A specific type of OFDM signal detection and identification method based on multi-domain feature fusion

Abstract

A specific type OFDM signal detection and identification method based on multi-domain feature fusion is provided.The received OFDM signal data is subjected to high-order cumulant feature extraction and time-frequency domain feature extraction, respectively, to obtain corresponding high-order cumulant and time-frequency domain features, which are further fused to form a multi-domain feature fusion vector.The obtained multi-dimensional fusion feature vector is input into an improved Transform network, the input signal is subjected to position coding, parallel input is realized, the connection between data internals is extracted, and high-dimensional abstract features are obtained.The high-dimensional abstract features are input into a classifier to identify specific type OFDM signals in a blind background.The present application has excellent anti-noise performance, improves the detection and identification performance, and reduces the calculation amount and complexity.

Description

Technical Field

[0001] This invention relates to a method for detecting and recognizing specific OFDM signals based on multi-domain feature fusion, specifically a method for detecting and recognizing specific OFDM signals based on multi-domain feature fusion and an improved Transformer, belonging to the field of artificial intelligence signal processing. Background Technology

[0002] In recent years, low Earth orbit (LEO) satellite communication systems have developed rapidly and attracted much attention. Currently, emerging mega-constellation projects such as Starlink, OneWeb, Kuiper, and Telesat are flourishing, with SpaceX's Starlink project being the largest. It has successfully deployed nearly 6,000 satellites in low Earth orbit at an altitude of 500-600 km, boasting the largest planned network scale and the most launched satellites, and independently owning and controlling the entire industry chain. As such a high-density LEO constellation, the radio frequency signal coverage and security of LEO constellations have attracted widespread attention. While the Federal Communications Commission (FCC) license database publicly records information such as the orbital parameters of various LEO communication system satellites, the operating frequency bands and bandwidths of communication signals, it does not describe detailed signal characteristics such as frame structure, modulation methods, and coding methods. Therefore, it is necessary to detect and analyze the signals of such communication systems.

[0003] Based on existing work on processing and analyzing the aforementioned low-Earth orbit satellite communication signals, it can be concluded that the signals of such systems mainly consist of two parts: OFDM signals (Orthogonal Frequency Division Multiplexing signals) and some special signals customized according to their own systems, such as the single-tone signal in the virtual subcarrier part of Starlink signals located in the OFDM center, and the special pilot structure of Kuiper. Among them, OFDM signals account for the majority. At present, most of the methods for effective detection and identification of OFDM signals with special formats are modeled after the methods for detecting and identifying non-cooperative signals in OFDM systems. These methods mainly include the following types of techniques: The first type is based on high-order statistics (Research on Reception and Processing Technology of Non-cooperative OFDM Signals [D], Dong Xiaoke, University of Electronic Science and Technology of China, 2019). The basic idea is to use the Gaussian statistical characteristics of OFDM signals, and the special property that its high-order cumulative quantity (generally fourth order and above) is 0, to detect and identify the signal. The problem with this type of technique is that the detection threshold is easily affected by the environment and interference. The first category is the signal factor, which leads to performance degradation. The second category is the method based on the cyclostationary characteristics of the signal (Research on Key Technologies of Blind Demodulation of OFDM Signal [D], Chai Xinhui, Nanchang University 2023). Its basic idea is to utilize the fact that OFDM signal is a cyclostationary random process with periodic characteristics, while the statistical characteristics of noise are completely random. Signal detection is achieved based on the difference between the statistical characteristics of signal and noise. The problem with this type of technology is that it has a large amount of computation, low computational efficiency, and requires a large amount of data to achieve effective detection. The third category is the eigenvalue detection method (M.Wang, J.Zhang, "OFDM Signal Detection and Parameter Estimation Using Covariance Matrix", IEEE Communication Letters, vol.17, no.6, pp.1080-1083, June, 2013). Its basic idea is to use the characteristics and eigenvalues ​​of the received signal covariance matrix to carry out signal detection. Signal detection and recognition are performed based on the structure of OFDM signal. However, the problem with this type of technology is that it is easily affected by strong background noise and interference, which leads to significant performance degradation. Summary of the Invention

[0004] The technical problem solved by this invention is to overcome the shortcomings of the prior art and provide a method for detecting and recognizing specific OFDM signals based on multi-domain feature fusion. By combining the unique features of the signal's time and frequency domains with the powerful capabilities of deep learning, this method can achieve high-precision detection and recognition of specific OFDM signals in non-cooperative contexts.

