An intelligent sorting method for parameter agile radar signals

Abstract

The application discloses a kind of parameter agile radar signal intelligent sorting methods, first, original radar signal is detected, and pulse description word sequence is obtained, the pulse description word sequence obtained is preprocessed, then feature extraction module based on multi-branch hollow convolution, feature fusion module based on attention mechanism, radiation source mapping module based on transpose convolution are cascaded, and intelligent deep sorting network is constructed, then the sequence after pre-processing is used as the input of intelligent deep sorting network, and the intelligent deep sorting network is trained, finally, the parameter agile radar signal is sorted using the intelligent deep sorting network after training is completed.The method of the application can realize the accurate sorting of parameter agile radar, effectively solve the "batch" problem, has the advantages of flexible, accurate and strong generalization ability, can adapt to parameter overlap, noise and pulse loss and other complex situations.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing technology, specifically relating to a method for intelligent sorting of parameter-agile radar signals. Background Technology

[0002] Radar signal sorting separates pulses received by a reconnaissance receiver from multiple radiation sources, reducing the pulse current density in the electromagnetic space and providing a basis for subsequent intercept signal processing and information acquisition.

[0003] With the application of new radar systems, signals in the electromagnetic space are becoming increasingly complex and variable, and pulse density is increasing dramatically. Radars with parameter agility can transmit pulse trains with different pulse repetition intervals (PRI) and carrier frequency (RF) modes. A single transmitter can be distributed into many clusters in the parameter space, making it difficult to determine the number of transmitters. Furthermore, parameters from different transmitters in a certain operating mode may overlap, easily leading to them being attributed to a single transmitter, severely reducing sorting accuracy. Therefore, parameter agility poses a significant challenge to radar sorting. Pulse trains are typically represented as pulse descriptors (PDWs), including angle of arrival (DOA), carrier frequency (RF), pulse width (PW), pulse amplitude (PA), and time of arrival (TOA). Existing sorting methods mainly fall into two categories: clustering-based methods and PRI-based methods. K-means clustering methods are widely used in sorting, but this method relies on the number of clusters as prior information. Density-based clustering methods such as DBSCAN and ADPC are gradually developing, which can adaptively determine cluster centers. However, parameter agility makes the estimation of cluster centers unreliable, severely degrading sorting performance. PRI-based methods include the Cumulative Difference Histogram (CDIF), Sequence Difference Histogram (SDIF), and PRI Transformation. These methods estimate the PRI value by statistically analyzing or transforming the PRI sequence, and are therefore only applicable to fixed or low-jitter PRI values. Furthermore, these methods are susceptible to measurement errors and struggle to achieve good performance in environments with pulse loss and noise. Therefore, existing methods are insufficient for sorting transmitters with variable parameters.

[0004] With the rapid development of deep learning, deep networks can effectively construct sequences. Deep learning-based methods can directly associate pulses with their respective transmitters by extracting high-dimensional features of the pulse detection wave (PDW). The paper "Z. Kang, Y. Zhong, Y. Wu and Y. Cai, 'Signal Deinterleaving Based on U-Net Networks,' 2023 8th International Conference on Computer and Communication Systems (ICCCS), Guangzhou, China, 2023, pp. 62-67" uses U-Net based on the fully convolutional network concept to achieve signal sorting. This method is very effective when the radar's PDW parameters are separable in any dimension, but it is not suitable for situations where the signal parameters of different radiation sources overlap. The paper “T. Chen, B. Yang and L. Guo, 'Radar Pulse StreamClustering Based on MaskRCNN Instance Segmentation Network,' in IEEE Signal Processing Letters, vol.30, pp.1022-1026, 2023” shows good performance under pulse loss and noise conditions through PDW mesh mapping; however, it cannot effectively correlate multiple sets of parameters of a transmitter when the parameter differences are large. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides an intelligent sorting method for parameter-agile radar signals. By analyzing the variation patterns of parameter-agile radar PDW sequences, deep learning methods are used to extract features, and the mapping relationship between PDW and radar radiation sources is fitted through feature reconstruction. This achieves accurate sorting of parameter-agile radar signals and effectively solves the "batch addition" problem.

