A method and related device for identifying individual imbalanced radiation sources based on temporal attention.

Abstract

This application discloses a method and related apparatus for identifying individuals from imbalanced radiation sources based on temporal attention, relating to the field of automatic identification. The method includes: extracting shallow temporal features from the acquired radiation source signal to obtain deep temporal high-order features; fusing multi-scale features from the deep temporal high-order features to obtain fused features; designing an identity-aware temporal attention module to generate identity-adaptive time-step attention weights; multiplying the fused features and the identity-adaptive time-step attention weights element-wise to obtain weighted features; aggregating global identity features from the weighted features to obtain aggregated global identity features; mapping the aggregated global identity features to obtain probability distribution features; and inputting the probability distribution features into a classifier to obtain the probability of the radiation source signal corresponding to an individual from the radiation source. This application can solve the problems of significant long-tailed data distribution and high-dimensionality sampling features in radiation source individual identification.

Description

Technical Field

[0001] This application relates to the field of automatic identification technology, and in particular to a method and related apparatus for identifying individuals of unbalanced radiation sources based on temporal attention. Background Technology

[0002] With the continuous expansion of global maritime transportation, the identification of non-cooperative targets has become a core requirement for traffic control and security. These targets often evade surveillance by tampering with identity broadcast information from Automatic Identification Systems (AIS) and Automatic Dependent Surveillance-Broadcast (ADS-B) systems, rendering target data directly interpreted by AIS / ADS-B systems unreliable. However, AIS, ADS-B, and other communication systems contain unique fingerprint features in their radio frequency signals, generated by inherent defects in the transmitter hardware circuitry (such as oscillator phase noise and power amplifier nonlinear distortion). These physical layer features are unaffected by upper-layer data tampering, providing crucial evidence for tracing the true identity of non-cooperative targets. Therefore, as a core technology for extracting and utilizing these features, Specific Emitter Identification (SEI) not only plays an irreplaceable role in civilian scenarios such as maritime traffic management and aviation surveillance, but also has significant research value and application prospects in fields such as spectrum management and cybersecurity.

[0003] For SEI (Search Engine Imaging) tasks, the academic community has proposed various methods, which can be broadly classified into two categories: traditional machine learning methods and deep learning methods. Traditional methods mainly rely on manually designed statistical features. For example, Gök et al. proposed a recognition method based on variational mode decomposition, which decomposes the signal envelope and instantaneous frequency into multiple modes and completes device differentiation by constructing a feature set; Sun et al., using manifold learning, first revealed the inherent low-dimensional nonlinear manifold structure of radio frequency signals and systematically analyzed the impact mechanism of the robustness and uniqueness of this structure on SEI performance; Xing et al. designed a robust radar identification scheme based on RFF (Radio Frequency Filter), first classifying the pulse operating mode and suppressing noise interference, and then achieving device identification by fusing transient features and modulation features; Sun et al., by verifying the locality and non-uniformity of radio frequency fingerprints, found that unintentional modulation of pulses is concentrated in a specific subspace of the RFF distribution region, and then proposed an SEI method based on dispersion feature index and radiation source specific information index. These traditional methods require a lot of human time to extract features and are difficult to adapt to nonlinear features caused by hardware defects, often resulting in limited performance in complex real-world scenarios.

[0004] With breakthroughs in deep learning technology in the field of automatic feature learning, SEI methods are gradually evolving towards a data-driven approach. For example, Tan Kaiwen et al. proposed IC-SGAN, an SEI method that integrates cost-sensitive learning and semi-supervised GAN, using adversarial generation for data augmentation to improve recognition accuracy under class imbalance and missing labels; Ding et al. proposed extracting bispectral features of USRP signals based on convolutional neural networks, and achieving radiation source identification after supervised dimensionality reduction, effectively mining fingerprint information in the higher-order statistical characteristics of the signal; Pan et al. converted the signal into a grayscale image through Hilbert transform, and used a deep residual network to mine visual differences to complete the SEI task, cleverly realizing the transformation from temporal signals to spatial features; Wong et al. extracted IQ imbalance features of signals through convolutional neural networks, providing a new feature dimension for radiation source identification, making full use of the inherent differences introduced by hardware nonlinearity; Wang et al. designed a sparse structure selection strategy and combined it with knowledge distillation technology to address the computational load problem of complex numerical convolutional networks, ensuring recognition accuracy while compressing model size, balancing efficiency and performance; Duan Kexin et al. designed a radar radiation source individual identification method that integrates bispectral features. Qian et al. introduced a channel attention mechanism into multi-scale neural networks, fused output features with the original signal to construct a multi-level dictionary, and achieved sparse representation recognition through principal component analysis, thus enhancing the representational ability of key features. Tu et al. first applied complex numerical networks to automatic modulation classification, comparing and verifying its performance advantage over traditional networks in the SEI task, providing a more suitable network architecture for complex signal processing. These methods all demonstrate superior recognition performance compared to traditional methods in different application scenarios due to their data-driven automatic feature extraction capabilities, but there is still room for optimization in complex real-world scenarios such as low signal-to-noise ratio and uneven distribution of individuals.

[0005] Although significant progress has been made in related research, SEI (Search Engine Identification) for signals from actual radiation sources still faces many challenges. First, the data distribution exhibits a significant long-tail characteristic. Due to the highly uneven operating patterns of ships or aircraft, there are ample samples of high-frequency individuals, while a large number of low-frequency individuals (tail-end individuals) are scarce, with some individuals accounting for less than 1% of the total sample. This long-tail distribution causes traditional classification models to easily overfit to head-end individuals, while the accuracy of identifying tail-end individuals drops sharply. Second, the signal features are high-dimensional and structurally complex. AIS and ADS-B signals exist in in-phase / quadrature (I / Q) complex time series, and their individual fingerprints are mainly reflected in the amplitude, phase, and subtle patterns of their time-varying changes. Existing classification methods based on two-dimensional images such as time-frequency maps and spectrograms introduce information loss during the conversion of one-dimensional I / Q signals into a two-dimensional representation and cannot fully capture temporal correlations. Finally, factors such as multipath fading, noise interference, and frequency shift in the actual electromagnetic environment further increase the difficulty of feature extraction, especially under low signal-to-noise ratio conditions, where the extraction of effective features is even more difficult.

[0006] To address the challenges of imbalance, high noise levels, and high dimensionality, existing methods primarily focus on three aspects: data processing, model design, and semi-supervised learning. For example, Tan Kaiwen et al. proposed the IC-SGAN method, which uses a cost-sensitive loss to mitigate gradient propagation imbalance caused by dominant samples, thus improving the classifier's recognition performance on imbalanced datasets. Fu et al. proposed a semi-supervised method for identifying specific radiation sources, utilizing metric adversarial training techniques to effectively leverage unlabeled samples and improve recognition performance on long-tailed data. Wang et al. proposed a self-supervised contrastive framework for radiation source identification under finite labeled data conditions, utilizing unlabeled samples to provide auxiliary information and regularization constraints, effectively mitigating the long-tailed distribution problem. Tao et al. addressed the radiation source identification problem under labeled noise conditions by proposing sample selection and regularization techniques, improving the model's recognition performance on noisy labeled data through the EM algorithm and loss function regularization. Liu et al. proposed a radiation source identification method using deep ensemble learning for multi-feature fusion, improving the algorithm's robustness to low signal-to-noise ratio environments. Xu et al. proposed a robust radiation source identification method based on deep residual shrinking networks, which can effectively handle noise and imbalanced data and improve identification capabilities under long-tailed distribution conditions. Duan Kexin et al. proposed an integrated solution for noise pollution problems, combining unsupervised clustering for identifying mislabeled data with supervised voting for correction. While these solutions perform well in targeted challenging scenarios, they are typically researched for single problems, with limited research on complex problems combining these challenges. Furthermore, they primarily design deep learning models from a data-driven perspective, neglecting the auxiliary role of identity information in feature learning.

