CSI prediction method based on influence of GCN-LSTM on multiple factors
By adopting the GCN-LSTM model in CSI prediction, combining wavelet transform and FIR filter technology to process low-frequency and high-frequency signals in channel state information, the problem of degradation of prediction accuracy in high-speed movement and multipath propagation environments is solved, and higher CSI prediction accuracy and a more comprehensive channel model are achieved.
Patent Information
- Application Number
- CN202411842018.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-13
- Publication Date
- 2025-03-21
AI Technical Summary
The existing CSI prediction model has reduced prediction accuracy in high-speed movement and multipath propagation environments, and cannot effectively capture the nonlinear correlation of the channel, especially the high-frequency aliasing components that interfere with the prediction.
Using a GCN-LSTM-based method, wavelet transform and decompose the historical channel state information, low-frequency and high-frequency signals are extracted, and high-frequency signals are processed through FIR filters, a GCN model with a graph structure is constructed for feature extraction, and a channel state information is predicted by combining the LSTM model.
It significantly improves the prediction accuracy of CSI, can effectively process high-frequency aliased data caused by multipath effect, ensures accurate prediction in complex channel environments, and optimizes the performance of the Internet of Things mobile communication system.
Smart Images
Figure CN119675808A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of wireless communications, and in particular to a CSI prediction method under the influence of multiple factors based on GCN-LSTM. Background Art
[0002] The sixth generation (6G) wireless communication technology is developing towards a faster, smarter and wider direction. Channel status information (CSI) has a serious negative impact on the performance of various adaptive transmission systems. Channel characteristics have become more dynamic and unpredictable, which puts higher requirements on existing channel technology.
[0003] The existing classic statistical linear prediction model has been widely used in CSI prediction. However, with the large-scale access of mobile terminals, channel state information often exhibits unstable characteristics such as nonlinearity and randomness. Especially in high-speed mobile scenarios and multipath propagation environments, channel characteristics become more dynamic and unpredictable. The prediction accuracy of traditional linear prediction models will drop significantly in such highly dynamic environments. At the same time, linear models often assume that the correlation between channels is linear, while actual channels have complex nonlinear correlations. These subtle correlations cannot be effectively captured by linear models.
[0004] Considering that the construction of prediction models must rely on more powerful data processing capabilities, deep learning has become the mainstream solution for network traffic prediction with its excellent data feature extraction capabilities. In CSI prediction, existing technologies do not take into account that channel state information is also affected by factors such as multipath effects and interference noise. Due to the complex propagation environment of the IoT mobile communication system and the dense deployment of IoT terminal devices, high-frequency aliasing components appear in the channel state information. The existence of these high-frequency components brings challenges to channel estimation and signal detection, making the existing prediction models still have limitations and cannot meet the needs of mobile Internet. Summary of the invention
[0005] In view of the above-mentioned deficiencies in the prior art, the present invention provides a CSI prediction method under the influence of multiple factors based on GCN-LSTM, which solves the problem that high-frequency aliasing components in channel state information will affect the CSI prediction accuracy.
[0006] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:
[0007] A CSI prediction method based on GCN-LSTM under the influence of multiple factors is provided, which includes the following steps:
[0008] S1, decomposing the historical channel state information time series data through a wavelet transform filter to obtain a low-frequency signal and a high-frequency signal with the same time span;
[0009] S2, performing discrete-time Fourier transform on the high-frequency signal and inputting the signal into the FIR filter, and performing discrete-time inverse Fourier transform on the output of the FIR filter to obtain a time-domain expression of the high-frequency signal;
[0010] S3, taking the features of the channel state information time series data as the nodes of the graph structure, obtaining the adjacency matrix between the nodes according to the correlation between the nodes of the graph structure, and then constructing the GCN model of the graph structure;
[0011] S4, taking the time domain expressions of the low-frequency signal and the high-frequency signal as the input of the GCN model of the graph structure, respectively, to obtain the feature vector corresponding to the low-frequency signal and the feature vector corresponding to the time domain expression of the high-frequency signal;
[0012] S5. Using the feature vector corresponding to the low-frequency signal as the input of the trained first LSTM model, and using the output of the trained first LSTM model as the channel state information corresponding to the low-frequency signal; using the feature vector corresponding to the time domain expression of the high-frequency signal as the input of the trained second LSTM model, and using the output of the trained second LSTM model as the channel state information corresponding to the high-frequency signal;
[0013] S6. Taking the sum of the channel state information corresponding to the low-frequency signal and the channel state information corresponding to the high-frequency signal as the final channel state information prediction result.
