Intelligent identification method for individual radar radiation sources based on multi-branch feature fusion

Abstract

This invention discloses an intelligent identification method for individual radar radiation sources based on multi-branch feature fusion, applied to the field of radar signal processing. Addressing the problem of low accuracy in identifying individual radiation sources due to the difficulty in extracting subtle individual differences in existing technologies, this invention achieves effective extraction of features representing individual differences by cumulatively superimposing multiple radar pulses. Based on the influence of subtle differences in radar transmitters on radar signal envelope fluctuations, carrier frequency deviations, and phase noise, the instantaneous envelope, instantaneous frequency, and instantaneous phase of the signal are extracted to reflect subtle individual differences. Based on the basic architecture of a residual convolutional neural network and combined with the distribution characteristics of individual differences in radiation sources, a feature extraction network for individual radar radiation sources is constructed. Simultaneously, an attention mechanism is introduced to fully consider the manifestation and proportion of subtle individual differences in radiation sources across different dimensions. Features are fused according to the importance of information from different channels to achieve accurate identification of individual radar radiation sources.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing, and specifically relates to a radar radiation source individual identification technology. Background Technology

[0002] Individual radar radiation source identification is a crucial technology in the field of electronic warfare. This process specifically refers to the use of inherent and stable individual characteristics of radar equipment during production and use, i.e., fingerprint features, to accurately identify and distinguish different individuals from signals received by radar reconnaissance receivers from multiple different radiation sources. It is the premise and foundation for achieving precise strikes against enemy targets and effectively avoiding potential risks in electronic warfare.

[0003] Early radar source identification techniques characterized the state of radar pulses by measuring five basic parameters: carrier frequency (RF), time of arrival (TOA), direction of arrival (DOA), pulse amplitude (PA), and pulse width (PW). Traditional template matching and shallow machine learning were used for source classification. However, with continuous optimization of radar equipment design and manufacturing processes, and the widespread use of more complex and advanced equipment in modern electronic warfare, the basic pulse description parameters have become insufficient to represent subtle individual differences. Traditional classification methods are also struggling to cope with the increasingly complex electronic warfare environment. Therefore, methods that mine parameter features carrying more individual information and employ deep neural networks for feature extraction and classification have been proposed.

[0004] The paper "J. Pan, L. Guo, Q. Chen, S. Zhang and J. Xiong, Specific Radar Emitter Identification Using 1D-CBAM-ResNet, 2022 14th International Conference on Wireless Communications and Signal Processing (WCSP), Nanjing, China, 2022, pp. 483-488" proposes a neural network architecture based on a one-dimensional convolutional residual neural network and a convolutional block attention module (1D-CBAM-ResNet). This architecture aggregates channel and spatial information to accurately capture fingerprint features in signal time series. The method achieves an overall identification accuracy of 93.03% on 10 transmitters of the same type. However, this method directly uses the signal time series and fails to effectively filter out a large amount of intentional modulation information and interference contained in the signal. The fingerprint features are still obscured within the principal components, limiting the upper limit of this identification method.

[0005] The paper "J. Su, H. Liu and L. Yang, Specific Emitter Identification Based on CNN via Variational Mode Decomposition and Bimodal Feature Fusion," 2023 IEEE 3rd International Conference on Power, Electronics and Computer Applications (ICPECA), Shenyang, China, 2023, pp. 539-543" proposes a radiation source individual identification scheme based on Variational Mode Decomposition (VMD) and bimodal feature fusion using a convolutional neural network (CNN). This method first extracts four spurious components from the LoRa signal using VMD, which contains unintentional modulation features of the original signal; then, it extracts time-frequency features (VMD-S) and temporal features (VMD-T) from the four spurious components. Therefore, a CNN-based dual-channel feature extraction network is constructed, which incorporates an attention module and a bimodal feature fusion strategy to achieve the fusion of VMD-S and VMD-T. In the low signal-to-noise ratio (SNR) region, this method significantly improves performance compared to networks without feature fusion. However, since this method relies on the variational mode decomposition results of the signal, it requires prior empirical preset of the number of modes K, and there is signal spectrum aliasing, resulting in poor generalization ability. Summary of the Invention

[0006] To address the aforementioned technical issues, this invention proposes an intelligent identification method for individual radar radiation sources based on multi-branch feature fusion. By highlighting subtle individual features through preprocessing, employing multi-branch feature fusion, and introducing a deep learning framework, the method effectively achieves accurate identification of individual radiation sources.

