A complex electromagnetic signal sorting method based on synchronous compression transformation and light-weight neural network

Abstract

This invention discloses a method for sorting complex electromagnetic signals based on synchronous compression transform and a lightweight neural network. First, the received time-domain sampled sequence is subjected to synchronous compression transform to obtain a highly clustered time-frequency complex matrix. Then, the magnitude, argument, phase, and spatial gradient features of this matrix are extracted. After logarithmic compression, phase mapping, and gradient normalization, three feature channels—amplitude, phase, and edge—are obtained. These channels are then concatenated to form a three-channel time-frequency feature tensor. Finally, this tensor is input into a lightweight neural network with GhostNet as its backbone. After network mapping, the signal category identification result is output. This invention overcomes the energy diffusion problem of traditional time-frequency analysis through synchronous compression transform, enriches the discriminative information through three-channel fusion, and improves accuracy while maintaining a lightweight design through pixel shifting and SimAM. It is suitable for sorting signals with low signal-to-noise ratio and complex aliasing.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to a method for sorting complex electromagnetic signals based on synchronous compression transform and lightweight neural networks. Background Technology

[0002] With the increasing complexity of the electromagnetic environment, various wireless communication signals and radar signals frequently overlap in the time and frequency domains, posing a significant challenge to signal monitoring and sorting. Signal sorting, as a key technology in electronic reconnaissance, spectrum management, and electromagnetic situational awareness, has the core task of accurately identifying the category and parameters of each signal source from massive overlapping pulse streams or continuous signals.

[0003] Traditional signal sorting methods mainly rely on pulse descriptor extraction and subsequent clustering algorithms, such as K-means and DBSCAN, or histogram methods based on pulse repetition intervals, such as CDIF and SDIF. However, these methods face significant limitations when dealing with modern complex modulated signals: First, the measurement of pulse parameters is severely distorted when multiple signals overlap, leading to a decrease in the reliability of PDW parameters; second, for signals with complex time-frequency structures, such as broadband orthogonal frequency division multiplexing signals, traditional energy-triggered pulse correlation mechanisms are difficult to effectively characterize; finally, under low signal-to-noise ratio conditions, the characteristics of weak signals are easily masked by background noise or sidelobes of high-power signals.

[0004] In recent years, artificial intelligence technologies, represented by deep neural networks, have provided a new technical approach for signal sorting. Through end-to-end learning of time-frequency images, models can automatically extract features from signals that are difficult to obtain manually. However, existing deep learning sorting methods still have the following shortcomings:

[0005] (1) In the preprocessing stage, traditional methods usually use Short Time Fourier Transform (STFT) to map one-dimensional time-series signals into two-dimensional time-frequency graphs. However, STFT is limited by the Heisenberg uncertainty principle, resulting in low time-frequency concentration, severe energy diffusion, and blurred time-frequency ridges under low signal-to-noise ratio. At the same time, the modulo operation loses the phase information of the signal, making it difficult for the network to capture the subtle features of phase-coded signals.

[0006] (2) In terms of network structure, high-performance deep models are usually accompanied by a large number of parameters and forward inference computation overhead, which is not conducive to deployment on resource-constrained embedded platforms. Although existing lightweight networks reduce the amount of computation, there is still room for improvement in recognition accuracy in complex electromagnetic environments, especially in multi-signal aliasing scenarios, where the adaptive modeling ability for long-distance spatial correlation and significant signal regions is relatively insufficient.

[0007] (3) Existing methods are relatively independent in the two stages of preprocessing and network design, lacking synergistic development and optimization for the time-frequency characteristics of the signal. There is a mutual constraint between the extraction quality of time-frequency features and the feature extraction capability of the network. It is difficult to obtain the best sorting performance by optimizing a single stage. There is an urgent need for a signal sorting method that can improve the recognition accuracy from the perspective of synergistic optimization of time-frequency representation and network structure while maintaining the lightweight model. Summary of the Invention

[0008] Objective: This invention addresses the problems of existing technologies by providing a method for sorting complex electromagnetic signals based on synchronous compression transform and a lightweight neural network. In the preprocessing stage, this method utilizes synchronous compression transform to sharpen time-frequency ridges and constructs three-channel feature tensors (amplitude, phase, and edge) to enrich the signal's discriminative information. The network structure employs a lightweight GhostNet as its backbone, integrating the SimAM parameterless attention mechanism and spatial pixel displacement mechanism to expand the receptive field and enhance salient features without increasing the number of parameters. Compared to traditional time-frequency analysis methods, this invention can effectively improve recognition accuracy in low signal-to-noise ratio and complex aliasing scenarios, while maintaining a lightweight model for easy deployment.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] A method for sorting complex electromagnetic signals based on synchronous compression transform and lightweight neural networks includes:

[0011] (1) Receive the complex electromagnetic signal to be sorted, wherein the complex electromagnetic signal is a time-domain sampling sequence containing at least one signal type;

[0012] (2) Perform synchronous compression transformation on the time-domain sampling sequence to obtain a highly clustered time-frequency complex matrix;

[0013] (3) Extract the magnitude, argument, phase and spatial gradient features of the time-frequency complex matrix respectively. After logarithmic compression, phase mapping and gradient normalization, obtain the amplitude intensity feature channel, phase feature channel and edge feature channel. Then, splice the three feature channels in the channel dimension to form a three-channel time-frequency feature tensor.

[0014] (4) Input the three-channel time-frequency feature tensor into the pre-trained lightweight neural network. The lightweight neural network uses GhostNet as its backbone. Its Ghost module embeds a spatial pixel displacement mechanism to expand the equivalent receptive field and introduces the SimAM parameterless attention mechanism to enhance salient features. After the network is mapped, the signal category recognition result is output.

[0015] Furthermore, the signal type mentioned in step (1) includes at least one of 5G NR signal, BPSK modulated signal and linear frequency modulated radar signal; the signal category identification result mentioned in step (4) is a single signal category, a multiple signal aliasing category or no signal.

[0016] Furthermore, step (2) specifically includes:

[0017] (2.1) Perform a windowed short-time Fourier transform on the time-domain sampled sequence to obtain the time-frequency complex matrix G. x (t,η);

[0018] (2.2) Estimate the instantaneous frequency at each time-frequency point based on the phase information of the time-frequency complex matrix. The instantaneous frequency estimation operator is calculated using the following formula:

[0019] ;

[0020] In the formula, For time and frequency points The estimated instantaneous frequency; Indicates taking the real part; for Regarding time The partial derivatives; This is the complex spectrum of the short-time Fourier transform after windowing.

[0021] (2.3) The energy of each frequency point in the time-frequency complex matrix is ​​shifted along the frequency axis to the position corresponding to the instantaneous frequency. The expression for the synchronous compression transformation is:

[0022] ;

[0023] In the formula, This is the complex time-frequency matrix after synchronous compression transformation; It is the Dirac function; This is the new frequency axis after compression; To provide the instantaneous frequency estimate, this formula incorporates all observed frequencies on the STFT plane. Energy transfer from point to instantaneous frequency The corresponding position.

[0024] Furthermore, step (3) specifically includes:

[0025] (3.1) After extracting the modulus of the time-frequency complex matrix, a logarithmic compression transformation is performed. The logarithmic transformation operator is:

[0026] ;

[0027] In the formula, The amplitude after logarithmic transformation; This is the regularization bias term.

[0028] After normalization and mapping, the amplitude intensity feature channel is obtained. The discrete pixel value matrix of the R channel is calculated using the following formula:

[0029] ;

[0030] In the formula, These are the discrete pixel values ​​of the R channel; and These are the current time-frequency matrices. The global minimum and maximum values; This is the rounding function.

[0031] (3.2) When extracting the argument phase of the time-frequency complex matrix, the phase estimation formula in the interval [-π, π] is:

[0032] ;

[0033] In the formula, This is the arctangent function in the four quadrants, and its return value is... interval; and These represent the imaginary and real parts of the complex number, respectively. This refers to the extracted instantaneous phase.

[0034] Subsequently, the phase feature channel is obtained by generating the pixel value matrix of the G channel through the quantization operator:

[0035] ;

[0036] In the formula, The pixel value for the G channel; Given by the phase estimation formula, after linear mapping, Convert to [0,255]

[0037] (3.3) The horizontal gradient of the amplitude intensity characteristic channel is calculated using the Sobel operator. and vertical gradient Define two Sobel convolution templates for estimating the horizontal and vertical gradients, respectively. and :

[0038] ;

[0039] In the formula, The Sobel template for horizontal gradient estimation; The Sobel template for vertical gradient estimation; both are Convolution kernel.

