A radiation source identification method and system for dynamic signal bandwidth

Abstract

This invention discloses a method for identifying radiation sources with dynamic signal bandwidth, comprising the following steps: S1. Obtaining Wi-Fi hotspot signals emitted by mobile phones, determining the digital spectrum persistence graph corresponding to the Wi-Fi hotspot signals, and constructing a target bandwidth diversity dataset; S2. Constructing a radiation source identification model using the YOLOv5 model as the backbone network and embedding a multi-scale convolution module; S3. Optimizing the loss function; and S4. Using the optimized radiation source identification model to identify radiation source signals with dynamic bandwidth. This method can effectively detect and identify signals under various bandwidth conditions in complex electromagnetic environments, while maintaining recognition rate and improving detection speed.

Description

Technical Field

[0001] The present invention relates to the technical field of radiation source signal identification, and in particular to a radiation source identification method and system for dynamic signal bandwidth. Background Art

[0002] Specific Emitter Identification (SEI), also known as radiation source "fingerprint" identification, is a technology that identifies individual radio transmitters by analyzing the external characteristics of intercepted electromagnetic signals. This technology is of great significance in military and civilian fields such as electronic countermeasure systems, radio communications, and battlefield friend-or-foe identification.

[0003] Traditional identification methods extract emitter fingerprint features using methods such as Hilbert spectrum, bispectral analysis, wavelet, power spectrum, and higher-order spectral transforms. These features are then used for classification and identification using methods such as Bayesian classification, KNN (K-Nearest Neighbor) nearest neighbor, and support vector machines. While traditional identification methods can update and adjust their knowledge bases based on specific problems, offering considerable flexibility, the process is cumbersome, time-consuming, and labor-intensive, making them inadequate for classification challenges arising from the increasing variety of emitter signals. To address this, researchers domestically and internationally have developed SEI methods based on deep learning (DL). These networks can autonomously learn abstract features from signals and make predictions, demonstrating strong adaptability. These methods include convolutional neural networks (CNNs) and the YOLO (You Only Look Once) family of algorithms. Shengliang Peng et al. converted complex signals into grayscale and three-channel image formats and used CNN models and frameworks for classification and identification. Chen Haiyong et al. utilized the YOLOv5 model to improve the poor performance of fine-grained SEI tasks for different individuals of the same model.

[0004] As the number of wireless devices and the variety of emitter devices continue to increase in future networks, receivers will need to dynamically adjust bandwidth to receive the full signal. Bandwidth changes pose significant challenges to SEI tasks. Existing SEI techniques have limited generalization capabilities, making it difficult to accurately identify individual emitters with varying bandwidths. Therefore, it is crucial to design a emitter identification method and system with strong generalization capabilities that can accurately identify signals even when bandwidth changes. Summary of the Invention

[0005] The present invention addresses the problem that the existing SEI technology has weak generalization ability and is difficult to accurately identify individual radiation sources of different bandwidths. The present invention provides a radiation source identification method for dynamic signal bandwidth. It can effectively detect and identify signals under various bandwidth conditions in complex electromagnetic environments. While ensuring the recognition rate, the detection speed is also improved.

[0006] In order to solve the above technical problems, the technical solution of the present invention is:

[0007] A radiation source identification method for dynamic signal bandwidth, characterized by comprising the following steps:

[0008] S1. Obtain the Wi-Fi hotspot signal emitted by the mobile phone, determine the digital spectrum persistence graph corresponding to the Wi-Fi hotspot signal, and construct a target bandwidth diversity data set;

[0009] S2. Build a radiation source recognition model, which uses the YOLOv5 model as the backbone network and embeds a multi-scale convolution module;

[0010] S3, optimize the loss function;

[0011] S4. Use the optimized radiation source identification model to identify the radiation source signal with dynamic bandwidth.

