A method for recovering intra-pulse relay interference signals based on multi-residual encoder-decoder networks

Abstract

This invention discloses a method for recovering intra-pulse relay interference signals based on a multi-residual codec network, relating to the field of radar technology. The method includes: receiving the interfered signal and preprocessing it to obtain input data; inputting the input data into a pre-trained multi-residual codec network, whereby the multi-residual codec network uses the encoder to extract features from the input data, compresses the feature dimensionality, increases the number of channels, and further uses a decoder connected to the encoder to perform signal recovery and dimensionality adjustment based on the input data to obtain the recovered signal. This method can reconstruct target echo signals in complex interference environments without relying on human experience, and exhibits good reconstruction performance under different types of relay interference.

Description

Technical Field

[0001] This invention belongs to the field of radar technology, specifically relating to a method for recovering intra-pulse relay interference signals based on a multi-residual codec network. Background Technology

[0002] In recent years, with the continuous advancement of radar technology, radar signal jamming technology has also developed rapidly. LFM (Linear Frequency Modulation) signals, due to their unique time-bandwidth product characteristics, high Doppler tolerance, and insensitivity to Doppler frequency shift, are widely used in radar and communication systems. However, LFM signals are often susceptible to jamming in actual combat, necessitating measures to conceal friendly radar information to mitigate or avoid the impact of enemy jamming.

[0003] For radar systems, improving anti-jamming capabilities not only enhances their survivability but also significantly strengthens their detection capabilities. Future radar systems will need to utilize more advanced signal processing technologies, such as high-speed parallel processing and deep learning, to more effectively combat various complex and novel types of interference. Simultaneously, to improve the confidentiality and anti-jamming capabilities of radar systems, encryption and decryption technologies for radar signals will become important research directions. The application of these cutting-edge technologies will open a new chapter in the development of radar technology and play a vital role in multiple fields, including military, aerospace, and meteorological observation.

[0004] Existing anti-interference methods, such as sidelobe cancellation, sidelobe blanking, frequency agility, waveform agility, and adaptive digital beamforming, face new challenges with the development of digital radio frequency memory and the emergence of interference sources with interception-store-forward capabilities.

[0005] To address these novel types of interference, researching new anti-interference strategies is crucial. Mainloole interference, pulse-based relay interference, and frequency-based interference are current research hotspots, with pulse-based relay interference attracting particular attention due to its powerful interference effect. Traditional anti-interference methods typically only distinguish between the target signal and the interfering signal in one dimension; once this dimension is blocked by the interfering signal, the anti-interference effect is severely compromised. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this invention provides a method for recovering intra-pulse forwarding interference signals based on a multi-residual codec network. The technical problem to be solved by this invention is achieved through the following technical solution:

[0007] This invention provides a method for recovering intra-pulse forwarding interference signals based on a multi-residual codec network, comprising:

[0008] Receive the interfered signal and preprocess the interfered signal to obtain input data;

[0009] The input data is fed into a pre-trained multi-residual codec network, which uses the encoder to extract features from the input data, compresses the dimension of the features, increases the number of channels, and then uses a decoder connected to the encoder to perform signal recovery and dimension adjustment based on the input data to obtain the recovered signal.

[0010] In one embodiment of the present invention, the step of receiving an interfered signal and preprocessing the interfered signal to obtain input data includes:

[0011] Receive the interfered signal S using a receiver R ;

[0012] The interfered signal is decomplexed, and its real and imaginary parts are split and concatenated to form a three-dimensional matrix.

[0013]

[0014] In the formula, Re(·) represents taking the real part, Im(·) represents taking the imaginary part, L represents the number of sampling points for a single pulse repetition period, and U represents the number of pulses;

[0015] The three-dimensional matrix is ​​normalized to obtain the input data:

[0016]

[0017] In the formula, i = 1, 2, i = 1 represents S In (i) is a three-dimensional matrix S In The real part of the interfered signal, i = 2 represents S In (i) is a three-dimensional matrix S In The imaginary part of the interfered signal, 1 L×U This represents an L×U matrix of all 1s.