[0014] Furthermore, the specific method of decomposing the historical channel state information time series data by wavelet transform filter in step S1 is:
[0015] The discrete wavelet transform is performed on the time series data of the historical channel state information through the wavelet transform filter to obtain the wavelet coefficients at different scales. The wavelet coefficients at different scales are divided into low-frequency part and high-frequency part by selecting the scale factor to obtain the low-frequency signal and high-frequency signal with the same time span.
[0016] Further, the cutoff frequency and order of the FIR filter are determined by the characteristics of the channel state information data.
[0017] Furthermore, the cutoff frequency of the FIR filter is 2.53 GHz, and the order of the FIR filter is 3.
[0018] Furthermore, the characteristics of the channel state information time series data in step S3 are the real part, real part fading, imaginary part, imaginary part fading, amplitude, phase and signal power of the channel state information time series data.
[0019] Furthermore, in step S3, the specific method for obtaining the adjacency matrix between nodes according to the correlation between nodes in the graph structure is:
[0020] Calculate the Pearson correlation coefficient between the nodes of the graph structure, set the adjacency matrix between two nodes with a correlation coefficient lower than 0.6 to 0, and set the adjacency matrix between two nodes with a correlation coefficient greater than or equal to 0.6 to 1.
[0021] A computer device is provided, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes a CSI prediction method under the influence of multiple factors based on GCN-LSTM.
[0022] A computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, the processor executes a CSI prediction method under the influence of multiple factors based on GCN-LSTM.
[0023] The beneficial effects of the present invention are:
[0024] 1. The present invention significantly improves the prediction accuracy of CSI by combining the powerful feature extraction capabilities of the GCN model and the LSTM model. Compared with the traditional model, the overall prediction accuracy is improved, and the performance of the IoT mobile communication system can be optimized.
[0025] 2. The present invention can effectively process high-frequency aliasing data caused by multipath effects. Through wavelet transform and FIR filter technology, clear low-frequency time domain components (low-frequency signals) and high-frequency frequency domain components (high-frequency signals) are extracted, ensuring accurate prediction in complex channel environments and avoiding the influence of high-frequency aliasing components on the CSI prediction accuracy.
[0026] 3. The present invention constructs a more comprehensive prediction model by considering the correlation between channel state information and multiple factors such as amplitude, phase, and signal power, which can more accurately capture the fluctuations and trends of CSI. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 is a schematic diagram of the process of this method;
[0028] Figure 2 This is a schematic diagram of data processing of this method;
[0029] Figure 3 This is the performance comparison of the mean square error (MSE) of the real part of CSI of different schemes in the ablation experiment;
[0030] Figure 4 This is a performance comparison chart of the mean square error (MSE) of the CSI imaginary part of different schemes in the ablation experiment;
[0031] Figure 5 This is a comparison diagram of the CSI real part prediction effect in the comparative experiment;
[0032] Figure 6 This is a comparison diagram of the CSI imaginary part prediction effect in the comparative experiment. DETAILED DESCRIPTION
[0033] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.
