Passive sensing radar target classification method and system based on electromagnetic detection feature extraction

Abstract

The application provides a passive sensing radar target classification method and system based on electromagnetic detection feature extraction, by acquiring an electromagnetic radiation signal set of a target area, performing multi-dimensional signal decomposition processing on the original scattering signal sequence, obtaining a time-frequency distribution feature set and a polarization response feature set corresponding to each signal collection period, according to a preset electromagnetic feature correlation rule, extracting a target electromagnetic feature combination from the time-frequency distribution feature set and the polarization response feature set, and generating a weight distribution parameter, performing weighted fusion processing on the target electromagnetic feature combination, inputting a fusion feature vector set into an adaptive classification network, performing hierarchical classification operation on the fusion feature vector set, outputting a class identification and a corresponding classification confidence of a target scatterer, and generating a radar detection optimization instruction. The application realizes high-precision differentiation of multiple types of scatterers under passive sensing conditions, while ensuring the environmental adaptability and computational efficiency of the classification process.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data processing, in particular to a passive perception radar target classification method and system based on electromagnetic detection feature extraction. BACKGROUND

[0002] In the field of electromagnetic perception technology, the passive radar target classification method realizes target identification by analyzing environmental electromagnetic radiation, and gradually becomes a key technology for electronic reconnaissance and remote sensing monitoring. The existing methods mostly use fixed mode feature extraction strategies, such as separately analyzing the time-frequency energy distribution or polarization response characteristics of target scattering signals, and inputting a general classifier for judgment after static weight fusion. Such methods have inherent limitations in complex electromagnetic environments: time-frequency features are difficult to distinguish different material targets with similar polarization response, and relying solely on polarization features cannot capture the dynamic scattering characteristics of the target; the static weight mechanism cannot adapt to the changes in feature reliability caused by environmental noise fluctuations, resulting in inaccurate feature fusion; traditional classifiers lack embedded modeling of electromagnetic physical mechanisms, and it is difficult to analyze effective features in weak scattering signals under passive perception conditions. The above defects seriously restrict the accurate identification ability of camouflage targets and composite material targets in complex battlefield environments and civil remote sensing scenes, and limit the practical application effectiveness of passive perception technology. SUMMARY

[0003] The present application provides a passive perception radar target classification method and system based on electromagnetic detection feature extraction.

[0004] In a first aspect, the present application embodiment provides a passive perception radar target classification method based on electromagnetic detection feature extraction, comprising:

[0005] Obtaining an electromagnetic radiation signal set of a target area, the electromagnetic radiation signal set containing a sequence of original scattering signals collected in multiple signal collection periods;

[0006] Performing multi-dimensional signal decomposition processing on the original scattering signal sequence to obtain a time-frequency distribution feature set and a polarization response feature set corresponding to each signal collection period;

[0007] According to a preset electromagnetic feature correlation rule, a target electromagnetic feature combination is extracted from the time-frequency distribution feature set and the polarization response feature set, and a weight distribution parameter is generated;

[0008] Based on the weight distribution parameter, the target electromagnetic feature combination is weighted and fused to generate a fused feature vector set, and the fused feature vector set is input into an adaptive classification network;

[0009] Perform a hierarchical classification operation on the fusion feature vector set through the adaptive classification network, output the class identification and corresponding classification confidence of the target scatterer, and generate a radar detection optimization instruction according to the classification confidence.

[0010] In a second aspect, an embodiment of the present application provides a computer system, comprising:

[0011] A memory, wherein the memory stores a computer program;

[0012] A processor for loading the computer program to implement the passive sensing radar target classification method based on electromagnetic detection feature extraction as described above.

[0013] The passive sensing radar target classification method based on electromagnetic detection feature extraction provided by the present application synchronously acquires the time-frequency distribution features and polarization response features of the target scatterer through multi-dimensional signal decomposition processing, and realizes the optimized fusion of multi-dimensional features based on the dynamic weight distribution strategy of electromagnetic physical mechanism, effectively overcoming the problem of limited classification precision caused by insufficient representation ability of traditional methods. The transient electromagnetic response characteristics of the target scatterer are accurately described using the time-frequency features, and the target material and geometric structure information are deeply analyzed using the polarization features, and the contribution weights of the two types of features in different electromagnetic environments are adaptively adjusted through dynamic weight distribution, which significantly enhances the robustness of feature representation in complex interference scenes. Further, the fusion features are progressively decided through the hierarchical adaptive classification network, the electromagnetic scattering physical law is embedded into the decision logic of the classification model, so that the classification process can not only fully utilize the physical interpretability of electromagnetic features, but also adapt to the time-varying characteristics in the actual detection environment, realizing high-precision differentiation of multiple types of scatterers under passive sensing conditions, while ensuring the environmental adaptability and computational efficiency of the classification process, providing reliable technical support for target identification in complex electromagnetic environments. BRIEF DESCRIPTION OF DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0015] Figure 1 is a flowchart of a passive sensing radar target classification method based on electromagnetic detection feature extraction provided by an embodiment of the present application.

[0016] Figure 2 is a composition schematic diagram of a computer system provided by an embodiment of the present application. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative effort belong to the scope of protection of the present application.

[0018] Please refer to Figure 1 , Figure 1 A flowchart of a passive sensing radar target classification method based on electromagnetic detection feature extraction is provided for an embodiment of the present application. The passive sensing radar target classification method based on electromagnetic detection feature extraction can be executed by a computer system. The passive sensing radar target classification method based on electromagnetic detection feature extraction can include the following steps:

[0019] Step S100: Obtain an electromagnetic radiation signal set of a target region, wherein the electromagnetic radiation signal set contains a sequence of original scattering signals collected in multiple signal collection periods.

[0020] The target region is a spatial range that needs to be detected by radar. Various target scatterers, such as airplanes, vehicles, and buildings, can exist in the region. The electromagnetic radiation signal set is a set of all electromagnetic radiation signals collected in the target region. These signals are collected in different signal collection periods. The sequence of original scattering signals is a sequence of electromagnetic signals scattered from the target region in chronological order. These signals contain various information of the target scatterers.

[0021] The electromagnetic radiation signal set of the target region can be obtained by deploying radar receiving devices around the target region. For example, when detecting a target region around an airport, multiple radar receiving nodes can be set up at different positions of the airport. Each node collects electromagnetic radiation signals of the target region in multiple signal collection periods, thereby obtaining a sequence of original scattering signals. Each signal collection period can be set according to actual needs, for example, set to 1 second. In each 1-second period, the radar receiving node continuously collects electromagnetic radiation signals. Finally, the signals collected in multiple such periods are combined into a sequence of original scattering signals, thereby forming an electromagnetic radiation signal set.

[0022] As an implementation, before the step S100 of obtaining the electromagnetic radiation signal set of the target region, the following signal preprocessing steps S101-S103 are further included:

[0023] Step S101: Deploy multiple passive radar receiving nodes, configure the frequency range coverage and polarization receiving mode of each receiving node, and form a spatial distribution detection array.

[0024] The passive radar receiving node is a device that does not actively emit electromagnetic signals and only receives electromagnetic signals scattered back by the target area. The frequency band coverage is the frequency range of the electromagnetic signals that the radar receiving node can receive. Different target scatterers may have more obvious scattering characteristics at different frequency bands, so reasonable configuration of the frequency band coverage can improve the detection ability of the target. The polarization receiving mode is the receiving mode of the radar receiving node for electromagnetic signals of different polarization directions. Common polarization modes include horizontal polarization and vertical polarization. Different polarization modes can obtain different polarization characteristic information of the target scatterer. The spatial distribution detection array is an array composed of multiple passive radar receiving nodes according to a set spatial layout. Through such an array, the target area can be detected from different angles and positions, so as to more comprehensively obtain the electromagnetic information of the target.

[0025] When deploying multiple passive radar receiving nodes, the number and position of the nodes need to be determined according to factors such as the size and shape of the target area and the types of targets that may appear. For example, in a target area of a city environment, in order to be able to cover the entire city range, passive radar receiving nodes can be deployed in various corners and key positions of the city. For the configuration of the frequency band coverage, the scattering characteristics of common target scatterers at different frequency bands can be selected. For example, for an airplane target, it may have more obvious scattering characteristics at certain high frequency bands, so the frequency band coverage of part of the receiving nodes can be set to these high frequency bands. For the configuration of the polarization receiving mode, both horizontal polarization and vertical polarization modes can be set, so that the scattering information of the target scatterer in different polarization directions can be obtained. By reasonably deploying and configuring these receiving nodes, an effective spatial distribution detection array can be formed.

[0026] Step S102: Synchronize the local clock sources of each receiving node to eliminate signal phase errors caused by time synchronization.

