Radiation source individual identification method based on time-frequency adaptive fusion of noise perception

Abstract

The application provides a noise perception-based time-frequency adaptive fusion radiation source individual identification method, and relates to the field of signal processing. The method comprises the following steps: improving a one-dimensional ResNet-18 network to extract features of an obtained original electromagnetic signal, and outputting a first enhanced feature after time attention module processing; mapping the original electromagnetic signal into a two-dimensional time-frequency feature map through short-time Fourier transform (STFT), extracting features through a lightweight ResNeSt-50 network, and outputting a second enhanced feature after frequency attention module processing; inputting the first enhanced feature and the second enhanced feature into a noise perception gate unit to process and output a time domain weight and a frequency domain weight; outputting an enhanced feature sequence under noise perception through adaptive feature weighting, and outputting a radiation source individual identification result through classification processing. The application realizes complementary utilization of time-frequency dual-domain features and noise robustness improvement, and ensures the stability and reliability of identification under different noise environments through adaptive weight distribution.

Description

Technical Field

[0001] This application relates to the field of signal processing technology, specifically to a method for individual radiation source identification based on noise-aware time-frequency adaptive fusion. Background Technology

[0002] Specific Emitter Identification (SEI) is a technology that distinguishes and identifies signal transmitters based on the differences in the radio frequency fingerprint (RFF) of the electromagnetic signals emitted by a radiation source. Due to inherent differences in the circuit design and component manufacturing of each signal transmitter, even transmitters of the same brand and model transmitting the same waveform will have subtle differences in their RRFF. SEI technology utilizes these subtle and unintentional differences to identify the transmitter. In the civilian sector, wireless communication has broadcast characteristics; wireless interfaces are public and easily shared by authorized and unauthorized users, making them more vulnerable to passive eavesdropping and active interference attacks compared to wired communication. With the rapid development of wireless devices and the Internet of Things (IoT), SEI technology is crucial for improving the security of wireless networks and the IoT. Therefore, improving the reliability of SEI algorithm models in real-world scenarios has become a critical issue that urgently needs to be addressed.

[0003] In related technologies, there are two main methods for identifying SEI (Self-Identifying Radiation Sources): Traditional SEI methods rely on manual extraction of signal features, specifically through bispectral analysis, cumulants, and time-frequency representation, and then input the extracted features into a classifier to complete the identification; With the development of deep learning technology, deep learning methods have been widely applied in the SEI field. Researchers use models such as Convolutional Neural Networks (CNN), Long Short-Term Memory Networks (LSTM), and Transformers to achieve automatic extraction of signal time-domain features. Compared with traditional methods, these methods significantly improve the recognition accuracy and robustness of SEI.

[0004] However, existing SEI (Self-Identification) technologies for individual radiation sources still have significant drawbacks: traditional methods rely on manual feature extraction, which is not only inefficient but also highly dependent on the operator's expertise, making it difficult to adapt to the needs of complex real-world scenarios; while deep learning methods have the advantage of automatic feature extraction, labeled signal data is severely insufficient in many practical applications, limiting the training effect of deep learning models and making it difficult to further improve recognition performance. Furthermore, existing technologies have not effectively solved the problem of fully utilizing unlabeled signal data, failing to meet the application requirements of high reliability and robustness of SEI technology in real-world scenarios. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this application provides a noise-aware time-frequency adaptive fusion method for identifying radiation sources. This method solves the problem that while deep learning can automatically extract features, existing radiation source identification technologies are limited by insufficient labeled data, making it difficult to fully utilize unlabeled data and meet the high reliability and robustness requirements of identification in real-world scenarios.

[0006] To achieve the above objectives, this application provides the following technical solution: In a first aspect, embodiments of this application provide a method for identifying individual radiation sources based on noise-aware time-frequency adaptive fusion. This method includes: constructing a time-domain feature extraction network and a frequency-domain feature extraction network; the time-domain feature extraction network includes an improved one-dimensional ResNet-18 network and a time attention module; the frequency-domain feature extraction network includes a lightweight ResNeSt-50 network and a frequency attention module; extracting features from the acquired raw electromagnetic signal using the improved one-dimensional ResNet-18 network, aggregating temporal information and suppressing redundant noise through the time attention module, modeling feature correlation, and outputting a first enhanced feature; and based on the constructed frequency-domain feature extraction network, performing short-time Fourier transform... STFT maps the original electromagnetic signal into a two-dimensional time-frequency feature map. Feature extraction is performed through a lightweight ResNeSt-50 network. Frequency dimension information is aggregated by a frequency attention module, and the feature evolution process is modeled to output a second enhanced feature. The first and second enhanced features are input into a noise-aware gating unit, which performs global average pooling and feature concatenation. Normalized time-domain and frequency-domain weights are learned and output through a linear layer and a Softmax activation function to determine the weight allocation under different noise conditions. Based on the time-domain and frequency-domain weights, adaptive feature weighting is performed through a channel attention module to output an enhanced feature sequence under noise awareness, so as to output the individual radiation source identification result through classification processing.

[0007] According to a first aspect of the embodiments of this application, the aforementioned feature extraction of the acquired original electromagnetic signal using an improved one-dimensional ResNet-18 network, the aggregation of temporal information and suppression of redundant noise by a time attention module, and the modeling of feature correlation to output a first enhanced feature, includes: using an improved one-dimensional ResNet-18 network to extract features from the original electromagnetic signal to obtain one-dimensional feature representation information of the signal; inputting the one-dimensional feature representation information into the time attention module, performing global average pooling and global max pooling respectively, and after concatenating and fusing the two pooling results, generating a first attention weight through 3×1 convolution and a Sigmoid activation function; weighting the one-dimensional feature representation information element-wise based on the first attention weight, suppressing temporal redundancy and random noise interference, and modeling feature correlation in the frequency dimension to enhance the target spectral features in the spectral structure with discrimination ability exceeding a preset discrimination threshold in the entire time range, and outputting the first enhanced feature.

