Radio interference identification method based on electromagnetic spectrum monitoring

Abstract

The invention discloses a radio interference identification method based on electromagnetic spectrum monitoring, and the method comprises the following steps: S1, collecting spectrum data in an electromagnetic environment through broadband spectrum monitoring equipment, and carrying out the digital sampling processing; s2, adopting a diffusion probability model to automatically remove environmental noise and non-interference signals; s3, automatically generating an expansion data set of edge and small sample interference signals by adopting an attention-enhanced data synthesis expansion technology; s4, extracting multi-dimensional time-frequency domain dynamic correlation characteristics of the spectrum interference signal based on a Transform structure; s5, adopting an attention enhancement self-supervision algorithm to generate a self-supervision soft label of an unknown interference type; s6, dynamic back diffusion iteration is carried out through the diffusion probability model, and an interference classification result is obtained; and S7, constructing a closed-loop dynamic feedback mechanism by using an interference classification result. According to the invention, accurate identification of radio interference signals is realized, and the accuracy, dynamic adaptability and automation degree of interference signal classification are improved.

Description

Technical Field

[0001] The present invention relates to the field of radio communication technology, and in particular to a radio interference identification method based on electromagnetic spectrum monitoring. Background Art

[0002] Radio communication technology plays a key role in fields such as communication navigation, radar detection, and electronic countermeasures. The electromagnetic spectrum serves as a fundamental resource for radio communications. In recent years, with the rapid growth of wireless communication devices, the electromagnetic environment has become increasingly complex and congested. Various types of radio interference have become frequent, posing a significant threat to the reliability and stability of wireless communication systems. Consequently, electromagnetic spectrum monitoring technology has gradually developed into a crucial means of ensuring the secure and stable operation of wireless communication systems, and has garnered widespread attention.

[0003] Currently, conventional electromagnetic spectrum monitoring interference identification technologies typically employ traditional signal processing methods, such as fast Fourier transforms, wavelet analysis, and classical pattern recognition algorithms. These methods primarily rely on feature extraction and statistical pattern matching based on time-frequency analysis, or are combined with traditional machine learning classifiers to identify and classify known types of interference. However, in complex actual electromagnetic environments, radio interference signals often exhibit dynamic changes and random characteristics. Existing static or semi-static feature extraction methods are difficult to adapt to the rapid changes in interference types, and are particularly difficult to efficiently identify edge-type interference signals and small-sample interference signals. Furthermore, traditional methods typically require extensive manual intervention for labeling and training, and lack automated dynamic feature learning capabilities, resulting in a significant gap between their degree of automation and actual application requirements.

[0004] While recent advances in deep learning methods, such as convolutional neural networks and recurrent neural networks, can improve interference identification performance to some extent, they still perform poorly for small sample sizes and for identifying unknown interference signals. Furthermore, deep learning-based models typically require large amounts of labeled data for training, making them difficult to quickly adapt to new interference patterns under unsupervised or self-supervised conditions. Furthermore, most existing deep learning methods ignore the long-range correlation characteristics of spectrum data in the time and frequency domains, as well as important information about weak interference signals, limiting their adaptability and classification accuracy in electromagnetic spectrum monitoring applications.

[0005] Therefore, how to provide a radio interference identification method based on electromagnetic spectrum monitoring is a problem that those skilled in the art urgently need to solve. Summary of the Invention

[0006] One purpose of the present invention is to propose a radio interference identification method based on electromagnetic spectrum monitoring. The present invention can effectively improve the accuracy and automation of radio interference identification, and is particularly suitable for the precise identification of weak interference signals and unknown interference types in complex and dynamic electromagnetic environments.

[0007] A radio interference identification method based on electromagnetic spectrum monitoring according to an embodiment of the present invention includes the following steps:

[0008] S1. Real-time collection of raw spectrum data in a complex electromagnetic environment using broadband spectrum monitoring equipment, and digital sampling and processing of the raw spectrum data in the time and frequency domains to obtain digital raw spectrum data;

[0009] S2. Perform unsupervised denoising on the digitized raw spectrum data using a self-supervised diffusion probability denoising model, automatically identify and remove environmental noise and non-interfering irrelevant signals, and output purified spectrum data;

[0010] S3. For the purified spectrum data, the attention-enhanced data synthesis and expansion technology is used to automatically generate an expanded spectrum feature dataset of edge-type interference signals and small-sample interference signals;

[0011] S4. Based on the multi-scale cross-time window Transformer structure, the multi-dimensional time-frequency domain long-range dynamic correlation features of the spectrum interference signal in the expanded spectrum feature data are extracted to obtain the interference signal feature coding data;

[0012] S5. Using the attention-enhanced self-supervised learning algorithm, based on the interference signal feature coding data, an unsupervised interference feature learning network is constructed to automatically generate self-supervised soft labels corresponding to unknown interference types, and a paired dataset is obtained that matches the feature coding data with the soft labels;

[0013] S6. Input the paired data set into the diffusion probability model, construct a multi-stage dynamic back-diffusion iterative process, classify the interference signal at a fine granularity in the high-dimensional probability space, and output the interference classification result;

[0014] S7. Based on the interference classification results, a bidirectional attention interaction feedback network is constructed between the Transformer structure to form a closed-loop dynamic feedback mechanism.

[0015] Optionally, the S1 specifically includes:

[0016] S11. Utilize a built-in ultra-wideband antenna array to capture broadband analog electromagnetic spectrum signals in a complex electromagnetic environment in real time using spatial diversity, and utilize a low insertion loss dielectric resonant bandpass filter to perform frequency band selection and out-of-band suppression on the analog electromagnetic spectrum signals to obtain a bandpass filtered analog electromagnetic spectrum signal.

[0017] S12, inputting the analog electromagnetic spectrum signal after bandpass filtering into a cascaded low-noise and high-linearity radio frequency preamplifier and a multi-channel radio frequency down-conversion circuit in sequence, performing radio frequency low-noise amplification and tuning processing, and converting the analog electromagnetic spectrum signal into an analog intermediate frequency signal in a predefined intermediate frequency range;

[0018] S13. Inputting the analog intermediate frequency signal into a high-speed analog-to-digital converter with a dual-channel interleaved sampling architecture, alternately completing analog-to-digital conversion using a dual-channel interleaved analog-to-digital conversion method, wherein the sampling rate of the analog-to-digital converter is set to be at least twice the bandwidth specified by the Nyquist sampling theorem, and calibrating the inconsistency error between the interleaved channels using a digital phase calibration method to obtain an initial digitized spectrum data sequence;

[0019] S14, inputting the initial digitized spectrum data sequence into a digital multi-stage cascade decimation filter, performing decimation, digital down-conversion, and anti-aliasing filtering operations in sequence, wherein the decimation multiple is a preset fixed integer value, the center frequency of the digital down-conversion is zero Hz, and outputting zero intermediate frequency baseband digital spectrum data;

[0020] S15, performing time domain sliding window segmentation processing on the baseband digital spectrum data, where the length of the sliding window is fixed, and the sliding step is set to half of the window length, to generate an overlapping data frame sequence;

[0021] S16. Performing a short-time Fourier transform on each data frame in the overlapping data frame sequence using a window function set to a Hamming window, with the window length fixed at 1024 sampling points and the overlap ratio between windows set to 50%. After each data frame is short-time Fourier transformed, a time-frequency matrix of digitized original spectrum data with a fixed spectrum resolution is obtained;

[0022] S17. Splicing the digitized original spectrum data time-frequency matrix into data frames according to a strict time sequence to form a complete and continuous digitized original spectrum data set.

[0023] Optionally, the S2 specifically includes:

[0024] S21, using the digitized original spectrum dataset as input data, and gradually performing a multi-scale random Gaussian noise superposition operation for a total of T stages on the input data based on a random Gaussian diffusion forward processing mechanism to obtain a data sequence after diffusion processing, where T is the total number of diffusion stages;

[0025] S22. Use a pre-trained unsupervised convolutional autoencoder network to perform initial feature encoding on the data sequence after diffusion processing. The convolutional autoencoder network structure adopts a symmetrical multi-layer convolution-transposed convolution stacked architecture. The convolution kernel sizes of the convolution layers are set to a combination of 3×3, 5×5, and 7×7, respectively. The activation function uses a linear rectification function to obtain the initial potential feature data representation;

[0026] S23, using the initial latent feature data representation as the initial value, constructing a T-stage dynamic back-diffusion denoising iterative process based on the Markov chain;

[0027] S26 , repeatedly executing steps S23 to S25 until all T stages of the reverse diffusion iteration process are completed, and the final reverse diffusion iteration output data is used as the purified spectrum data.

