Radiation source identification method based on multi-domain feature fusion and self-attention

Abstract

The invention relates to the technical field of image recognition processing, and discloses a radiation source recognition method based on multi-domain feature fusion and self-attention, comprising the following steps: receiving an original radiation source signal which is a one-dimensional complex signal containing an in-phase component and an orthogonal component; respectively inputting the original signal into a time domain branch and a frequency domain branch for processing, performing feature extraction on the time domain branch by adopting a long-short-term memory network in combination with a self-attention mechanism, converting the frequency domain branch into a time-frequency graph through short-time Fourier transform, and performing feature extraction by using a convolutional neural network; fusing the features extracted from the time domain branch and the frequency domain branch to obtain a comprehensive feature vector; and inputting the comprehensive feature vector into a classifier for category judgment. According to the invention, multi-domain feature fusion is carried out on the time domain features and the frequency domain features, so that more comprehensive and accurate capture of the radiation source signals is realized, and the identification accuracy is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to a radiation source recognition method based on multi-domain feature fusion and self-attention. Background Art

[0002] In the field of communication radiation source identification, convolutional neural networks (CNNs), as a typical representative of deep learning, have been widely used. Especially in image recognition, CNN can combine the wavelet transform, high-order spectrum features and IQ feature maps of the signal for image recognition, thereby effectively completing the radiation source identification task. Although these methods have achieved remarkable recognition results, they mainly rely on large-scale annotated data sets. In other words, the lack of annotated data sets directly limits the training effect of the model, especially in the actual environment where data is scarce, the generalization ability of such models is often insufficient.

[0003] In addition, radiation source identification based on convolutional neural networks still has its limitations. Its feature extraction methods often focus on a specific feature domain, whether it is the time domain or the frequency domain. This single-domain feature extraction method makes it impossible for the system to capture the multi-dimensional information of the signal in all directions when processing the signal, resulting in limited recognition capabilities for complex signals. In fact, radiation source signals usually have multi-dimensional characteristics, and feature extraction in a single domain cannot fully reflect the complete characteristics of the signal, thus affecting the accuracy of recognition.

[0004] The LSTM model provides a new idea, which directly feeds the original I / Q signal into the neural network for processing. Although this method avoids the complex steps of signal preprocessing, it is too dependent on the network design itself. Under different types of original signals, different network architectures usually need to be designed specifically, which places high demands on the flexibility and robustness of the model. More importantly, LSTM often faces the problem of gradient vanishing or gradient exploding when processing very long sequences, which makes its performance unstable when processing long time series signals and difficult to capture long-term dependent features.

[0005] In order to solve the shortcomings in time domain feature extraction, many methods try to extract the frequency domain characteristics of the radiation source waveform through time-frequency distribution (such as STFT, WVD and CWD). Although these methods have certain application effects in some scenarios, they still have a problem - the feature extraction is too single, limited to the characteristics of a specific domain, and cannot fully combine the multi-dimensional information of the time domain and frequency domain. The characteristics of the radiation source often rely not only on the single information of the time domain or the frequency domain, but on the combination of the two. Therefore, the existing time-frequency feature extraction methods still seem to be unable to capture the signal characteristics in an all-round way. Summary of the invention

[0006] To make up for the above deficiencies, the present invention provides a radiation source recognition method based on multi-domain feature fusion and self-attention, aiming to improve the problem that the existing single feature extraction method cannot fully extract sufficient information.

[0007] In the first aspect, the present invention provides the following technical solution. A radiation source recognition method based on multi-domain feature fusion and self-attention includes the following steps: Receive the original radiation source signal, which is a one-dimensional complex signal containing in-phase and quadrature components; Input the original signal into the time domain branch and the frequency domain branch for processing respectively. The time domain branch uses a long short-term memory network combined with a self-attention mechanism for feature extraction, and the frequency domain branch is transformed into a time-frequency diagram through a short-time Fourier transform and uses a convolutional neural network for feature extraction; Fuse the features extracted by the time domain branch and the frequency domain branch to obtain a comprehensive feature vector; Input the comprehensive feature vector into a classifier for class judgment, use the Softmax activation function to output the probabilities of each class, and select the class with the highest probability as the recognition result of the radiation source.