[0007] Therefore, how to solve the problems of significant long-tail distribution of data and high dimensionality of sampling features in the identification of individual radiation sources has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0008] The purpose of this application is to provide a method and related apparatus for identifying individual unbalanced radiation sources based on temporal attention, which can solve the problems of significant long-tail distribution of data and high dimensionality of sampling features in the identification of individual radiation sources.

[0009] To achieve the above objectives, this application provides the following solution.

[0010] In a first aspect, this application provides a method for identifying individuals of unbalanced radiation sources based on temporal attention, the method comprising the following steps.

[0011] Acquire radiation source signals.

[0012] Shallow temporal features are extracted from the radiation source signal to obtain deep temporal high-order features.

[0013] Multi-scale feature fusion is performed on the deep temporal high-order features to obtain fused features.

[0014] Design an identity-aware temporal attention module to generate identity-adaptive time-step attention weights.

[0015] The fused features and the identity-adaptive time-step attention weights are multiplied element-wise to obtain the weighted features.

[0016] The weighted features are then aggregated into global identity features to obtain aggregated global identity features.

[0017] The aggregated global identity features are mapped to obtain probability distribution features.

[0018] The probability distribution features are input into a classifier to obtain the probability of the radiation source signal corresponding to an individual radiation source.

[0019] In a second aspect, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the method for identifying individual unbalanced radiation sources based on temporal attention as described in the first aspect.

[0020] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for identifying individual unbalanced radiation sources based on temporal attention as described in the first aspect.

[0021] Based on the specific embodiments provided in this application, the following technical effects are disclosed.

[0022] This application provides a method and related apparatus for identifying individuals from imbalanced radiation sources based on temporal attention. The method includes: acquiring radiation source signals; extracting shallow temporal features from the radiation source signals to obtain deep temporal domain high-order features; transforming the original signal into effective features, mining the temporally relevant key information contained in the signal, and laying the foundation for subsequent feature processing; performing multi-scale feature fusion on the deep temporal domain high-order features to obtain fused features; integrating feature information at different scales to compensate for the limitations of single-scale features and improve the completeness and representational ability of the features; designing an identity-aware temporal attention module to generate identity-adaptive time-step attention weights; accurately focusing on key time-step information related to the individual's identity as a radiation source, reducing interference from irrelevant temporal information; and converting the obtained signals into effective features. The method involves element-wise multiplication of the fused features and the identity-adaptive time-step attention weights to obtain weighted features; enhancing the contribution of identity-related features to further highlight individual identity differences; aggregating the weighted features into global identity features to obtain aggregated global identity features; effectively integrating scattered features to form global features that comprehensively characterize the identity of radiation source individuals; mapping the aggregated global identity features to obtain probability distribution features; transforming high-dimensional identity features into easily classifiable probability distribution forms to provide suitable feature inputs for subsequent classification and recognition; inputting the probability distribution features into a classifier to obtain the probability of radiation source individuals corresponding to the radiation source signal; achieving accurate identification and probability determination of radiation source individuals, and outputting intuitive and reliable identification results. This method, through the progressive and collaborative steps, comprehensively covers the entire process of radiation source individual identification, from data acquisition, feature extraction, fusion optimization, attention focusing to final classification. It effectively mines key identity information in time-series features, strengthens the characterization of individual differences, reduces irrelevant interference, and significantly improves the accuracy, reliability, and robustness of radiation source individual identification in imbalanced scenarios. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is an application environment diagram of an unbalanced radiation source individual identification method based on temporal attention in one embodiment of this application.

[0025] Figure 2This is a flowchart illustrating an embodiment of an unbalanced radiation source individual identification method based on temporal attention provided in this application.

[0026] Figure 3 This is a schematic diagram of the overall framework of an unbalanced radiation source individual identification method based on temporal attention, provided in an embodiment of this application.

[0027] Figure 4 A diagram of a multi-scale feature fusion network structure provided in an embodiment of this application.

[0028] Figure 5 This is a schematic diagram illustrating the relationship between different individuals and the number of samples in an AIS dataset provided in an embodiment of this application.

[0029] Figure 6 This is a schematic diagram illustrating the number of individuals within different sample size ranges in an ADS-B dataset provided in an embodiment of this application.

[0030] Figure 7 This is a schematic diagram of the I and Q features of the collected AIS dataset provided in an embodiment of this application.

[0031] Figure 8 The graph shows the variation of loss and accuracy with the number of training iterations, as provided in one embodiment of this application.

[0032] Figure 9 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0033] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] To address the issues of significant long-tailed data distribution and high dimensionality of sampled features in radiation source individual identification, this application proposes an improved radiation source individual identification algorithm based on temporal attention. By constructing a multi-scale feature fusion module to integrate feature information at different temporal granularities, and designing an identity-aware temporal attention module, the model focuses on individuals with a small sample size and identity-related temporal features by generating and embedding identity-adaptive attention weights. Experiments on the real-world AIS dataset and the publicly available ADS-B dataset show that the proposed method achieves identification accuracies of 94.89% and 93.23%, respectively, with an accuracy decrease of less than 2% at a low signal-to-noise ratio of 3dB.

[0035] To address the aforementioned issues, this application proposes an individual radiation source identification algorithm oriented towards long-tailed distribution and high-dimensional features, which enables effective identification of individual radiation source data from actual sampling. The main contributions are as follows.

[0036] 1) To address the issues of traditional temporal attention mechanisms lacking identity information guidance and easily focusing on redundant regions, an identity-aware temporal attention mechanism is proposed. This mechanism transforms discrete radiation source identity labels into high-dimensional continuous embedding vectors, incorporating prior identity knowledge into the attention weight learning process. This mechanism not only strengthens key temporal features related to specific identities and suppresses noise interference, but also achieves implicit upsampling of a few individual samples in a long-tailed distribution through dynamic sample reweighting. This alleviates the training bias dominated by head individuals and improves the recognition performance of tail individuals.

[0037] 2) Based on the temporal characteristics of IQ signals, a multi-scale feature learning module adapted to IQ signals is designed and a four-branch parallel architecture is constructed. Through average pooling and linear upsampling operations of different time lengths, it covers multi-temporal granular features from the original scale to the 1 / 8 scale. This not only preserves the local instantaneous patterns of the signal but also captures global structural information, effectively enhancing the model's ability to distinguish fine-grained fingerprint features of radiation sources, while avoiding the information loss caused by traditional time-frequency conversion.

[0038] 3) Construct an efficient and collaborative end-to-end recognition framework to achieve collaboration between the 1D residual backbone network, identity perception temporal attention and multi-scale feature fusion module, and combine a cosine normalized classifier and cross-entropy loss to achieve accurate identification of radiation source individuals in high-dimensional, low signal-to-noise ratio, and imbalanced scenarios.

[0039] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0040] The method for identifying individual unbalanced radiation sources based on temporal attention provided in this application can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be set up independently, integrated into server 104, or placed in the cloud or on another server. Terminal 102 can send the acquired radiation source signal to server 104. After receiving the radiation source signal, server 104 performs shallow temporal feature extraction on the radiation source signal to obtain deep temporal domain high-order features; performs multi-scale feature fusion on the deep temporal domain high-order features to obtain fused features; designs an identity-aware temporal attention module to generate identity-adaptive time-step attention weights; multiplies the fused features and the identity-adaptive time-step attention weights element-wise to obtain weighted features; aggregates global identity features on the weighted features to obtain aggregated global identity features; maps the aggregated global identity features to obtain probability distribution features; and inputs the probability distribution features into a classifier to obtain the probability of the radiation source signal corresponding to the individual radiation source. Server 104 can feed back the probability of the obtained radiation source signal corresponding to the individual radiation source to terminal 102. Furthermore, in some embodiments, the imbalanced radiation source individual identification method based on temporal attention can also be implemented separately by server 104 or terminal 102. For example, terminal 102 can directly perform imbalanced radiation source individual identification based on temporal attention on the radiation source signal, or server 104 can obtain the radiation source signal from the data storage system and perform imbalanced radiation source individual identification based on temporal attention on the radiation source signal.

[0041] The terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, and tablets. The server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers, or it can be a cloud server.

