An open set radio source identification method and system based on learnable phase space reconstruction

Abstract

The application belongs to the technical field of radiation source identification, and discloses an open set radiation source identification method and system based on a learnable phase space reconstruction, which comprises the following steps: acquiring an emission signal of a radiation source from the ground, and pre-processing the emission signal to obtain a received signal; constructing a structure perception network, reconstructing a phase space feature and extracting a space-time feature of the received signal based on the structure perception network, and constructing a feature embedding space based on the extracted space-time feature; identifying the received signal corresponding to the space-time feature in the feature embedding space by using an open set identification algorithm, obtaining a known category radiation source and an unknown category radiation source, and clustering the unknown category radiation source to obtain an unknown category cluster; and outputting the known category radiation source and the unknown category cluster corresponding to the ground radiation source. The application can further distinguish the unknown category radiation source without prior knowledge of the number of unknown radiation sources, and solve the identification and distinction problem of multiple unknown radiation sources in a satellite-to-ground scene.

Description

Technical Field

[0001] This invention relates to the field of radiation source identification technology, and in particular to an open-set radiation source identification method and system based on learnable phase space reconstruction. Background Technology

[0002] Specific radiation source identification systems achieve non-invasive, high-precision differentiation and identification of radiation sources by mining subtle physical layer fingerprint features caused by hardware manufacturing defects. In recent years, low-Earth orbit satellites have been widely used for the detection and location of ground-based radiation sources. Compared with ground-based receiving equipment, satellite-borne receivers can cover a wider ground area, thus helping to discover and identify unknown radiation sources earlier and more accurately.

[0003] However, ground-based radiation source signals are inevitably affected by various distortion factors in the Earth-to-satellite propagation channel during their propagation to the satellite receiver. Although deep learning-based convolutional neural networks have demonstrated superior performance in radiation source identification tasks, the signal representation methods they rely on are insufficient to fully characterize the nonlinear characteristics of the radiation source system, exhibiting weak robustness under complex Earth-to-satellite propagation channel conditions. Furthermore, the radiation source populations observed by Earth-orbiting satellites typically exhibit high heterogeneity, with frequent occurrences of unknown or newly emerging signal sources. Therefore, traditional closed-set identification methods that rely on the completeness of the training dataset are no longer sufficient to meet the requirements of satellite-based ground-based radiation source identification tasks.

[0004] Open set identification frameworks can effectively reduce the model's dependence on the completeness of the training dataset by introducing an unknown category detection mechanism. However, existing open set radiation source identification methods are still mainly focused on simple unknown category discrimination. Although a few studies have attempted to further distinguish unknown radiation sources, they usually rely on prior information about unknown radiation sources. Therefore, the application of open set radiation source identification methods in complex Earth-planet propagation environments is significantly limited.

[0005] Existing technologies still suffer from limitations in the application of open-collector radiation source identification methods in complex Earth-planet propagation environments; therefore, existing technologies need further improvement. Summary of the Invention

[0006] The technical problem to be solved by the present invention is that, in view of the defects of the prior art, the present invention provides an open set radiation source identification method and system based on learnable phase space reconstruction, so as to solve the problem that the application of existing open set radiation source identification methods is significantly limited in complex Earth-planet propagation environments.

[0007] The technical solution adopted by this invention to solve the technical problem is as follows:

[0008] In a first aspect, the present invention provides an open-set radiation source identification method based on learnable phase space reconstruction, comprising:

[0009] The transmitted signal from the radiation source on the ground is acquired, and the transmitted signal is preprocessed to obtain the received signal;

[0010] A structure-aware network is constructed, and the received signal is reconstructed using phase space features and extracted using spatiotemporal features based on the structure-aware network. A feature embedding space is constructed based on the extracted spatiotemporal features.

[0011] The open set recognition algorithm is used to identify the received signals corresponding to the spatiotemporal features in the feature embedding space, to obtain known category radiation sources and unknown category radiation sources, and to cluster the unknown category radiation sources to obtain unknown category clusters;

[0012] Output the known category of radiation source and the unknown category cluster corresponding to the radiation source.

[0013] In one implementation, the preprocessing of the transmitted signal to obtain the received signal includes:

[0014] The radiation source is modeled as a nonlinear dynamic system, and the emitted signal is regarded as the system state observation value of the nonlinear dynamic system.

[0015] The system state observations are uniformly sampled to obtain a discrete complex baseband sequence;

[0016] Obtain the propagation channel parameters of the transmitted signal and calculate the equivalent ground-to-satellite propagation channel response;

[0017] The received signal corresponding to the transmitted signal is obtained based on the discrete complex baseband sequence and the equivalent ground-to-satellite propagation channel response.

[0018] In one implementation, the structure-aware network includes: a learnable phase space reconstruction module, a spatial structure enhancement module, a temporal dynamic modeling module, and a discriminative feature projection module.

[0019] In one implementation, the step of reconstructing phase space features and extracting spatiotemporal features from the received signal based on the structure-aware network, and constructing a feature embedding space based on the extracted spatiotemporal features, includes:

[0020] The received signal is adaptively mapped into a phase space trajectory through the learnable phase space reconstruction module.

[0021] The spatial features of the phase space trajectory are extracted using the spatial structure enhancement module.

[0022] The spatial features are modeled in time series by the time dynamic modeling module to obtain the spatiotemporal features corresponding to the received signal;

[0023] The discriminative feature projection module maps the spatiotemporal features to the feature space, constructing a feature embedding space containing all spatiotemporal features.

[0024] In one implementation, the step of reconstructing phase space features and extracting spatiotemporal features from the received signal based on the structure-aware network, and constructing a feature embedding space based on the extracted spatiotemporal features, further includes:

[0025] A joint discriminative learning objective function is constructed based on a joint loss function; wherein, the joint loss function includes: cross-entropy loss function, center loss function, and multi-relation boundary loss function;

[0026] The joint discriminative learning objective function is used to constrain and optimize the feature embedding space.

