A semi-supervised learning-based method for detecting co-channel interference between satellites.

Abstract

This invention provides a semi-supervised learning-based method for detecting co-channel interference between satellites. The method includes: performing time-frequency preprocessing on received satellite signals; constructing a dual-attention Transformer and performing self-supervised pre-training; employing a dual-attention mechanism based on higher-order statistics to statistically model the input features and extract enhanced features; using wavelet regularization technology for multi-task fine-tuning to obtain a trained model; preprocessing the real-time received satellite signals into a time-frequency graph, inputting it into the trained model, and the model outputting the interference detection result. This invention employs a two-stage hybrid learning architecture to achieve interference detection in low signal-to-noise ratio environments.

Description

Technical Field

[0001] This invention relates to the field of signal processing and interference monitoring technology in satellite communication systems, and more specifically, to a method for detecting co-channel interference between satellites based on semi-supervised learning. Background Technology

[0002] In recent years, the global satellite communications industry has been undergoing unprecedented changes, with massive non-geostationary orbit (NGSO) constellations, represented by Starlink and OneWeb, being deployed in low Earth orbit on a large scale. The high-speed operation of tens of thousands of NGSO satellites inevitably overlaps their beam coverage with existing geostationary orbit (GSO) satellite systems, leading to extremely congested limited radio spectrum resources and increasingly severe inter-satellite co-channel interference. To maintain the reliable coexistence of next-generation integrated space-air-ground networks and strictly comply with the radio regulations set by the International Telecommunication Union, communication receivers must possess all-weather, real-time, and high-precision interference detection capabilities to promptly trigger interference avoidance or resource scheduling mechanisms.

[0003] The defects and shortcomings of the existing technology are as follows:

[0004] 1. Existing reconstruction-based detection models (such as TrID and VAE) rely on a fixed, empirically set reconstruction error (RE) threshold to determine signal status during the inference phase. However, real-world satellite channels are highly dynamic, time-varying systems, with signal power fluctuating drastically due to factors such as beam pointing and atmospheric attenuation. This leads to significant overlap in the numerical range between the reconstruction error distribution of normal signals (caused by noise) and the reconstruction error distribution of weak interference signals (caused by interference). In this ambiguous distribution, a single fixed threshold cannot find the optimal segmentation point: a high threshold leads to missed detections of weak interference, while a low threshold results in an extremely high false alarm rate. For example, the TrID model has an FPR as high as 17.63% in typical scenarios, meaning that a large number of normal communications will be misjudged as interference interruptions, severely damaging link availability.

[0005] 2. Satellite interference signals often exhibit both burstiness in the time domain (e.g., pulse interference) and specific spectral characteristics in the frequency domain (e.g., narrowband leakage). However, existing training paradigms typically employ a divide-and-conquer strategy, inputting time-domain waveform data and frequency-domain spectral data into two separate network branches for isolated processing, or using only single-domain data. This fragmented approach ignores the potential cross-domain dependencies of the signal during time-frequency transformation, causing the model to fail to capture complex interference features that only manifest in the joint time-frequency distribution. Experimental data shows that, due to insufficient feature utilization, the current state-of-the-art models remain at bottleneck levels of 0.832 and 0.711 in the receiver operating characteristic curves in the time and frequency domains, respectively, making further improvement difficult.

[0006] 3. Satellite links typically operate in low signal-to-noise ratio (SNR) environments, and the power of interfering signals can be extremely weak. Existing Transformer-based architectures mainly utilize first-order dot product attention mechanisms to calculate correlations, lacking specific suppression mechanisms for high-power Gaussian white noise. In low SNR environments, the random fluctuations in background noise often mask the structural characteristics of interfering signals. Existing models lack the ability to model higher-order statistics, failing to distinguish between structured interference and unstructured random noise. They are prone to misclassifying high-power noise as interference or missing interference under strong noise masking, leading to a significant decrease in the robustness of the model under harsh conditions.

