A method for emitter individual identification based on adaptive iterative multi-order wavelet coefficients

Abstract

The application discloses a radiation source individual identification method based on adaptive iterative multi-order wavelet coefficients, which comprises the following steps: taking the modified radiation source I and Q sequences as signal samples, and performing multi-order wavelet decomposition on the I and Q sequences respectively; calculating the optimal number of wavelet decomposition layers for different signal samples to obtain the wavelet coefficients of the I and Q sequences with the optimal number of layers; longitudinally splicing the corresponding features of the I and Q sequences, and transversely splicing the feature information of all signal samples to obtain adaptive iterative multi-order wavelet coefficients of the radiation source individual; constructing a neural network model based on a Transform network and an inverted residual mode, taking the obtained adaptive iterative multi-order wavelet coefficients as input to perform model training, and finally performing classification and identification of the radiation source individual. The application has low calculation cost, high processing efficiency, strong radiation source fingerprint feature extraction capability, high classification precision and strong generalization ability of the identification system, and can identify radiation sources of various systems in complex actual application scenarios.

Description

Technical Field

[0001] This invention relates to electromagnetic wave and signal recognition technology, and in particular to a method for individual radiation source identification based on adaptive iterative multi-order wavelet coefficients. Background Technology

[0002] Individual radiation source identification technology can detect and identify target signals of interest without relying on intrinsic information, even in situations involving information encryption, missing intrinsic information, and dynamic allocation of channel resources. It then associates these signals with the radiation source target, its carrier platform, and the user's identity, thus possessing significant practical value. In the information age, individual radiation source identification is a crucial means of gaining information control. It primarily relies on the distinctive characteristics of different radiation source devices in their transmitted signals to determine which device the signal originates from, thereby enabling the tracking and identification of the radiation source device. Specifically, individual radiation source identification technology often utilizes the extracted fingerprint characteristics of individual radiation sources, including the spectral characteristics, temporal characteristics, spatial characteristics, and modulation characteristics of communication radiation sources. By analyzing and comparing these characteristics, the detection and control of communication signals can be achieved.

[0003] Feature extraction of radiation sources under cooperative conditions is relatively easy, as pre-defined pseudo-random sequences can be embedded in the transmitted signals for identification. However, in practical applications, most communication occurs under non-cooperative conditions, which greatly complicates radiation source fingerprint extraction. Researchers both domestically and internationally have studied a series of feature extraction methods focusing on the fingerprint characteristics of individual radiation sources. These methods have achieved relatively ideal experimental results under certain conditions; however, in complex electromagnetic environments and with real detection data, these methods are severely affected. The main reason is that existing feature extraction methods fail to extract robust, separable, and essential subtle features of individual radiation sources.

[0004] The problem of identifying individual radiation sources is logically a classification decision problem. Deep learning has some significant advantages over traditional machine learning methods in this field, such as automated feature extraction, efficient handling of complex patterns, strong generalization and adaptive learning. Therefore, using deep learning methods for identifying individual radiation sources can fully extract the multi-scale features of received signals in complex electromagnetic environments, enabling the network to obtain more separable fingerprint information. Summary of the Invention

[0005] The purpose of this invention is to provide a radiation source individual identification method that has low computational cost, high processing efficiency, strong ability to extract radiation source fingerprint features, high classification accuracy, strong generalization ability of the identification system, and can identify multiple radiation sources of different systems in complex practical application scenarios.

[0006] The technical solution to achieve the purpose of this invention is: a method for identifying individual radiation sources based on adaptive iterative multi-order wavelet coefficients, comprising the following steps:

[0007] Step 1: Perform data preprocessing on the received radiation source signal to remove unwanted peaks in the signal and obtain the corrected radiation source IQ sequence as a signal sample.

[0008] Step 2: Perform multi-order wavelet decomposition on the I and Q paths of the corrected IQ sequence, respectively;

[0009] Step 3: For different signal samples, calculate the optimal number of wavelet decomposition layers to obtain the wavelet coefficients of the optimal number of layers for the I-channel and Q-channel sequences;

[0010] Step 4: Vertically splice the corresponding features from the I-channel sequence and the Q-channel sequence, retain the inherent structural relationship between the I and Q sequences, and horizontally splice the feature information of all signal samples of the radiation source individual to obtain the adaptive iterative multi-order wavelet coefficients of the radiation source individual.

