Anti-interference robustness wireless channel prediction method

Abstract

The invention relates to an anti-interference robustness wireless channel prediction method. The method comprises the following steps: acquiring observation uplink CSI (Channel State Information) sequences of a plurality of time slots in the past; inputting the observation uplink CSI sequences of the plurality of time slots into an interference perception gating prediction model to obtain ideal downlink CSI sequences of a plurality of time slots in the future; wherein the interference sensing gating prediction model comprises an interference sensing module which is used for carrying out normalization operation on the observation uplink CSI sequence, detecting interfered time-frequency distribution of the observation uplink CSI sequence to generate an interference mask, and inhibiting an interfered area through a self-adaptive gating factor to obtain a purified CSI sequence; and the channel prediction backbone network is used for modeling long time sequence correlation for the purified CSI sequence and predicting an ideal downlink CSI sequence at a future moment. According to the invention, adaptive suppression of interference and accurate prediction of a channel can be realized.

Description

An anti-interference robust wireless channel prediction method Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to an interference-resistant robust wireless channel prediction method. Background Technology

[0002] As 5G-Advanced and 6G mobile communication systems evolve towards higher frequency bands, larger bandwidths, and higher speeds, the time-varying characteristics of wireless channels are becoming increasingly pronounced. In high-speed mobile scenarios, the channel coherence time is drastically shortened, and traditional methods based on instantaneous estimation and feedback face significant overhead and latency challenges. Therefore, channel prediction technology has become a research focus, aiming to directly infer future CSI using historical channel state information (CSI), thereby providing forward-looking information for precoding, resource allocation, and other functions, effectively overcoming the channel aging problem.

[0003] In recent years, deep learning sequence models, represented by Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) networks, and Transformers, have been widely applied to channel prediction tasks due to their powerful ability to extract nonlinear temporal features. These methods, trained end-to-end on large-scale channel data, can learn the "space-time-frequency" nonlinear mapping relationship from historical CSI sequences to future CSI sequences, demonstrating excellent prediction performance in ideal or mildly Gaussian white noise scenarios.

[0004] The performance of the aforementioned channel prediction models is highly dependent on the quality of the input data and is largely based on ideal assumptions of clean or mildly noisy environments. In real-world complex electromagnetic environments, uplink received signals inevitably suffer from strong non-Gaussian interference from various sources such as co-channel devices and malicious jammers, resulting in a large amount of structured contamination components in the uplink CSI acquired by the base station. Existing channel prediction models lack robustness to this type of structured interference. When the input data deviates from the ideal distribution of the training data, the model's generalization ability is insufficient, leading to a significant decrease in prediction accuracy. This has become a key bottleneck restricting their deployment in practical systems.

[0005] To enhance the anti-interference capability of systems, traditional communication systems typically employ a cascaded processing architecture. This involves deploying an independent interference management module (such as energy-detection-based threshold decision, frequency domain filtering, or frequency hopping mechanisms) before the channel prediction module to preprocess the received signal, and then feeding the processed signal into the channel prediction module. However, this modular, independently optimized design paradigm has inherent technical limitations: on the one hand, interference suppression modules usually prioritize maximizing the signal-to-interference-to-noise ratio (SINR) as their sole optimization objective, lacking awareness of subsequent prediction tasks. This can easily damage the high-frequency characteristics of the channel, which contain important time-varying information, while filtering out interference. On the other hand, the residual errors from the preceding modules and the introduced signal distortion are input as irreversible noise into the subsequent prediction model. Due to the lack of a gradient feedback mechanism for joint optimization among the cascaded modules, the prediction model cannot guide the preceding modules to correct their parameters, making it difficult for the system to converge to the global optimum. Therefore, designing a method that deeply integrates interference perception and channel prediction in strong interference environments and enables end-to-end collaborative optimization has become a pressing technical challenge in this field. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide an anti-interference robust wireless channel prediction method that can achieve adaptive suppression of interference and accurate prediction of the channel.

