Link 11 signal detection and identification method based on YOLOv8

Abstract

The invention discloses a Link 11 signal detection and identification method based on YOLOv8, and belongs to the field of signal detection and identification in the communication field. An existing signal detection method has the defects of complex calculation, poor performance and the like, and a signal recognition method has the problems of low training efficiency and poor recognition capability of an algorithm. The invention discloses a Link 11 signal detection and identification method based on YOLOv8, and the method is realized through the following steps: making a data set based on a signal generated by MATLAB, carrying out the preprocessing of the data set signal, so as to optimize the signal, and enabling the processed signal to be convenient for subsequent processing; time-frequency analysis processing is carried out, and a converted time-frequency graph is sent to a YOLO network; feature learning is carried out in the YOLO network, and the network is trained; and detecting and identifying the signal by using the trained network weight, and classifying and identifying the extracted feature vector. According to the method, the YOLOv8 model is used for detecting and identifying the Link11 data link communication signal, and the detection rate and the identification rate of the signal are further improved on the basis of the prior art.

Description

Technical Field

[0001] The present invention relates to a Link11 signal detection and recognition method, and in particular to a Link11 signal detection and recognition method based on YOLOv8. Background Art

[0002] Signal detection is a basic link in communication signal processing and the basis for subsequent tasks such as signal parameter estimation. Its essence is to perform statistical analysis on the received signal to determine whether the target signal exists. Its main function is to classify, demodulate, encode and modulate the signal. Since the generation and transmission of the signal will be affected by unstable factors such as other signals and channels, which will cause the instability of the received signal waveform, it is necessary to detect the signal first to determine whether there is a signal in the received data, and then identify the modulation method of the signal, etc., to facilitate subsequent demodulation, analysis and other signal processing tasks.

[0003] At present, the commonly used signal detection methods are mainly divided into matched filtering method, cyclostationary method, eigenvalue detection method and energy detection method as well as deep learning method. Figure 5 shown.

[0004] Although traditional methods have been widely used, they still have many defects. It can be seen that traditional signal detection methods have shortcomings such as complex calculations, many criteria, and poor performance at low signal-to-noise ratios. There is an urgent need for intelligent and effective detection methods.

[0005] The above four types of traditional signal detection algorithms only focus on the problem of signal existence determination, which can no longer meet today's technical requirements. In the past two years, the rapid development of deep learning in the field of signal detection has attracted widespread attention from scholars, but it is still in its infancy and few research results have been published.

[0006] The signal detection method based on deep learning can automatically extract signal features to complete detection and classification by building and training a network model, and the actual effect is better than the traditional method. The present invention will detect Link11 signals. How to efficiently construct a communication signal annotation sample set, and appropriately improve the network structure and training algorithm according to the characteristics of the communication signal are key issues in the application of deep learning in communication signal detection.

[0007] In addition, the current research on communication signal recognition methods can be roughly divided into two categories: traditional modulation recognition algorithms and machine learning-based methods. Figure 6 shown.

[0008] The traditional recognition algorithms of communication signals can be roughly divided into two categories: likelihood ratio recognition method based on decision theory and statistical pattern recognition method based on feature extraction.

[0009] The likelihood ratio recognition method refers to treating the signal modulation recognition problem as a hypothesis testing problem based on probability theory and hypothesis testing related theories, establishing the maximum likelihood discriminant of the signal, deriving the optimal decision threshold of the signal, and comparing a certain statistic of the signal with the threshold value to determine the signal modulation type. Methods based on likelihood ratio recognition generally need to rely on prior information such as the mean, variance, and covariance of the signal. In non-cooperative communications, these parameters are difficult to obtain. Therefore, fewer and fewer scholars are studying likelihood recognition methods.

[0010] The statistical pattern recognition method refers to extracting the modulation statistics of the signal, and then comparing it with the theoretical value to perform modulation recognition. This method has the advantages of requiring less prior knowledge, simple theoretical analysis, high efficiency, and matching each modulation mode, but it requires the design of appropriate feature parameters and a classifier with strong generalization ability, and the recognition performance is limited under low signal-to-noise ratio. The features used in the recognition method based on feature extraction mainly include amplitude distribution, cyclic autocorrelation function, cyclic spectrum, high-order cumulant, time-frequency diagram, quartic spectrum, constellation diagram, etc.

[0011] The existing technology realizes modulation mode recognition by extracting artificially designed shallow signal features, and the feature design depends on the experience of researchers, which is difficult to extract and has a small range of applicable modulation modes. Deep learning is a branch of machine learning, which is a machine learning method based on neural networks. Deep learning has the advantage of automatically extracting features from signals. By training and learning a large amount of sample data, a multi-layer deep neural network is formed to extract deep hidden features of samples, solving the problem of artificially designed features. In recent years, it has been widely used in the field of modulation signal recognition.

