A radar space target recognition method, a terminal device and a medium

Abstract

The application provides a radar space target identification method, a terminal device and a medium, comprising: obtaining radar echo data of a target to be measured; dividing the radar echo data by using a time sliding window to obtain a plurality of radar echo data segments; determining at least one effective polarization channel data from the plurality of radar echo data segments based on a Fisher separability criterion; inputting each effective polarization channel data into a pre-trained identification network model respectively to obtain an identification result of the target to be measured output by the identification network model. The application can improve the accuracy and efficiency of radar space target identification.

Description

Technical Field

[0001] This invention belongs to the field of radar space target recognition technology, specifically relating to a radar space target recognition method, terminal equipment, and medium. Background Technology

[0002] Existing radar space target recognition methods are mainly divided into methods based on physical feature extraction and classifier design, and deep learning methods based on deep neural network feature extraction and recognition.

[0003] Traditional radar target recognition methods first extract target features from radar detection data, and then design a suitable classifier to achieve target recognition. For the diverse data acquired by radar, numerous researchers have proposed various recognition methods with clear physical meanings based on the target's geometric structure, scattering characteristics, and other essential physical properties.

[0004] The radar cross section (RCS) contains a wealth of target feature information, including key parameters such as target size and surface material properties, and is widely used in target detection and identification tasks in traditional narrowband radar systems.

[0005] With the continuous advancement of radar information processing technology, high-resolution imaging and polarization scattering characteristic characterization have become important research directions in the field of modern radar target recognition. High-resolution one-dimensional range profiles (HRRP) can effectively characterize the electromagnetic scattering distribution characteristics of targets, and have advantages such as simple data acquisition and high real-time processing efficiency.

[0006] Target features extracted using traditional radar target recognition methods possess clear physical meaning, exhibiting strong interpretability and maintaining relatively stable recognition performance even with small sample sizes. However, the feature design of traditional methods relies on prior expert knowledge, which may lead to feature failure when novel or unknown targets are encountered. Furthermore, under low signal-to-noise ratio conditions, the data dimensionality reduction process based on physical feature extraction may amplify noise interference, affecting feature representation and thus reducing the robustness of traditional methods.

[0007] In recent years, deep learning technology has been widely used in the field of radar target recognition due to its powerful feature learning capabilities. Compared with traditional methods based on physical feature extraction and classifier design, deep learning methods, through an end-to-end learning mechanism, can adaptively extract discriminative deep feature representations from raw radar data. Based on recognition models such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTMs), deep learning methods have significantly improved the noise robustness and recognition performance of radar target recognition systems in complex environments.

[0008] Deep learning-based radar target recognition methods can automatically extract features and have strong noise robustness, outperforming traditional methods under large-scale sample conditions. However, deep learning methods typically have a large number of network parameters and high training costs, leading to a significant drop in recognition performance under small sample conditions. Furthermore, deep learning models are black-box characteristics, lacking interpretability in their feature extraction and decision-making processes. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to provide a radar space target identification method, terminal equipment and medium to improve the accuracy and efficiency of radar space target identification.

[0010] In a first aspect, the present invention provides a radar spatial target identification method, the method comprising the following steps:

[0011] Acquire radar echo data of the target under test; radar echo data includes fully polarized RCS data and fully polarized HRRP data;

[0012] The radar echo data is divided into multiple radar echo data segments by using a time sliding window; each radar echo data segment contains data from all polarization channels.

[0013] At least one effective polarization channel data is determined from multiple radar echo data segments based on Fisher's separability criterion.

[0014] Each effective polarization channel data is input into a pre-trained recognition network model to obtain the recognition result of the target object output by the recognition network model. The recognition network model includes a branch module and a fusion module connected in sequence. The branch module includes a first branch and a second branch, both of which are input to the effective polarization channel data. The first branch is used to extract the physical features of the effective polarization channel data and recognize these physical features to obtain a first recognition result. The second branch is used to extract the deep features of the effective polarization channel data and recognize these deep features to obtain a second recognition result. The fusion module is used to perform dynamic weighted fusion of the first and second recognition results based on entropy values ​​to output the recognition result. Physical features represent the physical properties of the corresponding target, including but not limited to geometric shape and material; deep features characterize complex nonlinear relationships.

[0015] Optionally, the fully polarized RCS data includes two co-polarized channel RCS data and two cross-polarized channel RCS data.

[0016] Optionally, at least one effective polarization channel data is determined from multiple radar echo data segments based on the Fisher separability criterion, including:

[0017] Based on Fisher's separability criterion, the separability values ​​corresponding to each polarization channel data in each radar echo data segment are calculated.

[0018] Polarization channel data with a separability value greater than a preset separability threshold are considered valid polarization channel data.

[0019] Optional physical features include RCS features, HRRP features, and narrowband polarization features;

[0020] RCS features include location features, scattering features, transformation features, and distribution features. Location features include the maximum, minimum, mean, truncated mean, median, and mode of the RCS data. Scattering features include the range, variance, third central moment, fourth central moment, and coefficient of variation of the RCS data. The third central moment of the RCS indicates the skewness of the scattering distribution; positive skewness indicates the presence of strong scattering points. The fourth central moment of the RCS indicates the kurtosis of the scattering; high values ​​correspond to targets with multiple scattering centers. The coefficient of variation of the RCS indicates the degree of normalized fluctuation, used to compare targets of different sizes. Transformation features include the mean of the spectrum, the mean of the power spectrum, and the mean of the correlation coefficient. Distribution features include the kurtosis coefficient and the skewness coefficient.

[0021] HRRP features include the first, second, and third central moments of HRRP data, descaled structural features, the distance of the highest peak relative to the leftmost peak, the distance of the highest peak relative to the rightmost peak, the symmetry of target scattering, the dispersion of target scattering, and the number of strong scattering centers of the target. Among them, the first central moment of HRRP data represents the centroid shift of the scattered energy distribution, the second central moment of HRRP data represents the dispersion of scattering points in the range dimension, and the third central moment of HRRP data represents the morphological asymmetry of the range profile.

[0022] Narrowband polarization characteristics include the trace of the power matrix, the determinant of the scattering matrix, the depolarization coefficient, the intrinsic polarization ellipticity, and the intrinsic polarization direction angle.

[0023] Optionally, after extracting the physical features of the effective polarization channel data and before identifying the physical features, the first branch also includes:

[0024] A multi-criteria fusion feature selection strategy based on entropy weight method is used to evaluate the separability of physical features and obtain the separability score of physical features.

[0025] Physical features with a separability score greater than a preset score threshold are identified as the physical features to be identified.

[0026] Optionally, the formula for calculating the separability score is as follows:

[0027]

[0028] Where EWFRM represents the separability score, j represents the j-th separability evaluation algorithm, j = 1, 2, 3 represent Fisher's algorithm, Relief algorithm, or MRMR algorithm, respectively, and ω j S′ represents the weight of the j-th separability evaluation algorithm. ij Let represent the separability score of the j-th separability evaluation algorithm after normalization for the i-th physical feature, where i = 1, 2, ..., N, and N represents the total number of physical features. j p represents the entropy value of the j-th separability evaluation algorithm. ij S represents the weight of the j-th separability evaluation algorithm. ij S represents the separability score of the j-th separability evaluation algorithm for the i-th physical feature. j Let represent the set of separability scores for all physical features by the j-th separability evaluation algorithm.

