Radar target and interference signal separation method based on lightweight neural network

Abstract

According to the radar target and interference signal separation method based on the lightweight neural network provided by the invention, the problems that most of methods for directly separating interference in the prior art cannot extract reliable pure target signals and perform post-processing on the reliable pure target signals to directly obtain target information, the feasibility is relatively low, and the generalization is relatively poor are solved; the method comprises the following steps: preprocessing a time domain echo signal to obtain a preprocessed signal; inputting the pre-processed signal into a trained mixed attention mechanism network to obtain a separation target echo signal; wherein the mixed attention mechanism network comprises an encoder module, a separation module, an inter-stage connection module and a decoder module; obtaining specific information of the estimated target according to the separated target echo signal; according to the method, the time domain signal is directly processed, and the method has relatively good separation precision and calculation efficiency.

Description

Technical Field

[0001] The present invention relates to the field of radar signal processing technology, and in particular to a radar target and interference signal separation method based on a lightweight neural network. Background Art

[0002] With the increasing diversification of modern air operations, the gradual application of drones, unmanned ships and unmanned boats, and the development of various electronic detection methods, radar has a significant advantage in electronic reconnaissance, target observation, counter-unmanned vehicle applications, etc., with its long detection distance, unaffected by fog, clouds and rain, and a certain penetration ability while having the significant advantage of working all day. Due to the emergence and development of digital radio frequency memory (DRFM), jammers can quickly intercept radar transmission signals and generate flexible and changeable interference signals based on the extracted radar technical parameters such as carrier frequency, repetition frequency, and wave direction, making the detection environment of modern radar increasingly complex. Forwarding jammers can enable active deception jamming systems to copy the collected radar transmission signals at a sub-second processing speed to generate false targets with high noise-to-signal ratio in radar echoes to deceive radars. When faced with deceptive jamming such as Intermittent Sampling Repeater Jamming (ISRJ), traditional time-frequency analysis methods have obvious limitations and difficulties in suppressing jamming and extracting radar echo signals. The resolution of short-time Fourier transform (STFT) and wavelet transform in the time-frequency domain usually has a certain trade-off, that is, it is impossible to achieve the best between frequency resolution and time resolution at the same time. This resolution limitation may not be able to effectively distinguish jamming and target echo signals. In addition, most of the existing radar anti-jamming methods are mainly based on designing a transmission waveform with a smaller jamming duty cycle in the echo, suppressing the jamming in the echo signal, etc., while most of the methods for directly separating the jamming cannot extract reliable pure target signals and directly obtain target information by post-processing them, and have low feasibility and poor generalization. In the face of complex electromagnetic interference environments with many jamming styles and combinations, there is no effective solution to the problem for the related anti-jamming methods for separating the jamming and extracting pure target signals. Summary of the invention

[0003] The present invention provides a radar target and interference signal separation method based on a lightweight neural network, which solves the problems that most of the methods of directly separating interference in the prior art cannot extract reliable pure target signals and post-process them to directly obtain target information, and have low feasibility and poor generalization. It realizes direct processing of time domain signals and has relatively good separation accuracy and computational efficiency.

[0004] The present invention provides a radar target and interference signal separation method based on a lightweight neural network, the method comprising:

[0005] Preprocessing the time domain echo signal to obtain a preprocessed signal and a full feature corresponding to the preprocessed signal;

[0006] The preprocessed signal is input into the trained hybrid attention mechanism network to obtain a separated target echo signal; wherein the hybrid attention mechanism network includes an encoder module, a separation module, an inter-stage connection module and a decoder module; the encoder module is used to encode the full features and extract the multi-scale features of the preprocessed signal; the separation module is used to fuse the multi-scale features to obtain fused features, and use the global attention module to calculate the global attention weight to obtain a global self-attention map; the inter-stage connection module is used to restore the accuracy of the multi-scale features to obtain restored multi-scale features; the decoder module is used to modulate the restored multi-scale feature pairs using the global self-attention map to obtain a separated target echo signal;

[0007] The specific information of the target is estimated based on the separated target echo signal.

[0008] In a possible implementation, preprocessing the time domain echo signal to obtain a preprocessed signal includes:

[0009] Loading the time domain echo signal and initializing variables;

[0010] Read the data file in the specified path, and load the intermittent sampling forwarding interference data and target echo data in plural form; wherein the target echo data includes noise data;

[0011] Performing data shape transformation on the intermittent sampling forwarding interference data and the target echo data respectively to obtain reshaped interference data and reshaped target echo data;

[0012] Respectively extracting the real part and the imaginary part of the reshaped interference data and the real part and the imaginary part of the reshaped target echo data;

[0013] The real part and the imaginary part of the reshaped interference data are alternately sorted by a cyclic interpolation sorting method to obtain a continuous interference data long vector, and the real part and the imaginary part of the reshaped target echo data are alternately sorted to obtain a continuous target echo data long vector;

[0014] The continuous interference data long vector is added to the continuous target echo data long vector to obtain a preprocessing signal.

