencies in the link network and improving the sensitivity of feature representation to network state changes.

[0009] As another technical solution of the present invention, the step of calling a pre-trained deep separable attention network to perform load balancing decision generation processing on the multi-link joint feature tensor and outputting the target load allocation coefficient and traffic scheduling path matrix for each link includes: performing spatiotemporal dimension separation processing on the multi-link joint feature tensor to generate spatial attention feature maps and temporal attention feature maps, so that spatial correlation features and temporal evolution features can be focused and extracted respectively; inputting the spatial attention feature map into the spatial coding layer of the deep separable attention network for channel-wise convolution and point convolution combination processing to output spatial context feature vectors, using deep separable convolution to reduce computational overhead while maintaining the ability to extract spatial structural information; inputting the temporal attention feature map into the deep separable attention network... The temporal encoding layer of the attention network performs a combination of causal convolution and gating units to output temporal dependency feature vectors, capturing the long-term dependencies of link states in the time dimension. Multi-head cross-attention fusion is applied to the spatial context feature vectors and temporal dependency feature vectors to generate fused feature representations, enabling full interaction between local spatial features and global temporal evolution information, thus enhancing the feature representation capabilities of complex network behaviors. The fused feature representations are input into the decision output layer of the depthwise separable attention network. Target load allocation coefficients are generated through fully connected layers and softmax normalization. Simultaneously, a traffic scheduling path matrix is ​​output through a path selection matrix generation module, thereby achieving joint decision-making for load allocation and path planning, avoiding performance losses caused by step-by-step decision-making.

[0010] Further preferably, the step of performing multi-head cross-attention fusion processing on the spatial context feature vector and the temporal dependency feature vector to generate a fused feature representation includes: mapping the spatial context feature vector to a first query matrix, and mapping the temporal dependency feature vector to a first key matrix and a first value matrix; calculating the dot product similarity between the first query matrix and the first key matrix, and performing scaling and softmax normalization to obtain the spatial attention weight matrix for temporal information; performing a weighted summation of the first value matrix based on the spatial attention weight matrix for temporal information to generate a first intermediate fused feature, which characterizes the key attention region of spatial features for temporal information; mapping the temporal dependency feature vector to a second query matrix, and mapping the spatial context feature vector to a first query matrix and the temporal dependency feature vector to a first key matrix and a first value matrix; calculating the dot product similarity between the first query matrix and the first key matrix, and performing scaling and softmax normalization to obtain the spatial attention weight matrix for temporal information; mapping the spatial context feature vector to a second query matrix, and mapping the spatial context feature vector to a first query matrix and the temporal dependency feature vector to a second query matrix; and mapping the spatial context feature vector to a first query matrix and the temporal dependency feature vector to a second query matrix. The text feature vector is mapped to a second key matrix and a second value matrix; the dot product similarity between the second query matrix and the second key matrix is ​​calculated, and scaling and softmax normalization are performed to obtain the attention weight matrix of temporal to spatial; the second value matrix is ​​weighted and summed based on the attention weight matrix of temporal to spatial to generate a second intermediate fusion feature, which extracts the association preference between temporal and spatial features; the first intermediate fusion feature and the second intermediate fusion feature are concatenated and mapped to a unified dimension through a linear projection layer to generate a fusion feature representation. The deep interactive fusion of spatial and temporal information is achieved through a bidirectional cross-attention mechanism, which effectively improves the perception accuracy of load balancing decisions for multi-dimensional dynamic changes.

[0011] As another technical solution of the present invention, the step of intelligently decomposing and mapping the original service traffic according to the target load allocation coefficient to generate a multi-link allocation strategy based on the service flow granularity includes: parsing the packet header of the original service traffic, extracting the source IP address, destination IP address, port number and protocol type of each service flow, and generating a service flow five-tuple identifier; calculating the carrying ratio of each link according to the target load allocation coefficient, and determining the set of optional transmission paths for each service flow in combination with the traffic scheduling path matrix, thereby providing diversified path options for the service flow while ensuring load balancing; and calling a pre-trained flow classification model to perform feature encoding on the service flow five-tuple identifier. The code processing obtains the type label and service quality requirement level of the service flow, enabling fine-grained service quality intent perception of the service flow. Based on the type label and service quality requirement level, a target path that meets the minimum bandwidth requirement and maximum latency constraint is matched from the set of optional transmission paths for each service flow, ensuring the service quality of various services during transmission. Service flows of the same type with the same target path are aggregated and encapsulated to form link-level transmission units, generating a multi-link allocation strategy based on service flow granularity that includes service flow mapping relationships and path identifiers. Flow aggregation reduces scheduling signaling overhead and improves link utilization while meeting personalized service quality requirements.

[0012] Preferably, the step of calling the pre-trained flow classification model to perform feature encoding processing on the five-tuple identifier of the business flow to obtain the type label and service quality requirement level of the business flow includes: inputting the five-tuple identifier of the business flow into the embedding layer of the flow classification model for discrete feature sparse encoding processing to generate an embedding vector representation, converting high-dimensional discrete information into low-dimensional dense representation; inputting the embedding vector representation into the convolutional feature extraction layer of the flow classification model for one-dimensional convolution and pooling combination processing to extract local sequence features to capture local patterns of continuous port distribution or protocol combination; inputting the local sequence features into the temporal feature extraction layer of the flow classification model for bidirectional LSTM processing to extract long-distance temporal dependency features and fully explore the contextual relationships in the flow features; inputting the long-distance temporal dependency features into the classification output layer of the flow classification model, generating type labels through a fully connected layer and a multi-class softmax function, and generating service quality requirement levels through a fully connected layer and a regression function, thereby simultaneously and accurately predicting the business type and service quality requirements of the flow.

[0013] In the above technical solution, the spatial coding layer of the deep separable attention network uses 3×3 channel-wise convolutions and 1×1 pointwise convolutions to effectively reduce the number of computational parameters, balancing spatial feature extraction capabilities with lightweight deployment requirements. The temporal coding layer uses dilated causal convolutions to expand the receptive field while maintaining temporal causality, adapting to the temporal prediction characteristics of link state data. The stream classification model uses the cross-entropy loss function to train type labels and the mean squared error loss function to train service quality requirement levels. Multi-task joint training improves the model's generalization ability and prediction consistency.