[0011] Step 5: Construct a neural network model based on Transformer network and inverted residuals. Use the adaptive iterative multi-order wavelet coefficients of the individual radiation sources obtained in Step 4 as input to train the neural network model.

[0012] Step 6: Use the trained neural network model to classify and identify individual radiation sources.

[0013] Furthermore, in step 2, the corrected IQ sequence is subjected to multi-order wavelet decomposition on both the I and Q paths, as follows:

[0014] Step 2.1: Discretely sample the wavelet function using the scaling factor a and the translation factor b;

[0015] Step 2.2: Obtain the wavelet coefficients of the signal sample through DWT transform, where DWT transform represents discrete wavelet transform;

[0016] Step 2.3: Pass the result after DWT transformation through a low-pass filter and a high-pass filter to obtain the low-frequency component cA of the DWT transformed signal. i and high-frequency component cD j i and j represent the decomposition levels;

[0017] Step 2.4: For each layer, the low-frequency component cA is obtained. i Perform DWT transformation to obtain the decomposition of the next layer;

[0018] Step 2.5: Calculate the required low-frequency components cA for each layer. i and high-frequency component cD jBy summing up the number of wavelet coefficients, we can obtain all the wavelet coefficient results after the DWT transform.

[0019] Further, in step 2.2, the wavelet coefficients of the signal samples are obtained through DWT transform, as follows:

[0020] Based on the different factor extraction methods, DWT transform is divided into two types: Redundant Wavelet Transform (RWT) and Multiple Resolution Analysis (MRA).

[0021] In the Redundant Wavelet Transform (RWT), the scaling factor α expands exponentially, i.e. Where a0>0, m∈Z, and Z represents the set of integers, for wide wavelets we want to translate with a larger step size, so we define a translation factor. Where b0 > 0, n ∈ Z; a0, b0, m, and n are all positive integers. Different initial values ​​are set according to different signals to form scaling and translation factors. In summary, the wavelet function and DWT formula are as follows:

[0022] Where m, n∈Z

[0023]

[0024] Among them Ψ m,n (k) represents the wavelet function in terms of m and n; Ψ(·) represents the mother wavelet function, and the corresponding wavelet basis function is obtained by scaling and translating the mother wavelet function; <·> represents the discrete wavelet transform, and f(k) represents the discrete signal;

[0025] In multiresolution analysis (MRA), the scaling factor a and the translation factor b are selected according to powers of 2, and downsampling is performed, i.e., let Where a0 = 2, b0 = 1, m, n ∈ Z, the relevant wavelet function and DWT formula are obtained as follows:

[0026] Where (m,n∈Z)

[0027]

[0028] Furthermore, in step 3, the optimal number of wavelet decomposition layers is calculated for different signal samples to obtain the wavelet coefficients of the optimal number of layers for the I-path and Q-path sequences, as detailed below:

[0029] Step 3.1: Assume a radiation source individual has N samples. The wavelet coefficients obtained by decomposing the I-path sequence of the j-th sample at the i-th level are... The original signal sequence is represented as f(k)={k1,k2,···,k n}, where n represents the total number of sample points, then the wavelet coefficients obtained from the I-way sequence of the j-th sample after decomposition at the i-th level are... The relative signal-to-noise ratio compared to the original signal sequence f(k) is:

[0030]

[0031] Step 3.2: Select the decomposition layer with the highest relative signal-to-noise ratio, and decompose the wavelet coefficients of that layer. The optimal fingerprint feature information for the j-th sample I-path sequence;

[0032] Step 3.3: Following the methods in Steps 3.1 to 3.2, obtain the optimal fingerprint feature information of the Q-path sequence of the j-th sample.