[0029] Optionally, the second branch includes an RCS branch for processing fully polarized RCS data and an HRRP branch for processing fully polarized HRRP data; wherein the RCS branch uses one-dimensional convolution and pooling processing, and the HRRP branch uses two-dimensional convolution and pooling processing.

[0030] Optionally, the fusion module includes multiple recognizers;

[0031] The first and second recognition results are dynamically weighted and fused based on entropy values ​​to output the recognition results, including:

[0032] The recognition capability of each recognizer is measured by its entropy value.

[0033] Based on the recognition capabilities, calculate the fusion weight coefficient of each recognizer during the dynamic weighted fusion process;

[0034] The first and second recognition results are fused according to the fusion weight coefficient to obtain the recognition result.

[0035] In a second aspect, the present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method.

[0036] Thirdly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0037] The beneficial effects of this invention are:

[0038] The radar spatial target recognition method provided by this invention overcomes the limitations of a single data source by acquiring the full polarization RCS and HRRP data of the target, thus improving the accuracy of radar spatial target recognition. Based on the Fisher separability criterion, it selects effective polarization channel data, reducing data redundancy and improving data quality, thereby increasing the efficiency of radar spatial target recognition. The first branch of the recognition network model extracts interpretable physical features, ensuring stability in small sample scenarios, while its second branch automatically learns noise robustness features to adapt to complex electromagnetic environments, improving the accuracy of radar spatial target recognition. Based on entropy-weighted fusion of the two branch results, the high-confidence branch has a larger weight, reducing the false positive rate and effectively improving the accuracy of radar spatial target recognition. Attached Figure Description

[0039] Figure 1 This is a flowchart of a radar spatial target identification method in one embodiment of this application;

[0040] Figure 2 This is a schematic diagram of radar echo data in one embodiment of this application, wherein, Figure 2 (a) represents the fully polarized RCS data. Figure 2 (b) represents fully polarized HRRP data;

[0041] Figure 3This is a graph showing the trend of Fisher separability values ​​of fully polarized RCS data at different flight times in one embodiment of this application.

[0042] Figure 4 This is a graph showing the variation of Fisher separability values ​​of fully polarized HRRP data at different flight times in one embodiment of this application.

[0043] Figure 5 This is a schematic diagram of the structure of the identification network model in one embodiment of this application;

[0044] Figure 6 This is a schematic diagram of the structure of the second branch in the network model identified in one embodiment of this application;

[0045] Figure 7 This is a schematic diagram of the confusion matrix and t-SNE visualization results of the first branch in one embodiment of this application, wherein, Figure 7 (a) is the recognition structure confusion matrix when the SNR is -20dB. Figure 7 (b) is a schematic diagram of the t-SNE visualization results when the SNR is -20dB. Figure 7 (c) is the recognition structure confusion matrix when the SNR is -4dB. Figure 7 (d) is a schematic diagram of the t-SNE visualization results when the SNR is -4dB;

[0046] Figure 8 This is a schematic diagram of the confusion matrix and t-SNE visualization results of the second branch in one embodiment of this application, wherein, Figure 8 (a) is the recognition structure confusion matrix when the SNR is -20dB. Figure 8 (b) is a schematic diagram of the t-SNE visualization results when the SNR is -20dB. Figure 8 (c) is the recognition structure confusion matrix when the SNR is -10dB. Figure 8 (d) is a schematic diagram of the t-SNE visualization results when the SNR is -10dB;

[0047] Figure 9 This is a graph showing the trend of the second branch recognition accuracy as a function of the number of iterations in one embodiment of this application.

[0048] Figure 10 This is a graph showing the trend of the loss value of the second branch as a function of the number of iterations in one embodiment of this application.

[0049] Figure 11 This is a graph showing the overall accuracy of different decision-level fusion methods as a function of signal-to-noise ratio in one embodiment of this application.

[0050] Figure 12This is the confusion matrix for dual-branch fusion identification in one embodiment of this application;

[0051] Figure 13 This is a comparison chart of the overall accuracy of different decision-level fusion methods under different training sample ratios in one embodiment of this application;

[0052] Figure 14 This is a graph showing the change in recognition accuracy as a function of SNR for different methods in one embodiment of this application.

[0053] Figure 15 The curves showing the recognition accuracy as a function of SNR for different feature fusion schemes in one embodiment of this application are shown.

[0054] Figure 16 This is a graph showing the network recognition accuracy as a function of SNR under different signal-to-noise ratios in one embodiment of this application.

[0055] Figure 17 This is a schematic diagram of the structure of a terminal device in one embodiment of this application. Detailed Implementation

[0056] Current radar space target identification methods face key challenges such as a lack of training samples for non-cooperative targets, low efficiency of multi-source data fusion, and insufficient interpretability of identification models.

[0057] To address the aforementioned challenges, this invention provides a radar spatial target identification method, terminal device, and medium. By acquiring the fully polarized RCS and HRRP data of the target under test, the limitations of a single data source can be overcome, which is beneficial to improving the accuracy of radar spatial target identification. Selecting effective polarization channel data based on the Fisher separability criterion can reduce data redundancy and improve data quality, thus improving the efficiency of radar spatial target identification. The first branch of the identification network model extracts interpretable physical features, ensuring stability in small sample scenarios, while its second branch can automatically learn noise robustness features to adapt to complex electromagnetic environments, improving the accuracy of radar spatial target identification. Based on entropy-weighted fusion of the two branch results, the high-confidence branch has a larger weight, reducing the false positive rate and effectively improving the accuracy of radar spatial target identification.

[0058] The radar space target identification method provided by the invention will be described below with reference to specific embodiments.

[0059] like Figure 1 As shown, the radar space target identification method provided by the invention includes steps 11 to 14.

[0060] Step 11: Obtain radar echo data of the target under test.

[0061] It should be noted that when the distance between the radar and the target is relatively large, it can be considered as the far-field region. In this case, both the incident and scattered waves can be considered plane waves, and the target scattering process can be considered a linear process. In the corresponding scattering coordinate system, there is a linear transformation relationship between the polarization components of the incident and scattered waves, which can be characterized by a two-dimensional complex matrix called the polarization scattering matrix. The polarization scattering matrix is ​​usually represented by a symbol and is related to factors such as the target's shape, size, structure, material, and attitude, as well as the radar frequency, containing information about the overall scattering characteristics of the target. Therefore, compared to single-polarization and dual-polarization radar echoes, fully polarized echoes can more comprehensively describe the target's electromagnetic scattering characteristics and extract more refined information from the target echo, thereby improving the performance of the identification system.