[0008] According to a first aspect of the embodiments of this application, the improved one-dimensional ResNet-18 network replaces the first convolution of the standard one-dimensional ResNet-18 network with a dilated convolution, and embeds a CBAM attention mechanism between the convolutional layers of each residual block; the dilation rate of the dilated convolution is set according to the signal feature extraction requirements to increase the receptive field of the convolution; the CBAM attention mechanism includes a channel attention module and a spatial attention module, which sequentially perform channel-dimensional and spatial-dimensional attention weighting on the feature map.

[0009] According to a first aspect of the embodiments of this application, the aforementioned frequency domain feature extraction network, through a Short Time Fourier Transform (STFT), maps the original electromagnetic signal into a two-dimensional time-frequency feature map, extracts features through a lightweight ResNeSt-50 network, aggregates frequency dimension information through a frequency attention module, models the feature evolution process, and outputs a second enhanced feature, including: mapping the original electromagnetic signal from a one-dimensional time series to a two-dimensional time-frequency representation through a Short Time Fourier Transform (STFT) to obtain a two-dimensional time-frequency feature map; and inputting the two-dimensional time-frequency feature map into the constructed lightweight ResNeSt-50 network for further processing. Feature extraction is performed to obtain three-dimensional feature representations of the signal in the channel, frequency, and time dimensions. The three-dimensional feature representations are then processed by a frequency attention module, performing global average pooling and global max pooling respectively. After concatenating and fusing the two pooling results, a second attention weight is generated by a 1×3 convolution and a sigmoid activation function. The three-dimensional feature representations are then weighted element-wise based on the second attention weight to statistically aggregate the frequency dimension information. The feature evolution process is modeled in the time dimension to enhance the dynamic representation ability of the spectral components as they change over time, and the second enhanced feature is output.

[0010] According to a first aspect of the embodiments of this application, the lightweight ResNeSt-50 network retains the original design of cardinality grouping and split subgroups of the ResNeSt-50 network to achieve multi-level feature extraction and attention-weighted fusion through a multi-branch structure, while retaining the original multi-branch feature aggregation capability; the lightweight ResNeSt-50 network reduces the number of convolutional layers in each residual block of the ResNeSt-50 network from 4 layers to 2 layers to reduce the number of parameters and computational complexity.

[0011] According to a first aspect of the embodiments of this application, the aforementioned method for individual radiation source identification based on noise-aware time-frequency adaptive fusion further includes: periodically performing preset floating-point multiplication and addition operations to perform floating-point operation self-checks, comparing the operation self-check results with preset theoretical values ​​to generate a precision risk index; embedding an independent digital logic module at the output end of at least one residual block of an improved one-dimensional ResNet-18 network and a lightweight ResNeSt-50 network to perform precision sensitivity assessment; receiving feature data generated by the previous residual block in real time through the independent digital logic module, analyzing the numerical range, variation amplitude, and distribution concentration of the feature data, and performing a sensitivity assessment of the current feature in conjunction with the precision risk index; and, if it is determined that the current feature exceeds a preset sensitivity threshold, performing numerical representation adjustment and feature smoothing processing on the residual block downstream of the independent digital logic module.

[0012] According to a first aspect of the embodiments of this application, the aforementioned numerical representation adjustment includes: sending a precision control signal to the residual block located downstream of the independent digital logic module, dynamically switching the operation precision mode, and using 16-bit floating-point numbers or fixed-point numbers for feature calculation; the aforementioned feature smoothing process includes: calling a smoothing operator to filter and reduce noise on the feature map output by the residual block upstream of the independent digital logic module, so as to eliminate the numerical deviation introduced by the precision loss of the mantissa of floating-point operations; wherein, the smoothing operator uses a moving average filter on the time-domain feature map and a 3×3 average pooling filter on the frequency-domain feature map.

[0013] Secondly, embodiments of this application provide a radiation source individual identification system based on noise perception and time-frequency adaptive fusion. The radiation source individual identification system includes: a network construction module, a first feature enhancement module, a second feature enhancement module, a weight allocation module, and a weighted classification module.

[0014] Specifically, the network construction module is used to construct a time-domain feature extraction network and a frequency-domain feature extraction network. The time-domain feature extraction network includes an improved one-dimensional ResNet-18 network and a time attention module; the frequency-domain feature extraction network includes a lightweight ResNeSt-50 network and a frequency attention module. The first feature enhancement module is used to extract features from the acquired raw electromagnetic signal through the improved one-dimensional ResNet-18 network, aggregate time-series information, suppress redundant noise, and model feature correlations through the time attention module, outputting the first enhanced feature. The second feature enhancement module is used to map the raw electromagnetic signal into a two-dimensional time-frequency feature based on the constructed frequency-domain feature extraction network through a short-time Fourier transform (STFT). The first feature extraction module extracts features using a lightweight ResNeSt-50 network, aggregates frequency dimension information via a frequency attention module, models the feature evolution process, and outputs a second enhanced feature. The second feature allocation module inputs the first and second enhanced features into a noise-aware gating unit, performs global average pooling and feature concatenation, and learns and outputs normalized temporal and frequency domain weights through a linear layer and a Softmax activation function to determine weight allocation under different noise conditions. The third feature weighting module, based on the temporal and frequency domain weights, performs adaptive feature weighting through a channel attention module, outputting an enhanced feature sequence under noise awareness, which is then used to classify and identify individual radiation sources.