[0014] This invention also provides a wireless communication multi-link load balancing and intelligent optimization system, including a processor and a memory connected to the processor. The memory stores programs, instructions, or code, and the processor executes the programs, instructions, or code in the memory to implement the aforementioned wireless communication multi-link load balancing and intelligent optimization method. This system integrates multi-dimensional link status awareness, spatiotemporal feature fusion, deep decision optimization, and service flow granular scheduling into one system, achieving efficient and accurate multi-link load balancing in complex wireless environments.

[0015] The technical effects and advantages provided by the present invention in the above technical solution are as follows: By inputting the temporal differential encoding of the channel state information sequence, the sliding window statistical features of the real-time load rate sequence, the one-hot encoding of the traffic flow type, and the spatial proximity matrix into a spatiotemporal graph convolutional network, node neighborhood feature aggregation and edge weight updates are performed on a dynamically optimized graph structure to generate a multi-link joint feature tensor that integrates inter-link interference coupling coefficients, load dynamic offset vectors, and traffic flow spatiotemporal distribution patterns. This feature tensor replaces traditional independent statistical indicators or simple concatenated vectors, simultaneously capturing the electromagnetic interference propagation path between links in the spatial dimension and the evolution trend of load state in the temporal dimension in the form of a single dense tensor. This allows the subsequent decision network to directly perceive the global correlation state of the multi-link system, eliminating decision bias caused by neglecting inter-link interference and load spatiotemporal coupling.

[0016] A deep separable attention network is employed to generate load decisions for multi-link joint feature tensors. The spatial coding layer of this network decouples channel-level and spatial-level features from the input tensor through a combination of channel-wise and pointwise convolutions, generating spatial context feature vectors with extremely low parameter overhead. The temporal coding layer utilizes a cascaded structure of dilated causal convolutions and gating units to capture long-range temporal dependencies with an exponentially expanded receptive field while maintaining strict temporal causality. The spatial and temporal feature vectors are then fused using multi-head cross-attention, bidirectionally querying the attention of spatial features to temporal features and vice versa. The concatenated vectors are then linearly projected to form a unified fused feature representation. This architecture decouples the spatial and temporal computation processes through deep separable convolutions and cross-attention mechanisms, significantly reducing the computational complexity of traditional 3D dense attention models. Furthermore, the bidirectional cross-fusion mechanism strengthens the feature interaction between service flow type preferences and path latency and bandwidth constraints. The final output target load allocation coefficients and traffic scheduling path matrix can match each service flow with a target transmission path that meets its quality of service requirements. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1 This is a flowchart of a method for multi-link load balancing and intelligent optimization in wireless communication. Figure 2 This is a flowchart of the multi-link joint feature tensor generation process; Figure 3 It is a flowchart for generating multi-link allocation strategies based on business flow granularity. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] See Figure 1 This invention provides a method for multi-link load balancing and intelligent optimization in wireless communication. The method includes: collecting multi-modal link status data of a target communication terminal under a multi-link concurrent transmission environment; performing spatiotemporal correlation analysis on the multi-modal link status data to generate a multi-link joint feature tensor, wherein the multi-link joint feature tensor includes inter-link interference coupling coefficients, load dynamic offset vectors, and spatiotemporal distribution patterns of service flows; calling a pre-trained deep separable attention network to perform load balancing decision generation processing on the multi-link joint feature tensor, outputting a target load allocation coefficient and a traffic scheduling path matrix for each link; and performing intelligent decomposition and mapping processing on the original service traffic based on the target load allocation coefficients to generate a multi-link allocation strategy based on service flow granularity.

[0021] Example 1: In specific implementation, please refer to Figure 2 The process of performing spatiotemporal correlation analysis on multimodal link state data to generate a multi-link joint feature tensor is as follows.

[0022] The collected multimodal link state data includes channel state information sequences, real-time load rate sequences, service flow type identifiers, and geospatial coordinates for each link. The channel state information sequences undergo temporal differential coding, subtracting channel state information values ​​at adjacent times to obtain a difference sequence. This difference sequence constitutes the channel quality fluctuation feature vector for each link. The real-time load rate sequences are processed using sliding window statistical feature extraction. The sliding window length and step size are pre-defined. Within each sliding window, the mean and variance of the real-time load rate sequences are calculated, and the ratio of the maximum load rate to the mean load rate within the window is extracted as a peak value. The mean, variance, and peak value are concatenated to generate a multidimensional load feature vector for each link. Service flow type identifiers are mapped to one-hot encoded vectors using a lookup table. The dimension of the one-hot encoded vector equals the preset total number of service flow types. Positions in the vector corresponding to the service flow type identifier are set to 1, and all other positions are set to 0. A spatial proximity matrix is ​​constructed by combining geospatial coordinates. The element in the i-th row and j-th column of the spatial proximity matrix represents the spatial proximity between the i-th link and the j-th link. The spatial proximity is determined by a Gaussian decay function with the Euclidean distance between geospatial coordinates as the variable. The scale parameter of the Gaussian decay function is set according to the average Euclidean distance between the geospatial coordinates of all links, and is set to 1.5 times the average Euclidean distance.

