Intelligent interference suppression method of aviation monitoring system based on generative adversarial network

Abstract

The invention discloses an intelligent interference suppression method for an aviation monitoring system based on a generative adversarial network, and relates to the technical field of aviation communication and electromagnetic interference intelligent processing, and the method comprises the steps: constructing a generative adversarial network interference suppression model; performing frequency domain preprocessing based on short-time Fourier transform (STFT) on the model; and carrying out generative adversarial network-based adversarial reconstruction on the model. According to the intelligent interference suppression method for the aviation monitoring system based on the generative adversarial network, frequency domain filtering and an attention mechanism are fused, the extraction and suppression capability of a model on interference characteristics in a complex electromagnetic environment is improved, an intelligent interference suppression platform suitable for the aviation monitoring system is constructed, and the interference suppression capability of the aviation monitoring system is improved by relying on the platform. The electromagnetic compatibility test and evaluation of the LDACS on the ADS-B system can be carried out, and the anti-interference performance and the communication reliability of the system are effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of aviation communication and electromagnetic interference intelligent processing, and in particular to an interference intelligent suppression method for an aviation surveillance system based on a generative adversarial network. Background Art

[0002] As global air transport continues to expand, aviation communication systems face unprecedented challenges. The rapid growth in air traffic has placed higher demands on communication systems' data transmission capabilities, spectrum scheduling efficiency, and anti-interference performance. However, the currently widely used aviation communication architecture has significant limitations in spectrum utilization, transmission speed, security, and interference mitigation, making it difficult to support the urgent need for intelligent and information-based modern air traffic management (ATM) systems.

[0003] To alleviate the increasingly severe shortage of spectrum resources and meet the future development trend of aviation data services, the International Telecommunications Regulatory Organization has allocated a portion of the L-band (960MHz to 1164MHz) for aeronautical mobile (route) service (AM(R)S). While this allocation improves spectrum utilization, it also places higher requirements on the electromagnetic compatibility of the new system, requiring that the operation of the new system must not cause harmful interference to existing aeronautical radio navigation systems within this frequency band.

[0004] Against this backdrop, the "Future Communications Infrastructure" strategy, jointly promoted by the European Organization for the Safety of Air Navigation (EUROCONTROL) and the U.S. Federal Aviation Administration (FAA), came into being. The project aims to integrate multiple data link technologies to build a unified and efficient aviation communications network architecture. Among them, the L-band Digital Aeronautical Communication System (LDACS) has been established as a key technical solution for air-to-ground data communications. Designed by the International Civil Aviation Organization (ICAO), the LDACS system features large bandwidth, high speed, low latency, and strong security, and can fully meet the core demands of modern flight missions for real-time, highly reliable communications.

[0005] However, the actual deployment of LDACS faces significant spectrum coexistence risks. Its operating frequency band overlaps or is in close proximity with several existing aviation communication, navigation, and surveillance (CNS) systems, including distance measuring equipment (DME), the global navigation satellite system (GNSS), the traffic collision avoidance system (TCAS), and the automatic dependent surveillance-broadcast system (ADS-B). The frequency difference between LDACS and ADS-B is only approximately 20 MHz, making out-of-band interference a particularly prominent issue. As the current mainstream air surveillance method, the ADS-B system undertakes critical tasks such as broadcasting aircraft position information and sharing navigation dynamics. Its communication reliability is directly related to the safety of air traffic control operations.

[0006] In this scenario, if the potential interference relationship between LDACS and ADS-B is not scientifically modeled and interference suppression is not performed, the safe deployment and promotion of LDACS will be seriously hindered. Therefore, it is urgent to develop a new interference suppression technology for the complex electromagnetic environment of the L-band to ensure stable communication of the LDACS system without compromising the normal operation of existing systems. In particular, in application scenarios with diverse interference types and complex spectral environments, a highly adaptable intelligent interference identification and suppression mechanism that does not require a priori interference models is required to effectively protect the communication integrity and data availability of critical systems such as ADS-B.

[0007] Existing interference suppression methods are mostly based on signal isolation technology in the time domain, frequency domain, code domain or spatial domain. Although they can reduce the impact of interference to a certain extent, they generally rely on prior information about the interference source, have insufficient generalization capabilities, and lack the flexibility to deal with time-varying interference. In addition, some studies have attempted to start from the perspective of physical layer optimization design, using redundancy design, adaptive filtering and other methods to perform active interference suppression. However, when faced with complex non-stationary interference or unknown interference types, there are still problems such as unstable performance and poor adaptability, making it difficult to meet the anti-interference requirements in actual aviation environments. Based on the above background, it is urgent to propose a new, highly robust and scalable intelligent interference suppression method to achieve high-fidelity recovery of ADS-B signals in LDACS interference environments and ensure the communication quality and operational safety of aviation surveillance systems.

[0008] In recent years, the development of artificial intelligence (AI) technology has provided new insights for interference suppression in complex environments. Deep learning models possess end-to-end feature learning capabilities, eliminating reliance on domain knowledge and prior interference models. They demonstrate strong robustness in dynamic RF interference environments. Interference identification and classification methods based on convolutional neural network architectures have achieved initial success, with some research demonstrating promising results in multi-label interference identification accuracy.

[0009] In order to improve the versatility and stability of the model in complex environments, existing studies have introduced an unsupervised learning framework to achieve feature decoupling of the interfered signal based on an autoencoder, thereby separating the interference and target signals. However, this method has certain limitations on the accuracy of feature reconstruction, and the effect is not ideal in data-scarce scenarios. In recent years, generative adversarial networks (GANs) have been introduced into interference suppression tasks as a game-based training mechanism. By constructing a two-stage structure of "adversarial generation-feature reconstruction", the model's adaptability to unknown interference environments and signal restoration capabilities have been effectively improved, becoming an important direction of intelligent interference suppression research. Therefore, the present invention provides an intelligent interference suppression method for aviation surveillance systems based on generative adversarial networks (GANs). Summary of the Invention

[0010] The purpose of the present invention is to provide an intelligent interference suppression method for an aviation surveillance system based on a generative adversarial network to solve the problems raised in the above background technology.

