A wireless network channel state prediction system based on a graph neural network

Abstract

The application discloses a wireless network channel state prediction system based on a graph neural network, comprising: a channel collection module for generating an access channel observation sequence; a state coding module for generating an access channel state vector; a spectral mapping module for inputting a Koopman operator algorithm introducing a manifold spectral stability verification mechanism; an initial graph construction module for constructing an initial switching channel graph; a spectral state verification module for generating a switching channel graph; a trajectory generation module for inputting an improved Graph WaveNet model, introducing a signal transition spectral kernel sequential connection mechanism, generating an adjacency matrix and a candidate channel state trajectory; and a trajectory sorting module for generating a channel state trajectory. The application improves the continuous stability and prediction accuracy of channel state prediction in a wireless network switching scenario, and reduces the trajectory jump rate and switching false trigger rate in a complex dynamic propagation environment.

Description

Technical Field

[0001] This invention relates to the field of wireless communication channel prediction technology, and in particular to a wireless network channel state prediction system based on graph neural networks. Background Technology

[0002] With the continuous development of 5G communication, vehicle-to-everything (V2X) communication, industrial wireless control, and mobile edge computing, the scenarios of high-speed terminal movement, frequent handover, and complex propagation in wireless network environments are gradually increasing. Dynamic prediction technology for wireless network channel states has attracted widespread attention. Existing wireless network systems typically rely on wireless channel parameters such as CSI and SINR to evaluate and schedule link quality; however, the following problems commonly exist in practical applications: During high-speed movement and frequent access handovers, wireless terminals exhibit significant nonlinear time-varying characteristics in CSI and SINR. Traditional prediction methods based on linear statistical models or single time series models struggle to accurately describe the dynamic relationships between wireless links, leading to drift and distortion in channel prediction results. Furthermore, the handover relationships between access nodes, candidate access relationships, and wireless link propagation relationships in wireless networks exhibit clear graph-structure coupling characteristics. Existing prediction techniques typically model only the temporal changes of a single link, failing to simultaneously characterize the continuous propagation and evolution relationships between multiple nodes, resulting in poor stability of channel state predictions in wireless network handover scenarios. For dynamic channel changes in complex wireless propagation environments, traditional graph neural network methods often employ fixed topological relationships or simple correlation calculations during adjacency matrix construction, making it difficult to accurately reflect the continuous channel evolution characteristics between adjacent sampling times. This can easily cause abrupt graph structure changes and discontinuous spatial propagation characteristics, affecting the continuity of candidate channel state trajectories and prediction accuracy. Furthermore, existing wireless network channel prediction methods often lack effective characterization of stable channel evolution regions and invariant propagation characteristics when processing continuous evolution data at multiple sampling times. This makes the channel state trajectory susceptible to local abnormal fluctuations and short-term interference, making it difficult to meet the application requirements for highly continuous and stable channel prediction in complex wireless network environments.

[0003] Therefore, how to provide a wireless network channel state prediction system based on graph neural networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a wireless network channel state prediction system based on graph neural networks. This invention fully utilizes the Koopman operator algorithm, the improved Graph WaveNet model, and the spatiotemporal propagation analysis technology of graph neural networks. It details the processes of constructing access channel observation sequences, extracting channel invariant subspace features, generating channel evolution mapping relationships, constructing handover channel graphs, spectral state verification, adjacency matrix generation, candidate channel state trajectory generation, and trajectory ordering. It has the advantages of strong adaptability to wireless network handover scenarios, strong ability to express channel propagation relationships, high continuity of channel state trajectories, and high accuracy of wireless channel prediction.

[0005] A wireless network channel state prediction system based on a graph neural network according to an embodiment of the present invention includes: The channel acquisition module is used to acquire the CSI, SINR, current access node identifier, and candidate access node identifier of the terminal at the sampling time, and generate the access channel observation sequence. The state coding module is used to extract CSI amplitude difference, CSI phase difference and SINR difference from the access channel observation sequence, perform embedding coding on the current access node identifier and the candidate access node identifier, and generate the access channel state vector. The spectral mapping module is used to introduce the access channel state vector input into the Koopman operator algorithm of the manifold spectral stability verification mechanism to generate channel invariant subspace features and channel evolution mapping relationship; The initial mapping module is used to construct an initial handover channel map based on the current access node identifier, candidate access node identifier, CSI amplitude difference, CSI phase difference, SINR difference, channel invariant subspace characteristics, and channel evolution mapping relationship; The spectral state verification module is used to perform spectral state verification on the candidate access relationships in the initial handover channel diagram based on the channel evolution mapping relationship, and generate the handover channel diagram. The trajectory generation module is used to input the switching channel graph into the improved Graph WaveNet model, introduce a signal-to-switching spectrum kernel sequential communication mechanism in the adaptive adjacency generation layer, generate an adjacency matrix, and perform diffusion graph convolution calculation and gated dilation time convolution calculation based on the adjacency matrix to generate candidate channel state trajectories. The trajectory ordering module is used to perform trajectory ordering processing on candidate channel state trajectories based on the channel evolution mapping relationship, and generate channel state trajectories.

[0006] Optionally, the channel acquisition module specifically comprises: Collect the CSI and SINR corresponding to the terminal according to the sampling time, and record the current access node identifier and candidate access node identifier corresponding to each sampling time; Pair CSI, SINR, current access node identifier, and candidate access node identifier according to the same sampling time to generate a single-time access channel observation item; Arrange the single-time access channel observation items in the order of sampling time to generate the access channel observation sequence.

[0007] Optionally, the step of extracting CSI amplitude difference, CSI phase difference, and SINR difference based on the access channel observation sequence specifically involves: The CSI and SINR in the access channel observation sequence are arranged in chronological order of sampling time. The CSI is split using complex number decomposition to generate the CSI amplitude sequence and the CSI phase sequence; Select the CSI amplitude corresponding to adjacent sampling times, calculate the amplitude change between sampling times, and generate the CSI amplitude difference; Select the CSI phase corresponding to adjacent sampling times, calculate the phase change between sampling times, and generate the CSI phase difference; Select the SINR corresponding to adjacent sampling times, calculate the SINR change between sampling times, and generate the SINR difference.

