A polarized feature assisted signal sorting method and system
The polarization-assisted signal sorting method utilizes polarization features and neural network clustering to solve the problem of signal sorting in complex electromagnetic environments, achieving high-accuracy signal sorting and multipath signal identification, and improving sorting performance.
Patent Information
- Application Number
- CN202611455580.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-18
- Publication Date
- 2026-11-03
AI Technical Summary
Existing signal sorting technologies struggle to distinguish signals from multiple radars in complex electromagnetic environments. Multipath effects lead to an increase in false targets, traditional five-dimensional parameter sorting windows are difficult to stabilize, and the increased agility of modern radar parameters further complicates the sorting process.
A polarization-assisted signal sorting method is adopted. The polarization features and polarization feature vectors of radar signals are obtained through dual-polarized antennas. Clustering is performed using self-organizing feature mapping (SOM) neural network and improved fuzzy C-means FCM algorithm. Signal pairing and sorting are performed by combining five-dimensional parameters, and multipath signals are identified by polarization feature differences.
It improves the accuracy of signal sorting, reduces false targets, and enhances sorting performance in complex environments. In particular, in the case of multiple radars of the same type and multipath signals, the sorting accuracy has increased from about 72% to 95%, and the multipath signal recognition rate exceeds 90%.
Smart Images
Figure CN123030892A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of radar signal processing technology, specifically to a polarization-assisted signal sorting method and system. Background Technology
[0002] Signal sorting is one of the core technologies in electronic reconnaissance. Its purpose is to separate signals from different radiation sources from dense, overlapping radar pulse streams, providing a foundation for subsequent radiation source identification and threat assessment. With the increasing complexity of the modern electronic warfare environment, the density, modulation methods, and parameter agility of radar signals are constantly improving, placing higher demands on signal sorting technology.
[0003] Existing signal sorting technologies primarily utilize the five-dimensional parameters in the Pulse Description Word (PDW): Angle of Arrival (DOA), Carrier Frequency (CF), Pulse Width (PW), Time of Arrival (TOA), and Pulse Amplitude (PA), for signal sorting and deinterleaving. Commonly used sorting algorithms include histogram-based methods based on time of arrival (such as SDIF and CDIF), multi-parameter joint clustering methods (such as K-means and DBSCAN), and intelligent sorting methods based on neural networks. These methods exhibit good sorting performance under normal electromagnetic environments.
[0004] However, existing technologies have the following shortcomings: First, in complex electromagnetic environments, the signals from multiple radars severely overlap in the five-dimensional parameter space. In particular, when multiple radars of the same type or with similar parameters operate simultaneously, their parameters such as angle of arrival, carrier frequency, and pulse width are highly similar, causing a sharp decrease in the distinguishability of traditional five-dimensional parameters and a significant reduction in sorting accuracy.
[0005] Second, multipath effect is a significant factor affecting signal sorting. Radar signals are reflected by the ground, buildings, sea surface, etc., forming multipath signals. These multipath signals are highly similar to direct wave signals in parameters such as angle of arrival, carrier frequency, and pulse width, making them easily misidentified as new radiation sources, resulting in a large number of false targets in the sorting results.
[0006] Third, modern multi-functional radars widely employ parameter agility technologies such as frequency agility, pulse group agility, and repetition rate variation, making it difficult to stably establish the sorting window of traditional five-dimensional parameters, further increasing the difficulty of signal sorting. Summary of the Invention
[0007] To address the shortcomings of existing technologies that only employ two-dimensional polarization parameters, supervised single-level classifiers, and lack multipath elimination mechanisms, this invention proposes a polarization feature-assisted signal sorting method and signal sorting system.
[0008] To solve the above-mentioned technical problems, the present invention adopts the following solution: A polarization-assisted signal sorting method includes: S1: Obtain the five-dimensional parameters of radar signal, including angle of arrival, carrier frequency, pulse width, time of arrival, and pulse amplitude. Receive radar signal through a dual-polarized antenna, process the radar signal, and obtain polarization features and polarization feature vectors. S2: Using the polarization feature vector as the input of the self-organizing feature map (SOM) neural network, train the SOM neural network to map radar signals with similar polarization features to the same output neuron, wherein the polarization features are the first polarization feature and / or the second polarization feature, and output polarization feature category labels. S3: Using the mean polarization feature corresponding to the polarization feature category label as the initial cluster center, construct an improved fuzzy C-means FCM objective function that introduces polarization physical constraints, refine the polarization feature vectors, and output the refined polarization clustering results. S4: Based on the refined polarization clustering results, the radar signal is paired and sorted using five-dimensional parameters: angle of arrival, carrier frequency, pulse width, time of arrival, and pulse amplitude. The final sorting result containing the pulse descriptor word (PDW) is output.