[0028] Optionally, the S3 specifically includes:

[0029] S31, dividing the purified spectrum data set into a mainstream interference data set, a marginal interference data set, and a small sample interference data set, wherein the marginal interference data set and the small sample interference data set are respectively determined by a preset interference category occurrence frequency threshold and an interference category sample quantity threshold;

[0030] S32, performing time-domain energy distribution statistical processing, frequency-domain power spectrum density calculation processing, and short-time Fourier transform processing on the spectrum data in the edge-type interference data set and the small-sample-type interference data set respectively to obtain the time-frequency domain amplitude-phase joint statistical distribution characteristics, and construct the initial spectrum feature vectors of the respective data sets;

[0031] S33, performing random time domain window position offset processing, frequency domain feature scale random scaling processing, and time-frequency domain feature rotation transformation processing on the initial spectrum feature vector of the edge type interference data set, to obtain multiple enhanced edge type interference spectrum feature vectors;

[0032] S34, performing spectrum amplitude random perturbation processing, spectrum power random energy redistribution processing, and spectrum phase random perturbation processing on the initial spectrum feature vector of the small sample type interference data set, respectively, to obtain multiple enhanced small sample type interference spectrum feature vectors;

[0033] S35. Based on the Transformer multi-head self-attention network, we construct an edge-type interference feature selection network and a small sample-type interference feature selection network, and input the enhanced feature vector and the original feature vector into their respective corresponding feature selection networks to obtain the attention weight of each feature vector.

[0034] S36. Based on the attention weights, perform feature-wise attention weighted fusion processing on the original feature vectors and enhanced feature vectors in the edge type interference dataset and the small sample type interference dataset, respectively, to generate edge type interference attention enhanced feature vectors and small sample type interference attention enhanced feature vectors;

[0035] S37. Merge the edge type interference attention enhancement feature vector, the small sample type interference attention enhancement feature vector, and the initial spectrum feature vector set in the mainstream type interference data set to obtain an expanded spectrum feature data set.

[0036] Optionally, the S34 specifically includes:

[0037] S341, using the initial spectrum feature vector of the small sample type interference data set as input data, performing spectrum amplitude random perturbation processing on each feature dimension, randomly selecting an amplitude perturbation scale factor to increase or decrease the amplitude of each feature dimension, and obtaining multiple spectrum feature vectors after amplitude perturbation;

[0038] S342. Performing a spectrum power random energy redistribution process on the multiple amplitude-perturbed spectrum feature vectors, with the constraint that the total amount of spectrum energy remains unchanged, randomly redistributing energy between the frequency points of the spectrum feature to obtain multiple spectrum feature vectors with differentiated energy distributions;

[0039] S343, performing a spectrum phase random perturbation process based on the multiple spectrum eigenvectors after energy redistribution, applying a random phase offset to the phase value of each frequency component of each eigenvector, and obtaining multiple spectrum eigenvectors after phase perturbation;

[0040] S344, respectively calculating the Euclidean distances between the plurality of feature vectors and the initial spectrum feature vector, and selecting the spectrum feature vectors whose Euclidean distances satisfy a preset feature disturbance threshold as effective enhanced candidate feature vectors;

[0041] S345. Calculate the feature correlation score between each valid enhanced candidate feature vector and the initial spectrum feature vector using the Transformer multi-head self-attention network, sort the feature vectors from high to low according to the feature correlation score, and select a preset number of enhanced feature vectors as the final enhanced spectrum feature vectors.

[0042] S346 , merging the finally selected enhanced spectrum feature vector with the initial spectrum feature vector in the small sample type interference data set to form an enhanced small sample type interference spectrum feature data set.

[0043] Optionally, the S4 specifically includes:

[0044] S41, dividing the expanded spectrum feature data set into multiple time series data segments according to three fixed time window lengths of different scales, to generate overlapping time series data segments of three scales;

[0045] S42, performing a short-time Fourier transform of a fixed length on each time series data segment within each scale, using a Hamming window as the window function, to generate multi-dimensional time-frequency domain feature representation data of different scales;

[0046] S43. Based on the Transformer position encoding rule, scale-time domain joint position encoding is added to the multi-dimensional time-frequency domain feature representation data in the three scales. Each position encoding consists of two parts.

[0047] S44. Using multi-dimensional time-frequency domain feature representation data containing scale-time domain joint position encoding as input data for the Transformer multi-head self-attention network, each self-attention head extracts the long-range dynamic correlation features of the spectral feature data in each scale in the time dimension and frequency dimension through the inner product attention calculation method, and outputs independent time-frequency feature encodings in multiple scales;

[0048] S45. Based on the time-frequency feature encoding output at each scale, cross-scale feature fusion is performed through a cross-scale Transformer multi-head self-attention interactive fusion network. The time-frequency feature encoding at each scale is simultaneously used as the query vector, key vector, and value vector of the self-attention network. The inner product attention is used to calculate the similarity correlation features between multiple scales to form a cross-scale time-frequency feature fusion encoding.

[0049] S46. The cross-scale time-frequency feature fusion encoding is input into a self-supervised two-layer feedforward neural network for nonlinear transformation. The input layer to the hidden layer and the hidden layer to the output layer of the feedforward neural network use linear mapping and ReLU activation function respectively to extract the deep correlation feature patterns of the spectrum interference signal in the time-frequency domain at different scales.

[0050] S47, performing a layer normalization operation on each feature dimension on the time-frequency feature coding that has undergone nonlinear transformation of the feedforward neural network, and adjusting the normalized coding data to interference signal feature coding data of a unified fixed dimension through a unified linear mapping process.

[0051] Optionally, the S5 specifically includes:

[0052] S51, dividing the interference signal characteristic coded data into multiple characteristic data batches;

[0053] S52, performing unsupervised clustering processing on each batch of feature data using a density peak clustering algorithm, and determining the cluster center within each batch based on the feature space density distribution;

[0054] S53. Taking the cluster center as the benchmark, the attention-enhanced self-supervised learning algorithm based on the Transformer multi-head self-attention structure is used to calculate the feature space similarity between the feature encoding data vector in each batch and the corresponding cluster center, and obtain the attention similarity coefficient of each feature encoding data vector;

[0055] S54, performing normalization processing based on the relative magnitude relationship of the attention similarity coefficients, and automatically generating a self-supervised soft label vector with a dimension corresponding to the number of cluster centers for each feature encoding data vector;

[0056] S55. For the soft label vector, a mutual information calculation method is used to perform a self-supervised consistency check, and the mutual information value between each feature encoding data vector and the corresponding soft label vector is calculated respectively, and the data pairs with a mutual information value greater than a preset threshold are selected as valid paired data;

[0057] S56. Construct an initial paired data set based on the valid paired data, and use the initial paired data set to perform dynamic gradient optimization update of the self-supervised network parameters;

[0058] S57. Loop through steps S52 to S56 until the loss function value of the attention-enhanced self-supervised learning algorithm is lower than the convergence threshold for multiple consecutive iterations, and then stop the iteration. The data set output from the last iteration is determined as the final paired data set that matches the interference signal feature encoding data with the self-supervised soft label.