[0008] Optionally, the step of performing embedding encoding on the current access node identifier and the candidate access node identifier to generate an access channel state vector specifically involves: The current access node identifier and the candidate access node identifier are transformed by index mapping to generate a node index sequence; based on the embedding parameter matrix corresponding to the node index sequence, an access node embedding vector is generated. Extract the CSI amplitude difference, CSI phase difference, SINR difference, and access node embedding vector corresponding to the same sampling time. The access channel state vector is generated by concatenating the CSI amplitude difference, CSI phase difference, SINR difference, and access node embedding vector in the order of feature arrangement.

[0009] Optionally, the spectral mapping module specifically comprises: Based on the CSI amplitude difference, CSI phase difference, SINR difference and access node embedding vector in the access channel state vector, the access channel state vector is mapped by an observable function to generate a channel observable state sequence. Access manifold coordinates are constructed from the access node embedding vectors, and the access manifold coordinates are written into the channel observable state sequence to generate the manifold channel observable sequence; The Koopman operator algorithm is used to perform a linear evolution mapping on the observable sequences of the manifold channel to generate candidate invariant subspace features; A manifold spectral stability verification mechanism is adopted to calculate the spectral change sequence between candidate invariant subspace features corresponding to adjacent sampling times; the spectral components are sorted according to the change amplitude of the spectral change sequence; the spectral component with the smallest change amplitude between sampling times is retained to generate channel invariant subspace features; Based on the channel-invariant subspace characteristics, the linear evolution relationship between adjacent sampling times is calculated to generate the channel evolution mapping relationship.

[0010] Optionally, the initial mapping module specifically includes: Generate a graph node set based on the current access node identifier, candidate access node identifier, and wireless link identifier; Generate connection graph edges based on the connection relationships corresponding to the current access node identifiers, generate candidate access graph edges based on the access relationships corresponding to the candidate access node identifiers, and generate evolution graph edges based on the channel evolution mapping relationship; Write the CSI amplitude difference, CSI phase difference, and SINR difference into the edge attributes of the connection graph and the candidate access graph; write the channel invariant subspace features and the channel evolution mapping relationship into the edge attributes of the evolution graph. An initial switching channel graph is constructed from the set of graph nodes, connecting graph edges, candidate access graph edges, evolving graph edges, and edge attributes.

[0011] Optionally, the spectral state verification module specifically comprises: Extract the CSI amplitude difference, CSI phase difference, and SINR difference corresponding to the candidate access graph edges to generate the candidate access edge channel difference sequence; extract the channel evolution mapping relationship corresponding to the evolution graph edges to generate the candidate access edge evolution sequence; Calculate the spectral state offset between the candidate access side channel differential sequence and the candidate access side evolution sequence; Candidate access graph edges are arranged in ascending order of spectral state offset to generate a candidate access graph edge sequence. The candidate access graph edge with the smallest spectral state offset in the candidate access graph edge sequence is selected, and the corresponding connection graph edge and evolution graph edge are retained to generate the switching channel graph.

[0012] Optionally, the improved Graph WaveNet model specifically includes a channel input layer, an adaptive adjacency generation layer, a diffusion graph convolutional layer, a gated dilation temporal convolutional layer, and a channel prediction decoding layer; The channel input layer converts the set of graph nodes, connecting graph edges, candidate access graph edges, evolving graph edges, and edge attributes in the switching channel graph into a graph structure input tensor. The edge attributes include CSI amplitude difference, CSI phase difference, SINR difference, channel invariant subspace features, and channel evolution mapping relationship. Under the sequential linking mechanism of the signal migration spectrum kernel, the adaptive adjacency generation layer assembles the CSI amplitude difference, CSI phase difference, SINR difference, channel invariant subspace features, and channel evolution mapping relationship into a signal migration spectrum kernel vector. The signal migration spectrum kernel vectors are arranged in the order of sampling time to generate a signal migration spectrum kernel sequence. The signal migration spectrum kernel vectors corresponding to adjacent sampling times are associated and concatenated to generate a sequential linking vector. The sequential linking vectors are written into the edges of the connection graph, candidate access graph, and evolution graph, respectively, forming the edges of the connection spectrum kernel graph, candidate access spectrum kernel graph, and evolution spectrum kernel graph. The graph edge association values ​​between the edges of the connection spectrum kernel graph, candidate access spectrum kernel graph, and evolution spectrum kernel graph are calculated, and an adjacency matrix is ​​generated by arranging the edges according to the graph edge association values. The diffusion graph convolutional layer constructs a forward diffusion matrix and a backward diffusion matrix based on the adjacency matrix. The graph structure input tensor is multiplied by the forward diffusion matrix and the backward diffusion matrix respectively to generate forward channel propagation features and backward channel propagation features. The forward channel propagation features and backward channel propagation features are then concatenated to generate spatial channel propagation features. The gated dilated temporal convolutional layer inputs the spatial channel propagation features into the dilated causal convolutional branch and the gated convolutional branch to generate temporal convolutional features and gated modulation features, and multiplies the temporal convolutional features and gated modulation features to generate temporal channel propagation features; The channel prediction decoding layer maps the time channel propagation characteristics to the CSI prediction sequence and SINR prediction sequence corresponding to the future sampling time, generating candidate channel state trajectories.

[0013] Optionally, the step of performing diffusion graph convolution calculation and gated dilation time convolution calculation based on the adjacency matrix to generate candidate channel state trajectories specifically involves: Extract the CSI prediction sequence and SINR prediction sequence from the candidate channel state trajectory; arrange the CSI prediction sequence and SINR prediction sequence according to the future sampling time order to generate a time-series channel trajectory sequence; Extract the CSI and SINR prediction values ​​corresponding to adjacent future sampling times from the time-series channel trajectory sequence; calculate the CSI trajectory change and SINR trajectory change between adjacent future sampling times; arrange the trajectory nodes in the time-series channel trajectory sequence according to the CSI trajectory change and SINR trajectory change. The trajectory nodes with continuously changing CSI trajectory changes and the trajectory nodes with continuously changing SINR trajectory changes are sequentially connected to generate continuous trajectory segments; the continuous trajectory segments are then spliced ​​together according to the order of future sampling times to generate the channel state trajectory.