[0023] The channel quality fluctuation feature vector, multidimensional load feature vector, one-hot encoded vector, and spatial proximity matrix are input into a pre-defined spatiotemporal graph convolutional network for node feature aggregation and edge weight update, outputting a multi-link joint feature tensor that integrates spatiotemporal correlation information. The processing procedure of the spatiotemporal graph convolutional network is as follows: each link is used as a graph node, and the geographical distance and signal-to-interference ratio (SIR) between links are used as initial edge weights to construct an initial link topology graph. The initial edge weights are obtained by weighted summation of geographical proximity and normalized SIR values. The geographical proximity is obtained by Gaussian attenuation calculation of geographical distance, and the normalized SIR value is obtained by mapping the SIR to the (0,1) interval using a sigmoid function. The weight coefficient used for the weighted summation is set to 0.5 to balance the contribution of geographical distance and SIR factors to the initial edge weights. For each node in the initial link topology graph, the channel quality fluctuation feature vector and multidimensional load feature vector of its first-order neighbor nodes are aggregated to generate aggregated node neighborhood features. The aggregation method involves calculating a weighted average of the channel quality fluctuation feature vectors of the first-order neighbor nodes based on the initial edge weights, and calculating a weighted average of the multidimensional load feature vectors of the first-order neighbor nodes based on the initial edge weights. The two weighted averages are then concatenated to form the aggregated node neighborhood features. The node's own features are updated based on the one-hot encoded vector. The update operation concatenates the channel quality fluctuation feature vector, multidimensional load feature vector, and one-hot encoded vector from the original node features, and then feeds them into a fully connected layer to map to a preset unified feature dimension, generating updated node features. The edge weights between adjacent nodes are dynamically adjusted using the spatial proximity matrix to generate an optimized graph structure that incorporates business type preferences. The adjustment method is as follows: for node pairs with edge connections, the cosine similarity of the one-hot encoded vectors of the two nodes is calculated. This cosine similarity is multiplied by the element value corresponding to that node pair in the spatial proximity matrix to obtain an adjustment coefficient. The initial edge weights are then multiplied by the adjustment coefficient, and the product is used as the optimized edge weights. The optimized graph structure and the updated node features are input into a multi-layer graph convolutional layer for iterative message passing and feature updating. The total number of graph convolutional layers is set to 3 to aggregate link information within the 3rd-order neighbor range. The feature update for each node v in the k-th layer graph convolution is performed according to the following formula: in, Represents a node In the Feature vectors output by the layer; Representing neighboring nodes The initial input feature vector; Represents a node The set of first-order neighbor nodes in an optimized graph structure; Represents a node With nodes The edge weights are dynamically adjusted. For the first The learnable weight matrix of the layer graph convolution has dimensions of , For the first The dimension of the layer feature vector; For the first The learnable bias vector of the layer graph convolution has a dimension of . ; To modify the activation function of the linear unit; The values ​​are 1, 2, and 3. After three layers of graph convolution iteration, the output feature vector of each node in the third layer is obtained. The output feature vectors of all links are arranged according to the link index to form a feature matrix, and a multi-link joint feature tensor is constructed based on the feature matrix. The multi-link joint feature tensor contains the inter-link interference coupling coefficient, the load dynamic offset vector, and the spatiotemporal distribution pattern of the service flow. The inter-link interference coupling coefficient is determined by the edge weights in the optimized graph structure. The load dynamic offset vector is represented by the corresponding components of the multidimensional load feature vector after multiple rounds of propagation, and the spatiotemporal distribution pattern of the business flow is characterized by the aggregation results of node features in the spatial and temporal dimensions.

[0024] Example 2: In practice, a pre-trained deep separable attention network is invoked to perform load balancing decision generation on the multi-link joint feature tensor. The deep separable attention network consists of a spatiotemporal dimension separation module, a spatial encoding layer, a temporal encoding layer, a multi-head cross-attention fusion module, and a decision output layer.

[0025] The spatiotemporal dimension separation module receives the multi-link joint feature tensor and separates it into a spatial attention feature map and a temporal attention feature map. The dimension of the multi-link joint feature tensor is denoted as... ,in Indicates the total number of links. Indicates the length of the time series. This represents the number of feature channels for each link at each time step. The spatiotemporal dimension separation module segments the joint feature tensor of multiple links along the time dimension, obtaining... Each size is A two-dimensional matrix, where each two-dimensional matrix corresponds to the spatial feature distribution of all links at a given time step, will... The two-dimensional matrices are arranged in chronological order to form a spatial attention feature map. Simultaneously, the spatiotemporal dimension separation module segments the multi-link joint feature tensor along the link dimension, resulting in... Each size is A two-dimensional matrix, each two-dimensional matrix corresponding to the temporal feature evolution of a link at all time steps, will... A time attention feature map is formed by arranging two-dimensional matrices according to the link index.

[0026] The spatial attention feature map is input into the spatial encoding layer of a depthwise separable attention network. The spatial encoding layer consists of a cascaded 3×3 channel-wise convolutional layer and a 1×1 pointwise convolutional layer. The 3×3 channel-wise convolutional layer contains the same number of 3×3 convolutional kernels as the feature channels of the spatial attention feature map. Each kernel operates on only one corresponding channel of the spatial attention feature map, independently extracting local spatial patterns within each channel and outputting an intermediate feature map that maintains the same number of channels. The 1×1 pointwise convolutional layer contains a predetermined number of 1×1 convolutional kernels. The 1×1 pointwise convolutional layer linearly combines all channels of the intermediate feature map to achieve cross-channel information interaction and output a spatial context feature vector. The number of 1×1 convolutional kernels is set to a predetermined value. Then the dimension of the spatial context feature vector is .

[0027] The temporal attention feature map is input into the temporal coding layer of a deep separable attention network. The temporal coding layer consists of a dilated causal convolutional layer connected in series with gated linear units. The dilated causal convolutional layer uses a one-dimensional dilated convolutional kernel with a kernel length of 3 and a dilation factor sequence of [1, 2, 4]. The convolution operation is forward-dependent only in the temporal dimension, ensuring that the output at the current time step does not contain information about future time steps. The output of the dilated causal convolutional layer is divided into two equal parts in the channel dimension. The first part is used to generate a gated signal via a sigmoid activation function, and the second part is used to generate a candidate signal via a tanh activation function. The gated signal and the candidate signal are multiplied element-wise to output a temporally dependent feature vector. The dimension of the temporally dependent feature vector is... .

[0028] The spatial context feature vector and temporal dependency feature vector are input into the multi-head cross-attention fusion module for multi-head cross-attention fusion processing to generate a fused feature representation. The multi-head cross-attention fusion module is configured with 8 attention heads. For the ... The attention head processes the following steps: inputting the spatial context feature vector into the first linear mapping layer to generate the first query matrix. The time-dependent feature vectors are input into the second linear mapping layer to generate the first key matrix. and the first value matrix Both the first and second linear mapping layers are fully connected layers, which do not change the feature dimensions. The first query matrix is ​​then calculated. With the first key matrix The scaled dot product attention is used to obtain the spatial attention weight matrix for temporal relationships. The calculation method is as follows: in, Indicates the first The attention weight matrix for spatial pairs in each attention head has dimensions of . ; Indicates the first The first query matrix, obtained by mapping spatial context feature vectors from each attention head, has dimensions of [missing information]. ; Indicates the first The first key matrix, obtained by mapping temporally dependent feature vectors in each attention head, has dimensions of [missing information]. ; Indicates the first The first value matrix obtained by mapping temporally dependent feature vectors in each attention head; This represents the dimension of the query vector and key vector in each attention head, with values ​​ranging from 1 to 2. ; This represents the softmax normalization function performed along the last dimension; This represents the transpose of the first bond matrix. Based on right Perform a weighted summation to generate the first... The first intermediate fusion feature of each attention head, the dimension of the first intermediate fusion feature is: .

