Radar radiation source signal sorting method based on adaptive integrated deep clustering
Through the adaptive integrated depth clustering method, the dual attention residual autoencoder and the adaptive integrated clustering function are used to solve the problem of radar radiation source signal sorting in complex electromagnetic environments, and efficient and accurate signal sorting is achieved.
Patent Information
- Application Number
- CN202510410910.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-02
- Publication Date
- 2025-07-25
AI Technical Summary
The existing radar radiation source signal sorting method faces the problems of signal surge, insufficient feature extraction, the need for prior information and poor sorting effect of unknown signals in complex electromagnetic environments.
Adaptive integrated depth clustering method is adopted to extract features through a dual attention residual autoencoder, optimize K-means clustering and Bayesian information criterion to optimize Gaussian hybrid model clustering, and fuse clustering results to achieve end-to-end signal sorting.
Without presetting key parameters, it can adaptively process unknown signals in high-density hybrid pulse flow, with robust noise immunity and strong adaptability, improving the accuracy and efficiency of signal sorting.
Smart Images
Figure CN120372282A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of radar, and relates to a method for sorting radar emitter signals based on adaptive integrated deep clustering. Background Art
[0002] Radar emitter signal sorting is a core technology in electronic reconnaissance and countermeasure, aiming to separate signals from different radar emitters in a complex electromagnetic environment, and obtain their parameter characteristics, etc., so as to support target recognition, threat assessment and tactical decision-making. During the process of radar emitter signal sorting, the received pulses are usually characterized by a Pulse Descriptive Word (PDW) containing 5-dimensional parameters, including: Direction Of Arrival (DOA), Time Of Arrival (TOA), Pulse Width (PW), Radio Frequency (RF), and Pulse Amplitude (PA).
[0003] At present, the electromagnetic environment presents characteristics such as multi-source signal overlap, crowded spectrum resources, and rapid dynamic changes. At present, radar emitter signal sorting faces great technical challenges. Radar signal sorting is a key link in the radar reconnaissance system, and its main function is to perform clustering processing on radar signals through pulse descriptive word parameters such as modulation type, direction of arrival, pulse width, and carrier frequency of radar signals.
[0004] Existing radar emitter signal sorting methods mainly perform parameter extraction on PDW parameters, such as carrier frequency, direction of arrival, pulse width, etc., and perform clustering processing on radar signals. Common clustering algorithms include partitioning clustering represented by k-means, hierarchical clustering represented by BIRCH (Balanced Iterative Reducing and Clustering Using Hierarchies), grid clustering represented by STING (Statistical Information Grid), density clustering represented by DBSCAN (Density-Based Spatial Clustering of Applications with Noise), and model-based clustering represented by Gaussian Mixture Model (GMM).
[0005] K-means clustering is only suitable for convex datasets, treating each feature dimension equally and unable to distinguish the importance of different feature dimensions. BIRCH clustering assumes that clusters are spherical and has poor clustering effect on high-dimensional feature data. If the distribution clusters of the dataset are not similar to hyperspheres, or in other words, not convex, the clustering effect will be poor. The quality of STING clustering depends on the granularity. If the granularity is relatively fine, the processing cost will increase significantly. However, if the granularity is too coarse, the quality of clustering analysis will be reduced. DBSCAN clustering requires setting the neighborhood radius and minimum number of points in advance, has high requirements for computing resources and performs poorly on datasets with uneven density. GMM requires the dataset to follow a Gaussian distribution. Especially in the current electromagnetic environment where the signal surge and the received signal data grow geometrically, due to the curse of dimensionality problem, these algorithms often fail. Moreover, these clustering methods require prior information, but in actual situations, prior information cannot be obtained. In the face of unknown signals, a pre-sorting method without prior information needs to be designed to meet the actual needs. Summary of the Invention
[0006] The present invention aims to solve the technical problems in the prior art, such as the surge of electromagnetic environment pulse streams, high requirements for existing software and hardware devices, less information contained in the features extracted by sorting algorithms, the need for prior information, poor sorting effect on unknown signals, and the mismatch between feature distribution and clustering algorithms. The present invention provides a method for sorting radar emitter signals based on adaptive integrated deep clustering, and the technical solution adopted is as follows:
[0007] A method for sorting radar emitter signals based on adaptive integrated deep clustering includes the following steps:
[0008] S1. Normalize the data to obtain normalized data excluding different parameter dimensions;
[0009] S2. Input the normalized data into a dual-attention residual autoencoder for feature extraction and output the extracted features;
[0010] S3. Input the extracted features into an adaptive integrated clustering function, optimize K-means clustering based on the silhouette coefficient, optimize Gaussian mixture model clustering based on the Bayesian information criterion, fuse the clustering results of the K-means clustering and Gaussian mixture model clustering through hierarchical clustering to obtain clustering labels, align the clustering labels with the true labels, evaluate the clustering effect, and complete the sorting of radar emitter signals.