[0027] In one implementation, the step of using an open-set recognition algorithm to identify the received signal corresponding to the spatiotemporal features in the feature embedding space, and obtaining known-category radiation sources and unknown-category radiation sources, includes:

[0028] Obtain the known features and classification confidence of known categories of radiation sources, and map the known features to the feature embedding space;

[0029] Calculate the deviation between the spatiotemporal feature and the known feature, and calculate the anomaly probability corresponding to the spatiotemporal feature based on the deviation;

[0030] Based on the classification confidence level and the anomaly probability, an unknown category confidence index is constructed;

[0031] Based on the classification confidence level and the unknown category confidence index, the spatiotemporal features are identified to obtain the spatiotemporal features corresponding to known category radiation sources and unknown category radiation sources;

[0032] Output the known and unknown category radiation sources corresponding to all spatiotemporal features in the feature embedding space.

[0033] In one implementation, clustering the unknown category of radiation sources includes:

[0034] Calculate the similarity between the spatiotemporal features corresponding to the unknown category of radiation sources;

[0035] Based on the similarity, unknown category radiation sources are clustered to obtain unknown category clusters.

[0036] Secondly, the present invention provides an open-set radiation source identification system based on learnable phase space reconstruction, comprising:

[0037] A signal receiving module is used to acquire the transmitted signal from a radiation source on the ground and preprocess the transmitted signal to obtain the received signal.

[0038] The feature extraction module is used to construct a structure-aware network, reconstruct phase space features and extract spatiotemporal features from the received signal based on the structure-aware network, and construct a feature embedding space based on the extracted spatiotemporal features.

[0039] The category identification module is used to identify the received signals corresponding to the spatiotemporal features in the feature embedding space using an open set identification algorithm, to obtain known category radiation sources and unknown category radiation sources, and to cluster the unknown category radiation sources to obtain unknown category clusters;

[0040] The result output module is used to output the known category of radiation source and the unknown category cluster corresponding to the radiation source.

[0041] Thirdly, the present invention provides a terminal, comprising: a processor and a memory, wherein the memory stores an open set radiation source identification program based on learnable phase space reconstruction, and the open set radiation source identification program based on learnable phase space reconstruction, when executed by the processor, is used to implement the operation of the open set radiation source identification method based on learnable phase space reconstruction as described in the first aspect.

[0042] Fourthly, the present invention also provides a computer-readable storage medium storing an open-set radiation source identification program based on learnable phase space reconstruction, wherein the open-set radiation source identification program based on learnable phase space reconstruction, when executed by a processor, is used to implement the operation of the open-set radiation source identification method based on learnable phase space reconstruction as described in the first aspect.

[0043] The present invention, by employing the above technical solution, has the following effects:

[0044] This invention provides an open-set radiation source identification method and system based on learnable phase space reconstruction, comprising: acquiring the transmitted signal from a radiation source on the ground; preprocessing the transmitted signal to obtain a received signal; constructing a structure-aware network; reconstructing phase space features and extracting spatiotemporal features from the received signal based on the structure-aware network; constructing a feature embedding space based on the extracted spatiotemporal features; using an open-set recognition algorithm to identify the received signal corresponding to the spatiotemporal features in the feature embedding space, obtaining known-category radiation sources and unknown-category radiation sources, and clustering the unknown-category radiation sources to obtain unknown-category clusters; and outputting the known-category radiation sources and the unknown-category clusters corresponding to the radiation sources. This invention can further distinguish unknown-category radiation sources without prior knowledge of the number of unknown radiation sources, solving the problem of identifying and distinguishing multiple unknown radiation sources in satellite-to-Earth scenarios. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0046] Figure 1 This is a flowchart of the open set radiation source identification method based on learnable phase space reconstruction in this invention.

[0047] Figure 2 This is a schematic diagram of the spatiotemporal structure sensing network structure in one implementation of the present invention.

[0048] Figure 3 This is a schematic diagram of the structure of an open set radiation source identification system based on learnable phase space reconstruction in one implementation of the present invention.

[0049] Figure 4 This is a functional schematic diagram of the terminal in one implementation of the present invention.

[0050] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0052] Exemplary methods

[0053] During the propagation of ground-based radiation source signals to satellite receivers, they are inevitably affected by various distortion factors in the Earth-to-satellite propagation channel, including path loss, shadowing fading, and Doppler shift. These effects are typically highly coupled and exhibit significant nonlinear characteristics, leading to a marked decrease in signal-to-noise ratio and strong nonstationarity of the signal waveform, thus severely weakening the performance of radiation source identification algorithms. In recent years, deep learning-based convolutional neural networks have demonstrated superior performance in radiation source identification tasks due to their excellent automatic feature learning capabilities. However, the signal representation forms relied upon by these methods, such as raw waveforms, time-spectrum diagrams, constellation diagrams, and higher-order statistical features, often fail to fully characterize the inherent nonlinear properties of the radiation source system. This insufficient representation capability limits the amount of information available to downstream learning models, resulting in weak robustness under complex Earth-to-satellite propagation channel conditions. Furthermore, due to the wide coverage and high speed of low-Earth orbit satellites, the observed radiation source groups typically exhibit high heterogeneity, with frequent occurrences of unknown or newly emerging signal sources. Therefore, traditional closed-set identification methods that rely on the completeness of the training dataset are no longer sufficient to meet the requirements of satellite-based ground-based radiation source identification tasks. Meanwhile, diverse spaceborne observation scenarios, such as different geographical regions and variable weather conditions, introduce significant differences in signal distribution, further increasing the complexity of radiation source feature modeling and extraction, blurring the discrimination boundaries between different radiation source signal categories, and thus hindering the accurate identification of unknown radiation source signals.