[0012] There are two main methods for recognition algorithms based on machine learning:

[0013] At present, signal recognition based on classifiers generally uses classifiers such as decision trees and support vector machines. Some technologies have designed support vector machine classifiers based on second-order cyclic spectrum features and amplitude statistical moment features for signal modulation recognition. This method has certain reference significance for in-depth research on data link signals. However, classification and recognition based on cyclic spectrum features requires a certain amount of data and a high signal-to-noise ratio, and is relatively complex to implement. Some technologies have proposed using high-order cumulants, differential high-order cumulants, second-order spectra and fourth-order spectra as feature parameters, and using CART decision trees to effectively identify Link16 data link signals. Some technologies have proposed a data link recognition and classification method based on a support vector machine (SVM) decision tree, and its convergence speed and recognition accuracy are better than the BP neural network algorithm.

[0014] In recent years, many scholars have used neural networks to study modulation recognition. Some researchers have proposed a parallel network structure of GRU and CNN to extract the high-order cumulative amount, wavelet transform features, signal-to-noise ratio (SNR) estimation features, instantaneous features and cyclic spectrum features of the signal, and then input them into the GRU and CNN networks for recognition. The final decision is made based on the recognition effect of the two networks. The recognition rate is greater than 90% when the signal-to-noise ratio is -6dB, and the recognition rate is 100% when the signal-to-noise ratio is -1dB. Some researchers have proposed an attention mechanism short-link convolutional long short-term memory deep neural network (ASCLDNN) recognition model. The network is optimized for modulation signal recognition and introduces an attention mechanism to improve recognition accuracy by increasing the weight of important signals. Experimental results show that ASCLDNN can recognize 11 types of signal modulation with high accuracy at a low signal-to-noise ratio and without confusion for specific signals. Some researchers have proposed a modulation recognition algorithm based on the TCNN-Bilstm (two-way convolutional neural network cascaded bidirectional long short-term memory) network. The identification signal is divided into I and Q components, and the upper and lower convolutional layer branches are connected in parallel to extract the multi-dimensional features of the I and Q components; the multi-dimensional features are fused and sent to the Bilstm network; finally, the softmax classifier is used to complete the recognition of 6 digital modulation signals, and the recognition rate reaches more than 92%, but the algorithm training efficiency needs to be improved. There is a technology that aims at the modulation recognition problem of frequency hopping signals in non-cooperative communication. First, the frequency hopping signal is subjected to short-time Fourier transform (STFT) and Wigner-Ville distribution (WVD) joint time-frequency transformation to obtain the time-frequency diagram, and then the time-frequency diagram is denoised by two-dimensional Wiener filtering, and then segmented and reset, and finally sent to CNN for classification and recognition. The algorithm can achieve an average recognition accuracy of 92.54% for 8 frequency hopping signals at a -4dB signal-to-noise ratio, but the algorithm has a large amount of calculation and its timeliness needs to be improved. In the technology, the vector diagram and I / Q eye diagram of the signal are used as features and input into the multi-terminal CNN to realize the recognition of 8 commonly used signals in satellite communications. The satellite amplitude and phase signals are converted into two-dimensional grayscale eye diagrams and vector diagrams with stronger ability to represent communication signals, and the prior knowledge carried by the signal itself is used to reduce the requirements for neural networks. The simulation results show that when there is no frequency offset, the algorithm can achieve an accuracy rate of more than 80% for the recognition of 8 types of signals at 0dB, which is better than other existing algorithms to a certain extent. However, the performance of the algorithm is greatly affected by frequency offset. A technology has proposed a signal modulation recognition method based on multi-feature fusion, and constructed a deep learning network with a dual-branch structure to extract the features of IQ signals and multi-channel constellations respectively. Through the complementary features of signals in different forms, a more complete signal feature representation can be constructed. At the same time, it can better alleviate the impact of noise and frequency offset on recognition performance, and effectively improve the classification accuracy of modulation recognition. However, the features of the constellation diagram are easily affected by frequency offset. In the case of high-frequency offset, the recognition ability of the algorithm is poor.

[0015] The present invention designs a network based on the working mode, message format, modulation mode and other characteristics of Link11 signals, selects the optimal network model, and improves the signal recognition rate. Summary of the invention

[0016] The purpose of the present invention is to solve the shortcomings of existing signal detection methods, such as complex calculations, more criteria, and poor performance at low signal-to-noise ratios, and an intelligent and effective detection method is urgently needed. The signal detection and recognition method has the problems of needing to improve the training efficiency, large amount of calculations, timeliness, and the performance of the algorithm being greatly affected by frequency deviation, and poor recognition ability of the algorithm in the process of signal detection and recognition, and proposes a Link11 signal detection and recognition method based on YOLOv8.