[0040] After obtaining the component gradients in both the horizontal and vertical dimensions, the Euclidean norm is used to calculate the edge feature channel of the combined gradient magnitude:

[0041] ;

[0042] In the formula, For the overall gradient magnitude; for The horizontal gradient value at the time-frequency point; This represents the gradient value in the vertical direction.

[0043] The edge feature channels are obtained after normalization:

[0044] (3.4) The amplitude intensity feature channel, phase feature channel and edge feature channel are spliced ​​together in the channel dimension to form a three-channel time-frequency feature tensor.

[0045] Furthermore, step (4) specifically includes:

[0046] (4.1) The lightweight neural network is based on GhostNet. Its entry point consists of a set of 3×3 standard convolutional layers, batch normalization layers and ReLU activation function. The stride is set to 2, and the input is downsampled to a scale of 112×112. Then it goes through five feature extraction stages from Stage 1 to Stage 5 in sequence. Stage 3 to Stage 5 introduce the SimAM attention mechanism. Finally, the classification result is output through global average pooling and fully connected layers.

[0047] (4.2) The spatial pixel displacement mechanism is as follows: the reference feature map generated in the first stage of the Ghost module is divided into five subgroups along the channel dimension, and the up, down, left, right and stationary operations are performed respectively. Then, the five subgroups after displacement are re-stitched in the channel dimension so that the subsequent depth separable convolution can obtain an equivalent 5×5 receptive field.

[0048] (4.3) The formula for calculating the minimum energy of a neuron in the SimAM attention mechanism is:

[0049] ;

[0050] In the formula, For target neurons The minimum energy value; These are the feature values ​​of the target neuron; and These are the mean and variance of the neighboring neurons, respectively; This is a regularization hyperparameter used to ensure numerical stability.

[0051] The Sigmoid function is used to perform nonlinear activation on the reciprocal of the energy function:

[0052] ;

[0053] In the formula, Input feature map; This is the output feature map after attention weighting; To summarize the energy values ​​of all neurons tensor; Use the Sigmoid activation function; This indicates element-wise multiplication.

[0054] (4.4) The lightweight neural network uses an improved Ghost bottleneck module as the basic feature extraction unit. The improved Ghost bottleneck module includes a first Ghost module, a SimAM attention module and a second Ghost module connected in sequence. When the stride is 1, the main branch is the above-mentioned cascaded structure, the shortcut branch is the identity mapping, and the output is the element-wise addition of the two. When the stride is 2, the main branch adds depthwise convolution downsampling, and the shortcut branch is a cascaded depthwise convolution and point convolution.

[0055] Beneficial effects: Compared with the prior art, the significant advantages of this invention are:

[0056] 1. First, in the preprocessing stage, synchronous compression transform is used to replace the traditional STFT. By rearranging and sharpening the energy in the time-frequency plane, the energy diffusion problem in time-frequency analysis is overcome, the sparsity and energy concentration of the time-frequency representation are improved, and higher quality input features are provided for the neural network.

[0057] 2. Simultaneously, a three-channel feature tensor integrating amplitude, phase, and edge was constructed. Compared with the traditional single-channel grayscale time-frequency map, this multi-feature representation can describe the time-frequency characteristics of the signal from multiple perspectives, making up for the defect of losing phase information in the traditional modulo operation and improving the separability of phase-coded signals and aliased signals under low signal-to-noise ratio conditions.

[0058] 3. Furthermore, based on the lightweight GhostNet network, a spatial pixel displacement mechanism and a SimAM parameterless attention mechanism are integrated. The spatial pixel displacement mechanism expands the equivalent receptive field through channel displacement, avoiding the sampling sparsity problem of dilated convolution. The SimAM attention mechanism enhances salient features and suppresses noise through neuron energy saliency measurement without increasing parameters. The combination of the two allows the network to improve recognition accuracy in low signal-to-noise ratio and aliasing scenarios while maintaining its lightweight design. Experimental results show that the proposed method has only 2.40M parameters and a computational cost of 0.11 GFLOPs, making it suitable for deployment on resource-constrained embedded platforms. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating an embodiment of the present invention. Detailed Implementation

[0060] This embodiment provides a method for sorting complex electromagnetic signals based on synchronous compression transform and lightweight neural networks, such as... Figure 1 As shown, it includes:

[0061] Step 1: Receive the complex electromagnetic signals to be sorted.

