A complex channel state prediction method based on time-frequency difference dynamic evolution

Abstract

The application relates to the technical field of wireless communication and artificial intelligence signal processing, and discloses a complex channel state prediction method based on time-frequency difference dynamic evolution, which comprises the following steps: S1, acquiring complex time-frequency features obtained by constructing or preprocessing historical complex channel state information; S2, performing causal exponential moving average smoothing on the complex time-frequency features along the time dimension to obtain smoothed complex features; S3, performing difference along the time and frequency dimensions to obtain a complex domain double-flow difference result; S4, calculating the difference energy of the two paths and constructing a dynamic intensity map; S5, inputting the lightweight two-dimensional convolution network to obtain a dynamic mapping map; S6, generating an adaptive gating coefficient according to the dynamic mapping map; S7, weighting the original complex time-frequency features by using the coefficient to obtain enhanced features; and S8, inputting the enhanced features into a prediction structure to output a channel prediction result. The complex channel state prediction method based on time-frequency difference dynamic evolution has the advantages of sufficient time-frequency joint dynamic feature extraction and strong non-stationary channel evolution representation capability.

Description

Technical Field

[0001] This invention relates to the fields of wireless communication and artificial intelligence signal processing technology, specifically to a complex channel state prediction method based on time-frequency differential dynamic evolution. Background Technology

[0002] Channel state information (CSO) is a core parameter characterizing the transmission characteristics of wireless communication links. It reflects the changing patterns of the channel in time, frequency, and space dimensions, and is widely used in communication processing such as channel prediction, precoding, beamforming, link adaptation, and wireless resource scheduling. In multi-antenna and multi-carrier wireless communication systems, CSO is usually represented in complex form, with its real and imaginary parts jointly determining the amplitude and phase characteristics of the channel. At the same time, the channel state exhibits differentiated changing characteristics with time evolution and frequency distribution. As mobile communication systems continue to develop towards higher speeds, lower latency, and higher reliability, accurately predicting the channel state at future moments has become an important technical direction for improving the timeliness of wireless link decision-making and transmission reliability.

[0003] Existing channel state prediction technologies are mainly divided into two implementation paths: model-driven and data-driven. Model-driven methods are usually based on mathematical models such as autoregression and Kalman filtering to describe the channel evolution process, and have a corresponding applicable basis in scenarios where channel changes are relatively stable and statistical characteristics are easy to match. Data-driven methods use deep learning networks to learn the nonlinear mapping relationship between historical channels and future channels, and can adapt to the channel prediction needs in complex propagation environments to a certain extent.

[0004] However, current technologies are somewhat inconvenient for fine-grained modeling of joint time-frequency dynamic features: First, they do not pay enough attention to local structural changes in the frequency dimension, making it difficult to fully reflect the joint dynamic evolution characteristics of the channel in the time-frequency plane; second, they lack explicit measurement and adaptive calibration mechanisms for the intensity of local dynamic changes in the channel, and their response capability to local abrupt changes and short-term disturbances in non-stationary channel scenarios needs to be improved; third, some complex channel processing methods treat the real and imaginary parts as independent features, which is not conducive to maintaining the intrinsic structural correlation between channel amplitude and phase. Therefore, a complex channel state prediction method based on time-frequency differential dynamic evolution is proposed to solve the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a complex channel state prediction method based on time-frequency differential dynamic evolution. It has the advantages of sufficient extraction of time-frequency joint dynamic features and strong ability to characterize non-stationary channel evolution, and solves the problem that existing methods do not make sufficient use of non-stationary dynamic changes, local time-frequency disturbances and complex amplitude-phase structures.

[0006] The technical solution of this invention to solve the above-mentioned technical problems is as follows: A complex channel state prediction method based on time-frequency differential dynamic evolution, comprising the following steps: S1. Obtain complex time-frequency characteristics, wherein the complex time-frequency characteristics are constructed from historical complex channel state information, or obtained from historical complex channel state information after time-frequency preprocessing; S2. Perform causal exponential moving average smoothing on the complex time-frequency features along the time dimension to obtain smoothed complex features; S3. Calculate the time difference result along the time direction and the frequency difference result along the frequency direction for the smoothed complex features, respectively, to obtain the complex domain dual-stream difference result containing both time difference and frequency difference features; S4. Calculate the time difference energy corresponding to the time difference result and the frequency difference energy corresponding to the frequency difference result respectively, and construct a dynamic intensity map based on the time difference energy and the frequency difference energy; S5. Input the dynamic intensity map into a lightweight two-dimensional convolutional mapping network for mapping processing, and output a dynamic mapping map. S6. Generate adaptive gating coefficients based on the dynamic mapping graph; S7. The adaptive gating coefficients are weighted element by element to the complex time-frequency feature to obtain the complex enhanced feature; S8. Input the complex enhancement features into the prediction mapping structure and output the complex channel state prediction results for future time periods.

