An intelligent radar signal sorting method

Abstract

The present invention discloses an intelligent radar signal sorting method, which first simulates and generates multiple radar pulse sequences. Secondly, the radar pulse sequence is input into a pulse width statistical histogram module to filter out noise interference in the radar pulse sequence. Then, a density-based spatial clustering algorithm DBSCAN and a time fusion Transformer network model TFT are used to construct a radar signal sorting model. Finally, the pulse sequence after filtering out noise interference is trained through a radar signal sorting model, and a radar signal sorting result is obtained according to the output probability value. The present invention reduces the interference of noise signals on pulse sorting, dilutes dense pulse flows to reduce sorting pressure, improves the performance and speed of pulse sorting, and can accurately sort pulse signals.

Description

Technical Field

[0001] The present invention belongs to the field of radar signal sorting, and specifically relates to a radar signal intelligent sorting method integrating clustering and a temporal fusion transformer network model (Temporal Fusion Transformer, TFT). Background Art

[0002] With the continuous escalation of electronic countermeasures, the signal environment is becoming increasingly complex. Rapid and accurate sorting of unknown radar signals has become a key task in modern electronic warfare. Radar signal sorting, as the core step of radar reconnaissance and signal processing technology, extracts the characteristics of radar radiation sources to achieve signal recognition and classification, providing important information support for decision-making. Advanced recognition technology can quickly detect and analyze the modulation mode and parameters of the radar to determine its function. With the rapid development of electronic technology and the continuous emergence of radars with complex modulation modes, the current electromagnetic and space environment has become more complex and diverse, which has brought challenges to the sorting and identification of radar signals. In such an electromagnetic environment, how to achieve real-time and accurate radar signal sorting and identification has become an important research direction in the field of radar reconnaissance.

[0003] The current mainstream radar signal sorting algorithms cover a variety of technical paths, ranging from traditional methods to the wide application of artificial intelligence. Pulse description word (PDW) classification algorithms, such as K-means clustering, fuzzy C-means clustering, and density clustering (such as DBSCAN), cluster the parameters of radar pulses, and then, based on clustering, a series of pulse sorting algorithms such as the dynamic extended association method based on pulse arrival time (TOA), the statistical histogram method, the cumulative difference histogram algorithm CDIF, the sequence difference histogram algorithm SDIF, the PRI transformation method, the PRI correction method, and the plane transformation method. In addition, time-frequency analysis techniques, such as short-time Fourier transform (STFT) and wavelet transform, can effectively extract the time-frequency characteristics of non-stationary signals and are suitable for radar signal sorting with complex modulation modes. With the rapid development of artificial intelligence, machine learning algorithms such as support vector machines (SVM), convolutional neural networks (CNN), recurrent neural networks (RNN), and graph neural networks (GNN) have been widely used in radar signal sorting tasks, further improving the accuracy and efficiency of sorting. At the same time, adaptive signal processing algorithms, such as adaptive filtering and Kalman filtering, can achieve accurate signal sorting in complex electromagnetic environments by dynamically adjusting filter parameters. The above methods are often used in combination to cope with today's complex and changing signal environments.

[0004] In a complex electromagnetic environment, radar signal sorting faces many challenges. A single sorting algorithm is difficult to deal with various problems, such as pulse loss and noise interference. In addition, existing signal sorting algorithms often have a relatively single adaptability range and fail to fully utilize all the characteristics of radar pulse descriptors. These problems will significantly affect the signal sorting effect in practical applications, resulting in false detection or missed detection. Summary of the invention

[0005] The present invention considers combining the three characteristics of pulse arrival angle (AOA), pulse width and pulse arrival time (TOA) as sorting parameters, integrates the respective advantages of clustering algorithm and TFT, and proposes a radar signal intelligent sorting method based on fusion clustering and TFT. By organically combining the machine learning algorithm with the large model and combining the advantages of the two methods, this method can better capture the timing and spatial characteristics of radar signals in complex electromagnetic environments, and significantly improve the accuracy and robustness of sorting. This method can not only effectively handle the complex interactions of multi-source radar signals, but also cope with various interference and signal loss problems in actual environments, showing excellent sorting performance.