[0007] The technical solution adopted by this invention to solve its technical problem is: to provide an anti-interference robust wireless channel prediction method, comprising the following steps:

[0008] Obtain the uplink CSI sequence of observations from several past time slots;

[0009] The observed uplink CSI sequences of the aforementioned time slots are input into an interference-aware gating prediction model to obtain ideal downlink CSI sequences for future time slots; wherein, the interference-aware gating prediction model includes:

[0010] The interference sensing module is used to normalize the observed uplink CSI sequence, detect the time-frequency distribution of interference to generate an interference mask, and suppress the interference region through an adaptive gating factor to obtain a purified CSI sequence.

[0011] The channel prediction backbone network is used to model long-term temporal correlations of the cleaned CSI sequence and predict the ideal downlink CSI sequence at future times.

[0012] The interference sensing module includes:

[0013] The interference detection section is used to perform interference detection on the observed uplink CSI sequence using a feature extraction network, identify the distribution of interference in the time-frequency feature dimension, and output the interference mask for each feature point.

[0014] The adaptive gating suppression part is used to perform weighted cleaning of the observed uplink CSI sequence based on the distribution of interference in the time-frequency feature dimension and the interference mask, so as to obtain the cleaned CSI sequence.

[0015] The feature extraction network adopts a bottleneck-type multilayer perceptron structure, which first compresses and then restores the input observation uplink CSI sequence to extract the nonlinear representation of interference.

[0016] The adaptive gating network takes the distribution of interference in the time-frequency feature dimension as input, maps the interference feature vector of each time slot to a low-dimensional hidden space, performs global interference feature aggregation to obtain hidden features, maps the hidden features to scalars, and constrains them within a preset interval through the Sigmoid function to obtain the adaptive gating factor for each time slot. The adaptive gating factor and interference mask are used to perform weighted cleaning of the observed uplink CSI sequence to obtain the cleaned CSI features.

[0017] The channel prediction backbone network includes:

[0018] The embedding module is used to construct a sequence representation of the purified CSI features that is adapted to the input of the backbone module of the large language model;

[0019] The large language model backbone module is used to model long-term temporal correlations based on the sequence representation;

[0020] The output projection module is used to convert long-term time-series correlations into ideal downlink CSI sequences for future time periods.

[0021] The embedded module includes:

[0022] The block unit is used to form frequency domain representation and delay domain representation of the cleaned CSI features respectively, and to divide the frequency domain representation and delay domain representation into blocks;

[0023] An enhanced coding unit is used to enhance the CSI features of the block through residual convolution and channel attention, and to superimpose positional encoding and Dropout regularization on the enhanced CSI features to form a sequence representation that is adapted to the input of the backbone module of the large language model.

[0024] The large language model backbone module is composed of several stacked Transformer blocks. Each Transformer block includes LayerNorm, multi-head self-attention sub-layer, feedforward network sub-layer, and residual connections. During training, the main parameters of the multi-head attention sub-layer and feedforward network sub-layer are frozen, and only LayerNorm is trained.

[0025] The loss function of the interference-aware gating prediction model is expressed as: ,in, The loss function for the interference-sensing gating prediction model is... The normalized mean square error between the predicted CSI feature sequence and the true CSI feature sequence is expressed as: , For the predicted future Time of the first Ideal downlink CSI sequence for each time slot For the real future Time of the first Ideal downlink CSI sequence for each time slot The number of future time slots. Indicates the total amount of time in the future. It is the Frobenius norm. This is a normalization operation; The binary cross-entropy loss of the interference mask is expressed as: , This represents the number of past time slots. To observe the dimension of the uplink CSI sequence, The real label is used to interfere with the mask. For the generated interference mask; This is the balance coefficient.

[0026] The technical solution adopted by the present invention to solve its technical problem is: to provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-mentioned anti-interference robust wireless channel prediction method.

[0027] The technical solution adopted by the present invention to solve its technical problem is: to provide a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, implements the steps of the above-mentioned anti-interference robust wireless channel prediction method.