[0007] 4. Existing models, such as TrID, typically follow a supervised learning paradigm during training, requiring massive amounts of high-quality paired data of clean and interfering signals. In real-world satellite networks, acquiring and accurately labeling this anomalous data is extremely expensive and difficult. Furthermore, to improve accuracy, existing models often employ complex decoupled architectures, resulting in enormous computational overhead during inference. Experiments show that their single inference latency often exceeds 100 milliseconds, while in NGSO systems, beam switching and time slot scheduling often occur within milliseconds. Existing high-latency models cannot meet the timeliness requirements of online real-time detection. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the present invention aims to provide a method for detecting co-channel interference between satellites based on semi-supervised learning. The present invention adopts a two-stage hybrid learning architecture to achieve interference detection in low signal-to-noise ratio environments.

[0009] The technical solution of the present invention is as follows:

[0010] A method for detecting co-channel interference between satellites based on semi-supervised learning includes the following steps:

[0011] Perform time-frequency preprocessing on the received satellite signals;

[0012] Construct a dual-attention Transformer and perform self-supervised pre-training;

[0013] A dual attention mechanism based on higher-order statistics is used to statistically model the input features and extract enhanced features;

[0014] Wavelet regularization is used for multi-task fine-tuning to obtain a well-trained model;

[0015] The real-time received satellite signals are preprocessed into a time-frequency graph, which is then input into the trained model, and the model outputs the interference detection results.

[0016] Preferably, the time-frequency preprocessing of the received satellite signal specifically includes: assuming the complex baseband signal received by the satellite downlink is x(t), a time-frequency image is generated using a pseudo-Wigner-Ville distribution, the calculation formula of which is as follows:

[0017]

[0018] in, Let f represent the time variable and f represent the frequency variable. For time delay variables, * denotes complex conjugate operation. It is a real-valued Gaussian smoothing window function used to suppress the inherent cross-term interference of the Wigner-Ville distribution in the time-frequency domain, thereby preserving the true energy distribution structure of the signal. The generated time-frequency plot is denoted as . As the input tensor for subsequent deep neural networks, the one-dimensional baseband signal is converted into a two-dimensional time-frequency tensor by performing the PWVD transformation.

[0019] Preferably, the construction of the dual-attention Transformer and the self-supervised pre-training specifically include:

[0020] Asymmetric architecture design: The model consists of a deep DAT encoder and a shallow decoder. The DAT encoder is used to extract robust latent features, while the decoder is only used for signal reconstruction during the pre-training stage.

[0021] Masking strategy: X the input time-frequency graph TF The image is divided into non-overlapping patches of size P×P. During training, a portion of the patches is randomly masked according to a preset ratio, and only the remaining visible patch sequence is input into the encoder.

[0022] Dual attention encoder structure: The encoder contains The network blocks are stacked, and each layer consists of a dilated convolution module and a dual attention module. The dilated convolution is used to extract local frequency domain features.

[0023] Dual attention mechanism: Self-attention is applied in parallel on the time axis and frequency axis respectively to capture long-distance signal dependencies;

[0024] Pre-training objective function: The goal of pre-training is to minimize the reconstruction error of the masked region, and the loss function is L. MAE Mean squared error is used, but only in the mask index set. Above calculation:

[0025]

[0026] in It is the reconstructed output of the decoder.

[0027] Preferably, the step of employing a dual attention mechanism based on higher-order statistics to statistically model the input features and extract enhanced features specifically includes:

[0028] Feature covariance calculation: Assume the input features of the attention layer are... N is the sequence length, and D is the feature dimension. First, calculate its feature covariance matrix. :

[0029]

[0030] in It is the eigenvalue mean vector, this matrix It encodes the global co-occurrence relationship between feature channels, which can effectively capture the statistical features of structured interference;

[0031]

[0032]

[0033] in For standard linear projection results, It is a learnable gated scalar;

[0034] Enhanced attention map generation: based on the corrected and Calculate attention weights:

[0035]

[0036] This design enables the model to suppress the weights of random noise based on global statistical properties, focusing on interference signal regions with structured characteristics.

