A radar signal deep feature extraction method and system based on an adversarial sample defense, an electronic device, and a storage medium

Abstract

The application provides a radar signal deep feature extraction method and system based on an adversarial sample defense, an electronic device and a storage medium, and relates to the technical field of radar signal processing. The application first collects a spacecraft synthetic aperture radar original echo signal, converts it into a time-frequency feature map, and extracts intra-pulse and inter-pulse features in parallel to generate a combination matrix. Then, a deception jamming signal is received by a polarization radar group, and a signal data body without polarization influence is generated through complex coherent superposition. Both are input into a feature decoupling adversarial network to separate the warhead micro-motion target feature data by maximizing mutual information constraints. Finally, the multi-scale local attention module is input to extract multi-band feature components, the adaptive weight coefficients are fused according to the power entropy, the polarization channel energy is corrected, and the anti-jamming fingerprint feature with micro-Doppler feature enhancement is output. The spacecraft synthetic aperture radar signal can be processed to resist interference, and the warhead target anti-jamming fingerprint feature with micro-Doppler feature enhancement can be effectively extracted.

Description

Technical Field

[0001] This application relates to the field of radar signal processing technology, and in particular to a method, system, electronic device and storage medium for extracting deep features of radar signals based on adversarial sample defense. Background Technology

[0002] During missile penetration, spacecraft synthetic aperture radar (SAR) must possess extremely high anti-jamming capabilities. During missile flight, the enemy may release mainlobe deception jamming to mislead radar target identification and tracking, resulting in false targets in the imaging and interfering with the acquisition of crucial information such as the actual warhead position and motion status. This necessitates radar signal processing methods capable of accurately extracting true target features from complex jamming environments, especially the subtle motion characteristics of the warhead, while simultaneously suppressing deception jamming to ensure the accuracy and reliability of imaging. This provides effective support for missile penetration and enhances its battlefield survivability.

[0003] Currently, a relatively advanced solution is an anti-spoofing jamming algorithm based on multi-channel synthetic aperture radar. This method first separates the jamming signal by canceling stationary targets using a split phase center antenna; then, it estimates the spatial location of the jamming signal using track interferometry; finally, based on the difference between the jamming signal and the target signal in the two-dimensional spatial and temporal distribution, it jointly employs a spatial filtering method based on zero-point constraint least squares error pattern synthesis and Doppler filtering to suppress the spoofing jamming.

[0004] This anti-spoofing jamming algorithm based on multi-channel synthetic aperture radar has certain limitations. Firstly, it has high hardware requirements, necessitating the coordinated operation of multiple channels, which increases system complexity and cost. Furthermore, it demands stringent synchronization and calibration accuracy between channels; any deviation significantly reduces its anti-jamming effectiveness. Secondly, in complex and variable jamming environments, especially when multiple jamming sources exist and the jamming patterns are complex, the algorithm's ability to separate and suppress jamming signals is unstable. For example, when the jamming signal and the target signal have similar two-dimensional spatial-temporal distribution characteristics, it is difficult to accurately distinguish between them, leading to ineffective jamming suppression. This affects the radar's extraction of true target features and reduces the reliability of radar signal processing during missile penetration. Summary of the Invention

[0005] The purpose of this application is to provide a method, system, electronic device and storage medium for extracting deep features of radar signals based on adversarial sample defense, so as to solve the problems of high hardware requirements, strict channel synchronization calibration accuracy and unstable interference separation and suppression effect and difficulty in distinguishing similar feature signals in the existing anti-spoofing interference algorithm based on multi-channel synthetic aperture radar.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides a method for extracting deep features of radar signals based on adversarial example defense, comprising:

[0007] The raw echo signal from the spacecraft's synthetic aperture radar is acquired, and the raw echo signal is converted into a time-frequency feature map.

[0008] The instantaneous frequency change curve of the intra-pulse modulation feature and the cumulative difference spectrum of the inter-pulse repetition interval are extracted in parallel from the time-frequency feature map. The derivative sequence of the instantaneous frequency change curve and the envelope feature of the cumulative difference spectrum are matrix-concatenated to generate a feature combination matrix.

[0009] The deception jamming signal is received by the polarization radar group, and a complex coherent superposition operation is performed on the horizontal and vertical signals of the deception jamming signal to generate a signal data volume that is not affected by polarization.

[0010] The feature combination matrix and the signal data volume unaffected by polarization are input into the feature decoupling adversarial network, and the target feature data representing the tiny motion of the warhead are separated by mutual information maximization constraint.

[0011] The target feature data is input into a multi-scale local attention module, and the spatial structure of the target feature data is scanned in parallel using a multi-level perception range convolution kernel group to extract multi-frequency feature components excited by projectile motion.

[0012] Adaptive weighting coefficients are calculated based on the power entropy distribution of the feature maps output by each convolutional kernel. The multi-band feature components are then fused at the pixel level according to the adaptive weighting coefficients. Polarization channel energy correction processing is then performed on the fused feature maps to eliminate the energy differences between channels caused by coherent superposition, so as to output an anti-interference fingerprint feature with micro-Doppler feature enhancement characteristics.

[0013] Optionally, the step of calculating adaptive weight coefficients based on the power entropy distribution of the output feature maps of each convolutional kernel, performing pixel-level fusion of 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-interference fingerprint feature with micro-Doppler feature enhancement characteristics, includes:

[0014] Calculate the power entropy distribution of the output feature map of each convolution kernel. The power entropy distribution is obtained by statistically analyzing the dispersion of the energy distribution of each pixel in each feature map.

[0015] Calculate the adaptive weight coefficients corresponding to each feature map based on the power entropy distribution;

[0016] The weighted summation of each pixel position in the feature map of the multi-band feature component is calculated according to the adaptive weighting coefficient to generate a preliminary fused feature map.

[0017] The preliminary fused feature map is subjected to polarization channel energy correction processing. By adjusting the energy distribution of each polarization channel, the energy difference between channels caused by coherent superposition is eliminated, so that the energy levels between different polarization channels are kept consistent. The energy-corrected preliminary fused feature map is output as an anti-interference fingerprint feature with micro-Doppler feature enhancement characteristics.

[0018] Optionally, the preliminary fused feature map is subjected to polarization channel energy correction processing. By adjusting the energy distribution of each polarization channel, the energy difference between channels caused by coherent superposition is eliminated, so that the energy levels between different polarization channels are kept consistent. The energy-corrected preliminary fused feature map is then output as an anti-interference fingerprint feature with micro-Doppler feature enhancement characteristics, including:

[0019] The average energy 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 based on 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 energy difference between channels caused by coherent superposition, so that the energy level of each polarization channel is consistent.

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

[0023] Optionally, the step of inputting the feature combination matrix and the polarization-independent signal data volume into the feature decoupling adversarial network, and separating the target feature data representing the small motions of the warhead through mutual information maximization constraints, includes:

[0024] The feature combination matrix and the polarization-independent signal data volume are input into the feature decoupling adversarial network as input data;

[0025] The encoder of the feature decoupling adversarial network performs nonlinear transformation and dimensionality enhancement on the feature combination matrix and the polarization-independent signal data volume to extract deep, high-dimensional feature representations.

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

[0027] The mutual information value between each feature subset and the input data is calculated by maximizing mutual information constraints, and the feature subset with the largest mutual information value is selected as the target feature data representing the tiny motion of the warhead.

[0028] Optionally, the step of inputting the target feature data into a multi-scale local attention module, and using a group of convolutional kernels with multi-level perception ranges to scan the spatial structure of the target feature data in parallel, and extracting multi-frequency feature components excited by projectile motion, includes:

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

[0030] In the multi-scale local attention module, a set of convolutional kernels with different receptive ranges is used to perform parallel convolution operations on the target feature data. The set of convolutional kernels contains three different sizes of convolutional kernels.

[0031] The fine spatial structure of the target feature data is scanned by the smallest size convolution kernel to extract high-frequency feature components that characterize high-frequency motion details;

[0032] By scanning the local spatial structure of the target feature data using a medium-sized convolution kernel, mid-frequency feature components characterizing mid-frequency motion changes are extracted.