[0011] In an embodiment of the present invention, the step S1 includes:
[0012] The normalization process is expressed as:
[0013]
[0014] In formula (1), p represents the parameter vector to be normalized.
[0015] In one embodiment of the present invention, the step S2 includes:
[0016] The dual attention residual autoencoder consists of an encoder part and a decoder part;
[0017] The normalized data is input into the encoder part, and the encoder part gradually compresses the input features into the latent space;
[0018] The decoder part reconstructs and gradually restores the feature dimension. At the same time, it jumps through the Linear function connection and ensures that the output is between 0 and 1 through the Sigmoid activation function.
[0019] In one embodiment of the present invention, the jump connection through the Linear function includes:
[0020] The output of the first encoder is adjusted in dimension through the Linear function and added to the output of the second decoder, and the output of the second encoder is adjusted in dimension through the Linear function and added to the output of the first decoder.
[0021] In one embodiment of the present invention, the encoder includes a first encoder, a second encoder, and a third encoder;
[0022] The first encoder includes a ResDSConv function, a dual attention mechanism, and a Linear function connected in sequence. The input end of the ResDSConv function of the first encoder is connected to the normalized data, and the output end of the Linear function of the first encoder is connected to the second encoder;
[0023] The second encoder includes a ResDSConv function and a dual attention mechanism. The input end of the ResDSConv function of the second encoder is connected to the output end of the first encoder. The output end of the ResDSConv function of the second encoder is connected to the input end of the dual attention mechanism of the second encoder, and the output end of the dual attention mechanism of the second encoder is connected to the input end of the third encoder;
[0024] The third encoder consists of a Linear function, and the output end of the third encoder is connected to the first decoder.
[0025] In one embodiment of the present invention, the decoder part includes a first decoder, a second decoder, and a third decoder;
[0026] The first decoder includes a Linear function and a ResDSConv function. The input end of the Linear function of the first decoder is connected to the output end of the third encoder. The output end of the Linear function of the first decoder is connected to the input end of the ResDSConv function of the first decoder. The output end of the ResDSConv function of the first decoder is connected to the input end of the second decoder;
[0027] The second decoder includes a Linear function, a dual attention mechanism, and a ResDSConv function connected in sequence. The input end of the Linear function of the second decoder is connected to the output end of the first decoder. The output end of the ResDSConv function of the second decoder is connected to the input end of the third decoder;
[0028] The third decoder includes a Sigmoid activation function and a Linear function. The input end of the Sigmoid activation function of the third decoder is connected to the output end of the second decoder. The output end of the Sigmoid activation function of the third decoder is connected to the input end of the Linear function of the third decoder. The output of the Linear function of the third decoder extracts features.
[0029] In one embodiment of the present invention, the dual attention mechanism includes a channel attention branch and a spatial attention branch;
[0030] In the channel attention branch, the input features are subjected to global average pooling, flattening, a fully connected layer, and ReLU activation, and then passed through a fully connected layer and a Sigmoid activation function to obtain channel attention weights;
[0031] In the spatial attention branch, the input features are activated through two 1x1 convolutional layers and ReLU, and finally passed through a 1x1 convolutional layer and Sigmoid activation to obtain spatial attention weights;
[0032] The channel attention weights and the spatial attention weights are multiplied by the input features respectively and then fused to form the final output features.
[0033] In one embodiment of the present invention, the optimization of K-means clustering based on the silhouette coefficient in step S3 includes:
[0034] The silhouette coefficient combines the within-cluster cohesion and the between-cluster separation. Set the maximum number of clusters max_clusters, traverse the number of clusters k ∈ [2, max_clusters], evaluate the clustering quality through the silhouette coefficient, and the value range is [-1, 1]. The higher the value, the better the clustering effect;
[0035] Select the clustering number with the highest silhouette coefficient to obtain the geometric cluster centers of the k-means clustering.