[0037] Preferably, the step of using wavelet regularization technology for multi-task fine-tuning to obtain the trained model specifically includes:

[0038] The interference detection head consists of two fully connected layers and a sigmoid activation function, and outputs the probability of interference. The modulation recognition head is used to identify the modulation type of a signal, helping the model understand the semantics of the signal;

[0039] Signal reconstruction head: Used to output the reconstructed time-frequency diagram. ;

[0040] Constraining the reconstructed signal using discrete wavelet transform on the original input and reconstructed output Perform separately Wavelet decomposition of each level yields the high-frequency detail coefficients of each level. and low-frequency approximation coefficient Wavelet regularization loss function Defined as the L1 distance of the wavelet domain coefficients:

[0041]

[0042] loss function This ensures that the reconstructed signal not only approximates the real signal in terms of overall waveform, but also maintains consistency in high-frequency details such as phase transitions, thereby feeding back into the main detection task;

[0043] Total loss function during fine-tuning phase The weighted combination of losses for each task:

[0044]

[0045] in For binary cross-entropy loss, For multi-class cross-entropy loss, It is a pixel-level L1 loss.

[0046] Preferably, the step of preprocessing the real-time received satellite signal into a time-frequency map, inputting it into the trained model, and having the model output the interference detection result specifically includes: in the online deployment stage, preprocessing the real-time received satellite signal into a time-frequency map, inputting it into the trained satellite interference detection model based on semi-supervised dual attention Transformer, and having the model directly output the interference probability through the interference detection head.

[0047] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0048] 1. This invention trains a direct binary interference detection head through a multi-task fine-tuning strategy, directly outputting the probability of interference presence. This design eliminates the dependence on unstable thresholds and establishes an accurate, learnable decision boundary. Experimental results show that on public datasets, this invention maintains a high AUC value of 0.9085, significantly improving link reliability in practical deployments.

[0049] 2. This invention innovatively introduces a high-order statistical enhancement module, which explicitly models the global correlation structure between channels by calculating the feature covariance matrix, and injects this statistical prior into the attention mechanism. Experimental verification shows that even under adverse conditions such as low INR and large off-axis angles, this invention can still maintain a high detection probability, demonstrating significantly better robustness than existing benchmark models.

[0050] 3. This invention introduces Discrete Wavelet Transform (DWT) constraints into the reconstruction task. By applying L1 penalties across multiple frequency scales in the wavelet domain, the model is forced not only to match the global waveform envelope but also to accurately recover transient high-frequency details crucial for distinguishing interference. This mechanism makes the phase evolution pattern and constellation diagram geometry of the reconstructed signal clearer, thereby improving the feature extraction capabilities of the main detection task.

[0051] 4. This invention uses a masked autoencoder (MAE) for self-supervised pre-training, and learns the general physical structure of signals using a large amount of unlabeled data. Only a small amount of labeled data is needed to achieve state-of-the-art performance through fine-tuning. This overcomes the strict dependence of existing technologies on large-scale high-quality labeled data. Attached Figure Description

[0052] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0053] Figure 1 This is a flowchart of the inter-satellite co-frequency interference detection method based on semi-supervised learning of the present invention;

[0054] Figure 2 A schematic diagram of the network architecture for a satellite co-frequency interference detection model;

[0055] Figure 3 This is a schematic diagram of the dual-attention Transformer encoder structure;

[0056] Figure 4 A block diagram illustrating the principle of the dual attention mechanism for enhancing higher-order statistics;

[0057] Figure 5 This is a schematic diagram of the wavelet regularization loss calculation process;

[0058] Figure 6 This is a comparison chart of ROC curves under mixed interference scenarios;

[0059] Figure 7 A trend chart comparing detection accuracy under different INR conditions. Detailed Implementation

[0060] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0061] Specifically, this invention provides a method for detecting co-channel interference between satellites based on semi-supervised learning, such as... Figure 1 As shown, the method includes the following steps:

[0062] S1: Perform time-frequency preprocessing on the received satellite signals;

[0063] Specifically, addressing the prevalent Doppler shift and non-stationary characteristics in satellite signals, the received signal is first subjected to time-frequency transformation. Let the complex baseband signal received by the satellite downlink be x(t). To simultaneously obtain high time and frequency resolution and suppress cross-term interference, this invention employs a pseudo-Wigner-Ville distribution (PWVD) to generate the time-frequency image. The calculation formula is as follows:

[0064]

[0065] in, Let f represent the time variable and f represent the frequency variable. For time delay variables, * denotes complex conjugate operation. It is a real-valued Gaussian smoothing window function used to suppress the cross-term interference inherent in the Wigner-Ville distribution in the time-frequency domain, thereby preserving the true energy distribution structure of the signal. The generated time-frequency plot is denoted as... This serves as the input tensor for subsequent deep neural networks. By performing the PWVD transform, the one-dimensional baseband signal is converted into a two-dimensional time-frequency tensor.

[0066] S2: Construct a dual-attention Transformer and perform self-supervised pre-training;

[0067] Specifically, to address the difficulty of labeling satellite interference data, this invention designs an asymmetric dual-attention Transformer (DAT) pre-training architecture, such as... Figure 2 As shown, it includes:

[0068] Asymmetric architecture design: such as Figure 3 As shown, the model consists of a deep DAT encoder and a shallow decoder. The DAT encoder is used to extract robust latent features, while the decoder is only used for signal reconstruction during the pre-training phase.

[0069] Masking strategy: X the input time-frequency graph TF The image is divided into non-overlapping patches of size P×P. During training, a portion of the patches is randomly masked according to a preset ratio, and only the remaining visible patch sequence is input into the encoder.

[0070] Dual attention encoder structure: The encoder contains The network blocks are stacked, and each layer consists of a dilated convolution module and a dual attention module. The dilated convolution is used to extract local frequency domain features.

[0071] Dual attention mechanism: Self-attention is applied in parallel on the time axis and frequency axis respectively to capture long-distance signal dependencies;

[0072] Pre-training objective function: The goal of pre-training is to minimize the reconstruction error of the masked region, and the loss function is L. MAE Mean squared error is used, but only in the mask index set. Above calculation:

[0073]

[0074] in It is the reconstructed output of the decoder.

[0075] S3: Employs a dual attention mechanism based on higher-order statistics to statistically model the input features and extract enhanced features;

[0076] Specifically, to address the difficulty in distinguishing between structured interference and Gaussian white noise in low INR environments, this invention proposes a dual attention mechanism enhanced by Higher-Order Statistics (HOS). The core improvement of this scheme lies in changing the standard Transformer's method of calculating attention weights using only first-order linear projection, and innovatively introducing second-order statistics (covariance) between feature channels to explicitly model global correlation.

[0077] Figure 4 This is a block diagram illustrating the principle of the dual attention mechanism enhanced by higher-order statistics (HOS) according to an embodiment of the present invention, such as... Figure 4 As shown, this example demonstrates how to use the feature covariance matrix to inject statistical priors into the query (Q) and key (K) in the attention mechanism to suppress Gaussian noise. This embodiment explicitly models global correlation by calculating the second-order covariance matrix between feature channels. The specific implementation process is as follows:

[0078] Feature covariance calculation: Assume the input features of the attention layer are... (N is the sequence length, D is the feature dimension), first calculate its feature covariance matrix. :

[0079]

[0080] in It is the eigenvalue mean vector. This matrix... It encodes the global co-occurrence relationship between feature channels, which can effectively capture the statistical features of structured interference.

[0081]

[0082]

[0083] in For standard linear projection results, It is a learnable gating scalar.

[0084] Enhanced attention map generation: based on the corrected and Calculate attention weights:

[0085]

[0086] This design enables the model to suppress the weights of random noise based on global statistical properties, focusing on interference signal regions with structured characteristics.