[0062] The receiver receives complex electromagnetic signals to be sorted. These complex electromagnetic signals are time-domain sampled sequences containing at least one signal type. In this embodiment, the signal types include at least one of 5G NR signals, BPSK modulated signals, and linear frequency modulated (LFM) radar signals, with a uniform sampling rate of 15.36MHz. For aliasing scenarios involving multiple signals, the signal components are directly superimposed in the time domain to form the signal to be sorted at the receiver. The dataset is constructed using a multi-label random combination strategy, covering eight combination states, including single signals and pairwise / triple aliasing, encompassing no-target signals, single-signal categories, and multiple-signal aliasing categories.

[0063] Step 2: Perform synchronous compression transformation on the time-domain sampling sequence to obtain a highly clustered time-frequency complex matrix.

[0064] Specifically, the time-domain sampled sequence is first subjected to a windowed short-time Fourier transform (STFT). Let the signal to be analyzed be... The window function is In this embodiment, a Hamming window is used as the window function, and the modified STFT is defined as follows:

[0065] ;

[0066] Due to the truncation effect of the window function, the energy is not concentrated at the true instantaneous frequency on the frequency axis, but rather distributed within a range limited by the bandwidth of the window function, resulting in blurred lines in the time-frequency plot. Therefore, SST uses the phase derivative of the time-frequency distribution to estimate the true instantaneous frequency of the signal, defining the instantaneous frequency estimation operator as:

[0067] ;

[0068] After obtaining the instantaneous frequency estimate, SST performs a nonlinear compression operation, which combines all frequencies on the STFT plane at the observed frequency. The energy at that location, based on its estimated value The expression for moving it to the corresponding time-frequency grid is:

[0069] ;

[0070] in For the Dirac function, This is the new frequency axis after compression. After this rearrangement and compression process, the signal energy that originally appeared as a banded dispersion in the STFT is squeezed to the vicinity of the true instantaneous frequency ridge, which significantly improves the sparsity and clustering of the time-frequency representation.

[0071] Step 3: Extract the magnitude, argument, phase, and spatial gradient features of the time-frequency complex matrix respectively. After logarithmic compression, phase mapping, and gradient normalization, obtain the amplitude intensity feature channel, phase feature channel, and edge feature channel. Then, concatenate the three feature channels along the channel dimension to form a three-channel time-frequency feature tensor.

[0072] Specifically, this step involves constructing three parallel channels:

[0073] (1) Amplitude intensity feature channel (R channel). Extract the synchronization compression matrix. The magnitude of the time-frequency amplitude response is used as the response value.

[0074] ;

[0075] To overcome the non-uniform energy distribution caused by differences in signal power, nonlinear dynamic range compression is introduced, followed by logarithmic compression:

[0076] ;

[0077] in As the regularization bias term, in this embodiment, we take... Then, after normalization mapping:

[0078] ;

[0079] The amplitude intensity characteristic channel is obtained.

[0080] (2) Phase characteristic channel (G channel). The synchronization compression matrix can be characterized as a time-varying amplitude and phase form:

[0081] ;

[0082] In the formula This refers to the instantaneous phase features that need to be extracted. The argument of each time-frequency grid point is extracted using the four-quadrant arctangent function:

[0083] ;

[0084] in and These represent the imaginary and real parts of the complex number, respectively. Due to the phase angle... The range of values ​​is After linear mapping:

[0085] ;

[0086] The phase characteristic channel is obtained.

[0087] (3) Edge feature channel (B channel). The spatial gradient of the R channel is calculated using the Sobel operator. The horizontal and vertical convolution templates are as follows:

[0088] ;

[0089] The horizontal gradient and the vertical gradient are respectively:

[0090] ;

[0091] The synthesized gradient magnitude is:

[0092] ;

[0093] Normalization:

[0094] ;

[0095] Obtain edge feature channels.

[0096] Then, the R, G, and B channels are concatenated along the channel dimension to form a three-channel time-frequency feature tensor:

[0097] ;

[0098] This tensor simultaneously contains information about the signal's energy distribution, phase changes, and geometric structure.

[0099] Step 4: Input the three-channel time-frequency feature tensor into a pre-trained lightweight neural network. The network uses GhostNet as its backbone. Its Ghost module embeds a spatial pixel displacement mechanism to expand the equivalent receptive field and introduces the SimAM parameterless attention mechanism to enhance salient features. After being mapped by the network, the signal category recognition result is output.

