Radar intra-pulse modulation type recognition method and system based on lightweight deep learning

Abstract

The present invention discloses a radar intra-pulse modulation type recognition method and system based on lightweight deep learning, belonging to the field of radar signal processing. The recognition method comprises: simulating different modulation types using simulation software to obtain an original data set; performing wavelet transform on the original data set to obtain a two-dimensional time-frequency feature image, and reducing the image size using a nearest neighbor interpolation algorithm; dividing the two-dimensional time-frequency feature image set of each modulation type into a training set and a test set; inputting the training set of each modulation type into a variational autoencoder model, balancing and expanding the training set of each modulation type; inputting the balanced training set into a lightweight deep learning network model for training; using the trained network-based model to test the test set of each modulation type to obtain the recognition accuracy; inputting the two-dimensional time-frequency feature image into the trained network model, and outputting the modulation type recognition result.

Description

Technical Field

[0001] The present invention belongs to the field of radar signal processing, and specifically relates to a radar intra-pulse modulation type recognition method and system based on lightweight deep learning. Background Art

[0002] One of the core elements of radar countermeasure information processing is radar signal recognition, which determines the true attributes of a given emitter by comparing its signal parameters with pre-stored parameters in a database. Radar signal recognition typically includes emitter identification, emitter carrier identification, threat level determination, and identification confidence estimation. Emitter recognition is crucial for ensuring the full effectiveness of radar countermeasure reconnaissance and jamming. In their paper "Radar Emitter Signal Recognition Based on Radial Basis Function Probabilistic Neural Networks," Zhang Gexiang et al. extracted Lernpel-Ziv complexity and fractal dimension features from radar emitter signals and then trained a radial basis function-based BP neural network for recognition. Simulations showed that for ten typical radar emitter signals—including conventional pulse signals, linear frequency modulation signals, frequency-coded signals, and binary phase-coded signals—they could achieve 90% accuracy at signal-to-noise ratios of 5 dB or above. However, this accuracy decreased significantly with decreasing signal-to-noise ratios. However, the complexity of the sea surface environment makes acquiring radar emitter signal samples extremely difficult. This results in insufficient and unbalanced sample data, which in turn affects the neural network's recognition performance.

[0003] In summary, the deficiencies in the prior art include:

[0004] (1) Existing neural network-based recognition methods rely on sufficient and balanced radar emitter signal sample data. The complexity of the sea surface environment and the development of radar technology can lead to insufficient radar emitter signal sample data, which greatly limits the classification and recognition performance of the neural network.

[0005] (2) Existing dataset expansion methods are relatively simple, and using only simple oversampling or undersampling is difficult to meet actual needs.

[0006] (3) If the existing neural network adopts a more complex network structure, although it can achieve relatively good recognition results, it will greatly increase the amount of training calculations and the time spent on training, making it difficult to meet the needs of real-time recognition. Summary of the Invention

[0007] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide a radar intra-pulse modulation type identification method and system based on lightweight deep learning, which solves the problems in the existing technology.

[0008] The purpose of the present invention can be achieved through the following technical solutions:

[0009] The radar intra-pulse modulation type recognition method based on lightweight deep learning includes the following steps:

[0010] Use simulation software to simulate different modulation types and sample the obtained pulse data to obtain the original data set;

[0011] The original data set is subjected to wavelet transformation to obtain a two-dimensional time-frequency feature image, and the size of the two-dimensional time-frequency feature image is reduced by the nearest neighbor interpolation algorithm;

[0012] The two-dimensional time-frequency feature atlas of each modulation type is randomly sampled and divided into a training set and a test set;

[0013] Feed the training sets of each modulation type into the variational autoencoder model, and balance and expand the training sets of each modulation type;

[0014] The balanced training set is fed into a lightweight deep learning network model based on ShuffleNet for training;

[0015] The trained ShuffleNet-based lightweight deep learning network model is used to test the test set of each modulation type to obtain the recognition accuracy of each modulation type;

[0016] The two-dimensional time-frequency feature map is input into the trained lightweight deep learning network model based on ShuffleNet, and the modulation type recognition result is output.

[0017] Furthermore, the modulation types include five types, and the data models of the five modulation types are:

[0018] 1) Conventional single pulse signal:

[0019]

[0020] 2) Linear frequency modulation signal:

[0021]

[0022] 3) Nonlinear FM signal:

[0023]

[0024] 4) Two-phase encoding signal:

[0025]

[0026] 5) Two-frequency coded signal:

[0027]

[0028] Where A is the signal amplitude, fc is the carrier frequency of the signal, is the initial phase of the signal, j is the imaginary unit; N is the number of code elements, is the initial phase of the signal; θ(t) is the nonlinear frequency modulation of the signal, f i is the carrier frequency of the signal.

