Lightweight radiation source signal identification method and system

Abstract

The application relates to the technical field of radiation source signal identification, and discloses a light-weight radiation source signal identification method and system, which comprises the following steps: acquiring multiple types of radiation source signal data, pre-processing the radiation source signal data to obtain a pre-processed data set, and dividing the pre-processed data set into a training data set and a verification data set according to a preset proportion; inputting the training data set into an attention mechanism module of a convolutional neural network to obtain a two-dimensional spatial attention graph of each sample in the training data set; inputting the two-dimensional spatial attention graph of each sample in the training data set into a teacher model to obtain a spectral relationship matrix; training a student model according to the training data set and the spectral relationship matrix, and simultaneously adaptively constructing an optimal decay point set in the student model training process according to the verification data set to obtain a trained student model, which is used for identifying radiation source signals. The application improves the accuracy of radiation source signal identification.

Description

Technical Field

[0001] This application relates to the field of radiation source signal identification technology, and in particular to a lightweight radiation source signal identification method and system. Background Technology

[0002] Traditional radiation source identification methods rely on manual feature extraction and simple classification algorithms, which severely limit their accuracy and robustness in complex electromagnetic environments, making them unsuitable for demanding applications. In recent years, deep learning-based identification methods have become a research hotspot due to their high intelligence, robustness, and strong generalization ability. These methods can automatically learn data feature representations, overcoming the problems of complex feature extraction and low accuracy associated with traditional methods. However, deep learning models have a large parameter scale and high computational complexity, making them difficult to deploy on resource-constrained edge computing or mobile devices. Training and inference require significant computational resources and time, potentially leading to excessive latency or inability to process signals in real time, thus limiting their practical application scope. Summary of the Invention

[0003] To address the issues of deployment efficiency and real-time performance of deep learning-based recognition models in resource-constrained environments, this application provides a lightweight radiation source signal recognition method and system, which improves the accuracy of radiation source signal recognition.

[0004] This application discloses a lightweight radiation source signal identification method, which includes: Step 1: Acquire various types of radiation source signal data, preprocess the radiation source signal data to obtain a preprocessed dataset, and divide it into a training dataset and a validation dataset according to a preset ratio; the radiation source signal data includes the amplitude, frequency, and phase of the radiation source signal; the preprocessing includes filtering and normalization; Step 2: Input the training dataset into the attention mechanism module of the convolutional neural network to obtain a two-dimensional spatial attention map of each sample in the training dataset; Step 3: Input the two-dimensional spatial attention map of each sample in the training dataset into the teacher model to obtain the spectral relation matrix; Step 4: Train the student model based on the training dataset and the spectral relation matrix, and simultaneously construct the optimal attenuation point set during the student model training process based on the validation dataset to obtain the trained student model. The trained student model is used to identify radiation source signals.

[0005] Further, step 2 includes: Step 21: Input each sample in the training dataset into a convolutional neural network to generate a three-dimensional tensor for each sample; Step 22: Construct a spatial attention mapping function based on the feature map channels, the 3D tensor channels, and the attention focusing parameters; Step 23: Based on the three-dimensional tensor of each sample and the spatial attention mapping function, obtain the two-dimensional spatial attention map of each sample.

[0006] Further, step 21 includes: The three-dimensional tensor T of each sample in the training dataset is generated using the following formula:

[0007] in, This is a feature extraction operator for convolutional neural networks, used to perform convolution operations on samples X in the training dataset to obtain a three-dimensional tensor T. Step 22 includes: The spatial attention mapping function is obtained through the following formula:

[0008] middle, Let C be the spatial attention mapping function, and C be the feature map channel. For the m-th channel of the three-dimensional tensor, For attention focusing parameters, To find the sign of the absolute value; Step 23 includes: Substitute the three-dimensional tensor T of each sample into the spatial attention mapping function. The intermediate matrix of each sample is obtained and normalized to generate a two-dimensional spatial attention map of each sample, wherein the two-dimensional spatial attention map is represented in the form of a two-dimensional matrix.

