Angular domain adaptive radar target identification method based on knowledge assistance and multi-pole fusion

Abstract

The invention discloses an angular domain adaptive radar target recognition method based on knowledge assistance and multi-pole fusion, relates to the technical field of radar recognition, and aims to solve the problems that the same target may show completely different scattering structures at different angles, and different targets may have feature overlapping at certain visual angles. According to the invention, the method can effectively improve the consistency of the scattering characteristics, can accurately recognize the change region of the scattering structure in a wide azimuth angle range, and achieves the stable segmentation of the scattering characteristics of various types of targets. According to the method and the device, the influence of feature drift and an unstable interval under a wide-angle-domain condition is effectively avoided, the problem of feature overlapping of different targets under certain view angles is avoided, and the radar target recognition accuracy in a wide-angle-domain scene is further improved.

Description

Technical Field

[0001] This application relates to the field of radar identification technology, specifically to an angle-domain adaptive radar target identification method based on knowledge assistance and multi-polar fusion. Background Technology

[0002] Radar target identification technology is a key component of modern electronic reconnaissance, intelligence gathering, and weapon systems. Radar acquires the physical properties of targets by emitting electromagnetic waves and receiving the echo signals scattered by the targets. Its scattering characteristics are typically influenced by factors such as the target's geometry, material, electromagnetic properties, attitude angle, and viewing angle. Under wide azimuth conditions, the visibility of the target's scattering center, the distribution of scattering intensity, and the phase structure change significantly with the angle, causing the target echo to exhibit highly nonlinear and azimuth-sensitive characteristics, placing higher demands on traditional identification methods.

[0003] With the development of millimeter-wave / wideband radar, range profile (HRRP), frequency domain scattering spectrum, and polarimetric scattering features have been widely used for target identification. However, in practical applications, the same target may exhibit completely different scattering structures at different angles, and different targets may show feature overlap at certain viewpoints, resulting in a significant decrease in the accuracy of traditional identification methods in wide-angle scenarios. Summary of the Invention

[0004] The purpose of this invention is to address the problem that the same target may exhibit completely different scattering structures at different angles, and different targets may have overlapping features at certain viewpoints, which leads to a significant decrease in the accuracy of traditional recognition methods in wide-angle scenarios. This invention provides an angle-domain adaptive radar target recognition method based on knowledge assistance and multi-polar fusion.

[0005] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0006] The knowledge-assisted and multi-polarity fusion-based angle-domain adaptive radar target recognition method includes the following steps:

[0007] The radar echo data to be identified and its corresponding azimuth angle are obtained, and its angular domain is determined based on the azimuth angle. The corresponding trained CNN model is then retrieved based on the angular domain. The amplitude spectrum of the radar echo data is then obtained, and finally the amplitude spectrum is input into the CNN model to obtain the classification result.

[0008] Furthermore, the trained CNN model is obtained through the following steps:

[0009] Step 1: Acquire radar echo data of four polarization channels at multiple azimuth angles for multiple targets in the same working band, and perform complex spectrum reconstruction on each radar echo data to obtain the complex scattering spectrum and amplitude spectrum of each polarization channel at each azimuth angle.

[0010] Step 2: For each complex scattering spectrum, obtain a one-dimensional range profile through inverse Fourier transform, and extract scattering features sensitive to azimuth angle based on the one-dimensional range profile;

[0011] Step 3: Concatenate the scattering features corresponding to each azimuth angle according to the dimension to obtain a multidimensional feature vector. Sort the multidimensional feature vectors corresponding to all azimuth angles of the same target and the same polarization, and calculate the difference measure based on Euclidean distance for adjacent multidimensional feature vectors to obtain the structural feature difference curve.

[0012] Step 4: Calculate the correlation coefficient difference between the HRRP sequence and the complex scattering spectrum between adjacent azimuth angle samples, and obtain the correlation difference curve based on HRRP and the correlation difference curve based on the complex scattering spectrum.

[0013] Step 5: Weight and superimpose the structural feature difference curve, the correlation difference curve based on HRRP, and the correlation difference curve based on complex scattering spectrum to obtain the single-polarization comprehensive difference curve corresponding to a single target.

[0014] Step 6: Repeat the above steps to obtain the four-polarization comprehensive difference curve corresponding to the same target, and perform weighted fusion of the four-polarization comprehensive difference curves corresponding to the same target to obtain the comprehensive difference curve under the four-polarization coupling condition corresponding to a single target.

[0015] Step 7: Repeat the above steps to obtain the comprehensive difference curves under the four-polar coupling conditions for all targets. Then, interpolate and normalize the comprehensive difference curves under the four-polar coupling conditions for all targets on the same azimuth grid, and superimpose them in an equal weight manner to obtain the global angular domain difference curves covering all target categories.

[0016] Step 8: Based on the global angular domain difference curve covering all target categories, the change point detection method is adopted, and under the preset constraints of the maximum number of change points and the minimum angular domain width, the preliminary angular domain division result covering the entire angular range is obtained;

[0017] Step 9: Based on the preliminary angular domain division results covering the entire angular range, and according to the angle threshold, determine the angular domain with excessive span, and use the starting azimuth angle in the angular domain as the reference point to calculate the Euclidean distance difference between each point in the angular domain and the starting azimuth angle in turn, so as to obtain the difference sequence, and recursively divide the angular domain according to the median position of the difference sequence.