[0028] Beneficial effects

[0029] Due to the adoption of the above technical solutions, this invention has the following advantages and positive effects compared with the prior art: This invention filters contaminated uplink CSI by interference detection and gating suppression. The entire process can identify various random interferences and obtain the purified CSI sequence without manual intervention. Furthermore, based on a large language model, it accurately captures the long-term evolution law of the channel in high-speed mobile scenarios and predicts future downlink CSI, thereby achieving adaptive suppression of interference and accurate prediction of the channel. Attached Figure Description

[0030] Figure 1 is a flowchart of the anti-interference robust wireless channel prediction method according to the first embodiment of the present invention;

[0031] Figure 2 is a structural block diagram of the interference sensing gating prediction model in the first embodiment of the present invention;

[0032] Figure 3 is a schematic diagram of the interference sensing module in the first embodiment of the present invention;

[0033] Figure 4 is a schematic diagram of the embedded module in the first embodiment of the present invention;

[0034] Figure 5 is a schematic diagram of the large language model backbone module in the first embodiment of the present invention;

[0035] Figure 6 is a performance comparison chart of the interference-aware gating prediction model and the traditional deep learning model in the embodiments of the present invention. Detailed Implementation

[0036] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0037] The first embodiment of this invention relates to an interference-resistant robust wireless channel prediction method. This method can predict a relatively clean downlink CSI at a future time using current and historical interfered uplink CSI. The prediction result will be used in transmission technologies such as downlink beamforming to improve communication speed and reliability.

[0038] Consider a MISO system consisting of a single base station (BS) and multiple user equipments (UEs), where the set of UEs is represented as: Base station equipment and settings The system has one transmitting antenna, and the user equipment is equipped with a single antenna. The system operates in Time Division Duplex (TDD) mode and is affected by multiple non-Gaussian interference sources, denoted as [missing information - likely a suffix]. The base station uses precoding based on matched filtering to maximize the rate and uses comb reference signals for CSI detection. In high-speed mobile scenarios, the channel coherence time is extremely short. Although the traditional channel reciprocity principle is theoretically valid, its application faces two major practical challenges: (1) Channel aging: High-speed movement of user equipment causes the channel to change rapidly over time, making it extremely difficult to estimate future CSI based on historical CSI. (2) Link interference: Complex interference on the communication link will contaminate the channel estimation results, resulting in a huge deviation in downlink prediction based on unclean uplink CSI. Therefore, the core task of this implementation is to extract clean channel features from the historical uplink CSI sequence contaminated by strong interference and accurately predict future downlink CSI.

[0039] Wireless signals typically travel from the transmitter to the receiver through reflection, scattering, and diffraction by obstacles, resulting in multiple propagation paths—a phenomenon known as "multipath propagation." These paths are not isolated in space but tend to appear in groups; each group of paths originating from a primary scatterer is called a "scattering cluster." This implementation uses a cluster-based multipath channel model to characterize the space-time-frequency characteristics of the wireless channel. Under interference-free conditions, the downlink channel at time... , subcarrier The frequency domain channel response can be expressed as:

[0040] (1)

[0041] in, Indicates the number of base station antennas. Indicates signal transmission path loss; This represents the total number of scattering clusters; Indicates the first The number of subpaths contained in a scattering cluster; Indicates the first The scattering cluster, the scattering cluster The complex gain of the strip path includes phase and amplitude gain information; and They represent the first The scattering cluster, the scattering cluster The time delay and Doppler shift of the strip path, the latter caused by user movement, directly lead to the time-varying characteristics of the channel and are the main cause of channel aging, as expressed as:

[0042] (2)

[0043] in, The speed of the user equipment relative to the base station. For carrier wavelength, The angle between the user equipment's direction of movement and the signal's direction of arrival reflects the channel's time selectivity. The faster the user moves, the stronger the channel's time-varying nature, and the more difficult it is to predict the CSI. The steering vector of the base station-side antenna array describes the phase relationship caused by the path difference when a signal is transmitted from different antennas. The base station uses a uniform planar array antenna (UPA), and the array response vector can be expressed as:

[0044] (3)

[0045] (4)

[0046] (5)

[0047] in, The horizontal response vector, The vertical response vector, This represents the Kronecker product. The antenna spacing is set to half a wavelength to avoid grating lobes; These represent the azimuth and elevation angles of the signal, respectively.

