Intelligent radar signal identification method under multipath fading condition

Abstract

The invention discloses an intelligent radar signal recognition method under a multipath fading condition, is applied to the field of radar signal processing, and provides a multipath suppression convolutional neural network model based on an attention mechanism for solving the problem that the recognition rate of intercepted radar signals is low due to multipath fading. The method is characterized in that signal energy is gathered through synchronous compression Fourier transform; extracting features of different scales through a multi-level feature extraction module; signal features are enhanced through the adaptive multipath effect suppression module, and the influence of components generated by the multipath effect on signal recognition is reduced, so that effective recognition of the radar signal modulation type under the multipath fading condition is realized, and certain robustness is achieved.

Description

Technical Field

[0001] The present invention belongs to the field of radar signal processing, and particularly relates to a radar emitter signal recognition technology in electronic countermeasure. Background Art

[0002] The recognition of radar signal modulation type aims to obtain the modulation information of the signal intercepted by the reconnaissance receiver, which is an important basis for sorting and emitter recognition, and has become one of the important research contents in the field of electronic reconnaissance.

[0003] The radar signal is affected by noise and multipath fading during the transmission process, resulting in distortion of the intercepted signal. Multipath fading is an inevitable problem in wireless signal transmission. After the signal arrives at the receiving end through multiple paths, the signal superposition forms a multipath effect, which will cause changes such as time shift, phase distortion, and amplitude attenuation, leading to a sharp decline in the recognition performance of existing methods. For communication signals, the research on modulation recognition under multipath fading can reduce the influence of the channel through methods such as equalization or channel compensation. However, the radar signal has a shorter duration and more diverse modulation methods. In actual application scenarios, the multipath fading effect also causes many difficulties in the recognition of radar signals.

[0004] The literature "Kong G, Jung M, Koivunen V. Waveform Recognition In Multipath Fading Using Autoencoder And CNN With Fourier Synchrosqueezing Transform. 2020 IEEE International Radar Conference (RADAR). IEEE, 2020: 612 - 617." equalizes the received signal to mitigate the influence of channel multipath fading, and then extracts and classifies the features of the equalized signal through a denoising autoencoder to achieve the recognition of radar signal modulation types. However, this method can recognize fewer signal types and has a large computational amount. The literature "Huynh - The T, Doan V S, Hua C H, et al. Accurate LPI Radar Waveform Recognition With CWD - TFA For Deep Convolutional Network. IEEE Wireless Communications Letters, 2021, 10(8): 1638 - 1642." effectively recognizes signals through the deep convolutional neural network by aggregating the extracted feature sets and implementing skip connections. However, this method only considers the influence of fading on the signal and does not consider the multipath effect. The literature "Chen Boze. Research on the Recognition Algorithm of FH Signal Modulation Modes under Multipath Channels. Harbin Engineering University, 2022." screens out the eigenvalue less affected by multipath interference through feature analysis and constructs a classifier through support vector machines to achieve the effective recognition of FH signals under multipath channels. However, this method is only for the recognition of FH signals and it is difficult to achieve the recognition of more modulation types. Therefore, the modulation recognition of radar signals under multipath fading conditions urgently needs to be studied. Summary of the Invention

[0005] To solve the above - mentioned technical problems, the present invention adopts an intelligent radar signal recognition method under multipath fading conditions, and realizes the effective recognition of radar signal modulation types under multipath fading conditions.

[0006] The technical solution adopted by the present invention is as follows: An intelligent radar signal recognition method under multipath fading conditions, including:

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

[0008] S2. Simulate the multipath fading channel and parameters;

[0009] S3. Pass the radar signal simulated in step S1 through the channel simulated in step S2 to generate a radar modulation signal under multipath fading;

[0010] S4. Extract the time-frequency domain features of the radar modulation signal obtained in step S3 to obtain the training set data;

[0011] S5. Construct an AMSCNN network, which successively includes a plurality of cascaded multi-level feature extraction modules and a probability classification sub-module;

