Wireless Mesh network flow prediction method based on multi-domain feature fusion deep learning

Abstract

The invention discloses a wireless Mesh network flow prediction method based on multi-domain feature fusion deep learning, and belongs to the technical field of communication networks. The method comprises the following steps: preprocessing data according to the collected data; constructing a multi-domain graph convolution model, and predicting network traffic; the multi-domain graph convolution model comprises a TimesNet time-frequency domain multi-periodic feature model, a space-time graph convolution network topology feature model and a Transform multi-domain feature adaptive model; and training the constructed convolution model divided into the multi-domain graphs, and outputting a network traffic identification result. According to the method, precise modeling of complex space-time correlation and non-stationarity in a topological domain is realized through the TE-GRU and the CCW-GCN; the three-domain features are subjected to adaptive alignment, so that the adaptability to a complex dynamic Mesh network scene is improved; the depth and accuracy of cross-domain feature fusion are improved, higher flexibility and robustness are shown in complex non-stationary Mesh network flow modeling, and efficient unified expression of the time-frequency domain and the topological domain is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of communication networks, and particularly relates to a method for predicting wireless Mesh network traffic based on multi-domain feature fusion deep learning. Background Technique

[0002] With the development of 6G technology and the continuous emergence of new application scenarios, requirements such as extended reality (XR), holographic communication, space-air-ground integrated network, and ultra-dense Internet of Things deployment have posed unprecedented challenges to communication networks. These applications require communication technologies to support characteristics such as ultra-large-scale device connection, ultra-low latency, and ultra-high reliability. However, traditional wireless communication technologies have shown significant limitations when facing these complex scenarios. For example, Wi-Fi has a limited coverage area and its signal is vulnerable to interference, making it difficult to meet the stable communication requirements of ultra-dense device environments; although cellular networks perform excellently in wide-area coverage, their high energy consumption and centralized architecture limit their applications in low-power consumption and dynamic expansion scenarios.

[0003] In this context, the Mesh network, with its self-organizing, self-healing capabilities, and multi-hop routing characteristics, has become an indispensable part of the 6G architecture. Its multi-hop routing can dynamically expand the coverage area, self-organize the routing, and data can also be re-routed through other nodes, significantly enhancing the robustness and reliability of the network. At the same time, the distributed architecture of the Mesh network reduces the dependence on centralized base stations and has significant advantages in scenarios with limited resources or high flexibility requirements. In addition, the Mesh network does not require a wired connection to be laid for each node, greatly reducing the construction cost and improving the deployment and expansion efficiency.

[0004] To fully unleash the potential of the Mesh network in 6G, network intelligence is the key. A crucial step is to accurately predict the traffic of each node in the Mesh network. Through traffic prediction, the network can perceive the load conditions of each node in advance, optimize resource allocation and scheduling, thereby significantly enhancing the stability and communication quality of the network, effectively reducing the risk of latency and data loss, and fully meeting the strict requirements for high reliability and low latency in 6G application scenarios.

[0005] However, traffic prediction in Mesh networks is more challenging than in other communication scenarios. Nodes in Mesh networks are interdependent. The traffic volume of a certain node depends not only on its own user needs, but also on multiple factors such as the communication behavior of surrounding nodes and channel interference. For example, in the industrial Internet of Things scenario, edge sensing nodes responsible for collecting various sensor data need to transmit this data to monitoring center nodes at a relatively long spatial distance for comprehensive analysis and recording; network nodes with a short spatial distance may perform similar data transmission tasks. This spatial correlation, both near and far, significantly increases the complexity of traffic prediction and must be fully considered in the traffic prediction model. In addition, changes in different user behaviors (such as daily activities, business peaks) and the external environment (such as weather, holidays) will cause traffic to show multi-periodic fluctuations. For example, on a daily or weekly time scale, traffic often shows regular fluctuations. On a daily cycle scale, traffic gradually climbs during the day, reaches a peak, and then gradually drops at night; on a weekly cycle scale, the traffic pattern may form predictable changes as user behaviors repeat periodically. Finally, unexpected events bring great pressure to traffic prediction. For example, the access of a large number of devices during disaster relief, the sharp increase in traffic during large-scale events, etc. These events often lead to drastic fluctuations in network traffic, and traditional time series methods are difficult to effectively handle such irregular changes.

[0006] Early traffic prediction models mainly focused on time series analysis. These methods predicted by learning the changing rules and trends hidden in historical data. Representative methods include the autoregressive integrated moving average model (ARIMA), vector autoregressive model (VAR), and Kalman filter model. Although these methods perform well in short-term traffic flow prediction tasks, they ignore the important spatial correlation existing in traffic data.

[0007] Graph Convolutional Network (GCN) has excellent feature extraction capabilities for non-Euclidean data. It models data using the spatial relationships in the graph and is very suitable for traffic prediction tasks. In recent traffic prediction research, graph convolutional networks and time-depth neural networks are usually combined to simultaneously capture the temporal and spatial relationships in traffic data. Spatio-Temporal Graph Convolutional Network (STGCN) and Diffusion Convolutional Recurrent Neural Network (DCRNN) are two GCN-based traffic prediction frameworks. Subsequent research works (such as T-GCN, ASTGCN, and Graph WaveNet) further explore the spatio-temporal relationships in traffic from different perspectives. However, none of these methods can simultaneously take into account the multi-periodicity, spatio-temporal correlation, and non-stationarity characteristics of mesh network traffic mentioned above, and it is difficult to achieve accurate prediction.

