Multi-antenna interference elimination method based on neural network model

Abstract

The invention belongs to the technical field of unmanned aerial vehicle anti-interference communication, and particularly relates to a multi-antenna interference elimination method based on a neural network model. According to the method disclosed by the invention, the signals are acquired by using the plurality of receiving antennas, and the space-time characteristics of the target signals and the interference signals are learned and modeled through the neural network model, so that the time-frequency synchronization of the signals is realized and the strong interference signals are effectively suppressed on the premise that channel state information of a transmitter or an interference source does not need to be known priori. Simulation experiments prove that the method can still keep a low bit error rate even under the extreme condition that the interference signal power is 40 dB higher than that of a target signal, and the communication reliability of the unmanned aerial vehicle in the complex electromagnetic environment is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of unmanned aerial vehicle (UAV) anti-interference communication technology, and in particular relates to a multi-antenna interference elimination method based on a neural network model. Background Art

[0002] Unmanned aerial vehicles (UAVs) are flexible, low-cost, and versatile flying platforms. They are currently widely used in a variety of fields, including aerial photography, disaster relief, material delivery, and combat support. UAVs are typically remotely controlled by ground or space-based base stations, with control commands transmitted from the base station to the UAV in flight via wireless communication links. Although there is typically good line-of-sight communication between the UAV and the base station, the long distance between them results in significant signal path loss, making the wireless link susceptible to external interference. Therefore, researching and developing anti-interference communication technologies for UAVs operating in interference environments is crucial for ensuring the stability of their communication links and the reliability of mission execution.

[0003] Currently, commonly used communication interference mitigation technologies primarily include spread spectrum and frequency hopping. However, these technologies often suffer from low spectrum utilization, making it difficult to achieve high communication rates given the same frequency resources. To improve spectrum efficiency, co-channel interference mitigation technologies based on multi-antenna systems have garnered widespread attention in recent years. Furthermore, with the development of artificial intelligence (AI), neural network-based multi-antenna interference mitigation methods are emerging as a solution and are beginning to demonstrate their unique advantages. Typical multi-antenna interference mitigation methods include adaptive beamforming and semi-blind source separation. Adaptive beamforming leverages the direction of arrival (DoA) of the target signal to dynamically adjust the beam direction, forming a narrow beam to enhance the target signal and suppress interference. However, in practical applications, accurate DoA information for the target signal is often difficult to obtain, limiting the applicability of this method. Another class of methods relies on channel state information or pilot sequences within the target signal for interference mitigation. Channel-information-based methods typically require estimating the communication channel at the receiver and extracting the interference subspace from the covariance matrix of the received signal. However, these methods rely on so-called "interference gaps," meaning that the interference signal must be tracked to form a time window useful for estimation. Performance degrades significantly when the interference gap is missing or too short. Semi-blind source separation techniques, on the other hand, utilize known pilot information to reconstruct and remove the target signal components from the received signal, thereby obtaining the interference subspace. However, this method requires strict time and frequency synchronization between the transmitter and receiver, which is often difficult to achieve in practical systems, limiting its practicality and deployment flexibility. Summary of the Invention

[0004] To address the above issues, the present invention proposes a neural network-based multi-antenna anti-interference technology. This method trains a neural network model to learn target signal characteristics and interference patterns in complex environments, achieving efficient interference suppression without the need for precise channel state information or prior time-frequency synchronization. The method of the present invention aims to utilize the preamble sequence to achieve both time / frequency synchronization and interference cancellation. Unlike existing methods, the proposed method does not require information about the communication and interference channels. Furthermore, the proposed method does not require complex matrix operations (such as matrix inversion) required by existing methods to solve the optimization problems of time / frequency synchronization and interference cancellation. Furthermore, the neural network structure is simple, requiring fewer parameters.