[0042] In one exemplary embodiment, such as Figure 2 As shown, a method for identifying individual imbalanced radiation sources based on temporal attention is provided. This method is executed by a computer device, specifically a terminal or server, or both. In this embodiment, the method is applied to... Figure 1 Taking server 104 as an example, the following steps are included.

[0043] S1: Acquire radiation source signal.

[0044] S2: Perform shallow temporal feature extraction on the radiation source signal to obtain deep temporal high-order features.

[0045] S3: Perform multi-scale feature fusion on the deep temporal high-order features to obtain fused features.

[0046] S4: Design an identity-aware temporal attention module to generate identity-adaptive time-step attention weights.

[0047] S5: Multiply the fused features and the identity-adaptive time-step attention weights element-wise to obtain the weighted features.

[0048] S6: Perform global identity feature aggregation on the weighted features to obtain aggregated global identity features.

[0049] S7: Map the aggregated global identity features to obtain probability distribution features.

[0050] S8: Input the probability distribution features into the classifier to obtain the probability of the radiation source signal corresponding to the individual radiation source.

[0051] Implementing steps S1 to S8 as described above can effectively extract key identity information from temporal features, strengthen the representation of individual differences, reduce irrelevant interference, and significantly improve the accuracy, reliability, and robustness of individual identification of radiation sources in imbalanced scenarios.

[0052] The following section provides a detailed description of the individual identification method for imbalanced radiation sources based on temporal attention proposed in this application.

[0053] (1) Feature extraction module design.

[0054] To address the challenges of adapting I / Q signal temporal characteristics, long-tailed data distribution, and strong noise interference in radiation source identification tasks, this application proposes an improved residual network model that integrates identity-aware temporal attention and multi-scale feature fusion. The model takes the original I / Q dual-channel signals as input and achieves end-to-end radiation source identification through shallow temporal feature extraction via a 1D residual backbone network, fine-grained information integration via a multi-scale feature fusion module, and key feature enhancement via an identity-aware temporal attention module. The overall framework is as follows: Figure 3 As shown.

[0055] 1.1 Shallow feature learning based on residual networks.

[0056] As an optional implementation, in step S2, shallow temporal feature extraction is performed on the radiation source signal to obtain deep temporal high-order features, specifically including the following steps.

[0057] S21: Use 1D convolutional layers, batch normalization, and ReLU activation function to extract features from the radiation source signal to obtain initial features.

[0058] S22: The initial features are downsampled using the MaxPool1d layer and then input into the four residual layers for progressive feature extraction to obtain deep temporal high-order features; each residual layer consists of two stacked 1D residual blocks, and each 1D residual block contains two-level feature transformations of convolution, batch normalization and ReLUctant.

[0059] Specifically, in order to extract higher-order representations from local instantaneous features to global structural features from the radiation source IQ signal layer by layer, the model adopts a 1D residual backbone network design. By using skip connections, it alleviates the gradient vanishing problem in deep network training and adapts to the long-range dependency modeling requirements of time series signals.

[0060] To capture the temporal characteristics of the IQ signal and avoid information loss caused by cross-domain conversion, the model uses 1D convolution to directly extract features from the temporal signal. The network takes the original IQ signal as input, and first performs preliminary temporal feature encoding, channel upscaling, and temporal downsampling on the dual-channel signal through an initial 1D convolutional layer to filter out redundant noise and enhance the effective temporal pattern. The formal expression is shown below.

[0061] (1).

[0062] in, Initial features; X IQ For radiation source signal; kernel size `core` is the kernel size; `stride` is the span; `padding` is the padding value; `out` channel This is the output channel.

[0063] The initial convolutional layer is a 1D structure to match the unidirectional characteristics of the temporal signal. The kernel size is set to 15 to cover a longer temporal window to capture local correlations and ensure that the temporal length is downsampled to half of the original length. Simultaneously, the number of channels is increased from 2D to 64D in the IQ signal, effectively increasing the feature dimension and giving the data richer representations of local temporal patterns. Then, batch normalization (BN) and ReLU activation functions are applied to finally obtain... Its time-series dimension is ;in, L This represents the original timing length. L 1 represents the time series length after the first downsampling.

[0064] Initial features Further downsampling via the MaxPool1d layer reduces the timing length to (L2 is the temporal length after the second downsampling), and then input to four residual layers. By expanding the receptive field layer by layer, progressive extraction from local instantaneous features to global structural features is achieved. Each residual layer consists of two stacked 1D residual blocks (BasicBlock1D). Each residual block contains two-level feature transformations: convolution, batch normalization, and ReLU. The formal expression is shown below.

[0065] (2).

[0066] Among them, F out Output features for residual blocks; and All are 1D convolutional layers with a kernel size of 3. Padding=1 is set to perform symmetrical padding to keep the temporal length unchanged, which is suitable for modeling the local correlation of temporal signals. For skip connection branches, when the number of input and output channels or the temporal length do not match (such as the first residual block in residual layers 2-4), downsampling is performed by setting a stride of 2 in the 1D convolution to achieve simultaneous dimensionality increase of the number of channels and simultaneous temporal length downsampling, ensuring the effectiveness of residual fusion. When the dimensions are completely consistent (such as the residual block in the first layer and the second residual block in each layer), identity mapping is directly applied to ensure the effectiveness of residual fusion and the smoothness of gradient propagation. The input features are for the residual block.

[0067] The number of channels in the four residual layers gradually increases with network depth, forming a hierarchical structure that progressively enhances feature abstraction. The first layer has 64 channels and uses two levels of residual blocks to refine the representation of local temporal features, maintaining the temporal length. The first residual block of the second layer is upsampled to 128 channels and downsampled temporally using a convolution with stride=2. The first layer compresses redundant temporal information while improving feature abstraction; the second residual block maintains the same dimensionality, deepening feature representation. The third layer further expands the number of channels to 256 through convolutions with a stride of 2, and downsamples the temporal length to [missing information]. This expands the receptive field to capture a wider range of temporal dependencies. The fourth layer completes the final feature upsampling and temporal downsampling. This forms a 512-dimensional high-dimensional abstract temporal feature, enhancing the ability to represent individual differences. L 3 represents the time series length after the third downsampling. L 4 represents the time series length after the fourth downsampling. L 5 represents the time series length after the fifth downsampling.

[0068] Through progressive feature extraction using four residual layers, the network ultimately outputs deep temporal high-order features. The 512-dimensional channels demonstrate the high-dimensional abstraction capability of features. L 5 represents the final time series length. Through the above design, the model fully learns the temporal structure information and individual-specific patterns of the IQ signal, achieving deep mining of temporal features based on 1D convolution and residual structures.

[0069] 1.2 Multi-scale feature fusion.

[0070] As an optional implementation, in step S3, multi-scale feature fusion is performed on the deep temporal high-order features to obtain fused features. Specifically, this includes: using four parallel branches to perform multi-scale processing on the deep temporal high-order features to obtain fused features; wherein, branch 0 directly performs convolution, batch normalization, and ReLU activation on the deep temporal high-order features to extract the original scale features; branches 1-3 respectively perform average pooling downsampling with strides of 2, 4, and 8, and then extract multi-scale features through convolution, batch normalization, and ReLU activation, and then restore them to the original length through linear upsampling interpolation. After feature concatenation, preliminary fused features are obtained; then, the channel dimension fusion and transformation are completed through 1×1 convolution, batch normalization, and ReLU activation functions to obtain the final fused features.

[0071] Specifically, to fully explore the multi-scale time-frequency features of the radiation source signal, a multi-scale feature fusion module is designed to integrate feature information from different scales and improve the model's ability to represent fine-grained individual differences. This module adopts a multi-branch parallel architecture to process the input features... Multi-scale processing is performed separately: four parallel branches are set up, such as... Figure 4 As shown, branch 0 directly performs convolution, batch normalization (BatchNorm), and ReLU activation on the input to extract the original scale features; branches 1-3 perform average pooling downsampling with strides of 2, 4, and 8 respectively, then extract multi-scale features through convolution, batch normalization, and ReLU activation, and finally restore the original length through linear upsampling interpolation. Let the first branch... i The output of each branch is , , C i For the first i Each embedding dimension is used; finally, the fused features are obtained through feature concatenation.