[0006] The technical solution adopted in this invention is: a method for intelligent sorting of parameter-agile radar signals, the specific steps of which are as follows:

[0007] S1. Detect the raw radar signal and obtain the pulse descriptor sequence;

[0008] The pulse descriptor sequence includes: pulse arrival time TOA, pulse amplitude PA, pulse carrier frequency RF, pulse width PW, and pulse arrival angle DOA.

[0009] S2. Preprocess the pulse descriptor sequence in step S1;

[0010] S3. Construct a feature extraction module based on multi-branch dilated convolution;

[0011] S4. Construct a feature fusion module based on an attention mechanism;

[0012] S5. Construct a radiation source mapping module based on transposed convolution;

[0013] S6. Concatenate the feature extraction module obtained in step S3, the feature fusion module obtained in step S4, and the radiation source mapping module obtained in step S5 to construct an intelligent deep sorting network.

[0014] S7. Use the sequence obtained in step S2 as the input of the intelligent deep sorting network constructed in step S6 to train the intelligent deep sorting network.

[0015] S8. Use the intelligent deep sorting network trained in step S7 to sort the parameter-agile radar signals.

[0016] Furthermore, step S2 is specifically as follows:

[0017] The TOA sequence is initialized, and the arrival time of each pulse relative to the first pulse is calculated; mean normalization is performed on RF and PW.

[0018] Given that the sequence contains N pulses, the pulse sequence expression is as follows:

[0019] PDW = {pdw0, pdw1, ..., pdw} N-1}

[0020] Where PDW represents the set of raw pulse descriptors detected, N represents the number of pulses, and pdw i (i = 0, ..., N-1) represents the pulse description word of the i-th pulse, i.e.:

[0021] pdw i ={toa i ,rf i ,pw i ,pa i doa i (i = 0, ..., N-1)

[0022] Among them, toa i (i = 0, ..., N-1) represents the arrival time of the i-th pulse, rf i (i = 0, ..., N-1) represents the carrier frequency of the i-th pulse, pw i (i = 0, ..., N-1) represents the pulse width of the i-th pulse, pa i (i = 0, ..., N-1) represents the amplitude of the i-th pulse, doa i(i = 0, ..., N-1) represents the arrival angle of the i-th pulse.

[0023] The expression for the preprocessed pulse sequence PDW' is as follows:

[0024] PDW' = {pdw'0,pdw'1,…,pdw'} N-1}

[0025] pdw' i ={toa' i ,rf' i ,pw' i}

[0026] Among them, pdw' i (i = 0, ..., N-1) represents the pulse descriptor of the i-th pulse after preprocessing, toa' i Let represent the arrival time of the i-th pulse after preprocessing, calculated as follows:

[0027]

[0028] rf' i The carrier frequency of the i-th pulse after preprocessing is represented by the following expression:

[0029]

[0030] pw' i Let represent the pulse width of the i-th pulse after preprocessing. The calculation expression is as follows:

[0031]

[0032] If the arrival angle ranges of all radiation sources are set to be similar, then DOA is not used as input and pulse amplitude PA is not considered. If DOA is discriminative, the DOA sequence is directly used as input.

[0033] Furthermore, step S3 is specifically as follows:

[0034] The feature extraction module includes: 1 convolutional layer, 4 basic modules, 4 parallel dilated convolutional branches, each basic module includes 2 residual layers, and each convolutional layer is followed by a batch regularization layer and a ReLU activation function.

[0035] Furthermore, step S4 is specifically as follows:

[0036] The feature fusion module includes: one channel attention layer and one basic module.

[0037] The channel attention layer includes: one global average pooling layer, two fully connected layers, and two activation function layers; the basic module includes: two residual layers, each followed by a batch regularization layer and a ReLU activation function.

[0038] Furthermore, step S5 is specifically as follows:

[0039] The radiation source mapping module includes: 3 transposed convolutional layers and 4 convolutional layers, wherein each of the 3 convolutional layers is followed by a batch regularization layer and a ReLU activation function.

[0040] Furthermore, in step S6, the intelligent deep sorting network uses a cross-entropy loss function to output the radiation source to which the pulse descriptor belongs at each time step.