[0012] Preferably, the step S1 of constructing the target bandwidth diversity data set specifically includes the following steps:

[0013] S11. Build a Wi-Fi hotspot signal acquisition system in an interference-free environment;

[0014] S12, taking a preset type of mobile phone model as the research object, collecting hotspot signals emitted in the 2.4 GHz frequency band;

[0015] S13. Divide the target bandwidth diversity data set according to different bandwidths, and determine a first bandwidth data set and a second bandwidth data set; the first bandwidth data set is a 40 MHz bandwidth data set, and the second bandwidth data set is a bandwidth data set greater than 40 MHz;

[0016] S14. Perform data enhancement processing on the labeled first bandwidth dataset images; the data enhancement processing includes compression and cropping and splicing at random positions.

[0017] Preferably, the original YOLOv5 model framework described in step S2 is modified to embed a multi-scale convolution module. Specifically, a multi-scale convolution (MSC) module is embedded in the backbone network. The MSC module uses multiple parallel convolution operations to cause the scale of the Wi-Fi hotspot signal to change in the horizontal direction, thereby integrating the spatial position coordinate information into the extracted feature map.

[0018] Preferably, the specific process of step S21 is as follows:

[0019] Preferably, the multi-scale convolution module input feature map first passes through 1*1 convolution and ReLU activation function, and then passes through 1*2, 1*3 and 1*5 convolution operations respectively, where the 1*2 convolution block has a horizontal step size of 2 and a vertical step size of 1, the 1*3 convolution block has a horizontal step size of 3 and a vertical step size of 1, and the 1*5 convolution block has a horizontal step size of 5 and a vertical step size of 1. The two 1*2 convolution parts use a residual block structure.

[0020] Preferably, in step S3, the CIOU loss function of the original YOLOv5 model is replaced by the EIOUs loss function, and an angle penalty term is introduced into the EIOUs loss function.

[0021] As a preferred option, the EIOUs loss function consists of three parts: IOU loss, width and height loss, and distance loss.

[0022]

[0023] As a preference, define b(b x ,b y )and Represents the center point coordinates of the predicted box and the real box respectively, o h represents the vertical distance between the center of the predicted box and the center of the real box, ρ represents the Euclidean distance between the two, and C w and C h are the width and height of the minimum bounding rectangle of the predicted box and the real box respectively. The distance loss is calculated as follows:

[0024]

[0025] in,

[0026]

[0027]

[0028] In the above formula, α is the deviation angle of the predicted box center relative to the ground-truth box center. During training, when α is less than π / 4, α is minimized; when α is π / 4 or 0, the angle loss is 0.

[0029] Preferably, step S4 includes the following steps:

[0030] S41, performance evaluation of the modified model;

[0031] S42, after the model training is completed, the trained optimal weight is used to identify the radiation source signal with dynamic bandwidth, and the recognition accuracy is obtained and compared.

[0032] The present invention also provides a radiation source identification system for dynamic signal bandwidth, comprising:

[0033] The data acquisition module is used to obtain Wi-Fi hotspot signals emitted by mobile phones, determine the digital spectrum persistence diagram corresponding to the Wi-Fi hotspot signals, and construct the first bandwidth-diverse dataset in China;

[0034] Model modification module, used to modify the original YOLOv5 model framework and embed a multi-scale convolution module;

[0035] Model optimization module, used to optimize the loss function;

[0036] The radiation source signal recognition module is used to identify radiation source signals with dynamic bandwidth using the improved YOLOv5 model.

[0037] The present invention has the following characteristics and beneficial effects:

[0038] By adopting the above technical solution, the unique fingerprint features displayed by the digital spectrum afterglow image are used to analyze the subtle differences between signals when the bandwidth changes. By introducing a multi-scale convolution module, the improved YOLOv5 network can accurately identify the subtle differences between the same signals with different bandwidths. By improving the loss function to increase the convergence speed of the model, the robustness and generalization ability of the network model are enhanced, focusing on solving the problem of insufficient accuracy in recognizing signals with dynamically changing bandwidth in SEI. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0040] Figure 1 This is a flow chart of the radiation source identification method for dynamic signal bandwidth in the present invention.