[0008] Preferably, the long short-term memory network in the time domain branch is used to capture the time dependence relationship of the signal, and the self-attention mechanism assigns different weights to different time steps of the time domain signal.

[0009] Preferably, the short-time Fourier transform in the frequency domain branch converts the signal from the time domain to the frequency domain to obtain a two-dimensional time-frequency diagram, and the convolutional neural network is used to extract frequency domain features from the time-frequency diagram to identify key patterns in the spectrum.

[0010] Preferably, the features extracted by the time domain branch and the frequency domain branch are fused through a concatenation operation, and the obtained fused feature vector contains the joint information of the time domain and the frequency domain.

[0011] Preferably, the classifier is a fully connected layer, and the Softmax activation function is used to calculate the recognition probability of each class, and the radiation source is recognized according to the one with the largest probability value.

[0012] Preferably, during the training process, the cross-entropy loss function and the Adam optimizer are used for optimization and the network parameters are adjusted through the backpropagation algorithm.

[0013] In the second aspect, the present invention provides the following technical solution. A radiation source recognition system based on multi-domain feature fusion and self-attention, the system includes: A signal input module for receiving the original radiation source signal; A time domain branch module for extracting features from the time domain signal by combining a long short-term memory network and a self-attention mechanism; Frequency domain branch module, which uses short-time Fourier transform and convolutional neural network to extract features from frequency domain signals; Feature fusion module, which is used to fuse the features of the time domain branch and the frequency domain branch to form a comprehensive feature vector; Classification module, which is used to classify the comprehensive feature vector through a fully connected layer and a Softmax activation function, and output the recognition result of the radiation source.

[0014] Preferably, the time domain branch module and the frequency domain branch module respectively process time domain and frequency domain signals, and perform joint processing through the feature fusion module, and finally predict the category of the radiation source through the classification module.

[0015] In a third aspect, the present invention provides the following technical solution. A computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the above-mentioned radiation source recognition method based on multi-domain feature fusion and self-attention.

[0016] In a fourth aspect, the present invention provides the following technical solution. A readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the above-mentioned radiation source recognition method based on multi-domain feature fusion and self-attention.

[0017] The present invention has the following beneficial effects: 1. In the present invention, by performing multi-domain feature fusion on time domain features and frequency domain features, more comprehensive and accurate capture of radiation source signals is achieved, and the recognition accuracy is significantly improved. By fully utilizing time domain and frequency domain information, the problem of incomplete identification caused by single feature extraction in traditional methods is overcome.

[0018] 2. In the present invention, by combining the long short-term memory network and the self-attention mechanism, deep modeling and dynamic attention to the timing features of the radiation source are achieved, and accurate capture of long-term dependencies and key time steps in the sequence is obtained.

[0019] 3. In the present invention, by adopting a training mode of multi-domain feature fusion, sufficient learning of radiation source signals is achieved, and the effect of maintaining a high recognition accuracy even in the case of scarce labeled data is obtained. This innovation greatly reduces the dependence on a large amount of labeled data and improves the application adaptability of the model in an actual non-cooperative environment. Description of the Drawings

[0020] Figure 1 It is a method flow chart of a radiation source recognition method based on multi-domain feature fusion and self-attention proposed by the present invention; Figure 2This is the system architecture diagram of a radiation source recognition system based on multi-domain feature fusion and self-attention proposed by the present invention. Detailed implementation mode

[0021] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0022] Embodiment 1: Refer to Figure 1 , in the first embodiment of the present invention, the present invention provides a radiation source recognition method based on multi-domain feature fusion and self-attention, including the following steps: Receive the original radiation source signal, and the signal is a one-dimensional complex signal including in-phase component and quadrature component; Specifically, first receive the original radiation source signal. The signal is in one-dimensional complex form and includes in-phase component and quadrature component. Represent this signal as: ; Among them, is the received complex signal, indicating the signal amplitude and phase information at time moment. This signal consists of two components, namely: is the in-phase component, usually used to represent the amplitude change of the signal. This is the real part of the complex signal, reflecting the change of signal intensity over time. In a communication system, is directly related to the baseband information of the transmitted signal. is the quadrature component, which is the imaginary part of the complex signal and is usually related to the phase change of the signal. It represents the component perpendicular to , and is often used for phase information in the modulated signal. is the imaginary unit, with the property of . It is responsible for separating the in-phase component and the quadrature component in the expression of the complex signal, enabling the signal to effectively represent the amplitude and phase information.