[0033] The global spatial structure of the target feature data is scanned by the largest size convolution kernel to extract low-frequency feature components that characterize low-frequency motion trends;

[0034] The high-frequency, mid-frequency, and low-frequency characteristic components are combined to form multi-band characteristic components for projectile motion excitation.

[0035] Optionally, the step of receiving deception jamming signals using a polarized radar array and performing complex coherent superposition operations on the horizontal and vertical signals of the deception jamming signals to generate a signal data volume unaffected by polarization includes:

[0036] The horizontal polarization channel of the polarization radar array is used to receive the horizontal direction signal of the deception jamming signal, and the vertical direction signal of the deception jamming signal is received through the vertical polarization channel.

[0037] The horizontal and vertical signals are phase aligned to ensure that the two signals are completely synchronized in time.

[0038] The phase-aligned horizontal and vertical signals are multiplied by a complex number to obtain a product that represents the coherence of the two signals. Simultaneously, a complex addition operation is performed on the product to obtain a composite complex signal.

[0039] The magnitude of the composite complex signal is calculated to obtain amplitude information characterizing the signal strength, and a signal data volume unaffected by polarization is generated based on the amplitude information.

[0040] Optionally, the step of extracting the instantaneous frequency change curve of the intra-pulse modulation feature and the cumulative difference spectrum of the inter-pulse repetition interval in parallel from the time-frequency feature map, and matrix-concatenating the derivative sequence of the instantaneous frequency change curve with the envelope feature of the cumulative difference spectrum to generate a feature combination matrix includes:

[0041] The first extraction process and the second extraction process are executed in parallel from the time-frequency feature map. The first extraction process identifies the frequency point with the strongest energy in each pulse period in the time-frequency feature map and connects these frequency points in chronological order to form an instantaneous frequency change curve.

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

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

[0044] The derivative sequence and the envelope feature are treated as two one-dimensional arrays and concatenated horizontally or vertically to generate a feature combination matrix.

[0045] Secondly, this application provides a radar signal depth feature extraction system based on adversarial example defense, comprising:

[0046] The acquisition module is used to acquire the raw echo signal of the spacecraft's synthetic aperture radar and convert the raw echo signal into a time-frequency feature map.

[0047] The first generation module is used to extract the instantaneous frequency change curve of the intra-pulse modulation feature and the cumulative difference spectrum of the inter-pulse repetition interval in parallel from the time-frequency feature map, and to perform matrix concatenation 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.

[0048] The second generation module is used to receive deception jamming signals using the polarization radar group, and perform complex coherent superposition operations on the horizontal and vertical signals of the deception jamming signals to generate a signal data volume that is not affected by polarization.

[0049] The separation module is used to input the feature combination matrix and the signal data volume that is not affected by polarization into the feature decoupling adversarial network, and separate the target feature data representing the small motion of the warhead through mutual information maximization constraints.

[0050] The extraction module is used to input the target feature data into the multi-scale local attention module, and use a group of convolutional kernels with multi-level perception range to scan the spatial structure of the target feature data in parallel to extract the multi-frequency feature components excited by the projectile motion.

[0051] The output module is used to calculate adaptive weight coefficients based on the power entropy distribution of the feature maps output by each convolution kernel, perform pixel-level fusion of the multi-band feature components according to the adaptive weight coefficients, and 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-interference fingerprint feature with micro-Doppler feature enhancement characteristics.

[0052] Thirdly, this application provides an electronic device, comprising:

[0053] Memory, used to store computer programs;

[0054] A processor, configured to execute the computer program to implement the steps of the radar signal depth feature extraction method based on adversarial sample defense as described in the first aspect above.

[0055] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the radar signal depth feature extraction method based on adversarial sample defense as described in the first aspect above.

[0056] The radar signal depth feature extraction method based on adversarial example defense provided in this application acquires the raw echo signal of a spacecraft's synthetic aperture radar and converts it into a time-frequency feature map, laying a precise data foundation for subsequent feature extraction. It achieves efficient integration of multi-dimensional features of the target signal by extracting key intra-pulse and inter-pulse features in parallel and generating a feature combination matrix through matrix concatenation. It utilizes a polarimetric radar array to receive deception jamming signals and generates a signal data volume unaffected by polarization through complex coherent superposition, effectively avoiding interference from polarization differences in signal processing. The feature combination matrix and the polarization-free signal data volume are input into a feature decoupling adversarial network, and the representative missile is accurately separated by maximizing mutual information constraints. The target feature data of minute head movements solves the problem of difficult target feature extraction under complex interference. Multi-scale local attention modules are used to extract multi-band feature components excited by the projectile motion, which broadens the feature perception range and ensures that no multi-dimensional target features are missed. Adaptive weight coefficients are calculated based on power entropy distribution for pixel-level fusion, and polarization channel energy correction is used to eliminate energy differences between channels. Finally, an anti-interference fingerprint feature with micro-Doppler feature enhancement is output, which not only achieves effective defense against main lobe deception interference, but also improves the recognition accuracy and reliability of real warhead targets, meeting the anti-interference and accurate target feature extraction requirements of spacecraft synthetic aperture radar during missile penetration.

[0057] Furthermore, the power entropy distribution of the output feature maps of each convolutional kernel is calculated, and then the adaptive weight coefficients corresponding to each feature map are determined based on this power entropy distribution. Subsequently, the weighted summation of each pixel position of each feature map in the multi-band feature components is performed according to these coefficients to generate a preliminary fused feature map. Finally, polarization channel energy correction processing is performed on the preliminary fused feature map. By adjusting the energy distribution of each polarization channel, the energy difference between channels caused by coherent superposition is eliminated, and the energy levels of different polarization channels are kept consistent. The energy-corrected preliminary fused feature map is then output as an anti-interference fingerprint feature with micro-Doppler feature enhancement characteristics. This method, through the calculation of adaptive weight coefficients based on power entropy distribution, allows for precise weight allocation of multi-band feature components according to the energy dispersion of the feature maps 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. The final output anti-interference fingerprint feature not only has a more stable micro-Doppler feature enhancement effect, but also reduces the interference of energy differences on target recognition, further improving the reliability and recognizability of the fingerprint feature. Attached Figure Description

[0058] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 A flowchart illustrating a radar signal depth feature extraction method based on adversarial sample defense provided in this application embodiment;

[0060] Figure 2 A schematic diagram illustrating a specific implementation of a radar signal depth feature extraction method based on adversarial sample defense provided in this application embodiment;

[0061] Figure 3 A schematic diagram illustrating a specific implementation of a radar signal depth feature extraction method based on adversarial sample defense provided in this application embodiment;

[0062] Figure 4 This is a schematic diagram of a radar signal depth feature extraction system based on adversarial sample defense, provided in an embodiment of this application. Detailed Implementation

[0063] In the scenario of main lobe deception jamming defense of spacecraft synthetic aperture radar during missile penetration, existing anti-deception jamming algorithms based on multi-channel synthetic aperture radar have obvious technical limitations: On the one hand, the algorithm relies on the collaborative work of multiple channels, which has high hardware configuration requirements, increasing system costs and imposing stringent requirements on the synchronization and calibration accuracy between channels. Once the channel synchronization is deviated or the calibration accuracy is insufficient, the anti-jamming effect will be directly weakened. On the other hand, in the face of complex and ever-changing jamming environments, especially when there are multiple jamming sources and the two-dimensional spatial and temporal distribution characteristics of the jamming signal and the target signal are similar, the algorithm's separation and suppression effect on the jamming signal is unstable, making it difficult to accurately distinguish between the two. This leads to the obstruction of the extraction of real warhead target features and fails to meet the radar's reliability requirements for target identification and tracking. Therefore, a better anti-jamming signal processing solution is urgently needed.