[0011] To achieve the above objectives, the present invention provides an intelligent interference suppression method for an aviation surveillance system based on a generative adversarial network, comprising the following steps:

[0012] S1. Construct a generative adversarial network interference suppression model under the interference of L-band digital aviation communication system LDACS;

[0013] S2, preprocessing the model in the frequency domain based on short-time Fourier transform (STFT);

[0014] S3. Perform adversarial reconstruction of the model based on the generative adversarial network, including the design of the convolutional block attention mechanism module, the discriminator and generator structure design, the adversarial loss function design, and the model training process design.

[0015] Preferably, the specific steps of S2 are:

[0016] S21. Perform short-time Fourier transform (STFT) on the discrete sampled baseband signal u(t) to obtain U(f,t), transforming the signal from the time domain to the time-frequency domain. The short-time Fourier transform (STFT) uses the fast Fourier transform (FFT) algorithm. The definition of the discrete STFT transform is:

[0017]

[0018] Where u[n] is the input discrete sampled baseband signal, h[n-mR] is the time domain window function, centered at time point mR, with a window length of N, R is the sliding step of the window function, usually 50% or 75% of the window length, k is the discrete frequency index, ranging from 0 ≤ k ≤ N-1, and m is the time frame index;

[0019] S22, set the spectrum mask M(f) of the signal frequency below 2MHz to 1 and above 2MHz to 0, and multiply the signal time-frequency domain data by the spectrum mask to perform digital frequency filtering to obtain the time-frequency domain data U of the filtered signal M (f,t);

[0020] S23, the time-frequency domain data U M (f, t) is obtained by discrete inverse transform ISTFT to obtain u M (t), when the window function and the overlap rate meet When , the inverse discrete transform ISTFT is defined as:

[0021]

[0022] Where A is the energy normalization factor of the window function, usually After inverse transformation, the original signal is restored by overlapping and adding each frame.

[0023] Preferably, the specific steps of S21 are:

[0024] S211, framing stage, dividing the discrete sampled baseband signal into short time frames of length N, with adjacent frames overlapping by L=NR points;

[0025] In the windowing stage, a window function h[n-mR] is applied to each frame of the signal to suppress spectrum leakage. The STFT uses a Hamming window, which is defined as follows:

[0026]

[0027] Where n is the sample point index and M is the window function length;

[0028] S213 , performing FFT calculation, performing Fast Fourier Transform (FFT) on the windowed frame to obtain a time-frequency matrix.

[0029] Preferably, the convolutional block attention mechanism module design in S3 includes the following specific steps:

[0030] S31. Construct a channel attention mechanism, which includes three stages: compression, excitation, and recalibration. In the compression stage, the input feature map is subjected to global average pooling (GAP) and global maximum pooling (GMP) in the spatial dimension to generate two tensors of shape (C, 1, 1). In the excitation stage, two fully connected layers are used to perform dimensionality reduction and dimensionality increase respectively. The ReLU activation function is embedded in the middle to improve the nonlinear expression ability. After the dimensionality increase, the Sigmoid function is introduced to generate the channel attention weight, which falls between 0 and 1. In the recalibration stage, the channel weight is multiplied by the input feature map channel by channel.

[0031] S32. Introduce the spatial attention mechanism, which includes three stages: compression, excitation, and recalibration. In the compression stage, global averaging and maximum pooling are performed on the channel dimension to obtain two feature maps of (1, H, W) shape. In the excitation stage, the two feature maps are superimposed and local neighborhood information is fused through a convolution operation with a convolution kernel size of 7×7. The spatial attention map is output through the Sigmoid function. In the recalibration stage, the spatial weight map is multiplied element-by-element with the channel-weighted feature map.

[0032] S33, the convolutional block attention mechanism module CBAM combines channel attention and spatial attention in a series manner, and embeds the series-connected convolutional block attention mechanism module CBAM into the middle layer of the generative adversarial network.

[0033] Preferably, the discriminator and generator structure design in S3 is as follows: the discriminator is composed of three one-dimensional convolutional layers, the convolution kernel size of each layer is set to 5×1, the number of convolution channels is 64, 128 and 1 respectively, each convolution operation is connected to the ReLU activation function, and a group of convolution block attention mechanism CBAM is inserted between the second convolution layer and the third convolution layer. The final output feature map is mapped to a normalized probability value through the fully connected layer, indicating the confidence that the input signal is a true automatic dependent surveillance broadcast system ADS-B signal;

[0034] The generator uses a four-layer one-dimensional convolutional network. The input is a mixed interference signal and a pure reference signal. The convolution kernel size is uniformly 5×1. The number of channels in each layer is 64, 128, 256 and 1, respectively. ReLU activation functions are inserted between each layer, and a CBAM attention module is embedded between the second and third convolutional layers. The final output is a generated signal, and the signal shape is consistent with the input pure ADS-B signal.

[0035] Preferably, the specific steps of designing the adversarial loss function in S3 are as follows:

[0036] JS divergence is used in the generative adversarial network. JS divergence is modified based on KL divergence and is defined as follows:

[0037]

[0038] Among them, p R (x) = p(D(x) = x i ) is the probability distribution of pure ADS-B signal data x, x i ∈X={x 1 ,x 2 ,…,x m}, P G (x) = p(G(x) = x i ) is the probability distribution of the generated signal data G(x), and the two probability distributions P R and P G The relative entropy is the KL divergence;

[0039] Construct the likelihood function:

[0040]

[0041] Among them, θ is an adjustable parameter, p G (x i ;θ)=p(D(x)|θ),x∈X={x 1 ,x 2 ,…,x m}, the range belongs to [0,1], which is the probability distribution of the generated signal data G(x);

[0042] The optimal parameter distribution θ * :

[0043]

[0044] We will find the θ that maximizes the likelihood function. * The problem is converted to finding the θ that minimizes the KL divergence * Problems;

[0045] Define the construction loss function V(G,D):

[0046]

[0047] Through the game between the generator G and the discriminator D, the optimal loss function V * (G,D) has:

[0048]

[0049] Optimal Discriminator D * (x) has:

[0050]

[0051] p G (x) = p R(x), the training standard V(G,D * ) to reach the global minimum.