[0048] In an ideal TDD system with no interference and no noise, the downlink channel and uplink channel The reciprocity property is satisfied, meaning that within coherent time, we have:

[0049] (6)

[0050] Therefore, the base station obtains the ideal uplink CSI and can estimate the downlink CSI by time extrapolation, which can then be directly used for downlink transmission. However, in actual communication, the uplink received signal at the base station not only contains the desired signal but is also affected by multipath channels exhibiting multiplicative interference and various additive noises. Therefore, the uplink CSI is not the pure physical channel response under ideal conditions but is contaminated by various noises and interferences. This implementation comprehensively considers additive white Gaussian noise (AWGN) and various typical artificial interferences commonly found in communication systems. In scenarios where interference exists, The time base station receives data from the user The actual uplink signal, its frequency domain model can be simplified to a multiplicative relationship on the subcarriers, and the received signal on the subcarriers... The above can be modeled as:

[0051] (7)

[0052] in, Given a known reference signal, The variance is A zero-mean additive white Gaussian noise vector; This indicates the set of artificial disturbances considered in this embodiment, including but not limited to...

[0053] (8)

[0054] To simulate the complexity of real electromagnetic environments, this implementation randomizes the parameters of the aforementioned interference signals. Specifically, for each type of interference, the interference-to-signal ratio (JSR) is randomly selected between -5dB and 25dB; its center frequency, bandwidth, pulse duration, and frequency hopping pattern all follow a specific random distribution. This highly randomized interference modeling ensures that the model obtained through subsequent training has the ability to generalize to unknown interference environments. After channel prediction, interference and AWGN will be superimposed on the CSI as residual errors, severely distorting the true characteristics of the channel response and causing a large deviation in the downlink CSI prediction based on the uplink CSI. To clarify the source of the deviation, consider the channel estimation process. The base station uses a standard channel estimator based on the reference signal... From the received signal CSI is extracted from the data. For ease of analysis, the widely used least squares (LS) channel estimation method is used as an example. For subcarrier... Using a known reference signal In the case of a linear receiver, the LS channel estimation process can be simplified to:

[0055] (9)

[0056] in, This represents the processing vector related to the reference signal. For LS estimation, the processing vector... Essentially, it is a pilot signal. The pseudo-inverse operation. Bias in base station observation CSI. Its statistical characteristics are highly dependent on the type and intensity of noise and interference. In strong interference scenarios (JSR>0), the LS channel estimation of observed uplink CSI... Completely dominated by interference errors, losing sight of the ideal channel. Its representational ability.

[0057] In OFDM systems, time-frequency resources are divided into resource grids. A resource cell (RE) is the basic unit of a subcarrier on an OFDM symbol. A reference signal is periodically placed on a specific RE for CSI prediction.

[0058] Assuming the system bandwidth includes Subcarriers, in time slots ,user The uplink CSI is represented as a complex matrix. In the channel prediction process, this implementation only considers the CSI at the reference RE position; the CSI between reference REs can be obtained through interpolation. Based on the aforementioned channel model and interference analysis, this implementation formalizes the anti-interference downlink channel prediction task into a multi-task joint optimization sequence generation problem. The goal is to establish an interference-aware gated prediction model, extract clean features from the interfered historical uplink CSI sequence, and predict the ideal downlink CSI sequence at future times through reciprocity mapping. .

[0059] As shown in Figure 1, the anti-interference robust wireless channel prediction method of this embodiment includes the following steps:

[0060] Step 1: Obtain the uplink CSI sequence of observations from the past several time slots;

[0061] Step 2: Input the observed uplink CSI sequences of the several time slots into the interference sensing gating prediction model to obtain the ideal downlink CSI sequences for the future several time slots.

[0062] As shown in Figure 2, this interference-aware gating prediction model can be built based on the Transformer architecture (GPT2), mainly consisting of an interference-aware module and a channel prediction backbone network, which are jointly trained end-to-end. The interference-aware module normalizes the observed uplink CSI sequence, detects its time-frequency distribution of interference to generate an interference mask, and suppresses the interfered region using an adaptive gating factor to obtain a cleaned CSI sequence. The channel prediction backbone network models the long-term temporal correlation of the cleaned CSI sequence and predicts the ideal downlink CSI sequence for future times.

[0063] To adapt to the real number operation characteristics of deep neural networks, this implementation decouples and reconstructs the real and imaginary parts of the complex CSI matrix; that is, for each time slot, the observation matrix is ​​extracted ( The real and imaginary parts of the integers are concatenated along the feature dimension to construct a real feature tensor: After flattening, the data for each time slot is represented as a dimension of . A real-number vector. This data representation not only satisfies the neural network's constraint on real-number inputs, but also fully preserves the channel's amplitude and phase information.