[0064] To address the aforementioned issues, this application proposes a radar signal depth feature extraction method based on adversarial example defense. The scheme first acquires the raw radar echo signal and converts it into a time-frequency feature map, extracting key features to generate a feature combination matrix. Simultaneously, it generates a signal data volume unaffected by polarization through the superposition of polarized radar arrays and complex coherent signals. Then, using a feature decoupling adversarial network and mutual information maximization constraints, it accurately separates target feature data representing the warhead's minute movements. Finally, a multi-scale local attention module extracts multi-band feature components, and through adaptive weight fusion and polarization channel energy correction, it outputs an anti-jamming fingerprint feature with micro-Doppler feature enhancement. This scheme does not rely on complex multi-channel hardware, avoiding the challenge of multi-channel synchronous calibration. Through mutual information maximization constraints and multi-scale feature processing, it can accurately separate targets and jamming signals even in complex jamming environments. Polarization channel energy correction solves the energy imbalance problem caused by coherent superposition, overcoming the shortcomings of existing algorithms such as high hardware requirements and unstable jamming separation, significantly improving the radar's recognition accuracy and anti-jamming reliability for real warhead targets.

[0065] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0066] The core of this application is to provide a method for extracting deep features of radar signals based on adversarial sample defense. A flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes:

[0067] S101. Acquire the raw echo signal of the spacecraft's synthetic aperture radar and convert the raw echo signal into a time-frequency characteristic map;

[0068] In the above scheme, the raw echo signal of the spacecraft's synthetic aperture radar refers to the initial signal reflected back and received by the radar after the synthetic aperture radar on board the spacecraft emits electromagnetic waves towards the target. A time-frequency characteristic map is a graphical data representation that simultaneously shows the characteristics of a signal in both time and frequency dimensions. It displays the frequency components of the signal at different time points and the corresponding energy levels of each frequency component.

[0069] In this embodiment, the synthetic aperture radar on the spacecraft first actively transmits electromagnetic waves of a specific frequency to the target. These electromagnetic waves are reflected when they encounter the target during propagation. The signal receiving device equipped with the radar captures these reflected electromagnetic waves. Then, the signal conversion module converts the captured electromagnetic wave physical signal into a processable electrical signal. This electrical signal is the original echo signal containing information such as the target's position and motion state. Moreover, this signal exists in the form of a time-series signal, that is, it continuously records data such as signal strength over time.

[0070] Next, the original echo signal is preprocessed to ensure signal stability. Then, a spectrum analysis technique such as Short-Time Fourier Transform (SFT) is used. First, a suitable time window parameter is set, dividing the continuous time series signal into multiple overlapping short time segments. Then, a SFT is performed on the signal within each time window. This algorithm calculates all frequency components contained in the signal within each time window, as well as the energy intensity corresponding to each frequency component, obtaining the spectrum data for each time window. Finally, the spectrum data of all time windows are arranged in chronological order to form a two-dimensional matrix. The columns of the matrix correspond to the time axis, representing the chronological order of different time windows, and the rows correspond to the frequency axis, representing the different frequency values ​​contained in the signal. The value of each cell in the matrix represents the energy intensity of the signal at the corresponding time window and frequency. This two-dimensional matrix is ​​a time-frequency characteristic map that simultaneously reflects the three dimensions of signal characteristics: time, frequency, and energy.

[0071] In practical applications, such as when a spacecraft's synthetic aperture radar emits electromagnetic waves to a missile, the receiving device captures the reflected waves and converts them into a time-series electrical signal (raw echo signal) that is recorded every 0.001 seconds from 0 to 10 seconds. Then, a time window of 0.5 seconds and an overlap ratio of 50% are set, and the 0-10 second signal is divided into multiple overlapping windows. For each window, a short-time Fourier transform is used to calculate the spectrum containing frequencies of 50-1000 Hz and corresponding energies. These spectra are then arranged in chronological order to form a two-dimensional matrix with columns corresponding to the 0-10 second time period, rows corresponding to the 50-1000 Hz frequency period, and cell values ​​representing energy. This completes the conversion of the time-frequency feature map, which is used for subsequent extraction of missile-related features.

[0072] The overall scheme of S101 described above uses the acquisition of raw echo signals as the foundation for target feature analysis. It provides initial data, preventing the inability to conduct subsequent analysis due to a lack of signal. After converting the raw echo signals into time-frequency feature maps, the changes in the signal in the time and frequency dimensions can be intuitively presented, more clearly reflecting the frequency variation pattern of the signal during target movement, and providing a more comprehensive basis for the extraction and identification of key target features.

[0073] S102. Extract the instantaneous frequency change curve of the intra-pulse modulation feature and the cumulative difference spectrum of the inter-pulse repetition interval in parallel from the time-frequency feature map. Then, perform matrix concatenation 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 may specifically include the following steps:

[0075] S1021. The first extraction process and the second extraction process are executed in parallel from the time-frequency feature map. The first extraction process identifies the frequency point with the strongest energy in each pulse period in the time-frequency feature map and connects these frequency points in time order to form an instantaneous frequency change curve.

[0076] S1022, The second extraction process detects the time position of the pulse energy peak in the time-frequency feature map, calculates the difference in the time interval of consecutive pulse peaks, and accumulates and sums these differences to generate a cumulative difference spectrum;

[0077] S1023. Calculate the rate of change of each point on the instantaneous frequency change curve to generate a derivative sequence, and simultaneously identify and connect the maximum points in the cumulative difference spectrum to form an envelope feature;

[0078] S1024. The derivative sequence and the envelope feature are treated as two one-dimensional arrays and are concatenated horizontally or vertically to generate a feature combination matrix.

[0079] In the above scheme, intra-pulse modulation characteristics refer to the modulation properties within each pulse of the radar signal, reflecting the signal variation pattern within the pulse. The instantaneous frequency change curve is formed by connecting the strongest frequency points within each pulse period in the time-frequency characteristic graph in chronological order, showing the frequency change over 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 chronological order, reflecting the speed of frequency change. The envelope feature of the cumulative difference spectrum is the contour formed by connecting the maxima in the cumulative difference spectrum, highlighting its overall trend. The feature combination matrix is ​​a matrix formed by concatenating the derivative sequence and the envelope feature, used to integrate feature information from two dimensions.

[0080] In this embodiment, step S1021 first determines the time range of each pulse period in the time-frequency feature map. This requires combining the known pulse repetition period parameters of the radar signal to divide the time axis of the time-frequency feature map into independent time intervals, each interval corresponding to one pulse period. Then, an energy detection algorithm is used to process the frequency axis data within each divided pulse period, comparing the energy intensity values ​​corresponding to different frequency points within that period, and finding the frequency point with the largest energy value. This point is the frequency point that best represents the signal characteristics within that pulse period. Finally, according to the chronological order of the pulse periods on the time axis, the strongest energy frequency points found in each period are connected sequentially with line segments to form a curve that reflects the frequency change over time, serving as the instantaneous frequency change curve of the intra-pulse modulation characteristics. For example, if 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, 10 pulse periods can be divided. Comparing the energy of all frequency points within the 0-0.2 second period, it is found that 180Hz has the largest energy. This process is repeated to find 10 frequency points, and connecting them in chronological order yields the instantaneous frequency change curve.

[0081] Secondly, in step S1022, pulse peak detection is performed in the time-frequency feature map. Using a peak detection algorithm, a suitable sliding window size is set, and the window moves point by point on the time axis of the time-frequency feature map. By comparing the energy values ​​at different time points within the window, the time position with the highest energy of each pulse signal is found. Then, based on these peak time positions, the time difference between two consecutive adjacent pulse peaks is calculated. This time difference is the inter-pulse repetition interval. For example, if the first peak is at 0.1 seconds and the second is at 0.3 seconds, the corresponding inter-pulse repetition interval is 0.2 seconds. Then, the difference between consecutive inter-pulse repetition intervals is calculated. For example, if the previous inter-pulse repetition interval is 0.2 seconds and the next is 0.22 seconds, the difference between the two is 0.02 seconds. Finally, these calculated differences are accumulated and summed sequentially in time order. Each new difference is added to the previous accumulated sum, and all accumulated results are arranged in order to form a data sequence, which is used as the cumulative difference spectrum of inter-pulse repetition intervals.

