A method for identifying a P-wave band burst signal modulation mode

Abstract

The application discloses a P-wave burst signal modulation mode recognition method. The method utilizes the preamble characteristics of the P-wave burst signal, solves the problem that the signal information segment characteristics are difficult to distinguish in a low signal-to-noise ratio environment, and avoids the dependence of the traditional signal preamble autocorrelation recognition algorithm on the signal preamble prior information and the influence of the frequency offset. The method first utilizes the signal preamble spectrum to complete image construction, and utilizes the spectrum characteristics formed by different burst signal preamble rule codes to distinguish; secondly, the multi-semantic feature fusion network fully utilizes the texture features such as signal preamble spectrum edge contour extracted by the low-layer residual network, and fuses the high-level residual network to extract the abstract complex high-level semantic features of the signal spectrum, so as to solve the problem that the residual network only utilizes the high-level abstract semantic features and ignores the low-level features, and improve the burst signal modulation mode recognition performance.

Description

Technical Field

[0001] This invention belongs to the field of deep learning and relates to a method for identifying the modulation mode of P-band burst signals based on a preamble image and a multi-semantic feature fusion network. Background Technology

[0002] Ultra-shortwave burst communication systems are wireless communication systems that transmit information in a short period of time. They are characterized by short signal exposure time during transmission, thus improving the system's anti-interference capability. Before acquiring the baseband information of the burst signal, it is necessary to first determine the modulation scheme of the received signal. Therefore, burst signal identification is a crucial step before signal demodulation and has significant research value.

[0003] During P-band burst signal transmission, a signal preamble exists. The signal preamble serves to identify the start position of the burst signal and assist in receiver loop synchronization. Furthermore, because the signal preamble is composed of regular codewords, it exhibits distinct time-frequency domain characteristics, allowing for accurate signal type differentiation through observation. Qian Ying, in her paper "Signal Frequency Identification Based on Autocorrelation Method," distinguishes signals by performing autocorrelation matching between a known local preamble sequence and the actual burst signal preamble, using the peak value of the matched sequence. However, this algorithm requires prior information about the preamble sequence, and the frequency offset causes the autocorrelation peak value of the burst signal preamble to decrease, resulting in poor identification performance.

[0004] In her paper "Research on Modulation Recognition Method of Communication Signal Based on Time-Frequency Image," Li Yuqian proposed an algorithm for time-frequency analysis of signals. This algorithm uses the signal's time-frequency image as a dataset and employs a Residual Network (ResNet) to extract features for signal classification. However, in the time-frequency image, the energy distribution of the signal in the frequency domain changes from amplitude to color intensity. In low signal-to-noise ratio (SNR) environments, this can lead to noise energy being comparable to signal energy, resulting in low color differentiation in the time-frequency image and a rapid decline in recognition performance. Huang Fengjie, in his paper "Research on Modulation Mode Recognition Technology Based on Deep Learning," proposed an algorithm that extracts the frequency domain image of the signal from the Fast Fourier Transform (FFT) as a network sample and uses a Convolutional Neural Network (CNN) to extract spectral features for signal classification. However, this algorithm extracts frequency domain features of signal information segments, and the frequency domain features of signal information segments with different modulation types show little difference in low SNR environments. The multi-semantic feature fusion network not only makes full use of the texture features such as the edge contour of the signal preamble spectrum extracted by the low-level residual network, but also integrates the abstract and complex high-level semantic features of the signal spectrum extracted by the high-level residual network. This solves the problem that the residual network only uses high-level abstract semantic features and ignores low-level features, thus improving the recognition performance of burst signal modulation mode. Summary of the Invention

[0005] The purpose of this invention is to improve the signal classification capability of 10 burst signal types in the P-band. These 10 burst signal types include 6 modulation types: Binary Phase Shift Keying (BPSK), Quadrature Phase Shift Keying (QPSK), Continuous Phase Frequency Shift Keying (CPFSK), Continuous Phase Modulation (CPM), Shaped Binary Phase Shift Keying (SBPSK), and Shaped Offset Quadrature Phase Shift Keying (SBPSK). Keying (SOQPSK) and six code rates {1200 Baud, 2400 Baud, 4800 Baud, 9600 Baud, 16000 Baud, 19200 Baud}, forming a set of 10 burst signals: {BPSK2400, BPSK4800, QPSK16000, CPM2400, CPM16000, CPFSK16000, SBPSK1200, SBPSK9600, SOQPSK2400, SOQPSK19200}.

