New untrained radiation source detection method based on multi-discriminator model

Abstract

The present invention discloses a new untrained radiation source detection method based on a multi-discriminator model. The method obtains a set of signal image samples of N identified radiation sources and divides them into a training set and a test set. A generator and N+1 discriminators are set up. The generator and each discriminator are sequentially constructed into a sub-generative adversarial network. The corresponding training set is used for training. The test set and the image sample set are input into the trained discriminator. The judgment interval and proportion parameter are statistically obtained. The signal image of the radiation source to be identified is input into each discriminator. The proportion of the output probability value within the judgment interval is statistically calculated. If the proportion is greater than the proportion parameter, it is determined that the radiation source to be identified has been identified by the discriminator. Finally, the new untrained radiation source detection is completed according to the judgment rule. The present invention uses the multi-discriminator model to judge the signal image of the radiation source, effectively realizing the detection of the new untrained radiation source.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communication device identification, and more particularly, relates to a new untrained radiation source detection method based on a multi-discriminator model. Background Art

[0002] Wireless communication device identification is an indispensable technology in modern life, with widespread application across various fields. It identifies the source device by emitting different radio frequency signals based on the hardware characteristics of each device. These differences are primarily due to discrepancies between the electrical parameters of electronic components and their nominal values ​​during production. This leads to manufacturing variations between components, which in turn creates hardware differences within different signal source systems. This leads to unchangeable errors in signal transmission and reception between devices.

[0003] In today's complex and ever-changing electronic warfare environments, radio frequency (RF) signal transmitters, such as communication radios and unmanned reconnaissance aircraft, play a crucial role. If we can train our own equipment and establish a database, and accurately determine the information of both enemy and friendly radiators using this database, we can determine their operating status and activity patterns, and predict their combat capabilities. This plays a crucial role in this era of self-evident importance of electronic warfare. Accurately classifying and identifying each radiator is a current research priority. Therefore, in-depth research into the extraction and classification of RF emitter fingerprints remains a hot topic in this field.

[0004] In recent years, with the continuous advancement of RF fingerprinting technology, it has been discovered that existing methods rely on training and testing on previously trained emitter samples to achieve classification and identification of undetected emitters. However, this approach presents a significant problem: it can only identify the emitter categories already in the training set. When new, untrained emitter samples are introduced, they are only classified into one of the previously labeled categories, resulting in incorrect identification. However, in actual electronic warfare, wireless communication equipment is complex and diverse, and enemy equipment is unlikely to provide opportunities for pre-training. Therefore, a method suitable for detecting newly added, untrained emitters is needed to meet real-world needs. Summary of the Invention

[0005] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a new untrained radiation source detection method based on a multi-discriminator model, which uses the multi-discriminator model to judge the signal image of the radiation source and effectively realizes the detection of the new untrained radiation source.

[0006] In order to achieve the above-mentioned object of the invention, the newly added untrained radiation source detection method based on the multi-discriminator model of the present invention includes the following steps:

[0007] S1: Determine N identified radiation sources according to the actual situation, collect a number of signals of duration T for each identified radiation source, and then image each signal to obtain the corresponding signal image; the signal image set φ of each identified radiation source n Divide into training set α n and the test set β n ,n=1,2,...,N,the signal image set φ of N identified radiation sources n The collection of signal images is recorded as the signal image set φ N+1 , the training set α of N identified radiation sources n The collection of is recorded as training set α N+1 , the test set β of N identified radiation sources n The collection of is recorded as test set β N+1 ;

[0008] S2: Set up a generator G and N+1 discriminators D n′ , n′=1,2,…,N+1, where the generator G is used to generate pseudo signal images, and the 1st to Nth discriminators D n They are used to identify the probability value of the input signal image belonging to the nth identified radiation source, the N+1th discriminator D N+1 The probability value used to identify that the input signal image belongs to the identified radiation source;

[0009] S3: Sequentially combine the generator G and each discriminator D n′ Constructed into a sub-generation adversarial network GAN n′ , for each sub-generation adversarial network GAN n′ , using the training set α constructed in step S1 n′ As a real image set, the pair generative adversarial network GAN n′ Train to obtain N+1 trained discriminators D n′ ;

