An individual identification method for aliasing radar based on PDW clustering

Abstract

The application discloses a kind of based on PDW clustering's aliasing radar individual identification method, belong to radar signal identification field.The application implementation method is: using carrier frequency in PDW, pulse width and direction of arrival are carried out k-means clustering to extract the features of hidden PRI and second-order PRI in radar signal;Seven features of PRI, second-order PRI, time of arrival, carrier frequency, pulse width, pulse amplitude, direction of arrival are used.Construct the deep convolutional neural network AAVGGPro model for radar signal identification, by increasing fully connected layer in adaptive channel attention submodule and adaptive spatial attention submodule respectively to obtain AAVGG model, by adjusting the number of channels and convolution kernel size in AAVGG model, increase fully connected layer to obtain AAVGGPro model.Improved loss function is used to train the deep convolutional neural network AAVGGPro model to extract the high-dimensional features of radar signal, and the improved loss function is composed of cross-entropy loss function, inter-class loss function and intra-class loss function;According to the convergence of AAVGGPro model after training, the input radar signal is classified to realize automatic identification of radar signal.

Description

Technical Field

[0001] This invention relates to a method for identifying individual radar entities in aliased radar based on PDW clustering, belonging to the field of radar signal identification. Background Technology

[0002] Radar signal recognition algorithms are based on the received overlapping radar signals. By analyzing and processing the signals, radar characteristic parameter information is obtained. Then, through classification and recognition algorithms, the correlation of parameters of the same radar signals and the differences of parameters of different radar signals are used to distinguish and identify radar signals.

[0003] Generally speaking, research on radar source signal identification can be mainly divided into three research directions: radar source identification based on inter-pulse modulation characteristics, radar source identification based on intra-pulse modulation characteristics, and radar source identification based on machine learning. Traditionally, signal sorting and identification are two sequential processes. However, with the innovative development of signal feature extraction and other processing technologies, the distinction between signal sorting and identification has become blurred. Often, signal identification is achieved simultaneously with radar signal sorting; conversely, signal identification often plays a crucial role as the primary sorting method, and the identification results are fed back to the sorting process, further improving its efficiency and accuracy.

[0004] Traditional signal identification methods rely on the inter-pulse characteristics of the full-pulse signal and the pulse repetition interval (PRI) extracted from the pulse arrival time for identification. Inter-pulse characteristics include time of arrival (TOA), radio frequency (RF), pulse width (PW), pulse amplitude (PA), and direction of arrival (DOA). These five parameters—TOA, RF, PW, PA, and DOA—are the five most typical instantaneous parameters of a pulse descriptive word, meaning they can be directly obtained through measurement. However, with the diversification of modulation methods and the increasing complexity of the battlefield electromagnetic environment, relying solely on inter-pulse characteristic information is insufficient to achieve ideal identification accuracy. Furthermore, traditional algorithms cannot meet the precision requirements of modern battlefields for identifying complex radiation source signals.

[0005] In recent years, deep learning has achieved great success in machine vision, natural language processing, and other fields. Some studies have combined processed one-dimensional radar signals with neural network algorithms to perform recognition using the signal's time-frequency image. However, the time-frequency image only reflects the intra-pulse modulation information of the radar signal and cannot represent the complex and variable inter-pulse modulation information of modern radar signals; furthermore, the time-frequency image involves image transformation, resulting in high computational complexity. In contrast, directly processing the signal and learning its features can better extract its characteristic information while shortening the training time. Summary of the Invention

[0006] To address the issue that traditional radar individual identification algorithms are insufficient in terms of accuracy to meet the requirements of identifying complex radar radiation source signals on modern battlefields, the main objective of this invention is to provide a mixed radar individual identification method based on PDW clustering. This method utilizes k-means clustering based on carrier frequency, pulse width, and direction of arrival (PDW) to extract hidden PRI and second-order PRI features from radar signals. A neural network model, AAVGGPro, is trained using seven features: PRI, second-order PRI, time of arrival, carrier frequency, pulse width, pulse amplitude, and direction of arrival. The seven features of each radar signal to be identified are input into the trained AAVGGPro model, and a predicted label for each radar signal is output. The output predicted label is used as the category corresponding to the radar signal to obtain the radar signal identification result, thus achieving automatic radar signal identification.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] This invention discloses a method for identifying individual radar signals in aliased systems based on PDW clustering. Building upon existing radar signal identification systems, it selects three inter-pulse modulation features—carrier frequency, pulse width, and direction of arrival—and uses a k-means clustering algorithm to cluster the radar signals. Within each cluster, the arrival time of all radar signals is differentially analyzed to obtain the PRI feature, and then the PRI is differentially analyzed again to obtain the second-order PRI feature. Finally, the five original features and the two calculated features within each cluster are normalized to satisfy a distribution with a mean of 0 and a standard deviation of 1, resulting in a preprocessed radar signal identification dataset. The five original features refer to arrival time, carrier frequency, pulse width, pulse amplitude, and direction of arrival, and the two features refer to PRI and second-order PRI. A deep convolutional neural network (AAVGGPro) model for radar signal recognition is constructed. Its basic network skeleton is VGG (Visual Geometry Group). The AAVGG model is obtained by adding fully connected layers to the adaptive channel attention submodule and the adaptive spatial attention submodule. AAVGGPro is obtained by adjusting the number of channels and the convolutional kernel size in the AAVGG model and adding fully connected layers, thus enhancing the AAVGGPro model's ability to extract nonlinear features from radar signals. The AAVGGPro model consists of convolutional layers, an adaptive attention mechanism layer, and fully connected layers. The adaptive attention mechanism layer is composed of the adaptive channel attention submodule and the adaptive spatial attention submodule connected in series. An improved loss function is used to train the AAVGGPro model to extract high-dimensional features from radar signals. This improved loss function consists of a cross-entropy loss function, an inter-class loss function, and an intra-class loss function. Based on the converged AAVGGPro model, the input radar signal is classified, achieving automatic radar signal recognition.

[0009] This invention discloses a method for identifying individual objects in aliased radar based on PDW clustering, comprising the following steps:

[0010] Step 1: Preprocess the dataset. The training dataset only includes clean, interference-free inter-pulse feature data from each radar. Based on the parameters of the AAVGGPro model constructed in Step 2, a mixed dataset is obtained by temporal aliasing. This mixed dataset simulates the individual identification situation of real-world aliased radars. For each aliased dataset, using the carrier frequency, pulse width, and direction of arrival (DAM) features from that dataset, the k-means algorithm is used to divide the dataset into k clusters.