[0062] Based on the above analysis, in this embodiment of the invention, the radar echo data includes fully polarized RCS data and fully polarized HRRP data. Specifically, the fully polarized RCS data includes two co-polarized channel RCS data and two cross-polarized channel RCS data. It can be represented as RCS-HH, RCS-VV, RCS-HV, and RCS-VH; among them, RCS-HH and RCS-VV are co-polarized channel RCS data, that is, both the transmit polarization and the receive polarization are horizontal or vertical. RCS-HH is suitable for symmetrical structures (such as metal planes, corner reflectors) or targets extending in the horizontal direction (such as ground vehicles), while RCS-VV is suitable for vertical structures (such as masts, trees) or targets extending in the vertical direction (such as buildings); RCS-HV and RCS-VH are cross-polarized channel RCS data; the transmit pole of RCS-HV is horizontally polarized and the receive pole is vertically polarized, which is suitable for detecting complex structures (such as vegetation, rough surfaces) or targets with significant depolarization effects (such as stealth materials); the transmit pole of RCS-VH is vertically polarized and the receive pole is horizontally polarized, which is usually used to verify polarization symmetry.

[0063] It should be understood that the range resolution ΔR of a radar reflects its ability to distinguish adjacent targets in a multi-target environment. It is primarily affected by the size of the target and the bandwidth of the radar's transmitted signal. The calculation formula is as follows:

[0064]

[0065] Where c is the speed of light, τ is the pulse width of the signal, and B is the bandwidth of the signal.

[0066] Let the size of the target along the radar's line-of-sight be L. When ΔR >> L, the echo signal is considered a point target echo, and the radar system is classified as a low-resolution radar. Conversely, when ΔR << L, the echo signal exhibits the target's characteristics extended in the range dimension, forming a "one-dimensional range profile," and such radar systems are classified as high-resolution radars. Compared to low-resolution radars, high-resolution radars contain richer target feature information, which is beneficial for improving the accuracy of target identification.

[0067] To acquire HRRP (High-Resolution One-Dimensional Range Profile) data, the radar system needs to operate in the high-frequency range, where ΔR << L. The scattering point model of the target can be considered as a set of multiple independent scattering centers (strong reflection points). Superimposing the echo vectors from these scattering centers yields the high-resolution one-dimensional range profile of the target.

[0068] In the absence of noise, the echo h of the i-th distance cell i It can be expressed as the superposition of all scattering center echoes within that range cell, and its calculation formula is:

[0069]

[0070] Where, N i Let a be the number of target scattering centers in the i-th range cell. i,k Let f be the scattering intensity at the k-th scattering center within the i-th range cell, and t be the radar operating frequency. i,k Let I(i) be the arrival time of the echo from the k-th scattering center within the i-th distance cell, where I(i) represents the in-phase component of the echo signal, Q(i) represents the quadrature component of the echo signal, and j″ represents the imaginary unit.

[0071] In one feasible implementation, the acquired fully polarized RCS data and fully polarized HRRP data are respectively as follows: Figure 2 (a) Figure 2 As shown in (b) Figure 2 In (a), the horizontal axis represents the polarization mode: VV, VH, HV, HH, and the time: 950s-960s. The vertical axis represents the normalized amplitude of the RCS. Figure 2 In (b), the horizontal axis represents time: 950s-960s, the distance unit: 0-100, and the vertical axis represents the normalized magnitude of HRRP.

[0072] Step 12: Divide the radar echo data using a time sliding window to obtain multiple radar echo data segments.

[0073] In one feasible implementation, the target to be tested is a flying intelligent agent. The flying intelligent agent flies for a total distance of 0 to 2000 seconds. The data is divided into 20 radar echo data segments at equal intervals, with a sliding window and sliding step size of 100 seconds.

[0074] In embodiments of the present invention, each radar echo data segment contains full polarization channel data, that is, each radar echo data segment contains HH polarization channel data (RCS-HH or HRRP-HH), HV polarization channel data (RCS-HV or HRRP-HV), VH polarization channel data (RCS-VH or HRRP-VH), and VV polarization channel data (RCS-VV or HRRP-VV).

[0075] Step 13: Determine at least one effective polarization channel data from multiple radar echo data segments based on Fisher's separability criterion.

[0076] To meet the real-time processing requirements of radar spatial target identification, this invention uses Fisher's separability criterion to quickly filter fully polarized RCS and fully polarized HRRP data to obtain the subset with the best separability.

[0077] The Fisher separability criterion is explained below.

[0078] Fisher's separability criterion evaluates feature separability by maximizing the ratio of between-class divergence to minimizing within-class divergence. In embodiments of the present invention, S represents the i-th sample of the c-th class (c = 1, 2, ..., N, where N = 4 in this embodiment, indicating that there are 4 possible classes of the target), and its within-class scatter matrix S. W Inter-class scatter matrix S B It can be represented as:

[0079]

[0080] Where, n c Let μ represent the number of samples for the c-th type of target, n represent the total number of samples, and μ represent the number of samples for the c-th type of target. c Let μ represent the mean of the target samples of class I, and μ represent the global mean. c The formulas for calculating μ are:

[0081] The criterion function for separability determination using the Fisher algorithm is: Where Tr(·) denotes finding the trace of the matrix. The denser the distribution of similar target samples in the feature space, the larger the within-class scatter matrix S. W The smaller the value, the greater the distance between different classes of target samples; the greater the distance between them, the smaller the inter-class scatter matrix S. B The larger the Fisher's ratio, the stronger the separability of the samples.

[0082] In one feasible implementation, the target flight time is 0 to 2000 seconds, and the entire data is divided into 20 segments at equal intervals with a sliding window and sliding step of 100 seconds.

[0083] The variation trend of Fisher separability values ​​in fully polarized RCS data at different flight periods is as follows: Figure 3 As shown. By Figure 3 It can be seen that the VH channel exhibits the best overall separability, with a peak value of 40.57 during the 1900-2000 second period. The 1600-2000 second period also shows high separability, containing multiple consecutive peaks. The HH polarization channel demonstrates the second best performance, with a separability value higher than the overall average for 50% of the flight time. The 700-800 second period represents the peak separability value for this channel at 14.45, while the 700-1100 second period shows continuous high separability. Considering the temporal continuity of space target flight, individual high-separability periods are easily affected by interference; continuous and stable high-separability intervals are more conducive to target identification. Therefore, considering the continuity of time periods, the magnitude of Fisher separability values, and the complementarity of polarization channels, RCS data from the 700-1100 second period of the HH polarization channel and the 1600-2000 second period of the VH polarization channel—two consecutive high-separability periods—are selected for target identification. The synergistic complementarity of the advantageous periods of the two channels improves the system's identification performance.

[0084] Table 2 shows that the average Fisher separability of the target's HRRP data across the four polarization channels throughout the entire flight is 1.33. Experimental data indicates that the Fisher separability value of the HH polarization channel is slightly higher than the overall average during the 1800-2000 second period, but lower than that of the VV channel during the same period. The HV channel exhibits high separability during the 1300-1600 second period, but its overall separability performance is worse than that of the VH channel during the same period. Therefore, the HH and HV polarization channels have limited ability to distinguish targets, and HRRP data from these two channels will not be used for target identification in this section.