[0010] S4: For each discriminator D trained in step S3 n′ , the test set β n′ Each signal image in the discriminator D n′ , get the output probability value set, and determine the minimum probability value γ based on the probability value set n′,min and the maximum probability value γ n′,max , thus obtaining the judgment interval [γ n′,min ,γ n′,max ]; Set the signal image set φ n′Each signal image in the discriminator D n′ , the statistical output probability value is in the judgment interval [γ n′,min ,γ n′,max ] ratio value z n′ , let the proportional parameter ω n′ =λ×z n′ ,λ is a preset constant, and its value range isλ∈[0.5,1);

[0011] S5: Collect a number of signals of duration T from the radiation source to be identified, generate a signal image of each signal using the same method as in step S1, and then input each signal image into the discriminator D n′ , the statistical output probability value is in the judgment interval [γ n′,min ,γ n′,max ] ratio ω′, if ω′>ω n′ , then it is determined that the radiation source to be identified is determined by the discriminator D n′ The identification is passed, otherwise the identification fails; then the following rules are used to determine whether the radiation source to be identified is a new radiation source:

[0012] If the discriminator D N+1 If the identification fails, further identification of the discriminators D1 to D N Whether all identifications are passed, if so, the radiation source to be identified is determined to be an identified radiation source, otherwise it is determined to be a newly added radiation source;

[0013] If the discriminator D N+1 If the identification is passed, the identification devices D1 to D N Whether all identifications fail, if so, the radiation source to be identified is determined to be a newly added radiation source, otherwise, the radiation source to be identified is determined to be an identified radiation source.

[0014] The present invention is based on a new untrained radiation source detection method of a multi-discriminator model. The method obtains a signal image sample set of N identified radiation sources and divides them into a training set and a test set respectively. A generator and N+1 discriminators are set. The generator and each discriminator are sequentially constructed into a sub-generative adversarial network, which is trained using a corresponding training set. The test set and the image sample set are input into the trained discriminator, and a judgment interval and a proportion parameter are obtained by statistics. The signal image of the radiation source to be identified is input into each discriminator, and the proportion of the output probability value within the judgment interval is calculated. If the proportion is greater than the proportion parameter, it is determined that the radiation source to be identified has passed the identification by the discriminator. Finally, the detection and judgment of the new untrained radiation source are completed according to the judgment rule.

[0015] The present invention trains a generative adversarial network based on a multi-discriminator model, uses the signal image of the identified radiation source to statistically calculate the proportion parameter of the judgment interval, and then uses the multi-discriminator model to judge the signal image of the radiation source, effectively realizing the detection of newly added untrained radiation sources. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flowchart of a specific implementation of the newly added untrained radiation source detection method based on the multi-discriminator model of the present invention;

[0017] Figure 2 is an example diagram of a signal image generated by the signal image generation method based on segmented mapping in this embodiment;

[0018] Figure 3 is a structural diagram of the discriminator in this embodiment;

[0019] Figure 4 Schematic diagram of the decision intervals of 15 discriminators in this embodiment;

[0020] Figure 5 1 is a curve diagram of the accuracy of detecting newly added untrained radiation sources according to the present invention in an embodiment;

[0021] Figure 6 This is a bar chart of the proportion of 14 intercoms that pass the discriminator when all of them are identified as radiation sources in this embodiment. DETAILED DESCRIPTION

[0022] The following describes the specific embodiments of the present invention in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when detailed descriptions of known functions and designs may dilute the main content of the present invention, such descriptions will be omitted here.

[0023] Example

[0024] Figure 1 This is a flowchart of a specific implementation of the newly added untrained radiation source detection method based on the multi-discriminator model of the present invention. Figure 1 As shown, the specific steps of the newly added untrained radiation source detection method based on the multi-discriminator model of the present invention include:

[0025] S101: Collecting signals from identified radiation sources:

[0026] According to the actual situation, determine N identified radiation sources, collect several signals of duration T for each identified radiation source, and then image each signal to obtain the corresponding signal image. n Divide into training set α n and the test set βn ,n=1,2,...,N,the signal image set φ of N identified radiation sources n The collection of signal images is recorded as the signal image set φ N+1 , the training set α of N identified radiation sources n The collection of is recorded as training set α N+1 , the test set β of N identified radiation sources n The collection of is recorded as test set β N+1 .