[0034] like Figure 1 and Figure 2 As shown, the CSI prediction method based on GCN-LSTM under the influence of multiple factors includes the following steps:
[0035] S1, decomposing the historical channel state information time series data through a wavelet transform filter to obtain a low-frequency signal and a high-frequency signal with the same time span;
[0036] S2, performing discrete-time Fourier transform on the high-frequency signal and inputting the signal into the FIR filter, and performing discrete-time inverse Fourier transform on the output of the FIR filter to obtain a time-domain expression of the high-frequency signal;
[0037] S3, taking the features of the channel state information time series data as the nodes of the graph structure, obtaining the adjacency matrix between the nodes according to the correlation between the nodes of the graph structure, and then constructing the GCN model of the graph structure;
[0038] S4, taking the time domain expressions of the low-frequency signal and the high-frequency signal as the input of the GCN model of the graph structure, respectively, to obtain the feature vector corresponding to the low-frequency signal and the feature vector corresponding to the time domain expression of the high-frequency signal;
[0039] S5. Using the feature vector corresponding to the low-frequency signal as the input of the trained first LSTM model, and using the output of the trained first LSTM model as the channel state information corresponding to the low-frequency signal; using the feature vector corresponding to the time domain expression of the high-frequency signal as the input of the trained second LSTM model, and using the output of the trained second LSTM model as the channel state information corresponding to the high-frequency signal;
[0040] S6. Taking the sum of the channel state information corresponding to the low-frequency signal and the channel state information corresponding to the high-frequency signal as the final channel state information prediction result.
[0041] A computer device is provided, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes a CSI prediction method under the influence of multiple factors based on GCN-LSTM.
[0042] A computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, the processor executes a CSI prediction method under the influence of multiple factors based on GCN-LSTM.
[0043] The specific method of decomposing the historical channel state information time series data by wavelet transform filter in step S1 is:
[0044] The discrete wavelet transform is performed on the time series data of the historical channel state information through the wavelet transform filter to obtain the wavelet coefficients at different scales. The wavelet coefficients at different scales are divided into low-frequency part and high-frequency part by selecting the scale factor to obtain the low-frequency signal and high-frequency signal with the same time span.
[0045] In the specific implementation process, the scale factor can be discretized in power order in combination with the characteristics of the channel state information data, and the translation factor τ can be uniformly discretized. The wavelet transform after discretization can be expressed as:
[0046]
[0047] where ψ j,k (t) represents the channel state information at time t; j represents the scale parameter, which is used to control the scale or frequency of the wavelet function; k represents the position parameter, which is used to control the translation position of the wavelet transform; τ0 represents the translation factor.
[0048] In this embodiment, the cutoff frequency and order of the FIR filter are determined by the characteristics of the channel state information data. The FIR filter multiplies the signal obtained by the discrete time Fourier transform by the transfer function, outputs the expression of the high-frequency signal in the frequency domain, and then obtains the time domain expression of the high-frequency signal by the discrete time Fourier inverse transform. In the specific implementation process, the order of the FIR filter is 3, the center frequency of the system is 2.61GHz, the bandwidth is 20MHz, and according to the Nyquist sampling theorem, the sampling frequency is set to at least 40MHz, and the cutoff frequency of the FIR filter should be at least 2.53GHz.
[0049] In the process of constructing the graph structure, CSI can be defined as a graph G = (V, E, A), where V = {v1, v2, …, v N} represents the set of all nodes; v i ∈V represents the i-th node; e ij ∈E represents node v i To node v jThe edge of ; A is the adjacency matrix, which is defined as the adjacent relationship between different nodes. The feature matrix of node V is expressed as M is the total number of node features, and the node features are the real part, real part fading, imaginary part, imaginary part fading, amplitude, phase and signal power of the channel state information time series data. On this basis, the specific method of obtaining the adjacency matrix between nodes according to the correlation between nodes of the graph structure in step S3 is: calculate the Pearson correlation coefficient between the nodes of the graph structure, set the adjacency matrix between two nodes with a correlation coefficient less than 0.6 to 0, and set the adjacency matrix between two nodes with a correlation coefficient greater than or equal to 0.6 to 1.