[0027] The local clock source is a clock system used for timing in each passive radar receiving node. Since the clock systems of the receiving nodes may have deviations, the times at which they collect signals are inconsistent. This time synchronization error will cause signal phase errors. Signal phase errors will affect subsequent signal processing and analysis, and may cause inaccurate extraction of target characteristics, thereby affecting the accuracy of target classification.

[0028] Synchronizing the local clock sources of each receiving node can be achieved in various ways, such as using global positioning system (GPS) clock synchronization technology. The GPS satellite can provide a high-precision time signal, and each passive radar receiving node can receive the GPS time signal through a GPS receiver and synchronize its local clock source with the GPS time. The specific implementation process is as follows: after the GPS receiver of each receiving node receives the GPS time signal, it compares it with the local clock and calculates the time difference between the two. Then, the receiving node adjusts the local clock according to this time difference to keep it consistent with the GPS time. In this way, the phase error of the signal caused by different time synchronization can be effectively eliminated.

[0029] Step S103: Perform local oscillator leakage suppression processing on the original signal collected by each receiving node, and use an adaptive filtering algorithm to eliminate the self-radiation interference of the device; wherein the execution steps of the adaptive filtering algorithm include: establishing a transfer function model of the hardware characteristics of the receiving node, and offsetting the leakage signal component through a reverse compensation operation.

[0030] Local oscillator leakage is the leakage of the local oscillator signal inside the radar receiving device into the receiving channel, which mixes with the target scattering signal, thereby interfering with the collection and processing of the target signal. Self-radiation interference of the device is the electromagnetic radiation generated by the radar receiving device during operation, which will affect the accurate collection of electromagnetic signals in the target area. The adaptive filtering algorithm is an algorithm that can automatically adjust the filter parameters according to the statistical characteristics of the input signal, which can effectively suppress the interference components in the signal.

[0031] Establishing a transfer function model of the hardware characteristics of the receiving node is the process of the adaptive filtering algorithm, and the transfer function model describes the transformation relationship from the input signal to the output signal of the receiving node hardware. It can be obtained by analyzing the circuit structure and working principle of the receiving node. For example, for a passive radar receiving node based on a radio frequency front-end circuit, its transfer function model can be established by measuring the relationship between its input and output signals using system identification methods. Specifically, a series of known test signals can be input to the receiving node, and its output signal is recorded, and then the parameters of the transfer function model are fitted by methods such as least squares, and the fitting process can refer to existing technologies.

[0032] After obtaining the transfer function model, the leakage signal component is offset through a reverse compensation operation. Specifically, the estimated value of the local oscillator leakage signal is calculated according to the transfer function model, and then the inverse signal of this estimated value is superimposed on the original signal to offset the influence of the leakage signal. For example, if the local oscillator leakage signal is calculated to be S leak by the transfer function model, then -S leak is superimposed on the original signal S rawThe signal S after the local oscillator leakage suppression processing is obtained processed =S raw -S leak Through the adaptive filtering algorithm, the device self-radiation interference can be effectively eliminated, and the signal quality can be improved.

[0033] Step S200: performing multi-dimensional signal decomposition processing on the original scattering signal sequence to obtain a time-frequency distribution feature set and a polarization response feature set corresponding to each signal acquisition period.

[0034] The multi-dimensional signal decomposition processing is to analyze and decompose the original scattering signal sequence from multiple different dimensions to extract more abundant signal features. The time-frequency distribution feature set is a set of features describing the distribution characteristics of the signal in the time and frequency dimensions, which can reflect the motion characteristics, electromagnetic characteristics and other information of the target scatterer. The polarization response feature set is a set of features describing the response characteristics of the signal in different polarization directions, which can reflect the shape, material and other information of the target scatterer.

[0035] The multi-dimensional signal decomposition processing on the original scattering signal sequence can be performed in various ways. For example, the time domain analysis and frequency domain analysis methods can be combined with the polarization analysis method. By analyzing the distribution of the signal in time and frequency, the time-frequency distribution features can be obtained; by analyzing the response of the signal in different polarization directions, the polarization response features can be obtained. In actual application, the original scattering signal sequence can be first converted in time domain and frequency domain, and then the corresponding feature parameters are extracted, and the signal is processed by polarization channel separation to extract the polarization response feature parameters, and finally the time-frequency distribution feature set and the polarization response feature set are formed respectively.

[0036] As an implementation manner, the step S200 of performing multi-dimensional signal decomposition processing on the original scattering signal sequence to obtain a time-frequency distribution feature set and a polarization response feature set corresponding to each signal acquisition period can specifically include the following steps S210-S240:

[0037] Step S210: performing time domain alignment operation on the original scattering signal sequence to eliminate the signal time sequence offset caused by environmental interference and generate a standardized time domain signal sequence.

[0038] Temporal alignment adjusts the original scattered signal sequence along the time axis to ensure temporal consistency. Environmental interference, including various electromagnetic interferences and noise around the target area, causes temporal shifts in the original scattered signal sequence, resulting in inaccurate signal arrival times. These temporal shifts affect subsequent signal analysis and processing, necessitating temporal alignment to eliminate them. A standardized temporal signal sequence, obtained after temporal alignment, possesses a unified temporal reference standard, facilitating subsequent feature extraction and analysis.

[0039] Temporal alignment can be performed using correlation analysis. The specific steps are as follows: First, select a reference signal, which can be a known standard signal or a signal acquired under interference-free conditions. Then, calculate the correlation between the original scattered signal sequence and the reference signal, finding the position with the highest correlation; the time offset corresponding to this position is the amount that needs adjustment. Finally, shift the original scattered signal sequence in time according to this time offset, thereby eliminating the temporal offset and generating a normalized time-domain signal sequence. For example, if the reference signal is S... ref (t), the original scattering signal sequence is S raw (t), by calculating the relevant function ,turn up The maximum value corresponding to , will S raw (t) time shift A standardized time-domain signal sequence is obtained. .

[0040] Step S220: The standardized time-domain signal sequence is divided into frequency bands using a multi-resolution orthogonal decomposition algorithm. The signal energy distribution parameters and phase continuity parameters in each frequency band are extracted to form an initial time-frequency feature set.

[0041] Multiresolution orthogonal decomposition (MOD) is an algorithm capable of decomposing a signal at different resolutions, breaking it down into sub-signals with different frequency components. Frequency banding divides the signal's frequency range into multiple distinct frequency bands, each corresponding to a different frequency interval. Signal energy distribution parameters describe the energy distribution of the signal within each frequency band, reflecting the scattering capability of the target scatterer at different frequencies. Phase continuity parameters describe the continuous phase change of the signal within each frequency band, reflecting information such as the motion characteristics of the target scatterer. The initial time-frequency feature set, composed of the signal energy distribution parameters and phase continuity parameters within each frequency band, forms the basis for further processing and analysis.

[0042] A possible multi-resolution orthogonal decomposition algorithm is, for example, a wavelet decomposition algorithm. Taking the wavelet decomposition algorithm as an example, the specific steps of frequency band division on the normalized time-domain signal sequence are as follows: first, a suitable wavelet basis function is selected, such as a Daubechies wavelet basis. Then, the normalized time-domain signal sequence is decomposed into wavelet coefficients of different scales and different positions by wavelet decomposition. The wavelet coefficients corresponding to each scale can be regarded as the components of the signal in different frequency bands. Next, the signal energy distribution parameters in each frequency band are calculated, for example, the signal energy in the frequency band can be obtained by calculating the sum of squares of the wavelet coefficients. For the phase continuity parameter, it can be extracted by analyzing the phase change of the wavelet coefficients. For example, the phase difference between adjacent wavelet coefficients can be calculated, and the phase continuity parameter can be obtained by analyzing the distribution of the phase differences. Finally, the signal energy distribution parameters and the phase continuity parameters in each frequency band are combined to form the initial time-frequency feature set.

[0043] Step S230: synchronously performing polarization channel separation processing on the original scattering signal sequence to obtain scattering intensity difference parameters and polarization coherence parameters under different polarization directions, and forming an initial polarization response set.

[0044] The polarization channel separation processing is to separate the original scattering signal sequence according to different polarization directions to obtain signal components in different polarization directions. The scattering intensity difference parameter is the difference of the scattering intensity of the signal under different polarization directions, which can reflect the shape, material and other information of the target scattering body. The polarization coherence parameter is the coherence degree between signals in different polarization directions, which can reflect the polarization characteristics of the target scattering body. The initial polarization response set is a feature set composed of scattering intensity difference parameters and polarization coherence parameters under different polarization directions, which can be used to describe the polarization response characteristics of the target scattering body.

[0045] The polarization channel separation processing can be performed by using polarization filtering. Specifically, for an original scattering signal sequence containing horizontal polarization and vertical polarization components, a horizontal polarization filter and a vertical polarization filter can be used to filter it respectively to obtain horizontal polarization components and vertical polarization components. Then, the scattering intensity difference parameters under different polarization directions are calculated, for example, the scattering intensity difference can be obtained by calculating the energy ratio of the horizontal polarization components and the vertical polarization components. For the polarization coherence parameter, the cross-correlation coefficient of the signals in different polarization directions can be calculated. For example, if the horizontal polarization component is and the vertical polarization component is , the polarization coherence parameter can be represented as , where represents the conjugate of the scattering intensity difference parameter and the polarization coherence parameter under different polarization directions to form an initial polarization response set.

