Radar weak target identification method and system fused with pulse neural network

Abstract

The invention provides a radar weak target identification method and system fused with a pulse neural network. The method comprises the following steps: extracting a polarization scattering entropy and a polarization ratio from projection sequence data to form an electromagnetic feature vector; performing spatial domain and polarization domain combined filtering operation on the projection sequence data to generate a polarization coherent feature map; processing the polarization coherent feature map by using a fusion pulse neural network, and outputting a pulse sequence; and performing weighting processing on the pulse sequence based on an attention weighting mechanism to obtain a weighted pulse sequence, and fusing the weighted pulse sequence and the polarization scattering entropy and the polarization ratio in the electromagnetic feature vector to generate a weak target recognition result. According to the technical scheme provided by the invention, the recognition precision and the anti-interference capability of the radar on the weak target are improved.

Description

Technical Field

[0001] This application relates to the field of radar identification technology, and in particular to a radar weak target identification method and system that integrates pulse neural networks. Background Technology

[0002] In radar detection under complex electromagnetic environments, the reliable identification of weak targets faces the following severe challenges: extremely low signal-to-noise ratio, insufficient utilization of polarization information, and lack of dynamic feature representation.

[0003] To improve the utilization of polarization information, existing solutions employ polarization feature extraction and deep classification networks to extract statistical features of the polarization scattering matrix from fully polarimetric radar echoes. Convolutional neural networks or recurrent neural networks are designed to perform end-to-end classification of the features and output the probability of the target's existence.

[0004] However, existing solutions have three main drawbacks: first, the bottleneck of feature solidification, with static polarization features losing dynamic correlation in the time dimension and being sensitive to noise; second, the problem of feature space distortion and computational redundancy caused by impulse noise interference at low signal-to-noise ratios; and third, the need to increase network depth to cover the possibility of weak targets, which increases real-time processing latency. Summary of the Invention

[0005] This application provides a radar weak target identification method and system that integrates pulse neural networks, in order to solve the problems of low radar weak target identification accuracy caused by the lack of time-dynamic correlation of static polarization features, the feature space being easily distorted by noise interference under low signal-to-noise ratio environment, and insufficient real-time performance due to computational redundancy in the prior art.

[0006] In a first aspect, this application provides a radar weak target identification method that integrates pulse neural networks, including:

[0007] The radar emits electromagnetic waves through a fully polarimetric radar, receives target echoes, and generates projection sequence data.

[0008] The polarization scattering entropy and polarization ratio are extracted from the projection sequence data to form an electromagnetic feature vector;

[0009] Perform a joint spatial and polarization domain filtering operation on the projected sequence data to generate a polarization coherence feature map;

[0010] The polarization coherence feature map is processed using a fused pulse neural network to output a pulse sequence;

[0011] The pulse sequence is weighted based on an attention weighting mechanism to obtain a weighted pulse sequence. The weighted pulse sequence, along with the polarization scattering entropy and polarization ratio in the electromagnetic feature vector, are then fused to generate a weak target identification result.

[0012] Optionally, the step of weighting the pulse sequence based on the attention weighting mechanism to obtain a weighted pulse sequence, and fusing the weighted pulse sequence with the polarization scattering entropy and polarization ratio in the electromagnetic feature vector to generate a weak target recognition result, includes:

[0013] Based on the pulse firing rate distribution of the pulse sequence within a preset pulse time window, a significance score is generated for each pulse event in the pulse sequence.

[0014] Based on the significance score, a corresponding dynamic weight coefficient is assigned to each pulse event to generate a weighted pulse sequence;

[0015] The weighted pulse sequence is compressed along the time dimension to generate a pulse feature vector of fixed dimension;

[0016] The pulse feature vector, the polarization scattering entropy, and the polarization ratio are concatenated to form a fused feature vector;

[0017] The fused feature vector is input into a multi-layer nonlinear decision unit to generate a binary identification identifier, and the binary identification identifier is used as the weak target identification result. The binary identification identifier is used to indicate whether the weak target exists.

[0018] Optionally, the step of inputting the fused feature vector into a multi-layer nonlinear decision unit to generate a binary identification label includes:

[0019] The fused feature vector is subjected to sparse activation processing. After sparse activation processing, feature components exceeding a preset threshold are retained to generate an activated feature vector.

[0020] The activated feature vector is subjected to multi-branch aggregation processing to generate aggregated features;

[0021] The aggregated features are mapped to corresponding binary identification identifiers by using a preset symbol function constraint.

[0022] Optionally, performing joint spatial and polarimetric filtering on the projected sequence data to generate a polarimetric coherence feature map includes:

[0023] Perform spatial filtering operations based on the projection sequence data to generate spatial filtering results;

[0024] The spatial domain filtering result is then subjected to polarization domain enhancement to generate a polarization domain enhanced result.

[0025] The spatial filtering result and the polarization enhancement result are subjected to polarization domain and spatial domain fusion operation to generate a polarization coherence feature map.

[0026] Optionally, the step of processing the polarization coherence feature map using a fused spiking neural network to output a pulse sequence includes:

[0027] The polarization coherence feature map is input into the fusion pulse neural network, and the multi-polarization channel features are extracted through the convolutional pulse layer of the fusion pulse neural network to generate a polarization feature response vector.

[0028] The polarization feature response vector is input into the pulse firing frequency modulation layer of the fused pulse neural network. Based on the amplitude component and phase gradient of the polarization feature response vector, the pulse firing threshold of the neurons in the pulse firing frequency modulation layer is dynamically adjusted to generate a primary pulse pattern.

[0029] Perform a cross-channel pulse coupling operation on the primary pulse mode to generate a synchronized pulse sequence;

[0030] The synchronized pulse sequence is input into the pulse time compression layer of the fused pulse neural network. The Poisson distribution characteristics of each polarization channel in the synchronized pulse sequence are statistically analyzed. Adjacent pulse events are merged according to the Poisson distribution characteristics, and the pulse sequence is output.

[0031] Optionally, performing a cross-channel pulse coupling operation on the primary pulse mode to generate a synchronized pulse sequence includes:

[0032] Calculate the phase difference between the pulse phase response of the horizontal polarization channel and the pulse phase response of the vertical polarization channel in the primary pulse mode;

[0033] If the phase difference exceeds the preset phase tolerance, the pulse emission event of the channel with the earlier pulse emission time in the horizontal polarization channel and the vertical polarization channel is delayed, and a phase adjustment pulse event is generated.

[0034] The phase adjustment pulse event is input into the pulse coupler, and a phase synchronization window constraint is applied through the pulse coupler to trigger a synchronized pulse delivery event that satisfies the phase synchronization window constraint;

[0035] Multiple synchronization pulse emission events are aggregated to generate a synchronization pulse sequence.

[0036] Optionally, the step of inputting the synchronized pulse sequence into the pulse time compression layer of the fused pulse neural network, statistically analyzing the Poisson distribution characteristics of each polarization channel in the synchronized pulse sequence, merging adjacent pulse events according to the Poisson distribution characteristics, and outputting a pulse sequence includes:

[0037] In the pulse time compression layer, based on the pulse firing interval of each polarization channel in the synchronized pulse sequence, the Poisson distribution characteristics of each polarization channel are statistically analyzed, and the Poisson distribution characteristics include mean parameter and variance parameter;

[0038] An adaptive time window width is generated based on the mean and variance parameters.