[0018] Step 10: Based on the results of recursive segmentation, determine the high-difference intervals where the average difference of each corner domain exceeds the global threshold according to the set global threshold, and segment them according to the pattern of difference change using median segmentation or peak segmentation strategy to obtain sub-intervals. Then, merge the sub-intervals into adjacent stable corner domains to obtain the final corner domain division result.

[0019] Step 11: Segment the amplitude spectra of all polarization channels at each azimuth angle to obtain the quadpolar amplitude spectrum;

[0020] Step 12: Based on the final angular domain division results, group all quadrature amplitude spectra according to the divided angular domains, and train a CNN model using each group of quadrature amplitude spectra.

[0021] Furthermore, the complex scattering spectrum is expressed as:

[0022] ,

[0023] in, Represents the complex scattering spectrum. Indicates frequency point azimuth polarization channel and target category Complex scattering coefficients under these conditions , The total number of distance cells. For the first One distance unit.

[0024] Furthermore, the amplitude spectrum is represented as:

[0025] ,

[0026] in, Represents the amplitude spectrum.

[0027] Furthermore, the sensitive scattering feature includes the number of strong scattering units. Scattering center distance spread Energy ratio of the first 25% distance window Energy ratio of the first 50% distance window First belt energy ratio Second band energy ratio Third band energy ratio HRRP frequency domain entropy and the first difference mean of HRRP .

[0028] Furthermore, in the sensitive scattering features, the number of strong scattering units... Based on the set energy threshold The distance units whose amplitude exceeds the threshold are counted and represented as:

[0029] ,

[0030] ,

[0031] in, For the distance image Energy per unit distance, The total energy of the entire distance image. This is a discrete sequence of one-dimensional distance images after norm normalization;

[0032] Scattering center distance spread Represented as:

[0033] ,

[0034] ,

[0035] ,

[0036] in, For distance axis, The distance-weighted average is the energy-weighted average. Energy normalization weight;

[0037] Represented as:

[0038] ,

[0039] Represented as:

[0040] ,

[0041] First belt energy ratio Represented as:

[0042] ,

[0043] ,

[0044] ,

[0045] ,

[0046] in, For frequency domain energy, The total energy of the spectrum of the entire one-dimensional range image. For discrete indices on the spectrum of HRRP, For the spectrum of HRRP, ;

[0047] Second energy ratio Represented as:

[0048] ,

[0049] Third zone energy ratio Represented as:

[0050] ,

[0051] in, , , The frequency domain index is divided into three sub-bands;

[0052] HRRP frequency domain entropy Represented as:

[0053] ,

[0054] ,

[0055] in, This represents the normalized frequency domain energy probability distribution. It is a very small positive number. The spectral sequence of HRRP;

[0056] HRRP first difference mean Represented as:

[0057] ,

[0058] ,

[0059] in, For the normalized first The sample values ​​of a one-dimensional distance image cell, This represents the magnitude of the jump between two adjacent distance units.

[0060] Furthermore, the structural feature difference curve is represented as follows:

[0061] ,

[0062] in, This is a curve showing the difference in structural characteristics. For known categories polarization mode Azimuth The structural feature vector of the sample, For known categories polarization mode azimuth angle is The structural feature vector of the sample, For known categories polarization mode Azimuth The first of the structural feature vectors of the sample The eigenvalues ​​of each feature.

[0063] Furthermore, the correlation difference curve based on HRRP is represented as follows:

[0064] ,

[0065] in, The correlation difference curve is based on HRRP. For known categories polarization mode Azimuth HRRP sequence of the sample, Adjacent azimuth angles Adjacent azimuth angles Pearson correlation coefficient, For known categories polarization mode Azimuth The HRRP sequence of the sample The value of each distance unit. For known categories polarization mode Azimuth The HRRP sequence of the sample after norm normalization is the first... The value of each distance unit.

[0066] Furthermore, the correlation difference curve based on the complex scattering spectrum is expressed as follows:

[0067] ,

[0068] in, The correlation difference curve is based on the complex scattering spectrum. Adjacent azimuth angles The multiple correlation coefficient, To represent the number of distance cells in a one-dimensional range image, For the azimuth angle is The complex scattering spectrum at time 1 Complex scattering coefficients at each frequency point For the azimuth angle is The complex scattering spectrum at time 1 Complex scattering coefficients at each frequency point For the azimuth angle is The complex conjugate form of the complex scattering spectrum at time is in the . Complex scattering coefficients at each frequency point.

[0069] Furthermore, the angle threshold in step 9 is 60 degrees.

[0070] The beneficial effects of this invention are:

[0071] This application can effectively improve the consistency of scattering characteristics, accurately identify regions of scattering structure change over a wide azimuth range, and achieve stable segmentation of scattering characteristics for multiple types of targets. This application effectively avoids the influence of feature drift and unstable regions under wide-angle conditions, and avoids feature overlap problems of different targets at certain viewpoints, thereby improving the accuracy of radar target recognition in wide-angle scenarios.