[0015] Thirdly, embodiments of this application provide an electronic device, which includes: a processor, a memory, and a program stored in the memory and executable on the processor. When the program is executed by the processor, it implements the noise-aware time-frequency adaptive fusion radiation source individual identification method described in the first aspect above.

[0016] Fourthly, embodiments of this application provide a computer-readable storage medium storing a program or instructions that, when executed by a processor, implement the noise-aware time-frequency adaptive fusion-based radiation source individual identification method described in the first aspect above.

[0017] This application provides a noise-aware time-frequency adaptive fusion method for individual radiation source identification. Compared with existing technologies, it has the following advantages: This application constructs a temporal feature extraction network comprising an improved one-dimensional ResNet-18 network and a temporal attention module, and a frequency domain feature extraction network comprising a lightweight ResNeSt-50 network and a frequency attention module. These networks can effectively extract temporal and frequency-dimensional features from the original electromagnetic signal. The temporal attention module aggregates temporal information, suppresses redundant noise, and models feature correlations, while the frequency attention module aggregates frequency-dimensional information and models the feature evolution process, achieving initial enhancement of both temporal and frequency domain features. Based on this, a noise-aware gating unit performs global average pooling and feature concatenation on the two enhanced features. Normalized temporal and frequency domain weights are learned through a linear layer and a Softmax activation function, enabling adaptive determination of the fusion weight allocation for the two features under different noise conditions. Furthermore, a channel attention module adaptively weights the two features, outputting an enhanced feature sequence under noise awareness for classification to obtain individual radiation source identification results. This achieves complementary utilization of temporal and frequency domain features and improved noise robustness, while the adaptive weight allocation ensures the stability and reliability of identification under different noise environments. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a method for identifying individual radiation sources based on noise-sensing time-frequency adaptive fusion, as provided in an embodiment of this application. Figure 2 This is an exemplary architecture diagram corresponding to a noise-aware time-frequency adaptive fusion-based radiation source individual identification method provided in this application embodiment; Figure 3 This is an exemplary structural diagram of a noise sensing gating unit provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of a radiation source individual identification system based on noise perception and time-frequency adaptive fusion provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0022] This application provides a noise-aware time-frequency adaptive fusion method for identifying individual radiation sources. This solves the problem in existing individual radiation source identification technologies that, although deep learning can automatically extract features, it is limited by insufficient labeled data, making it impossible to fully utilize unlabeled data and meet the high reliability and robustness requirements of identification in real-world scenarios.

[0023] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0024] The following section first introduces a method for individual radiation source identification based on noise perception and time-frequency adaptive fusion, provided in the embodiments of this application.

[0025] This application provides a flowchart illustrating a noise-aware time-frequency adaptive fusion-based radiation source individual identification method, as shown in the embodiments below. Figure 1 As shown, the noise-aware time-frequency adaptive fusion radiation source individual identification method may include the following steps S110-S150.

[0026] S110. Construct a temporal feature extraction network and a frequency domain feature extraction network. The temporal feature extraction network includes an improved one-dimensional ResNet-18 network and a temporal attention module; the frequency domain feature extraction network includes a lightweight ResNeSt-50 network and a frequency attention module.

[0027] Understandably, this application constructs a temporal feature extraction network that includes an improved one-dimensional ResNet-18 network and a temporal attention module, and a frequency domain feature extraction network that includes a lightweight ResNeSt-50 network and a frequency attention module. This provides a dedicated and adaptable network architecture for the subsequent extraction and enhancement of temporal and frequency domain features of the original electromagnetic signal. It clarifies the core module composition of dual-domain feature extraction, provides a basis for the complementary use of time and frequency dual-domain features, and the lightweight ResNeSt-50 network balances feature extraction capability and computational efficiency, while the improved one-dimensional ResNet-18 network adapts to the feature extraction requirements of one-dimensional original electromagnetic signals.

[0028] S120. The original electromagnetic signal is extracted by improving the one-dimensional ResNet-18 network. The temporal information is aggregated by the time attention module, redundant noise is suppressed, and feature correlation is modeled to output the first enhanced feature.

[0029] Understandably, this application improves the one-dimensional ResNet-18 network to extract features from the original electromagnetic signal, which can effectively capture the one-dimensional temporal features of the original electromagnetic signal. Then, by aggregating the temporal information through the time attention module, redundant noise in the signal can be effectively suppressed and interference from invalid information can be reduced. At the same time, by modeling feature correlation, the effective information in the temporal features can be strengthened, and the first enhanced feature is output, thus realizing the initial enhancement and purification of the temporal features and improving the effectiveness and discriminative ability of the temporal features.

[0030] S130. Based on the constructed frequency domain feature extraction network, the original electromagnetic signal is mapped into a two-dimensional time-frequency feature map through short-time Fourier transform (STFT). Feature extraction is performed through a lightweight ResNeSt-50 network. Frequency dimension information is aggregated through a frequency attention module, and the feature evolution process is modeled to output the second enhanced feature.

[0031] Understandably, this application maps the original electromagnetic signal into a two-dimensional time-frequency feature map through Short Time Fourier Transform (STFT), realizing the transformation from a one-dimensional original signal to a two-dimensional time-frequency feature map, which can fully preserve the frequency dimension information of the signal. Then, feature extraction is performed through a lightweight ResNeSt-50 network, which can efficiently capture the frequency dimension information in the two-dimensional time-frequency feature map, balancing feature extraction accuracy and computational efficiency. After aggregating the frequency dimension information and modeling the feature evolution process through the frequency attention module, the effective frequency dimension features can be strengthened, the evolution law of features over time can be captured, and a second enhanced feature can be output, realizing the initial enhancement and purification of frequency domain features, and improving the effectiveness and representation ability of frequency domain features.