[0082] Next, in step S1023, the time coordinates and corresponding frequency values ​​of each data point on the instantaneous frequency change curve are first determined. Then, a numerical differentiation algorithm is used to calculate the frequency change and time change between each point on the curve and the next adjacent point. The frequency change is divided by the time change to obtain the frequency change rate at that point. All derivative values ​​are arranged in the time order of the points on the curve to form the derivative sequence of the instantaneous frequency change curve. For example, if 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 calculated to be 100 Hz / second using the forward difference method, and the derivative of the second point is -50 Hz / second. The derivative sequence is obtained by arranging these two derivatives. The cumulative difference spectrum data sequence is simultaneously subjected to maximum detection. Using a maximum detection algorithm, each data point in the sequence is compared with the value of its two adjacent data points to find the data point that is larger than both of the adjacent data points. Then, according to the time order of the cumulative difference spectrum, these maximum points are connected sequentially with line segments. This contour line is the envelope feature of the cumulative difference spectrum. For example, if the cumulative difference spectrum data is 0.02, 0.04, 0.03, 0.05, 0.04, and the maximum points are 0.04 and 0.05, connecting these two points yields the envelope feature.

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

[0084] In practical applications, when spacecraft-borne synthetic aperture radar monitors missile penetration, it acquires the missile's time-frequency characteristic map and, combined with the radar's 0.3-second pulse repetition period parameter, divides the time axis of the time-frequency characteristic map into 10 pulse periods (0-0.3 seconds, 0.3-0.6 seconds...2.7-3.0 seconds). An energy detection algorithm is used to determine the frequency point with the strongest energy within each period, and these are connected in chronological order to form an instantaneous frequency change curve. A sliding window peak detection method is then employed, moving the window along the time axis of the time-frequency characteristic map to compare the energy levels and identify the peak frequency. The peak time of 0 pulses is used to calculate the inter-pulse repetition interval and the difference between adjacent intervals, and the summation is performed to form a cumulative difference spectrum. The time and frequency values ​​of 10 points on the instantaneous frequency change curve are determined, and the frequency change rate is calculated using the forward difference method to obtain the derivative sequence. At the same time, maximum value detection is performed on the cumulative difference spectrum to form the envelope feature. Finally, it is confirmed that the derivative sequence and the envelope feature are both one-dimensional arrays with the same length. Through horizontal splicing, the signal processing module combines the two into a 10-row, 2-column feature combination matrix to prepare for the subsequent separation of the missile's small moving target features.

[0085] The overall scheme of S102 described above extracts the instantaneous frequency change curve and the cumulative difference spectrum in parallel, which can simultaneously obtain the features inside and outside the pulse and avoid the limitations of single features; the calculation of the derivative sequence can reflect the rate of frequency change and capture the micro-motion information of the target; the generation of envelope features highlights the trend of the cumulative difference spectrum and reduces local interference; matrix splicing integrates the two types of features, providing more comprehensive information for subsequent processing, helping to accurately identify target features, and laying a solid foundation for defending against deception interference and separating the true target features.

[0086] S103. Receive the deception jamming signal using the polarization radar group, and perform complex coherent superposition operation on the horizontal and vertical signals of the deception jamming signal to generate a signal data volume that is not affected by polarization.

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

[0088] S1031. Receive the horizontal direction signal of the deception jamming signal using the horizontal polarization channel of the polarization radar group, and simultaneously receive the vertical direction signal of the deception jamming signal through the vertical polarization channel.

[0089] S1032. Perform phase alignment processing on the horizontal and vertical signals respectively to ensure that the two signals are completely synchronized in time.

[0090] S1033. Perform complex multiplication on the phase-aligned horizontal and vertical signals to obtain a product that characterizes the coherence of the two signals. Simultaneously perform complex addition on the product to obtain a composite complex signal.

[0091] S1034. Calculate the modulus of the composite complex signal to obtain amplitude information characterizing the signal strength, and generate a signal data volume unaffected by polarization based on the amplitude information.

[0092] In the above scheme, a polarimetric radar group refers to a combination of radar equipment containing horizontal and vertical polarization channels, used to acquire multi-polarization information of deception and jamming signals. Deception and jamming signals refer to false signals released by the enemy to mislead radar identification, containing signals with two different polarization directions: horizontal and vertical. Horizontal signals refer to signals received through the horizontal polarization channel of the polarimetric radar group, with their vibration direction parallel to the ground. Vertical signals refer to signals received through the vertical polarization channel of the polarimetric radar group, with their vibration direction perpendicular to the ground. The signal data volume refers to the processed signal data set no longer affected by differences in horizontal or vertical polarization directions; it contains information that stably characterizes signal strength and can be used for subsequent fusion processing with target feature data.

[0093] In this embodiment, the polarization radar group is first activated in step S1031 to ensure that its horizontal polarization channel and vertical polarization channel are in normal working condition. These two channels are components of the radar group specifically designed to receive signals in different polarization directions. The horizontal polarization channel is responsible for capturing the components in the signal whose vibration direction is parallel to the ground, and the vertical polarization channel is responsible for capturing the components 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 enemy's deception jamming signal to ensure that both channels can receive the deception jamming signal simultaneously. Subsequently, the two channels start receiving signals synchronously. The horizontal polarization channel converts the received horizontal direction signal into a complex form, and the vertical polarization channel similarly converts the received vertical direction signal into a complex form, ultimately obtaining the original complex signals in the two polarization directions.

[0094] Secondly, the horizontal and vertical signals obtained in step S1031 are acquired through step S1032. During transmission and reception, these two signals may experience time delays due to factors such as differences in channel hardware response speed, resulting in phase mismatch at the same moment. To solve this problem, a time synchronization algorithm is used. First, timestamp information representing the time characteristics of the signals is extracted from the horizontal and vertical signals. Then, the timestamps of the horizontal and vertical signals are compared point by point to find the time deviation between them at the same signal characteristic points. Finally, based on this time deviation, the time series of the vertical signal is adjusted so that the phase of the adjusted vertical signal is completely consistent with that of the horizontal signal at each corresponding time point, ensuring that subsequent operations can be performed based on the synchronized signal.

[0095] Next, step S1033 calls the horizontal and vertical signals after phase alignment in step S1032. These two signals are now completely synchronized in time and can be used for complex number operations. First, a complex multiplication operation is performed. According to the rules of complex multiplication, the complex numbers of the two signals at each corresponding time point are multiplied together. The resulting product reflects the coherence characteristics between the two signals, i.e., the degree of correlation between the signals. After the complex multiplication operation is completed, a complex addition operation is performed on the product. According to the rules of complex addition, the product is added to another set of complex signals at the same time point to obtain a composite complex signal that integrates the overall information of the two signals.

[0096] Finally, the composite complex signal is obtained through step S1034, and the magnitude of the composite complex signal at each time point is calculated. The formula for calculating the magnitude of the complex signal is used, and the magnitude directly represents the signal strength at the corresponding time point. After calculating the magnitude of all time points, these magnitudes are arranged in chronological order to form a continuous data set containing only signal strength information. This data set is the signal data volume that is not affected by polarization, because the magnitude is only related to the signal strength and is no longer affected by phase and amplitude deviations caused by differences in horizontal or vertical polarization directions. It can be directly used for subsequent fusion processing with the feature combination matrix.

[0097] In practical applications, in scenarios where spacecraft monitor missile penetration and encounter enemy deception and jamming signals, the spacecraft's onboard polarization radar is activated. After confirming that both the horizontal and vertical polarization channels are functioning normally, the radar's receiving direction is aligned with the source of the deception and jamming signal. Both channels receive signals synchronously. The horizontal polarization channel captures the horizontal signal with its vibration direction parallel to the ground and converts it into a complex number 4+7i. Simultaneously, the vertical polarization channel captures the vertical signal with its vibration direction perpendicular to the ground and converts it into a complex number 2+4i. Then, a time delay is detected between the horizontal and vertical signals due to differences in hardware response speed. A time synchronization algorithm based on signal feature point matching is used to extract the two signals with an accuracy of 0.0001 seconds. After comparing the timestamps point by point, it was found that the vertical signal was 0.0006 seconds later than the horizontal signal. Therefore, all data points of the vertical signal were shifted forward by 0.0006 seconds to ensure that the phases of the two signals were perfectly aligned at each corresponding time point. Then, the phase-aligned horizontal signal 4+7i and the vertical signal 2+4i were retrieved. First, their product was calculated according to the complex multiplication rule to obtain the product result representing the coherence characteristics of the two signals, i.e., -20+30i. Then, this product result was added to another set of signals at the same time point and with phase alignment result 5+8i according to the complex addition rule to calculate (-20+5)+(30+8)i=-15+38i, generating a composite complex signal. Finally, its modulus was calculated according to the complex modulus calculation formula. The modulus values ​​calculated at all time points are arranged in chronological order to form a signal data volume that contains only signal strength information and is not affected by polarization direction, which is used for subsequent fusion processing with the feature combination matrix.