[0029] Furthermore, the algorithm model of the wavelet transform is:

[0030]

[0031] Where σ is the attenuation coefficient, which is proportional to the amplitude of the cross term, t is time, ω is the angular frequency, u is the integral variable, τ is the time shift parameter, s(·) is the signal to be processed, and CWD s (t, ω) is the result obtained after the operation.

[0032] Furthermore, the nearest neighbor interpolation algorithm formula is:

[0033]

[0034] Among them, src x ,src y Indicates the coordinates in the original image, dst x ,dst y Indicates the coordinates in the target image, and scale indicates the zoom factor

[0035] Furthermore, during random sampling, random sampling without replacement is performed on the two-dimensional time-frequency feature image set of each modulation type until the number of sampled samples reaches 70% of the modulation type data set, and finally a training set and test set with a ratio of 7:3 is obtained.

[0036] Furthermore, the variational autoencoder model includes:

[0037] Encoder: Receives an image input of size 28×28×1, then connects two modules consisting of a convolutional layer and a ReLU activation function. The convolution kernel size is 3×3, the number of convolution kernels in the first convolution layer is 64, and the number of convolution kernels in the second convolution layer is 128. Finally, it passes through a fully connected layer and outputs two vectors: one is the mean vector μ of the latent variable, and the other represents the logarithmic variance vector logo of the latent variable. 2 ;

[0038] Decoder: Receives latent variables as input, followed by two modules consisting of transposed convolution and ReLU activation function. The size of the convolution kernel in the transposed convolution is 3×3. The number of convolution kernels in the first transposed convolution layer is 128, and the number of convolution kernels in the second transposed convolution layer is 64. Finally, a transposed convolution layer is used to restore the size of the latent variable to 28×28×1.

[0039] Furthermore, the lightweight deep learning network model based on ShuffleNet includes, in sequence: an input layer, a module consisting of a convolutional layer and a maximum pooling layer, three Shuffle modules, a convolutional layer, a global pooling layer and a fully connected layer; the Shuffle module includes two operations: grouped convolution and channel rearrangement.

[0040] Radar intra-pulse modulation type recognition system based on lightweight deep learning, including:

[0041] Simulation module: Use simulation software to simulate different modulation types and sample the obtained pulse data to obtain the original data set;

[0042] Feature extraction module: The original data set is subjected to wavelet transformation to obtain a two-dimensional time-frequency feature image, and the size of the two-dimensional time-frequency feature image is reduced by the nearest neighbor interpolation algorithm;

[0043] Dataset partitioning module: randomly samples the two-dimensional time-frequency feature map of each modulation type and divides it into training set and test set;

[0044] Dataset expansion module: The training sets of each modulation type are fed into the variational autoencoder model, and the training sets of each modulation type are balanced and expanded;

[0045] Model training module: Send the balanced training set to the lightweight deep learning network model based on ShuffleNet for training;

[0046] Model testing module: Use the trained ShuffleNet-based lightweight deep learning network model to test the test set of each modulation type to obtain the recognition accuracy of each modulation type;

[0047] And, the recognition module: inputs the two-dimensional time-frequency feature map into the trained ShuffleNet-based lightweight deep learning network model, and outputs the modulation type recognition result.

[0048] A computer storage medium stores a readable program, which, when running, can execute the above-mentioned radar intra-pulse modulation type identification method based on lightweight deep learning.

[0049] An electronic device comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus;

[0050] The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform operations corresponding to the above-mentioned radar intra-pulse modulation type identification method based on lightweight deep learning.

[0051] Beneficial effects of the present invention:

[0052] 1. The radar intra-pulse modulation type recognition method based on lightweight deep learning under sample imbalance provided by the present invention is intended to solve the problem of sample imbalance and sample shortage between various radar pulse modulation types, which in turn limits the technical problem of deep learning network performance. It includes: using simulation software to simulate different types of modulation types, sampling the obtained pulse data to obtain the original data set; performing wavelet transform on the obtained original data set to obtain a two-dimensional time-frequency feature grayscale image of uniform size, and randomly sampling the data set of each modulation type to divide it into a training set and a test set; sending the training set of each modulation type into a variational autoencoder model, and balancing the training set of each modulation type according to a certain method; sending the balanced training set into a lightweight deep learning network model to train the required neural network model; using the trained lightweight neural network model to test the test set of each modulation type to obtain the recognition accuracy of each modulation type;

[0053] 2. The present invention is suitable for radar modulation type recognition in complex environments, has good detection effect and robustness, and meets practical needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0055] Figure 1 This is a flow chart of the radar intra-pulse modulation type identification method of the present invention.