[0011] The K-means algorithm divides radar signals into k clusters based on the Euclidean distance between three inter-pulse modulation features: carrier frequency, pulse width, and direction of arrival. This ensures that radar signal features within a cluster are as close as possible to each other, while maximizing the distance between radar signals in different clusters. The specific implementation of the K-means algorithm for dividing radar signals into k clusters based on the Euclidean distance is as follows: Step 1: Randomly select k radar signal samples as cluster centers for radar signal identification; Step 2: Calculate the distance from each radar signal sample point to the center vector of each radar signal identification cluster and assign it to the closest cluster; Step 3: Update the center vector of each radar signal identification cluster; Step 4: Check if the new center vector of the radar signal identification cluster has changed. If it has changed, return to Step 2; if the center vector of the radar signal identification cluster no longer changes, stop and output the clustering result. A maximum iteration parameter `max_iters` is set to prevent the iteration from entering an infinite loop. If the number of iterations exceeds the `max_iters` parameter, the iteration automatically stops and the result is output. In the radar signal recognition training set, the true labels of radar signals within the same cluster are changed to the mode of the true labels in that cluster. In the validation and test sets, the AAVGGPro model predicts the category of the radar signals in that cluster as the mode of the prediction results. The Euclidean distance formula is expressed as follows:

[0012]

[0013] Where X = (x1, x2, ..., x...) n Y = (y1, y1, ..., y1) n ), where n is the length of vector X and vector Y.

[0014] For radar signal identification samples within the same cluster, the arrival time (TOA) is first differentially divided to obtain the pulse repetition interval (PRI), and then the PRI is differentially divided to obtain the second-order PRI. The calculation formula is as follows:

[0015] PRI i =|TOA i+1 -TOA i | (2)

[0016] 2nd order PRI i =|PRI i+1 -PRI i | (3)

[0017] After processing as described in equations (2) and (3), the feature set of each radar signal identification sample is features = {PRI, 2nd order PRI, TOA, RF, PW, PA, DOA}. The features of radar signal identification samples within the same cluster are normalized to ensure their feature distribution has a mean of 0 and a standard deviation of 1. The calculation formula is as follows:

[0018]

[0019] Here, feature is an element belonging to the feature set features above, μ is the mean of the feature, and σ is the standard deviation of the feature.

[0020] After preprocessing the radar signal identification data of the same cluster, seven normalized features are obtained. Then, the original radar signal identification data is transformed into a single data sample consisting of n consecutive data points, changing the feature shape of the radar signal sample from the original 1*7 to n*7. The seven features include five original features and two features calculated by difference. The five original features are arrival time, carrier frequency, pulse width, pulse amplitude, and direction of arrival. The two features are PRI and second-order PRI.

[0021] Step 2: Construct the AAVGGPro model for aliased radar signal recognition based on PDW clustering. The AAVGGPro model includes a VGG model, an adaptive attention mechanism layer, and a newly added high-dimensional feature fully connected layer. The adaptive attention mechanism layer consists of an adaptive channel attention submodule and an adaptive spatial attention submodule connected in series. The AAVGG model is obtained by adding fully connected layers to both the adaptive channel attention submodule and the adaptive spatial attention submodule. The AAVGGPro model is obtained by adjusting the number of channels, convolutional kernel size, and adding a high-dimensional feature fully connected layer within the AAVGG model, thus enhancing the AAVGGPro model's ability to extract nonlinear features from radar signals. The VGG model includes convolutional layers and fully connected layers. The high-dimensional feature fully connected layer is used to extract high-dimensional features for radar signal recognition. Each convolutional layer consists of a one-dimensional convolutional layer, an activation layer, and a pooling layer. The one-dimensional convolutional layer extracts information between radar signal positions, the activation layer enhances the nonlinear representation capability of radar signal features, and the pooling layer reduces the size of radar signal features and the computational load. The adaptive channel attention submodule captures the dependencies between channels by focusing on different channels of the seven features of the radar signal input. It automatically learns and determines the importance of the feature channels of the radar signal, thus assigning different weight coefficients to each feature channel. The adaptive spatial attention submodule focuses on key regions in the feature space dimension of the radar signal. By applying average pooling and max pooling operations along the channel axis, it generates effective radar signal feature descriptors, thereby determining the location of radar signal features containing key information, extracting the location information of radar signal features, enhancing the robustness of the AAVGGPro model, and preventing overfitting. The pooling operation here compares the values ​​between feature channels of different radar signals, rather than the values ​​between different regions of the same channel. Three fully connected layers are used for recognition and prediction. The middle fully connected layer is used for loss function optimization in step three. The output of the middle fully connected layer can obtain the high-dimensional features of the radar signal samples. The last fully connected layer is used for radar signal classification prediction, and the radar signal is automatically identified based on the predicted category.

[0022] A deep convolutional neural network (DCNN) model for radar individual identification based on PDW clustering is constructed. The DCNN model for radar individual identification is the AAVGGPro model, which extracts deep features of radar signals and achieves identification through multiple convolutional operations. The AAVGGPro model consists of five convolutional layers, each including a one-dimensional convolutional layer, a ReLU (Rectified Linear Unit) activation layer, and a max-pooling layer. After the input features of the radar signal pass through the one-dimensional convolutional layer, the padding is adjusted to maintain the feature shape, changing only the number of channels to ensure the effectiveness of the information contained in the radar signal features. After the radar signal features pass through the max-pooling layer, the feature shape size is halved while the number of channels remains unchanged, ensuring that the feature shape is reduced and computational complexity is lowered without sacrificing recognition accuracy, thus improving the efficiency of radar signal identification. The output channels of the convolutional layers contain different radar signal feature maps extracted using different convolutional kernels. A nonlinear activation function is added to the radar signal feature map to learn nonlinear feature combinations. The pooling layer achieves translation invariance and dimensionality reduction of the radar signal features, extracting the invariant main features. The repeated convolution-activation-pooling hierarchical structure can extract high-dimensional abstract features of radar signals that continuously enhance from local to global levels, thus achieving radar signal feature extraction and recognition tasks. The formula for the ReLU activation function is...

[0023] ReLU(x)max(0,x) (5)

[0024] Each convolutional layer is followed by an adaptive attention mechanism layer. The adaptive attention mechanism layer includes an adaptive channel attention submodule and an adaptive spatial attention submodule. The adaptive channel attention submodule selects the feature channel importance of the radar signal according to equation (6), and the adaptive spatial attention submodule selects the feature spatial importance of the radar signal according to equation (7).

[0025]

[0026]

[0027] in, Let M represent element-wise multiplication, X represent the input features of the adaptive attention mechanism layer, X1 represent the features after passing through the adaptive channel attention submodule, X2 represent the features after passing through the adaptive spatial attention submodule, and M represent the features after passing through the adaptive spatial attention submodule. c (x) represents the feature map obtained after the adaptive channel attention submodule, M s (X1) represents the feature map obtained after the adaptive spatial attention submodule.