[0064] This implementation can use a sliding time window mechanism to construct the input and output sequences, where the input sequence is the past... The observed uplink CSI sequence for each time slot is represented as follows:

[0065] (10)

[0066] Among them, the observed uplink CSI sequence It contains physical information about the channel, but also includes mixed noise components caused by additive white Gaussian noise and random time-varying interference.

[0067] Predicting the target is the future. The ideal downlink CSI sequence for each time slot can be expressed as:

[0068] (11)

[0069] Among them, the ideal downlink CSI sequence Leveraging the channel reciprocity of TDD systems, the target downlink CSI can be derived from the clean uplink CSI. Therefore, interference detection and suppression of the CSI sequence are necessary first. As shown in Figure 3, the observed sequence first enters the interference sensing module, which includes an interference detection part and an adaptive gating suppression part. The interference detection part uses a feature extraction network to detect interference in the observed uplink CSI sequence, identify the distribution of interference in the time-frequency feature dimension, and output the interference mask for each feature point to characterize the interference intensity. The adaptive gating suppression part, based on the distribution of interference in the time-frequency feature dimension and the interference mask, performs weighted cleaning of the observed uplink CSI sequence through an adaptive gating network to obtain the cleaned CSI sequence. The above interference sensing module can be abstractly represented as:

[0070] (12)

[0071] in, This represents an interference suppression operator, consisting of two steps: interference detection and adaptive gating suppression, used to suppress interference in the observed sequence. Transform into a purification sequence Purification sequence The channel prediction backbone network consists of an embedding module, a large language model backbone module, and an output projection module. The embedding module cleans the sequence... The sequences are constructed into frequency domain and delay domain representations respectively, and then patched. Position encoding is then used to form a sequence representation adapted to the input of the large language model backbone module. The large language model backbone module models long-term temporal correlations based on these sequence representations. An output projection module is used to convert the long-term temporal correlations into ideal downlink CSI sequences for future timeframes. The overall mapping relationship can be defined as:

[0072] (13)

[0073] in, This represents a channel prediction backbone network based on cleanup sequences. Generate prediction results.

[0074] Before entering the prediction backbone network, the observed uplink CSI sequence first passes through the interference suppression operator. The process involves purification. Figure 2 shows a schematic diagram of the interference sensing module, which consists of an interference detection section and an adaptive gating suppression section. The input is the interfered CSI sequence. The output is the purified CSI sequence. .

[0075] The following is a detailed introduction to the two parts:

[0076] Interference detection section: A lightweight feature extraction network is used. Interference detection is performed on the input CSI sequence to identify the distribution of interference along the time-frequency feature dimension, and a soft mask (i.e., interference mask) is output for the probability of each feature point being interfered with. The probability of each feature point being interfered with is expressed as:

[0077] (14)

[0078] in, The feature extraction network represents the Sigmoid activation function. A bottleneck-type multilayer perceptron (Bottleneck MLP) structure is employed to first compress and then restore the input features to extract the nonlinear representation of the interference. Its output is the interference log-odds matrix. , means as follows:

[0079] (15)

[0080] Wherein, the input feature dimension is Hidden layer dimensions , The matrix is ​​then compressed using equation (13). Mapped as an interference mask This is used to indicate the probability of interference in each feature dimension of each time slot.

[0081] The adaptive gating suppression part is based on the detected matrix With interference mask Introducing adaptive gating networks Through learnable adaptive gating factors The input features are weighted and cleaned to obtain a cleaned feature sequence. This mechanism includes an adaptive gated network (Gate MLP), which receives a matrix... As input, the output is a scalar gating factor. The process of suppressing the interfered area is simplified as follows:

[0082] Gated Networks The matrix output by the received interference detection head As input. For the... Interference feature vector of each time slot Gating network First, it is mapped to a low-dimensional hidden space, and then global perturbation feature aggregation is performed to obtain the hidden features, represented as:

[0083] (16)

[0084] in, and These are the projection parameters for the first layer.