[0012] The multi-level feature extraction module includes a cascaded basic feature extraction module and an adaptive multipath effect suppression sub-module; the basic feature extraction module includes a convolution module and a downsampling module connected in sequence; the convolution module includes two cascaded convolution blocks, each of which has the same structure and includes a convolution layer, a normalization layer, and an activation function connected in sequence; the downsampling module includes a cascaded convolution layer and an activation function; the adaptive multipath effect suppression sub-module includes a global average pooling layer, a first fully connected layer, a ReLU activation function, a second fully connected layer, and a Sigmoid activation function connected in sequence;

[0013] S6. Use the training data set processed in step S4 to train the AMSCNN network constructed in step S5;

[0014] S7. Input the time-frequency domain features of the radar signal to be recognized into the trained AMSCNN network to obtain the recognition result.

[0015] Advantages of the present invention: Aiming at the problem of radar signal modulation recognition under multipath fading that cannot be avoided in the actual propagation process, the present invention proposes a recognition method that adapts to signal distortion under multipath fading conditions. By constructing a multi-level feature extraction module to extract the features of distorted signals under multipath fading, and at the same time using an adaptive multipath effect suppression module to suppress the signal components caused by time delay, enhance the signal features, reduce the influence of multipath effects on the recognition result, and achieve effective recognition of various radar signals in a multipath fading environment. The method of the present invention has the advantages of flexibility, accuracy, and strong robustness. Description of the Drawings

[0016] Figure 1 It is a flow chart of a radar signal modulation type recognition scheme provided by an embodiment of the present invention;

[0017] Figure 2 It is a structural diagram of an AMSCNN network provided by an embodiment of the present invention;

[0018] Figure 3 It is a structural diagram of a multi-level feature extraction module provided by an embodiment of the present invention;

[0019] Figure 4 It is a structural diagram of an adaptive multipath effect sub-suppression module provided by an embodiment of the present invention;

[0020] Figure 5 This is the radar signal modulation recognition result diagram provided for the embodiments of the present invention under multipath fading. Specific Embodiments

[0021] To facilitate the understanding of the technical content of the present invention by those skilled in the art, the following further explains the content of the present invention with reference to the accompanying drawings.

[0022] As Figure 1 shown in the flowchart of the solution of the present invention, the technical solution of the present invention is: a method for radar signal modulation recognition based on a multipath suppression convolutional neural network, including:

[0023] S1. Simulate the original radar signal and modulation parameters. The original radar signal is an instantaneous complex signal in the time domain, and the parameters include bandwidth, pulse width, sampling frequency, etc.

[0024] S2. Simulate the multipath fading channel and parameters. The channel is a Rice multipath fading channel, and the channel parameters include delay, Doppler frequency shift, number of paths, etc.

[0025] S3. Pass the radar signal simulated in step S1 through the channel simulated in step S2 to generate a radar modulation signal under multipath fading, and generate training set data and test set data.

[0026] As shown in Tables 1 and 2 are the specific dataset parameter settings of the embodiments of the present invention. Table 1 is the parameters of the original radar signal. The present invention simulates eleven typical radar signal modulation types, including single carrier frequency, linear frequency modulation (LFM, Linear Frequency Modulation), binary frequency shift keying (Costas), binary phase shift keying (BPSK, Binary Phase Shift Keying), five polyphase codes (Frank, P1, P2, P3, and P4), composite modulation of linear frequency modulation and BPSK, and composite modulation of Costas and BPSK. The simulated signals are all baseband signals, the sampling frequency f s is 100 MHz, the number of sampling points is set to 1024, and the carrier frequency f c range is U[f s / 6, f s / 4], the bandwidth Band range is U[f s / 20, f s / 15], U represents a uniform distribution, and the signal-to-noise ratio varies in the range of -10 dB to 10 dB.

[0027] Table 1 Parameters of the original radar signal

[0028]

[0029] Table 2 shows the multipath fading channel parameters, including the number of paths N, path delay D, path gain G, Rice factor K, and maximum Doppler shift S.

