Radiation source individual identification method and device based on GRU-ResNet
By using GRU-ResNet network and Gram angle field transformation in individual recognition of radiation source, the problem of information loss during feature extraction and signal conversion in the prior art is solved, and a higher accuracy of individual recognition of radiation source and algorithm generalization ability are achieved.
Patent Information
- Application Number
- CN202411588890.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-08
- Publication Date
- 2025-04-08
AI Technical Summary
The prior art is difficult to effectively utilize unintentional modulation features in individual recognition of radiation sources, resulting in low generalization, and traditional methods are prone to loss of deep information during feature extraction and signal conversion.
The individual recognition method of radiation source based on GRU-ResNet is used to convert one-dimensional sequences into two-dimensional data through Gram Angle Field (GAF), and combined with the improved ResNet network and GRU network, feature extraction and signal modeling are performed to maintain the semantic information and sequence correlation of the signal to the greatest extent.
It improves the accuracy of individual recognition of radiation sources, reduces information loss during feature extraction, enhances the generalization ability of algorithms, and reduces the demand for hardware resources.
Smart Images
Figure CN119782775A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of radiation source identification, and in particular to a method and device for individual radiation source identification based on GRU-ResNet. Background Art
[0002] Specific Emitter Identification (SEI), also known as individual emitter identification, refers to the technology of extracting subtle differences from received signals to associate a single emitter. This technology is the basis for friend-or-foe discrimination, precise tracking, and attack, and is a key technical means in the field of electronic reconnaissance. In the early days, the emitter system was single, and individual identification could be completed through traditional signal modulation identification. In recent years, with the continuous development of the emitter system and signal modulation style, the traditional means of individual emitter identification through modulation style identification has become ineffective, and the technology of individual emitter identification has begun to attract the attention of relevant experts at home and abroad. The research on individual emitter identification has gradually increased, and related research has also gradually shifted from intentional modulation features to unintentional modulation features. Commonly used unintentional modulation features include transform domain features, higher order cumulant (HOC) features, etc. However, the extraction of these features is highly dependent on prior knowledge and transmission data, so the algorithm designed for individual identification based on simple feature extraction has low generalization.
[0003] As the application potential of deep learning (DL) in different fields is gradually developed, its excellent feature extraction ability in SEI problems has attracted widespread attention. So far, SEI research has evolved from traditional signal feature calculation to signal feature plus DL and direct use of DL to extract high-dimensional features. The current mainstream research focuses more on the combination of signal feature extraction and neural networks, but it is mostly focused on the recognition of a single domain and a single feature. Some feature fusion algorithms are also aimed at signal features of the same dimension. At present, the use of deep learning algorithms for individual identification of radiation sources can be roughly divided into three categories from a data perspective:
[0004] One is to extract features directly from the one-dimensional sequence. The advantage of this method is that it directly utilizes signal information. In theory, complete fingerprint features can be extracted through appropriate network design. The disadvantage is that since there is no complete mechanism explanation for the formation of radiation source fingerprints, it is difficult to design a network that takes into account both completeness and timeliness of fingerprint extraction.
[0005] The second is to convert the features in the sequence into two-dimensional images, and use the powerful classification ability of deep learning for individual identification. First, the DenseNet network is used to perform primary classification of the three types of image data sets, and then the Gradient Boosting Decision Tree (GBDT) classifier is used for secondary identification. After directly extracting multiple transformation features, the image data set is directly constructed for classification. The advantage of this method is that it can give full play to the image advantages of deep learning and facilitate the development of cutting-edge networks using computer vision. The disadvantage is that in the process of converting to images, images in different domains make different trade-offs for data features. Since there is no complete mechanism explanation for the formation of radiation source fingerprints, the deep information of the signal may be lost.