[0017] The above purpose is achieved through the following technical solutions:

[0018] A Link11 signal detection and recognition method based on YOLOv8 is implemented by the following steps:

[0019] Step 1: Create a data set based on the signal generated by MATLAB, and pre-process the data set signal to optimize the signal so that the processed signal is convenient for subsequent processing;

[0020] Step 2: Perform time-frequency analysis and send the converted time-frequency graph to the YOLO network;

[0021] Step 3: Perform feature learning in the YOLO network and train the network;

[0022] Step 4: Use the trained network weights to detect and identify the signal, and classify and identify the extracted feature vectors.

[0023] Furthermore, the steps of making a data set based on the signal generated by MATLAB specifically include:

[0024] Step 11, setting data link signal simulation parameters, the parameters include sampling rate, symbol rate, Gaussian white noise, relative frequency offset, multipath number, multipath delay and multipath fading;

[0025] Step 12: generating an information sequence representing information transmitted by the modulated signal;

[0026] Step 13: Use the set modulation type to perform a modulation operation on the random sequence to obtain signal data and complete signal modulation;

[0027] Step 14: Select different channel environments for simulation according to specific circumstances;

[0028] Step 15: Set the sampling rate within the range allowed by the sampling theorem to sample the modulated signal.

[0029] Furthermore, the step of preprocessing the data set signal in step 1 is specifically:

[0030] The signal is preprocessed using denoising methods;

[0031] In the case of low signal-to-noise ratio, denoising technology is used to remove part of the noise; wavelet basis and threshold are selected according to the specific signal characteristics and noise characteristics; if the signal characteristics are concentrated in the low-frequency area and the noise mainly exists in the high-frequency area, the wavelet basis function is selected, such as Daubechies wavelet;

[0032] If the spectral distributions of the signal and noise are similar, choose a wavelet basis function, such as Symlets or Coiflets.

[0033] Furthermore, the step of performing time-frequency analysis and sending the converted time-frequency graph to the YOLO network in step 2 is specifically:

[0034] First, the one-dimensional time-frequency signal is converted into a time-frequency diagram in a two-dimensional plane to display the energy distribution of the signal, and then the time-frequency diagram is subjected to feature detection and recognition; typical linear time-frequency analysis methods include short-time Fourier transform and smoothed pseudo-Wigner-Ville distribution;

[0035] a. Short-time Fourier transform STFT is defined as follows:

[0036]

[0037] Where x(t) is the signal, w(t) is the window function, t is time, f is frequency, and * represents the conjugate operation;

[0038] b. In the pseudo-Wigner-Ville distribution (PWVD), the window length determines the compromise between the time domain resolution and the frequency domain resolution. By smoothing the frequency direction, PWVD improves the accuracy of the time-frequency analysis results and reduces the influence of cross-term interference. The definition of PWVD is as follows:

[0039]

[0040] Where x(t) is the signal, * represents the conjugate operation, h(τ) is the window function, t is the time, and f is the frequency;

[0041] c. Smoothed pseudo-Wigner-Ville distribution:

[0042] The SPWVD time-frequency analysis method is formed by adding smoothing processing on the basis of PWVD. By performing weighted averaging on the time domain window and the frequency domain filter, the time-frequency characteristics of the signal become smoother and the influence of cross-term interference is reduced.

[0043] Assume that the time domain window function is h(τ), the frequency domain window function is g(ω), the time-frequency domain symmetric window function is p(τ, ω), and the autocorrelation function is smoothed using the time-frequency domain symmetric transformation window. The smoothed sequence of the discrete input signal is shown in the following formula:

[0044]

[0045] The discrete input signal time-frequency domain smooth sequence is transformed into the Fourier time-frequency domain, and then the signal time-frequency domain is double-integrated to obtain the SPWVD output time domain signal Z * (t) See the following formula:

[0046]

[0047] The output time domain signal Z * (t) is decomposed into a low-frequency band and a high-frequency band dual signal time domain sequence as shown in the following formula:

[0048] Z * (t) = Z 1* (t)+Z 2* (t)

[0049] Among them, Z 1* (t) is the low frequency signal, Z 2* (t) is the high frequency signal; the result after the signal is calculated by SPWVD is shown in the following formula:

[0050]

[0051] SPWVD(t, ω) is the SPWVD output signal in the time-frequency domain. It is the SPWVD low frequency band output signal. It is the high frequency band output signal of SPWVD. Output crossover frequency interference signal for SPWVD high and low frequency;

[0052] Based on the Wigner-Ville distribution, the time domain window function and the frequency domain window function are introduced, and the STFT-SPWVD time-frequency analysis method is designed. Combining STFT and SPWVD, STFT is used to extract the local characteristics of Link11 signal and SPWVD is used to smooth the Link11 signal.