[0059] Optionally, the S6 specifically includes:

[0060] S61, each feature encoding data vector and the corresponding self-supervised soft label vector in the paired data set are combined to form initial high-dimensional feature state data, and input into the diffusion probability model for initial probability state mapping;

[0061] S62. Construct a multi-stage dynamic reverse diffusion iterative process based on a Markov chain within the diffusion probability model, define the probability transfer relationship between each reverse diffusion stage, and use a preset dynamic state transition probability matrix to perform reverse state migration on the characteristic state data stage by stage;

[0062] S63. In each back-diffusion stage, the Transformer multi-head self-attention structure is used to calculate the conditional probability distribution between the feature state data of the current stage and the associated self-supervised soft label vector, and obtain a refined conditional probability representation of the feature state data of the current stage;

[0063] S64. Based on the refined conditional probability representation of the characteristic state data obtained in each back-diffusion stage, a deep neural network is used to perform probability space noise residual estimation of the current stage, and the back-diffusion state of the characteristic state data is dynamically corrected using the residual estimation result;

[0064] S65. Based on the dynamically corrected feature state data, perform state inverse mapping stage by stage, gradually reduce the probability space noise through the probability transfer inverse update mechanism, and gradually approach the pure feature state data in the feature space;

[0065] S66. At the end of the reverse diffusion iterative process, the multi-head self-attention fusion mechanism is used to integrate the intermediate state data of all diffusion stages to construct the final fused pure feature state data, and obtain the pure state feature representation of the interference signal in the high-dimensional probability space;

[0066] S67. Input the pure state feature representation into the pre-trained linear multi-classifier, perform fine-grained classification and recognition on the pure state feature in the high-dimensional probability space based on the classification decision boundary defined by the multi-classifier, and output the final interference classification result.

[0067] Optionally, the S7 specifically includes:

[0068] S71. Using the interference classification results as initial feedback data, a bidirectional attention interaction feedback network is constructed between the feedback data and the Transformer structure.

[0069] S72, using the feature vector corresponding to each interference category in the interference classification result as input data, respectively calculating the attention weight value between the feature vector and the interference signal feature encoding data output by the Transformer structure, and obtaining the association strength between the interference category and the feature encoding data;

[0070] S73. Representing the correlation strength between the interference category and the feature encoding data in the form of an attention weight matrix, obtaining feedback feature adjustment weights through multi-head self-attention calculation, and dynamically adjusting the parameters of the multi-head self-attention network in the feature extraction stage of the Transformer structure in real time;

[0071] S74, based on the real-time adjusted Transformer multi-head self-attention network parameters, re-execute the feature extraction process within the Transformer structure to generate updated feature encoding data;

[0072] S75, re-inputting the updated feature coding data into the diffusion probability model, executing the multi-stage dynamic reverse diffusion iterative process of the diffusion probability model, and generating an updated interference classification result;

[0073] S76, calculating a classification difference index between the updated interference classification result and the previous interference classification result based on the degree of category difference between the two classification results. When the classification difference index does not reach a preset stability threshold, continue to execute steps S72 to S75;

[0074] S77 , repeat steps S72 to S76 until the classification difference index reaches a stable threshold, output a stable interference classification result, and form a closed-loop dynamic feedback mechanism.

[0075] The beneficial effects of the present invention are:

[0076] (1) The present invention can automatically identify and remove environmental noise and non-interference signals through an automatic denoising method for spectrum data based on a diffusion probability model, effectively improving the signal-to-noise ratio and feature purity of interference signals, and enhancing the recognition accuracy and reliability of interference signals in complex electromagnetic environments.

[0077] (2) The present invention uses attention-enhanced data synthesis expansion technology and attention-enhanced self-supervised learning algorithm to achieve automatic high-quality feature synthesis and soft labeling of edge-type and small-sample-type interference signals, significantly improving the model's ability to classify and identify unknown or sample-scarce interference types, and showing better adaptability and classification effects in actual complex and changeable radio environments.

[0078] (3) In terms of the accuracy of dynamic feature extraction and classification of interference signals, the present invention effectively solves the problems in the existing technology that static feature extraction methods are difficult to capture long-range dynamic correlation information and lack real-time adaptive feedback mechanisms through a multi-scale cross-time window Transformer structure and a closed-loop dynamic feedback mechanism. It breaks through the bottleneck of existing interference identification methods in terms of dynamic adaptability and automation, and achieves a significant improvement in the interference signal classification performance, thereby effectively improving the interference identification and monitoring capabilities in complex electromagnetic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0080] Figure 1 This is a schematic diagram of the overall process of a radio interference identification method based on electromagnetic spectrum monitoring proposed by the present invention;

[0081] Figure 2 This is a schematic diagram of the internal structure of the Transformer and attention-enhanced self-supervised diffusion probability model for a radio interference identification method based on electromagnetic spectrum monitoring proposed in the present invention;

[0082] Figure 3 This is a flow chart of a closed-loop dynamic feedback mechanism of a radio interference identification method based on electromagnetic spectrum monitoring proposed by the present invention. DETAILED DESCRIPTION

[0083] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0084] refer to Figure 1-Figure 3 , a radio interference identification method based on electromagnetic spectrum monitoring, comprising the following steps:

[0085] S1. Real-time collection of raw spectrum data in a complex electromagnetic environment using broadband spectrum monitoring equipment, and digital sampling and processing of the raw spectrum data in the time and frequency domains to obtain digital raw spectrum data;

[0086] S2. Perform unsupervised denoising on the digitized raw spectrum data using a self-supervised diffusion probability denoising model, automatically identify and remove environmental noise and non-interfering irrelevant signals, and output purified spectrum data;

[0087] S3. For the purified spectrum data, the attention-enhanced data synthesis and expansion technology is used to automatically generate an expanded spectrum feature dataset of edge-type interference signals and small-sample interference signals;

[0088] S4. Based on the multi-scale cross-time window Transformer structure, the multi-dimensional time-frequency domain long-range dynamic correlation features of the spectrum interference signal in the expanded spectrum feature data are extracted to obtain the interference signal feature coding data;

[0089] S5. Using the attention-enhanced self-supervised learning algorithm, based on the interference signal feature coding data, an unsupervised interference feature learning network is constructed to automatically generate self-supervised soft labels corresponding to unknown interference types, and a paired dataset is obtained that matches the feature coding data with the soft labels;

[0090] S6. Input the paired data set into the diffusion probability model, construct a multi-stage dynamic back-diffusion iterative process, classify the interference signal at a fine granularity in the high-dimensional probability space, and output the interference classification result;

[0091] S7. Based on the interference classification results, a bidirectional attention interaction feedback network is constructed between the Transformer structure to form a closed-loop dynamic feedback mechanism.

[0092] The present invention uses a diffusion probability model to perform intelligent preprocessing on the original spectrum data, realizes the automatic identification and removal of environmental noise and non-interference signals, and significantly improves the quality and purity of the spectrum data; further utilizes the attention-enhanced data synthesis expansion technology to effectively solve the problem that traditional methods cannot effectively identify edge-type interference signals and small-sample interference signals; at the same time, the self-supervised learning algorithm based on the multi-scale cross-time window Transformer structure and attention enhancement is adopted to deeply explore the long-range dynamic correlation characteristics of interference signals in the time-frequency domain, thereby improving the accuracy and dynamic adaptability of interference signal feature extraction; in addition, the dynamic back-diffusion iterative process of the diffusion probability model is used to achieve accurate classification of interference signals, and the closed-loop dynamic feedback mechanism is combined to continuously optimize the model performance, so that the radio interference identification of the present invention in complex electromagnetic environments has higher accuracy, adaptability and automation level.

[0093] In this embodiment, S1 specifically includes:

[0094] S11. Utilize a built-in ultra-wideband antenna array to capture broadband analog electromagnetic spectrum signals in a complex electromagnetic environment in real time using spatial diversity, and utilize a low insertion loss dielectric resonant bandpass filter to perform frequency band selection and out-of-band suppression on the analog electromagnetic spectrum signals to obtain a bandpass filtered analog electromagnetic spectrum signal.

[0095] S12, inputting the analog electromagnetic spectrum signal after bandpass filtering into a cascaded low-noise and high-linearity radio frequency preamplifier and a multi-channel radio frequency down-conversion circuit in sequence, performing radio frequency low-noise amplification and tuning processing, and converting the analog electromagnetic spectrum signal into an analog intermediate frequency signal in a predefined intermediate frequency range;

[0096] S13. Inputting the analog intermediate frequency signal into a high-speed analog-to-digital converter with a dual-channel interleaved sampling architecture, alternately completing analog-to-digital conversion using a dual-channel interleaved analog-to-digital conversion method, wherein the sampling rate of the analog-to-digital converter is set to be at least twice the bandwidth specified by the Nyquist sampling theorem, and calibrating the inconsistency error between the interleaved channels using a digital phase calibration method to obtain an initial digitized spectrum data sequence;

[0097] S14, inputting the initial digitized spectrum data sequence into a digital multi-stage cascade decimation filter, performing decimation, digital down-conversion, and anti-aliasing filtering operations in sequence, wherein the decimation multiple is a preset fixed integer value, the center frequency of the digital down-conversion is zero Hz, and outputting zero intermediate frequency baseband digital spectrum data;

[0098] S15, performing time domain sliding window segmentation processing on the baseband digital spectrum data, where the length of the sliding window is fixed, and the sliding step is set to half of the window length, to generate an overlapping data frame sequence;

[0099] S16. Performing a short-time Fourier transform on each data frame in the overlapping data frame sequence using a window function set to a Hamming window, with the window length fixed at 1024 sampling points and the overlap ratio between windows set to 50%. After each data frame is short-time Fourier transformed, a time-frequency matrix of digitized original spectrum data with a fixed spectrum resolution is obtained;

[0100] S17. Splicing the digitized original spectrum data time-frequency matrix into data frames according to a strict time sequence to form a complete and continuous digitized original spectrum data set.