[0027] The specific method for signal visualization can be selected as needed. Generally, a signal image is obtained by mapping the signal's I / Q data. Additional processing may be performed to enhance the image's feature representation. This embodiment employs a signal image generation method based on segmented mapping. Specifically, the I / Q data of a signal of duration T is evenly divided into M segments, where the value of M is determined based on actual needs. Each segment of I / Q data is mapped to obtain a sub-signal image. The M sub-signal images are then superimposed, and the superimposed signal image is used as the signal image obtained by mapping the original signal.

[0028] Figure 2 : is an example of a signal image generated by the signal image generation method based on segmented mapping in this embodiment. Figure 2 As shown in FIG, the signal image generated by segmented mapping and superposition has richer features than the original image and can better distinguish different radiation sources.

[0029] S102: Constructing a generative adversarial network with multiple discriminators:

[0030] To achieve radiation source detection, the present invention draws on the idea of ​​generative adversarial networks to construct a radiation source detection model based on generative adversarial networks. In the prior art, anomaly detection based on generative adversarial networks typically focuses on the encoding and decoding process of the generator. By setting a threshold, the difference between the sample encoded and decoded by the generator is detected to determine whether it is an anomaly sample, while the role of the discriminator is minimized. The goal of the present invention is to detect whether there are new untrained radiation sources. This goal is different from the conventional anomaly detection method based on generative adversarial networks. Therefore, a generative adversarial network using multiple discriminators is proposed. The specific structure is as follows:

[0031] Set up a generator G and N+1 discriminators D n′ , n′=1,2,…,N+1, where the generator G is used to generate pseudo signal images, and the 1st to Nth discriminators D n They are used to identify the probability value of the input signal image belonging to the nth identified radiation source, the N+1th discriminator D N+1The probability value used to identify the input signal image belongs to the identified radiation source. In other words, the discriminator D N+1 Can be viewed as a hybrid discriminator.

[0032] According to the description of the generative adversarial network of the present invention, the discriminator is a key module that affects the detection performance of new radiation sources. In order to enable the discriminator to extract feature scalars faster and better, the discriminator is also improved in this embodiment by introducing an automatic selective convolution (Selective Kernel) module. Figure 3 This is the structural diagram of the discriminator in this embodiment. Figure 3 As shown, the discriminator in this embodiment includes a first convolutional layer, an automatic convolution kernel selection module, a second convolutional layer and a discrimination module, wherein:

[0033] The first convolutional layer is used to perform a 1×1 convolution operation on the input image and output the obtained feature image to the automatic convolution kernel selection module.

[0034] The automatic kernel selection module performs a convolution operation on the received feature image using an automatically selected convolution kernel, outputting the resulting feature image to the second convolutional layer. This module leverages its adaptive receptive field size, enabling the discriminator to adaptively adjust its convolution layer size based on the input data size, thereby better extracting signal image features. The automatic kernel selection convolution operation consists of three steps: split, fuse, and select. The split operation generates multiple paths with varying kernel sizes (corresponding to the different receptive field sizes of neurons). The fuse operation combines and aggregates information from multiple paths to obtain a global and comprehensive representation of selectivity weights. The select operation aggregates feature maps from kernels of different sizes based on the selection weights. The detailed principles and workings of the automatic kernel selection can be found in the paper "Li X, Wang W, Hu X, et al. Selective Kernel Networks [J]. IEEE, 2020."

[0035] The second convolutional layer is used to perform a 1×1 convolution operation on the received feature image and output the obtained feature image to the identification module.

[0036] The identification module is used to determine the authenticity of the received feature image and output a probability value that the input image is a real image. The identification module can use a conventional discriminator in a generative adversarial network.

[0037] S103: Training Generative Adversarial Networks:

[0038] Sequentially combine the generator G and each discriminator D n′ Constructed into a sub-generation adversarial network GANn′ , for each sub-generation adversarial network GAN n′ , using the training set α constructed in step S101 n′ As a real image set, the pair generative adversarial network GAN n′ Train to obtain N+1 trained discriminators D n′ .