[0028] In the adaptive channel attention submodule, each channel of the radar signal feature map is equivalent to a feature detector. The adaptive channel attention submodule focuses on meaningful features in the channel dimension of the input feature map of the radar signal. Max pooling and average pooling are used to effectively extract the channel attention feature map of the radar signal. The extracted features are fed into a multilayer perceptron (MLP). The output features of the MLP are element-wise summed and fed into a batch normalization (BN) layer to fuse the radar signal features through channels, improving the AAVGGPro model's ability to remember features in the channel dimension of the radar signal. The output features of the normalization layer are activated by the 2*Sigmoid function, and the adaptive channel attention feature map M is generated according to equations (8) and (9). c (x).

[0029] M c (X)=2Sigmoid(BN(MLP(AvgPool(X))+MLP(MaxPool(X)))) (8)

[0030] Sigmoid(x) = (1 + e -x ) -1 (9)

[0031] In the adaptive spatial attention submodule, unlike the adaptive channel attention submodule, the adaptive spatial attention submodule focuses on the spatial location information in the radar signal features. By calculating the maximum and average values ​​of the channel dimensions in the radar signal features, two different feature maps are obtained. These two different feature maps are concatenated and then fed into a convolutional layer. The output features of the convolutional layer are sequentially input into a normalization layer and a 2*Sigmoid activation function to generate feature map M according to equation (10). s (X1).

[0032] M s (X1)=2Sigmoid(BN(f w*w ([AvgPool(X1);MaxPool(X1)]))) (10)

[0033] Among them, f w*w This indicates the convolution calculation with a kernel of w*w.

[0034] After five convolutional layers and five adaptive attention layers, the radar signal feature map is flattened into a one-dimensional vector. Then, three fully connected layers are used to generate and predict the high-dimensional features of the radar signal. The SoftMax activation function is employed for radar signal recognition and prediction. The first fully connected layer is used for rapid dimensionality reduction after flattening the radar signal feature vector into a one-dimensional vector, reducing the complexity of the AAVGGPro model, mitigating overfitting, and reducing the computational cost of the AAVGGPro model, thus accelerating radar signal recognition. The second fully connected layer generates high-dimensional features from the radar signal input samples, which are used to optimize the loss function in step three. The third fully connected layer is used for final recognition and prediction. The output vector of the last fully connected layer is z, and the length of vector z is the number of radar recognition categories. The SoftMax activation function is then used to calculate the probability value, expressed as:

[0035]

[0036] Where 0≤i≤L-1, the output y i It is the probability that the current radar input signal is of class i. The sum of the probabilities of all radar identification classes is 1. The class corresponding to the highest identification probability is taken as the class predicted by the AAVGGPro model. The radar identification class is uniquely determined based on the predicted class.

[0037] Preferably, the AAVGGPro model includes five convolutional layers, five adaptive attention mechanism layers, and three fully connected layers.

[0038] Step 3: Based on the radar identification dataset generated in Step 1, train the AAVGGPro deep convolutional neural network model for aliased radar individual identification constructed in Step 2, which is based on PDW clustering. In Step 1, the shape of the radar signal identification data features was transformed from 1*5 to n*7 as the input of the AAVGGPro model. The cross-entropy loss function was used to measure the difference between the predicted and true labels of radar signal identification. The mean of the Euclidean distance between the radar signal features and the corresponding radar signal identification class centers was calculated to measure the intra-class loss. The sum of the Euclidean distances between the radar signal identification class centers was calculated and then normalized to measure the inter-class loss. The final loss value is obtained by a weighted combination of the classification loss, intra-class loss, and inter-class loss, where the weight parameters are controlled by parameters α and β. The purpose of this combined loss function is to simultaneously optimize the radar signal identification accuracy and the clustering effect of radar signals in the feature space, enhance the robustness of the AAVGGPro model to radar signal identification tasks, and make it more adaptable to complex and ever-changing real-world environments.

[0039] The formulas for calculating the classification loss function, intra-class loss function, inter-class loss function, and total loss function are as follows:

[0040] Classification loss =CrossEntropyLoss(outputs,labels) (12)

[0041]

[0042]

[0043]

[0044] Where N and K are the number of radar signal recognition samples, ||·||2 represents the Euclidean distance, α and β are weight parameters, outputs is the final recognition prediction result of the AAVGGPro model, labels are the true recognition labels of the radar signal recognition samples, features[i] is the high-dimensional feature vector output by the second fully connected layer mentioned in step two, and centers[labels[i]] is the category center of the i-th radar signal recognition sample.

[0045] The Adam optimization algorithm is used to adaptively adjust the learning rate, and the backpropagation method is used to update the parameters of the AAVGGPro model to minimize the total loss function. The batch size and initial learning rate are adjusted according to the size and sample length of the radar signal recognition dataset after preprocessing in step one. The AAVGGPro model is trained until convergence, and the AAVGGPro model parameters are saved. The accuracy of the AAVGGPro model for radar signal recognition is tested using a radar recognition test sample set. Based on the test results, the AAVGGPro model that meets the preset convergence requirements is selected, thus obtaining the trained AAVGGPro model. Different mixed and overlapping radar signal recognition sample data are input into the saved trained AAVGGPro model, and the predicted label for each radar signal is output. The output predicted label result is used as the category corresponding to the radar signal to obtain the radar signal recognition result. Based on the radar signal recognition result, the radar recognition accuracy in different mixed scenarios is calculated. Therefore, the AAVGGPro model based on the PDW clustering algorithm achieves automatic radar signal recognition.

[0046] Beneficial effects:

[0047] 1. This invention discloses a method for identifying aliased radar individuals based on PDW clustering. It achieves similarity aggregation of radar signals through a k-means clustering algorithm based on PDW, completes the redistribution of intra-cluster labels and deep extraction of PRI and second-order PRI features of radar signals, increases the feature dimension of radar signal identification sample data, and combines multiple consecutive radar identification samples into a new sample, realizing the transformation of the radar signal feature space from one-dimensional (1D) to two-dimensional (2D). This increases the temporal continuity of radar signal feature data while ensuring that the original information is not lost, thereby increasing the extraction of useful information for radar signal identification to a certain extent, reducing the risk of overfitting of the AAVGGPro model, and improving the recognition accuracy of radar signals in complex scenarios.