[0072] This application presents a four-polarization coupled CNN classification model based on the final angular domain partitioning. It independently learns local scattering features within each angular domain, significantly improving the recognition performance of deep learning under wide-angle and multi-polarization conditions. This model exhibits higher robustness, better convergence, and stronger practical engineering applicability, with overall recognition performance superior to traditional end-to-end unified training methods. Attached Figure Description

[0073] Figure 1 Flowchart for adaptive corner domain partitioning and target recognition;

[0074] Figure 2 Flowchart for calculating the global difference curve;

[0075] Figure 3 Comparison of the three types of difference curves with the final difference curve for objective 1 Figure 1 ;

[0076] Figure 4 Comparison of the three types of difference curves with the final difference curve for objective 1 Figure 2 ;

[0077] Figure 5 Comparison of the three types of difference curves with the final difference curve for objective 1 Figure 3 ;

[0078] Figure 6 Comparison of the three types of difference curves with the final difference curve for objective 1 Figure 4 ;

[0079] Figure 7 Comparison of difference curves under four scenarios for objective 1 Figure 1 ;

[0080] Figure 8Comparison of difference curves under four scenarios for objective 1 Figure 2 ;

[0081] Figure 9 Comparison of the four target and global difference curves Figure 1 ;

[0082] Figure 10 Comparison of the four target and global difference curves Figure 2 ;

[0083] Figure 11 Flowchart for fine-grained partitioning of the corner domain;

[0084] Figure 12 For excessively long and excessively high angular domains, the difference curves between the starting points are... Figure 1 ;

[0085] Figure 13 For excessively long and excessively high angular domains, the difference curves between the starting points are... Figure 2 ;

[0086] Figure 14 This is a schematic diagram of the final corner domain partitioning;

[0087] Figure 15 A schematic diagram of a deep learning network;

[0088] Figure 16 This is a schematic diagram of the CNN training results without corner domain partitioning;

[0089] Figure 17 This is a schematic diagram of the CNN training results for corner domain 1;

[0090] Figure 18 This is a schematic diagram of the CNN training results for corner domain 2;

[0091] Figure 19 This is a schematic diagram of the CNN training results for corner domain 3;

[0092] Figure 20 This is a schematic diagram of the CNN training results for angle domain 4. Detailed Implementation

[0093] It should be noted that, where there is no conflict, the various embodiments disclosed in this application can be combined with each other.

[0094] Specific Implementation Method 1: The corner-domain adaptive radar target recognition method based on knowledge assistance and multi-polar fusion described in this implementation method includes:

[0095] Step 1: Acquire four-polarized (HH, HV, VH, VV) radar echo data of multiple targets under the same working band at multiple azimuth angles, including the amplitude and phase information of each frequency point. Reconstruct the complex spectrum to obtain the complex scattering spectrum and amplitude spectrum of each azimuth angle, and store it in a data file in sample format as input for subsequent feature extraction and recognition model training.

[0096] Step 2: Extract the complex scattering spectrum from the data file and obtain the HRRP (one-dimensional range profile) for each sample using IFFT (Inverse Fourier Transform). Then, extract azimuth-sensitive structured physical features based on the HRRP. These azimuth-sensitive structured physical features include, but are not limited to: structural features describing the number of strong scattering centers; frequency domain energy ratio features reflecting the distribution of range profile frequency domain energy across different frequency bands; entropy features characterizing the complexity of range profile texture or frequency domain uncertainty; and differential statistical features characterizing the smoothness of range profile point sequence changes or the degree of local jumps.

[0097] Step 3: Concatenate the azimuth-sensitive scattering features according to dimensions to form a multi-dimensional feature vector for each sample. Sort the samples according to azimuth within the same target and the same polarization channel. Calculate the difference measure based on Euclidean distance for the feature vectors of adjacent azimuth samples to obtain a structural feature difference curve that varies with azimuth.

[0098] The correlation coefficient differences between HRRP sequences and complex scattering spectra in adjacent azimuth angle samples were calculated separately to form two types of difference curves. The three difference curves were then standardized and weighted and superimposed to form a comprehensive difference curve for single-target single-polarization.

[0099] Step 4: Apply preset polarization weights to the difference curves obtained from the four polarization channels of the same target and perform weighted fusion to construct a comprehensive difference curve under the single-target four-polarization coupling condition; then interpolate and normalize the comprehensive difference curves of all targets on the same azimuth grid, and superimpose them in an equal weight manner to form a global angular domain difference curve covering all target categories, which is used to characterize the overall change of scattering characteristics of multiple target categories as the azimuth changes.

[0100] Step 5: Apply the findchangepts method to the global difference curve obtained in Step 4 to perform preliminary angular domain division. Under the preset constraints of the maximum number of change points and the minimum angular domain width, a preliminary angular domain division covering the entire angular range is obtained. This coarse-grained division reflects the main changing trend of global scattering characteristics, but may contain long intervals with excessively large spans (processed in Step 6) or high-difference intervals that are generally unstable (processed in Step 7), requiring further processing.

[0101] Step 6: For angular regions with excessively large spans in the initial division (60-degree threshold), calculate the Euclidean distance difference degree with reference to the starting azimuth angle within the angular region. If the length of the angular region exceeds the preset threshold, recursively divide it according to the median position of the difference degree sequence to limit the maximum span of the angular region and avoid the cumulative drift of azimuth angle differences.

[0102] Step 7: On the global dissimilarity curve, based on the set global threshold, determine the high dissimilarity intervals where the average dissimilarity exceeds the global threshold. Then, use a median or peak segmentation strategy according to the dissimilarity variation pattern. Finally, merge the segmented sub-intervals into adjacent stable angular domains, ensuring that unstable intervals are not used as independent angular domains for training. This guarantees the consistency and stability of the scattering characteristics within the final angular domain. The final defined angular domain range can then be used for individual neural network learning.