[0029] The time-dependent feature vectors are input into the third linear mapping layer to generate the second query matrix. The spatial context feature vector is input into the fourth linear mapping layer to generate the second key matrix. Second-value matrix The attention weight matrix for the temporal pairs is calculated using the same method: in, Indicates the first The attention weight matrix for temporal pairs in each attention head has dimensions of . ; Indicates the first The second query matrix, obtained by mapping temporally dependent feature vectors in each attention head, has dimensions of [missing information]. ; Indicates the first The second key matrix, obtained by mapping the spatial context feature vector in each attention head, has dimensions of [missing information]. ; Indicates the first The second value matrix obtained by mapping the spatial context feature vectors in each attention head. Based on right Perform a weighted summation to generate the first... The second intermediate fusion feature of each attention head, the dimension of the second intermediate fusion feature is... .

[0030] The first intermediate fusion features generated by all eight attention heads are concatenated in the last dimension to obtain the dimension. The first fusion feature; the second intermediate fusion features generated by all 8 attention heads are concatenated in the last dimension to obtain the dimension. The second fusion feature. The first and second fusion features are then summed element-wise and fed into a linear projection layer, which maps the summation result to a dimension. This generates a fused feature representation. The fused feature representation retains the joint information of the spatial and temporal dimensions, with the first dimension being... .

[0031] The fused feature representation is used as the input to the decision output layer of a depthwise separable attention network. The decision output layer contains a load distribution branch and a path selection branch. The load distribution branch consists of two fully connected layers: the first fully connected layer contains 256 neurons and uses the ReLU activation function, and the second fully connected layer has the same number of neurons as the total number of links. The output value is normalized using softmax to generate the target load allocation coefficient, which has a length of [value missing]. The vector has components that sum to 1. The path selection branch includes a path selection matrix generation module, which consists of three fully connected layers. The first two fully connected layers contain 512 and 256 neurons respectively and both use the ReLU activation function. The third fully connected layer reshapes the output into a dimensionless vector. The three-dimensional tensor, in which Indicates the total number of service flows to be scheduled. This represents the maximum number of candidate paths on each link; softmax normalization is applied to the last dimension to obtain the traffic scheduling path matrix, where each element represents the probability of a service flow choosing a candidate path on a link.

[0032] The deep separable attention network undergoes pre-training before deployment. Historical link operation data is collected to construct a training sample set. Each training sample consists of an input multi-link joint feature tensor, labeled target load allocation coefficients, and a labeled traffic scheduling path matrix. The labeled target load allocation coefficients are determined by the link load ratio calculated based on the optimal traffic allocation algorithm, and the labeled traffic scheduling path matrix is ​​a path allocation matrix that matches the labeled load allocation coefficients. A combined loss function is then constructed. , ,in This represents the Kullback-Leibler divergence loss between the predicted target load distribution coefficient and the labeled target load distribution coefficient. This represents the cross-entropy loss between the predicted values ​​of the traffic scheduling path matrix and the labeled traffic scheduling path matrix. and The loss weights are all set to 1.0. The Adam optimizer is used to update the learnable parameters of the deep separable attention network, with an initial learning rate of 0.001. Each batch of data contains 32 training samples, and the training process iterates for 300 rounds until the combined loss function converges. The network parameters are then saved to obtain the pre-trained deep separable attention network. During the online inference phase, the multi-link joint feature tensor generated in real time is input into the pre-trained deep separable attention network. Through the sequential processing of the above modules, the target load allocation coefficients and traffic scheduling path matrix for each link are output.

[0033] Example 3: In specific implementation, please refer to Figure 3 The process of intelligently decomposing and mapping the original service traffic according to the target load allocation coefficient to generate a multi-link allocation strategy based on the service flow granularity is as follows.

[0034] The raw traffic is parsed packet by packet, and the header fields of each packet are read to extract five fields: source IP address, destination IP address, source port number, destination port number, and transport layer protocol type. These five fields are combined to form a five-tuple identifier for the traffic flow. For IPv4 packets, the source IP address and destination IP address fields are each 32 bits long; for IPv6 packets, the source IP address and destination IP address fields are each 128 bits long; the source port number and destination port number fields are each 16 bits long; and the protocol type field is 8 bits long. Packets belonging to the same five-tuple identifier for the traffic flow are grouped into one traffic flow.

[0035] Obtain the target load allocation coefficients and traffic scheduling path matrix output by the deep separable attention network. The vector, the first in the vector Each component Indicates the first The normalized traffic capacity ratio allocated to each link. The range of values ​​is ,all The sum of these is always 1. Multiply by the total rate of the original business traffic to get the first... The maximum carrying rate of the link, the first The carrying ratio of each link is The dimension of the traffic scheduling path matrix is... ,in The total number of service flows to be scheduled. The number of candidate paths pre-configured for each link. For the first link... The first business flow, the first in the traffic scheduling path matrix line, number The elements of the column are a column of length . The probability vector, the probability vector in which the th probability vector is... The component represents the first... The first business flow selection The first link The probability of the candidate path. Extract the indices of paths with a probability greater than 0 from the traffic scheduling path matrix to form the... The business flow in the first A set of optional transmission paths on a link.

