A knowledge and data dual-driven interference suppression method for ground-to-air communication systems

Abstract

The invention relates to a knowledge and data dual-driven interference suppression method for a ground-to-air communication system, and belongs to the technical field of wireless communication. The method of the invention comprises: step 1. constructing an optimization problem of a minimum mean square error problem of a received signal in an airport interference scenario; step 2. constructing a 3DConvTransformer model; using the 3DConvTransformer model to obtain a predicted interference plus noise IPN covariance matrix; obtaining an observed IPN covariance matrix of a subcarrier from the predicted interference plus noise IPN covariance matrix; step 3. constructing an anti-interference hybrid beamforming model based on K3D HBF based on the observed IPN covariance matrix of the subcarrier; and step 4. using the anti-interference hybrid beamforming model to obtain an interference suppression matrix, thereby improving prediction accuracy, achieving a lower mean square error, and significantly reducing the complexity of interference suppression.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communication technology, and specifically relates to a knowledge and data dual-driven interference suppression method for a ground-to-air communication system in response to the requirements of high timeliness and anti-interference in airport interference scenarios. Background Art

[0002] Due to the broadcasting characteristics of ground-to-air wireless communications, they are susceptible to various forms of interference. For example, the L-band Digital Aeronautical Communications System (LDACS) is embedded between the 960MHz-1164MHz frequency band distance measurement equipment (DME) systems, causing co-channel interference. Due to the out-of-band radiation of random DME pulses, the signal-to-interference power ratio of the LDACS receiver can reach -3.8 dB in a typical en-route scenario. This interference will significantly reduce the spectral efficiency and reliability of LDACS transmission. Currently, Multiple-Input Multiple-Output (MIMO) technology uses directional beams to adaptively protect communication systems from interference, but MIMO anti-interference processing remains a challenging problem under the rapidly changing airport surface communication channel conditions: First, the interference signal needs to be projected into the signal-free space to avoid mixing with the desired signal, resulting in unbearable pilot overhead and reduced communication efficiency. Second, the beamforming matrix should be updated with the channel changes determined by the coherence time, which requires the implementation of anti-interference beamforming technology at a relatively low computational complexity. Therefore, it is crucial to design a time-efficient MIMO anti-interference scheme to improve transmission reliability in environments with rapid channel changes.

[0003] At present, MIMO anti-interference technology mainly includes two aspects: interference estimation and interference suppression. For the interference estimation algorithm, the existing method is to deploy a specific subset of the received sample signal in the time domain, frequency domain or spatial domain to estimate the interference characteristics (such as the covariance matrix), which will lead to large pilot overhead and computational complexity. Based on the interference estimation, MIMO Hybrid Beamforming (HBF) technology can be used for interference suppression to simultaneously filter out interference and enhance the required signal strength. However, most of the existing anti-interference HBF designs based on iterative optimization have the problem of high computational complexity, and the anti-interference HBF design based on deep learning has the problems of poor interpretability and slow convergence speed. Therefore, it is not suitable for rapidly changing channels. Although the Knowledge Data Dual Driven (K3D) network is used in MIMO beamforming technology, it combines communication domain knowledge with deep learning to reduce the demand for computing resources and training time. However, there is currently a lack of research on the knowledge data dual-driven anti-interference (anti-interference hybrid beamforming model K3D HBF) method with interference. Summary of the invention

[0004] In view of the above problems, the present invention provides a knowledge-data dual-driven interference suppression method for ground-to-air communication systems. Through the anti-interference hybrid beamforming model of the anti-interference K3D HBF, the interference-plus-noise (IPN) covariance matrix and the legal transceiver channel are used as inputs, adaptively mapped to the HBF weight matrix and used as outputs. The prediction accuracy is improved, and the proposed anti-interference hybrid beamforming model of the anti-interference K3D HBF achieves a lower mean square error (MSE), significantly reducing the complexity of interference suppression.

[0005] The present invention provides a knowledge and data dual-driven ground-to-air communication system interference suppression method, the specific steps are as follows:

[0006] Step 1. Construct a mean square error (MSE) model; Based on the mean square error (MSE) model, construct an optimization problem for the minimum mean square error of the received signal in the airport interference scenario;

[0007] Step 2. Construct a 3DConvTransformer model; use the 3DConvTransformer model to obtain the predicted interference plus noise IPN covariance matrix; obtain the observed IPN covariance matrix of the subcarrier from the predicted interference plus noise IPN covariance matrix;

[0008] Step 3. Based on the observed IPN covariance matrix of the subcarriers, an anti-interference hybrid beamforming model based on K3D HBF is constructed;

[0009] Step 4: Use the anti-interference hybrid beamforming model to obtain the interference suppression matrix.