[0041] Furthermore, step S7 is specifically as follows:

[0042] The training samples processed in step S2 are input into the intelligent deep sorting network constructed in step S6 for forward propagation; the cost function value is calculated, using the cross-entropy loss function as the cost function, and the calculation expression is as follows:

[0043] loss = CEL(prob, label)

[0044] Where CEL represents the cross-entropy loss function, prob∈R 1×M Let label represent the output probability, R denote the real number field, and label ∈ R. 1×M This represents the true value of the radiation source in one-hot code form, where M represents the total number of radiation sources.

[0045] Then, the network parameters are updated using the backpropagation algorithm with momentum gradient descent; the backpropagation is performed iteratively until the cost function converges, resulting in a trained intelligent deep sorting network.

[0046] Furthermore, step S8 is specifically as follows:

[0047] The test samples preprocessed in step S2 are input into the intelligent deep sorting network trained in step S7 for forward propagation to obtain the posterior probability of each pulse in the test samples belonging to each radiation source. The magnitudes of the posterior probabilities of each radiation source are compared, and the radiation source corresponding to the maximum value is taken as the pulse sorting result.

[0048] The beneficial effects of this invention are as follows: The method of this invention first detects the original radar signal to obtain a pulse descriptor sequence. The obtained pulse descriptor sequence is preprocessed, and then a feature extraction module based on multi-branch dilated convolution, a feature fusion module based on an attention mechanism, and a radiation source mapping module based on transposed convolution are cascaded to construct an intelligent deep sorting network. The preprocessed sequence is then used as input to the intelligent deep sorting network for training. Finally, the trained intelligent deep sorting network is used to sort parameters-agile radar signals. This invention utilizes intercepted reconnaissance parameters, combined with radar spatiotemporal frequency domain characteristics, and uses pulse descriptor sequences to characterize the relative position and changes of pulse trains. Based on the variation law and characteristics of radar signal parameters, a multi-branch dilated convolution based on multiple receptive fields is constructed to extract parameter features. Based on the mapping relationship between pulse parameters and radar radiation sources, a feature reconstruction module is constructed to reconstruct pulse train information, achieving radar signal sorting under conditions of agile signal parameters. The method of the present invention can achieve accurate sorting of parameter-agile radar, effectively solve the "batch addition" problem, and has the advantages of flexibility, accuracy and strong generalization ability. It can adapt to complex situations such as parameter overlap, noise and pulse loss. Attached Figure Description

[0049] Figure 1 This is a flowchart of a parameter-agile radar signal intelligent sorting method according to the present invention.

[0050] Figure 2 This is a schematic diagram of the feature extraction module structure in an embodiment of the present invention.

[0051] Figure 3 This is a schematic diagram of the feature fusion module structure in an embodiment of the present invention.

[0052] Figure 4 This is a schematic diagram of the radiation source mapping module in an embodiment of the present invention.

[0053] Figure 5 This is a schematic diagram of the sorting and visualization results of parameter-agile radar signals in an embodiment of the present invention.

[0054] Figure 6 These are the recognition results in three scenarios according to embodiments of the present invention. Detailed Implementation

[0055] The method of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0056] like Figure 1 The flowchart of the intelligent sorting method for parameter-agile radar signals of the present invention is shown below, and the specific steps are as follows:

[0057] S1. Detect the raw radar signal and obtain the pulse descriptor sequence;

[0058] The pulse descriptor sequence includes: pulse arrival time TOA, pulse amplitude PA, pulse carrier frequency RF, pulse width PW, and pulse arrival angle DOA.

[0059] S2. Preprocess the pulse descriptor sequence in step S1;

[0060] S3. Construct a feature extraction module based on multi-branch dilated convolution;

[0061] S4. Construct a feature fusion module based on an attention mechanism;

[0062] S5. Construct a radiation source mapping module based on transposed convolution;

[0063] S6. Concatenate the feature extraction module obtained in step S3, the feature fusion module obtained in step S4, and the radiation source mapping module obtained in step S5 to construct an intelligent deep sorting network.

[0064] S7. Use the sequence obtained in step S2 as the input of the intelligent deep sorting network constructed in step S6 to train the intelligent deep sorting network.

[0065] S8. Use the intelligent deep sorting network trained in step S7 to sort the parameter-agile radar signals.

