Frequency hopping communication method fusing self-attention mechanism and lightweight convolution

Abstract

The invention relates to the technical field of wireless communication, and particularly discloses a frequency hopping communication method fusing a self-attention mechanism and lightweight convolution, which comprises the following steps: S1, a frequency hopping signal detection and parameter estimation step: S1-1, a data preprocessing step: performing short-time Fourier transform (STFT) on a received communication signal, and generating a time-frequency graph; for a given observation signal x (t), the short-time Fourier transform is # imgabs0 #, w (t) is a window function, and e-j2piftau is in a complex conjugate form; and after sampling at equal intervals, the discrete form is # imgabs1 # imgabs2 # frequency dimension k = 1, 2,..., K, and time dimension n = 1, 2,..., N. And S1-2, a feature extraction step: inputting the generated time-frequency graph into a lightweight convolutional layer, and extracting the time-frequency feature of the frequency hopping signal. And S1-3, a multi-task processing step: sharing the features extracted by the convolutional network to a frequency hopping signal detection branch and a parameter estimation branch. By adopting the technical scheme of the invention, the requirements of efficient and accurate processing of frequency hopping signals in a complex electromagnetic environment can be met.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technology, and in particular to a frequency hopping communication method integrating a self-attention mechanism and lightweight convolution. Background Art

[0002] In the field of modern communications, frequency-hopping communication plays an important role in military communications due to its strong anti-interference ability and low interception rate. However, with the development of communication technology, the electromagnetic environment is becoming increasingly complex, which brings many challenges to the processing of frequency-hopping signals.

[0003] While current research has achieved some success in frequency-hopping signal detection and parameter estimation, numerous challenges remain. While non-blind detection methods can achieve rapid detection when some prior information is known, obtaining that prior information is challenging. For example, methods like autocorrelation detection and likelihood detection require pre-setting parameters such as the frequency-hopping signal's hopping rate, hopping time, or candidate frequency set, significantly limiting their practical applications. While blind detection methods require no prior information, their performance degrades significantly in low signal-to-noise ratio environments, and signal separation accuracy is insufficient when multiple signals overlap. Methods based on time-frequency analysis, such as wavelet transforms and Wigner distributions, are susceptible to cross-term interference, resulting in incomplete and inaccurate parameter estimation. Deep learning-based methods, such as convolutional neural networks and long short-term memory networks, suffer from high algorithmic complexity and demanding training sample requirements.

[0004] In the field of underdetermined blind source separation, because the number of source signals outnumbers the number of observed signals, reconstructing the source signals requires solving an underdetermined system of equations, which has non-unique solutions and is therefore an ill-posed problem. Classic independent component analysis (ICA) algorithms cannot be directly applied. While the commonly used sparse component analysis method has some effectiveness, when processing overlapping frequency-hopping signals, the proportion of mixed points increases and the density of single source points decreases. Traditional single source point determination criteria are prone to failure, resulting in inaccurate mixing matrix estimation and limited source signal recovery accuracy.

[0005] In summary, the existing frequency hopping communication processing technology is difficult to meet the needs of efficient and accurate processing of frequency hopping signals in complex electromagnetic environments. There is an urgent need for a frequency hopping communication method that integrates self-attention mechanism and lightweight convolution. Summary of the Invention

[0006] The present invention provides a frequency hopping communication method that integrates self-attention mechanism and lightweight convolution, which can meet the requirements of efficient and accurate processing of frequency hopping signals in complex electromagnetic environments.

[0007] In order to solve the above technical problems, this application provides the following technical solutions:

[0008] A frequency hopping communication method integrating a self-attention mechanism and lightweight convolution includes the following steps:

[0009] The S1 frequency hopping signal detection and parameter estimation steps include the following:

[0010] S1-1 data preprocessing step, the received communication signal is subjected to short-time Fourier transform (STFT) to generate a time-frequency diagram. For a given observation signal x(t), its short-time Fourier transform is Where w(t) is the window function, e -j2πfτ is a complex conjugate form; after equal-interval sampling, its discrete form is Frequency dimension k = 1, 2, ..., K, time dimension n = 1, 2, ..., N.