[0015] In one possible implementation, the training of the hybrid attention mechanism network includes: dynamically adjusting the learning rate of the hybrid attention mechanism network according to an evaluation separation index.

[0016] In a possible implementation, it is characterized in that the calculation formula of the evaluation separation index is expressed as:

[0017]

[0018] in, represents the estimated target echo signal; s represents the pure source signal; S target represents the target signal part extracted from the estimated signal; e noise represents the noise part, that is, the difference between the estimated signal and the target signal; s target Represents the signal-to-noise ratio, which is used to evaluate the quality of the target signal.

[0019] In one possible implementation, the encoder module includes a first depth-wise separable convolutional layer, a second depth-wise separable convolutional layer, a third depth-wise separable convolutional layer, a first average pooling layer, a second average pooling layer, and a third average pooling layer connected in sequence;

[0020] The first depth-wise separable convolutional layer is used to extract features from the full features to obtain 1 / 2 scale features;

[0021] The second depth-wise separable convolutional layer is used to extract features from the 1 / 2 scale features to obtain 1 / 4 scale features;

[0022] The third depth-wise separable convolutional layer is used to extract features from the 1 / 4 scale features to obtain 1 / 8 scale features;

[0023] The first average pooling layer is used to perform an average pooling operation on the 1 / 4 scale feature to obtain a first 1 / 8 scale feature;

[0024] The second average pooling layer is used to perform an average pooling operation on the 1 / 4 scale feature to obtain a second 1 / 8 scale feature;

[0025] The third average pooling layer is used to perform average pooling on the full features to obtain a third 1 / 8 scale feature.

[0026] In a possible implementation, the inter-stage connection module includes: a first upsampling module, a second upsampling module and a third upsampling module that are connected in sequence;

[0027] The first upsampling module is used to perform a double upsampling on the dot product result of the fusion feature and the 1 / 8 scale feature according to the global self-attention map, and perform a dot product of the sampling result and the 1 / 4 scale feature to obtain a first restored multi-scale feature;

[0028] The second upsampling module is used to perform four-fold upsampling on the first restored multi-scale feature according to the global self-attention map, and perform dot multiplication of the sampling result and the 1 / 4 scale feature to obtain the second restored multi-scale feature;

[0029] The third upsampling module is used to upsample the second restored multi-scale feature eight times according to the global self-attention map, and perform dot multiplication of the sampling result with the full feature of the preprocessed signal to obtain the third restored multi-scale feature; wherein the 1 / 4 scale feature, the 1 / 4 scale feature and the 1 / 8 scale feature are the multi-scale features.

[0030] In one possible implementation, the decoder module includes a first cascaded local attention layer, a second cascaded local attention layer, and a third cascaded local attention layer connected in sequence;

[0031] The first cascaded local attention layer is used to extract local attention features according to the first restored multi-scale features and the second restored multi-scale features to obtain first local attention features;

[0032] The second cascaded local attention layer is used to extract local attention features according to the first local attention features and the third restored multi-scale features to obtain second local attention features;

[0033] The third cascaded local attention layer is used to extract local attention features based on the second local attention features and the full features of the preprocessed signal to obtain a third local attention feature.

[0034] In a possible implementation, the decoder module is used to modulate the restored multi-scale feature pair using the global self-attention map to obtain a separated target echo signal, including:

[0035] The third local attention feature is mapped to the same spatial dimension as the time domain echo signal to obtain a separated target echo signal.

[0036] In a possible implementation manner, obtaining the estimated specific information of the target according to the separated target echo signal includes:

[0037] Determine the signal type of the time domain echo signal, and perform pulse compression processing on the separated target echo signal according to the corresponding pulse compression method to obtain a pulse compression signal;

[0038] The pulse pressure signal is subjected to coherent processing to obtain a coherent processing signal; wherein the coherent processing signal includes specific information of the target.

[0039] One or more technical solutions provided in the present invention have at least the following technical effects or advantages:

[0040] The present invention processes the time domain signal directly without the traditional time-frequency domain transformation. Such processing can eliminate the shortcomings such as the phase and amplitude of the signal are not coupled, the time-frequency representation of multiple source signal separation is not ideal, and the delay time of the spectrum calculation is long. The fully convolutional network has better separation accuracy and calculation efficiency for the sequence model.