[0054] To address the aforementioned issues, the open set identification framework introduces an unknown category detection mechanism, effectively reducing the model's dependence on the completeness of the training dataset. Preliminary research has validated the feasibility of applying the open set paradigm to satellite-based ground-based radiation source identification tasks. However, existing open set radiation source identification methods primarily focus on simple unknown category discrimination. Although a few studies have attempted to further differentiate unknown radiation sources, they typically rely on prior information about the unknown sources, significantly limiting their application in the complex Earth-satellite propagation environment.

[0055] To address the above technical problems, this invention provides an open-set radiation source identification method based on learnable phase space reconstruction, comprising: acquiring the transmitted signal from a radiation source on the ground; preprocessing the transmitted signal to obtain a received signal; constructing a structure-aware network; reconstructing phase space features and extracting spatiotemporal features from the received signal based on the structure-aware network; constructing a feature embedding space based on the extracted spatiotemporal features; using an open-set recognition algorithm to identify the received signal corresponding to the spatiotemporal features in the feature embedding space, obtaining known-category radiation sources and unknown-category radiation sources, and clustering the unknown-category radiation sources to obtain unknown-category clusters; outputting the known-category radiation sources and the unknown-category clusters corresponding to the radiation sources; this invention can further distinguish unknown-category radiation sources without prior knowledge of the number of unknown radiation sources, solving the problem of identifying and distinguishing multiple unknown radiation sources in satellite-to-ground scenarios.

[0056] like Figure 1 As shown, this embodiment of the invention provides an open-set radiation source identification method based on learnable phase space reconstruction, comprising the following steps:

[0057] Step S100: Obtain the emission signal from the radiation source on the ground, and preprocess the emission signal to obtain the received signal.

[0058] It should be noted that the open set radiation source identification method based on learnable phase space reconstruction is applied to satellite specific emitter identification (SEI), that is, it is used by the satellite to identify the category of the transmitter's transmitted signal, determine whether the transmitter belongs to a known type of radiation source or an unknown type of radiation source, and further distinguish transmitters belonging to unknown type of radiation sources.

[0059] In this embodiment, the transmitter is regarded as a radiation source, and the signal emitted by the transmitter is the transmitted signal. The transmitted signal is preprocessed to obtain a complex baseband signal, which facilitates frequency domain analysis and feature extraction.

[0060] Specifically, in one implementation of this embodiment, step S100 includes the following steps:

[0061] Step S101: Obtain the emission signal from the radiation source on the ground.

[0062] In this embodiment, the satellite receives the transmission signal from a radiation source on the ground, wherein the radiation source on the ground is a radiation source of the type to be identified.

[0063] Step S102: The radiation source is modeled as a nonlinear dynamic system, and the emitted signal is regarded as the system state observation value of the nonlinear dynamic system.

[0064] It should be noted that the hardware differences of radiation sources can be reflected by the transmitted signals, but the receiving end (satellite) cannot directly observe the internal state of the radiation source. It can only infer its implicit dynamic evolution process through the externally observable I / Q signals.

[0065] In this embodiment, based on the inherent characteristics of the radiation source, the ground radiation source is modeled as a nonlinear dynamic system, and the internal state vector of the corresponding nonlinear dynamic system can be expressed as:

[0066] ;

[0067] in, Indicates time.

[0068] In this embodiment, the internal state of the nonlinear dynamic system satisfies:

[0069] .

[0070] Furthermore, if the transmitted signal is considered as an observation of the system state of the nonlinear dynamic system, the corresponding formula is as follows:

[0071] ;

[0072] in, This is an excitation signal.

[0073] Step S103: Uniformly sample the system state observations to obtain a discrete complex baseband sequence.

[0074] In this embodiment, the system state observations are uniformly sampled to obtain a discrete complex baseband sequence, and the corresponding calculation formula is as follows:

[0075] , .

[0076] Step S104: Obtain the propagation channel parameters of the transmitted signal and calculate the equivalent ground-to-satellite propagation channel response.

[0077] In this embodiment, the propagation channel parameters of the transmitted signal include the instantaneous propagation distance of the radiation source, shadowing fading, and Doppler frequency shift. The Earth-to-satellite propagation channel is mainly dominated by large-scale effects, and its equivalent Earth-to-satellite propagation channel response can be decomposed into the instantaneous propagation distance of the radiation source. Shadow decay Time-varying phase caused by the Doppler effect The formula for calculating the Earth-to-Satellite propagation channel response is as follows:

[0078] .

[0079] Specifically, the instantaneous propagation distance between the radiation source corresponding to the transmitted signal and the satellite is obtained, that is, the instantaneous propagation distance of the radiation source. Based on instantaneous propagation distance The formula for calculating path loss is as follows:

[0080] ;

[0081] in, For reference distance, This is the path loss index determined by the propagation environment.

[0082] Furthermore, shadow decay For log-normal shadowing fading, it satisfies:

[0083] ;

[0084] in, and Let represent the mean and variance of the shadow fading component in the logarithmic domain, respectively.

[0085] The formula for calculating the Doppler effect in the baseband domain is as follows:

[0086] ;

[0087] in, This indicates Doppler frequency shift.

[0088] Step S105: Based on the discrete complex baseband sequence and the equivalent ground-to-satellite propagation channel response, obtain the received signal corresponding to the transmitted signal.

[0089] Specifically, in the scenario of ground-based radiation source identification based on low Earth orbit satellites, the complex baseband signal received by the onboard receiver during its on-orbit operation can be represented by the following formula:

[0090] .

[0091] in, This represents the discrete complex baseband sequence corresponding to the received signal from a radiation source on the ground. For the equivalent Earth-to-satellite propagation channel response, It is additive white Gaussian noise.

[0092] In this embodiment, additive white Gaussian noise satisfy:

[0093] .

[0094] like Figure 1As shown, this embodiment of the invention provides an open-set radiation source identification method based on learnable phase space reconstruction, comprising the following steps:

[0095] Step S200: Construct a structure-aware network, reconstruct phase space features and extract spatiotemporal features from the received signal based on the structure-aware network, and construct a feature embedding space based on the extracted spatiotemporal features.