[0014] Optionally, the trajectory sequencing module specifically comprises: Extract the CSI prediction sequence and SINR prediction sequence corresponding to the future sampling time from the candidate channel state trajectory to generate the candidate trajectory node sequence; The CSI evolution and SINR evolution between adjacent future sampling times are calculated based on the channel evolution mapping relationship to generate an evolution reference sequence; Calculate the CSI prediction change and SINR prediction change between adjacent trajectory nodes in the candidate trajectory node sequence to generate a candidate trajectory change sequence; Calculate the trajectory offset between the candidate trajectory change sequence and the evolution reference sequence; Arrange the candidate trajectory nodes in ascending order of trajectory offset to generate an ordered trajectory node sequence; connect the trajectory nodes in the ordered trajectory node sequence to generate the channel state trajectory.

[0015] The beneficial effects of this invention are: This invention employs the Koopman operator algorithm with a manifold spectral stability verification mechanism. It constructs a manifold channel observable sequence using access manifold coordinates and a channel observable state sequence, and filters stable spectral components through spectral change sequences between candidate invariant subspace features. This achieves stable generation of channel invariant subspace features and channel evolution mapping relationship, thereby improving the ability to model the dynamic evolution of wireless channels. By constructing the initial handover channel graph and performing spectral state verification, and by filtering candidate access graph edges using spectral state offsets, the connection and evolution relationships in the handover channel graph are continuously maintained, thereby improving the stability of the graph structure in wireless network handover scenarios. By introducing a sequential connection mechanism for the signal migration spectrum kernel in the improved Graph WaveNet model, the signal migration spectrum kernel vectors corresponding to adjacent sampling times are constructed into sequential connection vectors. Based on the edges of the connecting spectrum kernel graph, the edges of the candidate access spectrum kernel graph, and the edges of the evolving spectrum kernel graph, an adjacency matrix is ​​generated to realize the joint expression of the multi-node propagation relationship and the channel evolution relationship in the wireless network, thereby improving the continuity of spatial channel propagation characteristics. By performing diffusion graph convolution calculation, gated dilation time convolution calculation, and trajectory ordering, spatial channel propagation features are extracted simultaneously using forward and backward diffusion matrices. The candidate trajectory node sequences are then ordered to generate continuous and stable channel state trajectories, thereby improving the accuracy of wireless network channel state prediction. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of a wireless network channel state prediction system based on a graph neural network proposed in this invention. Figure 2 This is a schematic diagram of the improved Graph WaveNet model proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-2 A wireless network channel state prediction system based on graph neural networks, comprising: The channel acquisition module is used to acquire the CSI, SINR, current access node identifier, and candidate access node identifier of the terminal at the sampling time, and generate the access channel observation sequence. The state coding module is used to extract CSI amplitude difference, CSI phase difference and SINR difference from the access channel observation sequence, perform embedding coding on the current access node identifier and the candidate access node identifier, and generate the access channel state vector. The spectral mapping module is used to introduce the access channel state vector input into the Koopman operator algorithm of the manifold spectral stability verification mechanism to generate channel invariant subspace features and channel evolution mapping relationship; The initial mapping module is used to construct an initial handover channel map based on the current access node identifier, candidate access node identifier, CSI amplitude difference, CSI phase difference, SINR difference, channel invariant subspace characteristics, and channel evolution mapping relationship; The spectral state verification module is used to perform spectral state verification on the candidate access relationships in the initial handover channel diagram based on the channel evolution mapping relationship, and generate the handover channel diagram. The trajectory generation module is used to input the switching channel graph into the improved Graph WaveNet model. It introduces a sequential communication mechanism with a signal-to-switching spectrum kernel in the adaptive adjacency generation layer to generate an adjacency matrix. Based on the adjacency matrix, it performs diffusion graph convolution calculation and gated dilation time convolution calculation to generate candidate channel state trajectories. The trajectory ordering module is used to perform trajectory ordering processing on candidate channel state trajectories based on the channel evolution mapping relationship, and generate channel state trajectories.

[0019] In this embodiment, the channel acquisition module specifically includes: During the sampling period in which the terminal and the wireless network access node maintain communication, the CSI and SINR corresponding to each sampling moment are collected; CSI includes the complex channel response value between the terminal and the access node, and SINR includes the signal-to-interference-plus-noise ratio value of the terminal at the same sampling moment; Record the identifier of the current access node that maintains a communication connection with the terminal at each sampling time, and record the identifiers of the candidate access nodes in the access candidate list; The CSI, SINR, current access node identifier, and candidate access node identifier corresponding to the same sampling time are grouped into the same data unit to generate a single-time access channel observation item; the single-time access channel observation item includes the sampling time, CSI, SINR, current access node identifier, and candidate access node identifier; Multiple single-time access channel observation items are arranged in order of sampling time from early to late to generate an access channel observation sequence.

[0020] In this embodiment, the CSI amplitude difference, CSI phase difference, and SINR difference are extracted from the access channel observation sequence, specifically as follows: The CSI and SINR in the access channel observation sequence are arranged in order from earliest to latest sampling time; CSI is a complex channel response value, which includes a real part and an imaginary part; For each sampling time, the CSI is decomposed into a complex number. The CSI amplitude is the square root of the sum of the squares of the real part and the squares of the imaginary part, and the CSI phase is the arctangent angle corresponding to the ratio of the imaginary part to the real part. Arrange the CSI amplitudes in the order of sampling time to generate a CSI amplitude sequence; arrange the CSI phases in the order of sampling time to generate a CSI phase sequence. Select the CSI amplitude corresponding to adjacent sampling times, and subtract the CSI amplitude of the previous sampling time from the CSI amplitude of the next sampling time to generate the CSI amplitude difference; select the CSI phase corresponding to adjacent sampling times, and subtract the CSI phase of the previous sampling time from the CSI phase of the next sampling time to generate the CSI phase difference; select the SINR corresponding to adjacent sampling times, and subtract the SINR of the previous sampling time from the SINR of the next sampling time to generate the SINR difference. The CSI amplitude difference, CSI phase difference, and SINR difference are arranged in order of sampling time.