[0098] The overall scheme of S103 described above can obtain complete polarization information by receiving dual-polarization signals through polarization radar group, avoiding the loss of information in a single channel; phase alignment realizes signal time synchronization, eliminates time delay interference, and ensures accurate calculation; complex multiplication and addition operations integrate signal coherence characteristics and overall information to form a composite signal; and the calculation of modulus generates polarization-independent signal data volume, reducing the impact of interference and improving the accuracy of target feature separation.

[0099] S104. Input the feature combination matrix and the signal data volume that is not affected by polarization into the feature decoupling adversarial network, and separate the target feature data representing the small motion of the warhead by maximizing mutual information constraint.

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

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

[0102] S1042. The encoder of the feature decoupling adversarial network performs nonlinear transformation and dimensionality enhancement on the feature combination matrix and the signal data volume that is not affected by polarization to extract deep high-dimensional feature representations.

[0103] S1043. In the decoupling layer of the feature decoupling adversarial network, a feature separation operation is performed on the high-dimensional feature representation to obtain multiple feature subsets;

[0104] S1044. Calculate the mutual information value between each feature subset and the input data through the mutual information maximization constraint, and select the feature subset with the maximum mutual information value as the target feature data representing the small motion of the warhead.

[0105] In the above scheme, the Feature Decoupling Adversarial Network (FAN) is an artificial intelligence network used to separate mixed features. It includes modules such as an encoder and a decoupling layer, and can be used to extract target-specific features from the mixed features. Deep high-dimensional feature representation is the high-dimensional data stream output by the encoder, containing deep features, which can more accurately reflect the essential characteristics of the data. Feature subsets are single-attribute feature fragments obtained after decoupling, which may represent target features, interference features, etc. Mutual information values ​​are numerical values ​​that quantify the degree of correlation and can be used as a basis for selecting target features. Target feature data is the final selected feature subset most closely related to the subtle movements of the warhead, which can be used for subsequent accurate target identification.

[0106] In this embodiment, step S1041 prepares a feature combination matrix and a signal data volume unaffected by polarization. These two types of data contain the basic features of the target and the processed interference signal strength information, respectively, and are the core of subsequent network processing. Next, the format and dimensions of the two types of data are checked for compatibility. If their time dimensions differ, a linear interpolation algorithm is used to adjust the length of the signal data volume to match the number of time points in the feature combination matrix. After confirming compatibility, the two types of data are concatenated and integrated along the time dimension. The integrated data is then input into a pre-built and trained feature decoupling adversarial network, providing the network with a complete input data source. For example, if the feature combination matrix is ​​15 rows and 2 columns, and the signal data volume unaffected by polarization is a one-dimensional array of 15 values, after interpolation to ensure consistent length, the signal data volume is added as the third column to the feature combination matrix, forming a 15-row, 3-column integrated data set, which is then input into the feature decoupling adversarial network.

[0107] Secondly, by calling the integrated data input into the feature decoupling adversarial network in step S1042, the encoder module of the network first processes this data. The encoder uses a fully connected neural network as its core technology. The first layer performs a nonlinear transformation and processes the integrated data through the ReLU activation function, breaking the original simple linear relationship of the data and making the data present a more complex change pattern that fits the real target features. For example, the integrated data contains values ​​such as -0.3, 0.5, -1.2, and 2.1, which become 0, 0.5, 0, and 2.1 after processing by the ReLU function. After completing the nonlinear transformation, the encoder performs dimensionality enhancement processing through subsequent multi-layer neural networks. Each layer of neurons increases the feature dimension of the data. For example, the integrated data of 15 rows and 3 columns is enhanced to 20 dimensions through the first layer of neurons and then to 50 dimensions through the second layer of neurons. Finally, the output is a high-dimensional feature representation with higher dimensions and containing deep feature information. For example, after the encoder processing, the original 3-dimensional data of 15 time points becomes 50-dimensional data of 15 time points. Each dimension corresponds to a feature that can reflect the deep pattern of the target or interference.

[0108] Next, the deep high-dimensional feature representation output by the encoder is obtained through step S1043. This data is a comprehensive feature set that mixes target features, interference features, and background features. It needs 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 that represents the small motion features of the warhead learned in the training data, it is determined whether the dimension is related to the target. Dimensions with high similarity are marked as "target related" and those with low similarity are marked as "non-target related". Then, according to the marking results, the deep high-dimensional feature representation is split into multiple independent feature subsets according to attributes. For example, in the 50-dimensional deep high-dimensional features, 20 dimensions are marked as "target related", 15 dimensions are marked as "interference related" and 15 dimensions are marked as "background related". These 20-dimensional, 15-dimensional and 15-dimensional features are integrated to form 3 feature subsets. Each subset contains only features of a single attribute to avoid mutual interference between features of different attributes.

[0109] Finally, in step S1044, multiple feature subsets output from the decoupling layer and the original integrated data input to the network are collected. Target feature data representing the warhead's minute motion is selected using the 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 correlation between the feature subset and the original data, and the more likely it is to contain key target information from the original data. For example, the mutual information values ​​of three feature subsets and the original integrated data are calculated to be 0.35, 0.88, and 0.22, respectively. 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's minute motion, i.e., the target feature data representing the warhead's minute motion, and is used in subsequent multi-band feature extraction steps.

[0110] In practical applications, in spacecraft monitoring missile penetration scenarios, a 12x2 feature combination matrix containing intra- and inter-pulse features of the missile at 12 time points is generated. This matrix, along with a generated one-dimensional signal data volume composed of intensity information from 12 time points of interference signals, is unaffected by polarization. After confirming that both have 12 time points, the signal data volume is added as the third column to the feature combination matrix, forming a 12x3 integrated data input to a pre-trained feature decoupling adversarial network. The network encoder uses a fully connected neural network, applying a ReLU activation function to the integrated data non-linearly. Two layers of neurons sequentially upscale the 3D features to 25 and 60 dimensions, outputting a 60-dimensional deep high-dimensional feature representation at 12 time points. The decoupling layer uses an attention mechanism to calculate the similarity between the 60-dimensional features and the "missile warhead micro-motion target feature template," labeling the 22, 18, and 20 dimensions as "target-related," "interference-related," and "background-related," respectively, and integrating them into three independent feature subsets. Finally, three feature subsets were collected and integrated with the original data. The mutual information value was calculated using the KL divergence estimation method. The calculation results were 0.4, 0.92, and 0.25, respectively. The "target correlation" feature subset corresponding to the maximum mutual information value of 0.92 was selected as the target feature data representing the small motion of the missile warhead for subsequent multi-band feature extraction.

[0111] The overall scheme of S104 described above inputs the feature combination matrix and the signal data volume unaffected by polarization into the network, integrating the target's basic features and information after interference processing to provide data for feature separation. The encoder extracts deep features through nonlinear transformation and dimensionality enhancement, overcoming the limitations of shallow layers; the decoupling layer separates the mixed high-dimensional features into subsets, facilitating selection. Mutual information maximization constraint filters features strongly correlated with the warhead's minute movements, eliminating interference and irrelevant background, improving feature purity and accuracy, and providing a basis for subsequent extraction of multi-band and generation of anti-interference fingerprint features.

[0112] S105. Input the target feature data into the multi-scale local attention module, and use a multi-level perception range convolution kernel group to scan the spatial structure of the target feature data in parallel to extract the multi-frequency feature components excited by the projectile motion.

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

[0114] S1051. Input the target feature data into the multi-scale local attention module;

[0115] S1052. In the multi-scale local attention module, a set of convolutional kernels with different receptive ranges is used to perform parallel convolution operations on the target feature data. The set of convolutional kernels contains three different sizes of convolutional kernels.

