Multi-view depth alignment unknown electromagnetic signal clustering method

Abstract

This invention proposes a multi-view depth-aligned clustering method for unknown electromagnetic signals, belonging to the field of electromagnetic signal recognition. It addresses the problems of existing unknown electromagnetic signal clustering methods relying on manual features, single representations, and pre-defined categories, leading to insufficient feature representation, poor clustering stability, and limited ability to discover unknown categories in complex electromagnetic environments. The method includes: acquiring each unknown electromagnetic signal sample and its corresponding perturbation views from multiple perspectives to construct a training sample set; selecting two fixed perturbation views for each sample in the training sample set, corresponding to the online branch and target branch of a two-branch self-distillation feature network, respectively; constructing consistency constraints using the corresponding output features of the two branches; training the feature network; and constructing an unknown electromagnetic signal clustering model based on this constraint; training the unknown electromagnetic signal clustering model using the training sample set; and performing actual clustering prediction using the unknown electromagnetic signal clustering model.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic signal recognition, and in particular to a multi-view depth-aligned method for clustering unknown electromagnetic signals. Background Technology

[0002] With the rapid development of wireless communication and IoT technologies, the parallel access of numerous heterogeneous terminals has led to increasingly complex electromagnetic environments. Not only are spectrum resources highly congested, but signal morphology is also constantly changing, posing higher demands on the identification and management of electromagnetic signals. In practical applications, receivers often need to automatically distinguish and classify electromagnetic signals of unknown origin and type without complete prior knowledge, providing support for tasks such as spectrum monitoring, interference detection, and situational awareness. Traditional supervised identification methods rely on a large number of labeled samples, but the communication environment is complex and ever-changing, and modulation methods and service protocols are constantly evolving. Not only is the acquisition of high-quality labels costly, but it is also difficult to promptly cover new or abnormal signals. Furthermore, supervised models are typically trained only on known categories, exhibiting insufficient adaptability and generalization ability for unknown signals. Therefore, there is an urgent need to develop a clustering method for unknown electromagnetic signals that can automatically mine signal structural differences and classify them under label-less conditions.

[0003] Existing unsupervised electromagnetic signal clustering methods are mainly divided into two categories: those based on manual features and those based on deep learning. The former relies on manual statistical features such as higher-order cumulants, cyclic spectra, and instantaneous features, combined with algorithms like K-means and spectral clustering to achieve signal segmentation. However, these methods are highly sensitive to channel fading, frequency offset, and noise, and their discriminative ability significantly decreases in scenarios with low signal-to-noise ratios or changes in signal structure, making them difficult to adapt to complex communication scenarios. The latter utilizes autoencoders, contrastive learning, and deep clustering networks to automatically learn features from the original IQ sequences, improving the clustering effect of unknown signals to some extent. However, most existing methods use a single input form or a fixed representation method for feature learning, making it difficult to fully utilize the complementary information between different representation domains such as the time domain, frequency domain, and instantaneous amplitude. Furthermore, differences in data distribution exist between different representation domains, and simply combining multi-domain information cannot guarantee that features are mapped to a unified semantic space, resulting in insufficient adaptability of the model to signal changes under complex observation conditions. Furthermore, most existing methods require pre-setting the number of cluster categories, but the number of categories of unknown signals in the actual electromagnetic environment is usually impossible to determine in advance. The constraint of a fixed number of categories limits the model's ability to discover and adapt to new categories.

[0004] Therefore, existing clustering techniques for unknown electromagnetic signals still suffer from problems such as insufficient utilization of multi-view representation, limited semantic consistency learning ability under different observation conditions, insufficient stability of clustering structure, and inability to adaptively determine the number of unknown categories. These issues make it difficult to meet the application requirements for automatic discovery and stable clustering of unknown signals in complex open electromagnetic environments. Summary of the Invention

[0005] Based on the above analysis, this invention provides a multi-view depth-aligned clustering method for unknown electromagnetic signals, aiming to solve the problems of existing unknown electromagnetic signal clustering methods, such as strong dependence on a single representation domain, poor adaptability to complex channel environments, and insufficient stability of clustering structures.