[0052] Preferably, the model training process of S3 is designed to adopt progressive three-stage adversarial training, including discriminator pre-training, adversarial game training and refined training. The adversarial game training includes a standard training stage and a cyclic sample training stage. The standard training stage adopts a traditional adversarial training mode, and the generator and the discriminator are optimized alternately to perform cyclic adversarial training. The cyclic sample training stage performs cyclic training by inputting random different pure signal data into the discriminator; the refined training discriminator inputs the initial pure signal data again, performs deep optimization, and performs cyclic training.

[0053] Preferably, the specific steps of parameter training of the discriminator are as follows:

[0054] Given a generator G unchanged, given the optimal loss function format, the discriminator loss function is:

[0055]

[0056] The Adam optimization algorithm is used to update the parameters of the generator G. The parameter optimization process of gradient descent is as follows:

[0057]

[0058] The specific steps of parameter training of the generator are as follows:

[0059] Given a constant discriminator D, define the generator loss function:

[0060]

[0061] The Adam optimization algorithm is used to update the parameters of the generator G. The parameter optimization process of gradient descent is as follows:

[0062]

[0063] Preferably, the cyclic adversarial training process comprises the following specific steps:

[0064] Given G0, minimize L D (D) to obtain Right now

[0065] fixed calculate In order to obtain the updated G1;

[0066] Fix G1 and maximize L G (G) To obtain Right now

[0067] fixed calculate In order to obtain the updated G2.

[0068] Preferably, the aviation surveillance system includes: an L-band digital aviation communication system LDACS and an automatic dependent surveillance broadcast system ADS-B electromagnetic compatibility data acquisition experimental platform, a convolutional block attention mechanism module CBAM, a discriminator and a generator. The LDACS and automatic phase ADS-B electromagnetic compatibility data acquisition experimental platform are used for end-to-end denoising of ADS-B signals in a complex electromagnetic environment. The convolutional block attention mechanism module CBAM integrates channel attention and spatial attention in series and is inserted in the middle layer between the generator and the discriminator.

[0069] Therefore, the present invention adopts the above-mentioned intelligent interference suppression method for aviation surveillance systems based on generative adversarial networks, which has the following beneficial effects:

[0070] 1. The present invention provides an intelligent and scalable technical path for suppressing interference of the L-band digital aviation communication system LDACS against the Automatic Dependent Surveillance-Broadcast System ADS-B. This technology breaks through the limitations of traditional interference suppression methods based on prior knowledge and can adapt to the complex and changing aviation electromagnetic environment.

[0071] 2. This paper proposes a phased progressive training strategy, which significantly improves the stability of the generative model through steps such as discriminator pre-training, adversarial training, and fine-tuning, and reduces training problems such as mode collapse and gradient disappearance.

[0072] 3. The present invention constructs a dual-stage interference suppression architecture of "frequency domain filtering + adversarial generation" to improve interference suppression performance.

[0073] 4. The present invention introduces the convolutional block attention mechanism to enhance the recognition ability of the model's key features, effectively improves the model's adaptability and convergence speed in the face of different interference scenarios, and enhances the robustness of the algorithm.

[0074] 5. The model proposed in this invention has strong portability.

[0075] 6. This invention provides theoretical reference and algorithmic support for the electromagnetic compatibility research and standard formulation of future aviation communication systems, and has a positive role in promoting the deployment and application of LDACS systems on a global scale.

[0076] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1This is a diagram of the serial structure of the convolutional block attention mechanism module CBAM of an embodiment of the interference intelligent suppression method for an aviation surveillance system based on a generative adversarial network of the present invention;

[0078] Figure 2 This is a framework diagram of a spatial attention mechanism module of an embodiment of an intelligent interference suppression method for an aviation surveillance system based on a generative adversarial network according to the present invention;

[0079] Figure 3 This is a discriminator structure diagram of an embodiment of an intelligent interference suppression method for an aviation surveillance system based on a generative adversarial network according to the present invention;

[0080] Figure 4 This is a structural diagram of a generator of an embodiment of an intelligent interference suppression method for an aviation surveillance system based on a generative adversarial network according to the present invention;

[0081] Figure 5 This is a schematic diagram of the LDACS and ADS-B electromagnetic compatibility data acquisition experimental platform for an embodiment of the intelligent interference suppression method for aviation surveillance systems based on a generative adversarial network of the present invention;

[0082] Figure 6 This is a flow chart of a generative adversarial network method according to an embodiment of the invention's intelligent interference suppression method for an aviation surveillance system based on a generative adversarial network. DETAILED DESCRIPTION

[0083] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0084] Unless otherwise defined, the technical or scientific terms used in the present invention shall have the usual meanings understood by persons of ordinary skill in the field to which the present invention belongs. The words "first", "second" and similar terms used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Words such as "include" or "comprise" mean that the elements or objects preceding the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Words such as "connect" or "connected" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0085] Example

[0086] See also Figure 1-6 The present invention provides an intelligent interference suppression method for an aviation surveillance system based on a generative adversarial network, comprising the following steps:

[0087] S1. A generative adversarial network interference suppression model constructed under the interference of the L-band digital aviation communication system LDACS is implemented through a two-stage processing flow combining frequency domain preprocessing and adversarial reconstruction. This model achieves high-fidelity reconstruction of the Automatic Dependent Surveillance-Broadcast System (ADS-B) signal under the interference background of the L-band digital aviation communication system LDACS.