[0066] In this embodiment, the multiple sets of parameters of the original radar signal (pulse stream) detected in step S1 may originate from the same radiation source, and some signal parameters may overlap. In this case, it is difficult to achieve signal sorting with a single pulse, and sorting is required based on the variation pattern of the parameters over a certain period of time. Therefore, a pulse descriptor sequence formed based on the pulse arrival time is used as the input of this method.

[0067] In this embodiment, step S2 is specifically as follows:

[0068] The TOA sequence is initialized, and the arrival time of each pulse relative to the first pulse is calculated; mean normalization is performed on RF and PW.

[0069] Given that the sequence contains N pulses (N is a positive integer), the pulse sequence expression is as follows:

[0070] PDW = {pdw0, pdw1, ..., pdw} N-1}

[0071] Where PDW represents the set of raw pulse descriptors detected, N represents the number of pulses, and pdw i (i = 0, ..., N-1) represents the pulse description word of the i-th pulse, i.e.:

[0072] pdwi ={toa i ,rf i ,pw i ,pa i doa i (i = 0, ..., N-1)

[0073] Among them, toa i (i = 0, ..., N-1) represents the arrival time of the i-th pulse, rf i (i = 0, ..., N-1) represents the carrier frequency of the i-th pulse, pw i (i = 0, ..., N-1) represents the pulse width of the i-th pulse, pa i (i = 0, ..., N-1) represents the amplitude of the i-th pulse, doa i (i = 0, ..., N-1) represents the arrival angle of the i-th pulse.

[0074] The expression for the preprocessed pulse sequence PDW' is as follows:

[0075] PDW' = {pdw'0,pdw'1,…,pdw'} N-1}

[0076] pdw' i ={toa' i ,rf' i ,pw' i}

[0077] Among them, pdw' i (i = 0, ..., N-1) represents the pulse descriptor of the i-th pulse after preprocessing, toa' i Let represent the arrival time of the i-th pulse after preprocessing, calculated as follows:

[0078]

[0079] rf' i The carrier frequency of the i-th pulse after preprocessing is represented by the following expression:

[0080]

[0081] pw' i Let represent the pulse width of the i-th pulse after preprocessing. The calculation expression is as follows:

[0082]

[0083] This embodiment addresses the sorting and identification problem of radiation sources in the same location. It assumes that the angle of arrival (DOA) ranges of all radiation sources are similar, therefore the DOA is not used as input. Since the amplitude of the signal intercepted by the receiver in non-cooperative situations is a relative value, amplitude is not considered. When the DOA is discriminative, the DOA sequence is directly used as input.

[0084] In this embodiment, step S3 is specifically as follows:

[0085] The feature extraction module includes: 1 convolutional layer, 4 basic modules, 4 parallel dilated convolutional branches, each basic module includes 2 residual layers, and each convolutional layer is followed by a batch regularization layer and a ReLU activation function.

[0086] like Figure 2 The diagram shows the structure of the feature extraction module in this embodiment. In the diagram, k represents the kernel size, d represents the dilation rate, and the convolutional layer is represented as "convolutional layer_(kernel size, output channel, stride)". The basic module includes two residual layers, represented as "basic module_(kernel size, output channel, stride)"; dilated convolution is represented as "dilated convolution_(kernel size, output channel, stride, dilation rate)".

[0087] To extract the variation patterns of agile parameters, the convolution kernel is designed as a one-dimensional temporal convolution, thus preserving the relative temporal relationship of the pulse parameters after convolution. Since adjacent pulses may originate from different radiation sources and have no specific relationship, pooling layers are not used in the feature extraction module; all dimensional transformations are achieved through convolution kernels with a stride of 2. In this embodiment, the convolution kernel size of the convolutional layer is set to 7, and the kernel size of the basic module is 3, with strides of 2, 2, 1, and 1 respectively. Because pulses from different radiation sources are intertwined, dilated convolutions with different receptive fields are used to extract features at different time scales. The feature extraction module contains four dilated convolution branches, each with a different dilation rate, thereby achieving feature extraction at different scales. The kernel size of the four dilated convolutional layers is 3, with dilation rates of 1, 6, 12, and 18 respectively.

[0088] In this embodiment, step S4 is specifically as follows:

[0089] The feature fusion module includes: one channel attention layer and one basic module.

[0090] The channel attention layer includes: one global average pooling layer, two fully connected layers, and two activation function layers; the basic module includes: two residual layers, each followed by a batch regularization layer and a ReLU activation function.