[0006] This invention proposes a method for identifying P-band burst signal modulation schemes based on a preamble image and a multi-semantic feature fusion network. This method constructs an image of the received signal, fully utilizing the frequency domain features of the P-band burst signal preamble. It retains the edge texture features of the signal preamble spectrum extracted by the low-level residual network and fuses abstract information such as spectral geometry extracted by the high-level residual network. Compared to the relatively intuitive and directly extractable basic features of the signal preamble spectrum, such as edges, contours, and textures, high-level semantic features provide a deeper understanding and expression of semantic information within the image. High-level semantic features imply a higher-level understanding of the structure and relationships within the signal preamble spectrum image. This algorithm proposes a multi-semantic feature fusion module to fuse semantic features at various levels, constructing a more complete multi-dimensional signal preamble frequency domain feature, improving the accuracy of burst signal modulation scheme identification, and achieving accurate differentiation of different burst signal modulation schemes. Under a 0dB signal-to-noise ratio environment, compared with the signal preamble autocorrelation algorithm, the recognition algorithm of the signal information segment spectrum image and ResNet50 network, and the recognition algorithm of the signal information segment time-frequency image and ResNet50 network, the algorithm proposed in this invention improves the recognition performance of these 10 P-band burst signals by 10.56%, 5.40%, and 3.76%, respectively.

[0007] The technical solution adopted by this invention to solve its technical problem includes the following steps:

[0008] Step 1: Obtain signal sample data including burst signals;

[0009] Step 2: Detect the starting point position of the burst signal leader using the starting point detection method based on the frequency domain ratio method on the acquired signal sample data;

[0010] Step 3: Select N from the starting point of the signal. p The data sample points of each length are used to construct an image, and a three-channel image of each burst signal is obtained.

[0011] Step 4: After preprocessing the constructed three-channel image by channel transformation and size reduction, a residual layer including several levels of nonlinearity is used as the network structure to extract the semantic features of the preamble spectrum of the burst signal.

[0012] Step 5: The semantic features of the extracted signal preamble spectrum are processed by a multi-semantic feature fusion module consisting of a feature preprocessing module, an average pooling layer, and a Flatten layer. The semantic features of the burst signal preamble spectrum extracted by the residual layers at each level in Step 4 are fused to obtain multi-semantic fused features. Then, a fully connected classifier is used to complete the classification and the loss is updated in reverse.

[0013] Preferably, in step 1:

[0014] The signal sample data is obtained from real receivers and simulations under different signal-to-noise ratios.

[0015] The burst signals include {BPSK2400, BPSK4800, QPSK16000, CPM2400, CPM16000, CPFSK16000, SBPSK1200, SBPSK9600, SOQPSK2400, SOQPSK19200}, a total of 6 modulation types. Based on whether their phase changes continuously, they are divided into phase shift keying signals {BPSK, QPSK} and continuous phase signals {CPFSK, CPM, SBPSK, SOQPSK}.

[0016] Phase Shift Keying (PSK) signals transmit information using changes in the phase of a carrier wave. The absolute phase can take on M possible values, representing M-ary information, while the carrier amplitude and frequency remain unchanged. The expression for a PSK signal is:

[0017]

[0018] In equation (1), a nLet f be the baseband information of the nth symbol ±1, T be the symbol period, g(t) be the baseband pulse with duration T, and f be the baseband information of the nth symbol ±1. c The center carrier frequency of the signal. This represents the initial phase of the signal, n0(t), which follows the expression N(0,σ). 2 Additive white Gaussian noise, θ k = (k-1)*2π / M represents a set of uniformly spaced sinusoidal carrier phases, determined by the baseband symbols;

[0019] CPM signals are generated by loading digital information onto the phase through nonlinear modulation with memory, ensuring that the phase is a continuous function of time. The expression for a continuous-phase CPM signal is:

[0020]

[0021] In equation (2), A(t) is the function of the change in signal amplitude. This represents the phase function of the information-carrying symbol over time t, expressed as:

[0022]

[0023] In the formula, h represents the modulation index, and α = [α0, α0, ..., α0]. N-1 ] represents the baseband sequence, and q(t) represents the phase pulse shaping function;

[0024] The CPFSK signal is modulated by a single continuously varying frequency carrier wave, with a modulation index h = 1 / 2, and the baseband sequence of the transmitted signal α. i ∈{±1};

[0025] SOQPSK signal is a constant envelope continuous phase modulated signal with a modulation index h = 1 / 2, α i It is a pre-encoded transmission sequence with values ​​ranging from {-1, 0, 1}.