[0103] Step 8: Reorganize the amplitude spectrum data corresponding to each azimuth angle according to the polarization channel. The amplitude spectra of the four polarizations HH, HV, VH, and VV under the same target and the same azimuth angle are spliced ​​in a fixed order to form the input sample of the four-polarization joint representation. On this basis, Gaussian white noise is introduced to perturb the amplitude spectrum to simulate the noise environment under actual radar measurement, thereby enhancing the robustness of the sample and constructing the final input data for neural network training.

[0104] Step 9: Based on the final angular domain division results, the four-polarization amplitude spectrum samples are assigned to their respective angular domains according to their azimuth. A convolutional neural network (CNN) classifier is independently constructed and trained for each angular domain, including random data partitioning, training / validation set construction, and multi-model training with consistent network structures. The CNN model for each angular domain learns the stable statistical laws governing target scattering within the local azimuth range. Ultimately, the models from all angular domains together constitute a complete recognizer, achieving automatic target recognition based on prior knowledge of the angular domain.

[0105] Step 10: In the identification stage, for any radar echo sample to be identified, first determine its angular domain based on its azimuth angle, then use the corresponding CNN model to perform forward inference of the four-polarization amplitude spectrum input features to obtain the output probability of each category, and make a category determination based on this to determine which target the echo to be identified comes from.

[0106] This application addresses various special flight targets and, for multi-polarization echo data of multiple targets acquired by bistatic radar over a wide azimuth range, overcomes the shortcomings of existing methods that generally rely on deep learning models for direct end-to-end training, resulting in slightly poor recognition performance over wide azimuth angles. It proposes an angular domain adaptive radar target recognition method based on knowledge assistance and multi-polarization fusion. Through a "difference degree-segmentation-local learning" approach, the complex scattering angle changes are transformed into several stable learning intervals, enabling the deep network to focus on locally consistent scattering patterns, thereby achieving higher recognition performance under conditions of multiple targets, multi-polarization, and wide azimuth angles.

[0107] Example:

[0108] like Figure 1 As shown, it includes the following steps:

[0109] Step 1: Acquire raw scattering measurement data of four types of targets (Target 1, Target 2, Target 3, and Target 4) in the C-band (5~6.5GHz), and analyze the azimuth range of each type of target. Inside, with The step size is calculated at each azimuth observation point. Below, four polarization channels were collected. The echo spectrum data, with a step frequency of The frequency sampling points are denoted as And record the corresponding amplitude and phase information for a target. In azimuth angle polarization channel The complex scattering spectrum under the given condition can be written as a column vector arranged according to the frequency sampling points:

[0110] ;

[0111] in, Indicates frequency point azimuth polarization channel and target category The complex scattering coefficients under the given conditions, and the corresponding amplitude spectral vectors, can be expressed as:

[0112] ;

[0113] Further norm normalization yields and ;

[0114] This eliminates the influence of dimensional differences and absolute magnitudes on subsequent feature analysis. (The same target...) Same polarization channel The complex scattering spectra of all azimuth angles are stacked column-wise and can be represented as a two-dimensional scattering matrix and amplitude spectrum matrix of frequency and azimuth angle:

[0115] ;

[0116] ;

[0117] Each column corresponds to a complex scattering spectrum or amplitude spectrum at a given azimuth angle. In engineering implementation, each complex spectrum is treated as a sample, and each sample is appended with a target category label, polarization type, and azimuth angle information to form a structured multi-target, multi-azimuth, and multi-polarization radar scattering dataset, which serves as the basic input for subsequent feature extraction, angular domain partitioning, and classification model training.

[0118] Step 2: Analyze the complex scattering spectrum obtained in Step 1. Perform an inverse Fourier transform to obtain the complex envelope of the one-dimensional distance image.

[0119] ;

[0120] in, The total number of distance cells. This represents the discrete index of the one-dimensional range image sequence, corresponding to different resolution units of the target in the range dimension. The corresponding sampling step size... The distance axis is To eliminate the absolute amplitude differences between different samples, peak normalization was performed on each HRRP line to obtain a normalized distance image. .

[0121] To simplify the expression, the following uses This is used to represent the normalized HRRP sequence under fixed conditions.

[0122] Since the distribution, intensity, and frequency domain structure of the target scattering center change significantly under different azimuth angles, the following structured features and two types of adjacent angle difference features can be constructed to comprehensively characterize the scattering variation with azimuth angle, providing a stable and quantifiable physical basis for subsequent angular domain division.

[0123] (1) Number of strong scattering units ;

[0124] The number of major scattering centers exposed by the target at the current azimuth angle reflects the significant differences in its key scattering structures, thus allowing the extraction of the number of strong scattering units. The range image energy is defined as:

[0125] ;

[0126] Set the energy threshold as Distance cells with amplitudes exceeding a threshold are considered strong scattering cells. And count.