[0033] Further, in step 4, the corresponding features from the I-channel and Q-channel sequences are vertically spliced ​​together, preserving the inherent structural relationship between the I and Q sequences, and the feature information of all signal samples of the radiation source individual is horizontally spliced ​​together to obtain the adaptive iterative multi-order wavelet coefficients of the radiation source individual, expressed as:

[0034] Where j = 1, 2, ..., N

[0035] X = h(X) 1 ,X 2 ,···,X N )

[0036] Where h(·) and v(·) represent the horizontal and vertical splicing operators, respectively, and X j X represents the vertical splicing result of the feature information of the j-th signal sample, and X is the adaptive iterative multi-order wavelet coefficient of a single radiation source.

[0037] Furthermore, in step 5, a neural network model is constructed based on the Transformer network and inverted residuals. The adaptive iterative multi-order wavelet coefficients of the individual radiation sources obtained in step 4 are used as input to the neural network model for training, as detailed below:

[0038] Step 5.1: Construct a neural network model based on Transformer networks and inverted residuals;

[0039] Step 5.2: Use the adaptive iterative multi-order wavelet coefficients of the individual radiation sources obtained in Step 4 as the input to the neural network model, and normalize the input X of the neural network model:

[0040]

[0041]

[0042] Where mean(·) and std(·) represent the mean operator and the standard deviation operator, respectively;

[0043] Step 5.3: Treat both the first and second parts of the neural network model as feature extractors, and name them as... Where x is the input to the network model, θ m Here are the network parameters, where Let c be a feature tensor of dimension c. After the input is processed by the first and second parts, it is represented as:

[0044] Step 5.4, let the fully connected layer be... The network prediction result y is obtained. pred for:

[0045]

[0046] Step 5.5: Given the true label y of the individual radiation source. true The network model is trained by optimizing the following classification loss function:

[0047]

[0048] Where, θ={θ m ,θ n}, L c =y true log(y pred ) represents the cross-entropy loss function.

[0049] Furthermore, in step 5.1, the neural network model includes a first part, a second part, and a fully connected layer;

[0050] In the first part, the network input first extracts features through a convolution, and then enters four adjacent inverted residual modules. In each inverted residual module, there is an up-dimensional convolution, a depthwise convolution and a down-dimensional convolution. After each convolution, a batch normalization operation is performed, and finally, the dimensionality is reduced by another convolution.

[0051] In the second part, the Patch Embedding layer adds positional encoding to the input, which is then fed into 8 adjacent Encoder modules for deep feature extraction. Each Encoder module contains two layer normalization operations, and after the two layer normalization operations, a multi-head attention layer and a multi-layer perceptron are connected to extract feature embeddings.

[0052] Finally, a fully connected layer is used to classify and identify individual radiation sources.

[0053] Compared with the prior art, the significant advantages of this invention are: (1) It uses multi-order wavelet transform for data decomposition, which can fully extract the multi-scale features of the received signal under complex electromagnetic environment, so that the network can obtain more separable fingerprint information; (2) It adopts a network model based on Transformer network and "inverted residual" idea, which can efficiently process the data stream from the front end under the premise of low computational cost and achieve high-precision classification; (3) It can identify individuals of radiation sources with different systems and has strong generalization ability. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the radiation source individual identification method based on adaptive iterative multi-order wavelet coefficients of the present invention.

[0055] Figure 2 This is a schematic diagram of the wavelet decomposition process in the embodiment.

[0056] Figure 3 This is a schematic diagram of the neural network structure constructed in the embodiment.

[0057] Figure 4 This is a schematic diagram of the data acquisition process in the embodiment.

[0058] Figure 5 The graphs show the recognition results of the three datasets in the example under different signal-to-noise ratios.

[0059] Figure 6 This is a comparison chart of the present invention and several existing algorithms on an ultra-shortwave dataset in the embodiments.

[0060] Figure 7 This is a comparison chart of the present invention and several existing algorithms on a shortwave dataset in the embodiments.

[0061] Figure 8 This is a comparison chart of the present invention and several existing algorithms in the example using a radio walkie-talkie dataset. Detailed Implementation

[0062] This invention discloses a method for identifying individual radiation sources based on adaptive iterative multi-order wavelet coefficients, comprising the following steps:

[0063] Step 1: Perform data preprocessing on the received radiation source signal to remove unwanted peaks in the signal and obtain the corrected radiation source IQ sequence as a signal sample.