[0048] 2. This invention discloses a method for identifying individual radar signals based on PDW clustering. The method utilizes the AAVGGPro model to identify radar signals. The AAVGGPro model includes a VGG model, an adaptive attention mechanism layer, and a newly added high-dimensional feature fully connected layer. The adaptive attention mechanism layer consists of an adaptive channel attention submodule and an adaptive spatial attention submodule connected in series. The AAVGGPro model uses VGG as its basic framework and introduces the adaptive attention mechanism layer to better utilize AAVGGPro's deep feature representation capabilities for seven features of radar signals: arrival time, carrier frequency, pulse width, pulse amplitude, direction of arrival, PRI, and second-order PRI. The adaptive channel attention submodule and the adaptive spatial attention submodule help improve the feature extraction capability of the AAVGGPro model, enabling it to adapt to complex radar signal identification environments. The adaptive attention mechanism layer can effectively enhance the feature representation capability of radar signal recognition tasks. It adaptively adjusts the weights of the AAVGGPro model to better capture key features at different locations of radar signals; improves the generalization ability of the AAVGGPro model by reducing its sensitivity to noise and non-recognition task-related information in the radar recognition environment; reduces the number of AAVGGPro model parameters by dynamically adjusting channel and spatial weights to reduce unnecessary redundant information in radar signal features, thereby reducing the computational burden of the AAVGGPro model; and enhances the AAVGGPro model's perception of radar signal feature scale information by using convolutional kernels of different sizes.

[0049] 3. The present invention discloses a method for identifying aliased radar individuals based on PDW clustering. The method adopts a combination loss function of classification loss, inter-class loss and intra-class loss, which can comprehensively consider multiple aspects of radar signal identification information. It covers the accuracy of radar signal identification task (classification loss), the internal similarity of radar signal features of the same class (intra-class loss) and the differences between radar signal features of different classes (inter-class loss). This helps to more comprehensively measure the performance of the AAVGGPro model. The intra-class loss function, by considering the feature similarity of samples with the same radar identification results, prompts the AAVGGPro model to learn more compact and discriminative feature representations, improving the aggregation effect of the AAVGGPro model on samples within the same radar identification category, making radar signal features more discriminative. The inter-class loss, by considering the feature differences between samples with different radar identification results, drives the AAVGGPro model to learn better radar identification category classification, enhancing the AAVGGPro model's discriminative ability between different radar identification categories and improving the accuracy of radar identification tasks. The loss function combining intra-class and inter-class losses enables the AAVGGPro model to learn the same identification signal with similar features, while ensuring that the features of samples from different radar identification categories have significant differences. By employing a loss function combining classification loss, intra-class loss, and inter-class loss, the performance of the AAVGGPro model is improved, making it more generalizable and adaptable, while eliminating the need for manual extraction of radar identification features, thus improving the efficiency of radar signal identification tasks.

[0050] 4. This invention discloses a radar individual identification method based on PDW clustering. First, PDW clustering is used to preprocess radar identification data, and then the AAVGGPro model is used to perform radar signal identification. Calculating PRI and second-order PRI using clustered groups is more efficient and accurate than traditional trial-and-error, histogram, and planar transformation methods. The AAVGGPro model has stronger identification capabilities, higher robustness, and better generalization than traditional algorithms. Furthermore, the radar signal depth identification method based on PDW clustering can construct radar identification datasets and adjust the AAVGGPro model according to the characteristics of radar aliasing signal identification in different scenarios. For radar aliasing scenarios of varying complexity, the converged AAVGGPro model based on PDW clustering can provide identification prediction results. Compared to traditional radar signal identification methods based on inter-pulse modulation features and intra-pulse modulation features, the radar signal depth identification method based on PDW clustering has better generalization performance, higher accuracy, and faster speed, and can handle various complex parameter settings and scenarios. Attached Figure Description

[0051] Figure 1This is a schematic diagram of the process of an aliased radar individual identification method based on PDW clustering as described in this invention;

[0052] Figure 2 This is a flowchart of the feature processing based on the PDW clustering algorithm described in this invention;

[0053] Figure 3 (a) is an architecture diagram of the deep convolutional neural network model of the aliased radar individual identification method based on PDW clustering described in this invention; (b) is a layer structure diagram of the adaptive attention mechanism; (c) is a structure diagram of the adaptive channel attention submodule; and (d) is a structure diagram of the adaptive spatial attention submodule.

[0054] Figure 4 This is a comparison of the performance curves of the deep convolutional neural network model and the traditional neural network recognition method of the aliased radar individual identification method based on PDW clustering described in this invention under different aliasing scenarios. Detailed Implementation

[0055] To better illustrate the purpose and advantages of the present invention, the invention will be further described below in conjunction with the accompanying drawings and examples.

[0056] Example 1:

[0057] like Figure 1As shown in this example, a method for identifying individual radar signals based on PDW clustering is disclosed. Building upon the existing radar signal identification system, three inter-pulse modulation features—carrier frequency, pulse width, and direction of arrival—are selected. The radar signals are then clustered using the k-means clustering algorithm. Within each cluster, the arrival time of all radar signals is differencing to obtain the PRI feature. The PRI is then differencing again to obtain the second-order PRI feature. Finally, the five original features and the two calculated features within each cluster are normalized to satisfy a distribution with a mean of 0 and a standard deviation of 1, resulting in a preprocessed radar signal identification dataset. The five original features refer to arrival time, carrier frequency, pulse width, pulse amplitude, and direction of arrival; the two features refer to PRI and second-order PRI. Then, a deep convolutional neural network (AAVGGPro) model for radar signal recognition is constructed. Its basic network skeleton is VGG (Visual Geometry Group). The AAVGG model is obtained by adding fully connected layers to the adaptive channel attention submodule and the adaptive spatial attention submodule. AAVGGPro is obtained by adjusting the number of channels and the convolutional kernel size in the AAVGG model and adding fully connected layers, thus enhancing the AAVGGPro model's ability to extract nonlinear features from radar signals. Specifically, the AAVGGPro model includes convolutional layers, an adaptive attention mechanism layer, and fully connected layers. The adaptive attention mechanism layer is composed of the adaptive channel attention submodule and the adaptive spatial attention submodule connected in series. The improved loss function is used to train the AAVGGPro deep convolutional neural network model to extract high-dimensional features of radar signals. The improved loss function consists of a cross-entropy loss function, an inter-class loss function, and an intra-class loss function. Based on the converged AAVGGPro model, the input radar signal is classified, achieving automatic radar signal recognition. The process of this embodiment is as follows: Figure 1 As shown, the experimental verification in this embodiment was carried out on a server. The specific configuration of the server is as follows: AMD EPYC 7513 32-Core Processor, CPU 2.60GHz, GPU NVIDIA A800 80G*2, operating system Ubuntu 22.04.2LTS.