[0006] The third is to convert the features of different dimensions into one-dimensional features and then splice them to form multi-dimensional joint features, and then use the classifier to identify individuals. For example, the literature constructs a multi-dimensional feature vector containing multi-domain features by performing traditional feature analysis on the time domain, decomposition domain, frequency domain, and high-dimensional feature domain of the signal. The essence of its multi-dimensional feature vector construction is to convert all features into one-dimensional vectors and then splice them. Finally, the GBDT classifier is used to identify the individual radiation source, but the decision differences that may be caused by different dimensional features are not considered in the simulation and principle description process. The advantage of this method is that it can make use of the features of different transformation domains of the signal as much as possible, but it inevitably causes the loss of feature information when extracting the one-dimensional feature vector of high-dimensional features. Summary of the invention
[0007] The present invention provides a method and device for individual radiation source identification based on GRU-ResNet to solve the problems raised in the above background technology.
[0008] In order to solve the above technical problems, the technical solution adopted by the present invention is:
[0009] A method and device for identifying individual radiation sources based on GRU-ResNet, comprising the following steps:
[0010] Step 1: Data feature extraction: divided into A1: Gramian Angular Field (GAF); A2: 2D sequence high-dimensional feature extraction; A3: 1D sequence feature extraction;
[0011] Step 2: Identification framework: divided into B1: radiation source signal modeling; B2: GRU-ResNet identification network: divided into B21: improved ResNet network; B22: GRU-ResNet identification network architecture;
[0012] Step 3: Algorithm verification: divided into C1: parameter setting; C2: algorithm performance.
[0013] Preferably, the Gramian Angular Field (GAF):
[0014] The characteristic Gram angle field is the sequence X={x1,x2,…,x N}, N is the sequence length, and the sequence is normalized by the following formula
[0015]
[0016] Using the inverse function, the sequence Converted to a two-dimensional GAF matrix, its expression is as follows:
[0017]
[0018] in, A' represents the transpose of A, I is the unit row vector; when GAF is constructed using cosine, it is the Gram Angle Sum Field (GASF), and its expression is as follows:
[0019]
[0020] in, Represents the inner product; it is used to intuitively display the effect of converting the sequence into GASF.
[0021] Preferably, the high-dimensional feature extraction of the two-dimensional sequence:
[0022] The feature extraction of two-dimensional sequences can directly use image classification models. With the development of computer vision, in order to achieve better recognition effects, convolutional neural networks (CNN) have developed a series of variant networks with their own characteristics. In order to better extract high-dimensional features in two-dimensional sequences, ResNet50 network is used for feature extraction, where N is the first dimension of the output data of the current stage and the first dimension of the input data of the next stage. ReLU represents the activation function layer. Assuming that P is the first dimension of the input data of the current stage, M = P, N = 4P in stage 1, and M = P / 2N = 2P in stage 2-stage 4;
[0023] Among them, Cov2d(3,7,64) represents a two-dimensional convolutional layer with an input dimension of 3, a convolution kernel size of 7, and an output dimension of 64, BN is a batch normalization layer, MaxPool2d is a pooling layer, and Linear(2048,1000) represents a linear layer with an input dimension of 2048 and an output dimension of 1000 for final classification.
[0024] Preferably, the one-dimensional sequence feature extraction is:
[0025] Commonly used deep learning models include multi-layer perceptron, CNN and its variant networks, recurrent neural network (RNN) and attention mechanism. The MLP model has poor ability to capture time dependencies and is limited by the network architecture. CNN and its variant networks are more widely used for image processing of time series. RNN is a type of network designed for sequence data processing. Its improved network gated recurrent unit (Gated Recurrent Unit, GRU) has a simple architecture and can be used for deep architectures to suppress gradient disappearance and feature explosion problems. At the same time, it is efficient in calculation and has strong generalization ability. The attention mechanism can capture long-distance dependencies and is mostly used in natural language processing tasks. Taking comprehensive considerations, the GRU network is selected to extract one-dimensional sequence features. The calculation process of GRU is as follows:
[0026]
[0027] Among them, r t j To reset the gate, To update the gate, Indicates alternative status information. is the output of GRU at time t.