[0006] This invention provides a multi-view depth-aligned clustering method for unknown electromagnetic signals, the method comprising: Acquire each unknown electromagnetic signal sample and its corresponding perturbation views from multiple perspectives to construct a training sample set; For each unknown electromagnetic signal sample in the training sample set, two perturbation views are fixedly selected, corresponding to the online branch and the target branch of the dual-branch self-distillation feature network, respectively. Consistency constraints are constructed using the predicted features output by the online branch and the target features output by the target branch, and the dual-branch self-distillation feature network is trained. A clustering model for unknown electromagnetic signals is constructed based on the trained dual-branch self-distillation feature network; the clustering model for unknown electromagnetic signals is trained using the training sample set; The unknown electromagnetic signal to be clustered is input into the trained unknown electromagnetic signal clustering model, and the clustering result is output.

[0007] Based on the above solution, the present invention also makes the following improvements: Furthermore, the dual-branch self-distillation feature network includes an online branch, a target branch, a shared encoder, and a prediction network; wherein, The online branch uses randomized representation domain projection for the received perturbation view. In each iteration, it independently extracts a domain from the time domain, frequency domain, and instantaneous domain as the target subspace, and then performs view representation and projection mapping in the corresponding domain in turn. The target branch performs temporal-domain view representation and projection mapping on the received perturbation view. The shared encoder performs shared encoding on the projection mapping results output by the online branch and the target branch, respectively, and outputs semantic features and the target features accordingly. A prediction network is used to predict the semantic features and output the predicted features.

[0008] Furthermore, consistency constraints are constructed using the predicted features of the output corresponding to the online branch and the target features of the output corresponding to the target branch, resulting in a consistency loss based on feature similarity. :

[0009] in, This indicates that the gradient operation is stopped. Indicates the first Predicted features corresponding to an unknown electromagnetic signal sample; Indicates the first Target features corresponding to an unknown electromagnetic signal sample; Represents the L2 norm, Indicates transpose. This represents the number of unknown electromagnetic signal samples in a training batch during the training of a dual-branch self-distillation feature network.

[0010] Furthermore, the projection mapping under each domain is implemented using an independent projection mapping module.

[0011] Furthermore, the training of the dual-branch self-distillation feature network is performed by: Backpropagation is performed based on the consistency loss function based on feature similarity. The gradient of the consistency loss function with respect to each trainable parameter in the dual-branch self-distillation feature network is calculated. The gradient is then used to optimize the network parameters of the shared encoder, the prediction network, and the projection mapping module in different domains. The training continues until the consistency loss reaches the preset convergence condition of the two-branch self-distillation feature network, thus obtaining the trained two-branch self-distillation feature network.

[0012] Furthermore, the unknown electromagnetic signal clustering model constructed based on the trained dual-branch self-distillation feature network is executed as follows: The online branches and their connected shared encoders in the trained dual-branch self-distillation feature network are used to construct a feature extraction network. After the feature extraction network, a clustering center module is connected to set multiple category centers in the semantic feature space. This results in the construction of an unknown electromagnetic signal clustering model.

[0013] Furthermore, the step of training the unknown electromagnetic signal clustering model using the training sample set involves: During each training process, the class centers and their quantities for unknown electromagnetic signals are adaptively adjusted based on the sample distribution relationship of different class centers at present. Based on the adaptively adjusted category centers of the unknown electromagnetic signals, the clustering model of the unknown electromagnetic signals is trained and optimized using a joint optimization objective function based on multi-view category consistency constraints and cluster distribution entropy regularization constraints. Determine whether the joint optimization objective function meets the preset convergence condition of the unknown electromagnetic signal clustering model. If not, proceed to the next training process; otherwise, save the training parameters of the unknown electromagnetic signal clustering model to obtain the successfully trained unknown electromagnetic signal clustering model.

[0014] Furthermore, jointly optimize the objective function. Represented as:

[0015] in, , These represent multi-perspective category consistency constraints and cluster distribution entropy regularization constraints, respectively. The weight coefficients represent the regularization constraints of the cluster distribution entropy.

[0016] further,

[0017] in, This represents the number of samples in a training batch in the clustering model of unknown electromagnetic signals. , They represent the first The class assignment results obtained by the first and second perturbation views of an unknown electromagnetic signal sample through time-domain representation under online branching; Indicates the temperature coefficient; This represents the similarity function.

[0018] further,

[0019] in, This indicates that the samples in the current training batch have been assigned to the th ... The average probability of each category; , Indicates the first The unknown electromagnetic signal sample belongs to the first The probability of each category.