[0006] The technical solution of the present invention is:

[0007] Step 1: Simulate and generate multiple radar pulse sequences, which are composed of pulse description words (PDWs). The pulse angle of arrival (AOA), pulse width (PW) and pulse arrival time (TOA) in the PDW parameters are extracted to construct A sequence combination of . represents the pulse arrival angle of the first pulse in the pulse train, represents the pulse width of the first pulse in the pulse train, Indicates the pulse arrival time of the first pulse in the pulse train. Similarly, the radar pulse train The CCP includes A pulse.

[0008] Step 2: Input the radar pulse sequence obtained in step 1 into the pulse width statistical histogram module to filter out the noise interference in the radar pulse sequence, that is, the output of the pulse width statistical histogram module is the filtered radar pulse sequence.

[0009] Step 3: Use the density-based spatial clustering algorithm DBSCAN and the time fusion Transformer network model TFT to build a radar signal sorting model.

[0010] Step 4: The pulse sequence after noise interference is filtered out in step 2 is trained through the radar signal sorting model constructed in step 3; the real radar pulse sequence is input for model testing to complete radar signal sorting, and the radar signal sorting result is obtained according to the output probability value.

[0011] The present invention proposes a radar signal intelligent sorting method that integrates clustering and TFT. By extracting the pulse arrival angle (AOA), pulse width (PW), pulse arrival time (TOA), and first-order TOA differential features of the signal, histogram statistics and fusion clustering are used to achieve denoising and dilution of pulse information, and TFT is used to predict the processed time series to generate a training model for radar signal sorting.

[0012] Furthermore, in the above step 1, the generated pulse sequence features include:

[0013] Pulse arrival angle (AOA), pulse width (PW), pulse arrival time (TOA), generate multiple pulse sequences as in step 1 In the figure, the pulse arrival angle (AOA), pulse width (PW) and pulse arrival time (TOA) of each radar pulse sequence are different. Multiple radar pulse sequences are arranged in ascending order according to the (TOA) value.

[0014] Furthermore, the radar pulse sequence after noise filtering is input into the clustering module, and the pulse arrival angle (AOA) is clustered and grouped using the DBSCAN algorithm (density-based spatial clustering algorithm) to dilute the pulse data stream. The grouped pulse sequence is digitally processed and input into the TFT model to perform signal sorting based on the pulse arrival time (TOA) and pulse width (PW).

[0015] Based on the pulse width statistical histogram module, in the actual signal detection process, due to the limitation of detection accuracy, the pulse width of the transmitted signal will have a certain range of deviation during actual measurement. Therefore, in the process of establishing the pulse width statistical histogram, the pulse width value with a deviation within 5% is set as Class 1 pulse width value, and its mean is taken to represent the pulse width value.

[0016] The statistical process is as follows, starting with the first pulse width (PW), marked as , search for pulse width values ​​with a deviation of less than 5%, and classify the pulse width values ​​that meet the conditions into category 1. Traverse the remaining pulse width values ​​that have not been classified, and repeat the above process to complete the statistical steps of the pulse width statistical histogram. The formula is expressed as follows:

[0017] Set the pulse width value to ; Initialize the unsorted collection ;

[0018] Retrieve the unsorted collection The first pulse width value , find Pulse width values ​​with a deviation within 5% are classified and compared with Classify into statistical combinations of pulse width values , otherwise Classify as isolated points and put into statistical combination of pulse width values , complete the first round of loop, take out the first value of the remaining pulse width values, and proceed to the next round of loop until the remaining unclassified pulse width values ​​are traversed, and the final statistical combination of pulse width values ​​is output. ;

[0019] For the statistical combination of pulse width values , set the threshold value to 3, the number of pulse width values ​​counted If it is less than 3, the pulse width value is filtered out. The effective pulse width value combination after statistical filtering is expressed as:

[0020]

[0021] According to the combination of valid pulse width values, the abnormal pulse width values ​​in the radar pulse sequence are eliminated.