[0032] S140. Input the first and second enhanced features into the constructed noise perception gating unit, perform global average pooling and feature concatenation, learn and output normalized time-domain weights and frequency-domain weights through a linear layer and a Softmax activation function, so as to determine the weight allocation under different noise conditions.

[0033] Please refer to the above as well. Figure 2 and Figure 3 Understandably, this application inputs the first and second enhancement features into the noise perception gating unit. By performing global average pooling, global information statistical analysis can be performed on the dual-domain enhancement features, and the weight allocation relationship between the two features under the current signal conditions can be adaptively learned, thereby dynamically adjusting the contribution of different time-frequency features in the final fusion process. Specifically, feature concatenation can integrate the global information of dual-domain features; then, through a linear layer and a Softmax activation function, normalized time-domain weights and frequency-domain weights are learned and output, which can adaptively determine the fusion weight allocation of time-domain and frequency-domain features under different noise conditions, making the weight allocation adaptable to the current noise environment, avoiding the dominance of a single feature branch in the fusion process, and enabling the model to obtain stable and reliable feature representations under different noise environments. This provides a reasonable weight basis for the effective fusion of subsequent dual-domain features, improving the targeting and adaptability of feature fusion.

[0034] S150: Based on time-domain weights and frequency-domain weights, adaptive feature weighting is performed through the channel attention module to output an enhanced feature sequence under noise perception, so as to output the individual identification result of the radiation source through classification processing.

[0035] Understandably, based on the time-domain and frequency-domain weights obtained from S140, adaptive feature weighting through the channel attention module can further enhance the effective features and suppress invalid noise in the dual-domain fusion process, outputting an enhanced feature sequence under noise perception, ensuring that the enhanced feature sequence has good effectiveness and robustness under different noise conditions; by classifying the enhanced feature sequence, the individual radiation source identification results can be accurately output, ensuring the reliability of the identification results and realizing the effective identification of individual radiation sources in noisy environments.

[0036] In some embodiments, the aforementioned feature extraction of the acquired raw electromagnetic signal using an improved one-dimensional ResNet-18 network, the aggregation of temporal information and suppression of redundant noise by a time attention module, and the modeling of feature correlation, outputting a first enhanced feature, that is, the aforementioned S120 may specifically include the following steps: S210. An improved one-dimensional ResNet-18 network is used to extract features from the original electromagnetic signal to obtain one-dimensional feature representation information of the signal. S220. Input the one-dimensional feature representation information into the temporal attention module, and perform global average pooling and global max pooling respectively. After concatenating and fusing the two pooling results, generate the first attention weight through 3×1 convolution and Sigmoid activation function. S230. The one-dimensional feature representation information is weighted element by element based on the first attention weight to suppress temporal redundancy and random noise interference. Feature correlation is modeled on the frequency dimension to enhance the target spectral features in the spectral structure of the whole time range that have a discrimination ability exceeding the preset discrimination threshold, and the first enhanced feature is output.

[0037] In the embodiments of this application, it is understood that this application improves the one-dimensional ResNet-18 network to extract features from the original electromagnetic signal, effectively capturing the one-dimensional features of the original electromagnetic signal and forming one-dimensional feature representation information, providing reliable basic feature support for subsequent time-domain feature enhancement processing. This application inputs the one-dimensional feature representation information into the time attention module, and by performing global average pooling and global max pooling respectively, it can comprehensively capture the global statistical information and key extreme value information in the one-dimensional feature representation information. The concatenation and fusion of the two pooling results can integrate multi-dimensional statistical features, and then generate the first attention weight through 3×1 convolution and Sigmoid activation function, providing accurate weight basis for subsequent feature weighting. Furthermore, this application performs element-wise weighting of the one-dimensional feature representation information based on the first attention weight, which can effectively suppress temporal redundancy information and random noise interference. At the same time, by modeling feature correlation in the frequency dimension, it strengthens the target spectral features in the spectral structure of the entire time range whose discrimination ability exceeds the preset discrimination threshold, and finally outputs the first enhanced feature, improving the effectiveness, purity and discrimination ability of the feature.

[0038] In some embodiments, the improved one-dimensional ResNet-18 network replaces the first convolution of the standard one-dimensional ResNet-18 network with a dilated convolution, and embeds a CBAM attention mechanism between the convolutional layers of each residual block; the dilation rate of the dilated convolution is set according to the signal feature extraction requirements to increase the receptive field of the convolution; the CBAM attention mechanism includes a channel attention module and a spatial attention module, which sequentially apply attention weights to the feature map in the channel dimension and the spatial dimension.

[0039] In the embodiments of this application, it is understood that by replacing the first convolution of the standard one-dimensional ResNet-18 network with a dilated convolution and setting the corresponding dilation rate according to the signal feature extraction requirements to increase the receptive field of the convolution, this application can capture the long-distance dependence and global feature information of the original electromagnetic signal more comprehensively without reducing the feature resolution or increasing the additional computational overhead. At the same time, a CBAM attention mechanism containing channel attention module and spatial attention module is embedded between the convolutional layers of each residual block. The feature map can be adaptively weighted by attention in the channel dimension and spatial dimension in turn, which can effectively strengthen key features, suppress redundant noise and invalid information, and further improve the feature extraction accuracy, expressive power and anti-interference performance of the one-dimensional ResNet-18 network, making the network more adaptable to the feature extraction requirements in complex electromagnetic signal scenarios.