[0011] In the S1-2 feature extraction step, the generated time-frequency graph is input into the lightweight convolution layer to extract the time-frequency features of the frequency hopping signal.

[0012] In the S1-3 multi-task processing step, the features extracted by the convolutional network are shared with the frequency hopping signal detection branch and the parameter estimation branch, and the Transformer is introduced to capture the long-term dependencies using the self-attention mechanism. The detection results and parameter estimation results of the frequency hopping signal are jointly output through the multi-task head. The parameters include the hopping time, hopping period, and frequency hopping frequency set.

[0013] The S2 underdetermined blind source separation step includes the following:

[0014] In the single source point detection step S2-1, a feature vector is constructed for the time-frequency point of the overlapping frequency hopping signal, and a classification model that has been trained a priori is introduced to determine whether the time-frequency point is a single source point.

[0015] In the S2-2 clustering and estimation step, the density-based spatial clustering with noise application (DBSCAN) algorithm is used to cluster the detected single source points by frequency, obtain the time-frequency single source point sets corresponding to different source signals, and complete the mixing matrix and wave direction of arrival estimation.

[0016] In the S2-3 source signal recovery step, after estimating the mixing matrix, the least squares method is used to minimize the sum of squared errors between the observed data and the estimated model to determine the model parameters. Then, through residual iterative optimization, the residuals are continuously calculated and the estimated values ​​are adjusted to improve the accuracy of source signal recovery.

[0017] The basic principles and benefits of this solution are as follows: In frequency-hopping signal detection and parameter estimation, the received signal is converted into a time-frequency diagram using a short-time Fourier transform (SFT), fully reflecting the non-stationary nature of frequency-hopping signals. Lightweight convolutional layers efficiently extract local features from the time-frequency diagram, reducing computational effort and improving processing efficiency. The self-attention mechanism within the Transformer captures long-term temporal dependencies and analyzes time-frequency features from a global perspective, enabling more accurate detection and parameter estimation. Multiple task heads jointly output detection and parameter estimation results, enabling end-to-end processing that balances accuracy and efficiency.

[0018] In the underdetermined blind source separation (BSS) component, a single-source point detection method based on multi-parameter fusion constructs a feature vector and utilizes a previously trained classification model to identify single sources. This method improves single-source point detection accuracy compared to traditional single-condition detection. The DBSCAN clustering algorithm clusters single sources by frequency, effectively processing overlapping frequency-hopping signals and accurately estimating the mixing matrix and direction of arrival. The least-squares method combined with iterative residual optimization continuously adjusts model parameters to improve source signal recovery accuracy.

[0019] In summary, the present invention can effectively improve the frequency hopping signal processing capability in a complex electromagnetic environment, achieve the purpose of efficient and accurate processing of frequency hopping signals, and meet the task requirements of drone swarm monitoring and interference.

[0020] Furthermore, the lightweight convolutional layer adopts the MobileNet network structure and reduces model parameters and computational complexity through depth-wise separable convolution.

[0021] Through the above method, the efficiency of feature extraction can be improved.

[0022] Furthermore, in Transformer, the self-attention mechanism uses the scaled dot product attention method to calculate the attention score by scaling the calculation results.

[0023] The above method can avoid the gradient vanishing or gradient exploding problem and improve the calculation stability and accuracy.

[0024] Furthermore, the classification model that has been trained a priori is a convolutional neural network (CNN) model, which is trained with a large number of time-frequency point samples to learn the feature differences between single source points and mixed points, thereby improving the accuracy of single source point detection.

[0025] Furthermore, when performing underdetermined blind source separation, a sparse transformation is first performed on the signal with poor time domain sparsity, and clustering and subsequent processing are performed after the signal sparsity is enhanced.

[0026] Furthermore, in the frequency hopping signal detection branch, a threshold judgment method is used to determine the existence of the frequency hopping signal. When the detection result exceeds a preset threshold, it is determined that the frequency hopping signal exists.

[0027] Furthermore, in the parameter estimation branch, the Kalman filter algorithm is used to dynamically update the frequency hopping signal parameters by combining historical parameter estimation results and current time-frequency characteristics.

[0028] Through the above method, the accuracy and real-time performance of parameter estimation can be improved.