[0028] Preferably, the radiation source signal is modeled as follows:
[0029] Assuming that the receiver receives sequences sent by K radiation source signal transmitters, each sequence length is N, then the received sequence of the kth radiation source can be expressed as:
[0030]
[0031] Its GASF matrix can be expressed as:
[0032]
[0033] Preferably, the GRU-ResNet recognition network includes the following:
[0034] B21: Improved ResNet network: Since the ResNet50 network is designed for images, the input is a three-channel two-dimensional image, while the two-dimensional sequence has only one channel, so the Cov2d(3,7,64) in the ResNet50 network stage0 is changed to Cov2d(1,7,64), and the last classification layer is removed. In addition, in order to better extract the high-dimensional features of the signal, the projection networks Pro1 and Pro2 are added after the ResNet50 network;
[0035] B22: GRU-ResNet recognition network architecture: Due to the use of two-dimensional sequences for recognition, the memory space occupied during program operation is greatly reduced. The recognition framework can be executed using the CPU, which can reduce the algorithm's hardware requirements and can complete quick recognition on an ordinary computer or laptop. In addition, when adapting the artificial intelligence algorithm to domestic hardware, the CPU adaptation is less difficult and domestic migration is convenient. Abandoning images can reduce the memory requirements during model operation, speed up computing speed and reduce energy consumption.
[0036] Preferably, the parameter settings are:
[0037] The dataset used in the simulation is USRP-B210s based on GNU-Radio and SDR, with a total of 12 targets. The debugging methods cover BPSK, QPSK and 16QAM, the carrier frequency is 2.4GHz, the bandwidth is 100kHz, and the data is collected in indoor channels with different transmission / reception distances; the number of samples for each transmitter at the same distance is 1.5e3, the points of a single sample are 512, and the ratio of training set to test set is 7:3; the experiment is carried out on a framework based on Tensorflow 2.4.0, the CPU is Intel(R) Core(TM) i7-10750H, the GPU is RTX 3080, and the RAM is 16gb; Adam is selected as the optimizer, the epoch is 100, the learning rate is 0.001, and the batchsize is 16.
[0038] Preferably, the algorithm performance: 12 radiation source recognition accuracy, when there are 12 radiation source individuals, the recognition accuracy is nearly 100% after 100 iterations.
[0039] A radiation source individual identification device based on GRU-ResNet, a radiation source individual identification device implemented by a radiation source individual identification method based on GRU-ResNet.
[0040] Due to the adoption of the above technical solution, the present invention has the following technical advances compared with the prior art:
[0041] 1. The present invention provides a method and device for individual radiation source identification based on GRU-ResNet, and adopts a strategy for radiation source identification using one-dimensional and two-dimensional features. The strategy simultaneously utilizes the advantages of one-dimensional and two-dimensional features and their feature extraction networks, and supplements the shortcomings of the current status of individual identification feature utilization. In addition, a joint fusion method that can reduce dimensional redundancy and maintain semantic information consistency to the greatest extent and a Gram Angular Sum Field (GASF) transform that can maintain sequence correlation to the greatest extent are selected to avoid possible decision differences between features of different dimensions.
[0042] 2. The present invention provides a method and device for individual radiation source identification based on GRU-ResNet, which uses the Gramian Angular Field (GAF) to directly convert a one-dimensional sequence into two-dimensional data and designs a projection network to improve the ResNet network, mapping the features to a high-dimensional space, making full use of the advantages of the original sequence features and machine learning to process two-dimensional data, and improving the accuracy of individual radiation source identification. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 It is a schematic diagram of the original sequence of the present invention;
[0044] Figure 2 It is a schematic diagram of the GASF of the present invention;
[0045] Figure 3 Schematic diagram of the basic residual identity module structure of ResNet50 of the present invention;
[0046] Figure 4 It is a schematic diagram of the basic residual convolution module structure of ResNet50 of the present invention;
[0047] Figure 5 Schematic diagram of a gated cycle unit of the present invention;
[0048] Figure 6 This is a schematic diagram of GRU-ResNet recognition of the present invention;
[0049] Figure 7 It is a schematic diagram of the recognition accuracy of 12 radiation sources of the present invention. DETAILED DESCRIPTION
[0050] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further explained below in conjunction with specific implementation methods.
