A lightweight method for radar signal modulation type identification networks

Abstract

This invention discloses a lightweight method for radar signal modulation type recognition networks, applied in the field of radar signal processing technology. The method includes: obtaining simulated radar signals; setting different signal-to-noise ratios and modulation parameters to generate a training dataset; extracting various transform domain features of the radar signals from the training dataset; designing a radar signal modulation type recognition network, adding L1 regularization terms to the loss of all channels of the network, and training the network using the training dataset; performing channel pruning on the trained network; parallelizing the residual modules in the channel-pruned network to obtain a lightweight radar signal modulation type recognition network; and training the lightweight network using the transform domain features to obtain the trained lightweight radar signal modulation type recognition network. This invention utilizes a parallel structure and channel pruning method to reduce network parameters and size, compressing the network model with minimal loss in recognition accuracy, thus achieving network lightweighting.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing technology, and relates to radar radiation source signal sorting technology in electronic warfare, specifically to a lightweight method for radar signal modulation type identification network. Background Technology

[0002] With the widespread application of electronic technology in the military field, electronic warfare has become an important component of modern warfare. Both sides use electronic reconnaissance equipment to search for and intercept each other's electromagnetic signals, estimate their parameters, and then analyze, identify, and locate the operating status of radiation sources. This provides strong support for defense and attack on the battlefield, and helps to seize the initiative and control of the war. Radar signal modulation type identification, as a key technology in electronic reconnaissance, directly affects the accuracy of enemy radiation source identification and the reliability of strategic decision-making.

[0003] Traditional radar signal modulation type identification algorithms mainly extract time-domain and frequency-domain feature parameters, resulting in poor noise resistance and difficulty in meeting practical needs. With the development of time-frequency analysis techniques and deep learning theory, methods based on time-frequency images and convolutional neural networks have gained popularity due to their high accuracy and strong noise resistance. However, neural networks consume large amounts of storage space and have high hardware requirements, making them inconvenient for engineering applications. To address this, researchers have proposed a series of methods to compress deep networks, mainly including low-rank approximation, network quantization, weight pruning, and structural sparsity. The paper "Imagenet Dataset Classification Based on Binary Convolutional Neural Networks" ("M. Rastegari, V. Ordonez, J. Redmon, and A. Farhadi. Xnornet: Imagenet classification using binary convolutional neural networks. In ECCV, 2016.") quantizes real-valued weights into binary weights (weight values ​​are restricted to {-1, 1}), significantly compressing the model size. However, this aggressive low-rank approximation method often leads to a loss of accuracy. The paper "Learning both weights and connections for efficient neural networks" ("S. Han, J. Pool, J. Tran, and W. Daily. Learning both weights and connections for efficient neural networks. In NIPS, pages 1135–1143, 2015") prunes unimportant connections with small weights from a trained neural network, resulting in a network with mostly zero weights. This allows for storage reduction by storing the model in a sparse format. However, this method can only be accelerated using dedicated sparse matrix operation libraries and / or hardware, and the runtime memory savings are very limited. Summary of the Invention

[0004] To address the issues of large storage space requirements and high hardware demands in existing radar signal modulation type identification networks, this invention proposes a lightweight method for radar signal modulation type identification networks. This method combines channel pruning with a parallel structure, significantly reducing network parameters and computational load while maintaining network sorting accuracy.

[0005] To solve the above problems, the present invention adopts the following technical solution:

[0006] A lightweight method for radar signal modulation type identification networks includes the following steps:

[0007] S1. Simulate the original radar signal and its modulation parameters to obtain the simulated radar signal;

[0008] S2. Set different signal-to-noise ratios and modulation parameters for the simulated radar signals to generate a training dataset;

[0009] S3. Extract various transform domain features of the radar signals in the training dataset;

[0010] S4. Design a radar signal modulation type recognition network. After adding an L1 regularization term with a scaling factor of γ to the loss of all channels of the radar signal modulation type recognition network, train the radar signal modulation type recognition network using the training dataset.

[0011] S5. Perform channel pruning on the trained radar signal modulation type recognition network;

[0012] S6. Parallelize the residual modules in the network after channel pruning to obtain a lightweight network for radar signal modulation type recognition.