[0007] The technical solution adopted in this invention is: an intelligent identification method for individual radar radiation sources based on multi-branch feature fusion, comprising:

[0008] S1. Detect the raw radar signal of each individual radiation source;

[0009] S2. Preprocess the original radar signals of each radiation source individual detected in step S1. Specifically, align the original radar signals of each radiation source individual based on the signal autocorrelation function for time synchronization, and then superimpose and calculate the average value.

[0010] S3. Extract the envelope, instantaneous frequency, and instantaneous phase features to obtain the dataset; divide the dataset into a training set and a validation set;

[0011] S4. Construct a residual convolutional neural network;

[0012] S5. Train the residual convolutional neural network constructed in step S4 based on the training set of step S3; specifically, the training dataset processed in step S3 is used as input data and fed into the AFF-SE-ResNet network built in step S4. The input data is propagated forward in the network, and the network parameters are optimized in a specified number of iterations to obtain the optimal feature weights.

[0013] S6. Based on the deep feature weights extracted from the input signal data in step S5, the AFF-SE-ResNet network trained in step S5 is used to complete the individual classification and identification of the radiation source. Specifically, the optimal feature weights from step S5 are loaded, the verification data processed in step S3 is used as input data and passed into the network trained in step S5 for forward propagation, the AFF feature fusion module fuses different forms of input features, and the predicted label is output through the fully connected layer to realize the individual identification of the radar radiation source.

[0014] The beneficial effects of this invention are as follows: The method of this invention utilizes intercepted reconnaissance parameters and radar signals, based on the generation mechanism of individual fingerprint features of radiation sources, to extract signal features that effectively characterize individual differences. An accumulation algorithm is used to enhance subtle individual features. Based on the basic architecture of a deep residual convolutional neural network, and combined with the distribution characteristics of individual differences in radar radiation sources, a network for extracting and recognizing subtle differences in individual radar radiation source features is constructed. This fully mines the deep individual features of radar radiation source signals, achieving accurate identification of individual radar radiation sources. The method of this invention has the advantages of accuracy, efficiency, and strong generalization ability. Attached Figure Description

[0015] Figure 1 A flowchart of a radar radiation source individual identification algorithm provided for an example of the present invention;

[0016] Figure 2 A schematic diagram illustrating the effect of the autocorrelation function synchronization signal provided in this invention example;

[0017] Figure 3 A comparative schematic diagram of cumulative averaging processing provided for an example of the present invention;

[0018] Figure 4 A flowchart of the radar radiation source individual feature extraction and identification network provided for an example of the present invention;

[0019] Figure 5 This is a block diagram of the basic components of the network architecture used in this invention example;

[0020] Among them, (A) and (B) are two basic residual units; (C) is the extrusion excitation module;

[0021] Figure 6This is a schematic diagram of the attention-based feature fusion structure used in an example of the present invention;

[0022] Figure 7 The ablation experiment results of the radar radiation source individual identification algorithm are provided as an example of the present invention.

[0023] Figure 8 This is a schematic diagram of the individual radar radiation source identification results provided as an example of the present invention. Detailed Implementation

[0024] To facilitate understanding of the technical content of this invention by those skilled in the art, the following description, in conjunction with the accompanying drawings, further illustrates the invention.

[0025] like Figure 1 The diagram shown is a flowchart of the present invention. The technical solution of the present invention is: an intelligent identification algorithm for individual radar radiation sources, comprising:

[0026] S1. Detect raw radar signals. Raw radar signals are time-domain signals, usually in the form of multi-pulse signals.

[0027] The raw radar signal detected in step S1 is usually composed of multiple pulses. A single pulse can be obtained for subsequent identification by setting a threshold and windowing. Under these conditions, the occurrence times of the single pulse signals are often inconsistent, requiring time alignment for subsequent processing. The purpose of preprocessing is to align and accumulate the time series of multiple pulses from different individuals, thereby amplifying the inherent and stable individual characteristics.

[0028] S2. Synchronizing signal pulses based on the Cross-Correlation function. In statistics, cross-correlation is used to represent the covariance between two random vectors X and Y.