[0064] Step 2: Perform multi-order wavelet decomposition on the I and Q paths of the corrected IQ sequence, respectively;

[0065] Step 3: For different signal samples, calculate the optimal number of wavelet decomposition layers to obtain the wavelet coefficients of the optimal number of layers for the I-channel and Q-channel sequences;

[0066] Step 4: Vertically splice the corresponding features from the I-channel sequence and the Q-channel sequence, retain the inherent structural relationship between the I and Q sequences, and horizontally splice the feature information of all signal samples of the radiation source individual to obtain the adaptive iterative multi-order wavelet coefficients of the radiation source individual.

[0067] Step 5: Construct a neural network model based on Transformer network and inverted residuals. Use the adaptive iterative multi-order wavelet coefficients of the individual radiation sources obtained in Step 4 as input to train the neural network model.

[0068] Step 6: Use the trained neural network model to classify and identify individual radiation sources.

[0069] As a specific example, in step 2, the corrected IQ sequence is subjected to multi-order wavelet decomposition on both the I and Q paths, as follows:

[0070] Step 2.1: Discretely sample the wavelet function using the scaling factor a and the translation factor b;

[0071] Step 2.2: Obtain the wavelet coefficients of the signal sample through DWT transform, where DWT transform represents discrete wavelet transform;

[0072] Step 2.3: Pass the result after DWT transformation through a low-pass filter and a high-pass filter to obtain the low-frequency component cA of the DWT transformed signal. i and high-frequency component cD j i and j represent the decomposition levels;

[0073] Step 2.4: For each layer, the low-frequency component cA is obtained. i Perform DWT transformation to obtain the decomposition of the next layer;

[0074] Step 2.5: Calculate the required low-frequency components cA for each layer. i and high-frequency component cD j By summing up the number of wavelet coefficients, we can obtain all the wavelet coefficient results after the DWT transform.

[0075] As a specific example, in step 2.2, the wavelet coefficients of the signal sample are obtained through DWT transform, as follows:

[0076] Based on the different factor extraction methods, DWT transform is divided into two types: Redundant Wavelet Transform (RWT) and Multiple Resolution Analysis (MRA).

[0077] In the Redundant Wavelet Transform (RWT), the scaling factor α expands exponentially, i.e. Where a0>0, m∈Z, and Z represents the set of integers, for wide wavelets we want to translate with a larger step size, so we define a translation factor. Where b0 > 0, n ∈ Z; a0, b0, m, and n are all positive integers. Different initial values ​​are set according to different signals to form scaling and translation factors. In summary, the wavelet function and DWT formula are as follows:

[0078] Where m, n∈Z

[0079]

[0080] Among them Ψ m,n (k) represents the wavelet function in terms of m and n; Ψ(·) represents the mother wavelet function, and the corresponding wavelet basis function is obtained by scaling and translating the mother wavelet function; <·> represents the discrete wavelet transform, and f(k) represents the discrete signal;

[0081] In multiresolution analysis (MRA), the scaling factor a and the translation factor b are selected according to powers of 2, and downsampling is performed, i.e., let Where a0 = 2, b0 = 1, m, n ∈ Z, the relevant wavelet function and DWT formula are obtained as follows:

[0082] Where (m,n∈Z)

[0083]

[0084] As a specific example, in step 3, the optimal number of wavelet decomposition layers is calculated for different signal samples to obtain the wavelet coefficients of the optimal number of layers for the I-path and Q-path sequences, as follows:

[0085] Step 3.1: Assume a radiation source individual has N samples. The wavelet coefficients obtained by decomposing the I-path sequence of the j-th sample at the i-th level are... The original signal sequence is represented as f(k)={k1,k2,···,k n}, where n represents the total number of sample points, then the wavelet coefficients obtained from the I-way sequence of the j-th sample after decomposition at the i-th level are... The relative signal-to-noise ratio compared to the original signal sequence f(k) is:

[0086]

[0087] Step 3.2: Select the decomposition layer with the highest relative signal-to-noise ratio, and decompose the wavelet coefficients of that layer. The optimal fingerprint feature information for the j-th sample I-path sequence;

[0088] Step 3.3: Following the methods in Steps 3.1 to 3.2, obtain the optimal fingerprint feature information of the Q-path sequence of the j-th sample.