[0053] Furthermore, the step of performing feature learning in the YOLO network described in step 3, the steps of training the network include:

[0054] Step 31: Determine the structure of the YOLO network and select YOLOv8. The YOLOv8 structure includes a backbone feature extraction network, a feature enhancement network and a detection head;

[0055] The backbone feature extraction network is used to extract features of the signal image, and the backbone feature extraction network includes modules such as Conv, C2f and SPPF; in the Conv module, operations include convolution, batch normalization and SiLU activation function; YOLOv8n obtains residual features, and at the same time, the SPPF module converts the feature map into a feature vector of fixed dimension;

[0056] The feature enhancement network establishes a connection between the backbone network and the detection layer, and the neck adopts FPN and PAN structures to capture contextual information;

[0057] The detection head is used to generate object detection predictions, and non-maximum suppression techniques are applied to filter out redundant results and retain the prediction results;

[0058] Step 32: Add the attention mechanism CA module to the YOLOv8 network to improve the YOLOv8 network;

[0059] The CA attention mechanism module is used to enhance the expressiveness of mobile network learning features. c ]∈R H×W×C , after transformation, the output is a tensor of the same size Y = [y1, y2, ..., y c ]∈R H×W×C ;

[0060] The CA module encodes channel relationships and long-range dependencies through precise location information, which includes the following steps: coordinate information embedding and coordinate attention generation.

[0061] The beneficial effects of the present invention are:

[0062] In recent years, the detection and identification of Link11 signals have adopted traditional methods, and neural networks have not yet been introduced into the detection and identification of the signal. For signal detection and identification, the present invention preprocesses the Link11 data set, sets conditions, and further optimizes the signal, so that the effective information in the data set is more prominent and the interference of invalid information is eliminated. The present invention applies YOLOv8 to the detection of Link11 signals. For signal detection and identification, most neural networks complete the detection and identification of signals in steps, while YOLO can complete the detection and identification of signals at the same time. First, the received signal is preprocessed, mainly to optimize the signal, so that the processed signal can facilitate subsequent time-frequency analysis and processing, and then the converted time-frequency diagram is sent to the YOLO network for feature learning. By improving a more efficient network structure, it can quickly identify the time-frequency diagram of the signal. Finally, the trained network weights are used to test the signal, and the extracted feature vectors are classified and identified. The present invention uses the YOLOv8 network of the deep learning framework to achieve high-precision signal detection and identification.

[0063] The data link signal is obtained through antennas, radars, reconnaissance satellites and various detection sensors. The present invention uses MATLAB for simulation. Due to the diversification of communication means and the increasingly complex electromagnetic environment, traditional signal detection and recognition methods cannot meet the needs of all parties. The present invention introduces deep learning into the detection and recognition of Link11 signals, making up for the shortcomings of traditional signal detection and recognition methods, making it of great value in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 is a flow chart of the method of the present invention;

[0065] Figure 2 It is a Link11 signal time-frequency feature extraction framework involved in the present invention;

[0066] Figure 3 The YOLOv8 structure involved in the present invention;

[0067] Figure 4 It is the CA attention mechanism module involved in the present invention;

[0068] Figure 5 It is a traditional signal detection method involved in the background technology of the present invention;

[0069] Figure 6 It is a signal recognition algorithm involved in the background technology of the present invention;

[0070] Figure 7 It is a diagram of the network control station call mode signal and production time domain waveform generated under -5dB-Link11 signal-to-noise ratio;

[0071] Figure 8 It is a diagram of the network control station call mode signal and production time domain waveform generated under 0dB-Link11 signal-to-noise ratio;

[0072] Fig. 9 It is a diagram of the network control station call mode signal and production time domain waveform generated under 2dB-Link11 signal-to-noise ratio;

[0073] Fig.10 It is a diagram of the -5dB-Link11 outpost reply mode signal and production time domain waveform;

[0074] Fig.11 It is a diagram of the 0dB-Link11 outpost response mode signal and production time domain waveform;

[0075] Fig.12 It is a diagram of the 2dB-Link11 outpost reply mode signal and production time domain waveform;

[0076] Fig.13 It is a diagram of the -5dB-Link11 network control station reporting mode signal and production time domain waveform;

[0077] Fig.14 It is a 0dB-Link11 network control station reporting mode signal and production time domain waveform diagram;

[0078] Fig.15 2dB-Link11 network control station report mode signal and production time domain waveform diagram. DETAILED DESCRIPTION

[0079] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Specific implementation method one:

[0081] A Link11 signal detection and recognition method based on YOLOv8 in this embodiment is as follows: Figure 1 As shown, the method is implemented by the following steps:

[0082] Step 1: For signal detection and recognition, most neural networks complete signal detection and recognition in steps, while YOLO can complete signal detection and recognition at the same time. Since Link11 signals in actual scenarios are difficult to collect on a large scale and cannot provide an accurate parameter range, the present invention makes a data set based on the signal generated by MATLAB, pre-processes the data set signal to optimize the signal, and makes the processed signal convenient for subsequent processing;

[0083] Step 2: Perform time-frequency analysis and send the converted time-frequency graph to the YOLO network;

[0084] Step 3: Perform feature learning in the YOLO network and train the network to improve a more efficient network structure so that it can quickly identify the time-frequency diagram of the signal;

[0085] Step 4: Use the trained network weights to detect and identify the signal, and classify and identify the extracted feature vectors. Specific implementation method 2:

[0087] Different from the first specific implementation mode, in the Link11 signal detection and recognition method based on YOLOv8 in this implementation mode, the step of preparing a data set based on the signal generated by MATLAB in step 1 specifically includes:

[0088] Step 11, setting data link signal simulation parameters, the parameters include sampling rate, symbol rate, Gaussian white noise, relative frequency offset, multipath number, multipath delay and multipath fading;

[0089] Step 12: generating an information sequence representing information transmitted by the modulated signal;

[0090] Step 13: Use the set modulation type to perform a modulation operation on the random sequence to obtain signal data and complete signal modulation;

[0091] Step 14: After simulating the channel, the modulated signal will encounter different channel environments during transmission. For example, there are different ranges of Gaussian white noise and frequency offset in the Gaussian channel, and there is the influence of multipath effect in the Rayleigh channel. Different channel environments are selected for simulation according to specific circumstances.

[0092] Step 15: Set a suitable sampling rate within the range allowed by the sampling theorem to sample the modulated signal. Specific implementation method three:

[0094] Different from the second specific implementation mode, in the Link11 signal detection and recognition method based on YOLOv8 in this implementation mode, the step of preprocessing the data set signal in step 1 is specifically:

[0095] Signal preprocessing refers to a series of preprocessing steps to improve the quality of the signal. The preprocessing methods used include: denoising, filtering, gain adjustment, etc. The present invention uses a denoising method to preprocess the signal;

[0096] In the case of low signal-to-noise ratio, the signal may be submerged by the noise, and it is difficult to detect the existence of the signal. Therefore, it is necessary to use denoising technology to remove part of the noise; thus, it is more conducive to signal detection. The principle of wavelet denoising is to use the multi-scale characteristics of wavelet transform to decompose the signal into multiple scales, then perform threshold processing on the noise at each scale, and finally reconstruct the denoised signal through inverse wavelet transform. When performing wavelet denoising, it is necessary to select a suitable wavelet basis and threshold strategy. In practical applications, the wavelet basis and threshold can be selected according to the specific signal characteristics and noise characteristics; if the signal characteristics are concentrated in the low-frequency area, and the noise mainly exists in the high-frequency area, then select a wavelet basis function with better frequency domain characteristics, such as Daubechies wavelet;

[0097] If the spectral distributions of the signal and noise are similar and better noise resistance is required, a wavelet basis function with better noise resistance is selected, such as Symlets wavelet or Coiflets wavelet.

[0098] The advantage of wavelet denoising is that it can consider the characteristics of the signal in both the time and frequency domains, effectively remove the noise in the signal, retain the important information of the signal, and improve the quality of the signal. Link11 signal detection and recognition. Specific implementation method four:

[0100] Different from the third embodiment, in the Link11 signal detection and recognition method based on YOLOv8 in this embodiment, the step of performing time-frequency analysis and processing and sending the converted time-frequency graph to the YOLO network in step 2 is specifically yes:

[0101] Due to the differences between communication signals of different modulation methods, the conventional Fourier transform method cannot describe the frequency characteristics at any moment and cannot comprehensively analyze the modulation type of the signal. In order to identify the signal characteristics and distinguish different modulation types, first, the one-dimensional time-frequency signal is converted into a time-frequency diagram in a two-dimensional plane to display the energy distribution of the signal, and then the time-frequency diagram is subjected to feature detection and recognition; typical linear time-frequency analysis methods include short-time Fourier transform (STFT) and smoothed pseudo-Wigner-Ville distribution (SPWVD);

[0102] a. The basic idea of ​​short-time Fourier transform is to use the local stationary characteristics of the signal to construct a time window function of limited length in the time domain to intercept the signal, translate the time window along the time axis to perform Fourier transform, and obtain the change of signal frequency over time, thereby mapping the signal from the time domain to the time-frequency domain and obtaining the distribution of the signal in the time-frequency domain; the definition of STFT is as follows:

[0103]

[0104] Among them, x(t) is the signal, w(t) is the window function, t is time, f is frequency, and * represents the conjugate operation; although the Link11 data link signal is a non-stationary signal, it can be considered to be stationary within a very small time τ. In this way, the non-stationary signal x(t) is equivalent to a linear combination of multiple stationary signals. The spectrum of the non-stationary signal x(t) can be obtained by linearly combining the spectrum of multiple stationary signals. As one of the linear time-frequency analysis methods, STFT is simple to calculate in actual processing. When processing multi-component signals, there are no cross terms and no interference will be caused in the time-frequency domain. However, STFT has certain limitations. The form and length of its time window are fixed. Once the time window changes, it will have a great impact on the performance of STFT.