[0085] The curves showing the variation of Fisher separability values ​​in fully polarized HRRP data at different flight stages are as follows: Figure 4 As shown. By Figure 4It can be seen that the VH channel exhibits the best overall separability, with a separability value higher than the overall average for 80% of the flight time, and continuous high separability during the 500-900 second period. The VV channel shows the second best separability, with the separability value continuously increasing from 1600 to 2000 seconds, reaching a peak of 5.95 during the 1900-2000 second period, playing a crucial role in the final stage of target recognition. Based on temporal continuity and maximizing separability, in this embodiment of the invention, HRRP data from the 500-900 second period of the VH polarization channel and the 1600-2000 second period of the VV polarization channel, two consecutive high separability periods, are comprehensively selected for target recognition. The 500-900 second period of the VH channel represents the stable flight phase of the target, providing a stable representation of HRRP; the peak value at the end of the VV channel corresponds to the abrupt change in target action. Combining the two can improve the anti-interference performance of the recognition system in complex environments.

[0086] Step 14: Input the data of each effective polarization channel into the pre-trained recognition network model to obtain the recognition result of the target to be tested output by the recognition network model.

[0087] like Figure 5 As shown, the recognition network model includes a branch module 501 and a fusion module 502 connected in sequence. Branch module 501 includes a first branch 501A and a second branch 501B. The inputs to both branches 501A and 501B are effective polarization channel data. The first branch 501A is used to extract the physical features of the effective polarization channel data and recognize these features to obtain a first recognition result. The second branch 501B is used to extract the deep features of the effective polarization channel data and recognize these deep features to obtain a second recognition result. The fusion module 502 is used to perform dynamic weighted fusion of the first and second recognition results based on entropy values, and output the recognition result.

[0088] The functional process of the first branch is described below, specifically including steps 51 to 54.

[0089] Step 51: Extract the physical features of the effective polarization channel data.

[0090] In this embodiment of the invention, the physical features include RCS features, HRRP features, and narrowband polarization features;

[0091] RCS features include location features, scattering features, transformation features, and distribution features. Location features include the maximum, minimum, mean, truncated mean, median, and mode of the RCS data. Scattering features include the range, variance, third central moment, fourth central moment, and coefficient of variation of the RCS data. The third central moment of the RCS indicates the skewness of the scattering distribution; positive skewness indicates the presence of strong scattering points. The fourth central moment of the RCS indicates the kurtosis of the scattering; high values ​​correspond to targets with multiple scattering centers. The coefficient of variation of the RCS indicates the degree of normalized fluctuation, used to compare targets of different sizes. Transformation features include the mean of the spectrum, the mean of the power spectrum, and the mean of the correlation coefficient. Distribution features include the kurtosis coefficient and the skewness coefficient.

[0092] HRRP features include the first, second, and third central moments of HRRP data, descaled structural features, the distance of the highest peak relative to the leftmost peak, the distance of the highest peak relative to the rightmost peak, the symmetry of target scattering, the dispersion of target scattering, and the number of strong scattering centers of the target. Among them, the first central moment of HRRP data represents the centroid shift of the scattered energy distribution, the second central moment of HRRP data represents the dispersion of scattering points in the range dimension, and the third central moment of HRRP data represents the morphological asymmetry of the range profile.

[0093] Narrowband polarization characteristics include the trace of the power matrix, the determinant of the scattering matrix, the depolarization coefficient, the intrinsic polarization ellipticity, and the intrinsic polarization direction angle.

[0094] Step 52: Using a multi-criteria fusion feature selection strategy based on entropy weight method, the separability of physical features is evaluated to obtain the separability score of physical features, and physical features with a separability score greater than a preset score threshold are identified as the physical features to be identified.

[0095] Specifically, the formula for calculating the separability score is as follows:

[0096]

[0097] Where EWFRM represents the separability score, j represents the j-th separability evaluation algorithm, j = 1, 2, 3 represent Fisher's algorithm, Relief algorithm, or MRMR algorithm, respectively, and ω j S′ represents the weight of the j-th separability evaluation algorithm. ij Let represent the separability score of the j-th separability evaluation algorithm after normalization for the i-th physical feature, where i = 1, 2, ..., N, and N represents the total number of physical features. j p represents the entropy value of the j-th separability evaluation algorithm. ij S represents the weight of the j-th separability evaluation algorithm. ij S represents the separability score of the j-th separability evaluation algorithm for the i-th physical feature.j Let represent the set of separability scores for all physical features by the j-th separability evaluation algorithm.

[0098] Step 53: After splicing the identified physical features, input them into the classifier to obtain the first identification result.

[0099] In one feasible implementation, the classifier is a random forest classifier. The random forest classifier outputs the class probability of the target to be tested, thus obtaining the first identification result.

[0100] The functional process of the second branch is explained below.

[0101] In this embodiment of the invention, the second branch utilizes a deep neural network to automatically mine deep features in the data, and achieves feature fusion recognition of multiple input branches based on a cross-modal fusion method. The specific structure is as follows: Figure 6 As shown. The second branch takes effective polarization channel data as input, including RCS and HRRP multi-input data branches under different polarization channels. Each branch contains a convolutional block, which consists of a convolutional layer, a ReLU activation function, and a batch normalization layer. The RCS branch uses one-dimensional convolution and pooling operations, while the HRRP branch uses two-dimensional convolution and pooling. Each convolutional block is followed by a max pooling layer to reduce the feature dimensionality, and a global average pooling layer is used to obtain the deep feature vectors of the multi-branch. The multi-branch features are concatenated across modalities in the second branch, and adaptive feature fusion is achieved through two fully connected layers. The first fully connected layer captures cross-modal correlation characteristics through a 100-dimensional feature map and introduces a batch normalization layer and a Dropout layer to improve the model's generalization ability. The second fully connected layer consists of 4 neurons, corresponding to the 4 possible types of the target to be tested. Finally, a Softmax classifier classifies the target to be tested, obtains the target class probability, and obtains the second recognition result.

[0102] In one feasible implementation, the model parameters in the second branch are shown in Table 1.

[0103] Table 1

[0104]

[0105] In Table 1, “Conv” represents a convolutional layer, “MP” represents a max pooling layer, “GAP” represents a global average pooling layer, “Concat” represents a concatenation layer, and “FC” represents a fully connected layer.

[0106] In this embodiment of the invention, the second branch uses the cross-entropy loss function and the Adaptive Moment Estimation (ADAM) optimizer for model training.

[0107] The fusion module is explained below.

[0108] In traditional decision-level fusion methods, the performance of weighted voting depends on the allocation of weights, and majority voting assumes that each classifier has the same classification ability. However, in complex spatial object recognition environments, the recognition ability of different classifiers for different samples fluctuates with changes in the environment, and assigning fixed weights to classifiers may lead to unsatisfactory fusion results.

[0109] In this embodiment of the invention, the fusion module includes multiple classifiers, including a random forest classifier and a Softmax classifier of a convolutional neural network. The functional process of the fusion module includes steps 41 to 43.

[0110] Step 41: Measure the recognition ability of each recognizer by the entropy value.

[0111] Step 42: Calculate the fusion weight coefficient of each recognizer in the dynamic weighted fusion process based on the recognition capability.

[0112] Step 43: Based on the fusion weight coefficient, fuse the first recognition result and the second recognition result to obtain the recognition result.