[0089] As a specific example, in step 4, the corresponding features from the I-channel and Q-channel sequences are vertically concatenated to preserve the inherent structural relationship between the I and Q sequences, and the feature information of all signal samples of the radiation source individual is horizontally concatenated to obtain the adaptive iterative multi-order wavelet coefficients of the radiation source individual, expressed as:

[0090] Where j = 1, 2, ..., N

[0091] X = h(X) 1 ,X 2 ,···,X N )

[0092] Where h(·) and v(·) represent the horizontal and vertical splicing operators, respectively, and X j X represents the vertical splicing result of the feature information of the j-th signal sample, and X is the adaptive iterative multi-order wavelet coefficient of a single radiation source.

[0093] As a specific example, in step 5, a neural network model is constructed based on the Transformer network and inverted residuals. The adaptive iterative multi-order wavelet coefficients of the individual radiation sources obtained in step 4 are used as input to the neural network model for training, as follows:

[0094] Step 5.1: Construct a neural network model based on Transformer networks and inverted residuals;

[0095] Step 5.2: Use the adaptive iterative multi-order wavelet coefficients of the individual radiation sources obtained in Step 4 as the input to the neural network model, and normalize the input X of the neural network model:

[0096]

[0097]

[0098] Where mean(·) and std(·) represent the mean operator and the standard deviation operator, respectively;

[0099] Step 5.3: Treat both the first and second parts of the neural network model as feature extractors, and name them as... Where x is the input to the network model, θ m Here are the network parameters, where Let c be a feature tensor of dimension c. After the input is processed by the first and second parts, it is represented as:

[0100] Step 5.4, let the fully connected layer be... The network prediction result y is obtained. pred for:

[0101]

[0102] Step 5.5: Given the true label y of the individual radiation source. true The network model is trained by optimizing the following classification loss function:

[0103]

[0104] Where, θ={θ m ,θ n}, L c =y true log(y pred ) represents the cross-entropy loss function.

[0105] As a specific example, in step 5.1, the neural network model includes a first part, a second part, and a fully connected layer;

[0106] In the first part, the network input first undergoes initial feature extraction through a single convolution, followed by four adjacent inverted residual modules. Each inverted residual module contains one dimensionality-up convolution, one depthwise convolution, and one dimensionality-reducing convolution, with batch normalization performed after each convolution. This ensures that the number of parameters remains constant while fully extracting the input features. Finally, dimensionality reduction is achieved through another convolution.

[0107] In the second part, the Patch Embedding layer adds positional encoding to the input, which is more conducive to the extraction of local features of the data. Then, it enters 8 adjacent Encoder modules for deep feature extraction. Each Encoder module contains two layer normalization operations, and after the two layer normalization operations, a multi-head attention layer and a multi-layer perceptron are connected respectively. The multi-head attention mechanism captures long-distance dependencies in the feature sequence to extract more generalizable and separable feature embeddings.

[0108] Finally, a fully connected layer is used to classify and identify individual radiation sources.

[0109] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0110] Example 1

[0111] Combination Figure 1 This invention discloses a method for identifying individual radiation sources based on adaptive iterative multi-order wavelet coefficients, comprising the following steps:

[0112] Step 1: Perform data preprocessing on the received radiation source signal to remove unnecessary peaks in the signal and obtain the corrected radiation source IQ sequence;

[0113] Step 2: Perform multi-order wavelet decomposition on the corrected radiation source IQ sequence in both the I-channel and Q-channel data, as follows:

[0114] The number of wavelet coefficients depends on the number of wavelet transform layers. Figure 2 This is a schematic diagram of the wavelet three-level decomposition results, where cA i Represents low-frequency components, cD j represents the high-frequency component, and i and j represent the decomposition layer number.

[0115] Step 2.1: Discretely sample the wavelet function using the scaling factor a and the translation factor b;

[0116] Step 2.2: Obtain the wavelet coefficients of the signal samples through Discrete Wavelet Transform (DWT), as follows:

[0117] Based on the different factor extraction methods, discrete wavelet transform is divided into two types: redundant wavelet transform (RWT) and multiple resolution analysis (MRA).