[0096] Specifically, in one implementation of this embodiment, step S200 includes the following steps:

[0097] Step S201: Construct a structure-aware network, which includes: a learnable phase space reconstruction module, a spatial structure enhancement module, a temporal dynamic modeling module, and a discriminative feature projection module.

[0098] In this embodiment, based on adaptive dynamic trajectory embedding, a spatiotemporal structure-aware representation is constructed to jointly model the spatial structure features and temporal evolution characteristics contained in the received signal, thereby obtaining a stable and discriminative feature representation.

[0099] In this embodiment, the constructed structure-aware network consists of two parts: dynamic trajectory embedding and spatiotemporal structure-aware network. The main body of the dynamic trajectory embedding part is a learnable phase space reconstruction module, which is used to adaptively reconstruct the phase space of the original IQ (In-phase and Quadrature) sequence to obtain the reconstructed phase space trajectory. The spatiotemporal structure-aware network part is used to extract the spatiotemporal features of the phase space trajectory and map the spatiotemporal features to the feature embedding space to realize the construction of the feature embedding space based on the spatiotemporal features.

[0100] like Figure 2 The diagram shown is a schematic of the spatiotemporal structure perception network in the spatiotemporal perception network constructed in this embodiment. The spatiotemporal structure perception network consists of three parts: a spatial structure enhancement module, a temporal dynamic modeling module, and a discriminative feature projection module, which are used for spatial structure enhancement, temporal dynamic modeling, and discriminative feature projection, respectively.

[0101] Step S202: The received signal is adaptively mapped into a phase space trajectory through the learnable phase space reconstruction module.

[0102] In this embodiment, to overcome the limitations of traditional signal representation methods, phase space reconstruction theory is introduced, reformulating the problem of identifying specific radiation sources as a nonlinear dynamical system discrimination problem. Within this modeling framework, each radiation source is considered a dynamic system evolving over time, and its emitted waveform is regarded as the output of the system's internal state evolution. Based on this, a phase space learning method is further proposed to achieve adaptive learning and optimization of the phase space dynamical representation.

[0103] Specifically, the Dynamic Trajectory Embedding (DTE) method is used, introducing the concept of phase space reconstruction. Observations at different times and their delays are combined into phase space coordinates, thereby effectively characterizing the nonlinear dynamics within the radiation source. Unlike traditional phase space reconstruction methods that use fixed time delay parameters, this embodiment introduces a learnable delay weighting mechanism under the premise of a fixed embedding dimension. Multiple candidate time delays are adaptively fused, allowing the phase space representation to dynamically adjust according to signal characteristics. The phase space embedding process can be uniformly represented as follows:

[0104] ;

[0105] in, This represents the candidate delay set consisting of the current sample and its multiple delayed samples. Indicates that for the first The delayed weight vector of each embedded component. By implementing phase space embedding, a phase space representation that is discriminative of the dynamic characteristics of radiation sources can be obtained, providing a feature basis for subsequent radiation source identification and discrimination of unknown radiation sources.

[0106] Step S203: Extract the spatial features of the phase space trajectory through the spatial structure enhancement module.

[0107] Specifically, the spatial structure enhancement module is used to characterize the geometric relationships within the local neighborhood of the phase space trajectory. It extracts the structural correlation between adjacent embedded components in a multi-scale manner and assigns adaptive weights to different feature components, thereby highlighting the structural information that is highly correlated with the hardware characteristics of the radiation source and suppressing the interference of redundant or noise components on the feature representation.

[0108] In this embodiment, after embedding multi-dimensional temporal features into the phase space trajectory, it is input into the spatial structure enhancement module, such as... Figure 2 As shown, the space structure enhancement module includes:

[0109] A 1×1 two-dimensional convolutional layer with 32 channels is used to extract local spatial structure information;

[0110] A 3×1 two-dimensional convolutional layer with 32 channels is used to extract local spatial structure information;

[0111] The channel stitching layer is used to stitch together extracted local spatial structure information at different scales along the channel dimension.

[0112] Channel attention layer, used for adaptive recalibration of spatial features.

[0113] In this embodiment, the channel attention layer consists of a two-level mapping of fully connected layers and also includes a channel-dimensional weighted fusion module, thereby adaptively enhancing the discriminative spatial structure features that are highly correlated with the non-ideal characteristics of the radiation source hardware and effectively suppressing redundant or noisy channel information.

[0114] Step S204: The phase space trajectory is modeled in time series by the time dynamic modeling module to obtain the spatiotemporal characteristics corresponding to the complex baseband signal.

[0115] Specifically, the temporal dynamic modeling module is used to describe the overall evolution of the phase space trajectory in the time dimension. This module performs time series modeling on the spatially enhanced phase space trajectory, i.e., the spatial features, fusing information from different time locations to capture the evolutionary behavior of the nonlinear dynamic system characteristics of the radiation source over long time scales. This avoids the instability caused by relying solely on instantaneous features and improves the robustness of the feature representation to channel disturbances and noise.

[0116] In this embodiment, the output of the spatial structure enhancement module is used as the input of the time dynamic modeling module, such as... Figure 2 As shown, the time dynamic modeling module includes:

[0117] The temporal downsampling layer is used to downsample the spatial features of the input in the temporal dimension, thereby reducing temporal redundancy.

[0118] The temporal flattening layer is used to compress the sampling results in the time dimension, turning them into a fixed-length feature vector, reducing temporal redundancy and compressing computational complexity.

[0119] Fully connected layers are used to perform feature mapping;

[0120] The position encoding layer is used for sinusoidal position encoding. First, it generates the corresponding position index information based on the time series index, and then maps the discrete time position into a continuous position encoding vector through sinusoidal position transformation. Finally, it concatenates the obtained position encoding with the original features in the feature dimension.

