Radar emitter type recognition method based on pulse imaging and migration learning
Patent Information
- Application Number
- CN202410304102.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-03-18
- Publication Date
- 2026-09-22
- Estimated Expiration
- 2044-03-18
AI Technical Summary
[0005]针对当前依赖单脉冲脉间特征参数的雷达辐射源类型识别方法难以准确识别复杂体制雷达类型的问题，本发明提供一种基于脉冲图像化与迁移学习的雷达辐射源类型识别方法，该方法通过聚合依据到达时间的多脉冲并映射为脉冲类图像，扩展辐射源信号表征维度；同时引入图像识别领域的基础预训练模型DINOv2提取脉冲类图像隐特征，进而构建辐射源类型分类器，缓解模型对于训练样本量的需求，实现辐射源类型的准确识别
1、通过模拟不同复杂程度的雷达辐射源脉间参数数据进行验证，结果显示本发明所提出的基于脉冲图像化与迁移学习的雷达辐射源类型识别方法能够准确用于复杂场景下的雷达辐射源类型识别。
Smart Images
Figure CN118262211B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of target recognition in electronic reconnaissance, and more specifically, to a method for identifying radar radiation source types based on pulse imaging and transfer learning. Background Technology
[0002] In recent years, new radar technologies such as phased array have developed rapidly, and the battlefield electromagnetic environment has become increasingly complex and changeable. Electromagnetic targets exhibit characteristics such as complex signal waveforms, overlapping parameters, and dynamic alternation. In addition, the actual detection environment has problems such as reflection, multipath, low signal-to-noise ratio, and receiver glitches. This has brought severe challenges to the accurate identification of radar radiation source types and the interpretation of complex target systems.
[0003] Currently, traditional radar radiation source type identification methods mainly rely on conventional inter-pulse characteristic parameters such as carrier frequency (RF), pulse width (PW), and repetition period (PRI). These methods identify radiation source types by constructing a classifier based on the relationship between inter-pulse characteristic parameters and radar model. While these methods are effective for conventional radar systems, their ability to identify complex radar systems is limited, resulting in low accuracy. This is due to two main reasons: first, relying solely on inter-pulse characteristic parameters of a single pulse, without considering the time accumulation effect between pulses, makes it difficult to characterize subtle signal differences, leading to low accuracy; second, the received data for some complex radar radiation sources is insufficient to support the retraining of the identification model from scratch.
[0004] To address the shortcomings of current methods, firstly, it is necessary to expand the dimensions of radiation source signal representation and increase the amount of input information during the training of the identification model; secondly, based on the idea of transfer learning, domain adaptation should be considered, and the need for training samples should be reduced by introducing a pre-trained model. Based on this, this invention proposes a radar radiation source type identification method based on pulse imaging and transfer learning. Summary of the Invention
[0005] To address the problem that current radar radiation source type identification methods relying on single-pulse inter-pulse feature parameters are insufficient for accurately identifying complex radar systems, this invention provides a radar radiation source type identification method based on pulse imaging and transfer learning. This method expands the dimension of radiation source signal representation by aggregating multiple pulses based on arrival time and mapping them to pulse-class images. Simultaneously, it introduces the DINOv2 pre-trained model, a fundamental model in image recognition, to extract latent features from the pulse-class images, thereby constructing a radiation source type classifier. This alleviates the model's requirement for a large number of training samples and enables accurate identification of radiation source types.
[0006] This invention provides a radar radiation source type identification method based on pulse imaging and transfer learning, comprising the following steps: S1, Input the raw pulse data of the radar radiation source and preprocess the raw pulse data; S2, based on projection matrix mapping and convolutional channel expansion, converts the preprocessed pulses into a three-channel image; S3 uses a pre-trained image model to extract the temporal pulse latent features of the three-channel image; S4. Construct and train a radar radiation source type classifier, and use the trained radar radiation source type classifier to complete the radar radiation source type identification.
[0007] Furthermore, in step S1, the input raw pulse data of the radar radiation source is represented as follows:
[0008] in, Indicates the number of pulse samples. This represents the inter-pulse parameters corresponding to a single pulse, considering four main parameters: RF, PW, PRI, and DOA. This represents the radar radiation source label from which the corresponding raw pulse data originated.
[0009] Furthermore, in step S1, the preprocessing of the raw pulse data includes: preprocessing the raw pulse data by setting a multi-condition threshold based on inter-pulse parameters, which can extract the single-target level signal from the raw pulse data.