[0008] Therefore, how to solve the problem of non-stationary traffic changes caused by the multi-periodic and precise response to emergencies in mesh network traffic and achieve high-precision traffic prediction in the Mesh network scenario is the technical problem to be solved by the present invention. Summary of the Invention

[0009] The purpose of the present invention is to provide a wireless Mesh network traffic prediction method based on multi-domain feature fusion deep learning to solve the problems proposed in the above background technology.

[0010] The purpose of the present invention is achieved as follows: A wireless Mesh network traffic prediction method based on multi-domain feature fusion deep learning, characterized in that: the method includes the following steps:

[0011] Step S1: Data collection;

[0012] Based on the large-scale Mesh network scenario, collect multi-dimensional data of network node traffic, channel bandwidth, and antenna gain;

[0013] Step S2: Preprocess the collected data;

[0014] Step S3: Construct a multi-domain graph convolutional model to predict network traffic;

[0015] The multi-domain graph convolutional model includes a TimesNet time-frequency domain multi-periodic feature model, a spatio-temporal graph convolutional network topology feature model, and a Transformer multi-domain feature adaptive model;

[0016] Step S4: Train the constructed multi-domain graph convolutional model and output the network traffic recognition result.

[0017] Preferably, the preprocessing of the data in step S2 specifically includes:

[0018] Step S2-1: Calculate the maximum capacity of each channel based on Shannon's formula, specifically:

[0019] Define the channel capacity C ij as:

[0020]

[0021] where B ij represents the channel bandwidth size between nodes i and j, P t represents the transmit power; G i and G j represent the antenna gains of the transmitter and receiver; λ represents the signal wavelength; N represents the noise power, and d ij represents the distance between nodes;

[0022] Step S2-2: Construct an undirected weighted adjacency matrix, and use the maximum channel capacity calculated between each pair of nodes as the weight of the edge;

[0023] Generate the corresponding undirected unweighted adjacency matrix A = [a ij according to the network topology:

[0024]

[0025] Take the channel capacity C ij as the weight and add it to the undirected unweighted adjacency matrix A to generate an undirected weighted adjacency matrix

[0026]

[0027] Preferably, the TimesNet time-frequency domain multi-periodicity feature model includes a time-frequency conversion module, a period calculation module, a dimension elevation module, a vision transformer module, and a dimension reduction module. The time-frequency conversion module is used to calculate the frequency components of the time series, convert the time series from the time domain to the frequency domain using the Fourier transform, and identify the frequency components. Specifically:

[0028] For the input one-dimensional time series X 1D ∈R T×C , where T is the time length and C is the channel dimension, perform a fast Fourier transform to calculate the amplitude Amp(FFT(X 1D )) and take the mean of the amplitude of the Fourier transform along the channel dimension C to obtain the frequency intensity vector A:

[0029] A = Avg(Amp(FFT(X 1D )));

[0030] The period calculation module is used to calculate the period length, select several frequencies with high intensity, and deduce the corresponding period length from the frequencies;

[0031] In the frequency intensity vector A, find the k frequencies {f1, f2,..., f k} with the highest intensity:

[0032]

[0033] where argTopK is the index corresponding to the k frequencies with the highest frequency intensity values, and the frequency range is only selected from 1 to

[0034] Calculate the corresponding period lengths {p1, p2,..., p k}:

[0035]

[0036] Preferably, the dimension elevation module is used to reshape the time series. Through the reshaping operation, the pattern of the time series in the periodic dimension is visually presented;

[0037] For each period length p i , first perform a padding operation on the time series X 1D , fill zeros at the end of the sequence to make its length divisible by p i :

[0038] Padding(X 1D );

[0039] Then convert the padded sequence into a two-dimensional matrix according to p i :

[0040]

[0041] where Padding represents the padding operation; is a two-dimensional matrix;

[0042] The visual Transformer module is used to extract two-dimensional information features and extract two-dimensional features through Vision Transformer;

[0043] The dimension reduction module aggregates the extracted two-dimensional features into a one-dimensional feature space, and reshapes each two-dimensional feature into a one-dimensional sequence according to the dimension of p i ×f i :

[0044]

[0045] Then perform a truncation operation to adjust the length of the feature and retain the key part:

[0046]

[0047] Finally, aggregate all one-dimensional features into a unified feature space and output the feature

[0048] Preferably, the spatio-temporal graph convolutional network topological feature model includes a time encoding GRU module TE-GRU and a channel capacity weight GCN module CCW-GCN. The time encoding GRU module TE-GRU is used to extract time features; the time encoding GRU module TE-GRU includes a time encoding module and a GRU module. The time encoding module encodes time through sine and cosine functions to obtain:

[0049]

[0050] Among them, Time represents the current moment, and T is the corresponding known period;

[0051] Expand the original input to:

[0052] x t ′ = [x t ,(Time sin ,Time cos )];

[0053] Among them, x t represents the input node traffic data;

[0054] The GRU module includes an update gate, a reset gate, and a hidden state. The calculation process is as follows:

[0055] z t = σ(W z · [h t-1 , x t ′]);

[0056] r t = σ(W r · [h t-1 , x t ′]);

[0057]

[0058] Among them, z t is the update gate, which is used to control the proportion of new information flowing into the hidden state; r t is the reset gate, which is used to determine how much information needs to be retained from the past hidden state; h t is the hidden state sequence at time t, which serves as the memory unit of the model and is used to capture the dependencies of the time series; σ and tanh represent the sigmoid function and the hyperbolic tangent function respectively; W is the weight matrix of the hidden state; W r is the weight matrix of the reset gate, and W z is the weight matrix of the update gate.

[0059] Preferably, the channel capacity weighted GCN module CCW-GCN is used to extract spatial features. The channel capacity weighted GCN module CCW-GCN introduces a weighted adjacency matrix A C , and in each layer of convolution operation, the weighted adjacency matrix replaces the traditional adjacency matrix to obtain

[0060] Normalize A c as follows:

[0061]

[0062] Among them, D C is the degree matrix of A C The CCW-GCN convolution update operation is as follows:

[0063]

[0064] Among them, H (l) is the node feature matrix of the l-th layer, W (l) is the weight matrix trained by the GCN module, and σ is the activation function.

[0065] Preferably, the Transformer multi-domain feature adaptive model is used for multi-domain feature adaptive alignment, and the specific steps are as follows:

[0066] Step S3-1: Receive the spatio-temporal features of the topological domain and the multi-periodic features of the time-frequency domain Among them, N is the number of network nodes;

[0067] Concatenate H (l) and T (l) along the feature dimension into a joint representation Defined as follows:

[0068] Y (l) = Ccontact((H (l) , T (l) ));

[0069] Among them, H (l) is the spatio-temporal feature of the topological domain, and T (l) is the multi-periodic feature of the time-frequency domain;

[0070] Step S3-2: Implement the alignment operation between feature domains through the multi-head attention mechanism. The specific operation is as follows:

[0071] Map H (l) to the query vector Q m , map T (l) to the key vector K m , and the joint representation Y (l) is used as the value vector V m . The calculation formula is:

[0072]

[0073] Among them, represents the learning parameter matrix of the query vector; represents the learning parameter matrix of the key vector; represents the learning parameter matrix of the value vector; d k and d vare the dimension sizes of the query and value vectors respectively;

[0074] For the attention weights of each head, the calculation formula is:

[0075]

[0076] Step S3-3: Generate a fused representation by concatenating the outputs of M attention heads and projecting them into a unified feature space

[0077]

[0078] F (l) = Contact(head1, head1, …, head M )W C ;

[0079] where is the weight matrix of the output projection, and d o is the dimension size of the final fused feature.

[0080] Preferably, in step S4, the constructed multi-domain graph convolutional model is trained as follows:

[0081] A multi-domain loss function L total for model training is adopted:

[0082] L total = αL TF + βL Topo ;

[0083] where α represents the weight parameter of the time-frequency domain feature, β represents the topological domain weight parameter; L TF represents the time-frequency domain loss, and L Topo represents the topological domain loss;

[0084]

[0085] Compared with the prior art, the present invention has the following improvements and advantages:

[0086] 1. The multi-periodic characteristics of traffic data are modeled in the time-frequency domain through the TimesNet model, and the complex spatio-temporal correlation and non-stationarity within the topological domain are accurately modeled through TE-GRU and CCW-GCN; the three-domain features are adaptively aligned, avoiding the feature coupling problem in the traditional tandem architecture, significantly improving the adaptability to complex dynamic Mesh network scenarios; enhancing the depth and accuracy of cross-domain feature fusion, showing higher flexibility and robustness in complex non-stationary Mesh network traffic modeling, and achieving an efficient unified expression of the time-frequency domain and the topological domain.

[0087] 2. By establishing the TE-GRU module in the multi-domain graph convolutional model, TE-GRU captures the periodic changes in traffic data by introducing explicit time encoding and responds more precisely to the non-stationary changes in traffic caused by emergencies; the design of the CCW-GCN channel capacity weighted adjacency matrix strengthens the ability to capture the complex spatial dependencies between nodes in the mesh network; through the multi-head self-attention mechanism of Transformer, the correlation matrix between feature domains is calculated to capture the dependencies between the time-frequency domain and the topological domain. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 FIG. is the overall flowchart of the method of the present invention.

[0089] Figure 2 FIG. is the structural diagram of the traffic prediction scenario under the intelligent wireless Mesh network.

[0090] Figure 3 FIG. is the comparative structural diagram of MeshHSTGT and multiple benchmark models at different prediction horizons.

[0091] Figure 4 FIG. is the comparative result diagram of the MeshHSTGT model in different traffic prediction scenarios.