[0121] The Mamba feature extraction layer contains three Mamba modules for deep modeling of temporal features.

[0122] Specifically, each Mamba module takes temporal features as input. First, it uses one-dimensional convolution operators 3×1 and 1×1 to linearly transform and mix the features within the local temporal neighborhood to enhance short-term dynamic responses. Then, LayerNorm is used to standardize the feature distribution, improving model training stability and suppressing the impact of temporal amplitude fluctuations. By cascading multiple Mamba modules, the long-term dynamic evolution characteristics of phase space trajectories can be fully characterized while maintaining linear computational complexity.

[0123] Step S205: The spatiotemporal features are mapped to the feature space through the discriminative feature projection module to construct a feature embedding space containing all spatiotemporal features.

[0124] Specifically, the discriminative feature projection module is used to map the phase space trajectory, i.e., the spatiotemporal feature representation, after spatiotemporal modeling to a compact and normalized feature space. This makes the features generated by the same radiation source more concentrated in the feature embedding space, while maintaining sufficient discriminability between the features corresponding to different radiation sources. This provides a reliable feature basis for subsequent identification of known transmitters and discrimination of unknown transmitters.

[0125] In this embodiment, the spatiotemporal features are mapped to the feature space by the discriminative feature projection module. The spatiotemporal features are high-level features obtained after spatiotemporal fusion, such as... Figure 2 As shown, the discriminative feature projection module includes:

[0126] A global average pooling layer is used to aggregate information along the time dimension;

[0127] The L2 normalization layer is used to constrain the feature scale through L2 normalization to form a discriminative feature vector with clear structure and good separability.

[0128] The Softmax classification layer is used for known class identification and provides stable feature representations for subsequent open-set decision and unknown class analysis.

[0129] Overall, the process by which the spatiotemporal structure-aware network in this embodiment extracts the spatiotemporal features of the phase space trajectory and maps these features to the feature embedding space can be abstracted as a mapping relationship from the phase space trajectory to the feature embedding vector, as shown in the following formula:

[0130] ;

[0131] in, This represents the spatiotemporal structure-aware mapping function, which is composed of spatial structure enhancement, temporal dynamic modeling, and discriminative projection. This is the final feature embedding vector.

[0132] Step S206: Construct a joint discriminative learning objective function based on the joint loss function; wherein the joint loss function includes: cross-entropy loss function, center loss function and multi-relation boundary loss function.

[0133] In this embodiment, to ensure that the learned feature embeddings possess both good class discriminativeness and a stable geometric structure, a discriminative learning objective function composed of multiple loss functions is used to constrain and optimize the feature representation learning process, specifically by constraining and optimizing the feature embedding space.

[0134] Specifically, the joint loss function in this embodiment consists of cross-entropy loss, center loss, and Masked Region Modeling (MRM) loss, which constrain the embedding space from three aspects: classification consistency, intra-class compactness, and inter-class separation. Cross-entropy loss constrains the consistency between the output radiation source category prediction and the known radiation source categories, ensuring that the learned features possess basic supervised classification capabilities. Center loss constrains the features of samples from the same radiation source category to cluster towards their corresponding category center, effectively reducing intra-class feature dispersion and improving the robustness of feature representation to noise and channel disturbances. MRM loss explicitly models the distance relationships between different radiation source categories, introducing category-level margin constraints in the embedding space to ensure sufficient separation of feature distributions corresponding to different radiation sources, thereby forming clear category decision boundaries.

[0135] In this embodiment, the joint discriminative learning objective function can be calculated using the following formula:

[0136] ;

[0137] in, Represents cross-entropy loss, Indicates the loss at the center. Indicates multi-relation boundary loss, and These are weighting coefficients used to balance the relative roles of different loss terms in the overall learning process.

[0138] Step S207: Use the joint discriminative learning objective function to constrain and optimize the feature embedding space.

[0139] In this embodiment, the joint discriminative learning objective function is used to constrain and optimize the feature embedding space. By jointly optimizing the above loss function, a feature embedding space that is compact within classes, separate between classes, and suitable for open set recognition can be constructed while ensuring the accuracy of known radiation source identification.

[0140] In this embodiment, by introducing the MRM loss function, the compactness of intra-class features and the separability between classes can be enhanced, thereby enabling the network to simultaneously capture discriminative spatial structural features and temporal dependencies in the phase space representation.

[0141] In one implementation of this embodiment, the phase space reconstruction and dynamic representation can also use other implementation forms, such as variable time delay, non-uniform embedding, statistical trajectory features and other time series modeling networks, to achieve the learning of the dynamic characteristics of the radiation source.

[0142] like Figure 1As shown, this embodiment of the invention provides an open-set radiation source identification method based on learnable phase space reconstruction, comprising the following steps:

[0143] Step S300: Use the open set recognition algorithm to identify radiation sources in the feature embedding space, obtain known category radiation sources and unknown category radiation sources, and cluster the unknown category radiation sources to obtain unknown category clusters.

[0144] In this embodiment, after obtaining the discriminative feature embedding vector Based on this, a radiation source identification algorithm for open set scenarios is proposed. This algorithm is used to identify known radiation sources, reject unknown radiation sources, and automatically distinguish unknown radiation sources when they exist. Specifically, the open set identification algorithm is used to identify radiation sources in the feature embedding space to obtain known category radiation sources and unknown category radiation sources. Then, the unknown category radiation sources are clustered to obtain unknown category clusters.

[0145] Specifically, in one implementation of this embodiment, step S300 includes the following steps:

[0146] Step S301: Obtain the known features and classification confidence of known category radiation sources, and map the known features into the feature embedding space.

[0147] In this embodiment, known features and classification confidence scores of known categories of radiation sources are obtained. Mapping these known features to a feature embedding space allows for the calculation of the deviation between the spatiotemporal features of the received signal to be judged and the known features in the feature embedding space. Furthermore, the deviation score and classification confidence score are combined to determine whether the received signal corresponding to the spatiotemporal features originates from a known category of radiation source, and if so, which category of known radiation source it belongs to.