[0041] Figure 2This is a framework diagram of the improved YOLOv5 model in the present invention.

[0042] Figure 3 Schematic diagram of multi-scale convolution in the present invention.

[0043] Figure 4 Schematic diagram of the signal generation and acquisition system in the present invention.

[0044] Figure 5 Schematic diagram of signal waveforms under different bandwidths collected in the present invention.

[0045] Figure 6 The embodiment of the present invention provides a comparison chart of detection effects of 60MHz and 160MHz for the iPhone 11 category as the detection object.

[0046] Figure 7 This is a detection effect diagram of the improved YOLOv5 provided in an embodiment of the present invention for various categories of 100MHz. DETAILED DESCRIPTION

[0047] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0048] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings.

[0049] On the contrary, the present invention covers any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention as defined by the claims. Furthermore, to facilitate a better understanding of the present invention, certain specific details are described in detail below in the detailed description of the present invention. Those skilled in the art will be able to fully understand the present invention without these details.

[0050] Example 1

[0051] This embodiment provides a radiation source identification method for dynamic signal bandwidth, such as Figure 1 As shown, the following steps are included:

[0052] S1, by acquiring the Wi-Fi hotspot signal emitted by the mobile phone, determining the digital spectrum persistence graph corresponding to the Wi-Fi hotspot signal and constructing a target bandwidth diversity dataset;

[0053] S2, building a radiation source recognition model, which uses the YOLOv5 model as the backbone network and embeds a multi-scale convolution module;

[0054] S3, optimize the loss function;

[0055] S4, using the optimized radiation source identification model to identify the radiation source signal with dynamic bandwidth.

[0056] Specifically, for step S1:

[0057] The first step is to build a data set and build a WIFI hotspot signal acquisition system in an interference-free environment. Figure 2 As shown in the figure, the system consists of an OmniLoG70600 receiving antenna, a receiver with digital persistence spectrum analysis, a computer with parameter analysis capabilities, and a waveform display. Six mobile phones each emit a Wi-Fi hotspot signal. The receiving antenna then transmits the captured signal to the receiver. The computer then adjusts parameters such as bandwidth and center frequency. The signal waveform, displayed as a digital spectrum persistence graph, is observed and recorded on the display.

[0058] Six models of mobile phones were used as the collection objects, and hotspot signals emitted in the frequency band of 2.4GHz (Wi-Fi signals usually occur at 2.4GHz and 5GHz) were collected. Signals with bandwidths of 40MHz, 60MHz, 80MHz, 100MHz, 120MHz, 140MHz and 160MHz were collected to construct the first domestic bandwidth diversified data set.

[0059] The data set is constructed with 40MHz bandwidth signals as training samples and validation samples, and 60MHz, 80MHz, 100MHz, 120MHz, 140MHz and 160MHz bandwidth data as test samples, totaling 2220 images, whose distribution is shown in Table 1. The digital spectrum persistence diagrams of hotspot signals of six types of mobile phone models at different bandwidths are shown in Table 1. Figure 3 shown.

[0060] Table 1 Dataset distribution

[0061]

[0062] Next, data annotation was performed using the labelimg annotation software, following the Pascal VOC 2007 standard format, to classify the collected 40MHz data images and select annotation boxes. This generated an Extensible Markup Language (XML) file. Finally, data augmentation operations were performed, including random position compression, cropping, and splicing. The image labels corresponding to the augmented 40MHz data were divided into training and validation sets in an 8:2 ratio to ensure that each set contained representative samples.

[0063] After processing the data image, the network model yaml file was modified. First, the label categories in the mydata.yaml file in the date folder were changed to signal 1, signal 2, signal 3, signal 4, signal 5, and signal 6. Since this invention uses six labels, the number of categories is set to 6. Next, the training and validation paths in the model yaml file were replaced with the paths in the current dataset. The weight file used for subsequent model training was YOLOv5s.pt downloaded from the official website. The environment and parameter settings for this invention are shown in Table 2.