[0005] The technical solution of this invention is: a method for detecting and recognizing specific OFDM signals based on multi-domain feature fusion, comprising:

[0006] The received OFDM signal data is subjected to high-order cumulant feature extraction and time-frequency domain feature extraction respectively to obtain the corresponding high-order cumulant and time-frequency domain features, which are then fused to form a multi-domain feature fusion vector.

[0007] The obtained multidimensional fused feature vector is input into the improved Transformer network to extract the internal relationships between the data in the input signal, thus obtaining high-dimensional abstract features;

[0008] High-dimensional abstract features are input into a classifier to identify custom OFDM signals in a blind background.

[0009] Preferably, higher-order cumulant feature extraction is performed on the received OFDM signal data. Specifically, the p-th and q-th order mixed moments of the OFDM signal x[n] are calculated using the moment generation function and the cumulant generation formula, and the higher-order cumulant feature vector C is further derived through formula derivation. 20 C 21 C 30 C 31 C 40 C 41 C 42 C 60 C 63}:

[0010] C 20 =M 20

[0011] C 21 =M 21

[0012] C 30 =M 30

[0013] C 31 =M 31

[0014]

[0015] C 41 =M 41 -3M 20 M 21

[0016]

[0017] Where: M p,q =E[x[n] p-q x * [n] q ], * denotes complex conjugation, p is the order, q is the number of sequences to take complex conjugation, and p>q; E[·] denotes the mean; p={2,3,4,6}, q={0,1,2,3}.

[0018] Preferably, the time-frequency domain features of the OFDM signal data are extracted by discrete orthogonal Stockwell transform, and the time-frequency domain features S are obtained as follows:

[0019] S=[S(0,τ0),S(1,τ1),…S(k,τ k ),…S(K-1,τ K-1 k = {0, 1, ..., K-1}

[0020] Specifically:

[0021]

[0022] Where X[f] is the Discrete Fourier Transform (DFT) of the OFDM signal x[n], the total length of the OFDM signal x[n] is N, k = {0, 1, ..., K-1} is the frequency band number when dividing the entire spectrum using the Dyadic structure, and K is the total number of frequency band segments, where the bandwidth of the k-th frequency band is β. k The center frequency is v k The time location index is τ k f represents the discrete frequency index; n represents the discrete time index; j is the standard imaginary unit.

[0023] Preferably, the multi-domain feature fusion vector fx is as follows:

[0024] fx = [C 20 C 21 C 30 C 31 C 40 C 41 C 42 C 60 C 63 ,S]

[0025] Among them, {C 20 C 21 C 30 C 31 C 40 C 41 C 42 C 60 C 63} represents the higher-order cumulant, and S represents the time-frequency domain feature; after normalizing fx, the normalized multidimensional fused feature vector fv is obtained.

[0026] The preferred, improved Transformer network comprises: a position encoding module and H cascaded sub-attention modules, where H ≥ 3;

[0027] After the multidimensional fused feature vector enters the improved Transformer network, it first passes through the position encoding module to obtain a position-encoded feature vector carrying position information.

[0028] Let alpha = {1, 2, ..., H} and be an integer. When alpha = 1, the input of the alpha-th sub-attention module is the feature vector after position encoding. When 2 ≤ alpha ≤ H-1, the input of the alpha-th sub-attention module is the feature vector output by the (alpha-1)-th sub-attention module. The H-th sub-attention module outputs the extracted high-dimensional abstract features.

[0029] Each sub-attention module includes a self-attention mechanism, a fully connected feedforward network, a first residual connection structure, a second residual connection structure, a first normalization layer structure, and a second normalization layer structure.

[0030] The attention mechanism extracts the relationships between data from the input feature vector, obtains the weighted aggregated features, and outputs them to the first residual connection structure.

[0031] The first residual connection structure extracts more abstract features from the weighted aggregated features and outputs the recognition result. After the recognition probability is processed by the first normalization structure, it is output to the feedforward neural network.