[0100] Specifically, the lightweight neural network uses GhostNet as its backbone. Its entry point consists of a set of 3×3 standard convolutional layers, batch normalization layers, and a ReLU activation function, with a stride of 2, downsampling the input image to a scale of 112×112. Subsequently, it passes through five feature extraction stages, from Stage 1 to Stage 5, each consisting of stacked improved Ghost bottleneck modules (IGB). Stages 3 to 5 introduce the SimAM attention mechanism. Finally, the network outputs the classification result after global average pooling and fully connected layers.

[0101] The improved Ghost bottleneck module includes a first Ghost module, a SimAM attention module, and a second Ghost module connected in sequence. When the stride is 1, the main branch is a series of the three modules, the shortcut branch is an identity mapping, and the output is the element-wise addition of the two modules. When the stride is 2, the main branch adds depthwise convolution for downsampling, and the shortcut branch is a series of depthwise convolution and pointwise convolution.

[0102] The spatial pixel displacement mechanism is as follows: Let the original feature map generated in the first stage of the Ghost module be... Group along the channel dimension:

[0103] ;

[0104] The first four subgroups are shifted up, down, left, and right respectively. The remaining channels are assigned to the stationary branch, and the shifted feature maps are reassembled as follows:

[0105] ;

[0106] This allows subsequent depthwise separable convolutions to achieve an equivalent 5×5 receptive field, thereby expanding the network's perceptual range without increasing parameters or computational cost.

[0107] The SimAM parameterless attention mechanism works as follows: for each neuron in the input feature map, its minimum energy function is calculated:

[0108] ;

[0109] in The neighborhood mean For neighborhood variance, As the regularization hyperparameter, in this embodiment, we take... The reciprocal of the minimum energy function, after being activated by the Sigmoid function, is used as the attention weight.

[0110] ;

[0111] in To summarize the energy tensor of the minimum energy function of all neurons, This represents element-wise multiplication. This mechanism does not introduce additional learnable parameters and can adaptively enhance salient signal regions while suppressing background noise.

[0112] The network was trained using the AdamW optimizer with an initial learning rate of 0.0001 and a cosine annealing learning rate scheduling strategy, with a total of 60 iterations. During training, data augmentation strategies such as random erasure, slight translation, and amplitude perturbation were applied to the input three-channel feature tensor to reduce the risk of overfitting.

[0113] Experimental results demonstrate that, under a signal-to-noise ratio (SNR) of -10 dB, the proposed method achieves a rigorous accuracy of 81.35%, representing an improvement of approximately 12.5% ​​compared to the traditional STFT preprocessing method. The model has only 2.40 M parameters and a computational cost of 0.11 GFLOPs, both lower than comparable networks. Experimental results indicate that this method can effectively identify signal categories in low SNR and complex aliasing scenarios while maintaining low storage and computational overhead, making it suitable for deployment on resource-constrained embedded platforms.

[0114] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A method for sorting complex electromagnetic signals based on synchronous compression transform and lightweight neural networks, characterized in that, include: (1) Receive complex electromagnetic signals to be sorted, wherein the complex electromagnetic signals are time-domain sampling sequences containing at least one signal type; (2) Perform synchronous compression transformation on the complex electromagnetic signal to obtain a highly concentrated time-frequency complex matrix; (3) Extract the magnitude, argument, phase and spatial gradient features of the time-frequency complex matrix respectively. After logarithmic compression, phase mapping and gradient normalization, obtain the amplitude intensity feature channel, phase feature channel and edge feature channel. Then, splice the three feature channels in the channel dimension to form a three-channel time-frequency feature tensor. (4) Input the three-channel time-frequency feature tensor into a pre-trained lightweight neural network. The lightweight neural network uses GhostNet as its backbone. Its Ghost module embeds a spatial pixel displacement mechanism to expand the equivalent receptive field and introduces a SimAM parameterless attention mechanism to enhance salient features. After being mapped by the lightweight neural network, the signal category recognition result is output.

2. The method for sorting complex electromagnetic signals based on synchronous compression transform and lightweight neural networks according to claim 1, characterized in that: The signal type mentioned in step (1) includes at least one of 5G NR signal, BPSK modulated signal and linear frequency modulated radar signal; the signal category identification result mentioned in step (4) is a single signal category or an aliasing category.

3. The method for sorting complex electromagnetic signals based on synchronous compression transform and lightweight neural networks according to claim 1, characterized in that: In step (2), a windowed short-time Fourier transform is performed on the complex electromagnetic signal to obtain... The energy is shifted along the frequency axis to the instantaneous frequency, and the formula for estimating the instantaneous frequency is as follows: ； In the formula, For time and frequency points The estimated instantaneous frequency; Indicates taking the real part; for Regarding time The partial derivatives; This is the complex spectrum of the short-time Fourier transform after windowing.