[0072] (3).

[0073] Subsequently, channel-dimensional fusion and transformation are completed through 1×1 convolution, batch normalization, and ReLU activation function, ultimately outputting features. This module enhances the ability to represent the features of radiation source signals at different temporal granularities through complementary fusion of multi-scale features, providing a rich multi-scale feature foundation for subsequent individual perception temporal attention mechanisms, thereby improving the model's ability to distinguish fine-grained differences in individual radiation source identification tasks.

[0074] 1.3 Deep Feature Learning for Identity Perception.

[0075] As an optional implementation, in step S4, an identity-aware temporal attention module is designed to generate identity-adaptive temporal step attention weights, specifically including the following steps.

[0076] S41: Map each identity tag to its corresponding identity embedding vector.

[0077] S42: Align each identity embedding vector with the temporal dimension to obtain the aligned identity embedding vector.

[0078] S43: The aligned identity embedding vector and the preliminary fusion feature are concatenated along the channel dimension to obtain the concatenated feature.

[0079] S44: Based on the concatenated features, the number of channels of the preliminary fused features is first compressed from 2C to the hidden dimension D through a 1×1 convolution. The channel dimension is compressed without destroying the temporal structure, reducing the computational complexity. Then, nonlinearity is introduced through batch normalization and ReLU activation. The number of channels is then compressed to 1 through a second 1×1 convolution to obtain the original attention score. The score is then mapped to the [0,1] interval through the Sigmoid activation function, and finally the identity-adaptive time-step attention weight is obtained.

[0080] Specifically, the key identification features of different radiation sources exhibit significant differences in their temporal distribution, and are further complicated by channel noise, environmental interference, and redundant temporal information. Traditional temporal attention mechanisms rely solely on the original features to adaptively learn weights, lacking explicit guidance from identity information. This can easily lead to attention focusing on non-specific redundant regions, reducing recognition robustness. Therefore, an identity-aware temporal attention module is designed to transform discrete radiation source identity labels into continuous high-dimensional embedding vectors. Through cross-modal interaction between identity information and temporal features, identity-adaptive time-step attention weights are generated to extract key temporal features of the target radiation source while suppressing interference from noise and redundant information. Specifically, this module comprises four parts.

[0081] (I) Identity Label Embedding. To transform discrete identity labels into continuous vectors that can participate in feature interactions, an embedding layer is designed to achieve high-dimensional space mapping of the labels. This layer learns a trainable embedding matrix. Each identity tag Mapped to the corresponding identity embedding vector .in, C =512 is the embedding dimension. The formal expression of the identity embedding vector is shown below.

[0082] (4).

[0083] in, This is the corresponding identity embedding vector; E For embedding matrix; , For identity tags The corresponding one-hot encoded vector, B For training batch size, N The total number of radiation source individuals; Through this mapping, discrete identity information is transformed into continuous high-dimensional features. During training, the embedding matrix E adaptively adjusts to bring the embedding vectors of the same individual radiation source closer together and those of different individuals further apart, thus encoding the identity-specific information of the radiation source. For the entire batch, the identity embedding output is obtained through matrix multiplication. It can be formally expressed as shown below.

[0084] (5).

[0085] Where y is the identity label; e y This is the identity embedding vector.

[0086] (ii) Temporal dimension alignment. Due to the identity embedding vector Lacking a time-series dimension, it cannot be directly compared with time-series features. To enable interaction, its dimensions are expanded through a temporal-dimensional broadcast alignment operation. This operation repeats the identity embedding vector of each sample along the temporal dimension. L This ensures that the embedding vector perfectly matches the dimension of the temporal features. This can be formally expressed as follows.

[0087] (6).

[0088] in, for L A dimensional vector of all 1s L This represents the original timing length. C For the embedded dimension; This represents the outer product operation. This represents repeated operations along the time-series dimension, ultimately expanding the identity embedding vector into a vector with the same dimension as the time-series features; This is the aligned identity embedding vector.

[0089] (III) Cross-modal feature interaction. To enable identity information to effectively guide the attention allocation of temporal features, a channel-dimensional concatenation strategy is adopted to achieve cross-modal interaction between identity embedding and temporal features. The aligned identity embedding vector is then... and By concatenating along the channel dimension, identity-specific guiding information is incorporated while preserving the detailed information of the original temporal features. The formal expression is shown below.

[0090] (7).

[0091] in, Features after splicing; To splice along the channel dimension; Features of fusion; T For the total time step; 2 C This represents the number of feature channels after fusion.

[0092] (iv) Adaptive Attention Generation. To adaptively learn the importance of each time step from the fused features, a lightweight convolutional attention generator is designed to generate time step attention weights through dimensionality compression and nonlinear mapping. First, a 1×1 convolution is used to compress the number of channels of the fused features from 2C to the hidden dimension D, compressing the channel dimension without destroying the temporal structure and reducing computational complexity. Then, batch normalization and ReLU activation are used to introduce nonlinearity, and a second 1×1 convolution is used to compress the number of channels to 1 to obtain the original attention score. Finally, the score is mapped to the [0,1] interval by the Sigmoid activation function to obtain the time step attention weight.

[0093] (8).

[0094] in, For the first i The first sample t Attention weights are assigned to each time step; a larger weight indicates that the time step contains richer radiation source identity-specific features. These attention weights are then compared with... Element-wise multiplication is performed to enhance key temporal features and suppress redundant information. The formal expression of the weighted features is shown below.

[0095] (9).

[0096] in, f The output features are fused and transformed using 1×1 convolution, batch normalization, and ReLU activation function to achieve channel-dimensional fusion. B This refers to the training batch size; C For the embedded dimension; T This represents the total time step. The weighted features will be used as input to the subsequent classifier, providing a more discriminative temporal feature representation for individual radiation source identification.

[0097] Based on the above steps, the designed module can inject prior identity information into the attention mechanism to guide the model to focus on discriminative temporal patterns related to individuals with specific identities. Then, through dynamic sample reweighting, implicit upsampling of a few individual samples is achieved, which alleviates the training bias caused by long-tail distribution, suppresses noisy periods, enhances key signal segments, improves the internal compactness and external separability of individual features, and ultimately improves the model's recognition performance on imbalanced individuals.

[0098] During the model inference testing phase, since the true labels of the test set are unavailable, the mean of all identity embedding vectors is constructed as the proxy identity information, which can be formally expressed as follows.

[0099] (10).

[0100] Among them, Embedding ( i ) is the first i Each identity embedding vector; N The total number of radiation source individuals; E avg This is the mean of the identity embedding vector.

[0101] This design ensures the feasibility and stability of the module during testing, while maintaining the statistical characteristics of identity semantic information.

[0102] 1.4 Global identity feature aggregation.

[0103] As an optional implementation, in step S6, the weighted features are aggregated into global identity features to obtain aggregated global identity features, which specifically includes the following steps.

[0104] S61: Adaptive average pooling is used to globally compress the weighted features to obtain the global feature vector after adaptive average pooling.

[0105] S62: The global feature vector after adaptive average pooling is further compressed by 1D convolution and activation function to obtain the aggregated global identity features.

[0106] Specifically, the weighted features It includes high-order discriminative features that are distributed along the temporal dimension and enhanced by identity information. To adapt to the global feature requirements of the final classification task, this embodiment compresses temporal redundancy and strengthens the representation of core identity features through multi-step aggregation and dimensional transformation operations.

[0107] First, convolution, batch normalization (BN), ReLU, and adaptive average pooling are used to... The temporal dimension is globally compressed, mapping temporal features of arbitrary length to a global feature vector of fixed dimension, in order to eliminate temporal dimension differences and retain core temporal structure information. The formal expression of the global feature vector after adaptive average pooling is shown below.

[0108] (11).

[0109] in, F pool This is the global feature vector after adaptive average pooling.

[0110] This operation integrates individual differences in information such as amplitude variation patterns and phase continuity across the entire I / Q signal timing, forming a preliminary global feature representation.