[0092] Figure 5 FIG. is the comparative result diagram of the MeshHSTGT model and different variant architectures (focusing on time-frequency modeling, emphasizing topological features, and cascading the two components). DETAILED DESCRIPTION OF THE EMBODIMENTS

[0093] The following further outlines the present invention with reference to the accompanying drawings.

[0094] As Figure 1 shown, a wireless Mesh network traffic prediction method based on multi-domain feature fusion deep learning, the method includes the following steps:

[0095] Step S1: Data collection;

[0096] Based on the large-scale Mesh network scenario, multi-dimensional data of network node traffic, channel bandwidth, and antenna gain are collected and organized into a standardized CSV format file.

[0097] Step S2: Preprocess the data according to the collected data, specifically including:

[0098] Step S2-1: Calculate the maximum capacity of each channel based on the Shannon formula, specifically:

[0099] Define the channel capacity C ij as:

[0100]

[0101] Among them, B ij represents the channel bandwidth size between nodes i and j, and P t represents the transmit power; G i and G j represent the antenna gains of the transmitter and receiver; λ represents the signal wavelength; N represents the noise power, and d ij represents the distance between nodes;

[0102] Step S2-2: Construct an undirected weighted adjacency matrix, and use the maximum channel capacity calculated between each pair of nodes as the weight of the edge;

[0103] Generate the corresponding undirected unweighted adjacency matrix A = [a ij according to the network topology:

[0104]

[0105] Use the channel capacity C ij as the weight and add it to the undirected unweighted adjacency matrix A to generate an undirected weighted adjacency matrix

[0106]

[0107]

[0108] Based on the collected data, use the Shannon formula to calculate the maximum capacity of the channel to quantitatively characterize the difference in channel capacity between different nodes. According to the network topology graph, first construct an undirected unweighted adjacency matrix to describe the basic connection relationship of the nodes in the Mesh network. On this basis, use the calculated maximum channel capacity as the weight of the edge between nodes to generate an undirected weighted adjacency matrix, which is used to more accurately characterize the network topology characteristics and the communication capabilities between nodes.

[0109] Step S3: Construct a multi-domain graph convolutional model MeshHSTGT to predict network traffic;

[0110] The multi-domain graph convolutional model includes the TimesNet time-frequency domain multi-periodic feature model, the spatio-temporal graph convolutional network topology feature model, and the Transformer multi-domain feature adaptive model;

[0111] The TimesNet time-frequency domain multi-periodic feature model includes a time-frequency conversion module, a period calculation module, a dimension elevation module, a visual transformer module, and a dimension reduction module. The time-frequency conversion module is used to calculate the frequency components of the time series, convert the time series from the time domain to the frequency domain using the Fourier transform, and identify the frequency components. Specifically:

[0112] For the input one-dimensional time series X1D ∈R T×C , where T is the time length and C is the channel dimension, perform a fast Fourier transform to calculate the amplitude Amp(FFT(X 1D )) and take the mean of the amplitudes of the Fourier transform along the channel dimension C to obtain the frequency intensity vector A:

[0113] A = Avg(Amp(FFT(X 1D )));

[0114] Time series usually contain periodic information, and periodic information can be revealed through frequency domain analysis; the Fourier transform can convert a time series from the time domain to the frequency domain to identify its main frequency components.

[0115] The period calculation module is used to calculate the period length, select several frequencies with high intensity, and deduce the corresponding period length from the frequencies;

[0116] In the frequency intensity vector A, find the k frequencies {f1, f2,..., f k} with the maximum intensity:

[0117]

[0118] where argTopK is the index corresponding to the k frequencies with the highest frequency intensity values. Here, the frequency range only selects from 1 to

[0119] Calculate the corresponding period lengths {p1, p2,..., p k}:

[0120]

[0121] In the frequency domain, different frequency components may have different significances; significant frequency components often correspond to key periodic features in the signal; select several frequencies with high intensity and deduce the corresponding period length from the frequencies.

[0122] The dimensionality increase module is used to reshape the time series. Through the reshaping operation, the pattern of the time series in the period dimension is visually presented;

[0123] For each period length p i , first perform a padding operation on the time series X 1D by appending zeros at the end of the sequence to make its length divisible by p i :

[0124] Padding(X 1D );

[0125] Then convert the padded sequence into a two-dimensional matrix according to p i :

[0126]

[0127] Among them, Padding represents the padding operation; is a two-dimensional matrix.

[0128] The visual transformer module is used to extract two-dimensional information features, and the two-dimensional features are extracted through Vision Transformer;

[0129] The dimensionality reduction module aggregates the extracted two-dimensional features into a one-dimensional feature space, and for each two-dimensional feature in accordance with p i ×f i is reshaped into a one-dimensional sequence of dimensions:

[0130]

[0131] Then a truncation operation is performed to adjust the length of the features and retain the key parts:

[0132]

[0133] Finally, all one-dimensional features are aggregated into a unified feature space, and the output features

[0134] The extracted two-dimensional features need to be mapped back to a one-dimensional representation for unified use in subsequent tasks; at the same time, the feature dimensions need to be adjusted to meet the input requirements of the model.