[0088] S2. Perform frequency domain preprocessing on the model based on short-time Fourier transform (STFT) to effectively filter out high-frequency mutation noise and LDACS out-of-band interference in non-stationary ADS-B signals. This serves as a preprocessing module before GAN input and defines the specific window function selection and mask construction method. The specific algorithm for frequency domain digital filtering based on short-time Fourier transform is as follows:

[0089] For stationary discrete signals whose statistical properties, such as mean, variance, and covariance, are independent of time, the Fast Fourier Transform (FFT) is often used as a key method for time-domain analysis and processing. However, because the statistical properties of discrete ADS-B and LDACS baseband sampling signals vary over time, both discrete baseband signals are non-stationary. The Fast Fourier Transform (FFT) only captures spectral information over the entire signal time period, losing information on spectral variations within local time periods. Consequently, it cannot accurately characterize the characteristics of non-stationary signals.

[0090] Because the ADS-B signals to be processed often contain short, high-frequency mutations of varying frequencies in the time domain, when short, high-frequency mutations are added to the signal and then subjected to a Fast Fourier Transform (FFT), it can be observed that the FFT interprets the mutations as the superposition of a series of low-component, high-frequency signals, failing to adequately reflect the changes brought about by the high-frequency mutation disturbances. This can, in turn, disperse the high-frequency mutation interference within the ADS-B baseband signal frequency range, weakening the interference suppression effect of the digital filter.

[0091] To solve the time-varying problem of signal statistical characteristics, time-frequency analysis technology is used instead of FFT. The FFT algorithm is used in the short-time Fourier transform (STFT), which has a computational complexity of O(N log2 N) and high computational efficiency. Its discrete STFT transform is defined as:

[0092]

[0093] Where u[n] is the input discrete sampled baseband signal; h[n-mR] is the time domain window function, centered at time point mR, with a window length of N; R is the sliding step size (frame shift) of the window function, usually 50% or 75% of the window length; k is the discrete frequency index, ranging from 0≤k≤N-1; and m is the time frame index.

[0094] The STFT transform calculation can be divided into three stages: framing, windowing, and FFT calculation. First, the framing stage divides the signal into short time frames of length N, with L = NR points overlapping between adjacent frames. Next, the windowing stage applies a window function h[n - mR] to each frame to suppress spectral leakage. Finally, the FFT calculation is performed, performing a Fast Fourier Transform (FFT) on the windowed frames to obtain a time-frequency matrix.

[0095] When the perfect reconstruction condition (COLA) is met, that is, the window function and the overlap rate meet:

[0096]

[0097] Its inverse discrete transform ISTFT is defined as:

[0098]

[0099] Where A is the energy normalization factor of the window function, usually After inverse transformation, the original signal is restored by overlapping and adding each frame.

[0100] Since the Hanning window and Hamming window can balance leakage suppression and resolution, they are suitable for quasi-stationary signals with little change in overall characteristics, such as the baseband signal in this scenario. However, the main lobe width of the Hamming window is narrower than that of the Hanning window, but the side lobe attenuation is greater. Therefore, the STFT adopts the Hamming window, which is defined as follows:

[0101]

[0102] Where n is the sample point index and M is the window function length;

[0103] If a short high-frequency mutation is added to the signal and STFT transformation is performed, the time-frequency components of the high-frequency mutation can be well identified, which is conducive to the digital frequency domain filter to filter out high-frequency interference.

[0104] The basic implementation steps of the frequency domain filter based on STFT are as follows: First, perform STFT transform on the discrete sampled baseband signal u(t) to obtain U(f,t), and transform the signal from the time domain to the time-frequency domain. Then, since the maximum frequency of the ADS-B baseband signal is 2MHz, a spectrum mask M(f) can be set with 1 below 2MHz and 0 above 2MHz, and the signal time-frequency domain data is multiplied by the spectrum mask to obtain the time-frequency domain data U of the signal after digital frequency filtering. M (f,t). Finally, U M (f, t) is transformed into u after ISTFT M (t).

[0105] S3. Perform adversarial reconstruction of the model based on the generative adversarial network, including the design of the convolutional block attention mechanism module, the discriminator and generator structure design, the adversarial loss function design, and the model training process design.

[0106] The convolutional block attention mechanism module design in S3 is as follows:

[0107] The present invention designs a convolutional block attention mechanism module (CBAM) suitable for signal interference suppression tasks. By connecting channel attention and spatial attention mechanisms in series, the module strengthens the intermediate feature expression capability through the channel-spatial dual-dimensional attention module, enhances the network's perception and modeling capabilities of key signal features, and improves the accuracy and stability of signal restoration.

[0108] S31. This module first constructs a channel attention mechanism to determine the importance of different channels in the feature map. Specifically, it includes three stages: compression, excitation, and recalibration. In the compression stage, the input feature map is subjected to global average pooling (GAP) and global maximum pooling (GMP) in the spatial dimension to generate two tensors of shape (C, 1, 1) for extracting overall and local saliency information. In the excitation stage, two fully connected layers are used to perform dimensionality reduction and dimensionality increase operations respectively. The ReLU activation function is embedded in the middle to improve the nonlinear expression ability, and the Sigmoid function is introduced after the dimensionality increase to generate the channel attention weight. The weight falls between 0 and 1 and has good trainability and physical interpretability. In the recalibration stage, the channel weight is multiplied by the input feature map channel by channel to achieve the enhancement of key channels and the suppression of non-key channels.

[0109] S32. Based on the channel attention mechanism, a spatial attention mechanism is further introduced to focus on the significant position areas in the feature map. Its processing flow also includes three stages: compression, excitation, and recalibration. In the compression stage, global averaging and maximum pooling are performed on the channel dimension to obtain two feature maps in the shape of (1, H, W). The two are then superimposed and the local neighborhood information is fused through a convolution operation with a convolution kernel size of 7×7. Finally, the spatial attention map is output through the Sigmoid function. In the recalibration stage, the spatial weight map is multiplied element by element with the channel-weighted feature map to further highlight the key spatial areas and suppress background interference.