[0113] In one feasible implementation, the entropy calculation formula is as follows:

[0114]

[0115] Among them, H i (x) represents the entropy value of the i-th classifier, H i (x) can reflect the classification ability of the i-th classifier, H i The larger (x) is, the weaker the classification ability of the i-th classifier, and vice versa. m represents the total number of classifiers, p ij′ (x) represents the probability that the i-th classifier predicts the target x as the j′-th class.

[0116] In one possible implementation, step 42 includes:

[0117] Through calculation formula

[0118]

[0119] Obtain the fusion weight coefficient ω of the i-th classifier i .

[0120] The fusion weight coefficient ω obtained from the above formula i Assigning classes to m classifiers, the adjusted probability matrix p′(x) can be expressed as:

[0121]

[0122] In one possible implementation, step 43 includes:

[0123] Through calculation formula

[0124] P fusion (C)=β·P RF (C)+(1-β)·P CNN (C)

[0125] The recognition result P was obtained. fusion (C); where P RF (C) represents the first identification result, P CNN (C) represents the second recognition result. In the formula, β corresponds to ω1 of the adjusted probability matrix, (1-β) corresponds to ω2, and formula P... fusion (C) is a decision-level fusion of the classification and identification results of the two branches.

[0126] Finally, by selecting the label with the highest category probability in the recognition results as the final target label, radar space target recognition based on dual-branch collaborative decision-level fusion is achieved.

[0127] The training process of the recognition network model is explained below.

[0128] Specifically, the pre-acquired samples were divided into training and testing sets in an 8:2 ratio. The data was randomly shuffled before each training or testing session to enhance the model's generalization ability. After multiple training parameter optimizations, the number of decision trees in the RF classifier was set to 100; the MBCNN model was trained 20 times with a batch size of 32. The initial learning rate of the MBCNN was set to 10%. -3 The training loss function uses cross-entropy loss; the network weights are randomly initialized, and the ADAM optimizer is used for weight updates during backpropagation.

[0129] In another embodiment of the present invention, the accuracy of the first recognition result output by the first branch was verified, and the process is as follows:

[0130] The first recognition result was input into a random forest classifier for target recognition. The first recognition results under different signal-to-noise ratio conditions are shown in Table 2. The confusion matrix and t-SNE visualization results of the first recognition results at signal-to-noise ratios of -20dB and -4dB are shown in Table 2. Figure 7 As shown.

[0131] Table 2

[0132] Signal-to-noise ratio OA (%) <![CDATA[PA T (%)]]> <![CDATA[PA F1 (%)]]> <![CDATA[PA F2 (%)]]> <![CDATA[PA F3 (%)]]> Average AUC -20 66.03 74.4 53.8 64.1 71.8 0.8759 -18 74.36 79.5 71.8 76.9 69.2 0.8920 -16 77.56 79.5 79.5 84.6 66.7 0.9280 -14 78.85 82.1 71.8 79.5 82.1 0.9380 -12 82.05 89.7 66.7 82.1 89.7 0.9575 -10 88.46 92.3 87.2 84.6 89.7 0.9811 -8 91.67 100 87.2 94.9 84.6 0.9903 -6 95.51 100 97.4 97.4 87.2 0.9962 -4 98.72 100 100 97.4 97.4 0.9997

[0133] As shown in Table 2, under extreme noise conditions with an SNR of -20dB, the overall accuracy of the system is 66.03%, with the accuracy of identifying true targets being higher than that of the other three types of false targets, effectively proving the accuracy of the first identification result.

[0134] from Figure 7 As shown in the confusion matrix (a), misjudgment is common among various targets. The probability of false target 1 being misjudged as a real target is as high as 25.6%, which will increase the false alarm processing load. Figure 7 (b) The t-SNE visualization results show that the feature distribution of the four types of targets has no obvious inter-class distance, indicating that the physical features are relatively sensitive to noise interference, and the classifier's recognition performance drops sharply under low signal-to-noise ratio conditions.

[0135] When the signal-to-noise ratio (SNR) is increased to above -8dB, the classifier's accuracy in recognizing true targets increases to over 90%, and the overall system accuracy improves by 22.43%. This indicates that, with effective suppression of noise interference, physical features can accurately reflect the essential attributes of the target. At an SNR of -4dB, the accuracy in recognizing true targets and false target 1 reaches 100%, while false targets 2 and 3, due to their extremely similar structures, exhibit a 2.6% misclassification rate. Figure 7 (d) The t-SNE visualization results show the overlapping area between the two, and the other targets are well separable. Compared with the traditional HRPnet method, the first branch shows better recognition performance when the signal-to-noise ratio is greater than -6dB, and the training time of a single model is optimized to within 5 seconds, which can ensure efficiency and accuracy.

[0136] In another embodiment of the present invention, the accuracy of the second recognition result output by the second branch was verified, and the process is as follows:

[0137] The second branch's recognition results under different signal-to-noise ratio (SNR) conditions are shown in Table 3. The confusion matrix and t-SNE visualization results for SNRs of -20dB and -10dB are shown in Table 3. Figure 8 As shown.

[0138] Table 3

[0139] Signal-to-noise ratio OA (%) <![CDATA[PA T (%)]]> <![CDATA[PA F1 (%)]]> <![CDATA[PA F2 (%)]]> <![CDATA[PA F3 (%)]]> Average AUC -20 87.18 94.9 84.6 84.6 84.6 0.9856 -18 89.75 100 87.2 87.2 84.6 0.9892 -16 91.67 100 97.4 79.5 89.7 0.9874 -14 92.95 100 92.3 84.6 94.9 0.9920 -12 97.44 100 94.9 94.9 100 0.9986 -10 98.72 100 100 94.9 100 0.9995

[0140] As shown in Table 3, the recognition performance of the second branch steadily improves as the signal-to-noise ratio (SNR) gradually increases from -20dB to -10dB. Under extreme noise conditions with an SNR of -20dB, the overall accuracy reaches 87.18%, which is 21.15% higher than that of the first branch. This indicates that the second branch can still maintain strong classification performance under noise interference and is more robust to noise.

[0141] Figure 8The confusion matrix shown in (a) indicates that there are still a small number of misclassifications among the targets; compared with the first branch Figure 7 (b) In comparison, Figure 8 (b) Visualization results show that in the second branch, the inter-class spacing of the target feature space is significantly increased, and the separability of the target is improved.

[0142] When the signal-to-noise ratio (SNR) is increased to above -18dB, the accuracy of real target recognition reaches 100%, while the accuracy of false target recognition shows a fluctuating increase. This indicates that deep neural networks can effectively extract deep features from effective polarization channel data, exhibiting better recognition performance for real targets with richer information and more obvious features. Figure 8 (c) It can be seen that when the SNR is -10dB, the overall accuracy of the second branch reaches 98.72%, and except for a 5.1% misclassification of false target 2, the other three types of targets are all identified 100%; Figure 8 In (d), the t-SNE visualization results remain consistent, with only false targets 2 and 3 showing partial overlap; the remaining targets are clearly separated. Compared to the first branch, the second branch effectively improves noise robustness in complex electromagnetic environments. At an SNR of -12dB, the trends of the second branch's recognition accuracy and loss function with the number of iterations are as follows: Figure 9 and Figure 10 As shown, after 60 iterations, the recognition accuracy of the second branch stabilized above 95%, and the loss function converged to below 0.01, indicating that the network has converged to a better recognition performance.