[0118] In RWT, the scaling factor 'a' expands exponentially, i.e. Where a0>0, m∈Z, because for wide wavelets we want to translate with a larger step size, hence the definition Where b0>0, n∈Z, the wavelet function and DWT formula in this case are as follows:

[0119]

[0120]

[0121] In MRA, the scaling factor a and the translation factor b are selected according to powers of 2, and downsampling is performed, that is, let Where a0 = 2, b0 = 1, m, n ∈ Z, the relevant wavelet function and DWT formula are obtained as follows:

[0122] Where (m,n∈Z)

[0123]

[0124] Step 2.3: Pass the result after DWT transformation through a low-pass or high-pass filter to obtain the low-frequency component cA of the wavelet decomposition signal. i Or high-frequency component cD j ;

[0125] Step 2.4: Sequentially process the obtained approximate components cA i Perform DWT to obtain the breakdown amount for the next level;

[0126] Step 2.5: Calculate the required CA for each layer. i and cD j By summing up the number of wavelet coefficients, we can obtain all the wavelet coefficient results after DWT decomposition.

[0127] Step 3: For different signal samples, calculate the optimal number of wavelet decomposition layers to obtain the wavelet coefficients of the optimal number of layers for the I-channel and Q-channel sequences, as follows:

[0128] Step 3.1: Assume that there are N samples of a certain radiation source individual, and the wavelet coefficients obtained by decomposing the I-path sequence of the j-th sample at the i-th level are... The original signal sequence is represented as f(k)={k1,k2,···,k n}, where n represents the total number of sample points, then the relative signal-to-noise ratio of the wavelet coefficients obtained after the i-th level decomposition of the I-path sequence of the j-th sample compared to the original signal sequence is:

[0129]

[0130] Step 3.2: Select the decomposition level with the highest relative signal-to-noise ratio, and decompose the wavelet coefficients of that level. The optimal fingerprint feature information of the j-th sample I-path sequence.

[0131] Step 4: Vertically concatenate the corresponding features from the I-path and Q-path sequences, preserving the inherent structural relationship of the original IQ sequences, and horizontally concatenate the feature information of all samples of the radiation source individual to obtain the adaptive iterative multi-order wavelet coefficients of the radiation source individual, which are used as the input of the neural network.

[0132] The adaptive iterative multi-order wavelet coefficients of an individual radiation source are expressed as follows:

[0133]

[0134] X = h(X) 1 ,X 2 ,···,X N )

[0135] Where h(·) and v(·) represent the horizontal and vertical splicing operators, respectively, and X is the adaptive iterative multi-order wavelet coefficient of a single radiation source.

[0136] Step 5: Based on the Transformer network and the idea of ​​"inverted residuals," construct and train a neural network. Use the trained neural network to classify and identify individual radiation sources, combined with... Figure 3 The details are as follows:

[0137] Step 5.2: Normalize the input X of the neural network:

[0138]

[0139]

[0140] Where mean(·) and std(·) represent the mean operator and the standard deviation operator, respectively;

[0141] Step 5.3: Treat both the first and second parts of the neural network as feature extractors, and name them as follows: Where x is the network input, θ m For the corresponding network parameters, the input after passing through the first and second parts of the network is represented as follows:

[0142] Step 5.4, let the fully connected layer be... The network prediction results are as follows:

[0143]

[0144] Step 5.5: Given the true label y of the individual radiation source. true The network is trained by optimizing the following classification loss function:

[0145]

[0146] Where, θ={θ m ,θ n}, L c =y true log(y pred () represents the cross-entropy loss function;

[0147] Step 5.6: Use the trained neural network to classify and identify individual radiation sources.