[0041] For the separated signals, the present invention can, on the one hand, estimate and analyze the separated interference signals, understand the type of interference and its specific parameters, so as to construct a waveform with a smaller interference signal duty cycle in the subsequent radar transmission waveform design; on the other hand, the separated relatively clean target echo signals can be processed including pulse pressure, coherence, etc., to estimate the specific position, speed and other information of the target. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A flowchart of the steps of a radar target and interference signal separation method based on a lightweight neural network provided in an embodiment of the present invention;

[0043] Figure 2 A schematic diagram of a hybrid attention mechanism network provided for an embodiment of the present invention;

[0044] Figure 3 A structural diagram of a hybrid attention mechanism network provided for an embodiment of the present invention;

[0045] Figure 4a A time-frequency diagram of a separated target signal with the interference type provided in an embodiment of the present invention set to intermittent sampling interference;

[0046] Figure 4b A separation echo signal restoration diagram in which the interference type provided in the embodiment of the present invention is set to intermittent sampling interference;

[0047] Figure 4c A pulse pressure diagram of a separated target signal with the interference type provided in an embodiment of the present invention set to intermittent sampling interference;

[0048] Figure 4d A coherent diagram of a separated target signal with the interference type provided in an embodiment of the present invention set to intermittent sampling interference;

[0049] Figure 5a A time-frequency diagram of a separated target signal with an SNR adjusted to -10 dBd provided in an embodiment of the present invention;

[0050] Figure 5b A restored image of a separated echo signal with the SNR adjusted to -10dBd provided in an embodiment of the present invention;

[0051] Figure 5cA pulse compression diagram of a separated target signal with an SNR adjusted to -10dBd provided in an embodiment of the present invention;

[0052] Figure 5d A coherence diagram of a separated target signal with an SNR adjusted to -10 dBd provided in an embodiment of the present invention;

[0053] Figure 6a A time-frequency diagram of a separated target signal when the interference distance unit provided by an embodiment of the present invention is the same as that of the target;

[0054] Figure 6b A recovery diagram of the separated echo signal when the interference distance unit and the target are the same as provided in an embodiment of the present invention;

[0055] Figure 6c A pulse pressure diagram of a separated target signal when the interference distance unit provided by an embodiment of the present invention is the same as the target;

[0056] Figure 6d A coherent diagram of a separated target signal when the interference distance unit provided by an embodiment of the present invention is the same as the target;

[0057] Figure 7a A time-frequency diagram of a separated target signal when a jammer is set at a distance of 3000m based on a target at a distance of 2000m provided in an embodiment of the present invention;

[0058] Figure 7b A pulse pressure diagram of a separated target signal when a jammer is set at a distance of 3000m based on a target at a distance of 2000m provided in an embodiment of the present invention;

[0059] Figure 7c A coherence diagram of separated target signals when a jammer is set at a distance of 3000m based on a target at a distance of 2000m provided in an embodiment of the present invention;

[0060] Figure 8a The time-frequency diagram of the original target signal whose signal type is LFM signal provided in the embodiment of the present invention;

[0061] Figure 8b A time-frequency diagram of a separated target signal whose signal type is an LFM signal provided in an embodiment of the present invention;

[0062] Figure 8c A separation echo signal restoration diagram of a LFM signal provided in an embodiment of the present invention;

[0063] Figure 8d The signal type provided in the embodiment of the present invention is a separated echo signal pulse pressure diagram of an LFM signal;

[0064] Figure 8eA separation target signal coherence diagram of a LFM signal provided in an embodiment of the present invention;

[0065] Figure 8f A separation interference signal coherence diagram of a LFM signal provided in an embodiment of the present invention;

[0066] Figure 9a The signal type provided in the embodiment of the present invention is a time-frequency diagram of a separated target signal in which the interference type of the LFM signal is set to intermittent sampling interference;

[0067] Figure 9b The signal type provided in the embodiment of the present invention is a time-frequency diagram of a separated echo signal in which the interference type of the LFM signal is set to intermittent sampling interference;

[0068] Fig.9c The signal type provided in the embodiment of the present invention is a separated target signal pulse pressure diagram in which the LFM signal interference type is set to intermittent sampling interference;

[0069] Figure 9d The signal type provided in the embodiment of the present invention is a separation target signal coherence diagram in which the LFM signal interference type is set to intermittent sampling interference;

[0070] Fig.10a The signal type provided in the embodiment of the present invention is a time-frequency diagram of a separated target signal with an LFM signal SNR set to 0 dB;

[0071] Fig.10b The signal type provided in the embodiment of the present invention is an LFM signal, and the time-frequency diagram of the separated echo signal with the SNR set to 0dB;

[0072] Fig.10c The signal type provided in the embodiment of the present invention is an LFM signal, and the separated target signal pulse pressure diagram with SNR set to 0dB;

[0073] Fig.10d The signal type provided in the embodiment of the present invention is an LFM signal, and the separation target signal coherence diagram with the SNR set to 0dB;

[0074] Fig.11a The signal type provided for the embodiment of the present invention is an LFM signal, and the interference forwarding times are 1-3 times of the random original target signal time-frequency diagram;