[0009] Furthermore, when the radar signal is a fixed polarization signal, the polarization feature is a first polarization feature, and the polarization feature vector includes the first polarization feature vector; Extract the horizontal and vertical polarization components of the fixed polarization signal, calculate the complete polarization feature vector containing all four Stokes parameters, and use it as the first polarization feature vector. Extract the instantaneous Stokes parameter vector as the first polarization feature.
[0010] Furthermore, when the radar signal is a fixed polarization signal and a variable polarization signal, the polarization feature is a first polarization feature and a second polarization feature, and the polarization feature vector is a first polarization feature vector and a second polarization feature vector. Extract the horizontal and vertical polarization components of the fixed polarization signal, calculate the complete polarization feature vector containing all four Stokes parameters, and use it as the first polarization feature vector; extract the instantaneous Stokes parameter vector as the first polarization feature; extract the continuous polarization state sequence of the variable polarization radar signal, extract the statistical features of the polarization state sequence, use the statistical features as the second polarization feature vector, and construct the polarization state sequence feature matrix as the second polarization feature.
[0011] Furthermore, in S1, the complete polarization eigenvector containing all four Stokes parameters is calculated, specifically as follows: Calculate the four Stokes parameters: S = [S0, S1, S2, S3] T ; After normalization, the polarization eigenvectors are obtained: P = [S1 / S0, S2 / S0, S3 / S0]; in, , , , ; and These are the complex envelopes of the horizontal and vertical polarization components, respectively. Indicates taking and The real part, Indicates taking and The imaginary part, This indicates the conjugate operation. This indicates the transpose operation.
[0012] Furthermore, S2 specifically includes: S201: SOM network initialization sub-step: Determine the topology and number of neurons in the SOM output layer, and initialize the weight vector of each neuron; S202: Polarization feature input sub-step: Normalize the complete polarization feature vector and input it into the SOM network; S203: Competitive learning sub-step: Calculate the Euclidean distance between the input vector and the weight vectors of each neuron, select the neuron with the smallest distance as the winning neuron, and update the weight vectors of the winning neuron and its neighboring neurons. S204: Category Mapping Sub-step: After training convergence, signals mapped to the same output neuron are grouped into one category, and the polarization feature category label is output.
[0013] Furthermore, in S203, the formula for updating the weight vector is: ; in, For the first j One neuron in t The weight vector at time +1; For the first j One neuron in t The weight vector at time step, for t The input polarization eigenvector at time t. The learning rate decays over training time. To win neurons c The neighborhood function centered at the center.
[0014] Furthermore, S3 specifically includes: S301: The number of polarization categories in S2 is used as the number of FCM clusters C, and the mean value of the polarization features of each category is used as the initial cluster center; S302: Construct an FCM objective function that incorporates polarization physical constraints, including polarization orthogonality constraints and polarization similarity constraints; S303: Iteratively optimize the objective function using the Lagrange multiplier method, and update the membership matrix and cluster center matrix; S304: Terminate the iteration when the change in the objective function is less than the preset threshold or the maximum number of iterations is reached, and output the refined clustering results.
[0015] Furthermore, in S3, the expression for the improved fuzzy C-means FCM objective function is: ; in, C The number of clusters, N The number of signal samples. For the first k The nth sample pair Membership degree of a class m For fuzzy index, This indicates taking the membership degree. m The power of the result is used as the weight. For the first k The sample to the first Euclidean distance between cluster centers Indicates the first k The sample to the first The square of the Euclidean distance between the cluster centers This means that the farther a sample is from the cluster center and the higher its membership degree, the greater the penalty to the objective function. and The first Class and First The normalized Stokes parameter mean vector of the class. For the first k The sample and the first Polarization similarity of polarization feature templates and These are the polarization orthogonality constraint weights and the polarization similarity constraint weights, respectively.
[0016] Furthermore, it also includes S5: using the difference in polarization characteristics to identify multipath signals, and identifying radar signals whose polarization state is inconsistent with the polarization state of the direct wave as multipath signals and eliminating them; Specifically, the polarization state difference between each radar signal and its neighboring signals is calculated. If the polarization state difference is less than a preset threshold and the signal parameters satisfy the multipath delay relationship, the radar signal is marked as a multipath signal and removed from the sorting results.
[0017] A polarization-assisted signal sorting system includes: The polarization feature extraction module is configured to acquire five-dimensional parameters of radar signal, including angle of arrival, carrier frequency, pulse width, time of arrival, and pulse amplitude. It receives radar signal through a dual-polarized antenna, processes the radar signal, and obtains polarization features and polarization feature vectors. The polarization-assisted SOM pre-sorting module is configured to use polarization feature vectors as input to the self-organizing feature map (SOM) neural network, train the SOM neural network to map radar signals with similar polarization features to the same output neuron, wherein the polarization features are a first polarization feature and / or a second polarization feature, and output polarization feature category labels. An improved FCM clustering refinement module is configured to use the mean of polarization features corresponding to the polarization feature category labels as the initial cluster centers. An improved fuzzy C-means FCM objective function with polarization physical constraints is constructed to refine the polarization feature vectors and output the refined polarization clustering results. The five-dimensional parameter fine sorting module is configured to perform signal pairing and sorting based on the refined polarization clustering results, using the radar signal's angle of arrival, carrier frequency, pulse width, time of arrival, and pulse amplitude. It outputs the final sorting results, which include the pulse descriptor word (PDW).
