Radiation source individual identification method based on signal feature analysis

Abstract

The invention relates to a radio frequency identification technology, discloses a radiation source individual identification method based on signal feature analysis, and provides a simplified CNN (Convolutional Neural Network) architecture for signal time domain multi-scale feature extraction in order to solve the problem that the existing CNN-based radiation source identification framework is difficult to effectively extract signal low-frequency features. According to the architecture, a SigNx module is designed, and hierarchical aggregation of signal multi-scale features is realized through parallel deployment of cavity convolution with an exponential incremental expansion rate; meanwhile, a mixed kernel configuration scheme is adopted to collaboratively model a signal long / short range time domain dependency mode, a large-expansion receptive field kernel captures global features, a small-expansion receptive field kernel extracts local details, and therefore the isomorphic multi-branch network with efficient parameters is constructed. Actual measurement data verification shows that the framework significantly improves the feature representation capability of radiation source individual identification.

Description

Technical Field

[0001] This invention relates to radio frequency identification (RFID) technology, and in particular to a radiation source individual identification technology based on signal feature analysis. Background Technology

[0002] With the rapid development of IoT and wireless communication technologies, device security authentication faces severe challenges. Traditional IP / MAC address-based authentication methods are vulnerable to spoofing attacks, while Radio Frequency Fingerprints (RFF) identification technology provides a more reliable hardware-level solution by analyzing signal characteristics generated by inherent hardware defects.

[0003] Current RFF identification methods are mainly divided into two categories: feature engineering and deep learning. On the one hand, traditional IP / MAC address authentication is vulnerable to spoofing attacks, and feature engineering methods rely on manually extracted features based on expert experience, which have poor feature stability. On the other hand, although deep learning methods can automatically learn features, CNNs are limited by their receptive field and have difficulty modeling long-term temporal dependencies, while Transformers face the problem of high computational complexity.

[0004] The current limitations of deep learning-based RFF recognition make it difficult for existing methods to achieve high-precision and high-efficiency radio frequency fingerprint recognition in dynamic electromagnetic environments. Summary of the Invention

[0005] The technical problem to be solved by this invention is to address the shortcomings of existing CNN models in capturing long-term dependence and low-frequency features of signals, as well as the high computational cost and insufficient high-frequency analysis capability of Transformer. In order to make radio frequency fingerprint (RFF) recognition more practical in complex electromagnetic environments and resource-constrained scenarios, this invention provides a novel architecture, SigSCNN, which integrates the advantages of CNN and Transformer, and a method for individual radiation source identification through a multi-dilation rate dilated convolution enhancement strategy.

[0006] The technical solution adopted by this invention to solve the above-mentioned technical problems is a method for individual identification of radiation sources based on signal feature analysis, comprising the following steps:

[0007] Signal preprocessing steps: The received radio frequency signal is segmented to obtain signal segments, and preliminary feature extraction is performed on each signal segment to obtain a signal vector;

[0008] Multi-scale feature extraction step: The signal vector is input into the SigSCNN model for radio frequency signal feature extraction. The SigSCNN model includes an extended receptive field stacking module, which includes several sets of convolutional kernels with large and small extended receptive fields. The large and small extended receptive field convolutional kernels respectively realize multi-scale feature representation through the SigNx module. The SigNx module adopts a parallel dilated convolution path and has several dilated convolutional layers with different dilation rates to extract multi-scale features of the signal and capture the long-term and short-term dependence features of the signal.

[0009] Classification steps: The extracted radio frequency signal features are classified using a classifier, and the individual radiation source identification results are output.

[0010] The SigNx module's processing procedure specifically involves: input signal vector The input is split into two paths: one to the shortcut connection and the other to the parallel dilated convolution module, which serves as the main path.

[0011] Quick connection to signal vector Transform into a shortcut output that matches the output dimension of the main path. ;

[0012] In the parallel dilated convolution module, each branch operates in parallel on the signal vector. First perform dilated convolution , For dilated convolution, the first The spacing between species, i.e., the dilation rate, is obtained after dilated convolution of each dilation rate. One mapping feature For each mapping feature Batch normalization (BN) operations are performed separately, followed by a convolutional feedforward (ConvFFN) network to obtain dilated convolution results at corresponding scales. The parallel output dilated convolution results at each scale are then concatenated, followed by spatial feature fusion and compression to obtain the multi-scale feature representation of the SigNx module within a specific receptive field. Then output the quick link. and Summing yields the final multi-scale features. .