[0039] Only one generator is used in the present invention because there is a certain similarity between the signal images of each radiation source. After the sub-generative adversarial network of the previous identified radiation source is trained, the parameters of its generator can be used as the initial parameters of the sub-generative adversarial network of the subsequent identified radiation source, so as to shorten the training time.

[0040] S104: Determine radiation source judgment parameters:

[0041] In actual scenarios, it is impossible to obtain any prior information about any newly added untrained radiation sources in advance, so when performing radiation source identification, it is necessary to rely on the identified radiation sources to determine the radiation source judgment parameters. An intuitive method to distinguish whether the radiation source is new is to compare the average value of the probability values ​​obtained by the discriminator for the signal images of the radiation source to be identified and the identified radiation source. However, this method can easily lead to an identified radiation source being identified as a new radiation source. Therefore, the present invention performs radiation source identification by using the value range of the probability value output by the discriminator. The method for determining the radiation source judgment parameters is as follows:

[0042] For each discriminator D trained in step S103 n′ , the test set β n′ Each signal image in the discriminator D n′ , get the output probability value set, and determine the minimum probability value γ based on the probability value set n′,min and the maximum probability value γ n′,max , thus obtaining the judgment interval [γ n′,min ,γ n′,max In this embodiment, the minimum probability value γ n′,min and the maximum probability value γ n′,max The method for determining γ is as follows: first, remove the abnormal maximum and minimum values ​​from the probability value set, then sort the remaining probability values ​​from large to small, and take the mean of the first K maximum values ​​as the maximum probability value γ n′,max , take the mean of the last K minimum values ​​as the minimum probability value γ n′,min , the value of K is determined according to actual needs.

[0043] However, there is another problem that cannot be ignored. The probability value of the signal image of the radiation source to be determined obtained by the discriminator may overlap with the identified radiation source. Therefore, the determination interval [γn′,min ,γ n′,max ] to determine the radiation source may result in a large error. Therefore, the present invention converts the signal image set φ n′ Each signal image in the discriminator D n′ , the statistical output probability value is in the judgment interval [γ n′,min ,γ n′,max ] ratio value z n′ , let the proportional parameter ω n′ =λ×z n′ , λ is a preset constant, and its value range is λ∈[0.5,1).

[0044] Then, the signal image of the radiation source to be identified can be input into the discriminator D n′ The output probability value is obtained, and the statistical output probability value is in the judgment interval [γ n′,min ,γ n′,max ] exceeds the proportional parameter ω n′ To determine whether the radiation source to be identified is the corresponding identified radiation source.

[0045] S105: Added radiation source detection:

[0046] The radiation source to be identified collects a number of signals of duration T, generates a signal image of each signal according to the same method as in step S101, and then inputs each signal image into the discriminator D n′ , the statistical output probability value is in the judgment interval [γ n′,min ,γ n′,max ] ratio ω′, if ω′>ω n′ , then it is determined that the radiation source to be identified is determined by the discriminator D n′ The identification is passed, otherwise the identification fails. Then the following rules are used to determine whether the radiation source to be identified is a new radiation source:

[0047] If the discriminator D N+1 If the identification fails, further identification of the discriminators D1 to D N Whether all identifications are passed, if so, the radiation source to be identified is determined to be an identified radiation source, otherwise it is determined to be a newly added radiation source.

[0048] If the discriminator D N+1 If the identification is passed, the identification devices D1 to D N Whether all identifications fail, if so, the radiation source to be identified is determined to be a newly added radiation source, otherwise, the radiation source to be identified is determined to be an identified radiation source.

[0049] This allows for more accurate detection of new radiation sources, avoiding situations where there are too many radiation sources, resulting in blurred boundaries and a probability value interval that is too large to be accurately detected.

[0050] To better illustrate the technical solution of the present invention, a specific example was used to experimentally verify the present invention. In this example, six different types of walkie-talkies were used for radiation source detection, labeled A (A1 / A2), B (B1 / B2), C (C1 / C2), D (D1 / D2), E (E1 / E2), and V (V1 / V2 / V3 / V4). A total of 14 walkie-talkies were used as radiation sources, and the walkie-talkie data acquisition sampling rate was set to 1MS / s. Signal samples from these 14 walkie-talkies were collected and visualized to obtain a set of signal images for each walkie-talkie. The images were then divided into a training set and a test set in a 9:1 ratio.