[0110] S33, the convolutional block attention mechanism module (CBAM) combines channel attention and spatial attention in series, resulting in lower computational overhead and parameter count compared to the complex dual attention network (DANet), making it suitable for aviation signal processing scenarios requiring high inference efficiency. In this invention, this module is embedded in the middle layer of the generative adversarial network, significantly enhancing the generator and discriminator's ability to extract important time-frequency domain features while ensuring real-time performance, improving the model's overall anti-interference performance.

[0111] The CBAM attention mechanism module designed in this paper has the characteristics of lightweight structure, efficient implementation, and strong interpretability. It is suitable for non-stationary signal interference suppression, feature extraction and reconstruction tasks, and is particularly suitable for interference recovery scenarios of aviation communication signals such as ADS-B and LDACS. It has significant engineering practical value and technology promotion potential.

[0112] The discriminator and generator structure design in S3 is as follows:

[0113] The present invention provides an aviation signal interference suppression structure based on a generative adversarial network (GAN). The core of the structure consists of two parts: a generator and a discriminator, which are used for denoising and reconstructing interference signals and distinguishing true from false signals, respectively.

[0114] The discriminator structure is designed to effectively distinguish between pure ADS-B signals and the reconstructed signals output by the generator. The discriminator consists of three layers of one-dimensional convolutional layers. The convolution kernel size of each layer is set to 5×1, and the number of convolution channels is 64, 128, and 1, respectively, to extract the time-frequency domain features of the interference signal layer by layer. Each convolution operation is followed by a ReLU activation function to enhance the nonlinear fitting ability of the network. A set of convolutional block attention mechanisms (CBAMs) is inserted between the second and third convolutional layers. This module combines channel attention and spatial attention weights to perform weighted enhancement on key dimensions in the feature map, effectively improving the recognition ability of weak signal features. The final output feature map is mapped to a normalized probability value through a fully connected layer, indicating the confidence that the input signal is a true ADS-B signal, which is used for adversarial optimization in the training phase.

[0115] The generator architecture is designed to construct a highly reliable target signal output to confuse the discriminator, enabling adaptive reconstruction of ADS-B signals in jammed environments. The generator utilizes a four-layer, one-dimensional convolutional network. Its input consists of a mixed jamming signal and a clean reference signal. The convolution kernel size is uniformly 5×1, and the number of channels in each layer is 64, 128, 256, and 1, respectively. This is used to gradually restore the time-frequency characteristics of the clean signal. Reinforced Luminance (ReLU) activation functions are inserted between layers to maintain the nonlinear transfer of feature information and enhance the model's expressiveness. A CBAM attention module is embedded between the second and third convolutional layers, enabling the model to focus on key information areas relevant to signal restoration, improving the fidelity and structural integrity of the generated results. The resulting generated signal, whose shape matches the clean ADS-B input, is used for error calculation and adversarial training with the original signal.

[0116] In summary, the generator and discriminator constructed in the present invention are structurally embedded with attention mechanism modules, taking into account both model efficiency and signal feature expression capabilities. They are suitable for processing the automatic restoration and interference identification of ADS-B signals interfered by LDACS, have good time-frequency feature extraction capabilities, robustness and trainability, and are suitable for deployment in aviation communication systems in complex electromagnetic environments.

[0117] The adversarial loss function in S3 is designed to optimize the loss function under the complex channel conditions of the aviation surveillance system. A composite loss function system combining adversarial loss and reconstruction loss is constructed. By integrating KL divergence, JS divergence, and maximum likelihood criterion, the generated signal is guaranteed to achieve an optimal balance between discrimination accuracy and restoration quality, so that the GAN output signal can approach the real ADS-B signal in both the time and frequency domains. The specific steps are as follows:

[0118] In information theory, the amount of information I(x) is used to characterize the amount of information required to eliminate the uncertainty of the random variable X at x, and is defined as:

[0119] I(x)=-log2 p(x);

[0120] Where p(x) is the probability distribution of the random variable X. The information entropy H(X) is the mathematical expectation of the amount of information, defined as:

[0121] H(X)=E[I(x)]=-∑p(x)·log2 p(x);

[0122] In a generative adversarial network, the generator outputs a distribution D(x): R n →[0,1], discriminator output distribution G(x): R n →R n , P R (x) = p(D(x) = x i) is the probability distribution of pure ADS-B signal data x, x i ∈X={x 1 ,x 2 ,…,x m}, P G (x) = p(G(x) = x i ) is the probability distribution of the generated signal data G(x). The purpose of the confrontation is to make P R (x) fully fit P G (x). And the cross entropy H(P R ,P G ) is used to measure P G In fitting P R In the process of eliminating uncertainty, the amount of information used is as follows:

[0123] H(P R ,P G )=-∑p R (x)·log2 p G (x);

[0124] And the two probability distributions P R and P G The relative entropy, also known as KL divergence, is used to characterize the probability distribution P G Fitting probability distribution P R The degree of G Fitting probability distribution P R In the process, if P G Can fully fit P R , naturally there is H(P R )=H(P R ,P G ), if P G The fitting P is insufficient, and information loss H(P R )-H(P R ,P G ), that is, P and P G The KL divergence between is defined as follows:

[0125]

[0126] In addition, the continuous KL divergence is defined as follows:

[0127]

[0128] But unlike many distance metrics, KL divergence is asymmetric, that is:

[0129] KL(p R (x)||p G (x))≠KL(p G(x)||p R (x));

[0130] The asymmetry is particularly evident at the "zero probability event" of the distribution, as shown by KL(p R (x)||p G (x)) as an example:

[0131] When p R (x)→0, that is, when the pure signal is mistakenly judged as the generated signal, as long as p G (x)≠0, even if the generated signal deviates greatly from the pure signal, p R (x)log2(p R (x) / p G (x))→0, for KL(p R (x)||p G The contribution of (x)) tends to 0, and the penalty in this case is extremely small, which will cause unexpected errors.

[0132] When p G (x)→0, that is, when the pure signal is mistakenly judged as the generated signal

[0133] This causes the generator to tend to generate repeated signals that are close to pure signals, resulting in insufficient diversity of the generated signals.