[0046] Step S240: performing nonlinear normalization processing on the initial time-frequency feature set and the initial polarization response set to eliminate dimension difference and retain feature distribution mode, to generate the time-frequency distribution feature set and the polarization response feature set; wherein the nonlinear normalization processing comprises: selecting a corresponding normalization function according to the statistical distribution characteristics of each feature parameter, performing truncation compensation operation on the feature values that exceed a preset fluctuation threshold, and retaining the shape of the normalized feature distribution curve.

[0047] The nonlinear normalization processing is a method of uniformly processing different feature parameters to have the same dimension and range. The dimension difference is that different feature parameters may have different physical units and numerical ranges, and such difference will affect the performance of subsequent feature analysis and classification algorithms. The feature distribution mode is the distribution law of the feature parameter in the value range, and retaining the feature distribution mode can ensure that the normalized feature can still reflect the essential information of the original feature.

[0048] A corresponding normalization function is selected according to the statistical distribution characteristics of each feature parameter. For example, for a feature parameter subject to normal distribution, a z - score normalization function can be selected, and its formula is where x is the original feature value, is the mean of the feature value, and σ is the standard deviation of the feature value. For a feature parameter not subject to normal distribution, a nonlinear normalization function such as a logarithmic normalization function .

[0049] The feature values that exceed the preset fluctuation threshold are subjected to truncation compensation operation. The preset fluctuation threshold is a feature value range set according to the actual situation, and if a feature value exceeds this range, it is considered to be an abnormal value. For abnormal values, a truncation compensation method can be used for processing. For example, if the preset fluctuation threshold is [a, b], if a feature value x is greater than b, it is truncated to b; if x is less than a, it is truncated to a. At the same time, in order to retain the shape of the normalized feature distribution curve, the feature values can be appropriately adjusted after the truncation compensation so as to still conform to the distribution law of the original feature. Through these operations, the initial time-frequency feature set and the initial polarization response set are subjected to nonlinear normalization processing, and finally the time-frequency distribution feature set and the polarization response feature set are generated.

[0050] Step S300: Extracting a target electromagnetic feature combination from the time-frequency distribution feature set and the polarization response feature set according to a preset electromagnetic feature correlation rule, and generating a weight distribution parameter.

[0051] The preset electromagnetic feature correlation rule is a rule that is preset to describe the correlation between the feature parameters in the time-frequency distribution feature set and the polarization response feature set. These rules can be formulated according to the electromagnetic characteristics of the target scatterer and the actual detection requirements. The target electromagnetic feature combination is a group of feature parameter combinations closely related to target classification selected from the time-frequency distribution feature set and the polarization response feature set. The weight distribution parameter is a weight value assigned to each feature parameter in the target electromagnetic feature combination. These weight values reflect the importance of each feature parameter in target classification.

[0052] According to the preset electromagnetic feature correlation rule, the target electromagnetic feature combination can be extracted from the time-frequency distribution feature set and the polarization response feature set using a feature selection algorithm. For example, a feature selection algorithm based on correlation analysis can be used to calculate the correlation between each feature parameter in the time-frequency distribution feature set and the polarization response feature set and the target classification label, and select the feature parameters with higher correlation to form the target electromagnetic feature combination. For generating the weight distribution parameter, the importance and stability of each feature parameter can be considered. For example, a feature importance evaluation method in machine learning, such as feature importance evaluation in random forest algorithm, can be used to train a random forest model, calculate the influence degree of each feature parameter on the model classification result, and assign a weight value to each feature parameter according to the influence degree.

[0053] As an implementation, the step S300, according to the preset electromagnetic feature correlation rule, extracts a target electromagnetic feature combination from the time-frequency distribution feature set and the polarization response feature set, and generates a weight distribution parameter, can specifically include the following steps S310-S340:

[0054] Step S310: Constructing a feature correlation degree evaluation matrix to calculate the mutual information and covariance correlation coefficient between each feature parameter in the time-frequency distribution feature set and each feature parameter in the polarization response feature set.

[0055] The feature correlation degree evaluation matrix is a matrix used to evaluate the correlation degree between the feature parameters in the time-frequency distribution feature set and the polarization response feature set. Mutual information is an index used to measure the dependence between two random variables, which can reflect the information sharing degree between two feature parameters. The covariance correlation coefficient is an index used to measure the linear relationship between two random variables, which can reflect the linear correlation between two feature parameters.

[0056] The specific steps of constructing the feature correlation evaluation matrix are as follows: first, all feature parameters in the time-frequency distribution feature set and the polarization response feature set are numbered. Then, for each pair of feature parameters, the mutual information and the covariance correlation coefficient between them are calculated. The mutual information can be calculated using the formula in information theory. If two feature parameters X and Y, their joint probability distribution is p(x, y), and the marginal probability distribution is p(x) and p(y) respectively, then the mutual information between them is The covariance correlation coefficient can be calculated using the formula where Cov(X, Y) is the covariance of X and Y, and Var(X) and Var(Y) are the variances of X and Y. Fill the calculated mutual information and covariance correlation coefficient into the matrix to form the feature correlation evaluation matrix.

[0057] Step S320: generating a feature redundancy index according to the mutual information and the covariance correlation coefficient, and performing a redundancy filtering operation on the time-frequency distribution feature set and the polarization response feature set based on the feature redundancy index to obtain a candidate feature subset.

[0058] The feature redundancy index is an index for measuring the degree of redundancy between feature parameters, which can be calculated according to the mutual information and the covariance correlation coefficient. The redundancy filtering operation is to remove those redundant feature parameters from the time-frequency distribution feature set and the polarization response feature set, and only keep those feature parameters with independent information, thereby obtaining a candidate feature subset.

[0059] Generating a feature redundancy index can use the method of weighted average, for example, the mutual information and the covariance correlation coefficient are weighted and summed to obtain the feature redundancy index. If the mutual information is I, the covariance correlation coefficient is , and the weights are w1 and w2 respectively, then the feature redundancy index is The specific steps of performing a redundancy filtering operation based on the feature redundancy index are as follows: set a redundancy threshold T, for each pair of feature parameters in the feature correlation evaluation matrix, if their feature redundancy index is greater than T, it is considered that the two feature parameters are redundant, and one of them is selected to be kept and the other is removed. In this way, the time-frequency distribution feature set and the polarization response feature set are subjected to a redundancy filtering operation, and finally a candidate feature subset is obtained.

[0060] Step S330: performing sensitivity analysis on the candidate feature subset to obtain the discrimination degree weight of each candidate feature for the target scatterer category, and generating an initial feature weight set based on the discrimination degree weight.

[0061] Sensitivity analysis is a method for analyzing the degree of influence of changes in input variables on output results. In this step, the degree of influence of each feature parameter in the candidate feature subset on the classification of the target scatterer is analyzed. The discrimination weight is the importance of each candidate feature in distinguishing different target scatterer categories. The initial feature weight set is a set composed of the discrimination weights of each candidate feature.

[0062] Sensitivity analysis of the candidate feature subset can be performed using experimental design methods. For example, a series of experiments can be designed, in each experiment, the value of a candidate feature is changed, and the change in the classification result of the target scatterer is observed. Through multiple experiments, the degree of influence of the change in the value of each candidate feature on the classification result is calculated, thereby obtaining the discrimination weight of each candidate feature. Specifically, the change in classification accuracy can be used to measure the discrimination weight. If the classification accuracy changes from A1 to A2 after changing the value of a candidate feature x i , the discrimination weight w i of the candidate feature is |A 2- A1|. The discrimination weights of each candidate feature are combined to generate the initial feature weight set.

[0063] Step S340: According to the physical property constraint condition of the target scatterer, the initial feature weight set is adjusted to eliminate the weight deviation caused by environmental noise, and the weight distribution parameter is generated; wherein the adjustment includes: introducing an external calibration signal source to verify the physical property correlation of the candidate feature subset, when the feature weight deviates from the standard weight of the calibration signal source by more than a preset threshold, starting the weight compensation algorithm to correct the deviation.

[0064] The physical property constraint condition of the target scatterer is some inherent physical properties of the target scatterer, such as shape, material, size, etc., which will affect the value and importance of the feature parameters. Environmental noise is various interference factors received during radar detection, which will cause the feature weight to deviate. The external calibration signal source is a signal source with known physical properties, which can be used to verify the physical property correlation of the candidate feature subset. The standard weight is the feature weight value of the candidate feature subset in an ideal case. The weight compensation algorithm is an algorithm for correcting the deviation of the feature weight.