[0148] Step S302: Calculate the deviation between the spatiotemporal feature and the known feature, and calculate the anomaly probability corresponding to the spatiotemporal feature based on the deviation.

[0149] In this embodiment, during the radiation source category discrimination stage, the classification confidence is recalibrated based on the OpenMax concept. To address the heteroscedasticity issue in the feature embedding space of different radiation source category feature distributions, Mahalanobis distance is introduced to measure the deviation between the sample and each known radiation source category. Specifically, the deviation between the spatiotemporal features and known features is calculated, and the corresponding calculation formula is as follows:

[0150] ;

[0151] in, Indicates the first The feature centers in the feature embedding space for known features of a known category of radiation source. Let be the covariance matrix of this known feature.

[0152] Furthermore, for each known category of radiation source, extreme value modeling is performed on the tail samples of the Mahalanobis distance distribution, and the Weibull distribution is used to characterize the probability of the sample deviating from the known distribution, thus obtaining the corresponding anomaly probability. .

[0153] Step S303: Construct an unknown category confidence index based on the classification confidence level and the anomaly probability.

[0154] In this embodiment, based on anomaly probability With classification confidence To construct a confidence index for the unknown category, the corresponding formula is as follows:

[0155] ;

[0156] in, This indicates that the spatiotemporal features are determined to be the first. Classification confidence for each known category This is the scaling factor. This represents the total number of known categories corresponding to all known features in the feature embedding space.

[0157] In this embodiment, the actual meaning of the unknown category confidence index is: an index used to measure whether spatiotemporal features belong to the known category corresponding to any known category radiation source.

[0158] Step S304: Based on the classification confidence level and the unknown category confidence level, the spatiotemporal features are identified to obtain the spatiotemporal features corresponding to known category radiation sources and the spatiotemporal features corresponding to unknown category radiation sources.

[0159] In this embodiment, the spatiotemporal features are identified based on the classification confidence score and the unknown category confidence score, resulting in spatiotemporal features corresponding to known category radiation sources and unknown category radiation sources. Taking into account both the geometric deviation of samples in the embedding space and the overall classification uncertainty, a dual-threshold decision criterion is used to achieve open set identification, resulting in:

[0160] ;

[0161] in, This represents a set of unclassified radiation sources. and This is a preset threshold.

[0162] Step S305: Output the known category radiation sources and unknown category radiation sources corresponding to all spatiotemporal features in the feature embedding space.

[0163] In this embodiment, based on the discrimination condition in step S304, we have:

[0164] For any spatiotemporal feature, when the spatiotemporal feature is related to the first... The maximum anomaly probability of each known category is greater than or equal to a preset threshold. When its unknown category confidence index is greater than or equal to a preset threshold If the result is true, the output of the discrimination result corresponding to the spatiotemporal feature is an unknown category radiation source; otherwise, the output of the discrimination result corresponding to the spatiotemporal feature is a known category radiation source, and the output of its known category is also provided.

[0165] Furthermore, the output feature embedding space contains known and unknown category radiation sources corresponding to all spatiotemporal features.

[0166] Step S306: Calculate the similarity between the spatiotemporal features corresponding to the unknown category of radiation sources.

[0167] In this embodiment, for all unknown category radiation sources in all outputs, clustering of unknown categories is used to further distinguish the unknown category radiation sources.

[0168] Specifically, the spatiotemporal features of radiation sources classified as unknown categories are remapped into a new feature space, constructing an unknown category feature embedding space, thus forming an unknown sample set containing the spatiotemporal features of all unknown category radiation sources. .in, This represents the total number of spatiotemporal characteristics of radiation sources of unknown categories.

[0169] Furthermore, to characterize the local density structure in the embedding space, cross-distance based on Mahalanobis distance is used to measure the similarity between unknown samples.

[0170] for Any two unknown samples and ( and The similarity calculation formula is as follows:

[0171] ;

[0172] in, Indicates the sample to its The Mahalanobis distance between the i-th nearest neighbor samples ( ), used to characterize the local density level of the region where the sample is located.

[0173] Step S307: Cluster the unknown category radiation sources based on the similarity to obtain unknown category clusters.

[0174] In this embodiment, unknown samples are clustered based on similarity, that is, unknown category radiation sources are clustered to obtain unknown category clusters. Specifically, a weighted graph structure of unknown samples is constructed based on the aforementioned distance, and compact connectivity is obtained by extracting the Minimum Spanning Tree (MST), resulting in:

[0175] ;

[0176] Subsequently, by gradually removing edges with larger distances in the spanning tree, unknown samples are divided at different density levels, and the clustering results are automatically determined based on the cluster stability criterion. This enables the automatic differentiation of multiple unknown transmitters without prior knowledge of the number of unknown radiation sources.

[0177] like Figure 1 As shown, this embodiment of the invention provides an open-set radiation source identification method based on learnable phase space reconstruction, comprising the following steps:

[0178] Step S400: Output the known category radiation source and the unknown category cluster corresponding to the radiation source.

[0179] In this embodiment, the final output results are known category radiation sources and unknown category clusters, that is, the output of the known category radiation source to which the emitted signal belongs, or the unknown category cluster to which the emitted signal belongs, thereby realizing the category identification of radiation sources.

[0180] In this embodiment, the radiation source identification algorithm for open set scenarios uses clearer feature boundaries to further distinguish unknown categories without requiring any prior information on the type or quantity of unknown transmitters / radiation sources. This effectively solves the problem of identifying and distinguishing multiple unknown transmitters in actual satellite-to-ground SEI scenarios.

[0181] In one implementation of this embodiment, the method for identifying and distinguishing unknown categories can be implemented by different statistical decisions, distance metrics, or unsupervised clustering methods, such as adaptive thresholds, probabilistic modeling, or graph structure clustering, as long as the discovery and distinction of unknown transmitters is completed under the condition that there is no prior knowledge of the number of unknown categories.