[0007] The beneficial effects of this invention are: 1) The complex channel state prediction method based on time-frequency differential dynamic evolution has the advantages of sufficient extraction of time-frequency joint dynamic features and strong ability to characterize non-stationary channel evolution.

[0008] 2) The complex channel state prediction method based on time-frequency differential dynamic evolution has the advantages of maintaining the integrity of the complex amplitude and phase structure, high stability of dynamic estimation under causal constraints, accurate adaptive differential calibration, and good suppression of original feature disturbances.

[0009] Based on the above technical solution, the present invention can be further improved as follows.

[0010] Furthermore, in step S2, the causal exponential moving average smoothing process calculates the smoothed complex feature at the t-th time position using the following formula. :

[0011] in, This is the smoothing coefficient, and its value range is... When t=0, the smoothed complex feature of the initial time position is equal to the input complex time-frequency feature of that position; the smoothing process only uses the feature data of the current and past time positions, and does not introduce the feature data of future time positions.

[0012] Furthermore, in step S3, the time difference result is the difference between the smoothed complex features corresponding to adjacent time positions; the frequency difference result is the difference between the smoothed complex features corresponding to adjacent frequency positions; for the time difference result of the first time position and the frequency difference result of the first frequency position, the difference is calculated using the smoothed complex features of the current position itself as the preceding reference.

[0013] Furthermore, step S4 specifically includes: S4.1. The average of the sum of the squares of the real and imaginary parts of the time difference results along the feature dimension is used to obtain the time difference energy. S4.2. The average of the sum of the squares of the real and imaginary parts of the frequency difference results along the feature dimension is taken to obtain the frequency difference energy. S4.3. Perform a logarithmic operation on the sum of the time difference energy and the frequency difference energy to obtain the dynamic intensity map.

[0014] Furthermore, in step S5, the lightweight two-dimensional convolutional mapping network includes at least one two-dimensional convolutional layer and a nonlinear activation function; the operation of the two-dimensional convolutional layer is performed on a time-frequency plane composed of time and frequency dimensions.

[0015] Furthermore, in step S6, the adaptive gating coefficient is generated according to the following formula:

[0016] in, Indicates the gating coefficient. This represents the Sigmoid function. This indicates an adjustable gating strength coefficient. This represents a dynamic mapping graph, where the adaptive gating coefficients are real-valued coefficients.

[0017] Furthermore, in step S7, the adaptive gating coefficient is multiplied element-wise with the complex time-frequency feature; at the same time-frequency position, the adaptive gating coefficient is applied simultaneously to the real and imaginary parts of the complex time-frequency feature at that position with the same real value.

[0018] Furthermore, in step S1, the complex time-frequency feature includes a batch dimension, a frequency dimension, a time dimension, and a feature dimension; the complex time-frequency feature maintains its complex form during the processing of steps S2 to S7, and the real part and imaginary part are not separated into independent channels for processing. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the structure of the present invention; Figure 2 This is a flowchart illustrating the specific steps of step S4 in this invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1 addresses the complex channel state prediction scenario in a multiple-input multiple-output (MIMO) orthogonal frequency division multiplexing (OFDM) wireless communication system. It is applicable to complex wireless environments characterized by high-speed mobility, frequent multipath scattering changes, and frequent obstruction switching. Through the time-frequency differential dynamic evolution processing flow of this example, local dynamic changes in the time-frequency two-dimensional plane can be explicitly extracted while maintaining the consistency of the amplitude and phase structure of the complex channel characteristics. This effectively improves the channel prediction's response to non-stationary changes, local abrupt changes, and frequency-selective fading, while reducing interference from local random disturbances. This makes the prediction features more suitable for the channel state prediction requirements in complex dynamic wireless environments.

[0022] Depend on Figure 1-2 A method for predicting the state of a complex channel based on time-frequency differential dynamic evolution is presented, comprising the following steps: S1. Obtain complex time-frequency characteristics The complex time-frequency characteristics are obtained by constructing them from historical complex channel state information or by obtaining them from historical complex channel state information after time-frequency preprocessing.

