CN-121984572-A - Inter-satellite co-channel interference detection method based on semi-supervised learning

Abstract

The invention provides a method for detecting inter-satellite co-channel interference based on semi-supervised learning, which comprises the steps of performing time-frequency preprocessing on received satellite signals, constructing a dual-attention transducer and performing self-supervised pre-training, performing statistical modeling on input features and extracting enhancement features by adopting a dual-attention mechanism based on high order statistic enhancement, performing multi-task fine tuning by adopting a wavelet regularization technology to obtain a trained model, preprocessing the satellite signals received in real time into a time-frequency diagram, inputting the trained model, and outputting an interference detection result by the model. The invention adopts a two-stage hybrid learning architecture to realize interference detection in a low signal-to-noise ratio environment.

Inventors

• GAO YUE
• YANG BOYU
• ZHAO JIN

Assignees

• 复旦大学

Dates

Publication Date

20260505

Application Date

20260331

Claims (6)

1. 1. The method for detecting the inter-satellite co-channel interference based on semi-supervised learning is characterized by comprising the following steps: Performing time-frequency preprocessing on the received satellite signals; constructing a dual-attention transducer and performing self-supervision pre-training; Adopting a dual-attention mechanism based on high order statistic enhancement to perform statistical modeling on input features and extract enhancement features; performing multitasking fine tuning by adopting a wavelet regularization technology to obtain a trained model; and preprocessing the satellite signals received in real time into a time-frequency diagram, inputting a trained model, and outputting an interference detection result by the model.
2. 2. The method for detecting inter-satellite co-channel interference based on semi-supervised learning as set forth in claim 1, wherein the performing time-frequency preprocessing on the received satellite signals specifically includes setting a complex baseband signal received by a satellite downlink as x (t), and generating a time-frequency image by using pseudo-Wigner-Ville distribution, wherein the calculation formula is as follows: Wherein, the Represents a time variable, f represents a frequency variable, Is a time delay variable, representing a complex conjugate operation, Is a real Gaussian smoothing window function for suppressing cross-term interference inherent in Wigner-Ville distribution in a time-frequency domain, thereby preserving the real energy distribution structure of the signal, and the generated time-frequency chart is recorded as As an input tensor for the subsequent deep neural network, the one-dimensional baseband signal is converted into a two-dimensional time-frequency tensor by performing PWVD transforms.
3. 3. The method for detecting inter-satellite co-channel interference based on semi-supervised learning as set forth in claim 1, wherein the constructing a dual-attention transducer and performing self-supervised pre-training specifically comprises: the asymmetric architecture design is that the model consists of a deep DAT encoder and a shallow decoder, wherein the DAT encoder is used for extracting robust potential characteristics, and the decoder is only used for signal reconstruction in a pre-training stage; The masking strategy is to divide the input time-frequency diagram X TF into non-overlapped image blocks Patch with the size of P multiplied by P, in the training process, randomly masking a part of Patch according to the preset proportion, and inputting the rest visible Patch sequence into an encoder; Dual-attention encoder structure-encoder includes A network block stacked in layers, each layer consisting of a hole convolution module and a dual-attention module, the hole convolution module being used for extracting local frequency domain features; A dual-attention mechanism, which applies self-attention on a time axis and a frequency axis in parallel respectively to capture long-distance signal dependency; Pre-training objective function the goal of pre-training is to minimize the reconstruction error of the mask region, the loss function L MAE employs the mean square error, but only at the mask index set And (3) calculating: Wherein the method comprises the steps of Is the reconstructed output of the decoder.
4. 4. The method for detecting inter-satellite co-channel interference based on semi-supervised learning as recited in claim 1, wherein the step of statistically modeling the input features and extracting the enhanced features using a dual-attention mechanism based on higher order statistic enhancement specifically comprises: feature covariance calculation assuming that the input features of the attention layer are N is the length of the sequence, D is the feature dimension, and the feature covariance matrix is calculated first : Wherein the method comprises the steps of Is a feature mean vector, the matrix The global co-occurrence relation between the characteristic channels is encoded, and the statistical characteristics of the structured interference can be effectively captured; Wherein the method comprises the steps of As a result of the standard linear projection, Is a learnable gating scalar; enhanced attention-seeking diagram generation based on corrected And Calculating attention weight: This design enables the model to suppress the weight of random noise based on global statistics, focusing on the region of the interference signal with structured features.
5. 5. The method for detecting inter-satellite co-channel interference based on semi-supervised learning as set forth in claim 1, wherein the step of performing multi-task fine tuning by using a wavelet regularization technique to obtain a trained model specifically includes: The interference detection head consists of two full-connection layers and a Sigmoid activation function, and outputs the probability of interference existence The modulation recognition head is used for recognizing the modulation type of the signal and assisting the model in understanding the signal semantics; the signal reconstruction head is used for outputting a reconstructed time-frequency diagram ; Constraining reconstructed signals using discrete wavelet transforms to original inputs And reconstructing the output Respectively do Stage wavelet decomposition to obtain high-frequency detail coefficients of each stage And low frequency approximation coefficients Wavelet regularization loss function L1 distance defined as wavelet domain coefficients: total loss function of fine tuning stage Weighted combination of losses for each task: Wherein the method comprises the steps of For the binary cross-entropy loss, For multi-class cross entropy loss, Lost for pixel level L1.
6. 6. The method for detecting the inter-satellite co-channel interference based on semi-supervised learning according to claim 1 is characterized in that the step of preprocessing a satellite signal received in real time into a time-frequency diagram, inputting a trained model, and outputting an interference detection result by the model specifically comprises the steps of preprocessing the satellite signal received in real time into the time-frequency diagram in an on-line deployment stage, inputting the trained satellite interference detection model based on a semi-supervised double-attention transducer, and outputting the interference probability by the model directly through an interference detection head.