[0116] S1053. Scan the fine spatial structure of the target feature data using the smallest size convolution kernel to extract high-frequency feature components that characterize high-frequency motion details;

[0117] S1054. Scan the local spatial structure of the target feature data using a medium-sized convolution kernel to extract mid-frequency feature components that characterize mid-frequency motion changes;

[0118] S1055. Scan the global spatial structure of the target feature data using the largest size convolution kernel to extract low-frequency feature components that characterize low-frequency motion trends;

[0119] S1056. The high-frequency characteristic components, mid-frequency characteristic components and low-frequency characteristic components are combined to form multi-band characteristic components for projectile motion excitation.

[0120] In the above scheme, the multi-scale local attention module is a processing module capable of capturing features from different ranges. It contains convolutional kernels of various sizes and can be used to comprehensively extract the spatial structure features of target data. A convolutional kernel group refers to a set of convolutional kernels of different sizes. Different sizes correspond to different feature perception ranges, and can extract fine-grained, medium-grained, and coarse-grained features respectively. High-frequency feature components are feature data characterizing the details of the warhead's high-frequency motion, reflecting rapid changes in motion. Local spatial structure is the spatial distribution information of local regions in the target feature data, corresponding to the medium-speed motion changes of the warhead. Mid-frequency feature components are feature data characterizing the mid-frequency motion changes of the warhead, reflecting regular changes in motion. Global spatial structure is the overall spatial distribution information in the target feature data, corresponding to the slow, overall motion trend of the warhead. Low-frequency feature components are feature data characterizing the low-frequency motion trend of the warhead, reflecting the overall law of motion. Multi-band feature components are a set formed by integrating high-frequency, mid-frequency, and low-frequency feature components, which can comprehensively reflect the different frequency band features of the projectile's motion.

[0121] In this embodiment, step S1051 first confirms whether the format of the target feature data meets the input requirements of the multi-scale local attention module. For example, if the module needs to receive three-dimensional data (time × feature dimension × number of channels), while the target feature data may be two-dimensional data, such as a 15×25 two-dimensional matrix with 15 time points × 25-dimensional features, then the reshape operation in the dimension adjustment algorithm is used to convert the two-dimensional data into three-dimensional data, such as 15×25×1, where the last "1" represents a single channel, to adapt to the module's input format. After confirming that the data format is compatible, 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 material to be analyzed". For example, if the target feature data is a 15×25 two-dimensional matrix, after being converted into 15×25×1 three-dimensional data by the reshape operation, it can be successfully input into the multi-scale local attention module, preparing for the subsequent extraction of multi-band features.

[0122] Secondly, in step S1052, a pre-configured multi-level perceptual range convolutional kernel group is invoked. This convolutional kernel group contains three different sizes of convolutional kernels. Then, parallel computing technology is used to allow multiple convolutional kernels to work simultaneously, avoiding the waste of time in sequential processing. These three sizes of convolutional kernels simultaneously perform convolution operations on the 3D target feature data input in step S1051: each convolutional kernel is like a "small window with filtering function," moving point by point on the data with a fixed stride. At each position, the product of all data in the window and the value at the corresponding position of the convolutional kernel is calculated, and all products are added together to obtain a new value, which is the feature value at that position. As the convolutional kernel continues to slide, a new feature matrix is ​​generated, which is the preliminary feature data of the convolutional kernel of the corresponding size.

[0123] Next, feature components for the corresponding frequency bands are extracted through steps S1053 to S1055: For the preliminary feature data generated by the smallest 3×3 convolution kernel, because the 3×3 convolution kernel covers a small area each time, it can accurately capture subtle changes in the data. Therefore, it is used to scan the fine spatial structure of the target feature data, such as the small positional shift of the warhead within 0.008 seconds, high-frequency jitter, and other details. Then, through feature selection rules, the parts with high frequency of numerical changes in the preliminary features are retained. High-frequency feature components that characterize high-frequency motion details are extracted from the preliminary features of the 3×3 convolution kernel. For example, feature values ​​reflecting 120 vibrations per second of the warhead are selected from the preliminary features to form high-frequency feature components. For the preliminary feature data generated by the medium-sized 5×5 convolution kernel, the 5×5 convolution kernel covers a moderate area and can capture the overall changes in local areas of the data. Therefore, it is used to scan the local spatial structure of the target feature data, such as the warhead's positional shift within 0.008 seconds, high-frequency jitter, and other details. Within 0.06 seconds, the local flight trajectory and moderate speed changes are analyzed. Then, using feature selection rules, the portion of the initial features with moderate frequency of numerical changes is retained. Mid-frequency feature components representing mid-frequency motion changes are extracted from the initial features of the 5×5 convolution kernel. For example, feature values ​​reflecting the warhead's speed change of 45 meters per second are selected to form mid-frequency feature components. For the initial feature data generated by the largest size 7×7 convolution kernel, the 7×7 kernel covers a large area and can capture the overall trend changes in the data. Therefore, it is used to scan the global spatial structure of the target feature data, such as the warhead's overall flight direction and slow attitude adjustment trend within 0.6 seconds. Then, using feature selection rules, the portion of the initial features with low frequency of numerical changes is retained. Low-frequency feature components representing low-frequency motion trends are extracted from the initial features of the 7×7 convolution kernel. For example, feature values ​​reflecting the warhead's continuous flight in a northeast direction are selected to form low-frequency feature components.

[0124] Finally, in step S1056, high-frequency feature components, mid-frequency feature components, and low-frequency feature components are collected. At this point, all 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, each of the high-frequency, mid-frequency, and low-frequency feature components corresponds to 15 time points. If a component is missing a time point due to convolution operations, linear interpolation is used to supplement the feature values ​​of the missing time points to ensure that the time dimensions of the three components are consistent. After time alignment, feature splicing technology is used to combine the three components in the direction of "feature dimension" rather than the time direction, integrating the high-frequency, mid-frequency, and low-frequency feature values ​​corresponding to each time point. For example, at the time point of 0.1 seconds, the high-frequency feature value is 0.8, the mid-frequency is 0.5, and the low-frequency is 0.3. These three values ​​are combined into the comprehensive feature of that time point. After all time points are combined, the resulting set containing the multi-band comprehensive features of each time point is the multi-band feature component excited by the projectile motion, such as a set of 15 time points, each containing 3 types of frequency band features, which can be used for subsequent feature fusion processing.

[0125] In practical applications, during spacecraft monitoring of missile penetration scenarios, when performing related operations, target feature data representing the minute movements of the missile warhead, filtered in step S104, is acquired, resulting in a 12×20 two-dimensional matrix of 12 time points × 20-dimensional features. Since the multi-scale local attention module needs to receive three-dimensional data, it is converted to 12×20×1 three-dimensional data through a reshape dimension adjustment operation. After confirming format compatibility, the data is input into the multi-scale local attention module. The module calls convolutional kernel groups containing three sizes: 3×3, 5×5, and 7×7. Parallel computing technology is used to allow the three convolutional kernels to simultaneously perform convolution operations on the 12×20×1 three-dimensional data. The 3×3 convolutional kernel slides with a stride of 1, covering a 3×3 region each time to calculate the product sum, generating a preliminary feature of 10×18×1; the 5×5 convolutional kernel generates a preliminary feature of 8×16×1; and the 7×7 convolutional kernel generates a preliminary feature of 6×14×1. The 3×3 convolution kernel, due to its small coverage area, accurately captures minute jitter details of the missile warhead within 0.006 seconds. Following the rule of "retaining features with high numerical change frequency," it extracts high-frequency feature components reflecting the warhead's vibration of 110 times per second. The 5×5 convolution kernel, with its moderate coverage area, captures local flight trajectory changes of the warhead within 0.05 seconds. Following the rule of "retaining features with medium numerical change frequency," it extracts mid-frequency feature components reflecting the warhead's movement of 40 meters per second. The 7×7 convolution kernel, with its large coverage area, captures the warhead's overall flight direction within 0.5 seconds. Following the rule of "retaining features with low numerical change frequency," it extracts low-frequency feature components reflecting the warhead's northwestward flight. Collecting feature components from the three frequency bands revealed that the low-frequency feature components generated by the 7×7 convolution kernel only have 6 time points. Linear interpolation was used to supplement the missing 6 time point feature values, ensuring that all three correspond to 12 time points, and then they were concatenated according to the "feature dimension" direction. For example, at time point 0.2, the high-frequency feature value is 0.7, the mid-frequency value is 0.4, and the low-frequency value is 0.2. These three values ​​are combined to form the comprehensive feature at that time point. After integrating all time points, a multi-band feature component is formed, which contains 12 time points and each time point contains three types of frequency band features. This component is then used for subsequent fusion processing.