[0023] Complex time-frequency features include batch dimension, frequency dimension, time dimension, and feature dimension; let the input complex time-frequency feature be denoted as X, and its dimensions are represented as follows: Where B represents the batch size, F represents the frequency dimension length, T represents the time dimension length, and D represents the feature dimension; The complex time-frequency characteristic consists of a real part and an imaginary part, and can be expressed as follows: ,in, Indicates the real part, The real part represents the imaginary part, and j represents the imaginary unit; the real part and the imaginary part together determine the amplitude and phase of the channel characteristics.

[0024] The complex time-frequency features remain in complex form throughout the subsequent steps S2 to S7. The real and imaginary parts are not separated into independent channels for processing, so as to avoid disrupting the inherent correspondence between channel amplitude and phase and to ensure the structural integrity of the complex channel features.

[0025] S2, Causal Exponential Moving Average Smoothing Process The complex time-frequency features are smoothed by performing a causal exponential moving average smoothing process along the time dimension to obtain the smoothed complex features.

[0026] This step is used to reduce the impact of local random disturbances on subsequent difference calculations. The smoothing process strictly adheres to causal constraints, using only feature data from the current and past time positions, and not introducing feature data from future time positions. Let the smoothed complex feature at time position t be... The calculation formula is as follows: ;in, This is the smoothing coefficient, and its value range is... .when At the initial time position, the smoothed complex feature at that initial time position is equal to the input complex time-frequency feature at that position, i.e.: .

[0027] The above causal smoothing process can suppress local rapid fluctuations in the input features, avoid excessive amplification of the differential calculation by random disturbances, and ensure that the processing conforms to the temporal logic of the channel prediction task, thus avoiding future information leakage problems.

[0028] S3, Complex Domain Two-Stream Difference Calculation The time difference results are calculated along the time direction and the frequency difference results are calculated along the frequency direction for the smoothed complex features, respectively, to obtain the complex domain two-stream difference results.

[0029] The time difference result is the difference between the smoothed complex features corresponding to adjacent time positions, used to characterize the evolution of the channel state at adjacent time points. Let the feature of the previous position in the time direction be... Then the time difference can be expressed as: ; For the time difference result of the first time position, the smoothed complex feature of the current position itself is used as the previous reference position for calculation. That is, the smoothed feature of the previous position at t=0 is equal to the feature of the current position to ensure the stability of the time domain boundary calculation.

[0030] The frequency difference result is the difference between the smoothed complex features corresponding to adjacent frequency positions, used to characterize the channel structure change between adjacent frequency subcarriers. Let the feature of the previous position in the frequency direction be... Then the frequency difference can be expressed as: ; For the frequency difference result of the first frequency position, the smoothed complex feature of the current position itself is used as the previous reference position for calculation. That is, when f=0, the smoothed feature of the previous position is equal to the feature of the current position, so as to ensure the stability of the frequency domain boundary calculation.

[0031] By jointly modeling the time and frequency dual-path differential, it is possible to simultaneously extract the evolution of channel features in the time direction and the local structural changes in the frequency direction. Compared with the method of using only time differential, dual-path differential can more fully reflect the dynamic evolution characteristics of complex channels in the time-frequency plane.

[0032] S4, Dynamic Intensity Map Construction The time difference energy corresponding to the time difference result and the frequency difference energy corresponding to the frequency difference result are calculated respectively. A dynamic intensity map is constructed based on the time difference energy and the frequency difference energy; specifically including: S4.1. The average of the sums of the squared real and squared imaginary parts of the time difference results along the feature dimension is used to obtain the time difference energy. The calculation formula is: , in, Indicates taking the real part, This indicates taking the imaginary part. This indicates an averaging operation along the feature dimension D; S4.2. The average of the sums of the squares of the real and imaginary parts of the frequency difference results along the characteristic dimension is taken to obtain the frequency difference energy. The calculation formula is: ; S4.3. Perform a logarithmic operation on the sum of the time-frequency difference energy and the frequency difference energy at each time-frequency position to obtain the dynamic intensity map S. The calculation formula is as follows: ; in, To prevent numerical instability caused by small constants, the dynamic intensity map can reflect the degree of local variation at each time and frequency location by jointly measuring the time difference energy and frequency difference energy. Using a logarithmic form can compress the influence of excessively large changes, making the dynamic intensity distribution more stable.