[0018] Also includes: The multipath rejection module is configured to identify multipath signals by utilizing differences in polarization characteristics, and to identify signals whose polarization state is inconsistent with that of the direct wave as multipath signals and reject them. Specifically, the polarization state difference between each radar signal and its neighboring signals is calculated. If the polarization state difference is less than a preset threshold and the signal parameters satisfy the multipath delay relationship, the radar signal is marked as a multipath signal and removed from the sorting results.
[0019] The beneficial effects of this invention are: First, this invention uses four complete Stokes parameters (S0, S1, S2, S3) to characterize the polarization state of radar signals, unlike existing technologies that only use two parameters: polarization amplitude ratio and polarization phase difference. The complete Stokes parameters can describe any polarization state (including partially polarized and fully polarized waves), providing a richer and more refined polarization discrimination capability than existing two-dimensional polarization parameters. In simulation experiments, when the five-dimensional parameters of multiple identical radars highly overlap, the sorting accuracy increased from approximately 72% to approximately 95% after introducing the complete Stokes polarization features.
[0020] Second, this invention adopts a hierarchical cascaded sorting architecture based on unsupervised learning, which differs from the existing technology that uses a supervised single-level classifier. The unsupervised approach does not require labeled training data, making it more suitable for real-world electronic reconnaissance scenarios. The cascaded architecture fully utilizes the independent guiding role of polarization features, first performing coarse sorting based on polarization features, and then performing fine sorting using traditional parameters based on the coarse sorting results, avoiding the problem of polarization information being overwhelmed due to direct feature concatenation.
[0021] Third, this invention introduces polarization physical constraints into the traditional FCM algorithm, including polarization orthogonality constraints and polarization similarity constraints. The polarization orthogonality constraint utilizes the physical law that the polarization states of different radar signals typically possess orthogonal or approximately orthogonal characteristics, forcing cluster centers of different categories to maintain a reasonable sorting degree in the polarization space. The polarization similarity constraint guides the polarization features of signals of the same category to approach the polarization template of that category. The introduction of these two constraints makes the clustering results more consistent with the physical nature of polarization features, improving the clustering accuracy by approximately 15% compared to traditional FCM.
[0022] Fourth, this invention utilizes differences in polarization characteristics to identify and eliminate multipath signals. Since multipath reflection causes significant changes in the signal polarization state (e.g., horizontal polarization becomes elliptical polarization after reflection), and traditional five-dimensional parameters are almost insensitive to this, this invention can effectively distinguish between direct waves and multipath signals. In typical multipath environments, the multipath signal identification rate exceeds 90%, significantly reducing the number of false targets in the sorting results.
[0023] Fifth, this invention supports both fixed-polarization and variable-polarization signals. For variable-polarization radar signals, by constructing a polarization state sequence matrix and extracting sequence statistical features, the time-varying patterns of polarization states can be effectively captured, thus expanding the applicability of the method. Attached Figure Description
[0024] Figure 1 This is a flowchart illustrating the overall process of the polarization-feature-assisted signal sorting method of the present invention. Figure 2 This is a flowchart of the polarization-assisted SOM pre-sorting module of the present invention; Figure 3 This is a flowchart of the multipath rejection module of the present invention; Figure 4 The graph shows a comparison of the sorting performance of the method of this invention with that of existing technologies. Figure 5 The graph shows a comparison of the performance improvement results between the method of this invention and the prior art. Figure 6 The diagram shows the comparison of the contribution of the sorting accuracy decomposition between the method of the present invention and the prior art. Detailed Implementation
[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0026] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values of the components and steps described in these embodiments do not limit the scope of the invention.
[0027] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.
[0028] Furthermore, for clarity and brevity, descriptions of well-known structures, functions, and configurations may have been omitted. Those skilled in the art will recognize that various changes and modifications can be made to the examples described herein without departing from the spirit and scope of this disclosure.
[0029] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.
[0030] In all examples shown and discussed herein, any specific values should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.
[0031] The present invention will now be described in detail with reference to the accompanying drawings and embodiments: Example 1 This embodiment uses the example of sorting signals from multiple identical fixed-polarization radars in a complex electromagnetic environment to illustrate the technical solution of the present invention in detail.