[0058] like Figure 1 As shown in this embodiment, the aliased radar individual depth identification method based on PDW clustering includes the following steps:

[0059] Step 1: Preprocess the dataset. The existing training dataset only includes clean, interference-free inter-pulse feature data from 12 radars. Based on the parameters of the AAVGGPro model constructed in Step 2, 300 different mixed datasets for various scenarios can be obtained by temporal aliasing. For each aliased dataset, using the three inter-pulse modulation features of wave frequency, pulse width, and direction of arrival from the dataset, the k-means algorithm is used to divide the dataset into 24 clusters.

[0060] The K-means algorithm divides radar signals into 24 clusters based on the Euclidean distance between three inter-pulse modulation features: carrier frequency, pulse width, and direction of arrival. This ensures that radar signal features within a cluster are as close as possible to each other, while maximizing the distance between radar signals in different clusters. The specific implementation of the K-means algorithm for dividing radar signals into 24 clusters based on the Euclidean distance is as follows: Step 1: Randomly select 24 radar signal samples as cluster centers for radar signal identification; Step 2: Calculate the distance from each other radar signal sample point to the center vector of each radar signal identification cluster and assign it to the nearest cluster; Step 3: Update the center vector of each radar signal identification cluster; Step 4: Check if the new center vector of the radar signal identification cluster has changed. If it has changed, return to Step 2; if the center vector of the radar signal identification cluster no longer changes, stop and output the clustering results. A maximum iteration parameter `max_iters = 200` is set to prevent the iteration from entering an infinite loop. If the number of iterations exceeds `max_iters = 200`, the iteration automatically stops and the results are output. In the radar signal recognition training set, the true labels of radar signals within the same cluster are changed to the mode of the true labels in that cluster. In the validation and test sets, the AAVGGPro model predicts the category of the radar signals in that cluster as the mode of the prediction results. The Euclidean distance formula is expressed as follows:

[0061]

[0062] Where X = (x1, x2, ..., x...) n Y = (y1, y2, ..., y) n ), where n is the length of vector X and vector Y.

[0063] like Figure 2 As shown, for radar signal identification samples in the same cluster, the arrival time (TOA) is first differentially divided to obtain the pulse repetition interval (PRI), and then the PRI is differentially divided to obtain the second-order PRI. The calculation formula is as follows:

[0064] PRI i =|TOA i+1 -TOA i| (2)

[0065] 2nd order PRI i =|PRI i+1 -PRI i | (3)

[0066] After processing as described in equations (2) and (3), the feature set of each radar signal identification sample is features = {PRI, 2nd order PRI, TOA, RF, PW, PA, DOA}. The features of radar signal identification samples within the same cluster are normalized to ensure their feature distribution has a mean of 0 and a standard deviation of 1. The calculation formula is as follows:

[0067]

[0068] Here, feature is an element belonging to the feature set features above, μ is the mean of the feature, and σ is the standard deviation of the feature.

[0069] After preprocessing the radar signal identification data of the same cluster, seven normalized features are obtained. Then, each radar signal identification data point is treated as a single data sample, and 256 consecutive data points are treated as a single data sample. The feature shape of the radar signal sample changes from the original 1*7 to 256*7. The seven features include five original features and two features calculated using difference. The five original features are arrival time, carrier frequency, pulse width, pulse amplitude, and direction of arrival. The two features are PRI and second-order PRI. The changes in sample data before and after using PDW-based clustering are shown in the table below:

[0070]

[0071] The dataset includes 300 different self-aliasing radar identification data, totaling 34 million 7*256 samples. The training and validation sets are divided in a 7:3 ratio. The test sets consist of three real-world radar aliasing datasets: the first set includes aliasing data from four radars (numbers 1, 3, 6, and 8), totaling 999,038 data points; the second set includes aliasing data from eight radars (numbers 1, 2, 3, 5, 6, 7, 8, and 9), totaling 1,762,620 data points; and the third set includes aliasing data from twelve radars, totaling 2,579,053 data points.

[0072] Step 2: Construct an AAVGGPro model for aliased radar individual identification based on PDW clustering. The AAVGGPro model in this embodiment includes a VGG model, an adaptive attention mechanism layer, and a newly added high-dimensional feature fully connected layer. The adaptive attention mechanism layer consists of an adaptive channel attention submodule and an adaptive spatial attention submodule connected in series. The AAVGG model is obtained by adding fully connected layers to both the adaptive channel attention submodule and the adaptive spatial attention submodule. The AAVGGPro model is obtained by adjusting the number of channels, convolutional kernel size, and adding a high-dimensional feature fully connected layer in the AAVGG model, thus enhancing the AAVGGPro model's ability to extract nonlinear features from radar identification signals. The VGG model includes convolutional layers and fully connected layers. The high-dimensional feature fully connected layer is used to extract high-dimensional features for radar signal identification. Each convolutional layer consists of a one-dimensional convolutional layer, an activation layer, and a pooling layer. The one-dimensional convolutional layer is used to extract information between radar signal positions, the activation layer is used to enhance the nonlinear representation capability of radar signal features, and the pooling layer is used to reduce the size of radar signal features and reduce computational load. The adaptive channel attention submodule captures the dependencies between channels by focusing on different channels of the seven features of the radar signal input. It automatically learns and determines the importance of the feature channels of the radar signal, thus assigning different weight coefficients to each feature channel. The adaptive spatial attention submodule focuses on key regions in the feature space dimension of the radar signal. By applying average pooling and max pooling operations along the channel axis, it generates effective radar signal feature descriptors, thereby determining the location of radar signal features containing key information, extracting the location information of radar signal features, enhancing the robustness of the AAVGGPro model, and preventing overfitting. The pooling operation here compares the values ​​between feature channels of different radar signals, rather than the values ​​between different regions of the same channel. Three fully connected layers are used for recognition and prediction. The middle fully connected layer is used for loss function optimization in step three. The output of the middle fully connected layer can obtain the 128-dimensional high-dimensional features of the radar signal sample. The last fully connected layer performs 12-class classification prediction of the radar signal, and the radar signal is automatically identified based on the predicted category.