[0039] Slide a time window with the width of the adaptive time window along the time axis, merge adjacent pulse events within the time window that meet the preset pulse density threshold, and output the merged pulse event sequence as a pulse sequence.

[0040] Secondly, this application provides a radar weak target identification system that integrates pulse neural networks, comprising:

[0041] The receiving module is used to transmit electromagnetic waves through a fully polarized radar, receive target echoes, and generate projection sequence data.

[0042] The extraction module is used to extract polarization scattering entropy and polarization ratio from the projection sequence data to form an electromagnetic feature vector;

[0043] The filtering module is used to perform joint spatial and polarization domain filtering operations on the projected sequence data to generate a polarization coherence feature map;

[0044] The output module is used to process the polarization coherence feature map using a fused pulse neural network and output a pulse sequence.

[0045] The fusion module is used to perform weighted processing on the pulse sequence based on the attention weighting mechanism to obtain a weighted pulse sequence, and to fuse the weighted pulse sequence with the polarization scattering entropy and polarization ratio in the electromagnetic feature vector to generate a weak target recognition result.

[0046] Thirdly, this application provides a computing device, including a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are to be invoked and executed by the processing component to implement a radar weak target identification method fused with a pulse neural network as described in any of the first aspects.

[0047] Fourthly, this application provides a computer storage medium storing a computer program, which, when executed by a computer, implements a radar weak target identification method fused with a pulse neural network as described in any of the first aspects.

[0048] This application provides a radar weak target identification method based on a fused pulse neural network. The method includes: transmitting electromagnetic waves using a fully polarimetric radar, receiving target echoes, and generating projection sequence data; extracting polarization scattering entropy and polarization ratio from the projection sequence data to form an electromagnetic feature vector; performing joint spatial and polarization domain filtering operations on the projection sequence data to generate a polarization coherence feature map; processing the polarization coherence feature map using a fused pulse neural network to output a pulse sequence; weighting the pulse sequence based on an attention weighting mechanism to obtain a weighted pulse sequence, and fusing the weighted pulse sequence with the polarization scattering entropy and polarization ratio from the electromagnetic feature vector to generate a weak target identification result.

[0049] This application utilizes a fully polarimetric radar transceiver mechanism to fully capture multidimensional scattering information of the target, providing high-fidelity data for subsequent processing; it constructs a highly discriminative electromagnetic feature vector by quantifying the randomness and channel differences of the target's polarimetric scattering; it collaboratively suppresses spatial noise and polarimetric interference to generate a high signal-to-noise ratio polarimetric coherent feature map; it efficiently extracts the target's dynamic spatiotemporal evolution features using pulse timing coding characteristics and outputs a pulse sequence; it focuses on key pulse events, enhances the weight of effective information, and combines static polarimetric features to achieve accurate identification of weak targets.

[0050] Furthermore, this application generates a score for each pulse event based on the firing rate distribution of the pulse sequence within a preset time window, and assigns dynamic weight coefficients accordingly to form a weighted pulse sequence. The weighted sequence is compressed along the time dimension to obtain a fixed-dimensional pulse feature vector, which is then concatenated with the polarization scattering entropy and polarization ratio to form a fused feature vector. After inputting into a multi-layer nonlinear decision unit, sparse activation processing is first used to retain super-threshold feature components to generate an activation feature vector, followed by multi-branch aggregation processing to generate aggregated features. Finally, a preset sign function constraint maps this to a binary identification identifier. Dynamic weight allocation strengthens the expression of key pulse events, while time compression reduces feature dimensional redundancy. The fusion of static polarization features and dynamic pulse features enhances information complementarity. Sparse activation filters noise interference, multi-branch aggregation improves feature discrimination, and sign function constraints ensure the robustness of the identification results, ultimately achieving a highly reliable binary determination of the existence of weak targets.

[0051] These or other aspects of this application will become more apparent in the following description of the embodiments. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 A flowchart illustrating a radar weak target identification method fused with a pulse neural network, provided in this application embodiment;

[0054] Figure 2 A schematic diagram of the structure of a radar weak target identification system fused with a pulse neural network provided in this application embodiment;

[0055] Figure 3 This is a schematic diagram of the structure of a computing device provided in an embodiment of this application. Detailed Implementation

[0056] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0057] In some of the processes described in the specification, claims, and accompanying drawings of this application, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as 11, 12, etc., are merely used to distinguish different operations and do not themselves represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0058] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0059] To address the low accuracy of radar weak target identification caused by the lack of temporal dynamic correlation of static polarization features, the susceptibility of feature space to noise interference and distortion under low signal-to-noise ratio (SNR) environments, and insufficient real-time performance due to computational redundancy in existing technologies, this application provides a radar weak target identification method incorporating a spiking neural network. This method employs the following concept: To address the issues of easy loss of polarization dynamic features and strong noise interference in low SNR environments for weak targets, a three-level collaborative architecture of "polarization-spatiotemporal-pulse" is provided. A static discrimination basis is constructed using polarization scattering entropy and the target's essential scattering characteristics; a joint filter in the spatial polarization domain is designed to suppress noise and enhance coherent feature expression; a spiking neural network is introduced to transform feature maps into pulse sequences to capture spatiotemporal dynamic evolution; and an attention mechanism is used to focus on key pulse events, fusing static polarization features to achieve complementary decision-making, ultimately overcoming the SNR bottleneck for reliable weak target identification.

[0060] Figure 1 A flowchart of a radar weak target identification method fused with a pulse neural network provided in this application embodiment is shown below. Figure 1 As shown, the method includes:

[0061] S11. Transmit electromagnetic waves through a fully polarized radar, receive target echoes, and generate projection sequence data.

[0062] Fully polarized radar refers to a radar system capable of simultaneously transmitting and receiving horizontally and vertically polarized electromagnetic waves, acquiring the complete target scattering matrix through orthogonal polarization channels. Electromagnetic waves can refer to specific frequency band radio waves radiated by the radar antenna, used to detect targets and generate echo signals. Target echoes can refer to the electromagnetic signals reflecting off a target after electromagnetic waves illuminate it, carrying the target's scattering characteristics. Projected sequence data can refer to a discrete data sequence containing azimuth, range, and polarization dimensions generated by temporal sampling and spatial projection of the target echoes.

[0063] In this embodiment, electromagnetic waves of a specific frequency band are first transmitted into the monitoring airspace by a fully polarized radar. When the electromagnetic waves encounter a target, they are reflected to form a target echo. Then, the target echo signal is received, and after analog-to-digital conversion and time-domain sampling, projection sequence data containing three-dimensional information of azimuth, range, and polarization is generated.

[0064] S12. Extract the polarization scattering entropy and polarization ratio from the projection sequence data to form an electromagnetic feature vector.