[0064] Table 2 Environment and parameter settings

[0065]

[0066] The present invention is further configured to:

[0067] The radiation source recognition model is obtained by modifying the YOLOv5 framework as the backbone network. Figure 4 As shown in the figure, it consists of three parts: the backbone network (Backbone), the neck network (Neck), and the prediction head (Head). The digital spectrum persistence map of the signal is used as input and first passes through the Backbone, where the MSC module is used to perform scale changes to capture a wider range of contextual information. The Conv module contains ordinary convolution, Bn layer, and SiLU activation function. The SiLU activation function makes the gradient calculation during the backpropagation process more stable. The C3 module realizes feature fusion through the residual structure. The SPPF module uses multiple small-sized pooling kernels in cascade to replace the single large-sized pooling kernel in the SPP module, thereby further improving the running speed while retaining the original function, that is, fusing feature maps with different receptive fields and enriching the expressive power of feature maps. Then, it enters the Neck layer and uses the Path Aggregation Network (PANet) to fuse and extract multi-scale feature representations to improve the accuracy and robustness of target detection. Finally, the signal is predicted and classified through the Head layer.

[0068] Schematic diagram of the embedded multi-scale convolution module, as shown in Figure 5As shown, the input feature map undergoes 1*2, 1*3, and 1*5 convolution operations, respectively. The 1*2 convolution block has a horizontal stride of 2 and a vertical stride of 1, the 1*3 convolution block has a horizontal stride of 3 and a vertical stride of 1, and the 1*5 convolution block has a horizontal stride of 5 and a vertical stride of 1. The two 1*2 convolutions use a residual block structure, which facilitates gradient propagation within the network, effectively alleviating the vanishing gradient problem and improving network convergence speed. The 1*1 convolution block reduces the dimensionality of the feature map, thereby reducing computational effort and memory consumption. Through these multi-scale convolution operations, the network can fuse information from deeper and shallower feature maps. Deeper features are beneficial for capturing global information, while shallower features provide more local details. Furthermore, by simultaneously leveraging global and local information at different levels, it is possible to detect and recognize objects of varying sizes.

[0069] Furthermore, for step S3:

[0070] This paper optimizes the EIOU loss function and introduces an angle penalty term based on the EIOU loss function, so that the prediction box can be quickly moved to the nearest axis, replacing the CIOU loss function in the original YOLOv5. The EIOUs loss function consists of three parts: IOU loss, width and height loss, and distance loss. The IOU loss formula is:

[0071]

[0072] Among them, AIB represents the area where the predicted box and the real box intersect, and AUB represents the area where the predicted box and the real box are combined.

[0073] The width and height loss calculates the length and width of the target box and the predicted box respectively. The calculation formula is:

[0074]

[0075] Among them, w and h represent the width and height of the prediction box respectively, w gt and h gt are the width and height of the real box respectively, C w and C h are the width and height of the minimum bounding rectangle of the predicted box and the true box, respectively.

[0076] In order to make the loss more sensitive to the size and position of the object, an angle penalty term is introduced on the distance loss, so that the prediction box can be quickly moved to the nearest axis, thereby improving the convergence speed of the model. x ,b y )and Represents the center point coordinates of the predicted box and the real box respectively, o hRepresents the vertical distance between the center of the predicted box and the center of the real box, ρ represents the Euclidean distance between the two, and the distance loss is calculated as follows:

[0077]

[0078] in,

[0079]

[0080] In the above formula, α is the deviation angle of the predicted box center relative to the ground-truth box center. During training, when α is less than π / 4, α is minimized; when α is π / 4 or 0, the angle loss is 0.

[0081] Total regression loss is the sum of the above three losses, and the formula is:

[0082]

[0083] For step S4:

[0084] Based on the trained and modified YOLOv5 model, the model was evaluated using metrics such as precision (P), recall (R), average detection accuracy (mAP), detection speed (FPS), and model parameter count (Params). The overall performance of the YOLOv5 network and the improved YOLOv5 network is shown in Table 3.