[0013] This invention designs an adaptive, optimized, and simplified CNN architecture, SigSCNN, for RFF (Recognition of Free-Frequency) signals. Through its unique SigNx module, it integrates multi-scale parallel dilated convolutional paths and employs an exponentially increasing dilation rate and a hierarchical aggregation mechanism to achieve the collaborative extraction of high-frequency details and low-frequency trends. Simultaneously, it introduces a hybrid kernel configuration scheme to fuse convolutional kernel combinations with different receptive fields, effectively modeling the short-term burst features and long-term variation patterns of signals, significantly improving recognition robustness across time and location scenarios. The SigSCNN architecture, through its parameter-efficient isomorphic multi-branch design and temporal optimization, enhances global feature perception capabilities while maintaining a lightweight architecture.

[0014] The beneficial effects of this invention are that the proposed SigSCNN model employs a multi-dilation rate dilated convolution strategy. By stacking feature modules with different receptive fields layer by layer, it achieves multi-scale feature learning while maintaining computational efficiency. This effectively captures both the local fine-grained features of radio frequency signals and models long-term temporal dependencies, thereby significantly improving recognition accuracy. The SigSCNN architecture performs excellently in RFF recognition tasks and is particularly suitable for deployment in complex electromagnetic environments and resource-constrained scenarios. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the SigNx module.

[0016] Figure 2 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0017] This invention proposes a specific emitter identification (SEI) method based on signal feature analysis, which is a dedicated modification for the RFF identification task. For the SEI identification task, this invention proposes a signal-specific CNN architecture—SigSCNN—based on dilated convolution enhancement, aiming to balance efficiency and performance to achieve effective temporal signal classification.

[0018] SigSCNN achieves multi-scale feature learning by systematically expanding the effective receptive field (ERF), breaking through the receptive field limitations of traditional CNNs while maintaining computational efficiency. Furthermore, it adopts a multi-kernel-size ERF module stacked layer by layer to achieve temporal feature aggregation from fine-grained local features to long-range dependencies, effectively balancing global perception and local feature extraction.

[0019] I. A description of the terms and their use in this scheme:

[0020] (1) Dential convolution

[0021] Dilated convolution is a special convolutional operation that expands the receptive field by introducing a controllable gap adjusted by a dilation parameter. It effectively expands the feature extraction range while maintaining spatial resolution, without significantly increasing computational complexity. Unlike standard convolution, dilated convolution achieves this characteristic by inserting gaps between kernel elements. This avoids information loss caused by pooling operations and overcomes the limited receptive field of traditional convolution, providing neural networks with more efficient multi-scale feature extraction capabilities. A detailed explanation follows.

[0022] Given signal and convolution kernel Its dilated convolution operation H can be expressed as:

[0023]

[0024] in, Indicates signal Length, Represents the real number field. Indicates the size of the convolution kernel. Let be the k-th element output by the dilated convolution operation, and d be the dilation rate. Indicates the input signal The elements that participate in the convolution operation, among which This is the index of the element. The dilation operation introduces a stride of d between adjacent convolutional kernel taps. When At that time, it degenerates into a regular convolution. Given a hierarchical dilation rate d, configure a set. For the initial expansion rate Baseline convolution kernel configuration (This mainly refers to 1-D time-series signal convolution), the corresponding convolution kernel sizes are respectively The signal receptive field is significantly increased.

[0025] Dilated convolutions enable neurons to capture exponentially expanding contextual information while maintaining the original resolution of feature maps by dynamically adjusting the dilation rate. This characteristic makes them superior to traditional downsampling methods in several ways: First, by avoiding spatial information loss, they effectively preserve high-frequency details and small-scale patterns, making them particularly suitable for dense prediction tasks such as semantic segmentation and object detection. Second, their ability to capture long-range dependencies between distant elements in the input data gives them a unique advantage in sequence data processing. Furthermore, the introduction of dilated convolutions allows for parameter-efficient design, enabling flexible integration into hierarchical network architectures and achieving efficient multi-scale feature fusion.