[0143] In another embodiment of the present invention, the accuracy of the recognition result of the fusion module was verified, and the process is as follows:

[0144] In this embodiment, a dynamic weighted fusion algorithm based on entropy values ​​performs decision-level fusion on the identification results of the first and second branches. The overall accuracy of different decision-level fusion methods varies with signal-to-noise ratio (SNR) in the range of -20dB to -10dB as shown below. Figure 11 As shown.

[0145] In the weighted voting method, the second branch, which has better recognition performance, is given a higher weight, while the weights of the first and second branches are set to 0.4 and 0.6, respectively. Experimental results show that the result of the weighted voting decision fusion is completely consistent with that of the second branch. This is mainly because the second branch performs better under most signal-to-noise ratio (SNR) conditions and dominates the final decision under a fixed weight allocation. The majority voting method assumes that the two branches have the same recognition ability, and the fusion recognition result always falls between the two branches, failing to fully utilize their complementary advantages. The entropy-based dynamic weighted fusion algorithm (ADWF decision fusion) provided in this invention effectively integrates the advantages of the two branches under different SNR conditions by dynamically adjusting the weights of each branch, resulting in an improved fusion recognition result compared to a single branch. At an SNR of -10dB, the recognition accuracy of the second branch reaches 98.72%, approaching its performance limit. At this point, the room for improvement in system performance by the entropy-based dynamic weighted fusion algorithm is limited, and the fusion result tends to be consistent with that of the second branch.

[0146] When the SNR is -20dB, after decision-level fusion based on the entropy-based dynamic weighted fusion algorithm provided by this invention, the confusion matrix of the dual-branch fusion identification is as follows: Figure 12 As shown, compared with the results of the second branch, the accuracy of true target recognition after fusion remains unchanged at 94.9%, while the accuracy of false target 1 and false target 2 increases by 2.6% and 7.7% respectively, and the accuracy of false target 3 decreases by 2.5% due to the influence of the first branch. This indicates that the discrimination ability of the first branch on specific interference features has been effectively integrated. Although some false judgments are introduced, the stability of key true target recognition is guaranteed. Through decision-level fusion, not only is the real-time processing advantage of the first branch retained, but the robustness of the system to noise is also enhanced through the second branch. Overall, the branch module of the recognition network model provided by this invention maintains an overall recognition accuracy of over 89% in complex electromagnetic environments, which is superior to the recognition performance of traditional single-branch models.

[0147] In another embodiment of the present invention, to further verify the effectiveness of the radar spatial target recognition method provided by the present invention, the impact of different training sample ratios on recognition performance at a signal-to-noise ratio of -10dB was compared. The overall accuracy trend with the training sample ratio is as follows: Figure 13 As shown.

[0148] In the weighted voting method, the first branch, which is relatively less sensitive to data, is given a higher weight. The weights of the first and second branches are set to 0.6 and 0.4, respectively. The first branch relies on physical separability features and is highly adaptable to small samples. When the proportion of training samples decreases from 80% to 10%, the overall accuracy only decreases by 7.47%. The second branch, however, requires a large number of samples to train the deep neural network, and its recognition performance drops sharply with small samples. When the proportion of training samples decreases from 80% to 10%, the overall system accuracy decreases by 38.15%. Figure 13 It can be seen that the ADWF decision fusion method provided by the present invention achieves a peak accuracy of 99.15% when the training sample ratio is 70%, and maintains the best recognition performance under different training sample ratios.

[0149] Analysis reveals that the recognition network model provided by this invention constructs a dynamically optimized target recognition mechanism by integrating the advantages of the first and second branches. In the initial stage of data scarcity, the first branch, leveraging its robust classification characteristics with small samples, can quickly complete coarse target classification within 5 seconds, improving the interpretability of the recognition network while ensuring real-time performance. The second branch employs a lightweight design with single-layer convolution, enabling rapid training within 2 minutes through adaptive feature extraction. As data accumulates, it gradually improves the system's recognition accuracy and noise robustness, achieving precise target classification. In practical applications, the two branches form a complementary and collaborative recognition system, enhancing the radar spatial target recognition capability in complex environments.

[0150] In another embodiment of the present invention, in order to further verify the effectiveness of the radar spatial target recognition method provided by the present invention, this embodiment compares and analyzes it with traditional radar target recognition methods, different feature fusion schemes and existing deep learning models.

[0151] Compared with traditional methods, the first branch of this invention extracts physical features from the optimal separability subset and selects a 28-dimensional optimal feature subset based on a physical feature selection algorithm. This subset is then input into a random forest classifier for recognition. To verify the effectiveness of data separability assessment and multi-channel feature fusion, the original feature parameter set obtained from physical feature extraction is directly input into SVM and KNN classifiers for recognition. Table 4 records the recognition accuracy of the network under different signal-to-noise ratio (SNR) conditions. The curves showing the change in recognition accuracy with SNR are as follows: Figure 14 As shown in the figure. The input feature size of the RCS-HH and RCS-VH channels is 196×5, and the input feature size of the HRRP-VH and HRRP-VV channels is 196×6. Narrowband polarization features for the period from 1600 to 2000 seconds are extracted.

[0152] Table 4

[0153]

[0154] Experimental results show that the recognition accuracy of the single-channel method gradually improves with the increase of SNR. The average Fisher separability of the RCS-VH channel from 1600 to 2000 seconds is 24.74, higher than the average separability of the RCS-HH channel from 700 to 1100 seconds (12.97). After inputting the RCS features of both channels into the classifier, the RCS-VH channel, with its higher Fisher separability, exhibits superior recognition performance, with an overall accuracy improvement of over 10% compared to the RCS-HH channel. The average separability of the HRRP VV channel from 1600 to 2000 seconds is 2.84, better than the average separability of the HRRP VH channel from 500 to 900 seconds (2.37). At a signal-to-noise ratio of -14dB, the SVM classifier based on HRRP-VV features achieves a 19.24% improvement in recognition accuracy compared to the HRRP-VH feature-based classifier. Compared to directly extracting features from the raw data, obtaining the optimal subset through data separability assessment before feature extraction improves network recognition performance. The above comparison validates the effectiveness of the data separability assessment module, which reduces the adverse effects of noise accumulation on recognition accuracy by eliminating low-discrimination data, thereby improving system recognition performance. Furthermore, from... Figure 14 It can be seen that the recognition accuracy based on RCS and HRRP features is better than that based on polarization features.

[0155] However, single-channel feature recognition methods have significant limitations. When the signal-to-noise ratio (SNR) is below -10dB, the recognition accuracy of all channels is below 60%, and the difference in recognition accuracy between different features is as high as 20%, reflecting the information loss problem of single-type features under strong noise conditions. In contrast, the first branch provided by this invention constructs an optimal feature subset through feature separability evaluation and multi-channel feature fusion. Under the extreme SNR condition of -14dB, the recognition accuracy is improved by more than 15.77% compared with traditional methods. The recognition network model adopts dual-branch collaborative recognition, which complements and fuses physical features and deep learning features. It can still achieve a recognition accuracy of 96.79% when the SNR is -14dB, which is more than 30% higher than the single-channel method. When the SNR is greater than -10dB, the recognition accuracy remains stable at over 99%, effectively demonstrating the efficiency and accuracy of the radar spatial target recognition method provided by this invention.