[0026] The phase pulse shaping function q(t) for CPM, CPFSK, and SOQPSK signals is expressed as follows:

[0027]

[0028] The phase change of SBPSK signals within a symbol is similar to that of SOQPSK signals; it can only be constant or ±π / 2. The modulation index of SBPSK signals is h = 1, α... i It is a pre-coded baseband sequence with values ​​ranging from {-1, 0, 1}; the phase pulse shaping function q(t) of the SBPSK signal has a width of half a symbol period, and its expression is:

[0029]

[0030] Preferably, in step 2, the starting point detection method based on the frequency domain ratio method specifically includes the following steps:

[0031] First, automatic gain control (AGC) is applied to the real signal data after front-end sampling to keep the burst signals and noise in the channel at similar low power levels.

[0032] Secondly, the data after AGC is completed is subjected to quadrature downconversion and low-pass filtering to remove the second harmonic component and convert it into complex signal baseband data.

[0033] Then, perform N on the obtained complex signal data. s Point sliding FFT is used to obtain the frequency domain features of the data and calculate N. s The frequency domain ratio sequence is obtained by comparing the peak frequency points of the FFT data sorted from smallest to largest with the mean frequency points of the in-band noise floor.

[0034] Finally, for the frequency domain ratio sequence, a sliding window of a preset size is selected, and each value in the sliding window is compared with a set threshold. If there are more than a specified number of frequency domain ratios in the sliding window, the starting point of the burst signal is determined.

[0035] Preferably, in step 3, the specific implementation of the image-based construction includes the following steps:

[0036] Let the data point at the starting point of the signal be r(0), then select N from the starting point. p The length data points are used as a set of signal preamble data points, denoted as preamble;

[0037] After performing an FFT operation on the preamble, a frequency domain sequence of the signal preamble data points is obtained. Then, a preprocessing method of spectrum filling is used to enhance edge features, resulting in a 224×224 three-channel image.

[0038] Preferably, in step 4, extracting the semantic features of the preamble spectrum of the burst signal specifically includes:

[0039] The preprocessed image is then processed through convolutional layers and max pooling layers to obtain an input image suitable for the basic residual unit.

[0040] The network structure includes four levels of residual network layers, each of which includes three convolutional layers. Each convolutional layer is accompanied by a ReLU activation function layer, and the residual networks at each level are connected in a skip connection to form a deep network.

[0041] Preferably, in step 5,

[0042] The multi-semantic feature fusion module consists of four feature preprocessing modules, an average pooling layer, and a Flatten layer;

[0043] The feature preprocessing module standardizes the semantic features of the burst signal preamble spectrum with different sizes and number of channels output by each residual layer to obtain standardized features; the standardized features are then dimensionality-concatenated and then subjected to dimensionality reduction processing by average pooling and Flatten layers to obtain multi-semantic fusion features.

[0044] The beneficial effects of this invention are as follows:

[0045] 1. It solves the problem of difficulty in distinguishing signal information segment features in low signal-to-noise ratio environments.

[0046] 2. It avoids the dependence on prior information of the signal preamble and the influence of frequency offset on traditional signal preamble autocorrelation identification algorithms.

[0047] 3. In a 0dB signal-to-noise ratio environment, compared with the signal preamble autocorrelation algorithm, the recognition algorithm of the signal information segment spectrum image and ResNet50 network, and the recognition algorithm of the signal information segment time-frequency image and ResNet50 network, the algorithm of this invention significantly improves the classification accuracy.

[0048] In summary, this invention features high recognition rate and strong robustness, and maintains a high recognition rate even at low signal-to-noise ratios. Attached Figure Description

[0049] Figure 1 Flowchart for detecting the starting point of a sudden signal;

[0050] Figure 2 Three-channel images of three different preamble spectra under different algorithms;

[0051] Figure 3 This is a signal preamble recognition network based on a multi-semantic feature fusion network;

[0052] Figure 4 It is a basic residual unit structure;

[0053] Figure 5 For feature preprocessing module;

[0054] Figure 6 Output feature maps for each level of the residual network for burst signals;

[0055] Figure 7 Performance of different algorithms for recognition at a signal-to-noise ratio of -10dB;

[0056] Figure 8 The average recognition performance of six algorithms at different signal-to-noise ratios is given. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0058] The specific implementation of the signal model received in step 1 is as follows:

[0059] Communication signal sample data under different signal-to-noise ratios were obtained through actual receivers and simulations. The sample set of burst signals includes {BPSK2400, BPSK4800, QPSK16000, CPM2400, CPM16000, CPFSK16000, SBPSK1200, SBPSK9600, SOQPSK2400, SOQPSK19200}, a total of 6 modulation types. Based on whether their phase changes continuously, they are divided into phase shift keying signals {BPSK, QPSK} and continuous phase signals {CPFSK, CPM, SBPSK, SOQPSK}.