[0013] S7. Input the transform domain features of each radar signal in the training dataset extracted in step S3 into the lightweight radar signal modulation type recognition network for training, and obtain the trained lightweight radar signal modulation type recognition network.

[0014] Compared with the prior art, the beneficial effects of the present invention are:

[0015] This invention provides a lightweight method for radar signal modulation type identification networks. By combining channel pruning and parallelization, it can significantly reduce network complexity while maintaining the high recognition accuracy of the original network, which is beneficial for the engineering application of radar signal modulation type identification networks. This invention is characterized by flexibility, high efficiency, and strong generalization ability. Attached Figure Description

[0016] Figure 1This is a flowchart of a lightweight method for radar signal modulation type identification network according to the present invention;

[0017] Figure 2 This is a flowchart of channel pruning provided in an embodiment of the present invention;

[0018] Figure 3 This is a schematic diagram of channel selection provided in an embodiment of the present invention;

[0019] Figure 4 This is a schematic diagram of SE-ResNet module channel pruning provided in an embodiment of the present invention;

[0020] Figure 5 This is a schematic diagram of the residual module before and after parallel processing provided in an embodiment of the present invention;

[0021] Figure 6 This is a comparison chart of the recognition rates of the original network and the lightweight network with a pruning rate of 60% provided in an embodiment of the present invention. Detailed Implementation

[0022] To facilitate understanding of the technical content of this invention by those skilled in the art, the following description, in conjunction with the accompanying drawings and specific embodiments, further illustrates the content of this invention.

[0023] like Figure 1 As shown in the figure, this embodiment of the invention provides a lightweight method for radar signal modulation type identification networks, which includes the following steps:

[0024] S1. Simulate the original radar signal and its modulation parameters to obtain the simulated radar signal. The original radar signal is a time-domain signal, and the modulation parameters include bandwidth, pulse width, sampling frequency, etc.

[0025] S2. Based on the simulated radar signal obtained in step 1, set different signal-to-noise ratios and modulation parameters to generate training and test datasets. The specific dataset parameter settings for the training and test datasets are shown in Table 1. This embodiment simulates six common radar signal modulation types, including linear frequency modulation (LFM), nonlinear frequency modulation (NLFM), conventional wave (CW) signals, frequency agile (FA) signals, Barker code signals, and Costas code signals. Among them, the NLFM signal is an S-mode frequency modulation, the agile frequency signal (FA) is agile five times within the pulse width, the Barker code signal is a phase modulation signal, and its symbol sequence is [1,1,1,0,0,1,0],[1,1,1,0,0,0,1,0,0,1,0],[1,1,1,1,1,0,0,1,1,0,1,0,1]; the Costas code signal is a frequency modulation signal, and its symbol sequence is [1,6,3,4,2,5],[1,5,6,3,2,4],[2,5,6,4,3,7,1],[1,2,7,4,6,5,3],[2,6,3,8,7,5,1,4]. The simulated signals are all baseband signals with a sampling frequency of 1 GHz, a bandwidth range of 100 to 600 MHz, a pulse width of 1 μs, and a signal-to-noise ratio range of -16 dB to 16 dB, increasing in 2 dB increments.

[0026] Table 1 Radar signal dataset parameter settings table

[0027]

[0028] S3. Extract various transform domain features from the radar signals in the training and test datasets obtained in step S2. These various transform domain features refer to time domain, frequency domain, autocorrelation domain, and time-frequency domain features, specifically:

[0029] S31, Frequency Domain Transformation:

[0030] The frequency domain describes the frequency composition of a signal, showing the relationship between frequency and frequency amplitude. Overlapping signals, due to their different bandwidths or frequency distribution characteristics, can be distinguished to a certain extent in the spectrum. Performing a Fast Fourier Transform (FFT) on the signal can obtain the frequency domain characteristics of the radar signal. Let the received signal be s(t), then the FFT expression for the signal is:

[0031]

[0032] Where ω represents frequency. For a discrete-time signal sequence, let the received signal be s[n], then the expression for the discrete-time Fourier transform (DTFT) of the signal is:

[0033]

[0034] S32. Autocorrelation domain transformation:

[0035] The autocorrelation function (ACF) quantitatively measures the similarity between a signal s(t) and its time-delayed signal s(t-τ), and its expression is:

[0036]

[0037] Where τ represents the time delay. For a discrete-time series, let the received sequence be s[n], then its autocorrelation function is expressed as:

[0038]

[0039] S33, Time-Frequency Domain Transformation:

[0040] Time-frequency analysis of a signal can reflect its instantaneous frequency and reveal the patterns of frequency variation over time. This embodiment uses the pseudo-Wigner-Ville distribution (PWVD) within the Cohen-type time-frequency distribution (CWD) to extract the time-frequency features of the radar signal. The expression for CWD is:

[0041]

[0042] Where σ is the attenuation coefficient, σ∈[0.1,10]. The kernel function of PWVD is the window function h(t), and the expression for PWVD is:

[0043]

[0044] The various transform domain features extracted in step S3 can be used individually or in combination. This embodiment of the invention combines features from the time domain, frequency domain, autocorrelation domain, and time-frequency domain to form multi-domain signal features.

[0045] S4. Design a radar signal modulation type recognition network. After adding an L1 regularization term with a scaling factor of γ to the loss of all channels of the radar signal modulation type recognition network, train the radar signal modulation type recognition network using the training dataset.

[0046] Based on the multi-domain features of radar signals extracted according to embodiments of the present invention, a radar signal modulation type identification network is designed. The radar signal modulation type identification network generally includes convolutional layers, pooling layers, batch normalization (BN) layers, and fully connected layers, and also includes residual modules (ResNet), squeezing-and-excitation (SE) modules, etc.

[0047] The radar signal modulation type recognition network consists of sequentially connected convolutional layers, three SE_ResNet modules, pooling layers, a feature fusion SE structure, and fully connected layers. The pooling layers use global average pooling. Each SE_ResNet module includes a convolutional layer Conv k_x and a subsequent compression and excitation module SE k_x. The convolutional layer Conv k_x is composed of three cascaded blocks, each block containing three convolutional layers, where k = 2, 3, and 4. The overall structure of the radar signal modulation type recognition network is shown in Table 2. Conv represents a convolutional layer, and its structure is represented by "kernel size, number of feature channels, stride". The second, third, and fourth convolutional layers are each composed of three identical SE-ResNet modules cascaded together. The structure of each block is represented in [·], where the i-th row represents the i-th convolutional layer in the SE-ResNet module (i = 1, 2, 3). The structure of the SE module is represented by "number of input feature channels, compression coefficient". Conv k_x and SE k_x together form the k-th SE-ResNet module (k = 2, 3, 4). Pool represents a pooling layer, using global average pooling. The last SE represents the SE structure of the feature fusion part, with a compression coefficient ratio = 16. FC represents a fully connected layer, outputting the recognition result. The number of parameters in this radar signal modulation type recognition network is approximately 1.76 × 10⁻⁶. 6 The computational workload is approximately 1.25 × 10⁻⁶. 7 At a signal-to-noise ratio of -4dB, the accuracy of radar signal identification reaches over 92%.

[0048] Table 2. Network Structure for Radar Signal Modulation Type Identification

[0049]

[0050] S5. Perform channel pruning on the radar signal modulation type recognition network trained in step S4.

[0051] The channel pruning described in this step is a type of feature pruning, where each channel is treated as a feature. The overall process is divided into three stages: training, pruning, and fine-tuning. The channel pruning process is as follows: Figure 2As shown, firstly, the importance of each channel in the convolutional layer is obtained by training the original network; then, channels with lower importance are pruned according to their importance; finally, the network is fine-tuned to improve the accuracy of the pruned network. Figure 2 As shown, the channel pruning process specifically includes the following steps:

[0052] S51. Obtain the scaling factor from the BN layer to measure the channel importance weight.

[0053] In this step, the scaling factor of the BN layer in the trained radar signal modulation type recognition network is obtained, and the channel importance weight is measured based on the scaling factor.