[0005] The technical solution of the present invention is:

[0006] A multi-antenna interference cancellation method based on a neural network model is used for downlink communication scenarios from a base station to an unmanned aerial vehicle (UAV). It is defined that there are K unknown interference sources in the environment. The base station and each interference source are equipped with only a single antenna, and the UAV is equipped with N antennas. ; The base station will periodically generate the preamble sequence , is the length of the leading sequence, As the target signal, it is transmitted from the base station to the UAV, where the K interference source signals are recorded as , each interference source Independent of each other and both with the target signal Unrelated; the UAV signal reception model is defined as:

[0007] ,

[0008] in, To receive the signal, represents the channel vector between the base station and the UAV, represents the channel vector between the kth interference source and the UAV, represents the Gaussian white noise vector, Indicates the frequency offset between the base station and the UAV. The target signal is affected by the form of represents the unknown delay between the transmitter and the receiver;

[0009] Interference cancellation methods include:

[0010] Construct a neural network for interference cancellation, which includes a spatial domain filter and frequency offset compensation module ,in The input is the received signal matrix , the output is the spatial filter coefficient :

[0011] ,

[0012] ,

[0013] in, is an underdetermined matrix or a full rank matrix;

[0014] The input is also the received signal matrix , the output is the frequency offset compensation value ,when When , the frequency offset is completely eliminated, and the corresponding optimization problem with frequency offset and interference is expressed as the first optimization problem:

[0015] ,

[0016] Input the received signal matrix to the neural network model , the model updates the module synchronously according to the value of the loss function and The network layer parameters in , thereby solving the first optimization problem, so that the frequency offset compensation Approaching the true value At the same time, the corresponding spatial filter coefficients can be obtained To remove interference signals;

[0017] Based on the obtained neural network, the unknown delay The value range is set to [1, ], is the maximum possible delay, according to The value range of adjusts the input and output elements of the neural network, specifically:

[0018] At the input of the neural network, the received signal dimension is expanded to:

[0019] ,

[0020] At the output of the neural network, Output coefficients of a single spatial filter Expanded to the spatial domain filter coefficient matrix, that is ; Similarly, The output frequency offset compensation is from Expand to , define the delay estimation matrix:

[0021] ,

[0022] Where d is an assumed delay value, , take out the corresponding output matrix of the neural network and Acts on ,get:

[0023] ,

[0024] At time t Expressed as:

[0025] ,

[0026] The recovered signal under the assumed delay d With the leading signal The mean square error between them is expressed as:

[0027] ,

[0028] Due to the existence When training a neural network, it is necessary to calculate indivual , these mean square errors are summed together as the final loss function of the neural network model:

[0029] ,

[0030] The corresponding optimization problem is the second optimization problem:

[0031] ,

[0032] when hour, and After successful alignment, the second optimization problem degenerates into the first optimization problem. and , can successfully restore the signal; thus, when there is a time delay, the neural network minimizes it by continuously updating the parameters , so that the composition of various can be minimized, and finally find the one with the smallest value after training. , and its corresponding label That is, it is considered to be equal to the delay value, and the estimated value is recorded as ;

[0033] After the neural network training is completed, according to the estimated delay value Select the corresponding and , and use both together to Get the recovery signal .

[0034] The beneficial effects of the present invention are: without requiring any prior knowledge of channel state information or interference sources, the method can achieve key functions such as time synchronization, frequency offset compensation, and interference suppression, relying solely on the preamble sequence and received signal. Experimental verification demonstrates that this method outperforms existing two-stage filtering anti-interference methods, achieving a low bit error rate even in the presence of strong interference. Furthermore, the designed neural network has a simple structure, a small parameter size, and a fast convergence speed, making it suitable for efficient deployment in practical systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is the initial frame structure of the transmitted signal.

[0036] Figure 2 This is a time domain diagram of the received signal.

[0037] Figure 3 Schematic diagram of the signal transmission processing flow in the communication system.

[0038] Figure 4 This is the corresponding signal receiving and processing flow in the communication system.

[0039] Figure 5 The bit error rate (BER) of the proposed method and the two-stage filtering method is demonstrated under different signal-to-interference ratio (SIR) and signal-to-noise ratio (SNR) values.