[0148] Example 2

[0149] This embodiment verifies the effectiveness and superiority of the proposed method on three different datasets: a VHF / UHF FM radio dataset, a HF / USB radio dataset, and a walkie-talkie dataset. Each dataset contains radio transmission signal data from three different speakers (A, B, and C). The VHF / UHF FM and HF / USB radio datasets include three signal propagation modes: short-range direct transmission, short-range diffraction, and long-range diffraction. The walkie-talkie dataset includes three signal propagation modes: short-range diffraction, short-range direct transmission, and long-range direct transmission. All datasets were collected to verify whether the proposed model is affected by factors such as shortwave and VHF frequency bands, FM and USB modulation styles, individual speaker information (A, B, and C), and signal propagation modes. Each dataset was sampled from the transmission signals of five VHF / UHF FM radios, HF / USB radios, and walkie-talkies of the same manufacturer, model, and operating mode. The specific data collection methods are as follows... Figure 4 As shown. In each dataset, each radio station contains three different information individuals and three different signal transmission methods corresponding to each information individual, resulting in 45 sample datasets from 5 radio stations. Based on these datasets, artificial noise is added to simulate noise data ranging from -8dB to 8dB with a 2dB interval, and these are then combined with the original datasets to form new datasets.

[0150] Figure 5 The graph shows the recognition rate of three different datasets as a function of signal-to-noise ratio. As can be seen from the graph, each dataset can achieve a recognition rate of nearly 100% when the signal-to-noise ratio is the highest. Moreover, the recognition rate does not drop significantly when the signal-to-noise ratio is low, which demonstrates the effectiveness of the method of this invention. Figure 6 , Figure 7 , Figure 8 The figures show the comparison results between the method of the present invention and several existing algorithms under ultra-shortwave datasets, shortwave datasets, and radio two-way radio datasets. As can be seen from the figures, the method of the present invention outperforms the existing methods under each signal-to-noise ratio condition, demonstrating the superiority of the present invention.

[0151] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for identifying individual radiation sources based on adaptive iterative multi-order wavelet coefficients, characterized in that, Includes the following steps: Step 1: Perform data preprocessing on the received radiation source signal to remove unwanted peaks in the signal and obtain the corrected radiation source IQ sequence as a signal sample. Step 2: Perform multi-order wavelet decomposition on the I and Q paths of the corrected radiation source IQ sequence, respectively; Step 3: For different signal samples, calculate the optimal number of wavelet decomposition layers to obtain the wavelet coefficients of the optimal number of layers for the I-channel and Q-channel sequences; In step 3, the optimal number of wavelet decomposition layers is calculated for different signal samples to obtain the wavelet coefficients of the optimal number of layers for the I-path and Q-path sequences, as detailed below: Step 3.1: Establish the existence of a radiation source entity. The nth sample, the th The I-path sequence of the nth sample in the nth... The wavelet coefficients obtained from the layer decomposition are The original signal sequence is represented as , Representing the total number of sample points, the i-th The I-path sequence of the nth sample in the nth... Wavelet coefficients obtained after layer decomposition Compared to the original signal sequence The relative signal-to-noise ratio is: Step 3.2: Select the decomposition layer with the highest relative signal-to-noise ratio, and decompose the wavelet coefficients of that layer. As the first The optimal fingerprint feature information of each sample I-path sequence; Step 3.3: Following the methods in steps 3.1 to 3.2, obtain the... Optimal fingerprint feature information of each sample Q-path sequence ; Step 4: Vertically splice the corresponding features from the I-channel sequence and the Q-channel sequence, retain the inherent structural relationship between the I and Q sequences, and horizontally splice the feature information of all signal samples of the radiation source individual to obtain the adaptive iterative multi-order wavelet coefficients of the radiation source individual. Step 5: Construct a neural network model based on Transformer network and inverted residuals. Use the adaptive iterative multi-order wavelet coefficients of the individual radiation sources obtained in Step 4 as input to train the neural network model. Step 6: Use the trained neural network model to classify and identify individual radiation sources.

2. The radiation source individual identification method based on adaptive iterative multi-order wavelet coefficients according to claim 1, characterized in that, In step 2, the corrected radiation source IQ sequence is subjected to multi-order wavelet decomposition on both the I and Q paths, as follows: Step 2.1: Scaling factor for the wavelet function Translation factor Perform discrete sampling; Step 2.2: Obtain the wavelet coefficients of the signal sample through DWT transform, where DWT transform represents discrete wavelet transform; Step 2.3: Pass the result after DWT transformation through a low-pass filter and a high-pass filter to obtain the low-frequency component of the DWT transformed signal. and high frequency components , and Represents the number of decomposition levels; Step 2.4: Process the low-frequency components obtained from each layer. Perform DWT transformation to obtain the decomposition of the next layer; Step 2.5: Select the required low-frequency components for each layer. and high frequency components By summing up the number of wavelet coefficients, we can obtain all the wavelet coefficient results after the DWT transform.