[0085] Table 3 Overall performance of YOLOv5 network and radiation source recognition model

[0086]

[0087] Compared to the YOLOv5 network, the modified YOLOv5 network (radiator identification model) achieved improvements in precision, recall, mAP50, mAP50-95, and FPS. The mAP50-95 score increased by 3.1%, and the FPS rose from 233.0 to 244.4. This demonstrates that the dynamic bandwidth-specific radiator identification method provided by the present invention enables the network to effectively learn signal characteristics related to the bandwidth portion, while also significantly improving training speed.

[0088] After training, save the trained weights in best.pt. Then, modify the weight path and the test image path in detect.py. Run the program and save the detected images in the runs folder. The recognition performance of YOLOv5 and the radiant source recognition model on the test set is shown in Tables 4 and 5. The numbers in the table represent recognition accuracy, and false positives indicate classification errors.

[0089] Table 4. Recognition effect of YOLOv5 on signals of different bandwidths in the test set

[0090]

[0091]

[0092] Table 5. Recognition effect of the radiation source recognition model on signals of different bandwidths in the test set

[0093]

[0094] Table 4 shows that when the signal bandwidth changes, only signal 2 can accurately identify the signal, while the other five categories will experience false detection, resulting in poor experimental results. Table 5 shows that when the signal bandwidth changes, signals 2, 5, and 6 can effectively identify samples across all bandwidths with high recognition accuracy. Signals 3 and 4 can effectively identify the vast majority of bandwidth samples, but false detection occurs at 160MHz bandwidth due to the significant feature variation of the 160MHz signal compared to the 40MHz signal used for training. Signal 1, on the other hand, can identify some signals under various bandwidth conditions.

[0095] The effects achieved by the present invention will be described with reference to the accompanying drawings. Figure 6 The recognition accuracy of the YOLOv5 model and the improved YOLOv5 model for signals of the same category with different bandwidths (60MHz and 160MHz) is demonstrated. Figure 6 (a) is the recognition effect of the YOLOv5 model. Figure 6 (b) is the recognition effect of the improved YOLOv5 model. Figure 6 It can be seen that when the same signal is presented in different bandwidth forms, the signal characteristics vary greatly. By training six categories of 40MHz data, the YOLOv5 model cannot accurately identify the 60MHz and 160MHz signals of signal 5, and mistakenly detects signal 5 as signal 2. However, the radiation source identification system for dynamic signal bandwidth of the present invention has a good detection effect for the 60MHz and 160MHz signals of signal 5, and the recognition accuracy is generally above 0.9.

[0096] The above scheme is used to identify different types of signals with the same bandwidth and observe Figure 7The improved YOLOv5 model's performance in identifying six categories at 100MHz demonstrates that, despite similar signal characteristics across different signals within the same bandwidth, the proposed radiation source identification system for dynamic signal bandwidth can still accurately identify signal categories with an accuracy exceeding 0.9. The final results demonstrate that, when trained on 40MHz signals, the proposed method can still accurately identify signals presented at other bandwidths. The trained model exhibits strong generalization and robustness, making it suitable for SEI tasks addressing dynamic bandwidth.

[0097] In this embodiment, the unique fingerprint features displayed by the digital spectrum persistence graph are used to analyze the subtle differences between signals when the bandwidth changes. By introducing a multi-scale convolution module, the improved YOLOv5 network can accurately identify the subtle differences between the same signals with different bandwidths. By improving the loss function to increase the convergence speed of the model, the robustness and generalization ability of the network model are improved, focusing on solving the problem of insufficient recognition accuracy of signals with dynamically changing bandwidth in SEI.

[0098] Example 2

[0099] This embodiment further provides a radiation source identification system for dynamic signal bandwidth, which is used to implement the method described in Example 1, and specifically includes:

[0100] The data acquisition module is used to obtain Wi-Fi hotspot signals emitted by mobile phones, determine the digital spectrum persistence diagram corresponding to the Wi-Fi hotspot signals, and construct the first bandwidth-diverse dataset in China;

[0101] Model modification module, used to modify the original YOLOv5 model framework and embed a multi-scale convolution module;

[0102] Model optimization module, used to optimize the loss function;

[0103] The radiation source signal recognition module is used to identify radiation source signals with dynamic bandwidth using the improved YOLOv5 model.