[0134] At the same time, when p G (x)→0 and p R (x)→1, p R (x)log2(p R (x) / p G (x))→∞, for KL(p R (x)||p G The contribution of (x)) tends to positive infinity, causing the gradient to explode or disappear.

[0135] Therefore, the asymmetric KL divergence is not used in the generative adversarial network, but the symmetric JS divergence is used. The JS divergence is modified based on the KL divergence and is defined as follows:

[0136]

[0137] Then the symmetry is satisfied:

[0138] JS(p R (x)||p G (x))=JS(p R (x)||p R (x));

[0139] For the generative adversarial network, the probability function p of the generator G(x) is controlled by the adjustable parameter θ. The probability distribution of the generated signal data G(x) can be expressed as p G (x i ;θ)=p(D(x)|θ), x∈X={x 1 ,x 2 ,…,x m}, the range belongs to [0,1]. Therefore, we can construct the likelihood function:

[0140]

[0141] If the generated signal distribution G(x) is close to the pure ADS-B signal distribution D(x), then p G (x i ;θ)=p(D(x)|θ) should be equal to P R (x) = p(x), the likelihood function L takes the maximum value, and the optimal parameter distribution θ when generating the signal distribution closest to the pure ADS-B signal distribution can be obtained by maximizing the likelihood function L. * :

[0142] The logarithmic function is monotonically increasing. To simplify the cumulative multiplication calculation into cumulative calculation, maximizing the likelihood function is equivalent to maximizing its logarithm, that is:

[0143]

[0144] Then we have:

[0145]

[0146] in Indicates that x follows the pure signal distribution P R (x), and the normal number It is independent of θ and can be ignored in the maximization process, that is:

[0147]

[0148] Because -∫ x∈X p R (x)log2 p R (x)dx has nothing to do with θ, so adding it to the likelihood function does not affect θ * The value of , that is:

[0149]

[0150] Then we have:

[0151]

[0152] Then find the θ that maximizes the likelihood function* The problem can be converted into finding the θ that minimizes the KL divergence * problem.

[0153] Define the loss function V(G,D), and then prove that when V(G,D) is maximized, the KL divergence is minimized:

[0154]

[0155] In the case of signal interference suppression, when the input signal x satisfies the pure signal distribution P R (x), D(x) represents the probability that the discriminator correctly judges it as a pure signal. When the input signal x satisfies the interference signal distribution P Z (x), 1-D(G(x)) represents the probability that the discriminator correctly judges it as an interfered signal.

[0156] The purpose of training the discriminator is to maximize V(G,D), the optimal discriminator D * have:

[0157]

[0158] The purpose of the generator D is to deceive the discriminator G to the greatest extent. The purpose of training the generator is to minimize V(G,D). The optimal generator G * have:

[0159]

[0160] Then through the game between the generator G and the discriminator D, the optimal loss function V * (G,D) has:

[0161]

[0162] Since the optimal loss function satisfies when x~P Z (x) when G(x)~P G (x), then:

[0163]

[0164] Then we have:

[0165] V(G,D)=∫ x∈X p R (x)log2(D(x))+p G (x)log2(1-D(x))dx;

[0166]

[0167] If the discriminator function D(x) satisfies y∈[0,1], the integrand f(y) can be written as:

[0168] f(y)=alog2(y)+b log2(1-y);

[0169] where a = p R (x), b=p G (x).

[0170] To find the maximum value of the integrand f(y), when a+b≠0, let f'(y)=0:

[0171]

[0172] The solution is:

[0173]

[0174] The second-order derivative of the integrand f(y) at the stationary point is:

[0175]

[0176] Then the second-order derivative of the stationary point is less than zero, which means is the maximum point.

[0177] Table 1 Monotonicity and extreme value analysis of integrand

[0178]

[0179] As shown in the above table, is the only maximum value and is greater than the boundary value, then is the only maximum value.

[0180] And because a=p R (x), b=p G (x), then the optimal discriminator D * (x) has:

[0181]

[0182] The following proves that if and only if p G (x) = p R (x), the training standard V(G,D * ) to reach the global minimum.

[0183] The goal of generative adversarial network training is to generate signal distribution p G (x) is equal to the pure signal distribution p R (x), that is, p G (x) = p G (x), then the optimal discriminator function D of the target * (x) has:

[0184]

[0185] This means that when the generator is trained, the discriminator is completely unable to distinguish between p G (x) and p R (x) The difference between the two distributions, so the probability that the output signal is pure is always 0.5.

[0186] When the optimal discriminator D * When the loss function V(G,D * )have:

[0187]

[0188] Next, find p G (x)≠p R (x) when the global minimum value is obtained, the optimal discriminator D * (x) is substituted into

[0189]

[0190] have:

[0191]

[0192] To prove that the global minimum is -2, we hope to * ), we can add or subtract 1 from each integral and multiply by the probability density, which is essentially just adding 0, specifically:

[0193]

[0194] Simplified:

[0195]

[0196] According to the definition of probability density, p R (x) and p G (x) is integrated over its integral domain and is equal to 1, that is:

[0197] -∫ x∈X p R (x)+p G (x)dx=-2;

[0198] Furthermore, by the definition of logarithm, we have:

[0199]

[0200] From the formula of continuous KL divergence in the previous article, we can find that each integral corresponds to KL divergence, then:

[0201]

[0202] Since KL divergence is non-negative, we can conclude that V(G,D * ) has a global minimum of -2.

[0203] According to the definition of JS divergence in the previous article, we can find that V(G,D * ) can form the JS divergence, namely:

[0204] V(G,D * )=-2+2·JS(p R (x)||p G (x));

[0205] Due to the symmetry of JS divergence, we know that if and only if p R (x) = p G (x), the JS divergence is 0, which also proves that if and only if p R (x) = p G (x), V(G,D * ) has a global minimum of -2.