[0018] In one embodiment of the present invention, the encoder includes a feature discovery module and a feature extraction module, wherein the feature discovery module includes a cascaded first convolutional layer, a second convolutional layer and a third convolutional layer, and the feature extraction module includes a cascaded first residual block, a second residual block, a third residual block and a fourth residual block;

[0019] The steps of inputting the input data into a pre-trained multi-residual codec network, enabling the multi-residual codec network to extract features from the input data using the encoder, and compressing the dimensionality of the features and increasing the number of channels, include:

[0020] The input data is fed into a pre-trained multi-residual codec network, so that the multi-residual codec network uses the feature discovery module to calculate the receptive field based on the input data to obtain a feature map, and uses the feature extraction module to extract features based on the feature map, and compresses the dimension of the feature and increases the number of channels.

[0021] In one embodiment of the present invention, the stride of the first convolutional layer is 2×2 and the number of channels is 48, and the stride of the second convolutional layer and the third convolutional layer is 1×1 and the number of channels is 64.

[0022] In one embodiment of the present invention, the first residual block, the second residual block, the third residual block, and the fourth residual block each include a cascaded first sub-convolutional layer and a second sub-convolutional layer, wherein the kernel size of the first sub-convolutional layer and the second sub-convolutional layer is 3×3 and the stride is 1×1. The second residual block, the third residual block, and the fourth residual block also include a third sub-convolutional layer, wherein the input end of the third sub-convolutional layer is connected to the input end of the first sub-convolutional layer and the output end is connected to the output end of the second sub-convolutional layer, and the kernel size of the third sub-convolutional layer is 1×1 and the stride is 2×2.

[0023] In one embodiment of the present invention, the decoder includes a signal reconstruction module and a dimension correction module, wherein the signal reconstruction module includes a cascaded first transposed residual block, a second transposed residual block, a third transposed residual block and a fourth transposed residual block, and the dimension correction module includes a cascaded first transposed convolutional layer, a second transposed convolutional layer and a third transposed convolutional layer.

[0024] The output of the feature discovery module is connected to the input of the fourth transposed residual block through a 64-dimensional channel. The output of the first residual block is connected to the input of the third transposed residual block through a 64-dimensional channel. The output of the second residual block is connected to the input of the second transposed residual block through a 128-dimensional channel. The output of the third residual block is connected to the first transposed residual block through a 256-dimensional channel.

[0025] In one embodiment of the present invention, the first transposed residual block, the second transposed residual block, the third transposed residual block, and the fourth transposed residual block each include a cascaded first sub-transposed convolutional layer and a second sub-transposed convolutional layer, wherein the kernel size of the first sub-transposed convolutional layer and the second sub-transposed convolutional layer is 3×3 and the stride is 1×1. The first transposed residual block, the second transposed residual block, and the third transposed residual block also include a third sub-transposed convolutional layer, wherein the input end of the third sub-transposed convolutional layer is connected to the input end of the first sub-transposed convolutional layer and the output end is connected to the output end of the second sub-transposed convolutional layer, and the kernel size of the sub-transposed convolutional layer is 1×1 and the stride is 2×2.

[0026] In one embodiment of the present invention, the step of enabling the multi-residual decoder network to perform signal recovery and dimension adjustment based on the input data using a decoder connected to the encoder jumper to obtain a recovered signal includes:

[0027] The features extracted by the feature extraction module, after compression of dimensions and increase of channel number, are input into the decoder. This allows the multi-residual decoder network to use the signal reconstruction module to perform dimension recovery based on the input features and the data input via jumpers. After rectification and integration by the dimension correction module, the recovered signal corresponding to the echo signal is obtained.

[0028] In one embodiment of the present invention, the signal reconstruction module performs dimensionality recovery based on the input features and the data input via a jumper connection, following these steps:

[0029] The first transposed residual block obtains the features after the feature extraction module has compressed the dimensions and increased the number of channels, performs a transposed convolution operation on it, and adds it to the data input by the third residual block to obtain the first reconstructed signal;

[0030] The second transposed residual block performs a transposed convolution operation on the first reconstructed signal and adds it to the data input to the second residual block to obtain the second reconstructed signal;

[0031] The third transposed residual block performs a transposed convolution operation on the second reconstructed signal and adds it to the data input by the first residual block to obtain the third reconstructed signal.