[0065] The specific steps of verifying the physical property correlation of the candidate feature subset by introducing an external calibration signal source are as follows: an external calibration signal source is transmitted to the target area, signals received by the radar receiving node are collected, and feature parameters of the candidate feature subset are extracted. According to the known physical properties of the external calibration signal source, the standard weight of each feature parameter is calculated. Then, the calculated standard weight is compared with the feature weight in the initial feature weight set. When it is detected that the feature weight deviates from the standard weight by more than a preset threshold, a weight compensation algorithm is started to correct the deviation. For example, the least square method or other methods can be used to adjust the feature weight, so that the adjusted feature weight is closer to the standard weight. Through these operations, the initial feature weight set is adjusted, and finally the weight distribution parameter is generated.

[0066] Step S400: performing weighted fusion processing on the target electromagnetic feature combination based on the weight distribution parameter, generating a fusion feature vector set, and inputting the fusion feature vector set into an adaptive classification network.

[0067] The weighted fusion processing is to perform weighted summation on each feature parameter in the target electromagnetic feature combination according to the weight distribution parameter, and fuse multiple feature parameters into a comprehensive feature vector. The fusion feature vector set is a set composed of multiple weighted fused feature vectors, which contains more comprehensive information of the target scatterer. The adaptive classification network is a neural network that can automatically adjust the classification rule according to the input data, which can classify the fusion feature vector set and output the class label of the target scatterer.

[0068] The specific steps of performing weighted fusion processing on the target electromagnetic feature combination based on the weight distribution parameter are as follows: if there are n feature parameters in the target electromagnetic feature combination , the corresponding weight distribution parameter is , the weighted fused feature vector is . The weighted fusion processing is performed on each feature parameter combination in the target electromagnetic feature combination, and multiple fusion feature vectors are obtained. These fusion feature vectors are combined to generate a fusion feature vector set. Then, the fusion feature vector set is input into the adaptive classification network, and the network classifies the fusion feature vector according to its internal classification rule and outputs the class label of the target scatterer.

[0069] As an implementation manner, the step S400 of performing weighted fusion processing on the target electromagnetic feature combination based on the weight distribution parameter to generate a fusion feature vector set can specifically include the following steps S410-S440:

[0070] Step S410: Construct a feature fusion equation according to the weight distribution parameters, and respectively weight and superimpose each feature parameter in the time-frequency distribution feature set and each feature parameter in the polarization response feature set.

[0071] The feature fusion equation is an equation for describing how to fuse the feature parameters in the time-frequency distribution feature set and the polarization response feature set. It determines the weight coefficients of each feature parameter according to the weight distribution parameters. The weight superposition is to multiply each feature parameter by its corresponding weight coefficient, and then add the products to obtain a comprehensive feature value.

[0072] The specific steps of constructing the feature fusion equation according to the weight distribution parameters are as follows: if there are m feature parameters in the time-frequency distribution feature set , the corresponding weight distribution parameters are ; if there are k feature parameters in the polarization response feature set , the corresponding weight distribution parameters are . Then the feature fusion equation can be expressed as . Each feature parameter in the time-frequency distribution feature set and the polarization response feature set is substituted into the feature fusion equation for weight superposition to obtain a comprehensive feature value. For each feature parameter combination, such weight superposition processing is performed to obtain multiple comprehensive feature values, which can further form a fusion feature vector.

[0073] Step S420: Perform energy consistency detection on the candidate feature combinations after weight superposition, eliminate abnormal feature combinations caused by weight distribution mismatch, and retain effective feature combinations that satisfy the energy conservation criterion.

[0074] Energy consistency detection is to check whether the candidate feature combinations after weight superposition satisfy the energy conservation criterion, i.e., whether the energy of the feature combinations is within a reasonable range. Weight distribution mismatch is that due to unreasonable weight distribution parameter setting, the energy of some feature combinations is abnormal during weight superposition. Abnormal feature combinations are those that do not satisfy the energy conservation criterion, and these feature combinations may affect the subsequent classification results, so they need to be eliminated. Effective feature combinations are those that satisfy the energy conservation criterion, and they can be used for subsequent processing and analysis.

[0075] The energy consistency detection on the candidate feature combinations after weight superposition can adopt the method of energy calculation and threshold judgment. First, calculate the energy of each candidate feature combination, for example, the energy value can be obtained by calculating the sum of squares of the feature parameters in the feature combination. Then, set an energy threshold range [E min , E maxIf the energy value of a candidate feature combination is not within the range, the feature combination is considered to be an abnormal feature combination and is removed. If the energy value is within the range, the feature combination is considered to be an effective feature combination and is retained. Through the energy consistency detection, abnormal feature combinations caused by weight distribution mismatch can be removed, and the quality of the feature combinations can be improved.

[0076] Step S430: performing dimension compression processing on the effective feature combinations, mapping the high-dimensional feature space to a low-dimensional feature space by using a principal component preservation algorithm to generate a set of reduced-dimensional feature vectors; wherein the execution steps of the principal component preservation algorithm include: calculating the information entropy ratio of each effective feature combination before and after dimension reduction, and when the information entropy ratio is lower than a preset information loss threshold, enabling an alternative dimension reduction channel to compensate for the information loss amount.

[0077] The dimension compression processing is to convert the high-dimensional effective feature combinations into low-dimensional feature vectors to reduce the complexity and computational amount of data. The principal component preservation algorithm is an algorithm that can preserve the principal component information of the original data as much as possible during dimension reduction. The information entropy ratio is the ratio of the information entropy of the effective feature combination before and after dimension reduction, which can reflect the degree of information loss during dimension reduction. The preset information loss threshold is a lower limit of the information entropy ratio set according to actual needs. If the information entropy ratio is lower than this threshold, it is considered that the information loss during dimension reduction is too large, and the alternative dimension reduction channel needs to be enabled for compensation. The alternative dimension reduction channel is a backup dimension reduction method, which can be used when the information loss of the main dimension reduction channel is too large to reduce the information loss.

[0078] The specific steps of mapping the high-dimensional feature space to the low-dimensional feature space by using the principal component preservation algorithm are as follows: first, standardizing the effective feature combinations to have the same mean and variance. Then, the covariance matrix of the effective feature combinations is calculated, and the principal components are obtained by solving the eigenvalues and eigenvectors of the covariance matrix. The first n principal components with larger eigenvalues are selected as the main features, and the original effective feature combinations are projected onto these principal components to obtain the reduced-dimensional feature vectors. During dimension reduction, the information entropy ratio of each effective feature combination before and after dimension reduction is calculated. When the information entropy ratio is lower than the preset information loss threshold, the alternative dimension reduction channel is enabled, for example, the local linear embedding algorithm can be used for dimension reduction to compensate for the information loss amount. Finally, the reduced-dimensional feature vectors are combined to generate a set of reduced-dimensional feature vectors.

[0079] Step S440: serializing and recombining the set of reduced-dimensional feature vectors to generate the set of fusion feature vectors with spatiotemporal correlation characteristics.

[0080] The serialization and recombination of the dimension-reduced feature vector set is to recombine the feature vectors in the dimension-reduced feature vector set according to an order and a rule, so that the feature vectors have a spatio-temporal correlation characteristic. The spatio-temporal correlation characteristic is a correlation between the feature vectors in time and space, which can reflect the motion characteristic and the spatial distribution characteristic of the target scatterer.

[0081] The serialization and recombination of the dimension-reduced feature vector set can adopt a time series analysis and a spatial correlation analysis method. For example, the feature vectors can be sorted according to the time sequence of collection, and then adjacent feature vectors are combined to form a feature vector sequence with a time correlation. At the same time, the feature vectors collected at different positions can be analyzed according to the spatial position information of the radar receiving nodes, and the feature vectors with a spatial correlation are combined. Through these operations, the serialization and recombination of the dimension-reduced feature vector set is performed, and finally a fusion feature vector set with a spatio-temporal correlation characteristic is generated.

[0082] Step S500: performing a hierarchical classification operation on the fusion feature vector set by the adaptive classification network, outputting a class identification of the target scatterer and a corresponding classification confidence, and generating a radar detection optimization instruction according to the classification confidence.

[0083] The hierarchical classification operation is that the adaptive classification network classifies the fusion feature vector set according to a hierarchical structure, gradually narrows the classification range, and finally determines the class of the target scatterer. The class identification is a label used to represent the class to which the target scatterer belongs, such as an airplane, a vehicle, a building, etc. The classification confidence is the reliability of the classification result, which can be represented by a probability value. The radar detection optimization instruction is an instruction generated according to the classification confidence for optimizing the radar detection process, such as adjusting the transmission power, frequency, detection angle, etc. of the radar.