[0029] Furthermore, when using the DBSCAN clustering algorithm, the neighborhood radius and the minimum number of samples parameters are adjusted to adapt to overlapping frequency hopping signal scenarios of different complexities and optimize the clustering effect.

[0030] Furthermore, in the source signal recovery process, compressed sensing technology is used to reduce the amount of observation data required while ensuring the accuracy of signal reconstruction, thereby improving processing efficiency.

[0031] Furthermore, it also includes the verification and feedback links of the processing results, comparing and verifying the frequency hopping signal detection and parameter estimation results, and the source signal after underdetermined blind source separation with known samples or actual conditions. If the error exceeds the allowable range, the relevant algorithm parameters are adjusted and reprocessed. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a flowchart of a frequency hopping communication method that integrates self-attention mechanism and lightweight convolution. DETAILED DESCRIPTION

[0033] The following is further described in detail through specific implementation methods:

[0034] A frequency hopping communication method that integrates self-attention mechanism and lightweight convolution (such as Figure 1 ), comprising the following steps:

[0035] The S1 frequency hopping signal detection and parameter estimation steps include the following:

[0036] S1-1 data preprocessing step, the received communication signal is subjected to short-time Fourier transform (STFT) to generate a time-frequency diagram. For a given observation signal x(t), its short-time Fourier transform is Where w(t) is the window function, e -j2πfτ is a complex conjugate form; after equal-interval sampling, its discrete form is Frequency dimension k = 1, 2, ..., K, time dimension n = 1, 2, ..., N.

[0037] In the S1-2 feature extraction step, the generated time-frequency graph is input into the lightweight convolution layer to extract the time-frequency features of the frequency hopping signal.

[0038] In the S1-3 multi-task processing step, the features extracted by the convolutional network are shared with the frequency hopping signal detection branch and the parameter estimation branch, and the Transformer is introduced to capture the long-term dependencies using the self-attention mechanism. The detection results and parameter estimation results of the frequency hopping signal are jointly output through the multi-task head. The parameters include the hopping time, hopping period, and frequency hopping frequency set.

[0039] The S2 underdetermined blind source separation step includes the following:

[0040] In the single source point detection step S2-1, a feature vector is constructed for the time-frequency point of the overlapping frequency hopping signal, and a classification model that has been trained a priori is introduced to determine whether the time-frequency point is a single source point.

[0041] In the S2-2 clustering and estimation step, the density-based spatial clustering with noise application (DBSCAN) algorithm is used to cluster the detected single source points by frequency, obtain the time-frequency single source point sets corresponding to different source signals, and complete the mixing matrix and wave direction of arrival estimation.

[0042] In the S2-3 source signal recovery step, after estimating the mixing matrix, the least squares method is used to minimize the sum of squared errors between the observed data and the estimated model to determine the model parameters. Then, through residual iterative optimization, the residuals are continuously calculated and the estimated values ​​are adjusted to improve the accuracy of source signal recovery.

[0043] Specifically, the MobileNet network structure is used as the lightweight convolutional layer. Its depth-wise separable convolution can significantly reduce model parameters and computational complexity while ensuring the feature extraction effect, thereby improving the overall processing speed.

[0044] The Transformer uses scaled dot product attention to calculate the attention score, which optimizes the calculation process of the self-attention mechanism, ensures stable model training, and improves the ability to capture time-frequency features.

[0045] The classification model is trained using a convolutional neural network. Its powerful feature learning ability can better identify the feature differences between single-source points and mixed points, providing a guarantee for accurate judgment of single-source points.

[0046] Sparse transformation is performed on signals with poor time domain sparsity to enhance their sparsity, making subsequent clustering and processing effects better and improving the accuracy of underdetermined blind source separation.

[0047] The threshold judgment method is used in the frequency hopping signal detection branch to simply and efficiently determine the existence of the frequency hopping signal, providing a basis for subsequent processing.

[0048] The parameter estimation branch combines the Kalman filter algorithm to dynamically update the parameter estimation results based on historical and current information, adapt to signal changes, and improve the real-time performance and accuracy of parameter estimation.

[0049] By adjusting the neighborhood radius and minimum sample number parameters of the DBSCAN clustering algorithm, the clustering effect can be optimized according to different overlapping frequency hopping signal scenarios and the reliability of the mixing matrix estimation can be improved.