[0036] In one embodiment of the present invention, the optimizing the Gaussian mixture model clustering based on the Bayesian information criterion in step S3 includes:
[0037] Set the maximum number of clusters max_clusters, traverse the number of clusters k ∈ [2, max_clusters], and evaluate the clustering quality through the Bayesian information criterion, where the Bayesian information criterion is expressed as:
[0038] BIC = ln(n)k - 2ln(L) (2)
[0039] In formula (2), n is the sample size, i.e., the number of data points, k is the number of free parameters in the model, including the intercept term, and L is the maximum likelihood estimate value of the model on the data;
[0040] Select the model with the minimum Bayesian information criterion value. The smaller the Bayesian information criterion value, the better the model, and obtain the probability distribution center of the Gaussian mixture model clustering.
[0041] In one embodiment of the present invention, step S3 further includes:
[0042] By stacking the geometric cluster centers of the k-means clustering and the probability distribution centers of the Gaussian mixture model clustering, increase the diversity of candidate centers and cover different distribution characteristics of the data;
[0043] Perform hierarchical clustering fusion on the set of geometric cluster centers of the k-means clustering and the probability distribution centers of the Gaussian mixture model clustering. Merge similar centers according to a preset distance threshold, and divide the samples into the nearest clusters by calculating the Euclidean distance from the samples to each fused cluster center, forming an adaptive clustering result that takes into account the advantages of density distribution and geometric partitioning;
[0044] According to the fused centers, calculate the Euclidean distance from all samples to each center, assign the samples to the nearest center to form clustering labels, align the clustering labels with the true labels, evaluate the clustering effect, and complete the sorting of radar emitter signals.
[0045] The beneficial effects of the present invention:
[0046] The method for sorting radar emitter signals based on adaptive integrated deep clustering of the present invention uses a dual-attention residual autoencoder to learn the data distribution characteristics, and then uses an adaptive ensemble clustering function to complete the sorting of signals. Through the multi-layer non-linear transformation of a deep neural network, a joint optimization model of "feature learning - clustering and sorting" is constructed. It does not require pre-setting of key parameters (such as the number of clusters, neighborhood radius, etc.) and does not rely on artificially designed features. Facing unknown signals in a high-density mixed pulse stream, it can learn the signal data distribution, capture complex data distributions using multi-layer non-linear transformations, unify deep feature learning and clustering objectives in an end-to-end framework, break through the limitations of traditional staged processing, adapt to complex data distributions through the attention mechanism and clustering optimization, have robust noise resistance, and have strong adaptability. Description of the Drawings
[0047] Figure 1 is a flowchart of a method for sorting radar emitter signals based on adaptive integrated deep clustering provided by an embodiment of the present invention;
[0048] Figure 2 is a schematic structural diagram of a dual-attention residual autoencoder provided by an embodiment of the present invention;
[0049] Figure 3 is a schematic structural diagram of a DA network provided by an embodiment of the present invention. Detailed Embodiments
[0050] The present invention will be described in detail below with reference to the drawings and specific embodiments.
[0051] The present invention provides a method for sorting radar emitter signals based on adaptive integrated deep clustering. The method used is to use a dual-attention residual autoencoder (DARA) to learn the data distribution characteristics, and then use an adaptive ensemble clustering (AEC) function to complete the sorting of signals. This method is accurate and efficient.
[0052] Refer to the attached Figure 1 , the method for sorting radar emitter signals based on adaptive integrated deep clustering includes the steps:
[0053] S1. Normalize the data to obtain normalized data excluding different parameter dimensions;
[0054] S2. Input the normalized data into the dual-attention residual autoencoder for feature extraction, and output the extracted features with deep regularities;
[0055] S3. Input the extracted features into the adaptive integrated clustering function, optimize K-means clustering based on the silhouette coefficient, optimize Gaussian mixture model clustering based on the Bayesian information criterion, fuse the clustering results of K-means clustering and Gaussian mixture model clustering through hierarchical clustering, obtain cluster labels, align the cluster labels with the true labels, evaluate the clustering effect, and complete the radar radiation source signal sorting.