[0010] Optionally, the expression for constructing the mean square error MSE model in step 1 is:

[0011]

[0012] in, Indicates The mean square error of subcarriers; represents the mean square value; Indicated in The signal vector transmitted on the subcarriers; Indicates Scaling factor for a given transmit power and noise power on a subcarrier; Indicated in The received signal vector on the subcarriers; It means finding the trace of a matrix; Indicates The ground-to-air channel matrix of subcarriers; and They represent the precoding matrix of the LGS end and the combination matrix of the AC end respectively; Indicated in The IPN covariance matrix of antenna domain reception on subcarriers; H represents conjugate transpose; The dimension is The identity matrix of .

[0013] Optionally, , , Indicates the first The baseband precoding matrix of subcarriers is represents the RF precoding matrix at the LGS end, Indicates the AC end The baseband combination matrix of subcarriers, Indicates the RF combination matrix at the AC end.

[0014] Optionally, the optimization problem of the minimum mean square error problem of the received signal in the airport interference scenario is expressed as:

[0015]

[0016]

[0017]

[0018]

[0019] in, Indicates the LGS end radio frequency precoding matrix The constant modulus constraint, They represent the LGS end radio frequency precoding matrix respectively No. Line The index of the column element; Represents the AC end RF combination matrix The constant modulus constraint, Respectively represent the AC end RF combination matrix No. Row, No. The index of the column element; X Indicates the total number of subcarriers.

[0020] Optionally, the specific steps for constructing the 3DconvTransformer model in step 2 are:

[0021] Get the sampled IPN covariance matrix;

[0022] Obtaining a subcarrier sampled IPN covariance matrix of a sampled IPN covariance matrix;

[0023] Obtaining a frame sampling IPN covariance matrix based on a subcarrier sampling IPN covariance matrix;

[0024] Get the simulated IPN covariance matrix;

[0025] The 3DConvTransformer model is obtained based on the simulated IPN covariance matrix and the sampled IPN covariance matrix training.

[0026] Optionally, the 3DConvTransformer model includes a 3D convolutional neural network and a Transformer network.

[0027] Optionally, the 3D convolutional neural network includes a 3D convolutional neural network encoder and a 3D convolutional neural network decoder.

[0028] Optionally, the Transformer network includes an embedding module, a transformer encoder, and a transformer decoder.

[0029] Compared with the prior art, the present invention has at least the following beneficial effects:

[0030] (1) The IPN covariance matrix prediction method of 3DConvTransformer proposed in the interference suppression method of the present invention can reduce the sampling error and improve the IPN covariance matrix prediction accuracy of multiple future frame lengths after obtaining the historical sampling interference covariance;

[0031] (2) The anti-interference beamforming scheme proposed by the interference suppression method of the present invention adopts the anti-interference K3D HBF method, which realizes the joint hybrid beamforming of transmission and reception with a lower number of iterations, and is suitable for interference scenarios with rapid channel changes;

[0032] (3) The interference suppression method of the present invention proposes an anti-interference beamforming scheme that adopts the anti-interference K3D HBF method, which replaces the step size of the gradient descent obtained by traditional iterative convergence with a learnable step size, thereby reducing the complexity and being interpretable. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The drawings are only for the purpose of illustrating particular embodiments and are not to be construed as limiting the invention.

[0034] Figure 1 Schematic diagram of a 3DConvTransformer model in the interference suppression method of the present invention;

[0035] FIG2( a ) is a schematic diagram of an anti-interference hybrid beamforming model based on K3D HBF of the interference suppression method of the present invention;

[0036] FIG2( b ) is a schematic diagram of obtaining a step size by gradient descent of a radio frequency precoder in the interference suppression method of the present invention;

[0037] FIG2( c ) is a schematic diagram of obtaining a step size by gradient descent of a radio frequency combiner in the interference suppression method of the present invention;

[0038] Figure 3 It is a flow chart of the interference suppression method of the present invention. DETAILED DESCRIPTION

[0039] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. In addition, the present invention can also be implemented in other ways different from those described herein, and therefore, the protection scope of the present invention is not limited by the specific embodiments disclosed below.

[0040] A specific embodiment of the present invention, as Figure 1-Figure 3 , a knowledge and data dual-driven interference suppression method for ground-to-air communication system, the specific steps are as follows:

[0041] Step 1. Construct a mean square error (MSE) model; Based on the mean square error (MSE) model, construct an optimization problem for the minimum mean square error of the received signal in the airport interference scenario;

[0042] Specifically, the receiving signal model of the AC end (i.e., the aircraft end) on the subcarrier in the data transmission stage is constructed, and the expression is:

[0043] (1)

[0044] in, Representative Frame No. The received signal at the AC end on the subcarrier; and Respectively represent Frame No. The baseband combination matrix and RF combination matrix at the AC end of each subcarrier, , , represents a complex value, Indicates the number of RF chains at the AC end; Indicates the number of antennas at the AC end; Indicates Frame No. LGS-AC (i.e., legal ground station to aircraft) channel matrix on subcarriers, , Indicates the number of antennas of LGS (i.e. legal ground station); and Respectively represent Frame No. The baseband precoding matrix and RF precoding matrix of the LGS end on the subcarriers, , , Indicates the number of radio frequency chains at LGS; , and Respectively represent Frame No. The legal signal on the LGS end on the subcarrier, the interference signal received on the AC end and the zero mean and variance are Gaussian white noise, , , ; Indicates the number of data streams transmitted on the subcarrier; represents the variance of the noise, express The identity matrix of H represents the conjugate transpose.