[0032] The fourth transposed residual block performs a transposed convolution operation on the third reconstructed signal and adds it to the data input by the feature discovery module to obtain the fourth reconstructed signal.

[0033] In one embodiment of the present invention, the loss function used by the multi-residual codec network during training is the root mean square error (RMSE) or the Pearson coefficient.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0035] This invention provides a method for recovering transmitted signals subject to intra-pulse relay interference based on a multi-residual codec network. First, the transmitted signal subject to relay interference is decomplexed and normalized. Then, the pre-processed input data is input into a pre-trained multi-residual codec network for feature extraction and reconstruction of the signal. This method can reconstruct the target echo signal in complex interference environments without relying on human experience, and it exhibits good reconstruction performance under different types of relay interference.

[0036] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0037] Figure 1 This is a flowchart of a method for recovering intra-pulse forwarding interference signals based on a multi-residual codec network provided in an embodiment of the present invention;

[0038] Figure 2a This is a time-frequency relationship diagram of the transmitted linear frequency modulated signal provided in an embodiment of the present invention;

[0039] Figure 2b This is a time-frequency relationship diagram of the SMSP interference signal provided in an embodiment of the present invention;

[0040] Figure 2c This is a time-frequency relationship diagram of the ISRJ interference signal provided in an embodiment of the present invention;

[0041] Figure 2d This is a time-frequency relationship diagram of the INUSRJ interference signal provided in an embodiment of the present invention;

[0042] Figure 3 This is a schematic diagram of the inverse complexization process provided in an embodiment of the present invention;

[0043] Figure 4 This is a schematic diagram of the structure of the multi-residual codec network provided in an embodiment of the present invention;

[0044] Figures 5a-5c This is a diagram showing the recovery result of the multi-residual codec network provided in this embodiment of the invention for the transmitted signal affected by ISRJ interference during the training process;

[0045] Figure 6 This is a graph showing the relationship between the number of iterations and RMSE provided in an embodiment of the present invention. Detailed Implementation

[0046] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0047] Figure 1 This is a flowchart of a method for recovering intra-pulse forwarding interference signals based on a multi-residual codec network, provided in an embodiment of the present invention. Figure 1 As shown, this embodiment of the invention provides a method for recovering intra-pulse forwarding interference signals based on a multi-residual codec network, comprising:

[0048] S1. Receive the interfered signal and preprocess the interfered signal to obtain the input data;

[0049] S2. Input the input data into a pre-trained multi-residual codec network, so that the multi-residual codec network can use the encoder to extract the features of the input data, compress the dimension of the features and increase the number of channels, and then use the decoder connected to the encoder to perform signal recovery and dimension adjustment based on the input data to obtain the recovered signal.

[0050] Specifically, in this embodiment, the transmitted signal is a linear frequency modulated signal, such as... Figure 2a As shown, its form is:

[0051]

[0052] In the formula, B represents bandwidth, and T e F represents the pulse width. s The sampling frequency is represented by L = T, and the number of signal sampling points is L = T. e ×F s Then time t = [1 / F s ,2 / F s ,...,L / F s ] T , operator This represents the product of corresponding terms in a vector.

[0053] Optionally, the basic parameters are set as follows: carrier frequency f0 = 4GHz, sampling rate F s =5MHz, pulse repetition period T r =1ms, the pulse width T of the linear frequency modulated signal e =160μs, number of transmitted pulses U=64, signal-to-noise ratio SNR=30dB, signal-to-interference ratio JSR=10dB, target velocity v=30m / s, frequency modulation μ=B / T e The bandwidth B is constrained by the distance resolution and sampling rate, and is set to |μ|∈[3×10]. 9 3×10 10 If the sampling rate is 100 Hz / s, then the smaller the μ value, the worse the ranging resolution; however, if the μ value is too large, it may result in insufficient sampling rate. It should be noted that the magnitude of μ is constrained in this embodiment, which means that μ can be positive or negative.

[0054] During the training of a multi-residual encoder-decoder network, |μ| of the training data in the training set is in μ min to μ max The μ values ​​are randomly distributed within a certain range to ensure the comprehensiveness of the dataset. Similarly, the |μ| values ​​of the 1000 test data points in the test set are uniformly selected from the same region. Furthermore, to ensure no overlap between the training and test datasets, any μ values ​​selected for the test data are excluded from the training data.