[0073] like Figure 3As shown in (a), a deep convolutional neural network model for aliased radar individual recognition based on PDW clustering is constructed. The radar recognition deep convolutional neural network model in this embodiment is the AAVGGPro model. The AAVGGPro model extracts deep features of radar signals and achieves recognition through multiple convolutional operations. In this embodiment, the AAVGGPro model includes five convolutional layers, each consisting of a one-dimensional convolutional layer, a ReLU (Rectified Linear Unit) activation layer, and a max-pooling layer. After the input features of the radar signal pass through the one-dimensional convolutional layer, the padding is adjusted to keep the feature shape unchanged, only changing the number of channels to ensure the effectiveness of the information contained in the radar signal features. After the radar signal features pass through the max-pooling layer, the feature shape size is halved while the number of channels remains unchanged, ensuring that the feature shape is reduced and the computational complexity is lowered without sacrificing recognition accuracy, thus improving the efficiency of radar signal recognition. The output channels of the convolutional layers contain different radar signal feature maps extracted using different convolutional kernels. A nonlinear activation function is added to the radar signal feature map to learn nonlinear feature combinations. The pooling layer achieves translation invariance and dimensionality reduction of the radar signal features, extracting the invariant main features. The repeated convolution-activation-pooling hierarchical structure can extract high-dimensional abstract features of radar signals that continuously enhance from local to global levels, thus achieving radar signal feature extraction and recognition tasks. The formula for the ReLU activation function is...

[0074] ReLU(x) = max(0,x) (5)

[0075] like Figure 3 As shown in (b), each convolutional layer is followed by an adaptive attention mechanism layer. The adaptive attention mechanism layer includes an adaptive channel attention submodule and an adaptive spatial attention submodule. The adaptive channel attention submodule selects the importance of the feature channels of the radar signal according to equation (6), and the adaptive spatial attention submodule selects the importance of the feature spaces of the radar signal according to equation (7).

[0076]

[0077]

[0078] in, This represents element-wise multiplication, where X represents the input feature of the adaptive attention mechanism layer, X1 represents the feature after passing through the adaptive channel attention mechanism module, X2 represents the feature after passing through the adaptive spatial attention mechanism module, and M represents the feature after passing through the adaptive spatial attention mechanism module. c (x) represents the feature map obtained after the adaptive channel attention submodule, M s (X1) represents the feature map obtained after the adaptive spatial attention submodule.

[0079] like Figure 3 As shown in (c), in the adaptive channel attention submodule, each channel of the radar signal feature map is equivalent to a feature detector. The adaptive channel attention submodule focuses on meaningful features in the channel dimension of the input feature map of the radar signal. Max pooling and average pooling are used to effectively extract the channel attention feature map of the radar signal. The extracted features are fed into a multilayer perceptron (MLP). The output features of the MLP are element-wise summed and fed into a batch normalization (BN) layer to fuse the radar signal features through channels, thereby improving the AAVGGPro model's ability to remember features in the channel dimension of the radar signal. The output features of the normalization layer are activated by the 2*Sigmoid function, and the adaptive channel attention feature map M is generated according to equations (8) and (9). c (x).

[0080] M c (X)=2Sigmoid(BN(MLP(AvgPool(X))+MLP(MaxPool(X)))) (8)

[0081] Sigmoid(x) = (1 + e -x ) -1 (9)

[0082] like Figure 3 As shown in (d), unlike the adaptive channel attention submodule, the adaptive spatial attention submodule focuses on the spatial location information in the radar signal features. Two different feature maps are obtained by calculating the maximum and average values ​​of the channel dimensions in the radar signal features. These two different feature maps are then concatenated and fed into a convolutional layer. The output features of the convolutional layer are sequentially input into a normalization layer and a 2*Sigmoid activation function to generate feature map M according to equation (10). s (X1).

[0083] M s (X1)=2Sigmoid(BN(f 7*7 ([AvgPool(X1);MaxPool(X1)]))) (10)

[0084] Among them, f 7*7 This indicates a convolution calculation with a 7x7 kernel.

[0085] After five convolutional layers and five adaptive attention layers, the radar signal feature map is flattened into a one-dimensional vector of length 512. Then, three fully connected layers are used to generate and predict the high-dimensional features of the radar signal. The SoftMax activation function is used for radar signal identification and prediction. The first fully connected layer is used for rapid dimensionality reduction after flattening the radar signal feature vector into a one-dimensional vector, reducing the complexity of the AAVGGPro model, mitigating overfitting, and reducing the computational cost of the AAVGGPro model. The second fully connected layer generates the high-dimensional feature vector of the radar signal input sample, which is used to optimize the loss function in step three. The third fully connected layer is used for the final classification prediction. The output vector z of the last fully connected layer has a length of 12, which represents the number of radar identification categories. The SoftMax activation function is then used to calculate the probability value, and its function expression is:

[0086]

[0087] Where 0≤i≤11, the output is y. i It is the probability that the current input is of class i. The sum of the probabilities of all radar recognition classes is 1. The class corresponding to the highest recognition probability is taken as the final predicted class of the AAVGGPro model. The radar recognition class is uniquely determined based on the predicted class.

[0088] Step 3: Based on the radar identification dataset generated in Step 1, train the AAVGGPro deep convolutional neural network model for aliased radar individual identification constructed in Step 2, which is based on PDW clustering. In Step 1, the shape of the radar signal identification data features was transformed from 1*5 to 256*7 as input to the AAVGGPro model. The cross-entropy loss function was used to measure the difference between the predicted and true labels of radar signal identification. The mean of the Euclidean distance between the radar signal features and the corresponding radar signal identification class centers was calculated to measure the intra-class loss. The sum of the Euclidean distances between the radar signal identification class centers was calculated and then normalized to measure the inter-class loss. The final loss value is obtained by a weighted combination of the classification loss, intra-class loss, and inter-class loss, where the weight parameters are controlled by parameters α and β. The purpose of this combined loss function is to simultaneously optimize the radar signal identification accuracy and the clustering effect of radar signals in the feature space, enhance the robustness of the AAVGGPro model to radar signal identification tasks, and make it more adaptable to complex and ever-changing real-world environments.

[0089] The formulas for calculating the classification loss function, intra-class loss function, inter-class loss function, and total loss function are as follows:

[0090] Classification loss =CrossEntropyLoss(outputs,labels) (12)

[0091]

[0092]

[0093]

[0094] Where N and K are the number of radar signal recognition samples, both 23.8 million; ||·||2 represents the Euclidean distance, α and β are weight parameters, both 0.001; outputs is the final recognition prediction result of the AAVGGPro model, labels are the true labels of the radar signal recognition samples, features[i] is the high-dimensional feature vector output by the second fully connected layer mentioned in step two, and centers[labels[i]] is the category center of the i-th radar signal recognition sample.

[0095] The Adam optimization algorithm is used to adaptively adjust the learning rate, and the backpropagation method is used to update the parameters of the AAVGGPro model to minimize the total loss function. Based on the size and sample length of the preprocessed dataset in this embodiment, the batch size is set to 256, and the initial learning rate is 0.0001. The AAVGGPro model is trained until convergence, and the AAVGGPro model parameters are saved. The accuracy of the AAVGGPro model in radar signal recognition is tested using a radar recognition test sample set. Based on the test results, the AAVGGPro model that meets the preset convergence requirements is selected, thus obtaining the trained AAVGGPro model. Different mixed overlapping radar signal recognition sample data are input into the saved trained AAVGGPro model, and the predicted label of each radar signal is output. The output predicted label result is used as the category corresponding to the radar signal to obtain the radar signal recognition result. Based on the radar signal recognition result, the radar recognition accuracy in different mixed scenarios is calculated, as shown in the appendix. Figure 4 As shown.