[0045] Further, ;The ground-to-air communication system divides time resources into multiple frames, each frame contains three stages, namely the interference estimation stage, the channel estimation stage and the data transmission stage.

[0046] Further, Frame No. LGS-AC channel matrix on subcarriers The expression is:

[0047] (2)

[0048] in, and Respectively represent Frame No. The complex gain of the LoS path (i.e., direct path) on the subcarrier and the The complex gain of the NLoS path (i.e., non-direct path); The steering vector representing the angle of arrival, ; The steering vector representing the departure angle, , H represents conjugate transpose; and They represent the azimuth and elevation of the LoS path departure angle at the LGS end, and They represent the azimuth and elevation angles of the LoS path arrival angle at the AC end respectively; represents the total number of NLoS paths, and Respectively represent the LGS end The azimuth and elevation angles of the NLoS path departure angle, and Respectively represent the AC end The azimuth and elevation of the NLoS path departure angle; is the Ricean factor, which is the power ratio between the LoS path and the NLoS path.

[0049] Further, Frame No. Complex gain of the LoS path on the subcarrier and the complex gain of the NLoS path The expression is:

[0050] (3)

[0051] (4)

[0052] in, and Respectively represent The NLoS path is Large-scale fading gain and delay of frames; represents an index; represents an imaginary unit; Indicates the symbol period of Orthogonal Frequency Division Multiplexing (OFDM); Indicates the total number of subcarriers; and Respectively represent the Los path in the Large-scale fading gain and delay of frames.

[0053] Furthermore, based on Frame No. LGS-AC channel matrix on subcarriers , baseband precoding matrix at the LGS end , RF precoding matrix at the LGS end , AC end baseband combination matrix and RF combination matrix , construct the mean square error MSE model, the expression is:

[0054] (5)

[0055] in, Indicates The mean square error of subcarriers; represents the mean square value; Indicated in The signal vector transmitted on the subcarriers is ; Indicates Scaling factor for a given transmit power and noise power on a subcarrier; Indicated in The received signal vector on the subcarriers; It means finding the trace of a matrix; Indicates The ground-to-air channel matrix of subcarriers; and They represent the precoding matrix of the LGS end and the combination matrix of the AC end, , , Indicates the first The baseband precoding matrix of subcarriers is represents the RF precoding matrix at the LGS end, Indicates the AC end The baseband combination matrix of subcarriers, Represents the RF combination matrix at the AC end; Indicates The IPN covariance matrix of antenna domain reception on subcarriers is: , Indicates Interference plus noise signal vector of subcarriers; H represents conjugate transpose; The dimension is The identity matrix of .

[0056] Further, , Indicates Interference signal vector received by AC end on subcarriers; Indicates subcarriers have zero mean and variance Gaussian white noise vector.

[0057] Based on the mean square error (MSE) model, the optimization problem of the minimum mean square error of the received signal in the airport interference scenario is constructed, and the expression is:

[0058] (6a)

[0059] (6b)

[0060] (6c)

[0061] (6d)

[0062] in, Indicates the LGS end radio frequency precoding matrix The constant modulus constraint, They represent the LGS end radio frequency precoding matrix respectively No. Line The index of the column element; Represents the AC end RF combination matrix Constant modulus constraint. Respectively represent the AC end RF combination matrix No. Row, No. The index of the column element; X Indicates the total number of subcarriers.

[0063] The present invention uses the forward link transmission of the Ground to Air (G2A) broadband MIMO system as the communication with a rapidly changing channel, and an antenna number is The number of legal ground stations (LGS) and an antenna is The legitimate ground station LGS and the aircraft AC both use antenna units formed by uniform planar arrays (UPA) and HBF (hybrid beamforming) architecture. The legitimate ground station LGS uses OFDM (orthogonal frequency division multiplexing) communication scheme. The subcarriers transmit independent data streams, The signal vector transmitted on the subcarrier ,satisfy The number of data streams transmitted on each subcarrier satisfies To ensure the effectiveness of communication. The present invention considers the high-power random interference signal generated by the interfering ground station IGS, the duration of which is approximately equal to the inverse of the signal bandwidth, that is, the narrow random pulse of the time domain sampling interval of the OFDM symbol to model. By modeling the minimum mean square error problem between the received signal and the transmitted signal, the joint beamforming technology of transmission and reception is used for interference suppression, and the damage of random interference signals to useful signals is reduced to improve the reliability of communication.