[0055] Smeared Spectrum (SMSP) jamming is a dense multi-false-target jamming specifically designed for linear frequency modulated pulse compression radar. For example, taking SMSP jamming as an example... Figure 2b As shown, its form is:

[0056]

[0057] Where K represents the number of forwards, A SMSP Indicates the amplitude, t = [1 / F] s ,2 / F s ,...,T e / F s ] represents the time vector of each sampling point, operator Let Kronecker product be represented. If the transmitted signal, interference signal, and noise after time delay are represented as follows: When this happens, the interference signal received by the radar receiver can be expressed as:

[0058]

[0059] in, It is the phase difference between pulses, and the u-th Doppler unit f u =F r ((u-1) / (U-1)-0.5),u=1,…,U,F r This indicates the pulse repetition frequency.

[0060] Of course, in other embodiments of this application, such as Figure 2c , 2d As shown, the interference patterns can also be ISRJ (Interrupted-sampling Repeater Jamming) and INUSRJ (Interrupted Non-uniform Sampling Repeater Jamming).

[0061] For ISRJ, a segment of the transmitted signal needs to be extracted M times, and each extraction needs to be retransmitted K times. The extracted pulse train of the transmitted signal It is expressed as follows:

[0062]

[0063] Intercepted signals If the intercept time is T intercept Then N = T intercept F s .

[0064] Interference signals can be represented as:

[0065]

[0066] Among them, A ISRJ Indicates the magnitude.

[0067] Since INURJ is a generalization of ISRJ, its generation pattern is consistent with ISRJ, including initial interception of the transmitted signal, followed by storage and forwarding. However, in INURJ, the number of interceptions... Number of samples for the i-th sample and the number of forwards for the i-th forward. These are random integers, and these random integers are subject to the following specific constraints:

[0068]

[0069] In the above constraints, the first two inequalities are used to limit the lower bound of these random integers. The first inequality guarantees that each interception will forward at least once, and the second inequality guarantees the minimum sampling time T. min Since the pulse length L is constant, the maximum and minimum values ​​of M can respectively limit n. i and k i The lower and upper limits of M. Furthermore, M... max The choice of should take into account the value of L. The last equation of the constraint conditions establishes n i k i and l i The quantitative relationship between them. In integer n i and k i Cannot satisfy (k) i +1)n i =l i In special cases, the k-th forward is truncated to ensure... Therefore, the i-th INUSRJ signal It can be represented as:

[0070]

[0071] here Indicates the i-th intercepted transmission signal:

[0072]

[0073] When i=1, l0=0.

[0074] Finally, the INUSRJ interference signal will pass through The following can be obtained by cascading:

[0075]

[0076] Here, ISRJ interference is selected, the number of forwardings is set to 4, and the sampling time is 8μs.

[0077] Figure 3 This is a schematic diagram of the inverse complex number conversion process provided in an embodiment of the present invention. Please refer to [link / reference]. Figure 3 Step S1, which involves receiving the interfered signal and preprocessing it to obtain the input data, includes:

[0078] S101, Receive the interfered signal S using a receiver. R ;

[0079] S102. Perform decomplexification on the interfered signal, splitting and concatenating the real and imaginary parts of the interfered signal to form a three-dimensional matrix:

[0080]

[0081] In the formula, Re(·) represents taking the real part, Im(·) represents taking the imaginary part, L represents the number of sampling points for a single pulse repetition period, and U represents the number of pulses;

[0082] S103. Normalize the three-dimensional matrix to obtain the input data:

[0083]

[0084] In the formula, i = 1, 2, i = 1 represents S In (i) is a three-dimensional matrix S In The real part of the interfered signal, i = 2 represents S In (i) is a three-dimensional matrix S In The imaginary part of the interfered signal, 1 L×U This represents an L×U matrix of all 1s.