[0022] Furthermore, the radar pulse sequence after filtering out interference through pulse width statistical histogram is input into the clustering module, and clustering is performed based on the pulse angle of arrival (AOA) parameter. The DBSCAN clustering algorithm is used to effectively separate the different pulse angle of arrival (AOA) parameters in the radar pulse sequence, and distinguish the pulse signals from different radars.

[0023] Furthermore, DBSCAN relies on two important parameters: ε (the neighborhood radius of the sample point) and MinPts (the minimum number of points).

[0024] ε: defines the neighborhood radius of a sample point, also called the "epsilon neighborhood".

[0025] MinPts: defines the minimum number of neighborhood points required for a point to be considered a core point.

[0026] When clustering radar pulse sequences, the setting of two important parameters is particularly important. However, when sorting unknown radar signals, the setting of fixed parameters has limitations for clustering unknown radar signals. Therefore, the present invention adopts a method of adaptive parameter adjustment, that is, automatically adjusting the ε and MinPts parameters, and selecting the optimal parameters through cross-validation or statistical analysis based on data distribution. This can be achieved by introducing dynamic ε and setting different ε values ​​based on the AOA distribution characteristics.

[0027] The DBSCAN algorithm includes a neighborhood search. Using an efficient neighborhood search algorithm can reduce the complexity of the algorithm and speed up clustering. In the present invention, KD-Tree is selected to accelerate the nearest neighbor query. KD-Tree is a partitioning plane binary tree used to store point data in k-dimensional space to support fast nearest neighbor search. KD-Tree recursively divides the data space into smaller areas, and divides the data into left and right subspaces through the plane (in each dimension).

[0028] Furthermore, this method can dilute the data flow in the preprocessing stage, reduce data redundancy, and effectively reduce the computational pressure of pulse sorting and recognition in the subsequent stage, thereby improving the efficiency and accuracy of the overall sorting system. In this way, the system can achieve more efficient radar signal sorting in a complex electromagnetic environment, providing more accurate and clear input data for subsequent deep learning model processing.

[0029] The radar pulse sequence processed by the DBSCAN clustering algorithm is further divided into multiple groups of radar pulse sequences, and the grouped radar pulse sequences are digitally processed.

[0030] First, a set of pulse trains after cluster sorting is expressed as:

[0031]

[0032] Calculate the first-order TOA difference PRI of the radar pulse sequence and replace the TOA value, that is, the pulse repetition interval:

[0033]

[0034] After the first-order difference It is expressed as:

[0035]

[0036] Right now , is the difference between the pulse arrival time of the second sample point and the pulse arrival time of the first sample point.

[0037] say For the sequence feature combination, the feature combination is digitized and the quantization factor is set to , that is, the feature composition after digital processing is expressed as:

[0038]

[0039] The feature combination after digital processing brings features of different scales into a similar range, preventing certain features from dominating the model's learning during training due to their large values, causing the model to be unable to learn useful patterns.

[0040] Furthermore, the TFT model is composed of an input layer, a long short-term memory network (LSTM), a temporal fusion layer (TemporalFusion Layer), a static enhancement layer, a skip connection gating mechanism, and a full output layer.

[0041] First, the digitally processed features are combined Input to the input layer of the TFT model, and then enter the time series encoding layer. TFT does not directly use the position encoding of the traditional Transformer (such as the sinusoidal position encoding in the original Transformer), because the time series encoding layer uses LSTM to process the dependencies of the time series. LSTM itself has the ability to perceive order, because it processes sequence data recursively, and the information of each time step depends on the previous time step. In the input pulse sequence, after the pulse arrival angle (AOA) is clustered and sorted, the feature sequences of the same group of inputs have basically the same AOA parameters. Therefore, this parameter can be processed using a static variable layer, which is encoded by static features to assist the output and attention behavior of LSTM.