[0056] Compared with traditional methods, the present invention does not require pre-setting of key parameters (number of clusters, domain radius, etc.) and does not rely on manually designed features. When faced with unknown signals in high-density mixed pulse streams, the present invention can automatically learn highly discriminative features through neural networks, use multi-layer nonlinear transformations to capture complex data distributions, and unify deep feature learning with clustering goals in an end-to-end framework, breaking through the limitations of traditional staged processing; through attention mechanisms and clustering optimization, it can adapt to complex data distributions, has robust noise resistance, and has strong adaptability.
[0057] The dimensions of different parameters in the pulse descriptive word (PDW) are usually very different, so it is necessary to preprocess the data. Normalize the data to eliminate the influence of different parameter dimensions, avoid affecting the training of the network, and affect the convergence speed. Step S1 of the present invention includes:
[0058] The normalization process is expressed as:
[0059]
[0060] In formula (1), p represents the parameter vector that needs to be normalized.
[0061] The normalized data is sent to DARA for feature extraction to explore the deep laws of the data. Traditional feature extraction networks such as autoencoders perform well in processing local features, but have defects in capturing sequence data or complex data structures with long-distance dependencies, resulting in inaccurate and incomplete extracted data features; while variational autoencoders usually assume that data obeys a certain distribution, such as Gaussian distribution, and have higher requirements for data distribution.
[0062] The Dual Attention Residual Autoencoder (DARA) adds a residual layer and a dual attention mechanism (Dual Attention, DA network) on the basis of the traditional autoencoder. The residual structure can optimize the training process, improve the feature extraction ability, reduce the overfitting phenomenon during the training process, and improve the generalization ability of the model. Most importantly, the residual layer can combine features at different levels, and the fusion of this multi-scale information can further improve the performance of the network, enabling it to better capture the essential features of the data. The channel attention branch of the DA network adopts a 1D pooling + fully connected structure, uses a linear layer to achieve cross-channel interaction, and the spatial attention branch uses 1D convolution for channel dimensionality reduction and constructs spatial weights through pointwise convolution, which is more suitable for processing sequence data. Finally, through the broadcast mechanism, the parallel superposition of channel and spatial attention is realized, and the multi-dimensional feature capture improves the feature extraction and representation ability of the model.
[0063] In one embodiment of the present invention, step S2 includes:
[0064] The dual attention residual autoencoder consists of multiple encoder parts and decoder parts. Referring to the appendix Figure 2 , the present invention adopts three encoders and three decoders. The network contains 4 depthwise separable convolutions, 2 residual connection blocks and 3 dual attention mechanisms (DA network).
[0065] The normalized data is input to the encoder part, and the encoder part maps the input data. The encoder part of the present invention gradually compresses the input features into the latent space in three stages. The decoder part gradually restores the feature dimension through three-stage reconstruction. At the same time, through the Linear function skip connection and the Sigmoid activation function, it is ensured that the output is between 0 and 1. The specific process of the Linear function skip connection is as follows: the output of the first encoder is adjusted in dimension through the Linear function and added to the output of the second decoder, and the output of the second encoder is adjusted in dimension through the Linear function and added to the output of the first decoder.
[0066] The Sigmoid function has two functions. One is to introduce non-linear changes and enhance the expression ability of the input data. The input features are normalized data between (0, 1), ensuring that the output data is also between (0, 1). This concept comes from the autoencoder, which aims to encode the input data and then reconstruct the original input from the encoding result.
[0067] Preferably, the encoder includes a first encoder, a second encoder, and a third encoder. The first encoder includes a ResDSConv function, a dual attention mechanism, and a Linear function connected in sequence. The input end of the ResDSConv function of the first encoder is connected to the normalized data, and the output end of the Linear function of the first encoder is connected to the second encoder. The second encoder includes a ResDSConv function and a dual attention mechanism. The input end of the ResDSConv function of the second encoder is connected to the output end of the first encoder. The output end of the ResDSConv function of the second encoder is connected to the input end of the dual attention mechanism of the second encoder. The output end of the dual attention mechanism of the second encoder is connected to the input end of the third encoder. The third encoder is composed of a Linear function, and the output end of the third encoder is connected to the first decoder.