[0040] In some embodiments, the aforementioned frequency domain feature extraction network maps the original electromagnetic signal into a two-dimensional time-frequency feature map through a short-time Fourier transform (STFT), performs feature extraction through a lightweight ResNeSt-50 network, aggregates frequency dimension information through a frequency attention module, models the feature evolution process, and outputs a second enhanced feature. Specifically, the aforementioned S130 may include the following steps: S310. The original electromagnetic signal is mapped from a one-dimensional time series to a two-dimensional time-frequency representation by the short-time Fourier transform (STFT), resulting in a two-dimensional time-frequency feature map. S320. Input the two-dimensional time-frequency feature map into the constructed lightweight ResNeSt-50 network for feature extraction to obtain the three-dimensional feature representation information of the signal in the channel dimension, frequency dimension and time dimension. S330: Based on the frequency attention module, the three-dimensional feature representation information is processed by performing global average pooling and global max pooling respectively. After the two pooling results are concatenated and fused, a second attention weight is generated by 1×3 convolution and Sigmoid activation function. S340. Based on the second attention weight, the three-dimensional feature representation information is weighted element by element to statistically aggregate the frequency dimension information, and the feature evolution process is modeled in the time dimension to enhance the dynamic representation ability of the spectral components changing over time, and the second enhanced feature is output.

[0041] In the embodiments of this application, it is understood that this application uses Short Time Fourier Transform (STFT) to map a one-dimensional original electromagnetic signal into a two-dimensional time-frequency feature map, which can simultaneously retain the detailed information of the signal in the time domain and frequency domain, and realize the effective conversion from one-dimensional time series to two-dimensional joint representation. This application inputs the two-dimensional time-frequency feature map into a lightweight ResNeSt-50 network for feature extraction, which can efficiently output three-dimensional feature representation information that simultaneously contains channel, frequency and time dimensions, taking into account the advantages of multi-scale aggregation capability and computational lightweight.

[0042] Furthermore, in the frequency attention module, this application performs global average pooling and global max pooling on the three-dimensional feature representation information and concatenates and fuses them. A second attention weight is generated through 1×3 convolution and sigmoid activation, which can accurately capture key information on global frequency statistics and extreme values, forming a stable and adaptively adjustable frequency weight distribution. Then, based on the second attention weight, this application weights the three-dimensional feature representation information element-wise, achieving effective statistical aggregation of frequency-dimensional information and modeling the feature evolution process in the time dimension, enhancing the dynamic representation ability of spectral components changing over time. Finally, it outputs a highly robust and discriminative second enhanced feature, effectively improving the utilization rate of dynamic features in the frequency domain and noise adaptability.

[0043] In some embodiments, the lightweight ResNeSt-50 network retains the original cardinality grouping and split subgrouping design of the ResNeSt-50 network to achieve multi-level feature extraction and attention-weighted fusion through a multi-branch structure, while retaining the original multi-branch feature aggregation capability. The lightweight ResNeSt-50 network reduces the number of convolutional layers in each residual block of the ResNeSt-50 network from 4 layers to 2 layers to reduce the number of parameters and computational complexity.

[0044] In the embodiments of this application, it is understood that by retaining the original cardinality grouping and split subgroup design of ResNeSt-50, the lightweight ResNeSt-50 network can achieve multi-level feature extraction and attention-weighted fusion based on the multi-branch structure, fully retaining the original network's multi-scale and multi-semantic feature aggregation capabilities in channel splitting and cross-branch interaction; at the same time, the number of convolutional layers in each residual block is reduced from 4 layers to 2 layers, thereby reducing the number of parameters and computational complexity. Without destroying the multi-branch feature aggregation framework and expression upper limit, it effectively reduces the number of network parameters and floating-point computation, reduces inference storage overhead and latency, and balances strong representation ability, high feature utilization and low computational complexity, making it more suitable for edge or real-time electromagnetic signal feature extraction scenarios.

[0045] In some embodiments, the aforementioned noise-aware time-frequency adaptive fusion-based radiation source individual identification method may further include the following steps: S410. Periodically perform preset floating-point multiplication and addition operations to perform floating-point operation self-checks, and compare the operation self-check results with preset theoretical values ​​to generate an accuracy risk index; S420. An independent digital logic module is embedded at the output of at least one residual block of the improved one-dimensional ResNet-18 network and the lightweight ResNeSt-50 network to perform accuracy sensitivity evaluation. S430: Receives feature data generated by the previous residual block in real time through an independent digital logic module, analyzes the numerical range, variation amplitude, and distribution concentration of the feature data, and performs a sensitivity assessment of the current feature in conjunction with the accuracy risk index; S440. If it is determined that the current feature exceeds the preset sensitivity threshold, the residual block downstream of the independent digital logic module is subjected to numerical representation adjustment and feature smoothing processing.

[0046] In this embodiment, it should be noted that, under the background of long-term uninterrupted operation of the radiation source individual identification system, the electrical characteristics of key internal components of the power management unit voltage regulator module, which provides power to the digital signal processor and its internal floating-point arithmetic unit, as well as the feature extraction network dedicated computing chip, may experience slight drift. This drift is not a sudden fault, but a slow, cumulative performance degradation, the magnitude of which may be much smaller than the fault threshold set by the system, and therefore difficult to detect in a timely manner during routine system health monitoring. This slow, cumulative performance degradation causes slight fluctuations in the power supply voltage that exceed the design tolerance. These slight voltage fluctuations, under specific high load or high temperature environments, directly affect the stable operating point of the floating-point arithmetic unit inside the computing chip. The design of the floating-point arithmetic unit has strict requirements for the stability of the power supply voltage; even slight deviations in voltage may cause subtle changes in the switching characteristics, timing margins, and analog reference voltage of its internal transistors. This, in turn, leads to a systematic loss of the last digit precision of the floating-point arithmetic result.