Description

Inter-satellite co-channel interference detection method based on semi-supervised learning Technical Field The invention relates to the technical field of signal processing and interference monitoring in a satellite communication system, in particular to a method for detecting inter-satellite co-channel interference based on semi-supervised learning. Background In recent years, the global satellite communications industry is experiencing an unprecedented revolution in which Non-stationary orbit (Non-Geostationary Orbit, NGSO) giant constellations represented by the star chain and OneWeb are being deployed densely on a large scale in low earth orbit. The high speed operation of thousands of NGSO satellites inevitably overlaps the beam coverage of the existing stationary orbit (Geostationary Orbit, GSO) satellite system, resulting in extremely crowded limited radio spectrum resources and more severe inter-satellite co-channel interference problems. In order to maintain the reliable coexistence of the next generation air-space-ground integrated network and strictly conform to the radio rule formulated by the international telecommunication union, the communication receiving end must have all-weather, real-time and high-precision interference detection capability so as to trigger an interference avoidance or resource scheduling mechanism in time. The drawbacks and deficiencies of the prior art are shown below: 1. Existing reconstruction-based detection models (e.g., trID, VAE) must rely on a fixed, empirically set Reconstruction Error (RE) threshold to determine signal state during the inference phase. However, the actual satellite channel is a highly dynamic time-varying system, and the signal power fluctuates dramatically due to factors such as beam pointing, atmospheric attenuation, etc. This results in a severe overlap in the value interval between the reconstruction error distribution of the normal signal (caused by noise) and the reconstruction error distribution of the weak interference signal (caused by interference). Under the condition of fuzzy distribution, a single fixed threshold cannot find an optimal division point, namely, a threshold is high, so that weak interference is missed, and a threshold is low, so that a very high false alarm rate is caused. For example, the TrID model FPR in a typical scenario is as high as 17.63%, which means that a large number of normal communications can be misinterpreted as an interference break, severely compromising the availability of the link. 2. Satellite interference signals tend to have both burstiness in the time domain (e.g., impulse interference) and specific spectral characteristics in the frequency domain (e.g., narrowband leakage). However, existing training patterns typically employ a divide-and-conquer strategy, with time domain waveform data and frequency domain spectral data being input into two separate network branches for isolation processing, or with only a single domain data. The processing mode of the splitting ignores potential cross-domain dependence of the signal in the time-frequency transformation process, so that the model cannot capture complex interference features only appearing in the time-frequency joint distribution. Experimental data indicate that the area under the receiver operating characteristic curve of the existing state-of-the-art model in the time and frequency domains stays at the bottleneck level of 0.832 and 0.711, respectively, and is difficult to further promote due to insufficient characteristic utilization. 3. Satellite links typically operate in low signal-to-noise ratio (SNR) environments and the interfering signal power may be extremely weak. The existing architecture based on the transducer mainly utilizes a first-order dot product attention mechanism to calculate the correlation, and lacks a special suppression mechanism for high-power Gaussian white noise. Random fluctuation energy of background noise tends to mask structural features of interfering signals when in a low signal-to-interference-and-noise environment. The existing model lacks high order statistic modeling capability, cannot distinguish structured interference from unstructured random noise, and is easy to misjudge high-power noise as interference or miss detection interference under the masking of strong noise, so that the robustness of the model under severe working conditions is obviously reduced. 4. The training process of the existing model represented by TrID usually follows a supervised learning paradigm, massive and high-quality pure signals and paired data containing interference signals are required to train, and in an actual satellite network, acquiring and accurately labeling the abnormal data is extremely expensive and difficult. In addition, in order to improve accuracy, the existing model often stacks complex decoupling architectures, which results in huge computational overhead in reasoning. Experiments show that the single reas