[0068] The decoder part includes a first decoder, a second decoder, and a third decoder. The first decoder includes a Linear function and a ResDSConv function. The input end of the Linear function of the first decoder is connected to the output end of the third encoder. The output end of the Linear function of the first decoder is connected to the input end of the ResDSConv function of the first decoder. The output end of the ResDSConv function of the first decoder is connected to the input end of the second decoder. The second decoder includes a Linear function, a dual attention mechanism, and a ResDSConv function connected in sequence. The input end of the Linear function of the second decoder is connected to the output end of the first decoder. The output end of the ResDSConv function of the second decoder is connected to the input end of the third decoder. The third decoder includes a Sigmoid activation function and a Linear function. The input end of the Sigmoid activation function of the third decoder is connected to the output end of the second decoder. The output end of the Sigmoid activation function of the third decoder is connected to the input end of the Linear function of the third decoder. The output of the Linear function of the third decoder extracts features.
[0069] The schematic diagram of the DA network structure is shown in Appendix Figure 3 . The DA network consists of two branches: channel attention and spatial attention. In the channel attention branch, the input features are subjected to global average pooling, flattening, a fully connected layer, and ReLU activation, and then passed through another fully connected layer and Sigmoid activation to obtain the channel attention weights. In the spatial attention branch, the input features pass through two 1x1 convolutional layers and ReLU activation, and finally pass through a 1x1 convolutional layer and Sigmoid activation to obtain the spatial attention weights. The channel attention weights and the spatial attention weights are multiplied by the input features respectively and then fused to form the final output features. This design enables the network to simultaneously focus on the channel and spatial information of the input features, thereby enhancing the richness and accuracy of feature representation.
[0070] The radar emitter signal feature extraction network proposed by the present invention, namely DARA, focuses on the time-domain correlation pattern of the pulse sequence through spatial attention in the DualAttention module, enhances the key parameter dimension through channel attention, the depthwise separable module can reduce the network complexity and computational amount in the face of high-density pulse streams, and the multi-level residual design retains the original pulse information through skip connections.
[0071] The extracted features obtained in step S2 are fed into the AEC function. Based on the Silhouette Score and the Bayesian Information Criterion (BIC), the number of clusters of K-means clustering and Gaussian Mixture Model clustering (GMM) are optimized respectively, and the two clustering results are fused through hierarchical clustering, aiming to combine the advantages of both to improve the clustering robustness and solve the limitations of a single algorithm.
[0072] The AEC function of the present invention adopts a multi-model collaborative mechanism, and simultaneously extracts the geometric cluster center of K-means clustering and the probability distribution center of Gaussian mixture model clustering for dual-center generation, overcoming the defect that a single model is sensitive to the data distribution form; automatically selects the optimal k value of K-means clustering by maximizing the silhouette coefficient, determines the optimal number of Gaussian components by minimizing BIC, avoids the subjectivity of artificially setting parameters, and enhances the adaptability to unknown data distributions; merges the central points of K-means clustering and GMM and then performs hierarchical clustering (Hierarchical Clustering, HAC) fusion to reduce the computational complexity. The label reassignment strategy is based on the nearest neighbor assignment of the merged new center, which not only retains the interpretability of hierarchical clustering but also avoids the memory bottleneck when HAC directly processes large-scale data. Compared with traditional clustering, it is more suitable for multi-modal mixed data, shows stronger adaptability and robustness in complex data scenarios, and is especially suitable for application scenarios that require automatic parameter adjustment and multi-modal feature fusion.
[0073] Optimizing K-means clustering based on the silhouette coefficient in step S3 of the present invention includes: The silhouette coefficient combines the intra-cluster cohesion and the inter-cluster separation. Set the maximum number of clusters max_clusters, traverse the number of clusters k∈[2,max_clusters], and evaluate the clustering quality through the silhouette coefficient. The value range of the silhouette coefficient is [-1,1], and the higher the value, the better the clustering effect. The data points within each cluster in the clustering under this number of clusters are closer. Select the number of clusters with the highest silhouette coefficient. After K-means clustering, the cluster centers of each cluster will be obtained, that is, the geometric cluster center of K-means clustering.
[0074] In step S3 of the present invention, optimizing the number of clusters of GMM based on the Bayesian Information Criterion includes: setting the maximum number of clusters max_clusters, traversing the number of clusters k ∈ [2, max_clusters], evaluating the clustering quality through the Bayesian Information Criterion, and the Bayesian Information Criterion is expressed as:
[0075] BIC = ln(n)k - 2ln(L) (2)
[0076] In formula (2), n is the sample size, that is, the number of data points, k is the number of free parameters in the model, including the intercept term, and L is the maximum likelihood estimate value of the model on the data.