[0135] The spatio-temporal graph convolutional network topological feature model includes a time-encoding GRU module TE-GRU and a channel capacity weight GCN module CCW-GCN;

[0136] In a mesh network with emergencies and complex topological structures, there are multiple connection paths between nodes. The time features of traffic data not only come from the periodic changes of user behavior but also from uncertain events inside the network, such as node interruptions and severe congestion during specific peak hours. It is difficult to predict the non-stationary changes in traffic caused by emergencies relying solely on traditional GRU;

[0137] The time-encoding GRU module TE-GRU introduces time-encoding features and encodes time through sine and cosine functions to obtain:

[0138]

[0139] Among them, Time represents the current moment, and T is the corresponding known period;

[0140] The original input is extended to:

[0141] x t ′ = [x t ,(Time sin ,Time cos )];

[0142] Among them, x t represents the input node traffic data;

[0143] Finally, the extended x t ′ is input into the GRU module for feature extraction. The GRU module includes an update gate, a reset gate, and a hidden state. The calculation process is

[0144] z t = σ(W z · [h t-1 , x t ′]);

[0145] r t = σ(W r · [h t-1 , x t ′]);

[0146]

[0147] Among them, z t is the update gate, used to control the proportion of new information flowing into the hidden state; r t is the reset gate, used to determine how much information needs to be retained from the past hidden state; h t is the hidden state sequence at time t, serving as the memory unit of the model, used to capture the dependencies of the time series; σ and tanh represent the sigmoid function and the hyperbolic tangent function respectively; W is the weight matrix of the hidden state; W r is the weight matrix of the reset gate, and W z is the weight matrix of the update gate.

[0148] The channel capacity weight GCN module CCW-GCN is used to extract spatial features. The traditional GCN constructs the adjacency matrix and the Laplacian matrix to represent the topological structure of the network. This simple representation method cannot reflect the complex characteristics that the traffic size of nodes in the Mesh communication network is affected by various communication characteristics (such as signal strength, distance attenuation, signal-to-noise ratio); the channel capacity weight GCN module CCW-GCN introduces a weighted adjacency matrix A C , and replaces the traditional adjacency matrix with the weighted adjacency matrix in each layer of the convolution operation to obtain

[0149] For Ac Perform normalization processing:

[0150]

[0151] Among them, D C is the degree matrix of A C The CCW-GCN convolution update operation is:

[0152]

[0153] Among them, H (l) is the node feature matrix of the l-th layer, W (l) is the weight matrix trained by the GCN module, and σ is the activation function.

[0154] To effectively fuse the information in different feature domains in Mesh network traffic prediction, a multi-domain feature adaptive alignment method based on Transformer is designed. The Transformer multi-domain feature adaptive model is used for multi-domain feature adaptive alignment. The specific steps are as follows:

[0155] Step S3-1: Receive the spatio-temporal features of the topological domain and the multi-periodic features of the time-frequency domain Among them, N is the number of network nodes;

[0156] Concatenate H (l) and T (l) along the feature dimension into a joint representation defined as follows:

[0157] Y (l) = Ccontact((H (l) , T (l) ));

[0158] Among them, H (l) is the spatio-temporal feature of the topological domain, and T (l) is the multi-periodic feature of the time-frequency domain;

[0159] Step S3-2: Implement the alignment operation between feature domains through the multi-head attention mechanism. The specific operation is:

[0160] Map H (l) to the query vector Q m , map T (l) to the key vector K m , and the joint representation Y (l) is used as the value vector V m . The calculation formula is:

[0161]

[0162] Among them, The learning parameter matrix representing the query vector; The learning parameter matrix representing the key vector; The learning parameter matrix representing the value vector; d k and d v are respectively the dimensionality sizes of the query and value vectors;

[0163] For the attention weights of each head, the calculation formula is:

[0164]

[0165] Step S3-3: Generate a fused representation by concatenating the outputs of M attention heads and projecting them into a unified feature space

[0166]

[0167] F (l) = Contact(head1, head1, …, head M )W C ;

[0168] Among them, is the weight matrix of the output projection, d o is the dimensionality size of the final fused feature.

[0169] To improve the training accuracy of the multi-domain graph convolutional model and reduce the risk of overfitting, a multi-domain loss function L total is designed for model training, L total :

[0170] L total = αL TF + βL Topo ;

[0171] Among them, α represents the weight parameter of the time-frequency domain feature, β represents the topological domain weight parameter; L TF represents the time-frequency domain loss, L Topo represents the topological domain loss;

[0172]

[0173] To prove the effectiveness of the present invention, the experiments are set as follows:

[0174] Set the training configuration. The dataset is divided into a training set, a test set, and a validation set in a ratio of 7:2:1. The multi-domain graph convolutional model is trained for a total of 200 epochs. To mitigate overfitting and improve generalization ability, a dropout rate of 0.3 is adopted. The RMSProp optimizer is used to train the multi-domain graph convolutional model, and the learning rate is set to 0.001. The spatio-temporal graph convolutional network topology feature model contains 3 groups of TE-GRU and CCW-GCN modules. The dimension of the entire spatio-temporal graph convolutional network topology feature model is 32→64→64→32→128→128→128→128.