[0042] Furthermore, the time fusion layer in TFT uses a multi-head attention mechanism, which is used to capture long-term dependencies in radar pulse sequences, that is, similar PRI values. The attention mechanism can identify these changes and make adjustments. After being processed by LSTM and the attention mechanism, the features are further extracted through GRN to extract nonlinear relationships. The gating mechanism of GRN can filter out unimportant information, while the residual connection can ensure that the original features are not completely ignored. Finally, the features processed by each module will be summarized by the output layer to output the pulse sequence prediction value.

[0043] The overall training process of the radar signal intelligent sorting model integrating clustering and TFT is as follows: the radar pulse sequence after interference is filtered out by the pulse width statistical histogram is input into the clustering sorting diluted pulse sequence, and the diluted pulse sequence is input into the TFT model for pulse sequence prediction. The loss function of the model output result and the real pulse sequence is estimated, iterative optimization is performed, and the Adam optimizer is used to save the training model.

[0044] Use the generated pulse sequence to verify and evaluate the model effect:

[0045] After the model training is completed, the corresponding .pth file will be generated. By loading the best performing model in the training phase, the weight parameters of the network can be imported, and the test set can be input into the model for inference. The test results obtained after running will be used to evaluate the performance of the model and calculate indicators such as accuracy and success rate. In order to verify the improvement effect of the proposed method in terms of success rate and accuracy, a comparative analysis will be conducted with the current mainstream algorithm, and the corresponding curve graph will be drawn to show the results.

[0046] Beneficial effects of the present invention: (1) In view of the difficulty of radar signal sorting in a complex electromagnetic environment, the present invention proposes a radar signal sorting structure with multi-processing fusion, which reduces the interference of noise signals on pulse sorting and dilutes dense pulse streams to reduce sorting pressure. (2) In order to solve the problem of excessive dependence of pulse sequences in long time series, the TFT model is introduced to predict the pulse sequence. Different from the serial input of RNN, TFT adopts parallel input, thereby improving the speed and performance of pulse sorting. In the processing of long time series, RNN will have the problem of gradient disappearance or gradient explosion. TFT introduces a self-attention mechanism to reduce the correlation dependency between sequences, so that the effect is better in the process of processing long sequences. (3) TFT is an improvement of Transformer. Different from Transformer, which needs to add position encoding information to the input sequence, the time encoder layer of TFT adopts LSTM to process the dependency of time series. Thereby improving the performance and speed of pulse sorting. The radar signal sorting model fused by multi-processing methods has enhanced anti-interference ability, adaptability to dense pulse data stream, and can accurately sort pulse signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flow chart of the method proposed by the present invention;

[0048] Figure 2 It is a sequence comparison diagram of radar pulse sequence after pulse width statistical histogram filtering;

[0049] Figure 3 It is a schematic diagram of the radar pulse sequence after clustering and sorting;

[0050] Figure 4 This is the overall structure diagram of the TFT model;

[0051] Figure 5 is the loss curve and training accuracy of radar signal sorting model training;

[0052] Figure 6 It is the test result curve of radar signal sorting model. DETAILED DESCRIPTION

[0053] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0054] The present invention proposes a radar signal intelligent sorting method based on fusion clustering and TFT to solve the problems of low sorting accuracy, signal batch increase, and poor anti-interference ability in existing signal sorting technologies. Existing signal sorting algorithms often have the disadvantages of a relatively single adaptation range and not fully utilizing all the characteristics of radar pulse descriptors. Therefore, the present invention considers combining the three characteristics of pulse arrival angle (AOA), pulse width (PW) and pulse arrival time (TOA) as sorting parameters, integrating the respective advantages of histogram statistics, clustering algorithm and TFT to achieve intelligent sorting of radar signals.