[0085] Figure 4 This is a schematic diagram of the structure of the multi-residual codec network provided in an embodiment of the present invention. Figure 4 As shown, in this embodiment, the encoder (layers 0-4) includes a feature discovery module (layer 0) and a feature extraction module (layers 1-4). The feature discovery module includes a cascaded first convolutional layer, a second convolutional layer, and a third convolutional layer, and the feature extraction module includes a cascaded first residual block (layer 1), a second residual block (layer 2), a third residual block (layer 3), and a fourth residual block (layer 4).

[0086] Step S2 involves feeding the input data into a pre-trained multi-residual codec network, enabling the network to extract features from the input data using the encoder, and compressing the feature dimensionality and increasing the number of channels. This includes:

[0087] The input data is fed into a pre-trained multi-residual codec network, which uses the feature discovery module to calculate the receptive field based on the input data to obtain a feature map. The feature extraction module then extracts features based on the feature map, compresses the dimension of the feature, and increases the number of channels.

[0088] Specifically, the feature discovery module is used to calculate the receptive field using the input data to obtain the feature map. The first convolutional layer has a stride of 2×2 and 48 channels, which reduces the data size compared to traditional convolutional neural networks. In addition, the stride of the second and third convolutional layers is 1×1, which maintains the data size, while the number of channels is increased to 64 through kernel configuration.

[0089] As can be seen, the feature discovery module reduces network parameters and expands the receptive field by cascading multiple 3×3 convolutional layers. In addition, due to its convolutional input ports, it can accommodate input data of various sizes, which is particularly important in radar signal processing, because radar parameters affect data length and the number of pulses is directly related to the coherence time of the signal.

[0090] Furthermore, the feature extraction module is used to extract features from the feature map using the residual structure, compress the dimensionality of the obtained features, and increase the number of channels. Optionally, in the feature extraction module, the first residual block, the second residual block, the third residual block, and the fourth residual block all include cascaded first and second sub-convolutional layers, wherein the kernel size of the first and second sub-convolutional layers is 3×3 and the stride is 1×1. The second, third, and fourth residual blocks also include a third sub-convolutional layer, the input of which is connected to the input of the first sub-convolutional layer and the output of which is connected to the output of the second sub-convolutional layer, and the kernel size of the third sub-convolutional layer is 1×1 and the stride is 2×2.

[0091] It should be understood that the third sub-convolutional layer in the second, third, and fourth residual blocks is used for jumps between layers. Compared with the previous residual block, the size of the output data in each residual block is halved, while the number of channels is doubled.

[0092] Furthermore, in the aforementioned multi-residual codec, the decoder includes a signal reconstruction module and a dimension correction module. The signal reconstruction module includes a cascaded first transposed residual block, a second transposed residual block, a third transposed residual block, and a fourth transposed residual block. The dimension correction module includes a cascaded first transposed convolutional layer, a second transposed convolutional layer, and a third transposed convolutional layer.

[0093] The output of the feature discovery module is connected to the input of the fourth transposed residual block through a 64-dimensional channel. The output of the first residual block is connected to the input of the third transposed residual block through a 64-dimensional channel. The output of the second residual block is connected to the input of the second transposed residual block through a 128-dimensional channel. The output of the third residual block is connected to the first transposed residual block through a 256-dimensional channel.

[0094] It should be understood that this embodiment, by establishing direct connections between different layers of the network, allows information to skip some layers and be directly passed to later layers, which helps to speed up training and prevent [failure to prevent data loss]. Between the encoder and decoder, block-level shortcuts facilitate the transfer of data from earlier modules to later modules after convolution (or transposed convolution) and batch normalization operations. Therefore, block-level shortcuts do not use convolution (or transposed convolution). Instead, data from earlier blocks is batch normalized, activated, and then transferred to later blocks.

[0095] Optionally, in step S2, the step of enabling the multi-residual decoder network to perform signal recovery and dimension adjustment based on the input data using a decoder connected to the encoder to obtain the recovered signal includes:

[0096] The feature extraction module compresses the dimension and increases the number of channels before inputting the features into the decoder. This allows the multi-residual decoder network to use the signal reconstruction module to recover the dimension based on the input features and the data connected via jumpers. After rectification and integration by the dimension correction module, the recovered signal corresponding to the echo signal is obtained.

[0097] Signal reconstruction is essentially an inverse feature extraction process. The signal reconstruction module is used to upsample the input data through transposed convolution operations, and restore the data dimension based on the transposed residual block according to the features to obtain the module output with the same dimension as the feature map output of layer 0.