[0032] Four radar signals coexist in an electromagnetic environment. The angle of arrival (DOA), carrier frequency (CF), and pulse width (PW) parameters in their pulse description words (PDW) are set as follows: Radar A (DOA=15.2°, CF=9.41GHz, PW=12.3μs), Radar B (DOA=15.3°, CF=9.38GHz, PW=11.9μs), Radar C (DOA=15.1°, CF=9.45GHz, PW=12.1μs), and Radar D (DOA=15.4°, CF=9.36GHz, PW=12.5μs). The time of arrival (TOA) of each radar overlaps randomly, the signal density is 500,000 pulses per second, and the signal-to-noise ratio is 15dB.
[0033] The five-dimensional parameters of the above four radars are highly overlapping, making it difficult for traditional sorting methods based on PDW parameters (such as SDIF histogram method and K-means clustering method) to effectively distinguish them. The sorting accuracy is usually less than 75%, and it is very easy to misclassify the pulse of one radar into the pulse sequence of another radar, resulting in additional or missing batches.
[0034] This embodiment utilizes the different polarization states of each radar to achieve signal sorting. The polarization states of the four radars are as follows: radar A is horizontally polarized, radar B is vertically polarized, radar C is left-hand circularly polarized, and radar D is 45° linearly polarized. These polarization states have different representations in the Stokes parameter space, providing a physical basis for the hierarchical sorting of this invention.
[0035] like Figure 1 As shown, S1: Polarization Feature Extraction: An orthogonal dual-polarized antenna (polarization isolation greater than 30dB, operating frequency band 2-18GHz) is used to simultaneously receive the horizontal and vertical polarization components of the signal. The dual-channel receiver performs synchronous sampling at a sampling rate of 2GS / s, with a quantization bit depth of 12 bits.
[0036] For each pulse signal, extract the intra-pulse sampling data and obtain the complex envelope of the horizontal polarization component. Complex envelope of vertical polarization component According to the formula S = [S0, S1, S2, S3] T Calculate the complete Stokes parameter vector, where: , representing the total signal power; This represents the difference between the horizontal and vertical polarization components; , representing the difference between the +45° and -45° linear polarization components; , representing the difference between the right-handed and left-handed circular polarization components.
[0037] in, Indicates taking the real part, This indicates taking the imaginary part. This indicates the conjugate operation. This indicates the transpose operation.
[0038] Power normalization of the Stokes parameter vector yields the polarization eigenvector P=[S1 / S0,S2 / S0,S3 / S0], which lies on the unit Poincaré sphere and fully characterizes the polarization state of the signal.
[0039] Existing technologies only extract two parameters: polarization amplitude ratio and polarization phase difference. On the Poincaré sphere, only one curve can be determined, failing to uniquely identify the polarization state. This invention employs complete three-dimensional normalized Stokes parameters, defining a unique point on the Poincaré sphere. This achieves an unambiguous description of any polarization state (including fully polarized and partially polarized waves), laying the foundation for subsequent high-precision sorting.
[0040] S2: Polarization-assisted SOM pre-sorting: such as Figure 2 As shown, it includes the following steps: SOM network initialization sub-step: Determine the topology and number of neurons in the SOM output layer, and initialize the weight vector of each neuron; Polarization feature input sub-step: Normalize the complete polarization feature vector and input it into the SOM network; Competitive learning sub-step: Calculate the Euclidean distance between the input vector and the weight vector of each neuron, select the neuron with the smallest distance as the winning neuron, and update the weight vectors of the winning neuron and its neighboring neurons; Class mapping sub-step: After training convergence, classify the signals mapped to the same output neuron into one class and output the polarization feature class label.
[0041] A self-organizing feature map (SOM) neural network was constructed, with the output layer adopting a two-dimensional hexagonal topology and a size of 10×10, totaling 100 neurons. This size achieves a good balance between clustering resolution and computational complexity. The weight vectors of each neuron were randomly initialized.
[0042] Input the normalized polarization feature vector P obtained from S1 into the SOM network. For each input vector, calculate its Euclidean distance to the weight vectors of all neurons, and select the neuron with the smallest distance as the winning neuron. Update the weights of the winning neuron and its neighboring neurons according to the following rules: ; in, For the first j One neuron in t The weight vector at each time step; for t The input polarization eigenvector at time t; The learning rate decays over training time. The value decays exponentially from 0.1 to 0.01 over time. To win neurons c The neighborhood function centered on the center A Gaussian function is used; the neighborhood radius is gradually reduced from the initial 5 neurons to 0. This parameter decay strategy ensures that the network converges quickly in the early stages of training and is finely adjusted in the later stages.
[0043] After training convergence, signals mapped to the same output neuron or adjacent neurons are grouped into one category, and polarization feature category labels are output. In this embodiment, the SOM network automatically divides signals into 5 polarization categories. Specifically: Radar B (vertical linear polarization, S1≈-1) and Radar A (horizontal linear polarization, S1≈+1) are located at opposite poles on the S1 axis, each forming an independent cluster; Radar C (left-hand circular polarization, S3≈-1) forms an independent cluster on the S3 axis; Radar A and Radar D (45° linear polarization, S2≈+1), although both are linearly polarized, have different projections on the S2 axis and are mapped to adjacent but different neuron regions by the topology-preserving property of the SOM network, forming two independent categories; the fifth category consists of a small number of boundary signals in the polarization transition region.