[0104] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. It will be apparent to those skilled in the art that various changes, modifications, substitutions, and variations of these embodiments, including components, without departing from the principles and spirit of the present invention are still within the scope of protection of the present invention.

Claims

1. A radiation source identification method for dynamic signal bandwidth, characterized in that: The steps include: Step 1: Obtain the Wi-Fi hotspot signal emitted by the mobile phone, determine the digital spectrum persistence graph corresponding to the Wi-Fi hotspot signal, and then construct a target bandwidth diversity dataset; Step 2: Build a radiation source recognition model, which uses the YOLOv5 model as the backbone network and embeds a multi-scale convolution module; Step 3: Optimize the loss function and introduce the angle penalty term into the EIOU loss function to construct the EIOUs loss function and replace the CIOU loss function of the original YOLOv5 model. The EIOUs loss function consists of IOU loss, width and height loss, and distance loss: In step 3, the distance loss function after introducing the angle penalty term is expressed as follows: Where, define b(b x ,b y )and Represents the center point coordinates of the predicted box and the real box respectively, o h represents the vertical distance between the center of the predicted box and the center of the real box, ρ represents the Euclidean distance between the two, and C w and C h are the width and height of the minimum bounding rectangle of the predicted box and the real box respectively, and α is the deviation angle of the center of the predicted box relative to the center of the real box; Step 4: Use the optimized radiation source identification model to identify the radiation source signal with dynamic bandwidth.

2. The radiation source identification method for dynamic signal bandwidth according to claim 1, characterized in that: In step 1, a WI-FI hotspot signal collection system is used to collect WI-FI hotspot signals emitted in the 2.4 GHz frequency band in an interference-free environment.

3. The radiation source identification method for dynamic signal bandwidth according to claim 1, characterized in that: The target bandwidth diversity data set includes a first bandwidth data set and a second bandwidth data set, the first bandwidth data set is a 40 MHz bandwidth data set, and the second bandwidth data set is a bandwidth data set greater than 40 MHz.

4. The radiation source identification method for dynamic signal bandwidth according to claim 3, characterized in that: In step 1, data enhancement processing is performed on the labeled first bandwidth dataset images; the data enhancement processing includes compression at different ratios and cropping and splicing at random positions.

5. The radiation source identification method for dynamic signal bandwidth according to claim 3, characterized in that: The embedded multi-scale convolution module inputs the feature map through 1*1 convolution and Relu activation function, and then passes through 1*2, 1*3 and 1*5 convolution operations respectively, where the horizontal step size of the 1*2 convolution block is 2 and the vertical step size is 1; the horizontal step size of the 1*3 convolution block is 3 and the vertical step size is 1; the horizontal step size of the 1*5 convolution block is 5 and the vertical step size is 1; the two 1*2 convolution parts use the residual block structure.

6. The radiation source identification method for dynamic signal bandwidth according to claim 1, characterized in that: During the training process of the EIOUs loss function, when α is less than π / 4, α is minimized; when α is π / 4 or 0, the angle loss is 0.

7. A radiation source identification system for dynamic signal bandwidth, used to implement the radiation source identification method for dynamic signal bandwidth according to any one of claims 1 to 6, characterized in that: include: A data acquisition module is used to obtain Wi-Fi hotspot signals emitted by mobile phones, determine the digital spectrum persistence graph corresponding to the Wi-Fi hotspot signals, and construct a target bandwidth diversified data set; The model modification module is used to modify the original YOLOv5 model framework and embed a multi-scale convolution module to obtain a radiation source recognition model; Model optimization module, used to optimize the loss function; The radiation source signal identification module is used to identify the radiation source signal with dynamic bandwidth using the radiation source identification model.

Images