[0127] (2) Scattering center distance spread ;

[0128] Scattering center distance spread This refers to the dispersion of scattered energy along the range direction. It describes the spatial distribution of scattered points at different angles and characterizes the dispersion and spread of the main scattered energy along the range direction. Based on energy normalization weights... , with distance axis Let be the independent variable, and define the energy-weighted distance mean as . The distance spread of the scattering center can then be expressed as:

[0129] ;

[0130] (3) Energy ratio of the first 25% and 50% distance windows ;

[0131] This represents the proportion of total energy contained in the first quarter of the range image, used to characterize the forward concentration of target scattered energy in the range dimension and its variation with azimuth. Similarly, It can also reflect the changes in the target's scattering structure at different azimuth angles, showing the overall forward or backward shift of the scattering center along the range axis. These two characteristics can be obtained by the following formula:

[0132] ;

[0133] (4) HRRP frequency domain three-band energy ratio;

[0134] Treating normalized HRRP as a discrete sequence The spectrum of HRRP is obtained by performing a one-dimensional Fourier transform on it:

[0135] ;

[0136] Define frequency domain energy as The frequency domain index is divided into three sub-bands. The energy ratio of the first band As a low-frequency feature, it typically corresponds to the overall layout of the scattering centers, which varies significantly with the angle and is used to characterize the smoothness and contour changes of HRRP; the second band energy ratio Local fluctuations reflecting the scattering intensity distribution are used to characterize the mid-frequency texture structure of the range image; third-band energy ratio. It represents high-frequency structures such as edge scattering and corner scattering, and is extremely sensitive to changes in angle, used to measure high-frequency sharp components.

[0137] (5) HRRP frequency domain entropy ;

[0138] This feature reflects the uniformity of the HRRP spectrum distribution and is related to the complexity of the scattering structure. It can quantify the complexity and energy uncertainty of the range image spectrum at different angles. After normalizing the frequency domain amplitude, a frequency domain uncertainty measure similar to Shannon entropy can be constructed.

[0139] ;

[0140] ;

[0141] It is a very small positive number (e.g.) ), used to avoid when Occasionally Numerical overflow caused;

[0142] (6) HRRP first-order difference mean ;

[0143] This feature refers to the average absolute value of the first-order difference of the HRRP sequence, which characterizes the smoothness of the distance image on the distance axis and the intensity of local jumps.

[0144] ;

[0145] ;

[0146] In summary, the above nine features can be combined into a nine-dimensional structural feature vector; this serves as a structured physical feature vector sensitive to azimuth angles.

[0147] ;

[0148] Step 3, Figure 2 The flowchart for calculating global dissimilarity in steps three and four is demonstrated. Within the same target and polarization channel, existing samples are sorted by azimuth angle from smallest to largest, and the multi-source dissimilarity metric between adjacent azimuth angle samples is calculated. Specifically, it includes the following three parts:

[0149] (1) Euclidean distance dissimilarity based on structural feature vectors ;

[0150] The known category is The polarization mode is azimuth angle is The structural feature vector of the sample is Then it is relative to the next azimuth angle The structural feature difference is defined as the Euclidean distance between adjacent angles:

[0151] ;

[0152] This index reflects the overall variation of the scattering structure characteristics of adjacent azimuth angles. If the structural characteristics differ greatly between adjacent angles, it indicates that the scattering structure in this region has undergone significant changes.

[0153] (2) Relevance differences based on HRRP ;

[0154] The known category is The polarization mode is azimuth angle is The HRRP sequence of the sample is Calculate the azimuth angle with adjacent azimuth angles. Pearson correlation coefficient And define the degree of difference for:

[0155] ;

[0156] When Pearson correlation coefficient Decrease, difference When the value increases, it indicates that the scattering intensity distribution or multipath structure of the target near that azimuth angle is changing rapidly.

[0157] (3) Based on Multiple correlation difference ;

[0158] Similar to Definition, calculation of adjacent azimuth angles multiple correlation coefficient The degree of multicorrelation difference can be obtained. for:

[0159] ;

[0160] This index reflects the combined change in amplitude and phase of the scattering center at adjacent azimuth angles in the frequency domain.

[0161] The three dissimilarity curves mentioned above are standardized using z-scores and weighted accordingly. By performing weighted superposition, a single-target single-polarization comprehensive difference curve is obtained. , Figure 3 , Figure 4 Figure 5 as well as Figure 6 A comparison of the three types of difference curves and the final difference curve under the target polarization channel HH.

[0162] Step four: For the difference curves under the four polarization conditions of a single target, perform weighted fusion of the four polarization difference curves according to preset weights to obtain the four polarization coupling difference curves of the single target. Different types of polarization channels exhibit significant differences in their sensitivity to target scattering characteristics. Same-polarization channels can more directly reflect the strong scattering characteristics of the target's main structure, while cross-polarization channels are more likely to capture weak scattering or additional scattering information. Figure 7 and Figure 8 The comparison of the difference curves under the four polarization channels of target 1 shows that the signal-to-noise ratio of the cross-polarization channel is low, and its scattering characteristics are greatly affected by noise. Therefore, in the process of angular domain division, the weight of the similarity index of HV / VH channels should be appropriately reduced, or used only as an auxiliary judgment.

[0163] Next, the four-polarization coupling difference curves for each target are superimposed with equal weights to obtain the final global difference curve. , Figure 9 It compares the difference curves of the four types of targets with the global difference curve.