[0010] Furthermore, setting multi-condition thresholds based on inter-pulse parameters includes:
[0011] in, These represent the average values of carrier frequency, pulse width, angle of arrival, and repetition period of signals merged into the same CPI level.
[0012] Furthermore, during training, it is necessary to remove discrete abnormal pulses that appear after preprocessing.
[0013] Furthermore, step S2 includes the following sub-steps: S21, obtains a length of [length] from the preprocessed pulse through equal-interval sampling. The sampling pulse sequence; S22, Calculate the image mapping matrix of the sampled pulse sequence based on the projection matrix mapping method. ; S23, a convolutional neural network is used to map the image of the sampled pulse sequence to a matrix. Expand the dimension to a two-channel matrix, and then map the two-channel matrix to the corresponding image mapping matrix. The images are then merged to obtain a three-channel image.
[0014] Furthermore, in step S21: make If the sampling pulse sequence is represented, then the equally spaced sampling process is represented as follows:
[0015] in, Indicates the sampling interval. This represents a sampling pulse, which ensures that the size of the image mapped to it is consistent by sampling at equal intervals.
[0016] Furthermore, in step S22, the image mapping matrix , is represented as:
[0017] in, The image mapping matrix represents the sampled pulse sequence, and its size is... , Represents the 2-norm of a matrix; sampling pulses in a sampling pulse sequence. Arranged by arrival time.
[0018] Furthermore, in step S23, the convolutional neural network includes: One convolutional layer, one batch normalization layer, and one non-linear function activation layer; Convolutional layers are used to extract discriminative temporal pulse map features from the input; Batch normalization layers are used to alleviate the potential gradient vanishing problem caused by network depth and accelerate the convergence speed of the model. Nonlinear function activation layers are used to enhance the representation of nonlinear features of time-series pulse graphs.
[0019] Furthermore, the formulaic description of the nonlinear function activation layer is as follows:
[0020] in, The activation function is called the parameterized modified linear unit activation function; the nonlinear function activation layer dynamically updates its coefficients using gradient descent via backpropagation. The value of .
[0021] Furthermore, the pre-trained image model uses the DINOv2 pre-trained model.
[0022] Furthermore, in step S4, constructing the radar radiation source type classifier includes: using average pooling to extract the latent features of temporal pulses. The channel dimensions are unified; based on this, the processed temporal pulse latent features are flattened, and a two-stage linear layer is added to reduce the dimensionality and output the classification results.
[0023] Furthermore, in step S4, the constructed radar radiation source type classifier is represented as:
[0024] in, Representing the latent features of time-series pulses ,in, Representing the The input feature map corresponding to Zhang San's three-channel image. This represents the number of channels in the input feature map. Indicates the height of the input feature map. This indicates the width of the input feature map. Representative input Output features after average pooling and flattening Represents channel average pooling. Represents the flattening operation of characteristic vectors; This represents the output of the first linear layer. This represents the weight vector of the first linear layer. This represents the bias of the first linear layer; This represents the output of the second linear layer. This represents the weight vector of the second linear layer. This represents the bias of the second linear layer.
[0025] Furthermore, in step S4, training the radar radiation source type classifier includes: Let the number of radar source categories corresponding to the training pulse data be . , This represents the label values of the training samples corresponding to the three-channel image. Indicates the sequence number of the three-channel image. Indicates the number of three-channel images. This indicates rounding down; let the parameter vector of all added networks be... The formula for calculating the category results using the SoftMax loss function is as follows:
[0026] in, This is an indicator function; it outputs 1 if the input expression is true, and 0 otherwise. This represents the first fully connected layer output vector. One element, This represents the loss of the radar radiation source type classifier, which is used to train the radar radiation source type classifier parameters and achieve the identification of known radiation source types.
[0027] Furthermore, the radar radiation source type identification method based on pulse imaging and transfer learning also includes: S5 evaluates the performance of the radar source type identification model by using the three-channel image of the test pulse mapping and its corresponding radar source label.
[0028] Furthermore, step S5 includes:
[0029] in, This represents the overall classification accuracy, used to measure the performance in identifying known radar radiation source types. The original labels are given as the size of the test sample set. The corresponding predicted label is .
[0030] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. The method proposed in this invention, based on pulse imaging and transfer learning, is validated by simulating inter-pulse parameter data of radar radiation sources with different levels of complexity. The results show that the radar radiation source type identification method can be accurately used for radar radiation source type identification in complex scenarios.