[0085] Then, hide the features Mapped to a scalar and constrained by the Sigmoid function. Within the interval, the adaptive gating factor for that time slot is obtained. , is represented as:

[0086] (17)

[0087] in, and These are the output layer parameters. Here, the adaptive gating factor is used. It is generated per-timestep, therefore, at times of severe sequence perturbation, the model can output an adaptive gating factor close to 1. For moments when there is no disturbance, the output has an adaptive gating factor close to 0. This enables fine-grained temporal adaptive suppression.

[0088] Finally, the generated gating factor is used. and interference mask The original observation features are weighted and cleaned to obtain the cleaned CSI features, which are represented as follows:

[0089] (18)

[0090] in, It is the first The soft mask vector of the time slot, This represents the Hadamard Product. Repeating the above operation for all time slots yields the cleaned sequence. This serves as the input for the subsequent channel prediction backbone network.

[0091] After processing by the above interference suppression module, the purified feature sequence is obtained. To capture the complex evolution of the channel over long time, this implementation uses a pre-trained large language model-based architecture as the channel prediction backbone network. This channel prediction backbone network mainly consists of an embedding module, a large language model backbone network module (GPT2), and an output projection module.

[0092] The embedding module is used to construct a sequence representation of the purified CSI sequence that fits the input of the transformer. In this embodiment, the embedding module is a generalized representation, as shown in Figure 4. This module includes a block unit and an enhancement coding unit. The block unit forms frequency domain and delay domain representations for the purified CSI features, and then patches these representations to enhance local structure modeling capabilities. The enhancement coding unit enhances the CSI representation through residual convolution and channel attention, and then superimposes positional encoding and dropout to form the final token sequence representation, which serves as the input to the large language model backbone module. The residual convolution in Figure 3 further incorporates a channel attention mechanism, which adaptively weights the channel-dimensional features to highlight effective features related to the channel state and suppress redundant features, thereby improving the robustness of subsequent sequence modeling and prediction.

[0093] As shown in Figure 5, the backbone module of the large language model consists of several stacked Transformer blocks. Each Transformer block includes LayerNorm, a multi-head self-attention sub-layer, a feedforward network sub-layer, and residual connections. This implementation adopts a lightweight fine-tuning strategy, that is, freezing the main parameters of the multi-head attention sub-layer and the feedforward network sub-layer, and training only a small number of parameters such as LayerNorm. This preserves the general representation ability learned by the pre-trained model in large-scale sequence modeling, reduces the fine-tuning overhead for channel prediction scenarios, and improves stability in scenarios with small samples and strong interference.

[0094] The output projection module is used to convert the output of the large language model backbone module into an ideal downlink CSI sequence for future time moments.

[0095] To simultaneously ensure the accuracy of interference detection and suppression, as well as the precision of channel prediction, the interference-aware gating prediction model in this embodiment uses a multi-task joint loss function to train the model parameters. This loss function can be expressed as:

[0096] (19)

[0097] For ease of description, the simplified form is as follows:

[0098] (20)

[0099] in:

[0100] (twenty one)

[0101] (twenty two)

[0102] To predict the normalized mean square error (NMSE) between the predicted CSI feature sequence and the true CSI feature sequence, which is used to guide channel prediction. For the predicted future Time of the first Ideal downlink CSI sequence for each time slot For the real future Time of the first Ideal downlink CSI sequence for each time slot Denotes the Frobenius norm; The binary cross-entropy loss (BCE) of the interference mask is used to supervise the learning of the interference detection module. This is a balancing factor used to adjust the weights of the two tasks. Here, This represents the ground truth of the interference mask. During the data construction process in the model training phase, the system generates interfered samples while recording the distribution of the interference signal on the time-frequency resource grid, generating a binary matrix with the same dimension as the input. Interference-affected resource units are marked as 1, and undisturbed units are marked as 0, thus providing a clear monitoring signal for the detection head.