[0065] Among them, polarization scattering entropy refers to the entropy value calculated based on the eigenvalues ​​of the target scattering matrix, used to quantify the randomness of polarization scattering, with a value range of 0 to 1. Polarization ratio can refer to the ratio of the echo energy of the horizontal polarization channel to that of the vertical polarization channel, used to reflect the degree of polarization anisotropy. Electromagnetic eigenvector is a two-dimensional vector composed of polarization scattering entropy and polarization ratio, used to characterize the static electromagnetic properties of the target.

[0066] In this embodiment, firstly, the horizontal and vertical polarization channel components are decomposed from the projection sequence data; secondly, the polarization scattering entropy is calculated based on the eigenvalues ​​of the polarization scattering matrix to quantify the randomness of target scattering; simultaneously, the polarization ratio is obtained by calculating the energy ratio of the horizontal and vertical polarization channels; finally, the polarization scattering entropy and the polarization ratio are combined to form an electromagnetic eigenvector characterizing the electromagnetic properties of the target.

[0067] S13. Perform joint spatial and polarization domain filtering on the projected sequence data to generate a polarization coherence feature map.

[0068] In this context, the spatial domain refers to the spatial dimension of radar data, comprised of azimuth and range. The polarization domain refers to the electromagnetic scattering response dimension of radar data, comprised of different polarization channels. Joint filtering refers to a filtering process executed simultaneously in both the spatial and polarization domains, improving the signal-to-noise ratio through spatial clutter suppression and polarization coherence enhancement. The polarization coherence feature map refers to a two-dimensional image generated by joint filtering, where the horizontal and vertical axes represent spatial location and polarization channels, respectively, and pixel values ​​reflect the target coherence intensity.

[0069] In this embodiment, spatial filtering is first performed on the projection sequence data, and a Wiener filter with an adaptive window length is used to suppress spatial clutter; secondly, coherence enhancement is performed in the polarization domain, and the target scattering characteristics are enhanced by eigenvalue decomposition of the polarization coherence matrix; finally, the spatial filtering result and the polarization enhancement result are fused by tensor to generate a polarization coherence feature map that preserves the coherence structure of the target.

[0070] S14. Use a fused pulse neural network to process the polarization coherence feature map and output a pulse sequence.

[0071] Among them, the pulse sequence refers to the discrete-time event sequence that integrates the output of the spiking neural network. Each pulse event can contain a timestamp and a channel identifier.

[0072] In this embodiment, the polarization coherence feature map is first input into the convolutional pulse layer of the fused spiking neural network, and the multi-polarization channel features are extracted by the pulse convolution kernel to generate a polarization feature response vector. Secondly, in the pulse firing frequency modulation layer, the neuron firing threshold is dynamically adjusted according to the amplitude gradient of the response vector to output the primary pulse pattern. Subsequently, the pulse phases of the horizontal and vertical polarization channels are aligned by a cross-channel pulse coupler to generate a synchronized pulse sequence. Finally, adjacent pulse events under the Poisson distribution are merged in the pulse time compression layer to output a dimension-reduced pulse sequence.

[0073] S15. The pulse sequence is weighted based on the attention weighting mechanism to obtain a weighted pulse sequence. The weighted pulse sequence, along with the polarization scattering entropy and polarization ratio in the electromagnetic feature vector, are then fused to generate a weak target recognition result.

[0074] The attention-weighted mechanism refers to an algorithm that calculates saliency scores based on the pulse firing rate distribution and assigns dynamic weights to reinforce key pulse events. The weighted pulse sequence refers to the attention-weighted pulse sequence, where each pulse event carries a weight coefficient representing its importance. The weak target identification result refers to the final generated binary decision label; for example, 0 indicates no target, and 1 indicates the presence of a weak target.

[0075] In this embodiment, firstly, the significance score of each pulse event is calculated based on the firing rate distribution of the pulse sequence within the time window; secondly, a weighted pulse sequence is generated by assigning dynamic weights according to the significance score; then, the weighted pulse sequence is compressed along the time axis to obtain a pulse feature vector; next, the pulse feature vector is spliced ​​with the polarization scattering entropy and polarization ratio in the electromagnetic feature vector to form a fused feature vector; finally, a binarized weak target identification result representing the existence of the target is output through a multi-layer nonlinear decision unit.

[0076] Here is a specific example: First, the monitoring station's fully polarimetric radar transmits X-band electromagnetic waves into the suspected airspace, receiving the target echoes reflected by the stealth UAV. This data is then sampled to generate a three-dimensional projection sequence of azimuth-range-polarization. Next, polarimetric scattering entropy (0.82) and polarization ratio (2.3) are extracted from the projection sequence data to form an electromagnetic feature vector. Then, spatial adaptive Wiener filtering and polarimetric coherence matrix decomposition are performed on the projection sequence data to generate a polarimetric coherence feature map with a 12dB improved signal-to-noise ratio. This feature map is then input into a fusion pulse neural network. After feature extraction via a convolutional pulse layer, pulse events in the horizontal and vertical polarization channels are synchronized using a pulse coupler, outputting a pulse sequence containing timestamps. Finally, a significance score is calculated based on the pulse firing rate distribution. High-significance pulses are weighted, compressed, and concatenated with the electromagnetic feature vector. The decision layer outputs identification marker 1, confirming the presence of a weak target.

[0077] By executing S11 to S15, this embodiment of the application captures complete target scattering information using fully polarimetric radar data, and constructs static identification features using polarization entropy and polarization ratio; joint filtering operation synergistically suppresses noise in the spatial and polarimetric domains, improving the signal-to-noise ratio of the feature map; the spiking neural network transforms the feature map into a pulse sequence to encode the target's dynamic characteristics; the attention mechanism can enhance key pulse events, and the fusion of static polarization features achieves complementary decision-making, ultimately improving the reliability of weak target identification in low signal-to-noise ratio environments.

[0078] In one possible embodiment, S15, the pulse sequence is weighted based on an attention weighting mechanism to obtain a weighted pulse sequence, and the weighted pulse sequence, along with the polarization scattering entropy and polarization ratio in the electromagnetic feature vector, are fused to generate a weak target identification result, including:

[0079] Step 151: Based on the pulse firing rate distribution of the pulse sequence within the preset pulse time window, generate a significance score for each pulse event in the pulse sequence.

[0080] The pulse time window can refer to a preset fixed time interval used to statistically analyze the pulse firing rate distribution. The window length is set according to the target motion characteristics, for example, 1 minute, 1 second, 50 ms, etc. The pulse firing rate distribution refers to the statistical distribution of the number of pulse events occurring per unit time within the pulse time window, reflecting the temporal clustering characteristics of pulse firing. A pulse event refers to a discrete event point in the pulse sequence carrying a timestamp and channel identifier, representing the activation state of a neuron at a specific moment. The significance score is a quantitative value calculated based on the pulse firing rate distribution, representing the importance of the pulse event in the time dimension; the higher the firing rate, the higher the score.

[0081] In this embodiment, a fixed-duration pulse time window is first set, and the pulse firing rate distribution of the pulse sequence within the window is statistically analyzed. Then, the significance score of each pulse event is calculated based on the time distance between the pulse event and the peak firing rate. The region with the higher firing rate corresponds to the larger significance score of the pulse event.

[0082] Step 152: Based on the significance score, assign a corresponding dynamic weight coefficient to each pulse event to generate a weighted pulse sequence.