[0036] The pre-trained flow classification model is used to encode features for each business flow quintuple identifier. The pre-trained flow classification model consists of an embedding layer, a convolutional feature extraction layer, a temporal feature extraction layer, and a classification output layer connected sequentially. The embedding layer receives the discrete numerical values ​​of the business flow quintuple identifier and maps them to a dense vector representation. The convolutional feature extraction layer performs one-dimensional convolution and pooling operations on the dense vector representation to extract local sequence features. The temporal feature extraction layer processes the local sequence features using a bidirectional long short-term memory recurrent network to obtain long-distance temporal dependency features. The classification output layer contains two parallel output branches: the first output branch consists of a fully connected layer and a multi-class softmax function, outputting the business flow type label; the second output branch consists of a fully connected layer and a linear regression function, outputting the service quality demand level. The pre-trained flow classification model is trained on historical labeled business flow data. During training, the cross-entropy loss function is used to optimize type label prediction, and the mean squared error loss function is used to optimize service quality demand level prediction. The weighted sum of the two loss functions is used as the overall optimization objective, with both weight coefficients set to 1.0. After inputting the business flow quintuple identifier into the pre-trained flow classification model, the type label of the business flow is obtained from the first output branch. The type label is one of the preset business categories, which includes five types: voice service, video streaming service, web browsing service, file transfer service, and real-time interactive service. The service quality requirement level is obtained from the second output branch. The service quality requirement level is a continuous real value between 0 and 1. The higher the value, the more stringent the service quality requirement.

[0037] After obtaining the type label and service quality requirement level for each service flow, the minimum bandwidth requirement and maximum latency constraint are determined for each service flow. A predefined service type service quality mapping table is consulted based on the service flow's type label. The predefined table sets the minimum bandwidth requirement at 64 kilobits per second and the maximum latency constraint at 150 milliseconds for voice services; 2 megabits per second and 400 milliseconds for video streaming services; 500 kilobits per second and 2 seconds for web browsing services; 1 megabit per second and 5 seconds for file transfer services; and 256 kilobits per second and 200 milliseconds for real-time interactive services. The minimum bandwidth requirement obtained from the table lookup is multiplied by the service quality requirement level to obtain the adjusted minimum bandwidth requirement; the maximum latency constraint obtained from the table lookup is divided by the service quality requirement level to obtain the adjusted maximum latency constraint. If the service quality requirement level is 0.8, the adjusted minimum bandwidth requirement increases to 1.25 times the lookup value, and the adjusted maximum latency constraint is shortened to 0.8 times the lookup value; if the service quality requirement level is 0.5, the adjusted minimum bandwidth requirement is 1.0 times the lookup value, and the adjusted maximum latency constraint is 1.0 times the lookup value.

[0038] For each service flow, a target path is determined from the set of available transmission paths based on the adjusted minimum bandwidth requirement and the adjusted maximum latency constraint. The current available bandwidth and estimated end-to-end latency of each available transmission path are queried. The current available bandwidth is obtained by subtracting the allocated bandwidth usage of the service flow from the total link bandwidth. The estimated end-to-end latency is calculated by summing the link propagation latency, the predicted queue latency, and the processing latency of the forwarding nodes in the path. Available transmission paths with current available bandwidth less than the adjusted minimum bandwidth requirement of the service flow are excluded. Then, available transmission paths with estimated end-to-end latency greater than the adjusted maximum latency constraint of the service flow are excluded from the remaining available transmission paths. If multiple candidate paths still exist after exclusion, a path score is calculated for each candidate path. in, Indicates the first The scores for each candidate path range from [value range missing]. ; Indicates the first The current available bandwidth of the candidate paths, in bits per second; This represents the maximum available bandwidth among all candidate paths. Indicates the first The estimated end-to-end latency of the candidate paths, in milliseconds; This represents the maximum estimated end-to-end delay among all candidate paths; The bandwidth weighting factor is set to 0.7 to reflect the higher priority given to bandwidth over latency in path selection. (Score selection) The candidate path with the highest value is selected as the target path.

[0039] All service flows are grouped according to the link identifier and path identifier of the target path. Service flows with the same target path identifier and the same type label are grouped into the same transmission aggregation group. Service flow data packets within each transmission aggregation group are encapsulated, and a link-level transmission unit header is added to the outer layer of the data packet. The link-level transmission unit header contains a link identifier field, a path identifier field, a service flow type label field, and aggregated service flow count information. The aggregated and encapsulated link-level transmission units are transmitted through the corresponding links. The final multi-link allocation strategy based on service flow granularity includes service flow mapping relationships and path identifiers. The service flow mapping relationship records the correspondence between each original service flow five-tuple identifier and its mapped target path identifier. The path identifier includes a combination of link index and candidate path index.

[0040] Example 4: In practice, the specific process of calling the pre-trained flow classification model to perform feature encoding on the business flow quintuple identifier is as follows.

[0041] The pre-trained flow classification model consists of four sequentially connected parts: an embedding layer, a convolutional feature extraction layer, a temporal feature extraction layer, and a classification output layer. The embedding layer receives five fields from the traffic flow's quintuple identifier: source IP address, destination IP address, source port number, destination port number, and protocol type. The embedding layer maintains an independent embedding matrix for each of the five fields. The source IP address embedding matrix has the following dimensions: , This represents the total number of possible values ​​for the source IP address, for IPv4 addresses. For IPv6 addresses, the value is 2 to the power of 32. It is 2 to the power of 128. The dimension of the IP address embedding vector is set to 32; the dimension of the destination IP address embedding matrix is... , Indicates the total number of possible values ​​for the destination IP address, and The values ​​are the same. Set to 32; the dimension of the source port number embedding matrix is... , This indicates the total number of possible values ​​for the port number, with a possible value of 65536. The dimension of the port number embedding vector is set to 16; the dimension of the destination port number embedding matrix is... , Set to 16; the dimension of the protocol type embedding matrix is... , This indicates the total number of possible values ​​for the protocol type, with a maximum value of 256. The dimension of the protocol type embedding vector is set to 8. The embedding layer maps each discrete field value in the business flow 5-tuple identifier to a corresponding dense embedding vector using a lookup table. The five embedding vectors are then concatenated sequentially to generate an embedding vector representation, the dimension of which is... .