[0040] Figure 6 When eliminating interference from a certain frame signal, the neural network loss function is Schematic diagram of changes with training.

[0041] Figure 7 When performing interference elimination on a certain frame signal, When , the corresponding mean square error Schematic diagram of the changes in .

[0042] Figure 8 The BER of the two methods is shown at different SNRs and preamble lengths when SIR = -30dB. DETAILED DESCRIPTION

[0043] The present invention will be described in detail below with reference to the accompanying drawings.

[0044] Signal sending method:

[0045] Consider the downlink communication scenario from the base station to the drone. There are K unknown interference sources in the environment. The base station and each interference source are equipped with only a single antenna, while the drone is equipped with N antennas. The number of antennas meets In this scenario, the base station will periodically generate a preamble sequence , is the length of the preamble sequence, and the target signal Transmitted from the base station to the UAV. Among them, the K interference source signals are recorded as , each interference source Independent of each other and both with the target signal Not related to each other.

[0046] Signal reception method:

[0047] The present invention considers the case of single-carrier narrowband. The received signal model in this scenario can be expressed as:

[0048]

[0049] in To receive the signal, represents the channel vector between the base station and the UAV, represents the channel vector between the kth interference source and the UAV, Represents the Gaussian white noise vector. In the above formula, Indicates the frequency offset between the base station and the UAV. The target signal is affected by the target signal. Represents the unknown delay between the transmitter and the receiver.

[0050] The present invention is based on the fact that the UAV receiver cannot know the target signal channel in advance. and interference channels In this scenario, assuming that the channel vector remains unchanged within a coherent time slot, the signal-to-noise ratio (SNR) is assumed to be at a high level.

[0051] Signal recovery method:

[0052] The present invention uses the preamble sequence as a reference signal and uses a neural network to design a space-time filter with frequency offset compensation to make the filtered signal as close to the reference signal as possible. The specific implementation steps are as follows:

[0053] Considering that the received signal is affected by time delay, frequency deviation, and strong interference, we start with the simplest case, where only the interference signal exists. The receiving model can be simplified as follows:

[0054]

[0055] It is not difficult to find that the way to eliminate interference is to design a spatial filter to project the channel vector of the interference source signal to the null space, while retaining the target signal. , and its effect on the received signal can be expressed as:

[0056]

[0057] Where H represents the conjugate transpose of the vector. The purpose of the spatial filter is to make the output after filtering Get as close to the target signal as possible ,but The following conditions must be met:

[0058]

[0059] When the noise effect is ignored, The signal restored by the spatial domain filter is .

[0060] The conditions can be equivalently expressed as:

[0061]

[0062] in .because , is an underdetermined matrix or a full rank matrix, then the equation is satisfied The spatial filter coefficients of There must be a solution, that is, it must exist It can eliminate interference while preserving the target signal.

[0063] Define the leading signal vector as , the received signal matrix corresponding to its length is .

[0064] We design the first module of the neural network model and denote it as . The input is the received signal matrix , the output is the spatial filter coefficient .Will Acts on The interference-eliminating signal vector can be obtained , by calculating the signal With the leading sequence The mean square error between , we can get the loss function of the neural network model :

[0065]

[0066] The neural network model is based on the loss function The size of is updated by the back propagation algorithm The network layer parameters are used to minimize the filtered recovery signal With the leading sequence The difference between them is calculated, and a spatial filter that can effectively eliminate interference is output. The optimization problem corresponding to the above process is:

[0067]

[0068] Then, a frequency deviation is introduced into the transmission process of the target signal. , the received signal can be expressed as:

[0069]

[0070] The impact of frequency offset on the target signal cannot be eliminated by linear filtering, so frequency offset compensation is also required for the signal during spatial filtering. , the neural network model adds a second module , used to achieve signal frequency deviation compensation.