[0060] Phase Shift Keying (PSK) signals transmit information using changes in the phase of a carrier wave. The absolute phase can take on M possible values, representing M-ary information, while the carrier amplitude and frequency remain unchanged. The expression for a PSK signal is:

[0061]

[0062] In equation (6), a n Let f be the baseband information of the nth symbol ±1, T be the symbol period, g(t) be the baseband pulse with duration T, and f be the baseband information of the nth symbol ±1. c The center carrier frequency of the signal. This represents the initial phase of the signal, n0(t), which follows the expression N(0,σ). 2 Additive white Gaussian noise, θ k = (k-1)*2π / M represents a set of uniformly spaced sinusoidal carrier phases, determined by the baseband symbols.

[0063] Continuous phase modulation (CPM) is a nonlinear modulation technique with memory. It loads digital information onto the phase while ensuring that the phase is a continuous function of time, expressed as:

[0064]

[0065] In equation (7), A(t) is the function of the change in signal amplitude. This represents the phase function of the information-carrying symbol over time t, expressed as:

[0066]

[0067] In the formula, h represents the modulation index, and α = [α0, α0, ..., α0]. N-1 ] represents the baseband sequence, and q(t) represents the phase pulse shaping function.

[0068] The CPM signal preamble modulation index h = 1 / 2 in this invention, and the baseband sequence α of the transmitted signal... i ∈{±1,±3}.

[0069] Unlike Frequency Shift Keying (FSK) signals, which transmit information through carrier frequency variations, CPFSK signals are modulated by a single continuously varying carrier with a modulation index h = 1 / 2. The baseband sequence of the transmitted signal is α. i ∈{±1}.

[0070] SOQPSK is a constant-envelope continuous-phase modulated signal. Unlike QPSK, SOQPSK's phase change has a slope, and the phase within each symbol is a function of time. Furthermore, SOQPSK differs from ordinary CPM signals in that its phase change within a symbol can only be constant or ±π / 2. The modulation index of SOQPSK is h = 1 / 2, α... i It is a pre-encoded transmission sequence with values ​​ranging from {-1, 0, 1}.

[0071] The phase pulse shaping function q(t) for CPM, CPFSK, and SOQPSK signals is expressed as follows:

[0072]

[0073] The phase change of SBPSK signals within a symbol is similar to that of SOQPSK signals; it can only remain constant or be ±π / 2. The modulation index of an SBPSK signal is h = 1, and α... i It is a pre-coded baseband sequence with values ​​ranging from {-1, 0, 1}. However, the phase pulse shaping function q(t) of the SBPSK signal has a width of half a symbol period, and its expression is:

[0074]

[0075] Step 2, starting point detection based on the frequency domain ratio method, is implemented as follows:

[0076] Detecting the start point of a burst signal is the first preprocessing operation required for the image-based construction of the signal preamble spectrum. By accurately identifying the start point, the starting position of the burst signal preamble can be precisely located, and the image-based construction can then be achieved using the preamble spectrum.

[0077] The burst signal start point detection is achieved using the frequency domain ratio method. This algorithm utilizes the power distribution of the burst signal's leading segment and noise segment within the same channel to analyze the differences between them in the frequency domain, thereby realizing the start point detection of the VHF burst signal.

[0078] First, Automatic Gain Control (AGC) is applied to the sampled real signal data to keep burst signals and noise in the channel at similar, low power levels. Second, the AGC-processed data undergoes quadrature down-conversion and low-pass filtering to remove second harmonic components, converting it into complex baseband signal data. Then, the resulting complex signal data is processed using N... s Point sliding FFT is used to obtain the frequency domain features of the data and calculate N. s The ratio of the peak frequency of the FFT data points sorted from smallest to largest to the mean frequency of the in-band noise floor is calculated. Finally, each calculated ratio within the sliding window is compared with a set threshold. If there are more than a specified number of frequency domain ratios within the sliding window that are higher than the set threshold, the starting point of the burst signal is determined.

[0079] Step 3, signal visualization construction, is implemented as follows:

[0080] The preamble of a burst signal consists of a set of cyclic codewords, exhibiting a power concentration distribution across several frequency points in the frequency domain. Due to differences in preamble modulation type and code rate, the preamble spectra of these 10 burst signals show significant differences in the frequency points of power distribution. Therefore, the preamble spectrum is extracted as an identification feature of the burst signal. Table 1 shows the preamble specifications for the 10 burst signals.