[0055] Reference Figure 1 It can be known that the specific process of the present invention is as follows: Step 1: Simulate and generate multiple radar pulse sequences, the pulse sequence consists of a pulse description word (PDW), extract the pulse arrival angle (AOA), pulse width (PW) and pulse arrival time (TOA) in the PDW parameters to construct

[0056] A sequence combination of . represents the pulse arrival angle of the first pulse in the pulse train, represents the pulse width of the first pulse in the pulse train, It represents the pulse arrival time of the first pulse in the pulse sequence. The radar pulse sequence x contains n pulses in total.

[0057] Step 2: Input the radar pulse sequence obtained in step 1 into the pulse width statistical histogram module to filter out the noise interference in the radar pulse sequence. That is, the output of the pulse width (PW) statistical histogram module is the filtered radar pulse sequence. The experimental comparison results are shown in Figure 2 shown.

[0058] Step 3: Integrate the pulse clustering algorithm and TFT network model to build a radar signal intelligent sorting model.

[0059] Step 4: The pulse sequence after noise interference is filtered out in step 2 is trained through the radar signal sorting model constructed in step 3; the real radar pulse sequence is input for model testing to complete radar signal sorting.

[0060] Furthermore, in the above step 1, the specific implementation steps and sequence characteristics include the following steps:

[0061] Generate multiple pulse trains with three parameters: pulse angle of arrival (AOA), pulse width (PW), and pulse arrival time (TOA):

[0062]

[0063] …

[0064] Combine multiple pulse sequences and arrange them in ascending order of TOA.

[0065] Furthermore, the pulse width histogram statistical filtering is performed on the radar pulse sequence arranged in ascending order. In the actual signal detection process, due to the limitation of detection accuracy, the pulse width of the transmitted signal will have a certain range of deviation in actual measurement. Therefore, in the process of establishing the statistical histogram of the pulse width, the pulse width value with a deviation within 5% is set as Class 1 pulse width value, and its mean is taken to represent the pulse width value.

[0066] The statistical process is as follows, starting with the first pulse width (PW), marked as , search for pulse width values ​​with a deviation of less than 5%, and classify the pulse width values ​​that meet the conditions into category 1. Traverse the remaining pulse width values ​​that have not been classified, and repeat the above process to complete the histogram statistics step of the pulse width. The formula is expressed as follows:

[0067] Set the pulse width value to , initialize the unsorted collection ;

[0068] Loop through the classification and take out the first unclassified pulse width value (For example ), find Pulse width values ​​with a deviation within 5% are classified and compared with Classify into statistical combinations of pulse width values , otherwise Classify as isolated points and put into statistical combination of pulse width values , complete the first round of loop, take out the first value of the remaining pulse width values, and proceed to the next round of loop until the remaining unclassified pulse width values ​​are traversed, and the final statistical combination of pulse width values ​​is output. .

[0069] For the statistical combination of pulse width values , set the threshold value to 3, the number of pulse width values ​​counted If it is less than 3, the pulse width value is filtered out. The effective pulse width value combination after statistical filtering is expressed as:

[0070]

[0071] That is, only those pulse width categories whose number is greater than or equal to 3 are retained in the statistical results, and the abnormal pulse width values ​​in the radar pulse sequence are eliminated based on the combination of valid pulse width values.

[0072] Furthermore, the radar pulse sequence after filtering out interference is clustered using the DBSCAN algorithm. The results are as follows: Figure 3 As shown in the figure, this method adopts the method of adaptive parameter adjustment, that is, automatically adjusts the ε and MinPts parameters, and selects the optimal parameters through cross-validation or statistical analysis based on data distribution. This can be achieved by introducing dynamic ε and setting different ε values ​​based on the characteristics of AOA distribution. An efficient neighborhood search algorithm can reduce the complexity of the algorithm and speed up clustering. In this method, KD-Tree is selected to accelerate the nearest neighbor query. KD-Tree is a partition plane binary tree used to store point data in k-dimensional space to support fast nearest neighbor search.