[0064] Step 2. Construct a 3DConvTransformer model; use the 3DConvTransformer model to obtain the predicted interference plus noise IPN covariance matrix; obtain the observed IPN covariance matrix of the subcarrier from the predicted interference plus noise IPN covariance matrix.

[0065] The specific steps are:

[0066] Get the sampled IPN covariance matrix;

[0067] The subcarrier sampling IPN covariance matrix of the sampling IPN covariance matrix is ​​obtained, and the expression is:

[0068]

[0069] in, Indicates Frame No. Subcarrier sampling IPN covariance matrix; is the total number of samples; Indicated in The first Subsampled IPN covariance matrix.

[0070] It is understandable that as the total number of samples The increase in, under the stationary and ergodic assumptions, Converges to the simulated IPN covariance matrix .

[0071] Obtaining a frame sampling IPN covariance matrix based on a subcarrier sampling IPN covariance matrix;

[0072] The simulated IPN covariance matrix is ​​obtained, and the expression is:

[0073] (8)

[0074] in, Indicates Frame simulation IPN covariance matrix; Indicates The frame sampled IPN covariance matrix of the frame; Indicates Estimation error of the frame;

[0075] Further, , is subject to zero mean and variance Gaussian distribution. Since With time domain correlation brought by time-varying channels, The elements in are related to each other in the antenna domain through the steering vectors.

[0076] The 3DConvTransformer model is obtained based on the simulated IPN covariance matrix and the sampled IPN covariance matrix. The predicted interference plus noise IPN covariance matrix is ​​obtained using the 3DConvTransformer model. The IPN covariance matrix of antenna domain reception on subcarriers .

[0077] Specifically, the 3DConvTransformer model consists of a 3D convolutional neural network (3DCNN) network cascaded with a Transformer network.

[0078] The present invention adopts methods such as interpolation or compressed sensing to restore the full-space observation, overcoming the problem that the existing research deploys empty subcarriers in each OFDM symbol to estimate the IPN covariance matrix during the interference estimation stage, making the projection subspace smaller than the total subcarrier space and causing the interference projection to be lacking in rank.

[0079] Based on the sampled IPN covariance matrix obtained above, the 3DConvTransformer model of the present invention can accurately predict the interference plus noise IPN covariance matrix in advance, thereby reducing the pilot overhead caused by the IPN covariance matrix estimation.

[0080] See also Figure 1 , the 3D convolutional neural network includes a 3D convolutional neural network encoder and a 3D convolutional neural network decoder.

[0081] Furthermore, the 3D convolutional neural network encoder consists of an input convolutional layer and 5 residual blocks (ResBlock).

[0082] Furthermore, the input convolutional layer includes filters, batch normalization units (Batch Normalization, BN) and rectified linear units (Rectified Linear Unit, ReLU).

[0083] Among them, the filter adopts the convolution kernel size of of Sub-filters, Sub-filters generate feature maps and use the rectified linear unit (ReLU) to prevent the gradient from disappearing. Each residual block consists of three convolutional layers, each of which consists of 64, 128, and The convolution kernel size is The sub-filters are composed of;

[0084] In the batch normalization unit (Batch Normalization, BN) and the rectified linear unit (Rectified Linear Unit, ReLU) of the present invention, BN standardizes the activation values ​​of each layer so that they have zero mean and unit variance, which is used to speed up the training speed and enhance the denoising effect. ReLU avoids the gradient vanishing problem by introducing a nonlinear activation function.

[0085] The 3D convolutional neural network decoder consists of The convolution kernel size is The output convolution layer of the sub-filters is used to reconstruct the IPN covariance matrix.

[0086] When using, The frame-by-frame IPN covariance matrix Input 3D convolutional neural network to obtain the Frame denoising frame sampling IPN covariance matrix ; Frame denoising frame sampling IPN covariance matrix Input the parallel Transformer network to obtain the L-frame prediction interference plus noise IPN covariance matrix.

[0087] Furthermore, if Figure 1As shown in Figure 1, the parallel Transformer network consists of an embedding module, a transformer encoder, and a transformer decoder. The embedding module is placed before the encoder and decoder stack to convert the discrete Frame denoising frame sampling IPN covariance matrix is converted into a continuous vector containing position information. Then, the transformer encoder starts from Frame denoising frame sampling IPN covariance matrix Extract features from the vector and capture the Frame denoising frame sampling IPN covariance matrix The interrelationships and dependencies between vector elements. By combining the transformer encoder features with the input reference sequence, the transformer decoder can more accurately predict the interference plus noise IPN covariance matrix.