[0077] Select the model with the smallest Bayesian Information Criterion value. The smaller the Bayesian Information Criterion value, the better the model, and the probability distribution center of the Gaussian mixture model clustering is obtained.
[0078] Step S3 of the present invention further includes:
[0079] By stacking the geometric cluster centers of k-means clustering and the probability distribution centers of Gaussian mixture model clustering, the diversity of candidate centers is increased, covering different distribution characteristics of the data. Subsequently, hierarchical clustering (HAC) fusion is performed on the set of geometric cluster centers of k-means clustering and the probability distribution centers of Gaussian mixture model clustering. Similar centers are merged according to a preset distance threshold. Finally, by calculating the Euclidean distance from the sample to each fused cluster center, the sample is divided into the nearest cluster, forming an adaptive clustering result that takes into account the advantages of density distribution and geometric partitioning. According to the fused centers, calculate the Euclidean distance from all samples to each center, assign the samples to the nearest center, form clustering labels, align the clustering labels with the true labels, evaluate the clustering effect, and complete the sorting of radar emitter signals.
[0080] The adaptive ensemble clustering proposed by the present invention, that is, the AEC function, improves the quality and robustness of the clustering result by fusing multiple clustering results and combining different clustering advantages; supports dynamic adjustment of the number of clusters, is fully automated from cluster number selection to final merging, reduces manual intervention, and is applicable to unknown signal scenarios with parameter overlap and dynamic changes in complex electromagnetic environments.
[0081] The performance of the present invention is verified through simulation experiments below.
[0082] 1. Experimental settings
[0083] The known class signal data set is set with reference to Table 1.
[0084] Table 1
[0085]
[0086] 2. Simulation Settings
[0087] The specific environment of the experiment of the present invention is shown in Table 2. The model is constructed using the PyTorch deep learning framework and relevant experiments are completed. The experimental platform consists of an Inter I5-13400F processor (10 cores), 16GB of memory, and an NVIDIA GeForce RTX3060. The optimizer is set to the Adam optimizer, and the loss function of the experiment is set to the mean squared error loss function. The formula is as follows:
[0088]
[0089] In formula (3), y_pred represents the predicted value of the model, y_true represents the true value, and the smaller the MSELoss, the smaller the difference between the predicted result of the model and the true value.
[0090] 3. Baseline Methods
[0091] To further evaluate the effectiveness of the present invention, clustering methods such as k-means, DBSCAN, MeanShift, GMM, Birch, Hierarchical, MiniBatchKMeans, BayesianGaussianMixture, OPTICS, etc. are selected, and the parameters of these algorithms are obtained using adaptive algorithms. A radar emitter signal sorting method based on deep clustering is selected for comparison. The feature extraction network selects an autoencoder, and the clustering selects adaptive k-means and adaptive GMM in the above traditional clustering.
[0092] 4. Evaluation Metrics
[0093] The evaluation metrics of the experiment adopt three metrics of the evaluation metrics of clustering: Adjusted Rand Index (ARI), Normalized Mutual Information (NMI), and Purity.
[0094] The Adjusted Rand Index is a metric used to evaluate the similarity between the clustering result and the true label. The value range of ARI is [-1,1], and the larger the value, the better the agreement with the true result, that is, the better the clustering effect. The Normalized Mutual Information is commonly used in clustering to measure the similarity between two clustering results. The value range is [0,1]. Purity calculates the proportion of samples of the dominant true class in each predicted cluster, and the overall purity is obtained after weighted averaging. The value range is [0,1]. The closer the value is to 1, the higher the consistency between the clustering result and the true label, that is, the better the clustering effect.
[0095] 5. Experimental Results
[0096] Refer to Table 2 for the experimental results comparison with traditional clustering.
[0097] Table 2
[0098] Method ARI MNI Purity k-means 0.040 0.060 0.030 DBSCAN 0.056 0.186 0.103 Meanshift 0.064 0.206 0.130 GMM 0.135 0.247 0.276 Birch 0.069 0.213 0.146 Hierarchical 0.076 0.229 0.197 MiniBatchKMeans 0.084 0.230 0.200 BayesianGaussianMixture 0.098 0.256 0.283 OPTICS 0.105 0.298 0.312 Ours 0.497 0.670 0.733
[0099] As can be seen from Table 2, the present invention is 0.457, 0.441, 0.433, 0.362, 0.428, 0.421, 0.413, 0.399, 0.392 higher than clustering methods such as k-means, DBSCAN, MeanShift, GMM, Birch, Hierarchical, MiniBatchKMeans, BayesianGaussianMixture, OPTICS in terms of the ARI index; 0.61, 0.484, 0.464, 0.423, 0.457, 0.441, 0.44, 0.414, 0.372 higher in terms of the MNI index; and 0.703, 0.63, 0.603, 0.457, 0.587, 0.536, 0.533, 0.45, 0.421 higher in terms of the Purity index.