[0042] The embedding vector representation is used in the convolutional feature extraction layer of the input stream classification model. The convolutional feature extraction layer contains three parallel one-dimensional convolutional branches with kernel lengths of 3, 5, and 7, respectively, to capture local combination patterns of quintuple fields at different granularities. Each one-dimensional convolutional branch contains a one-dimensional convolutional layer and a max-pooling layer. The one-dimensional convolutional layer has 1 input channel and treats the embedding vector representation as a one-dimensional sequence of length 104. In the first one-dimensional convolutional branch, the kernel length is 3, the output channels are 64, the stride is 1, and the padding consists of one zero at each end of the sequence. In the second one-dimensional convolutional branch, the kernel length is 5, the output channels are 64, the stride is 1, and the padding consists of two zeros at each end of the sequence. In the third one-dimensional convolutional branch, the kernel length is 7, the output channels are 64, the stride is 1, and the padding consists of three zeros at each end of the sequence. The outputs of the three one-dimensional convolutional layers are each activated by ReLU and then fed into their respective max-pooling layers. The pooling window length and stride of each of the three max-pooling layers are set to 2. The output vectors of the three max-pooling layers are concatenated along the channel dimension to obtain the local sequence features, which have the following dimensions: ,in This indicates the length of the sequence after pooling, and 192 is the sum of the 64 output channels of the three branches.

[0043] The local sequence features are input into the temporal feature extraction layer of the stream classification model. The temporal feature extraction layer employs a bidirectional LSTM structure, containing one forward LSTM sublayer and one backward LSTM sublayer. The forward LSTM sublayer is processed step-by-step from... arrive Process the local sequence features sequentially. Feature vectors at each time step; backward LSTM sublayers from... arrive In reverse processing of local sequence features, the first The feature vectors for each time step are set. The number of hidden units in both the forward and backward LSTM sub-layers is set to 128. The forward LSTM sub-layer outputs a forward hidden state vector of dimension 128 at each time step; the backward LSTM sub-layer outputs a backward hidden state vector of dimension 128 at each time step. The forward and backward hidden state vectors at the same time step are concatenated to obtain a bidirectional hidden state vector of dimension 256. The bidirectional hidden state vector from the last time step is extracted as the long-range temporal dependency feature, which has a dimension of 256.

[0044] The classification output layer of the stream classification model is input with long-range temporal dependency features. The classification output layer contains a first output branch and a second output branch, both sharing the long-range temporal dependency features. The first output branch consists of a first fully connected layer and a first activation function. The first fully connected layer has an input dimension of 256 and an output dimension of 5. The output dimension of 5 corresponds to the total number of preset service categories: voice service, video streaming service, web browsing service, file transfer service, and real-time interactive service. The first activation function is a multi-class softmax function, which normalizes the 5-dimensional vector output by the first fully connected layer into a probability distribution. Each component in the probability distribution represents the predicted probability that the service stream belongs to the corresponding category, and the index of the maximum probability is taken as the output type label. The second output branch consists of a second fully connected layer and a second activation function. The second fully connected layer has an input dimension of 256 and an output dimension of 1. The second activation function is the sigmoid function, which maps the scalar value output by the second fully connected layer to the (0,1) interval. The mapped value is used as the service quality requirement level. The closer the service quality requirement level is to 1, the stricter the service quality requirement is, and the closer the value is to 0, the more lenient the service quality requirement is.

[0045] The training process of the pre-trained flow classification model is as follows: Historical network traffic data is collected and manually labeled to determine the five-tuple identifier for each business flow, along with its corresponding true type label and true service quality requirement level. The labeled data is divided into training and validation datasets in an 8:2 ratio. During the training initialization phase, the embedding matrix parameters of the embedding layers are randomly initialized using a normal distribution with a mean of 0 and a standard deviation of 0.1. The weight parameters of all fully connected layers and LSTM units in the convolutional and temporal feature extraction layers are initialized using a Xavier uniform distribution. The bias parameters of all layers are initialized to 0. The training loss function of the flow classification model consists of two parts: the first part is the type label prediction loss, calculated using the cross-entropy loss function. The first part is the cross-entropy between the one-hot encoded vector of the true type label and the softmax output probability distribution; the second part is the service quality demand level prediction loss, which is calculated using the mean squared error loss function. The calculation method is the square of the difference between the predicted service quality demand level and the actual service quality demand level. The overall loss function is: in, The overall loss function value of the stream classification model is represented by a scalar. The cross-entropy loss value representing the type label prediction is a scalar. The mean squared error loss value representing the forecast of service quality demand level is a scalar. The balance coefficient between the two losses is set to 0.5 to ensure that the type label prediction loss and the service quality demand level prediction loss contribute equally to the total loss. The training process uses the Adam optimizer with an initial learning rate of 0.0005. Each batch contains 64 training samples, and the training is performed for 200 iterations. After each iteration, the validation loss is calculated using the validation dataset. When the validation loss no longer decreases for 10 consecutive iterations, the learning rate is reduced to 0.5 times the current rate. When the validation loss no longer decreases for 20 consecutive iterations, training is terminated, and the model parameters with the lowest validation loss are saved as the pre-trained stream classification model parameters.

[0046] Example 5: In practice, the spatial and temporal coding layers of the deep separable attention network adopt specific convolutional structures, and the stream classification model is trained using a specific loss function. The above configuration is described in detail below.

[0047] The spatial encoding layer of the deep separable attention network consists of a cascaded 3×3 channel-wise convolutional layer and a 1×1 pointwise convolutional layer. The 3×3 channel-wise convolutional layer receives the spatial attention feature map as input, and the spatial attention feature map has a dimension of 1. ,in Indicates the length of the time series. This indicates the number of feature channels. The number of 3×3 channel-wise convolutional layers configured is... The convolutional kernels are 3×3 in size, and each kernel operates on only one corresponding channel in the spatial attention feature map. The stride of the 3×3 channel-wise convolution operation is set to 1, and the padding method is set to pad outward by 1 pixel at each of the height and width edges of the spatial attention feature map, with the padding pixel value uniformly set to 0 to ensure that the spatial dimensions remain unchanged before and after the convolution operation. The output of the 3×3 channel-wise convolutional layer is preserved. The intermediate spatial feature map of each channel. A 1×1 point convolutional layer receives the intermediate spatial feature map as input, and the 1×1 point convolutional layer configuration is... There are 1×1×1 convolutional kernels, each with a size of 1×1×1. Where 1×1 represents the size of the convolution kernel in the spatial dimension. This indicates that the number of channels in the convolutional kernel perfectly matches the number of channels in the intermediate spatial feature map. The stride is set to 128 to provide sufficient spatial feature representation. The stride of the 1×1 point convolution operation is set to 1, with no padding. The 1×1 point convolution layer handles all intermediate spatial feature maps. The pixel values ​​of each channel at the same spatial location are linearly weighted and combined, and cross-channel information interaction is performed independently at each 1×1 spatial location. The output of the 1×1 point convolutional layer is a spatial context feature vector, and the dimension of the spatial context feature vector is... .