[0021] In this embodiment, embedding encoding is performed on the current access node identifier and the candidate access node identifier to generate an access channel state vector, specifically as follows: Establish an access node identifier table, which records the one-to-one correspondence between access node identifiers and node indexes in the wireless network; Search the access node identifier table for the node index corresponding to the current access node identifier and the candidate access node identifier, and generate a node index sequence; The embedding parameter matrix includes multiple embedding parameter rows corresponding to node indices, and each embedding parameter row includes node representation values ​​of a fixed dimension. The embedding parameter rows corresponding to the node index sequence are extracted from the embedding parameter matrix to generate the embedding vector of the current access node and the embedding vector of the candidate access node. The access node embedding vectors are generated by arranging the current access node embedding vectors first and the candidate access node embedding vectors last; the CSI amplitude difference, CSI phase difference, SINR difference, and access node embedding vectors corresponding to the same sampling time are extracted. The vectors are concatenated according to the characteristic arrangement order of CSI amplitude difference, CSI phase difference, SINR difference, and access node embedded vector to generate a single-time access channel state vector. Arrange the single-time access channel state vectors in order of sampling time from early to late to generate the access channel state vector.

[0022] In this embodiment, the spectral mapping module specifically includes: The access channel state vector includes CSI amplitude difference, CSI phase difference, SINR difference, and access node embedding vector arranged according to sampling time. The access channel state vector is represented in higher dimensions using observable functions, which include CSI amplitude difference, CSI phase difference, SINR difference, access node embedding vector, squared CSI amplitude difference, squared CSI phase difference, squared SINR difference, product of CSI amplitude difference and SINR difference, and product of CSI phase difference and SINR difference. The observable function output values ​​are arranged in order from earliest to latest according to the sampling time to generate a channel observable state sequence; the access node embedding vector includes the current access node embedding vector and the candidate access node embedding vector; Calculate the vector difference between the candidate access node embedding vector and the current access node embedding vector, and generate access manifold coordinates; By appending the access manifold coordinates to the end of the channel observable state sequence corresponding to the same sampling time, a manifold channel observable sequence is generated. Select the manifold channel observable vectors corresponding to adjacent sampling times in the manifold channel observable sequence to generate the previous time observable vector set and the next time observable vector set; The Koopman operator algorithm is used to establish a linear transformation relationship between the set of observable vectors in the previous time step and the set of observable vectors in the next time step, generating a linear evolution matrix; Eigenvalue decomposition is performed on the linear evolution matrix to generate candidate spectral components and candidate invariant subspace features; The spectral change between candidate spectral components at adjacent sampling times is calculated using a manifold spectral stability check mechanism. The spectral change is obtained by subtracting the candidate spectral component at the previous sampling time from the candidate spectral component at the next sampling time and taking the absolute value. Candidate spectral components are arranged in ascending order of spectral change to generate a spectral component sequence; the candidate invariant subspace feature corresponding to the candidate spectral component with the smallest spectral change in the spectral component sequence is selected to generate the channel invariant subspace feature. Based on the channel-invariant subspace characteristics, a linear transformation matrix between adjacent sampling times is constructed to generate the channel evolution mapping relationship.

[0023] In this embodiment, an observable function is used to represent the access channel state vector in a higher dimension. This maps the nonlinear changes in CSI amplitude difference, CSI phase difference, SINR difference, and the access node embedding vector to a high-dimensional state space, reducing the impact of local channel disturbances on the channel state evolution process during wireless network handover. The observable function is constructed using the squared terms of CSI amplitude difference, CSI phase difference, SINR difference, the product of CSI amplitude difference and SINR difference, and the product of CSI phase difference and SINR difference. This enhances the ability to express the correlation between different channel state characteristics and improves the time-series representation of channel change trends. Simultaneously, the current access node is described using access manifold coordinates. The spatial migration relationship between the embedded vectors of incoming nodes and the embedded vectors of candidate access nodes can improve the continuous representation of the access node change process in handover scenarios. Furthermore, the Koopman operator algorithm is used to establish a linear transformation relationship between the set of observable vectors at previous time steps and the set of observable vectors at subsequent time steps, which can transform the complex channel evolution process into a linear evolution process and improve the stability of the channel evolution relationship. At the same time, the manifold spectral stability verification mechanism can filter out candidate spectral components with large spectral changes, which can reduce the spectral drift phenomenon caused by sudden channel fluctuations and improve the stability of the channel invariant subspace features. Constructing a channel evolution mapping relationship based on the channel invariant subspace features can improve the consistency of graph edge evolution in the subsequent handover channel graph construction process.

[0024] In this embodiment, the initial composition module specifically includes: Obtain the current access node identifier and candidate access node identifier from the access channel observation sequence, and obtain the wireless link identifier from the communication link record between the terminal and the access node; The current access node identifier, the candidate access node identifier, and the wireless link identifier are used as graph nodes to generate a graph node set; The communication relationship between the terminal and the access node corresponding to the current access node identifier at the same sampling time is recorded as the current connection relationship, and a connection graph edge is established between the graph node corresponding to the current access node identifier and the graph node corresponding to the wireless link identifier; the access relationship between the terminal and the access node corresponding to the candidate access node identifier at the same sampling time is recorded as the candidate access relationship, and a candidate access graph edge is established between the graph node corresponding to the candidate access node identifier and the graph node corresponding to the wireless link identifier; the linear transformation matrix corresponding to the channel evolution mapping relationship is recorded as the evolution graph edge attribute, and an evolution graph edge is established between the graph node corresponding to the current access node identifier and the graph node corresponding to the candidate access node identifier. Write the CSI amplitude difference, CSI phase difference, and SINR difference corresponding to the same sampling time into the edge attributes of the connection graph; write the CSI amplitude difference, CSI phase difference, and SINR difference corresponding to the same sampling time into the edge attributes of the candidate access graph; write the channel invariant subspace features and channel evolution mapping relationship into the edge attributes of the evolution graph. An initial switching channel graph is constructed from the set of graph nodes, connecting graph edges, candidate access graph edges, evolving graph edges, and the edge attributes corresponding to the connecting graph edges, candidate access graph edges, and evolving graph edges.