[0084] The specific process of performing a hierarchical classification operation on the fusion feature vector set by the adaptive classification network is as follows: the adaptive classification network first receives the fusion feature vector set, and then starts classification of the fusion feature vector from the topmost layer according to the hierarchical classification structure inside it. At each level, the network judges the fusion feature vector according to the preset classification rule, selects the most suitable sub-class for classification at the next level, and continues until the bottommost class is reached. In the classification process, the network calculates the confidence of each classification result, for example, the posterior probability can be used to represent the classification confidence. Finally, the network outputs the class identification of the target scatterer and the corresponding classification confidence. According to the classification confidence, a radar detection optimization instruction is generated. If the classification confidence is low, it means that the reliability of the radar detection result is not high, and at this time, an instruction can be generated to adjust the parameters such as the transmission power and frequency of the radar, so as to improve the accuracy of detection.

[0085] As an implementation, the step S500 outputs the class label of the target scatterer and the corresponding classification confidence by performing a hierarchical classification operation on the set of fusion feature vectors through the adaptive classification network, and can specifically include the following steps S510-S540.

[0086] Step S510: Construct a multi-level classification decision tree, wherein each non-leaf node is configured with a feature discrimination model based on electromagnetic scattering mechanism, and each leaf node is associated with a preset class label of the target scatterer.

[0087] The multi-level classification decision tree is a tree-structured classification model composed of multiple levels of nodes, each node representing a classification decision point. The non-leaf node is a node in the decision tree other than the leaf node, which is used for classification judgment. The feature discrimination model based on electromagnetic scattering mechanism is a model established according to the electromagnetic scattering characteristics of the target scatterer, which can determine the subcategory of the target scatterer according to the input feature vector. The leaf node is the bottommost node of the decision tree, and each leaf node is associated with a preset class label of the target scatterer, used to represent the final classification result. The specific steps of constructing the multi-level classification decision tree are as follows: first, determine the hierarchical structure and node number of the decision tree according to the class and feature information of the target scatterer. Then, for each non-leaf node, establish a feature discrimination model based on electromagnetic scattering mechanism. For example, support vector machines, decision tree algorithms, etc. can be used to establish the feature discrimination model. For each leaf node, associate a preset class label of the target scatterer, such as an airplane, a vehicle, etc. Through these steps, the multi-level classification decision tree is constructed.

[0088] Step S520: input the set of fusion feature vectors into the root node of the multi-level classification decision tree, and calculate the matching degree between the fusion feature vector and the feature template of each child node according to the feature discrimination model corresponding to the current node; wherein the calculation steps of the feature discrimination model include: using a template updating mechanism to adjust the matching threshold of the feature template according to the real-time classification result feedback, to ensure the adaptability of the classification decision tree to the time-varying electromagnetic environment.

[0089] The feature template is a feature vector template corresponding to each child node, which represents the typical features of the target scatterer category corresponding to the child node. The matching degree is the similarity between the fusion feature vector and the feature template, which can be represented by a numerical value. The template updating mechanism is a mechanism for adjusting the matching threshold of the feature template according to the real-time classification result feedback, which can make the classification decision tree adapt to the time-varying electromagnetic environment.

[0090] After the fusion feature vector set is input into the root node of the multi-level classification decision tree, the matching degrees of the fusion feature vector and the feature templates of each child node are calculated according to the feature discrimination model corresponding to the current node. The specific calculation process is as follows: first, for the feature template of each child node, the distance between the fusion feature vector and the template is calculated, for example, the Euclidean distance, cosine similarity, etc. can be used to calculate the distance. Then, the distance value obtained by the feature discrimination model is converted into the matching degree. For example, if the Euclidean distance is used, the smaller the distance, the higher the matching degree. In the calculation process, the feature discrimination model adopts a template updating mechanism to ensure adaptability to the time-varying electromagnetic environment. Specifically, after completing a classification, feedback will be performed according to the real-time classification result. If the classification accuracy of a certain feature template is low, it means that the template may not be suitable for the current electromagnetic environment, at which time the matching threshold of the feature template will be adjusted. For example, if the original matching threshold is set to 0.8, and if there are many classification errors under this template, the threshold can be appropriately reduced to 0.7 to expand the matching range; on the contrary, if the classification accuracy is high, the threshold can be appropriately increased. In this way, by continuously adjusting the matching threshold according to the real-time classification result, the classification decision tree can better adapt to the changes in the electromagnetic environment.

[0091] Step S530: Select the child node with the highest matching degree as the next level decision node, and iteratively perform matching degree calculation and node selection operations until the leaf node is reached.

[0092] After the matching degrees of the fusion feature vector and the feature templates of each child node are calculated, the child node with the highest matching degree is selected from these child nodes as the decision node of the next level. Taking specific numerical values as an example, if the current node has three child nodes, and the matching degrees of the fusion feature vector and their feature templates are 0.6, 0.8 and 0.7 respectively, then the child node with a matching degree of 0.8 is selected as the decision node of the next level. Next, the fusion feature vector is input into this new decision node, and the matching degrees of the fusion feature vector and the feature templates of each child node under this node are calculated again according to the feature discrimination model corresponding to the node, and then the child node with the highest matching degree is selected as the decision node of the next level, and the matching degree calculation and node selection operations are iteratively performed. As the level continues to deepen, the leaf node of the decision tree will eventually be reached, at which time a complete classification path is completed.

[0093] Step S540: Calculate the classification confidence according to the historical classification accuracy corresponding to the leaf node and the current matching degree deviation, and generate a classification result set associated with the preset class identifier.

[0094] The historical classification accuracy is the proportion of correct classification of the leaf node in the past multiple classification processes, which reflects the reliability of the class identifier represented by the leaf node in the past classification. The current matching degree deviation is the difference between the matching degree of the fusion feature vector and the feature template of the leaf node and the average matching degree of the leaf node, which reflects the difference degree between this classification and the past situation. The classification confidence is an index that comprehensively considers the historical classification accuracy and the current matching degree deviation, and is used to measure the reliability of the classification result.

[0095] The specific way of calculating the classification confidence is as follows: first, set a weight coefficient, and weight the historical classification accuracy and the current matching degree deviation respectively. If the weight of the historical classification accuracy is 0.7 and the weight of the current matching degree deviation is 0.3. If the historical classification accuracy of a leaf node is 0.9 and the current matching degree deviation is 0.1 (indicating that this matching degree is 0.1 lower than the average matching degree), then the classification confidence = 0.7 x 0.9 + 0.3 x (1 - 0.1) = 0.63 + 0.27 = 0.9. After obtaining the classification confidence, it is combined with the preset class identifier associated with the leaf node to generate a classification result set. For example, the preset class identifier is "airplane" and the classification confidence is 0.9, so the classification result set contains information such as "airplane, 0.9".

[0096] As an implementation mode, the method further includes an optimization process of the classification model, which can specifically include steps S550-S580:

[0097] Step S550: Collecting error classification samples in the historical classification result set, and extracting abnormal fusion feature vectors and environmental interference parameters corresponding to the error classification samples.

[0098] The error classification sample is a sample that is incorrectly classified in the historical classification process. These samples contain information about possible problems of the classification model. The abnormal fusion feature vector is the fusion feature vector corresponding to the error classification sample, which has a different feature pattern from the correct classification sample, which may be caused by environmental interference or model defects. The environmental interference parameter is an environmental parameter when collecting these error classification samples, such as atmospheric humidity, temperature, electromagnetic noise intensity, etc., which may have an impact on the classification result.

[0099] The error classification samples in the collection history classification result set can be obtained by comparing the classification result with the actual target category. For example, the category identifier output by the classification model is compared with the actual category obtained by other reliable means (such as optical observation, manual confirmation, etc.), and the samples with classification errors are found out. For these error classification samples, the corresponding abnormal fusion feature vector is extracted, which can be specifically found from the stored fusion feature vector data according to the identification information of the sample. At the same time, the environmental interference parameters when collecting these samples are recorded, and the atmospheric humidity, temperature and other parameters can be obtained by setting environmental monitoring sensors around the radar detection equipment, and the electromagnetic noise intensity and other parameters can be obtained by electromagnetic monitoring equipment.

[0100] Step S560: constructing an interference feature simulator according to the environmental interference parameters to generate a simulated interference feature set with similar statistical characteristics to the abnormal fusion feature vector.

[0101] The interference feature simulator is a model for generating simulated interference features according to environmental interference parameters. Similar statistical characteristics refer to that the simulated interference feature set is similar to the abnormal fusion feature vector in statistical indicators such as mean, variance and distribution form.

[0102] The specific steps of constructing the interference feature simulator are as follows: first, analyze the relationship between the environmental interference parameters and the abnormal fusion feature vector. Regression analysis and other methods can be used to establish a regression model between the environmental interference parameters and the dimensions of the abnormal fusion feature vector. For example, through linear regression analysis, a linear relationship expression between atmospheric humidity and a dimension of the abnormal fusion feature vector is obtained. Then, according to this relationship model, a random number generator and other tools are used to generate simulated interference features with similar statistical characteristics within a given range of environmental interference parameters. For example, according to the regression model, the simulated interference features that meet the feature distribution under the condition of the current atmospheric humidity value are randomly generated on the basis of the current atmospheric humidity value. By continuously repeating this process, a large number of simulated interference features are generated to form a simulated interference feature set.