[0029] cov(X,Y)=E((XE(X)(YE(Y))=E(X·Y)-E(X)E(Y) (1)

[0030]

[0031] Where cov(X,Y) is the covariance of X and Y, var(·) represents the variance of the variables, and η represents the correlation between the two. Covariance describes the degree of similarity or difference in the changes between two variables.

[0032] In this example, a cross-correlation function is used to represent the similarity between two signals f[i] and g[j], where f[i] and g[j] are the two signals to be synchronized, and i and j are the number of time sampling points. Cross-correlation is used to find the offset that makes the two signals most similar, thus achieving signal synchronization. For discrete signals f[i] and g[j], the cross-correlation function... Defined as:

[0033]

[0034] Here, "" indicates a conjugate operation on the signal, f[t] represents the result of the conjugate operation on signal f[i] at parameter t, and m represents an offset of m units on signal g[j]. Based on the cross-correlation function of the two signals, the offset that maximizes their time correlation is found to adjust the signals and achieve signal synchronization. Specifically, the following steps are included:

[0035] S21. The first pulse of the radiation source individual is used as the reference signal, and the other pulses of the radiation source individual are used as synchronization signals.

[0036] S22. Calculate the cross-correlation function between the signal to be synchronized and the reference signal according to equation (3);

[0037] S23. Find the offset that maximizes the cross-correlation function and adjust the signal to be synchronized.

[0038] S24. Repeat S22 to S23 to complete the synchronization of all pulses within the individual radiation source.

[0039] Figure 2 The diagram illustrates the changes before and after the synchronization of the two signal pulses.

[0040] After all individuals have completed time synchronization, the average value is accumulated to reduce the impact of noise and highlight subtle individual differences. Figure 3 This study demonstrates a comparison of the differences between two volume signals from the perspective of the rising edge of the temporal envelope, before and after multiple superpositions and averaging. The example shown has a signal-to-noise ratio of 14 dB and 10 superpositions. By superposition, the proportion of information contained in random noise is significantly reduced, and the significance of the differences in the rising edge changes of the individual envelopes is also improved, which is helpful for further feature extraction.

[0041] S3. Extract envelope, instantaneous frequency, and instantaneous phase features. Specifically: First, based on the Hilbert transform, convert the individual radar signals synchronized in S2 into analytic signals, including amplitude, phase, and frequency information; extract the combination of peaks and valleys in the amplitude spectrum to obtain the envelope, extract the phase information to obtain the instantaneous phase, and perform differential calculation on the phase to obtain the instantaneous frequency. 20% of the total samples are used as the training dataset, and 80% as the validation dataset. The specific steps are as follows:

[0042] S31. Perform a Hilbert transform on the signal x(t), that is:

[0043]

[0044] Where τ is the time delay parameter, This represents the result of performing the Hilbert transform on x(t). H[x(t)] represents the convolution operation, and H[x(t)] represents the Hilbert transform of the signal x(t).

[0045] Based on the Hilbert transform, the original signal is converted into an analytic form, which gives:

[0046]

[0047] in,

[0048] S32. The analytical form of a signal includes amplitude and angle information. The signal envelope feature is obtained through |z(t)|, θ(t) is the instantaneous phase feature, and the frequency is obtained by differentiating the phase, i.e.

[0049]

[0050] Where ω(t) is the instantaneous frequency characteristic.

[0051] S33. The acquired envelope, instantaneous frequency, and instantaneous phase features are used as a dataset and split in an 8:2 ratio for training and validation to obtain the input data for the subsequent network.

[0052] S4. The residual convolutional neural network structure constructed in step S4 uses ResNet, a residual convolutional neural network with excellent deep feature extraction performance, as its backbone network. Based on this backbone network, high-dimensional features representing individual differences of radiation sources are extracted from the envelope, instantaneous frequency, and instantaneous phase. A squeeze-and-excitation (SE) module is added to the basic residual unit of ResNet to adaptively adjust the feature weights between different channels, enhancing the ability to represent differences in features. Different forms of individual difference features obtained from the envelope, frequency, and phase are combined and summarized using an attentional feature fusion (AFF) module. Finally, the predicted label results are output through a fully connected layer. Figure 4 The diagram shows the specific network structure of an embodiment of the present invention, which is mainly divided into three stages: preprocessing part, network body, feature fusion and fully connected layer.