[0164] Step 5: Analyze the global dissimilarity curve obtained in Step 4. A preliminary angular domain division is performed using the findchangepts method. This method can automatically identify significant transition points where the difference in azimuth changes, reflecting the "structural change" locations where the scattering mechanism shifts from one stable state to another. Under preset constraints on the maximum number of change points and the minimum angular domain width, a preliminary angular domain division covering the entire omnidirectional angular range is obtained. This coarse-grained division reflects the main changing trends of global scattering characteristics, such as... Figure 10 As shown, the findchangepts method found three boundary points. The initial points of the full-angle domain are respectively This allows us to divide the area into four regions. :

[0165] ;

[0166] Observation reveals that the third corner domain The overall difference is clearly very high, belonging to the region of drastic change in the target scattering mechanism, and is not suitable for being divided into a separate angular domain. The fourth angular domain... Although the overall dissimilarity is low and the transition is gradual, the large corner domain range leads to a slightly worse similarity result within the actual corner domain due to the accumulation of dissimilarity. Therefore, it is necessary to perform [further processing] on the coarsely divided corner domain. Figure 11 The fine division process is shown.

[0167] Step Six, for To avoid cumulative drift, such a large angular domain needs to be divided into two or more angular domains. Then, on the same azimuth grid, take the set of all indices belonging to that angular domain. Let the starting point index of this corner domain be . The starting reference difference sequence is defined as follows:

[0168] ;

[0169] This metric represents the "degree of deviation of scattering characteristics" between any azimuth angle within the angular domain and the starting azimuth angle. It exhibits a cumulative increasing trend and can clearly indicate the behavior of a gradual shift in the scattering mechanism caused by an excessively large azimuth angle span. This invention adopts the median difference principle, taking... The azimuth angle corresponding to the median value in The original corner domain is divided into two new corner domains using this as the new dividing point. , If any of the sub-corner domains still exceeds the judgment threshold, the above segmentation is recursively performed until all corner domains meet the condition.

[0170] After the above steps, the azimuth angle corresponding to the median difference is obtained. At this time, the original corner domain Become .

[0171] Step 7, after the long-angle domain is cut, the... Such high-discrepancy intervals, where the average discrepancy exceeds the global threshold, are not suitable as separate corner domains for classification because high discrepancy indicates drastic changes in the scattering mechanism. Instead, they should be assigned to adjacent corner domains to ensure the similarity of the scattering mechanism within each corner domain. Regarding the selection of the cut-off point, since it cannot be determined whether the discrepancy within the corner domain relative to the starting point is increasing, it is necessary to observe the discrepancy curve within that corner domain relative to the starting point. If the curve shows... Figure 12 If the change is monotonically alternating, then follow step six to find the dividing point and assign the two divided parts to adjacent regions respectively; if the change curve is as follows... Figure 13 If there is no monotonic variation, then a peak-cutting strategy should be adopted, that is, finding the global difference curve. Peak value within this angular domain As a dividing point, it is then assigned to adjacent regions.

[0172] Depend on Figure 12 and Figure 13 It can be seen that, This case does not exhibit a monotonic variation pattern; therefore, a peak-cutting strategy is employed for partitioning. After partitioning using this strategy, the following results are obtained. The final corner domain partitioning after steps seven and eight iterations is as follows: Figure 14 As shown.

[0173] ;

[0174] Step 8: Based on the final angular domain division results, the amplitude spectrum data corresponding to each azimuth angle are reorganized according to the polarization channel, and Gaussian noise is introduced to enhance the neural network input features.

[0175] Four polarization amplitude spectra for the same target:

[0176] , The number of frequency points is represented by a matrix arranged according to polarization, which serves as the four-dimensional input to the deep network classifier.

[0177] ;

[0178] Meanwhile, to enhance the robustness of the model under actual radar noise conditions, an additive white Gaussian noise model is applied to the above matrix:

[0179] ;

[0180] In the above formula, , The noise amplitude factor is usually taken as... This allows the noise amplitude to adaptively adjust relative to the sample itself, resulting in a four-polarization amplitude spectrum matrix after noise enhancement. This serves as the final input feature for the convolutional neural network model within the corresponding angular domain of that azimuth angle.

[0181] Step nine, based on the final corner region divided in step seven. The quadrature amplitude spectrum constructed in step eight Each azimuth angle is assigned to its corresponding angular domain, and each angular domain is independently constructed and trained. Figure 15 The convolutional neural network (CNN) classifier shown is used to achieve automatic target recognition based on angular domain prior knowledge.

[0182] According to azimuth Divide the interval with the final angular domain The relationship divides the samples into Subset:

[0183] ;

[0184] In the above formula This is the quadrature amplitude spectrum matrix after Gaussian noise enhancement. For the corresponding target category label. In each corner domain Within this process, the sample index is randomly shuffled and then adjusted according to a preset ratio. Within this corner domain, the data is randomly divided into training sets. With the validation set For angular regions with insufficient sample size that do not meet training requirements, additional data augmentation strategies can be employed to supplement the partitioning. Finally, the four-polarization amplitude spectrum matrix of each sample is reconstructed into a two-dimensional tensor. The first dimension consists of four polarization channels, the second dimension is the frequency point axis, and the third dimension is a single-channel "pseudo-space dimension," which satisfies the input format requirements of a two-dimensional convolutional network.

[0185] By training a CNN model using the training set, a set of network parameters corresponding to each corner domain can be obtained after training, forming a dedicated CNN classifier for that corner domain. Finally, the classifiers of each corner domain are combined to form a high-precision automatic radar target identification model based on corner domain prior knowledge.