[0182] This embodiment achieves the following technical effects through the above technical solution:

[0183] This embodiment of the open-set radiation source identification method based on learnable phase space reconstruction elevates transmitter signal identification from a system modeling perspective, transforming it from a traditional static feature matching problem into a comprehensive discrimination problem involving the signal generation mechanism and its temporal evolution characteristics. This overcomes the technical bottleneck of existing methods' limited identification capabilities under complex channel conditions and unknown transmitter scenarios. This design emphasizes characterizing the intrinsic structure and evolution of the signal, rather than relying on specific signal forms or prior parameter settings.

[0184] The open-set radiation source identification method based on learnable phase space reconstruction in this embodiment also has good versatility and scalability. Its core idea can be extended to signal identification and classification tasks under different communication systems, signal bandwidths, and sensing platform conditions. At the same time, the proposed structured representation and open-set discrimination framework has modular characteristics, which facilitates adjustment and expansion according to specific application requirements, providing ample design space for subsequent functional enhancements and engineering implementation.

[0185] Exemplary device

[0186] Based on the above embodiments, the present invention also provides an open-set radiation source identification system based on learnable phase space reconstruction, such as... Figure 3 As shown, the open-set radiation source identification system based on learnable phase space reconstruction in this embodiment includes:

[0187] A signal receiving module is used to acquire the transmitted signal from a radiation source on the ground and preprocess the transmitted signal to obtain the received signal.

[0188] The feature extraction module is used to construct a structure-aware network, reconstruct phase space features and extract spatiotemporal features from the received signal based on the structure-aware network, and construct a feature embedding space based on the extracted spatiotemporal features.

[0189] The category identification module is used to identify the received signals corresponding to the spatiotemporal features in the feature embedding space using an open set identification algorithm, to obtain known category radiation sources and unknown category radiation sources, and to cluster the unknown category radiation sources to obtain unknown category clusters;

[0190] The result output module is used to output the known category of radiation source and the unknown category cluster corresponding to the radiation source.

[0191] In this embodiment, the feature extraction module is a structure-aware representation network, which consists of two parts: dynamic trajectory embedding and spatiotemporal structure-aware network. The main part of the dynamic trajectory embedding part is a learnable phase space reconstruction module, which is used to reconstruct the input received signal into a phase space trajectory; the spatiotemporal structure-aware network consists of three parts: a spatial structure enhancement module, a temporal dynamic modeling module, and a discriminative feature projection module.

[0192] In this embodiment, the category identification module is OPEN-S³. This module takes the spatiotemporal features extracted by the feature extraction module as input. First, it constructs a statistical discriminant model for known categories, selects the t samples farthest from the category center as tail samples, and fits a Weibull cumulative distribution function based on the distance distribution of these samples to characterize the boundary characteristics of the category in the feature space. For the sample to be identified, it calculates its distance to each known category center and obtains the corresponding tail probability through the Weibull model. This probability is then fused with the classification confidence score to form an unknown category confidence index, thereby determining the category. The algorithm determines whether a sample deviates from the distribution of all known categories. If the decision condition meets the threshold constraint, the sample is rejected as an unknown category; otherwise, it is directly output as the corresponding known category. Subsequently, all rejected unknown samples are uniformly mapped into the feature space. A structural similarity measure between samples is constructed based on the cross-distance, and the HDBSCAN density clustering method is used to perform unsupervised clustering of unknown samples. This automatically discovers multiple structurally stable unknown category clusters without any prior knowledge of the number of unknown categories, ultimately achieving reliable identification of known category radiation sources and further differentiation and classification of unknown category radiation sources.

[0193] Based on the above embodiments, the present invention also provides a terminal, the principle block diagram of which can be as follows: Figure 4 As shown.

[0194] The terminal includes: a processor, a memory, an interface, a display screen, and a communication module connected via a system bus; wherein, the processor of the terminal provides computing and control capabilities; the memory of the terminal includes a computer-readable storage medium and internal memory; the computer-readable storage medium stores an operating system and computer programs; the internal memory provides an environment for the operation of the operating system and computer programs in the computer-readable storage medium; the interface is used to connect to external devices; the display screen is used to display relevant information; and the communication module is used to communicate with a cloud server or other devices.

[0195] When executed by the processor, this computer program is used to implement the operation of an open-set radiation source identification method based on learnable phase space reconstruction.

[0196] It will be understood by those skilled in the art that Figure 4 The schematic diagram shown is merely a partial structural diagram related to the present invention and does not constitute a limitation on the terminal to which the present invention is applied. A specific terminal may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0197] In one embodiment, a terminal is provided, comprising: a processor and a memory, the memory storing an open-set radiation source identification program based on learnable phase space reconstruction, the open-set radiation source identification program based on learnable phase space reconstruction being executed by the processor to implement the operation of the open-set radiation source identification method based on learnable phase space reconstruction as described above.

[0198] In one embodiment, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores an open-set radiation source identification program based on learnable phase space reconstruction, the open-set radiation source identification program based on learnable phase space reconstruction being executed by a processor to implement the operation of the above-described open-set radiation source identification method based on learnable phase space reconstruction.

[0199] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, database, or other media used in the embodiments provided by this invention can include both non-volatile and volatile memory.

[0200] In summary, this invention provides an open-set radiation source identification method and system based on learnable phase space reconstruction, comprising: acquiring received signals from radiation sources on the ground; preprocessing the received signals to obtain complex baseband signals; using a sensing network to reconstruct and extract phase space features from the complex baseband signals to obtain spatiotemporal features corresponding to the complex baseband signals, and constructing a feature embedding space containing all spatiotemporal features; using an open-set recognition algorithm to identify radiation sources in the feature embedding space to obtain known-category radiation sources and unknown-category radiation sources; clustering the unknown-category radiation sources, and outputting known-category radiation sources, unknown-category clusters, and radiation sources contained in the clusters; this invention can further distinguish unknown-category radiation sources without prior knowledge of the number of unknown radiation sources, solving the problem of identifying and distinguishing multiple unknown radiation sources in satellite-to-Earth scenarios.