[0033] Meanwhile, the constructed dynamic intensity map can quantitatively reflect the strength of local changes at each time-frequency position, clearly distinguishing between areas of drastic change and relatively stable areas, providing an intensity basis for subsequent adaptive gating calibration.

[0034] S5, Lightweight 2D Convolutional Mapping Processing The dynamic intensity map is input into a lightweight 2D convolutional mapping network for mapping processing, and the output is a dynamic mapping map.

[0035] The lightweight 2D convolutional mapping network includes at least one 2D convolutional layer and a nonlinear activation function. The operation of the 2D convolutional layer operates on the time-frequency plane, which consists of the frequency and time dimensions, to extract dynamic correlation patterns between adjacent time-frequency locations. The expression for the mapping process is: ; in, Represents a dynamic mapping graph. This represents a lightweight two-dimensional convolutional mapping network.

[0036] This network extracts features from the real-valued dynamic intensity map without directly modifying the original complex channel features. It can mine the local spatial correlation patterns of dynamic intensity with lower computational complexity, making the basis for subsequent gating calibration more context-consistent.

[0037] S6. Generate adaptive gating coefficients Generate adaptive gating coefficients based on the dynamic mapping graph.

[0038] The adaptive gating coefficient is generated by the following formula: ; in, Indicates the gating coefficient. This represents the Sigmoid function. This represents an adjustable gating strength coefficient; at the same time, the gating coefficient is a real-valued coefficient, which acts on both the real and imaginary parts of the complex feature; thus, while adjusting the strength of the feature response, the correspondence between the real and imaginary parts of the complex feature can be maintained.

[0039] To avoid causing excessive disturbance to the original complex features in the early stages of processing, the gating strength coefficient can be set to a small initial value, so that the gating calibration is close to the identity mapping in the initial state. As the parameters are adjusted in the subsequent stages, the gating coefficient can gradually enhance the response to the key time and frequency regions according to the dynamic changes in strength.

[0040] Through adaptive gating calibration, complex features can be enhanced differently based on the degree of dynamic change at different time and frequency positions. Positions with more significant changes can obtain stronger feature responses, while positions with weaker changes remain relatively stable, thus avoiding the application of the same processing intensity to all time and frequency positions.

[0041] 7. Generate complex enhancement features By applying adaptive gating coefficients to complex time-frequency features, complex enhanced features are obtained.

[0042] Specifically, the adaptive gating coefficients are multiplied element-wise with the complex time-frequency features; at the same time-frequency position, the adaptive gating coefficients act on both the real and imaginary parts of the complex time-frequency features with the same value.

[0043] Let the output feature be Y, then it can be expressed as: ; in, This represents element-wise multiplication; because The gating coefficients are real-valued and act on both the real and imaginary parts of the input complex features, thus ensuring that the output features retain their complex form; specifically, they can be expressed as: .

[0044] This processing method, while adjusting the strength of the characteristic response in a differentiated manner, fully maintains the correspondence between the real and imaginary parts of the complex feature, avoiding disruption of the amplitude and phase structure of the channel; time-frequency positions with higher intensity of change will receive stronger feature enhancement, while positions with gentler changes will maintain the original feature level, achieving adaptive dynamic calibration.

[0045] S8, Output complex channel state prediction results The complex enhanced features are input into the prediction mapping structure, and the complex channel state prediction results for future time periods are output.

[0046] The obtained complex enhancement features are input into the prediction mapping structure, and the complex channel state prediction results for future time steps are generated through nonlinear mapping. Let the prediction result be... Then it can be expressed as: ;in, This represents the prediction mapping function.

[0047] The prediction result can include the channel state at a single future time point, or it can include the channel state at J consecutive future time points, i.e. .

[0048] The final complex channel state prediction results can be directly used in subsequent communication processing steps such as precoding, beamforming, link adaptation, or wireless resource scheduling.

[0049] In summary, this embodiment, through the coordinated use of causal smoothing, dual-stream differential, dynamic intensity modeling, and adaptive gating, explicitly extracts local dynamic change information in the time-frequency two-dimensional domain while maintaining the consistency of the complex channel amplitude-phase structure. This effectively improves the channel prediction's response to non-stationary changes, local abrupt changes, and frequency-selective fading. The smoothing process using causal constraints reduces the interference of random disturbances on the differential results, improving the stability of dynamic intensity estimation. Lightweight two-dimensional convolution mines time-frequency neighborhood correlations, ensuring the smoothness of gating calibration with low computational overhead. The design of synchronously applying real-valued gating to both the real and imaginary parts balances dynamic enhancement and amplitude-phase structure integrity. The overall scheme is well-suited for typical wireless communication scenarios such as high-speed mobility, obstruction handover, and complex multipath structures, providing high-quality feature support for applications such as channel state prediction, predictive precoding, and intelligent resource scheduling.