[0050] Applying compressed sensing technology in source signal recovery can reduce data volume requirements, ensure signal reconstruction accuracy under resource-constrained conditions, and improve processing efficiency.

[0051] In addition, the verification and feedback links of the processing results are added. Through comparative verification, processing errors can be discovered in time, and the algorithm parameters can be adjusted and reprocessed to ensure the accuracy and reliability of the processing results.

[0052] In practical applications, consider the processing of frequency-hopping communication signals from drone swarms in military communications. During the frequency-hopping signal detection and parameter estimation phase, the receiving device acquires the communication signal and performs a short-time Fourier transform (SFT) to convert the signal into a time-frequency plot. Assuming a sampling frequency of fs = 10 MHz, a Hanning window function with a window length of M = 1024 and a sampling point count of N = 2048, the time-frequency plot is calculated using the following formula.

[0053] The time-frequency graph is fed into a lightweight convolutional layer using the MobileNet architecture, where depth-wise separable convolutions can quickly extract local time-frequency features. The features then enter the Transformer, where the self-attention mechanism calculates the attention score with a scaling factor of (d k is the feature dimension), effectively capturing long-term temporal dependencies.

[0054] Finally, the multi-task head outputs the detection results and parameter estimation results. If the detection result exceeds the preset threshold T = 0.5, a frequency hopping signal is determined to be present, and parameters such as the hopping time, hopping period, and hopping frequency set are output. In the underdetermined blind source separation stage: for overlapping frequency hopping signals, a feature vector containing frequency, amplitude, phase, and other information is constructed for the time-frequency point. This feature vector is then input into a convolutional neural network classification model trained with a large number of samples to determine the single source point. When using the DBSCAN clustering algorithm, the neighborhood radius ∈ = 0.1 and the minimum number of samples MinPts = 5 are set based on the signal characteristics. The mixing matrix and direction of arrival are estimated for the single source point clusters. The least squares method is combined with residual iterative optimization to recover the source signal. The number of iterations is set to 50 to continuously improve the recovery accuracy.

[0055] After processing is completed, the detection, parameter estimation and source signal recovery results are compared and verified with the known drone swarm frequency hopping communication parameters and signal samples. If the error exceeds the allowable range, such as the frequency hopping frequency estimation error exceeds 10kHz, the relevant algorithm parameters are adjusted and reprocessed.

[0056] The specific verification of the detection results is to perform cross-correlation matching on the time-frequency diagram of the detected frequency hopping signal and a known template (such as the typical time-frequency pattern of the frequency hopping signal stored in the database). If the correlation coefficient is less than 0.7, it is judged as a false alarm and the detection threshold τ is adjusted (for example, from 0.6 to 0.7).

[0057] Parameter estimation verification: compare the frequency stability of the estimated frequency hopping frequency set with the historical data (for example, the frequency hopping interval should meet the preset hopping rate range). If the error is >10%, the re-detection mechanism is triggered and the STFT window length is increased to 1.5 times the original one.

[0058] Source signal separation verification: calculate the signal-to-noise ratio (SNR) of the separated signal and the original source signal. If the SNR is less than 10 dB, adjust the ∈ parameter of DBSCAN (e.g., from 0.2 to 0.15) and perform clustering again.