[0081]

[0082] Once the burst signal sample passes the starting point detection and the starting point position is accurately determined, the signal sample is re-attached to the carrier frequency, and samples are continuously selected from the starting point. Perform an FFT on each data sample point to obtain the frequency domain data of this data segment. Since P-band burst signals have a preamble, the data after FFT belongs to the frequency domain information of the burst signal preamble.

[0083] Let the data point at the starting point of the signal be r(0), then select N from the starting point. p The length data points are the preamble of the signal's leading data points. After performing an FFT on the preamble, the frequency domain sequence of the data set is obtained. To avoid information loss during spectrum conversion, minimize conversion loss, and ensure the integrity of the signal's leading frequency domain characteristics, a preprocessing method of spectrum padding is used to enhance edge features, resulting in a 224×224 three-channel image.

[0084] Step 4, multi-semantic feature extraction, is implemented as follows:

[0085] To ensure the network can obtain signal burst preamble spectral image features and achieve the fusion of multiple semantic features, this invention proposes using nonlinear basic residual units as the network structure to extract signal features. To better adapt to feature extraction in residual networks, the single-channel input image, after preprocessing, undergoes channel transformation and size reduction before being passed through convolutional layers and max-pooling layers to obtain an input image suitable for the basic residual units.

[0086] To extract semantic features from the signal preamble spectrum image as much as possible, this invention utilizes a four-level residual network in the feature extraction module, with the basic residual units of each layer distributed as (3, 4, 6, 3). Each residual unit passes the input through three convolutional layers and one 1×1 convolutional layer, and then sums the outputs to compensate for the feature information lost during the three convolutional layers. Furthermore, the basic residual unit of this invention uses three 1×1 convolutional kernels and one 3×3 convolutional kernel. The 1×1 convolutional kernel, by changing the number of input channels, first reduces the dimensionality and then increases it, effectively reducing the computational load of the residual network and the number of network parameters in the basic residual unit, thus improving computational efficiency for deep residual networks. The 3×3 convolutional kernel, on the other hand, reduces the information size, utilizing a smaller receptive field to avoid the detail loss in the signal preamble spectrum image caused by a larger receptive field, enabling the network to accurately extract more precise semantic features from the spectrum. Meanwhile, by utilizing the nonlinear characteristics of the ReLU function, a deep network is formed in the basic residual unit structure at each level, increasing nonlinear convolution operations, enabling the network to extract features from multiple angles, and improving the performance of the signal recognition algorithm.

[0087] Step 5, multi-semantic feature fusion, is implemented as follows:

[0088] A residual network composed of multiple basic residual units extracts signal preamble spectral features. The lower convolutional layers of the network have relatively small receptive fields, primarily extracting local features such as spectral texture and edges. However, as the number of network layers increases, the convolutional layers can extract more abstract and complex semantic features of the signal preamble spectrum. Therefore, to enable the network to simultaneously acquire semantic features from each level, this invention employs a multi-semantic feature fusion module to fuse the semantic features extracted by the residual network at each level, obtaining multi-semantic features. The multi-semantic feature fusion module consists of four feature preprocessing modules, an average pooling layer, and a Flatten layer.

[0089] To ensure the consistency of the semantic feature contribution ratio of each layer of the residual network, it is necessary to use the feature preprocessing module to preprocess and change the channels and sizes of the output feature maps of each level of the residual network.

[0090] The feature preprocessing module standardizes the semantic features of the output feature maps of each residual network layer from different channel numbers and sizes. The standardized semantic features are then dimensionality-concatenated and further processed by average pooling and Flatten layers to obtain multi-semantic fusion features of the signal preamble spectrum image. The fused signal features not only retain the abstract and complex signal features output from higher convolutional layers but also incorporate more detailed features such as signal preamble spectrum edges and textures. This improves upon the current problem that deep learning networks rely solely on high-level semantic features for burst signal classification, enhancing the network's ability to identify burst signal types. Finally, a fully connected layer completes the burst signal type classification, and the recognition accuracy is statistically analyzed.

[0091] Example 1

[0092] 1. Ten burst signals were obtained through real receivers and simulations, including {BPSK2400, BPSK4800, QPSK16000, CPM2400, CPM16000, CPFSK16000, SBPSK1200, SBPSK9600, SOQPSK2400, SOQPSK19200}. The sampling rate of the burst signals was 96kHz, the carrier frequency was 24kHz, and the frequency offset was random within ±500Hz.

