Radar signal depth feature extraction method and system based on adversarial sample defense, electronic equipment and storage medium

Abstract

The invention provides a radar signal depth feature extraction method and system based on adversarial sample defense, electronic equipment and a storage medium, and relates to the technical field of radar signal processing.The radar signal depth feature extraction method comprises the steps that a spacecraft synthetic aperture radar original echo signal is collected and converted into a time-frequency feature map, and intra-pulse and inter-pulse features are extracted in parallel to generate a combined matrix; receiving a deception jamming signal by using a polarization radar group, generating a signal data body without polarization influence through complex coherent superposition, inputting the deception jamming signal and the signal data body into a feature decoupling adversarial network, and separating bullet micro-motion target feature data by means of mutual information maximization constraint; and finally, inputting into a multi-scale local attention module to extract multi-band characteristic components, performing adaptive weight coefficient fusion according to power entropy, performing polarization channel energy correction, and outputting anti-interference fingerprint characteristics with micro-Doppler characteristic enhancement, so that anti-interference processing of the spacecraft synthetic aperture radar signals can be realized. And warhead target anti-interference fingerprint features with micro-Doppler feature enhancement characteristics are effectively extracted.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radar signal processing, and particularly relates to a radar signal deep feature extraction method and system based on adversarial sample defense, an electronic device and a storage medium. BACKGROUND

[0002] In the process of missile penetration, the spacecraft synthetic aperture radar needs to have very high anti-interference capability. When the missile is flying, the enemy will release main lobe deception jamming to try to mislead the radar to identify and track the target, resulting in false targets in imaging, and interfering with the acquisition of key information such as the position and motion state of the real warhead. This requires that the radar signal processing method can accurately extract the real target features from the complex interference environment, especially the small motion features of the warhead, while suppressing the deception jamming to ensure the accuracy and reliability of the imaging, providing effective support for missile penetration and improving its battlefield survival probability.

[0003] At present, a relatively advanced countermeasure is an anti-deception jamming algorithm based on a multi-channel synthetic aperture radar. This method first separates the interference signal by canceling the stationary scene target through the split phase center antenna; then estimates the spatial position of the interference signal by using the along-track interference technology; finally, according to the difference in the space-time two-dimensional distribution of the interference signal and the target signal, the spatial filtering method of least square error pattern synthesis based on zero constraint and Doppler filtering are jointly used to realize the suppression of deception jamming.

[0004] This anti-deception jamming algorithm based on a multi-channel synthetic aperture radar has certain limitations. On the one hand, it has high requirements for hardware, which requires multiple channels to work together, increasing the complexity and cost of the system, and the synchronization and calibration accuracy between channels are strict. Once there is a deviation, the anti-interference effect will be greatly reduced. On the other hand, in a complex and variable interference environment, especially when there are multiple interference sources and the interference pattern is complex, the separation and suppression effect of the algorithm on the interference signal is unstable. For example, when the interference signal and the target signal have similar space-time two-dimensional distribution characteristics, it is difficult to accurately distinguish between the two, resulting in ineffective suppression of the interference, affecting the extraction of real target features by the radar, and reducing the reliability of radar signal processing in the process of missile penetration. SUMMARY

[0005] The present application aims to provide a radar signal deep feature extraction method and system based on adversarial sample defense, an electronic device and a storage medium, to solve the problem that the anti-deception jamming algorithm based on a multi-channel synthetic aperture radar in the prior art has high requirements for hardware, strict channel synchronization and calibration accuracy, and unstable interference separation and suppression effect in a complex interference environment, and it is difficult to distinguish similar characteristic signals.

[0006] To solve the above technical problems, in a first aspect, the application provides a radar signal deep feature extraction method based on adversarial sample defense, comprising:

[0007] Collecting original echo signals of a spacecraft synthetic aperture radar, and converting the original echo signals into time-frequency feature maps;

[0008] Parallelly extracting an instantaneous frequency variation curve of intra-pulse modulation features and a cumulative difference spectrum of inter-pulse repetition intervals from the time-frequency feature maps, performing matrix splicing on a derivative sequence of the instantaneous frequency variation curve and an envelope feature of the cumulative difference spectrum, and generating a feature combination matrix;

[0009] Receiving a deception jamming signal by using a polarized radar group, performing complex coherent superposition operation on horizontal direction signals and vertical direction signals of the deception jamming signal, and generating a signal data body not affected by polarization;

[0010] Inputting the feature combination matrix and the signal data body not affected by polarization into a feature decoupling adversarial network, and separating out target feature data representing a small movement of a warhead by mutual information maximization constraint;

[0011] Inputting the target feature data into a multi-scale local attention module, parallelly scanning a spatial structure of the target feature data by using a convolution kernel group with multiple levels of perception ranges, and extracting multi-band feature components excited by a projectile body movement;

[0012] Calculating adaptive weight coefficients according to power entropy distributions of feature maps output by each convolution kernel, performing pixel-level fusion on the multi-band feature components according to the adaptive weight coefficients, performing polarization channel energy correction processing on the fused feature maps, eliminating energy differences between channels caused by coherent superposition, and outputting anti-jamming fingerprint features with micro-Doppler feature enhancement characteristics.

[0013] Optionally, the calculation of adaptive weight coefficients according to power entropy distributions of feature maps output by each convolution kernel, the pixel-level fusion on the multi-band feature components according to the adaptive weight coefficients, and the polarization channel energy correction processing on the fused feature maps to eliminate energy differences between channels caused by coherent superposition to output anti-jamming fingerprint features with micro-Doppler feature enhancement characteristics, comprises:

[0014] Calculating power entropy distributions of feature maps output by each convolution kernel, the power entropy distributions being obtained by statistics of energy distribution discrete degrees of each pixel point in each feature map;

[0015] Calculating adaptive weight coefficients corresponding to each feature map according to the power entropy distributions;

[0016] The pixel positions of each feature map in the multi-band feature component are weighted and summed according to the adaptive weight coefficient to generate a preliminary fusion feature map;

[0017] The polarization channel energy correction processing is performed on the preliminary fusion feature map to eliminate the inter-channel energy difference caused by coherent superposition by adjusting the energy distribution of each polarization channel, so that the energy levels of different polarization channels are kept consistent, and the preliminary fusion feature map after energy correction is output as an anti-interference fingerprint feature with micro-Doppler feature enhancement characteristics.

[0018] Optionally, the polarization channel energy correction processing is performed on the preliminary fusion feature map to eliminate the inter-channel energy difference caused by coherent superposition by adjusting the energy distribution of each polarization channel, so that the energy levels of different polarization channels are kept consistent, and the preliminary fusion feature map after energy correction is output as an anti-interference fingerprint feature with micro-Doppler feature enhancement characteristics, comprising:

[0019] The energy mean of each polarization channel in the preliminary fusion feature map is calculated as the current energy level of each polarization channel, and a unified target energy level is determined based on the current energy levels of all polarization channels;

[0020] The energy adjustment coefficient corresponding to each polarization channel is calculated according to the ratio of the current energy level of each polarization channel to the target energy level;

[0021] The pixel values in each polarization channel are scaled based on the energy adjustment coefficient to eliminate the inter-channel energy difference caused by coherent superposition, so that the energy levels of each polarization channel are consistent;

[0022] The preliminary fusion feature map after energy correction is output as an anti-interference fingerprint feature with micro-Doppler feature enhancement characteristics.

[0023] Optionally, the feature combination matrix and the signal data body not affected by polarization are input into the feature decoupling adversarial network to separate out the target feature data representing the slight movement of the bullet by mutual information maximization constraint, comprising:

[0024] The feature combination matrix and the signal data body not affected by polarization are input into the feature decoupling adversarial network as input data;

[0025] The feature combination matrix and the signal data body not affected by polarization are subjected to nonlinear transformation and dimension lifting processing by the encoder of the feature decoupling adversarial network to extract deep high-dimensional feature representation;

[0026] The feature separation operation is performed on the high-dimensional feature representation in the decoupling layer of the feature decoupling adversarial network to obtain a plurality of feature subsets;

[0027] The mutual information value between each feature subset and the input data is calculated by mutual information maximization constraint, and the feature subset with the maximum mutual information value is selected as the target feature data representing the warhead micro-motion.

[0028] Optionally, the target feature data is input into a multi-scale local attention module, a plurality of convolution kernel groups with different perception ranges are used to scan the spatial structure of the target feature data in parallel, and multi-band feature components excited by the projectile motion are extracted, including:

[0029] The target feature data is input into a multi-scale local attention module;

[0030] In the multi-scale local attention module, a plurality of convolution kernel groups with different perception ranges are used to perform parallel convolution operations on the target feature data, and the convolution kernel groups include convolution kernels with three different sizes;

[0031] The fine spatial structure of the target feature data is scanned by the smallest size convolution kernel, and high-frequency feature components representing high-frequency motion details are extracted;

[0032] The local spatial structure of the target feature data is scanned by the medium size convolution kernel, and the medium-frequency feature components representing the medium-frequency motion change are extracted;

[0033] The global spatial structure of the target feature data is scanned by the largest size convolution kernel, and the low-frequency feature components representing the low-frequency motion trend are extracted;

[0034] The high-frequency feature components, medium-frequency feature components and low-frequency feature components are combined to form multi-band feature components excited by the projectile motion.

[0035] Optionally, the deception jamming signal is received by the polarization radar group, complex coherent superposition operation is performed on the horizontal direction signal and the vertical direction signal of the deception jamming signal to generate a signal data body not affected by polarization, including:

[0036] The horizontal direction signal of the deception jamming signal is received by the horizontal polarization channel of the polarization radar group, and the vertical direction signal of the deception jamming signal is received by the vertical polarization channel;

[0037] The horizontal direction signal and the vertical direction signal are respectively subjected to phase alignment processing to ensure that the two signals are completely synchronized in time;

[0038] The horizontal direction signal and the vertical direction signal after phase alignment are subjected to complex multiplication operation to obtain a product result representing the coherence characteristics of the two signals, and the product result is subjected to complex addition operation to obtain a complex complex signal;

[0039] Calculating the modulus of the complex composite signal obtains amplitude information representing signal intensity, and signal data volume not affected by polarization is generated based on the amplitude information.