[0101] The present invention effectively improves the capture quality of analog electromagnetic spectrum signals through spatial diversity capture and bandpass filtering of the built-in ultra-wideband antenna array and dielectric resonant bandpass filter; improves the amplification and frequency conversion quality of analog signals through cascade processing of low-noise and high-linearity RF preamplifier and multi-channel RF down-conversion circuit; further ensures the accuracy and consistency of digital sampling through a dual-channel interleaved high-speed analog-to-digital converter and a digital phase calibration method; at the same time, digital down-conversion and anti-aliasing filtering are achieved through multi-stage cascade extraction filters, effectively improving the quality of digitized spectrum data; and utilizes short-time Fourier transform and overlapping data frame splicing methods to ensure the continuity and time-frequency characteristic stability of digitized spectrum data, thereby achieving high-quality and accurate acquisition and processing of spectrum data in complex electromagnetic environments, providing a solid technical guarantee for subsequent interference identification.

[0102] In this embodiment, S2 specifically includes:

[0103] S21, using the digitized original spectrum dataset as input data, and gradually performing a multi-scale random Gaussian noise superposition operation for a total of T stages on the input data based on a random Gaussian diffusion forward processing mechanism to obtain a data sequence after diffusion processing, where T is the total number of diffusion stages;

[0104] S22. Use a pre-trained unsupervised convolutional autoencoder network to perform initial feature encoding on the data sequence after diffusion processing. The convolutional autoencoder network structure adopts a symmetrical multi-layer convolution-transposed convolution stacked architecture. The convolution kernel sizes of the convolution layers are set to a combination of 3×3, 5×5, and 7×7, respectively. The activation function uses a linear rectification function to obtain the initial potential feature data representation;

[0105] S23. Using the initial potential feature data representation as the initial value, a T-stage dynamic back-diffusion denoising iterative process based on the Markov chain is constructed. In the t-stage back-diffusion denoising process:

[0106]

[0107] Among them, X t Represents the spectrum data matrix of the back diffusion input in the tth stage, Xt-1 represents the denoised spectrum data matrix output at the (t-1) stage, γ t Represents the diffusion intensity adjustment factor at stage t:

[0108] γ t =e -δt ;

[0109] Where δ is the diffusion rate control constant, is the cumulative diffusion attenuation factor up to stage t, φ ψ (X t ,E t ) is the self-supervised denoising feature estimation function implemented by the multi-head self-attention neural network based on the Transformer structure, where ψ is the network model parameter set, E t is the Transformer position encoding matrix at stage t;

[0110] The formula describes the core computational principle of the dynamic reverse diffusion denoising iterative process in the present invention. By iteratively updating the feature data matrix stage by stage, refined denoising and classification of interference signals are achieved. The formula takes the spectrum data matrix of the tth stage as input, and uses the self-supervised denoising feature estimation function implemented by the Transformer multi-head self-attention neural network to estimate and eliminate the noise component in the input data. Then, the diffusion intensity adjustment factor and the random noise amplitude adjustment factor are used to dynamically adjust the data state through reverse diffusion. Through such a reverse diffusion denoising mechanism, the feature purity of the data is gradually improved, thereby significantly improving the accuracy and stability of spectrum interference identification, ensuring that the final output spectrum data has lower noise interference and higher feature clarity.

[0111] λ t is the random noise amplitude adjustment factor:

[0112]

[0113] Among them, W t is a random noise matrix that obeys the standard Gaussian distribution N(0,I), and X t The dimensions are the same;

[0114] S24, calculating the reconstruction error between the data of adjacent diffusion stages based on the back-diffusion denoising output data of each stage, and obtaining the reconstruction error value of each stage;

[0115] S25. Using the reconstruction error as the back-propagation error signal, the neural network model parameter ψ is dynamically updated and optimized based on the adaptive moment estimation optimization algorithm:

[0116]

[0117] Among them, μ t represents the dynamic learning rate parameter, which is defined as an adaptive learning rate that decreases gradually with the number of training stages t:

[0118]

[0119] Among them, μ0 is the initial learning rate, κ is the learning rate decay exponent;

[0120] is the first moment estimate of the gradient:

[0121]

[0122] is the second moment estimate of the gradient:

[0123]

[0124] Among them, ξ1,ξ2 are the gradient moment estimation attenuation coefficients, ρ is a small constant to prevent division by zero operation, is the model parameter ψ relative to the reconstruction error L at stage t t gradient;

[0125] S26 , repeatedly executing steps S23 to S25 until all T stages of the reverse diffusion iteration process are completed, and the final reverse diffusion iteration output data is used as the purified spectrum data.

[0126] The present invention can effectively improve the initial feature expression quality of digitized spectrum data by combining the random Gaussian diffusion mechanism with the convolutional autoencoder; based on the dynamic back diffusion denoising process of the Markov chain, the multi-head self-attention network with the Transformer structure is used for noise estimation, which achieves more accurate noise feature estimation and effectively improves the accuracy of feature denoising processing; further, the model parameters are dynamically optimized and updated through the adaptive moment estimation algorithm based on the reconstruction error, which effectively ensures the stability and convergence of the diffusion denoising process, so that the present invention can obtain higher quality purified spectrum data and significantly improve the classification and recognition performance of interference signals.

[0127] In this embodiment, S3 specifically includes:

[0128] S31, dividing the purified spectrum data set into a mainstream interference data set, a marginal interference data set, and a small sample interference data set, wherein the marginal interference data set and the small sample interference data set are respectively determined by a preset interference category occurrence frequency threshold and an interference category sample quantity threshold;

[0129] S32, performing time-domain energy distribution statistical processing, frequency-domain power spectrum density calculation processing, and short-time Fourier transform processing on the spectrum data in the edge-type interference data set and the small-sample-type interference data set respectively to obtain the time-frequency domain amplitude-phase joint statistical distribution characteristics, and construct the initial spectrum feature vectors of the respective data sets;

[0130] S33, performing random time domain window position offset processing, frequency domain feature scale random scaling processing, and time-frequency domain feature rotation transformation processing on the initial spectrum feature vector of the edge type interference data set, to obtain multiple enhanced edge type interference spectrum feature vectors;

[0131] S34, performing spectrum amplitude random perturbation processing, spectrum power random energy redistribution processing, and spectrum phase random perturbation processing on the initial spectrum feature vector of the small sample type interference data set, respectively, to obtain multiple enhanced small sample type interference spectrum feature vectors;

[0132] S35. Based on the Transformer multi-head self-attention network, we construct an edge-type interference feature selection network and a small sample-type interference feature selection network, and input the enhanced feature vector and the original feature vector into their respective corresponding feature selection networks to obtain the attention weight of each feature vector.

[0133] S36. Based on the attention weights, perform feature-wise attention weighted fusion processing on the original feature vectors and enhanced feature vectors in the edge type interference dataset and the small sample type interference dataset, respectively, to generate edge type interference attention enhanced feature vectors and small sample type interference attention enhanced feature vectors;

[0134] S37. Merge the edge type interference attention enhancement feature vector, the small sample type interference attention enhancement feature vector, and the initial spectrum feature vector set in the mainstream type interference data set to obtain an expanded spectrum feature data set.

[0135] The present invention effectively expands the feature diversity of the data set by performing random perturbation and enhancement processing in the time domain, frequency domain and time-frequency domain on the edge type and small sample type interference data set; selects and weightedly fuses the original and enhanced features through the Transformer multi-head self-attention network, significantly improving the correlation and representativeness of the features, and effectively solving the problem of insufficient recognition of edge type and small sample type interference signals in the existing technology; thereby improving the generalization ability and classification accuracy of the interference signal recognition model in complex electromagnetic environments.