[0075] Fig.11b The signal type provided for the embodiment of the present invention is an LFM signal, and the interference forwarding times are 1-3 times of random separation target signal time-frequency diagram;

[0076] Fig.11cThe signal type provided for the embodiment of the present invention is an LFM signal, and the interference forwarding times are 1-3 times of random separation echo signal restoration diagram;

[0077] Fig.11d The signal type provided for the embodiment of the present invention is an LFM signal, and the interference forwarding times are 1-3 times of random separation echo signal pulse pressure diagram;

[0078] Fig.11e The signal type provided for the embodiment of the present invention is an LFM signal, and the interference forwarding times are 1-3 times of random separation target signal coherence diagram;

[0079] Fig.11f The signal type provided for the embodiment of the present invention is an LFM signal, and the interference forwarding times are 1-3 times of random separation interference signal coherence diagram;

[0080] Fig.12 The originally generated target echo signal provided by the embodiment of the present invention is superimposed with an interference image. DETAILED DESCRIPTION

[0081] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are 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 creative work are within the scope of protection of the present invention.

[0082] The present invention provides a radar target and interference signal separation method based on a lightweight neural network, such as Figure 1 As shown, the method includes the following steps S101 to S103.

[0083] S101, preprocessing the time domain echo signal to obtain a preprocessed signal and a full feature corresponding to the preprocessed signal;

[0084] Specifically, in step S101, the time domain echo signal is preprocessed to obtain a preprocessed signal, including the following steps S1011 to S1016.

[0085] S1011, loading the time domain echo signal and initializing the variables.

[0086] S1012, reading a data file in a designated path, and loading intermittent sampling forwarding interference data and target echo data in a complex form; wherein the target echo data includes noise data.

[0087] S1013, performing data shape transformation on the intermittently sampled forwarding interference data and the target echo data respectively to obtain reshaped interference data and reshaped target echo data.

[0088] S1014, extracting the real part and imaginary part of the reshaped interference data and the real part and imaginary part of the reshaped target echo data respectively.

[0089] S1015, alternately sorting the real part and the imaginary part of the reshaped interference data by a cyclic interpolation sorting method to obtain a long continuous interference data vector, and alternately sorting the real part and the imaginary part of the reshaped target echo data to obtain a long continuous target echo data vector.

[0090] S1016, adding the continuous interference data long vector and the continuous target echo data long vector to obtain a preprocessing signal.

[0091] Exemplarily, data loading and variable initialization are first performed: by reading the .mat file stored in the specified path, the intermittent sampling forwarding interference data and target echo data in complex form are loaded; wherein the target echo data includes noise data. The data dimensions of the intermittent sampling forwarding interference data and target echo data are determined to ensure that they conform to the predetermined shape (for example, 10 frames of echo data are used in the experiment, each frame has 64 pulses, and each frame of data has 6400 range sampling points). <nrn>). Then the data is reshaped and combined: the data shape is converted from a multi-frame format to a continuous data stream, so that the originally scattered data segments are rearranged into continuous long vectors. The reshaped interference data and the reshaped target echo data are expanded so that their dimensions meet the input requirements of the model. Including extracting the real and imaginary parts from the complex matrices of the reshaped interference data and the reshaped target echo data respectively; data type conversion: after the extracted real and imaginary data are converted into data types, tensor processing is performed in PyTorch. The data type is converted into a format suitable for deep learning framework processing, which ensures the numerical accuracy of the data while allowing the data to smoothly participate in the training and verification process of the model; then cyclic interpolation sorting is performed: in order to effectively process time series data and utilize the characteristics of complex signals, the real and imaginary data are alternately sorted (Alternating Sorting). Specifically, for each sample, the real data is even-indexed and the imaginary data is odd-indexed. This alternating sorting method can maintain the time series coherence of the data, while enabling the subsequent model to better learn the phase information in the data and reduce the impact of data redundancy and noise to a certain extent. Finally, data mixing is performed: the continuous interference data long vector is added to the continuous target echo data long vector to form the final mixed data, and the data validity is checked to ensure that each data segment meets the model input requirements, filter out data with insufficient length and inconsistent dimensions, and return debug information. The mixed data that meets the requirements will be used in subsequent separation tasks.

[0092] S102, input the preprocessed signal into the trained hybrid attention mechanism network to obtain the separated target echo signal. Figure 2 As shown, the hybrid attention mechanism network includes an encoder module, a separation module, an inter-stage connection module and a decoder module.

[0093] The encoder module is used to encode all features and extract multi-scale features of the preprocessed signal; specifically, Figure 3 As shown, the encoder module includes a first depth-separable convolution layer (DWConv), a second depth-separable convolution layer (DWConv), a third depth-separable convolution layer (DWConv), a first average pooling layer (Average Pooling), a second average pooling layer (Average Pooling) and a third average pooling layer (Average Pooling) connected in sequence.