[0071] Module The input is also the received signal matrix , the output is the frequency offset compensation value Applying spatial filtering and frequency offset compensation to the received signal yields:

[0072]

[0073] Depend on It can be seen that when When the frequency offset is completely eliminated, the optimization problem degenerates into The loss function of the neural network model is updated as follows:

[0074]

[0075] The above optimization problem in the presence of frequency offset and interference can be expressed as:

[0076]

[0077] Input the received signal matrix to the neural network model After that, the model updates the module synchronously according to the value of the loss function and The network layer parameters in , thus solving the problem , so that the frequency offset compensation Approaching the true value At the same time, the corresponding spatial filter coefficients can be obtained To remove interference signals.

[0078] After solving the problems of frequency deviation and strong interference, we further introduce unknown delay into the transmission process. , the received signal is expressed as Without loss of generality, assume is a random signal.

[0079] Considering that the interference signal and Gaussian white noise always exist and are random, the expression Interference in and Gaussian white noise vector The impact of delay is not discussed.

[0080] Due to the influence of time delay, it is difficult to align the received signal with the target signal in terms of timing. Therefore, even if the spatial filter coefficients and frequency offset can be accurately obtained, there will still be a large bit error rate in the decision stage.

[0081] Although the UAV does not know specific values, but The dynamic change range is assumed to be determinable, and this range is recorded as [1, Based on this, the present invention further adjusts the input and output elements of the neural network model to effectively deal with Different delay situations.

[0082] When there is no delay, the input is a received signal with the same length as the preamble sequence ; But considering the delay Afterwards, originally with Corresponding , will move to Therefore, at the input of the neural network, the dimension of the received signal is expanded to At this time It must include the unknown delay The signal portion that can be strictly aligned with the preamble sequence in timing.

[0083] At the output, The coefficients of the original output single spatial filter Expanded to the spatial domain filter coefficient matrix, that is ; Similarly, The output frequency deviation compensation is changed from the original Expand to .

[0084] Next is the design of the loss function in the neural network:

[0085] Define the delay estimation matrix , Each d corresponds to a delay value assumed by the interference cancellation model, and the corresponding value is taken from the output of the neural network. and Acts on , we can get:

[0086] .

[0087] Among them, at time t It can be expressed as:

[0088]

[0089] The restored signal under the current delay assumption With the leading signal The mean square error between can be expressed as:

[0090]

[0091] Due to the existence There are delay situations, so neural network training needs to calculate indivual , these mean square errors are summed together as the final loss function of the neural network model:

[0092]

[0093] The corresponding optimization problem is:

[0094]

[0095] By expression It can be seen that the timing misalignment cannot be corrected by spatial filters and frequency offset compensation, which means that when When training, no matter what and , the recovered signal can never be aligned with the target signal. will remain at a high level.

[0096] On the contrary, when hour, and Successful alignment, optimization problem It degenerates into a problem , by continuously optimizing the corresponding and , can successfully realize the signal recovery, so Eventually it will drop to a very small value.

[0097] Based on the above discussion, when there is a time delay, the neural network minimizes the delay by continuously updating the parameters. , so that the composition of various ( ) can be minimized, but only one of them It can eventually approach 0 and find the smallest value after training. , and its corresponding label It can be considered equal to the delay value, and the estimated value is recorded as .

[0098] After the neural network training is completed, according to the estimated delay value Select the corresponding and , and use both together to The target signal can be restored .

[0099] Therefore, for the expression In the form of the received signal, the complete process of realizing signal time-frequency synchronization and interference elimination based on the neural network model in the present invention can be summarized as Algorithm 1:

[0100]

[0101] Simulation experiment

[0102] The entire interference cancellation simulation process combined MATLAB software and the PyTorch deep learning library in the Python environment. Specifically, the communication transmission system was built using MATLAB to generate all relevant experimental signals. The received signal and preamble sequence required for interference cancellation were then fed into the neural network model within the PyTorch framework. Steps 3 through 15 of Algorithm 1 were then executed to obtain the recovered signal. Finally, the signal was compared with the target signal to calculate the bit error rate.