[0040] Optionally, a first extraction process and a second extraction process are performed in parallel from the time-frequency feature map, the first extraction process identifies the frequency points with the strongest energy in each pulse period in the time-frequency feature map, and connects these frequency points in time sequence to form a instantaneous frequency variation curve;

[0041] A first extraction process and a second extraction process are performed in parallel from the time-frequency feature map, the first extraction process identifies the frequency points with the strongest energy in each pulse period in the time-frequency feature map, and connects these frequency points in time sequence to form a instantaneous frequency variation curve;

[0042] The second extraction process detects the time position of the pulse energy peak value in the time-frequency feature map and calculates the difference of the continuous pulse peak time interval and accumulates the sum of the differences to generate a cumulative difference spectrum;

[0043] The rate of change of each point on the instantaneous frequency variation curve is calculated to generate a derivative sequence, and the maximum value points in the cumulative difference spectrum are identified and connected synchronously to form an envelope feature;

[0044] The derivative sequence and the envelope feature are transversely or longitudinally matrix-spliced as two one-dimensional arrays to generate a feature combination matrix.

[0045] In a second aspect, the present application provides a radar signal deep feature extraction system based on adversarial sample defense, comprising:

[0046] The acquisition module is configured to acquire the original echo signal of the spacecraft synthetic aperture radar, and convert the original echo signal into a time-frequency feature map;

[0047] The first generation module is configured to extract, in parallel from the time-frequency feature map, an instantaneous frequency variation curve of the intra-pulse modulation feature and a cumulative difference spectrum of the inter-pulse repetition interval, and matrix-splice a derivative sequence of the instantaneous frequency variation curve and an envelope feature of the cumulative difference spectrum to generate a feature combination matrix;

[0048] The second generation module is configured to receive a spoofing jamming signal by using a polarized radar group, perform complex coherent superposition operation on the horizontal direction signal and the vertical direction signal of the spoofing jamming signal, and generate a signal data volume not affected by polarization;

[0049] The separation module is configured to input the feature combination matrix and the signal data volume not affected by polarization into a feature decoupling adversarial network, and separate out target feature data representing the micro-motion of the warhead by mutual information maximization constraint;

[0050] The extraction module is configured to input the target feature data into a multi-scale local attention module, utilize a plurality of convolution kernel groups with multi-level perception ranges to scan spatial structures of the target feature data in parallel, and extract multi-band feature components excited by the motion of the projectile body;

[0051] The output module is configured to calculate adaptive weight coefficients according to power entropy distributions of feature maps output by the respective convolution kernels, perform pixel-level fusion on the multi-band feature components according to the adaptive weight coefficients, perform polarization channel energy correction processing on the fused feature maps, eliminate energy differences between channels caused by coherent superposition, and output an anti-interference fingerprint feature with enhanced micro-Doppler feature characteristics.

[0052] In a third aspect, the present application provides an electronic device, comprising:

[0053] a memory configured to store a computer program;

[0054] a processor configured to implement steps of the radar signal deep feature extraction method based on adversarial sample defense according to the first aspect when executing the computer program.

[0055] In a fourth aspect, the present application provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program is executable by a processor to implement steps of the radar signal deep feature extraction method based on adversarial sample defense according to the first aspect.

[0056] The radar signal deep feature extraction method based on adversarial sample defense provided in the application lays a precise data foundation for subsequent feature extraction by collecting spacecraft synthetic aperture radar original echo signals and converting them into time-frequency feature maps; through parallel extraction of intra-pulse and inter-pulse key features and matrix splicing to generate a feature combination matrix, efficient integration of multi-dimensional target signal features is realized; the polarization radar group receives deception jamming signals and generates signal data bodies that are not affected by polarization through complex coherent superposition, effectively avoiding the interference of polarization differences on signal processing; the feature combination matrix and the signal data body without polarization influence are input into the feature decoupling adversarial network, and the target feature data representing the small movement of the warhead is accurately separated by means of mutual information maximization constraint, solving the problem of difficult extraction of target features under complex interference; multi-band feature components excited by the movement of the projectile body are extracted through a multi-scale local attention module, widening the feature perception range and ensuring that no multi-dimensional target features are missed; adaptive weight coefficients are calculated according to the power entropy distribution for pixel-level fusion, and the energy difference between channels is eliminated through polarization channel energy correction, finally outputting anti-jamming fingerprint features with micro-Doppler feature enhancement characteristics, which not only realizes effective defense against main lobe deception jamming, but also improves the recognition accuracy and reliability of real warhead targets, meeting the anti-jamming and precise target feature extraction requirements of spacecraft synthetic aperture radars during missile penetration.

[0057] Further, the power entropy distribution of each convolution kernel output feature map is calculated, and then the adaptive weight coefficient corresponding to each feature map is determined according to the power entropy distribution; subsequently, each pixel position of each feature map in the multi-band feature component is weighted and summed according to the coefficient, to generate a preliminary fusion feature map; finally, polarization channel energy correction processing is performed on the preliminary fusion feature map, the energy difference between channels caused by coherent superposition is eliminated by adjusting the energy distribution of each polarization channel, so that the energy levels of different polarization channels remain consistent, and the preliminary fusion feature map after energy correction is output as anti-jamming fingerprint features with micro-Doppler feature enhancement characteristics. Through adaptive weight coefficient calculation based on power entropy distribution, the multi-band feature component can accurately allocate weights according to the energy dispersion degree of the feature map during pixel-level fusion, improving the retention of key target information in the fused feature map; the polarization channel energy correction processing effectively solves the problem of energy imbalance between channels caused by coherent superposition, ensuring the energy consistency of the fused feature map, and the finally output anti-jamming fingerprint features not only have more stable micro-Doppler feature enhancement effect, but also can reduce the interference of energy difference on target recognition, further improving the reliability and discriminability of the fingerprint features. BRIEF DESCRIPTION OF DRAWINGS

[0058] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced. Obviously, the accompanying drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0059] Figure 1 A flowchart of a radar signal deep feature extraction method based on adversarial sample defense provided by an embodiment of the present application;

[0060] Figure 2 A specific implementation diagram of a radar signal deep feature extraction method based on adversarial sample defense provided by an embodiment of the present application;

[0061] Figure 3 A specific implementation diagram of a radar signal deep feature extraction method based on adversarial sample defense provided by an embodiment of the present application;

[0062] Figure 4 A structure diagram of a radar signal deep feature extraction system based on adversarial sample defense provided by an embodiment of the present application. DETAILED DESCRIPTION

[0063] In the main lobe deception jamming defense scene of spacecraft synthetic aperture radar in the process of missile penetration, the existing anti-deception jamming algorithm based on multi-channel synthetic aperture radar has obvious technical limitations: on the one hand, the algorithm relies on the cooperative work of multiple channels, and has high requirements for hardware configuration, which not only increases the system cost, but also puts forward strict requirements for the synchronization and calibration accuracy between channels. Once the channel synchronization deviates or the calibration accuracy is insufficient, the anti-jamming effect will be directly weakened. On the other hand, in the face of complex and variable jamming environment, especially when there are multiple interference sources and the time-space two-dimensional distribution characteristics of interference signals and target signals are similar, the separation and suppression effect of the algorithm on interference signals is unstable, and it is difficult to accurately distinguish between the two, which leads to the obstruction of the extraction of real missile target features, and the algorithm cannot meet the reliability requirements of radar target recognition and tracking. Therefore, a better anti-jamming signal processing scheme is urgently needed.

[0064] To solve the above problems, the application provides a radar signal deep feature extraction method based on adversarial sample defense. The method first collects radar original echo signals and converts them into time-frequency feature maps, extracts key features to generate a feature combination matrix, and simultaneously generates a signal data body that is not affected by polarization through polarization radar groups and complex coherence superposition. Then, with the help of feature decoupling adversarial network and mutual information maximization constraint, the target feature data representing the small movement of the warhead is accurately separated. Finally, multi-scale local attention modules are used to extract multi-band feature components, and adaptive weight fusion and polarization channel energy correction are used to output anti-interference fingerprint features with micro-Doppler feature enhancement characteristics. The scheme does not need to rely on complex multi-channel hardware, avoiding the problem of multi-channel synchronization calibration. Through mutual information maximization constraint and multi-scale feature processing, the target and interference signals can be accurately separated even in a complex interference environment. The polarization channel energy correction solves the problem of energy imbalance caused by coherence superposition, overcoming the defects of high hardware requirements and unstable interference separation of existing algorithms, and significantly improving the recognition accuracy and anti-interference reliability of the radar for real warhead targets.

[0065] In order for those skilled in the art to better understand the scheme of the present application, the present application will be further described in detail below in combination with the drawings and specific embodiments. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0066] The core of the present application is to provide a radar signal deep feature extraction method based on adversarial sample defense, and a specific implementation process of the method is shown in Figure 1 The method comprises:

[0067] S101, collecting the original echo signal of the spacecraft synthetic aperture radar, and converting the original echo signal into a time-frequency feature map;

[0068] In the above scheme, the original echo signal of the spacecraft synthetic aperture radar refers to the initial signal received by the radar after the electromagnetic wave emitted by the synthetic aperture radar carried on the spacecraft to the target. The time-frequency feature map is a graphical data that can reflect the characteristics of the signal in both time and frequency dimensions, in which the frequency components contained by the signal at different time points and the energy intensity of each frequency component can be seen.