[0103] The invention is further illustrated below with an example. This example uses the widely used wireless channel simulation platform QuaDRiGa to generate a time-varying CSI dataset conforming to the 3GPP standard, used to verify the effectiveness and robustness of the interference-aware gating prediction model in this embodiment under high-speed movement and strong interference scenarios. The simulation considers a MISO-OFDM communication system, with a dual-polarized uniform planar array (UPA) configured on the base station side, and the array dimension is... A single omnidirectional antenna is configured on the user side; the element spacing is half the wavelength corresponding to the center frequency. The center frequency is set to 3.5 GHz, and the 3GPP 38.901 Urban Macro (UMa) Non-Line-of-Sight (NLoS) scene model is adopted; the user's initial position is randomly generated, the movement trajectory is a straight line, and the user speed is uniformly distributed between 5 km / h and 140 km / h. The dataset is divided as follows: the training set contains 9000 samples, the validation set contains 1000 samples; the test set is generated by uniformly selecting 10 speed points, with 1000 samples generated for each speed point, to evaluate the model's generalization ability under different movement speeds.

[0104] The system generates continuous CSI sequence segments at 0.5 ms sampling intervals to simulate the time-domain sampling process of the comb reference signal on the channel. K=64 reference frequencies are selected in the frequency domain as effective CSI features. In this embodiment, considering TDD mode, uplink and downlink CSIs are defined as sequences in different time slots within the same frequency band. To adapt to the real-number operations of deep neural networks, complex CSIs are concatenated along the feature dimension according to the aforementioned process and flattened to form a real-number feature vector for each time slot. Based on this, a sliding time window sequence sample is constructed: the interfered uplink CSI sequences of P=16 historical time slots are taken as model input, and the target CSI sequences of L=4 future time slots are predicted as output.

[0105] To simulate the complexity of real electromagnetic environments, this embodiment superimposes additive white Gaussian noise and random time-varying artificial interference onto the input CSI during the training phase. To simulate the interference uncertainty in real electromagnetic environments, random interference is injected into historical CSI sequences during the training data construction phase, and supervisory labels for the interference distribution are generated simultaneously. Specifically, artificial interference is injected into each training sample with a 40% probability; once interference injection is triggered, one or two types are randomly selected from a preset set of interference types and superimposed onto the sample, and the interference-to-signal ratio (JSR) is randomly selected within the range of [-5, 12] dB to cover typical operating conditions from weak to medium-strong interference. The interference parameters are shown in Table 1.

[0106] Table 1 Injection Interference Parameter Table

[0107]

[0108] This embodiment uses a Transformer backbone based on pre-trained GPT2 for long-term modeling, and the residual convolution module is set as follows: The patch size is set to ( ), dividing the historical sequence into ( ) non-overlapping patches; SE module reduction ratio set to 2, convolution kernel size set to ( The GPT2 backbone uses a small configuration with a feature dimension of 768 and selects ( The Transformer block is layered. To reduce training overhead and improve stability in scenarios with few samples and strong interference, this embodiment adopts a lightweight fine-tuning strategy, freezing the main parameters of the multi-head attention and feedforward networks in the GPT2 backbone. Only a few parameters such as LayerNorm and the embedding and output projection parts related to task adaptation are trained. Simultaneously, parameters related to interference detection and gating suppression are trained to achieve end-to-end joint optimization. The training objective uses a multi-task joint loss function, with binary cross-entropy loss of the interference mask supervising the learning of the detection head, and balancing coefficients are used. Balancing the contributions of the two tasks; in this embodiment .

[0109] The hyperparameter settings for model training are shown in Table 2.

[0110] Table 2 Model Training Hyperparameter Table

[0111]

[0112] After training, the optimal model parameters on the validation set are saved, and inference verification is performed on an independent test set. The NMSE is used to characterize the channel prediction error. Figure 6 shows the performance comparison between the interference-aware gated prediction model in this embodiment and the traditional deep learning model under the introduced interference environment. It can be seen that the prediction performance of this embodiment is optimal, and the NMSE is reduced by about 1.5dB compared with the traditional prediction model.

[0113] It is easy to see that this invention filters contaminated uplink CSI by interference detection and gating suppression. The entire process can identify various random interferences and obtain purified CSI sequences without manual intervention. Furthermore, based on a large language model, it accurately captures the long-term evolution of the channel in high-speed mobile scenarios and predicts future downlink CSI, thereby achieving adaptive suppression of interference and accurate prediction of the channel.

[0114] A second embodiment of the present invention relates to an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the interference-resistant robust wireless channel prediction method of the first embodiment.

[0115] The third embodiment of the present invention relates to a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the interference-resistant robust wireless channel prediction method of the first embodiment.

[0116] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0117] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or one or more blocks of the block diagrams.