[0156] In another embodiment of the present invention, the effectiveness of the radar spatial target recognition method provided by the present invention is verified by comparing the recognition performance of different feature fusion schemes in a noisy environment. The experiment uses a random forest classifier for recognition, with 100 decision trees. The input feature set is the first recognition result obtained based on the first branch, where the input feature size for the RCS-HH and RCS-VH channels is 196×8, the input feature size for the HRRP-VH and HRRP-VV channels is 196×5, and the input feature size for the POL channel is 196×2.

[0157] Table 5 records the recognition accuracy under different signal-to-noise ratio (SNR) conditions. The curves showing the change in recognition accuracy with SNR are as follows: Figure 15 As shown.

[0158] Table 5

[0159]

[0160] In single-channel recognition, compared with the scheme of directly inputting the original feature parameter set, the radar spatial target recognition method provided by this invention exhibits superior recognition performance using 8-dimensional RCS features. At an SNR of -14dB, the single-channel recognition accuracy of RCS-HH after feature selection is 42.95%, an improvement of 4.49% compared to the original feature set; at an SNR of -4dB, the single-channel recognition accuracy of RCS-VH increases to 80.13%, indicating that the first branch effectively removes redundant noise in the RCS features. However, during feature selection, the HRRP and POL channels suffer from partial information loss due to feature dimension compression, resulting in lower recognition accuracy than the original feature parameter set. After reducing the feature dimension of the HRRP-VH channel from 10 to 5, the recognition accuracy at an SNR of -4dB decreases from 86.54% to 82.05%; after reducing the POL feature dimension from 5 to 2, the recognition accuracy decreases by 20.39% at an SNR of -4dB. This result indicates that feature selection requires retaining appropriate feature dimensions to fully characterize the target's properties.

[0161] Regarding feature fusion, this embodiment compares and analyzes two methods: weighted feature fusion and concatenation fusion. The input feature sizes after weighted fusion of the RCS-HH and RCS-VH channels, and the HRRP-VH and HRRP-VV channels, are and , respectively; after concatenation fusion, the feature sizes are and , respectively. For example... Figure 15As shown, since splicing fusion can preserve the original feature information, its recognition performance improvement effect is better than weighted fusion. When the SNR is -4dB, the recognition accuracy of RCS dual-channel and HRRP dual-channel is improved by 8.97% and 6.41% respectively compared with single channel. Experimental results show that appropriate feature selection helps to improve recognition accuracy, but for low-dimensional feature data, multi-feature splicing fusion can achieve better recognition results. The EWFRM-RF branch fuses three types of features: RCS, HRRP, and POL. The input feature size is 196×28. When the SNR is -4dB, the recognition accuracy reaches 98.72%, which is improved by 14.1% and 7.69% respectively compared with single-channel and dual-channel fusion recognition accuracy, fully demonstrating the complementary advantages of multi-dimensional feature fusion. To address the problem of insufficient robustness of EWFRM-RF branch in noisy environments, the radar spatial target recognition method provided by this invention achieves optimal recognition performance by combining the interpretability of physical features and the noise resistance capability of deep learning.

[0162] In another embodiment of the present invention, a comparative experiment is conducted between the radar spatial target recognition method provided by the present invention and five radar target recognition methods based on deep learning. The comparative methods include: the BiGRU_RCS method for RCS recognition based on bidirectional gated recurrent units, the 1DCNN_HRRP method for HRRP recognition based on a one-dimensional convolutional neural network, the LSTM_RCS_POL method for fully polarized RCS recognition based on a long short-term memory network, and the LSTM_HRRP_POL method for fully polarized HRRP recognition based on LSTM. The experimental dataset specifically includes data from the 700-1100 second period of the RCS-HH polarization channel and the 500-900 second period of the HRRP-VH polarization channel, as well as the corresponding fully polarized data. The hyperparameter settings of all comparative methods are consistent with the original literature. Table 6 records the recognition accuracy of each method under different signal-to-noise ratio conditions, and Table 7 records the average training time of each method based on five repeated experiments. The curves showing the network recognition accuracy versus SNR are shown below. Figure 16 As shown.

[0163] Table 6

[0164]

[0165] Table 7

[0166]

[0167]

[0168] Table 6 shows that the performance of deep learning-based target recognition networks depends on the network architecture design and the quality of the input data. The BiGRU_RCS and 1DCNN_HRRP methods, using single-polarization channel RCS and HRRP data as input, did not show a significant performance improvement compared to the single-channel recognition methods RCS-HH+RF and HRRP-VH+RF, which are based on "physical feature extraction + classifier." At an SNR of -4dB, the recognition accuracy of the 1DCNN_HRRP method even decreased by 5.21% compared to the HRRP-VH+RF method. This phenomenon indicates that if deep neural networks cannot effectively extract the essential features of the target, their performance may be inferior to traditional methods based on physical feature extraction.

[0169] Depend on Figure 16 It can be seen that the LSTM_RCS_POL and LSTM_HRRP_POL methods, by fusing fully polarized RCS and HRRP data, outperform the BiGRU_RCS and 1DCNN_HRRP methods under the same signal-to-noise ratio (SNR) conditions. At an SNR of -4dB, their recognition accuracy is improved by 13.24% and 11.88% respectively compared to single-polarization recognition, validating the effectiveness of polarization information fusion. However, such fully polarized fusion methods require simultaneous processing of multi-channel data; the LSTM_HRRP_POL method has a training time as high as 272.58s, which is not conducive to real-time processing. Compared with cross-time-period, cross-polarization channel splicing fusion methods based on RCS and HRRP data, although the computational complexity increases, the recognition performance is not further improved. In contrast, the radar spatial target recognition method provided by this invention fuses data from different polarization channels with highly separable time periods, such as fusion of data from the RCS-HH polarization channel (700-1100 seconds) and the RCS-VH polarization channel (1600-2000 seconds). This effectively reduces redundancy in all polarization data while achieving complementarity of cross-time period and cross-polarization features. Experimental results show that the radar spatial target recognition method provided by this invention outperforms existing deep learning methods in terms of recognition performance. Even under extreme signal-to-noise ratio conditions with an SNR of -14dB, it maintains a recognition accuracy of over 90%, representing an improvement of more than 20% compared to existing methods, thus verifying its recognition advantages in complex environments.

[0170] In summary, the radar spatial target recognition method provided by this invention overcomes the limitations of a single data source by acquiring the full polarization RCS and HRRP data of the target, thus improving the accuracy of radar spatial target recognition. Selecting effective polarization channel data based on the Fisher separability criterion reduces data redundancy and improves data quality, thereby increasing the efficiency of radar spatial target recognition. The first branch of the recognition network model extracts interpretable physical features, ensuring stability in small sample scenarios, while its second branch automatically learns noise robustness features to adapt to complex electromagnetic environments, improving the accuracy of radar spatial target recognition. By fusing the results of the two branches using entropy weighting, the high-confidence branch has a larger weight, reducing the false positive rate and effectively improving the accuracy of radar spatial target recognition.