[0047] When these feature maps with slight precision loss enter the temporal or frequency attention module for global average pooling, since global average pooling essentially averages all elements of the feature map, if each element has a systematic loss of last-order precision, this tendency to be small will accumulate and be reflected in the final descriptor vector after summing and averaging. Therefore, these small accumulated errors are amplified, leading to a systematic bias in the global statistics of the extracted temporal and frequency domain descriptor vectors. Ultimately, the noise-aware gating unit, upon receiving these systematically biased descriptor vectors, will make slight misjudgments, causing the output fusion weights to no longer fully and accurately reflect the true reliability of the temporal and frequency domain features. This results in the model failing to fully achieve its noise-aware purpose in low signal-to-noise ratio or complex noise environments, leading to intermittent or cumulative decreases in recognition accuracy and robustness.

[0048] Based on this, this application periodically performs preset floating-point operations for self-checking and compares the results to generate a precision risk index. This allows for real-time quantification of the numerical stability and error accumulation trend of the underlying computing units, providing a global benchmark for subsequent precision sensitivity assessment. This application embeds an independent digital logic module at the output of at least one residual block, which can accurately locate key detection nodes in the feature propagation chain and achieve hierarchical and controllable online monitoring. The independent digital logic module receives the preceding feature data in real time and analyzes the numerical range, variation amplitude, and distribution concentration. Simultaneously, it combines the precision risk index to conduct joint sensitivity assessment, which can take into account both hardware floating-point drift and feature space perturbation, improving the reliability of anomaly identification. This allows for adjustment processing when the current feature exceeds the preset sensitivity threshold.

[0049] Furthermore, by making real-time and flexible adjustments during the feature generation stage, the feature maps output by the time-domain and frequency-domain feature extraction networks will have stronger error resistance from the outset. Even if there is a slight drift in the voltage regulation module of the power management unit, resulting in a systematic loss of accuracy, the statistical information carried by these "self-protected" feature maps will be closer to the real situation when they enter the global average pooling and noise-aware gating units. This allows the noise-aware gating units to more accurately judge the authenticity and reliability of the time-domain and frequency-domain features and generate appropriate fusion weights. Ultimately, this ensures that the system can maintain a high level of recognition accuracy and robustness even in low signal-to-noise ratio and complex noise environments.

[0050] In some embodiments, numerical representation adjustment includes: sending a precision control signal to the residual block located downstream of the independent digital logic module, dynamically switching the operation precision mode, and using 16-bit floating-point numbers or fixed-point numbers for feature calculation.

[0051] Feature smoothing processing includes: calling a smoothing operator to filter and reduce noise on the feature map output by the residual block upstream of the independent digital logic module in order to eliminate the numerical deviation introduced by the loss of precision of the mantissa in floating-point operations; wherein, the smoothing operator uses a moving average filter on the time domain feature map and a 3×3 average pooling filter on the frequency domain feature map.

[0052] In the embodiments of this application, it is understood that this application sends a precision control signal to the residual block located downstream of the independent digital logic module. This signal instructs subsequent computing units to preferentially use a numerical representation method that is less sensitive to precision loss when performing floating-point operations such as multiplication and addition. For example, if the hardware supports it, standard 32-bit floating-point operations can be dynamically switched to 16-bit floating-point operations, or in certain specific scenarios, floating-point numbers can be converted to fixed-point numbers for operations. Although 16-bit floating-point numbers or fixed-point numbers have relatively low representation range and precision, their least significant bits are less affected by small drifts when the values ​​are very small or do not change much, and the calculation speed is faster, which helps to avoid the cumulative effect of precision loss in the last bit.

[0053] Furthermore, this application employs a smoothing operator to perform targeted filtering and noise reduction on the feature maps output by the upstream residual block. Moving average filtering is used for the time-domain feature maps, and 3×3 average pooling filtering is used for the frequency-domain feature maps. This effectively smooths out high-frequency numerical glitches and local offset deviations introduced by the loss of floating-point mantissa precision, suppresses the layer-by-layer amplification of errors and the activation of pseudo-textures, stabilizes the feature distribution, and improves the network's numerical robustness and recognition consistency under low-precision inference and noise disturbances. This feature smoothing process is like "smoothing out the edges" of the feature data, effectively averaging out the small, isolated numerical deviations caused by the loss of last-digit precision. This makes the overall trend and macroscopic information of the feature maps more stable, reducing the amplifying effect of global average pooling operations on these small errors.

[0054] In some embodiments, this application provides a noise-aware time-frequency adaptive fusion-based radiation source individual identification system 500, such as... Figure 4 As shown, the noise-aware time-frequency adaptive fusion radiation source individual identification system 500 may include the following modules: Network building module 510 is used to build a temporal feature extraction network and a frequency domain feature extraction network. The temporal feature extraction network includes an improved one-dimensional ResNet-18 network and a temporal attention module; the frequency domain feature extraction network includes a lightweight ResNeSt-50 network and a frequency attention module. The first feature enhancement module 520 is used to extract features from the acquired original electromagnetic signal by improving the one-dimensional ResNet-18 network, aggregate temporal information and suppress redundant noise through the time attention module, model feature correlation, and output the first enhanced feature. The second feature enhancement module 530 is used to map the original electromagnetic signal into a two-dimensional time-frequency feature map through the short-time Fourier transform (STFT) based on the constructed frequency domain feature extraction network, extract features through a lightweight ResNeSt-50 network, aggregate frequency dimension information through the frequency attention module, model the feature evolution process, and output the second enhanced feature. The weight allocation module 540 is used to input the first and second enhanced features into the noise-aware gating unit, perform global average pooling and feature concatenation, learn and output normalized time-domain weights and frequency-domain weights through a linear layer and a Softmax activation function, so as to determine the weight allocation under different noise conditions. The weighted classification module 550 is used to perform adaptive feature weighting based on time domain weights and frequency domain weights through the channel attention module, and output an enhanced feature sequence under noise perception, so as to output the individual identification result of radiation source through classification processing.