[0009] Further, step 3 includes: Step 31: In the same layer of the teacher model, the two-dimensional spatial attention map of each sample in the training dataset is expanded into a one-dimensional vector in row order to obtain the feature vector of the sample. Step 32: Calculate the Euclidean distance between the feature vectors of each sample; Step 33: Calculate the spectral relation matrix between each sample in the same layer of the teacher model based on the Euclidean distance; wherein, the value in the i-th row and j-th column of the spectral relation matrix is ​​the similarity between the i-th row and sample j. The specific calculation method of the similarity is to first calculate the Euclidean distance between sample i and sample j, and then use the Gaussian kernel function to map the Euclidean distance to convert the Euclidean distance into a similarity.

[0010] Further, step 33 includes: The teacher model is obtained through the following formula. Spectral relation matrix of the layer Similarity of the i-th row and j-th column:

[0011] in, The teacher model is represented by the first... Spectral relation matrix of the layer The similarity between the i-th sample and the j-th sample, This represents the Euclidean distance between the i-th sample and the j-th sample in layer l; This represents the smoothing parameter, which determines the degree of smoothing of the similarity between samples at different distances; This represents the total number of samples in the training dataset.

[0012] Further, step 4 includes: Step 41: Input the training dataset into the student model to obtain high-level semantic feature information, wherein the high-level semantic feature information refers to the feature vector or feature matrix obtained by pooling or flattening the output of the last three feature extraction layers of the student model. Step 42: Based on the high-level semantic feature information and the spectral relation matrix, calculate the loss value of the preset loss function, which includes classification loss, knowledge distillation loss and structural similarity loss; Step 43: Optimize the student model based on the loss value, and update the learning rate by setting a multi-step learning rate decay interval strategy based on the validation dataset to obtain the trained student model.

[0013] Further, step 42 includes: Step 421: Obtain the classification loss based on the true and predicted labels of the student model; Step 422: Calculate the KL divergence between the student model and the teacher model, i.e., the knowledge distillation loss; Step 423: Obtain the structural similarity loss based on the high-level semantic features and the spectral relation matrix; Step 424: Obtain the loss value based on the classification loss, the knowledge distillation loss, and the structural similarity loss.

[0014] Further, step 421 includes: The classification loss is obtained using the following formula. :

[0015] in, This represents the u-th true label of the student model. This represents the u-th predicted label of the student model. This represents the total number of actual tags; Step 422 includes: The knowledge distillation loss is obtained using the following formula. :

[0016] in, Let represent the probability distribution of the k-th sample output by the teacher model. Let represent the probability distribution of the k-th sample output by the student model. This represents the total number of samples in the training dataset; Step 423 includes: The structural similarity loss is obtained using the following formula. :

[0017] in, This refers to the last three layers of the student and teacher models. Representative Teacher Model The first in the layer The first sample and the first The similarity of individual samples and These refer to the two feature vectors resulting from the expansion of the spatial attention maps of the i-th and j-th samples in the l-th layer of the student model, respectively. To find the modulus sign; Step 424 includes: The loss value is obtained using the following formula. :

[0018] in, and All represent weights.

[0019] Further, step 43 includes: Step 431: Set different learning rate decay interval strategies, including learning rate decay every B1 rounds, decay every B2 rounds, and decay every B3 rounds; and for each learning rate decay interval strategy, set a corresponding learning rate decay round as a candidate decay point; B2 is greater than B1 and less than B3. Step 432: For each learning rate decay interval strategy, the student model is trained using the training dataset and the loss value calculated in step 42. During the training process, the student model is backpropagated and its parameters are updated based on the loss value, and the learning rate is decayed when the corresponding candidate decay point is reached. After training, an ablation experiment is conducted for each learning rate decay interval strategy. The recognition effect of the student model using different learning rate decay interval strategies is tested on the validation dataset. The recognition accuracy of the radiation source signal is calculated and compared to obtain the accuracy information corresponding to different learning rate decay interval strategies. Step 433: Based on the accuracy information of different learning rate decay interval strategies, determine the decay point corresponding to the highest recognition accuracy for each learning rate decay interval strategy. Combine the decay points of different learning rate decay interval strategies to construct an adaptive optimal decay point set for the student model, so that the learning rate adjustment can adapt to the parameter optimization needs of different training stages of the model, and improve the convergence stability and parameter optimization effect of the model. Step 434: Based on the set of optimal decay points, iteratively train the student model; in each round of iterative training, backpropagate and update the parameters of the student model according to the loss value, and perform learning rate decay when the training round reaches the decay point in the set of optimal decay points, until the preset training termination condition is met, and finally obtain the trained student model.