[0048] The temporal encoding layer of a deep separable attention network consists of a dilated causal convolutional layer and gated linear units cascaded together. The dilated causal convolutional layer receives a temporal attention feature map as input, and the temporal attention feature map has a dimension of [missing information]. ,in Indicates the total number of links. This represents the number of feature channels. The dilated causal convolutional layer uses a one-dimensional dilated convolution operation with a kernel length of 3 and a dilation factor sequence of [1, 2, 4]. Three layers of dilated causal convolutions are stacked, with dilation factors of 1, 2, and 4 for each layer. The convolutional layer with a dilation factor of 1 provides a local receptive field of length 3; the convolutional layer with a dilation factor of 2 provides an effective receptive field of length 5 by spacing adjacent sampling points within the receptive field by 1 time step; and the convolutional layer with a dilation factor of 4 provides an effective receptive field of length 9 by spacing adjacent sampling points within the receptive field by 3 time steps. The number of kernels in all three dilated causal convolutional layers is set to 256, and the stride is set to 1. To ensure causality, the one-dimensional dilated convolution operation is padded only in the initial direction of the time dimension. The padding amount is equal to the product of the dilation factor and the kernel length minus 1, and the padding amount is calculated layer by layer according to the dilation factor sequence. The output dimension of the dilated causal convolutional layer is... The gated linear unit (GLU) divides the output of the dilated causal convolutional layer into two equal parts along the channel dimension: the first part contains 128 channels, and the second part contains 128 channels. A sigmoid activation function is applied to the first part to output a gated signal, with each component ranging from (0,1). A tanh activation function is applied to the second part to output a candidate signal, with each component ranging from (-1,1). The gated signal and the candidate signal are then multiplied element-wise to obtain the output of the GLU, which is the time-dependent feature vector. The dimension of the time-dependent feature vector is... .

[0049] The streaming classification model is trained using a composite loss function, which is a weighted combination of the cross-entropy loss function and the mean squared error loss function. The cross-entropy loss function is used to train the type label prediction branch, measuring the difference between the probability distribution of the type labels predicted by the streaming classification model and the one-hot encoding distribution of the true type labels. Let the total number of samples in the training sample set be... For the first The training samples and their true type labels are processed by one-hot encoding to form a dataset of length [length missing]. The vector, Indicates the total number of preset business categories. The value is 5. In the one-hot encoded vector, the element corresponding to the true type position has a value of 1, and the other positions have a value of 0. Let... Indicates the first Training samples in category The actual label value on the screen. The softmax function representing the first output branch of the stream classification model is used to represent the first... Each training sample outputs its category. The predicted probability value. The formula for calculating the cross-entropy loss function is: in, The value of the cross-entropy loss function is a scalar. This indicates the number of samples in the training sample set that participate in a single loss calculation; Indicates the total number of preset business categories. The value is 5; Indicates the first Training samples in category The actual label value on the screen can be either 0 or 1. This indicates that the first output branch of the stream classification model is related to the first... Training samples in category The predicted probability value is in the range (0,1); Let be the numerical stability constant, set to . This is used to avoid the case where the value is 0 when calculating logarithms; Let represent the logarithmic function with the natural constant e as the base. The mean squared error loss function is used to train the service quality demand level prediction branch. The mean squared error loss function measures the squared deviation between the service quality demand level predicted by the flow classification model and the actual service quality demand level. Let Indicates the first The actual service quality demand level value of each training sample is in the range of [0,1]. This indicates that the second output branch of the flow classification model is related to the first... The predicted service quality demand level value output by each training sample takes a value in the range (0,1). The formula for calculating the mean squared error loss function is: in, The value of the mean squared error loss function is a scalar. Indicates the first The true service quality demand level values ​​of each training sample; This indicates that the second output branch of the flow classification model is related to the first... The predicted service quality demand level value output from each training sample. The overall training loss function of the stream classification model is a weighted combination of the cross-entropy loss function and the mean squared error loss function, with weighting coefficients set to 0.5 and 0.5 respectively, to ensure a balanced optimization effort between the type label prediction task and the service quality demand level prediction task during training. The expression for the overall training loss function is as follows: During training, a mini-batch stochastic gradient descent algorithm is used, with a minimum number of samples per batch. Set the value to 64, use the Adam optimizer, set the initial learning rate to 0.0005, and perform 200 iterations to update all learnable parameters of the stream classification model.

[0050] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for multi-link load balancing and intelligent optimization in wireless communication, characterized in that, The method includes: Collect multimodal link status data of the target communication terminal in a multi-link concurrent transmission environment; The multimodal link status data is subjected to spatiotemporal correlation analysis to generate a multi-link joint feature tensor, which includes the inter-link interference coupling coefficient, load dynamic offset vector, and service flow spatiotemporal distribution pattern. A pre-trained deep separable attention network is invoked to perform load balancing decision generation on the multi-link joint feature tensor, outputting the target load allocation coefficient and traffic scheduling path matrix for each link. Based on the target load allocation coefficient, the original service traffic is intelligently decomposed and mapped to generate a multi-link allocation strategy based on service flow granularity.

2. The wireless communication multi-link load balancing and intelligent optimization method according to claim 1, characterized in that, The step of performing spatiotemporal correlation analysis on the multimodal link state data to generate a multi-link joint feature tensor includes: The multimodal link status data includes a sequence of channel status information, a real-time load rate sequence, a service flow type identifier, and geospatial location coordinates for each link. The channel state information sequence is subjected to time-domain differential coding to obtain the channel quality fluctuation feature vector for each link; The real-time load rate sequence is subjected to sliding window statistical feature extraction processing to generate a multi-dimensional load feature vector with mean, variance and peak distribution of load rate; The business flow type identifier is mapped to a one-hot encoded vector, and a spatial proximity matrix is ​​constructed by combining the geospatial location coordinates; The channel quality fluctuation feature vector, the multidimensional load feature vector, the one-hot encoded vector, and the spatial proximity matrix are input into a preset spatiotemporal graph convolutional network for node feature aggregation and edge weight update processing, and the multi-link joint feature tensor with spatiotemporal correlation information is output.