[0103] Step S570: injecting the simulated interference feature set into the training data set of the adaptive classification network to retrain the feature discrimination model in the multi-level classification decision tree.

[0104] Injecting the simulated interference feature set into the training data set of the adaptive classification network is to enable the classification model to learn the feature pattern in the presence of environmental interference and improve the robustness of the model. Re-training the feature discrimination model in the multi-level classification decision tree adjusts the parameters of the model to better adapt to input data containing interference features.

[0105] In a specific implementation, the simulated interference feature set is merged with the original training data set. For example, if the original training data set has 1000 samples and the simulated interference feature set has 200 samples, the new training data set has 1200 samples after merging. Then, the feature discrimination model in the multi-level classification decision tree is retrained using the new training data set. During the training process, optimization algorithms such as gradient descent are used to continuously adjust the parameters of the model to minimize the classification error of the model on the training data.

[0106] As an implementation, the step S570 of retraining the feature discrimination model in the multi-level classification decision tree can specifically include the following steps S571-S574:

[0107] Step S571: Perform node vulnerability analysis on the multi-level classification decision tree to identify a set of decision nodes that are susceptible to interference features.

[0108] Node vulnerability analysis is to evaluate the stability and reliability of each decision node in the multi-level classification decision tree when facing interference features. Decision nodes susceptible to interference features are those nodes whose classification results are prone to errors when inputting data containing interference features.

[0109] The method of performing node vulnerability analysis can be to input a series of test data containing different levels of interference features into the decision tree and observe the changes in the classification results of each node. For example, gradually increase the intensity of interference features in the test data and record the classification accuracy of each node under different interference intensities. If the classification accuracy of a node decreases significantly with the increase of interference feature intensity, then the node is considered to be susceptible to interference features. By analyzing all nodes in this way, a set of decision nodes susceptible to interference features is identified.

[0110] Step S572: For each vulnerable decision node, generate an adversarial training sample set containing adversarial feature vectors generated by perturbing the original feature vector.

[0111] The adversarial training sample set is a collection of training samples generated to enhance the robustness of the vulnerable decision node. The adversarial feature vector is a vector obtained by perturbing the original feature vector, and these perturbations simulate the impact of environmental interference on the feature vector.

[0112] The specific way of generating the set of adversarial training samples is as follows: for each fragile decision node, select its corresponding original feature vector. Then, perturb the original feature vectors. For example, a random noise value can be added to each dimension of the original feature vector, and the range of the noise value is set according to the actual situation. If the original feature vector is [x1, x2, x3], a random noise in the range of [-0.1, 0.1] is added to each dimension to obtain the adversarial feature vector [x1+n1, x2+n2, x3+n3], where n1, n2, n3 are random noise values generated in the range of [-0.1, 0.1]. By perturbing a plurality of original feature vectors in this way, a large number of adversarial feature vectors are generated to form a set of adversarial training samples.

[0113] Step S573: Merge the set of adversarial training samples and the set of simulated interference features to construct an enhanced training data set.

[0114] The set of adversarial training samples and the set of simulated interference features are merged to enable the retrained feature discrimination model to learn the information of both adversarial samples and simulated interference features, further improving the robustness of the model.

[0115] The specific merging process is, for example, to directly combine the samples in the set of adversarial training samples and the set of simulated interference features.

[0116] Step S574: locally retrain the feature discrimination model corresponding to the fragile decision node using an elastic weight solidification algorithm to prevent overfitting and maintain global classification stability; wherein the execution steps of the elastic weight solidification algorithm include: calculating the importance weight of the model parameters, limiting the adjustment range of the important parameters, and allowing the secondary parameters to be updated according to the new training data.

[0117] The elastic weight solidification algorithm is an algorithm for balancing model parameter updates and preventing overfitting. The importance weight of the model parameters is the contribution degree of each model parameter to the performance of the model, the important parameters are the parameters that have a greater impact on the performance of the model, and the secondary parameters are the parameters that have a relatively smaller impact on the performance of the model.

[0118] The specific steps of performing the elastic weight solidification algorithm are as follows: first, the importance weight of each parameter in the feature discrimination model corresponding to the fragile decision node is calculated. Gradient-based methods can be used, such as measuring the importance of a parameter by calculating the absolute value of its gradient. The larger the absolute value of the gradient, the greater the impact of the parameter on the model performance, and the higher the importance weight. Then, in the local retraining process, for important parameters, limit their adjustment range. For example, set an upper limit for the adjustment range, and if the update amount of the important parameter exceeds this upper limit, limit the update amount within the upper limit. For secondary parameters, allow them to be updated normally according to the new training data. In this way, the model can be adjusted according to the new training data, and overfitting caused by excessive adjustment of important parameters can be prevented, thereby maintaining the global classification stability.

[0119] Step S580: Evaluate the optimization effect by comparing the change in classification accuracy before and after retraining. If the change in classification accuracy does not reach the preset optimization threshold, increase the complexity of the simulated interference features and iteratively perform the injection and training operations. The retraining step includes: using an incremental learning strategy to retain the original classification knowledge, and adjusting the splitting criteria and feature weight parameters of the decision tree nodes according to the simulated interference features.

[0120] The change in classification accuracy is the difference between the classification accuracy of the retrained model and the classification accuracy before retraining, which reflects the degree of improvement in model performance after retraining. The preset optimization threshold is a lower limit of the change in classification accuracy set according to actual needs. If the change in classification accuracy does not reach this threshold, the optimization effect is not obvious and needs to be further adjusted. The complexity of the simulated interference features can be improved by increasing the dimension of the interference features or changing the distribution form of the interference features.

[0121] The change in classification accuracy before and after retraining can be tested by using the same test data set to test the model before and after retraining, respectively, and then calculating the difference. For example, the classification accuracy of the model before retraining is 0.8, and the classification accuracy after retraining is 0.82, then the classification accuracy change is 0.02. If the preset optimization threshold is 0.05, the classification accuracy change does not reach the threshold at this time, and the complexity of the simulated interference features needs to be increased. For example, the original simulated interference features have only 3 dimensions, which can be increased to 5 dimensions, or the distribution form can be changed from uniform distribution to normal distribution, etc. Then, the new simulated interference feature set is injected into the training data set, and the incremental learning strategy is used for retraining. The incremental learning strategy is to adjust the splitting criteria and feature weight parameters of the decision tree nodes based on the new simulated interference features while preserving the original classification knowledge. For example, when splitting the decision tree node, consider the influence of the new interference features on the splitting, and adjust the splitting conditions; for the feature weight parameters, redistribute the weights according to the importance of each feature in the new data. By continuously iterating the injection and training operations, the classification accuracy change reaches the preset optimization threshold.

[0122] As an implementation, the method further comprises the process of environmental calibration, specifically comprising steps S600-S900:

[0123] Step S600: Deploy reference scatterers in the target area, which have known electromagnetic scattering characteristics and spatial position information.

[0124] The reference scatterer is an object used for environmental calibration, whose electromagnetic scattering characteristics (such as scattering coefficient, polarization characteristics, etc.) and spatial position information are known. The purpose of deploying reference scatterers is to establish a known electromagnetic scattering reference point in the target area, so as to calibrate the subsequent radar detection data.

[0125] When deploying reference scatterers in the target area, the number and position of reference scatterers need to be determined according to the size, shape of the target area and the requirements of radar detection. For example, in a larger target area, reference scatterers can be deployed at the four corners and the center of the area. Reference scatterers can be selected to have stable electromagnetic scattering characteristics, such as metal balls, corner reflectors, etc. During deployment, the spatial position information of the reference scatterers needs to be accurately recorded, which can be positioned using global positioning system (GPS) and other devices.

[0126] Step S700: Collect the calibration signal sequence of the reference scatterer at multiple observation angles, extract the calibration feature vector and construct the environmental disturbance compensation database.

[0127] The calibration signal sequence is a sequence of electromagnetic signals received by the radar at different observation angles of a reference scatterer. The calibration feature vector is a representative feature vector extracted from the calibration signal sequence, which contains electromagnetic scattering feature information of the reference scatterer in a preset environment. The environmental disturbance compensation database is a database storing the calibration feature vectors and related environmental information, which is used for subsequent environmental disturbance compensation of actual detection data.

[0128] The calibration signal sequence of the reference scatterer at multiple observation angles can be collected by rotating the radar antenna or moving the reference scatterer to change the observation angle. For example, the radar antenna is rotated at a set angle interval (e.g., 10 degrees), and a calibration signal sequence is collected at each angle for a period of time. Then, the collected calibration signal sequence is feature extracted, and the extracted features can include signal energy, frequency distribution, polarization characteristics, etc. These features are combined into a calibration feature vector. At the same time, the environmental information when the calibration signal sequence is collected is recorded, such as atmospheric conditions, surrounding object distribution, etc. The calibration feature vector and the corresponding environmental information are stored in the environmental disturbance compensation database for subsequent use.