[0091] like Figure 3The diagram shows the feature fusion module structure of this embodiment. In the diagram, k represents the convolution kernel size, the fully connected layer is represented as "fully connected layer_(number of output neurons)", the activation function layer is represented as "activation function name activation function layer", and the module is represented as "basic module_(convolution kernel size, output channel, stride)".

[0092] In this embodiment, to obtain adaptive weights for each channel's feature map, a global average pooling layer is used to compress the feature map. Then, two fully connected layers are used to obtain the weights for each feature channel. Finally, the weights are multiplied channel-by-channel by the original feature map, thus allowing the network model to focus more on important features. The number of output neurons in the fully connected layers of the attention mechanism layer are 1024 / 16 and 1024, respectively, with ReLU and Sigmoid activation functions. The convolutional layers have a kernel size of 1 and 256 output channels.

[0093] In this embodiment, step S5 is specifically as follows:

[0094] The radiation source mapping module includes: 3 transposed convolutional layers and 4 convolutional layers, wherein each of the 3 convolutional layers is followed by a batch regularization layer and a ReLU activation function.

[0095] like Figure 4 The diagram shows the structure of the radiation source mapping module in this embodiment. In the diagram, k represents the kernel size, M represents the number of radiation sources, the transposed convolutional layer is represented as "transposed convolution_(kernel size, output channels, stride)", and the convolutional layer is represented as "convolutional layer_(kernel size, output channels, stride)". Each transposed convolutional layer is followed by a convolutional layer, a batch normalization layer, and an activation function layer in sequence. The kernel size of the last convolutional layer is 1, and the number of output channels is equal to the number of radiation sources.

[0096] In this embodiment, considering the complex variation patterns of intertwined pulse parameters, a learnable transposed convolution is used to accurately map the pulse parameters to the radiation sources, gradually reconstructing the pulse features. Each transposed convolutional layer has a kernel size of 3, and each layer is followed by another convolutional layer, thus better learning and reconstructing the features. This module consists of three transposed convolutional layers. The last layer is a convolutional layer with a kernel size of 1, and the output channel dimension is 4, which corresponds to the number of radiation sources.

[0097] In this embodiment, in step S6, the intelligent deep sorting network uses the cross-entropy loss function to output the radiation source to which the pulse descriptor belongs at each time moment.

[0098] In this embodiment, step S7 is specifically as follows:

[0099] The training samples processed in step S2 are input into the intelligent deep sorting network constructed in step S6 for forward propagation; the cost function value is calculated, using the cross-entropy loss function as the cost function, and the calculation expression is as follows:

[0100] loss = CEL(prob, label)

[0101] Where CEL represents the cross-entropy loss function, prob∈R 1×M Let label represent the output probability, R denote the real number field, and label ∈ R. 1×M This represents the true value of the radiation source in one-hot code form, where M represents the total number of radiation sources.

[0102] Then, the network parameters are updated using the backpropagation algorithm with momentum gradient descent; the backpropagation is performed iteratively until the cost function converges, resulting in a trained intelligent deep sorting network.

[0103] In this embodiment, step S8 is specifically as follows:

[0104] The test samples preprocessed in step S2 are input into the intelligent deep sorting network trained in step S7 for forward propagation to obtain the posterior probability of each pulse in the test samples belonging to each radiation source. The magnitudes of the posterior probabilities of each radiation source are compared, and the radiation source corresponding to the maximum value is taken as the pulse sorting result.

[0105] In this embodiment, the method of the present invention further includes performing a parameter-agile radar sorting performance test on the intelligent deep sorting network trained in step S7, as follows:

[0106] After preprocessing the intercepted signals according to step S2, they are input into the network trained in step S7 for forward propagation to obtain the posterior probability of each pulse in the test sample belonging to each radiation source. The category corresponding to the maximum posterior probability is taken as the radiation source to which the pulse belongs, thus verifying the sorting performance of the intelligent depth recognition network for parameter-agile radar signals.

[0107] Table 1 shows the simulation parameters of the agile radar signal used in this embodiment, including: 4 radar radiation sources and 7 operating modes, determined by the distribution characteristics of PRI modulation mode, PW and RF.