[0051] like Figure 1-Figure 7 As shown, a method and device for identifying individual radiation sources based on GRU-ResNet include the following steps:
[0052] Step 1: Data feature extraction: divided into A1: Gramian Angular Field (GAF); A2: 2D sequence high-dimensional feature extraction; A3: 1D sequence feature extraction;
[0053] Step 2: Identification framework: divided into B1: radiation source signal modeling; B2: GRU-ResNet identification network: divided into B21: improved ResNet network; B22: GRU-ResNet identification network architecture;
[0054] Step 3: Algorithm verification: divided into C1: parameter setting; C2: algorithm performance.
[0055] Among them, Gramian Angular Field (GAF):
[0056] The characteristic Gram angle field is the sequence X={x1,x2,…,x N}, N is the sequence length, and the sequence is normalized by the following formula
[0057]
[0058] Using the inverse function, the sequence Converted to a two-dimensional GAF matrix, its expression is as follows:
[0059]
[0060] in, A' represents the transpose of A, I is the unit row vector; when GAF is constructed using cosine, it is the Gram Angle Sum Field (GASF), and its expression is as follows:
[0061]
[0062] in, Represents the inner product; it is used to intuitively display the effect of converting the sequence into GASF.
[0063] Converting a one-dimensional sequence to a two-dimensional sequence through GASF can effectively preserve the temporal correlation and spatial relationship within the sequence, which is conducive to deep learning to extract the relationship between variables and features. However, this will also lead to the loss of pattern information in the original time series, reduce the accuracy of data classification, and a large number of images will increase the memory requirements during program operation. In order to reduce the overhead of the program operation process, images are used as features to directly extract high-dimensional features from the two-dimensional sequence. In order to make up for the loss of information such as patterns caused by the conversion of two-dimensional sequences, features are directly extracted from the one-dimensional sequence as a supplement. Since the two-dimensional sequence transformed by GASF maintains the consistency of correlation with the original sequence to the greatest extent, there is no need to consider data alignment and complex fusion methods after feature extraction, and the features can be directly spliced.
[0064] Among them, high-dimensional feature extraction of two-dimensional sequence:
[0065] The feature extraction of two-dimensional sequences can directly use image classification models. With the development of computer vision, in order to achieve better recognition effects, convolutional neural networks (CNN) have developed a series of variant networks with their own characteristics. In order to better extract high-dimensional features in two-dimensional sequences, ResNet50 network is used for feature extraction, where N is the first dimension of the output data of the current stage and the first dimension of the input data of the next stage. ReLU represents the activation function layer. Assuming that P is the first dimension of the input data of the current stage, M = P, N = 4P in stage 1, and M = P / 2N = 2P in stage 2-stage 4;
[0066] Among them, Cov2d(3,7,64) represents a two-dimensional convolutional layer with an input dimension of 3, a convolution kernel size of 7, and an output dimension of 64, BN is a batch normalization layer, MaxPool2d is a pooling layer, and Linear(2048,1000) represents a linear layer with an input dimension of 2048 and an output dimension of 1000 for final classification.
[0067] Table 1 ResNet50 network structure
[0068]
[0069]
[0070] Among them, one-dimensional sequence feature extraction:
[0071] Commonly used deep learning models include multi-layer perceptron, CNN and its variant networks, recurrent neural network (RNN) and attention mechanism. The MLP model has poor ability to capture time dependencies and is limited by the network architecture. CNN and its variant networks are more widely used for image processing of time series. RNN is a type of network designed for sequence data processing. Its improved network gated recurrent unit (Gated Recurrent Unit, GRU) has a simple architecture and can be used for deep architectures to suppress gradient disappearance and feature explosion problems. At the same time, it is efficient in calculation and has strong generalization ability. The attention mechanism can capture long-distance dependencies and is mostly used in natural language processing tasks. Taking comprehensive considerations, the GRU network is selected to extract one-dimensional sequence features. The calculation process of GRU is as follows:
[0072]
[0073] Among them, r t j To reset the gate, To update the gate, Indicates alternative status information. is the output of GRU at time t.