[0129] Step S800: In the actual detection process, the current environmental parameters are matched with the calibration feature vectors in the environmental disturbance compensation database in real time, and the environmental disturbance correction coefficient is calculated; wherein the real-time matching includes: using the nearest neighbor search algorithm to find the calibration sample closest to the current environmental parameters, and generating a continuous correction coefficient curve through interpolation calculation.

[0130] The current environmental parameters are environmental related parameters acquired in real time in the actual detection process, such as atmospheric humidity, temperature, electromagnetic noise intensity, etc. The environmental disturbance correction coefficient is a coefficient used to compensate for the influence of environmental disturbance on radar detection data. The nearest neighbor search algorithm is an algorithm for finding the nearest sample to the given data in the database.

[0131] In the actual detection process, the current environmental parameters are first acquired in real time. Then, the nearest neighbor search algorithm is used to find the calibration sample closest to the current environmental parameters in the environmental disturbance compensation database. For example, the distance (such as Euclidean distance) between the current environmental parameters and the environmental parameters of each calibration sample in the database is calculated, and the calibration sample with the smallest distance is selected as the nearest neighbor sample. Then, according to the calibration feature vectors of the nearest neighbor sample and its adjacent samples, a continuous correction coefficient curve is generated through interpolation calculation. For example, a linear interpolation method is used to calculate the correction coefficient under the current environmental parameters according to the correction coefficient values of the nearest neighbor sample and the adjacent samples. Through this real-time matching and interpolation calculation, the environmental disturbance correction coefficient is obtained.

[0132] Step S900: Apply the environmental disturbance correction coefficient to the fusion feature vector set to eliminate feature distortion caused by atmospheric attenuation and multipath effect.

[0133] Atmospheric attenuation is a phenomenon that electromagnetic signals weaken in intensity due to absorption, scattering, and other reasons when propagating in the atmosphere. Multipath effect is a phenomenon that electromagnetic signals encounter multiple reflectors during propagation, causing the signals to pass through multiple paths to reach the radar receiving device, thus producing signal distortion. Feature distortion is caused by environmental factors such as atmospheric attenuation and multipath effect, which changes the eigenvalues of the fusion feature vector and affects the accuracy of the classification result.

[0134] The specific way to apply the environmental disturbance correction coefficient to the fusion feature vector set is to multiply each feature vector in the fusion feature vector set by the corresponding environmental disturbance correction coefficient. For example, if the fusion feature vector is [x1, x2, x3] and the environmental disturbance correction coefficient is k, then the corrected fusion feature vector is [kx1, kx2, kx3]. In this way, the fusion feature vector set is corrected to eliminate feature distortion caused by atmospheric attenuation and multipath effect, improving the accuracy of subsequent classification.

[0135] As an implementation, the method further includes a process of classification result verification, which can specifically include steps S1000-S1200:

[0136] Step S1000: Establish a multi-source information fusion verification platform to access optical remote sensing data, infrared detection data, and geographic information system data.

[0137] The multi-source information fusion verification platform is a platform for integrating multiple data sources. It can fuse and analyze different types of data to verify the accuracy of radar classification results. Optical remote sensing data is image data of the target area obtained by optical remote sensing equipment (such as satellites, drones, etc.), which can provide information such as the appearance and shape of the target. Infrared detection data is infrared radiation information of the target obtained by infrared detection equipment, which can reflect the temperature distribution and other characteristics of the target. Geographic information system data is data containing geographic information of the target area (such as terrain, topography, building distribution, etc.).

[0138] Establishing a multi-source information fusion verification platform requires building the corresponding hardware and software environment. In terms of hardware, servers, storage devices, etc. are needed. In terms of software, data access interfaces, data fusion algorithms, etc. need to be developed. When accessing optical remote sensing data, infrared detection data, and geographic information system data, ensure that the format and quality of the data meet the requirements of the platform. For example, for optical remote sensing data, image preprocessing such as denoising and correction should be performed to improve data quality.

[0139] Step S1100: Spatial position matching and attribute consistency checking of the category identification of the target scatterer with the associated data in the multi-source information fusion verification platform.

[0140] Spatial position matching is to compare the spatial position information of the target scatterer with the spatial position information of the associated data in the multi-source information fusion verification platform, to ensure that they correspond in space. Attribute consistency checking is to check whether the attributes corresponding to the category identification of the target scatterer are consistent with the attributes of the target in the associated data. For example, if the category identification of the target scatterer in the radar classification result is "aircraft", it is checked whether the target at the corresponding position in the optical remote sensing data and the infrared detection data has the attributes of an aircraft (such as shape, infrared radiation characteristics, etc.).

[0141] When performing spatial position matching, geographic information system (GIS) technology can be used to unify the spatial position information of the target scatterer and the spatial position information of the associated data to the same coordinate system, and then compare them. For attribute consistency checking, an attribute feature library can be established to store the attribute features of different categories of targets. For example, the attribute features of an aircraft can include fuselage shape, wing structure, target infrared radiation pattern, etc. The category identification of the target scatterer is compared with the attribute feature library, and at the same time, the associated data in the multi-source information fusion verification platform is combined to check whether the attributes are consistent.

[0142] Step S1200: When attribute conflict is detected between the classification result and the multi-source data, a classification backtracking mechanism is triggered to re-execute the feature extraction and classification decision operation; wherein the execution steps of the classification backtracking mechanism include adjusting the weight distribution strategy in the electromagnetic feature association rule according to the attribute conflict type, and performing credibility weighted voting decision on the historical classification results.

[0143] Attribute conflict is a situation where the attributes corresponding to the category identification of the target scatterer are inconsistent with the attributes reflected by the associated data in the multi-source information fusion verification platform. Classification backtracking mechanism is a mechanism for handling classification errors, which can re-execute feature extraction and classification decision operation to improve the accuracy of classification.

[0144] When a property conflict between the classification result and the multi-source data is detected, the type of the property conflict is first analyzed. For example, is it a shape property conflict, a temperature property conflict, or other property conflicts. Then, the weight distribution strategy in the electromagnetic feature association rule is adjusted according to the type of the property conflict. If it is a shape property conflict, the weight of the feature parameter related to the shape may need to be increased. Next, the historical classification results are subjected to a credibility weighted voting decision. For each historical classification result, a credibility weight is assigned according to factors such as its classification confidence and relevance to the current situation. For example, a historical classification result with high classification confidence and similar to the current environment is given a higher credibility weight. All historical classification results are subjected to weighted voting, and the classification result with the highest votes is selected as the new classification result. Then, the feature extraction and classification decision operation is re-executed using the adjusted weight distribution strategy to obtain a more accurate classification result.

[0145] It can be understood that the various algorithms involved in the above introduction of the embodiments of the present application, such as the adaptive filtering algorithm, the wavelet decomposition algorithm, the weight compensation algorithm, etc., can be known from the related content in the prior art. In order to save space, the above-mentioned algorithms will not be expanded in the embodiments of the present application. In addition, those skilled in the art can supplement the details according to the common knowledge in the art when implementing the scheme of the present application. For example, according to the common knowledge in the art, the dimensional conflict before feature fusion can be eliminated by normalization, the dimension difference can be eliminated by interpolation, the threshold can be reasonably set by combining historical data, experience or business scene requirements, the model can be trained based on a general model training method, etc. The present application will not make redundant introduction to the too detailed implementation process.

[0146] Please refer to Figure 2 , Figure 2A structural schematic diagram of a computer system provided by an embodiment of the present application is shown in FIG. 1. The computer system includes at least a processor 101, a communication interface 102, and a memory 103. The processor 101, the communication interface 102, and the memory 103 can be connected by a bus or other means. The processor 101 (also referred to as a central processing unit (CPU)) is the computing core and control core of the computer system, which can parse various instructions in the computer system and process various data of the computer system. The communication interface 102 can optionally include a standard wired interface, a wireless interface (such as WI-FI, a mobile communication interface, etc.), and can be used to send and receive data under the control of the processor 101; the communication interface 102 can also be used for transmission and interaction of data within the computer system. The memory 103 is a memory device in the computer system, used to store programs and data. It can be understood that the memory 103 can include a built-in memory of the computer system, and of course can also include an extended memory supported by the computer system. The memory 103 provides a storage space that stores an operating system of the computer system, which can include but is not limited to: an Android system, an iOS system, a Windows Phone system, etc., and the present application is not limited thereto.

[0147] In one embodiment, the processor 101 executes the computer program in the memory 103 to perform the passive sensing radar target classification method based on electromagnetic detection feature extraction provided by the above embodiment of the present application.