[0094] The first depth-wise separable convolution layer is used to perform feature extraction on all features to obtain 1 / 2-scale features; the second depth-wise separable convolution layer is used to perform feature extraction on 1 / 2-scale features to obtain 1 / 4-scale features; the third depth-wise separable convolution layer is used to perform feature extraction on 1 / 4-scale features to obtain 1 / 8-scale features; the first average pooling layer is used to perform average pooling operations on 1 / 4-scale features to obtain first 1 / 8-scale features; the second average pooling layer is used to perform average pooling operations on 1 / 2-scale features to obtain second 1 / 8-scale features; the third average pooling layer is used to perform average pooling operations on all features to obtain third 1 / 8-scale features.

[0095] The separation module is used to fuse multi-scale features to obtain fused features, and use the global attention module to calculate the global attention weight to obtain a global self-attention map; specifically, the separation module includes a global attention module.

[0096] For example, the input of the global attention module is the multi-scale features [F 1 ,F 2 ,…,F n ] The multi-scale features are fused to form a comprehensive feature representation. The global attention weight is calculated and the attention map A is obtained through the softmax function. GA , A GA =softmax(f(F 1 ,F 2 ,…,F n )); where f(·) represents the function for calculating the similarity between features.

[0097] The global attention module (Global Attention) includes: TasNet Layer can modulate the features of different scales in the encoder from top to bottom using the attention signals obtained from multi-scale features.

[0098] The inter-level connection module is used to restore the accuracy of multi-scale features to obtain restored multi-scale features; specifically, Figure 3 As shown, the inter-stage connection module includes: three sequentially connected first up-sampling modules (Up-sample 2×), second up-sampling modules (Up-sample 4×) and third up-sampling modules (Up-sample 8×).

[0099] The first upsampling module is used to perform a two-fold upsampling of the dot product result of the fusion feature and the 1 / 8 scale feature according to the global self-attention map, and perform a dot product of the sampling result with the 1 / 4 scale feature to obtain a first restored multi-scale feature; the second upsampling module is used to perform a four-fold upsampling of the first restored multi-scale feature according to the global self-attention map, and perform a dot product of the sampling result with the 1 / 2 scale feature to obtain a second restored multi-scale feature;

[0100] The third upsampling module is used to upsample the second restored multi-scale feature eight times according to the global self-attention map, and perform dot multiplication of the sampling result with the full feature of the preprocessed signal to obtain the third restored multi-scale feature.

[0101] The decoder module is used to modulate the restored multi-scale feature pairs using the global self-attention map to obtain the separated target echo signal; wherein the multi-scale features include: 1 / 2 scale features, 1 / 4 scale features and 1 / 8 scale features; the restored multi-scale features include: first restored multi-scale features, second restored multi-scale features and third restored multi-scale features.

[0102] Specifically, Figure 3 As shown, the decoder module includes a first cascaded local attention layer (LALayer), a second cascaded local attention layer (LA Layer) and a third cascaded local attention layer (LA Layer) connected in sequence;

[0103] The first cascaded local attention layer is used to extract the local attention feature according to the first restored multi-scale feature and the second restored multi-scale feature to obtain the first local attention feature;

[0104] The second cascaded local attention layer is used to extract the local attention feature according to the first local attention feature and the third restored multi-scale feature to obtain the second local attention feature;

[0105] The third cascaded local attention layer is used to extract local attention features based on the second local attention features and the full features of the preprocessed signal to obtain a third local attention feature.

[0106] Specifically, the decoder module is used to use the global self-attention map to modulate the restored multi-scale feature pairs to obtain a separated target echo signal, including: mapping the third local attention feature to the same spatial dimension as the time domain echo signal to obtain a separated target echo signal.

[0107] like Figure 3 As shown, a local attention layer is used in the decoder to restore the features to high resolution.

[0108] Specifically, in step S102, the training of the hybrid attention mechanism network includes: dynamically adjusting the learning rate of the hybrid attention mechanism network according to the evaluation separation index.

[0109] Here, the calculation formula for evaluating the separation index is expressed as:

[0110] in, represents the estimated target echo signal; s represents the pure source signal; S target represents the target signal part extracted from the estimated signal; e noise represents the noise part, that is, the difference between the estimated signal and the target signal; s target Represents the signal-to-noise ratio, which is used to evaluate the quality of the target signal.

[0111] The similarity can be better estimated through the projection form of the inner product. Since the loss in the network decreases from large to small, the negative value of SI-SNR should be taken. The improvement of SI-SDR can indicate the improvement of separation performance.

[0112] S103, obtaining specific information of the estimated target based on the separated target echo signal.

[0113] Specifically, in step S103, specific information of the target is estimated based on the separated target echo signal, including the following steps S1031 to S1032.