[0020] This application also discloses a lightweight radiation source signal identification system, which implements the aforementioned lightweight radiation source signal identification method, comprising: A dataset construction module is used to acquire various types of radiation source signal data, preprocess the radiation source signal data to obtain a preprocessed dataset, and divide it into a training dataset and a validation dataset according to a preset ratio; the radiation source signal data includes the amplitude, frequency, and phase of the radiation source signal; the preprocessing includes filtering and normalization; The spatial attention map acquisition module is used to input the training dataset into the attention mechanism module of the convolutional neural network to obtain a two-dimensional spatial attention map of each sample in the training dataset. The spectral relation matrix acquisition module is used to input the two-dimensional spatial attention map of each sample in the training dataset into the teacher model to obtain the spectral relation matrix; The student model training module is used to train the student model based on the training dataset and the spectral relation matrix, and to adaptively construct the optimal attenuation point set during the student model training process based on the validation dataset, thereby obtaining the trained student model, which is used to identify radiation source signals.

[0021] By adopting the above technical solution, this application has the following advantages: Addressing the deployment efficiency and real-time performance issues of deep learning-based recognition models in resource-constrained environments, this application first acquires and preprocesses radiation source signal data to provide a high-quality data foundation for subsequent processing. The preprocessed dataset is input into the attention mechanism module, which generates a three-dimensional feature tensor through a convolutional neural network and calculates a two-dimensional spatial attention map focusing on key features using a spatial attention mapping function, thereby improving signal representation capabilities. Subsequently, this attention map is input into the teacher model to construct a spectral relation matrix, achieving knowledge transfer at the feature level. Next, the student model is trained using this spectral relation matrix in conjunction with the preprocessed dataset, and a multi-step learning rate decay strategy is employed to optimize model updates. Through this method, while reducing the number of model parameters and computational overhead and improving operational efficiency, the student model effectively inherits the generalization ability of the teacher model, enhancing its adaptability to scarce samples and changes in data distribution. Furthermore, this method supports automatic learning of deep features, reducing reliance on manual prior knowledge and effectively improving the accuracy of radiation source signal recognition in complex environments. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments recorded in the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings.

[0023] Figure 1 This is a flowchart illustrating a lightweight radiation source signal identification method according to an embodiment of this application. Detailed Implementation

[0024] The present application will be further described in conjunction with the accompanying drawings and embodiments. The described embodiments are only some, not all, of the embodiments of the present application. All other embodiments obtained by those skilled in the art should fall within the protection scope of the embodiments of the present application.

[0025] See Figure 1 This application provides an embodiment of a lightweight radiation source signal identification method, which includes: Step 1: Acquire various types of radiation source signal data, preprocess the radiation source signal data to obtain a preprocessed dataset, and divide it into a training dataset and a validation dataset according to a preset ratio (e.g., 8:2); the radiation source signal data includes the amplitude, frequency, and phase of the radiation source signal; the preprocessing includes filtering and normalization; the radiation source signals include, but are not limited to, radar signals and communication signals; Step 2: Input the training dataset into the attention mechanism module of the convolutional neural network to obtain a two-dimensional spatial attention map of each sample in the training dataset; Step 3: Input the two-dimensional spatial attention map of each sample in the training dataset into the teacher model to obtain the spectral relation matrix; Step 4: Train the student model based on the training dataset and the spectral relation matrix, and simultaneously construct the optimal attenuation point set during the student model training process based on the validation dataset to obtain the trained student model. The trained student model is used to identify radiation source signals.