[0032] After mapping the recognition probability, the feedforward neural network outputs the feedforward recognition probability to the second residual connection structure;

[0033] The second residual connection structure extracts more abstract features from the feedforward recognition probability and outputs the recognition result. After being processed by the second normalization structure, it becomes a high-dimensional abstract feature, which is then used as the output of the sub-attention module.

[0034] Preferably, when the position encoding module performs position encoding on the normalized multidimensional fused feature vector fv, it specifically does so as follows:

[0035] x a =fv a +PE a

[0036] When a is an even number:

[0037] When a is an odd number:

[0038] Among them, fv a The element at position a in the normalized multi-domain feature fusion vector fv, with dimension d, where a = {0, 1, ..., d-1} is the position index, PE aLet g be the position encoding vector corresponding to position a, where g represents the semi-order index of the dimension, g = {0, 1, ..., d / 2} and rounded down; the final feature vector input to the first sub-attention module of the backend is X1 = {x1, x2, ..., x a+1 ,…,x d}

[0039] Preferably, after inputting the feature vector X1 into the attention mechanism, the output result of each position i in the weighted aggregated features is as follows:

[0040]

[0041] Where i represents the center position of the local window attention; j represents the center position of the global attention. Let α be the set of other locations that position i can focus on. ij The attention weights are calculated as follows:

[0042]

[0043] Q, K, and V are the query matrix, key matrix, and value matrix obtained by linear transformation of the input feature vector X1, respectively. k For the dimensions of Q and K, e ik Let w represent the scoring function, and w represent the position that position i can focus on.

[0044] Preferably, both the first residual connection structure and the second residual connection structure adopt the form of adding the input and the mapping function, and the mapping function adopts the convolution function or the activation function;

[0045] The first and second normalization structures adjust the input vector to a vector with a mean of 0 and a variance of 1.

[0046] Preferably, the feedforward neural network performs two linear transformations on the input vector LN, then uses the ReLU function to activate and map the first recognition probability, before outputting the feedforward recognition probability, specifically:

[0047] FN = max(0, LNW1+b1)W2+b2

[0048] Where W1,b1,W2,b2 are linear transformation coefficients, and FN represents the feedforward recognition probability.

[0049] Preferably, when high-dimensional abstract features are input into the classifier to identify custom OFDM signals in a blind background:

[0050] The classifier is a structure consisting of two convolutional layers, two fully connected layers, and a SoftMax activation function connected in sequence. The classifier classifies high-dimensional abstract features and obtains the signal category recognition probability result.

[0051] Compared with the prior art, the present invention has the following advantages:

[0052] This method avoids performance degradation caused by other factors affecting the detection threshold; this method

[0053] (1) This invention integrates time domain, frequency domain and statistical domain features, and uses deep networks to extract features with high robustness and high resolution, which can effectively improve the performance of detecting and recognizing customized OFDM signals, and can be applied to a wider range of scenarios, while also having excellent anti-noise performance.

[0054] (2) The improved Transformer architecture of this invention can reduce the computational overhead of the self-attention mechanism. Compared with the cyclic stationary method, it requires less data and has lower complexity, improving detection and recognition performance while reducing computational load and complexity. Attached Figure Description

[0055] Figure 1 This is a flowchart of the method of the present invention;

[0056] Figure 2 This is a schematic diagram of the improved Transformer network structure of the present invention;

[0057] Figure 3 This is a schematic diagram of the improved self-attention mechanism of the present invention;

[0058] Figure 4 This is a schematic diagram of the residual network structure of the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described embodiments are merely some embodiments of the invention, not all embodiments. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0060] This invention provides a method for detecting and recognizing specific OFDM signals based on multi-domain feature fusion and an improved Transformer, such as... Figure 1 As shown, it includes the following steps:

[0061] (1) Perform multi-domain feature extraction on the received OFDM signal data, including high-order cumulants and time-frequency features, and fuse them to form a multi-domain feature fusion vector;

[0062] i. Extraction of higher-order cumulants

[0063] The extraction of higher-order cumulants includes:

[0064] The raw OFDM signal data x(t) is acquired, and the continuous OFDM signal data is denoted as x(t). Using the moment generation function and the cumulant generation formula, the p-th and q-th order mixed moments of the random signal x[n] are:

[0065] M p,q =E[x[n] p-q x * [n] q ]

[0066] In the formula, * denotes complex conjugation, p is the order (p>q), and q is the number of sequences to be complex conjugated; E[·] denotes the mean. In this patent, p={2,3,4,6}, q={0,1,2,3}.