[0026] (2) Large extended receptive field convolution

[0027] In recent years, the field of computer vision has undergone a paradigm shift from CNNs to Visual Transformers (ViTs), the latter demonstrating superior performance in long-range dependency modeling thanks to its multi-head self-attention mechanism. However, the high computational cost of ViTs has prompted researchers to re-examine the potential of CNNs—their limitations do not stem from the architecture itself, but rather from the traditional over-reliance on small convolutional kernels (such as 3×3). Inspired by the Transformer, this invention proposes a hierarchical hybrid scaling strategy. Through an innovative combination of structured sparse convolutional kernels and adaptive dilation modes, it achieves multi-scale receptive field expansion while maintaining the local inductive bias of CNNs. This method avoids the parameter explosion and low-frequency feature dominance problems caused by simply enlarging the kernel size, while simulating the global modeling capabilities of the Transformer while maintaining computational efficiency, and retaining the inherent translational equivariance advantages of CNNs.

[0028] Effective deployment of large convolutional kernels in modern architectures requires adherence to systematic design principles: First, identity shortcut connections enable dynamic fusion of multi-scale features, balancing local details with global context; second, kernel reparameterization techniques are employed during training to maintain strict translational equivariance while obtaining equivalent large kernel representations through structural reparameterization; finally, large convolutional kernels are applied in deeper layers of hierarchical networks, utilizing boundary padding to introduce controllable positional encoding. These designs demonstrate that in tasks requiring long-range dependency modeling, such as signal processing and temporal analysis, the optimized large-kernel CNN scheme significantly outperforms traditional kernel-small architectures, providing a new direction for visual coding.

[0029] II. Technical Details of the Specific Implementation of the Invention

[0030] This invention innovatively designs a signal-adaptive SigSCNN architecture for the task of identifying individual radiation sources, and achieves multi-scale feature extraction and temporal dependency modeling through the following key technologies:

[0031] First, we propose that the SigNx module adopts a parallel convolution structure to integrate dilated convolutions with exponentially increasing dilation rates, thereby achieving hierarchical aggregation of multi-level signal features.

[0032] Secondly, a hybrid core configuration scheme that integrates large and small effective receptive fields is designed to simultaneously capture short-term local features and long-term global dependency patterns of signals.

[0033] This architecture, through a carefully designed combination of modules, enables efficient characterization of the multi-scale characteristics and complex timing patterns of radio frequency signals. The specific implementation scheme and technical details will be elaborated in the following text.

[0034] (1) Multi-scale characteristic characterization of radiation source signals

[0035] To better characterize the multi-scale features of signals, a SigNx module is designed. This module adopts an Inception-like architecture, integrating parallel dilated convolutional modules with heterogeneous dilation rates. Each dilated convolutional layer is followed by a Convolutional Feedforward Network (ConvFFN), which, by mimicking the structural paradigm of the Transformer module, utilizes nonlinear projection to optimize scale-specific representations. The ConvFFN is implemented by two pointwise convolutional layers, whose channel dimensions are proportional to the coefficients. Scaling: The first pointwise convolution reduces the number of channels from the original number of channels. Expand to ( After GELU activation, the second layer performs pointwise convolution to compress it back to the original number of channels. .

[0036] This design achieves nonlinear transformations while maintaining consistency in input and output dimensions. Within each SigNX module, convolutions with different dilation rates are executed in parallel and form a residual structure with shortcut connections. Finally, multi-scale feature interaction and fusion are achieved through channel convolutions.

[0037] Figure 2 The structure of the SigNx module is shown:

[0038] Input characteristics of signal data The input is split into two paths: one to a shortcut connection, and the other to a parallel dilated convolution module; let the given size of the input feature f be... B represents the batch size. Where L is the number of input channels and L is the feature / sample length, and the number of output channels is... .

[0039] The quick connection contains two convolutional layers; the convolutional layer is performed first. Then, batch normalization (BN) is performed, followed by convolutional layers. Get quick connection output ;in, The convolution scale is determined by the receptive field of the current SigNx module. This indicates that there are no gaps between the elements of the convolution kernel, which is a normal convolution.