[0020] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: 1. This invention constructs a feature learning mechanism based on multi-view representation projection and bi-branch self-distillation, which maps different forms of signal representation such as time domain, frequency domain and instantaneous domain to a unified feature space, and uses consistency constraints between different input views to learn stable semantic features, thereby improving the feature representation capability of unknown electromagnetic signals.

[0021] 2. This invention enhances the stability of the clustering structure by constructing a soft-assignment clustering model based on category centers, and combining category assignment consistency constraints under different representation perspectives and cluster distribution entropy regularization constraints. It avoids problems such as unbalanced category assignment, category center degradation, and fuzzy cluster boundaries in traditional clustering methods, thereby improving the accuracy of discovering unknown electromagnetic signal categories.

[0022] 3. This invention introduces a dynamic cluster evolution mechanism, which adaptively performs category splitting, category merging, and category center updating based on the compactness of samples within a category and the similarity between categories. It does not require presetting the number of unknown signal categories and can automatically adjust the clustering structure according to the changes in sample distribution in the actual electromagnetic environment, thereby improving the clustering capability of unknown electromagnetic signals in open environments.

[0023] As mentioned above, this method outputs clustering results based on the dynamically updated category structure, thereby enabling the automatic discovery and clustering of unknown electromagnetic signals.

[0024] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0025] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 This is a flowchart of the multi-view depth-aligned unknown electromagnetic signal clustering method provided in the embodiments of the present invention; Detailed Implementation Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0026] A specific embodiment of the present invention provides a multi-view depth-aligned clustering method for unknown electromagnetic signals, the flowchart of which is shown below. Figure 1 As shown, the specific implementation process is explained below.

[0027] Step S1: Obtain each unknown electromagnetic signal sample and its corresponding perturbation views from multiple perspectives to construct a training sample set; Step S2: For each unknown electromagnetic signal sample in the training sample set, two perturbation views are fixedly selected, corresponding to the online branch and the target branch of the dual-branch self-distillation feature network, respectively. Consistency constraints are constructed using the predicted features output by the online branch and the target features output by the target branch, and the dual-branch self-distillation feature network is trained. Step S3: Construct an unknown electromagnetic signal clustering model based on the trained dual-branch self-distillation feature network; train the unknown electromagnetic signal clustering model using the training sample set; Step S4: Input the unknown electromagnetic signal to be clustered into the trained unknown electromagnetic signal clustering model, and output the clustering result.

[0028] The specific implementation process of each step is described in detail below.

[0029] In step S1, the unknown electromagnetic signal sample is represented as a discrete-time complex baseband signal. For example, for unknown electromagnetic signal data, which is a discrete-time complex baseband sequence, it is represented as:

[0030] in, Indicating the first unknown electromagnetic signal data Complex values ​​of each sampling point , They are respectively In-phase signal components and quadrature signal components, The imaginary unit, The length of the signal sequence for the unknown electromagnetic signal data.

[0031] In the specific implementation process, various perturbation enhancement processes are applied to unknown electromagnetic signal samples. For example, by simulating non-ideal propagation and reception factors such as frequency offset, phase offset, noise, and channel fading, perturbation views from multiple perspectives are constructed to maintain semantic consistency of signal categories.

[0032] For example, for By introducing non-ideal factors under different propagation environments and reception conditions, a construct is made. The Perturbation view :

[0033] in, , , , , They represent the first The amplitude of the disturbance, the disturbance coefficient, the channel response, the frequency offset, the phase offset, and the additive noise disturbance; This represents the convolution operation. Indicates the sampling period. This indicates the sequence number of the discrete sampling point.

[0034] In practice, by randomly adjusting the amplitude perturbation coefficient, channel response, frequency offset, phase offset, and noise intensity parameters, multiple perspectives of perturbation views corresponding to the same unknown electromagnetic signal can be generated.

[0035] It should be noted that in this embodiment, the perturbation views from multiple perspectives change the signal observation conditions while keeping the category semantic information of the original unknown electromagnetic signal unchanged. This information is then used for subsequent bi-branch self-distillation feature learning, enabling the model to learn stable and consistent signal semantic features under different reception states.