[0087] S4: Wavelet regularization is used for multi-task fine-tuning to obtain a trained model;

[0088] Specifically, using a small amount of labeled data, pre-trained weights are loaded, and three parallel task heads are added for fine-tuning. To prevent feature forgetting during the fine-tuning process and to preserve signal details, this embodiment introduces wavelet regularization techniques, such as... Figure 5 As shown, Figure 5 This is a schematic diagram of the wavelet regularization loss calculation process according to an embodiment of the present invention, illustrating the process of multi-scale decomposition and L1 constraint of the original signal and the reconstructed signal using discrete wavelet transform.

[0089] The interference detection head consists of two fully connected layers (MLP) and a sigmoid activation function, outputting the probability of interference presence. The modulation identification head identifies the modulation type of a signal, assisting the model in understanding the signal semantics. The signal reconstruction head outputs the reconstructed time-frequency graph. .

[0090] To force the model to maintain structural fidelity across different frequency scales, this embodiment utilizes Discrete Wavelet Transform (DWT) to constrain the reconstructed signal. For the original input... and reconstructed output Perform separately Wavelet decomposition of each level yields the high-frequency detail coefficients of each level. and low-frequency approximation coefficient Wavelet regularization loss function Defined as the L1 distance of the wavelet domain coefficients:

[0091]

[0092] This loss function ensures that the reconstructed signal not only approximates the real signal in terms of the overall waveform, but also maintains consistency in high-frequency details such as phase transitions, thereby feeding back into the main detection task.

[0093] Total loss function during fine-tuning phase The weighted combination of losses for each task:

[0094]

[0095] in For binary cross-entropy loss, For multi-class cross-entropy loss, It is a pixel-level L1 loss.

[0096] S5: Preprocess the real-time received satellite signals into a time-frequency graph, input it into the trained model, and the model outputs the interference detection results.

[0097] Specifically, during the online deployment phase, the real-time received satellite signals are preprocessed into a time-frequency map, which is then input into a pre-trained satellite interference detection model based on a semi-supervised dual-attention Transformer. The model directly outputs the interference probability through the interference detection head. .

[0098] The judgment rules are set as follows:

[0099]

[0100] in The confidence threshold is typically set to 0.5. Unlike existing technologies that rely on unstable reconstruction error thresholds, the output of this invention is a calibrated probability value with clear physical meaning and higher robustness.

[0101] To verify the effectiveness of the proposed semi-supervised dual-attention Transformer-based satellite interference detection method, this embodiment conducted extensive simulation experiments in a simulated non-geostationary satellite communication scenario and compared it with existing mainstream detection models such as TrID, VAE, and CNN-AE. The experimental dataset covers satellite downlink signals under various channel conditions. The channel model incorporates Gaussian white noise, Rayleigh fading, and Doppler shift caused by high-speed satellite motion to cover the entire spectrum from weak to strong interference. The experiments used the area under the receiver operating characteristic curve (AUC), detection accuracy, false alarm rate, and inference delay as the main evaluation metrics.

[0102] The comprehensive testing performance comparison results show that... Figure 6 As shown, this invention significantly outperforms the comparative model in detection accuracy. The satellite interference detection model based on semi-supervised dual-attention Transformer achieves an AUC of 0.9085, an improvement of approximately 18 percentage points compared to the current state-of-the-art TrID model. Robustness analysis under low signal-to-interference-plus-noise ratio (SNR) is also demonstrated. Figure 7 As shown, in an extremely low environment with a signal-to-interference-plus-noise ratio of -10 dB, the detection accuracy of traditional models such as TrID drops sharply to below 80%, indicating that they are unable to distinguish weak interference features in strong noise. However, thanks to the high-order statistics enhancement module's modeling of global covariance, this invention can still maintain an accuracy of nearly 90% under equally harsh conditions, proving that the introduction of second-order statistics can effectively suppress the influence of Gaussian white noise.