[0111] Subsequently, 1D convolution and activation functions are used to further compact the features and reduce dimensionality redundancy. Specifically, the first-level 1D convolution compresses the 512-dimensional channel features to 128 dimensions, followed by a ReLU activation function to introduce nonlinearity and enhance the feature discrimination capability. The formal expression of the aggregated global identity features is shown below.

[0112] (12).

[0113] in, F conv1 The aggregated global identity features; in is the input; out is the output.

[0114] (2) Design of classification loss function.

[0115] 2.1 Classifier Design.

[0116] To adapt to the significant long-tail distribution characteristics of radiation source data and the recognition requirements in low signal-to-noise ratio scenarios, improve the inter-class discriminative power of features of a few individual samples, and enhance the distinguishability of classification boundaries, a cosine normalized classifier is introduced as the classification head. Specifically, firstly, 1D convolution is used to further map the 128-dimensional features to 4D, and the probability distribution features are initially generated through the Softmax activation function, as shown in the following formula.

[0117] (13).

[0118] in, F conv2 This represents the characteristics of a probability distribution.

[0119] Then, features F conv2After dimensionality compression to remove redundant single-temporal axes, the data is input into a cosine normalized classifier to predict the radiation source identity. This classifier enhances the inter-class separation of features through a cosine normalization mechanism, mapping 4-dimensional features to... N dimensional identity probability distribution N That is, the total number of radiation source individuals, as shown in the formula below.

[0120] (14).

[0121] in, Indicates the first b The sample belongs to the first c The probability of an individual radiation source satisfies The model optimizes the parameters of the entire feature extraction and classification network through the cross-entropy loss function, achieving [the desired result]. N Identification of individual radiation sources.

[0122] 2.2 Cross-entropy loss function.

[0123] As an optional implementation, the method for identifying imbalanced radiation sources based on temporal attention further includes: optimizing the parameters of the entire feature extraction and classifier using a cross-entropy loss function.

[0124] Specifically, for the individual identity prediction task in radiation source identification, the model uses the standard cross-entropy loss function as the training optimization objective to directly measure the difference between the model's predicted identity probability distribution and the true identity label distribution, guiding the model to learn highly discriminative individual fingerprint features of radiation sources. In particular, the cross-entropy loss directly measures the predicted probability distribution. P Distribution of real labels Y The differences guide the model to learn highly discriminative fingerprint features of radiation sources, as shown in the following mathematical expression.

[0125] (15).

[0126] in, The cross-entropy loss function; B This refers to the training batch size; N The total number of radiation source individuals; For the first b The true one-hot label of each sample, corresponding to the element position of the real individual. c The value is 1, and the rest are 0; This is a minimum value, used to avoid... A numerical error occurred during logarithmic operations.

[0127] 2.3 Model optimization steps.

[0128] The pseudocode of the algorithm proposed in this embodiment is shown in Algorithm 1.

[0129] Algorithm 1: Radiation source identification training algorithm based on attention residual network.

[0130] Input: Radiation source I / Q dataset, training batch size, initial learning rate, number of training rounds.

[0131] Output: The trained radiation source identification model.

[0132] 1) Initialize the Model (including a 1D residual backbone network, a temporal attention module, and a linear classifier).

[0133] 2) Initialize the optimizer Optimizer(Adam, learning rate).

[0134] 3) Initialize the loss function.

[0135] 4) while do.

[0136] 5) Sample from batches of 64 to obtain batch data and labels.

[0137] 6) Forward propagation to obtain the predicted probability labels of individuals in the input batch.

[0138] 7) Calculate the loss and perform backpropagation.

[0139] 8) The optimizer updates the model parameters.

[0140] 9) Calculate the validation accuracy on the test set.

[0141] 10) If the current model is better than the historical best, then save the current model as the best model.

[0142] 11) t = t + 1.

[0143] 12)end while.

[0144] 13) Return the best model.

[0145] (3) Experimental analysis.

[0146] To verify the effectiveness of the algorithm proposed in this application, we first describe in detail the two types of real-world datasets used in the experiment, then clarify the selection and parameter configuration of the comparison algorithm, and provide the hardware and software environment and core evaluation indicators of the experiment. Finally, we analyze and discuss the experimental results from the dimensions of performance comparison of different algorithms, noise robustness, long-tail sample recognition effect, and ablation experiment.

[0147] 3.1 Introduction to the dataset.

[0148] To comprehensively evaluate the performance of the designed radiation source identification model, experiments were conducted using radiation source signal datasets collected in two real-world scenarios: a self-collected long-tailed identification dataset of ship AIS radiation sources and a publicly available aircraft ADS-B dataset. The sample size distribution is as follows: Figure 5 and Figure 6 As shown.

[0149] The details are as follows.

[0150] (1) Data set for identifying long tails of ship radiation sources.

[0151] The data acquisition device is an SM200C signal acquisition unit, specifically designed to capture AIS radiation source signals from civilian vessels. The acquisition frequency band is set to 161.975MHz, and the signal bandwidth is 25kHz. After acquisition, the raw signal data is manually labeled and cleaned to ensure the accuracy and reliability of the data tags. To obtain structured radiation source information, this embodiment implements an automatic detection and demodulation algorithm for AIS radiation source signals. The acquired raw signals are demodulated according to the AIS data encoding protocol, ultimately yielding multi-dimensional data information, including the ship's AIS radiation source radio frequency signals stored in IQ complex number data format.

[0152] The constructed dataset contains 30 different civilian ship radiation sources. The sample size of each source exhibits a typical long-tail distribution characteristic in terms of frequency and conforms to the Zipf distribution pattern, such as... Figure 5 As shown, the largest number of individuals in the dataset is 7309, while the smallest number is only 59. Individuals at the bottom of the dataset account for less than 1% of the total sample size. Each sample in the dataset contains two classes of complex signal features, I and Q, with a dimension of 11520. A schematic diagram of the I and Q features of the collected AIS dataset is shown below. Figure 7 As shown. Furthermore, due to building obstruction and other signal interference at the receiving point, the signal-to-noise ratio of the collected data is low. Therefore, this is a typical imbalanced, high-dimensional, low-signal-to-noise-ratio radiation source individual identification problem. To ensure the independence and distribution consistency of the training and testing data, the training and testing sets are randomly divided in a 6:4 ratio according to the number of samples per individual, and there is no overlap of continuous signal segments between the two.

[0153] (2) ADS-B real-world dataset.

[0154] This dataset is a publicly available, real-world dataset published on the Science Data Bank platform. The dataset was collected on October 5, 2022, with a signal sampling rate of 40MHz. The collected data consists of Aircraft Automatic Dependent Surveillance-Broadcast (ADS-B) signals in DF=17 format. The dataset contains 216 data files, each corresponding to the I and Q signal data of one aircraft, with a sampling feature length of 4800 points. Considering the large number of individuals, the distribution of individuals with different sample sizes has been plotted, as shown below. Figure 6 As shown, there are 5 individuals with fewer than 10 samples and 2 individuals with more than 270 samples. Therefore, this is a typical problem of identifying radiation source individuals with high dimensionality and obvious imbalance. The training and test sets are divided in a 7:3 ratio, and other partitioning methods are the same as those used for the AIS data.

[0155] 3.2 Introduction to the comparison algorithm.