[0040] In the parallel dilated convolution module, each branch operates in parallel on the input features. First perform dilated convolution , For dilated convolution, the nth spacing interval is used. List of expansion rates Input features After dilation at various dilation rates, n mapped features are obtained. , Each mapping feature Batch normalization (BN) is performed separately, followed by a convolutional feedforward (Conv FFN) network to obtain dilated convolution results at the corresponding scales. The parallel output dilated convolution results at each scale are then concatenated along the channel dimension using a concat operation before entering the channel convolution network (Channels Conv). After performing a tensor transpose operation, it enters the feature convolution. The multi-scale feature representation of the SigNx module under a specific receptive field was then obtained. Then output the quick link. and Summing yields the final multi-scale features. .

[0041] By employing a modular "split-transform-fusion" paradigm, the SigNx module constructs a resolution-invariant multi-scale feature learning foundation, effectively bridging the gap between local detail preservation and global pattern integration. Through strategic aggregation of these heterogeneous representations, the model then synergizes the geometric accuracy of a small receptive field with the contextual awareness of a large receptive field, optimizing performance in tasks requiring joint localization and parsing. This enhances the network's ability to decouple overlapping spectral features and diverse features across scales. Unlike pooling-based methods (which irreversibly discard high-frequency information), dilated convolutions use spacing. This configuration operates on time-series sequences. It achieves an exponential expansion of the receptive field while maintaining temporal resolution, enabling hierarchical feature maps to maintain full resolution while progressively integrating long-range contextual information.

[0042] (2) Signal length-time dependent feature learning

[0043] Learning the long- and short-term dependencies of a signal is achieved by stacking different receptive fields. Specifically, since the Transformer's self-attention architecture exhibits a spectral bias towards low-frequency modes, stemming from its inherent long-range context-aware mechanism—achieving a panoramic receptive field through global dependency modeling—this contrasts with the local processing of CNNs. Inspired by this, we propose expanding the receptive field of CNNs by increasing the convolutional kernel size, enabling them to capture long-range dependencies like the Transformer, while deliberately preserving the inherent advantages of CNNs in local feature learning.

[0044] By employing paired convolutions with large and small extended receptive fields, the model achieves simultaneous learning of high- and low-frequency features of the signal. Feature extraction for each receptive field is implemented through the SigNx module. Each pair of paired convolutions with large and small extended receptive fields corresponds to a set of parallel convolutions with large and small convolution kernels, and feature interaction is achieved through pointwise convolutions.

[0045] Specifically set the large and small convolutional kernel groups , as the network deepens, the size of the large convolutional kernel gradually decreases, and the smallest convolutional kernel group is , and its value is related to the signal length. The small convolutional kernel always uses size. To implement a large and small convolutional kernel group , 2 SigNx modules are required. The convolutional scale in one SigNx module is , and the convolutional scale in the other SigNx module is . A total of K groups of large and small convolutional kernel groups are set.

[0046] Specifically, for the input feature and the number of output channels , each convolutional kernel group will generate a corresponding pair of output features , , where represents the output feature of the large extended receptive field, represents the output feature of the small extended receptive field. The final mapping feature is generated through pointwise convolutional operations .

[0047] (3) Signal periodic pattern feature analysis

[0048] The received signal can be regarded as time-series data with a fixed frame length. Let L be the fixed frame length, then the input sample size for emitter recognition is / / Here the specific content of the formula should be filled in according to the actual situation. Different from traditional time-series analysis (which needs to consider both intra-period and inter-period changes), in the problem of radio frequency fingerprint recognition, inter-period features are of primary importance. Therefore, the minimum period length of the signal frame is selected as the basic unit of the model training sample, which contains unique fingerprint information, and the signal block size is set to the minimum period value.

[0049] Due to the non-stationary characteristics of the received signal, the Fourier transform cannot be directly used for period detection (it cannot fully reflect the frequency time-varying characteristics). The weighted autocorrelation method is used for signal period detection. This method is especially suitable for pitch detection and performs excellently when the signal is affected by noise or shows non-stationary characteristics. By weighting the autocorrelation function, the signal period can be more accurately identified.