[0093] 2. The procedure for detecting the starting point of a sudden signal is as follows: Figure 1 As shown. The received 10 burst signals are converted into complex baseband data through AGC, quadrature downconversion, and low-pass filtering. N is then selected. s Perform a sliding FFT on the points and calculate N. s The ratio of the peak frequency of the FFT data after sorting the points from smallest to largest to the mean frequency of the in-band noise floor is used. Finally, each calculated ratio within the sliding window is compared with a set threshold. If there are more than a specified number of frequency domain ratios within the sliding window that are higher than the set threshold, the starting point of the UHF burst signal is determined.

[0094] 3. The signal samples are re-loaded with a carrier frequency. 2048 points are selected from the detected starting point as the signal preamble data set. An FFT is performed to obtain the frequency domain sequence of the data set. Then, this frequency domain sequence is padded with spectral fill to enhance its edge features, resulting in a three-channel image with a resolution of 224×224. Figure 2 The images show three-channel images of three different preamble spectra under different algorithms. From top to bottom, they are the spectra of BPSK2400, BPSK4800, and SOQPSK2400 signals at 10dB and -10dB. Figure 2As shown, due to differences in modulation type and code rate, the number and bandwidth of peak spectral lines in the preamble spectrum, the bandwidth and number of sidelobes in the information segment spectrum, and the bandwidth and phase transition in the time-frequency plot of the three burst signals all differ under a high signal-to-noise ratio (SNR) of 10dB. However, the characteristics of the signal preamble spectrum are more pronounced compared to the information segment spectrum and the time-frequency plot. In an environment with a low SNR of -10dB, there are no significant differences in the characteristics of the information segment spectrum and the time-frequency plot of the different burst signals, while the differences in the number and distribution of spectral lines can still be observed in the preamble spectrum image, maintaining a relatively ideal feature discrimination.

[0095] For each type of burst signal, 4000 images were used as the training set and 1000 images were used as the dataset, and the samples were labeled with the corresponding type.

[0096] 4. Figure 3 The overall structure of the signal preamble recognition network based on multi-semantic feature fusion is shown below. First, the input image is processed through a 7×7 dimensional input convolutional layer, an instance normalization layer, and a ReLU layer to obtain a signal image of (112×112×1). Then, it is processed through a max pooling layer to obtain an input image of (56×56×64). Finally, the language features of each level of the residual network are extracted through a 4-level residual unit network with a distribution of (3, 4, 6, 3). Figure 4 The basic residual unit structure used for feature extraction.

[0097] 5. Semantic features extracted from residual networks at each level are processed through... Figure 5 The feature preprocessing module shown achieves semantic feature standardization, obtaining a 4-level semantic feature of (7×7×2048). Then, it performs dimensionality concatenation (7×7×8192) and then performs dimensionality reduction processing with average pooling layer and Flatten layer to obtain the multi-semantic fusion feature of the signal preamble spectrum image (1×8192). Finally, the output feature is classified through a fully connected layer, and the recognition accuracy is calculated.

[0098] Example 2:

[0099] 1. Dataset

[0100] The experiment used 10 burst signal sets, including {BPSK2400, BPSK4800, QPSK16000, CPM2400, CPM16000, CPFSK16000, SBPSK1200, SBPSK9600, SOQPSK2400, and SOQPSK19200}. The samples obtained were acquired through a real signal receiver and simulation. Samples with a signal-to-noise ratio (SNR) of 5 dB or higher were collected by the receiver; however, for samples with a SNR below 5 dB, due to the limited number collected by the receiver, additive white Gaussian noise was added to the higher SNR signals using Matlab 2022a. The burst signal preamble dataset has a sampling rate of 96kHz, a carrier frequency of 24kHz, and a random frequency offset within ±500Hz. The signal-to-noise ratio range is {-20dB, -17dB, -15dB, -13dB, -10dB, -5dB, 0dB, 5dB, 10dB, 15dB}. Each burst signal preamble has 2048 sampling points. The selected preamble complex signal samples are represented as a signal preamble spectrum image with a resolution of 224×224 for three channels. For each burst signal type, 4000 images are used as the training set and 1000 images as the dataset, with corresponding labeling for each image sample. The dataset uses a preprocessing method of spectrum padding to increase the network's sensitivity to spectral edge features.

[0101] 2. Experimental Environment and Network Configuration

[0102] The server GPU used in the simulation experiment was an NVIDIA TITAN RTX 3090, the deep learning framework was PyTorch, the batch size during model training was 32, the training epochs were 64, and the loss function was optimized by the Adaptive Momentum Estimation (ADAM) optimizer, with a learning rate of 3e-4, a weight decay coefficient of 5e-4, and the cross-entropy loss function.

[0103] 3. Performance Simulation

[0104] In addition to the network proposed in this invention, this example also uses the ResNet-50 network from "Deep residual learning for image recognition" (Reference 1).