[0103] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for inter-satellite co-frequency interference detection based on semi-supervised learning, characterized in that, The method includes the following steps: Perform time-frequency preprocessing on the received satellite signals; Constructing a dual-attention Transformer and performing self-supervised pre-training specifically includes: Asymmetric architecture design: The model consists of a deep DAT encoder and a shallow decoder. The DAT encoder is used to extract robust latent features, while the decoder is only used for signal reconstruction during the pre-training stage. Mask strategy: divide the input time-frequency map X TF into non-overlapping image patches Patch of size P × P; during the training process, randomly mask a portion of Patch according to a preset proportion, and only input the remaining visible Patch sequence into the encoder; Dual attention encoder structure: the encoder comprises A network block of layer stacks, each layer being composed of a dual attention module and a dilated convolution module, the dilated convolution being used to extract local frequency domain features; Dual attention mechanism: Self-attention is applied in parallel on the time axis and frequency axis respectively to capture long-distance signal dependencies; Pre-training objective function: The goal of pre-training is to minimize the reconstruction error of the masked region, and the loss function is L. MAE Mean squared error is used, but only in the mask index set. Above calculation: in It is the reconstructed output of the decoder; A dual attention mechanism based on higher-order statistics is employed to statistically model the input features and extract enhanced features, specifically including: Feature covariance calculation: Assume the input features of the attention layer are... N is the sequence length, and D is the feature dimension. First, calculate its feature covariance matrix. : in It is the eigenvalue mean vector, this matrix It encodes the global co-occurrence relationship between feature channels, which can effectively capture the statistical features of structured interference; in For standard linear projection results, It is a learnable gated scalar; Enhanced attention map generation: based on the corrected and Calculate attention weights: This design enables the model to suppress random noise weights based on global statistical properties, focusing on interference signal regions with structured characteristics. Wavelet regularization is used for multi-task fine-tuning to obtain a trained model, specifically including: The interference detection head consists of two fully connected layers and a sigmoid activation function, and outputs the probability of interference. The modulation recognition head is used to identify the modulation type of a signal, helping the model understand the semantics of the signal; Signal reconstruction head: Used to output the reconstructed time-frequency diagram. ; Constraining the reconstructed signal using discrete wavelet transform on the original input and reconstructed output Perform separately Wavelet decomposition of each level yields the high-frequency detail coefficients of each level. and low-frequency approximation coefficients Wavelet regularization loss function Defined as the L1 distance of the wavelet domain coefficients: Total loss function during fine-tuning phase The weighted combination of losses for each task: in For binary cross-entropy loss, For multi-class cross-entropy loss, For pixel-level L1 loss; The real-time received satellite signals are preprocessed into a time-frequency graph, which is then input into the trained model, and the model outputs the interference detection results.

2. The method for detecting co-channel interference between satellites based on semi-supervised learning according to claim 1, characterized in that, The time-frequency preprocessing of the received satellite signal specifically includes: assuming the complex baseband signal received by the satellite downlink is x(t), a time-frequency image is generated using a pseudo-Wigner-Ville distribution, calculated as follows: in, Let f represent the time variable and f represent the frequency variable. For time delay variables, * denotes complex conjugate operation. It is a real-valued Gaussian smoothing window function used to suppress the inherent cross-term interference of the Wigner-Ville distribution in the time-frequency domain, thereby preserving the true energy distribution structure of the signal. The generated time-frequency plot is denoted as . As the input tensor for subsequent deep neural networks, the one-dimensional baseband signal is converted into a two-dimensional time-frequency tensor by performing the PWVD transformation.

3. The method for detecting co-channel interference between satellites based on semi-supervised learning according to claim 1, characterized in that, The steps of preprocessing the real-time received satellite signals into a time-frequency map, inputting it into a trained model, and having the model output interference detection results specifically include: in the online deployment stage, preprocessing the real-time received satellite signals into a time-frequency map, inputting it into a trained satellite interference detection model based on a semi-supervised dual-attention Transformer, and having the model directly output interference probability through the interference detection head.

Images