[0103] Figure 1 Shows the initial frame sequence of the transmitted signal. The initial sequence of each frame includes bits, of which the first T bits are used to generate the preamble sequence, and the remaining V bits correspond to the length of the data sequence. After considering the effects of interference, noise, and delay, the corresponding received signal in the time domain is as follows Figure 2In the communication transmission system simulated by MATLAB, the number of receiver antennas is set to , the number of interference sources The interference target signal adopts QPSK modulation mode, and the signal processing flow in the communication system is as follows: Figure 3 and Figure 4 As shown in the figure, a root raised cosine finite impulse response filter is selected as the shaping filter and matched filter in the signal processing. The filter length includes 49 sampling points, the number of samples per symbol is 8, and the roll-off coefficient is set to 0.5. The frequency offset between the transmitter and the receiver is set to 760Hz, and an unknown delay is randomly generated. , the delay range corresponds to sampling points, the duration is .

[0104] The baseband communication signal is randomly generated by a 0 / 1 binary bit sequence. The symbol rate is set to 0.5Mb / s, the upsampling ratio is 8, and the receiving end sampling rate is 4MHz. Due to the symbol mapping and upsampling operations during the modulation process, each frame has a total of sampling points.

[0105] The interference signal is a randomly generated zero-mean Gaussian narrowband signal that is received together with the target signal at the receiving end. and interference channels and Gaussian white noise vector are all Gaussian random vectors, where SIR is defined as:

[0106]

[0107] The neural network in this invention adopts a feedforward neural network model called Multilayer Perceptron (MLP), in which The module contains three hidden layers, with the number of neurons in each layer being 640, 512, and 256, respectively; The module contains two hidden layers with 512 and 64 neurons respectively; Leaky ReLU is used between layers as a nonlinear activation function; the model is trained using the Adam (Adaptive Moment Estimation) optimizer, with an initial learning rate of , the total number of training epochs = 3000. It should be emphasized that, considering that the constancy of the signal channel and the interference channel is only applicable within the same frame period, when performing interference cancellation on signals in different frames, Algorithm 1 needs to be repeatedly executed to ensure that accurate interference suppression parameters are obtained.

[0108] In the present invention, the simulation experiment includes two parts:

[0109] Experiment 1 discusses the performance difference between the proposed method and the two-stage filtering method under different SIR and SNR conditions. , After symbol mapping and upsampling, each frame has 656 samples. The SIR range is [-40dB, -30dB, -20dB, -10dB, 0dB]; the SNR range is [10dB, 15dB, 20dB]. Since the preamble sequence is already known to the receiver, when discussing BER, only the data segment is used for BER calculation.

[0110] The bit error rates of the two methods are expressed by Figure 5 As shown in the figure, it can be seen that under all given SIR and SNR conditions, the BER of the proposed method is significantly lower than that of the two-stage filtering method, and the performance is stable. Even when the SIR is -40dB, the proposed method still has good anti-interference performance. It is worth noting that when the SIR is in the range of -30dB to -10dB, the BER of the two-stage filtering method increases significantly. This is attributed to the fact that the minimum eigenvector of the received signal covariance matrix can no longer effectively serve as an interference cancellation filter as the SIR increases.

[0111] Figure 6 Shows the loss function of the neural network when training the signal of a certain frame to eliminate interference ( ) changes, the random delay generated by the system is ; Figure 6 yes The change of this mean square error during training. It can be seen that After a period of decline, it remained at a high level (approximately ), except for can eventually approach 0, in other cases ( ) cannot successfully recover the signal because they cannot eliminate the time delay, so the mean square error is large. This result verifies that only when the time domain is successfully aligned, that is, When , the corresponding mean square error Only then will it have a chance to drop to a very small value, otherwise it will be difficult to drop further, so according to It is feasible to estimate the delay value based on the size of Figure 7 It also shows that the neural network has a faster convergence speed during training.