[0023] To achieve efficient signal processing and frequency estimation, a dual-branch network structure is adopted. The dual-branch network consists of two branches, which are respectively used to process the in-phase component (I branch) and the quadrature component (Q branch). Each branch independently processes different components of the signal to ensure that the amplitude and phase information of the signal can be accurately captured and analyzed.

[0024] Including: in-phase branch (I branch): This branch mainly processes the real part , which is used to extract the amplitude information of the signal. By processing the in-phase component, the intensity change of the signal can be identified, and further time-frequency analysis of the signal can be carried out to obtain the instantaneous amplitude.

[0025] Orthogonal branch (Q branch): This branch processes the imaginary part of the signal , which is mainly used to capture the phase information of the signal. By independently analyzing the orthogonal components, the phase change of the signal can be accurately obtained, and then phase recovery and frequency estimation can be carried out.

[0026] The acquisition process of the complex signal utilizes the sampling technology in digital signal processing. This method can simultaneously obtain the amplitude and phase of the signal, avoiding signal loss that may occur in traditional single-channel sampling. Such a signal not only contains amplitude information but also retains phase information, so it can comprehensively reflect the characteristics of the radiation source. For example, instantaneous frequency analysis and time-domain waveform reconstruction of the signal can be performed using this complex signal.

[0027] Furthermore, in order to completely describe the time variation of the signal, the complex signal can be analyzed using the autocorrelation function, expressed as: ; In this formula: is the autocorrelation function of the signal, which describes the similarity of the signal at different time delays ; is the conjugate complex number of, that is , representing the complex conjugate of the signal; represents the expectation operation, that is, the long-term average of the signal.

[0028] By calculating the autocorrelation function, we can extract features such as the periodicity and similarity of the signal and use them for signal characteristic analysis. Since the complex signal contains in-phase and orthogonal components, its autocorrelation function can more accurately capture the frequency and phase information in the signal, thus providing a data basis for subsequent frequency estimation and signal detection.

[0029] The original signal is respectively input into the time-domain branch and the frequency-domain branch for processing. The time-domain branch uses a long short-term memory network combined with a self-attention mechanism for feature extraction, and the frequency-domain branch is transformed into a time-frequency diagram through a short-time Fourier transform and uses a convolutional neural network for feature extraction; the long short-term memory network in the time-domain branch is used to capture the time-dependent relationship of the signal, and the self-attention mechanism assigns different weights to different time steps of the time-domain signal.

[0030] Specifically, in the specific implementation manner of the present invention, the original signal is first divided into a time-domain branch and a frequency-domain branch for processing. Each branch adopts different network structures and processing methods to extract the key information of the signal.

[0031] First, enter the time-domain branch processing. The main task of the time-domain branch is to capture the time-dependent relationship of the signal. For this purpose, a long short-term memory (LSTM) network combined with a self-attention mechanism is adopted. LSTM is a special recurrent neural network that can effectively capture long-term dependencies and has significant advantages in processing sequence data. In the present invention, LSTM is used to analyze the law of change of the original signal over time, especially to identify complex patterns that change over time in the signal.

[0032] At the same time, the self-attention mechanism is introduced into the time-domain branch, mainly to enhance the feature extraction ability of LSTM. The self-attention mechanism calculates the correlation between different time steps in the signal sequence and automatically assigns weights to each time step. In actual operation, different time steps of the time-domain signal will be given different importance, and more important time steps will obtain higher weights, thus effectively highlighting the key time features in the signal.

[0033] The core calculation formula of its self-attention mechanism is as follows: ; where: Q is the query matrix, representing the query information of the time step; K is the key matrix, representing the key information of the time step; V is the value matrix, representing the value information of the time step; is the dimension of the key, used for scaling the calculation to prevent the inner product value from being too large.

[0034] Through the self-attention mechanism, the time-domain branch can process time information more flexibly and automatically focus on important moments in the signal, thus enhancing the ability to model the long-term dependence of the signal.