[0156] The designed individual radiation source identification algorithm was compared with typical traditional methods and deep learning methods. In terms of traditional algorithms, in order to identify radiation source signals, the feature extraction process takes the radiation source I / Q complex signal as the core input. Based on the physical nature and mode differences of the signal, a four-dimensional feature system of statistics, spectrum, time domain, and modulation was constructed to realize the transformation of the original high-dimensional sampling points into low-dimensional effective features. Among them, statistical features include amplitude distribution (mean, skewness, etc.), phase fluctuation (mean of phase difference, standard deviation, etc.), I / Q component correlation (covariance, correlation coefficient, etc.), and higher-order moments (3rd moment, 4th moment), capturing the overall statistical regularity of the signal; spectral features are based on 1024-point FFT, extracting amplitude spectrum energy, phase spectrum distribution, spectral centroid, roll-off rate, and dominant frequency information to characterize the frequency domain structure differences of the signal; time domain features reflect the time dynamic changes of the signal through autocorrelation analysis (maximum autocorrelation value of I / Q components, tail mean), zero-crossing rate (number of times I / Q components cross zero per unit time), energy distribution, and envelope peak characteristics; modulation features target the modulation characteristics of the radiation source, extracting amplitude variation coefficient, phase jump parameters, instantaneous frequency fluctuations, and constellation diagram statistical attributes (symbol balance, I / Q correlation, etc.) to distinguish the signal differences of different modulation modes. The entire extraction process involves both manual screening to remove redundant information and comprehensive coverage of the signal's multi-dimensional physical characteristics, ultimately yielding 72-dimensional statistical features. This provides input features with both discriminative power and robustness for subsequent classification models such as Random Forest and XGBoost. Specific algorithm comparisons are described below.

[0157] (1) RandomForest: Random Forest is an ensemble learning algorithm based on multiple decision trees. Its core features are strong resistance to overfitting and outstanding robustness. It can efficiently process high-dimensional data and capture nonlinear feature associations. In this embodiment, the number of decision trees is set to 200 to balance training efficiency and generalization performance. The maximum tree depth is limited to 20 and the minimum number of sample splits is set to 5 to suppress noise interference. Finally, the classification result is output through a majority voting mechanism.

[0158] (2) SVM: Support Vector Machine is based on statistical learning theory. It performs well in high-dimensional data processing by utilizing the nonlinear mapping capability of the kernel function. Its core feature is that it can achieve recognition by finding the optimal separating hyperplane and has strong generalization ability. It is a commonly used algorithm for processing high-dimensional data. In this embodiment, the RBF kernel function is used and the scale mode is set to automatically match the feature scale difference. By setting the regularization parameter C=1.0, the fitting accuracy and generalization ability are balanced. By optimizing the kernel function and the regularization parameter, the weak linear correlation of signal features can be explored, avoiding the overfitting problem caused by high-dimensional data.

[0159] (3) XGboost: XGboost is an ensemble algorithm based on gradient boosting trees. It combines regularization mechanisms and sampling techniques to capture fine-grained feature differences in data, resulting in high training efficiency and strong generalization ability. In this embodiment, 200 decision trees are set up and the maximum depth is limited to 10 to control model complexity. A learning rate of 0.1 is used for iterative optimization. Subsampling and column sampling of 0.8 are used to reduce feature redundancy, and L1 and L2 regularization are combined to constrain model weights.

[0160] (4) Ensemble: Ensemble models reduce the bias and variance of a single model by combining the advantages of multiple base models, thereby improving classification stability and generalization performance. In this embodiment, a soft voting mechanism is used to integrate three base models: Random Forest, SVM, and XGboost. The optimal configuration of each base model is used to ensure the upper limit of performance. The probability distributions of the three outputs are integrated with equal weights, and the highest probability is taken as the final classification result.

[0161] (5) MLP: Multilayer Perceptron is a feedforward neural network with deep feature extraction capabilities. It has strong nonlinear fitting ability and can adaptively learn the abstract correlation of data through hidden layers. In this embodiment, a three-layer hidden layer structure of 128-64-32 is designed. The ReLU activation function is used to solve the gradient vanishing problem. At the same time, L2 regularization is used to suppress overfitting. The Adam optimizer is selected to train the model with a learning rate of 0.001 and 100 iterations.

[0162] (6) IC-SGAN: This algorithm targets the imbalanced individual SEI task and utilizes the binary game structure of generative adversarial networks. The generator generates signal samples that approximate the true distribution, while the discriminator simultaneously performs the dual tasks of distinguishing between real and fake samples and classifying radiation source individuals, effectively using unlabeled data to supplement feature information. Multi-scale topology modules and residual units are embedded in the discriminator network to enhance the multi-dimensional resolution feature extraction of time-domain signals. An imbalanced individual loss (ICL) is designed to alleviate the gradient dominance problem of majority class samples through adaptive weight adjustment, thereby addressing the performance degradation caused by the long-tail distribution of radiation source data.

[0163] (7) RSNI-SEI: This algorithm focuses on the robust SEI requirements in low signal-to-noise ratio environments, with a deep residual shrinkage network as its core architecture. By introducing a soft threshold function in the nonlinear transform layer, it automatically suppresses noise features in the signal and reduces the impact of interference on RF fingerprint extraction. An adaptive threshold subnet is designed to dynamically allocate the optimal threshold for each IQ signal sample, eliminating the need for manual parameter tuning based on expert knowledge. During the training phase, a noise injection strategy is adopted to add Gaussian white noise with different SNRs to the input signal, improving the model's generalization ability to complex channel environments.

[0164] 3.3 Software and hardware environment settings.

[0165] The hardware and software environment configuration for this embodiment is as follows: the operating system is Ubuntu 20.04 LTS, the programming language is Python 3.8, the deep learning framework is PyTorch 1.7.0, the computing power is provided by an i9-13800K and an NVIDIA GeForce RTX3080Ti GPU, the optimizer is Adam, the learning rate is 0.0002, the batch size is 64, and the number of training epochs is 100.

[0166] 3.4 Evaluation Indicators.

[0167] To objectively evaluate the performance of each model in the radiation source identification task, considering the long-tailed distribution and strong noise characteristics of the two datasets, accuracy (ACC), precision, recall, and F1 score were selected as four core metrics for performance quantification. Assuming TP represents correctly predicted samples that are actually radiation sources, TN represents correctly excluded samples that are not actually radiation sources, FP represents misclassified samples that are not actually radiation sources, and FN represents misclassified samples that are actually radiation sources, the relevant metrics are calculated as follows.

[0168] (1) Accuracy: Reflects the overall correctness of the model's classification of all samples. The formula is shown below.

[0169] (16).

[0170] This metric is used to measure the overall recognition accuracy, but in long-tail data, the top individuals may increase the ACC, so it is necessary to combine it with other metrics to analyze the performance of the tail samples.

[0171] (2) Accuracy: reflects the reliability of the model's prediction results to avoid misjudgment. The formula is shown below.

[0172] (17).

[0173] Improving this indicator can reduce the misjudgment rate of individuals at the stern of a ship.

[0174] (3) Recall: Reflects the model’s coverage of the real target to avoid missing the target. The formula is shown below.

[0175] (18).

[0176] This metric is crucial for adapting to long-tail scenarios, measuring the model's ability to capture small samples at the tail end, and preventing missed identifications that could lead to identity loss.

[0177] (4) F1 score: used to reconcile average precision and recall, and to balance accuracy and coverage, as shown in the formula below.

[0178] (19).

[0179] As a core indicator of comprehensive performance, it avoids bias by focusing on a single indicator and adapts to the unified performance measurement of long-tail and noisy datasets.

[0180] 3.5 Analysis of experimental results.

[0181] (1) Comparison of experimental results of different algorithms.

[0182] The experimental results of different algorithms are shown in Table 1. As can be seen from the table, traditional machine learning algorithms generally outperform deep learning algorithms on both datasets. The best-performing XGBoost achieves an ACC of 78.18% on the AIS dataset and 78.91% on the ADS-B dataset, significantly lagging behind deep learning algorithms. This is mainly due to the high dimensionality and large number of individuals in these datasets, making it difficult for traditional algorithms to capture the temporal correlation of IQ signals and the fine-grained differences in the hardware fingerprints of radiation sources. Furthermore, in the long-tailed distribution scenario of radiation source identification, traditional algorithms are prone to overfitting of head individuals and underfitting of tail individuals due to sample imbalance, resulting in insufficient overall generalization ability and limited algorithm performance.