[0044] The unsupervised learning nature of the SOM network allows it to automatically discover cluster structures in the polarization feature space without any prior labeled data, which aligns perfectly with the reality of electronic reconnaissance, where labeled samples are often lacking. Simultaneously, the topology-preserving property of SOM ensures that classes with similar polarization states are adjacent in the output layer, providing favorable initial conditions for subsequent cluster refinement. This invention does not rely on a large number of labeled samples for model training.
[0045] S3: Improve FCM clustering refinement: Use the 5 polarization categories output in step S2 as the initial number of clusters C=5 for FCM, and use the mean of the polarization features of each category as the initial cluster centers. Set the fuzziness index m=2.
[0046] Construct an improved FCM objective function that incorporates polarization physical constraints: The expression for the objective function of the improved fuzzy C-means FCM is: ; in, C The number of clusters, N The number of signal samples. For the first k The nth sample pair Membership degree of a class m For fuzzy index, This indicates taking the membership degree. m The power of the result is used as the weight. For the first k The sample to the first Euclidean distance between cluster centers Indicates the first k The sample to the first The square of the Euclidean distance between the cluster centers This means that the farther a sample is from the cluster center and the higher its membership degree, the greater the penalty to the objective function. and The first Class and First The normalized Stokes parameter mean vector of the class. For the first k The sample and the first Polarization similarity of polarization feature templates and These are the polarization orthogonality constraint weights and the polarization similarity constraint weights, respectively.
[0047] The objective function consists of three parts: The first term is the weighted sum of squared errors of the standard FCM, which ensures that the clustering results have a good fit to the data itself.
[0048] The second term is the polarization orthogonality constraint term. and The first Class and First j The normalized Stokes parameter mean vector for each class penalizes cluster centers of different classes that are too close together in polarization space. The physical basis for this is that due to differences in antenna design, mounting platform, and signal system, the polarization states of different radar sources are usually orthogonal or approximately orthogonal, tending towards a radial distribution on the Poincaré sphere. Weights The value was set to 0.3. This value was determined experimentally. If it is too small, the constraint effect will be insignificant, and if it is too large, it may interfere with the distribution characteristics of the data itself.
[0049] The third term is the polarization similarity constraint term. For the first k The sample and the first The polarization similarity of the polarization feature templates (measured by the cosine of the angle between the Stokes vectors) guides the polarization features of signals of the same class to move closer to the same polarization template, enhancing the intra-class compactness of the clustering results. Weights Set to 0.2. In this embodiment, the polarization feature template is preferably the mean vector of polarization features of the current category, used to measure the similarity between a sample and its category.
[0050] The objective function is iteratively optimized using the Lagrange multiplier method, with membership degrees updated alternately. And the cluster center matrix. The iteration termination condition is: the change in the objective function value between two consecutive iterations is less than 10⁻ 4 Or the number of iterations reaches the preset maximum number of iterations, 100 times.
[0051] In this embodiment, the objective function converges after 12 iterations. The refined clustering results correctly classify radar A (horizontal linear polarization) and radar D (45° linear polarization) into two independent categories, ultimately outputting 5 polarization clusters. Compared with the pre-sorting results in step S2, the membership assignment of the boundary signals is more accurate.
[0052] Traditional FCM algorithms cluster data solely based on data distribution, which can easily lead to confusion when different categories have similar polarization states. This invention introduces polarization orthogonality constraints, leveraging the physical principle in electromagnetism that the polarization states of different radiation sources tend to be orthogonal. Polarization similarity constraints further enhance intra-cluster consistency; together, they make the clustering results more consistent with the physical nature of polarization. Experiments show that introducing polarization physical constraints improves clustering accuracy by approximately 15% compared to traditional FCM.
[0053] S4: Five-dimensional parameter fine sorting: Within each polarization cluster output in step S3, the pulse repetition interval (PRI) of the pulse sequence is extracted using the cumulative difference histogram method based on arrival time (CDIF / SDIF). Specifically, the pulses within the cluster are sorted by arrival time, the time difference between adjacent pulses is calculated, and the difference histograms at each level are statistically analyzed. The difference corresponding to the peak value exceeding the preset threshold is the candidate PRI value.
[0054] Pulse sequence searching is performed based on the extracted PRI values, and joint matching verification is performed using carrier frequency and pulse width parameters. The matching criterion is that the difference between the carrier frequency and pulse width of the candidate pulse and the central value of the parameters of the established pulse sequence is within a preset tolerance range (in this embodiment, the carrier frequency tolerance is ±5MHz, and the pulse width tolerance is ±0.5μs). For each polarization cluster, pulse sequences of each radiation source are established, and the final sorting result containing a complete pulse descriptor word (PDW) is output.