[0073] The clustered pulse sequence is further digitized to adapt to the input vector characteristics of the TFT.

[0074] First, a group of radar pulse sequences after clustering and sorting is expressed as

[0075]

[0076] Calculate the first-order TOA difference PRI of the radar pulse sequence, that is, the pulse repetition interval:

[0077]

[0078] After the first-order difference It is expressed as:

[0079]

[0080] Right now , is the difference between the pulse arrival time of the second sample point and the pulse arrival time of the first sample point. The feature combination is a sequence feature combination, and the feature combination is digitized. The quantization factor is selected as , , , that is, the feature composition after digital processing is expressed as:

[0081]

[0082] The feature combination after digital processing brings features of different scales into a similar range, preventing certain features from dominating the model's learning during training due to their large values, causing the model to be unable to learn useful patterns.

[0083] The digitized pulse sequence is input into the TFT model. The structure of the TFT model is as follows: Figure 4 As shown in Figure 1, the TFT model consists of an input layer, a long short-term memory network (LSTM), a time fusion layer, a static enhancement layer, a skip connection gating mechanism, and a full output layer. The model level parameters are shown in Table 1 below:

[0084] Table 1 Model level parameters

[0085]

[0086] The embedding input dimension of the input layer is set to (64, 64, 3), where each parameter is defined as batch size, time step and input feature dimension. The batch is selected as 64, that is, 64 data are input for training each time. Because the 64 data are sequential sequences, the time step is 64. Each sample in the sequence contains three feature values: pulse arrival angle (AOA), pulse width (PW), and pulse arrival time (TOA), so the feature dimension is set to 3. The output dimension of the embedding layer is (64, 64, 32), 32 is the high-dimensional mapping of the feature dimension, because the input pulse sequence contains multiple radar signals, and the two sample values ​​between the sequences may belong to different radars. High-dimensional mapping of the feature dimension can help the model better capture the complex patterns of the data. Provide richer input for the next level LSTM Encoder.

[0087] The input tensor after feature dimension mapping is input to the LSTM Encoder, and its output dimension is (64, 64, 64), which represents the input batch size of 64 and the input time step of 64. The last dimension of the output after the bidirectional LSTM encoding layer is the number of LSTM units, which is set to 64 in this method. The input batch is small and the time step is short. Using a smaller number of LSTM units can avoid overfitting.

[0088] The LSTM encoding layer uses bidirectional LSTM through two LSTM layers, forward and backward, to capture past and future dependencies at the same time, enhance the understanding of the dependency of data before and after a certain data sample point, better capture the temporal dependency, and thus make reasonable predictions for the sample points.

[0089] The feature tensor processed by the bidirectional LSTM is input to the time fusion layer of the temporal interpretation network. It is mapped to three vectors: query Q, key K and value V.

[0090]

[0091] are the weight matrices for query, key, and value respectively. Compute the correlation between time steps and then scale them (by Normalized), and then converted into a set of weights through the Softmax function. The weights will be used as thresholds to filter features. Use these weights to perform weighted summation on the value vector V to obtain the attention output.

[0092]

[0093] The temporal fusion layer contains a multi-head attention mechanism, concatenating the outputs of all heads and then passing through a linear transformation Get the final output.

[0094]

[0095] In order to stabilize training and prevent gradient disappearance or gradient explosion, residual connections and layer normalization are usually added after the multi-head attention layer:

[0096]

[0097] A residual connection is formed by adding the input X to the multi-head attention output.

[0098] At the same time, layer normalization is introduced to ensure that the output of each layer has a stable distribution

[0099]

[0100] in and are the mean and standard deviation, is a tiny constant that prevents division by zero.

[0101] After multiple layers of temporal fusion processing, the output feature dimension is (64, 64, 64).