[0186] Step 10: In the identification phase, for any radar echo sample to be identified, first determine its azimuth angle. Determine its corresponding angle range Then use the corresponding CNN model. Forward inference is performed on the four-polarization amplitude spectrum input features to obtain the output probabilities of each category, and category decision is made accordingly:

[0187] ;

[0188] Figure 16 The training results and loss curves for deep network feature learning without corner domain partitioning are shown. The accuracy of the validation set obtained after training is [missing value]. Furthermore, it can be clearly observed that a slight overfit phenomenon occurs as the number of training iterations increases.

[0189] Figures 17 to 20 These are the training result loss curves for each of the four corner regions after corner region partitioning, and the validation set accuracies obtained from the training are respectively... , , , The accuracy of the validation set obtained by weighting by the corner domain length is Compared to training without corner domain segmentation, the accuracy of unknown echo identification was improved. The improvement in left and right accuracy enhances recognition performance, and the accuracy curves of the four corner domains show no obvious overfitting phenomenon, indicating that this application can improve the performance of echo recognition.

[0190] By dividing the angular domain and learning CNNs independently within the angular domain, this application avoids the performance degradation of a single full-angular domain model when faced with drastic changes in scattering statistics over a wide azimuth angle. This allows each CNN model to fit only a relatively stable set of scattering patterns within its corresponding angular domain. The angular domain models together constitute a complete recognizer, realizing high-precision automatic radar target recognition based on angular domain prior knowledge under wide azimuth angle multi-polarization conditions.

[0191] It should be noted that the specific embodiments are merely explanations and illustrations of the technical solution of the present invention and should not be used to limit the scope of protection. Any modifications made in accordance with the claims and specification of the present invention that are only partial should still fall within the protection scope of the present invention.

Claims

1. A knowledge-based auxiliary and multi-pole fusion-based angular domain adaptive radar target recognition method, characterized in that The method comprises the following steps: The radar echo data to be identified and its corresponding azimuth angle are obtained, and the angle domain of the radar echo data is determined according to the azimuth angle, and the corresponding trained CNN model is called according to the angle domain, then the amplitude spectrum of the radar echo data is obtained, and finally the amplitude spectrum is input into the CNN model to obtain a classification result.

2. The knowledge-based auxiliary and multi-pole fusion based angular domain adaptive radar target recognition method according to claim 1, characterized in that The trained CNN model is obtained through the following steps: Step 1: Obtain radar echo data of multiple targets in multiple azimuth angles of four types of polarization channels under the same working wave band, and perform complex spectrum reconstruction on each radar echo data to obtain complex scattering spectrum and amplitude spectrum of each type of polarization channel under each azimuth angle; Step 2: For each complex scattering spectrum, a one-dimensional range image is obtained by inverse Fourier transform, and a scattering feature sensitive to the azimuth angle is extracted according to the one-dimensional range image; Step 3: The scattering features corresponding to each azimuth angle are spliced according to the dimension to obtain a multi-dimensional feature vector, and the multi-dimensional feature vectors corresponding to all azimuth angles of the same target and the same polarization are sorted, and the difference degree measurement based on the Euclidean distance is calculated for adjacent multi-dimensional feature vectors, and then a structural feature difference degree curve is obtained; Step 4: The correlation coefficient difference between HRRP sequences and complex scattering spectrums between adjacent azimuth angle samples is calculated respectively to obtain a correlation difference degree curve based on HRRP and a correlation difference degree curve based on complex scattering spectrum; Step 5: The structural feature difference degree curve, the correlation difference degree curve based on HRRP and the correlation difference degree curve based on complex scattering spectrum are weighted and superimposed to obtain a single polarization comprehensive difference degree curve corresponding to a single target; Step 6: Repeat the above steps to obtain a four polarization comprehensive difference degree curve corresponding to the same target, and the four polarization comprehensive difference degree curves corresponding to the same target are weighted and fused to obtain a comprehensive difference degree curve under the four polarization coupling condition corresponding to a single target; Step 7: Repeat the above steps to obtain a comprehensive difference degree curve under the four polarization coupling condition corresponding to all targets, and the comprehensive difference degree curves under the four polarization coupling condition corresponding to all targets are interpolated and amplitude normalized on the same azimuth angle grid, and are superimposed in an equal weight manner to obtain a global angle domain difference degree curve covering all target categories; Step 8: Based on the global angle domain difference degree curve covering all target categories, a change point detection method is used, and under the constraints of a preset maximum number of change points and a minimum angle domain width, a preliminary angle domain division result covering the full azimuth angle range is obtained; Step 9: Based on the preliminary angle domain division result covering the full azimuth angle range, and according to an angle threshold, an angle domain with too large span is determined, and the Euclidean distance difference between each point in the angle domain and the starting point is calculated in turn with the starting azimuth angle in the angle domain as a reference point, so as to obtain a difference degree sequence, and the angle domain is recursively divided according to the median position of the difference degree sequence; Step 10: Based on the result after recursive division, a high difference interval in which the average difference degree of each angle domain exceeds a global threshold is determined according to a set global threshold, and the high difference interval is divided by using a median division or peak value division strategy according to the difference degree change pattern to obtain a subinterval, and then the subinterval is integrated into a neighboring stable angle domain to obtain a final angle domain division result. Step 11: Splice the amplitude spectrum of all class polarization channels under each azimuth angle to obtain a four-polarization amplitude spectrum; Step 12: Group all four-polarization amplitude spectra according to the final angular domain division result, and train a CNN model using each group of four-polarization amplitude spectra.