[0126] The overall solution of S105 described above ensures smooth input of features into the multi-scale module by reshaping the data format. It utilizes parallel computation with small, medium, and large convolutional kernels to capture high-frequency details, mid-frequency variations, and low-frequency trends, respectively, thus fully preserving the projectile information. Time alignment processing eliminates dimensionality differences, generating multi-band features to provide data for subsequent processing and improve the accuracy and reliability of anti-interference fingerprint features.

[0127] S106. Calculate adaptive weight coefficients based on the power entropy distribution of the feature maps output by each convolution kernel, perform pixel-level fusion of 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 energy difference between channels caused by coherent superposition, so as to output an anti-interference fingerprint feature with micro-Doppler feature enhancement characteristics.

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

[0129] S1061. Calculate the power entropy distribution of the output feature map of each convolution kernel. The power entropy distribution is obtained by statistically analyzing the dispersion of the energy distribution of each pixel in each feature map.

[0130] S1062. Calculate the adaptive weight coefficients corresponding to each feature map based on the power entropy distribution;

[0131] S1063. Perform weighted summation calculation on each pixel position of each feature map in the multi-band feature component according to the adaptive weighting coefficient to generate a preliminary fused feature map;

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

[0133] Specifically, step S1064 includes the following processes: calculating the average energy 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; scaling the pixel values ​​in each polarization channel based on the energy adjustment coefficient to eliminate the energy difference between channels caused by coherent superposition, so that the energy levels of each polarization channel are 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 a data distribution obtained by statistically analyzing the dispersion of energy distribution at each pixel in each feature map. The higher the dispersion, the larger the power entropy value, which can be used to determine the effective information content of the feature map. The adaptive weight coefficient is a value calculated based on the power entropy distribution and used to allocate the fusion weights of each feature map. The higher the power entropy value, the smaller the weight, ensuring a higher proportion of effective features. The preliminary fused feature map is a feature map generated after pixel-level fusion without energy correction, containing multi-band integrated features. The energy difference between channels is the inconsistent energy level phenomenon caused by coherent superposition of different polarization channels, which affects the stability of the features. The anti-interference fingerprint feature is the final output feature data that can highlight the Doppler features of the warhead's minute movements and has strong anti-interference capabilities, which can be used for accurate target identification.

[0135] In the embodiments of this application, such as Figure 2 As shown, firstly, feature maps output by three different sizes of convolutional kernels in step S105 are obtained through step S1061. Then, the power entropy value of each feature map is calculated sequentially using the entropy calculation method, thus forming a power entropy distribution: First, the energy values ​​of all pixels in a single feature map are statistically analyzed to determine the energy level of each pixel; second, the total energy of the feature map is calculated by summing the energy values ​​of all pixels in the map; third, the proportion p of the energy of each pixel in the total energy of the feature map is calculated. i This reflects the proportion of energy of a single pixel in the overall feature map; the fourth step substitutes the energy proportion of each pixel into the power entropy formula. 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; following the same four-step process as above, the power entropy values ​​of the mid-frequency feature map and the low-frequency feature map are calculated respectively, and finally the power entropy values ​​of the high-frequency, mid-frequency and low-frequency feature maps are arranged in a fixed order to obtain the power entropy distribution of the output feature map of each convolution kernel.

[0136] Secondly, the power entropy distribution calculated in step S1061 is called through step S1062, and then the adaptive weight coefficients corresponding to each feature map are calculated using a function whose weights are inversely proportional to the power entropy: the selected function is... Where w represents the adaptive weight coefficient of a 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. Its function is to avoid the denominator being 0 when H is 0, ensuring that the calculation process proceeds normally. According to this function, the power entropy values ​​of high-frequency, mid-frequency, and low-frequency feature maps are substituted into the calculation in turn to obtain the adaptive weight coefficient corresponding to each feature map. It can be seen from the calculation results that the higher the power entropy value of the feature map, the smaller the calculated adaptive weight coefficient. These calculated adaptive weight coefficients will serve as the core basis for subsequent multi-band feature component fusion.

[0137] Next, the multi-band feature components generated in step S105 are obtained through step S1063, and the adaptive weight coefficients calculated in step S1062 are called. A preliminary fusion feature map is generated according to the rule of pixel-by-pixel weighted summation: For all pixels with the same position in the three feature maps, the energy value of the pixel in the high-frequency, mid-frequency, and low-frequency feature maps is 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 different frequency band feature maps; then the three contribution values ​​at the same pixel position are summed to obtain the energy value of the pixel after fusion; according to the above method, weighted summation is performed on all corresponding pixel positions in the multi-band feature components one by one, and finally the fused energy values ​​of all pixels are arranged according to the original pixel positions to form a new feature map, which is the preliminary fusion feature map.

[0138] Finally, the implementation process of step S1064 involves first obtaining a preliminary fused feature map, followed by four steps of polarization channel energy correction processing: 1) Calculating the average energy of each polarization channel, counting the energy values ​​of all pixels within each polarization channel, summing these energy values ​​and dividing by the total number of pixels to obtain the average value as the current energy level of that polarization channel; 2) Determining a unified target energy level, calculating the average of the current energy levels of all polarization channels, which is the target energy level that all polarization channels need to be adjusted to; 3) Calculating the energy adjustment coefficient for each polarization channel, dividing the target energy level by the current energy level of each polarization channel, and obtaining the ratio as the corresponding energy adjustment coefficient for that polarization channel; 4) Scaling the pixel values ​​of each polarization channel, multiplying the energy values ​​of all pixels within each polarization channel by the energy adjustment coefficient of that channel to make the adjusted average energy of each polarization channel consistent with the target energy level, thereby eliminating the energy differences between channels caused by coherent superposition; 5) After completing the adjustment of all channels, outputting the preliminary fused feature map after energy correction, which is an anti-interference fingerprint feature with micro-Doppler feature enhancement characteristics.

[0139] In practical applications, in scenarios where spacecraft monitor missile penetration, high-frequency, mid-frequency, and low-frequency feature maps are acquired. Each pixel in one feature map contains an energy value. Using an entropy calculation method, the energy values ​​of all pixels in each of the three feature maps are first statistically analyzed, and the total energy of each map is calculated. Then, the proportion of energy in each pixel to the total energy is calculated. Finally, the power entropy values ​​of the high-frequency, mid-frequency, and low-frequency feature maps are calculated using the power entropy formula, and arranged in this order to form a power entropy distribution. This power entropy distribution is then used... The function (ε is set to 0.01) is used to successively substitute the power entropy values ​​of the three feature maps to calculate the corresponding adaptive weight coefficients, ensuring that the feature map with a high power entropy value has a smaller weight coefficient. The multi-band feature components and the previously calculated weight coefficients are obtained. For all pixels at the same position in the three feature maps, their energy values ​​are multiplied by the corresponding weight coefficients to obtain the contribution values. The three contribution values ​​at the same position are summed to obtain the fused pixel energy value. After all pixels are calculated, they are arranged in their original positions to generate a preliminary fused feature map. The preliminary fused feature map is obtained by first counting the energy values ​​of all pixels in the two channels, calculating the average energy of the two channels as the current energy level, and then taking the average of the current energy levels of the two channels as the target energy level. The target energy level is divided by the current energy level of each channel to obtain the energy adjustment coefficient. The energy values ​​of all pixels in each channel are multiplied by the corresponding adjustment coefficient to make the average energy of the two channels consistent with the target energy level, eliminating the energy difference between channels. Finally, the energy-corrected feature map is output as the micro-Doppler feature enhancement anti-interference fingerprint feature of the missile.