[0036] In step S2, the dual-branch self-distillation feature network includes an online branch, a target branch, a shared encoder, and a prediction network. The online branch projects the received perturbed view onto a randomized representation domain. In each iteration, it independently extracts one domain from the time domain, frequency domain, and instantaneous domain as the target subspace, and then sequentially performs view representation and projection mapping in the corresponding domain. The target branch performs fixed time-domain view representation and projection mapping on the received perturbed view. The shared encoder performs shared encoding on the projection mapping results output by the online branch and the target branch, respectively outputting semantic features and the target feature. The prediction network is used to predict the semantic features and output the predicted features. It should be noted that in this embodiment, the projection mappings corresponding to different representation domains of the online branch and the time-domain projection mapping are all projected onto a unified feature space; and the target feature and the predicted feature have the same dimension, that is, the input dimension and output dimension of the prediction network are the same. Specific details are as follows.

[0037] Assuming that the two perturbation views are fixedly selected, respectively, the first... Perturbation view and the first A perturbation view. The first unknown electromagnetic signal sample Perturbation view, the first The perturbation views are represented as follows:

[0038] in, , They represent the first species, first The transformation process of a perturbation.

[0039] (1) Online branches and their corresponding predictive feature generation process In practice, the online branch can randomly select a view representation format for the domain (viewpoint) based on a preset strategy. A visual representation of an unknown electromagnetic signal sample satisfy:

[0040] in, , , These represent the time-domain view representation, the frequency-domain view representation, and the instantaneous-domain view representation, respectively.

[0041] A time-domain view can represent the time-domain waveform information of an unknown electromagnetic signal, used to describe the dynamic characteristics of the signal amplitude changing over time. Time-domain view representation of an unknown electromagnetic signal sample Represented as:

[0042] Frequency domain views are used to describe the spectral structure and frequency distribution characteristics of a signal. In the frequency domain view representation of the unknown electromagnetic signal sample, the first... Frequency components Represented as:

[0043] in, Indicates the first In the unknown electromagnetic signal sample, the first Temporal view representation of each sampling point.

[0044] The instantaneous domain view characterizes the dynamic characteristics of the amplitude and phase of an unknown electromagnetic signal as a function of time. Instantaneous domain view representation of an unknown electromagnetic signal sample Represented as:

[0045] in, , They represent the first In the unknown electromagnetic signal sample, the first Instantaneous amplitude and instantaneous phase at each sampling point; , , , They represent the first In the unknown electromagnetic signal sample, the first The in-phase signal components and quadrature signal components at each sampling point.

[0046] Because the data formats and distributions differ across representation domains, feature space transformation is required using the corresponding domain's projection mapping module. The unknown electromagnetic signal sample at the th View representation of each domain Unified spatial features after projection mapping to a unified space Represented as:

[0047] in, Indicates the first Each representation domain corresponds to a projection mapping module.

[0048] Will Input the shared encoder to get the first Semantic features of each representation domain :

[0049] in, This indicates a shared encoder, meaning the feature encoder is shared by the online branch and the target branch. It should be noted that for each electromagnetic signal sample at a location, the online branch randomly selects only one processing method within the representation domain. Therefore, that is, the... Semantic features of an unknown electromagnetic signal sample .

[0050] Will Input the prediction network to obtain the first... Predictive features of an unknown electromagnetic signal sample :

[0051] It should be noted that, in this embodiment, the prediction network is a cross-view semantic alignment module used to enhance representation consistency and prevent degradation during self-supervised training. The prediction network can break the symmetry between the online network and the target network, promoting the model to learn discriminative representations rather than simply copying input features.

[0052] (2) Target branch and its corresponding target feature generation process In this embodiment, the target branch is used to fix the received alternative perturbation view using time-domain complex baseband representation and time-domain projection mapping. It should be noted that in this embodiment, the online branch and the target branch can share the time-domain projection mapping module.

[0053] For example, the first Time-domain view representation of an unknown electromagnetic signal sample The unified spatial features after projection mapping onto a unified spatial space Represented as:

[0054] Then, Input the shared encoder to obtain the first Target features corresponding to an unknown electromagnetic signal sample :

[0055] Then, consistency constraints can be constructed between the predicted features output by the online branch and the target features output by the target branch. This optimizes the dual-branch self-distillation feature network, ensuring that the same unknown electromagnetic signal maintains a consistent semantic representation under different input views, thereby obtaining discriminative features of the unknown electromagnetic signal. The specific implementation process is described below.