[0051] As for the generative adversarial network, this embodiment includes a generator and 15 discriminators. The generator G used in this embodiment is mainly divided into three layers, consisting of the generative encoder G E , generate decoder G D The encoder E is composed of an encoding-decoding process, which generates a fake image and then downsamples it to a low-dimensional tensor. Table 1 is the structural parameter table of the generator G in this embodiment.

[0052]

[0053]

[0054] Table 1

[0055] The discriminator D used in this embodiment is a discriminator that introduces an automatic convolution kernel selection module.

[0056] Table 2 is a table of structural parameters of the discriminator D in this embodiment.

[0057]

[0058] Table 2

[0059] As shown in Table 2, in the automatic convolution kernel selection module, there is a crucial parameter, m, which determines the number of paths, that is, the number of convolution kernels of different sizes selected for fusion. g is used to control the cardinality of each path, and r is used to control the number of parameters in the fuse operation.

[0060] The generative adversarial network is trained using a training set of 14 walkie-talkies to obtain 15 discriminators, which are then input into the test set to determine the decision interval [γ n′,min ,γ n′,max ]. Figure 4 This is a schematic diagram of the decision intervals of the 15 discriminators in this embodiment. Figure 4As shown in Figure 2, the thresholds of different types of radiation source devices will overlap, such as A2 and B1, B2 and V1, etc. Then, the signal image set of 14 walkie-talkies is input into 15 discriminators, and the statistical output probability value is within the judgment interval [γ n′,min ,γ n′,max ] ratio value, and then determine the ratio parameter. In order to facilitate the experiment, the ratio parameter of each discriminator in this embodiment is set to 3 / 4.

[0061] Among the 14 walkie-talkies, different numbers of walkie-talkies are randomly selected as identified radiation sources, and the remaining walkie-talkies are used as newly added untrained radiation sources. The present invention is used to detect the newly added untrained radiation sources, and the detection accuracy is statistically analyzed. Figure 5 : is a curve diagram of the accuracy of the present invention for detecting newly added untrained radiation sources in the embodiment. Figure 5 As shown, when the number of identified radiation sources, N, is 1, the detection accuracy for the remaining 13 categories of new radiation sources is the best, reaching over 98%. This is consistent with the minimal overlap between A1's determination and the other categories. As the number of identified radiation sources, N, increases, the detection accuracy of new radiation sources stabilizes at around 86% when N is between 5 and 9. When N is ≥ 12, the detection accuracy for new radiation sources drops below 75%.

[0062] When the number of identified radiation sources N ≥ 14, this is a special case. Under this condition, the number of newly added radiation source devices is 0, and the radiation source signal data needs to be re-collected to obtain new test data to verify the effect of the identifier. Figure 6 This is a bar chart of the proportion of 14 intercoms that pass the discriminator when all of them are identified as radiation sources. Figure 6 As shown, the proportion of walkie-talkie V1 that passes the corresponding identification is less than 75%, which is less than the proportion parameter 3 / 4. Walkie-talkie V1 will be misidentified as a new radiation source, while the remaining 13 categories will be identified as identified radiation sources, and the misidentification rate is only 1 / 14.

[0063] In summary, the present invention can achieve a detection accuracy of more than 90% when the number of identified radiation sources is less than 5, and the detection accuracy is still 80% when the number is increased to less than 12, which proves the feasibility of the present invention in the field of newly added untrained radiation source detection.

[0064] Although the above describes the illustrative specific embodiments of the present invention to facilitate understanding of the present invention by those skilled in the art, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concepts of the present invention are protected.