[0069] In the embodiments of the present application, first, the synthetic aperture radar carried on the spacecraft actively emits electromagnetic waves of a specific frequency to the target. These electromagnetic waves will be reflected after encountering the target during propagation. The signal receiving device of the radar will capture the reflected electromagnetic waves. Then, the signal conversion module converts the captured electromagnetic wave physical signals into processable electrical signals. This electrical signal is the original echo signal containing information such as the position and motion state of the target, and the signal exists in the form of a time series signal, that is, the signal intensity data is recorded continuously over time.

[0070] Next, the original echo signal is preprocessed to ensure the stability of the signal, and then a short-time Fourier transform spectrum analysis technique is used. First, appropriate time window parameters are set. The continuous time series signal is divided into multiple overlapping small time periods according to the set time window. Then, the short-time Fourier transform is performed on the signal in each time window. Through this algorithm, all frequency components contained in the signal in each time window and the energy intensity corresponding to each frequency component are calculated to obtain the frequency spectrum data corresponding to each time window. Finally, the frequency spectrum data of all time windows are arranged in chronological order to form a two-dimensional matrix. The column direction of the matrix corresponds to the time axis, representing the chronological order of different time windows. The row direction of the matrix corresponds to the frequency axis, representing different frequency values contained in the signal. The numerical value of each cell in the matrix represents the energy intensity of the signal at the corresponding time window and corresponding frequency. This two-dimensional matrix is a time-frequency feature map that can simultaneously reflect the time, frequency, and energy characteristics of the signal.

[0071] In actual applications, for example, the synthetic aperture radar of the spacecraft emits electromagnetic waves to the missile. After the receiving device captures the reflected waves, it is converted into a time series electrical signal (original echo signal) of 0-10 seconds, recorded once every 0.001 seconds. Then, the time window length is set to 0.5 seconds, and the overlap ratio is 50%. The 0-10 second signal is divided into multiple overlapping windows. The short-time Fourier transform is used to calculate the frequency spectrum containing 50-1000 Hz frequency and corresponding energy for each window. Then, these frequency spectrums are arranged in chronological order to form a two-dimensional matrix with columns corresponding to 0-10 seconds of time, rows corresponding to 50-1000 Hz of frequency, and cell values as energy. That is, the conversion of the time-frequency feature map is completed, which is used for subsequent extraction of missile-related features.

[0072] The overall scheme of S101 described above, the acquisition of the original echo signal is the basis for target feature analysis, which provides initial data and avoids the lack of signals to carry out subsequent analysis. After the original echo signal is converted into a time-frequency feature map, the changes of the signal in the time and frequency dimensions can be intuitively presented, and the change rule of the signal frequency when the target moves can be more clearly reflected, providing a more comprehensive basis for target key feature extraction and identification.

[0073] S102, parallel extraction of the instantaneous frequency change curve of the intra-pulse modulation feature and the cumulative difference spectrum of the inter-pulse repetition interval from the time-frequency feature map, matrix splicing of the derivative sequence of the instantaneous frequency change curve and the envelope feature of the cumulative difference spectrum to generate a feature combination matrix;

[0074] Optionally, step S102 can specifically include the following steps:

[0075] S1021, parallel execution of a first extraction process and a second extraction process from the time-frequency feature map, the first extraction process identifying the frequency points with the strongest energy in each pulse period in the time-frequency feature map and connecting these frequency points in time sequence to form an instantaneous frequency change curve;

[0076] S1022, the second extraction process detecting the time position of the pulse energy peak value in the time-frequency feature map and calculating the difference of the continuous pulse peak time interval and cumulatively summing the differences to generate a cumulative difference spectrum;

[0077] S1023, calculating the change rate of each point on the instantaneous frequency change curve to generate a derivative sequence, and synchronously identifying and connecting the maximum value points in the cumulative difference spectrum to form an envelope feature;

[0078] S1024, matrix splicing of the derivative sequence and the envelope feature as two one-dimensional arrays in horizontal or vertical direction to generate a feature combination matrix.

[0079] In the above scheme, the intra-pulse modulation feature refers to the modulation characteristics of each pulse inside the radar signal, reflecting the signal change rule inside the pulse. The instantaneous frequency change curve is a curve formed by connecting the frequency points with the strongest energy in each pulse period in the time-frequency feature map in time sequence, which can reflect the change of frequency with time. The derivative sequence of the instantaneous frequency change curve is a sequence formed by arranging the frequency change rate of each point on the curve in time, which reflects the speed of frequency change. The envelope feature of the cumulative difference spectrum is the outline formed by connecting the maximum value points in the cumulative difference spectrum, which highlights the overall change trend. The feature combination matrix is a matrix formed by splicing the derivative sequence and the envelope feature, which is used to integrate the feature information of two dimensions.

[0080] In the embodiment of the present application, step S1021 first determines the time range of each pulse period in the time-frequency feature map, which requires combining the known pulse repetition period parameter of the radar signal to divide independent time intervals on the time axis of the time-frequency feature map, and each interval corresponds to a pulse period; then the energy detection algorithm is used to process the frequency axis data in each divided pulse period, and the energy intensity values corresponding to different frequency points in the period are compared one by one to find the frequency point with the maximum energy value, which is the frequency point that best represents the signal characteristics in the pulse period; finally, according to the order of the pulse periods on the time axis, the frequency points with the strongest energy found in each period are connected in sequence by line segments to form a curve that reflects the change of frequency with time, which is the instantaneous frequency change curve of the intrapulse modulation characteristics. For example, the time axis of the time-frequency feature map is from 0 seconds to 2 seconds, and the known pulse repetition period is 0.2 seconds, which can be divided into 10 pulse periods. In the 0-0.2 second period, the energies of all frequency points are compared, and it is found that the energy of 180 Hz is the largest, and so on. The 10 frequency points are found, and the instantaneous frequency change curve is obtained by connecting them in time order.

[0081] Secondly, through step S1022, pulse peak detection is performed in the time-frequency feature map. A suitable sliding window size is set, and the window is moved point by point on the time axis of the time-frequency feature map. By comparing the energy values of different time points in the window, the time position with the highest energy of each pulse signal is found. Then, according to these peak time positions, the time difference between two adjacent pulse peaks is calculated, and this time difference is the inter-pulse repetition interval. For example, the first peak is at 0.1 seconds and the second peak is at 0.3 seconds, and the corresponding inter-pulse repetition interval is 0.2 seconds. Then, the difference between consecutive inter-pulse repetition intervals is calculated, for example, the first inter-pulse repetition interval is 0.2 seconds and the second is 0.22 seconds, and the difference between them is 0.02 seconds. Finally, the calculated difference values are sequentially accumulated and summed in time order. Each new difference value is added to the previous cumulative sum, and the data sequence formed by arranging all the cumulative results in order is the cumulative difference spectrum of the inter-pulse repetition interval.

[0082] Then, the time coordinate and the corresponding frequency value of each data point on the instantaneous frequency change curve are determined by step S1023, and then the numerical differentiation algorithm is used to calculate the frequency change and the time change between each point and the next adjacent point on the curve, and the frequency change rate of the point is obtained by dividing the frequency change by the time change. The derivative values of all points on the curve are arranged in time order to form the derivative sequence of the instantaneous frequency change curve, for example, the points on the instantaneous frequency change curve are (0.1 seconds, 180 Hz), (0.3 seconds, 200 Hz), and (0.5 seconds, 190 Hz). The derivative of the first point is 100 Hz / s, and the derivative of the second point is -50 Hz / s. Arranging the two derivatives obtains the derivative sequence. The data sequence of the cumulative difference spectrum is subjected to maximum value detection, and the maximum value detection algorithm is used to compare the numerical values of each data point and its adjacent two data points in the sequence, and find the data point which is greater than the two adjacent data points. Then, according to the time sequence of the cumulative difference spectrum, the maximum value points are connected in turn by line segments, and the contour line is the envelope feature of the cumulative difference spectrum, for example, the data of the cumulative difference spectrum is 0.02, 0.04, 0.03, 0.05, and 0.04, the maximum value points are 0.04 and 0.05, and the envelope feature is obtained by connecting the two points.

[0083] Finally, the data form of the derivative sequence of the instantaneous frequency change curve and the envelope feature of the cumulative difference spectrum is determined by step S1024, both of which are one-dimensional arrays, and the lengths of the two arrays need to be consistent. If the lengths are different, the lengths need to be adjusted to the same length by interpolation or other methods to meet the splicing conditions. Then, the matrix splicing method is selected, and horizontal splicing or vertical splicing is selected according to the subsequent data processing requirements. Finally, the matrix splicing operation is performed, and the two one-dimensional arrays are combined according to the selected method by using matrix operation tools or programs to form a two-dimensional matrix, which is the feature combination matrix.

[0084] In practical applications, when a synthetic aperture radar carried by a spacecraft monitors the penetration process of a missile, after obtaining a time-frequency feature map of the missile, combined with the radar 0.3-second pulse repetition period parameter, 10 pulse periods (0-0.3 seconds, 0.3-0.6 seconds, …, 2.7-3.0 seconds) are divided on the time axis of the time-frequency feature map, an energy detection algorithm is used to determine the strongest frequency point in each period, and the time sequence is connected to form an instantaneous frequency change curve; a sliding window peak detection method is used to move the window on the time axis of the time-frequency feature map to compare the energy, find the peak time of the 10 pulses, calculate the pulse repetition interval, the difference value of adjacent intervals, and accumulate the sum to form a cumulative difference spectrum; the time and frequency values of the 10 points on the instantaneous frequency change curve are determined, the frequency change rate is calculated by using the forward difference method to obtain a derivative sequence, and the maximum value detection is performed on the cumulative difference spectrum to form an envelope feature; finally, it is confirmed that the derivative sequence and the envelope feature are both one-dimensional arrays and have the same length, and the two are combined into a 10-row 2-column feature combination matrix by the signal processing module through a horizontal splicing method, so as to prepare for subsequent separation of the micro-motion target feature of the missile.