[0056] Figure 2 This is an operational flow chart for implementing radar intra-pulse modulation type identification in the present invention.

[0057] Figure 3 It is a structural diagram of the radar intra-pulse modulation type identification system of the present invention. DETAILED DESCRIPTION

[0058] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0059] Example 1

[0060] like Figure 1 As shown, the radar intra-pulse modulation type recognition method based on lightweight deep learning includes the following steps:

[0061] S1, using Matlab simulation software to simulate different modulation types, and sampling the obtained pulse data to obtain the original data set;

[0062] For modulation type simulation, the steps for sampling pulse data include:

[0063] S11, establishing a mathematical model for each modulation type signal;

[0064] 1) For a conventional single pulse signal, the mathematical model formula is as follows:

[0065]

[0066] Where A is the signal amplitude, f c is the carrier frequency of the signal, is the initial phase of the signal, and j is the imaginary unit.

[0067] 2) For linear frequency modulation signals, the mathematical model formula is as follows:

[0068]

[0069] Where A is the signal amplitude, f c is the carrier frequency of the signal, is the initial phase of the signal, and j is the imaginary unit.

[0070] 3) For nonlinear frequency modulation signals, the mathematical model formula is as follows:

[0071]

[0072] Where A is the signal amplitude, f c is the carrier frequency of the signal, is the initial phase of the signal, θ(t) is the nonlinear frequency modulation of the signal, and j is the imaginary unit.

[0073] 4) For the two-phase coded signal, the mathematical model formula is as follows:

[0074]

[0075] Where A is the signal amplitude, f c is the carrier frequency of the signal, N is the number of code elements, is the initial phase of the signal, j is an imaginary unit, and in the present invention, the value of the phase is determined by the encoded code element, and the code element in the present invention adopts the 13-bit Barker code 1111100110101.

[0076] 5) For the dual-frequency coded signal, the mathematical model formula is as follows:

[0077]

[0078] Where A is the signal amplitude, is the initial phase of the signal, N is the number of code elements, f i is the carrier frequency of the signal, j is an imaginary unit, and in the present invention, the value of the carrier frequency is determined by the coded code element, and the code element in the present invention adopts the 13-bit Barker code 1111100110101.

[0079] According to the mathematical models of the above five modulation types of signals, the signals are simulated based on simulation software to obtain the original data set.

[0080] S12, according to the mathematical models of the five modulation types mentioned above, the signals are simulated based on the simulation software. For each modulation type, the sampling rate is unified to f s , the original data set is obtained after sampling.

[0081] S2, performing wavelet transform on the original data set to obtain a two-dimensional time-frequency feature image, and reducing the size of the two-dimensional time-frequency feature image by the nearest neighbor interpolation algorithm;

[0082] The specific steps are as follows:

[0083] S21, after obtaining the original data set, it is necessary to extract the features of the signal. The method adopted by the present invention is the Continuous Wavelet Transform (CWT) algorithm. The wavelet transform algorithm model is as follows:

[0084]

[0085] Where σ is the attenuation coefficient, which is proportional to the amplitude of the cross term, t is time, ω is the angular frequency, u is the integral variable, τ is the time shift parameter, s(·) is the signal to be processed, and CWT s (t,ω) is the result obtained after the operation.

[0086] After wavelet transform, the original one-dimensional signal data will become a two-dimensional time-frequency feature image, which reflects the curve of the signal frequency changing with time.

[0087] S22, the wavelet transform algorithm will obtain a large-sized time-frequency feature image according to the length of the signal sequence and the algorithm parameters. In order to reduce the difficulty of subsequent experiments, it is necessary to reduce the size of the time-frequency feature image through the nearest neighbor interpolation algorithm. The nearest neighbor interpolation algorithm formula is as follows:

[0088]

[0089] Among them, src x ,src y Indicates the coordinates in the original image, dst x ,dst y Indicates the coordinates in the target image, and scale indicates the zoom factor. In the present invention, the size of all two-dimensional time-frequency feature images is fixed at 28×28×1.