[0055] According to embodiments of this application, any and multiple modules among the network construction module 510, the first feature enhancement module 520, the second feature enhancement module 530, the weight allocation module 540, and the weighted classification module 550 can be merged into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module.

[0056] Figure 4 Each module in the system shown has the function of implementing each step in the aforementioned noise-aware time-frequency adaptive fusion radiation source individual identification method, and can achieve its corresponding technical effect. For the sake of brevity, it will not be elaborated here.

[0057] In some embodiments, this application provides an electronic device, the structural schematic of which is shown below. Figure 5 As shown.

[0058] The electronic device may include a processor 610 and a memory 620 storing computer program instructions.

[0059] Specifically, the processor 610 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0060] Memory 620 may include mass storage for data or instructions. For example, and not limitingly, memory 620 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 620 may include removable or non-removable (or fixed) media. Where appropriate, memory 620 may be internal or external to the integrated gateway disaster recovery device. In a particular embodiment, memory 620 is non-volatile solid-state memory.

[0061] Memory 620 may include read-only memory (ROM), random access memory (RAM), disk storage media device, optical storage media device, flash memory device, electrical, optical, or other physical / tangible memory storage device. Therefore, typically, memory 620 includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it can perform the operations described in any of the noise-aware time-frequency adaptive fusion-based radiation source individual identification methods in the above embodiments.

[0062] The processor 610 reads and executes computer program instructions stored in the memory 620 to implement any of the noise-aware time-frequency adaptive fusion radiation source individual identification methods in the above embodiments.

[0063] In one example, the electronic device may also include a communication interface 630 and a bus 600. Wherein, such as Figure 5 As shown, the processor 610, memory 620, and communication interface 630 are connected via bus 600 and communicate with each other.

[0064] The communication interface 630 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application. Bus 600 includes hardware, software, or both, that couples components of an online data traffic metering device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 600 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, any suitable bus or interconnect is contemplated herein.

[0065] Furthermore, in conjunction with the noise-aware time-frequency adaptive fusion-based radiation source individual identification method in the above embodiments, this application embodiment can provide a computer storage medium for implementation. The computer storage medium stores computer program instructions; when these computer program instructions are executed by a processor, they implement any of the noise-aware time-frequency adaptive fusion-based radiation source individual identification methods in the above embodiments.

[0066] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0067] The functional blocks shown in the above block diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0068] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0069] The aspects of this disclosure have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by special-purpose hardware performing the specified functions or actions, or can be implemented by a combination of special-purpose hardware and computer instructions.

[0070] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for individual radiation source identification based on noise-aware time-frequency adaptive fusion, characterized in that, include: Construct a temporal feature extraction network and a frequency domain feature extraction network, wherein the temporal feature extraction network includes an improved one-dimensional ResNet-18 network and a temporal attention module; The frequency domain feature extraction network includes a lightweight ResNeSt-50 network and a frequency attention module; The improved one-dimensional ResNet-18 network is used to extract features from the acquired raw electromagnetic signal. The temporal information is aggregated and redundant noise is suppressed by the time attention module. The feature correlation is modeled and the first enhanced feature is output. Based on the constructed frequency domain feature extraction network, the original electromagnetic signal is mapped into a two-dimensional time-frequency feature map through short-time Fourier transform (STFT). Feature extraction is performed through a lightweight ResNeSt-50 network. Frequency dimension information is aggregated through a frequency attention module, and the feature evolution process is modeled to output a second enhanced feature. The noise-aware gating unit is constructed by inputting the first and second enhanced features, and performs global average pooling and feature concatenation. It learns and outputs normalized time-domain weights and frequency-domain weights through a linear layer and a Softmax activation function to determine the weight allocation under different noise conditions. Based on the time-domain weights and the frequency-domain weights, adaptive feature weighting is performed through the channel attention module to output an enhanced feature sequence under noise perception, so as to output the individual identification result of the radiation source through classification processing.

2. The radiation source individual identification method based on noise perception and time-frequency adaptive fusion as described in claim 1, characterized in that, The improved one-dimensional ResNet-18 network is used to extract features from the acquired raw electromagnetic signal. A time-attention module aggregates temporal information, suppresses redundant noise, and models feature correlations to output the first enhanced feature, including: An improved one-dimensional ResNet-18 network is used to extract features from the original electromagnetic signal to obtain one-dimensional feature representation information of the signal. The one-dimensional feature representation information is input into the temporal attention module, and global average pooling and global max pooling are performed respectively. After the two pooling results are concatenated and fused, the first attention weight is generated by 3×1 convolution and Sigmoid activation function. The one-dimensional feature representation information is weighted element-wise based on the first attention weight to suppress temporal redundancy and random noise interference, and feature correlation is modeled on the frequency dimension to enhance the target spectral features in the spectral structure with discrimination ability exceeding the preset discrimination threshold in the whole time range, and output the first enhanced feature.