[0054] After adding an L1 regularization term with a scaling factor of γ to the loss of all channels of the radar signal modulation type recognition network and training it, the importance of each channel is determined by the magnitude of this scaling factor γ. The network loss result after adding the scaling factor is as follows:

[0055]

[0056] Where (x,y) represents the training input and recognition result; l(x,y) represents the loss function without scaling; λ is the regularization coefficient, which is generally a very small value close to 0; L1 regularization term L1(γ)=|γ|, where γ represents the scaling coefficient of each channel. It can be seen from formula (11) that the closer γ is to 0, the lower the importance of the corresponding channel. At this time, γ is optimized together with the network loss function.

[0057] Adding a batch normalization (BN) layer to a network can optimize convergence speed and suppress gradient problems through data normalization. Furthermore, the scaling factor γ in the BN layer can be directly used as the basis for network channel pruning without increasing the number of network layers. The formula for linearly transforming the output of the BN layer is:

[0058]

[0059] Where γ and β represent scaling and translation factors, respectively; This represents the normalized result for the current batch. Let the current layer be a Batch Normalization (BN) layer, and let there be N samples in a batch. The mean of the current batch is μ. B The variance is After standardization of the current batch, the output of the BN layer... for:

[0060]

[0061] Where, x iFor the input of the BN layer, ε is a very small positive number to prevent the denominator from being 0.

[0062] Adding a scaling factor to the network layers is typically done when calculating the gradient of the Batch Normalization (BN) layer. Since the loss function needs to be differentiated during backpropagation, the added L1 regularization term, after differentiation, becomes:

[0063]

[0064] Therefore, when calculating the gradient, it is only necessary to add sign(γ) to the parameters of the BN layer.

[0065] S52, Channel Selection

[0066] In this step, all scaling factors γ are sorted, and channels with smaller γ values ​​are pruned. After counting and sorting all scaling factors γ, a threshold for channel selection is set based on the pruning rate set for the network model. Then, the scaling factor γ of each BN layer is compared with the threshold. If the scaling factor γ of the current channel is greater than the threshold, the channel is retained; if the scaling factor γ of the current channel is less than the threshold, the channel is pruned. When a channel is pruned, its upper-layer connections and lower-layer connections are also deleted. If the number of channels before pruning is C, the number of channels after pruning is C0. γ Channel selection diagram as shown below Figure 3 As shown, the lighter the grayscale in the γ legend, the smaller the value of the scaling factor γ of the current channel; by marking the channels that need to be deleted as 0 and the remaining channels as 1, the current feature channels can be binarized into [1,0,1,1,0,1,1,1].

[0067] S53, Model Reconstruction and Fine-tuning

[0068] After pruning and reconstruction, the model will lose recognition accuracy. Therefore, it is necessary to fine-tune the model reconstructed in step S52 to gradually restore the signal recognition accuracy and retrain it to obtain the optimal weights of the channel-pruned network. Specifically, the model after channel pruning is retrained to obtain the optimal weights and improve signal recognition accuracy. During fine-tuning, the same optimization settings (learning rate, etc.) as in the training process can be used, or they can be reset according to constraints such as training time.

[0069] After one pruning cycle, if it is necessary to further reduce the number of network parameters, the scaling factor of the BN layer can be obtained again after obtaining the optimal weights of the network after channel pruning, and secondary or even multiple pruning cycles can be performed. The importance of the feature channels can then be recalculated using the fine-tuned network model.

[0070] For the residual network in this embodiment of the invention, let the number of output channels of the previous residual block after pruning be... The number of input channels for the current residual block is the same as that of the original block, which is also... Number of output channels The network structure and the number of channels in each network layer after pruning are as follows: Figure 4 As shown, the Batch Normalization (BN) layer following the three convolutional layers on the left side of the residual block is used to obtain scaling factors and perform channel selection. The number of channels after pruning is as follows: The sampling convolutional layer in the right residual "shortcut" is only used to adjust the size of the input feature map, and the subsequent BN layer is not used as a basis for channel selection.

[0071] For the SE module, compared to convolutional layers, the SE module has a very small number of parameters and computational cost. Therefore, in this embodiment of the invention, no additional pruning is performed on the SE module; only the input dimension of the SE module changes with the number of feature channels after pruning. However, since the fully connected layers in the SE module have a process of reducing the feature dimension to its recovery state, when the input dimension after pruning is small, to ensure that the compressed output dimension is not zero, the number of output channels is:

[0072]

[0073] Where ceil(·) represents rounding up; ratio represents the compression factor in the SE module.