[0083] Among them, the dynamic weight coefficient can refer to the weight value that is dynamically adjusted according to the significance score and the pulse timestamp. The weight decays faster for earlier pulses and slower for later pulses.

[0084] In this embodiment, the significance score is first input into the S-function and normalized to a weight coefficient between 0 and 1; then, the decay rate of the weight coefficient is dynamically adjusted according to the timestamp of the pulse event, with earlier pulses decaying faster and later pulses decaying slower; finally, a weighted pulse sequence is generated by assigning a dynamic weight coefficient to each pulse event.

[0085] Step 153: Compress the weighted pulse sequence along the time dimension to generate a pulse feature vector with fixed dimensions.

[0086] The time dimension refers to the time axis of the pulse sequence, along which the temporal evolution of pulse events can be analyzed. The pulse feature vector is a fixed-length vector generated by compressing and weighting the pulse sequence along the time dimension, with each dimension representing the weighted pulse integral value of a sub-time window. The fixed dimension enables the conversion of the pulse sequence into a structured vector. First, the weighted pulse sequence is divided into 10 equal-length intervals along the time axis; second, the sum of the weights of all pulse events within each interval is calculated; then, the statistical values ​​of the 10 intervals are concatenated in chronological order to form a 10-dimensional feature vector; finally, a pulse feature vector with a constant dimension of 10 is generated. This dimension strictly corresponds to the number of time intervals, ensuring that the subsequent classifier receives input with a uniform dimension.

[0087] In this embodiment, the weighted pulse sequence is first divided into equal-length segments along the time dimension; then the amplitude integral of the weighted pulse in each segment is calculated; finally, the integral values ​​are arranged in time order to generate a pulse feature vector of fixed dimension.

[0088] Step 154: Concatenate the pulse feature vector, polarization scattering entropy, and polarization ratio to form a fused feature vector.

[0089] Among them, the fused feature vector refers to the multi-dimensional vector formed by splicing the pulse feature vector with the polarization scattering entropy and polarization ratio, which has both dynamic and static features.

[0090] In this embodiment, the pre-stored polarization scattering entropy and polarization ratio are first extracted from the electromagnetic feature vector; then the pulse feature vector, polarization scattering entropy, and polarization ratio are concatenated according to the channel dimension; finally, the dimensions of each channel are normalized to generate a fused feature vector.

[0091] Step 155: Input the fused feature vector into the multi-layer nonlinear decision unit to generate a binary identification label, and use the binary identification label as the weak target identification result. The binary identification label is used to indicate whether the weak target exists.

[0092] The multi-layer nonlinear decision unit refers to a classification module composed of a sparse activation layer, a multi-branch fully connected layer, and a sign function output layer, capable of feature selection, aggregation, and binary mapping. An exemplary binary identification identifier refers to the final output decision result of 0 or 1, where 0 indicates no target detected and 1 indicates the presence of a weak target, used to trigger a radar warning system.

[0093] In this embodiment, the fused feature vector is first input into a multi-layer nonlinear decision unit, and sparsification is performed through an activation function to retain the super-threshold features; secondly, the sparse features are input into three parallel fully connected layers for multi-branch aggregation; finally, the aggregation result is mapped to a binary identification label of 0 or 1 through a sign function.

[0094] Here is a specific example: First, for the stealth UAV pulse sequence received by radar, the firing rate distribution is statistically analyzed within a 50ms pulse time window, and the significance score of each pulse event is calculated. Next, dynamic weighting coefficients are generated based on the significance scores, assigning a weight of 0.9 to high-significance late pulses and a weight of 0.2 to early pulses, forming a weighted pulse sequence. Then, the sequence is divided into 10ms segments along the time axis, and the weighted pulse integral value of each segment is calculated to generate a 16-dimensional pulse feature vector. This vector is then concatenated with pre-stored polarization scattering entropy (0.82) and polarization ratio (2.3) to form an 18-dimensional fused feature vector. Finally, this vector is input into a multi-layer nonlinear decision unit, sparsified to retain 12 feature components, aggregated by a multi-branch fully connected layer, and output as a sign function to confirm the existence of the micro-target (identification identifier 1).

[0095] By executing steps 151 to 155, this embodiment of the application quantifies the importance of events through pulse firing rate analysis, strengthens key information through dynamic weight allocation, reduces feature dimension redundancy through time compression, enhances information complementarity by fusing static polarization features and dynamic pulse features, and improves feature discrimination power through sparse activation and multi-branch aggregation, ultimately achieving efficient and reliable determination of the existence of weak targets.

[0096] In one possible embodiment, step 155, inputting the fused feature vector into a multi-layer nonlinear decision unit to generate a binary identification label, includes:

[0097] Step a1: Perform sparse activation processing on the fused feature vector. After sparse activation processing, retain feature components that exceed a preset threshold to generate an activated feature vector.

[0098] Sparse activation processing refers to the process of calculating activation values ​​for feature vectors using nonlinear functions and filtering out low-response components. A preset threshold is used to retain high-discriminative features and suppress redundant information interference. The preset threshold is a feature selection threshold determined based on historical data training, used to distinguish between effective and noisy features. Feature components exceeding this value are retained. This embodiment does not specifically limit the size of this threshold. A feature component refers to a single-dimensional element in the feature vector, representing the quantized value of a certain attribute of the target. The activated feature vector is a vector composed of non-zero feature components retained after sparse activation processing; its dimension is smaller than the original feature vector, but its information density is higher.

[0099] In this embodiment, the fused feature vector is first input into the sparse activation layer, and the activation value of each feature component is calculated by the activation function. Then, a preset threshold is set as a filtering threshold, and the feature components with activation values ​​less than the threshold are set to zero. Finally, the non-zero feature components with activation values ​​exceeding the preset threshold are retained to generate the dimensionality-reduced activation feature vector.

[0100] Step a2: Perform multi-branch aggregation processing on the activated feature vector to generate aggregated features.

[0101] Among them, multi-branch aggregation processing refers to a feature enhancement method that uses parallel network branches to perform differential transformation on features, and then integrates multi-perspective information through weighted fusion.

[0102] In this embodiment, the activation feature vector is first copied to three independent fully connected branches; then, a nonlinear transformation is performed in each branch using a different weight matrix; subsequently, the output vectors of the three branches are summed element-wise with weights; and finally, aggregated features are generated through normalization.

[0103] Step a3: Map the aggregated features to the corresponding binary identification identifiers by using a preset symbol function constraint.

[0104] The formula for pre-setting the symbolic function constraint is as follows: Where x is the multi-branch aggregated feature value, θ is the preset decision threshold, and f(x) is the binarized identification identifier. The mapping rule states that if the binarized identification identifier is 1, the weak target exists; if the binarized identification identifier is 0, the weak target does not exist.

[0105] In this embodiment, the decision boundary threshold of the symbolic function constraint is first defined; then the difference between the aggregated feature and the threshold is calculated; finally, a binary identification identifier is output based on the sign of the difference: 1 is output when the difference is greater than the preset decision threshold, and 0 is output when the difference is less than or equal to the preset decision threshold.