[0055] S5: Multipath Removal Steps: (e.g.) Figure 3 As shown, the polarization state difference between each signal and signals within the same polarization cluster whose arrival angle difference is less than a first preset threshold is calculated. If the polarization state difference is greater than a second preset threshold (indicating that reflection causes a change in polarization) and the signal parameters (such as arrival time difference) satisfy a predetermined multipath delay relationship, then the signal with a larger delay is marked as a multipath signal and removed; otherwise, the signal is retained as a normal signal, and the signal that does not meet the conditions is retained.
[0056] For each signal sorted in step S4, calculate the polarization state difference between each signal and its neighboring signals within the same polarization cluster (e.g., signals with an angle of arrival difference less than a preset threshold, such as 3°). ,in and These are the normalized Stokes polarization eigenvectors of the two signals, respectively. The value ranges from [0,2], and the smaller the value, the more similar the polarization states are.
[0057] Multipath determination is performed based on the time delay relationship of signal parameters. The preset conditions are: (1) polarization state difference. (1) Less than the preset threshold of 0.15 (this threshold is determined by Monte Carlo simulation and can distinguish between direct waves and multipath waves with a 95% confidence level); (2) The arrival time difference between the candidate multipath signal and the direct signal. satisfy ,in To predict multipath delay based on scene geometry, For tolerance.
[0058] Signals that simultaneously meet both of the above conditions are marked as multipath signals and removed from the sorting results. Once multipath signals are removed, they will no longer participate in subsequent radiation source identification and threat assessment processes.
[0059] Multipath reflections (such as ground reflections, sea surface reflections, and building reflections) cause significant changes in the polarization state of signals. For example, a horizontally polarized signal becomes elliptical polarized after reflection due to depolarization. This invention utilizes this physical property to effectively distinguish between direct waves and multipath waves through polarization state differences. Traditional five-dimensional parameters (DOA, CF, PW, etc.) are almost insensitive to multipath effects, thus failing to achieve effective multipath identification. The multipath elimination step of this invention fills this technological gap. It should be noted that the polarization changes caused by multipath reflections usually manifest as depolarization or partial polarization rotation. After S3 clustering refinement, multipath signals may still fall into the polarization cluster of the direct wave or its neighboring clusters. Therefore, multipath signals with altered polarization states can be identified within polarization clusters through polarization state differences.
[0060] Experimental results show that the average sorting accuracy of four identical radar signals obtained using the method of this invention reaches 96.3%, while the sorting accuracy of the control experiment using the traditional five-dimensional parameter combined with the SDIF method is only 71.8%, representing an improvement of approximately 24.5 percentage points. The multipath signal recognition rate reaches 91.5%, and the number of false targets in the sorting results is reduced by approximately 86% compared to the traditional method.
[0061] Example 2 This embodiment uses a mixed scenario containing variable polarization radar signals as an example to illustrate the adaptability of the present invention to variable polarization signals.
[0062] Three radar signals exist in the electromagnetic environment: Radar A is fixed horizontally polarized, Radar B is fixed right-hand circularly polarized, and Radar C is a variable polarization radar, whose polarization state alternates between horizontal and vertical polarization at a period of 100 μs. The signal density is 300,000 pulses per second, and the signal-to-noise ratio is 12 dB.
[0063] The polarization state of a variable polarization signal changes dynamically over time. If only instantaneous Stokes parameters are extracted as polarization features, pulses from the same variable polarization radar at different times will be mapped to different positions on the Poincaré sphere. This can easily lead to incorrect classification into multiple radiation sources during clustering, causing batching problems. This invention solves this technical problem by extracting statistical features of the polarization state sequence.
[0064] Step S1: Polarization Feature Extraction For radars A and B with fixed polarization, the instantaneous normalized Stokes parameter vector is extracted as polarization features, similar to Example 1.
[0065] For a variable polarization radar C, using pulses as the time unit, the polarization state sequence of L consecutive pulses is extracted (in this embodiment, L is 200, corresponding to approximately 20 complete polarization switching cycles), and a polarization state sequence matrix M=[P1,P2,...,P] is constructed. L The following statistical features are extracted from the sequence matrix as the polarization features of radar C: Mean polarization state: reflects the equivalent central polarization state of a variable polarization signal; Polarization state variance: reflects the degree of dispersion of the polarization state around the mean, and is used to distinguish between fixed polarization signals (variance close to zero) and variable polarization signals (variance is relatively large). Switching frequency: By performing a Fourier transform on the polarization state time series, the main frequency component is extracted to reflect the switching period of the polarization state; Dwell time: The number of continuous pulses of the statistical signal in each polarization state (horizontal polarization and vertical polarization in this embodiment), reflecting the temporal structure of the polarization state.