[0206] Since the loss function V(G,D * ) contains JS divergence. According to the previous conclusion, when V(G,D * ) takes the minimum value, then the JS divergence takes the minimum, and then the KL divergence takes the minimum, then the maximum likelihood function takes the minimum, then the pure signal distribution p R (x) and the generated signal distribution p G (x) is closest. Therefore, the training goal of the model can be set as modifying the model to pursue the global minimum of the loss function.

[0207] The S3 model training process is designed to adopt a progressive three-stage adversarial training, including discriminator pre-training, adversarial game training and refined training. It effectively alleviates problems such as gradient vanishing and mode collapse in traditional GAN ​​training, and significantly improves the stability and generation effect of model training.

[0208] For the parameter training process, when training the discriminator, a generator G is given and remains unchanged. According to the optimal loss function format mentioned above, the discriminator loss function is given:

[0209]

[0210] The Adam optimization algorithm is used to update the parameters of the generator G. The parameter optimization process of gradient descent is as follows:

[0211]

[0212] When training the generator, given a constant discriminator D, since the generator only needs to deceive the discriminator, the generator loss function is defined as:

[0213]

[0214] The Adam optimization algorithm is used to update the parameters of the generator G. The parameter optimization process of gradient descent is as follows:

[0215]

[0216] The specific detailed cyclic adversarial training process is as follows:

[0217] Given G0, minimize L D (D) to obtain Right now

[0218] fixed calculate In order to obtain the updated G1;

[0219] Fix G1 and maximize L G (G) To obtain Right now

[0220] fixed calculate In order to obtain the updated G2;

[0221] And so on...

[0222] The actual training network framework is as follows:

[0223] The generative adversarial network model training process adopts a four-stage progressive strategy. In the pre-training phase, the discriminator is trained independently for 100 cycles to establish preliminary signal discrimination capabilities, laying a solid foundation for subsequent adversarial training. In the standard training phase, the generator and discriminator are alternately optimized using a traditional adversarial training model, with 1000 cycles. In the cyclic sample training phase, the discriminator is fed with randomized, different clean signal data each round, enhancing the model's generalization capabilities. In the refined training phase, the discriminator is fed with the initial clean signal data again for 250 cycles for further optimization. This phase helps the model converge to a more optimal local minimum.

[0224] Regarding optimizer selection, the algorithm uses the Adam optimizer, which addresses core issues in GAN training, such as gradient sparsity, oscillation, and initialization sensitivity, through adaptive learning rates, momentum mechanisms, and bias correction. Its robustness and ease of use make it the preferred optimizer for GAN implementations.

[0225] The aviation surveillance system includes: an L-band digital aviation communication system LDACS and automatic dependent surveillance broadcast system ADS-B electromagnetic compatibility data acquisition experimental platform, a convolutional block attention mechanism module CBAM, a discriminator and a generator. The LDACS and automatic phase ADS-B electromagnetic compatibility data acquisition experimental platform is used for end-to-end denoising of ADS-B signals in complex electromagnetic environments. The convolutional block attention mechanism module CBAM integrates channel attention and spatial attention in series and is inserted in the middle layer between the generator and the discriminator.

[0226] Therefore, the present invention adopts the above-mentioned intelligent interference suppression method for aviation surveillance systems based on generative adversarial networks, and proposes an end-to-end denoising scheme for ADS-B signals in complex electromagnetic environments based on the generative adversarial network architecture. This scheme does not rely on specific prior parameters, has the ability to perform generalized reconstruction under various unknown interference backgrounds, and is suitable for filtering, restoration and quality assessment of ADS-B signals in aviation surveillance systems, including signal quality assessment through time-frequency analysis and PPM demodulation, and using bit error rate BER to analyze the signal filtering effect.

[0227] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An intelligent interference suppression method for aviation surveillance systems based on generative adversarial networks, characterized in that: The following steps are involved: S1. Construct a generative adversarial network interference suppression model under the interference of L-band digital aviation communication system LDACS; S2, preprocessing the model in the frequency domain based on short-time Fourier transform (STFT); S3. Perform adversarial reconstruction of the model based on the generative adversarial network, including the design of the convolutional block attention mechanism module, the discriminator and generator structure design, the adversarial loss function design, and the model training process design.

2. The intelligent interference suppression method for aviation surveillance system based on generative adversarial network according to claim 1 is characterized in that: The specific steps of S2 are as follows: S21. Perform short-time Fourier transform (STFT) on the discrete sampled baseband signal u(t) to obtain U(f,t), transforming the signal from the time domain to the time-frequency domain. The short-time Fourier transform (STFT) uses the fast Fourier transform (FFT) algorithm. The definition of the discrete STFT transform is: Where u[n] is the input discrete sampled baseband signal, h[n-mR] is the time domain window function, centered at time point mR, with a window length of N, R is the sliding step of the window function, usually 50% or 75% of the window length, k is the discrete frequency index, ranging from 0 ≤ k ≤ N-1, and m is the time frame index; S22, set the spectrum mask M(f) of the signal frequency below 2MHz to 1 and above 2MHz to 0, and multiply the signal time-frequency domain data by the spectrum mask to perform digital frequency filtering to obtain the time-frequency domain data U of the filtered signal M (f,t); S23, the time-frequency domain data U M (f, t) is obtained by discrete inverse transform ISTFT to obtain u M (t), when the window function and the overlap rate meet When , the inverse discrete transform ISTFT is defined as: Where A is the energy normalization factor of the window function, usually After inverse transformation, the original signal is restored by overlapping and adding each frame.

3. The intelligent interference suppression method for aviation surveillance systems based on generative adversarial networks according to claim 2 is characterized in that: The specific steps of S21 are as follows: S211, framing stage, dividing the discrete sampled baseband signal into short time frames of length N, with adjacent frames overlapping by L=NR points; In the windowing stage, a window function h[n-mR] is applied to each frame of the signal to suppress spectrum leakage. The STFT uses a Hamming window, which is defined as follows: Where n is the sample point index and M is the window function length; S213 , performing FFT calculation, performing Fast Fourier Transform (FFT) on the windowed frame to obtain a time-frequency matrix.