[0098] Specifically, the first transposed residual block, the second transposed residual block, the third transposed residual block, and the fourth transposed residual block all include a cascaded first sub-transposed convolutional layer and a second sub-transposed convolutional layer. The kernel size of the first sub-transposed convolutional layer and the second sub-transposed convolutional layer is 3×3, and the stride is 1×1. The first transposed residual block, the second transposed residual block, and the third transposed residual block also include a third sub-transposed convolutional layer. The input of the third sub-transposed convolutional layer is connected to the input of the first sub-transposed convolutional layer, and the output is connected to the output of the second sub-transposed convolutional layer. The kernel size of the sub-transposed convolutional layer is 1×1, and the stride is 2×2.

[0099] As can be seen, layers 5-8 of the multi-residual codec network share the same kernel size, stride, and output size as layers 4-1, which is an inverse feature extraction process. The data size doubles every 2×2 steps, while the number of channels is halved. Furthermore, within the signal reconstruction module, the hierarchical jumper method connects smaller data to larger data, thus upsampling is necessary.

[0100] Optionally, in step S2 above, the signal reconstruction module performs dimensionality recovery based on the input features and the data input via jumpers, following these steps:

[0101] The first transposed residual block obtains the features after the feature extraction module compresses the dimensions and increases the number of channels. After transposing and convolving the features, it is added to the input data of the third residual block to obtain the first reconstructed signal.

[0102] The second transposed residual block performs a transposed convolution operation on the first reconstructed signal and adds it to the data input to the second residual block to obtain the second reconstructed signal.

[0103] The third transposed residual block performs a transposed convolution operation on the second reconstructed signal and adds it to the data input from the first residual block to obtain the third reconstructed signal.

[0104] The fourth transposed residual block performs a transposed convolution operation on the third reconstructed signal and adds it to the data input by the feature discovery module to obtain the fourth reconstructed signal.

[0105] The dimension correction module adjusts the data dimension to ensure the reconstructed signal output by the network matches the input data. Layer 9 plays a crucial role in this adjustment. Similar to Layer 0, Layer 9 includes a first transposed convolutional layer, a second transposed convolutional layer, and a third transposed convolutional layer, with 3×3 kernels and strides of 1×1, 1×1, and 2×2, respectively. Furthermore, the second transposed convolutional layer has 48 transposed kernels, producing an output of 48×400×32. Due to the 2×2 stride, the third transposed convolutional layer increases the output size to 800×64. Finally, by rectifying the signal dimension and integrating the corrected reconstructed data, the dimension of the output reconstructed signal is made identical to the input data of Layer 0.

[0106] It should be noted that during the training process of a multi-residual codec network, the expected signal corresponding to the input sample data... and its normalized signal Transmitted signal without interference Build:

[0107] Where i = 1, 2;

[0108] The loss function used is the root mean square error or the Pearson coefficient. For example, the root mean square error is expressed as:

[0109]

[0110] In the formula, S' Out S' represents the output signal of the multi-residual codec network during one iteration. Label This is the normalized expected signal.

[0111] Furthermore, the Pearson correlation coefficient between the recovered signal and the transmitted signal... Represented as:

[0112]

[0113] Where, μ LFM =E[vec(S) LFM )],μ Rec =E[vec(S) Rec )],1 LU×1 S represents a matrix of dimension LU×1 consisting entirely of 1s. Rec Indicates a recovery signal:

[0114] S Rec =S Out (1)+jS Out (2);

[0115] Among them, S Out Indicates S' Out The signal after denormalization:

[0116] S Out =(max(S) In (i))-min(S In (i)))S' Out +min(S In (i))×1 L×U i = 1, 2;

[0117] It should be noted that the solution normalization process yields S. Out Subsequently, further investigation into S Out By performing decomplexization, the recovered signal S can be obtained. Rec .

[0118] Although RMSE is often used to evaluate recovery performance under normal conditions, it is considered too stringent for radar signal processing, especially in point-to-point recovery, where the primary goal of a multi-residual codec network is to eliminate interference signals in the input, and subsequent operations include matched filters to improve the signal-to-noise ratio.