[0050] In the specific implementation process, the GCN model is mainly composed of convolutional layers, pooling layers and fully connected layers. In order to make the GCN network nonlinear and stable, the present invention uses the ReLU activation function and adds a Dropout layer after the ReLU activation layer to prevent overfitting, thereby achieving a regularization effect. The graph structure is constructed by nodes and adjacency matrices. The convolution operation of the GCN model transfers information based on the connection relationship between nodes. Each node uses the features of its neighboring nodes to update its own feature representation. Through multi-layer convolution operations, the GCN model can perform multiple information aggregation and feature extraction on the graph structure, thereby better capturing the relationship and complex patterns between nodes in the graph. The operator of graph convolution can be expressed as:
[0051]
[0052] Where X is the node features in the input data, namely the real part, real part decay, imaginary part, imaginary part decay, amplitude, phase and signal power in the time domain expression of low-frequency signal and high-frequency signal; It is an approximate graph convolution filter; The degree matrix is used to represent the number of connections of a node. It can represent the importance of a node in the graph and is a diagonal matrix. W represents the weight matrix that needs to be trained on the neural network. ReLU(.) is the ReLU activation function. I N is an N-dimensional identity matrix; GCN(A,X) represents the output of graph convolution.
[0053] The output of the GCN model is a feature vector containing the corresponding information of the real part, real part fading, imaginary part, imaginary part fading, amplitude, phase and signal power. The feature vector is input into the LSTM model, and the time series is predicted through the forget gate and input gate mechanism of the LSTM model, so as to effectively predict and memorize the long-term dependencies in the CSI data, and then effectively predict the CSI.
[0054] The forget gate of the LSTM model can be expressed as:
[0055] ft =σ(ω f ·[h t-1 ,x t ]+b f )
[0056] where x t is the input feature at the current time t; h t-1 is the result output at time t-1; ω f and b f are the forget gate weight matrix and bias vector respectively; σ(.) is the Sigmiod function; f t is the activation vector of the forget gate at time t. The closer its value is to 0, the more information is forgotten.
[0057] The memory cell is the core component of the LSTM model, which is similar to a memory used to store and manage information in the network. The memory cell allows the LSTM model to memorize information for a long time and pass the information to other time steps when needed, effectively solving the problems of gradient vanishing or exploding defects in recurrent neural networks (RNNs) in long sequences. How much information to store in the memory cell is determined by the input gate, which can be expressed as:
[0058] i t =σ(ω i ·[h t-1 ,x t ]+b i )
[0059]
[0060] where ω i and ω c are the weight matrices of the input gate; b i and b c is the bias vector of the input gate; tanh(.) is the hyperbolic tangent activation function; i t is the activation vector of the input gate. The larger its value is, the more important the new information is, and the unimportant information will be deleted. is the current memory unit. Before the input gate outputs, it will determine the updated information, which can be expressed as:
[0061]
[0062] Among them C t is the new memory unit output at time t; C t-1 is the new memory unit output at time t-1. After determining the information, the output gate will output to the external state, which can be expressed as:
[0063] ot =σ(ω o ·[h t-1 ,x t ]+b o )
[0064] where ω o and b o are the weight matrix and bias vector of the output gate respectively.
[0065] In one embodiment of the present invention, in order to verify the effectiveness of the present method, an ablation experiment is conducted on the present method (GCN-LSTM) and the prediction performance is evaluated with the models mentioned in the present method (GCN model and LSTM model). According to the architecture of the model, the data is decomposed into low-frequency data and high-frequency data, and prediction simulation is performed in four aspects such as the real part / imaginary part of the low-frequency and high-frequency data sets. The real part of the CSI data represents the strength or energy of the signal, which determines the degree of attenuation of the signal during transmission. The imaginary part reflects the phase change of the signal during transmission, which is related to the delay and frequency of the signal. At the same time, 80% of the historical channel state information data is used as the training data of the model, and the remaining 20% is used as the test set of the model.
[0066] like Figure 3 and Figure 4 As shown in the figure, in the MSE performance comparison of the real and imaginary parts of CSI of different schemes in the ablation experiment, this method can reach the maximum convergence the fastest. The simulation results further prove that this method can achieve good prediction accuracy for the prediction results of the real and imaginary parts of CSI data, showing strong adaptability and generalization ability.