[0085] The overall scheme of S102 can simultaneously obtain the intra-pulse and extra-pulse features by parallel extracting the instantaneous frequency change curve and the cumulative difference spectrum, avoid the limitation of a single feature, calculate the derivative sequence to reflect the frequency change rate and capture the target micro-motion information, generate the envelope feature to highlight the trend of the cumulative difference spectrum and reduce local interference, integrate the two types of features by matrix splicing to provide more comprehensive information for subsequent processing, and help to accurately identify the target feature and lay a solid foundation for defense against deception jamming and separation of real target features.

[0086] S103, receiving a deception jamming signal by using a polarized radar group, performing complex coherent superposition operation on horizontal direction signals and vertical direction signals of the deception jamming signal to generate a signal data body not affected by polarization;

[0087] Optionally, step S103 can specifically include the following steps:

[0088] S1031, receiving horizontal direction signals of the deception jamming signal by using a horizontal polarization channel of the polarized radar group, and receiving vertical direction signals of the deception jamming signal by using a vertical polarization channel;

[0089] S1032, performing phase alignment processing on the horizontal direction signals and the vertical direction signals respectively to ensure that the two signals are completely synchronized in time;

[0090] S1033, performing complex multiplication operation on the phase-aligned horizontal direction signals and vertical direction signals to obtain a product result representing the coherent characteristics of the two signals, and performing complex addition operation on the product result to obtain a complex complex signal;

[0091] S1034, calculate the modulus value of the complex composite signal to obtain amplitude information representing signal intensity, and generate signal data volume not affected by polarization based on the amplitude information.

[0092] In the above scheme, the polarization radar group refers to a combination of radar devices including a horizontal polarization channel and a vertical polarization channel, which is used to obtain multi-polarization information of the deception jamming signal. The deception jamming signal refers to a false signal released by the enemy to mislead the radar recognition, which includes signals of two different polarization directions, i.e., horizontal direction signals and vertical direction signals. The horizontal direction signal refers to a signal received by the horizontal polarization channel of the polarization radar group, whose vibration direction is parallel to the ground. The vertical direction signal refers to a signal received by the vertical polarization channel of the polarization radar group, whose vibration direction is perpendicular to the ground. The signal data volume refers to a signal data set after processing, which is no longer interfered by the difference between horizontal and vertical polarization directions, and includes information that can stably represent signal intensity, which can be used for subsequent fusion processing with target feature data.

[0093] In the embodiments of the present application, first, the polarization radar group is started through step S1031 to ensure that the horizontal polarization channel and the vertical polarization channel are in a normal working state. The two channels are components of the radar group specially used to receive signals of different polarization directions. The horizontal polarization channel is responsible for capturing the component of the signal whose vibration direction is parallel to the ground, and the vertical polarization channel is responsible for capturing the component whose vibration direction is perpendicular to the ground. Then, the receiving direction of the polarization radar group is aligned with the source direction of the deception jamming signal released by the enemy to ensure that the two channels can simultaneously receive the deception jamming signal. Then, the two channels start to receive signals synchronously. The horizontal polarization channel converts the received horizontal direction signal into a complex number form, and the vertical polarization channel also converts the received vertical direction signal into a complex number form, and finally obtains the original complex number signals of the two polarization directions.

[0094] Secondly, the horizontal direction signal and the vertical direction signal obtained in step S1031 are acquired through step S1032. In the transmission and reception process, the two signals may have a time delay due to the difference in response speed of the channel hardware and other factors, resulting in a mismatch in the phase of the signals at the same time. To solve this problem, a time synchronization algorithm is used. First, the timestamp information representing the time characteristics of the signals is extracted from the horizontal direction signal and the vertical direction signal. Then, the timestamps of the horizontal direction signal and the vertical direction signal are compared point by point to find the time deviation of the two signals at the same signal characteristic point. Finally, according to the time deviation, the time sequence of the vertical direction signal is adjusted so that the phase of the adjusted vertical direction signal and the horizontal direction signal at each corresponding time point is completely consistent, ensuring that the subsequent operation can be based on synchronized signals.

[0095] Then, the horizontal direction signal and the vertical direction signal after phase alignment in step S1032 are called by step S1033, and the two signals are completely synchronized in time at this time, and complex operation can be performed; first, complex multiplication operation is performed, according to complex multiplication rule, the complex numbers of each corresponding time point of the two signals are multiplied respectively, and the product result obtained can reflect the coherence characteristics between the two signals, that is, the correlation degree between the signals; after completing the complex multiplication operation, the product result is subjected to complex addition operation, according to complex addition rule, the product result is added to another group of complex signals at the same time point, and a composite complex signal which can integrate the overall information of the two signals is obtained.

[0096] Finally, the composite complex signal is obtained by step S1034, the modulus value of each time point of the composite complex signal is calculated, and the complex modulus value calculation formula is used, and the size of the modulus value directly represents the intensity of the signal at the corresponding time point; after calculating the modulus values of all time points, the modulus values are arranged in time sequence in turn to form a continuous data set containing only signal intensity information, and this data set is a signal data body which is not affected by polarization, because the modulus value is only related to the signal intensity, and is no longer disturbed by the phase and amplitude deviation caused by the difference between the horizontal and vertical polarization directions, and can be directly used for subsequent fusion processing with the characteristic combination matrix.

[0097] In actual application, in the scene of spacecraft monitoring missile penetration and encountering enemy released deception jamming signal, the polarization radar group carried by the spacecraft is started, it is confirmed that the horizontal polarization channel and the vertical polarization channel are in normal working state, the radar receiving direction is aligned to the source of the deception jamming signal, the two channels synchronously receive the signal, the horizontal polarization channel captures the horizontal direction signal with the vibration direction parallel to the ground and converts it into complex number 4+7i, and the vertical polarization channel simultaneously captures the vertical direction signal with the vibration direction perpendicular to the ground and converts it into complex number 2+4i; then it is monitored that due to the difference in hardware response speed, the horizontal direction signal and the vertical direction signal have time delay, the time synchronization algorithm based on signal feature point matching is used to extract the time stamp of the two signals accurate to 0.0001 seconds, and after point-by-point comparison, it is found that the vertical direction signal is 0.0006 seconds later than the horizontal direction signal, so all data points of the vertical direction signal are moved forward by 0.0006 seconds, so that the phases of the two signals at each corresponding time point are completely aligned; then the horizontal direction signal 4+7i and the vertical direction signal 2+4i after phase alignment are called, the product of the two is calculated according to the complex multiplication rule, the product result representing the coherence characteristics of the two signals is obtained, that is, -20+30i, then the product result is added to another group of signals at the same time point and after phase alignment according to the complex addition rule, -15+38i is calculated, and a composite complex signal is generated; finally, the modulus value of the composite complex signal is calculated according to the complex modulus value calculation formula And all the time point calculated modulus is arranged in time order, forming a signal data body containing only signal strength information, not affected by the polarization direction, for subsequent fusion processing with the feature combination matrix.

[0098] The overall scheme of S103 above can obtain complete polarization information by receiving dual-polarized signals through the polarized radar group, avoid missing of single channel information; phase alignment can realize signal time synchronization, eliminate time delay interference, and ensure accurate operation; complex multiplication and addition operation can integrate signal coherence characteristics and overall information to form a composite signal; calculating modulus can generate polarization-independent signal data body, reduce interference influence, and improve target feature separation accuracy.

[0099] S104, input the feature combination matrix and the polarization-independent signal data body into a feature decoupling adversarial network, and separate out target feature data representing the small movement of the bullet head by mutual information maximization constraint;

[0100] Optionally, step S104 can specifically include the following steps:

[0101] S1041, input the feature combination matrix and the polarization-independent signal data body into the feature decoupling adversarial network as input data;

[0102] S1042, perform nonlinear transformation and dimension lifting processing on the feature combination matrix and the polarization-independent signal data body by the encoder of the feature decoupling adversarial network, and extract deep high-dimensional feature representation;

[0103] S1043, perform feature separation operation on the high-dimensional feature representation in the decoupling layer of the feature decoupling adversarial network, and obtain a plurality of feature subsets;

[0104] S1044, calculate the mutual information value between each feature subset and the input data by mutual information maximization constraint, and select the feature subset with the maximum mutual information value as the target feature data representing the small movement of the bullet head.

[0105] In the above scheme, the feature decoupling adversarial network is an artificial intelligence network for separating mixed features, which includes an encoder, a decoupling layer and the like, and can be used to extract target exclusive features from mixed features. The deep high-dimensional feature representation is a data stream output by the encoder, which has high dimension and contains deep features, and can more accurately reflect the essential features of the data. The feature subset is a single-attribute feature fragment obtained after decoupling, which may represent target features, interference features, etc. The mutual information value is a numerical value quantifying the degree of association, which can be used as a basis for selecting target features. The target feature data is the feature subset that is finally selected and most closely related to the small movement of the bullet head, which can be used for subsequent accurate identification of the target.

[0106] In the embodiments of the present application, the feature combination matrix and the signal data body not affected by polarization are prepared through step S1041, and the two kinds of data respectively contain the basic features of the target and the processed interference signal strength information, which are the core of subsequent network processing; then check whether the format and dimension of the two kinds of data are suitable, if the time dimensions of the two are different, linear interpolation algorithm is used to adjust the length of the signal data body, so that it matches the number of time points of the feature combination matrix; after confirming the adaptation, the two kinds of data are spliced and integrated according to the time dimension, and then the integrated data is input into the feature decoupling adversarial network which is pre-built and trained, to provide complete input data source for the network. For example, the feature combination matrix is 15 rows and 2 columns, and the signal data body not affected by polarization is a one-dimensional array of 15 numerical values. After the length is adjusted by interpolation, the signal data body is added as the third column to the feature combination matrix to form 15 rows and 3 columns of integrated data, which is input into the feature decoupling adversarial network.