[0183] The generative model IC-SGAN exhibits significant differences across the two datasets, with its overall performance on the ADS-B dataset outperforming that of the AIS dataset. This difference primarily stems from the characteristics of the datasets. IC-SGAN employs a generative approach, validating its effectiveness on simulated data with 1000 dimensions and 5 individuals, as well as on real data with 10 individuals. The AIS dataset, however, has 11520 dimensions, 30 individuals, and strong noise interference. During training, the generative model struggles to fully learn the true feature distribution of the radiation source individuals, resulting in insufficient quality of generated samples and limiting recognition performance. Furthermore, its training in real-world tests is unstable and prone to crashing. In contrast, the ADS-B dataset has 4800 dimensions and a high signal-to-noise ratio, similar in dimensionality to the dataset processed in the IC-SGAN paper. Its generative architecture and multi-scale topological feature learning effectively enhance feature representation capabilities, thus resulting in better performance.

[0184] The algorithm proposed in this application demonstrates significant advantages on both the AIS and ADS-B datasets, validating the effectiveness of the designed method. This shows that by dynamically enhancing the effective features of the IQ signal and suppressing redundancy and noise interference, the algorithm achieves high recognition accuracy and robustness in long-tailed distribution and high-dimensional feature scenarios.

[0185] Table 1 Experimental results of different algorithms

[0186] (2) Noise robustness analysis.

[0187] During actual propagation, radiation source signals are subject to channel interference such as multipath fading and electromagnetic noise, and signal-to-noise ratio fluctuations directly affect the reliability of feature extraction. Considering that Gaussian white noise is the most typical background interference in communication systems, this embodiment uses the publicly available ADS-B dataset as the object to construct an interference model based on Gaussian white noise. By controlling the noise power to achieve different signal-to-noise ratio conditions, the performance variation of the model is analyzed.

[0188] The experiment used additive white Gaussian noise, whose probability density function follows a normal distribution and whose power spectral density is uniformly distributed throughout the frequency domain, consistent with the natural electromagnetic interference characteristics faced by the radiation source signal in the environmental channel. The noise addition process strictly followed the definition of signal-to-noise ratio.

[0189] (20).

[0190] in, The average power of the raw ADS-B IQ signal. The added Gaussian white noise power. Adjusted... This enabled the signal-to-noise ratios of the datasets to reach 14dB (corresponding to a noise ratio of 20%), 7dB (corresponding to a noise ratio of 40%), and 3dB (corresponding to a noise ratio of 60%), respectively, covering three typical high, medium, and low signal-to-noise ratio interference environments in practical applications.

[0191] The accuracy and F1 score of the main algorithms are shown in Table 2. The results show that the proposed algorithm exhibits optimal and stable anti-interference performance under various noise intensities. Specifically, under SNR conditions of 14dB and 7dB, the accuracy of the proposed algorithm reaches 93.94% and 92.87%, respectively, with F1 scores remaining above 92%. Even in extreme scenarios with a low SNR of 3dB, the proposed algorithm maintains an ACC of 91.96% and an F1 score of 91.87%, with performance degradation controlled within 2%. In contrast, while RSNI-SEI shows some robustness, its ACC is about 6% lower than the proposed algorithm, possibly due to its pre-set soft-threshold denoising mechanism excessively suppressing weak features. XGboost, limited by the sensitivity of traditional feature engineering to noise, experiences a continuous decline in performance as the SNR decreases. In summary, the proposed algorithm effectively enhances feature robustness in complex electromagnetic environments by dynamically focusing on key temporal patterns and suppressing noise interference through an identity-aware attention mechanism, combined with a multi-scale feature network to extract fine-grained features of IQ signals.

[0192] Table 2. ADS-B Individual Identification Accuracy under Different Noise Levels

[0193] (3) Analysis of individual identification accuracy with different sample sizes.

[0194] Considering the large number of individuals in the dataset, to further analyze the effectiveness of the designed algorithm, individuals were sorted from highest to lowest sample size, and then divided into four groups: 0-25% (head, the 25% with the most samples), 25%-50%, 50%-75%, and 75%-25% (tail). Based on the AIS dataset and the ADS-B dataset containing 60% noise, the recognition performance was compared with XGboost and RSNI-SEI. The accuracy of each group is shown in Table 3.

[0195] Table 3. Identification accuracy of individuals with different sample sizes

[0196] As can be seen from the table, the latter two methods, which specifically address imbalanced individuals, exhibit a pattern of high performance in the middle two groups, followed by the head group, and lowest performance in the tail group. This is related to the imbalanced individual handling strategy adopted by the algorithm to achieve higher overall accuracy. The tail samples, due to their dispersed nature, large number, but small sample size, are difficult to distinguish. Overall, the designed method demonstrates good balanced recognition ability across all groups, especially showing a significant performance advantage in the tail class. In the AIS tail class, the F1 score reaches 94.27%, exceeding existing methods by more than 5%. This is mainly due to the dynamic reweighting effect of the identity-aware attention mechanism, which allows tail class samples to obtain higher attention weights in feature interactions, effectively alleviating underfitting caused by sample scarcity.

[0197] The F1 score better reflects the balance of individual recognition. Our method's F1 score difference on the AIS dataset is 3.60%, lower than the other two methods, demonstrating that it does not sacrifice minority class performance for overall accuracy and exhibits good balance. On the 3dB noisy ADS-B dataset, the tail class F1 score still reaches 90.09%, with a difference of only 2.87% compared to the head class. This indicates that the attention mechanism and residual structure achieve an effective balance between noise suppression and feature preservation. It shows that the designed method does not favor the head class with more samples, but also achieves good recognition of the tail class with fewer samples, avoiding the learning bias of sacrificing minority class performance for overall accuracy.

[0198] Furthermore, the proposed algorithm maintains high stability even on the ADS-B dataset containing 60% noise. In this low signal-to-noise ratio scenario, the performance degradation of the designed algorithm in each group is far less than that of the comparative algorithms: while maintaining leading accuracy and F1 score for head and middle individuals, the accuracy and F1 score for tail individuals are significantly higher than existing algorithms, with the lowest degradation, proving that the algorithm can still stably capture the features of tail individuals under the dual challenges of long-tailed distribution and noise interference.

[0199] In summary, the designed method performs well in identifying radiation source individuals with different sample sizes under both AIS and ADS-B typical task scenarios. In particular, the accuracy of identifying the tail minority class is close to that of the head individual, indicating the balance and stability of the algorithm, which can meet the identification needs of radiation source sample size imbalance in real-world scenarios.

[0200] (4) Ablation experiment analysis.

[0201] To quantify the contributions of the identity-aware temporal attention module and the multi-scale feature fusion module in the proposed model, an ablation experiment was designed based on the AIS dataset. As shown in Table 4, the basic model based on the residual network achieved a certain accuracy. When only the identity-aware temporal attention module was introduced, the model performance improved significantly, with an ACC improvement of 3.84% and an F1 score improvement of 4.47%. This indicates that by dynamically modeling the feature importance weights of the temporal dimension, this method can adaptively focus on key temporal segments strongly correlated with the radiation source identity, such as signal modulation details and hardware nonlinear distortion features, while suppressing random noise and redundant information in the electromagnetic environment. This effectively solves the problems of large temporal span and sparse effective features in AIS signals.

[0202] Table 4 Ablation Experiment

[0203] Based on the introduction of the identity-aware temporal attention module, the addition of a multi-scale feature fusion module further improved the model accuracy by 1.48%, indicating that the multi-scale feature fusion module and the identity-aware temporal attention module can form an effective synergy. The former makes up for the limitations of single-scale features in representing complex radiation source signals by extracting features under different receptive fields in parallel, while the latter assigns dynamic attention weights to multi-scale features to ensure that effective features at key scales are given priority.

[0204] (5) Training stability analysis.

[0205] The training accuracy versus loss curves of the proposed algorithm are as follows: Figure 8 As shown in the figure, the designed model converges quickly, and the accuracy stabilizes at over 90% after approximately 30 training epochs. The loss curve decreases rapidly and flattens out without significant fluctuations, and there is no gradient vanishing phenomenon during training. This is mainly due to the synergistic effect of multi-scale feature fusion and the loss function, which ensures stable training of the deep network. In addition, the accuracy on the test set fluctuates little under different training epochs, indicating that the model has strong generalization ability and good adaptability and stability.