[0096] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying individual radar instances based on PDW clustering, characterized in that: Includes the following steps: Step 1: Preprocess the dataset; The training dataset contains only clean and interference-free inter-pulse feature data from each radar. Based on the parameters of the AAVGGPro model constructed in Step 2, a mixed dataset is obtained by temporal aliasing. For each mixed dataset, the three inter-pulse modulation features of carrier frequency, pulse width, and direction of arrival in the dataset are used to divide the dataset into k clusters using the k-means algorithm. Step 2: Construct an AAVGGPro model for aliased radar signal recognition based on PDW clustering. The AAVGGPro model includes a VGG model, an adaptive attention mechanism layer, and a newly added high-dimensional feature fully connected layer. The adaptive attention mechanism layer consists of an adaptive channel attention submodule and an adaptive spatial attention submodule connected in series. The AAVGG model is obtained by adding fully connected layers to both the adaptive channel attention submodule and the adaptive spatial attention submodule. The AAVGGPro model is obtained by adjusting the number of channels, convolutional kernel size, and adding a high-dimensional feature fully connected layer in the AAVGG model, thereby enhancing the AAVGGPro model's ability to extract nonlinear features from radar signals. The VGG model includes convolutional layers and fully connected layers. The high-dimensional feature fully connected layer is used to extract high-dimensional features for radar signal recognition. Each convolutional layer consists of a one-dimensional convolutional layer, an activation layer, and a pooling layer. The one-dimensional convolutional layer is used to extract information between radar signal positions, the activation layer is used to enhance the nonlinear representation ability of radar signal features, and the pooling layer is used to reduce the size of radar signal features and reduce computational load. The adaptive channel attention submodule captures the inter-channel dependencies by focusing on different channels of the seven features of the radar identification signal input, automatically learns and determines the importance of the feature channels of the radar signal, and assigns different weight coefficients to each feature channel of the radar signal. The adaptive spatial attention submodule focuses on key regions in the feature space dimension of radar signals. By applying average pooling and max pooling operations along the channel axis, it generates effective radar signal feature descriptors, thereby determining the location of radar signal features containing key information, extracting the location information of radar signal features, enhancing the robustness of the AAVGGPro model, and preventing overfitting. The pooling operation here compares the values ​​between feature channels of different radar signals, rather than the values ​​between different regions of the same channel. Three fully connected layers are used for recognition and prediction. The middle fully connected layer is used for loss function optimization in step three. The output of the middle fully connected layer can obtain the high-dimensional features of the radar signal samples. The last fully connected layer is used for radar signal classification prediction, and the radar signal is automatically identified based on the predicted category. Step 3: Based on the radar identification dataset generated in Step 1, train the AAVGGPro deep convolutional neural network model for aliased radar individual identification constructed in Step 2, which is based on PDW clustering. In Step 1, the shape of the radar signal identification data features was converted from 1*5 to n*7 as the input of the AAVGGPro model. The cross-entropy loss function was used to measure the difference between the predicted label and the true label of radar signal identification. The mean of the Euclidean distance between the radar signal features and the corresponding radar signal identification class center was calculated using Euclidean distance to measure the intra-class loss. The sum of the Euclidean distances between the radar signal identification class centers was calculated using Euclidean distance and then normalized to measure the inter-class loss. The final loss value is obtained by a weighted combination of classification loss, intra-class loss and inter-class loss, where the weight parameters are controlled by parameters α and β. The purpose of this combined loss function is to simultaneously optimize the radar signal identification accuracy and the clustering effect of radar signals in the feature space, thereby enhancing the robustness of the AAVGGPro model to the radar signal identification task. The Adam optimization algorithm is used to adaptively adjust the learning rate, and the backpropagation method is used to update the parameters of the AAVGGPro model to minimize the total loss function. The batch size and initial learning rate are adjusted according to the size and sample length of the radar signal recognition dataset after preprocessing in step one. The AAVGGPro model is trained until convergence, and the AAVGGPro model parameters are saved. The accuracy of the AAVGGPro model in radar signal recognition is tested using a radar recognition test sample set. Based on the test results, the AAVGGPro model that meets the preset convergence requirements is selected, thus obtaining the trained AAVGGPro model. Different mixed and overlapping radar signal recognition sample data are input into the saved trained AAVGGPro model, and the predicted label for each radar signal is output. The output predicted label is used as the category corresponding to the radar signal to obtain the radar signal recognition result. Based on the radar signal recognition result, the radar recognition accuracy in different mixed scenarios is calculated. Therefore, the AAVGGPro model based on the PDW clustering algorithm achieves automatic radar signal recognition.

2. The method for identifying individual aliased radars based on PDW clustering as described in claim 1, characterized in that: In step one, The K-means algorithm divides radar signals into k clusters based on the Euclidean distance between three inter-pulse modulation features: carrier frequency, pulse width, and direction of arrival. This ensures that radar signal features within a cluster are as close as possible to each other, while maximizing the distance between radar signals in different clusters. The specific implementation of the K-means algorithm based on the Euclidean distance between radar signals is as follows: Step 1, randomly select k radar signal samples as cluster centers for radar signal identification; Step 2, calculate the distance from each other radar signal sample point to the center vector of each radar signal identification cluster, and assign it to the closest cluster; Step 3, further refine... The new center vectors of each radar signal identification class; Step 4, determine if the center vectors of the new radar signal identification classes have changed. If they have changed, go back to Step 2. If the center vectors of the radar signal identification classes no longer change, stop and output the clustering results; a maximum iteration parameter `max_iters` is set to prevent iteration from entering an infinite loop. If the number of iterations exceeds the `max_iters` parameter, the iteration will automatically stop and the results will be output; in the radar signal identification training set, the true labels of radar signals in the same cluster are changed to the mode of the true labels in this cluster; in the validation and test sets, the AAVGGPro model predicts the class of the radar signals in this cluster as the mode of the prediction results; the Euclidean distance formula is expressed as... Where X = (x1, x2, ..., x...) n Y = (y1, y2, ..., y) n ), where n is the length of vector X and vector Y; For radar signal identification samples within the same cluster, the arrival time (TOA) is first differentially divided to obtain the pulse repetition interval (PRI), and then the PRI is differentially divided to obtain the second-order PRI. The calculation formula is as follows: PRI i =|TOA i+1 -TOA i | (2) 2nd price i =|PRICE i+1 -PRICE i | (3) After processing as described in equations (2) and (3), the feature set of each radar signal identification sample is features = {PRI, 2nd order PRI, TOA, RF, PW, PA, DOA}. The features of radar signal identification samples in the same cluster are normalized so that their feature distribution satisfies a mean of 0 and a standard deviation of 1. The calculation formula is as follows: Where feature is an element belonging to the feature set features above, μ is the mean of the feature, and σ is the standard deviation of the feature. After preprocessing the radar signal identification data of the same cluster, seven normalized features are obtained. Then, the original radar signal identification data is transformed into a data sample of n consecutive data, and the feature shape of the radar signal sample is changed from the original 1*7 to n*7. The seven features include five original features and two features calculated by difference. The five original features are arrival time, carrier frequency, pulse width, pulse amplitude and direction of arrival. The two features are PRI and second-order PRI.