[0107] Secondly, the integrated data input into the feature decoupling adversarial network is called through step S1042, and the encoder module of the network will first process these data. The encoder uses a fully connected neural network as the core technology, and the first layer first performs nonlinear transformation and processes the integrated data through the ReLU activation function, breaks the original simple linear correlation of the data, and makes the data present a more complex change rule that is closer to the real target features, for example, there are -0.3, 0.5, -1.2, 2.1 and other numerical values in the integrated data, which become 0, 0.5, 0, 2.1 after ReLU function processing; after completing the nonlinear transformation, the encoder processes the dimension elevation through the subsequent multilayer neuron network, and each layer of neurons increases the feature dimension of the data, for example, the integrated data of 15 rows and 3 columns is elevated to 20 dimensions through the first layer of neurons, and then elevated to 50 dimensions through the second layer of neurons, and finally a deep high-dimensional feature representation with higher dimension and containing deep feature information is output, for example, after the encoder processing, the original 3-dimensional data of 15 time points becomes 50-dimensional data of 15 time points, and each dimension corresponds to a feature that can reflect the deep law of the target or interference.

[0108] Next, the deep high-dimensional feature representation output by the encoder is obtained through step S1043. These data are a comprehensive feature set mixed with target features, interference features, and background features, which need to be separated by the decoupling layer of the feature decoupling adversarial network. The decoupling layer uses an attention mechanism as the core algorithm to evaluate the importance of each dimension of the deep high-dimensional feature representation. By calculating the similarity between each dimension feature and the reference template representing the warhead micro-motion feature learned through the training data, it is determined whether the dimension is related to the target. The dimensions with high similarity are marked as "target-related", and the dimensions with low similarity are marked as "non-target-related". Next, according to the marking results, the deep high-dimensional feature representation is split into multiple independent feature subsets according to the attributes. For example, in a 50-dimensional deep high-dimensional feature, 20 dimensions are marked as "target-related", 15 dimensions are marked as "interference-related", and 15 dimensions are marked as "background-related". These 20 dimensions, 15 dimensions, and 15 dimensions are integrated to form three feature subsets, each containing only single-attribute features, avoiding interference between different attribute features.

[0109] Finally, through step S1044, multiple feature subsets output by the decoupling layer and the original integrated data input into the network are collected, and the target feature data representing the warhead micro-motion is selected through mutual information maximization constraint. The KL divergence estimation method is used to calculate the mutual information value between each feature subset and the original integrated data. The larger the mutual information value, the closer the association between the feature subset and the original data, and the more likely it contains the key target information in the original data. For example, the mutual information values of the three feature subsets and the original integrated data are 0.35, 0.88, and 0.22, respectively. By comparing these three values, the feature subset corresponding to the largest mutual information value of 0.88 is selected. This subset is the feature set most closely related to the warhead micro-motion, i.e., the target feature data representing the warhead micro-motion, which is used for subsequent multi-band feature extraction steps.

[0110] In practical applications, in the scenario of spacecraft monitoring missile penetration, the generated 12-row 2-column feature combination matrix containing 12 time point missile intra-pulse and inter-pulse features is combined with the generated one-dimensional signal data body composed of 12 time point interference signal processing strength information which is not affected by polarization. After confirming that the time dimensions of both are 12 time points, the signal data body is added to the feature combination matrix as the third column to form a 12-row 3-column integrated data input into the pre-built feature decoupling adversarial network trained. The network encoder adopts a fully connected neural network, and the integrated data is nonlinearly processed through a ReLU activation function. The 3-dimensional features are sequentially lifted to 25 dimensions and 60 dimensions through two-layer neural network, and 60-dimensional deep high-dimensional feature representation of 12 time points is output. The decoupling layer uses attention mechanism to calculate the similarity of 60-dimensional features and "missile warhead micro-movement target feature template", and 22-dimensional, 18-dimensional and 20-dimensional features are marked as "target related", "interference related" and "background related" respectively and integrated into three independent feature subsets. Finally, the three feature subsets and the original integrated data are collected, and the mutual information value is calculated by using the KL divergence estimation method. The calculation results are 0.4, 0.92 and 0.25 respectively. The "target related" feature subset corresponding to the maximum mutual information value 0.92 is selected as the target feature data representing the missile warhead micro-movement, which is used for subsequent multi-band feature extraction.

[0111] The overall scheme of S104 combines the feature combination matrix and the signal data body not affected by polarization into the network, integrates the target basic features and the interference processed information, and provides data for feature separation. The encoder extracts deep features through nonlinear transformation and dimension lifting, breaking through the limitation of shallow layer; the decoupling layer separates the mixed high-dimensional features into subsets, facilitating screening. The mutual information maximization constraint selects the strong correlation features of the warhead micro-movement, excludes interference and irrelevant background, improves the purity and accuracy of the features, and provides a basis for subsequent extraction of multi-band and generation of anti-interference fingerprint features.

[0112] S105, inputting the target feature data into a multi-scale local attention module, using a group of convolution kernels with different perception ranges to scan the spatial structure of the target feature data in parallel, and extracting multi-band feature components excited by the missile body movement;

[0113] Optionally, step S105 can specifically include the following steps:

[0114] S1051, inputting the target feature data into a multi-scale local attention module;

[0115] S1052, in the multi-scale local attention module, a group of convolution kernels with different perception ranges are used to perform parallel convolution operation on the target feature data, and the group of convolution kernels includes three different sizes of convolution kernels;

[0116] S1053, scanning the fine spatial structure of the target feature data by a minimum size convolution kernel to extract a high-frequency feature component representing high-frequency motion details;

[0117] S1054, scanning the local spatial structure of the target feature data by a medium size convolution kernel to extract a medium-frequency feature component representing medium-frequency motion changes;

[0118] S1055, scanning the global spatial structure of the target feature data by a maximum size convolution kernel to extract a low-frequency feature component representing low-frequency motion trends;

[0119] S1056, combining the high-frequency feature component, the medium-frequency feature component, and the low-frequency feature component to form a multi-band feature component excited by the projectile motion.

[0120] In the above scheme, the multi-scale local attention module is a processing module that can capture features from different ranges, including multiple size convolution kernels, which can be used to comprehensively extract the spatial structure features of the target data. The convolution kernel group refers to a set composed of convolution kernels of different sizes, which correspond to different feature perception ranges and can extract fine-grained, medium-grained, and coarse-grained features. The high-frequency feature component is feature data representing high-frequency motion details of the projectile head, reflecting rapid changes in motion. The local spatial structure is the spatial distribution information of the local region in the target feature data, corresponding to the medium-speed motion changes of the projectile head. The medium-frequency feature component is feature data representing medium-frequency motion changes of the projectile head, reflecting regular changes in motion. The global spatial structure is the overall spatial distribution information of the target feature data, corresponding to the slow and overall motion trend of the projectile head. The low-frequency feature component is feature data representing low-frequency motion trends of the projectile head, reflecting the overall regularity of motion. The multi-band feature component is a set formed by integrating high-frequency, medium-frequency, and low-frequency feature components, which can comprehensively reflect different frequency band features of projectile motion.

[0121] In the embodiments of the present application, first, it is confirmed whether the format of the target feature data meets the input requirements of the multi-scale local attention module, such as the module receiving three-dimensional data (time x feature dimension x channel number), while the target feature data may be two-dimensional data, such as a 15x25 two-dimensional matrix of 15 time points x 25-dimensional features. At this time, the reshape operation in the dimension adjustment algorithm is used to convert the two-dimensional data into three-dimensional data, such as 15x25x1, and the last "1" represents a single channel, which adapts the module input format. After confirming the data format adaptation, the adjusted target feature data is input into the pre-built multi-scale local attention module, which is equivalent to providing the module with "target motion feature materials to be analyzed". For example, after the reshape operation, the two-dimensional matrix of 15x25 is converted into three-dimensional data of 15x25x1, which is successfully input into the multi-scale local attention module, preparing for the subsequent extraction of multi-band features.

[0122] Secondly, a pre-configured multi-level perception range convolution kernel group is called through step S1052, which fixedly contains three different size convolution kernels; then parallel computing technology is adopted to make multiple convolution kernels work at the same time, avoiding time waste caused by sequential processing, and making the three size convolution kernels simultaneously perform convolution operation on the three-dimensional target feature data input in step S1051: each convolution kernel is like a "small window with filtering function", moving point by point on the data with fixed step, and calculating the product of all data in the window and the corresponding position value of the convolution kernel every time it moves to a position, and then adding all the products to obtain a new value, which is the feature value of the position; with the continuous sliding of the convolution kernel, a new feature matrix is generated, which is the preliminary feature data corresponding to the size of the convolution kernel.

[0123] Then, the feature components corresponding to the frequency bands are extracted through steps S1053-S1055: for the preliminary feature data generated by the smallest 3x3 convolution kernel, the 3x3 convolution kernel can accurately capture subtle changes in the data because it covers a small area each time, so it is used to scan the fine spatial structure of the target feature data, such as the small position shift and high-frequency jitter of the warhead within 0.008 seconds, and then through the feature selection rule, the part with high frequency of numerical value change in the preliminary feature is retained, and the high-frequency feature component representing high-frequency motion details is extracted from the preliminary feature of the 3x3 convolution kernel, for example, the feature value reflecting the vibration of the warhead 120 times per second is selected from the preliminary feature to form the high-frequency feature component; for the preliminary feature data generated by the medium size 5x5 convolution kernel, the 5x5 convolution kernel can capture the overall change of the local area of the data because it covers a moderate area, so it is used to scan the local spatial structure of the target feature data, such as the local flight trajectory and medium amplitude change of the warhead within 0.06 seconds, and then through the feature selection rule, the part with medium frequency of numerical value change in the preliminary feature is retained, and the medium-frequency feature component representing medium-frequency motion change is extracted from the preliminary feature of the 5x5 convolution kernel, for example, the feature value reflecting the movement of the warhead 45 meters per second is selected to form the medium-frequency feature component; for the preliminary feature data generated by the largest 7x7 convolution kernel, the 7x7 convolution kernel can capture the overall trend change of the data because it covers a large area, so it is used to scan the global spatial structure of the target feature data, such as the overall flight direction and slow attitude adjustment trend of the warhead within 0.6 seconds, and then through the feature selection rule, the part with low frequency of numerical value change in the preliminary feature is retained, and the low-frequency feature component representing low-frequency motion trend is extracted from the preliminary feature of the 7x7 convolution kernel, for example, the trend feature value reflecting the continuous flight of the warhead to the northeast is selected to form the low-frequency feature component.