3. The wireless communication multi-link load balancing and intelligent optimization method according to claim 2, characterized in that, The step of inputting the channel quality fluctuation feature vector, the multidimensional load feature vector, the one-hot encoded vector, and the spatial proximity matrix into a preset spatiotemporal graph convolutional network for node feature aggregation and edge weight update processing includes: Using each link as a graph node, and the geographical distance and signal interference ratio between links as initial edge weights, an initial link topology graph is constructed. For each node in the initial link topology graph, the channel quality fluctuation feature vector and multidimensional load feature vector of its first-order neighbor nodes are aggregated to generate the aggregated node neighborhood feature. The node's own features are updated based on the one-hot encoding vector, and the edge weights between adjacent nodes are dynamically adjusted in conjunction with the spatial proximity matrix to generate an optimized graph structure that includes business type preferences. The optimized graph structure and corresponding node features are input into multiple layers of graph convolutional layers for iterative message passing and feature updating, and the multi-link joint feature tensor is output.

4. The wireless communication multi-link load balancing and intelligent optimization method according to claim 1, characterized in that, The process of calling a pre-trained deep separable attention network to perform load balancing decision generation on the multi-link joint feature tensor, and outputting the target load allocation coefficient and traffic scheduling path matrix for each link, including: The spatiotemporal dimension separation process is performed on the multi-link joint feature tensor to generate spatial attention feature maps and temporal attention feature maps; The spatial attention feature map is input into the spatial coding layer of the depthwise separable attention network and processed by a combination of channel-wise convolution and point convolution to output a spatial context feature vector. The temporal attention feature map is input into the temporal coding layer of the depthwise separable attention network for causal convolution and gating unit combination processing, and the temporal dependent feature vector is output. Multi-head cross-attention fusion processing is performed on the spatial context feature vector and the temporal dependency feature vector to generate a fused feature representation; The fused feature representation is input into the decision output layer of the deep separable attention network, and the target load allocation coefficient is generated through a fully connected layer and softmax normalization. At the same time, the traffic scheduling path matrix is ​​output through the path selection matrix generation module.

5. The wireless communication multi-link load balancing and intelligent optimization method according to claim 4, characterized in that, The step of performing multi-head cross-attention fusion processing on the spatial context feature vector and the temporal dependency feature vector to generate a fused feature representation includes: The spatial context feature vector is mapped to a first query matrix, and the temporal dependency feature vector is mapped to a first key matrix and a first value matrix; Calculate the dot product similarity between the first query matrix and the first key matrix, and perform scaling and softmax normalization to obtain the spatial attention weight matrix for temporal relationships. The first value matrix is ​​weighted and summed based on the attention weight matrix of the spatial time series to generate the first intermediate fusion feature; The temporal dependency feature vector is mapped to a second query matrix, and the spatial context feature vector is mapped to a second key matrix and a second value matrix; Calculate the dot product similarity between the second query matrix and the second key matrix, and perform scaling and softmax normalization to obtain the attention weight matrix of the temporal pair space. The second value matrix is ​​weighted and summed based on the attention weight matrix of the temporal sequence to generate a second intermediate fusion feature; The first intermediate fusion feature and the second intermediate fusion feature are concatenated and mapped to a unified dimension through a linear projection layer to generate the fusion feature representation.

6. The wireless communication multi-link load balancing and intelligent optimization method according to claim 1, characterized in that, The step of intelligently decomposing and mapping the original service traffic according to the target load allocation coefficient to generate a multi-link allocation strategy based on service flow granularity includes: Parse the packet header of the original service traffic, extract the source IP address, destination IP address, port number and protocol type of each service flow, and generate a five-tuple identifier for the service flow; The carrying ratio of each link is calculated based on the target load allocation coefficient, and the set of optional transmission paths for each service flow is determined in combination with the traffic scheduling path matrix. The pre-trained flow classification model is invoked to perform feature encoding on the five-tuple identifier of the business flow to obtain the type label and service quality requirement level of the business flow; Based on the type label and the service quality requirement level, a target path that meets the minimum bandwidth requirement and maximum latency constraint is matched from the set of optional transmission paths for each service flow; Aggregate and encapsulate similar service flows with the same target path to form a link-level transmission unit, and generate the multi-link allocation strategy based on service flow granularity, which includes service flow mapping relationships and path identifiers.

7. The wireless communication multi-link load balancing and intelligent optimization method according to claim 6, characterized in that, The pre-trained flow classification model is invoked to perform feature encoding on the five-tuple identifier of the business flow, resulting in the type label and service quality requirement level of the business flow, including: The business flow quintuple identifier is input into the embedding layer of the flow classification model for discrete feature sparse encoding to generate an embedding vector representation; The embedded vector representation is input into the convolutional feature extraction layer of the stream classification model and subjected to one-dimensional convolution and pooling combination processing to extract local sequence features; The local sequence features are input into the temporal feature extraction layer of the stream classification model for bidirectional LSTM processing to extract long-distance temporal dependency features; The long-distance temporal dependency features are input into the classification output layer of the stream classification model. The type labels are generated through a fully connected layer and a multi-class softmax function. The service quality requirement level is generated through a fully connected layer and a regression function.

8. The wireless communication multi-link load balancing and intelligent optimization method according to claim 4, characterized in that, The spatial coding layer of the depthwise separable attention network uses 3×3 channel-wise convolutions and 1×1 pointwise convolutions, while the temporal coding layer uses dilated causal convolutions.

9. The wireless communication multi-link load balancing and intelligent optimization method according to claim 6, characterized in that, The stream classification model uses the cross-entropy loss function to train type labels and the mean squared error loss function to train service quality demand levels.

10. A wireless communication multi-link load balancing and intelligent optimization system, characterized in that, The system includes a processor and a memory connected to the processor. The memory is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the memory to implement the wireless communication multi-link load balancing and intelligent optimization method according to any one of claims 1-9.

Images