[0114] S1031, determining the signal type of the time domain echo signal, and performing pulse compression processing on the separated target echo signal according to the corresponding pulse compression method to obtain a pulse compression signal;

[0115] S1032, performing coherent processing on the pulse pressure signal to obtain a coherent processing signal; wherein the coherent processing signal includes specific information of the target.

[0116] Exemplarily, the separated target echo signal is imported into Matlab, and the separated waveform data of corresponding length is imported according to the actual data settings. For example, one frame has N pulses. Since the dual module performs cyclic interpolation of the real and imaginary parts, to restore a frame of echo signal, it is necessary to import separated data with a length of 2N. By calculating the stored row index, it is ensured that the real and imaginary parts are spliced ​​in sequence, and they are merged into a complex matrix to store the signal and interference separately.

[0117] When it is a frequency-agile signal, it is not possible to directly use a matched filter for pulse compression. It is necessary to construct M sub-filters to complete the pulse compression processing. For the reconstructed pulse signal Each sub-pulse is subjected to segmented pulse compression processing.

[0118] Specifically, the agile frequency signal is expressed as:

[0119] When the agile frequency signal is subjected to pulse compression processing, the pulse compression signal of each sub-pulse is expressed as:

[0120]

[0121] Where rect(·) represents the rectangular pulse signal for the ratio of time series to pulse; t represents the time series, defined as m represents the index of the current pulse number; T S represents the width of the pulse; b(m) represents the frequency offset (modulation parameter) associated with the mth pulse; Δf s represents the frequency increment; u(t) represents the complex envelope of the signal, Indicates fast time, t m Indicates slow time.

[0122] Then, all sub-pulses are accumulated within the pulse, with relatively small pulse pressure loss. When using linear frequency modulation (LFM) signals, segmented pulse compression is not required. The input signal is directly subjected to Fourier transform (FFT) and multiplied by the conjugate spectrum H of the reference chirp signal, which is matched filtering. Coherent accumulation uses a more traditional algorithm, that is, coherent superposition is performed at the same position after different pulses are aligned. Since the processing is performed in the frequency domain, FFT operation is required. It should be noted that when the separation result is good, that is, when the phase information loss is small, a more ideal coherent processing result can be obtained.

[0123] The LFM signal is expressed as:

[0124] Among them, t represents the time series, A s represents the amplitude, f 0 represents the center frequency, T p represents pulse and B represents bandwidth.

[0125] In a specific embodiment provided by the present invention, for intermittent sampling forwarding interference, the experiment first uses an agile waveform, and in the first experiment group, the transmission waveform is first set to a fixed coding sequence signal within a pulse.

[0126] Main parameter settings of Experiment ① in Experiment Group 1:

[0127] Table 1 Dataset parameter settings

[0128]

[0129]

[0130] The interference type is set to intermittent sampling interference (sampling 1 to 3), the SNR is set to 0dB, and the interference-to-noise ratio (JNR) and interference-to-signal ratio (CTR) are calculated to be 22.92dB and 22.92dB respectively.

[0131] After the hybrid attention mechanism network provided by the present invention, the results obtained are as follows Figures 4a to 4d shown.

[0132] like Figures 4a to 4d It can be seen that when the SNR is fixed at 0dB and the CTR is 22.92dB, the network has good separation performance for targets at distances of 2000m and 3000m respectively when the interference is one-to-three. Further increase the noise, adjust the SNR to -10dB, and the CTR remains unchanged at 22.92dB. The following experiment ②, the experimental results are as follows Figure 5a to Figure 5d shown.

[0133] like Figure 5a As shown in Figure 2, the separated signal is not very clear due to the increase of noise. Figure 4c The pulse pressure effect is not as good as before, but the speed of the target signal can still be easily obtained through coherent accumulation, which is consistent with the preset parameters.

[0134] Considering whether the interference and target can be separated relatively cleanly in similar distance units or even in the same distance unit, based on experiment ②, the interference distance unit is set to be the same as the target, and the following experiment ③ is performed:

[0135] like Figures 6a to 6d As shown in the figure, the effect is similar to that of experiment ②. We further consider whether we can separate echo signals with larger interference-to-signal ratio and interference-to-noise ratio. At this time, the target normalized amplitude is 1, and the interference is 100, that is, the interference-to-signal ratio is 40dB, and the corresponding SNR remains unchanged at -10dB, then the interference-to-noise ratio is 30dB. Based on the target distance of 2000m, we set the experiment ④ with the jammer distance of 3000m. The results are as follows Figures 7a to 7c As shown:

[0136] Then we tried to use LFM as the radar transmission signal. At the same time, we further improved the randomness of the experiment and conducted the second experiment.