[0056] In this embodiment, to constrain the semantic consistency of the same unknown electromagnetic signal under different input views, a consistency constraint is constructed using the predicted features of the output corresponding to the online branch and the target features of the output corresponding to the target branch, resulting in a consistency loss based on feature similarity. :

[0057] in, Represents the L2 norm, Indicates transpose. This represents the number of unknown electromagnetic signal samples in a training batch during the training of a dual-branch self-distillation feature network.

[0058] Preferably, during the training of the two-branch self-distillation feature network, backpropagation is performed based on a consistency loss function based on feature similarity. The gradient of the consistency loss function with respect to each trainable parameter in the two-branch self-distillation feature network is calculated, and the gradient is used to optimize the network parameters of the shared encoder, prediction network, and projection mapping modules in different domains. This process continues until the consistency loss reaches a preset convergence condition, resulting in a successfully trained two-branch self-distillation feature network. Therefore, in specific implementations, minimizing the consistency loss function... By optimizing the network parameters of the shared encoder, prediction network, and projection mapping module in different domains, the same unknown electromagnetic signal from different perturbation views and different signal representations remains close in the feature space, thus obtaining a trained dual-branch self-distillation feature network.

[0059] In step S3, the online branches in the trained dual-branch self-distillation feature network and their connected shared encoders are used to construct a feature extraction network; a clustering center module is connected after the feature extraction network to set multiple category centers in the semantic feature space; and an unknown electromagnetic signal clustering model is constructed.

[0060] Preferably, in this embodiment, the unknown electromagnetic signal clustering model is trained using a training sample set, and the following steps are performed: Step S31: In each training process, the class center and its quantity of the unknown electromagnetic signal are adaptively adjusted according to the sample distribution relationship of different class centers at present; Step S32: Using the adaptively adjusted category centers of the unknown electromagnetic signals as a benchmark, the clustering model of the unknown electromagnetic signals is trained and optimized using a joint optimization objective function based on multi-view category consistency constraints and cluster distribution entropy regularization constraints. Step S33: Determine whether the joint optimization objective function has reached the preset convergence condition of the unknown electromagnetic signal clustering model (depending on the situation). If not, proceed to the next training process; otherwise, save the training parameters of the unknown electromagnetic signal clustering model to obtain the trained unknown electromagnetic signal clustering model.

[0061] The specific implementation process is described below.

[0062] First, step S31 will be explained as follows. In this embodiment, the clustering center module sets in the semantic feature space. Each category center:

[0063] in, This indicates the number of cluster categories currently being represented. Indicates the first The category centers corresponding to each potential unknown category are used to describe the distribution location of that category in the semantic feature space.

[0064] For each unknown electromagnetic signal sample, calculate its similarity to each category center, and establish the soft assignment probability of the sample belonging to different categories. The unknown electromagnetic signal sample belongs to the first The probability of each category Represented as:

[0065] in, Represents the similarity function; This represents the temperature coefficient, used to adjust the smoothness of the category probability distribution. The soft probability assignment described above establishes a correlation between the semantic features of a sample and its category center, allowing unknown electromagnetic signals with similar semantic structures to obtain higher category assignment probabilities and gradually cluster towards their corresponding category centers.

[0066] During the clustering model training process, the number of cluster categories is adaptively adjusted based on the sample distribution relationship in the current cluster structure to achieve adaptive adjustment of the number of cluster categories for unknown electromagnetic signals. The process is as follows: For each cluster category, calculate the compactness of samples within that category. Intra-category sample compactness of a potentially unknown category Represented as:

[0067] in, Indicates the first The number of unknown electromagnetic signal samples included in each category Indicates belonging to the first A collection of unknown electromagnetic signal samples of a class. This indicates the degree of dispersion in the distribution of unknown electromagnetic signal samples within this category.

[0068] When the sample distribution within a certain category is relatively dispersed, and satisfies When this indicates that the current category may contain multiple unknown signals with different semantic structures, a category splitting operation is performed on that category. This generates new category centers.

[0069] Further calculate the similarity between different category centers :

[0070] in, Indicates the first Class and the Category similarity between classes.

[0071] When the similarity between two category centers meets the following condition When this occurs, it indicates that the two categories have high semantic consistency, and a merging operation is performed on the corresponding categories. ,in, , These represent the category splitting threshold and the category merging threshold, respectively.