[0102] The pulse arrival angle (AOA) in the input feature vector is a stable parameter, so it can be used as a static feature of the model and fused with the feature vector output by the time fusion layer to inject static information into the time series feature to enhance the contextual information between time steps. The output feature dimension remains unchanged and is still (64,64,64).

[0103] Gated skip connection, this layer allows information to jump from the previous layer to the next layer through a gating mechanism, and selectively passes useful information through the gating mechanism. Skip connections are usually used to prevent information loss, gradient vanishing or gradient explosion problems, and the gating mechanism provides a filter to control which information should be retained. In the present invention, the tensor after static feature fusion is selected as the current input layer, the level of LSTM output is selected as the previous layer, and the gating vector g is used to selectively control the information from the previous layer and the current layer. The final output is a weighted combination between the two inputs:

[0104]

[0105] It indicates that the gating vector g controls the flow of features in the previous layer. Indicates that the gate vector 1-g controls the retention of the current layer features.

[0106] The tensor output by the gated jump connection is input to a fully connected layer to output the predicted value. According to the output pulse sequence prediction value, the predicted next pulse value and the actual corresponding next pulse value are calculated using the loss function, and optimized using the Adam optimizer. The learning rate is set to 0.003, and the model prediction results are optimized through iterative training. After multiple rounds of iterations, the model can more accurately predict the next pulse value. According to the predicted pulse value, the radar pulse sequence that meets the requirements is eliminated, and the radar pulse sorting is completed to obtain the final recognition model. Then the test features are put into the trained recognition model for target testing and recognition.

[0107] The specific embodiments are:

[0108] 1: Generate a training set, specifically:

[0109] The radar pulse data set is generated by software for experimental verification. The data set contains 6 different radar pulse sequences, which contain various complex situations such as false pulses, repetition rate PRI, pulse loss, etc. The arrival time range of various pulses is 0~20000us, as shown in Table 2.

[0110] Table 2 Radar full pulse data table

[0111]

[0112] 2: Data preprocessing, specifically:

[0113] The mixed radar pulse sequence is arranged in ascending order of TOA. It is input into the pulse width histogram statistics module to filter out noise. After filtering, the radar pulse sequence is input into the clustering module to cluster and sort the pulse sequence according to AOA. The first-order difference of TOA is calculated for each group of pulse sequences after clustering and sorting.

[0114] 3: Network parameter settings, specifically: the size of the input layer is set to 64×64×3, and the output size is 64×64×32. The output feature vector is encoded by bidirectional LSTM to capture the timing information, and the output size is 64×64×64. After passing through the temporal interpretation network, the output size is 64×64×64. After passing through the static feature coverage layer and the gated skip connection network, the output size is 64×64×64. Finally, after passing through the fully connected layer, the pulse sequence prediction value is output.

[0115] 4: Training parameter settings, specifically: training learning rate is 0.003, the number of rounds is 100, the loss function uses the cross loss function, the training optimizer selects Adam, and the batch size is 64.

[0116] 5: Identification result analysis, specifically: run the program file and evaluate the test results, such as Figure 5 As shown, observe the changes in the training loss function and the success rate curve, save and import the training model, input a complex repetition frequency pulse sequence, perform radar signal sorting, and compare the sorting accuracy results with the current mainstream algorithm. Figure 6 The actual sorting accuracy is measured and compared with other algorithms, the present invention is improved. The comparison results are shown in Table 3 below:

[0117] Table 3 Comparison results

[0118]