[0105] 3-1. Visualization of Network Features

[0106] After the signal preamble spectral image is passed through the residual networks of the multi-semantic feature fusion network, the spectral features extracted by each layer are saved to obtain the feature image. Because the output features of each residual network layer have different sizes, the output feature maps of each layer are normalized to the same scale to facilitate observation of feature differences. Figure 6 This represents the output characteristics of each layer of the residual network in the preamble spectrum image of some burst signals.

[0107] Figure 6 In equations 1.a to 1.c, the semantic features of the preamble spectra of P-band burst signals {BPSK2400, BPSK4800, SOQPSK2400} are extracted through the first-level residual network. The output features still reveal the energy distribution characteristics of the burst preamble in the frequency domain. Compared to the abstract features of higher-level networks, lower-level residual networks are more sensitive to texture features such as image edges and contours, reflecting the local features of the preamble spectra of different burst signals. From equations 2.a to 2.c and 3.a to 3.c, we find that as the residual network deepens, it extracts more abstract semantic features and some texture features such as contour edges. Observing equations 4.a to 4.c, at the fourth layer of the residual network, it can extract abstract semantic features with discriminative power. With the increase of network layers, the output of higher-level residual networks has more abstract and advanced semantic features, making different burst signals more identifiable. Therefore, by extracting signal features at different levels through residual networks and using a feature fusion module, multi-semantic features that fuse features from each layer can be obtained.

[0108] 3-2. Performance of Different Burst Signal Classification and Recognition Algorithms

[0109] Ten burst signal types were selected under a -10dB signal-to-noise ratio environment as the experimental subjects to test the classification performance of different algorithms for these ten burst signals. The test set for each burst signal type consisted of 1000 samples. Figure 7 This refers to the recognition performance of various burst signals under different algorithms at a signal-to-noise ratio of -10dB.

[0110] Figure 7 It can be seen that the signal preamble and multi-semantic feature fusion network modulation method proposed in this invention makes full use of the preamble features of P-band burst signals compared with other algorithms, and does not require prior information of the preamble sequence, thus achieving a more ideal classification effect for various burst signals.

[0111] 3-3. Classification Performance Analysis

[0112] Assuming the training and test sets are identical, the average classification performance of 10 burst signals under different signal-to-noise ratios (SNRs) is statistically analyzed. The classification results of six recognition algorithms under different SNRs are as follows: Figure 8 As shown.

[0113] Depend on Figure 8 It can be seen that the signal multi-semantic feature fusion network using the preamble spectrum image of the burst signal proposed in this invention significantly improves the average recognition rate under different signal-to-noise ratios compared to the signal bandwidth estimation and constellation diagram feature recognition algorithm, the signal preamble autocorrelation algorithm, the burst signal information segment spectrum image and ResNet50 network recognition algorithm, the signal information segment time-frequency image and ResNet50 network recognition algorithm, and the signal preamble image and ResNet50 network recognition algorithm. At a signal-to-noise ratio of -15dB, the recognition algorithm proposed in this invention improves the average recognition rate by 65.56%, 20.88%, 30.83%, 60.39%, and 4.21% respectively compared to the other five recognition algorithms.

[0114] Finally, it should be noted that the purpose of disclosing the embodiments is to help further understand the present invention. However, those skilled in the art will understand that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the embodiments, and the scope of protection of the present invention is defined by the claims.

Claims

1. A method for identifying the modulation mode of a P-band burst signal, the method being based on a preamble image and a multi-semantic feature fusion network, characterized in that, Includes the following steps: Step 1: Obtain signal sample data including burst signals; Step 2: Detect the starting point position of the burst signal leader using the starting point detection method based on the frequency domain ratio method on the acquired signal sample data; Step 3: Select from the starting point of the signal The data sample points of each length are used to construct an image, and a three-channel image of each burst signal is obtained. Step 4: After preprocessing the constructed three-channel image by channel transformation and size reduction, a residual layer including several levels of nonlinearity is used as the network structure to extract the semantic features of the preamble spectrum of the burst signal. Step 5: The semantic features of the extracted signal preamble spectrum are processed by a multi-semantic feature fusion module consisting of a feature preprocessing module, an average pooling layer, and a Flatten layer. The semantic features of the burst signal preamble spectrum extracted by the residual layers at each level in Step 4 are fused to obtain multi-semantic fused features. Then, a fully connected classifier is used to complete the classification and the loss is updated in reverse.