[0201] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for identifying open-set radiation sources based on learnable phase space reconstruction, characterized in that, include: The transmitted signal from the radiation source on the ground is acquired, and the transmitted signal is preprocessed to obtain the received signal; A structure-aware network is constructed, and the received signal is reconstructed using phase space features and extracted using spatiotemporal features based on the structure-aware network. A feature embedding space is constructed based on the extracted spatiotemporal features. The open set recognition algorithm is used to identify the received signals corresponding to the spatiotemporal features in the feature embedding space, to obtain known category radiation sources and unknown category radiation sources, and to cluster the unknown category radiation sources to obtain unknown category clusters; Output the known category of radiation source and the unknown category cluster corresponding to the radiation source; The structure-aware network includes: a learnable phase space reconstruction module, a spatial structure enhancement module, a temporal dynamic modeling module, and a discriminative feature projection module; The process of reconstructing phase space features and extracting spatiotemporal features from the received signal based on the structure-aware network, and constructing a feature embedding space based on the extracted spatiotemporal features, includes: The received signal is adaptively mapped into a phase space trajectory through the learnable phase space reconstruction module. The spatial features of the phase space trajectory are extracted using the spatial structure enhancement module. The spatial features are modeled in time series by the time dynamic modeling module to obtain the spatiotemporal features corresponding to the received signal; The discriminative feature projection module maps the spatiotemporal features to the feature space, thereby constructing a feature embedding space that contains all spatiotemporal features. The step of adaptively mapping the received signal into a phase space trajectory through the learnable phase space reconstruction module includes: using an adaptive dynamic trajectory embedding method, introducing the concept of phase space reconstruction, and combining different times and corresponding delayed observations into phase space coordinates; The spatial structure enhancement module includes: a 1×1 two-dimensional convolutional layer, a 3×1 two-dimensional convolutional layer, a channel splicing layer, and a channel attention layer; The time dynamic modeling module includes: a time downsampling layer, a time dimension flattening layer, a fully connected layer, a position encoding layer, and a Mamba feature extraction layer; The step of reconstructing phase space features and extracting spatiotemporal features from the received signal based on the structure-aware network, and constructing a feature embedding space based on the extracted spatiotemporal features, further includes: A joint discriminative learning objective function is constructed based on a joint loss function; wherein, the joint loss function includes: cross-entropy loss function, center loss function, and multi-relation boundary loss function; The joint discriminative learning objective function is used to constrain and optimize the feature embedding space.

2. The method for identifying open-set radiation sources based on learnable phase space reconstruction according to claim 1, characterized in that, The process of preprocessing the transmitted signal to obtain the received signal includes: The radiation source is modeled as a nonlinear dynamic system, and the emitted signal is regarded as the system state observation value of the nonlinear dynamic system. The system state observations are uniformly sampled to obtain a discrete complex baseband sequence; Obtain the propagation channel parameters of the transmitted signal and calculate the equivalent ground-to-satellite propagation channel response; The received signal corresponding to the transmitted signal is obtained based on the discrete complex baseband sequence and the equivalent ground-to-satellite propagation channel response.

3. The method for identifying open-set radiation sources based on learnable phase space reconstruction according to claim 1, characterized in that, The step of using an open-set recognition algorithm to identify the received signal corresponding to the spatiotemporal features in the feature embedding space, and obtaining known-class radiation sources and unknown-class radiation sources, includes: Obtain the known features and classification confidence of known categories of radiation sources, and map the known features to the feature embedding space; Calculate the deviation between the spatiotemporal feature and the known feature, and calculate the anomaly probability corresponding to the spatiotemporal feature based on the deviation; Based on the classification confidence level and the anomaly probability, an unknown category confidence index is constructed; Based on the classification confidence level and the unknown category confidence index, the spatiotemporal features are identified to obtain the spatiotemporal features corresponding to known category radiation sources and unknown category radiation sources; Output the known and unknown category radiation sources corresponding to all spatiotemporal features in the feature embedding space.

4. The method for identifying open-set radiation sources based on learnable phase space reconstruction according to claim 1, characterized in that, The clustering of the unknown category of radiation sources includes: Calculate the similarity between the spatiotemporal features corresponding to the unknown category of radiation sources; Based on the similarity, unknown category radiation sources are clustered to obtain unknown category clusters.

5. An open-set radiation source identification system based on learnable phase space reconstruction, used to implement the open-set radiation source identification method based on learnable phase space reconstruction as described in any one of claims 1-4, characterized in that, include: A signal receiving module is used to acquire the transmitted signal from a radiation source on the ground and preprocess the transmitted signal to obtain the received signal. The feature extraction module is used to construct a structure-aware network, reconstruct phase space features and extract spatiotemporal features from the received signal based on the structure-aware network, and construct a feature embedding space based on the extracted spatiotemporal features. The category identification module is used to identify the received signals corresponding to the spatiotemporal features in the feature embedding space using an open set identification algorithm, to obtain known category radiation sources and unknown category radiation sources, and to cluster the unknown category radiation sources to obtain unknown category clusters; The result output module is used to output the known category of radiation source and the unknown category cluster corresponding to the radiation source.

6. A terminal, characterized in that, include: The processor and memory, wherein the memory stores an open-set radiation source identification program based on learnable phase space reconstruction, the open-set radiation source identification program based on learnable phase space reconstruction being executed by the processor to implement the operation of the open-set radiation source identification method based on learnable phase space reconstruction as described in any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an open-set radiation source identification program based on learnable phase space reconstruction, which, when executed by a processor, is used to implement the operation of the open-set radiation source identification method based on learnable phase space reconstruction as described in any one of claims 1-4.

Images