[0105] b. Pseudo Wigner-Ville Distribution (PWVD). PWVD is a time-frequency analysis method. Its main feature is the use of a rectangular window function to apply a smooth window to the time-frequency distribution in the frequency direction. PWVD uses the idea of ​​window function expansion units, which are rectangular windows composed of sampling points. In the pseudo-Wigner-Ville distribution PWVD, the window length determines the compromise between the time domain resolution and the frequency domain resolution. As the window length increases, the time domain and frequency domain resolutions will decrease, and the cross-term interference will also decrease. Therefore, by smoothing the frequency direction, PWVD improves the accuracy of the time-frequency analysis results and reduces the impact of cross-term interference; the definition of PWVD is shown in the following formula:

[0106]

[0107] Where x(t) is the signal, * represents the conjugate operation, h(τ) is the window function, t is the time, and f is the frequency;

[0108] c. Smoothed Pseudo-Wigner-Ville Distribution (SPWVD)

[0109] The SPWVD time-frequency analysis method is formed by adding smoothing processing on the basis of PWVD. By performing weighted averaging on the time domain window and the frequency domain filter, the time-frequency characteristics of the signal become smoother and the influence of cross-term interference is reduced.

[0110] Assume that the time domain window function is h(τ), the frequency domain window function is g(ω), the time-frequency domain symmetric window function is p(τ, ω), and the autocorrelation function is smoothed using the time-frequency domain symmetric transformation window. The smoothed sequence of the discrete input signal is shown in the following formula:

[0111]

[0112] The discrete input signal time-frequency domain smooth sequence is transformed into the Fourier time-frequency domain, and then the signal time-frequency domain is double-integrated to obtain the SPWVD output time domain signal Z * (t) See the following formula:

[0113]

[0114] The output time domain signal Z * (t) is decomposed into a low-frequency band and a high-frequency band dual signal time domain sequence as shown in the following formula:

[0115] Z * (t) = Z 1* (t)+Z 2* (t)

[0116] Among them, Z 1*(t) is the low frequency signal, Z 2* (t) is the high frequency signal; the result after the signal is calculated by SPWVD is shown in the following formula:

[0117]

[0118] SPWVD(t, ω) is the SPWVD output signal in the time-frequency domain. It is the SPWVD low frequency band output signal. It is the high frequency band output signal of SPWVD. Output crossover frequency interference signal for SPWVD high and low frequency;

[0119] In order to overcome the problem of inaccurate time-frequency diagrams caused by noise interference, the time domain window function and the frequency domain window function are introduced on the basis of Wigner-Ville distribution to effectively suppress the cross-interference items in the time-frequency diagram, making it more advantageous when processing Link11 data link signals. In order to improve the time-frequency resolution of Link11 signals, a time-frequency analysis method based on STFT-SPWVD is designed, such as Figure 2 The core idea is to combine STFT and SPWVD, using STFT to extract the local features of Link11 signals, and SPWVD to smooth Link11 signals to reduce cross-interference, remove low energy, noise and irrelevant components in the signal, and present obvious Link11 signal features in the time-frequency diagram. Specific implementation method five:

[0121] Different from the fourth specific implementation, in the Link11 signal detection and recognition method based on YOLOv8 in this implementation, the feature learning is performed in the YOLO network in step 3, and the steps of training the network include:

[0122] Step 3: Determine the structure of the YOLO network. Select YOLOv8. The YOLOv8 structure includes a backbone feature extraction network (Backbone), a feature enhancement network (Neck) and a detection head (Detect); the YOLOv8 structure is as follows: Figure 3 As shown,

[0123] The backbone feature extraction network is used to extract the features of the signal image. The backbone feature extraction network includes modules such as Conv, C2f and SPPF. In the Conv module, operations include convolution, batch normalization (BN) and SiLU activation function. YOLOv8n introduces an innovative C2f structure to obtain residual features, thereby maintaining gradient flow information while ensuring lightweight design. At the same time, the SPPF module (called spatial pyramid pooling) converts the feature map into a feature vector of fixed dimension.