[0124] Finally, the high-frequency feature component, the medium-frequency feature component, and the low-frequency feature component are collected through step S1056. At this time, the three components contain time point information corresponding to the target feature data, but it is necessary to ensure that the time points are completely aligned. For example, the high-frequency, medium-frequency, and low-frequency feature components correspond to 15 time points. If a component is missing time points due to convolution operation, the missing time points are supplemented with linear interpolation method to ensure that the time dimensions of the three components are consistent. After time alignment, the feature splicing technology is used to combine the three components in the "feature dimension" direction rather than the time direction, and the high-frequency, medium-frequency, and low-frequency feature values corresponding to each time point are integrated together. For example, at the 0.1 second time point, the high-frequency feature value is 0.8, the medium-frequency is 0.5, and the low-frequency is 0.3. The three values are combined as the comprehensive feature of the time point. After all the time points are combined, the set containing the multi-frequency band comprehensive features of each time point is formed, which is the multi-frequency band feature component excited by the projectile motion. For example, the set of 15 time points, each containing 3 types of frequency band features, can be used for subsequent feature fusion processing.

[0125] In practical applications, in the scenario of spacecraft monitoring missile penetration, when performing related operations, the target feature data representing the small motion of the missile warhead obtained in the filtering step S104 is a 12x20 two-dimensional matrix of 12 time points x 20 dimensions. Since the multi-scale local attention module needs to receive three-dimensional data, it is converted to 12x20x1 three-dimensional data through reshape dimension adjustment operation, and after confirming the format adaptation, it is input into the multi-scale local attention module. The module calls a convolution kernel group containing 3x3, 5x5, and 7x7 sizes, and uses parallel computing technology to allow the three convolution kernels to perform convolution operations on the 12x20x1 three-dimensional data at the same time. The 3x3 convolution kernel slides with a step size of 1, covering a 3x3 area each time to calculate the product sum and generate a 10x18x1 preliminary feature; the 5x5 convolution kernel generates a 8x16x1 preliminary feature; and the 7x7 convolution kernel generates a 6x14x1 preliminary feature. The preliminary feature of the 3x3 convolution kernel has a small coverage area and accurately captures the small jitter details of the missile warhead within 0.006 seconds, and according to the "preserve features with high value change frequency" rule, the high-frequency feature component reflecting the 110 times per second vibration of the warhead is extracted; the preliminary feature of the 5x5 convolution kernel has a moderate coverage area and captures the local flight trajectory changes of the warhead within 0.05 seconds, and according to the "preserve features with medium value change frequency" rule, the medium-frequency feature component reflecting the 40 meters per second movement of the warhead is extracted; and the preliminary feature of the 7x7 convolution kernel has a large coverage area and captures the overall flight direction of the warhead within 0.5 seconds, and according to the "preserve features with low value change frequency" rule, the low-frequency feature component reflecting the northwest flight direction of the warhead is extracted. Collecting the three frequency band feature components, it is found that the low-frequency feature component generated by the 7x7 convolution kernel has only 6 time points, and the missing 6 time point feature values are supplemented by linear interpolation method, so that the three correspond to 12 time points, and then they are spliced according to the "feature dimension" direction. For example, at the 0.2 second time point, the high-frequency feature value is 0.7, the medium-frequency is 0.4, and the low-frequency is 0.2, which are combined as the comprehensive feature of this time point. After integrating all time points, a multi-band feature component containing 12 time points and three types of frequency band features at each time point is formed, which is used for subsequent fusion processing.

[0126] The overall scheme of S105 above ensures that the features are smoothly input into the multi-scale module by reshaping the data format. The parallel operation of small, medium and large convolution kernels captures high-frequency details, medium-frequency changes and low-frequency trends, respectively, and completely preserves the missile body information. After time alignment processing to eliminate dimension differences, a multi-band feature is generated, providing data for subsequent processing and improving the accuracy and reliability of the anti-interference fingerprint feature.

[0127] S106, calculate adaptive weight coefficients according to the power entropy distribution of the feature maps output by each convolution kernel, perform pixel-level fusion on the multi-band feature components according to the adaptive weight coefficients, and perform polarization channel energy correction processing on the fused feature maps to eliminate the inter-channel energy difference caused by coherent superposition, to output an anti-interference fingerprint feature with micro-Doppler feature enhancement characteristics.

[0128] Optionally, step S106 can specifically include the following steps:

[0129] S1061, calculate the power entropy distribution of the feature maps output by each convolution kernel, the power entropy distribution being obtained by counting the energy distribution dispersion degree of each pixel point in each feature map;

[0130] S1062, calculate the adaptive weight coefficients corresponding to each feature map according to the power entropy distribution;

[0131] S1063, perform weighted summation calculation on each pixel position of each feature map in the multi-band feature components according to the adaptive weight coefficients, to generate a preliminary fused feature map;

[0132] S1064, perform polarization channel energy correction processing on the preliminary fused feature map, eliminate the inter-channel energy difference caused by coherent superposition by adjusting the energy distribution of each polarization channel, make the energy levels of different polarization channels consistent, and output the energy-corrected preliminary fused feature map as an anti-interference fingerprint feature with micro-Doppler feature enhancement characteristics.

[0133] Among them, step S1064 specifically includes the following processes: calculating the energy mean value of each polarization channel in the preliminary fused feature map as the current energy level of each polarization channel, and determining a unified target energy level based on the current energy levels of all polarization channels; calculating the energy adjustment coefficient corresponding to each polarization channel according to the ratio of the current energy level of each polarization channel to the target energy level; performing scaling processing on the pixel values in each polarization channel based on the energy adjustment coefficient, eliminating the inter-channel energy difference caused by coherent superposition, and making the energy levels of each polarization channel consistent; and outputting the preliminary fused feature map after energy correction as an anti-interference fingerprint feature with micro-Doppler feature enhancement characteristics.

[0134] In the above scheme, the power entropy distribution is obtained by statistics of the energy distribution of each pixel point in each feature map. The higher the dispersion degree, the greater the power entropy value, which can be used to judge the effective information content of the feature map. The adaptive weight coefficient is a value calculated according to the power entropy distribution, which is used to distribute the fusion weight of each feature map. The higher the power entropy value, the smaller the weight, which can ensure a higher proportion of effective features. The preliminary fusion feature map is a feature map generated after pixel-level fusion and not subjected to energy correction, which contains multi-band integrated features. The inter-channel energy difference is the phenomenon that different polarization channels have inconsistent energy levels due to coherent superposition, which affects the stability of the features. The anti-interference fingerprint feature is the final output feature data that can highlight the Doppler feature of the warhead micro-motion and has strong anti-interference ability, which can be used for accurate target recognition.

[0135] In the embodiments of the present application, as shown in Figure 2 first, the feature maps output by the three different size convolution kernels in step S105 are obtained through step S1061, and then the power entropy value of each feature map is calculated in turn using an entropy value calculation method, and a power entropy distribution is formed. First, the energy values of all pixel points in a single feature map are counted to determine the energy size of each pixel. Second, the total energy of the feature map is calculated by summing the energy values of all pixel points in the map. Third, the proportion p of the energy of each pixel point to the total energy of the feature map is calculated, which reflects the proportion of the energy of a single pixel in the overall feature map. Fourth, the proportion of the energy of each pixel is substituted into the power entropy formula i where H represents the power entropy value, and n represents the total number of pixels in the feature map. The power entropy value of the feature map is obtained by calculation. According to the same four-step process described above, the power entropy values of the intermediate frequency feature map and the low frequency feature map are calculated respectively. Finally, the power entropy values of the high frequency, intermediate frequency and low frequency feature maps are arranged in a fixed order, and the power entropy distribution of the feature maps output by the three convolution kernels is obtained.

[0136] Secondly, the power entropy distribution calculated by step S1061 is called through step S1062, and a function inversely proportional to the weight and the power entropy is used to calculate the adaptive weight coefficient corresponding to each feature map. The selected function is where w represents the adaptive weight coefficient of a certain feature map, H represents the power entropy value of the feature map, and ε is a very small positive number, usually close to 0, such as 0.01, which is used to avoid the case that the denominator is 0 when H is 0, and to ensure the normal calculation process. According to the function, the power entropy values of the high frequency, intermediate frequency and low frequency feature maps are substituted in turn to calculate the adaptive weight coefficient corresponding to each feature map. As can be seen from the calculation result, the higher the power entropy value of the feature map, the smaller the adaptive weight coefficient calculated, and these calculated adaptive weight coefficients will be the core basis for the subsequent fusion of multi-band feature components. ​

[0137] Then, the multi-band feature components generated in step S105 are obtained by step S1063, and the adaptive weight coefficients calculated in step S1062 are called again to generate a preliminary fusion feature map according to the rule of pixel-by-pixel weighted summation: for all pixels with the same position in the three feature maps, the energy values of the pixel in the high-frequency, medium-frequency and low-frequency feature maps are obtained respectively; then each energy value is multiplied by the adaptive weight coefficient of the corresponding feature map to obtain the contribution value of the pixel in the different frequency band feature maps; then the three contribution values of the same pixel position are summed to obtain the energy value of the pixel after fusion; all corresponding pixel positions in the multi-band feature components are sequentially executed by the weighted summation calculation according to the above method, and finally the energy values of all pixels after fusion are arranged according to the original pixel position to form a new feature map, which is the preliminary fusion feature map.