[0175] The dimensions of Q, K, and V in the Transformer attention module are all set to 8, and each attention module contains 8 heads. In the multi-domain loss function, if the task is more sensitive to time-frequency domain features, the hyperparameters are set to α = 0.7, β = 0.3; if the task is more sensitive to topology domain features, the hyperparameters are set to α = 0.3, β = 0.7.

[0176] Regardless of whether the prediction time is 15 minutes, 30 minutes, 45 minutes, or 60 minutes, the input sequence (time window) is 60 minutes, that is, 12 time steps.

[0177] Train the model according to the above parameter configuration and loss function.

[0178] All experiments are conducted on NVIDIA Tesla V100 GPUs, and consistent random seeds are used to ensure reproducibility;

[0179] The multi-domain graph convolutional model MeshHSTGT is compared with multiple baseline models under different prediction horizons (10, 30, and 60 minutes). The results demonstrate that the architecture we proposed has superior performance in capturing complex mesh network traffic patterns; as Figure 3As shown, the traditional neural architectures FNN and FC-LSTM exhibit significant limitations due to their single-domain modeling methods. FNN performs poorly in 60-minute predictions, with an MAE of 813.28; while FC-LSTM is slightly improved through its gating mechanism, but due to the lack of spatial modeling ability in its fully connected architecture, the MAE is still high at 781.52, especially in long-term predictions. These results validate our hypothesis that relying solely on time features is insufficient for effective mesh network traffic prediction. Methods based on graph neural networks exhibit improved multi-domain modeling capabilities but also encounter the problem of feature entanglement. Although GWN innovatively combines an adaptive dependency matrix and dilated convolution, it shows suboptimal performance in 60-minute predictions, with an MAE of 513.93, 31.4% higher than MeshHSTGT. This limitation stems from the fact that its graph convolution module fails to consider changes in channel capacity. The Chebyshev polynomial approximation of ST-ChebNet, although computationally efficient, performs poorly in handling bursty traffic patterns, with a MAPE of 0.20 for 30-minute predictions. The local spatio-temporal graph structure of STSGCN enables synchronous feature extraction, but its cascaded architecture leads to feature entanglement, resulting in an RMSE of 561.79 for 60 minutes, 19.5% higher than our model. Recent advances in feature interaction modeling, such as STFGNN and TSGAN, although improved, are still limited by their feature fusion strategies. STFGNN's dynamic time graph based on DTW achieved an MAE of 212.53 in 30-minute predictions, but its fixed-weight feature fusion mechanism is insufficient to handle the dynamic changes in mesh networks. TSGAN's attention-based adjacency matrix construction performs well in short-term predictions (MAE of 94.71 in the 10-minute horizon), but due to its single-stage feature extraction, it performs poorly in long-term predictions, with a MAPE of 0.14 for 60 minutes, 40% higher than MeshHSTGT. Our proposed MeshHSTGT performs excellently in all metrics, and its innovative feature re-extraction architecture greatly improves performance. The time-frequency dual-domain modeling of the TimesNet component effectively separates mixed periodic patterns, achieving an MAE of 89.62 in 10-minute predictions, a 5.4% improvement over TSGAN. The channel capacity weighting mechanism of the enhanced CCW-GCN module demonstrates enhanced robustness to topological changes, with an RMSE of 470.25 in 60-minute predictions, 13.3% lower than STFGNN. The self-attention mechanism of the adaptive alignment module enables dynamic feature fusion, significantly improving the stability of long-term predictions. These results validate the effectiveness of our parallel feature re-extraction architecture in solving the problem of feature entanglement commonly present in traditional methods, establishing a new paradigm for complex mesh network traffic prediction.

[0180] To verify the generalization ability of our proposed MeshHSTGT model in different traffic prediction scenarios, we conducted extensive experiments on the Milan cellular traffic dataset provided by Telecom Italia. This dataset includes Internet activity call records sampled at 10-minute intervals over 62 days, providing rich temporal patterns in an urban context and posing unique challenges for traffic prediction models. As Figure 4 shown, accurately predicting urban cellular traffic patterns is crucial for government agencies and network operators to optimize resource allocation and foresee potential congestion issues. This is particularly critical in large urban areas, as the accuracy of predictions directly impacts traffic management efficiency. We evaluated the performance of MeshHSTGT against multiple baseline methods (FNN, FC-LSTM, GWN, ST-ChebNet, STSGCN, STFGNN, and TSGAN) under different prediction horizons. The results demonstrated the robust performance of MeshHSTGT in capturing changes in network demand across different urban areas. The success of our model can be attributed to its effective integration of time-frequency domain features and topological relationships through a parallel feature re-extraction architecture. Although TSGAN showed competitive performance in short-term predictions, MeshHSTGT maintained superior accuracy across all prediction horizons, with an approximately 14.8% improvement in MAE in long-term predictions (60-minute horizon). These findings validate the effectiveness of MeshHSTGT in real-world urban scenarios, demonstrating its ability to capture complex spatio-temporal dependencies in cellular network traffic. The strong performance of the model on this independent dataset confirms its generalization ability beyond mesh networks, indicating its potential applications in diverse network traffic prediction scenarios.