[0066] The instantaneous Stokes parameter vector with fixed polarization and the sequence statistical features with variable polarization are combined to form a polarization feature vector, which is then input into the subsequent processing stage.
[0067] The execution process of SOM pre-sorting (S2), improved FCM clustering refinement (S3), five-dimensional parameter fine sorting (S4) and multipath elimination (S5) is basically the same as in Example 1, and will not be repeated here.
[0068] It should be noted that due to the variable polarization characteristics of radar C, its polarization feature vector is mapped to an independent output neuron region during the SOM pre-sorting stage, and is explicitly sorted from radars A and B. In the improved FCM clustering refinement stage, radar C, due to its large polarization state variance feature, exhibits a high polarization similarity to fixed polarization signals. If the value is low, the polarization similarity constraint automatically guides it into a separate cluster.
[0069] The average sorting accuracy of the three radar signals was 97.8%, with the sorting accuracy of the variable polarization radar C being 96.5%, demonstrating that the present invention has excellent adaptability and sorting accuracy for both fixed polarization and variable polarization signals.
[0070] Comparative Example like Figures 4-6 As shown, to verify the contribution of each technical feature of the present invention, an ablation experiment was conducted, and the results are shown in Table 1 below: Table 1 Comparison of Ablation Experiment Results
[0071] Based on the experimental results in the table above, it can be seen that the complete Stokes parameters improve accuracy by approximately 7 percentage points compared to the two-dimensional polarization parameters; the improved FCM with the introduction of polarization physical constraints improves accuracy by approximately 7 percentage points compared to the standard FCM; and the multipath elimination step increases the multipath recognition rate from less than 30% to 91.5%. These technical features work synergistically to achieve the beneficial effects of this invention.
[0072] The system of this invention is implemented based on a heterogeneous computing architecture of FPGA+DSP+GPU. The FPGA uses a Xilinx Kintex-7 series processor, responsible for performing 2GS / s dual-channel synchronous sampling and real-time pipelined calculation of front-end Stokes parameters; the DSP processor is responsible for the training and inference of the polarization-assisted SOM pre-sorting module; and the GPU is responsible for the parallel iterative computation of the improved FCM objective function, achieving fast clustering under polarization physical constraints. The total system processing latency is less than 1ms, providing real-time sorting capability. The core algorithm is implemented based on C++ and CUDA, providing a standard API interface that can be seamlessly integrated with existing electronic reconnaissance systems.
[0073] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
[0074] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Based on the technical essence of the present invention, any simple modifications, equivalent substitutions, and improvements made to the above embodiments within the spirit and principles of the present invention shall still fall within the protection scope of the present invention.
Claims
1. A polarization-characteristic-assisted signal sorting method, characterized in that, Includes the following steps: S1: Obtain the five-dimensional parameters of radar signal, including angle of arrival, carrier frequency, pulse width, time of arrival, and pulse amplitude. Receive radar signal through a dual-polarized antenna, process the radar signal, and obtain polarization features and polarization feature vectors. S2: Using polarization feature vectors as input to the self-organizing feature map (SOM) neural network, the SOM neural network is trained to map radar signals with similar polarization features to the same output neuron and output polarization feature category labels. S3: Using the mean polarization feature corresponding to the polarization feature category label as the initial cluster center, construct an improved fuzzy C-means FCM objective function that introduces polarization physical constraints, refine the polarization feature vectors, and output the refined polarization clustering results. S4: Based on the refined polarization clustering results, the radar signal is paired and sorted using five-dimensional parameters: angle of arrival, carrier frequency, pulse width, time of arrival, and pulse amplitude. The final sorting result containing the pulse descriptor word (PDW) is output.
2. The polarization-assisted signal sorting method according to claim 1, characterized in that, When the radar signal is a fixed polarization signal, the polarization feature is a first polarization feature, and the polarization feature vector includes the first polarization feature vector; Extract the horizontal and vertical polarization components of the fixed polarization signal, calculate the complete polarization feature vector containing all four Stokes parameters, and use it as the first polarization feature vector. Extract the instantaneous Stokes parameter vector as the first polarization feature.
3. The polarization-assisted signal sorting method according to claim 1, characterized in that, When the radar signal is a fixed polarization signal and a variable polarization signal, the polarization feature is a first polarization feature and a second polarization feature, and the polarization feature vector is a first polarization feature vector and a second polarization feature vector; Extract the horizontal and vertical polarization components of the fixed polarization signal, calculate the complete polarization feature vector containing all four Stokes parameters, and use it as the first polarization feature vector; extract the instantaneous Stokes parameter vector as the first polarization feature. Extract the continuous polarization state sequence of the variable polarization radar signal, extract the statistical features of the polarization state sequence, use the statistical features as the second polarization feature vector, and construct the polarization state sequence feature matrix as the second polarization feature.