[0108] Table 1

[0109]

[0110] The carrier frequency and pulse width are uniformly distributed within a certain range or vary rapidly between several values. PRI modes include fixed repetition frequency, jittered repetition frequency, and staggered repetition frequency. Parameters from different transmitters overlap. Measurement errors of 1% to 3% are added to all parameters. Four transmitters are assumed to be located in the same direction, and overlapping pulse trains composed of pulses from different transmitters randomly contain 2 to 7 modes. Table 2 shows three scenarios set up in this embodiment to test the performance of the method of the present invention under different conditions: ideal scenario, typical scenario, and extreme scenario.

[0111] Table 2

[0112] Scene Measurement error / % Loss rate / % Ideal scenario 0 0 Typical scenarios 5 25 Extreme scenarios 10 50

[0113] Figure 5 The results of sorting and visualizing parameter-agile radar signals using the method of this invention are presented. The pulse descriptor measurement error of the test samples is 1-3%, and the loss rate is 5%. Figure 5 (a) shows the distribution of RF and PW of the original burst with TOA. Due to the overlap of pulses from multiple transmitters, RF jumps through multiple values, and PW is spuriously distributed within a specific range. The parameters of the original burst appear very chaotic, making it difficult to determine the number of transmitters and correlate the agility parameters with the transmitters. Figure 5 (b) shows the distribution of the sorted pulses in the RF-PW plane. The pulse trains were effectively sorted with an accuracy of 93.46%.

[0114] To further demonstrate the performance of the intelligent signal sorting network proposed in this invention, this embodiment compares the sorting accuracy of the method with that of a network structure that does not include a feature fusion module (Comparison Method 1) and that uses a residual structure without holes instead of multi-branch dilated convolutions (Comparison Method 2) in three scenarios. Figure 6 The image shows the sorting results in three scenarios in this embodiment. The sorting accuracy is calculated as follows:

[0115]

[0116] Where, N c N represents the pulse that was correctly sorted. p This indicates the total number of pulses.

[0117] from Figure 6It can be seen that the method of the present invention has the highest sorting accuracy in all three scenarios, and the accuracy reaches 87.13% with a measurement error of 5% and a pulse loss rate of 25%. Even in the extreme scenario with a measurement error of 10% and a pulse loss rate of 50%, the accuracy is still higher than 73%. As the measurement error and loss rate increase, the multi-branch dilated convolution and feature fusion module can effectively improve the sorting performance of the method of the present invention, and the method of the present invention has strong robustness under complex conditions such as measurement error, pulse loss, and parameter overlap.

[0118] In summary, the method of this invention addresses the problem of low sorting accuracy of parameter-agile radar in complex electromagnetic environments using existing technologies. It uses pulse descriptor sequences as a representation of radar pulses within the observation time. Based on a deep convolutional network architecture and combined with the variation law of the pulse descriptor sequences of parameter-agile radar, multi-branch dilated convolution is used to extract interleaved pulse parameter features, achieving an abstract representation of the pulse sequence in high-dimensional space. Based on a transposed convolution module, a one-to-one mapping between sequence features and their respective radiation sources is achieved step-by-step. The method of this invention can achieve accurate sorting of parameter-agile radar, effectively solving the "batch expansion" problem, and can also adapt to complex scenarios such as parameter overlap and pulse loss.

[0119] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. A method for intelligent sorting of parameter-agile radar signals, the specific steps of which are as follows: S1. Detect the raw radar signal and obtain the pulse descriptor sequence; in, The pulse descriptor sequence includes: pulse arrival time TOA, pulse amplitude PA, pulse carrier frequency RF, pulse width PW, and pulse arrival angle DOA; S2. Preprocess the pulse descriptor sequence in step S1; S3. Construct a feature extraction module based on multi-branch dilated convolution; S4. Construct a feature fusion module based on an attention mechanism; S5. Construct a radiation source mapping module based on transposed convolution; S6. Concatenate the feature extraction module obtained in step S3, the feature fusion module obtained in step S4, and the radiation source mapping module obtained in step S5 to construct an intelligent deep sorting network. S7. Use the sequence obtained in step S2 as the input of the intelligent deep sorting network constructed in step S6 to train the intelligent deep sorting network. S8. Use the intelligent deep sorting network trained in step S7 to sort the parameter-agile radar signals.