[0035] Then enter the frequency-domain branch processing. The frequency-domain branch converts the signal from the time domain to the frequency domain through the short-time Fourier transform (STFT) to obtain a time-frequency diagram. The short-time Fourier transform can decompose the signal into frequency components within several short-time windows and is usually used to analyze the frequency characteristics of non-stationary signals. After the signal passes through the short-time Fourier transform, we obtain a two-dimensional time-frequency diagram, where the horizontal axis represents time, the vertical axis represents frequency, and each point in the diagram represents the signal intensity at that time and frequency.

[0036] After obtaining the time-frequency diagram, the frequency-domain branch will use a convolutional neural network (CNN) for feature extraction. The convolutional neural network automatically extracts local features in the image (in this scheme, the time-frequency diagram) through convolutional layers. Through multiple convolutional and pooling operations, CNN can gradually extract high-level features in the time-frequency diagram, such as spectral morphology, signal periodicity, etc. The formula for the convolutional operation is as follows: ; where: represents the input time-frequency diagram signal; is the convolution kernel, i.e., the weight matrix for feature extraction; is the output of the convolution operation, representing the extracted feature map.

[0037] Through CNN, the frequency-domain branch can capture the detailed features of the signal in the frequency domain, which are crucial for tasks such as signal modulation recognition and frequency offset analysis.

[0038] Fuse the features extracted by the time-domain branch and the frequency-domain branch to obtain a comprehensive feature vector; the short-time Fourier transform in the frequency-domain branch converts the signal from the time domain to the frequency domain to obtain a two-dimensional time-frequency diagram, and the convolutional neural network is used to extract frequency-domain features from the time-frequency diagram to identify key patterns in the spectrum.

[0039] Combining the time-domain and frequency-domain branches enables the system to comprehensively analyze the signal from two different perspectives simultaneously. The time-domain branch captures the temporal correlation of the signal through LSTM and self-attention mechanisms, effectively mining the important features of the signal in the time domain; the frequency-domain branch extracts the frequency-domain features of the signal through STFT and CNN, further enhancing the perception ability of frequency and modulation information.

[0040] The advantage of this dual-processing architecture is that it can utilize both time and frequency information simultaneously, thereby improving the accuracy and robustness of signal feature extraction. Especially in complex signal environments, the time-domain branch can capture the temporal dependence of the signal, and the frequency-domain branch can reveal the frequency patterns in the signal. After combining the two, the system can perform deep learning on the signal in multiple dimensions, thereby enhancing the performance of signal detection and analysis.

[0041] Specifically, the time-domain branch is particularly effective for long-term changes and short-term dependencies of low-frequency signals, while the frequency-domain branch can accurately identify the spectral features of the signal. Through this fusion, the system not only has the advantage of high time resolution but also can perform more accurate analysis of frequency offset and signal perturbation, ensuring the efficient operation of the system in various complex environments.

[0042] This method provides stronger robustness for complex signal analysis, especially suitable for situations with multipath propagation, frequency drift, or high background noise. By combining LSTM and self-attention mechanisms, the time-domain branch can better capture the potential patterns in the signal, while by combining STFT and CNN, the frequency-domain branch can accurately analyze the spectral features of the signal.

[0043] The features extracted by the time-domain branch and the frequency-domain branch are fused through a concatenation operation, and the resulting fused feature vector contains the joint information of the time domain and the frequency domain.

[0044] Specifically, during the implementation of the present invention, the features extracted by the time-domain branch and the frequency-domain branch are fused through a concatenation operation. The core purpose of this process is to combine the different features obtained from the time-domain and frequency-domain branches to form a comprehensive feature vector that can comprehensively reflect the time-domain and frequency-domain information of the signal.

[0045] First, the time-domain branch extracts the time-dependent features of the signal through LSTM and the self-attention mechanism, which can capture the changing patterns of the signal over time; the frequency-domain branch extracts the frequency features of the signal through the short-time Fourier transform and the convolutional neural network, which can reveal the changes of the signal in frequency. Fusing the features extracted by these two branches helps to comprehensively utilize the time and frequency-domain information, thereby improving the performance of the system.