[0171] like Figure 17 As shown, embodiments of the present invention provide a terminal device, such as... Figure 17 As shown, the terminal device D10 of this embodiment includes: at least one processor D100 ( Figure 17 The diagram shows only one processor, a memory D101, and a computer program D102 stored in the memory D101 and executable on the at least one processor D100, wherein the processor D100 executes the computer program D102 to implement the steps in any of the above method embodiments.

[0172] Specifically, when the processor D100 executes the computer program D102, it acquires radar echo data of the target to be tested; it divides the radar echo data using a time sliding window to obtain multiple radar echo data segments; it determines at least one effective polarization channel data from the multiple radar echo data segments based on the Fisher separability criterion; and it inputs each effective polarization channel data into a pre-trained recognition network model to obtain the recognition result of the target to be tested output by the recognition network model. Among these methods, acquiring the full polarization RCS and HRRP data of the target under test can overcome the limitations of a single data source, thus improving the accuracy of radar spatial target identification. Selecting effective polarization channel data based on the Fisher separability criterion reduces data redundancy and improves data quality, thereby increasing the efficiency of radar spatial target identification. The first branch of the identification network model extracts interpretable physical features, ensuring stability in small sample scenarios, while its second branch can automatically learn noise robustness features to adapt to complex electromagnetic environments, improving the accuracy of radar spatial target identification. By fusing the results of the two branches using entropy weighting, the high-confidence branch has a larger weight, reducing the false positive rate and effectively improving the accuracy of radar spatial target identification.

[0173] The processor D100 can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0174] In some embodiments, the memory D101 may be an internal storage unit of the terminal device D10, such as a hard disk or memory of the terminal device D10. In other embodiments, the memory D101 may be an external storage device of the terminal device D10, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the terminal device D10. Furthermore, the memory D101 may include both internal and external storage units of the terminal device D10. The memory D101 is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of the computer program. The memory D101 can also be used to temporarily store data that has been output or will be output.

[0175] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the various method embodiments above.

[0176] This application provides a computer program product that, when run on a terminal device, enables the terminal device to implement the steps described in the above-described method embodiments.

[0177] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0178] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.

Claims

1. A radar spatial target identification method, characterized in that, include: Acquire radar echo data of the target under test; The radar echo data includes fully polarized RCS data and fully polarized HRRP data; The radar echo data is divided into multiple radar echo data segments using a time sliding window; each radar echo data segment contains data from all polarization channels. At least one effective polarization channel data is determined from the plurality of radar echo data segments based on Fisher's separability criterion; Each effective polarization channel data is input into a pre-trained recognition network model to obtain the recognition result of the target output by the recognition network model. The recognition network model includes a branch module and a fusion module connected in sequence. The branch module includes a first branch and a second branch, both of which are input to the effective polarization channel data. The first branch extracts the physical features of the effective polarization channel data and identifies these physical features to obtain a first recognition result. The second branch extracts the deep features of the effective polarization channel data and identifies these deep features to obtain a second recognition result. The fusion module performs a dynamic weighted fusion of the first and second recognition results based on entropy values ​​to output the recognition result. The physical features represent the physical properties of the corresponding target, including geometry and material. The deep features characterize complex nonlinear relationships. The physical characteristics include RCS characteristics, HRRP characteristics, and narrowband polarization characteristics; The RCS features include location features, scattering features, transformation features, and distribution features. The location features include the maximum, minimum, mean, truncated mean, median, and mode of the RCS data. The scattering features include the range, variance, third central moment, fourth central moment, and coefficient of variation of the RCS data. The third central moment of the RCS represents the skewness of the scattering distribution; a positive skewness indicates the presence of strong scattering points. The fourth central moment of the RCS represents the kurtosis of the scattering; high values ​​correspond to targets with multiple scattering centers. The coefficient of variation of the RCS represents the normalized fluctuation degree, comparing targets of different sizes. The transformation features include the mean of the spectrum, the mean of the power spectrum, and the mean of the correlation coefficient. The distribution features include the kurtosis coefficient and the skewness coefficient. The HRRP features include the first, second, and third central moments of the HRRP data, descaled structural features, the distance of the highest peak relative to the leftmost peak, the distance of the highest peak relative to the rightmost peak, the symmetry of target scattering, the dispersion of target scattering, and the number of strong scattering centers of the target. Among these, the first central moment of the HRRP data represents the centroid shift of the scattered energy distribution, the second central moment of the HRRP data represents the dispersion of scattering points in the range dimension, and the third central moment of the HRRP data represents the morphological asymmetry of the range profile. The narrowband polarization characteristics include the trace of the power matrix of the narrowband polarization, the determinant of the scattering matrix, the depolarization coefficient, the intrinsic polarization ellipticity, and the intrinsic polarization direction angle. After extracting the physical features of the effective polarization channel data and before identifying the physical features, the first branch further includes: A multi-criteria fusion feature selection strategy based on entropy weighting is used to evaluate the separability of the physical features, resulting in a separability score for each physical feature. The calculation expression for the separability score is as follows: in, Indicates the score for separability. Indicates the first Algorithms for evaluating separability They represent algorithm, Algorithm or algorithm, Indicates the first Weights of separability evaluation algorithms Represents the normalized i-th A separability evaluation algorithm for the first Separability score of each physical feature , Indicates the total number of physical features. Indicates the first The entropy value of a separability evaluation algorithm. Indicates the first The proportion of different separability evaluation algorithms, Indicates the first A separability evaluation algorithm for the first Separability score of each physical feature Indicates the first A set of separability scores for all physical features from a separability evaluation algorithm; Physical features whose separability score is greater than a preset score threshold are identified as the physical features to be identified.

2. The radar spatial target identification method according to claim 1, characterized in that, The fully polarized RCS data includes two co-polarized channel RCS data and two cross-polarized channel RCS data.

3. The radar spatial target identification method according to claim 2, characterized in that, The determination of at least one effective polarization channel data from the plurality of radar echo data segments based on the Fisher separability criterion includes: Based on Fisher's separability criterion, the separability value corresponding to each polarization channel data in each radar echo data segment is calculated; Polarization channel data with a separability value greater than a preset separability threshold are considered as valid polarization channel data.

4. The radar spatial target identification method according to claim 1, characterized in that, The second branch includes an RCS branch for processing fully polarized RCS data and an HRRP branch for processing fully polarized HRRP data; wherein the RCS branch employs one-dimensional convolution and pooling processing, and the HRRP branch employs two-dimensional convolution and pooling processing.

5. The radar spatial target identification method according to claim 4, characterized in that, The fusion module includes multiple recognizers; The step of dynamically weighting and fusing the first recognition result and the second recognition result based on entropy value, and outputting the recognition result, includes: The recognition capability of each of the aforementioned recognizers is measured by its entropy value; Based on the recognition capability, the fusion weight coefficient of each of the recognizers is calculated during the dynamic weighted fusion process; The first recognition result and the second recognition result are fused according to the fusion weight coefficient to obtain the recognition result.

6. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 5.

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 5.

Images