[0124] The feature enhancement network establishes a connection between the backbone network and the detection layer, promoting the integration and fusion of features of different scales and abstraction levels. The neck adopts FPN and PAN structures to capture contextual information to improve the representation ability of objects of different sizes and positions;

[0125] The detection head is used to generate object detection predictions and apply non-maximum suppression (NMS) technology to filter out redundant results, thereby retaining reliable and accurate prediction results;

[0126] Step 32: Add the attention mechanism CA module (Coordinate Attention, referred to as CA module) to the YOLOv8 network to improve the YOLOv8 network;

[0127] The attention mechanism has the following advantages:

[0128] It can capture direction-aware and position-aware information, which can help the model locate and identify targets of interest more accurately, and can bring huge gains to downstream tasks based on lightweight networks, especially those tasks with dense predictions.

[0129] The CA attention mechanism module is used to enhance the expressiveness of mobile network learning features. c ]∈R H×W×C , after transformation, the output is a tensor of the same size Y = [y1, y2, ..., y c ]∈R H×W×C ;

[0130] The CA module encodes channel relationships and long-range dependencies through precise location information. Similar to the SE module, it includes the following steps: coordinate information embedding and coordinate attention generation. The attention mechanism module is as follows: Figure 4 shown.

[0131] The method of the present invention is verified by simulation:

[0132] At present, we have completed the collection of Link11 related materials and literature reading. From the perspective of the characteristics of the time and space information of the Link11 signal, we have conducted a systematic study on the papers published by the top people in the field at home and abroad. We have mastered the more efficient prediction methods in the field of Link-type signal detection and recognition, and focused on understanding the theoretical research work on the topic of Link11 signal detection and recognition.

[0133] Link11 dataset construction

[0134] Due to the lack of public Link11 signal datasets, we are faced with the challenge of obtaining representative actual Link11 signal data, so we use a self-made dataset. Generate multiple Link11 signal data composed of three modes of signals: network control station call, outpost response, and network control station report, to ensure that the signal data is large and the coverage is wide. Introduce noise with different signal-to-noise ratios to make the Link11 signal close to the real scene and construct a Link11 signal dataset. According to the data transmission message format, the Link11 data link signal uses Hamming coding and π / 4-DQPSK modulation to generate the corresponding network control station call mode, outpost response mode, and network control station report data link signals.

[0135] Under the condition of the same noise amplitude, the Link11 network control station call mode, outpost response mode and network control station report signal data under different signal-to-noise ratios are selected to generate time domain waveforms, such as Figure 7-15 The Link data link has three modes of signals as shown;

[0136] Figure 7 It is a diagram of the network control station call mode signal and production time domain waveform generated under -5dB-Link11 signal-to-noise ratio;

[0137] Figure 8 It is a diagram of the network control station call mode signal and production time domain waveform generated under 0dB-Link11 signal-to-noise ratio;

[0138] Fig. 9 It is a diagram of the network control station call mode signal and production time domain waveform generated under 2dB-Link11 signal-to-noise ratio;

[0139] Fig.10 It is a diagram of the -5dB-Link11 outpost reply mode signal and production time domain waveform;

[0140] Fig.11 It is a diagram of the 0dB-Link11 outpost response mode signal and production time domain waveform;

[0141] Fig.12 It is a diagram of the 2dB-Link11 outpost reply mode signal and production time domain waveform;

[0142] Fig.13 It is a diagram of the -5dB-Link11 network control station reporting mode signal and production time domain waveform;

[0143] Fig.14 It is a 0dB-Link11 network control station reporting mode signal and production time domain waveform diagram;

[0144] Fig.152dB-Link11 network control station report mode signal and production time domain waveform diagram.

[0145] The embodiments of the present invention disclose preferred embodiments, but are not limited thereto. A person skilled in the art can easily understand the spirit of the present invention based on the above embodiments and make different extensions and changes. However, as long as they do not deviate from the spirit of the present invention, they are all within the protection scope of the present invention.

Claims

1. A Link11 signal detection and recognition method based on YOLOv8, characterized in that: The method is implemented by the following steps: Step 1: Create a data set based on the signal generated by MATLAB, and pre-process the data set signal to optimize the signal so that the processed signal is convenient for subsequent processing; Step 2: Perform time-frequency analysis and send the converted time-frequency graph to the YOLO network; Step 3: Perform feature learning in the YOLO network and train the network; Step 4: Use the trained network weights to detect and identify the signal, and classify and identify the extracted feature vectors.

2. According to a Link11 signal detection and recognition method based on YOLOv8 according to claim 1, it is characterized in that: The steps to create a data set based on the signal generated by MATLAB include: Step 11, setting data link signal simulation parameters, the parameters include sampling rate, symbol rate, Gaussian white noise, relative frequency offset, multipath number, multipath delay and multipath fading; Step 12: generating an information sequence representing information transmitted by the modulated signal; Step 13: Use the set modulation type to perform a modulation operation on the random sequence to obtain signal data and complete signal modulation; Step 14: Select different channel environments for simulation according to specific circumstances; Step 15: Set the sampling rate within the range allowed by the sampling theorem to sample the modulated signal.