[0067] like Figure 5 and Figure 6 As shown in the figure, the predicted values of the real and imaginary parts of CSI in the comparative experiment are almost consistent with the actual values. This prediction result also confirms that the method of using wavelet transform filter to decompose CSI data and then predicting low-frequency signals and high-frequency signals respectively and adding them as the final result is feasible. This method improves the accuracy of CSI prediction while maintaining the integrity of the data.
Claims
1. A CSI prediction method based on GCN-LSTM under the influence of multiple factors, characterized in that: The following steps are involved: S1, decomposing the historical channel state information time series data through a wavelet transform filter to obtain a low-frequency signal and a high-frequency signal with the same time span; S2, performing discrete-time Fourier transform on the high-frequency signal and inputting the signal into the FIR filter, and performing discrete-time inverse Fourier transform on the output of the FIR filter to obtain a time-domain expression of the high-frequency signal; S3, taking the features of the channel state information time series data as the nodes of the graph structure, obtaining the adjacency matrix between the nodes according to the correlation between the nodes of the graph structure, and then constructing the GCN model of the graph structure; S4, taking the time domain expressions of the low-frequency signal and the high-frequency signal as the input of the GCN model of the graph structure, respectively, to obtain the feature vector corresponding to the low-frequency signal and the feature vector corresponding to the time domain expression of the high-frequency signal; S5. Using the feature vector corresponding to the low-frequency signal as the input of the trained first LSTM model, and using the output of the trained first LSTM model as the channel state information corresponding to the low-frequency signal; using the feature vector corresponding to the time domain expression of the high-frequency signal as the input of the trained second LSTM model, and using the output of the trained second LSTM model as the channel state information corresponding to the high-frequency signal; S6. Taking the sum of the channel state information corresponding to the low-frequency signal and the channel state information corresponding to the high-frequency signal as the final channel state information prediction result.
2. According to claim 1, the CSI prediction method based on GCN-LSTM under the influence of multiple factors is characterized in that: The specific method of decomposing the historical channel state information time series data by wavelet transform filter in step S1 is: The discrete wavelet transform is performed on the time series data of the historical channel state information through the wavelet transform filter to obtain the wavelet coefficients at different scales. The wavelet coefficients at different scales are divided into low-frequency part and high-frequency part by selecting the scale factor to obtain the low-frequency signal and high-frequency signal with the same time span.
3. The CSI prediction method under the influence of multiple factors based on GCN-LSTM according to claim 1 is characterized in that: The cutoff frequency and order of the FIR filter are determined by the characteristics of the channel state information data.
4. The CSI prediction method under the influence of multiple factors based on GCN-LSTM according to claim 3 is characterized in that: The cutoff frequency of the FIR filter is 2.53 GHz, and the order of the FIR filter is 3.
5. The CSI prediction method under the influence of multiple factors based on GCN-LSTM according to claim 1 is characterized in that: The characteristics of the channel state information time series data in step S3 are the real part, real part fading, imaginary part, imaginary part fading, amplitude, phase and signal power of the channel state information time series data.
6. The CSI prediction method under the influence of multiple factors based on GCN-LSTM according to claim 1 is characterized in that: The specific method of obtaining the adjacency matrix between nodes according to the correlation between nodes in the graph structure in step S3 is: Calculate the Pearson correlation coefficient between the nodes of the graph structure, set the adjacency matrix between two nodes with a correlation coefficient lower than 0.6 to 0, and set the adjacency matrix between two nodes with a correlation coefficient greater than or equal to 0.6 to 1.
7. A computer device, characterized in that: The invention comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the CSI prediction method under the influence of multiple factors based on GCN-LSTM as described in any one of claims 1 to 6.
8. A computer-readable storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a processor, the processor executes the CSI prediction method under the influence of multiple factors based on GCN-LSTM as described in any one of claims 1 to 6.