2. The intelligent sorting method for parameter-agile radar signals according to claim 1, characterized in that, Step S2 is as follows: The TOA sequence is initialized, and the arrival time of each pulse relative to the first pulse is calculated; mean normalization is performed on RF and PW. Given that the sequence contains N pulses, the pulse sequence expression is as follows: PDW={pdw0,pdw1,…,pdw N-1 } Where PDW represents the set of raw pulse descriptors detected, N represents the number of pulses, and pdw i (i = 0, ..., N-1) represents the pulse description word of the i-th pulse, i.e.: pdw i ={winner i ,rf i ,pw i ,pa i ,prayer i }(i=0,…N-1) Among them, toa i (i = 0, ..., N-1) represents the arrival time of the i-th pulse, rf i (i = 0, ..., N-1) represents the carrier frequency of the i-th pulse, pw i (i = 0, ..., N-1) represents the pulse width of the i-th pulse, pa i (i = 0, ..., N-1) represents the amplitude of the i-th pulse, doa i (i = 0, ..., N-1) represents the arrival angle of the i-th pulse; The expression for the preprocessed pulse sequence PDW' is as follows: PDW'={pdw'0,pdw'1,…,pdw' N-1 } pdw' i ={toa' i ,rf' i ,pw' i } Among them, pdw' i (i = 0, ..., N-1) represents the pulse descriptor of the i-th pulse after preprocessing, toa' i Let represent the arrival time of the i-th pulse after preprocessing, calculated as follows: rf' i The carrier frequency of the i-th pulse after preprocessing is represented by the following expression: pw' i Let represent the pulse width of the i-th pulse after preprocessing. The calculation expression is as follows: If the arrival angle ranges of all radiation sources are set to be similar, then DOA is not used as input and pulse amplitude PA is not considered. If DOA is discriminative, the DOA sequence is directly used as input.

3. The intelligent sorting method for parameter-agile radar signals according to claim 1, characterized in that, Step S3 is as follows: The feature extraction module includes: 1 convolutional layer, 4 basic modules, 4 parallel dilated convolutional branches, each basic module includes 2 residual layers, and each convolutional layer is followed by a batch regularization layer and a ReLU activation function.

4. The intelligent sorting method for parameter-agile radar signals according to claim 1, characterized in that, Step S4 is as follows: The feature fusion module includes: one channel attention layer and one basic module; The channel attention layer includes: one global average pooling layer, two fully connected layers, and two activation function layers; the basic module includes: two residual layers, each followed by a batch regularization layer and a ReLU activation function.

5. The intelligent sorting method for parameter-agile radar signals according to claim 1, characterized in that, Step S5 is as follows: The radiation source mapping module includes: 3 transposed convolutional layers and 4 convolutional layers, wherein each of the 3 convolutional layers is followed by a batch regularization layer and a ReLU activation function.

6. The intelligent sorting method for parameter-agile radar signals according to claim 1, characterized in that, In step S6, the intelligent deep sorting network uses the cross-entropy loss function to output the radiation source to which the pulse descriptor belongs at each time moment.

7. The intelligent sorting method for parameter-agile radar signals according to claim 1, characterized in that, Step S7 is as follows: The training samples processed in step S2 are input into the intelligent deep sorting network constructed in step S6 for forward propagation; the cost function value is calculated, using the cross-entropy loss function as the cost function, and the calculation expression is as follows: loss = CEL(prob, label) Where CEL represents the cross-entropy loss function, prob∈R 1×M Let label represent the output probability, R denote the real number field, and label ∈ R. 1×M This represents the true value of a radiation source in one-hot code form, where M represents the total number of radiation sources. Then, the network parameters are updated using the backpropagation algorithm with momentum gradient descent; the backpropagation is performed iteratively until the cost function converges, resulting in a trained intelligent deep sorting network.

8. The intelligent sorting method for parameter-agile radar signals according to claim 1, characterized in that, Step S8 is as follows: The test samples preprocessed in step S2 are input into the intelligent deep sorting network trained in step S7 for forward propagation to obtain the posterior probability of each pulse in the test samples belonging to each radiation source. The magnitudes of the posterior probabilities of each radiation source are compared, and the radiation source corresponding to the maximum value is taken as the pulse sorting result.

Images