[0059] The above are only embodiments of the present invention. The invention is not limited to the fields involved in this implementation case. Common knowledge such as the known specific structures and characteristics in the scheme is not described in detail here. Ordinary technicians in the relevant field are aware of all common technical knowledge in the technical field to which the invention belongs before the application date or priority date, can obtain all existing technologies in the field, and have the ability to apply conventional experimental means before that date. Ordinary technicians in the relevant field can improve and implement this scheme in combination with their own abilities under the inspiration given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the relevant field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several variations and improvements can be made, which should also be regarded as the scope of protection of the present invention. These will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A frequency hopping communication method integrating self-attention mechanism and lightweight convolution, characterized in that: The steps include: The S1 frequency hopping signal detection and parameter estimation steps include the following: S1-1 data preprocessing step, the received communication signal is short-time Fourier transform to generate a time-frequency diagram; for a given observation signal x(t), its short-time Fourier transform is Where w(t) is the window function, e -j2πfτ is a complex conjugate form; after equal-interval sampling, its discrete form is Frequency dimension k = 1, 2, ..., K, time dimension n = 1, 2, ..., N; S1-2 feature extraction step, inputting the generated time-frequency map into the lightweight convolution layer to extract the time-frequency features of the frequency hopping signal; In the multi-task processing step S1-3, the features extracted by the convolutional network are shared with the frequency hopping signal detection branch and the parameter estimation branch. The Transformer is introduced to capture long-term dependencies using the self-attention mechanism. The multi-task head jointly outputs the frequency hopping signal detection results and parameter estimation results. The parameters include the hopping time, hopping period, and the frequency hopping frequency set. The S2 underdetermined blind source separation step includes the following: In the single source point detection step S2-1, for overlapping frequency hopping signals, a feature vector is constructed for the time-frequency point, and a classification model trained a priori is introduced to determine whether the time-frequency point is a single source point. In the S2-2 clustering and estimation step, the density-based spatial clustering with noise (DBSCAN) algorithm is used to cluster the detected single source points by frequency, obtain the time-frequency single source point sets corresponding to different source signals, and complete the mixing matrix and wave direction of arrival estimation; In the S2-3 source signal recovery step, after estimating the mixing matrix, the least squares method is used to minimize the sum of squared errors between the observed data and the estimated model to determine the model parameters. Then, through residual iterative optimization, the residuals are continuously calculated and the estimated values ​​are adjusted to improve the accuracy of source signal recovery.

2. The frequency hopping communication method integrating self-attention mechanism and lightweight convolution according to claim 1, characterized in that: The lightweight convolution layer adopts the MobileNet network structure and reduces model parameters and computational complexity through depth-wise separable convolution.

3. The frequency hopping communication method integrating self-attention mechanism and lightweight convolution according to claim 2, characterized in that: In Transformer, the self-attention mechanism uses scaled dot product attention to calculate the attention score by scaling the calculation results.

4. The frequency hopping communication method integrating self-attention mechanism and lightweight convolution according to claim 3, characterized in that: The classification model that has been trained a priori is a convolutional neural network model. Through training with a large number of time-frequency point samples, it learns the feature differences between single source points and mixed points, thereby improving the accuracy of single source point detection.

5. The frequency hopping communication method integrating self-attention mechanism and lightweight convolution according to claim 4, characterized in that: When performing underdetermined blind source separation, signals with poor time-domain sparsity are first subjected to sparse transformation, and then clustering and subsequent processing are performed after the signal sparsity is enhanced.

6. The frequency hopping communication method integrating self-attention mechanism and lightweight convolution according to claim 5, characterized in that: In the frequency hopping signal detection branch, a threshold judgment method is used to determine the existence of the frequency hopping signal. When the detection result exceeds a preset threshold, it is determined that the frequency hopping signal exists.

7. The frequency hopping communication method integrating self-attention mechanism and lightweight convolution according to claim 6, characterized in that: In the parameter estimation branch, the Kalman filter algorithm is used to dynamically update the frequency hopping signal parameters by combining the historical parameter estimation results and the current time-frequency characteristics.

8. The frequency hopping communication method integrating self-attention mechanism and lightweight convolution according to claim 7, characterized in that: When using the DBSCAN clustering algorithm, the neighborhood radius and the minimum number of samples are adjusted to adapt to overlapping frequency hopping signal scenarios of different complexities and optimize the clustering effect.

9. The frequency hopping communication method integrating self-attention mechanism and lightweight convolution according to claim 8, characterized in that: In the process of source signal recovery, compressed sensing technology is used to reduce the amount of required observation data while ensuring the accuracy of signal reconstruction.

10. A frequency hopping communication method integrating self-attention mechanism and lightweight convolution according to claim 9, characterized in that: It also includes the verification and feedback of the processing results, comparing and verifying the frequency hopping signal detection and parameter estimation results, and the source signal after underdetermined blind source separation with known samples or actual conditions. If the error exceeds the allowable range, the relevant algorithm parameters are adjusted and reprocessed.

Images