[0136] In this embodiment, the S34 specifically includes:

[0137] S341, using the initial spectrum feature vector of the small sample type interference data set as input data, performing spectrum amplitude random perturbation processing on each feature dimension, randomly selecting an amplitude perturbation scale factor to increase or decrease the amplitude of each feature dimension, and obtaining multiple spectrum feature vectors after amplitude perturbation;

[0138] S342. Performing a spectrum power random energy redistribution process on the multiple amplitude-perturbed spectrum feature vectors, with the constraint that the total amount of spectrum energy remains unchanged, randomly redistributing energy between the frequency points of the spectrum feature to obtain multiple spectrum feature vectors with differentiated energy distributions;

[0139] S343, performing a spectrum phase random perturbation process based on the multiple spectrum eigenvectors after energy redistribution, applying a random phase offset to the phase value of each frequency component of each eigenvector, and obtaining multiple spectrum eigenvectors after phase perturbation;

[0140] S344, respectively calculating the Euclidean distances between the plurality of feature vectors and the initial spectrum feature vector, and selecting the spectrum feature vectors whose Euclidean distances satisfy a preset feature disturbance threshold as effective enhanced candidate feature vectors;

[0141] S345. Calculate the feature correlation score between each valid enhanced candidate feature vector and the initial spectrum feature vector using the Transformer multi-head self-attention network, sort the feature vectors from high to low according to the feature correlation score, and select a preset number of enhanced feature vectors as the final enhanced spectrum feature vectors.

[0142] S346 , merging the finally selected enhanced spectrum feature vector with the initial spectrum feature vector in the small sample type interference data set to form an enhanced small sample type interference spectrum feature data set.

[0143] The present invention effectively improves the feature diversity and representativeness of small sample data by adopting random perturbation of spectral amplitude, random energy redistribution of spectral power and random perturbation of spectral phase for small sample type interference data sets; further utilizes the Transformer multi-head self-attention mechanism to screen and fuse enhanced features, so that the generated enhanced feature vector has higher correlation and consistency with the original data, thereby significantly improving the recognition accuracy of small sample type interference signals and the generalization performance of the model, and effectively solving the problem of insufficient recognition accuracy of small sample type interference by traditional technologies.

[0144] In this embodiment, the S4 specifically includes:

[0145] S41, dividing the expanded spectrum feature data set into multiple time series data segments according to three fixed time window lengths of different scales, to generate overlapping time series data segments of three scales;

[0146] S42, performing a short-time Fourier transform of a fixed length on each time series data segment within each scale, using a Hamming window as the window function, to generate multi-dimensional time-frequency domain feature representation data of different scales;

[0147] S43. Based on the Transformer position encoding rule, scale-time domain joint position encoding is added to the multi-dimensional time-frequency domain feature representation data in the three scales. Each position encoding consists of two parts.

[0148] S44. Using multi-dimensional time-frequency domain feature representation data containing scale-time domain joint position encoding as input data for the Transformer multi-head self-attention network, each self-attention head extracts the long-range dynamic correlation features of the spectral feature data in each scale in the time dimension and frequency dimension through the inner product attention calculation method, and outputs independent time-frequency feature encodings in multiple scales;

[0149] S45. Based on the time-frequency feature encoding output at each scale, cross-scale feature fusion is performed through a cross-scale Transformer multi-head self-attention interactive fusion network. The time-frequency feature encoding at each scale is simultaneously used as the query vector, key vector, and value vector of the self-attention network. The inner product attention is used to calculate the similarity correlation features between multiple scales to form a cross-scale time-frequency feature fusion encoding.

[0150] S46. The cross-scale time-frequency feature fusion encoding is input into a self-supervised two-layer feedforward neural network for nonlinear transformation. The input layer to the hidden layer and the hidden layer to the output layer of the feedforward neural network use linear mapping and ReLU activation function respectively to extract the deep correlation feature patterns of the spectrum interference signal in the time-frequency domain at different scales.

[0151] S47, performing a layer normalization operation on each feature dimension on the time-frequency feature coding that has undergone nonlinear transformation of the feedforward neural network, and adjusting the normalized coding data to interference signal feature coding data of a unified fixed dimension through a unified linear mapping process.

[0152] The present invention realizes the effective representation of data in the time domain and frequency domain by performing multi-scale overlapping window segmentation and short-time Fourier transform processing on the expanded spectral feature data; further based on the multi-head self-attention network and scale-time domain joint position coding of the Transformer structure, it deeply explores the long-range dynamic correlation characteristics of spectral data in the time domain and frequency domain, and adopts a cross-scale Transformer multi-head self-attention interactive fusion network to realize the effective fusion of features of different scales, which significantly improves the representativeness and robustness of spectral features; finally, the feedforward neural network and layer normalization technology are used for nonlinear transformation and standardization processing, which makes the feature encoding of spectral interference signals more stable and accurate, effectively overcoming the problem that traditional methods are insufficient in identifying the characteristics of interference signals in complex electromagnetic environments.

[0153] In this embodiment, the S5 specifically includes:

[0154] S51, dividing the interference signal characteristic coded data into multiple characteristic data batches;

[0155] S52, performing unsupervised clustering processing on each batch of feature data using a density peak clustering algorithm, and determining the cluster center within each batch based on the feature space density distribution;

[0156] S53. Taking the cluster center as the benchmark, the attention-enhanced self-supervised learning algorithm based on the Transformer multi-head self-attention structure is used to calculate the feature space similarity between the feature encoding data vector in each batch and the corresponding cluster center, and obtain the attention similarity coefficient of each feature encoding data vector;

[0157] S54, performing normalization processing based on the relative magnitude relationship of the attention similarity coefficients, and automatically generating a self-supervised soft label vector with a dimension corresponding to the number of cluster centers for each feature encoding data vector;

[0158] S55. For the soft label vector, a mutual information calculation method is used to perform a self-supervised consistency check, and the mutual information value between each feature encoding data vector and the corresponding soft label vector is calculated respectively, and the data pairs with a mutual information value greater than a preset threshold are selected as valid paired data;

[0159] S56. Construct an initial paired data set based on the valid paired data, and use the initial paired data set to perform dynamic gradient optimization update of the self-supervised network parameters;

[0160] S57. Loop through steps S52 to S56 until the loss function value of the attention-enhanced self-supervised learning algorithm is lower than the convergence threshold for multiple consecutive iterations, and then stop the iteration. The data set output from the last iteration is determined as the final paired data set that matches the interference signal feature encoding data with the self-supervised soft label.

[0161] The present invention uses a density peak clustering algorithm to perform unsupervised clustering processing on batches of feature data, and quickly and accurately determines the cluster centers within each batch; further, based on the attention-enhanced self-supervised learning algorithm of the Transformer multi-head self-attention structure, it accurately calculates the attention similarity coefficient between the feature data and the cluster center, and automatically generates self-supervised soft labels based on the coefficient; at the same time, through the mutual information consistency test, it effectively ensures the consistency of the quality of the soft label with the feature data, thereby constructing a high-quality paired data set, and using the data set to dynamically update and optimize the model parameters, effectively improving the model's ability to automatically recognize and classify unknown types of interference signals, and solving the problem of insufficient recognition of unknown interference signals by traditional methods.