4. The method for sorting complex electromagnetic signals based on synchronous compression transform and lightweight neural networks according to claim 3, characterized in that: Step (2) also includes SST performing a nonlinear compression operation, which moves all the energy at the observation frequency on the STFT plane to the corresponding time-frequency grid according to its estimated value, and its continuous definition is as follows: ； In the formula, This is the complex time-frequency matrix after synchronous compression transformation; It is the Dirac function; This is the new frequency axis after compression; To provide the instantaneous frequency estimate, this formula incorporates all observed frequencies on the STFT plane. Energy transfer from point to instantaneous frequency The corresponding position.

5. The method for sorting complex electromagnetic signals based on synchronous compression transform and lightweight neural networks according to claim 1, characterized in that: In step (3), after extracting the modulus of the time-frequency complex matrix, a logarithmic compression transformation is performed. The logarithmic transformation operator is: ； In the formula, The amplitude after logarithmic transformation; This is a regularization bias term; After normalization and mapping, the amplitude intensity feature channel is obtained. The discrete pixel value matrix of the R channel is calculated using the following formula: ； In the formula, These are the discrete pixel values ​​of the R channel; and These are the current time-frequency matrices. The global minimum and maximum values; This is the rounding function.

6. The method for sorting complex electromagnetic signals based on synchronous compression transform and lightweight neural networks according to claim 1, characterized in that: In step (3), when extracting the argument phase of the time-frequency complex matrix, the phase estimation formula in the interval [-π, π] is: ； In the formula, This is the arctangent function in the four quadrants, and its return value is... interval; and These represent the imaginary and real parts of the complex number, respectively. That is, the extracted instantaneous phase; Subsequently, the phase feature channel is obtained by generating the pixel value matrix of the G channel through the quantization operator: ； In the formula, The pixel value for the G channel; Given by the phase estimation formula, after linear mapping, Convert to [0,255].

7. The method for sorting complex electromagnetic signals based on synchronous compression transform and lightweight neural networks according to claim 1, characterized in that: In step (3), the Sobel operator is used to calculate the horizontal gradient of the amplitude intensity feature channel. and vertical gradient Define two Sobel convolution templates for estimating the horizontal and vertical gradients, respectively. and : ； In the formula, The Sobel template for horizontal gradient estimation; The Sobel template for vertical gradient estimation; both are Convolution kernel; After obtaining the component gradients in both the horizontal and vertical dimensions, the Euclidean norm is used to calculate the edge feature channel of the combined gradient magnitude: ； In the formula, For the overall gradient magnitude; for The horizontal gradient value at the time-frequency point; This represents the gradient value in the vertical direction.

8. The method for sorting complex electromagnetic signals based on synchronous compression transform and lightweight neural networks according to claim 1, characterized in that: The spatial pixel displacement mechanism described in step (4) is as follows: the reference feature map of the Ghost module is divided into five groups along the channel, and moved up, down, left, right and stationary respectively. After re-splicing, the depth-separable convolution is used to obtain an equivalent 5×5 receptive field.

9. A method for sorting complex electromagnetic signals based on synchronous compression transform and lightweight neural networks according to claim 1, characterized in that: The formula for calculating the minimum energy of a neuron in the SimAM nonparametric attention mechanism described in step (4) is as follows: ； In the formula, For target neurons The minimum energy value; These are the feature values ​​of the target neuron; and These are the mean and variance of the neighboring neurons, respectively; This is a regularization hyperparameter used to ensure numerical stability; The Sigmoid function is used to perform nonlinear activation on the reciprocal of the energy function: ； In the formula, Input feature map; This is the output feature map after attention weighting; To summarize the energy values ​​of all neurons tensor; Use the Sigmoid activation function; This indicates element-wise multiplication.

10. A method for sorting complex electromagnetic signals based on synchronous compression transform and lightweight neural networks according to claim 1, characterized in that: The lightweight neural network described in step (4) adopts an improved Ghost bottleneck module, which includes a first Ghost module, a SimAM attention module, and a second Ghost module; when the stride is 1, the backbone is the three concatenated and the shortcut is the identity mapping; when the stride is 2, the backbone adds downsampling convolution and the shortcut is the concatenation of depth convolution and point convolution.

Images