[0106] Here is a specific example: First, an activation function is applied to the 18-dimensional fused feature vector containing impulse and polarization features. After filtering with a preset threshold of 0.35, 12 effective feature components are retained to generate an activated feature vector. Next, this vector is input into three fully connected branches, which are transformed using 256-dimensional, 128-dimensional, and 64-dimensional weight matrices respectively. The weighted sum is then used to generate an aggregated feature. Finally, a decision boundary threshold of 0.7, constrained by a preset sign function, is applied for mapping. When the aggregated feature value is 0.83, a binary identification identifier 1 is output, confirming the existence of the micro-target.

[0107] By executing steps a1 to a3, the embodiments of this application reduce noise interference by filtering low-value features through sparse activation; enhance feature discrimination by multi-branch aggregation; and achieve disturbance-resistant binarization decision-making by sign function constraints, ultimately improving the stability and reliability of weak target recognition results.

[0108] In one possible embodiment, S13, performing a joint spatial and polarization domain filtering operation on the projected sequence data to generate a polarization coherence feature map, including:

[0109] Step 131: Perform spatial filtering operation based on the projection sequence data to generate spatial filtering results.

[0110] Spatial filtering refers to noise suppression processing of the spatial dimension of radar projection sequence data. This includes extracting local features using two-dimensional convolutional kernels, adaptively adjusting filter coefficients based on a background statistical model, and ultimately improving the target signal-to-noise ratio. The spatial filtering result refers to the intermediate data generated after the projection sequence data has undergone spatial filtering. It retains the target's spatial structure characteristics while suppressing clutter interference, serving as the input for polarization domain enhancement operations.

[0111] In this embodiment, firstly, based on the projection sequence data received by the fully polarimetric radar, a spatial filtering algorithm is used to suppress noise in the spatial dimension of the radar echo. Specifically, a two-dimensional convolution kernel is designed to perform sliding window convolution calculations on the azimuth and range directions of the projection sequence data to extract local spatial correlation features. Secondly, the filtering coefficients are adaptively adjusted according to the statistical characteristics of background clutter to eliminate interference from non-target scattering points. Finally, the spatial filtering result that preserves the target's spatial structure is output.

[0112] Step 132: Perform polarization domain enhancement on the spatial domain filtering results to generate polarization domain enhanced results.

[0113] Polarization domain enhancement refers to the process of strengthening the polarization difference characteristics of a target based on the amplitude and phase characteristics of the polarization scattering matrix through techniques such as polarization decomposition, cross-channel coherence calculation, and phase calibration. The polarization domain enhancement result refers to the intermediate data generated after the spatial filtering result has been polarized domain enhanced, highlighting the target's polarization scattering characteristics and used for subsequent fusion operations.

[0114] In this embodiment, the spatial filtering result is first input into the polarization domain enhancement module to analyze the amplitude and phase information of its polarization scattering matrix. Specifically, the odd-order and even-order scattering components are separated using a polarization decomposition algorithm; secondly, the target polarization characteristics are enhanced by calculating the coherence ratio of the cross-polarization channels; subsequently, the enhanced polarization components are phase-calibrated and amplitude-normalized; finally, a polarization domain enhancement result that highlights the target polarization differences is generated.

[0115] Step 133: Perform polarization domain and spatial domain fusion operations on the spatial domain filtering results and polarization domain enhancement results to generate a polarization coherence feature map.

[0116] The polarization domain and spatial domain fusion operation refers to the process of aligning the spatial domain filtering results with the polarization domain enhancement results, and then generating a polarization coherent feature map that retains both spatial domain structure information and polarization characteristics through weighted summation and resampling techniques.

[0117] In this embodiment, the spatial domain filtering result and the polarization domain enhancement result are first aligned by pixel position to construct a dual-channel feature tensor. Secondly, a weighted fusion strategy is adopted: spatial confidence weights are applied to the spatial domain features, and polarization scattering entropy weights are applied to the polarization features; then, feature complementarity is achieved through element-wise weighted summation; finally, bilinear interpolation is used to resample the fusion result to a standard size, generating a polarization coherent feature map that simultaneously contains spatial structure information and polarization characteristics.

[0118] Here is a specific example: First, a projection sequence of data (2048 points in the azimuth direction × 1024 points in the range direction) is acquired using a fully polarimetric radar system. After DC component removal and channel equalization preprocessing, a 7×7 Gaussian convolution kernel is used to perform spatial filtering on the data, generating a clutter-suppressed spatial filtering result. Next, polarimetric decomposition is performed on the spatial filtering result, calculating the coherence ratio of the horizontal and vertical polarization channels. Target scattering characteristics are enhanced through phase calibration, outputting a polarimetric enhancement result. Then, the spatial filtering result and the polarimetric enhancement result are aligned by pixel coordinates. Spatial features are weighted by 0.6, and polarization features by 0.4, and weighted point-by-point fusion is performed. Finally, bilinear interpolation is used to resample to a 128×128 resolution, generating a polarimetric coherence feature map that fuses the spatial structure and polarization characteristics.

[0119] By executing steps 131 to 133, this embodiment of the application effectively suppresses background clutter interference while preserving the target's spatial structure features through spatial domain filtering; it also deeply mines the target's polarization scattering characteristics through polarization domain enhancement; and finally, it generates a highly discriminative feature map through dual-domain fusion. This process significantly improves the distinguishability of weak targets in complex environments, provides robust input for subsequent neural network recognition, and ensures that the complementary advantages of spatial structure and polarization features are maximized.

[0120] In one possible embodiment, S14, processing the polarization coherence feature map using a fused pulse neural network to output a pulse sequence includes:

[0121] Step 141: Input the polarization coherence feature map into the fusion spiking neural network, extract multi-polarization channel features through the convolutional pulse layer of the fusion spiking neural network, and generate polarization feature response vectors.

[0122] The convolutional pulse layer refers to the hierarchical structure in the spiking neural network that extracts spatial features of multiple polarization channels through convolutional kernels. This involves convolutional operations on horizontal, vertical, and cross-polarization channels, followed by concatenation to output a three-dimensional feature vector. Multi-polarization channel features refer to the set of spatial features of horizontal, vertical, and cross-polarization channels separated from the polarization coherence feature map, reflecting the scattering response characteristics of the target under different polarization states. The polarization feature response vector is the three-dimensional feature tensor output by the convolutional pulse layer, containing the spatial convolutional response results of the multiple polarization channels, and serves as the input carrier for the pulse firing frequency modulation layer.

[0123] In this embodiment, the polarization coherence feature map is first input into the convolutional pulse layer of the fused spiking neural network. Multiple sets of learnable convolutional kernels are used to perform convolution operations on the horizontal polarization channel, vertical polarization channel, and cross polarization channel of the feature map. Then, the spatial texture features and boundary responses of each polarization channel are extracted. Subsequently, the convolutional outputs of different polarization channels are concatenated according to the channel dimension. Finally, a three-dimensional polarization feature response vector that fuses the features of multiple polarization channels is generated.

[0124] Step 142: Input the polarization feature response vector into the pulse firing frequency modulation layer of the fused spiking neural network. Based on the amplitude component and phase gradient of the polarization feature response vector, dynamically adjust the pulse firing threshold of the neurons in the pulse firing frequency modulation layer to generate the primary pulse pattern.