[0206] In summary, this application addresses the challenges of significant long-tailed data distribution and low efficiency in high-dimensional feature extraction encountered in practical applications of individual radiation source identification. It proposes a method for identifying individual imbalanced radiation sources based on temporal attention and multi-scale feature fusion. This method uses residual units as the backbone architecture, integrating identity-aware temporal attention and multi-scale feature fusion modules. It captures the temporal correlation of IQ signals through 1D convolution and residual connections, while leveraging identity-prior-guided attention mechanisms to focus on key features and suppress noise redundancy. Furthermore, it integrates information at different temporal granularities through a multi-branch parallel architecture to enhance fine-grained individual difference representation. Cross-entropy loss and cosine normalization classifiers are introduced to mitigate training bias caused by the long-tailed distribution. Multi-dimensional experiments based on real-world high-dimensional, low signal-to-noise ratio (SNR) AIS datasets and publicly available ADS-B datasets demonstrate that the proposed method achieves performance improvements of over 5% compared to mainstream algorithms such as XGboost and RSNI-SEI in core metrics such as accuracy. Furthermore, the algorithm exhibits minimal performance fluctuations under high noise conditions, demonstrating significant robustness advantages. This fully validates the effectiveness and practicality of the designed model in high-dimensional, low SNR, and unbalanced radiation source identification tasks.

[0207] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 9 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores radiation source signals. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a temporal attention-based method for identifying individual unbalanced radiation sources.

[0208] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0209] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method embodiments.

[0210] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the above-described method embodiments.

[0211] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of the relevant data are carried out in compliance with the relevant data protection laws and policies of the country where the location is located, and with the authorization granted by the owner of the corresponding device.

[0212] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0213] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0214] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0215] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for identifying individual imbalanced radiation sources based on temporal attention, characterized in that, The method for identifying individual imbalanced radiation sources based on temporal attention includes: Acquire radiation source signals; Shallow temporal feature extraction is performed on the radiation source signal to obtain deep temporal high-order features; Multi-scale feature fusion is performed on the deep temporal high-order features to obtain fused features; Design an identity-aware temporal attention module to generate identity-adaptive temporal step attention weights; The fused features and the identity-adaptive time-step attention weights are multiplied element-wise to obtain the weighted features; The weighted features are then aggregated into global identity features to obtain aggregated global identity features. The aggregated global identity features are mapped to obtain probability distribution features; The probability distribution features are input into the classifier to obtain the probability of the radiation source signal corresponding to the individual radiation source.

2. The method for identifying individual unbalanced radiation sources based on temporal attention according to claim 1, characterized in that, The method for identifying individual imbalanced radiation sources based on temporal attention also includes: The parameters of the entire feature extraction and classifier are optimized using the cross-entropy loss function; the expression for the cross-entropy loss function is: ； in, The cross-entropy loss function; B This refers to the training batch size; N The total number of radiation source individuals; For the first The true one-hot label of each sample, corresponding to the element position of the real individual. c The value is 1, and the rest are 0; It is the minimum value; Indicates the first b The sample belongs to the first c The probability of an individual radiation source.

3. The method for identifying individual imbalanced radiation sources based on temporal attention according to claim 1, characterized in that, Shallow temporal feature extraction is performed on the radiation source signal to obtain deep temporal high-order features, specifically including: The radiation source signal is subjected to feature extraction using 1D convolutional layers, batch normalization, and the ReLU activation function to obtain initial features; the formal expression of the initial features is as follows: ； in, Initial features; X IQ For radiation source signal; kernel size `core` is the kernel size; `stride` is the span; `padding` is the padding value; `out` channel For output channels; The initial features are downsampled using a MaxPool1d layer and then input into a four-layer residual layer for progressive feature extraction, yielding deep temporal high-order features. Each residual layer consists of two stacked 1D residual blocks, each containing a two-stage feature transformation of convolution, batch normalization, and ReLUctant. The formal expression of the residual block is as follows: ； Among them, F out Output features for residual blocks; and All are 1D convolutional layers with a kernel size of 3; For jump connection branches; The input features are for the residual block.

4. The method for identifying individual unbalanced radiation sources based on temporal attention according to claim 1, characterized in that, Multi-scale feature fusion is performed on the deep temporal high-order features to obtain fused features, specifically including: Four parallel branches are used to perform multi-scale processing on the deep temporal high-order features to obtain fused features. Branch 0 directly performs convolution, batch normalization, and ReLU activation on the deep temporal high-order features to extract the original scale features. Branches 1-3 perform average pooling downsampling with strides of 2, 4, and 8, respectively, and then extract multi-scale features through convolution, batch normalization, and ReLU activation. The features are then restored to their original length through linear upsampling interpolation. After feature concatenation, a preliminary fused feature is obtained. Finally, the channel dimension is fused and transformed through 1×1 convolution, batch normalization, and ReLU activation to obtain the final fused feature.

5. The method for identifying individual imbalanced radiation sources based on temporal attention according to claim 4, characterized in that, Design an identity-aware temporal attention module to generate identity-adaptive time-step attention weights, specifically including: Each identity tag is mapped to a corresponding identity embedding vector; the formal expression of the identity embedding vector is as follows: ； in, , For identity tags The corresponding one-hot encoded vector, N The total number of radiation source individuals. B This refers to the training batch size; This is the corresponding identity embedding vector; E For embedding matrix; Each identity embedding vector is aligned along the temporal dimension to obtain an aligned identity embedding vector; the formal expression of the aligned identity embedding vector is as follows: ； in, for L A dimensional vector of all 1s L This represents the original timing length. C For the embedded dimension; This represents the outer product operation. This indicates repeated operations based on a time sequence. The aligned identity embedding vector; The aligned identity embedding vector and the preliminary fused features are concatenated along the channel dimension to obtain the concatenated features; the formal expression of the concatenated features is as follows: ； in, Features after splicing; To splice along the channel dimension; Features of fusion; T For the total time step; 2 C This represents the number of feature channels after fusion. Based on the concatenated features, the number of channels in the preliminary fused features is first compressed from 2C to the hidden dimension D using a 1×1 convolution. This compression of the channel dimension reduces computational complexity without disrupting the temporal structure. Then, batch normalization and ReLU activation introduce non-linearity, followed by a second 1×1 convolution to compress the number of channels to 1, yielding the original attention score. The score is then mapped to the [0,1] interval using a Sigmoid activation function, ultimately resulting in the identity-adaptive time-step attention weights. The formal expression of these identity-adaptive time-step attention weights is as follows: ； in, For the first i The first sample t Attention weights for each time step.

6. The method for identifying individual imbalanced radiation sources based on temporal attention according to claim 1, characterized in that, The formal expression of the weighted features is as follows: ； in, The features are weighted; For the first i The first sample t Attention weights for each time step; f The output features are fused and transformed using 1×1 convolution, batch normalization, and ReLU activation function to achieve channel-dimensional fusion. B This refers to the training batch size; C For the embedded dimension; T This represents the total time step.

7. The method for identifying individual unbalanced radiation sources based on temporal attention according to claim 1, characterized in that, The weighted features are then aggregated into global identity features to obtain aggregated global identity features, specifically including: The weighted features are globally compressed using adaptive average pooling to obtain a global feature vector after adaptive average pooling; the formal expression of the global feature vector after adaptive average pooling is as follows: ； in, F pool This is the global feature vector after adaptive average pooling; The features are weighted; B This refers to the training batch size; The adaptive average pooling global feature vector is further compressed using 1D convolution and activation functions to obtain the aggregated global identity features; the formal expression of the aggregated global identity features is as follows: ； in, F conv1 The aggregated global identity features; in is the input; out is the output.

8. The method for identifying individual imbalanced radiation sources based on temporal attention according to claim 1, characterized in that, The formal expression of the probability distribution feature is as follows: ； in, F conv2 Characteristics of probability distribution; F conv1 The aggregated global identity features; in is the input; out is the output; B This is the training batch size.

9. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement the method for identifying individual unbalanced radiation sources based on temporal attention as described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the method for identifying individual unbalanced radiation sources based on temporal attention as described in any one of claims 1-8.

Images