3. A Link11 signal detection and recognition method based on YOLOv8 according to claim 1 or 2, characterized in that: The step of preprocessing the data set signal in step 1 is specifically: The signal is preprocessed by using denoising method; In the case of low signal-to-noise ratio, denoising technology is used to remove part of the noise; wavelet basis and threshold are selected according to the specific signal characteristics and noise characteristics; if the signal characteristics are concentrated in the low-frequency area and the noise mainly exists in the high-frequency area, the wavelet basis function is selected, such as Daubechies wavelet; If the spectral distributions of the signal and noise are similar, choose a wavelet basis function, such as Symlets or Coiflets.

4. The Link11 signal detection and recognition method based on YOLOv8 according to claim 3, characterized in that: The steps of performing time-frequency analysis and sending the converted time-frequency graph to the YOLO network in step 2 are as follows: First, the one-dimensional time-frequency signal is converted into a time-frequency diagram in a two-dimensional plane to display the energy distribution of the signal, and then the time-frequency diagram is subjected to feature detection and recognition; typical linear time-frequency analysis methods include short-time Fourier transform and smoothed pseudo-Wigner-Ville distribution; a. Short-time Fourier transform STFT is defined as follows: Where x(t) is the signal, w(t) is the window function, t is time, f is frequency, and * represents the conjugate operation; b. In the pseudo-Wigner-Ville distribution (PWVD), the window length determines the compromise between the time domain resolution and the frequency domain resolution. By smoothing the frequency direction, PWVD improves the accuracy of the time-frequency analysis results and reduces the influence of cross-term interference. The definition of PWVD is as follows: Where x(t) is the signal, * represents the conjugate operation, h(τ) is the window function, t is the time, and f is the frequency; c. Smoothed pseudo-Wigner-Ville distribution: The SPWVD time-frequency analysis method is formed by adding smoothing processing on the basis of PWVD. By performing weighted averaging on the time domain window and the frequency domain filter, the time-frequency characteristics of the signal become smoother and the influence of cross-term interference is reduced. Assume that the time domain window function is h(τ), the frequency domain window function is g(ω), the time-frequency domain symmetric window function is p(τ, ω), and the autocorrelation function is smoothed using the time-frequency domain symmetric transformation window. The smoothed sequence of the discrete input signal is shown in the following formula: The discrete input signal time-frequency domain smooth sequence is transformed into the Fourier time-frequency domain, and then the signal time-frequency domain is double-integrated to obtain the SPWVD output time domain signal Z * (t) See the following formula: The output time domain signal Z * (t) is decomposed into a low-frequency band and a high-frequency band dual signal time domain sequence as shown in the following formula: Z * (t)=Z 1* (t)+Z 2* (t) Among them, Z 1* (t) is the low frequency signal, Z 2* (t) is the high frequency signal; the result after the signal is calculated by SPWVD is shown in the following formula: SPWVD(t, ω) is the SPWVD output signal in the time-frequency domain. It is the SPWVD low frequency band output signal. It is the high frequency band output signal of SPWVD. Output crossover frequency interference signal for SPWVD high and low frequency; Based on the Wigner-Ville distribution, the time domain window function and the frequency domain window function are introduced, and the STFT-SPWVD time-frequency analysis method is designed. Combining STFT and SPWVD, STFT is used to extract the local characteristics of Link11 signal and SPWVD is used to smooth the Link11 signal.

5. The Link11 signal detection and recognition method based on YOLOv8 according to claim 4, characterized in that: In step 3, feature learning is performed in the YOLO network, and the steps of training the network include: Step 31: Determine the structure of the YOLO network and select YOLOv8. The YOLOv8 structure includes a backbone feature extraction network, a feature enhancement network and a detection head; The backbone feature extraction network is used to extract features of the signal image, and the backbone feature extraction network includes modules such as Conv, C2f and SPPF; in the Conv module, operations include convolution, batch normalization and SiLU activation function; YOLOv8n obtains residual features, and at the same time, the SPPF module converts the feature map into a feature vector of fixed dimension; The feature enhancement network establishes a connection between the backbone network and the detection layer, and the neck adopts FPN and PAN structures to capture contextual information; The detection head is used to generate object detection predictions, and non-maximum suppression techniques are applied to filter out redundant results and retain the prediction results; Step 32: Add the attention mechanism CA module to the YOLOv8 network to improve the YOLOv8 network; The CA attention mechanism module is used to enhance the expressiveness of mobile network learning features. c ]∈R H×W×C , after transformation, the output is a tensor of the same size Y = [y1, y2, ..., y c ]∈R H×W×C ; The CA module encodes channel relationships and long-range dependencies through precise location information, which includes the following steps: coordinate information embedding and coordinate attention generation.

Images