[0067] Define the bivariate moment generating function M(t1,t2) and the cumulant generating function K(t1,t2) of the complex signal as follows:

[0068]

[0069] The values ​​of the multiple partial derivatives of the cumulant generating function at (t1,t2) = (0,0) are the cumulants of each order. For example:

[0070]

[0071] Using the Taylor expansion of M(t1,t2) and K(t1,t2) and the Faa di Bruno formula, the above multiple partial derivatives can be mapped to a combination of p- and q-order mixed moments, thus obtaining the various higher-order cumulants.

[0072] The final higher-order cumulative is:

[0073] C 20 =M 20

[0074] C 21 =M 21

[0075] C 30 =M 30

[0076] C 31 =M 31

[0077]

[0078] C 41 =M 41 -3M 20 M 21

[0079]

[0080]

[0081] ii. Time-frequency domain feature extraction

[0082] The time-spectrum feature extraction includes:

[0083] Time-frequency domain features are extracted from the received signal in the time domain using the Discrete Orthogonal Stockwell Transform (DOST). DOST combines the absolute phase-preserving properties of the Short Time Fourier Transform (STFT) with the frequency correlation analysis and multi-resolution characteristics of the Discrete Wavelet Transform (DWT), making it a superior time-frequency domain transformation method.

[0084] Suppose that the received OFDM signal x[n] has a continuous Stockwell transform with frequency f as follows:

[0085]

[0086] Converted to discrete transformation as follows:

[0087]

[0088] Where X[f] is the Discrete Fourier Transform (DFT) of the signal, and the total length of the signal is N:

[0089]

[0090] The entire spectrum is divided into frequency bands using the Dyadic structure, where k = {0, 1, ..., K-1} is the frequency band number, K is the total number of frequency band segments, and the bandwidth of the k-th frequency band is β. k The center frequency is v k The time location index is τ k .

[0091] The specific division method is as follows:

[0092] First band: center frequency v0 = 0, width β0 = 1;

[0093] Second band: center frequency v1 = 1, width β1 = 1;

[0094] Starting from the second band, the frequency is allocated progressively as follows: Center frequency v p =2 p-1 Width β p =2 p-1 Time position τ p =0,1,…,β p -1, where the frequency points included in the p-th frequency band are:

[0095] DOST is a non-redundant and efficient version of the continuous Stockwell transform, solving the problems of high computational cost and complexity. DOST reduces redundancy by dividing the output of the continuous Stockwell transform into K regions, each represented by a sparse representation. DOST can be viewed as a set of orthogonal basis functions that yield the Fourier spectrum of the signal while preserving the effective phase properties of the continuous transform. The coefficients of DOST can be considered as inner products between the sampled signal and a set of orthogonal basis functions. The Stockwell basis functions are defined as phase-modulated synthesis of the DFT basis over a frequency band in the frequency domain.

[0096]

[0097] Projecting the discrete signal x[n] onto the aforementioned basis functions yields the in-band spectral values ​​based on phase e. j2πβτf / N Weighted and normalized synthesis yields the final DOST transform output:

[0098]

[0099] The final output of DOST is S = [S(0,τ0),S(1,τ1),…S(k,τ). k ),…S(K-1,τ K-1 ]], k={0,1,…,K-1}, is an N×1 complex vector that effectively represents the characteristics of the time-frequency domain, and the elements of the vector are independent of each other. The main advantages of DOST are high computational efficiency, low memory requirements, high robustness to target noise, and effectiveness in analyzing signals with multiple components and compositions.

[0100] iii. Multidimensional feature fusion

[0101] By integrating the aforementioned time-frequency domain DOST vector S with nine parameters of higher-order statistics, a multidimensional fused feature vector fx = [C] is constructed. 20 C 21 C 30 C 31 C 40 C 41 C 42 C 60 C 63 [S], and after normalization, the normalized multidimensional fusion feature vector fv is obtained.