[0088] Specifically, the embedding module is used to sample the IPN covariance matrix for each denoised frame: The real and imaginary parts are extracted, vectorized and concatenated into a vector. The embedding module operates by including the original input embedding and the position embedding, where the original input embedding is used to sample the IPN covariance matrix of the denoised frame output by the 3D convolutional neural network. Convert to a real vector and the imaginary vector Vector set of composition Convert to a The set of original embedding vectors ,in, , , represents a real number. The position embedding is obtained by sampling the IPN covariance matrix of the denoised frame The position information of enables the Transformer network to take advantage of the order of the sequence. Therefore, the embedding module finally outputs indivual The encoder matrix composed of dimensional vectors and G+L decoder matrices ,in, Indicates the output of the embedded module indivual dimensional encoder vector, Indicates the output of the embedded module indivual dimensional decoder vector, The 0 in is a zero-padded frame, and the number of zero-padded frames is the same as the number of predicted frames L.

[0089] Specifically, the transformer encoder includes an encoder attention mechanism layer, a fully connected layer, and an encoder residual connection and normalization layer; when used, the transformer encoder converts the encoder matrix Feed it into the encoder attention mechanism layer; the encoder attention mechanism layer is used to capture the encoder matrix The global dependencies in the encoder matrix Convert to key matrix , query matrix Sum Matrix Three trainable weight matrices; then, calculate the different encoder matrices Vector The attention score between the encoder matrix The importance of each vector in makes the Transformer network pay more attention to the parts that have a significant impact on the prediction of future IPN covariance vectors. Then, these attention scores are combined with the value matrix Combine to get the encoder matrix The final expression .

[0090] Furthermore, in order to prevent the gradient vanishing problem caused by the encoder attention mechanism, the final representation is The final encoder output is obtained through encoder residual connection and normalization layer processing. .

[0091] Specifically, the transformer decoder includes a masked attention mechanism layer, a full attention mechanism layer, and a decoder residual connection and normalization layer.

[0092] When used, the transformer decoder first converts the decoder matrix Input mask attention mechanism layer obtains feature matrix , the masked attention mechanism ensures that the transformer decoder focuses on known frames and ignores zero-padded frames; the feature matrix Send it to the full attention mechanism layer to hide the variables With the feature matrix Combine outputs and hidden variables ; Next, combine the hidden variables The final decoder output is obtained through the decoder residual connection and normalization layer ; Finally, the final decoder output The real and imaginary parts of the mixed IPN covariance matrix are combined to obtain the mixed IPN covariance matrix, and then the The predicted interference plus noise IPN covariance matrix is ​​obtained by column.

[0093] Specifically, the specific steps for training the 3DConvTransformer model are as follows:

[0094] S1. Obtain a multi-frame sampled IPN covariance matrix training set from the sampled IPN covariance matrix, and set the total number of training times to , the learning rate is .

[0095] S2. Order n =0, when n =0, it is the initial 3DConvTransformer model;

[0096] S3. Input the multi-frame sampled IPN covariance matrix training set into the first n Train the 3DConvTransformer model again; extract the real and imaginary parts of the sampled IPN covariance matrix training set of the input multi-frame;

[0097] S4. Use n The 3D convolutional neural network of the 3DConvTransformer model is trained to extract features;

[0098] S5. Use n The encoder output of the Transformer network of the trained 3DConvTransformer model Final encoder output , decoder output Final decoder output ;

[0099] S6. Output the final decoder The real and imaginary parts of the mixed IPN covariance matrix are combined to obtain the mixed IPN covariance matrix. Get the first n The predicted interference plus noise IPN covariance matrix of the 3DConvTransformer model trained this time.

[0100] S7. Determine the number of training times n With the total number of training times N, if n If it is less than N, return to step 3. If the total number of training times is equal to N, the 3DConvTransformer model converges and stops training.

[0101] Furthermore, the 3DConvTransformer model is trained by minimizing the normalized mean square error between the predicted interference plus noise IPN covariance matrix output by the 3DConvTransformer model and the simulated IPN covariance matrix. When the 3DConvTransformer model converges and stops training, the loss function of minimizing the 3DConvTransformer model is expressed as:

[0102]

[0103] in, is the first t The predicted interference plus noise IPN covariance matrix of the frame, .

[0104] From the derivation of the mean square error MSE model given in the optimization problem (6), we can see that only in the The IPN covariance matrix of antenna domain reception on subcarriers and The ground-to-air channel matrix of subcarriers The present invention assumes that The ground-to-air channel matrix of subcarriers The matrix is ​​known, and we only focus on the effect of random interference. The IPN covariance matrix of antenna domain reception on subcarriers Obviously, it will take up a lot of pilot overhead, thus reducing communication efficiency. In addition, since each The IPN covariance matrix of antenna domain reception on subcarriers and The ground-to-air channel matrix of subcarriers The optimization problem (6) must be solved whenever there are changes, so the proposed solution method should work under relatively low computational complexity. However, the optimization problem (6) is difficult to handle due to the highly coupled optimization variables and constant mode constraints, and the existing iteration-based optimizers require a lot of calculations to converge. In order to address the above two challenges, the present invention proposes an interference statistics information prediction strategy in the second step, which uses the past observation set to predict the IPN covariance matrix of multiple future frames, with low pilot overhead and high robustness. Then in the third step, based on the The IPN covariance matrix of antenna domain reception on subcarriers , , execute the K3D interference suppression algorithm to optimize the HBF matrix, thereby achieving low-complexity and high-robustness interference suppression performance within a predetermined number of iterations.