[0025] In this embodiment, the spectral verification module specifically includes: Extract candidate access graph edges, connection graph edges, and evolution graph edges from the initial handover channel graph; extract CSI amplitude difference, CSI phase difference, and SINR difference from the edge attributes of the candidate access graph edges, and arrange them in the order of sampling time to generate a candidate access edge channel difference sequence; extract the channel evolution mapping relationship from the edge attributes of the evolution graph edges, and arrange them in the order of sampling time to generate a candidate access edge evolution sequence. The CSI amplitude difference, CSI phase difference, and SINR difference in the candidate access side channel difference sequence are used to form a channel difference vector; the channel evolution mapping relationship in the candidate access side evolution sequence is applied to the channel difference vector corresponding to the same sampling time to generate an evolved channel difference vector; the difference between the evolved channel difference vector and the channel difference vector corresponding to the next sampling time is calculated to generate a spectral state offset vector. Take the absolute values ​​of the components corresponding to the CSI amplitude difference, CSI phase difference, and SINR difference in the spectral state offset vector, and sum them to generate the spectral state offset; arrange the candidate access graph edges in ascending order of spectral state offset to generate the candidate access graph edge sequence. The candidate access graph edge with the smallest spectral state offset in the candidate access graph edge sequence is selected, and the corresponding connection graph edge and evolution graph edge are retained to generate the switching channel graph.

[0026] In this embodiment, the improved Graph WaveNet model specifically includes a channel input layer, an adaptive adjacency generation layer, a diffusion graph convolutional layer, a gated dilation temporal convolutional layer, and a channel prediction decoding layer; The channel input layer arranges the set of graph nodes in the switching channel graph into a node sequence, and arranges the connecting graph edges, candidate access graph edges, and evolving graph edges into a graph edge sequence. It extracts CSI amplitude difference, CSI phase difference, SINR difference, channel invariant subspace features, and channel evolution mapping relationships from the edge attributes of the connecting graph edges, candidate access graph edges, and evolving graph edges. A graph structure input tensor is generated according to the correspondence between the graph node set, graph edge sequence, and edge attributes. The graph structure input tensor includes node dimension, sampling time dimension, and edge attribute dimension. Under the sequential communication mechanism of the signal-to-sense transition (STS) spectrum kernel, the adaptive adjacency generation layer concatenates the CSI amplitude difference, CSI phase difference, SINR difference, channel-invariant subspace features, and channel evolution mapping relationships corresponding to the same sampling time in the feature arrangement order to generate a single-time STS spectrum kernel vector. The single-time STS spectrum kernel vectors are arranged in ascending order of sampling time to generate a STS spectrum kernel sequence. The single-time STS spectrum kernel vectors corresponding to adjacent sampling times are selected, with the kernel vector corresponding to the previous sampling time placed first, and the kernel vector corresponding to the next sampling time placed last. The single-time signal migration kernel vectors are arranged to generate sequential connection vectors. The sequential connection vectors are then written into the edges of the connection graph, the candidate access graph, and the evolution graph, respectively, to generate the edges of the connection kernel graph, the candidate access kernel graph, and the evolution kernel graph. The inner vector products between the edges of the connection kernel graph and the candidate access kernel graph, the inner vector products between the candidate access kernel graph and the evolution kernel graph, and the inner vector products between the edges of the connection kernel graph and the evolution kernel graph are calculated to generate graph edge association values. The graph edge association values ​​are then filled into the matrix positions between the corresponding graph nodes to generate the adjacency matrix. The diffusion graph convolutional layer sums the row elements of the adjacency matrix and divides each row element by the sum of its corresponding row elements to generate the forward diffusion matrix; it sums the column elements of the adjacency matrix and divides each column element by the sum of its corresponding column elements to generate the backward diffusion matrix; the graph structure input tensor is multiplied by the forward diffusion matrix to generate the forward channel propagation feature; the graph structure input tensor is multiplied by the backward diffusion matrix to generate the backward channel propagation feature; the forward channel propagation feature and the backward channel propagation feature are concatenated along the feature dimension in the order of forward channel propagation feature first and backward channel propagation feature last to generate the spatial channel propagation feature. The gated dilated temporal convolutional layer uses dilated causal convolutional branches to process spatial channel propagation features and generate temporal convolutional features. The dilated causal convolutional branches take interval sampling points in the order of sampling time to participate in the convolution operation, and keep future sampling time points from participating in the convolution input of existing sampling time points. The gated convolutional branches are used to process spatial channel propagation features and generate gated modulation features. Each component in the gated modulation features corresponds to the modulation coefficient at the same position in the spatial channel propagation features. The components at the same position in the temporal convolutional features and the gated modulation features are multiplied together to generate the temporal channel propagation features. The channel prediction decoding layer maps the time channel propagation characteristics to CSI prediction sequences and SINR prediction sequences corresponding to future sampling times; it then arranges the CSI prediction sequences and SINR prediction sequences in the order of future sampling times to generate candidate channel state trajectories.

[0027] In this embodiment, both the improved Graph WaveNet model and the Graph WaveNet model employ graph structure modeling to describe the spatiotemporal relationships in wireless networks. Both also use a diffuse graph convolutional structure to extract spatial propagation features and a gated dilated temporal convolutional structure to extract temporal propagation features. The WaveNet model introduces a sequential communication mechanism with a signal-to-signal (SNS) kernel in its adaptive adjacency generation layer. This mechanism combines CSI amplitude difference, CSI phase difference, SINR difference, channel-invariant subspace features, and channel evolution mapping relationships into a SNS kernel vector. A SNS kernel sequence is constructed according to sampling times, enhancing the continuous correlation representation between adjacent sampling times during wireless network handover. Furthermore, a sequential communication vector is constructed using the single-time SNS kernel vectors corresponding to adjacent sampling times. This sequential communication vector is then written into the edges of the connection graph, candidate access graph, and evolution graph, simultaneously describing the handover relationship between access nodes, the propagation relationship of wireless links, and the channel evolution relationship, improving the adjacency matrix's ability to characterize the temporal sequence of dynamic handover processes. Simultaneously, the generation of graph edge correlation values ​​through the vector inner product between the edges of the connection kernel graph, candidate access kernel graph, and evolution kernel graph improves the continuous propagation capability of correlation relationships between graph nodes, reducing the instability of the adjacency matrix caused by abrupt changes in graph structure during wireless network handover. Furthermore, the graph structure is improved. The WaveNet model constructs forward and backward diffusion matrices in the diffusion graph convolutional layer, which can simultaneously describe the bidirectional diffusion relationship in the wireless channel propagation process and improve the integrity of spatial channel propagation features. At the same time, by performing convolution operations on interval sampling points through expanded causal convolution branches and restricting future sampling times from participating in the convolution input of existing sampling times, it can improve the predictive ability of temporal channel propagation features for future channel change trends. Furthermore, by performing position-corresponding modulation on temporal convolution features through gated modulation features, it can reduce the interference of local abnormal channel fluctuations on temporal propagation features and improve the continuous stability of candidate channel state trajectories.