[0102] (2) The obtained multidimensional fused feature vector f is fed into the improved Transformer network to perform position encoding on the input signal, realizing parallel input of the input feature vector. This network extracts the internal relationships between data through a self-attention mechanism to obtain high-dimensional abstract features:

[0103] The multi-domain feature fusion vector f is fed into an improved Transformer network for high-dimensional feature extraction, including:

[0104] The improved Transformer network consists of a positional encoding module and H cascaded sub-attention modules, where H ≥ 3. After the multidimensional fused feature vector enters the improved Transformer network, it first passes through the positional encoding module to obtain the positionally encoded feature vector, which is then input to the improved attention mechanism in the first sub-attention module. The input of subsequent sub-attention modules is the output of the previous sub-attention module: Let alpha = {1, 2, ..., H}. When alpha = 1, the input of the alpha-th sub-attention module is the positionally encoded feature vector. When 2 ≤ alpha ≤ H-1, the input of the alpha-th sub-attention module is the feature vector output by the (alpha-1)-th sub-attention module. The H-th sub-attention module outputs the extracted high-dimensional feature.

[0105] Each sub-attention module consists of a self-attention mechanism, a fully connected feedforward network, two residual connection structures (first residual connection structure and second residual connection structure), and two layer normalization structures (first layer normalization structure and second layer normalization structure), as follows: Figure 2 As shown:

[0106] The improved attention mechanism extracts the relationships between data from the input feature vectors, obtains weighted aggregated features, and outputs them to the first residual connection structure.

[0107] The first residual connection structure extracts more abstract features from the weighted aggregated features and outputs the recognition result. After the recognition probability is processed by the first normalization structure, it is output to the feedforward neural network.

[0108] After mapping the recognition probability, the feedforward neural network outputs the feedforward recognition probability to the second residual connection structure;

[0109] The second residual connection structure extracts more abstract features from the feedforward recognition probability and outputs the recognition result. After being processed by the second normalization structure, it becomes a high-dimensional abstract feature, which is then used as the output of the sub-attention module.

[0110] i. Location encoding

[0111] For the input normalized multi-domain feature fusion vector fv, position encoding is performed to carry position information while maintaining the parallelism of the original data.

[0112] The location coding process is as follows:

[0113] x a =fv a +PE a

[0114]

[0115] Among them, fv a The element at position a in the normalized multi-domain feature fusion vector fv, with dimension d, where a = {0, 1, ..., d-1} is the position index, PE a Let PE be the position encoding vector corresponding to position i, where g represents the semi-order index of the dimension, g = {0, 1, ..., d / 2} and rounded down; according to the above formula, PE a All even dimensions use sin (i.e., sin when a is even), and odd dimensions use cos (i.e., cos when a is odd); the final feature vector input to the first sub-attention module of the backend is X1={x1,x2,…,x a+1 ,…,x d}

[0116] Position encoding ensures that the parallel input data retains the correlation between previous and subsequent data, and enhances long-distance learning capabilities.

[0117] ii. Self-attention mechanism

[0118] Self-attention mechanisms are an improvement on attention mechanisms, which do not rely on external information and are more likely to obtain connections between internal data.

[0119] This invention improves upon the high complexity of attention in traditional Transformers. It enables the attention matrix after linear transformation to perform sliding local attention only on the input data within a fixed-size window, while adding global attention at fixed points for specific tasks. To better extract useful feature information, a small sliding window is typically used at the bottom layer of the model to establish local relationships, and then the window size is gradually increased to expand the receptive field. For classification problems, this invention employs a sliding window attention mechanism that integrates global information, adding global attention only at the label level. The improved self-attention mechanism is visualized as follows: Figure 3 As shown.

[0120] The specific process is as follows: For the input vector X1 = {x1, x2, ..., x...} a+1 ,…,x d},

[0121] Local window attention: Each position i only pays attention to other positions within a window of size w surrounding it.

[0122]

[0123] Global attention: Enabled only at certain positions j, it can pay attention to all other positions.

[0124]

[0125] Where Q = X1W Q K = X1W K V = X1W V These are the query matrix, key matrix, and value matrix obtained by linear transformation of the input vector X1, respectively. k For the dimensions of Q and K.

[0126] Finally, after X passes through the improved self-attention mechanism, the output at each position i is:

[0127]

[0128] in, Let α be the set of other locations that can be of interest at location i. ij The attention weights are calculated as follows:

[0129]

[0130] e ik Let w represent the scoring function, and w represent the position that position i can focus on.