[0162] In this embodiment, S6 specifically includes:

[0163] S61, each feature encoding data vector and the corresponding self-supervised soft label vector in the paired data set are combined to form initial high-dimensional feature state data, and input into the diffusion probability model for initial probability state mapping;

[0164] S62. Construct a multi-stage dynamic reverse diffusion iterative process based on a Markov chain within the diffusion probability model, define the probability transfer relationship between each reverse diffusion stage, and use a preset dynamic state transition probability matrix to perform reverse state migration on the characteristic state data stage by stage;

[0165] S63. In each back-diffusion stage, the Transformer multi-head self-attention structure is used to calculate the conditional probability distribution between the feature state data of the current stage and the associated self-supervised soft label vector, and obtain a refined conditional probability representation of the feature state data of the current stage;

[0166] S64. Based on the refined conditional probability representation of the characteristic state data obtained in each back-diffusion stage, a deep neural network is used to perform probability space noise residual estimation of the current stage, and the back-diffusion state of the characteristic state data is dynamically corrected using the residual estimation result;

[0167] S65. Based on the dynamically corrected feature state data, perform state inverse mapping stage by stage, gradually reduce the probability space noise through the probability transfer inverse update mechanism, and gradually approach the pure feature state data in the feature space;

[0168] S66. At the end of the reverse diffusion iterative process, the multi-head self-attention fusion mechanism is used to integrate the intermediate state data of all diffusion stages to construct the final fused pure feature state data, and obtain the pure state feature representation of the interference signal in the high-dimensional probability space;

[0169] S67. Input the pure state feature representation into the pre-trained linear multi-classifier, perform fine-grained classification and recognition on the pure state feature in the high-dimensional probability space based on the classification decision boundary defined by the multi-classifier, and output the final interference classification result.

[0170] The present invention effectively realizes the accurate conditional probability representation of high-dimensional feature state data and self-supervised soft labels through a multi-stage dynamic back-diffusion probability model based on Markov chain and a Transformer multi-head self-attention structure; further combines deep neural networks to perform refined estimation of probability space noise residuals and dynamic correction of back-diffusion, significantly improving the stability and accuracy of the diffusion iteration process; at the same time, through the multi-head self-attention fusion mechanism, the intermediate state data of multiple diffusion stages are integrated to obtain high-quality pure feature state data, effectively improving the fine-grained classification performance of interference signals in high-dimensional probability space, overcoming the problems of insufficient classification accuracy and stability of traditional methods, and realizing accurate identification and classification of radio interference signals in complex electromagnetic environments.

[0171] In this embodiment, the S7 specifically includes:

[0172] S71. Using the interference classification results as initial feedback data, a bidirectional attention interaction feedback network is constructed between the feedback data and the Transformer structure.

[0173] S72, using the feature vector corresponding to each interference category in the interference classification result as input data, respectively calculating the attention weight value between the feature vector and the interference signal feature encoding data output by the Transformer structure, and obtaining the association strength between the interference category and the feature encoding data;

[0174] S73. Representing the correlation strength between the interference category and the feature encoding data in the form of an attention weight matrix, obtaining feedback feature adjustment weights through multi-head self-attention calculation, and dynamically adjusting the parameters of the multi-head self-attention network in the feature extraction stage of the Transformer structure in real time;

[0175] S74, based on the real-time adjusted Transformer multi-head self-attention network parameters, re-execute the feature extraction process within the Transformer structure to generate updated feature encoding data;

[0176] S75, re-inputting the updated feature coding data into the diffusion probability model, executing the multi-stage dynamic reverse diffusion iterative process of the diffusion probability model, and generating an updated interference classification result;

[0177] S76, calculating a classification difference index between the updated interference classification result and the previous interference classification result based on the degree of category difference between the two classification results. When the classification difference index does not reach a preset stability threshold, continue to execute steps S72 to S75;

[0178] S77 , repeat steps S72 to S76 until the classification difference index reaches a stable threshold, output a stable interference classification result, and form a closed-loop dynamic feedback mechanism.

[0179] By constructing a bidirectional attention interactive feedback network between the interference classification results and the Transformer structure, the present invention can calculate the precise correlation strength between the classification results and the feature coding data in real time, thereby realizing the dynamic adjustment of the multi-head self-attention network parameters of the Transformer structure; and continuously updating the feature coding data and the interference classification results through the dynamic back-diffusion iteration mechanism of the diffusion probability model, effectively ensuring the real-time adaptability of the interference classification process; further, the termination condition of the closed-loop feedback is determined by analyzing the stability index of the classification results, thereby realizing efficient and stable convergence of the interference classification results, overcoming the problem of insufficient classification performance caused by the lack of a dynamic feedback mechanism in traditional methods, and significantly improving the accuracy and stability of radio interference signal recognition and classification in complex electromagnetic environments.

[0180] Example 1:

[0181] In order to verify the feasibility and practical application effect of the present invention, we selected a large-scale communication service enterprise with a complex electromagnetic environment as the application scenario of the present invention. This enterprise has long provided radio communication services to users in various industries and faced the problem of difficulty in identifying spectrum interference in complex electromagnetic environments. Recently, with the rapid expansion of wireless equipment and the increasing complexity of the electromagnetic environment, traditional spectrum monitoring and interference identification methods have gradually become difficult to meet the company's high-precision and high-automation identification requirements, especially difficult to effectively deal with weak interference, unknown type interference and small sample type interference signals. To solve these problems, the company decided to deploy the radio interference identification method based on electromagnetic spectrum monitoring proposed in this invention.

[0182] The company first deployed broadband spectrum monitoring equipment in its core business areas to collect real-time raw spectrum data in complex electromagnetic environments. To ensure data quality, the equipment is equipped with a built-in ultra-wideband antenna array and a low-insertion-loss dielectric resonant bandpass filter. This effectively captures high-quality analog electromagnetic spectrum signals. These signals are then converted to analog intermediate frequency (IF) signals through a low-noise, high-linearity RF preamplifier and a multi-channel RF downconversion circuit. These IF signals undergo dual-channel interleaved high-speed analog-to-digital conversion and digital phase calibration technology to generate a continuous, high-quality digitized raw spectrum dataset.

[0183] The company applies a diffusion probability model to automatically denoise the real-time spectrum data, effectively removing ambient noise and non-interfering signals, resulting in purified, high-quality spectrum data. Furthermore, to better identify rare edge-type interference signals and small-sample interference signals, the system further employs attention-enhanced data synthesis and expansion technology. This technology uses random amplitude perturbations, random energy redistributions, and random phase perturbations to generate a diverse, high-quality, expanded spectrum feature dataset.

[0184] Based on the expanded spectral feature data set, the company adopted the multi-scale cross-time window Transformer structure proposed in this invention to deeply extract the long-range dynamic correlation characteristics of interference signals in the time and frequency domains in the spectral data. Through the Transformer multi-head self-attention network and scale-time domain joint position encoding, the key features of interference signals in complex electromagnetic environments are effectively captured, and interference signal feature encoding data is generated. Subsequently, using the attention-enhanced self-supervised learning algorithm, based on the cluster center determined by density peak clustering, self-supervised soft labels are automatically generated for unknown types of interference signals, realizing automatic matching of feature encoding data and soft labels.

[0185] To further improve the accuracy of interference signal classification, the company fed feature encoding data and soft label pairing datasets into a dynamic backdiffusion probability model. Through multi-stage backdiffusion iterations, the accuracy of interference signal classification was significantly improved. During the classification process, the system leverages the bidirectional attention interaction feedback network between the Transformer structure and the classification results, dynamically adjusting the Transformer parameters in real time to continuously optimize feature extraction and classification accuracy, ultimately achieving accurate classification and stable output of interference signals.

[0186] Before and after the deployment of this embodiment, the enterprise conducted rigorous statistics and comparisons on the radio interference identification performance indicators for three consecutive months. The specific statistical data are shown in Table 1:

[0187] Table 1 Comparison of spectrum interference recognition performance before and after implementation of the present invention

[0188] Performance indicators Traditional method (before implementation) The method of the present invention (after implementation) Average recognition accuracy of interference signals 81.2％ 97.4％ Weak interference signal recognition accuracy 65.7％ 93.8％ Edge type interference recognition accuracy 58.3％ 92.5％ Accuracy of small sample interference recognition 54.9％ 91.7％ Automatic recognition rate of unknown interference signals 39.2％ 89.5％ Average response time for interference signal classification 42 minutes 5 minutes Interference signal false alarm rate 9.6％ 1.8％ Automatic labeling accuracy 32.7％ 90.1％

[0189] It can be clearly seen from Table 1 that after deploying the method of the present invention, the average recognition accuracy of the company's interference signals increased from the original 81.2% to 97.4%. In particular, the recognition accuracy of weak interference signals, edge-type interference signals and small-sample interference signals that are difficult to identify with traditional methods has been greatly improved. Among them, the recognition accuracy of weak interference signals has increased by 28.1%, the accuracy of edge-type interference signals has increased by 34.2%, and the accuracy of small-sample type interference signals has increased by 36.8%. In addition, the present invention also performs outstandingly in the automatic recognition of unknown types of interference signals, with the accuracy rate increased from 39.2% to 89.5%, greatly enhancing the system's adaptability to new types of unknown interference signals. At the same time, because the system has an automated self-supervised soft label annotation and feedback mechanism, the accuracy of automated annotation has also increased from 32.7% before implementation to 90.1%, significantly reducing the workload of manual intervention and manual annotation.