[0026] The data collected in this embodiment should include multi-dimensional features such as amplitude, frequency, and phase to comprehensively reflect the essential attributes of the radiation source signal. Ensuring data coverage and diversity provides a solid foundation of raw data for subsequent signal identification and processing, ensuring the effectiveness of model training and identification. Preprocessing of the radiation source signal data includes operations such as filtering and normalization to eliminate noise and interference and improve data quality.

[0027] Optionally, step 2 includes: Step 21: Input each sample in the training dataset into a convolutional neural network to generate a three-dimensional tensor for each sample; Step 22: Construct a spatial attention mapping function based on the feature map channels, the 3D tensor channels, and the attention focusing parameters; Step 23: Based on the three-dimensional tensor of each sample and the spatial attention mapping function, obtain the two-dimensional spatial attention map of each sample.

[0028] In this embodiment, an attention mechanism is used to focus on key feature regions in the signal, suppress irrelevant information, and generate a two-dimensional spatial attention map that highlights important features, thereby improving the effective representation of the signal. By automatically learning the deep features of the radiation source signal through CNN, it adapts to the characteristics of signal variations in complex electromagnetic environments, reduces reliance on prior knowledge, and improves the comprehensiveness and robustness of feature extraction.

[0029] Optionally, step 21 includes: The three-dimensional tensor T of each sample in the training dataset is generated using the following formula:

[0030] in, This is a feature extraction operator for convolutional neural networks, used to perform convolution operations on samples X in the training dataset to obtain a three-dimensional tensor T. Step 22 includes: The spatial attention mapping function is obtained through the following formula:

[0031] in, Let C be the spatial attention mapping function, and C be the feature map channel. For the m-th channel of the three-dimensional tensor, For attention focusing parameters, To determine the sign of the absolute value, in practice, it can be determined through ablation experiments. The possible values ​​of ; Step 23 includes: Substitute the three-dimensional tensor T of each sample into the spatial attention mapping function. The intermediate matrix of each sample is obtained and normalized to generate a two-dimensional spatial attention map of each sample, wherein the two-dimensional spatial attention map is represented in the form of a two-dimensional matrix.

[0032] This application's embodiments introduce a spatial attention mechanism, enabling the model to adaptively focus on key feature regions in the signal, suppressing interference from noise and irrelevant information. Compared to traditional methods, this mechanism enhances the model's ability to capture features of weak and complex modulated signals, thereby improving recognition accuracy. The generated two-dimensional attention map provides focusing guidance for subsequent feature learning in both the teacher and student models. In radiation source identification, this map helps the model ignore background noise and focus on learning the unique feature patterns of the radiation source. Simultaneously, by reducing the feature dimension, the model becomes lightweight, reducing computational overhead and making it more suitable for deployment on edge devices.

[0033] Optionally, step 3 includes: Step 31: In the same layer of the teacher model, the two-dimensional spatial attention map of each sample in the training dataset is expanded into a one-dimensional vector in row order to obtain the feature vector of the sample. Step 32: Calculate the Euclidean distance between the feature vectors of each sample; Step 33: Calculate the spectral relation matrix between each sample in the same layer of the teacher model based on the Euclidean distance; wherein, the value in the i-th row and j-th column of the spectral relation matrix is ​​the similarity between the i-th row and sample j. The specific calculation method of the similarity is to first calculate the Euclidean distance between sample i and sample j, and then use the Gaussian kernel function to map the Euclidean distance to convert the Euclidean distance into a similarity.