[0046] The feature fusion operation is carried out through a concatenation operation. Specifically, the feature vector T obtained from the time-domain branch and the feature vector F obtained from the frequency-domain branch are concatenated according to the dimension to obtain a new feature vector X, which is expressed as: ; where: T is the feature vector extracted by the time-domain branch, which contains the information of the signal in the time dimension; F is the feature vector extracted by the frequency-domain branch, which contains the information of the signal in the frequency dimension; | represents the concatenation operation of the feature vectors.

[0047] Through concatenation, all the information of the time-domain features and the frequency-domain features is integrated into one vector, thus ensuring the joint representation of the time and frequency-domain information. This fused feature vector can provide a more comprehensive and accurate input for subsequent signal analysis and classification tasks.

[0048] By concatenating the time-domain and frequency-domain features, the new feature vector can comprehensively capture the information of different dimensions of the signal. The time-domain branch focuses on the changes of the signal over time, while the frequency-domain branch provides the frequency features of the signal. The combination of the two can provide a more comprehensive understanding of the complexity of the signal. And because the time-domain and frequency-domain information is complementary, the concatenated feature vector can provide more valuable features for subsequent model training, thereby improving the accuracy of classification, detection and other tasks. Especially when dealing with complex signals, the fused features help the model to learn and judge from multiple perspectives, reducing the errors caused by insufficient single information dimension. In addition, different types of signals may exhibit different features in the time domain or the frequency domain. By simultaneously using the time-domain and frequency-domain features, the system can better adapt to various signal environments. For example, for signals with large frequency drifts, the role of the frequency-domain features is particularly important; while for signals with obvious temporal changes, the time-domain features are more critical. The fusion of the two makes the system more adaptable to various complex signal environments.

[0049] Furthermore, during the feature concatenation process, the feature vectors in the time domain and frequency domain are not simply concatenated. Their dimensions and information content may be different. Therefore, the feature vectors obtained after concatenation will fuse information from more dimensions. Suppose the dimension of the time-domain feature vector is , and the dimension of the frequency-domain feature vector is . Then the dimension of the concatenated feature vector is:

[0050] where: is the dimension of the time-domain feature vector ; is the dimension of the frequency-domain feature vector ; is the dimension of the feature vector obtained after concatenation.

[0051] This fused feature vector will be used as the input to subsequent models (such as fully connected layers, classifiers, or regressors) to provide feature supply for further processing of the task.

[0052] The comprehensive feature vector will be input into the classifier for class judgment. The Softmax activation function is used to output the probabilities of each class, and the class with the highest probability value is selected as the recognition result of the radiation source.

[0053] The classifier is a fully connected layer, and the Softmax activation function is used to calculate the recognition probability of each class. The radiation source is recognized according to the maximum probability value. During the training process, the cross-entropy loss function and the Adam optimizer are used for optimization, and the network parameters are adjusted through the backpropagation algorithm.

[0054] Specifically, in this embodiment, the fused feature vector is input into the classifier module for the final determination of the radiation source category. The classifier adopts a fully connected neural network structure to process the input comprehensive features and generate discrimination values for each target class. This structure has the ability of non-linear mapping and can adapt to the class boundaries in the high-dimensional complex feature space. The Softmax activation function is introduced at the end of the network to convert the output results of each class into probability values. This method can ensure that the sum of the outputs of all classes is 1, so that each dimension of the output can be interpreted as the prediction probability of its corresponding class.

[0055] The above function can perform a normalization operation on all class outputs to maintain the consistency of the probability distribution. In practical applications, in the finally output probability vector, the class corresponding to the maximum probability value is taken as the prediction result, that is: ; where: is the finally recognized radiation source category; is the Softmax output probability for the i-th class.

[0056] The network training process adopts a supervised learning method, and the goal is to minimize the difference between the predicted class and the true class. The cross-entropy loss is selected as the loss function, which performs stably in multi-class classification tasks and can punish the situation where the predicted probability deviates from the true label. The form of the cross-entropy loss function is as follows: ; where: L represents the overall loss; is the i-th one-hot encoded vector of the true label; is the predicted probability of the i-th class of the Softmax output.

[0057] When the model output deviates from the true class, the corresponding log(Pi) term approaches negative infinity, the loss function will increase, and then guide the adjustment of the direction of the network parameters. To optimize this loss function, the Adam optimizer is used for parameter iterative update.