[0031] 2. By comparing the model classification performance under different inter-pulse parameter inputs of radiation sources, the results show that: on the one hand, the radar radiation source type identification method based on pulse imaging and transfer learning proposed in this invention has good feature compatibility; on the other hand, by adding parameters with pulse temporal correlation characteristics, such as repetition period and duty cycle, the model classification performance can be significantly improved, which indirectly confirms the rationality of combining and mapping temporal pulses into images in this invention. Attached Figure Description
[0032] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0033] Figure 1 This is a flowchart of a radar radiation source type identification method based on pulse imaging and transfer learning in an embodiment of the present invention.
[0034] Figure 2 This is a flowchart of a time-pulse single-channel image extension to three channels in an embodiment of the present invention.
[0035] Figure 3 This is a schematic diagram of the label-free self-distillation architecture of DINOv2 in an embodiment of the present invention.
[0036] Figure 4This is an example diagram illustrating the extraction of temporal pulse latent features using the DINOv2 pre-trained model in an embodiment of the present invention.
[0037] Figure 5 The graphs show the loss convergence curve and classification performance evaluation curve during the training process of the radar radiation source type identification method based on pulse imaging and transfer learning in this embodiment of the invention.
[0038] Figure 6 This is a diagram illustrating the type recognition performance of the radar radiation source type recognition method based on pulse imaging and transfer learning in this embodiment of the invention at different input feature dimensions. Detailed Implementation
[0039] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0040] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0041] Example like Figure 1 As shown, this embodiment proposes a radar radiation source type identification method based on pulse imaging and transfer learning, including the following steps: S1, Input the raw pulse data of the radar radiation source and preprocess the raw pulse data; By setting multi-condition thresholds based on inter-pulse parameters to preprocess the raw pulse data, single-target level signals can be extracted from the raw pulse data. Assume the input raw pulse data from the radar radiation source is represented as follows:
[0042] in, Indicates the number of pulse samples. This represents the inter-pulse parameters corresponding to a single pulse. In this embodiment, we mainly consider four parameters: RF, PW, PRI, and DOA. This represents the radar radiation source label from which the corresponding raw pulse data originated.
[0043] This embodiment processes the input radar radiation source pulse data based on the following judgment logic:
[0044] in, These represent the average carrier frequency, pulse width, angle of arrival, and repetition period of the signal merged into the same CPI level. The specific values for each condition can be set as needed. Furthermore, since a single pulse... Radar radiation source tag Since this is known, any discrete abnormal pulses that appear after the above preprocessing, i.e., pulses whose corresponding labels are inconsistent with nearby pulses, are removed. By performing the above preprocessing on the original pulse data, it is ensured that the time-series pulse data subsequently input into the model comes from the same target.
[0045] S2, based on projection matrix mapping and convolution channel expansion, converts the preprocessed pulses into a three-channel image, expanding the dimension of the radiation source signal representation.
[0046] First, to standardize the size of the input image, a length of [length missing] is obtained from the preprocessed pulse using an equally spaced sampling method. The sampling pulse sequence.
[0047] make If we represent a sampling pulse sequence, then the equally spaced sampling process can be represented as follows:
[0048] in, Indicates the sampling interval. This represents a sampling pulse, which ensures that the size of the image mapped to it is consistent by sampling at equal intervals.
[0049] Then, the image mapping matrix of the time-series pulses is calculated based on the sampled pulses, where Arranged according to arrival time, this embodiment calculates the image mapping matrix of the pulse sequence based on projection matrix mapping, as follows:
[0050] in, The image mapping matrix represents the sampled pulse sequence, and its size is... , This represents the L2 norm of the matrix. Through the above method, the original pulse parameter characteristic matrix is transformed into... A single-channel image matrix.
[0051] Finally, to further expand the dimension of the radiation source signal representation and adapt to the input of the pre-trained image model, this embodiment uses a convolutional neural network to expand the image mapping matrix of the sampling pulse sequence before the image pre-training model. The number of channels is determined to obtain a three-channel image.