4. The intelligent interference suppression method for aviation surveillance systems based on generative adversarial networks according to claim 3 is characterized in that: The convolutional block attention mechanism module design in S3 is as follows: S31. Construct a channel attention mechanism, which includes three stages: compression, excitation, and recalibration. In the compression stage, the input feature map is subjected to global average pooling (GAP) and global maximum pooling (GMP) in the spatial dimension to generate two tensors of shape (C, 1, 1). In the excitation stage, two fully connected layers are used to perform dimensionality reduction and dimensionality increase respectively. The ReLU activation function is embedded in the middle to improve the nonlinear expression ability. After the dimensionality increase, the Sigmoid function is introduced to generate the channel attention weight, which falls between 0 and 1. In the recalibration stage, the channel weight is multiplied by the input feature map channel by channel. S32. Introduce the spatial attention mechanism, which includes three stages: compression, excitation, and recalibration. In the compression stage, global averaging and maximum pooling are performed on the channel dimension to obtain two feature maps of (1, H, W) shape. In the excitation stage, the two feature maps are superimposed and local neighborhood information is fused through a convolution operation with a convolution kernel size of 7×7. The spatial attention map is output through the Sigmoid function. In the recalibration stage, the spatial weight map is multiplied element-by-element with the channel-weighted feature map. S33, the convolutional block attention mechanism module CBAM combines channel attention and spatial attention in a series manner, and embeds the series-connected convolutional block attention mechanism module CBAM into the middle layer of the generative adversarial network.

5. The intelligent interference suppression method for aviation surveillance system based on generative adversarial network according to claim 4 is characterized in that: The discriminator and generator structure design in S3 is as follows: the discriminator consists of three layers of one-dimensional convolutional layers, the convolution kernel size of each layer is set to 5×1, the number of convolution channels is 64, 128 and 1 respectively, each convolution operation is connected to the ReLU activation function, and a set of convolution block attention mechanism CBAM is inserted between the second convolution layer and the third convolution layer. The final output feature map is mapped to a normalized probability value through the fully connected layer, indicating the confidence that the input signal is a true automatic dependent surveillance broadcast system ADS-B signal; The generator uses a four-layer one-dimensional convolutional network. The input is a mixed interference signal and a pure reference signal. The convolution kernel size is uniformly 5×1. The number of channels in each layer is 64, 128, 256 and 1, respectively. ReLU activation functions are inserted between each layer, and a CBAM attention module is embedded between the second and third convolutional layers. The final output is a generated signal, and the signal shape is consistent with the input pure ADS-B signal.

6. The intelligent interference suppression method for aviation surveillance systems based on generative adversarial networks according to claim 5 is characterized in that: The specific steps for designing the adversarial loss function in S3 are as follows: JS divergence is used in the generative adversarial network. JS divergence is modified based on KL divergence and is defined as follows: Among them, p R (x) = p(D(x) = x i ) is the probability distribution of pure ADS-B signal data x, x i ∈X={x 1 ,x 2 ,…,x m }, P G (x) = p(G(x) = x i ) is the probability distribution of the generated signal data G(x), and the two probability distributions P R and P G The relative entropy is the KL divergence; Construct the likelihood function: Among them, θ is an adjustable parameter, p G (x i ;θ)=p(D(x)|θ), x∈X={x 1 ,x 2 ,…,x m }, the range belongs to [0,1], which is the probability distribution of the generated signal data G(x); The optimal parameter distribution θ * : We will find the θ that maximizes the likelihood function. * The problem is converted to finding the θ that minimizes the KL divergence * Problems; Define the construction loss function V(G,D): Through the game between the generator G and the discriminator D, the optimal loss function V * (G,D) has: Optimal Discriminator D * (x) has: p G (x) = p R (x), the training standard V(G,D * ) to reach the global minimum.

7. The intelligent interference suppression method for an aviation surveillance system based on a generative adversarial network according to claim 6, characterized in that: The S3 model training process is designed to adopt a progressive three-stage adversarial training, including discriminator pre-training, adversarial game training and refined training. The adversarial game training includes a standard training stage and a cyclic sample training stage. The standard training stage adopts the traditional adversarial training mode, and the generator and the discriminator are optimized alternately to perform cyclic adversarial training. The cyclic sample training stage performs cyclic training by inputting random different pure signal data into the discriminator; the refined training discriminator inputs the initial pure signal data again, performs deep optimization, and performs cyclic training.

8. The intelligent interference suppression method for aviation surveillance system based on generative adversarial network according to claim 7, characterized in that: The specific steps of parameter training of the discriminator are as follows: Given a generator G unchanged, given the optimal loss function format, the discriminator loss function is: The Adam optimization algorithm is used to update the parameters of the generator G. The parameter optimization process of gradient descent is as follows: The specific steps of parameter training of the generator are as follows: Given a constant discriminator D, define the generator loss function: The Adam optimization algorithm is used to update the parameters of the generator G. The parameter optimization process of gradient descent is as follows:

9. The intelligent interference suppression method for aviation surveillance system based on generative adversarial network according to claim 8, characterized in that: The specific steps of the cyclic adversarial training process are as follows: Given G0, minimize L D (D) to obtain Right now fixed calculate In order to obtain the updated G1; Fix G1 and maximize L G (G) To obtain Right now fixed calculate In order to obtain the updated G2.

10. The intelligent interference suppression method for aviation surveillance system based on generative adversarial network according to claim 9, characterized in that: The aviation surveillance system includes: an L-band digital aviation communication system LDACS and automatic dependent surveillance broadcast system ADS-B electromagnetic compatibility data acquisition experimental platform, a convolutional block attention mechanism module CBAM, a discriminator and a generator. The LDACS and automatic phase ADS-B electromagnetic compatibility data acquisition experimental platform is used for end-to-end denoising of ADS-B signals in complex electromagnetic environments. The convolutional block attention mechanism module CBAM integrates channel attention and spatial attention in series and is inserted in the middle layer between the generator and the discriminator.

Images