[0138] Finally, the implementation process of step S1064 is to obtain the preliminary fusion feature map first, and then execute the polarization channel energy correction processing in four steps: calculate the energy mean value of each polarization channel, count the energy values of all pixels in each polarization channel, divide the sum of these energy values by the total number of pixels to obtain the average value as the current energy level of the polarization channel; determine the unified target energy level, calculate the average value of the current energy levels of all polarization channels, and the average value is the target energy level to which all polarization channels need to be adjusted; calculate the energy adjustment coefficient of each polarization channel, divide the target energy level by the current energy level of each polarization channel to obtain the ratio, which is the energy adjustment coefficient corresponding to the polarization channel; scale the pixel values of each polarization channel, multiply the energy values of all pixels in each polarization channel by the energy adjustment coefficient of the channel, so that the adjusted energy mean values of each polarization channel are consistent with the target energy level, thereby eliminating the energy difference between channels caused by coherent superposition; after adjusting all channels, the preliminary fusion feature map after energy correction is output, and the feature map is the anti-interference fingerprint feature with micro-Doppler feature enhancement characteristics.

[0139] In actual application, in the scene of monitoring missile penetration of spacecraft, high-frequency, medium-frequency and low-frequency feature maps are obtained, and each pixel of the feature map contains an energy value. The energy values of all pixels of the three feature maps are counted respectively by using the entropy calculation method, the total energy of each feature map is calculated, the proportion of the energy of each pixel to the total energy is calculated, and finally the power entropy values of the high-frequency, medium-frequency and low-frequency feature maps are calculated by substituting the proportion into the power entropy formula to form a power entropy distribution in this order. The power entropy values of the three feature maps are sequentially substituted into the function of (ε takes 0.01) to calculate the respective adaptive weight coefficients, so as to ensure that the weight coefficient of the feature map with a high power entropy value is smaller; the multi-band feature components and the weight coefficients calculated in the foregoing are obtained, the contribution values of all pixels at the same position in the three feature maps are obtained by multiplying the energy values of the pixels by the corresponding weight coefficients, the three contribution values at the same position are summed to obtain the energy value of the pixel after fusion, and the preliminary fusion feature map is generated by arranging all the pixels according to the original positions after the calculation of all the pixels is completed; the preliminary fusion feature map is obtained, the energy values of all the pixels in the two channels are counted, the energy mean values of the two channels are calculated as the current energy levels, the average value of the current energy levels of the two channels is taken as the target energy level, the energy adjustment coefficients are obtained by dividing the target energy level by the current energy levels of the channels, the energy values of all the pixels in the channels are multiplied by the corresponding adjustment coefficients, so that the energy mean values of the two channels are consistent with the target energy level, the energy difference between the channels is eliminated, and finally the feature map after energy correction is output as the micro-Doppler feature enhancement anti-interference fingerprint feature of the missile.

[0140] The overall scheme of the above S106 assigns weights according to the calculation of power entropy to judge the information disorder degree of the feature map. Higher weights are set for the feature map with low power entropy to suppress the disorder information. Pixel-level fusion is performed according to the weights, multi-band features are integrated, high-frequency details, medium-frequency changes and low-frequency trends are considered, energy correction of the polarization channel is performed to eliminate the energy difference, the enhanced micro-Doppler feature and the target recognition fingerprint with strong anti-interference ability are output, and the recognition stability and precision are improved.

[0141] The following is a complete example for steps 101-106, as shown in Figure 3 In the scene of spacecraft monitoring missile penetration, the spacecraft carries a synthetic aperture radar to emit electromagnetic waves to the missile, receives the original echo signal reflected back, divides the signal into overlapping time windows, performs frequency spectrum analysis on each window to calculate the frequency components, arranges all the window spectra in time sequence to form a 256x5000 two-dimensional time-frequency feature map, in which the row represents the frequency, the column represents the time, and the value represents the energy intensity. The features are extracted from the time-frequency feature map in parallel: the region is divided according to the pulse period of 0.2 seconds, the energy detection algorithm is used to find the frequency point with the strongest energy in each period, and the time connection is formed to form an instantaneous frequency change curve; the peak value detection algorithm is used to find the peak values of 12 pulses (0.1 seconds, 0.3 seconds…2.3 seconds), calculate the continuous peak value intervals (0.2 seconds, 0.2 seconds…0.22 seconds) and the adjacent interval differences (0, 0…0.02 seconds), and accumulate the sum to generate a cumulative difference spectrum; the derivative sequence is obtained by using the numerical differentiation method to calculate the change rate of the instantaneous frequency curve, and the envelope feature is formed by connecting the maximum value points of the cumulative difference spectrum using the maximum value detection algorithm; the derivative sequence and the envelope feature are transversely spliced to generate a 12x2 feature combination matrix.

[0142] Then the horizontal direction signal 3+4i of the deception jamming signal is received through the horizontal polarization channel, and the vertical direction signal 1+2i is received through the vertical polarization channel at the same time, the time synchronization algorithm is used to find that the vertical direction signal is delayed by 0.001 seconds, and the phase alignment is realized after adjustment; the complex multiplication-5+10i and addition 4+6i are performed on the aligned signals, and the modulus value of the addition result is calculated The signal data body not affected by polarization is generated in time sequence. The 12*2 feature combination matrix is integrated with the 12-value signal data body to form an input data of 12*3, which is input into the feature decoupling adversarial network. The encoder is processed by the ReLU function and the 3-dimensional feature is lifted to 60 dimensions, and a deep high-dimensional feature is output. The decoupling layer uses the attention mechanism to split the 60-dimensional feature into three feature subsets of "target correlation", "interference correlation" and "background correlation". The mutual information value (0.4, 0.9 and 0.2 respectively) is estimated by KL divergence, and the "target correlation" subset corresponding to 0.9 is selected as the target feature data (12*22 dimensions) representing the micro-motion of the missile warhead.

[0143] Finally, the 12*22-dimensional target feature data is reshaped into 12*22*1 three-dimensional data and input into the multi-scale local attention module. The 3*3, 5*5 and 7*7 convolution kernels in the module operate in parallel to generate 10*20*1, 8*18*1 and 6*16*1 preliminary features respectively. The 3*3 convolution kernel extracts high-frequency feature components representing the high-frequency jitter of the warhead, the 5*5 convolution kernel extracts medium-frequency feature components representing the local trajectory change, and the 7*7 convolution kernel extracts low-frequency feature components representing the overall flight trend. The time points of the low-frequency components are supplemented to 12 by linear interpolation, and the multi-band feature components are spliced. The power entropy distribution of the three feature maps is calculated, the high frequency is 1.2, the medium frequency is 0.9, and the low frequency is 0.7, and the corresponding weights are 0.826, 1.1 and 1.414; the multi-band feature components are weighted and summed pixel by pixel to generate a 12*20 preliminary fusion feature map; the average energy of the horizontal channel is 14 and the average energy of the vertical channel is 9, and 11.5 is taken as the target energy level. The adjustment coefficients (0.82, 1.28) are calculated and the pixel values are scaled to make the energy of the two channels consistent, and finally the anti-jamming fingerprint feature with micro-Doppler feature enhancement characteristics is output, which is used for accurate identification of missiles.

[0144] Figure 4 A specific implementation structure of a radar signal deep feature extraction system based on adversarial sample defense provided by an embodiment of the present application is shown in the figure Figure 4 The system can include:

[0145] The acquisition module 41 is configured to acquire the original echo signal of the synthetic aperture radar of the spacecraft, and convert the original echo signal into a time-frequency feature map.

[0146] The first generation module 42 is configured to extract the instantaneous frequency variation curve of the intra-pulse modulation feature and the cumulative difference spectrum of the inter-pulse repetition interval from the time-frequency feature map in parallel, to perform matrix splicing on the derivative sequence of the instantaneous frequency variation curve and the envelope feature of the cumulative difference spectrum, and to generate a feature combination matrix;

[0147] The second generation module 43 is configured to receive a deception jamming signal by using a polarized radar group, to perform complex coherent superposition operation on the horizontal direction signal and the vertical direction signal of the deception jamming signal, and to generate a signal data body not affected by polarization;

[0148] The separation module 44 is configured to input the feature combination matrix and the signal data body not affected by polarization into a feature decoupling countermeasure network, and to separate out target feature data representing the micro motion of the warhead by mutual information maximization constraint;

[0149] The extraction module 45 is configured to input the target feature data into a multi-scale local attention module, to use a convolution kernel group with multiple levels of perception range to scan the spatial structure of the target feature data in parallel, and to extract multi-band feature components excited by the motion of the projectile body;

[0150] The output module 46 is configured to calculate adaptive weight coefficients according to the power entropy distribution of the feature map output by each convolution kernel, to perform pixel-level fusion on the multi-band feature components according to the adaptive weight coefficients, and to perform polarization channel energy correction processing on the fused feature map to eliminate the energy difference between channels caused by coherent superposition, so as to output an anti-jamming fingerprint feature with micro-Doppler feature enhancement characteristics.

[0151] The radar signal deep feature extraction system based on adversarial sample defense provided in the embodiments of the present application is used to implement the radar signal deep feature extraction method based on adversarial sample defense described above, and therefore the specific embodiments in the radar signal deep feature extraction system based on adversarial sample defense can be seen from the embodiment part of the radar signal deep feature extraction method based on adversarial sample defense in the foregoing, and the specific embodiments can be referred to the description of the corresponding embodiment part, which will not be described here.

[0152] The present application also provides an electronic device, comprising: a memory for storing a computer program; a processor for executing the computer program to implement the steps of the radar signal deep feature extraction method based on adversarial sample defense described above.

[0153] The present application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the radar signal deep feature extraction method based on adversarial sample defense described above.