3. The radiation source individual identification method based on adaptive iterative multi-order wavelet coefficients according to claim 2, characterized in that, In step 2.2, the wavelet coefficients of the signal samples are obtained through DWT transform, as follows: Based on the different factor extraction methods, DWT transform is divided into two types: Redundant Wavelet Transform (RWT) and Multiple Resolution Analysis (MRA). In the Redundant Wavelet Transform (RWT), the scaling factor With exponential expansion, that is ,in , To represent the set of integers, for wide wavelets we want to translate with a larger step size, so we define a translation factor. ,in ; , , and All are positive integers. Different initial values ​​are set according to different signals to form scaling and translation factors. In summary, the wavelet function and DWT formula are as follows: in For use and The wavelet function represents; The mother wavelet function is represented by the wavelet basis function, which is obtained by scaling and translating the mother wavelet function. Represents discrete wavelet transform. Represents discrete signals; In multiresolution analysis (MRA), scaling factor Translation factor Select according to powers of 2, and perform downsampling, that is, let ,in Thus, the relevant wavelet function and DWT formula are obtained as follows: 。 4. The radiation source individual identification method based on adaptive iterative multi-order wavelet coefficients according to claim 3, characterized in that, In step 4, the corresponding features from the I-channel and Q-channel sequences are vertically concatenated, preserving the inherent structural relationship between the I and Q sequences. Then, the feature information of all signal samples from this individual radiation source is horizontally concatenated to obtain the adaptive iterative multi-order wavelet coefficients of the individual radiation source, expressed as: in, and These represent the horizontal and vertical splicing operators, respectively. Indicates the first The result of vertically splicing the feature information of each signal sample For individual radiation sources, the adaptive iterative multi-order wavelet coefficients are used.

5. The radiation source individual identification method based on adaptive iterative multi-order wavelet coefficients according to claim 4, characterized in that, In step 5, a neural network model is constructed based on the Transformer network and inverted residuals. The adaptive iterative multi-order wavelet coefficients of the individual radiation sources obtained in step 4 are used as input to the neural network model for training, as detailed below: Step 5.1: Construct a neural network model based on Transformer networks and inverted residuals; Step 5.2: Use the adaptive iterative multi-order wavelet coefficients of the individual radiation sources obtained in Step 4 as the input to the neural network model. Perform normalization: in, and These represent the mean operator and the standard deviation operator, respectively. Step 5.3: Treat both the first and second parts of the neural network model as feature extractors, and name them as... ,in As input to the network model, Here are the network parameters, where The dimension is The feature tensor, after the input is processed by the first and second parts, is represented as ; Step 5.4, let the fully connected layer be... The network prediction results were obtained. for: Step 5.5: Provide the true label of the individual radiation source. The network model is trained by optimizing the following classification loss function: in, , This represents the cross-entropy loss function.

6. The radiation source individual identification method based on adaptive iterative multi-order wavelet coefficients according to claim 5, characterized in that, In step 5.1, the neural network model includes a first part, a second part, and a fully connected layer; In the first part, the network input first extracts features through a convolution, and then enters four adjacent inverted residual modules. In each inverted residual module, there is an up-dimensional convolution, a depthwise convolution and a down-dimensional convolution. After each convolution, a batch normalization operation is performed, and finally, the dimensionality is reduced by another convolution. In the second part, the Patch Embedding layer adds positional encoding to the input, which is then fed into eight adjacent Encoder modules for deep feature extraction; Each Encoder module contains two layer normalization operations, and after the two layer normalization operations, a multi-head attention layer and a multi-layer perceptron are connected to extract feature embeddings. Finally, a fully connected layer is used to classify and identify individual radiation sources.

Images