[0074] S6. Parallelize the channel-pruned network obtained in step S5. This step mainly targets the residual modules in the network, resulting in a lightweight radar signal modulation type recognition network. The parallelization process for the residual modules includes the following steps:

[0075] S61. First, determine the number of parallel inputs c of the residual modules, and divide the high-dimensional input with M channels into multiple low-dimensional inputs with M / c channels.

[0076] S62. Next, perform the same convolution operation on multiple parallel low-dimensional inputs.

[0077] S63. Finally, the outputs of all low-dimensional inputs are stacked along the feature channel dimension to restore the number of channels, resulting in the residual module after parallel processing. A schematic diagram of the residual module before and after parallel processing is shown below. Figure 5 As shown, where Figure 5 (a) is a schematic diagram of the residual module before parallel processing. Figure 5 (b) is a schematic diagram of the residual module when the number of parallel operations is 8.

[0078] S7. Training the lightweight radar signal modulation type recognition network obtained in step S6: Specifically, input the transform domain features of each radar signal in the training dataset extracted in step S3 into the lightweight radar signal modulation type recognition network obtained in step S6 for training, and obtain the trained lightweight radar signal modulation type recognition network.

[0079] In this step, training the lightweight network for radar signal modulation type recognition includes the following steps:

[0080] S71. First, input the transform domain features of each radar signal in the training dataset into the lightweight radar signal modulation type recognition network for forward propagation, and calculate the cost function value.

[0081] In this embodiment of the invention, the cross-entropy function is used as the network loss function, and its expression is:

[0082]

[0083] Where W and b represent the matrices of the weights and biases to be trained in the network, respectively; ρ i p represents the true probability that the current sample belongs to the i-th class. i This represents the probability that the current sample predicted by the network belongs to the i-th class, and M represents the number of channels.

[0084] S72. Then, the parameters of the lightweight network for radar signal modulation type identification are updated using a gradient descent-based backpropagation algorithm.

[0085] S73. Finally, the backpropagation process is repeated until the cost function converges, thus obtaining the trained lightweight network for radar signal modulation type recognition.

[0086] Furthermore, to verify the classification and recognition performance of the lightweight radar signal modulation type recognition network after training, the following steps are included after step S7:

[0087] S8. The lightweight radar signal modulation type recognition network trained in step S7 is used to identify the test dataset obtained in step S2. Specifically, after the original radar signals in the test dataset generated in step S2 are processed by extracting multiple transform domain features in step S3, the transform domain features of each original radar signal in the test dataset are input into the trained lightweight radar signal modulation type recognition network for forward propagation to obtain the classification and recognition results of the test samples.

[0088] This invention provides a lightweight method for radar signal modulation type identification networks. By combining channel pruning and parallelization, it can significantly reduce network complexity while maintaining the high recognition accuracy of the original network, which is beneficial for the engineering application of radar signal modulation type identification networks. This invention is characterized by flexibility, efficiency, and strong generalization ability.

[0089] To demonstrate the superior performance of the proposed lightweight network method, this invention conducts performance experiments to verify the lightweight radar signal modulation type recognition network after channel pruning and parallelization. With a parallelization quantity c=8, the network parameter quantity and computational complexity under pruning rates of 40% to 70% are shown in Table 3. The recognition accuracy comparison for signals from -8dB to 4dB is shown in Table 4. When the pruning rate is less than 50%, the recognition accuracy is not less than or slightly less than that without pruning; when the pruning rate is greater than 60%, although the recognition accuracy decreases slightly, the network complexity is further reduced. When the pruning rate is 60%, the recognition rate comparison curves for the two residual networks—the original network (SE-ResNet) and the lightweight network (SE-ResNeXt)—from -12dB to 8dB are shown in Table 4. Figure 6 As shown in the experiment, after network channel pruning and the use of parallel residual modules, the network complexity is further reduced, with an overall parameter compression rate of 82.1% and a computational compression rate of 81.6%. While the recognition rate is reduced by 5.6% at -4dB, the recognition accuracy remains above 92% at levels above -2dB. Therefore, the lightweight radar signal modulation type recognition network method proposed in this invention can significantly reduce network complexity while maintaining network accuracy.