[0190] Furthermore, after implementing the method, the average response time for the enterprise's radio interference signal classification dropped from 42 minutes to just 5 minutes, significantly improving real-time response and interference suppression capabilities, effectively ensuring the stability and security of the enterprise's wireless communication services. Simultaneously, the false alarm rate for interference signals was significantly reduced from 9.6% to 1.8%, significantly reducing unnecessary manual review and processing work caused by false alarms and greatly improving the efficiency and reliability of spectrum monitoring.

[0191] Through the practical application of this embodiment, the high precision, high adaptability and high automation advantages of the electromagnetic spectrum monitoring radio interference identification method based on Transformer and attention-enhanced self-supervised diffusion probability model proposed in the present invention in complex electromagnetic environments are fully verified. It effectively solves the problems of insufficient recognition of edge-type interference, small sample interference and unknown interference signals, low accuracy and slow response speed of traditional technical solutions, and provides reliable and efficient technical guarantee for spectrum monitoring and interference management in the field of modern radio communications.

[0192] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A radio interference identification method based on electromagnetic spectrum monitoring, characterized in that: The steps include: S1. Real-time collection of raw spectrum data in a complex electromagnetic environment using broadband spectrum monitoring equipment, and digital sampling and processing of the raw spectrum data in the time and frequency domains to obtain digital raw spectrum data; S2. Perform unsupervised denoising on the digitized raw spectrum data using a self-supervised diffusion probability denoising model, automatically identify and remove environmental noise and non-interfering irrelevant signals, and output purified spectrum data; S3. For the purified spectrum data, the attention-enhanced data synthesis and expansion technology is used to automatically generate an expanded spectrum feature dataset of edge-type interference signals and small-sample interference signals; S4. Based on the multi-scale cross-time window Transformer structure, the multi-dimensional time-frequency domain long-range dynamic correlation features of the spectrum interference signal in the expanded spectrum feature data are extracted to obtain the interference signal feature coding data; S5. Using the attention-enhanced self-supervised learning algorithm, based on the interference signal feature coding data, an unsupervised interference feature learning network is constructed to automatically generate self-supervised soft labels corresponding to unknown interference types, and a paired dataset is obtained that matches the feature coding data with the soft labels; S6. Input the paired data set into the diffusion probability model, construct a multi-stage dynamic back-diffusion iterative process, classify the interference signal at a fine granularity in the high-dimensional probability space, and output the interference classification result; S7. Based on the interference classification results, a bidirectional attention interaction feedback network is constructed between the Transformer structure to form a closed-loop dynamic feedback mechanism.

2. The radio interference identification method based on electromagnetic spectrum monitoring according to claim 1, characterized in that: Said S1 specifically includes: S11. Utilize a built-in ultra-wideband antenna array to capture broadband analog electromagnetic spectrum signals in a complex electromagnetic environment in real time using spatial diversity, and utilize a low insertion loss dielectric resonant bandpass filter to perform frequency band selection and out-of-band suppression on the analog electromagnetic spectrum signals to obtain a bandpass filtered analog electromagnetic spectrum signal. S12, inputting the analog electromagnetic spectrum signal after bandpass filtering into a cascaded low-noise and high-linearity radio frequency preamplifier and a multi-channel radio frequency down-conversion circuit in sequence, performing radio frequency low-noise amplification and tuning processing, and converting the analog electromagnetic spectrum signal into an analog intermediate frequency signal in a predefined intermediate frequency range; S13. Inputting the analog intermediate frequency signal into a high-speed analog-to-digital converter with a dual-channel interleaved sampling architecture, alternately completing analog-to-digital conversion using a dual-channel interleaved analog-to-digital conversion method, wherein the sampling rate of the analog-to-digital converter is set to be at least twice the bandwidth specified by the Nyquist sampling theorem, and calibrating the inconsistency error between the interleaved channels using a digital phase calibration method to obtain an initial digitized spectrum data sequence; S14, inputting the initial digitized spectrum data sequence into a digital multi-stage cascade decimation filter, performing decimation, digital down-conversion, and anti-aliasing filtering operations in sequence, wherein the decimation multiple is a preset fixed integer value, the center frequency of the digital down-conversion is zero Hz, and outputting zero intermediate frequency baseband digital spectrum data; S15, performing time domain sliding window segmentation processing on the baseband digital spectrum data, where the length of the sliding window is fixed, and the sliding step is set to half of the window length, to generate an overlapping data frame sequence; S16. Performing a short-time Fourier transform on each data frame in the overlapping data frame sequence using a window function set to a Hamming window, with the window length fixed at 1024 sampling points and the overlap ratio between windows set to 50%. After each data frame is short-time Fourier transformed, a time-frequency matrix of digitized original spectrum data with a fixed spectrum resolution is obtained; S17. Splicing the digitized original spectrum data time-frequency matrix into data frames according to a strict time sequence to form a complete and continuous digitized original spectrum data set.

3. The radio interference identification method based on electromagnetic spectrum monitoring according to claim 1, characterized in that: The S2 specifically includes: S21, using the digitized original spectrum dataset as input data, and gradually performing a multi-scale random Gaussian noise superposition operation for a total of T stages on the input data based on a random Gaussian diffusion forward processing mechanism to obtain a data sequence after diffusion processing, where T is the total number of diffusion stages; S22. Use a pre-trained unsupervised convolutional autoencoder network to perform initial feature encoding on the data sequence after diffusion processing. The convolutional autoencoder network structure adopts a symmetrical multi-layer convolution-transposed convolution stacked architecture. The convolution kernel sizes of the convolution layers are set to a combination of 3×3, 5×5, and 7×7, respectively. The activation function uses a linear rectification function to obtain the initial potential feature data representation; S23, using the initial latent feature data representation as the initial value, constructing a T-stage dynamic back-diffusion denoising iterative process based on the Markov chain; S26 , repeatedly executing steps S23 to S25 until all T stages of the reverse diffusion iteration process are completed, and the final reverse diffusion iteration output data is used as the purified spectrum data.

4. The radio interference identification method based on electromagnetic spectrum monitoring according to claim 1, characterized in that: The S3 specifically includes: S31, dividing the purified spectrum data set into a mainstream interference data set, a marginal interference data set, and a small sample interference data set, wherein the marginal interference data set and the small sample interference data set are respectively determined by a preset interference category occurrence frequency threshold and an interference category sample quantity threshold; S32, performing time-domain energy distribution statistical processing, frequency-domain power spectrum density calculation processing, and short-time Fourier transform processing on the spectrum data in the edge-type interference data set and the small-sample-type interference data set respectively to obtain the time-frequency domain amplitude-phase joint statistical distribution characteristics, and construct the initial spectrum feature vectors of the respective data sets; S33, performing random time domain window position offset processing, frequency domain feature scale random scaling processing, and time-frequency domain feature rotation transformation processing on the initial spectrum feature vector of the edge type interference data set, to obtain multiple enhanced edge type interference spectrum feature vectors; S34, performing spectrum amplitude random perturbation processing, spectrum power random energy redistribution processing, and spectrum phase random perturbation processing on the initial spectrum feature vector of the small sample type interference data set, respectively, to obtain multiple enhanced small sample type interference spectrum feature vectors; S35. Based on the Transformer multi-head self-attention network, we construct an edge-type interference feature selection network and a small sample-type interference feature selection network, and input the enhanced feature vector and the original feature vector into their respective corresponding feature selection networks to obtain the attention weight of each feature vector. S36. Based on the attention weights, perform feature-wise attention weighted fusion processing on the original feature vectors and enhanced feature vectors in the edge type interference dataset and the small sample type interference dataset, respectively, to generate edge type interference attention enhanced feature vectors and small sample type interference attention enhanced feature vectors; S37. Merge the edge type interference attention enhancement feature vector, the small sample type interference attention enhancement feature vector, and the initial spectrum feature vector set in the mainstream type interference data set to obtain an expanded spectrum feature data set.