[0125] The pulse firing frequency modulation layer is a network layer that dynamically adjusts the firing threshold of neurons based on the amplitude component and phase gradient of the input features. Pulse patterns are generated by increasing the base frequency through amplitude and regulating the threshold through phase gradient. The amplitude component refers to the absolute value of the activation intensity of a single neuron in the polarization feature response vector, used to directly regulate the pulse firing frequency. The phase gradient refers to the rate of change of the difference in pulse triggering time between adjacent neurons, reflecting the continuity of the feature space, and is used to dynamically adjust the sensitivity of the firing threshold. The pulse firing threshold is the minimum membrane potential threshold required for a neuron to trigger a pulse, which is dynamically updated in the pulse firing frequency modulation layer based on the phase gradient. The primary pulse pattern refers to the set of original pulse events output by the pulse firing frequency modulation layer, including discrete pulse firing events that have not been synchronized across channels.

[0126] In this embodiment, the polarization feature response vector is first input into the pulse firing frequency modulation layer to analyze the amplitude component of each neuron and the phase gradient between adjacent neurons. Then, the pulse firing base frequency is increased proportionally according to the magnitude of the amplitude component. At the same time, the firing threshold is dynamically reduced according to the rate of change of the phase gradient. Subsequently, the updated pulse firing threshold is calculated independently for each neuron. Finally, when a neuron whose membrane potential exceeds the dynamic threshold triggers a pulse event, a primary pulse pattern containing spatiotemporal firing patterns is generated.

[0127] Step 143: Perform cross-channel pulse coupling operation on the primary pulse mode to generate a synchronized pulse sequence.

[0128] Cross-channel pulse coupling operation refers to the process of phase alignment of multi-polarization channel pulse events, including phase difference calculation, delay adjustment, and joint pulse triggering under synchronization window constraints. Synchronized pulse sequence refers to the set of pulse events generated after cross-channel pulse coupling operation, in which the horizontal and vertical polarization channel pulses achieve phase synchronization in the time dimension.

[0129] In this embodiment, the pulse phase response difference between the horizontally polarized channel and the vertically polarized channel in the primary pulse mode is first calculated; then, if the phase difference between the channels exceeds a preset tolerance, the channel event with the earlier pulse emission time is delayed; then, the phase-adjusted pulse is input into the pulse coupler; next, a phase synchronization window constraint with a fixed duration is applied; finally, when multiple channel pulses arrive within the synchronization window, a cross-channel joint emission event is triggered, and all synchronization events are aggregated to generate a synchronized pulse sequence.

[0130] Step 144: Input the synchronized pulse sequence into the pulse time compression layer of the fused pulse neural network, statistically analyze the Poisson distribution characteristics of each polarization channel in the synchronized pulse sequence, merge adjacent pulse events according to the Poisson distribution characteristics, and output the pulse sequence.

[0131] Cross-channel pulse coupling refers to the process of phase alignment of multi-polarization channel pulse events, including phase difference calculation, delay adjustment, and joint pulse triggering under synchronization window constraints. A synchronized pulse sequence refers to the set of pulse events generated after cross-channel pulse coupling, where the horizontal and vertical polarization channel pulses are phase-synchronized in the time dimension. The Poisson distribution describes the statistical regularity of the number of pulse firings per unit time, rather than directly describing the pulse firing time or interval; it is applicable to counting pulse events within a unit time window.

[0132] In this embodiment, firstly, the pulse firing interval of each polarization channel of the synchronized pulse sequence is statistically analyzed in the pulse time compression layer; secondly, the Poisson distribution characteristics of the firing interval are fitted and the mean and variance parameters are extracted; then, the adaptive time window width is calculated based on the parameters; next, the time window is slid along the time axis to detect the pulse density within the window; finally, when the density of adjacent pulse events exceeds a preset threshold, they are merged into a single pulse event, and the compressed pulse sequence is output.

[0133] Here is a specific example: First, a 128×128 resolution polarimetric coherence feature map is input into a convolutional pulse layer. Three sets of 5×5 convolutional kernels are used to process the horizontal, vertical, and cross-polarization channels respectively, outputting a 64-channel polarimetric feature response vector. Next, in the pulse firing frequency modulation layer, the base firing frequency is increased by multiplying the feature vector amplitude component by a scaling factor. Simultaneously, the firing threshold is dynamically reduced based on the phase gradient divided by a normalization factor, generating a primary pulse pattern containing 1024 neurons. Then, the pulse phase difference between the horizontal and vertical polarization channels is calculated, and nanosecond-level delays are applied to channels exceeding the limit. A 20-nanosecond synchronization window triggers cross-channel joint firing events, outputting a synchronized pulse sequence. Finally, in the pulse time compression layer, the Poisson distribution of the pulse firing interval for each channel is statistically analyzed. An adaptive time window is generated based on the mean and variance, and high-density pulse events within the window are merged to output a compressed pulse sequence.

[0134] By executing steps 141 to 144, this embodiment of the application fully exploits the spatial response characteristics of polarization features through convolutional pulse layers, utilizes a dynamic threshold mechanism to achieve adaptive matching between pulse firing and feature intensity, enhances the synergy of multi-polarization information through cross-channel synchronization, and finally compresses redundant pulses based on statistical characteristics. This process improves the spatiotemporal representation efficiency of pulse sequences, strengthens the discriminative power of pulse firing patterns for weak targets, and provides a high-information-density input sequence for attention-weighted processing.

[0135] In one possible embodiment, step 143, performing a cross-channel pulse coupling operation on the primary pulse mode to generate a synchronized pulse sequence, includes:

[0136] Step b1: Calculate the phase difference between the pulse phase response of the horizontal polarization channel and the pulse phase response of the vertical polarization channel in the primary pulse mode.

[0137] The horizontal polarization channel refers to the polarization component channel in the radar echo where the electric field vector is parallel to the horizontal direction, reflecting the scattering characteristics of the target for horizontally polarized electromagnetic waves. The pulse phase response refers to the phase angle value corresponding to the trigger moment of the pulse event, used to characterize the periodic position of the pulse in the time-domain waveform. The vertical polarization channel refers to the polarization component channel in the radar echo where the electric field vector is perpendicular to the horizontal direction, reflecting the scattering characteristics of the target for vertically polarized electromagnetic waves. The phase difference refers to the angular difference between the phase responses of corresponding pulse events in the horizontal and vertical polarization channels, used to quantify the degree of time mismatch between the two channels.

[0138] Step b2: If the phase difference exceeds the preset phase tolerance, delay the pulse emission event of the channel with the earlier pulse emission time in the horizontal polarization channel and the vertical polarization channel, and generate a phase adjustment pulse event.

[0139] The preset phase tolerance refers to the maximum threshold allowed for phase difference between the two-channel pulses; exceeding this value will trigger a time delay adjustment operation. The pulse delivery time refers to the precise timestamp at which a single pulse event is triggered, determining the event's temporal position in the pulse sequence. The phase adjustment pulse event refers to the set of pulse events after time delay compensation, where the phase difference between the horizontal and vertical polarization channel pulses is controlled within the preset tolerance.