[0052] First, a convolutional layer is added to expand the input single-channel image matrix to a two-channel matrix. Second, to preserve the features of the original image matrix, the expanded two-channel matrix is merged with the original single-channel matrix to form a three-channel image, which is then used as the subsequent input to the model. This represents the three-channel image output after passing through a convolutional neural network. The mapping function representing the pre-convolution is expressed in the following formula:
[0053] This convolutional neural network mainly consists of a convolutional layer (Conv_1), a batch normalization layer (BatchNorm_1), and a non-linear activation layer (PReLU). The specific computation process is as follows: Figure 2 As shown. Convolutional layers are used to extract discriminative temporal pulse map features from the input; batch normalization layers are used to alleviate the potential gradient vanishing problem caused by network depth, accelerating the convergence speed of the model; nonlinear function activation layers are used to enhance the representation ability of nonlinear features of temporal pulse maps, and their formulaic description is as follows:
[0054] in, The activation function is called the Parametric Modified Linear Unit (PReLU) activation function. Compared to the traditional ReLU function, its advantage lies in that it does not lose input on the negative half-axis, avoiding the "neuron death" phenomenon. Building upon this, the PReLU function uses gradient descent via backpropagation to dynamically update its coefficients. The value of is adjusted according to different inputs during the training phase, while the learned parameters remain unchanged during the testing phase.
[0055] S3 uses a pre-trained image model to extract the temporal pulse latent features of the three-channel image; In this embodiment, the image pre-training model uses the DINOv2 pre-training model. The DINOv2 pre-training model is based on the large-scale visual model ViT (Vision Transformer) and is obtained through unlabeled self-distillation. The total number of network parameters is approximately 21 million, and it achieves general visual processing capabilities through pre-training on over 100 million images. Testing shows that the DINOv2 pre-training model achieves 80% Top-1 accuracy on the ImageNet dataset, far exceeding most supervised models trained from scratch, fully demonstrating that its extracted latent image features can be used for downstream tasks such as image classification.
[0056] The key to the DINOv2 pre-trained model's general visual processing capabilities lies in its label-free self-distillation mechanism, such as... Figure 3As shown. Assume the input image of the DINOv2 pre-trained model is... By applying two different random variations, such as adding noise with different distributions, a pair of views is generated. Then the view pair is passed to the Student network and the Teacher network respectively, with the parameters for the two networks being... and The two networks share the same architecture but differ in parameters. The teacher network's output is centered on the batch-computed average, and each network outputs a latent feature, normalized using Softmax along the feature dimension. Their similarity is measured using a cross-entropy loss function. Specifically, a stop-gradient (SG) operator is applied to the teacher network, propagating gradients only through the student network. The teacher network parameters are then updated based on an exponential mean shift (EMA) of the student network parameters, thus achieving a pre-training process of unlabeled distillation. Figure 4 This demonstrates the practical application of the DINOv2 pre-trained model for image latent feature extraction, showing that the DINOv2 pre-trained model can resolve the key target contours of the input image. In this embodiment, temporal pulse latent features are extracted by inputting a three-channel image into the DINOv2 pre-trained model, and the obtained temporal pulse latent features are used as input to the subsequent radar radiation source type classification network. Let... This represents the mapping function used by the DINOv2 pre-trained model to extract temporal impulse latent features. To represent the latent features of time-series pulses, the process can be formulated as follows:
[0057] S4. Construct and train a radar radiation source type classifier, and use the trained radar radiation source type classifier to complete radar radiation source type identification. The temporal pulse latent features extracted from the DINOv2 pre-trained model are mapped to the latent space, and an image classification layer is added to complete the radar radiation source type identification output. To construct an end-to-end radar radiation source type classifier, in this embodiment, average pooling is first used to map the temporal pulse latent features... The channel dimensions are unified so that subsequent feature dimensions can be embedded into the low-dimensional space. Based on this, the processed temporal pulse latent features are flattened, and a two-stage linear layer is added for dimensionality reduction to output the classification result. Assume... Representing the latent features of time-series pulses ,in, Representing the The input feature map corresponding to Zhang San's three-channel image. This represents the number of channels in the input feature map. Indicates the height of the input feature map. Indicates the width of the input feature map:
[0058] in, Representative input Output features after average pooling and flattening Represents channel average pooling. Represents the flattening operation of characteristic vectors; This represents the output of the first linear layer. This represents the weight vector of the first linear layer. This represents the bias of the first linear layer; This represents the output of the second linear layer. This represents the weight vector of the second linear layer. This represents the bias of the second linear layer.
[0059] Assume the number of radar source categories corresponding to the training pulse data is . , This represents the label values of the training samples corresponding to the three-channel image. Indicates the sequence number of the three-channel image. Indicates the number of three-channel images. This indicates rounding down. Let the parameter vector for adding all networks in this embodiment be... The formula for calculating the category results using the SoftMax loss function is as follows:
[0060] in, This is an indicator function; it outputs 1 if the input expression is true, and 0 otherwise. This represents the first fully connected layer output vector. One element, This represents the loss of the radar radiation source type classifier, which is used to train the radar radiation source type classifier parameters and achieve the identification of known radiation source types.