[0090] S3, randomly sample the two-dimensional time-frequency feature image dataset of each modulation type and divide it into training set and test set

[0091] The specific content of random sampling is: after the wavelet transform algorithm and image preprocessing operations, random sampling without replacement is performed in the two-dimensional time-frequency feature image set of each modulation type until the number of sampled samples reaches 70% of the modulation type data set, and finally a training set and test set with a ratio of 7:3 is obtained.

[0092] S4, feeding the training sets of each modulation type into the variational autoencoder model, and balancing and expanding the training sets of each modulation type;

[0093] The steps to balance and expand the training set include:

[0094] S41, build a variational autoencoder model, including: encoder and decoder

[0095] Encoder: Receives an image input of size 28×28×1, then connects two modules consisting of a convolutional layer and a ReLU activation function. The convolution kernel size is 3×3, the number of convolution kernels in the first convolution layer is 64, and the number of convolution kernels in the second convolution layer is 128. Finally, it passes through a fully connected layer and outputs two vectors: one is the mean vector μ of the latent variable, and the other represents the logarithmic variance vector logo of the latent variable. 2 Since the final output is the concatenation of two vectors, the length of the fully connected layer in the encoder is twice the length of the latent variable.

[0096] Decoder: It takes the latent variable as input, followed by two modules consisting of transposed convolution and ReLU activation function. The size of the convolution kernel in the transposed convolution is 3×3. The number of convolution kernels in the first transposed convolution layer is 128, and the number of convolution kernels in the second transposed convolution layer is 64. Finally, a transposed convolution layer is used to restore the size of the latent variable to 28×28×1.

[0097] The role of the encoder is to map the original image input to a low-dimensional latent variable space through operations such as convolution and pooling. The latent variable has a mean of μ and a variance of logo. 2 The decoder has the opposite effect of the encoder and restores the latent variables to the original input image size through operations such as transposed convolution.

[0098] S42, sending the training set of each modulation type to the constructed variational autoencoder model for training, and then using the trained variational autoencoder model to generate feature images of the corresponding modulation type to achieve expansion and balance of the training set.

[0099] S5, sending the balanced training set to the lightweight deep learning network model based on ShuffleNet to train the lightweight deep learning network model based on ShuffleNet;

[0100] The steps for training a lightweight deep learning network model include:

[0101] S51, build a lightweight deep learning network model based on ShuffleNet, which includes: an input layer (size 28×28×1), a module consisting of a convolutional layer (size 3×3) and a max pooling layer, three shuffle modules, a convolutional layer, a global pooling layer, and a fully connected layer;

[0102] The Shuffle module is the optimization core of the entire network, including two operations: group convolution and channel rearrangement.

[0103] Grouped convolution: In general convolution operations, if the number of input feature images is N and the number of convolution kernels used is M, then N×M convolution operations are required. If both the feature images and the convolution kernels are divided into g groups, the required convolution operations become This reduces the amount of computation, but the problem is that the information learned by the network will be limited.

[0104] Channel shuffling: This operation addresses the limitations of group convolution. After the feature image and convolution kernel are grouped, they are randomly shuffled before the convolution operation. This shuffling ensures that each group convolution learns as much feature image information as possible, preserving the features while reducing computational overhead.

[0105] S52, use the two-dimensional time-frequency feature images in the training set of each modulation type to train the lightweight deep learning network model, and the label is the one-hot encoding of each modulation type; after the training is completed, the required network model can be obtained.

[0106] S6, use the trained ShuffleNet-based lightweight deep learning network model to test the test set of each modulation type and obtain the recognition accuracy of each modulation type;

[0107] The steps for obtaining the recognition accuracy of each modulation type include:

[0108] For each modulation type test set, we put it into the lightweight deep learning network model after training, and do not add labels to the two-dimensional time-frequency feature image. The network model will return the modulation type that the model thinks. If the modulation type returned by the model is the same as the actual modulation type, the network model is considered to have identified it correctly, otherwise it is considered to have identified it incorrectly. For each modulation type, the number of samples Q i ,i=1,2,...,5, if the number of samples correctly identified by the lightweight deep learning network model is q i , then the recognition accuracy is:

[0109]

[0110] Where, P i is the recognition accuracy.

[0111] S7, input the two-dimensional time-frequency feature map into the trained ShuffleNet-based lightweight deep learning network model, and output the modulation type recognition result.