[0140] Step b3: Input the phase adjustment pulse event into the pulse coupler, apply the phase synchronization window constraint through the pulse coupler, and trigger the synchronization pulse delivery event that satisfies the phase synchronization window constraint.

[0141] The preset phase tolerance refers to the maximum threshold allowed for phase difference between the two-channel pulses; exceeding this value will trigger a time delay adjustment operation. The pulse delivery time is the precise timestamp at which a single pulse event is triggered, determining its temporal position within the pulse sequence. A phase-adjusted pulse event refers to a set of pulse events that have undergone time delay compensation, where the phase difference between the horizontal and vertical polarization channel pulses is controlled within the preset tolerance.

[0142] Step b4: Aggregate multiple synchronization pulse emission events to generate a synchronization pulse sequence.

[0143] Here is a specific example: First, the phase difference between the pulse phase responses of the horizontally polarized channel and the vertically polarized channel is calculated. When the phase difference exceeds a preset tolerance, a nanosecond-level delay adjustment is applied to the channel that emitted the earlier pulse, generating a phase adjustment pulse event. The adjusted event is then input into a pulse coupler, which detects the arrival of pulses from both channels within a set time window. If pulses from both the horizontal and vertically polarized channels appear simultaneously within the window, a synchronization pulse emission event is triggered. Finally, all synchronization events triggered within the window are aggregated and sorted along the time axis to generate a synchronization pulse sequence with cross-channel synchronization characteristics.

[0144] By executing steps b1 to b4, this embodiment of the application eliminates the time mismatch of multi-channel signals by dynamically compensating for the phase deviation between polarization channels; it enhances the correlation of polarization information by forcibly triggering cross-channel joint pulses using synchronization window constraints; and finally generates a pulse sequence with strict spatiotemporal synchronization characteristics. This process significantly improves the synergy of multi-polarization feature fusion, providing a highly consistent data foundation for subsequent pulse compression and target recognition.

[0145] In one possible embodiment, step 144 involves inputting the synchronized pulse sequence into the pulse time compression layer of the fused pulse neural network, statistically analyzing the Poisson distribution characteristics of each polarization channel in the synchronized pulse sequence, merging adjacent pulse events based on the Poisson distribution characteristics, and outputting the pulse sequence, including:

[0146] Step c1: In the pulse time compression layer, based on the pulse firing interval of each polarization channel in the synchronized pulse sequence, statistically analyze the Poisson distribution characteristics of each polarization channel. The Poisson distribution characteristics include the mean parameter and the variance parameter.

[0147] The pulse firing interval refers to the time difference between the trigger times of two adjacent pulse events within the same polarization channel in a synchronized pulse sequence, reflecting the temporal regularity of pulse firing. The mean parameter is the statistical value of the average time interval of the Poisson distribution of the pulse firing interval, used to characterize the reference frequency characteristics of pulse firing. The variance parameter is the statistical value of the dispersion of the Poisson distribution of the pulse firing interval, reflecting the fluctuation range of pulse firing time.

[0148] Step c2: Generate the adaptive time window width based on the mean and variance parameters.

[0149] The adaptive time window width refers to the window width value dynamically calculated based on the mean and variance parameters, which is adaptively adjusted according to the pulse firing stability to achieve efficient compression.

[0150] Step c3: Slide a time window with an adaptive time window width along the time axis, merge adjacent pulse events that meet the preset pulse density threshold within the time window, and output the merged pulse event sequence as a pulse sequence.

[0151] The time axis refers to a one-dimensional coordinate axis that arranges pulse events in chronological order, serving as the basic temporal framework for pulse sequence operations. The time window is a variable-length interval that slides along the time axis, used to detect the density of pulse events within a local time period. The preset pulse density threshold is the minimum number of events required to trigger a pulse merging operation; event merging is performed when the number of pulses within the window exceeds this value.

[0152] Here's a specific example: First, the firing interval of adjacent pulses in the synchronized pulse sequence of the horizontal polarization channel is measured in the pulse time compression layer, and its Poisson distribution mean and variance parameters are calculated. Next, an adaptive time window width is generated by multiplying the mean parameter by a unit coefficient, adding the variance parameter, and dividing by a correction factor. Then, this time window is slid along the time axis. When the number of pulse events within a certain window exceeds a preset density threshold, multiple adjacent pulses in that dense region are merged into a single representative event. Finally, the compressed pulse sequence is output, which retains the original firing pattern characteristics while significantly reducing the number of events.

[0153] By executing steps c1 to c3, this embodiment of the application accurately models the pulse firing pattern using Poisson distribution statistics, adapts the pulse distribution characteristics of different density segments using dynamic window width, and triggers intelligent merging operations based on density thresholds. This process effectively reduces data redundancy while preserving key pulse patterns, improves the temporal compactness and information transmission efficiency of the pulse sequence, and provides optimized input for subsequent recognition modules.

[0154] Figure 2 A schematic diagram of a radar weak target identification system fused with a pulse neural network, provided in an embodiment of this application, is shown below. Figure 2 As shown, the system includes:

[0155] The receiving module 21 is used to transmit electromagnetic waves through a fully polarized radar, receive target echoes, and generate projection sequence data.

[0156] Extraction module 22 is used to extract polarization scattering entropy and polarization ratio from the projection sequence data to form electromagnetic feature vectors.

[0157] The filtering module 23 is used to perform joint spatial and polarization domain filtering operations on the projected sequence data to generate a polarization coherence feature map.

[0158] Output module 24 is used to process the polarization coherence feature map using a fused pulse neural network and output a pulse sequence.

[0159] The fusion module 25 is used to perform weighted processing on the pulse sequence based on the attention weighting mechanism to obtain a weighted pulse sequence, and to fuse the weighted pulse sequence, as well as the polarization scattering entropy and polarization ratio in the electromagnetic feature vector, to generate a weak target recognition result.

[0160] Figure 2 The radar weak target recognition system that integrates pulse neural networks can perform... Figure 1 The implementation principle and technical effects of the radar weak target identification method fused with a pulse neural network described in the illustrated embodiment will not be repeated here. The specific operation methods of each module and unit in the radar weak target identification system fused with a pulse neural network in the above embodiments have been described in detail in the embodiments related to this method, and will not be elaborated upon here.

[0161] In one possible design, Figure 2 The radar weak target identification system fused with pulse neural networks in the illustrated embodiment can be implemented as a computing device, such as... Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32.

[0162] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are invoked and executed by the processing component 32.

[0163] The processing component 32 is used to perform the following processes: transmitting electromagnetic waves via a fully polarimetric radar, receiving target echoes, and generating projection sequence data. Extracting polarization scattering entropy and polarization ratio from the projection sequence data to form an electromagnetic feature vector. Performing joint spatial and polarization domain filtering operations on the projection sequence data to generate a polarization coherence feature map. Processing the polarization coherence feature map using a fused pulse neural network to output a pulse sequence. Weighting the pulse sequence based on an attention weighting mechanism to obtain a weighted pulse sequence, and fusing the weighted pulse sequence with the polarization scattering entropy and polarization ratio from the electromagnetic feature vector to generate a weak target identification result.