[0030] Table 2 Multipath Fading Channel Parameters

[0031] Channel parameter Value N {1，2，3} D U[2e-7, 5e-7] G {-8，-7，-6，-5，-4} K U[5，15] S U[1000，3000]

[0032] S4. Extract the time-frequency domain features of the overlapping radar signals in step S3. Specifically, for a given training set, use the synchrosqueezed Fourier transform (FSST) algorithm to extract the time-frequency features. The FSST calculation formula is:

[0033]

[0034] where t represents time, f represents the frequency before FSST reassignment, ω represents the frequency after FSST reassignment, S(t, f) represents the short-time Fourier transform, g(0) represents the value of the sliding window function g(t) at time 0, δ is the Dirac delta function, is the instantaneous frequency, defined as:

[0035]

[0036] where Re(·) represents taking the real part, and i represents the imaginary unit.

[0037] For the radar signals affected by multipath fading generated in step S3, due to time delay, the signals propagated along each path are intercepted out of sync, and due to the fading phenomenon, there will be a superposition of original signal components with different intensities. Also, due to the Doppler frequency shift generated by the reflection path and the direct path, there is a frequency difference, and the instantaneous frequency will show a certain deviation. Compared with analyzing the time domain or frequency domain of the signals affected by multipath fading separately, time-frequency domain analysis can better display the signal features. Using the synchrosqueezed Fourier transform (FSST) to transform the radar signals, compared with the short-time Fourier transform (STFT), the FSST time-frequency features are clearer, the energy is concentrated, the frequency resolution is high, and it is not easily affected by the signal components generated by multipath, making it suitable for identifying the modulation types of signals under multipath fading.

[0038] S5. Construct a multi-level feature extraction basic module, specifically: construct a multi-level feature extraction basic module that extracts features at different levels, and extract the feature representations at different levels in the time-frequency map. Among them, the convolutional sub-module contains two convolutional blocks. The first convolutional block realizes the increase in the number of channels, helping the network extract more image features in order to obtain more variation and detail information. The second convolutional kernel further extracts and strengthens features while maintaining the spatial size, enhancing the expression ability of the network. Each convolutional block consists of a convolutional layer, a normalization layer, and an activation function. By continuously stacking these two types of convolutional blocks, the network continuously extracts and combines the multi-level features extracted. Through the continuous deepening of the feature depth, the effective extraction of data features at different levels is achieved. The downsampling sub-module contains a 3×3 convolutional layer with a stride of 2 and an activation function, which realizes the extraction and compression of high-level features of the image while reducing the spatial resolution of the feature map.

[0039] As Figure 3 shown is the specific structure of the multi-level feature extraction basic module of the embodiment of the present invention. Among them, Conv represents the convolutional layer, BN represents batch normalization, and LeakyReLU represents the LeakyReLU activation function. The detailed processing process is as follows:

[0040] S51. Input the time-frequency map of the given sample into the multi-level feature extraction basic module. First, input it into the convolutional sub-module, and pass through two convolutional layers with a convolutional kernel size of 3×3, a BN layer, and a LeakyReLU activation function;

[0041] S52. Input the output of the convolutional sub-module in step S51 into the downsampling sub-module to obtain the output result of the basic module;

[0042] S6. Construct an adaptive multipath effect suppression sub-module, specifically: a global average pooling layer, 2 fully connected layers, a ReLU activation function, and a Sigmoid activation function. Take the output of the downsampling sub-module in step S5 as the input to obtain an updated feature map.

[0043] The signal is affected by the multipath effect during transmission, carrying multiple signal components, and at the same time, phenomena such as phase distortion and frequency offset occur. There are signal components with different time delays and blurring phenomena caused by different frequency offsets on the time-frequency map, which seriously affect signal recognition. In order to suppress the influence of the multipath effect, the present invention designs an Adaptive multipath effect suppression module (AMS) based on the attention mechanism. This module can automatically adjust the weights of the feature map through iterative learning, enhance the feature map useful for signal recognition, and at the same time suppress the useless features caused by noise and multipath fading.