[0053] S41. The network has a three-branch structure, taking the envelope, instantaneous frequency, and instantaneous phase datasets as inputs for feature extraction and prediction training. First, a preprocessing module is built to preprocess the input data, including an adaptive average pooling, a convolutional layer with a kernel size of 3×3 and 3 input channels, and a normalization layer. The purpose is to transform the input into a form suitable for the designed AFF-SE-ResNet processing.

[0054] S42. Constructing Basic Residual Units: The basic residual unit ResBlock is a major component of the network backbone structure of residual convolutional networks, such as... Figure 5 As shown, the residual blocks have two design structures: ResBlock(A) and ResBlock(B). The number of residual blocks in the four stages are 3, 4, 3, and 6, respectively. The main structure of the residual block contains three convolutional layers with kernel sizes of 1×1, 3×3, and 1×1, along with their corresponding BN normalization layers and ReLU activation functions. In the residual connections, ResBlock(A) adds a 1×1 convolutional kernel to achieve feature addition across different dimensions, while ResBlock(B) does not require this design.

[0055] S43. Construct a feature enhancement module: Cascade a feature enhancement module after each residual unit to add attention fusion between different channels to the features, thereby enhancing the performance of each feature. For example... Figure 5 As shown, the specific structure of SEBlock is to first pass through a global pooling layer, and then output features through two fully connected layers and their respective activation functions.

[0056] S44. Constructing the Network Backbone Structure: This section describes the core phase of the SE-ResNet network, which consists of modules constructed in S42 and S43. The network core is divided into four phases, each composed of 3, 4, 3, and 6 basic residual units connected in series. Envelope, instantaneous frequency, and instantaneous phase features are extracted through branches within the network.

[0057] S45. Construct the feature fusion module and classification output. The three input branches fuse features through the AFF feature fusion module, and then a fully connected layer containing three linear functions performs softmax classification to output the predicted label. Figure 6 As shown, the AFF module is described as follows: The module is implemented in two parallel paths. One path consists of global pooling, two convolutional layers, and the ReLU activation function. The other path consists of two convolutional layers and the ReLU activation function. The two branches extract channel attention weights at different scales. The former extracts global feature attention through global pooling, while the latter directly extracts local feature channel attention through pointwise convolution. The outputs are then merged using the Sigmoid activation function.

[0058] S46. The sub-modules constructed in steps S42 to S46 are combined into an AFF-SE-ResNet. Specifically, the envelope, instantaneous frequency, and instantaneous phase are output as three branches. First, the output is processed by a preprocessing module to a form suitable for entering the main body of the constructed network. Then, it is processed by a preprocessing module containing... Figure 5 The basic residual structures shown in (A)(B) and as follows Figure 5(C) The network backbone of the excitation squeezing module extracts three types of features, which are then fused by the AFF attention feature fusion module and the predicted label results are output by the fully connected layer through Softmax classification.

[0059] S5. Using the data obtained in step S3 as input to the residual convolutional neural network constructed in S4, the residual convolutional neural network is trained. Specifically, the training dataset obtained after preprocessing the original signal is input into the constructed network model, and forward propagation is performed through the network. In each residual block constituting the residual convolutional network, the input data undergoes operations such as convolution and batch normalization, and is then added to the output through skip connections to form residual learning. The loss function is calculated based on the difference between the network output and the true label value. The gradient of the loss function with respect to the network parameters is calculated using the backpropagation algorithm. These gradients are used to update the network parameters to minimize the loss function. The network parameters are updated based on the gradient using an optimization algorithm (such as stochastic gradient descent). This step gradually adjusts the network weights to more suitable values. By repeatedly executing the aforementioned steps, backpropagation and parameter updates, and multiple traversals of the training dataset, the network performance is continuously optimized. By observing that the loss function no longer decreases until it becomes smooth, the training of the constructed network model is finally completed, and the optimal weights are obtained. Specifically, the following steps are included:

[0060] S51, the forward propagation process involves convolution, pooling, fully connected layers, and output operations.

[0061] If the first The layer is a convolutional layer, with For the first The j-th feature map of the layer, then

[0062]

[0063] Among them, M j Let be the set of feature maps connected to it in the layer. This represents the convolution kernel that connects the i-th input feature map and the j-th output feature map. The symbol "*" represents the bias, f(·) represents the non-linear activation function, and the symbol "*" represents two-dimensional convolution.

[0064] If the first If the layer is a pooling layer, then its output is

[0065]

[0066] Where down(·) represents the downsampling function, These represent multiplicative bias and additive bias, respectively.