[0112] Experiment 2 mainly discusses the BER of the two methods under different preamble sequence lengths and SNR sizes. , , SIR = -30dB, SNR = [10dB, 15dB, 20dB].

[0113] The results of Experiment 2 are as follows Figure 8 As shown in Figure 2, the length of the preamble sequence directly determines the interference elimination effect of the proposed method. By increasing the length of the preamble sequence, the proposed method can achieve significant performance improvement. In addition, Figure 7 This once again verifies that the method proposed in the present invention is significantly better than the two-stage filtering method in terms of interference elimination.

[0114] In summary, the present invention proposes a multi-antenna UAV anti-interference communication method based on a neural network model. Through theoretical analysis, this method can achieve key functions such as time synchronization, frequency offset compensation, and interference suppression by relying solely on the preamble sequence and the received signal, without the need to obtain any channel state information or prior knowledge of the interference source. Experimental verification shows that this method is superior to existing two-stage filtering anti-interference methods and can still achieve a low bit error rate under strong interference conditions. In addition, the designed neural network has a simple structure, a small parameter scale, and a fast convergence speed, making it suitable for efficient deployment in actual systems.

Claims

1. A multi-antenna interference cancellation method based on a neural network model is used for downlink communication from a base station to an unmanned aerial vehicle (UAV). The environment contains K unknown interference sources. The base station and each interference source are equipped with only a single antenna, while the UAV is equipped with N antennas. ; The base station will periodically generate the preamble sequence , is the length of the leading sequence, As the target signal is transmitted from the base station to the UAV, The K interference source signals are denoted as , each interference source Independent of each other and both with the target signal Unrelated; the UAV signal reception model is defined as: ， in, To receive the signal, represents the channel vector between the base station and the UAV, represents the channel vector between the kth interference source and the UAV, represents the Gaussian white noise vector, Indicates the frequency offset between the base station and the UAV. The target signal is affected by the form of represents the unknown delay between the transmitter and the receiver; The interference elimination method comprises: Construct a neural network for interference cancellation, which includes a spatial domain filter and frequency offset compensation module ,in The input is the received signal matrix , the output is the spatial filter coefficient : ， ， in, is an underdetermined matrix or a full rank matrix; The input is also the received signal matrix , the output is the frequency offset compensation value ,when When , the frequency offset is completely eliminated, and the corresponding optimization problem with frequency offset and interference is expressed as the first optimization problem: ， Input the received signal matrix to the neural network model , the model updates the module synchronously according to the value of the loss function and The network layer parameters in , thereby solving the first optimization problem, so that the frequency offset compensation Approaching the true value At the same time, the corresponding spatial filter coefficients can be obtained To remove interference signals; Based on the obtained neural network, the unknown delay The value range is set to [1, ], is the maximum possible delay, according to The value range of adjusts the input and output elements of the neural network, specifically: At the input of the neural network, the received signal dimension is expanded to: ， At the output of the neural network, Output coefficients of a single spatial filter Expanded to the spatial domain filter coefficient matrix, that is ; Similarly, The output frequency offset compensation is from Expand to , define the delay estimation matrix: ， Where d is an assumed delay value, , take out the corresponding output matrix of the neural network and Acts on ,get: ， At time t Expressed as: ， The recovered signal under the assumed delay d With the leading signal The mean square error between them is expressed as: ， Due to the existence When training a neural network, it is necessary to calculate indivual , these mean square errors are summed together as the final loss function of the neural network model: ， The corresponding optimization problem is the second optimization problem: ， when hour, and After successful alignment, the second optimization problem degenerates into the first optimization problem. and , can successfully restore the signal; thus, when there is a time delay, the neural network minimizes it by continuously updating the parameters , so that the composition of various can be minimized, and finally find the one with the smallest value after training. , and its corresponding label That is, it is considered to be equal to the delay value, and the estimated value is recorded as ; After the neural network training is completed, according to the estimated delay value Select the corresponding and , and use both together to Get the recovery signal .

Images