[0061] S5 evaluates the performance of the radar source type identification model using a three-channel image of the test pulse map and its corresponding radar source label. For a size of... The test sample set, whose original labels are The corresponding predicted label is Then use the overall classification accuracy. Measure the performance of identifying known radar radiation source types:
[0062] Example: The performance of the radar source type classifier was tested by simulating radar source type identification data from two sets of data with different levels of complexity. Group 1: The total number of radar source types was 15, with 239,874 pulses. The inter-pulse parameter variations were relatively simple, but the number of radar source types was large. Group 2: The total number of radar source types was 7, with 4,324,943 pulses. The inter-pulse parameter variations were complex, with influencing factors such as pulse aliasing and background interference, but the number of radar source types was smaller. In this embodiment, the number of sampled pulses... The specific implementation process is as follows: (1) Dividing training and test samples. In order to verify the radar radiation source type identification method based on pulse imaging and transfer learning proposed in this embodiment, the training and test samples are divided into completely independent training samples and test samples. Specifically, about 80% of the total number of pulse samples are used as the training set, and about 20% of the total number of pulse samples are used as the test set.
[0063] (2) Input training samples to train the radar radiation source type classifier parameters. The training samples were divided into mini-batch groups for radar radiation source type classifier parameter optimization. The training epochs were set to 50, the learning rate to 0.001, and the class loss was optimized using the Adam optimizer. The experimental hardware platform was: Intel(R) Core(TM) i9-13900K CPU 4.70GHz, 32GB RAM; the software platform was: Python 3.9, PyTorch 2.0, CUDA 11.7.
[0064] Figure 5 This paper presents the loss convergence curve and classification performance evaluation curve of the radar radiation source type recognition method based on pulse imaging and transfer learning proposed in this invention during the training process. Specifically, the loss value for each batch is recorded, and the classification accuracy of training and test samples is evaluated based on the parameters of the last iteration of each epoch. The results show that, overall, the radar radiation source type recognition method based on pulse imaging and transfer learning proposed in this invention exhibits convergence, and the accuracy of radar radiation source type recognition gradually improves as the training error decreases. Specifically, for group one, the type recognition accuracy on the test set eventually converges to approximately 95.2%; for group two, the type recognition accuracy on the test set eventually converges to approximately 80%. The effectiveness of the proposed radar radiation source type recognition method based on pulse imaging and transfer learning in radar radiation source type recognition is demonstrated by applying it to radar radiation source type recognition datasets with different complexities.
[0065] Figure 6The recognition performance of the proposed radar source type identification method based on pulse imaging and transfer learning was evaluated when using different feature parameters of the radar source, i.e., different input feature dimensions. For the test data in Group 1, different inter-pulse parameters were used to evaluate the type recognition performance of the proposed radar source type classifier. "Unprocessed" indicates that the input data has not undergone the preprocessing process proposed in this invention. RF represents carrier frequency, PW represents pulse width, DTOA represents strictly calculated repetition period (PRI), PA represents amplitude, and ZKB represents duty cycle. Experimental results show that preprocessing significantly enhances the proposed method. Furthermore, by adding pulse temporal correlation features, such as DTOA and ZKB, the information content after mapping the original pulses to images can be significantly improved, which helps the downstream model extract better latent features and thus improve classification performance.
[0066] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A radar radiation source type identification method based on pulse imaging and transfer learning, characterized in that, Includes the following steps: S1, Input the raw pulse data of the radar radiation source and preprocess the raw pulse data; Preprocessing the raw pulse data includes setting multi-condition thresholds based on inter-pulse parameters to preprocess the raw pulse data, which can extract single-target level signals from the raw pulse data; S2, based on projection matrix mapping and convolutional channel expansion, converts the preprocessed pulses into a three-channel image; S3, a pre-trained image model is used to extract the temporal pulse latent features of the three-channel image; the pre-trained image model is the DINOv2 pre-trained model; S4. Construct and train a radar radiation source type classifier, and use the trained radar radiation source type classifier to complete radar radiation source type identification. Step S2 includes the following sub-steps: S21, obtains a length of [length] from the preprocessed pulse through equal-interval sampling. The sampling pulse sequence; S22, Calculate the image mapping matrix of the sampled pulse sequence based on the projection matrix mapping method. ; S23, a convolutional neural network is used to map the image of the sampled pulse sequence to a matrix. Expand the dimension to a two-channel matrix, and then map the two-channel matrix to the corresponding image mapping matrix. The images are merged to obtain a three-channel image; the convolutional neural network includes a convolutional layer, a batch normalization layer, and a nonlinear function activation layer.