[0058] Adam combines the momentum method and the adaptive learning rate method, which can accelerate convergence and avoid falling into local optima at the same time. The formula for its parameter update is as follows: ; where: represents the network parameters at the current moment; is the learning rate; and are the bias correction values of the first-order moment estimate and the second-order moment estimate respectively; is a small constant to prevent division by zero; represents the current iteration round.

[0059] During the training process, through the backpropagation algorithm, the gradients of the loss function with respect to the parameters of each layer of the model are calculated, and the parameters are gradually adjusted according to the above update rules, so that the output of the network is getting closer and closer to the true label. Backpropagation involves the chain rule, and the error signal is propagated forward from the output layer in turn, and the gradient information of each layer is accumulated and calculated.

[0060] This training process not only has the characteristics of good stability and high training efficiency, but also has a certain ability to suppress gradient oscillation, making the final model have good generalization ability. Especially in the identification of multi-class radiation sources, this Softmax output framework based on fused features can effectively extract differential features and make accurate judgments. Through continuous iterative optimization, the model can gradually converge to the global optimum, improve the recognition accuracy, and enhance the adaptability to complex signal environments.

[0061] Example 2: Refer to Figure 2, in the second embodiment of the present invention, the present invention provides a radiation source recognition system based on multi-domain feature fusion and self-attention. The system includes: A signal input module for receiving the original radiation source signal; A time-domain branch module that combines a long short-term memory network and a self-attention mechanism to extract features from the time-domain signal; A frequency-domain branch module that uses the short-time Fourier transform and a convolutional neural network to extract features from the frequency-domain signal; A feature fusion module for fusing the features of the time-domain branch and the frequency-domain branch to form a comprehensive feature vector; A classification module for classifying the comprehensive feature vector through a fully connected layer and a Softmax activation function, and outputting the recognition result of the radiation source.

[0062] Specifically, the signal input module receives the original radiation source signal. Its main task is to preprocess the signal, including operations such as denoising and normalization, to ensure that the signal is suitable for the subsequent time-domain and frequency-domain feature extraction modules. This module transmits the original signal to the time-domain branch and the frequency-domain branch respectively for subsequent feature extraction.

[0063] The time-domain branch module receives the time-domain signal from the signal input module and processes the time-domain signal through a long short-term memory (LSTM) network to capture time-dependent features. The LSTM network can effectively extract the temporal variation law of the signal. Immediately afterwards, the self-attention mechanism is applied to the features output by the LSTM. By calculating the correlation between different time steps, different weights are assigned to the time steps of the signal, thereby enhancing the attention to important moments. Finally, the time-domain branch outputs a feature vector representing the key features of the signal in the time domain.

[0064] The frequency-domain branch module receives the signal from the signal input module and first converts the signal from the time domain to the frequency domain through the short-time Fourier transform (STFT) to obtain a time-frequency diagram. Then, a convolutional neural network (CNN) is used to extract features from the time-frequency diagram. The convolution operation can automatically identify local features in the time-frequency diagram, such as spectral morphology and periodicity. The frequency-domain branch module outputs a frequency-domain feature vector, which contains important information of the signal in the frequency domain, especially the frequency pattern and variation law of the signal.

[0065] The feature fusion module receives the feature vectors from the time-domain branch and the frequency-domain branch, and combines the two into a comprehensive feature vector through a concatenation operation. This fusion operation can combine the information in the time domain and the frequency domain, providing more comprehensive and accurate features for the subsequent classification module. The concatenated comprehensive feature vector combines all the key information of the signal in the time and frequency dimensions, ensuring that the subsequent classification operation can make judgments from more comprehensive features.

[0066] The classification module receives the comprehensive feature vector from the feature fusion module and further processes and classifies the input features through a fully-connected layer. The output of the fully-connected layer is processed by the Softmax activation function and transformed into a probability distribution, representing the predicted probabilities of each class. Finally, the class with the highest probability is selected as the recognition result of the radiation source, thus completing the classification task of the radiation source. This module provides the final recognition decision by outputting the signal class.