5. The radio interference identification method based on electromagnetic spectrum monitoring according to claim 4 is characterized in that: The S34 specifically includes: S341, using the initial spectrum feature vector of the small sample type interference data set as input data, performing spectrum amplitude random perturbation processing on each feature dimension, randomly selecting an amplitude perturbation scale factor to increase or decrease the amplitude of each feature dimension, and obtaining multiple spectrum feature vectors after amplitude perturbation; S342. Performing a spectrum power random energy redistribution process on the multiple amplitude-perturbed spectrum feature vectors, with the constraint that the total amount of spectrum energy remains unchanged, randomly redistributing energy between the frequency points of the spectrum feature to obtain multiple spectrum feature vectors with differentiated energy distributions; S343, performing a spectrum phase random perturbation process based on the multiple spectrum eigenvectors after energy redistribution, applying a random phase offset to the phase value of each frequency component of each eigenvector, and obtaining multiple spectrum eigenvectors after phase perturbation; S344, respectively calculating the Euclidean distances between the plurality of feature vectors and the initial spectrum feature vector, and selecting the spectrum feature vectors whose Euclidean distances satisfy a preset feature disturbance threshold as effective enhanced candidate feature vectors; S345. Calculate the feature correlation score between each valid enhanced candidate feature vector and the initial spectrum feature vector using the Transformer multi-head self-attention network, sort the feature vectors from high to low according to the feature correlation score, and select a preset number of enhanced feature vectors as the final enhanced spectrum feature vectors. S346 , merging the finally selected enhanced spectrum feature vector with the initial spectrum feature vector in the small sample type interference data set to form an enhanced small sample type interference spectrum feature data set.

6. The radio interference identification method based on electromagnetic spectrum monitoring according to claim 1, characterized in that: The S4 specifically includes: S41, dividing the expanded spectrum feature data set into multiple time series data segments according to three fixed time window lengths of different scales, to generate overlapping time series data segments of three scales; S42, performing a short-time Fourier transform of a fixed length on each time series data segment within each scale, using a Hamming window as the window function, to generate multi-dimensional time-frequency domain feature representation data of different scales; S43. Based on the Transformer position encoding rule, scale-time domain joint position encoding is added to the multi-dimensional time-frequency domain feature representation data in the three scales. Each position encoding consists of two parts. S44. Using multi-dimensional time-frequency domain feature representation data containing scale-time domain joint position encoding as input data for the Transformer multi-head self-attention network, each self-attention head extracts the long-range dynamic correlation features of the spectral feature data in each scale in the time dimension and frequency dimension through the inner product attention calculation method, and outputs independent time-frequency feature encodings in multiple scales; S45. Based on the time-frequency feature encoding output at each scale, cross-scale feature fusion is performed through a cross-scale Transformer multi-head self-attention interactive fusion network. The time-frequency feature encoding at each scale is simultaneously used as the query vector, key vector, and value vector of the self-attention network. The inner product attention is used to calculate the similarity correlation features between multiple scales to form a cross-scale time-frequency feature fusion encoding. S46. The cross-scale time-frequency feature fusion encoding is input into a self-supervised two-layer feedforward neural network for nonlinear transformation. The input layer to the hidden layer and the hidden layer to the output layer of the feedforward neural network use linear mapping and ReLU activation function respectively to extract the deep correlation feature patterns of the spectrum interference signal in the time-frequency domain at different scales. S47, performing a layer normalization operation on each feature dimension on the time-frequency feature coding that has undergone nonlinear transformation of the feedforward neural network, and adjusting the normalized coding data to interference signal feature coding data of a unified fixed dimension through a unified linear mapping process.

7. The method for identifying radio interference based on electromagnetic spectrum monitoring according to claim 1, characterized in that: The S5 specifically includes: S51, dividing the interference signal characteristic coded data into multiple characteristic data batches; S52, performing unsupervised clustering processing on each batch of feature data using a density peak clustering algorithm, and determining the cluster center within each batch based on the feature space density distribution; S53. Taking the cluster center as the benchmark, the attention-enhanced self-supervised learning algorithm based on the Transformer multi-head self-attention structure is used to calculate the feature space similarity between the feature encoding data vector in each batch and the corresponding cluster center, and obtain the attention similarity coefficient of each feature encoding data vector; S54, performing normalization processing based on the relative magnitude relationship of the attention similarity coefficients, and automatically generating a self-supervised soft label vector with a dimension corresponding to the number of cluster centers for each feature encoding data vector; S55. For the soft label vector, a mutual information calculation method is used to perform a self-supervised consistency check, and the mutual information value between each feature encoding data vector and the corresponding soft label vector is calculated respectively, and the data pairs with a mutual information value greater than a preset threshold are selected as valid paired data; S56. Construct an initial paired data set based on the valid paired data, and use the initial paired data set to perform dynamic gradient optimization update of the self-supervised network parameters; S57. Loop through steps S52 to S56 until the loss function value of the attention-enhanced self-supervised learning algorithm is lower than the convergence threshold for multiple consecutive iterations, and then stop the iteration. The data set output from the last iteration is determined as the final paired data set that matches the interference signal feature encoding data with the self-supervised soft label.

8. The radio interference identification method based on electromagnetic spectrum monitoring according to claim 1, characterized in that: The S6 specifically includes: S61, each feature encoding data vector and the corresponding self-supervised soft label vector in the paired data set are combined to form initial high-dimensional feature state data, and input into the diffusion probability model for initial probability state mapping; S62. Construct a multi-stage dynamic reverse diffusion iterative process based on a Markov chain within the diffusion probability model, define the probability transfer relationship between each reverse diffusion stage, and use a preset dynamic state transition probability matrix to perform reverse state migration on the characteristic state data stage by stage; S63. In each back-diffusion stage, the Transformer multi-head self-attention structure is used to calculate the conditional probability distribution between the feature state data of the current stage and the associated self-supervised soft label vector, and obtain a refined conditional probability representation of the feature state data of the current stage; S64. Based on the refined conditional probability representation of the characteristic state data obtained in each back-diffusion stage, a deep neural network is used to perform probability space noise residual estimation of the current stage, and the back-diffusion state of the characteristic state data is dynamically corrected using the residual estimation result; S65. Based on the dynamically corrected feature state data, perform state inverse mapping stage by stage, gradually reduce the probability space noise through the probability transfer inverse update mechanism, and gradually approach the pure feature state data in the feature space; S66. At the end of the reverse diffusion iterative process, the multi-head self-attention fusion mechanism is used to integrate the intermediate state data of all diffusion stages to construct the final fused pure feature state data, and obtain the pure state feature representation of the interference signal in the high-dimensional probability space; S67. Input the pure state feature representation into the pre-trained linear multi-classifier, perform fine-grained classification and recognition on the pure state feature in the high-dimensional probability space based on the classification decision boundary defined by the multi-classifier, and output the final interference classification result.

9. The radio interference identification method based on electromagnetic spectrum monitoring according to claim 1, characterized in that: The S7 specifically includes: S71. Using the interference classification results as initial feedback data, a bidirectional attention interaction feedback network is constructed between the feedback data and the Transformer structure. S72, using the feature vector corresponding to each interference category in the interference classification result as input data, respectively calculating the attention weight value between the feature vector and the interference signal feature encoding data output by the Transformer structure, and obtaining the association strength between the interference category and the feature encoding data; S73. Representing the correlation strength between the interference category and the feature encoding data in the form of an attention weight matrix, obtaining feedback feature adjustment weights through multi-head self-attention calculation, and dynamically adjusting the parameters of the multi-head self-attention network in the feature extraction stage of the Transformer structure in real time; S74, based on the real-time adjusted Transformer multi-head self-attention network parameters, re-execute the feature extraction process within the Transformer structure to generate updated feature encoding data; S75, re-inputting the updated feature coding data into the diffusion probability model, executing the multi-stage dynamic reverse diffusion iterative process of the diffusion probability model, and generating an updated interference classification result; S76, calculating a classification difference index between the updated interference classification result and the previous interference classification result based on the degree of category difference between the two classification results. When the classification difference index does not reach a preset stability threshold, continue to execute steps S72 to S75; S77 , repeat steps S72 to S76 until the classification difference index reaches a stable threshold, output a stable interference classification result, and form a closed-loop dynamic feedback mechanism.

Images