[0164] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above-described method. Alternatively, the processing component may be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above-described method.

[0165] Storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as Random Access Memory (RAM), Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read Only Memory (PROM), Read Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0166] Of course, computing devices may also include other components, such as input / output interfaces, display components, communication components, etc.

[0167] Input / output interfaces provide interfaces between processing components and peripheral interface modules, which can be output devices, input devices, etc.

[0168] The communication components are configured to facilitate wired or wireless communication between computing devices and other devices.

[0169] The computing device can be a physical device or an elastic computing host provided by a cloud computing platform. In this case, the computing device can refer to a cloud server, and the aforementioned processing components, storage components, etc., can be basic server resources rented or purchased from the cloud computing platform.

[0170] This application also provides a computer storage medium storing a computer program, which, when executed by a computer, can perform the above-described functions. Figure 1 The illustrated embodiment presents a radar weak target identification method that incorporates a pulse neural network.

[0171] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0172] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0173] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0174] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A radar weak target identification method incorporating pulse neural networks, characterized in that, include: The radar emits electromagnetic waves through a fully polarimetric radar, receives target echoes, and generates projection sequence data. The polarization scattering entropy and polarization ratio are extracted from the projection sequence data to form an electromagnetic feature vector; Perform a joint spatial and polarization domain filtering operation on the projected sequence data to generate a polarization coherence feature map; The polarization coherence feature map is processed using a fused pulse neural network to output a pulse sequence; The pulse sequence is weighted based on an attention weighting mechanism to obtain a weighted pulse sequence. The weighted pulse sequence, along with the polarization scattering entropy and polarization ratio in the electromagnetic feature vector, are then fused to generate a weak target identification result.

2. The method according to claim 1, characterized in that, The pulse sequence is weighted using an attention-weighted mechanism to obtain a weighted pulse sequence. This weighted pulse sequence, along with the polarization scattering entropy and polarization ratio in the electromagnetic feature vector, is then fused to generate a weak target identification result, including: Based on the pulse firing rate distribution of the pulse sequence within a preset pulse time window, a significance score is generated for each pulse event in the pulse sequence. Based on the significance score, a corresponding dynamic weight coefficient is assigned to each pulse event to generate a weighted pulse sequence; The weighted pulse sequence is compressed along the time dimension to generate a pulse feature vector of fixed dimension; The pulse feature vector, the polarization scattering entropy, and the polarization ratio are concatenated to form a fused feature vector; The fused feature vector is input into a multi-layer nonlinear decision unit to generate a binary identification identifier, and the binary identification identifier is used as the weak target identification result. The binary identification identifier is used to indicate whether the weak target exists.

3. The method according to claim 2, characterized in that, The step of inputting the fused feature vector into a multi-layer nonlinear decision unit to generate a binary identification label includes: The fused feature vector is subjected to sparse activation processing. After sparse activation processing, feature components exceeding a preset threshold are retained to generate an activated feature vector. The activated feature vector is subjected to multi-branch aggregation processing to generate aggregated features; The aggregated features are mapped to corresponding binary identification identifiers by using a preset symbol function constraint.

4. The method according to claim 1, characterized in that, The step of performing joint spatial and polarimetric filtering on the projected sequence data to generate a polarimetric coherence feature map includes: Perform spatial filtering operations based on the projection sequence data to generate spatial filtering results; The spatial domain filtering result is then subjected to polarization domain enhancement to generate a polarization domain enhanced result. The spatial filtering result and the polarization enhancement result are subjected to polarization domain and spatial domain fusion operation to generate a polarization coherence feature map.

5. The method according to claim 1, characterized in that, The process of processing the polarization coherence feature map using a fused pulse neural network to output a pulse sequence includes: The polarization coherence feature map is input into the fusion pulse neural network, and the multi-polarization channel features are extracted through the convolutional pulse layer of the fusion pulse neural network to generate a polarization feature response vector. The polarization feature response vector is input into the pulse firing frequency modulation layer of the fused pulse neural network. Based on the amplitude component and phase gradient of the polarization feature response vector, the pulse firing threshold of the neurons in the pulse firing frequency modulation layer is dynamically adjusted to generate a primary pulse pattern. Perform a cross-channel pulse coupling operation on the primary pulse mode to generate a synchronized pulse sequence; The synchronized pulse sequence is input into the pulse time compression layer of the fused pulse neural network. The Poisson distribution characteristics of each polarization channel in the synchronized pulse sequence are statistically analyzed. Adjacent pulse events are merged according to the Poisson distribution characteristics, and the pulse sequence is output.

6. The method according to claim 5, characterized in that, The step of performing a cross-channel pulse coupling operation on the primary pulse mode to generate a synchronized pulse sequence includes: Calculate the phase difference between the pulse phase response of the horizontal polarization channel and the pulse phase response of the vertical polarization channel in the primary pulse mode; If the phase difference exceeds the preset phase tolerance, the pulse emission event of the channel with the earlier pulse emission time in the horizontal polarization channel and the vertical polarization channel is delayed, and a phase adjustment pulse event is generated. The phase adjustment pulse event is input into the pulse coupler, and a phase synchronization window constraint is applied through the pulse coupler to trigger a synchronized pulse delivery event that satisfies the phase synchronization window constraint; Multiple synchronization pulse emission events are aggregated to generate a synchronization pulse sequence.

7. The method according to claim 5, characterized in that, The process of inputting the synchronized pulse sequence into the pulse time compression layer of the fused pulse neural network, statistically analyzing the Poisson distribution characteristics of each polarization channel in the synchronized pulse sequence, merging adjacent pulse events based on the Poisson distribution characteristics, and outputting a pulse sequence includes: In the pulse time compression layer, based on the pulse firing interval of each polarization channel in the synchronized pulse sequence, the Poisson distribution characteristics of each polarization channel are statistically analyzed, and the Poisson distribution characteristics include mean parameter and variance parameter; An adaptive time window width is generated based on the mean and variance parameters. Slide a time window with the width of the adaptive time window along the time axis, merge adjacent pulse events within the time window that meet the preset pulse density threshold, and output the merged pulse event sequence as a pulse sequence.

8. A radar weak target identification system integrating pulse neural networks, characterized in that, include: The receiving module is used to transmit electromagnetic waves through a fully polarized radar, receive target echoes, and generate projection sequence data. The extraction module is used to extract polarization scattering entropy and polarization ratio from the projection sequence data to form an electromagnetic feature vector; The filtering module is used to perform joint spatial and polarization domain filtering operations on the projected sequence data to generate a polarization coherence feature map; The output module is used to process the polarization coherence feature map using a fused pulse neural network and output a pulse sequence. The fusion module is used to perform weighted processing on the pulse sequence based on the attention weighting mechanism to obtain a weighted pulse sequence, and to fuse the weighted pulse sequence with the polarization scattering entropy and polarization ratio in the electromagnetic feature vector to generate a weak target recognition result.

9. A computing device, characterized in that, It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked and executed by the processing component to implement a radar weak target identification method fused with a pulse neural network as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that, The system contains a computer program that, when executed by a computer, implements a radar weak target identification method fused with a pulse neural network as described in any one of claims 1 to 7.

Images