3. The method for identifying individual aliased radars based on PDW clustering as described in claim 2, characterized in that: A deep convolutional neural network model for aliased radar individual recognition based on PDW clustering is constructed. This model is an AAVGGPro model, which extracts deep features of radar signals and achieves recognition through multiple convolutional operations. The AAVGGPro model consists of five convolutional layers, each including a one-dimensional convolutional layer, a ReLU activation layer, and a max-pooling layer. After the input features of the radar signal pass through the one-dimensional convolutional layer, the padding is adjusted to maintain the feature shape, only changing the number of channels to ensure the effectiveness of the information contained in the radar signal features. The radar signal features are then subjected to max pooling. The convolutional layer halves the feature shape size while maintaining the same number of channels, ensuring that the feature shape is reduced and computational complexity is lowered without sacrificing recognition accuracy, thus improving the efficiency of radar signal recognition. The output channels of the convolutional layer contain different radar signal feature maps extracted using different convolutional kernels. A non-linear activation function is added to the radar signal feature map to learn non-linear feature combinations. The pooling layer achieves translation invariance and dimensionality reduction of radar signal features, extracting invariant principal features. The repeated convolution-activation-pooling hierarchical structure can extract increasingly higher-dimensional abstract features of radar signals from local to global levels, realizing the task of radar signal feature extraction and recognition. The formula for the ReLU activation function is... ReLU(x) = max(0, x) (5) Each convolutional layer is followed by an adaptive attention mechanism layer; The adaptive attention mechanism layer includes an adaptive channel attention submodule and an adaptive spatial attention submodule. The adaptive channel attention submodule selects the feature channel importance of the radar signal according to equation (6), and the adaptive spatial attention submodule selects the feature spatial importance of the radar signal according to equation (7). in, Let M represent element-wise multiplication, X represent the input features of the adaptive attention mechanism layer, X1 represent the features after passing through the adaptive channel attention submodule, X2 represent the features after passing through the adaptive spatial attention submodule, and M represent the features after passing through the adaptive spatial attention submodule. c (x) represents the feature map obtained after the adaptive channel attention submodule, M s (X1) represents the feature map obtained after the adaptive spatial attention submodule; In the adaptive channel attention submodule, each channel of the radar signal feature map is equivalent to a feature detector; the adaptive channel attention submodule focuses on meaningful features in the channel dimension of the input feature map of the radar signal; max pooling and average pooling are used to effectively extract the channel attention feature map of the radar signal; the extracted features are fed into a multilayer perceptron (MLP); the output features of the MLP are element-wise summed and fed into a batch normalization (BN) layer to perform channel fusion of the radar signal features, thereby improving the AAVGGPro model's ability to remember features in the channel dimension of the radar signal; the output features of the normalization layer are activated by the 2*Sigmoid function, and the adaptive channel attention feature map M is generated according to equations (8) and (9). c (x); M c (x)=2Sigmoid(BN(MLP(AvgPool(X))+MLP(MaxPool(X)))) (8) Sigmoid(x)=(1+e -x ) -1 (9) In the adaptive spatial attention submodule, unlike the adaptive channel attention submodule, the adaptive spatial attention submodule focuses on the spatial location information in the radar signal features. By calculating the maximum and average values ​​of the channel dimensions in the radar signal features, two different feature maps are obtained. These two different feature maps are concatenated and then fed into the convolutional layer. The output features of the convolutional layer are sequentially input into the normalization layer and the 2*Sigmoid activation function to generate feature map M according to equation (10). s (X1); M s (X1)=2Sigmoid(BN(f w*w ([AvgPool(X1);MaxPool(X1)]))) (10) Among them, f w*w This indicates the convolution calculation with a kernel of w*w; After five convolutional layers and five adaptive attention layers, the radar signal feature map is flattened into a one-dimensional vector. Then, three fully connected layers are used to generate and predict the high-dimensional features of the radar signal. The SoftMax activation function is employed for radar signal recognition and prediction. The first fully connected layer is used for rapid dimensionality reduction after flattening the radar signal feature vector into a one-dimensional vector, reducing the complexity of the AAVGGPro model, mitigating overfitting, and decreasing the computational cost of the AAVGGPro model, thus accelerating radar signal recognition. The second fully connected layer generates high-dimensional features from the radar signal input samples, which are used to optimize the loss function in step three. The third fully connected layer is used for the final recognition and prediction. The output vector of the last fully connected layer is z, and the length of vector z is the number of radar recognition categories. The SoftMax activation function is then used to calculate the probability value, expressed as: Where 0≤i≤L-1, the output y i It is the probability that the current radar input signal is of class i. The sum of the probabilities of all radar identification classes is 1. The class corresponding to the highest identification probability is taken as the class predicted by the AAVGGPro model. The radar identification class is uniquely determined based on the predicted class.

4. The method for identifying individual aliased radars based on PDW clustering as described in claim 3, characterized in that: The AAVGGPro model consists of five convolutional layers, five adaptive attention mechanism layers, and three fully connected layers.

5. A method for identifying individual aliased radars based on PDW clustering as described in claim 3 or 4, characterized in that: In step three, The formulas for calculating the classification loss function, intra-class loss function, inter-class loss function, and total loss function are as follows: classification loss =CrossEntropyLoss(outputs,labels) (12) Where N and K are the number of radar signal recognition samples, ||·||2 represents the Euclidean distance, α and β are weight parameters, outputs is the final recognition prediction result of the AAVGGPro model, labels are the true recognition labels of the radar signal recognition samples, features[i] is the high-dimensional feature vector output by the second fully connected layer mentioned in step two, and centers[labels[i]] is the category center of the i-th radar signal recognition sample.

Images