[0118] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction methods that implement the functions specified in one or more flowcharts and / or one or more block diagrams.

[0119] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

[0120] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A robust wireless channel prediction method against interference, characterized in that, Includes the following steps: The process involves acquiring observed uplink CSI sequences from several past time slots; inputting these sequences into an interference-aware gating prediction model to obtain ideal downlink CSI sequences for several future time slots; wherein the interference-aware gating prediction model includes: an interference-aware module, used to normalize the observed uplink CSI sequences, detect the time-frequency distribution of interference to generate an interference mask, and suppress the interference region through an adaptive gating factor to obtain a cleaned CSI sequence; and a channel prediction backbone network, used to model long-term temporal correlations of the cleaned CSI sequences and predict the ideal downlink CSI sequences for future times.

2. The anti-interference robust wireless channel prediction method according to claim 1, characterized in that, The interference sensing module includes: an interference detection part, which uses a feature extraction network to detect interference in the observed uplink CSI sequence, identify the distribution of interference in the time-frequency feature dimension, and output the interference mask for each feature point; and an adaptive gating suppression part, which uses an adaptive gating network to perform weighted cleaning on the observed uplink CSI sequence based on the distribution of interference in the time-frequency feature dimension and the interference mask, to obtain a cleaned CSI sequence.

3. The anti-interference robust wireless channel prediction method according to claim 2, characterized in that, The feature extraction network adopts a bottleneck-type multilayer perceptron structure, which first compresses and then restores the input observation uplink CSI sequence to extract the nonlinear representation of interference.

4. The anti-interference robust wireless channel prediction method according to claim 2, characterized in that, The adaptive gating network takes the distribution of interference in the time-frequency feature dimension as input, maps the interference feature vector of each time slot to a low-dimensional hidden space, performs global interference feature aggregation to obtain hidden features, maps the hidden features to scalars, and constrains them within a preset interval through the Sigmoid function to obtain the adaptive gating factor for each time slot. The adaptive gating factor and interference mask are used to perform weighted cleaning of the observed uplink CSI sequence to obtain the cleaned CSI features.

5. The anti-interference robust wireless channel prediction method according to claim 1, characterized in that, The channel prediction backbone network includes: an embedding module for constructing a sequence representation of the purified CSI features that is adapted to the input of the large language model backbone network module; a large language model backbone network module for modeling long-term temporal correlations based on the sequence representations; and an output projection module for converting the long-term temporal correlations into an ideal downlink CSI sequence for future time periods.

6. The anti-interference robust wireless channel prediction method according to claim 5, characterized in that, The embedding module includes: a block segmentation unit, used to form frequency domain representation and delay domain representation of the purified CSI features respectively, and to segment the frequency domain representation and delay domain representation into blocks; and an enhancement coding unit, used to enhance the CSI features of the blocks through residual convolution and channel attention, and to superimpose position coding and Dropout regularization on the enhanced CSI features to form a sequence representation adapted to the input of the large language model backbone module.

7. The anti-interference robust wireless channel prediction method according to claim 5, characterized in that, The large language model backbone module is composed of several stacked Transformer blocks. Each Transformer block includes LayerNorm, multi-head self-attention sub-layer, feedforward network sub-layer, and residual connections. During training, the main parameters of the multi-head attention sub-layer and feedforward network sub-layer are frozen, and only LayerNorm is trained.

8. The anti-interference robust wireless channel prediction method according to claim 1, characterized in that, The loss function of the interference-aware gating prediction model is expressed as: ,in, The loss function for the interference-sensing gating prediction model is... The normalized mean square error between the predicted CSI feature sequence and the true CSI feature sequence is expressed as: ， For the predicted future Time of the first Ideal downlink CSI sequence for each time slot For the real future Time of the first Ideal downlink CSI sequence for each time slot The number of future time slots. Indicates the total amount of time in the future. It is the Frobenius norm. This is a normalization operation; The binary cross-entropy loss of the interference mask is expressed as: ， This represents the number of past time slots. To observe the dimension of the uplink CSI sequence, The real label is used to interfere with the mask. For the generated interference mask; This is the balance coefficient.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the anti-interference robust wireless channel prediction method as described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the anti-interference robust wireless channel prediction method as described in any one of claims 1-8.

Images