3. The radiation source individual identification method based on noise perception and time-frequency adaptive fusion as described in claim 2, characterized in that, The improved one-dimensional ResNet-18 network replaces the first convolution of the standard one-dimensional ResNet-18 network with a dilated convolution, and embeds a CBAM attention mechanism between the convolutional layers of each residual block; the dilation rate of the dilated convolution is set according to the signal feature extraction requirements to increase the receptive field of the convolution; the CBAM attention mechanism includes a channel attention module and a spatial attention module, which sequentially apply attention weights to the feature map in the channel dimension and the spatial dimension.

4. The radiation source individual identification method based on noise perception and time-frequency adaptive fusion as described in claim 1, characterized in that, The constructed frequency domain feature extraction network maps the original electromagnetic signal into a two-dimensional time-frequency feature map through Short Time Fourier Transform (STFT), extracts features through a lightweight ResNeSt-50 network, aggregates frequency dimension information through a frequency attention module, models the feature evolution process, and outputs a second enhanced feature, including: The original electromagnetic signal is mapped from a one-dimensional time series to a two-dimensional time-frequency representation by using the short-time Fourier transform (STFT), resulting in a two-dimensional time-frequency feature map. The two-dimensional time-frequency feature map is input into the constructed lightweight ResNeSt-50 network for feature extraction, and the three-dimensional feature representation information of the signal corresponding to the channel dimension, frequency dimension and time dimension is obtained. The three-dimensional feature representation information is processed based on the frequency attention module, and global average pooling and global max pooling are performed respectively. After the two pooling results are concatenated and fused, a second attention weight is generated by 1×3 convolution and Sigmoid activation function. The three-dimensional feature representation information is weighted element-wise based on the second attention weight to statistically aggregate the frequency dimension information, and the feature evolution process is modeled in the time dimension to enhance the dynamic representation ability of the spectral components changing over time, and output the second enhanced feature.

5. The radiation source individual identification method based on noise perception and time-frequency adaptive fusion as described in claim 4, characterized in that, The lightweight ResNeSt-50 network retains the original cardinality grouping and split subgrouping design of the ResNeSt-50 network to achieve multi-level feature extraction and attention-weighted fusion through a multi-branch structure, while retaining the original multi-branch feature aggregation capability. The lightweight ResNeSt-50 network reduces the number of convolutional layers per residual block of the ResNeSt-50 network from 4 to 2, thereby reducing the number of parameters and computational complexity.

6. The radiation source individual identification method based on noise perception and time-frequency adaptive fusion as described in claim 1, characterized in that, Also includes: Periodically perform preset floating-point multiplication and addition operations to perform floating-point operation self-checks, and compare the operation self-check results with preset theoretical values ​​to generate an accuracy risk index; Independent digital logic modules are embedded at the output of at least one residual block of the improved one-dimensional ResNet-18 network and the lightweight ResNeSt-50 network to perform accuracy sensitivity evaluation. The independent digital logic module receives feature data generated by the previous residual block in real time, analyzes the numerical range, variation amplitude and distribution concentration of the feature data, and performs a sensitivity assessment of the current feature in conjunction with the accuracy risk index. If it is determined that the current feature exceeds the preset sensitivity threshold, the residual block downstream of the independent digital logic module is subjected to numerical representation adjustment and feature smoothing processing.

7. The radiation source individual identification method based on noise perception and time-frequency adaptive fusion as described in claim 6, characterized in that, The numerical representation adjustment includes: sending a precision control signal to the residual block located downstream of the independent digital logic module, dynamically switching the operation precision mode, and using 16-bit floating-point numbers or fixed-point numbers for feature calculation; The feature smoothing process includes: calling a smoothing operator to filter and reduce noise on the feature map output by the residual block upstream of the independent digital logic module, so as to eliminate the numerical deviation introduced by the loss of precision of the mantissa in floating-point operations; wherein, the smoothing operator uses a moving average filter on the time domain feature map and a 3×3 average pooling filter on the frequency domain feature map.

8. A radiation source individual identification system based on noise perception and time-frequency adaptive fusion, characterized in that, include: The network construction module is used to construct a temporal feature extraction network and a frequency domain feature extraction network. The temporal feature extraction network includes an improved one-dimensional ResNet-18 network and a temporal attention module; the frequency domain feature extraction network includes a lightweight ResNeSt-50 network and a frequency attention module. The first feature enhancement module is used to extract features from the acquired original electromagnetic signal through the improved one-dimensional ResNet-18 network, aggregate temporal information and suppress redundant noise through the time attention module, model feature correlation, and output the first enhanced feature. The second feature enhancement module is used to map the original electromagnetic signal into a two-dimensional time-frequency feature map through the short-time Fourier transform (STFT) based on the constructed frequency domain feature extraction network, extract features through a lightweight ResNeSt-50 network, aggregate frequency dimension information through the frequency attention module, model the feature evolution process, and output the second enhanced feature. The weight allocation module is used to input the first enhanced feature and the second enhanced feature into the noise-aware gating unit, perform global average pooling and feature concatenation, learn and output normalized time-domain weights and frequency-domain weights through a linear layer and a Softmax activation function, so as to determine the weight allocation under different noise conditions. The weighted classification module is used to perform adaptive feature weighting based on the time domain weights and the frequency domain weights through the channel attention module, and output an enhanced feature sequence under noise perception, so as to output the individual identification result of the radiation source through classification processing.

9. An electronic device, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the noise-aware time-frequency adaptive fusion-based radiation source individual identification method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program or instructions that, when executed by a processor, implement the noise-aware time-frequency adaptive fusion radiation source individual identification method as described in any one of claims 1 to 7.

Images