[0028] In this embodiment, diffusion graph convolution calculation and gated dilation time convolution calculation are performed based on the adjacency matrix to generate candidate channel state trajectories, specifically as follows: The matrix elements in the adjacency matrix represent the graph edge association values ​​between graph nodes; summing the matrix elements in the same row of the adjacency matrix generates the row element sum value; Divide each element of the adjacency matrix by the sum of the elements in the corresponding row to generate a forward diffusion matrix; sum the elements in the same column of the adjacency matrix to generate a column element sum; divide each element of the adjacency matrix by the sum of the elements in the corresponding column to generate a backward diffusion matrix. Multiply the node features corresponding to the same sampling time in the graph structure input tensor with the forward diffusion matrix to generate forward channel propagation features; multiply the node features corresponding to the same sampling time in the graph structure input tensor with the reverse diffusion matrix to generate reverse channel propagation features; arrange the forward channel propagation features and reverse channel propagation features in the feature dimension according to the order of forward channel propagation features first and reverse channel propagation features last to generate spatial channel propagation features; The dilated causal convolution branch selects interval sampling points in the spatial channel propagation features according to the sampling time sequence. The distance between the interval sampling points is determined by the convolution dilation rate. The convolution kernel slides sequentially over the spatial channel propagation features corresponding to the interval sampling points to generate temporal convolution features. The gated convolution branch uses a convolution kernel at the same sampling position as the dilated causal convolution branch to process the spatial channel propagation features and generate gated modulation features; the feature component at each position in the temporal convolution feature is multiplied by the modulation coefficient at the same position in the gated modulation feature to generate the temporal channel propagation features. The channel prediction decoding layer maps the time channel propagation characteristics to the CSI prediction value and SINR prediction value corresponding to the future sampling time; arranges the CSI prediction value and SINR prediction value in the order of the future sampling time from early to late to generate the CSI prediction sequence and SINR prediction sequence; forms a trajectory node by combining the CSI prediction value and SINR prediction value corresponding to the same future sampling time; and arranges the trajectory node in the order of the future sampling time to generate the candidate channel state trajectory.

[0029] In this embodiment, the trajectory sequencing module specifically includes: The candidate channel state trajectory includes trajectory nodes arranged according to future sampling times. Each trajectory node includes a CSI prediction value and a SINR prediction value. The CSI prediction value and SINR prediction value corresponding to each future sampling time are extracted from the candidate channel state trajectory to generate a candidate trajectory node sequence. The channel evolution mapping relationship is applied to the trajectory node corresponding to the previous future sampling time in adjacent future sampling times to generate the reference trajectory node corresponding to the next future sampling time. The CSI evolution is generated by subtracting the predicted CSI value corresponding to the previous future sampling time from the CSI reference value in the reference trajectory node; the SINR evolution is generated by subtracting the predicted SINR value corresponding to the previous future sampling time from the SINR reference value in the reference trajectory node; the CSI evolution and SINR evolution are arranged in the order of future sampling times to generate the evolution reference sequence. Select the trajectory nodes corresponding to adjacent future sampling times in the candidate trajectory node sequence; subtract the CSI prediction value corresponding to the previous future sampling time from the CSI prediction value corresponding to the next future sampling time to generate the CSI prediction change; subtract the SINR prediction value corresponding to the previous future sampling time from the SINR prediction value corresponding to the next future sampling time to generate the SINR prediction change. Arrange the CSI prediction changes and SINR prediction changes in the order of future sampling times to generate a candidate trajectory change sequence; subtract the CSI evolution from the CSI prediction changes and take the absolute value to generate the CSI trajectory offset component; subtract the SINR evolution from the SINR prediction changes and take the absolute value to generate the SINR trajectory offset component; add the CSI trajectory offset component and the SINR trajectory offset component to generate the trajectory offset; arrange the candidate trajectory node sequence in ascending order of trajectory offset to generate an ordered trajectory node sequence; connect the CSI prediction values ​​and SINR prediction values ​​in the order of the trajectory nodes in the ordered trajectory node sequence to generate the channel state trajectory.

[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to a wireless network channel state prediction scenario in a vehicle-to-everything (V2X) test area on a city expressway. This area covers three roadside access nodes and an elevated road section of approximately 2.8 km. Test vehicles traveled continuously at speeds ranging from 40 km / h to 85 km / h, and the onboard terminals underwent multiple candidate access handovers between access nodes. During the test, the system collected the terminal's CSI, SINR, current access node identifier, and candidate access node identifiers at a 20 ms sampling period, generating an access channel observation sequence. Because the CSI amplitude and SINR fluctuate rapidly when vehicles pass through areas obstructed by bridge piers, noise barriers, and large buses, traditional single-link time-series prediction methods easily misjudge short-term fluctuations as long-term channel fading, leading to significant jumps in the predicted trajectory.

[0031] The channel acquisition module pairs the CSI, SINR, current access node identifier, and candidate access node identifier corresponding to the same sampling time into single-time access channel observation items, and arranges them in the order of sampling time to form an access channel observation sequence. The state coding module extracts the CSI amplitude difference, CSI phase difference, and SINR difference, and performs embedding coding on the current access node identifier and candidate access node identifier to generate an access channel state vector. The spectral mapping module introduces the access channel state vector into the Koopman operator algorithm of the manifold spectral stability verification mechanism to generate channel invariant subspace features and channel evolution mapping relationship. The initial graph construction module constructs an initial switching channel graph, and the spectral state verification module filters candidate access graph edges through spectral state offset to generate a switching channel graph. The trajectory generation module inputs the switching channel graph into the improved Graph WaveNet model, introduces the signal-to-transfer spectral kernel sequential connection mechanism in the adaptive adjacency generation layer to generate an adjacency matrix, and generates candidate channel state trajectories through diffusion graph convolution calculation and gated dilation time convolution calculation. The trajectory ordering module performs trajectory ordering processing on the candidate channel state trajectories based on the channel evolution mapping relationship to generate the channel state trajectory. The test selected six sets of data from the morning peak, midday off-peak, and evening peak, with each set lasting 12 minutes, collecting approximately 216,000 sampling points in total, and compared them with the traditional LSTM prediction method and the ordinary Graph WaveNet model.