[0072] After adjusting the number of categories, recalculate the category centers based on the updated soft allocation results:

[0073] in, Indicates the first The unknown electromagnetic signal sample belongs to the first The probability of each category It represents the semantic features of the samples. By iteratively performing category quantity evaluation, category generation, category merging, and category center update during the clustering training process, the clustering model does not need to know the true number of categories of unknown electromagnetic signals in advance, and can automatically discover potential unknown categories based on changes in sample distribution.

[0074] In step S32, a joint optimization objective function based on multi-view category consistency constraints and cluster distribution entropy regularization constraints can be constructed according to the soft assignment relationship between the semantic features of the unknown electromagnetic signal samples and the category centers, and the clustering model of the unknown electromagnetic signals can be trained.

[0075] Furthermore, to enhance the consistency of clustering results under different representation perspectives, multi-perspective category consistency constraints are constructed using the category assignment results obtained through time-domain representation under online branching for the first and second perturbation views of each unknown electromagnetic signal sample. , represented as:

[0076] in, This represents the number of samples in a training batch in the clustering model of unknown electromagnetic signals. , They represent the first The first and second perturbation views of an unknown electromagnetic signal sample are obtained through time-domain representation under online branching, resulting in class assignments. The first and second perturbation views can be selected from the multi-view perturbation views in step S1. This is achieved by minimizing the class consistency loss. This allows the same unknown electromagnetic signal sample to maintain a consistent category classification under different representation perspectives.

[0077] Furthermore, to avoid a large number of samples being concentrated in a few categories during the clustering process, which could prevent the effective discovery of some potential categories, a clustering distribution entropy regularization constraint is introduced. , represented as:

[0078] in, This indicates that the samples in the current training batch have been assigned to the th ... The average probability of each category; By minimizing the entropy regularization constraint, a reasonable sample distribution is obtained for each category, avoiding the collapse of clustering results towards a few categories and improving the ability to discover the structure of unknown categories.

[0079] Finally, the joint optimization objective function for the unknown electromagnetic signal clustering model is constructed as follows. , represented as:

[0080] in, The weight coefficients represent the regularization constraints of the cluster distribution entropy.

[0081] In the specific implementation process, the joint optimization objective function is minimized, while simultaneously optimizing the category centers, sample category assignment relationships, and cluster distribution structure. By executing the above training process, a trained clustering model for unknown electromagnetic signals can be obtained. It is important to emphasize that after training is complete, the cluster categories and category centers in the cluster center module are immediately determined.

[0082] In the specific execution of step S4, the electromagnetic signal to be clustered is input into the trained unknown electromagnetic signal clustering model. Based on the clustering assignment results output by the trained unknown electromagnetic signal clustering model, the cluster to which the electromagnetic signal belongs is determined, and the clustering result of the radio signal is obtained. The specific implementation process is described below.

[0083] For an unknown electromagnetic signal to be clustered, its time-domain complex baseband signal is input into a time-domain projection mapping module and a shared encoder to obtain the semantic features of the signal to be clustered. Represented as:

[0084] in, This indicates that the shared encoder has completed training. This indicates the projection mapping module.

[0085] Based on the final category center after training is complete Calculate the probability distribution of the unknown electromagnetic signal to be clustered into each category:

[0086] in, This indicates that the unknown electromagnetic signal to be clustered belongs to the first... The probability of each category.

[0087] The final clustering result is determined based on the category probability distribution. :

[0088] Through the above process, the trained shared encoder and the dynamically updated category structure are used to achieve automatic clustering of unknown electromagnetic signals.

[0089] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0090] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-view depth-aligned clustering method for unknown electromagnetic signals, characterized in that, The method includes: Acquire each unknown electromagnetic signal sample and its corresponding perturbation views from multiple perspectives to construct a training sample set; For each unknown electromagnetic signal sample in the training sample set, two perturbation views are fixedly selected, corresponding to the online branch and the target branch of the dual-branch self-distillation feature network, respectively. Consistency constraints are constructed using the predicted features output by the online branch and the target features output by the target branch, and the dual-branch self-distillation feature network is trained. A clustering model for unknown electromagnetic signals is constructed based on the trained dual-branch self-distillation feature network; the clustering model for unknown electromagnetic signals is trained using the training sample set; The unknown electromagnetic signal to be clustered is input into the trained unknown electromagnetic signal clustering model, and the clustering result is output.