[0100] Refer to Table 3 for the experimental results comparison with existing deep clustering.
[0101] Table 3
[0102] Method ARI MNI Purity Autoencoder + k-means 0.151 0.497 0.343 Autoencoder + GMM 0.401 0.521 0.623 Autoencoder + AEC 0.448 0.613 0.662 DARA + k-means 0.257 0.564 0.393 DARAE + GMM 0.436 0.573 0.679 DARA + AEC 0.497 0.670 0.733
[0103] From the results in Table 3, it can be seen that DARA + k-means is 0.106, 0.067, and 0.05 higher than Autoencoder + k-means in terms of ARI, MNI, and Purity respectively; DARA + GMM is 0.035, 0.052, and 0.056 higher than Autoencoder + GMM in terms of ARI, MNI, and Purity respectively; DARA + Adaptive Hybrid Clustering is 0.049, 0.057, and 0.071 higher than Autoencoder + Adaptive Hybrid Clustering in terms of ARI, MNI, and Purity respectively. The experimental results show that compared with the autoencoder, DARA can enhance important features and suppress redundant information because it has a DA module. The depthwise separable convolution module significantly reduces the number of parameters and computational complexity, has high computational efficiency, and has a stronger feature extraction ability, and can discover the potential laws between data. Adaptive ensemble clustering can combine the advantages of two clustering methods and adapt to multimodal data distributions.
[0104] Based on the above results analysis, the method of the present invention can more comprehensively extract the potential features of radar signals, discover deep laws in radar sorting. Compared with traditional clustering and some existing deep clustering methods, the method of the present invention has better sorting effect and can accurately and efficiently complete the sorting task.
[0105] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, any modification, equivalent replacement, and improvement made within the spirit and principle of the present invention shall be covered by the protection scope of the present invention.
Claims
1. An adaptive integrated deep clustering method for radar emitter signal sorting, characterized in that Including the steps: S1. Perform normalization processing on the data to obtain normalized data excluding different parameter dimensions; S2. Input the normalized data into a dual-attention residual autoencoder for feature extraction and output the extracted features; S3. Input the extracted features into an adaptive ensemble clustering function, optimize K-means clustering based on the silhouette coefficient, optimize Gaussian mixture model clustering based on the Bayesian information criterion, fuse the clustering results of the K-means clustering and Gaussian mixture model clustering through hierarchical clustering to obtain clustering labels, align the clustering labels with the true labels, evaluate the clustering effect, and complete the sorting of radar emitter signals.
2. The method for sorting radar emitter signals by adaptive integrated deep clustering according to claim 1, wherein The step S1 includes: The normalization processing is expressed as: In formula (1), p represents the parameter vector to be normalized.
3. The method for sorting radar emitter signals based on adaptive integrated deep clustering according to claim 1, wherein The step S2 includes: The dual-attention residual autoencoder consists of an encoder part and a decoder part; The normalized data is input into the encoder part, and the encoder part gradually compresses the input features into the latent space; The decoder part reconstructs and gradually restores the feature dimensions. At the same time, it has a skip connection through the Linear function and ensures that the output is between 0 and 1 through the Sigmoid activation function.
4. An adaptive integrated deep clustering-based radar emitter signal sorting method according to claim 3, characterized in that, The skip connection through the Linear function includes: The output of the first encoder is adjusted in dimension through the Linear function and added to the output of the second decoder, and the output of the second encoder is adjusted in dimension through the Linear function and added to the output of the first decoder.
5. An adaptive integrated deep clustering-based radar emitter signal sorting method according to claim 4, characterized in that, The encoder includes a first encoder, a second encoder, and a third encoder; The first encoder includes a ResDSConv function, a dual-attention mechanism, and a Linear function connected in sequence. The input end of the ResDSConv function of the first encoder is connected to the normalized data, and the output end of the Linear function of the first encoder is connected to the second encoder; The second encoder includes a ResDSConv function and a dual-attention mechanism. The input end of the ResDSConv function of the second encoder is connected to the output end of the first encoder, the output end of the ResDSConv function of the second encoder is connected to the input end of the dual-attention mechanism of the second encoder, and the output end of the dual-attention mechanism of the second encoder is connected to the input end of the third encoder; The third encoder consists of a Linear function, and the output end of the third encoder is connected to the first decoder.