2. The method for identifying P-band burst signal modulation mode according to claim 1, characterized in that, In step 1: The signal sample data is obtained from real receivers and simulations under different signal-to-noise ratios. The burst signals include {BPSK2400, BPSK4800, QPSK16000, CPM2400, CPM16000, CPFSK16000, SBPSK1200, SBPSK9600, SOQPSK2400, SOQPSK19200}, a total of 6 modulation types. Based on whether their phase changes continuously, they are divided into phase shift keying signals {BPSK, QPSK} and continuous phase signals {CPFSK, CPM, SBPSK, SOQPSK}. Phase Shift Keying (PSK) signals transmit information using changes in the phase of a carrier wave; their absolute phase can be determined. 10 possible values, representing The information is in a different number system, but its carrier amplitude and frequency remain unchanged; the expression for a PSK signal is: （1） In equation (1), For the first Baseband information of each symbol , For symbol period, The duration is The baseband pulse, The center carrier frequency of the signal. Indicates the initial phase of the signal. To obey Additive white Gaussian noise, It represents a set of uniformly spaced sinusoidal carrier phases, determined by the baseband symbols; CPM signals are generated by loading digital information onto the phase through nonlinear modulation with memory, ensuring that the phase is a continuous function of time. The expression for a continuous-phase CPM signal is: (2) In equation (2), It is a function of the change in signal amplitude. This represents the information-carrying symbols over time. The phase function is expressed as: （3） In the formula, Indicates the modulation index. Indicates the baseband sequence. Represents the phase pulse shaping function; A CPFSK signal is modulated by a single continuously varying frequency carrier wave, with a modulation index of [missing information]. Transmitted signal baseband sequence ; SOQPSK signal is a constant envelope continuous phase modulated signal with a modulation index of 100%. , It is a pre-encoded transmission sequence, with values ​​ranging from 1 to 2. ; Phase pulse shaping functions for CPM, CPFSK and SOQPSK signals The expression is: （4）； The phase changes of SBPSK and SOQPSK signals within a symbol are similar; they can only remain unchanged or... SBPSK signal modulation index , It is a pre-coded baseband sequence, with values ​​ranging from 1 to 2. Phase pulse shaping function of SBPSK signal The width is half a symbol period, and the expression is: （5）。 3. The method for identifying P-band burst signal modulation mode according to claim 1, characterized in that, Step 2, the starting point detection method based on the frequency domain ratio method, specifically includes the following steps: First, automatic gain control (AGC) is applied to the real signal data after front-end sampling to ensure that burst signals and noise in the channel are at similar power levels. Secondly, the data after AGC is completed is subjected to quadrature downconversion and low-pass filtering to remove the second harmonic component and convert it into complex signal baseband data. Then, the obtained complex signal data is processed. Point-sliding FFT is used to obtain the frequency domain features of the data and calculate... The frequency domain ratio sequence is obtained by comparing the peak frequency points of the FFT data sorted from smallest to largest with the mean frequency points of the in-band noise floor. Finally, for the frequency domain ratio sequence, a sliding window of a preset size is selected, and each value in the sliding window is compared with a set threshold. If there are more than a specified number of frequency domain ratios in the sliding window, the starting point of the burst signal is determined.

4. The method for identifying P-band burst signal modulation mode according to claim 1, characterized in that, Step 3, the specific implementation of the image-based construction includes the following steps: Let the data point at the starting point of the signal be . Then select from the starting point. The length data points are used as the set of signal preamble data points, denoted as . ; Will After performing the FFT operation, a frequency domain sequence of the signal preamble data points is obtained. Then, a preprocessing method of spectrum filling is used to enhance edge features, resulting in a 224×224 three-channel image.

5. The method for identifying P-band burst signal modulation mode according to claim 4, characterized in that, Step 4, specifically, involves extracting the semantic features of the preamble spectrum of the burst signal, including: The preprocessed image is then processed through convolutional layers and max pooling layers to obtain an input image suitable for the basic residual unit. The network structure includes four levels of residual network layers, each of which includes three convolutional layers. Each convolutional layer is accompanied by a ReLU activation function layer, and the residual networks at each level are connected in a skip connection to form a deep network.

6. The method for identifying P-band burst signal modulation mode according to claim 1, characterized in that, In step 5, the multi-semantic feature fusion module consists of four feature preprocessing modules, an average pooling layer, and a Flatten layer; The feature preprocessing module standardizes the semantic features of the preamble spectrum of burst signals with different sizes and number of channels output by each residual layer to obtain standardized features; The standardized features are then subjected to dimensionality concatenation followed by dimensionality reduction processing using an average pooling layer and a Flatten layer to obtain multi-semantic fusion features.

Images