[0074] Among them, the radiation source signal is modeled:
[0075] Assuming that the receiver receives sequences sent by K radiation source signal transmitters, each sequence length is N, then the received sequence of the kth radiation source can be expressed as:
[0076]
[0077] Its GASF matrix can be expressed as:
[0078]
[0079] Among them, the GRU-ResNet recognition network includes the following:
[0080] B21: Improved ResNet network: Since the ResNet50 network is designed for images, the input is a three-channel two-dimensional image, while the two-dimensional sequence has only one channel, so the Cov2d(3,7,64) in the ResNet50 network stage0 is changed to Cov2d(1,7,64), and the last classification layer is removed. In addition, in order to better extract the high-dimensional features of the signal, the projection networks Pro1 and Pro2 are added after the ResNet50 network;
[0081] Table 2 Improved ResNet network structure
[0082]
[0083] B22: GRU-ResNet recognition network architecture: Due to the use of two-dimensional sequences for recognition, the memory space occupied during program operation is greatly reduced. The recognition framework can be executed using the CPU, which can reduce the algorithm's hardware requirements and can complete quick recognition on an ordinary computer or laptop. In addition, when adapting the artificial intelligence algorithm to domestic hardware, the CPU adaptation is less difficult and domestic migration is convenient. Abandoning images can reduce the memory requirements during model operation, speed up computing speed and reduce energy consumption.
[0084] Among them, the parameter settings are:
[0085] The dataset used in the simulation is USRP-B210s based on GNU-Radio and SDR, with a total of 12 targets. The debugging methods cover BPSK, QPSK and 16QAM, the carrier frequency is 2.4GHz, the bandwidth is 100kHz, and the data is collected in indoor channels with different transmission / reception distances; the number of samples for each transmitter at the same distance is 1.5e3, the points of a single sample are 512, and the ratio of training set to test set is 7:3; the experiment is carried out on a framework based on Tensorflow 2.4.0, the CPU is Intel(R) Core(TM) i7-10750H, the GPU is RTX 3080, and the RAM is 16gb; Adam is selected as the optimizer, the epoch is 100, the learning rate is 0.001, and the batchsize is 16.
[0086] The algorithm performance: the recognition accuracy of 12 radiation sources is close to 100% after 100 iterations when there are 12 radiation sources.
[0087] A radiation source individual identification device implemented by a radiation source individual identification method based on GRU-ResNet.
[0088] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.
Claims
1. A method for identifying radiation sources based on GRU-ResNet, characterized in that: The following steps are involved: Step 1: Data feature extraction: divided into A1: Gram angle field; A2: 2D sequence high-dimensional feature extraction; A3: One-dimensional sequence feature extraction; Step 2: Identification framework: divided into B1: radiation source signal modeling; B2: GRU-ResNet recognition network: divided into B21: improved ResNet network; B22: GRU-ResNet recognition network architecture; Step 3: Algorithm verification: divided into C1: parameter setting; C2: algorithm performance.
2. According to claim 1, a method for identifying radiation sources based on GRU-ResNet is characterized in that: The Gram Corner Field: The characteristic Gram angle field is the sequence X={x1,x2,…,x N }, N is the sequence length, and the sequence is normalized by the following formula Using the inverse function, the sequence Converted to a two-dimensional GAF matrix, its expression is as follows: in, A' represents the transpose of A, I is the unit row vector; when GAF is constructed using cosine, it is the Gram Angle Sum Field (GASF), and its expression is as follows: in, Represents the inner product; it is used to intuitively display the effect of converting the sequence into GASF.