[0137] (2) Main parameter settings of Experiment ① of Experiment Group 2:

[0138] Table 2 Dataset parameter settings

[0139] parameter Numeric parameter Numeric Target amplitude 1 Interference amplitude 100 Target distance / m 2000 Interference initial distance / m 2300 Target speed / m / s 50 Jammer speed / m / s 75

[0140] The transmission signal type is changed to LFM signal, the interference type is set to intermittent sampling interference (sampling 1 to 1), the SNR is set to -10dB, and the interference-to-noise ratio (JNR) and interference-to-signal ratio (CTR) are calculated to be 30dB and 40dB respectively.

[0141] like Figures 8a to 8d The figure shows the post-processing diagram of the separated signal ① in Experiment 2; Figures 8a to 8d It was found that LFM signals have better separation performance than agile frequency signals under approximate conditions. Since the network structure is similar to long short-term memory and is more sensitive to causality, and the continuity of LFM signals in the time-frequency domain is much greater than that of agile frequency signals, its causality will also be better.

[0142] The main parameter settings of Experiment ② of Experiment Group 2, where the jammer sampling time is random:

[0143] Table 3 Dataset parameter settings

[0144]

[0145]

[0146] The transmission signal type is still LFM signal, the interference type is set to intermittent sampling interference ① (sampling 1 to 1), intermittent sampling interference ② (sampling 1 to N, N is a random value), the noise mark position is set to 0, that is, no noise is added, and the signal-to-interference ratio (CTR) is calculated to be 40dB.

[0147] like Figures 9a to 9d As shown in the figure, the post-processing diagram of the separated signal ② in Experiment 2. Experiment ② is just a multi-interference situation under high interference-to-signal ratio. In order to verify the generalization ability and usability of the network, the speed and initial position of the target and interference are set to random values. The specific parameters are shown in Experiment ③ as follows:

[0148] Table 4 Dataset parameter settings

[0149]

[0150] The transmission signal type is still LFM signal, the interference type is set to intermittent sampling interference (sampling 1 to N, N is a random value, and the sampling time is random), the SNR is set to 0dB, and the dry-to-noise ratio (JNR) and dry-to-signal ratio (CTR) are calculated to be 20dB and 20dB respectively.

[0151] The experimental results of experiment ③ are as follows Figures 10a to 10d As shown by Fig.10d It is found that the coherence result still has tiny side lobes, which are judged to be residual interference signals. Since the original data generated at this time is relatively random, that is, each frame of data sent to the network training is basically different, and the total number of training samples is not large, it is normal that the separation effect is weakened. Figures 11a to 11f The figure shows a time-frequency diagram of part of the input data.

[0152] The difference of the original generated data can be clearly seen in the figure above. The starting positions of the target echo and the interference are different, and the random speed information cannot be judged from the figure above. In addition, the sampling time and forwarding times of the interference are also random. If it exceeds the fixed range, it will be truncated. The original generated target echo signal superimposed with the interference is as follows Fig.12 shown.

[0153] Compare with the above picture Fig.10a It is found that the interference residue is quite obvious. The reason is as analyzed above, which comes from the differences between the input data and the limitation of the number of samples. It should be noted that all time-frequency diagrams need to be uniformly down-converted.

[0154] It can be known from the description of the above implementation mode that a person skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary hardware. Based on such an understanding, the technical solution of the present invention, in essence or in other words, the part that contributes to the prior art, can be embodied in the form of a software product, or can be embodied in the implementation process of data migration. The computer software product can be stored in a storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for a computer device (which can be a personal computer, a mobile terminal, a server, or a network device, etc.) to execute the methods described in the various embodiments of the present invention or certain parts of the embodiments.

[0155] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. All or part of the present invention can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld devices or portable devices, tablet devices, mobile communication terminals, multi-processor systems, microprocessor-based systems, programmable electronic devices, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc.

[0156] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some or all of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the present invention.< / nrn>

Claims

1. A radar target and interference signal separation method based on lightweight neural network, characterized in that: include: Preprocessing the time domain echo signal to obtain a preprocessed signal and a full feature corresponding to the preprocessed signal; The preprocessed signal is input into a trained hybrid attention mechanism network to obtain a separated target echo signal; wherein the hybrid attention mechanism network includes an encoder module, a separation module, an inter-stage connection module and a decoder module; the encoder module is used to encode the full features and extract multi-scale features of the preprocessed signal; The separation module is used to fuse the multi-scale features to obtain fused features, and use the global attention module to calculate the global attention weight to obtain a global self-attention map; the inter-level connection module is used to restore the accuracy of the multi-scale features to obtain restored multi-scale features; the decoder module is used to modulate the restored multi-scale feature pair using the global self-attention map to obtain a separated target echo signal; The specific information of the target is estimated based on the separated target echo signal.