2. The radar radiation source type identification method based on pulse imaging and transfer learning according to claim 1, characterized in that, In step S1, the input raw pulse data of the radar radiation source is represented as follows: in, Indicates the number of pulse samples. This represents the inter-pulse parameters corresponding to a single pulse, considering four main parameters: RF, PW, PRI, and DOA. This represents the radar radiation source label from which the corresponding raw pulse data originated.
3. The radar radiation source type identification method based on pulse imaging and transfer learning according to claim 2, characterized in that, Setting multi-condition thresholds based on inter-pulse parameters includes: in, These represent the average values of carrier frequency, pulse width, angle of arrival, and repetition period of signals merged into the same CPI level.
4. The radar radiation source type identification method based on pulse imaging and transfer learning according to claim 2, characterized in that, During training, discrete abnormal pulses that appear after preprocessing need to be removed.
5. The radar radiation source type identification method based on pulse imaging and transfer learning according to claim 2, characterized in that, In step S21: make If the sampling pulse sequence is represented, then the equally spaced sampling process is represented as follows: in, Indicates the sampling interval. This represents a sampling pulse, which ensures that the size of the image mapped to it is consistent by sampling at equal intervals.
6. The radar radiation source type identification method based on pulse imaging and transfer learning according to claim 5, characterized in that, In step S22, the image mapping matrix , represented as: in, The image mapping matrix represents the sampled pulse sequence, and its size is... , Represents the 2-norm of a matrix; sampling pulses in a sampling pulse sequence. Arranged by arrival time.
7. The radar radiation source type identification method based on pulse imaging and transfer learning according to claim 6, characterized in that, The formulaic description of the nonlinear function activation layer is as follows: in, The activation function is called the parameterized modified linear unit activation function; the nonlinear function activation layer dynamically updates its coefficients using gradient descent via backpropagation. The value of .
8. The radar radiation source type identification method based on pulse imaging and transfer learning according to claim 1, characterized in that, In step S4, constructing the radar radiation source type classifier includes: using average pooling to extract the latent features of temporal pulses. The channel dimensions are unified; based on this, the processed temporal pulse latent features are flattened, and a two-stage linear layer is added to reduce the dimensionality and output the classification results.
9. The radar radiation source type identification method based on pulse imaging and transfer learning according to claim 8, characterized in that, In step S4, the constructed radar radiation source type classifier is represented as follows: in, Representing the latent features of time-series pulses ,in, Representing the The input feature map corresponding to Zhang San's three-channel image. This represents the number of channels in the input feature map. Indicates the height of the input feature map. This indicates the width of the input feature map. Representative input Output features after average pooling and flattening Represents channel average pooling. Represents the flattening operation of feature vectors; This represents the output of the first linear layer. This represents the weight vector of the first linear layer. This represents the bias of the first linear layer; This represents the output of the second linear layer. This represents the weight vector of the second linear layer. This represents the bias of the second linear layer.
10. The radar radiation source type identification method based on pulse imaging and transfer learning according to claim 1, characterized in that, In step S4, training the radar radiation source type classifier includes: Let the number of radar source categories corresponding to the training pulse data be . , This represents the label values of the training samples corresponding to the three-channel image. Indicates the sequence number of the three-channel image. Indicates the number of three-channel images. This indicates rounding down; let the parameter vector of all added networks be... The formula for calculating the category results using the SoftMax loss function is as follows: in, This is an indicator function; it outputs 1 if the input expression is true, and 0 otherwise. This represents the first fully connected layer output vector. One element, This represents the loss of the radar radiation source type classifier, which is used to train the radar radiation source type classifier parameters and achieve the identification of known radiation source types.
11. The radar radiation source type identification method based on pulse imaging and transfer learning according to claim 1, characterized in that, Also includes: S5 evaluates the performance of the radar source type identification model by using the three-channel image of the test pulse mapping and its corresponding radar source label.
12. The radar radiation source type identification method based on pulse imaging and transfer learning according to claim 11, characterized in that, Step S5 includes: in, This represents the overall classification accuracy, used to measure the performance in identifying known radar radiation source types. The original labels are given as the size of the test sample set. The corresponding predicted label is .