[0119] The above is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any simple modification, change and equivalent change made to the above embodiment according to the technical essence of the invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. An intelligent radar signal sorting method, characterized in that: The following steps are involved: Step 1: Simulate and generate multiple radar pulse sequences; The radar pulse sequence is composed of a pulse description word PDW, and the pulse arrival angle AOA, pulse width PW and pulse arrival time TOA in the PDW parameters are extracted to construct x=[[aoa1,pw1,toa1],[aoa2,pw2,toa2],…,[aoa n ,pw n ,toa n ]]; where aoa1 represents the pulse arrival angle of the first pulse in the pulse sequence, pw1 represents the pulse width of the first pulse in the pulse sequence, toa1 represents the pulse arrival time of the first pulse in the pulse sequence, and so on. The radar pulse sequence x contains n pulses in total; Step 2: Input the radar pulse sequence into the pulse width statistical histogram module to filter out the noise interference in the radar pulse sequence; Step 3: Use the density-based spatial clustering algorithm DBSCAN and the time fusion Transformer network model TFT to build a radar signal sorting model; The radar signal sorting model is specifically implemented as follows: the radar pulse sequence after filtering out noise interference in the radar pulse sequence is input into the clustering module, the pulse arrival angle AOA is clustered and grouped using the density-based spatial clustering DBSCAN algorithm, and the grouped pulse sequence is digitally processed and input into the TFT model for signal sorting; Step 4: The pulse sequence after filtering out noise interference is trained through the radar signal sorting model, and the radar signal sorting result is obtained according to the output probability value.

2. The intelligent radar signal sorting method according to claim 1 is characterized in that: The radar pulse sequence x also includes arranging multiple radar pulse sequences in ascending order according to TOA values.

3. The intelligent radar signal sorting method according to claim 2 is characterized in that: The pulse width statistical histogram module implementation process is as follows: Set the pulse width value set to W = {w1, w2, ..., w n }; Initialize the unclassified set U = W; Take out the first pulse width value w of the unclassified set U j , find the j The pulse width values ​​with a deviation within 5% are classified and compared with w j into a category and put into the statistical combination C of pulse width values, otherwise w j The isolated points are put into the statistical combination C of pulse width values, the first cycle is completed, the first value of the remaining pulse width values ​​is taken out, and the next cycle is performed until the remaining pulse width values ​​that have not been classified are traversed, and the final statistical combination C of pulse width values ​​is output; For the statistical combination C of pulse width values, the threshold is set to 3, and the statistical number of pulse width values ​​is n. j If the pulse width is less than 3, the pulse width value is filtered out. The effective pulse width value combination after statistical filtering is expressed as: C effective = {C|n j ≥3}; According to the combination of valid pulse width values, the abnormal pulse width values ​​in the radar pulse sequence are eliminated.

4. The intelligent radar signal sorting method according to claim 3 is characterized in that: The density-based spatial clustering algorithm relies on two parameters: the neighborhood radius ε of the sample point and the minimum number of points MinPts; ε defines the neighborhood radius of a sample point; MinPts defines the minimum number of neighborhood points required for a point to be considered a core point; Automatically adjust ε and MinPts parameters, and select the optimal parameters through cross-validation or statistical analysis based on data distribution.

5. The intelligent radar signal sorting method according to claim 4 is characterized in that: The digital processing process is as follows: A set of pulse trains after clustering is expressed as y1=[[aoa1,pw1,toa1],[aoa2,pw2,toa2],…,[aoa n ,pw n ,hero n ]]; Calculate the first-order TOA difference PRI of the radar pulse sequence and replace the TOA value, that is, the pulse repetition interval PRI i =hero i+1 -win i ,i=1,…n; Then y1 after the first-order difference is expressed as: y1=[[aoa1,pw1,pri1],[aoa2,pw2,pri2],…,[aoa n ,pw n ,at n ]]; That is, pri1 = toa2-toa1, which is the difference between the pulse arrival time of the second sample point and the pulse arrival time of the first sample point; We call y1 the sequence feature combination, digitize the feature combination, and set the quantization factor to d aoa ,d pw and d pri , the feature combination after digital processing is expressed as: y 11 =[[aoa1 / d aoa ,pw1 / d pw ,pri1 / d pri ],…,[[aoa n / d aoa ,pw n / d pw ,first n / d pri ]]]。

Images