2. The radar target and interference signal separation method based on lightweight neural network according to claim 1 is characterized in that: The preprocessing of the time domain echo signal to obtain a preprocessed signal includes: Loading the time domain echo signal and initializing variables; Read the data file in the specified path, and load the intermittent sampling forwarding interference data and target echo data in plural form; wherein the target echo data includes noise data; Performing data shape transformation on the intermittent sampling forwarding interference data and the target echo data respectively to obtain reshaped interference data and reshaped target echo data; Respectively extracting the real part and the imaginary part of the reshaped interference data and the real part and the imaginary part of the reshaped target echo data; The real part and the imaginary part of the reshaped interference data are alternately sorted by a cyclic interpolation sorting method to obtain a continuous interference data long vector, and the real part and the imaginary part of the reshaped target echo data are alternately sorted to obtain a continuous target echo data long vector; The continuous interference data long vector is added to the continuous target echo data long vector to obtain a preprocessing signal.

3. The radar target and interference signal separation method based on lightweight neural network according to claim 1 is characterized in that: The training of the hybrid attention mechanism network includes: dynamically adjusting the learning rate of the hybrid attention mechanism network according to the evaluation separation index.

4. The radar target and interference signal separation method based on lightweight neural network according to claim 3 is characterized in that: The calculation formula for evaluating the separation index is expressed as: in, Represents the estimated target echo signal; s represents the pure source signal; Starget represents the target signal part extracted from the estimated signal; enoise represents the noise part, that is, the difference between the estimated signal and the target signal; starget represents the signal enhancement signal-to-noise ratio, which is used to evaluate the quality of the target signal.

5. The radar target and interference signal separation method based on lightweight neural network according to claim 1 is characterized in that: The encoder module includes a first depthwise separable convolutional layer, a second depthwise separable convolutional layer, a third depthwise separable convolutional layer, a first average pooling layer, a second average pooling layer, and a third average pooling layer connected in sequence; The first depth-separable convolutional layer is used to extract features from the full features to obtain 12 scale features; The second depth-wise separable convolutional layer is used to extract the / 12 scale feature to obtain a / 14 scale feature; The third depth-wise separable convolutional layer is used to extract the / 14 scale feature to obtain a / 18 scale feature; The first average pooling layer is used to perform an average pooling operation on the / 14 scale feature to obtain a first 18 scale feature; The second average pooling layer is used to perform an average pooling operation on the 14-scale feature to obtain a second 18-scale feature; The third average pooling layer is used to perform average pooling on the entire feature to obtain a third 18-scale feature.

6. The radar target and interference signal separation method based on lightweight neural network according to claim 1 is characterized in that: The inter-stage connection module comprises: a first up-sampling module, a second up-sampling module and a third up-sampling module which are connected in sequence; The first upsampling module is used to perform a double upsampling on the dot product result of the fusion feature and the 18-scale feature according to the global self-attention map, and perform a dot multiplication on the sampling result and the 14-scale feature to obtain a first restored multi-scale feature; The second upsampling module is used to perform four-fold upsampling on the first restored multi-scale feature according to the global self-attention map, and perform dot multiplication of the sampling result and the 14-scale feature to obtain the second restored multi-scale feature; The third upsampling module is used to upsample the second restored multi-scale feature by eight times according to the global self-attention map, and perform dot multiplication of the sampling result with the full feature of the preprocessed signal to obtain the third restored multi-scale feature; wherein the 14-scale feature, the 14-scale feature and the / 18-scale feature are the multi-scale features.

7. The radar target and interference signal separation method based on lightweight neural network according to claim 1 is characterized in that: The decoder module includes a first cascaded local attention layer, a second cascaded local attention layer, and a third cascaded local attention layer connected in sequence; The first cascaded local attention layer is used to extract local attention features according to the first restored multi-scale features and the second restored multi-scale features to obtain first local attention features; The second cascaded local attention layer is used to extract local attention features according to the first local attention features and the third restored multi-scale features to obtain second local attention features; The third cascaded local attention layer is used to extract local attention features based on the second local attention features and the full features of the preprocessed signal to obtain a third local attention feature.

8. The radar target and interference signal separation method based on lightweight neural network according to claim 7 is characterized in that: The decoder module is used to modulate the restored multi-scale feature pair using the global self-attention map to obtain a separated target echo signal, including: The third local attention feature is mapped to the same spatial dimension as the time domain echo signal to obtain a separated target echo signal.

9. The radar target and interference signal separation method based on lightweight neural network according to claim 1 is characterized in that: The step of estimating specific information of the target according to the separated target echo signal comprises: Determine the signal type of the time domain echo signal, and perform pulse compression processing on the separated target echo signal according to the corresponding pulse compression method to obtain a pulse compression signal; The pulse pressure signal is subjected to coherent processing to obtain a coherent processing signal; wherein the coherent processing signal includes specific information of the target.

Images