[0044] As Figure 4This is the structural diagram of the specific adaptive multipath effect suppression module of the present invention. Taking the output of the downsampling sub-module in step S5 as the input, it sequentially passes through the global average pooling layer GAP, the fully connected layer FC, the ReLU activation function, the fully connected layer FC, and the Sigmoid activation function, and outputs the updated feature map. The detailed processing process is as follows:

[0045] S61. Input the output of the downsampling sub-module in the basic feature extraction module into the adaptive multipath effect suppression module, and obtain the global information of each channel feature map through global average pooling operation;

[0046] S62. Input the output of global average pooling in step S61 into the fully connected layer, and increase the non-linearity through the ReLU activation function;

[0047] S63. Input the output in step S62 into the second fully connected layer, and perform normalization through the Sigmoid function;

[0048] S64. Update the feature map through the adaptive suppression operation on the output in step S63 to obtain the result after suppressing multipath.

[0049] S7. Construct a multi-level feature extraction module, specifically: the multi-level feature extraction module includes a basic feature extraction module and an adaptive multipath effect suppression sub-module.

[0050] Due to the influence of noise and multipath fading, the time-frequency map of the radar signal is not energy-concentrated compared with the ideal signal, and the components caused by noise and multipath are scattered in the time-frequency matrix. Therefore, how to effectively extract signal features is the key to recognition. In order to reduce the influence of noise and multipath fading and obtain the high-dimensional representation of the time-frequency features of the signal, the multi-level feature extraction module extracts features of different levels through the stacking of convolutional layers, gradually reduces the feature dimension, and extracts more abstract and deeper feature representations; and cascades a multipath effect suppression module after the convolutional layer to adaptively enhance the important information in the feature map.

[0051] S8. Construct a probability classification sub-module, specifically: adopt the structure of a convolutional block, a fully connected layer, a ReLU activation function, and a Sigmoid activation function, and take the output of the last multi-level feature extraction module in step S7 as the input to obtain the posterior probability.

[0052] S9. Combine the sub-modules constructed in steps S5 - S8 into the AMSCNN network, specifically: cascade five multi-level feature extraction modules including the adaptive multipath effect suppression module, and finally cascade the probability classification sub-module to form the AMSCNN network. The input is the time-frequency map of the original radar signal under the multipath fading condition in step S4, and the output is the classification result.

[0053] Such as Figure 1The flowchart of the solution of the present invention is shown. After the original radar signal time series passes through the multipath fading channel in step S3 and is processed in step S4, a two-dimensional time-frequency diagram is obtained. After the time-frequency diagram is input into the AMSCNN network, the posterior probability corresponding to each category can be obtained through the probability sub-module, and the predicted category is obtained through the fully connected layer. As Figure 2 It is the structural diagram of the AMSCNN network.

[0054] S10. Train the AMSCNN network, specifically: the time-frequency diagrams of the radar signals under each multipath fading in the training set processed by S4 are input into the AMSCNN network constructed in step S9 for forward propagation, and the cost function value is calculated; the parameters of the AMSCNN network are updated using the backpropagation algorithm based on gradient descent; the backpropagation process is iterated until the cost function converges, so as to obtain the trained AMSCNN network. Specifically, it includes the following steps:

[0055] S101. Forward propagation;

[0056] S102. Calculate the cost function value. In the example of the present invention, the cross-entropy loss function is used as the cost function, and the calculation method is:

[0057]

[0058] where p i is the true label of the i-th category, q i is the probability that the model predicts the i-th category, and C is the number of categories.

[0059] S103. Update the network parameters using the backpropagation algorithm based on gradient descent.

[0060] S11. Use the AMSCNN network trained in step S10 to identify the modulation type of the radar signals under the multipath fading in the test set processed in step S4, so as to perform a performance test on the trained model. Specifically: the radar signals under the multipath fading generated in step S3 are processed in step S4 and then input into the trained AMSCNN network for forward propagation to obtain the classification result. The specific performance test process is common knowledge in machine learning and will not be elaborated in the present invention.

[0061] In practical applications, the time-frequency domain features of the radar signals are input into the trained AMSCNN network to obtain the corresponding recognition results.