[0105] Step 3. Based on the observed IPN covariance matrix of the subcarriers, an anti-interference hybrid beamforming model based on K3D HBF is constructed.

[0106] Based on the observed IPN covariance matrix of the subcarriers, in order to solve the MSE minimization problem given in the optimization problem (6) of the minimum mean square error problem of the received signal in the airport interference scenario, the solution is based on the iterative optimization (AO) idea. , , , Decouple and iterate the solution. It is processed hierarchically through outer and inner loops, where the outer loop iterates between the precoder and combiner, and the inner loop iterates between baseband and RF processing.

[0107] The present invention focuses on the design of the algorithm for interference suppression by the AC-side combiner, and the LGS-side precoder design problem can be solved in the same way.

[0108] Specifically, the AC end combiner subproblem is designed, and the expression is:

[0109]

[0110] (10b)

[0111] In the formula, Represents the AC end RF combination matrix; Indicates the AC end Subcarrier baseband combination matrix; Represents the matrix trace symbol; Indicates the AC end The conjugate transpose of the subcarrier baseband combination matrix; represents the conjugate transpose of the AC-side RF combination matrix; Indicates Scaling factor for a given transmit power and noise power on a subcarrier; Indicates The negative first power of the scaling factor of the subcarriers, the ground-to-air channel matrix and the LGS precoding matrix The product of the three; express Conjugate transpose of a matrix; Indicates The observed IPN covariance matrix of subcarriers; The dimension is The identity matrix of and Respectively represent the AC end RF combination matrix No. Row, No. The index of the column element.

[0112] Further, .

[0113] The objective function (10a) in the AC-side combiner subproblem is relative to Subcarrier baseband combining matrix is convex, and we get The closed-form solution of the baseband combination matrix on the subcarriers is expressed as:

[0114] (11)

[0115] in, express Multiply The conjugate transpose of ;

[0116] Further, .

[0117] Baseband Combining Matrix Substituting into the objective function (10a) in the AC end combiner subproblem, we obtain the rewritten mean square error MSE model, which is expressed as:

[0118] (12)

[0119] in, Represents the rewritten mean square error MSE model about the The AC-side RF combination matrix of subcarriers Function of Indicates The conjugate transpose of the RF combination matrix at the AC end of the subcarriers , No. The IPN covariance matrix of antenna domain reception on subcarriers Combined with AC end RF matrix The product of three.

[0120] Further, .

[0121] Accordingly, the AC end combiner subproblem (10) is simplified and transformed into the RF combination matrix The optimization problem is expressed as:

[0122] (13a)

[0123] (13b)

[0124] Furthermore, the constant modulus constraint (13b) of the above RF combination matrix causes the optimization problem to be non-convex. The traditional manifold optimization algorithm (MO) method is very time-consuming to search the step size through the Armijo backtracking line, and is not suitable for the rapidly changing ground-to-air channel scenario. The present invention constructs an interference-suppressed K3D HBF network (anti-interference hybrid beamforming model), regards the step size of gradient descent as a trainable parameter, and can improve performance within a limited number of iterations.

[0125] Specifically, the K3D HBF network (i.e., the anti-interference hybrid beamforming model based on K3D HBF) Subcarrier ground-to-air channel matrix and The subcarrier observed IPN covariance matrix As input, the step size is intelligently learned to minimize the loss function, namely the mean square error MSE model defined in Equation (6). Compared with Armijo backtracking line search, the learning-based method can find a certain set of step sizes to achieve optimal performance under a predefined number of external and internal iterations. In addition, in order to prevent the gradient descent algorithm from falling into a local optimum, the K3D HBF network avoids this problem by "empiricism" obtained by inputting a large number of training samples.

[0126] The interference suppression K3D HBF network constructed by the present invention is shown in FIG2 (a). The structure of the AO algorithm is used as the network model, and a large amount of training data is used. Learn the gradient descent step size. Map the outer iteration to layers, each of which contains a precoder and a combiner. The operation of the gradient descent step of the RF precoder and RF combiner is shown in Figure 2(b) and Figure 2(c), respectively: It can be clearly observed that compared with the data-driven "black box" algorithm, K3D HBF has a simpler structure and fewer trainable parameters. Specifically, the external iteration Layer The baseband precoder, RF precoder, baseband combiner and RF combiner updated by step gradient descent are represented as baseband precoding matrices , RF precoding matrix , baseband combination matrix , RF combination matrix , and use represents a set of trainable parameters, where Indicates the first External iterations The set of step sizes for gradient descent steps, Indicates the first Layer external iteration The step size set of step gradient descent. The anti-interference hybrid beamforming model based on K3D HBF includes four modules: RF precoder, baseband precoder, RF combiner, and baseband combiner, which are described in detail as follows:

[0127] 1) RF precoder module:

[0128] This module contains Step gradient update, Layer external iteration The RF precoding matrix of the LGS end obtained by step gradient descent As input, the RF precoder is Step Riemannian Gradient Descent Update until The step gradient update ends.