3. The knowledge-based auxiliary and multi-pole fusion based angular domain adaptive radar target recognition method according to claim 2, characterized in that The complex scattering spectrum is represented as: ， wherein, denotes the complex scattering spectrum, denotes the complex scattering coefficient at frequency point , azimuth angle , polarization channel and target class , , is the total number of distance units, is the distance unit.

4. The knowledge-based auxiliary and multi-pole fusion based angular domain adaptive radar target recognition method according to claim 3, characterized in that The amplitude spectrum is represented as: ， wherein represents the amplitude spectrum.

5. The knowledge-based auxiliary and multi-pole fusion based angular domain adaptive radar target recognition method according to claim 4, characterized in that The sensitive scattering features include the number of strong scattering units , scattering center distance spread , energy proportion of the first 25% distance window , energy proportion of the first 50% distance window , energy proportion of the first band , energy proportion of the second band , energy proportion of the third band , HRRP frequency domain entropy , and HRRP first-order difference mean .

6. The knowledge-based auxiliary and multi-pole fusion based angular domain adaptive radar target recognition method according to claim 5, characterized in that In the sensitive scattering feature, Strong scattering unit number According to the set energy threshold And the distance unit whose amplitude exceeds the threshold is counted, which is expressed as: ， ， wherein, is the energy of the distance image of the first distance unit, is the total energy of the entire distance image, is the one-dimensional distance image discrete sequence after norm normalization; Scattering center distance spread is expressed as: ， ， ， wherein, is the distance to the axis, is the energy-weighted distance mean, is the energy-normalized weight; is represented as: ， is represented as: ， First band energy ratio is represented as: ， ， ， ， wherein, is the frequency energy, is the total energy of the spectrum of the entire one-dimensional range profile, is a discrete index on the spectrum of the HRRP, is the spectrum of the HRRP, ; Second band energy proportion is represented as: ， Third band energy proportion is represented as: ， wherein , , are three subbands equally divided in frequency domain index; HRRP frequency domain entropy is represented as: ， ， wherein, is the normalized frequency domain energy probability distribution, is a very small positive number, is the frequency spectrum sequence of the HRRP; HRRP first-order difference mean is represented as: ， ， wherein, is the sample value of the normalized one-dimensional distance image cell, is the degree of amplitude jump between the two adjacent distance cells.

7. The knowledge-based auxiliary and multi-pole fusion based angular domain adaptive radar target recognition method according to claim 6, characterized in that The structure feature difference degree curve is represented as: ， wherein, is a structure feature difference curve, is a structure feature vector of a sample of a known class , polarization mode , azimuth angle , a structure feature vector of a sample of a known class , polarization mode , azimuth angle , a structure feature vector of a sample of a known class , a structure feature vector of a sample of a known class , polarization mode , azimuth angle , a structure feature vector of a sample of a known class , a structure feature vector of a sample of a known class th feature value of the th feature among the structure feature vectors of the samples of the known class 8. The knowledge-based auxiliary and multi-pole fusion based angular domain adaptive radar target recognition method according to claim 7, characterized in that The correlation difference degree curve based on HRRP is represented as: ， wherein, is a correlation difference curve based on HRRP, is a known class , polarization mode , azimuth HRRP sequence of the sample, is an adjacent azimuth, is an adjacent azimuth Pearson correlation coefficient of the is a known class , polarization mode , azimuth value of the distance unit in the HRRP sequence of the sample, is a known class , polarization mode , azimuth value of the distance unit in the HRRP sequence of the sample after norm normalization.

9. The knowledge-based auxiliary and multi-pole fusion based angular domain adaptive radar target recognition method according to claim 8, characterized in that The correlation difference degree curve based on the complex scattering spectrum is represented as: ， wherein, is a complex correlation coefficient of the complex scattering spectrum based on a correlation difference degree curve, is a complex correlation coefficient of the complex scattering spectrum based on a correlation difference degree curve, is a complex correlation coefficient of the complex scattering spectrum based on a correlation difference degree curve, is a distance unit number representing a one-dimensional range profile, is a complex scattering coefficient of the complex scattering spectrum at the mth frequency point when the azimuth angle is is a complex scattering coefficient of the complex scattering spectrum at the mth frequency point when the azimuth angle is is a complex scattering coefficient of the complex scattering spectrum at the mth frequency point when the azimuth angle is is a complex scattering coefficient of the complex scattering spectrum at the mth frequency point when the azimuth angle is is a complex scattering coefficient of the complex scattering spectrum at the mth frequency point when the azimuth angle is is a complex scattering coefficient of the complex scattering spectrum at the mth frequency point when the azimuth angle is is a complex scattering coefficient of the complex scattering spectrum at the mth frequency point when the azimuth angle is is a complex scattering coefficient of the complex scattering spectrum at the mth frequency point when the azimuth angle is is a complex scattering coefficient of the complex scattering spectrum at the mth frequency point when the azimuth angle is 10. The knowledge-based auxiliary and multi-pole fusion based angular domain adaptive radar target recognition method according to claim 1, characterized in that The angle threshold in step 9 is 60 degrees.

Images