[0062] In order to show the recognition performance of the AMSCNN network proposed by the present invention, the present invention compares the recognition accuracies of AMSCNN with AlexNet, VGG, and GoogleNet. As Figure 5 The recognition result of the modulation type in the example of the present invention is shown.

[0063] From Figure 5 It can be seen that compared with the three classic networks VGG, AlexNet, and GooLeNet, AMSCNN has a higher overall recognition rate for signals in the case of multipath fading, and reaches a recognition accuracy of 100% at 2 dB. When the signal-to-noise ratio is -6 dB and above, the accuracy exceeds 95%. Even under the influence of multipath fading and Gaussian white noise with a signal-to-noise ratio of -10 dB, the accuracy can still reach more than 88%, exceeding the accuracy of the other three algorithms by more than 2%. Therefore, compared with the three classic networks, this method has better performance in the case of low signal-to-noise ratio. The recognition results of the multi-path fading signals by the networks with different structures mentioned above show that the adaptive multi-path effect suppression module with an attention mechanism can effectively suppress the interference of multi-path fading and Gaussian white noise on the signals, and reduce the influence of channel conditions on signal recognition; the multi-level feature extraction module can better enhance the network's ability to express features, effectively extract the signal features of the shallow and deep layers, and achieve a high recognition accuracy of the network for signals.

[0064] Those of ordinary skill in the art will realize that the embodiments described herein are for helping the reader understand the principles of the present invention, and it should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. For those skilled in the art, various changes and modifications can be made to the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the scope of the claims of the present invention.

Claims

1. An intelligent radar signal recognition method under multipath fading conditions, characterized in that: include: S1, simulating the original radar signal and modulation parameters to obtain a simulated radar signal; S2, simulate multipath fading channels and parameters; S3, passing the radar signal simulated in step S1 through the channel simulated in step S2 to generate a radar modulated signal under multipath fading; S4, extracting the time-frequency domain features of the radar modulation signal obtained in step S3, thereby obtaining training set data; S5. Construct an AMSCNN network, wherein the AMSCNN network sequentially includes a plurality of cascaded multi-level feature extraction modules and a probability classification submodule; The multi-level feature extraction module includes a cascaded feature extraction basic module and an adaptive multipath effect suppression submodule; the feature extraction basic module includes a convolution module and a downsampling module connected in sequence; the convolution module includes two cascaded convolution blocks, each of which has the same structure and includes a convolution layer, a normalization layer and an activation function cascaded in sequence; the downsampling module includes a cascaded convolution layer and an activation function; the adaptive multipath effect suppression submodule includes a global average pooling layer, a first fully connected layer, a ReLU activation function, a second fully connected layer and a Sigmoid activation function cascaded in sequence; S6, using the training data set processed in step S4 to train the AMSCNN network constructed in step S5; S7. Input the time-frequency domain features of the radar signal to be identified into the trained AMSCNN network to obtain the identification result.

2. The intelligent radar signal recognition method under multipath fading conditions according to claim 1 is characterized in that: In step S1, multiple radar signal modulation types are simulated, including but not limited to: single carrier frequency, linear frequency modulation, binary frequency shift keying, binary phase shift keying, polyphase code, linear frequency modulation and BPSK composite modulation, and Costas and BPSK composite modulation.

3. The intelligent radar signal recognition method under multipath fading conditions according to claim 2 is characterized in that: The simulated channel in step S2 is a Rice multipath fading channel.

4. The intelligent radar signal recognition method under multipath fading conditions according to claim 3 is characterized in that: In step S4 and step S7, the time-frequency features are extracted using a synchronous compressed Fourier transform algorithm.

5. The intelligent radar signal recognition method under multipath fading conditions according to claim 4 is characterized in that: Step S6 uses the cross entropy loss function as the cost function during the training process.

6. The intelligent radar signal recognition method under multipath fading conditions according to claim 5 is characterized in that: Step S6 updates the network parameters based on the back propagation algorithm of gradient descent during the training process.

Images