Claims

1. A passive sensing radar target classification method based on electromagnetic detection feature extraction, characterized in that, include: Acquire a set of electromagnetic radiation signals for the target area, wherein the set of electromagnetic radiation signals includes sequences of raw scattered signals acquired within multiple signal acquisition cycles; The original scattering signal sequence is subjected to multi-dimensional signal decomposition processing to obtain the time-frequency distribution feature set and polarization response feature set corresponding to each signal acquisition cycle; According to the preset electromagnetic feature association rules, the target electromagnetic feature combination is extracted from the time-frequency distribution feature set and the polarization response feature set, and weight allocation parameters are generated. The target electromagnetic feature combination is weighted and fused based on the weight allocation parameters to generate a fused feature vector set, and the fused feature vector set is input into an adaptive classification network. The adaptive classification network performs hierarchical classification on the fused feature vector set, outputs the category identifier of the target scatterer and the corresponding classification confidence score, and generates radar detection optimization instructions based on the classification confidence score; Based on preset electromagnetic feature association rules, target electromagnetic feature combinations are extracted from the time-frequency distribution feature set and the polarization response feature set, and weight allocation parameters are generated, including: Construct a feature correlation evaluation matrix, and calculate the mutual information and covariance correlation coefficient between each feature parameter in the time-frequency distribution feature set and each feature parameter in the polarization response feature set; A feature redundancy index is generated based on the mutual information and the covariance correlation coefficient, and a redundancy filtering operation is performed on the time-frequency distribution feature set and the polarization response feature set based on the feature redundancy index to obtain a candidate feature subset; Sensitivity analysis is performed on the candidate feature subset to obtain the discrimination weight of each candidate feature to the target scatterer category, and an initial feature weight set is generated based on the discrimination weight. Based on the physical property constraints of the target scatterer, the initial feature weight set is adjusted to eliminate weight deviation caused by environmental noise and generate the weight allocation parameters; wherein, the adjustment includes: introducing an external calibration signal source to verify the correlation of the physical properties of the candidate feature subset, and when the feature weight is detected to deviate from the standard weight of the calibration signal source by more than a preset threshold, a weight compensation algorithm is started to correct the deviation.

2. The method according to claim 1, characterized in that, The original scattering signal sequence is subjected to multi-dimensional signal decomposition processing to obtain the time-frequency distribution feature set and polarization response feature set corresponding to each signal acquisition cycle, including: A time-domain alignment operation is performed on the original scattered signal sequence to eliminate signal timing offsets caused by environmental interference and generate a standardized time-domain signal sequence. The standardized time-domain signal sequence is divided into frequency bands using a multi-resolution orthogonal decomposition algorithm. Signal energy distribution parameters and phase continuity parameters within each frequency band are extracted to form an initial time-frequency feature set. Simultaneously, the original scattering signal sequence is subjected to polarization channel separation processing to obtain the scattering intensity difference parameters and polarization coherence parameters under different polarization directions, forming an initial polarization response set; The initial time-frequency feature set and the initial polarization response set are subjected to nonlinear normalization processing to eliminate dimensional differences and retain feature distribution patterns, thereby generating the time-frequency distribution feature set and the polarization response feature set. The nonlinear normalization processing includes: selecting the corresponding normalization function according to the statistical distribution characteristics of each feature parameter, performing truncation compensation operation on feature values ​​that exceed a preset fluctuation threshold, and retaining the shape of the normalized feature distribution curve.

3. The method according to claim 1, characterized in that, The target electromagnetic feature combination is weighted and fused based on the weight allocation parameters to generate a fused feature vector set, including: A feature fusion equation is constructed based on the weight allocation parameters, and each feature parameter in the time-frequency distribution feature set is weighted and superimposed with each feature parameter in the polarization response feature set. Energy consistency detection is performed on the weighted superposition of candidate feature combinations to eliminate abnormal feature combinations caused by weight mismatch and retain valid feature combinations that satisfy the energy conservation criterion. The effective feature combination is subjected to dimensionality compression processing, and the principal component preservation algorithm is used to map the high-dimensional feature space to the low-dimensional feature space to generate a set of dimensionality-reduced feature vectors. The reduced feature vector set is serialized and recombined to generate the fused feature vector set with spatiotemporal correlation characteristics; The execution steps of the principal component preservation algorithm include: calculating the information entropy ratio of each effective feature combination before and after dimensionality reduction; and when the information entropy ratio is lower than a preset information loss threshold, enabling alternative dimensionality reduction channels to compensate for the amount of information loss.

4. The method according to claim 3, characterized in that, The adaptive classification network performs hierarchical classification on the fused feature vector set, outputting the category identifier of the target scatterer and the corresponding classification confidence score, including: Construct a multi-level classification decision tree, in which each non-leaf node is configured with a feature discrimination model based on electromagnetic scattering mechanism, and each leaf node is associated with a preset category label of the target scatterer; The fused feature vector set is input into the root node of the multi-level classification decision tree, and the matching degree between the fused feature vector and the feature template of each child node is calculated according to the feature discrimination model corresponding to the current node. Select the child node with the highest matching degree as the next level decision node, and iteratively execute the matching degree calculation and node selection operations until the leaf node is reached; The classification confidence is calculated based on the historical classification accuracy and current matching deviation of the leaf nodes, and a classification result set is generated by associating the preset category identifiers. The calculation steps of the feature discrimination model include: adopting a template update mechanism to adjust the matching threshold of the feature template based on real-time classification results to ensure the adaptability of the classification decision tree to time-varying electromagnetic environments.

5. The method according to claim 4, characterized in that, The method also includes an optimization process for the classification model, including: Collect misclassified samples from the historical classification result set, and extract the abnormal fusion feature vector and environmental interference parameters corresponding to the misclassified samples; An interference feature simulator is constructed based on the environmental interference parameters to generate a set of simulated interference features that have similar statistical properties to the anomaly fusion feature vector. The simulated interference feature set is injected into the training dataset of the adaptive classification network to retrain the feature discrimination model in the multi-level classification decision tree; The optimization effect is evaluated by comparing the change in classification accuracy before and after retraining. If the change in classification accuracy does not reach the preset optimization threshold, the complexity of the simulated interference features is increased and the injection and training operations are performed iteratively. The retraining steps include: using an incremental learning strategy to retain the original classification knowledge, while adjusting the splitting criteria and feature weight parameters of the decision tree nodes according to simulated interference features.

6. The method according to claim 5, characterized in that, Retraining the feature discrimination model in the multi-level classification decision tree includes: Perform node vulnerability analysis on the multi-level classification decision tree to identify the set of decision nodes that are susceptible to interference features; For each vulnerable decision node, an adversarial training sample set is generated, which contains adversarial feature vectors generated by perturbing the original feature vectors; The adversarial training sample set is merged with the simulated interference feature set to construct an enhanced training dataset; The elastic weight solidification algorithm is used to locally retrain the feature discrimination model corresponding to the fragile decision node to prevent overfitting and maintain global classification stability. The execution steps of the elastic weight solidification algorithm include: calculating the importance weights of the model parameters, limiting the adjustment range of important parameters, and allowing minor parameters to be updated according to new training data.

7. The method according to claim 1, characterized in that, Before acquiring the electromagnetic radiation signal set of the target area, a signal preprocessing step is also included: Multiple passive radar receiving nodes are deployed, and the frequency band coverage and polarization receiving mode of each receiving node are configured to form a spatially distributed detection array; Synchronize the local clock source of each receiving node to eliminate signal phase error caused by time asynchrony; The raw signals acquired by each receiving node are subjected to local oscillator leakage suppression processing, and an adaptive filtering algorithm is used to eliminate the equipment's own radiation interference. The execution steps of the adaptive filtering algorithm include: establishing a transfer function model of the hardware characteristics of the receiving node, and canceling the leakage signal components through reverse compensation operation.

8. The method according to claim 7, characterized in that, The method also includes an environmental calibration process, including: A reference scatterer is deployed in the target area, the reference scatterer having known electromagnetic scattering characteristics and spatial location information; The calibration signal sequence of the reference scatterer is collected at multiple observation angles, the calibration feature vector is extracted, and an environmental disturbance compensation database is constructed. During the detection process, the current environmental parameters are matched with the calibration feature vectors in the environmental disturbance compensation database in real time, and the environmental disturbance correction coefficients are calculated. The real-time matching includes: using the nearest neighbor search algorithm to find the calibration sample that is closest to the current environmental parameters, and generating a continuous correction coefficient curve by interpolation. The environmental disturbance correction coefficients are applied to the fused feature vector set to eliminate feature distortion caused by atmospheric attenuation and multipath effects.

9. A computer system, characterized in that, include: A memory, wherein a computer program is stored; A processor is configured to load the computer program to implement the passive sensing radar target classification method based on electromagnetic detection feature extraction as described in any one of claims 1-8.

Images