[0050] Given the signal x, the mathematical expression of the weighted autocorrelation function / / Here the specific content of the formula should be filled in according to the actual situation is:

[0051] [[ID=5p3]]

[0052] where k is the delay parameter, satisfying / / Here the specific content of the formula should be filled in according to the actual situation , N is the length of the signal x, / / Here the specific content of the formula should be filled in according to the actual situation is the delay-related weight coefficient at time t and satisfies The signal periodic block processing flow for weighted autocorrelation is as follows:

[0053] First, calculate the autocorrelation function between signal periods. This is achieved by calculating the mean of the product of the signal and its time-delayed signal, with the corresponding time delay weights determined by... Sure;

[0054] Secondly, the designed weight function Applied to the autocorrelation function, it calculates the weighted autocorrelation function value;

[0055] Then, the peak points of the weighted autocorrelation function are detected, and these peaks usually correspond to the period of the signal;

[0056] Finally, the signal period is determined based on the peak position—the delay corresponding to the first significant peak (excluding the zero-delay peak) is the signal period.

[0057] III. SigSCNN Architecture Design

[0058] like Figure 1 As shown, the deep learning architecture SigSCNN for signal recognition mainly consists of three parts: the base layer, the core stage, and the classifier.

[0059] (1) Base layer:

[0060] The periodic signal patch is processed according to the signal periodic pattern feature analysis described above. Patch embedding is used to transform the signal patch into a feature form suitable for subsequent processing. The right side shows the process of determining the signal period through autocorrelation for period, providing periodic information for patch embedding.

[0061] The base layer, serving as the initial layer of the network, is responsible for signal preprocessing, converting raw signal slices into period-length samples for feature mapping. The minimum period length of the signal frame is selected as the basic unit for model training samples through signal periodic pattern feature analysis. This layer uses a one-dimensional convolution with a kernel size of 5, 64 output channels, and a stride of 2 to achieve segment embedding (embedding dimension = 512) for learning low-level pattern features. To avoid loss of original signal information, downsampling is not performed in this process.

[0062] (2) Core Stage:

[0063] This is implemented using an extended receptive field stacking ERFs module, which includes K convolutional kernels with a large extended receptive field (Large ERF) and K convolutional kernels with a small extended receptive field (Small ERF). Each extended receptive field is implemented based on the aforementioned SigNx module. Multi-scale information is captured through features in different dilations, as shown on the right, with different convolutional feature maps at dilation rates d=1, 2, 4, and 8. These feature maps are then processed by channel convolution (ChannelsConv) before entering the classifier part.

[0064] The model employs a combination of large and small convolutional kernels to perceive the short- and long-term dependence features of signals, stacking features from different receptive fields to achieve interactive fusion of high- and low-frequency signal features. By using paired convolutions with large and small extended receptive fields, the model achieves simultaneous learning of high- and low-frequency signal features. Specifically, the signal feature learning utilizes a SigNx module with parallel dilated convolutions of heterogeneous dilation rates, designed based on multi-dilation rate dilated convolutions, to learn multi-scale feature representations of the signal.

[0065] The core stage is used for feature mapping. Each stage contains paired convolutional modules with K convolutional kernels of different sizes running in parallel. Each module consists of multiple parallel dilated SigNX blocks connected via shortcuts. The large extended receptive field and multi-scale information aggregation capabilities of SigSCNN are provided by the SigNX blocks. To enhance nonlinear characteristics and cross-channel information interaction, the feature map is reshaped at the end of the SigNX block through channel convolution. It should be noted that the SigNX block contains a series of topologically similar convolutional operations; therefore, reparameterization techniques can be considered during the inference stage.

[0066] (3) Classifier:

[0067] The signal processing flow from feature mapping to target classification is completed through global pooling, full layers for feature flattening, and a softmax full connected layer, and finally the recognition result (Identity Results) is output.

[0068] SigSCNN combines block embedding and periodic analysis to capture signal features through multi-scale dilated convolution, achieving effective signal classification and recognition.