[0067] If the first If the layer is a fully connected layer, then its output is

[0068]

[0069] in, Indicates the first Feature map of the layer Indicates the first Layer weight, b (l) This is the bias term for this layer;

[0070] If the first If layer i is the output layer, then the posterior probability that the current sample belongs to class i is:

[0071]

[0072] in, This represents the input to the output layer, and C represents the total number of categories.

[0073] definition For the first Layer error sensitivity term, Let the input of the nonlinear activation function be the first... If the layer is the output layer, then its sensitivity term is:

[0074]

[0075] Among them, symbols This indicates dot product.

[0076] S52. For the typical forward propagation, the residual element is additionally designed with a Shortcut identity transformation connection, namely:

[0077] x l+1 =x l +F(x l W l (12)

[0078] The feature representation of any deep unit L is:

[0079]

[0080] Where F(·) is the residual function, x l and x L This represents the feature representation of the corresponding layer.

[0081] Specifically, when calculating the residual function at different dimensional levels, a linear mapping needs to be performed on x to achieve dimensional matching, i.e.

[0082]

[0083] Where y represents a feature representation with a different dimension than x, and σ represents a linear mapping.

[0084] S53. In the example, cross-entropy loss is used to characterize the similarity between the actual output and the expected output, i.e.

[0085]

[0086] Where L represents the current loss function, and Y and f(x) represent the actual class and the predicted class, respectively.

[0087] S54. In the example, the backpropagation algorithm based on gradient descent updates the network parameters. The layer update transformation calculation process is as follows:

[0088]

[0089] Where η is the learning rate. Indicates and The corresponding position at (u,v) Zhongyu The region where convolution operations are performed.

[0090] S6. Extract deep individual features of radar radiation sources using the network trained in step S5. Specifically, the training dataset processed in step S3 is used as input data and fed into the AFF-SE-ResNet network built in step S4. The input data is propagated forward in the network, and the network parameters are optimized for a specified number of iterations to obtain the optimal feature weights.

[0091] S7. Based on the deep feature weights obtained from steps S5-S6, the AFF-SE-ResNet network constructed in S4 is used to complete the individual classification and identification of radiation sources. Specifically, the feature weights obtained from training the network in steps S5-S6 on the training set are loaded. The envelope, instantaneous frequency, and instantaneous phase features from the validation data processed in step S3 are respectively used as input data and fed into the network trained in step S5 for forward propagation. The AFF feature fusion module fuses the different forms of input features, and the predicted label is output through the fully connected layer to achieve individual identification of radar radiation sources. In this invention example, the AFF calculation method used is as follows:

[0092]

[0093] Where “⊕” represents a weighted average, This represents a convolution operation, where X and Y are different forms of feature inputs to be fused.

[0094] Finally, the present invention also includes:

[0095] (1) Test the recognition performance of the radar radiation source signal under different signal-to-noise ratios on the network trained in step S5. Specifically, input the test samples with signal-to-noise ratios of 5 to 14 dB into the network trained in S5 for forward propagation, obtain the posterior probability of the test sample belonging to each category, compare the size of the posterior probability of each category, and take the category corresponding to the largest one as the final prediction result to verify the recognition performance of the proposed network for the individual radiation source type of the radar signal under different signal-to-noise ratios.

[0096] (2) The invention design of the proposed radar radiation source individual identification algorithm is compared with the recognition performance of the classic residual convolutional network in this field: Specifically, signal data with a signal-to-noise ratio of 4 to 14 dB are used as samples, and the proposed preprocessing method is used to enhance the features of the signal. The preprocessed signal envelope, instantaneous frequency, and instantaneous phase are then input into the designed network structure to obtain the prediction results. After removing different functional modules of the network structure, the test experiments are repeated to realize the comparison and verification of the recognition performance of the proposed individual identification algorithm with other networks, such as... Figure 7 As shown, the effectiveness of the design content of this invention is verified.

[0097] Table 1 shows the signal dataset used for network training and testing in this embodiment, including eight different individual signals of two modulation types and the same signal bandwidth, specifically linear frequency modulation (LFM) and continuous wave (CW) signals. To verify that the extracted features effectively represent the inherent subtle differences in the radiation source, the signal parameters were not differentiated. The bandwidth of the single carrier frequency signal was 480MHz, and the bandwidth of the linear frequency modulation signal ranged from 240MHz to 480MHz, with a pulse width of 0.8µs for both. The convolutional neural network was trained with 6400 training samples and tested with 1600 test samples, with the sample signal-to-noise ratio (SNR) uniformly distributed from 5dB to 14dB. Figure 8 As shown, the overall accuracy rate of the test samples reached 98.6%.