[0181] To systematically evaluate the contribution of each architectural component to the mesh network traffic prediction performance, we conducted a comprehensive ablation study, comparing our full MeshHSTGT model with three variant architectures: TimesNet (focusing on time-frequency modeling), STGCN (emphasizing topological features), and TimesNet_STGCN (cascading and combining both components). The experimental results revealed the effectiveness of different modeling methods and validated the rationality of our architecture design. As Figure 5As shown, although the individual TimesNet module can effectively capture the time-frequency patterns and periodic fluctuations in network traffic, it shows limitations in dealing with complex node interactions due to the lack of topological modeling. This defect is particularly evident when dealing with network events involving sudden traffic changes or node failures, resulting in poor performance metrics under different prediction horizons (MAE: 139.38, 270.51, 564.62; RMSE: 105.13, 218.31, 438.46). In contrast, the STGCN variant shows improvements in modeling spatial dependencies and dealing with sudden traffic patterns by introducing the TE-GRU and CCW-GCN components. However, it fails to capture multi-period characteristics, leading to poor performance (MAE: 136.75, 282.42, 551.19; MAPE: 0.17, 0.17, 0.18), although it represents an improvement over the TimesNet architecture. The cascaded TimesNet_STGCN architecture attempts to bridge this gap by sequentially combining time-frequency modeling and topological modeling. Although this approach can theoretically capture all necessary feature dimensions, the sequential nature of feature processing introduces the problem of feature entanglement, resulting in performance metrics (MAE: 105.13, 218.31, 438.46; MAPE: 0.14, 0.16, 0.16) inferior to the capabilities of our complete model. This observation emphasizes a key insight: while comprehensively capturing features is crucial, the method of feature fusion significantly affects prediction accuracy. These ablation results decisively validate the superiority of the parallel architecture we proposed in MeshHSTGT. By implementing independent but synchronous modeling of time-frequency and topological features, and through feature re-extraction and adaptive alignment, our method effectively addresses the limitations observed in each variant. The parallel architecture successfully alleviates the problem of feature entanglement while maintaining the advantages of multi-domain modeling, thus achieving higher prediction accuracy and robustness under all evaluation metrics and prediction horizons. This comprehensive analysis not only justifies our architectural choice but also provides valuable insights for future research on network traffic prediction modeling.

[0182] The above description is only for the embodiments of the present invention and is not intended to limit the present invention. For those skilled in the art, various changes and modifications can be made to the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the scope of the claims of the present invention.

Claims

1. A wireless Mesh network traffic prediction method based on multi-domain feature fusion deep learning, characterized by: The method comprises the following steps: Step S1: data collection; Based on large-scale Mesh network scenarios, collect multi-dimensional data on network node traffic, channel bandwidth, and antenna gain; Step S2: pre-processing the data according to the collected data; Step S3: construct a multi-domain graph convolution model to predict network traffic; The multi-domain graph convolution model includes TimesNet time-frequency domain multi-periodic feature model, spatiotemporal graph convolution network topological feature model and Transformer multi-domain feature adaptation model; Step S4: Train the constructed multi-domain graph convolution model and output the network traffic recognition result.

2. According to claim 1, a wireless mesh network traffic prediction method based on multi-domain feature fusion deep learning is characterized by: The data is preprocessed in step S2, specifically including: Step S2-1: Calculate the maximum capacity of each channel based on the Shannon formula, specifically: Define the channel capacity C ij for: Among them, B ij represents the channel bandwidth between nodes i and j, P t Represents the transmission power; G i and G j represents the antenna gain of the transmitting and receiving ends; λ represents the signal wavelength; N represents the noise power, d ij Represents the distance between nodes; Step S2-2: construct an undirected weighted adjacency matrix, and use the maximum channel capacity calculated between each node as the weight of the edge; Generate the corresponding undirected and unweighted adjacency matrix A=[a ij ]: The channel capacity C ij As a weight, add it to the undirected unweighted adjacency matrix A to generate an undirected weighted adjacency matrix 3. According to claim 1, a wireless mesh network traffic prediction method based on multi-domain feature fusion deep learning is characterized by: The TimesNet time-frequency domain multi-periodic feature model includes a time-frequency conversion module, a period calculation module, a dimension increase module, a visual transformer module, and a dimension reduction module. The time-frequency conversion module is used to calculate the frequency components of the time series, and use Fourier transform to convert the time series from the time domain to the frequency domain to identify the frequency components, specifically: For the input one-dimensional time series X 1D ∈R T×C , T is the time length, C is the channel dimension, perform fast Fourier transform, calculate the amplitude Amp(FFT(X 1D )), take the average of the Fourier transform amplitude along the channel dimension C to obtain the frequency intensity vector A: A=Avg(Amp(FFT(X 1D ))); The cycle calculation module is used to calculate the cycle length, select a number of frequencies with high intensity, and derive the corresponding cycle length through the frequency; In the frequency intensity vector A, find the k frequencies {f1,f2,…,f k }: Among them, argTopK is the index corresponding to the k frequencies with the highest frequency intensity value. Here, the frequency range is only selected from 1 to Calculate the corresponding period length {p1,p2,…,p k }:

4. According to claim 3, a wireless mesh network traffic prediction method based on multi-domain feature fusion deep learning is characterized by: The dimension-increasing module is used to reshape the time series, and through the reshaping operation, the pattern of the time series in the period dimension is intuitively presented; For each period length p i , firstly, the time series X 1D Perform a padding operation to add zeros to the end of the sequence so that its length can be p i , divisible by: Padding(X 1D ); Then the filled sequence is divided into i Convert to a two-dimensional matrix: Among them, Padding represents the filling operation; is a two-dimensional matrix; The visual transformer module is used to extract two-dimensional information features, and the two-dimensional features are extracted through the Vision Transformer; The dimension reduction module aggregates the extracted two-dimensional features into a one-dimensional feature space, and According to p i ×f i Reshape the dimension into a one-dimensional sequence: Then perform a truncation operation to adjust the length of the feature and retain the key part: Finally, all one-dimensional features are aggregated into a unified feature space and the output feature 5. According to claim 1, a wireless mesh network traffic prediction method based on multi-domain feature fusion deep learning is characterized by: The spatiotemporal graph convolutional network topology feature model includes a time coding GRU module TE-GRU and a channel capacity weight GCN module CCW-GCN. The time coding GRU module TE-GRU is used to extract time features. The time coding GRU module TE-GRU includes a time coding module and a GRU module. The time coding module encodes time through sine and cosine functions to obtain: Among them, Time represents the current time, and T is the corresponding known period; Expands the original input to: x t ′=[x t ,(Time sin ,Time cos )]; Among them, x t Represents the input node traffic data; The GRU module includes an update gate, a reset gate, and a hidden state, and the calculation process is: z t =σ(W z ·[h t-1 ,x t ′]); r t =σ(W r ·[h t-1 ,x t ′]); Among them, z t is the update gate, which is used to control the ratio of new information flowing into the hidden state; r t is the reset gate, which is used to decide how much information needs to be retained from the past hidden state; h t is the hidden state sequence at time t, which serves as the memory unit of the model to capture the dependency of the time series; σ and tanh represent the sigmoid function and the hyperbolic tangent function respectively; W is the weight matrix of the hidden state; W r is the weight matrix of the reset gate, W z is the weight matrix of the update gate.

6. According to claim 5, a wireless mesh network traffic prediction method based on multi-domain feature fusion deep learning is characterized by: The channel capacity weighted GCN module CCW-GCN is used to extract spatial features. The channel capacity weighted GCN module CCW-GCN introduces a weighted adjacency matrix A based on channel capacity. C , in the convolution operation of each layer, the weighted adjacency matrix Replace the traditional adjacency matrix get A c Perform normalization: Among them, D C A C The degree matrix of CCW-GCN convolution update operation is: Among them, H (l) is the node feature matrix of the lth layer, W (l) is the weight matrix of GCN module training, and σ is the activation function.

7. According to claim 1, a wireless mesh network traffic prediction method based on multi-domain feature fusion deep learning is characterized by: The Transformer multi-domain feature adaptation model is used for multi-domain feature adaptive alignment. The specific steps are as follows: Step S3-1: Receive the spatiotemporal features H of the topological domain (l) ∈R N×dH and multi-periodic characteristics in time and frequency domains Where N is the number of network nodes; H (l) and T (l) Concatenate along the feature dimension into a joint representation The definition is as follows: Y (l) =Ccontact((H (l) ,T (l) ); Among them, H (l) is the spatiotemporal characteristics of the topological domain, T (l) Multi-periodic characteristics in time-frequency domain; Step S3-2: Alignment between feature domains is achieved through a multi-head attention mechanism. The specific operations are as follows: H (l) Mapped to query vector Q m , T (l) Mapped to key vector K m , and the joint expression Y (l) Used as value vector V m , the calculation formula is: in, a matrix of learned parameters representing the query vector; The learning parameter matrix representing the key vector; The learning parameter matrix representing the value vector; d k and d v are the dimension sizes of query and value vectors respectively; For the attention weight of each head, the calculation formula is: Step S3-3: Generate a fused representation by concatenating the outputs of M attention heads and projecting them into a unified feature space F (l) =Contact(head1,head1,…,head M )W C ; in, is the weight matrix of the output projection, d o is the dimension size of the final fusion feature.

8. According to claim 1, a wireless mesh network traffic prediction method based on multi-domain feature fusion deep learning is characterized by: In step S4, the constructed multi-domain graph convolution model is trained, specifically: Using multi-domain loss function L total For model training, L total : L total =αL TF +βL Topo ; Among them, α represents the weight parameter of the time-frequency domain feature, β represents the weight parameter of the topological domain; L TF represents the time-frequency domain loss, L Topo represents the topological domain loss;

Images