[0119] Figures 5a-5cThis is a diagram showing the recovery result of the transmitted signal subjected to ISRJ interference during the training process of the multi-residual codec network provided in this embodiment of the invention. Figures 5a-5c The different modulation frequencies shown indicate that, with increasing iterations, low-frequency signals are recovered preferentially over high-frequency signals. However, since the sampling and forwarding times are constant, the regular distribution of interference segments is maintained. Furthermore, the results of the first generation effectively mitigated some interference segments, demonstrating that the multi-residual codec network provided by this invention accurately estimates the starting position and sample forwarding time of the incoming interference segments. Although the transmitted signal was not fully recovered in the first generation, the location of the interfering data segments was successfully located. Figures 5a-5c As shown, locating large μ samples requires additional iterations, resulting in a slower recovery progress compared to small μ samples. Larger μ values ​​correspond to wider bandwidths, and as the frequency increases, the temporal changes of the data occur more rapidly, making the location of interfering data segments more challenging.

[0120] Figure 6 This is a graph showing the relationship between the number of iterations and RMSE provided in an embodiment of the present invention. After appropriate training, the transmitted signal was recovered in the 5th and 10th generations, as shown... Figure 6 As shown, their epochs involve 120,000 iterations, resulting in an RMSE exceeding 0.1. Therefore, discrepancies still exist between these results and their corresponding labels. Furthermore, Figure 5b and Figure 5c The results show significant trailing distortion at generations 5 and 10, which aligns with the final forwarding position. Further analysis of the numerical relationship between pulse width and ISRJ parameters indicates that the forwarding is incomplete, violating the integer multiple rule and disrupting the regularity of ISRJ. Therefore, further training is needed to obtain this additional information.

[0121] exist Figure 6 In the middle, the result of the 400th generation recovery is compared with the expected signal S. Label Precise matching eliminates trailing distortion, and high- and low-frequency data are accurately recovered with an RMSE error of less than 0.02. Therefore, the aforementioned multi-residual codec network exhibits excellent recovery performance against ISRJ interference.

[0122] As can be seen from the above embodiments, the beneficial effects of the present invention are as follows:

[0123] This invention provides a method for recovering transmitted signals subject to intra-pulse relay interference based on a multi-residual codec network. First, the transmitted signal subject to relay interference is decomplexed and normalized. Then, the pre-processed input data is input into a pre-trained multi-residual codec network for feature extraction and reconstruction of the signal. This method can reconstruct the target echo signal in complex interference environments without relying on human experience, and it exhibits good reconstruction performance under different types of relay interference.

[0124] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0125] The use of terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples" indicates that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0126] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for recovering intra-pulse relay interference signals based on a multi-residual codec network, characterized in that, include: Receive the interfered signal and preprocess the interfered signal to obtain input data; The input data is fed into a pre-trained multi-residual codec network, which uses the encoder to extract features from the input data, compresses the dimension of the features and increases the number of channels, and then uses a decoder connected to the encoder to perform signal recovery and dimension adjustment based on the input data to obtain the recovered signal. The steps of receiving the interfered signal and preprocessing the interfered signal to obtain input data include: Receive interference signals using a receiver ; The interfered signal is decomplexed, and its real and imaginary parts are split and concatenated to form a three-dimensional matrix. ； In the formula, Indicates taking the real part, This indicates taking the imaginary part. This indicates the number of sampling points for a single pulse repetition cycle. Indicates the number of pulses; The three-dimensional matrix is ​​normalized to obtain the input data: ； In the formula, , express A three-dimensional matrix The real part of the interfered signal, express A three-dimensional matrix The imaginary part of the signal that is being interfered with. The dimension is A matrix of all ones; The encoder includes a feature discovery module and a feature extraction module. The feature discovery module includes a cascaded first convolutional layer, a second convolutional layer, and a third convolutional layer. The feature extraction module includes a cascaded first residual block, a second residual block, a third residual block, and a fourth residual block. The steps of inputting the input data into a pre-trained multi-residual codec network, enabling the multi-residual codec network to extract features from the input data using the encoder, and compressing the dimensionality of the features and increasing the number of channels, include: The input data is fed into a pre-trained multi-residual codec network, so that the multi-residual codec network uses the feature discovery module to calculate the receptive field based on the input data to obtain a feature map, and uses the feature extraction module to extract features based on the feature map, and compresses the dimension of the feature and increases the number of channels; The decoder includes a signal reconstruction module and a dimension correction module. The signal reconstruction module includes a cascaded first transposed residual block, a second transposed residual block, a third transposed residual block, and a fourth transposed residual block. The dimension correction module includes a cascaded first transposed convolutional layer, a second transposed convolutional layer, and a third transposed convolutional layer. The output of the feature discovery module is connected to the input of the fourth transposed residual block via a 64-dimensional channel. The output of the first residual block is connected to the input of the third transposed residual block via a 64-dimensional channel. The output of the second residual block is connected to the input of the second transposed residual block via a 128-dimensional channel. The output of the third residual block is connected to the first transposed residual block via a 256-dimensional channel.