[0067] Embodiment III: In the third embodiment of the present invention, based on the same inventive concept, a computer-readable storage medium is proposed. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the steps of a method for identifying a radiation source based on multi-domain feature fusion and self-attention in the above embodiment.

[0068] Embodiment IV: In the fourth embodiment of the present invention, based on the same inventive concept, a computer is proposed. The computer includes: a processor and a memory; the processor and the memory communicate with each other; the memory is used to store instructions; the processor is used to execute the instructions in the memory and implement a method for identifying a radiation source based on multi-domain feature fusion and self-attention in the above embodiment.

[0069] It should be understood that each part of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following well-known technologies in the art can be used: discrete logic circuits with logic gate circuits for implementing logical functions on data signals, application-specific integrated circuits with appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0070] Finally, it should be noted that the above are only preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, for those skilled in the art, they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A radiation source recognition method based on multi-domain feature fusion and self-attention, characterized in that It includes the following steps: Receive the original radiation source signal, which is a one-dimensional complex signal containing in-phase and quadrature components; Input the original signal into the time domain branch and the frequency domain branch respectively for processing. The time domain branch uses a long short-term memory network combined with a self-attention mechanism for feature extraction, and the frequency domain branch is transformed into a time-frequency diagram through a short-time Fourier transform and uses a convolutional neural network for feature extraction; Fuse the features extracted by the time domain branch and the frequency domain branch to obtain a comprehensive feature vector; Input the comprehensive feature vector into a classifier for category judgment, use the Softmax activation function to output the probabilities of each category, and select the category with the highest probability as the recognition result of the radiation source.

2. The radiation source recognition method based on multi-domain feature fusion and self-attention according to claim 1, wherein The long short-term memory network in the time domain branch is used to capture the time dependence relationship of the signal, and the self-attention mechanism assigns different weights to different time steps of the time domain signal.

3. The method for identifying radiation sources based on multi-domain feature fusion and self-attention according to claim 1, wherein The short-time Fourier transform in the frequency domain branch transforms the signal from the time domain to the frequency domain to obtain a two-dimensional time-frequency diagram, and the convolutional neural network is used to extract frequency domain features from the time-frequency diagram to identify key patterns in the spectrum.

4. A radiation source recognition method based on multi-domain feature fusion and self-attention according to claim 1, characterized in that The features extracted by the time domain branch and the frequency domain branch are fused through a concatenation operation, and the obtained fused feature vector contains the joint information of the time domain and the frequency domain.

5. The radiation source recognition method based on multi-domain feature fusion and self-attention according to claim 1, characterized in that The classifier is a fully connected layer, and the Softmax activation function is used to calculate the recognition probability of each category, and the radiation source is identified according to the maximum probability value.

6. The radiation source recognition method based on multi-domain feature fusion and self-attention according to claim 1, characterized in that, During the training process, the cross-entropy loss function and the Adam optimizer are used for optimization, and the network parameters are adjusted through the backpropagation algorithm.

7. A radiation source recognition system based on multi-domain feature fusion and self-attention, characterized in that, For a radiation source recognition method based on multi-domain feature fusion and self-attention according to any one of claims 1-7, the system includes: A signal input module for receiving the original radiation source signal; A time domain branch module for extracting features from the time domain signal by combining a long short-term memory network and a self-attention mechanism; A frequency domain branch module for extracting features from the frequency domain signal using a short-time Fourier transform and a convolutional neural network; A feature fusion module for fusing the features of the time domain branch and the frequency domain branch to form a comprehensive feature vector; A classification module for classifying the comprehensive feature vector through a fully connected layer and a Softmax activation function and outputting the recognition result of the radiation source.

8. The radiation source recognition system based on multi-domain feature fusion and self-attention according to claim 7, wherein, The time domain branch module and the frequency domain branch module process the time domain and frequency domain signals respectively, and perform joint processing through the feature fusion module, and finally predict the radiation source category through the classification module.

9. A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements a radiation source recognition method based on multi-domain feature fusion and self-attention according to any one of claims 1 to 7.

10. A readable storage medium, characterized in that, A computer program is stored on the readable storage medium, and when the computer program is executed by the processor, it implements a radiation source recognition method based on multi-domain feature fusion and self-attention according to any one of claims 1 to 7.

Images