Claims

1. A method for identifying a radiation source individual based on signal feature analysis, characterized in that, The method comprises the following steps: A signal preprocessing step: segmenting the received radio frequency signal to obtain signal segments, and performing preliminary feature extraction on each signal segment to obtain a signal vector; A multi-scale feature extraction step: inputting the signal vector into a SigSCNN model to extract radio frequency signal features, wherein the SigSCNN model comprises an extended receptive field stacking module, and the extended receptive field stacking module comprises a plurality of groups of large and small extended receptive field convolution kernels, and the large and small extended receptive field convolution kernels respectively realize multi-scale feature representation through SigNx modules; the SigNx module adopts a parallel hollow convolution path and has a plurality of hollow convolution layers with different expansion rates, which are used to extract multi-scale features of the signal and capture long and short time dependence features of the signal; A classification step: classifying the extracted radio frequency signal features through a classifier to output a radiation source individual identification result; Wherein, the processing procedure of the SigNx module is specifically as follows: the input signal vector is divided into two paths, one of which is input to the shortcut connection, and the other is input to the parallel hollow convolution module as the main path. Fast connection pair signal vectors Transform to fast connection output matching the main path output dimension ; In the parallel dilated convolution module, each branch operates in parallel on the signal vector. First perform dilated convolution , For dilated convolution, the first The spacing between species, i.e., the dilation rate, is obtained after dilated convolution of each dilation rate. One mapping feature For each mapping feature Batch normalization (BN) operations are performed separately, followed by a convolutional feedforward (ConvFFN) network to obtain dilated convolution results at corresponding scales. The parallel output dilated convolution results at each scale are then concatenated, followed by spatial feature fusion and compression to obtain the multi-scale feature representation of the SigNx module within a specific receptive field. Then output the quick link. and Summing yields the final multi-scale features. .

2. The method of claim 1 wherein, The shortcut connection includes two convolution layers, and the signal vector The convolution layer is performed first The batch normalization operation is performed afterwards, and then the convolution layer is performed The shortcut connection output is obtained ; wherein, is a convolution scale determined by the receptive field of the current SigNx module, represents that there is no interval between the convolution kernel elements, which is a normal convolution.

3. The method of claim 2 wherein, The spatial dimension feature fusion and compression processing is to input the size hole convolution result after splicing into channel convolution Channels Conv , and then perform a tensor dimension transpose operation to enter feature convolution Features Conv , and obtain the multi-scale feature representation of the SigNx module under a specific receptive field .

4. The method of claim 1, wherein, In the signal preprocessing step, the segmentation of the radio frequency signal comprises: determining a signal period based on signal period pattern feature analysis, and dividing the signal segments with the signal period as the length; The signal period pattern feature analysis adopts a weighted autocorrelation method, which comprises: calculating the autocorrelation function of the signal; applying a time delay weight coefficient to the autocorrelation function to obtain a weighted autocorrelation function; detecting the peak value of the weighted autocorrelation function to determine the signal period; The time delay weight coefficient is wherein x is a signal, N is a signal length, k is a time delay parameter, is a time delay related weight coefficient at time t and satisfies ; Weighted autocorrelation function under the delay parameter k is: 。 5. The method of claim 1, wherein, an expansion rate of the dilated convolution in the parallel dilated convolution path follows an exponential increase, , is an expansion rate list including an expansion rate .​ 6. The method of claim 1, wherein, The convolutional feedforward network comprises two pointwise convolution layers, wherein the first pointwise convolution layer expands the channel number to γ times the original channel number, γ≥1, and after GELU activation function, the second pointwise convolution layer compresses the channel number back to the original channel number.

7. The method of claim 1, wherein, In the group of large and small extended receptive field convolution kernels, the size of the large convolution kernel gradually decreases with the deepening of the network, and the size of the small convolution kernel remains 3×3.

8. The method of claim 7, wherein, The size of the large convolution kernel gradually decreases from 47×1 to 17×1.

9. The method of claim 1, wherein, The preliminary feature extraction realizes the embedding of the signal segments to obtain the signal vector through one-dimensional convolution.

10. The method of claim 1, wherein, The classifier comprises a global pooling layer, a fully connected layer and a softmax layer.

Images