[0140] The overall scheme of S106 described above determines the clutter level of feature map information by calculating power entropy and assigns weights accordingly. Higher weights are assigned to low-power-entropy feature maps to suppress clutter. Pixel-level fusion is performed according to the weights, integrating multi-frequency band features while taking into account high-frequency details, mid-frequency variations, and low-frequency trends. Polarization channel energy correction eliminates energy differences, outputting a target recognition fingerprint with enhanced micro-Doppler features and strong anti-interference capabilities, thus improving recognition stability and accuracy.

[0141] The following is a complete example for steps 101 to 106, such as Figure 3 As shown, in the scenario of spacecraft monitoring missile penetration, the synthetic aperture radar on the spacecraft emits electromagnetic waves towards the missile, receives the reflected raw echo signal, divides the signal into overlapping time windows, performs spectrum analysis to calculate the frequency components of each window, and arranges all window spectra in chronological order to form a 256×5000 two-dimensional time-frequency characteristic map where rows represent frequency, columns represent time, and values ​​represent energy intensity. Features are extracted in parallel from the time-frequency feature map: the region is divided according to a pulse period of 0.2 seconds, and the frequency point with the strongest energy in each period is found using an energy detection algorithm. The points are then connected by time to form an instantaneous frequency change curve. The peak time of 12 pulses (0.1 seconds, 0.3 seconds...2.3 seconds) is found using a peak detection algorithm. The intervals between consecutive peaks (0.2 seconds, 0.2 seconds...0.22 seconds) and the differences between adjacent intervals (0, 0...0.02 seconds) are calculated and summed to generate a cumulative difference spectrum. The rate of change of the instantaneous frequency curve is calculated using numerical differentiation to obtain the derivative sequence. The maximum point of the cumulative difference spectrum is connected using a maximum detection algorithm to form an envelope feature. The derivative sequence and the envelope feature are horizontally concatenated to generate a 12×2 feature combination matrix.

[0142] Subsequently, the horizontal direction signal 3+4i of the deception interference signal was received through the horizontal polarization channel, while the vertical direction signal 1+2i was received simultaneously through the vertical polarization channel. A time synchronization algorithm revealed a 0.001-second delay in the vertical direction signal, which was adjusted to achieve phase alignment. Complex multiplication (-5+10i) and addition (4+6i) were then performed on the aligned signal, and the modulus of the addition result was calculated. The signal data volume, unaffected by polarization, is generated by arranging the data over time. A 12×2 feature combination matrix and the 12 numerical signal data volume are integrated into a 12×3 input data set, which is then used to decouple the input features into an adversarial network. The encoder performs nonlinear processing using the ReLU function, upscaling the 3D features to 60 dimensions and outputting deep, high-dimensional features. The decoupling layer uses an attention mechanism to split the 60-dimensional features into three subsets: "target-related," "interference-related," and "background-related." Mutual information values ​​are estimated using KL divergence (0.4, 0.9, and 0.2 respectively), and the subset corresponding to 0.9 ("target-related") is selected as the target feature data (12×22-dimensional) representing the minute motions 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. Within the module, three types of convolutional kernels, 3×3, 5×5, and 7×7, are used in parallel to generate preliminary features of 10×20×1, 8×18×1, and 6×16×1, respectively. The 3×3 convolutional kernel extracts the high-frequency feature components representing the high-frequency jitter of the warhead, the 5×5 convolutional kernel extracts the mid-frequency feature components representing local trajectory changes, and the 7×7 convolutional kernel extracts the low-frequency feature components representing the overall flight trend. Linear interpolation is used to supplement the time points of the low-frequency components to 12, and they are spliced ​​together to form multi-band feature components. The power entropy distribution of the three feature maps is calculated: 1.2 for high frequency, 0.9 for mid frequency, and 0.7 for low frequency, with corresponding weights of 0.826, 1.1, and 1.414. The multi-band feature components are weighted and summed pixel by pixel to generate a preliminary fusion feature map of 12×20. The average energy of the horizontal channel is 14 and the average energy of the vertical channel is 9. The target energy level is set to 11.5. Adjustment coefficients (0.82 and 1.28) are calculated and the pixel values ​​are scaled to make the energy of the two channels consistent. Finally, an anti-interference fingerprint feature with micro-Doppler feature enhancement is output for accurate missile identification.

[0144] Figure 4 This is a schematic diagram illustrating a specific implementation of a radar signal depth feature extraction system based on adversarial sample defense, as provided in this application. (Refer to...) Figure 4 The system may include:

[0145] Acquisition module 41 is used to acquire the raw echo signal of the spacecraft's synthetic aperture radar and convert the raw echo signal into a time-frequency feature map;

[0146] The first generation module 42 is used to extract the instantaneous frequency change curve of the intra-pulse modulation feature and the cumulative difference spectrum of the inter-pulse repetition interval in parallel from the time-frequency feature map, and to perform matrix concatenation 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.

[0147] The second generation module 43 is used to receive deception jamming signals using the polarization radar group, and perform complex coherent superposition operations on the horizontal and vertical signals of the deception jamming signals to generate a signal data volume that is not affected by polarization.

[0148] Separation module 44 is used to input the feature combination matrix and the signal data volume unaffected by polarization into the feature decoupling adversarial network, and separate the target feature data representing the small motion of the warhead through mutual information maximization constraint;

[0149] Extraction module 45 is used to input the target feature data into the multi-scale local attention module, and use a multi-level perception range convolution kernel group to scan the spatial structure of the target feature data in parallel to extract the multi-frequency feature components excited by the projectile motion.

[0150] The output module 46 is used to calculate adaptive weight coefficients based on the power entropy distribution of the feature maps output by each convolution kernel, perform pixel-level fusion of the multi-band feature components according to the adaptive weight coefficients, and 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-interference fingerprint feature with micro-Doppler feature enhancement characteristics.

[0151] The radar signal depth feature extraction system based on adversarial sample defense in this application embodiment is used to implement the aforementioned radar signal depth feature extraction method based on adversarial sample defense. Therefore, the specific implementation of the radar signal depth feature extraction system based on adversarial sample defense can be found in the embodiment section of the radar signal depth feature extraction method based on adversarial sample defense mentioned above. The specific implementation can be referred to the description of the corresponding embodiments, and will not be repeated here.

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

[0153] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the above-described methods for extracting deep features of radar signals based on adversarial sample defense.

[0154] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0155] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the embodiments of the radar signal depth feature extraction method based on adversarial sample defense.

[0156] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0157] The foregoing has provided a detailed description of a radar signal depth feature extraction method, system, electronic device, and storage medium based on adversarial example defense provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this 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 countermeasure network to separate 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, 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, and performing polarization channel energy correction processing on the fused feature maps to eliminate energy differences between channels caused by coherent superposition, so as to output 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, and performing polarization channel energy correction processing on the fused feature maps to eliminate energy differences between channels caused by coherent superposition, so as to output anti-jamming fingerprint features with micro-Doppler feature enhancement characteristics, comprises: 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 to eliminate energy differences between channels caused by coherent superposition by adjusting energy distributions of each polarization channel, so that energy levels between different polarization channels remain 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 to eliminate energy differences between channels caused by coherent superposition by adjusting energy distributions of each polarization channel, so that energy levels between different polarization channels remain consistent, and outputting the energy-corrected preliminary fused feature map as anti-jamming fingerprint features with micro-Doppler feature enhancement characteristics, comprises: 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 instantaneous frequency variation curve and the cumulative difference spectrum of the inter-pulse repetition interval of the intra-pulse modulation feature are extracted from the time-frequency feature map in parallel, the derivative sequence of the instantaneous frequency variation curve is matrix-pasted with the envelope feature of the cumulative difference spectrum to generate a feature combination matrix, including: 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 position of the pulse energy peak value in the time-frequency feature map and calculates the difference of the time interval of the continuous pulse peak value, and the cumulative sum of the differences is generated to form 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 matrix-pasted as two one-dimensional arrays in horizontal or vertical direction to generate a feature combination matrix.

8. A radar signal deep feature extraction system based on an adversarial sample defense, characterized in that, It includes: The acquisition module is used for acquiring the original echo signal of the synthetic aperture radar of the spacecraft, 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 with 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.

9. An electronic device, comprising: It includes: 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