6. The method for sorting radar emitter signals by adaptive integrated deep clustering according to claim 5, wherein, The decoder part includes a first decoder, a second decoder, and a third decoder; The first decoder includes a Linear function and a ResDSConv function. The input end of the Linear function of the first decoder is connected to the output end of the third encoder, the output end of the Linear function of the first decoder is connected to the input end of the ResDSConv function of the first decoder, and the output end of the ResDSConv function of the first decoder is connected to the input end of the second decoder; The second decoder includes a Linear function, a dual attention mechanism, and a ResDSConv function connected in sequence. The input end of the Linear function of the second decoder is connected to the output end of the first decoder, and the output end of the ResDSConv function of the second decoder is connected to the input end of the third decoder; The third decoder includes a Sigmoid activation function and a Linear function. The input end of the Sigmoid activation function of the third decoder is connected to the output end of the second decoder. The output end of the Sigmoid activation function of the third decoder is connected to the input end of the Linear function of the third decoder, and the output of the Linear function of the third decoder extracts features.
7. An adaptive integrated deep clustering-based radar emitter signal sorting method according to claim 6, characterized in that, The dual attention mechanism includes a channel attention branch and a spatial attention branch; In the channel attention branch, the input features are subjected to global average pooling, flattening, a fully connected layer, and ReLU activation, and then a channel attention weight is obtained through a fully connected layer and a Sigmoid activation function; In the spatial attention branch, the input features are passed through two 1x1 convolutional layers and ReLU activation, and finally a spatial attention weight is obtained through a 1x1 convolutional layer and Sigmoid activation; The channel attention weight and the spatial attention weight are multiplied by the input features respectively and then fused to form the final output features.
8. An adaptive integrated deep clustering-based radar emitter signal sorting method according to claim 1, characterized in that The optimization of K-means clustering based on the silhouette coefficient in step S3 includes: The silhouette coefficient combines the within-cluster cohesion and the between-cluster separation. Set the maximum number of clusters max_clusters, traverse the number of clusters k ∈ [2, max_clusters], and evaluate the clustering quality through the silhouette coefficient. The value range is [-1, 1], and the higher the value, the better the clustering effect; Select the number of clusters with the highest silhouette coefficient to obtain the geometric cluster centers of the k-means clustering.
9. An adaptive integrated deep clustering-based radar emitter signal sorting method according to claim 8, characterized in that The optimization of Gaussian mixture model clustering based on the Bayesian information criterion in step S3 includes: Set the maximum number of clusters max_clusters, traverse the number of clusters k ∈ [2, max_clusters], and evaluate the clustering quality through the Bayesian information criterion. The Bayesian information criterion is expressed as: BIC = ln(n)k - 2ln(L) (2) In formula (2), n is the sample size, that is, the number of data points, k is the number of free parameters in the model, including the intercept term, and L is the maximum likelihood estimate value of the model on the data; Select the model with the smallest Bayesian information criterion value. The smaller the Bayesian information criterion value, the better the model, and obtain the probability distribution center of the Gaussian mixture model clustering.
10. An adaptive integrated deep clustering-based radar emitter signal sorting method according to claim 9, characterized in that, Step S3 further includes: By stacking the geometric cluster centers of the k-means clustering and the probability distribution centers of the Gaussian mixture model clustering, the diversity of candidate centers is increased to cover different distribution characteristics of the data; Perform hierarchical clustering fusion on the set of geometric cluster centers of the k-means clustering and the probability distribution centers of the Gaussian mixture model clustering, merge similar centers according to a preset distance threshold, and divide the samples into the nearest clusters by calculating the Euclidean distances from the samples to each cluster center after fusion, so as to form an adaptive clustering result that takes into account the advantages of density distribution and geometric partitioning; According to the centers after fusion, calculate the Euclidean distances from all samples to each center, assign the samples to the nearest center to form clustering labels, align the clustering labels with the true labels, evaluate the clustering effect, and complete the sorting of radar emitter signals.