Claims

1. A new untrained radiation source detection method based on a multi-discriminator model, characterized in that: The following steps are involved: S1: Determine N identified radiation sources according to the actual situation, collect a number of signals of duration T for each identified radiation source, and then image each signal to obtain the corresponding signal image; the signal image set φ of each identified radiation source n Divide into training set α n and the test set β n ,n=1,2,...,N,the signal image set φ of N identified radiation sources n The collection of signal images is recorded as the signal image set φ N+1 , the training set α of N identified radiation sources n The collection of is recorded as training set α N+1 , the test set β of N identified radiation sources n The collection of is recorded as test set β N+1 ; S2: Set up a generator G and N+1 discriminators D n′ , n′=1,2,…,N+1, where the generator G is used to generate pseudo signal images, and the 1st to Nth discriminators D n They are used to identify the probability value of the input signal image belonging to the nth identified radiation source, the N+1th discriminator D N+1 The probability value used to identify that the input signal image belongs to the identified radiation source; S3: Sequentially combine the generator G and each discriminator D n′ Constructed into a sub-generation adversarial network GAN n , for each sub-generation adversarial network GAN n′ , using the training set α constructed in step S1 n′ As a real image set, the pair generative adversarial network GAN n′ Train to obtain N+1 trained discriminators D n′ ; S4: For each discriminator D trained in step S3 n′ , the test set β n′ Each signal image in the discriminator D n′ , get the output probability value set, and determine the minimum probability value γ based on the probability value set n′,min and the maximum probability value γ n′,max , thus obtaining the judgment interval [γ n′,min ,γ n′,max ]; Set the signal image set φ n′ Each signal image in the discriminator D n′ , the statistical output probability value is in the judgment interval [γ n′,min ,γ n′,max ] ratio value z n′ , let the proportional parameter ω n′ =λ×z n′ ,λ is a preset constant, and its value range isλ∈[0.5,1); S5: Collect a number of signals of duration T from the radiation source to be identified, generate a signal image of each signal using the same method as in step S1, and then input each signal image into the discriminator D n′ , the statistical output probability value is in the judgment interval [γ n′,min ,γ n′,max ] ratio ω′, if ω′>ω n′ , then it is determined that the radiation source to be identified is determined by the discriminator D n′ The identification is passed, otherwise the identification fails; then the following rules are used to determine whether the radiation source to be identified is a new radiation source: If the discriminator D N+1 If the identification fails, further identification of the discriminators D1 to D N Whether all identifications are passed, if so, the radiation source to be identified is determined to be an identified radiation source, otherwise it is determined to be a newly added radiation source; If the discriminator D N+1 If the identification is passed, the identification devices D1 to D N Whether all identifications fail, if so, the radiation source to be identified is determined to be a newly added radiation source, otherwise, the radiation source to be identified is determined to be an identified radiation source.

2. The newly added untrained radiation source detection method according to claim 1, characterized in that: In step S1, the signal is visualized using a signal image generation method based on segmented mapping. The specific method is: the I / Q data of the signal with a duration of T is evenly divided into M segments, the value of M is determined according to actual needs, each segment of I / Q data is mapped to obtain a sub-signal image, and then the obtained M segments of sub-signal images are superimposed, and the superimposed signal image is used as the signal image obtained by mapping the original signal.

3. The newly added untrained radiation source detection method according to claim 1, characterized in that: The discriminator in step S2 includes a first convolutional layer, an automatic convolution kernel selection module, a second convolutional layer and an identification module, wherein: The first convolutional layer is used to perform a 1×1 convolution operation on the input image and output the obtained feature image to the automatic convolution kernel selection module; The automatic convolution kernel selection module is used to perform a convolution operation of the received feature image by automatically selecting the convolution kernel, and output the obtained feature image to the second convolution layer; The second convolutional layer is used to perform a 1×1 convolution operation on the received feature image and output the obtained feature image to the identification module; The identification module is used to determine the authenticity of the received feature image and output the probability value that the input image is a real image.

4. The newly added untrained radiation source detection method according to claim 1, characterized in that: The minimum probability value γ in step S4 n′,min and the maximum probability value γ n′,max The method for determining γ is as follows: first, remove the abnormal maximum and minimum values ​​from the probability value set, then sort the remaining probability values ​​from large to small, and take the mean of the first K maximum values ​​as the maximum probability value γ n′,max , take the mean of the last K minimum values ​​as the minimum probability value γ n′,min , the value of K is determined according to actual needs.

Images