[0050] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0051] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A complex channel state prediction method based on time-frequency differential dynamic evolution, characterized in that: Includes the following steps: S1. Obtain complex time-frequency characteristics, wherein the complex time-frequency characteristics are constructed from historical complex channel state information, or obtained from historical complex channel state information after time-frequency preprocessing; S2. Perform causal exponential moving average smoothing on the complex time-frequency features along the time dimension to obtain smoothed complex features; S3. Calculate the time difference result along the time direction and the frequency difference result along the frequency direction for the smoothed complex features, respectively, to obtain the complex domain dual-stream difference result containing both time difference and frequency difference features; S4. Calculate the time difference energy corresponding to the time difference result and the frequency difference energy corresponding to the frequency difference result respectively, and construct a dynamic intensity map based on the time difference energy and the frequency difference energy; S5. Input the dynamic intensity map into a lightweight two-dimensional convolutional mapping network for mapping processing, and output a dynamic mapping map. S6. Generate adaptive gating coefficients based on the dynamic mapping graph; S7. The adaptive gating coefficients are weighted element by element to the complex time-frequency feature to obtain the complex enhanced feature; S8. Input the complex enhancement features into the prediction mapping structure and output the complex channel state prediction results for future time periods.

2. The complex channel state prediction method based on time-frequency differential dynamic evolution according to claim 1, characterized in that: In step S2, the causal exponential moving average smoothing process calculates the smoothed complex feature at time position t using the following formula. : ;in, This is the smoothing coefficient, and its value range is... When t=0, the smoothed complex feature of the initial time position is equal to the input complex time-frequency feature of that position; the smoothing process only uses the feature data of the current and past time positions, and does not introduce the feature data of future time positions.

3. The complex channel state prediction method based on time-frequency differential dynamic evolution according to claim 1, characterized in that: In step S3, the time difference result is the difference between the smoothed complex features corresponding to adjacent time positions; The frequency difference result is the difference between the smoothed complex features corresponding to adjacent frequency positions; For the time difference result of the first time position and the frequency difference result of the first frequency position, the difference is calculated using the smoothed complex feature of the current position itself as the preceding reference.

4. The complex channel state prediction method based on time-frequency differential dynamic evolution according to claim 1, characterized in that: Step S4 specifically includes: S4.

1. The average of the sum of the squares of the real and imaginary parts of the time difference results along the feature dimension is used to obtain the time difference energy. S4.

2. The average of the sum of the squares of the real and imaginary parts of the frequency difference results along the feature dimension is taken to obtain the frequency difference energy. S4.

3. Perform a logarithmic operation on the sum of the time difference energy and the frequency difference energy to obtain the dynamic intensity map.

5. The complex channel state prediction method based on time-frequency differential dynamic evolution according to claim 1, characterized in that: In step S5, the lightweight two-dimensional convolutional mapping network includes at least one two-dimensional convolutional layer and a nonlinear activation function; the operation of the two-dimensional convolutional layer is performed on the time-frequency plane composed of the time dimension and the frequency dimension.

6. The complex channel state prediction method based on time-frequency differential dynamic evolution according to claim 1, characterized in that: In step S6, the adaptive gating coefficient is generated according to the following formula: ;in, Indicates the gating coefficient. This represents the Sigmoid function. This indicates an adjustable gating strength coefficient. This represents a dynamic mapping graph, where the adaptive gating coefficients are real-valued coefficients.

7. The complex channel state prediction method based on time-frequency differential dynamic evolution according to claim 6, characterized in that: In step S7, the adaptive gating coefficient is multiplied element-wise with the complex time-frequency feature; at the same time-frequency position, the adaptive gating coefficient is applied simultaneously to the real and imaginary parts of the complex time-frequency feature at that position with the same real value.

8. The complex channel state prediction method based on time-frequency differential dynamic evolution according to claim 1, characterized in that: In step S1, the complex time-frequency feature includes a batch dimension, a frequency dimension, a time dimension, and a feature dimension; the complex time-frequency feature maintains its complex form during the processing of steps S2 to S7, and the real part and imaginary part are not separated into independent channels for processing.

Images