2. The method for recovering intra-pulse forwarding interference signals based on a multi-residual codec network according to claim 1, characterized in that, The first convolutional layer has a stride of 2×2 and 48 channels, while the second and third convolutional layers have a stride of 1×1 and 64 channels.

3. The method for recovering intra-pulse forwarding interference signals based on a multi-residual codec network according to claim 2, characterized in that, The first residual block, the second residual block, the third residual block, and the fourth residual block each include a cascaded first sub-convolutional layer and a second sub-convolutional layer. The kernel size of the first sub-convolutional layer and the second sub-convolutional layer is 3×3, and the stride is 1×1. The second residual block, the third residual block, and the fourth residual block also include a third sub-convolutional layer. The input of the third sub-convolutional layer is connected to the input of the first sub-convolutional layer, and the output is connected to the output of the second sub-convolutional layer. The kernel size of the third sub-convolutional layer is 1×1, and the stride is 2×2.

4. The method for recovering intra-pulse forwarding interference signals based on a multi-residual codec network according to claim 1, characterized in that, The first, second, third, and fourth transposed residual blocks each include a cascaded first sub-transposed convolutional layer and a second sub-transposed convolutional layer. The kernel size of the first and second sub-transposed convolutional layers is 3×3, and the stride is 1×1. The first, second, and third transposed residual blocks also include a third sub-transposed convolutional layer. The input of the third sub-transposed convolutional layer is connected to the input of the first sub-transposed convolutional layer, and the output is connected to the output of the second sub-transposed convolutional layer. The kernel size of the sub-transposed convolutional layer is 1×1, and the stride is 2×2.

5. The method for recovering intra-pulse forwarding interference signals based on a multi-residual codec network according to claim 4, characterized in that, The steps of enabling the multi-residual codec network to perform signal recovery and dimension adjustment based on the input data using a decoder connected to the encoder to obtain the recovered signal include: The features extracted by the feature extraction module, after compression of dimensions and increase of channel number, are input into the decoder. This allows the multi-residual codec network to use the signal reconstruction module to perform dimension recovery based on the input features and the data input via jumpers. After rectification and integration by the dimension correction module, the recovered signal corresponding to the echo signal is obtained.

6. The method for recovering intra-pulse forwarding interference signals based on a multi-residual codec network according to claim 5, characterized in that, The signal reconstruction module performs dimensionality recovery based on the input features and the data input via jumpers, following these steps: The first transposed residual block obtains the features after the feature extraction module has compressed the dimensions and increased the number of channels, performs a transposed convolution operation on it, and adds it to the data input by the third residual block to obtain the first reconstructed signal; The second transposed residual block performs a transposed convolution operation on the first reconstructed signal and adds it to the data input to the second residual block to obtain the second reconstructed signal; The third transposed residual block performs a transposed convolution operation on the second reconstructed signal and adds it to the data input by the first residual block to obtain the third reconstructed signal. The fourth transposed residual block performs a transposed convolution operation on the third reconstructed signal and adds it to the data input by the feature discovery module to obtain the fourth reconstructed signal.

7. The method for recovering intra-pulse forwarding interference signals based on a multi-residual codec network according to claim 1, characterized in that, The loss function used by the multi-residual encoder-decoder network during training is the root mean square error (RMSE) or the Pearson coefficient.

Images