[0032] Table 1 Comparison of Channel State Prediction Accuracy

[0033] As shown in Table 1, this invention exhibits lower CSI amplitude prediction MAE, CSI phase prediction MAE, and SINR prediction MAE across all three types of road wireless communication scenarios. Especially during frequent handover periods in the evening rush hour, where vehicles are significantly affected by obstructions, lane changes, and access node switching, the SINR prediction MAE of the LSTM prediction method reaches 2.58 dB. While the ordinary Graph WaveNet model can express access relationships using a graph structure, the number of trajectory jumps is still 38. This invention generates channel-invariant subspace features and channel evolution mapping relationships through a manifold spectral stability verification mechanism, and filters candidate access graph edges with the smallest spectral state offset using a spectral state verification module, making the handover channel graph more stable. Simultaneously, the sequential connection mechanism of the spectral kernel incorporates CSI amplitude difference, CSI phase difference, SINR difference, and channel evolution mapping relationships into the adjacency matrix generation process, making the channel propagation relationship between adjacent sampling times more continuous. Therefore, the number of trajectory jumps is reduced to 14 during frequent handover periods in the evening rush hour, demonstrating its adaptability to complex handover scenarios.

[0034] Table 2 Comparison of System Performance in Access Switching Scenarios

[0035] As shown in Table 2, this invention not only improves the accuracy of channel state prediction but also enhances its engineering application performance during wireless network access handover. The average prediction time for a single prediction using the ordinary Graph WaveNet model is 18.6 ms. Due to the addition of manifold spectral stability verification, spectral state verification, and trajectory ordering processing, the prediction time of this invention increases to 22.4 ms, but it still meets the real-time requirements of vehicular network channel state prediction in a 12-minute batch prediction task. Compared to the ordinary Graph WaveNet model, this invention reduces the average link interruption duration from 142.7 ms to 91.8 ms and increases the resource scheduling early hit rate from 78.6% to 88.9%, indicating that the channel state trajectory can reflect the changing trend of access nodes earlier. The channel state trajectory continuity rate reaches 95.4%, and the handover false trigger rate is reduced to 3.1%, indicating that the spectral state verification module and trajectory ordering module can suppress the interference of short-term abnormal fluctuations on candidate channel state trajectories, enabling the wireless resource scheduling unit and link handover unit to obtain more stable and continuous prediction data.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A wireless network channel state prediction system based on graph neural networks, characterized in that, include: The channel acquisition module is used to acquire the CSI, SINR, current access node identifier, and candidate access node identifier of the terminal at the sampling time, and generate the access channel observation sequence. The state coding module is used to extract CSI amplitude difference, CSI phase difference and SINR difference from the access channel observation sequence, perform embedding coding on the current access node identifier and the candidate access node identifier, and generate the access channel state vector. The spectral mapping module is used to introduce the access channel state vector input into the Koopman operator algorithm of the manifold spectral stability verification mechanism to generate channel invariant subspace features and channel evolution mapping relationship; The initial mapping module is used to construct an initial handover channel map based on the current access node identifier, candidate access node identifier, CSI amplitude difference, CSI phase difference, SINR difference, channel invariant subspace characteristics, and channel evolution mapping relationship; The spectral state verification module is used to perform spectral state verification on the candidate access relationships in the initial handover channel diagram based on the channel evolution mapping relationship, and generate the handover channel diagram. The trajectory generation module is used to input the switching channel graph into the improved Graph WaveNet model, introduce a signal-to-switching spectrum kernel sequential communication mechanism in the adaptive adjacency generation layer, generate an adjacency matrix, and perform diffusion graph convolution calculation and gated dilation time convolution calculation based on the adjacency matrix to generate candidate channel state trajectories. The trajectory ordering module is used to perform trajectory ordering processing on candidate channel state trajectories based on the channel evolution mapping relationship, and generate channel state trajectories.

2. The wireless network channel state prediction system based on graph neural networks according to claim 1, characterized in that, The channel acquisition module is specifically: Collect the CSI and SINR corresponding to the terminal according to the sampling time, and record the current access node identifier and candidate access node identifier corresponding to each sampling time; Pair CSI, SINR, current access node identifier, and candidate access node identifier according to the same sampling time to generate a single-time access channel observation item; Arrange the single-time access channel observation items in the order of sampling time to generate the access channel observation sequence.

3. The wireless network channel state prediction system based on graph neural networks according to claim 1, characterized in that, The extraction of CSI amplitude difference, CSI phase difference, and SINR difference based on the access channel observation sequence is specifically as follows: The CSI and SINR in the access channel observation sequence are arranged in chronological order of sampling time. The CSI is split using complex number decomposition to generate the CSI amplitude sequence and the CSI phase sequence; Select the CSI amplitude corresponding to adjacent sampling times, calculate the amplitude change between sampling times, and generate the CSI amplitude difference; Select the CSI phase corresponding to adjacent sampling times, calculate the phase change between sampling times, and generate the CSI phase difference; Select the SINR corresponding to adjacent sampling times, calculate the SINR change between sampling times, and generate the SINR difference.

4. The wireless network channel state prediction system based on graph neural networks according to claim 1, characterized in that, The process of performing embedding encoding on the current access node identifier and the candidate access node identifier to generate the access channel state vector is as follows: The current access node identifier and the candidate access node identifier are transformed by index mapping to generate a node index sequence; based on the embedding parameter matrix corresponding to the node index sequence, an access node embedding vector is generated. Extract the CSI amplitude difference, CSI phase difference, SINR difference, and access node embedding vector corresponding to the same sampling time. The access channel state vector is generated by concatenating the CSI amplitude difference, CSI phase difference, SINR difference, and access node embedding vector in the order of feature arrangement.