[0112] Example 2

[0113] In this embodiment, a specific experiment is given to verify the effect of the recognition method of the present invention;

[0114] The experimental data used was simulated using Matlab software based on a given mathematical model. It includes five types of radar intra-pulse modulation signals. To closely simulate a realistic environment, the signal-to-noise ratio (SNR) ranged from -5 to 5 dB, and the carrier frequency of each radar intra-pulse modulation signal ranged from 1000 to 4500 MHz. The data set distribution, obtained by applying a wavelet transform algorithm and randomly sampling at a 7:3 ratio, is shown in Table 1.

[0115] Table 1 Dataset distribution statistics

[0116]

[0117] In order to demonstrate the performance of the present invention, three comparative algorithms were also adopted. The first one did not perform any processing on the data set. The second one used a simple oversampling method to balance the training set to the minimum number of samples. The third one used a simple undersampling method to balance the training set to the maximum number of samples. The last one was the method adopted by the present invention. The experimental flow chart is shown in the figure below. Figure 2 The recognition accuracy of each modulation signal is shown in Table 2:

[0118] Table 2 Statistics of recognition accuracy of modulation signals

[0119]

[0120] The last column of Table 2 shows the average recognition accuracy of five types of modulation signals. It can be seen from the table that the variational autoencoder model combined with the lightweight deep learning network model based on ShuffleNet adopted in the present invention is superior to other algorithms in identifying radar intra-pulse modulation types in complex environments.

[0121] Example 3

[0122] In this embodiment, a radar intra-pulse modulation type recognition system based on lightweight deep learning is proposed, including:

[0123] Simulation module: Use simulation software to simulate different modulation types and sample the obtained pulse data to obtain the original data set;

[0124] Feature extraction module: The original data set is subjected to wavelet transformation to obtain a two-dimensional time-frequency feature image, and the size of the two-dimensional time-frequency feature image is reduced by the nearest neighbor interpolation algorithm;

[0125] Dataset partitioning module: randomly samples the two-dimensional time-frequency feature map of each modulation type and divides it into training set and test set;

[0126] Dataset expansion module: The training sets of each modulation type are fed into the variational autoencoder model, and the training sets of each modulation type are balanced and expanded;

[0127] Model training module: Send the balanced training set to the lightweight deep learning network model based on ShuffleNet to train the lightweight deep learning network model based on ShuffleNet;

[0128] Model testing module: Use the trained ShuffleNet-based lightweight deep learning network model to test the test set of each modulation type to obtain the recognition accuracy of each modulation type;

[0129] And, the recognition module: inputs the two-dimensional time-frequency feature map into the trained ShuffleNet-based lightweight deep learning network model, and outputs the modulation type recognition result.

[0130] The method of the present invention can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CDROM, RAM, floppy disk, hard disk or magneto-optical disk), or as computer code that is originally stored in a remote recording medium or a non-temporary machine-readable medium downloaded over a network and will be stored in a local recording medium, so that the method described herein can be stored in such software processing on a recording medium using a general-purpose computer, a special-purpose processor or programmable or special-purpose hardware (such as an ASIC or FPGA). It will be understood that a computer, a processor, a microprocessor controller or programmable hardware includes a storage component (e.g., RAM, ROM, flash memory, etc.) that can store or receive software or computer code, and when the software or computer code is accessed and executed by a computer, a processor or hardware, the method described herein is implemented. In addition, when a general-purpose computer accesses the code for implementing the method shown here, the execution of the code converts the general-purpose computer into a special-purpose computer for executing the method shown here.

[0131] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention, and such changes and modifications fall within the scope of the invention as claimed.

Claims

1. A radar intra-pulse modulation type recognition method based on lightweight deep learning, characterized by: The following steps are involved: Use simulation software to simulate different modulation types and sample the obtained pulse data to obtain the original data set; The original data set is subjected to wavelet transformation to obtain a two-dimensional time-frequency feature image, and the size of the two-dimensional time-frequency feature image is reduced by the nearest neighbor interpolation algorithm; The two-dimensional time-frequency feature atlas of each modulation type is randomly sampled and divided into a training set and a test set; Feed the training sets of each modulation type into the variational autoencoder model, and balance and expand the training sets of each modulation type; The balanced training set is fed into a lightweight deep learning network model based on ShuffleNet for training; The trained ShuffleNet-based lightweight deep learning network model is used to test the test set of each modulation type to obtain the recognition accuracy of each modulation type; The two-dimensional time-frequency feature map is input into the trained lightweight deep learning network model based on ShuffleNet, and the modulation type recognition result is output.