[0131] iii. Residual Network

[0132] Residual networks create shortcuts that allow shallow input information to directly participate in deeper layers. This residual connection increases network depth while preventing network degradation. Its structure is as follows: Figure 4 As shown.

[0133] When network degradation occurs, the output of the residual network equals 0. Then, by setting the overall mapping of the residual network modules equal to the feature mapping of the residual network input, the network can be brought to its optimal state. The formula for residual connections is shown below:

[0134] H(y)=F(y)+y

[0135] Where F(y) represents a mapping function, which is either a convolution or activation operation, y represents the input information, and attn is the output of the improved attention mechanism. iBy adding F(y) to obtain the output H(y), residual connections can directly pass the input information from shallow layers to deep layers, avoiding the gradient vanishing problem and thus improving the robustness and accuracy of the network.

[0136] iv. Layer normalization

[0137] Layer normalization calculates the mean and variance across all dimensions of a single input sample, transforming it into data with a mean of 0 and a variance of 1. This layer normalization stabilizes the distribution of each layer, allowing subsequent layers to learn from the earlier ones. The calculation formula is:

[0138]

[0139] In the formula, ρ,∈ are trainable parameters, μ L , Let be the mean and variance, and ε be a very small positive constant.

[0140] v. Feedforward Neural Network

[0141] After the data undergoes residual processing and normalization, it passes through a feedforward neural network. Two linear transformations are applied to the input LN, followed by activation using the ReLU function, as shown in the following formula:

[0142] FN = max(0, LNW1+b1)W2+b2

[0143] W1, b1, W2, b2 are linear transformation coefficients.

[0144] The multi-domain fusion feature vector is fed into H sequential network structures composed of the above modules, and finally outputs a high-dimensional abstract feature σ.

[0145] (3) The high-dimensional abstract feature σ is fed into a classifier consisting of convolutional layers and fully connected layers. After passing through two convolutional layers and two fully connected layers, the output of the fully connected layer is classified using the SoftMax activation function to obtain the signal category recognition probability result, thereby enabling the identification of specific OFDM signals in a blind background.

[0146] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

[0147] In the physical layer architecture of most current low-Earth orbit satellite transmission systems, there exists a unique, proprietary structure distinct from classic OFDM transmission systems such as LTE and 5G-NR. This structure can be used to achieve high-precision detection and identification of downlink signals. Furthermore, since the main component of this system is still OFDM, various features from existing traditional OFDM signal blind detection and identification methods can also be used as aids to further improve signal detection and identification performance. This invention combines the unique time-frequency domain features of signals with the powerful capabilities of deep learning to achieve high-precision detection and identification of specific OFDM signals in blind environments.

[0148] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A method for detecting and recognizing specific OFDM signals based on multi-domain feature fusion, characterized in that... include: The high-order cumulant is obtained by extracting high-order cumulant features from the received OFDM signal data. The time-frequency domain features are extracted from the OFDM signal data by discrete orthogonal Stockwell transform. The high-order cumulant and time-frequency domain features are further fused to form a multi-domain feature fusion vector. The obtained multidimensional fused feature vector is input into the improved Transformer network to extract the internal relationships between the data in the input signal, thus obtaining high-dimensional abstract features; The improved Transformer network includes: a position encoding module and A series of interconnected sub-attention modules, Each sub-attention module includes a self-attention mechanism, a fully connected feedforward network, a first residual connection structure, a second residual connection structure, a first-layer normalization structure, and a second-layer normalization structure. The self-attention mechanism adopts a sliding window attention mechanism that integrates global information. Each position i only pays attention to other positions within a window of size w around itself, and only adds a global attention at the label that can pay attention to all other positions. High-dimensional abstract features are input into a classifier to identify custom OFDM signals in a blind background.

2. The method for detecting and recognizing specific OFDM signals based on multi-domain feature fusion according to claim 1, characterized in that: The high-order cumulant feature is extracted from the received OFDM signal data, specifically by calculating the OFDM signal using the moment generation function and the cumulant generation formula. of The first-order mixture moment is further derived through formula to obtain the higher-order cumulant eigenvectors. : in: * indicates complex conjugation. For order, To determine the number of complex conjugate sequences, and ; This indicates calculating the mean. ， .