5. A wireless network channel state prediction system based on graph neural networks according to claim 1, characterized in that, The spectral mapping module is specifically as follows: Based on the CSI amplitude difference, CSI phase difference, SINR difference and access node embedding vector in the access channel state vector, the access channel state vector is mapped by an observable function to generate a channel observable state sequence. Access manifold coordinates are constructed from the access node embedding vectors, and the access manifold coordinates are written into the channel observable state sequence to generate the manifold channel observable sequence; The Koopman operator algorithm is used to perform a linear evolution mapping on the observable sequences of the manifold channel to generate candidate invariant subspace features; A manifold spectral stability verification mechanism is adopted to calculate the spectral change sequence between candidate invariant subspace features corresponding to adjacent sampling times; the spectral components are sorted according to the change amplitude of the spectral change sequence; the spectral component with the smallest change amplitude between sampling times is retained to generate channel invariant subspace features; Based on the channel-invariant subspace characteristics, the linear evolution relationship between adjacent sampling times is calculated to generate the channel evolution mapping relationship.

6. The wireless network channel state prediction system based on graph neural networks according to claim 1, characterized in that, The initial mapping module is specifically as follows: Generate a graph node set based on the current access node identifier, candidate access node identifier, and wireless link identifier; Generate connection graph edges based on the connection relationships corresponding to the current access node identifiers, generate candidate access graph edges based on the access relationships corresponding to the candidate access node identifiers, and generate evolution graph edges based on the channel evolution mapping relationship; Write the CSI amplitude difference, CSI phase difference, and SINR difference into the edge attributes of the connection graph and the candidate access graph; write the channel invariant subspace features and the channel evolution mapping relationship into the edge attributes of the evolution graph. An initial switching channel graph is constructed from the set of graph nodes, connecting graph edges, candidate access graph edges, evolving graph edges, and edge attributes.

7. The wireless network channel state prediction system based on graph neural networks according to claim 1, characterized in that, The spectral verification module is specifically as follows: Extract the CSI amplitude difference, CSI phase difference, and SINR difference corresponding to the candidate access graph edges to generate the candidate access edge channel difference sequence; extract the channel evolution mapping relationship corresponding to the evolution graph edges to generate the candidate access edge evolution sequence; Calculate the spectral state offset between the candidate access side channel differential sequence and the candidate access side evolution sequence; Candidate access graph edges are arranged in ascending order of spectral state offset to generate a candidate access graph edge sequence. The candidate access graph edge with the smallest spectral state offset in the candidate access graph edge sequence is selected, and the corresponding connection graph edge and evolution graph edge are retained to generate the switching channel graph.

8. A wireless network channel state prediction system based on graph neural networks according to claim 1, characterized in that, The improved Graph WaveNet model specifically includes a channel input layer, an adaptive adjacency generation layer, a diffusion graph convolutional layer, a gated dilation temporal convolutional layer, and a channel prediction decoding layer; The channel input layer converts the set of graph nodes, connecting graph edges, candidate access graph edges, evolving graph edges, and edge attributes in the switching channel graph into a graph structure input tensor. The edge attributes include CSI amplitude difference, CSI phase difference, SINR difference, channel invariant subspace features, and channel evolution mapping relationship. Under the sequential communication mechanism of the signal migration spectrum kernel, the adaptive adjacency generation layer combines CSI amplitude difference, CSI phase difference, SINR difference, channel invariant subspace features and channel evolution mapping relationship into a signal migration spectrum kernel vector; Arrange the signal migration spectrum kernel vectors in the order of sampling time to generate a signal migration spectrum kernel sequence; concatenate the signal migration spectrum kernel vectors corresponding to adjacent sampling times to generate a sequential connection vector; write the sequential connection vector into the edges of the connection graph, the candidate access graph, and the evolution graph respectively to form the edges of the connection spectrum kernel graph, the candidate access spectrum kernel graph, and the evolution spectrum kernel graph; calculate the graph edge association values ​​between the edges of the connection spectrum kernel graph, the candidate access spectrum kernel graph, and the evolution spectrum kernel graph, and arrange them according to the graph edge association values ​​to generate an adjacency matrix; The diffusion graph convolutional layer constructs a forward diffusion matrix and a backward diffusion matrix based on the adjacency matrix. The graph structure input tensor is multiplied by the forward diffusion matrix and the backward diffusion matrix respectively to generate forward channel propagation features and backward channel propagation features. The forward channel propagation features and backward channel propagation features are then concatenated to generate spatial channel propagation features. The gated dilated temporal convolutional layer inputs the spatial channel propagation features into the dilated causal convolutional branch and the gated convolutional branch to generate temporal convolutional features and gated modulation features, and multiplies the temporal convolutional features and gated modulation features to generate temporal channel propagation features; The channel prediction decoding layer maps the time channel propagation characteristics to the CSI prediction sequence and SINR prediction sequence corresponding to the future sampling time, generating candidate channel state trajectories.

9. A wireless network channel state prediction system based on graph neural networks according to claim 1, characterized in that, The process of performing diffusion graph convolution calculation and gated dilation time convolution calculation based on the adjacency matrix to generate candidate channel state trajectories is as follows: Extract the CSI prediction sequence and SINR prediction sequence from the candidate channel state trajectory; arrange the CSI prediction sequence and SINR prediction sequence according to the future sampling time order to generate a time-series channel trajectory sequence; Extract the CSI and SINR prediction values ​​corresponding to adjacent future sampling times from the time-series channel trajectory sequence; calculate the CSI trajectory change and SINR trajectory change between adjacent future sampling times; arrange the trajectory nodes in the time-series channel trajectory sequence according to the CSI trajectory change and SINR trajectory change. The trajectory nodes with continuously changing CSI trajectory changes and the trajectory nodes with continuously changing SINR trajectory changes are sequentially connected to generate continuous trajectory segments; the continuous trajectory segments are then spliced ​​together according to the order of future sampling times to generate the channel state trajectory.

10. A wireless network channel state prediction system based on a graph neural network according to claim 1, characterized in that, The trajectory sequencing module specifically comprises: Extract the CSI prediction sequence and SINR prediction sequence corresponding to the future sampling time from the candidate channel state trajectory to generate the candidate trajectory node sequence; The CSI evolution and SINR evolution between adjacent future sampling times are calculated based on the channel evolution mapping relationship to generate an evolution reference sequence; Calculate the CSI prediction change and SINR prediction change between adjacent trajectory nodes in the candidate trajectory node sequence to generate a candidate trajectory change sequence; Calculate the trajectory offset between the candidate trajectory change sequence and the evolution reference sequence; Arrange the candidate trajectory nodes in ascending order of trajectory offset to generate an ordered trajectory node sequence; connect the trajectory nodes in the ordered trajectory node sequence to generate the channel state trajectory.

Images