[0090] Table 3. Network complexity under different pruning rates when the number of parallel processes is 8.

[0091]

[0092] Table 4 Comparison of recognition accuracy under different pruning rates when the number of parallel operations is 8.

[0093]

[0094] Those skilled in the art will recognize that the examples described herein are intended to help the reader understand the principles of the invention and should be understood as, but not limited to, the specific statements and embodiments described. Various modifications and variations can be made to this invention by those skilled in the art. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention are included within the scope of the claims of this invention.

[0095] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0096] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for radar signal modulation type recognition network lightweight, characterized in that, The method comprises the following steps: S1, simulating original radar signals and modulation parameters thereof to obtain simulated radar signals; S2, setting different signal-to-noise ratios and modulation parameters for the simulated radar signals to generate a training data set; S3, extracting a plurality of transform domain features of the radar signals in the training data set; S4, designing a radar signal modulation type identification network, adding an L1 regularization term with a scaling factor of γ to the loss of all channels of the radar signal modulation type identification network, and training the radar signal modulation type identification network by using the training data set; S5, performing channel pruning on the trained radar signal modulation type identification network; S6, performing parallel processing on the residual modules in the channel-pruned network to obtain a radar signal modulation type identification lightweight network; S7, inputting the transform domain features of each radar signal in the training data set extracted in step S3 into the radar signal modulation type identification lightweight network for training to obtain a trained radar signal modulation type identification lightweight network.

2. The method of claim 1, wherein, The radar signal modulation type identification network comprises a convolutional layer, a pooling layer, a batch normalization layer and a fully connected layer, and further comprises a residual module, a compression and excitation module.

3. The method of claim 2, wherein, The radar signal modulation type identification network comprises a convolutional layer, three SE_ResNet modules, a pooling layer, a feature fusion part SE structure and a fully connected layer connected in sequence, the pooling layer adopts global average pooling, each SE_ResNet module comprises a convolutional layer Conv k_x and a compression and excitation module SE k_x located behind the convolutional layer, the convolutional layer Conv k_x is cascaded by three blocks, and each block comprises three convolutional layers, wherein k=2, 3, 4.

4. The method of claim 1, wherein, The process of performing channel pruning on the trained radar signal modulation type identification network comprises the following steps: S51, obtaining the scaling factors of the BN layers in the trained radar signal modulation type identification network; S52, statistically analyzing and sorting all the scaling factors, setting a channel selection threshold according to a set pruning rate, and then comparing the scaling factors of each BN layer with the threshold in sequence, if the scaling factor of the current channel is greater than the threshold, the channel is retained; if the scaling factor of the current channel is less than the threshold, the channel is pruned, and when a channel is pruned, the upper and lower connections of the channel are also deleted; S53, fine-tuning the model reconstructed in step S52 and retraining to obtain the optimal weights of the channel-pruned network.

5. The method of claim 4, wherein, After obtaining the optimal weights of the channel-pruned network, the scaling factors of the BN layers are reacquired for secondary pruning.

6. The method of claim 1, wherein, The process of performing parallel processing on the residual modules in the channel-pruned network comprises the following steps: S61, determining the parallel number c of the residual modules, and dividing the high-dimensional input with a channel number of M into a plurality of low-dimensional inputs with a channel number of M / c; S62, performing the same convolution operation on the plurality of parallel low-dimensional inputs; S63, stacking the output results of all the low-dimensional inputs in the feature channel dimension to restore the channel number and obtain the parallel processed residual module.

7. The method of claim 1, wherein, Step S7 comprises the following steps: S71, input the transformed domain features of each radar signal in the training data set extracted in step S3 into the radar signal modulation type identification lightweight network for forward propagation, and calculate the cost function value; S72, update the parameters of the radar signal modulation type identification lightweight network using the gradient descent-based back propagation algorithm; S73, iterate the back propagation process until the cost function converges, thereby obtaining the trained radar signal modulation type identification lightweight network.

Images