2. The radar intra-pulse modulation type identification method based on lightweight deep learning according to claim 1 is characterized in that: There are five modulation types, and the data models of the five modulation types are: 1) Conventional single pulse signal: 2) Linear frequency modulation signal: 3) Nonlinear FM signal: 4) Two-phase encoding signal: 5) Two-frequency coded signal: Where A is the signal amplitude, f c is the carrier frequency of the signal, is the initial phase of the signal, j is the imaginary unit; N is the number of code elements, is the initial phase of the signal; θ(t) is the nonlinear frequency modulation of the signal, f i is the carrier frequency of the signal.

3. The radar intra-pulse modulation type identification method based on lightweight deep learning according to claim 1 is characterized in that: The algorithm model of the wavelet transform is: Where σ is the attenuation coefficient, which is proportional to the amplitude of the cross term, t is time, ω is the angular frequency, u is the integral variable, τ is the time shift parameter, s(·) is the signal to be processed, and CWD s (t, ω) is the result obtained after the operation.

4. The radar intra-pulse modulation type identification method based on lightweight deep learning according to claim 1 is characterized in that: The nearest neighbor interpolation algorithm formula is: Among them, src x ,src y Indicates the coordinates in the original image, dst x ,dst y Represents the coordinates in the target image, and scale represents the zoom factor.

5. The radar intra-pulse modulation type identification method based on lightweight deep learning according to claim 1 is characterized in that: During random sampling, random sampling without replacement is performed on the two-dimensional time-frequency feature image set of each modulation type until the number of sampled samples reaches 70% of the modulation type data set, and finally a training set and test set with a ratio of 7:3 is obtained.

6. The radar intra-pulse modulation type identification method based on lightweight deep learning according to claim 1 is characterized in that: The variational autoencoder model includes: Encoder: Receives an image input of size 28×28×1, then connects two modules consisting of a convolutional layer and a ReLU activation function. The convolution kernel size is 3×3, the number of convolution kernels in the first convolution layer is 64, and the number of convolution kernels in the second convolution layer is 128. Finally, it passes through a fully connected layer and outputs two vectors: one is the mean vector μ of the latent variable, and the other represents the logarithmic variance vector logo of the latent variable. 2 ; Decoder: It takes the latent variable as input, followed by two modules consisting of transposed convolution and ReLU activation function. The size of the convolution kernel in the transposed convolution is 3×3. The number of convolution kernels in the first transposed convolution layer is 128, and the number of convolution kernels in the second transposed convolution layer is 64. Finally, a transposed convolution layer is used to restore the size of the latent variable to 28×28×1.

7. The radar intra-pulse modulation type identification method based on lightweight deep learning according to claim 1 is characterized in that: The lightweight deep learning network model based on ShuffleNet includes, in sequence: an input layer, a module consisting of a convolutional layer and a maximum pooling layer, three Shuffle modules, a convolutional layer, a global pooling layer, and a fully connected layer; the Shuffle module includes two operations: grouped convolution and channel rearrangement.

8. Radar intra-pulse modulation type recognition system based on lightweight deep learning, characterized by: include: Simulation module: Use simulation software to simulate different modulation types and sample the obtained pulse data to obtain the original data set; Feature extraction module: The original data set is subjected to wavelet transformation to obtain a two-dimensional time-frequency feature image, and the size of the two-dimensional time-frequency feature image is reduced by the nearest neighbor interpolation algorithm; Dataset partitioning module: randomly samples the two-dimensional time-frequency feature map of each modulation type and divides it into training set and test set; Dataset expansion module: The training sets of each modulation type are fed into the variational autoencoder model, and the training sets of each modulation type are balanced and expanded; Model training module: Send the balanced training set to the lightweight deep learning network model based on ShuffleNet for training; Model testing module: Use the trained ShuffleNet-based lightweight deep learning network model to test the test set of each modulation type to obtain the recognition accuracy of each modulation type; And, the recognition module: inputs the two-dimensional time-frequency feature map into the trained ShuffleNet-based lightweight deep learning network model, and outputs the modulation type recognition result.

9. A computer storage medium storing a readable program, characterized in that: When the program is running, it can execute the radar intra-pulse modulation type identification method based on lightweight deep learning as described in any one of claims 1 to 7.

10. An electronic device, characterized in that: include: A processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform operations corresponding to the radar intra-pulse modulation type identification method based on lightweight deep learning as described in any one of claims 1 to 7.

Images