4. The polarization-assisted signal sorting method according to claim 1, characterized in that, The complete polarization eigenvector containing all four Stokes parameters is calculated as follows: Calculate the four Stokes parameters: S = [S0, S1, S2, S3] T ; After normalization, the polarization eigenvector is obtained: P = [S1 / S0, S2 / S0, S3 / S0]; in, , , , ; and These are the complex envelopes of the horizontal and vertical polarization components, respectively. Indicates taking and The real part, Indicates taking and The imaginary part, This indicates the conjugate operation. This indicates the transpose operation.
5. The polarization-assisted signal sorting method according to claim 1, characterized in that, S2 specifically includes: S201: Determine the topology and number of neurons in the SOM output layer, and initialize the weight vector of each neuron; S202: Normalize the complete polarization eigenvectors and input them into the SOM network; S203: Calculate the Euclidean distance between the input vector and the weight vectors of each neuron, select the neuron with the smallest distance as the winning neuron, and update the weight vectors of the winning neuron and its neighboring neurons. S204: After training converges, signals mapped to the same output neuron are grouped into one category, and the polarization feature category label is output.
6. The polarization-assisted signal sorting method according to claim 5, characterized in that, In S203, the formula for updating the weight vector is: ； in, For the first j One neuron in t The weight vector at time +1; For the first j One neuron in t The weight vector at time step, for t The input polarization eigenvector at time t. The learning rate decays over training time. To win neurons c The neighborhood function centered at the center.
7. The polarization-assisted signal sorting method according to claim 1, characterized in that, S3 specifically includes: S301: The number of polarization categories in S2 is used as the number of FCM clusters C, and the mean value of the polarization features of each category is used as the initial cluster center; S302: Construct an FCM objective function that incorporates polarization physical constraints, including polarization orthogonality constraints and polarization similarity constraints; S303: Iteratively optimize the objective function using the Lagrange multiplier method, and update the membership matrix and cluster center matrix; S304: Terminate the iteration when the change in the objective function is less than the preset threshold or the maximum number of iterations is reached, and output the refined clustering results.
8. The polarization-assisted signal sorting method according to claim 7, characterized in that, In S3, the expression for the objective function of the improved fuzzy C-means FCM is: ； in, C The number of clusters, N The number of signal samples. For the first k The nth sample pair Membership degree of a class m For fuzzy index, This indicates taking the membership degree. m The power of the result is used as the weight. For the first k The sample to the first Euclidean distance between cluster centers Indicates the first k The sample to the first The square of the Euclidean distance between the cluster centers This means that the farther a sample is from the cluster center and the higher its membership degree, the greater the penalty to the objective function. and The first Class and First The normalized Stokes parameter mean vector of the class. For the first k The sample and the first Polarization similarity of polarization feature templates and These are the polarization orthogonality constraint weights and the polarization similarity constraint weights, respectively.
9. The polarization-assisted signal sorting method according to claim 1, characterized in that, It also includes S5: using polarization feature differences to identify multipath signals, and identifying radar signals whose polarization state is inconsistent with the polarization state of the direct wave as multipath signals and eliminating them; Specifically, the polarization state difference between each radar signal and its neighboring signals is calculated. If the polarization state difference is less than a preset threshold and the signal parameters satisfy the multipath delay relationship, the radar signal is marked as a multipath signal and removed from the sorting results.
10. A polarization-assisted signal sorting system, characterized in that, include: The polarization feature extraction module is configured to acquire five-dimensional parameters of radar signal, including angle of arrival, carrier frequency, pulse width, time of arrival, and pulse amplitude. It receives radar signal through a dual-polarized antenna, processes the radar signal, and obtains polarization feature vector and polarization features. The polarization-assisted SOM pre-sorting module is configured to use polarization feature vectors as input to the self-organizing feature map (SOM) neural network, train the SOM neural network to map radar signals with similar polarization features to the same output neuron, wherein the polarization features are a first polarization feature and / or a second polarization feature, and output polarization feature category labels. An improved FCM clustering refinement module is configured to use the mean of polarization features corresponding to the polarization feature category labels as the initial cluster centers. An improved fuzzy C-means FCM objective function with polarization physical constraints is constructed to refine the polarization feature vectors and output the refined polarization clustering results. The five-dimensional parameter fine sorting module is configured to perform signal pairing and sorting based on the refined polarization clustering results, using the five-dimensional parameters of radar signal's angle of arrival, carrier frequency, pulse width, time of arrival, and pulse amplitude, and outputs the final sorting result containing the pulse descriptor word (PDW). The multipath rejection module is configured to identify multipath signals by utilizing differences in polarization characteristics, and to identify signals whose polarization state is inconsistent with that of the direct wave as multipath signals and reject them. Specifically, the polarization state difference between each radar signal and its neighboring signals is calculated. If the polarization state difference is less than a preset threshold and the signal parameters satisfy the multipath delay relationship, the radar signal is marked as a multipath signal and removed from the sorting results.