[0129] The expression of the update matrix of the RF coding matrix at the LGS end is:

[0130] (14)

[0131] In the formula, Indicates i Layer external iteration j +1 step of gradient descent to obtain the updated matrix of the RF precoding matrix at the LGS end; Indicates External iterations Indicates Layer external iteration The step size obtained by step gradient descent, Indicates Layer external iteration The RF precoding matrix of the LGS end obtained by step gradient descent The Riemann gradient of Represents the RF precoding matrix The Riemann gradient of is used to find the Frobenius norm.

[0132] Then, the i Layer external iteration j +1 step gradient descent to obtain the updated matrix of the RF precoding matrix at the LGS end Back to the complex circular manifold, we get i Layer external iteration j The RF precoding matrix at the LGS end obtained by +1 step gradient descent is expressed as:

[0133] (15)

[0134] In the formula, Indicates i Layer external iteration j +1 step gradient descent to obtain the updated matrix of the RF precoding matrix at the LGS end No. Line Elements of a column; Indicates a retraction operation. express No. Line The modulus of the elements of the column; Indicates Layer external iteration j +1 step gradient descent to obtain the RF precoding matrix of the LGS end Line Elements of a column;

[0135] Until the External iterations After the step gradient update, the output will be Layer external iteration The RF precoding matrix of the LGS end obtained by step gradient descent .

[0136] 2) Baseband precoder module:

[0137] This module directly applies the closed-form solution In the external iteration The baseband precoding matrix of the LGS end on all subcarriers .

[0138] 3) RF combiner module:

[0139] The RF combination module is similar to the RF precoding module. External iterations +1 step of gradient descent to update the RF combination matrix at the AC end , until After the gradient descent step, the final AC-side RF combination matrix is ​​output. .

[0140] 4) Baseband combiner module:

[0141] Similar to the baseband precoder module, this module uses a closed-form solution for the In the external iteration The AC end baseband combination matrix of all subcarriers .

[0142] Furthermore, in order to improve the performance of K3D HBF within a limited number of iterations, full digital anti-interference hybrid beamforming is used for initial setting, because the optimal anti-interference hybrid beamforming will eventually be close to full digital anti-interference beamforming. minimize The loss function is the mean square error MSE model. The loss function is expressed as:

[0143] (16)

[0144] Based on the loss function defined above, the step size is updated through the back-propagation mechanism. In order to improve the training efficiency, a layered training strategy is adopted. Specifically, the specific steps of the training process of the anti-interference hybrid beamforming model K3D HBF are as follows:

[0145] Step 31. Given a training set of all subcarriers;

[0146] The training set of subcarriers includes Subcarrier Ground-to-Air Channel and The IPN covariance matrix of antenna domain reception on subcarriers ;

[0147] Step 32. Initialize trainable parameters , the number of training epochs , the number of training batches ;

[0148] Step 33. Initialize the baseband precoding matrix for all subcarriers ; According to the first External iterations The step size of the gradient descent step Calculate the External iterations The RF precoding matrix obtained by step gradient descent and the baseband precoding matrix ; According to the first External iterations The step size of the gradient descent step Calculate the External iterations The RF combination matrix obtained by step gradient descent and baseband combination matrix ;

[0149] Step 34. Calculate the loss value of the anti-interference hybrid beamforming model based on K3D HBF according to formula (16), and update the training parameters of the neural network through the back-propagation mechanism. ;

[0150] Step 35. Determine whether the loss value converges. If so, end the training and save the training parameters , otherwise repeat steps 32-34.

[0151] Step 4: Use the anti-interference hybrid beamforming model to obtain the interference suppression matrix.

[0152] The interference suppression matrix includes the External iterations The final baseband precoding matrix obtained after step gradient descent is , the final RF pre-coding matrix , the final baseband combination matrix And the final RF combination matrix .