3. The method for individual radiation source identification based on GRU-ResNet according to claim 1, characterized in that: The two-dimensional sequence high-dimensional feature extraction: The feature extraction of two-dimensional sequences can directly use image classification models. With the development of computer vision, in order to achieve better recognition effects, convolutional neural networks have developed a series of variant networks with their own characteristics. In order to better extract high-dimensional features in two-dimensional sequences, ResNet50 network is used for feature extraction, where N is the first dimension of the output data of the current stage and the first dimension of the input data of the next stage. ReLU represents the activation function layer. Assuming that P is the first dimension of the input data of the current stage, M=P, N=4P in stage1, and M=P / 2N=2P in stage2-stage4; Among them, Cov2d(3,7,64) represents a two-dimensional convolutional layer with an input dimension of 3, a convolution kernel size of 7, and an output dimension of 64, BN is a batch normalization layer, MaxPool 2d is a pooling layer, and Linear(2048,1000) represents a linear layer with an input dimension of 2048 and an output dimension of 1000 for the final classification.
4. The method for individual radiation source identification based on GRU-ResNet according to claim 1, characterized in that: The one-dimensional sequence feature extraction: Commonly used deep learning models include multi-layer perceptron, CNN and its variant networks, recurrent neural network (RNN) and attention mechanism. The MLP model has poor ability to capture time dependencies and is limited by the network architecture. CNN and its variant networks are more widely used for image processing of time series. RNN is a type of network designed for sequence data processing. Its improved network gated recurrent unit has a simple architecture and can be used for deep architectures to suppress gradient disappearance and feature explosion problems. It is also efficient in calculation and has strong generalization ability. The attention mechanism can capture long-distance dependencies and is mostly used for natural language processing tasks. Taking all factors into consideration, the GRU network is selected to extract one-dimensional sequence features. The calculation process of GRU is as follows: Among them, r t j To reset the gate, To update the gate, Indicates alternative status information. is the output of GRU at time t.
5. The method for identifying radiation sources based on GRU-ResNet according to claim 1, characterized in that: The radiation source signal modeling: Assuming that the receiver receives sequences sent by K radiation source signal transmitters, each sequence length is N, then the received sequence of the kth radiation source can be expressed as: Its GASF matrix can be expressed as:
6. The method for individual radiation source identification based on GRU-ResNet according to claim 1, characterized in that: The GRU-ResNet recognition network includes the following: B21: Improved ResNet network: Since the ResNet50 network is designed for images, the input is a three-channel two-dimensional image, while the two-dimensional sequence has only one channel, so the Cov2d(3,7,64) in the ResNet50 network stage0 is changed to Cov2d(1,7,64), and the last classification layer is removed. In addition, in order to better extract the high-dimensional features of the signal, the projection networks Pro1 and Pro2 are added after the ResNet50 network; B22: GRU-ResNet recognition network architecture: Due to the use of two-dimensional sequences for recognition, the memory space occupied during program operation is greatly reduced. The recognition framework can be executed using the CPU, which can reduce the algorithm's hardware requirements and can complete quick recognition on an ordinary computer or laptop. In addition, when adapting the artificial intelligence algorithm to domestic hardware, the CPU adaptation is less difficult and domestic migration is convenient. Abandoning images can reduce the memory requirements during model operation, speed up computing speed and reduce energy consumption.
7. The method for individual radiation source identification based on GRU-ResNet according to claim 1, characterized in that: The parameter settings are: The dataset used in the simulation is USRP-B210s based on GNU-Radio and SDR, with a total of 12 targets. The debugging methods cover BPSK, QPSK and 16QAM, the carrier frequency is 2.4GHz, the bandwidth is 100kHz, and the data is collected in indoor channels with different transmission / reception distances; the number of samples for each transmitter at the same distance is 1.5e3, the points of a single sample are 512, and the ratio of training set to test set is 7:3; the experiment is carried out on a framework based on Tensorflow 2.4.0, the CPU is Intel(R) Core(TM) i7-10750H, the GPU is RTX 3080, and the RAM is 16gb; Adam is selected as the optimizer, the epoch is 100, the learning rate is 0.001, and the batchsize is 16.
8. The method for individual radiation source identification based on GRU-ResNet according to claim 1, characterized in that: The algorithm performance: the recognition accuracy of 12 radiation sources is close to 100% after 100 iterations when there are 12 radiation sources.
9. A radiation source individual identification device based on GRU-ResNet, characterized in that: A radiation source individual identification device implemented by the radiation source individual identification method based on GRU-ResNet as described in any one of claims 1-8.