[0154] In an example embodiment, the computer readable storage medium described above can include, but is not limited to, a U disk, a read-only memory, a random access memory, a mobile hard disk, a magnetic disk or an optical disk, and various media that can store computer programs.

[0155] Embodiments of the present application also provide a computer program product, the computer program product comprising a computer program, the computer program being executed by a processor to implement the steps of any of the above radar signal deep feature extraction methods based on adversarial sample defense.

[0156] The skilled person can further realize that the units and algorithm steps of the examples described in connection with the embodiments disclosed herein can be realized in electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of the examples have been described in general terms in the above description. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0157] The above provides a radar signal deep feature extraction method based on adversarial sample defense, system, electronic device and storage medium. The principle and implementation of the present application are described in this paper. The above examples are used to help understand the method and its core idea. It should be pointed out that for ordinary skilled person in the art, without departing from the principle of the present application, some improvements and modifications can be made to the present application, and these improvements and modifications also fall within the protection scope of the present application.

Claims

1. A radar signal deep feature extraction method based on an adversarial sample defense, characterized in that, The method comprises the following steps: Collecting original echo signals of a space vehicle synthetic aperture radar, and converting the original echo signals into time-frequency feature maps; Parallelly extracting an instantaneous frequency variation curve of intra-pulse modulation features and a cumulative difference spectrum of inter-pulse repetition intervals from the time-frequency feature maps, matrix splicing a derivative sequence of the instantaneous frequency variation curve and an envelope feature of the cumulative difference spectrum to generate a feature combination matrix; Receiving a deception jamming signal by using a polarized radar group, performing complex coherent superposition operation on horizontal direction signals and vertical direction signals of the deception jamming signal to generate a signal data body not affected by polarization; Inputting the feature combination matrix and the signal data body not affected by polarization into a feature decoupling adversarial network, and separating out target feature data representing a warhead micro-motion by mutual information maximization constraint; Inputting the target feature data into a multi-scale local attention module, and parallelly scanning spatial structures of the target feature data by using a multi-level perception range of a convolution kernel group to extract multi-band feature components excited by a projectile motion; Calculating adaptive weight coefficients according to power entropy distributions of feature maps output by each convolution kernel, performing pixel-level fusion on the multi-band feature components according to the adaptive weight coefficients, performing polarization channel energy correction processing on the fused feature maps to eliminate energy differences between channels caused by coherent superposition, and outputting anti-jamming fingerprint features with micro-Doppler feature enhancement characteristics.

2. The method of claim 1, wherein, The method of calculating adaptive weight coefficients according to power entropy distributions of feature maps output by each convolution kernel, performing pixel-level fusion on the multi-band feature components according to the adaptive weight coefficients, performing polarization channel energy correction processing on the fused feature maps to eliminate energy differences between channels caused by coherent superposition, and outputting anti-jamming fingerprint features with micro-Doppler feature enhancement characteristics, comprises the following steps: Calculating power entropy distributions of feature maps output by each convolution kernel, wherein the power entropy distributions are obtained by statistically calculating energy distribution discrete degrees of each pixel point in each feature map; Calculating adaptive weight coefficients corresponding to each feature map according to the power entropy distributions; Performing weighted summation calculation on each pixel position of each feature map in the multi-band feature components according to the adaptive weight coefficients to generate a preliminary fused feature map; Performing polarization channel energy correction processing on the preliminary fused feature map, eliminating energy differences between channels caused by coherent superposition by adjusting energy distributions of each polarization channel, keeping energy levels between different polarization channels consistent, and outputting the energy-corrected preliminary fused feature map as anti-jamming fingerprint features with micro-Doppler feature enhancement characteristics.

3. The method of claim 2, wherein, The method of performing polarization channel energy correction processing on the preliminary fused feature map, eliminating energy differences between channels caused by coherent superposition by adjusting energy distributions of each polarization channel, keeping energy levels between different polarization channels consistent, and outputting the energy-corrected preliminary fused feature map as anti-jamming fingerprint features with micro-Doppler feature enhancement characteristics, comprises the following steps: Calculating energy means of each polarization channel in the preliminary fused feature map as current energy levels of each polarization channel, and determining a unified target energy level based on the current energy levels of all polarization channels; calculate an energy adjustment coefficient corresponding to each polarization channel according to a ratio of a current energy level of the polarization channel to the target energy level; scale pixel values in each polarization channel based on the energy adjustment coefficient, eliminate channel-to-channel energy differences caused by coherent superposition, and make energy levels of each polarization channel consistent; output the anti-interference fingerprint feature with the micro-Doppler feature enhancement characteristic.

4. The method of claim 1, wherein, The feature combination matrix and the signal data body not affected by polarization are input into the feature decoupling generative adversarial network, and target feature data representing the small movement of the warhead is separated out by maximizing mutual information constraint, including: The feature combination matrix and the signal data body not affected by polarization are input into the feature decoupling generative adversarial network as input data; The feature combination matrix and the signal data body not affected by polarization are subjected to nonlinear transformation and dimension lifting processing by the encoder of the feature decoupling generative adversarial network, and deep high-dimensional feature representation is extracted; Feature separation operation is performed on the high-dimensional feature representation in the decoupling layer of the feature decoupling generative adversarial network, and a plurality of feature subsets are obtained; The mutual information value between each feature subset and the input data is calculated by maximizing mutual information constraint, and the feature subset with the maximum mutual information value is selected as the target feature data representing the small movement of the warhead.

5. The method of claim 1, wherein, The target feature data is input into the multi-scale local attention module, a plurality of convolution kernel groups with different perception ranges are used to scan the spatial structure of the target feature data in parallel, and multi-band feature components excited by the movement of the projectile body are extracted, including: The target feature data is input into the multi-scale local attention module; In the multi-scale local attention module, convolution kernel groups with different perception ranges are used to perform parallel convolution operation on the target feature data, and the convolution kernel groups include convolution kernels with three different sizes; The fine spatial structure of the target feature data is scanned by the smallest size convolution kernel, and high-frequency feature components representing high-frequency movement details are extracted; The local spatial structure of the target feature data is scanned by the medium size convolution kernel, and the medium-frequency feature components representing medium-frequency movement changes are extracted; The global spatial structure of the target feature data is scanned by the largest size convolution kernel, and the low-frequency feature components representing low-frequency movement trends are extracted; The high-frequency feature components, medium-frequency feature components and low-frequency feature components are combined to form multi-band feature components excited by the movement of the projectile body.

6. The method of claim 1, wherein, The horizontal polarization channel of the polarization radar group is used to receive the horizontal direction signal of the deception jamming signal, and the vertical polarization channel is used to receive the vertical direction signal of the deception jamming signal at the same time; The horizontal direction signal and the vertical direction signal are subjected to phase alignment processing respectively, and the two signals are completely synchronized in time; ​ The phase-aligned horizontal direction signal and the vertical direction signal are subjected to complex multiplication operation to obtain a product result representing coherence characteristics of the two signals, and the product result is subjected to complex addition operation to obtain a complex composite signal; The modulus of the complex composite signal is calculated to obtain amplitude information representing signal intensity, and a signal data volume not affected by polarization is generated based on the amplitude information.

7. The method of claim 1, wherein, The first extraction process and the second extraction process are performed in parallel from the time-frequency feature map, the first extraction process identifies the frequency points with the strongest energy in each pulse period in the time-frequency feature map, and connects these frequency points in time sequence to form an instantaneous frequency variation curve; The second extraction process detects the time positions of the pulse energy peak values in the time-frequency feature map and calculates the differences of the time intervals of the continuous pulse peak values, and performs cumulative summation on the differences to generate a cumulative difference spectrum; The rate of change of each point on the instantaneous frequency variation curve is calculated to generate a derivative sequence, and the maximum value points in the cumulative difference spectrum are identified and connected synchronously to form an envelope feature; The derivative sequence and the envelope feature are transversely or longitudinally matrix-pasted as two one-dimensional arrays to generate a feature combination matrix. It comprises:

8. A radar signal deep feature extraction system based on an adversarial sample defense, characterized in that, The acquisition module is used for acquiring the original echo signal of the spacecraft synthetic aperture radar, and converting the original echo signal into a time-frequency feature map; The first generation module is used for extracting the instantaneous frequency variation curve of the intra-pulse modulation feature and the cumulative difference spectrum of the inter-pulse repetition interval from the time-frequency feature map in parallel, and matrix-pasting the derivative sequence of the instantaneous frequency variation curve and the envelope feature of the cumulative difference spectrum to generate a feature combination matrix; The second generation module is used for receiving a deception jamming signal by using a polarized radar group, performing complex coherent superposition operation on the horizontal direction signal and the vertical direction signal of the deception jamming signal, and generating a signal data volume not affected by polarization; The separation module is used for inputting the feature combination matrix and the signal data volume not affected by polarization into a feature decoupling countermeasure network, and separating out target feature data representing the micro-motion of the warhead by mutual information maximization constraint; The extraction module is used for inputting the target feature data into a multi-scale local attention module, and using a convolution kernel group with multiple levels of perception range to scan the spatial structure of the target feature data in parallel, and extracting multi-band feature components excited by the motion of the projectile body; The output module is used for calculating adaptive weight coefficients according to the power entropy distribution of the feature maps output by each convolution kernel, performing pixel-level fusion on the multi-band feature components according to the adaptive weight coefficients, and performing polarization channel energy correction processing on the fused feature map to eliminate the energy difference between channels caused by coherent superposition, so as to output an anti-jamming fingerprint feature with micro-Doppler feature enhancement characteristics. It comprises:

9. An electronic device, comprising: The memory is used for storing a computer program; ​ A processor is configured to implement the steps of the radar signal deep feature extraction method based on adversarial sample defense according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that, The computer program is stored in the computer readable storage medium and is configured to implement the steps of the radar signal deep feature extraction method based on adversarial sample defense according to any one of claims 1 to 7 when executed by the processor.

Images