[0153] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A knowledge data dual-driven K3D ground-to-air communication system interference suppression method, characterized in that: The specific steps are as follows: Step 1. Construct a mean square error (MSE) model; Based on the mean square error (MSE) model, construct an optimization problem for the minimum mean square error of the received signal in the airport interference scenario; Step 2. Build a 3DConvTransformer model and transform The frame-by-frame IPN covariance matrix Input the 3DConvTransformer model to obtain the predicted interference plus noise IPN covariance matrix of the future L frames; Step 3. Based on the observed IPN covariance matrix of the subcarriers, an anti-interference hybrid beamforming model based on K3D hybrid beamforming HBF is constructed; The anti-interference hybrid beamforming model based on K3D hybrid beamforming HBF includes RF precoder, baseband precoder, RF combiner and baseband combiner; The RF precoder module includes Step gradient update, Layer external iteration The RF precoding matrix of the legal ground station LGS end obtained by step gradient descent As the The input of the external iteration of the layer, the RF precoder is first Step Riemannian Gradient Descent , update the RF precoding matrix of the LGS end until Step gradient update ends; The baseband precoder module uses a closed-form solution In the external iteration The baseband precoding matrix of the LGS end on all subcarriers is gradually decreased ; The RF combiner module uses the first External iterations +1 step of gradient descent to update the RF combination matrix at the AC end of the aircraft , until After the gradient descent step, the final AC-side RF combination matrix is ​​output. ; The baseband combiner module uses a closed-form solution In the external iteration Step gradient descent AC end baseband combination matrix of all subcarriers ; The specific steps of the training process of the anti-interference hybrid beamforming model based on K3D HBF are as follows: Step 31. Given a training set of all subcarriers; The training set of subcarriers includes Subcarrier Ground-to-Air Channel and The IPN covariance matrix of antenna domain reception on subcarriers ; Step 32. Initialize trainable parameters , the number of training epochs , the number of training batches ,in, Indicates the first Layer external iteration The step size set of step gradient descent, Indicates the first External iterations The set of step sizes for the gradient descent steps; Step 33. Initialize the baseband precoding matrix for all subcarriers , baseband combination matrix , RF precoding matrix and RF combination matrix ; According to the first External iterations Step size of the gradient descent step Calculate the External iterations The RF precoding matrix obtained by step gradient descent and the baseband precoding matrix ; According to the first External iterations Step size of the gradient descent step Calculate the External iterations The RF combination matrix obtained by step gradient descent and baseband combination matrix ; Step 34. According to the loss function Calculate the loss value of the anti-interference hybrid beamforming model based on K3D HBF, and update the training parameters of the neural network through the back-propagation mechanism ; Step 35. Determine whether the loss value converges. If so, end the training and save the training parameters , otherwise continue training; Step 4. Obtain the interference suppression matrix using the anti-interference hybrid beamforming model; The expression for constructing the mean square error MSE model in step 1 is: in, Indicates The mean square error of subcarriers; represents the mean square value; Indicated in The signal vector transmitted on the subcarriers; Indicates Scaling factor for a given transmit power and noise power on a subcarrier; Indicated in Received signal vector on subcarriers; It means finding the trace of a matrix; Indicates The ground-to-air channel matrix of subcarriers; and They represent the precoding matrix of the LGS end and the combination matrix of the AC end respectively; Indicated in The IPN covariance matrix of antenna domain reception on subcarriers; H represents conjugate transpose; The dimension is The identity matrix of Indicates the number of data streams transmitted on the subcarrier; , , Indicates the first The baseband precoding matrix of subcarriers is represents the RF precoding matrix at the LGS end, Indicates the AC end The baseband combination matrix of subcarriers, Represents the RF combination matrix at the AC end; The optimization expression of the minimum mean square error problem of the received signal in the airport interference scenario is: in, Indicates the LGS end radio frequency precoding matrix The constant modulus constraint, They represent the LGS end radio frequency precoding matrix respectively No. Line The index of the column element; Represents the AC end RF combination matrix The constant modulus constraint, Respectively represent the AC end RF combination matrix No. Row, No. The index of the column element; Indicates the total number of subcarriers; The specific steps to build the 3DconvTransformer model in step 2 are: Get the sampled IPN covariance matrix; Obtaining a subcarrier sampled IPN covariance matrix of a sampled IPN covariance matrix; Obtaining a frame sampling IPN covariance matrix based on a subcarrier sampling IPN covariance matrix; Add the frame sampling IPN covariance matrix to the estimated error to obtain the simulated IPN covariance matrix; The 3DConvTransformer model is obtained by training based on the simulated IPN covariance matrix and the sampled IPN covariance matrix; the 3DConvTransformer model is trained by minimizing the normalized mean square error between the predicted interference plus noise IPN covariance matrix output by the 3DConvTransformer model and the simulated IPN covariance matrix.

2. The interference suppression method of the ground-to-air communication system according to claim 1, characterized in that: The 3DConvTransformer model includes a 3D convolutional neural network and a Transformer network.

3. The interference suppression method of the ground-to-air communication system according to claim 2, characterized in that: The 3D convolutional neural network includes a 3D convolutional neural network encoder and a 3D convolutional neural network decoder.

4. The interference suppression method of the ground-to-air communication system according to claim 2, characterized in that: The Transformer network consists of an embedding module, a transformer encoder, and a transformer decoder.

Images