[0098] Table 1. Parameter Correspondence Table for Individual Signal Datasets of Radiation Sources

[0099]

[0100] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. A method for intelligent identification of individual radar radiation sources based on multi-branch feature fusion, characterized in that, include: S1. Detect the raw radar signal of each individual radiation source; S2. Preprocess the original radar signals of each radiation source individual detected in step S1. Specifically, align the original radar signals of each radiation source individual based on the signal autocorrelation function for time synchronization, and then superimpose and calculate the average value. S3. Extract the envelope, instantaneous frequency, and instantaneous phase features to obtain the dataset; divide the dataset into a training set and a validation set; S4. Construct a residual convolutional neural network; S5. Train the residual convolutional neural network constructed in step S4 based on the training set of step S3; specifically, the training dataset processed in step S3 is used as input data and fed into the AFF-SE-ResNet network built in step S4. The input data is propagated forward in the network, and the network parameters are optimized in a specified number of iterations to obtain the optimal feature weights. S6. Based on the deep feature weights extracted from the input signal data in step S5, the AFF-SE-ResNet network trained in step S5 is used to complete the individual classification and identification of the radiation source. Specifically, the optimal feature weights from step S5 are loaded, the verification data processed in step S3 is used as input data and passed into the network trained in step S5 for forward propagation, the AFF feature fusion module fuses different forms of input features, and the predicted label is output through the fully connected layer to realize the individual identification of the radar radiation source.

2. The intelligent identification method for individual radar radiation sources based on multi-branch feature fusion according to claim 1, characterized in that, Step S2 is as follows: S21. The first pulse of the radiation source individual is used as the reference signal, and the other pulses of the radiation source individual are used as synchronization signals. S22. Calculate the cross-correlation function between the signal to be synchronized and the reference signal; S23. Find the offset that maximizes the cross-correlation function and adjust the signal to be synchronized. S24. Repeat S22 to S23 to complete the synchronization of all pulses within the individual radiation source.

3. The intelligent identification method for individual radar radiation sources based on multi-branch feature fusion according to claim 2, characterized in that, Step S3 is as follows: S31. Perform a Hilbert transform on the signal processed in step S2, i.e.: Where x(t) represents the signal after processing in step S2, and τ is the time delay parameter. This represents the result of performing the Hilbert transform on x(t). H[x(t)] represents the convolution operation, and H[x(t)] represents the Hilbert transform of the signal x(t). Based on the Hilbert transform, the original signal is converted into an analytic form, which gives: in, S32. The analytical form of a signal contains amplitude and angle information. The signal envelope feature is obtained through |z(t)|, θ(t) is the instantaneous phase feature, and the frequency is obtained by differentiating the phase. Where ω(t) is the instantaneous frequency characteristic; S33. The acquired envelope, instantaneous frequency, and instantaneous phase features are used as a dataset, and the dataset is divided into a training set and a validation set according to the proportion.

4. The intelligent identification method for individual radar radiation sources based on multi-branch feature fusion according to claim 3, characterized in that, The residual convolutional neural network includes: a preprocessing module, a network backbone containing a basic residual structure and an excitation squeezing module, an AFF attention feature fusion module, and a fully connected layer; the envelope, instantaneous frequency, and instantaneous phase are used as three-branch inputs. After being processed by the preprocessing module, the three types of features are extracted by the network backbone containing the basic residual structure. Then, after being fused by the AFF attention feature fusion module, the fully connected layer outputs the recognition result through Softmax classification.

5. The intelligent identification method for individual radar radiation sources based on multi-branch feature fusion according to claim 4, characterized in that, The network backbone, which includes the basic residual structure and the excitation squeezing module, is specifically a four-stage residual block structure: each of the four stages of residual blocks consists of 3, 4, 3, and 6 basic residual units connected in series; it also includes a feature enhancement module cascaded after each residual unit.

6. The intelligent identification method for individual radar radiation sources based on multi-branch feature fusion according to claim 5, characterized in that, The feature enhancement module includes: a global pooling layer and two fully connected layers.

Images