3. The method for detecting and recognizing specific OFDM signals based on multi-domain feature fusion according to claim 1, characterized in that: Time-frequency domain features for: Specifically: in, OFDM signal Discrete Fourier Transform (DFT), OFDM signal The total length is , This refers to the frequency band numbering used when dividing the entire spectrum using the Dyadic structure. The total number of frequency band segments, where the first segment is the number of segments in the frequency band. The bandwidth of each frequency band is The center frequency is The time location index is ; Indicates a discrete frequency index; Indicates a discrete-time index; It is the standard imaginary unit.

4. The method for detecting and recognizing specific OFDM signals based on multi-domain feature fusion according to claim 1, characterized in that: Multi-domain feature fusion vector Specifically: in, For higher-order cumulative quantities, Features in the time and frequency domains; After normalization and post-processing, the normalized multidimensional fused feature vector is obtained. .

5. The method for detecting and recognizing specific OFDM signals based on multi-domain feature fusion according to claim 1, characterized in that: After the multidimensional fused feature vector enters the improved Transformer network, it first passes through the position encoding module to obtain a position-encoded feature vector carrying position information. Let alpha = {1, 2, ..., H} and be an integer. When alpha = 1, the input of the alpha-th sub-attention module is the feature vector after position encoding. When 2 ≤ alpha ≤ H-1, the input of the alpha-th sub-attention module is the feature vector output by the (alpha-1)-th sub-attention module. The H-th sub-attention module outputs the extracted high-dimensional abstract features. The attention mechanism extracts the relationships between data from the input feature vector, obtains the weighted aggregated features, and outputs them to the first residual connection structure. The first residual connection structure extracts more abstract features from the weighted aggregated features and outputs the recognition result. After the recognition probability is processed by the first normalization structure, it is output to the feedforward neural network. After mapping the recognition probability, the feedforward neural network outputs the feedforward recognition probability to the second residual connection structure; The second residual connection structure extracts more abstract features from the feedforward recognition probability and outputs the recognition result. After being processed by the second normalization structure, it becomes a high-dimensional abstract feature, which is then used as the output of the sub-attention module.

6. The method for detecting and recognizing specific OFDM signals based on multi-domain feature fusion according to claim 5, characterized in that: The position encoding module processes the normalized multidimensional fused feature vector. When performing position encoding, the specific steps are as follows: when When it is even: ; when When it is an odd number: ; in, Normalized multi-domain feature fusion vector No. The element at position n, with dimension 1 , For location index, For position The corresponding positional encoding vector, Represents a semi-order index of a dimension. And round down; the final feature vector input to the first sub-attention module at the back end is .

7. The method for detecting and recognizing specific OFDM signals based on multi-domain feature fusion according to claim 5, characterized in that: Input feature vectors into the attention mechanism Then, each position in the weighted aggregated features is obtained. The output result is: in, Indicates the center position of the local window's attention; This indicates the central location of global attention. For position The set of other locations that can be focused on. The attention weights are calculated as follows: , , These are the input feature vectors. The query matrix, key matrix, and value matrix are obtained through linear transformation mapping. for Dimensions This represents the scoring function. Indicates position The locations that can be monitored.

8. The method for detecting and recognizing specific OFDM signals based on multi-domain feature fusion according to claim 5, characterized in that: Both the first residual connection structure and the second residual connection structure adopt the form of adding the input and the mapping function, and the mapping function is a convolution function or an activation function; The first and second normalization structures adjust the input vector to a vector with a mean of 0 and a variance of 1.

9. The method for detecting and recognizing specific OFDM signals based on multi-domain feature fusion according to claim 5, characterized in that: Feedforward neural networks process the input vector Perform two linear transformations, then use the ReLU function to activate and map the first recognition probability, and output the feedforward recognition probability, specifically: in, These are the linear transformation coefficients. This represents the probability of feedforward recognition.

10. The method for detecting and recognizing specific OFDM signals based on multi-domain feature fusion according to claim 1, characterized in that: When high-dimensional abstract features are input into a classifier to identify custom OFDM signals in a blind background: The classifier is a structure consisting of two convolutional layers, two fully connected layers, and a SoftMax activation function connected in sequence. The classifier classifies high-dimensional abstract features and obtains the signal category recognition probability result.

Images