2. The multi-view depth-aligned unknown electromagnetic signal clustering method according to claim 1, characterized in that, The dual-branch self-distillation feature network includes an online branch, a target branch, a shared encoder, and a prediction network; wherein... The online branch uses randomized representation domain projection for the received perturbation view. In each iteration, it independently extracts a domain from the time domain, frequency domain, and instantaneous domain as the target subspace, and then performs view representation and projection mapping in the corresponding domain in turn. The target branch performs temporal-domain view representation and projection mapping on the received perturbation view. The shared encoder performs shared encoding on the projection mapping results output by the online branch and the target branch, respectively, and outputs semantic features and the target features accordingly. A prediction network is used to predict the semantic features and output the predicted features.

3. The multi-view depth-aligned unknown electromagnetic signal clustering method according to claim 2, characterized in that, By constructing consistency constraints using the predicted features of the output corresponding to the online branch and the target features of the output corresponding to the target branch, a consistency loss based on feature similarity is obtained. : in, This indicates that the gradient operation is stopped. Indicates the first Predicted features corresponding to an unknown electromagnetic signal sample; Indicates the first Target features corresponding to an unknown electromagnetic signal sample; Represents the L2 norm, Indicates transpose. This represents the number of unknown electromagnetic signal samples in a training batch during the training of a dual-branch self-distillation feature network.

4. The multi-view depth-aligned unknown electromagnetic signal clustering method according to claim 3, characterized in that, The projection mapping for each domain is implemented using a separate projection mapping module.

5. The multi-view depth-aligned unknown electromagnetic signal clustering method according to claim 4, characterized in that, The training of the dual-branch self-distillation feature network is performed as follows: Backpropagation is performed based on the consistency loss function based on feature similarity. The gradient of the consistency loss function with respect to each trainable parameter in the dual-branch self-distillation feature network is calculated. The gradient is then used to optimize the network parameters of the shared encoder, the prediction network, and the projection mapping module in different domains. The training continues until the consistency loss reaches the preset convergence condition of the two-branch self-distillation feature network, thus obtaining the trained two-branch self-distillation feature network.

6. The multi-view depth-aligned unknown electromagnetic signal clustering method according to claim 5, characterized in that, The unknown electromagnetic signal clustering model is constructed based on the trained dual-branch self-distillation feature network, and the following steps are performed: The online branches and their connected shared encoders in the trained dual-branch self-distillation feature network are used to construct a feature extraction network; A clustering center module is connected after the feature extraction network to set multiple category centers in the semantic feature space; thus, an unknown electromagnetic signal clustering model is constructed.

7. The multi-view depth-aligned unknown electromagnetic signal clustering method according to claim 6, characterized in that, The process of training the clustering model for unknown electromagnetic signals using a training sample set is as follows: During each training process, the class centers and their quantities for unknown electromagnetic signals are adaptively adjusted based on the sample distribution relationship of different class centers at present. Based on the adaptively adjusted category centers of the unknown electromagnetic signals, the clustering model of the unknown electromagnetic signals is trained and optimized using a joint optimization objective function based on multi-view category consistency constraints and cluster distribution entropy regularization constraints. Determine whether the joint optimization objective function meets the preset convergence condition of the unknown electromagnetic signal clustering model. If not, proceed to the next training process; otherwise, save the training parameters of the unknown electromagnetic signal clustering model to obtain the successfully trained unknown electromagnetic signal clustering model.

8. The multi-view depth-aligned unknown electromagnetic signal clustering method according to claim 7, characterized in that, Joint optimization objective function Represented as: in, , These represent multi-perspective category consistency constraints and cluster distribution entropy regularization constraints, respectively. The weight coefficients represent the regularization constraints of the cluster distribution entropy.

9. The multi-view depth-aligned clustering method for unknown electromagnetic signals according to claim 8, characterized in that, in, This represents the number of samples in a training batch in the clustering model of unknown electromagnetic signals. , They represent the first The class assignment results obtained by the first and second perturbation views of an unknown electromagnetic signal sample through time-domain representation under online branching; Indicates the temperature coefficient; This represents the similarity function.

10. The multi-view depth-aligned unknown electromagnetic signal clustering method according to claim 9, characterized in that, in, This indicates that the samples in the current training batch have been assigned to the th ... The average probability of each category; , Indicates the first The unknown electromagnetic signal sample belongs to the first The probability of each category.

Images