[0034] The feature vectors in this embodiment can capture the time-frequency domain features or modulation mode features of the signal, making the feature vectors of similar radiation source signals closer in space, and providing a standardized feature representation for constructing a sample similarity structure. The calculation of Euclidean distance directly reflects the degree of difference between different radiation source signals in the feature space of the corresponding network layer, providing a numerical basis for constructing a sample similarity matrix. In complex electromagnetic environments, even if the signal is affected by noise interference, Euclidean distance can still stably characterize the feature similarity of similar radiation sources, helping the model filter noise interference, focus on essential feature differences, and improve recognition robustness.

[0035] The spectral relation matrix in this embodiment encodes the similarity distribution pattern of all samples in the feature space of each layer of the teacher model, transforming the teacher model's deep semantic understanding of radiation source signals into transferable structured knowledge. By using the spectral relation matrix to achieve sample structure knowledge distillation, the lightweight student model can learn the radiation source signal distribution pattern implicit in the teacher model, thereby improving the student model's generalization recognition ability of radiation source signals. At the same time, by compressing the feature space dimension through the matrix, it helps to achieve lightweight model deployment.

[0036] Optionally, step 33 includes: The teacher model is obtained through the following formula. Spectral relation matrix of the layer Similarity of the i-th row and j-th column:

[0037] in, The teacher model is represented by the first... Spectral relation matrix of the layer The similarity between the i-th sample and the j-th sample, This represents the Euclidean distance between the i-th sample and the j-th sample in layer l; The smoothing parameter (which can take the value 1) determines the smoothness of the similarity between samples at different distances. This represents the total number of samples in the training dataset.

[0038] Optionally, step 4 includes: Step 41: Input the training dataset into the student model to obtain high-level semantic feature information, wherein the high-level semantic feature information refers to the feature vector or feature matrix obtained by pooling or flattening the output of the last three feature extraction layers of the student model. Step 42: Based on the high-level semantic feature information and the spectral relation matrix, calculate the loss value of the preset loss function, which includes classification loss, knowledge distillation loss and structural similarity loss; Step 43: Optimize the student model based on the loss value, and update the learning rate by setting a multi-step learning rate decay interval strategy based on the validation dataset to obtain the trained student model.

[0039] This application's embodiments utilize a complex teacher model to extract the similarity distribution patterns between signal samples and generate a spectral relationship matrix to achieve knowledge transfer and provide guidance for training the student model. It should be noted that the teacher model includes, but is not limited to, ResNet110 and VGG19.

[0040] Optionally, step 42 includes: Step 421: Obtain the classification loss based on the true and predicted labels of the student model; Step 422: Calculate the KL divergence between the student model and the teacher model, i.e., the knowledge distillation loss; Step 423: Obtain the structural similarity loss based on the high-level semantic features and the spectral relation matrix; Step 424: Obtain the loss value based on the classification loss, the knowledge distillation loss, and the structural similarity loss.

[0041] Optionally, step 421 includes: The classification loss is obtained using the following formula. :

[0042] in, This represents the u-th true label of the student model. This represents the u-th predicted label of the student model. This represents the total number of actual tags; Step 422 includes: The knowledge distillation loss is obtained using the following formula. :

[0043] in, Let represent the probability distribution of the k-th sample output by the teacher model. Let represent the probability distribution of the k-th sample output by the student model. This represents the total number of samples in the training dataset; Step 423 includes: The structural similarity loss is obtained using the following formula. :

[0044] in, This refers to the last three layers of the student and teacher models. Representative Teacher Model The first in the layer The first sample and the first The similarity of individual samples and These refer to the two feature vectors resulting from the expansion of the spatial attention maps of the i-th and j-th samples in the l-th layer of the student model, respectively. To find the modulus sign; Step 424 includes: The loss value is obtained using the following formula. :

[0045] in, and All represent weights.

[0046] This application embodiment measures the deviation between the student model's prediction (predicted label) of the radiation source signal type and the true label by using classification loss, guiding the student model to learn the discrimination boundary of different radiation sources, thereby improving the recognition accuracy of known signal types.

[0047] This application's embodiments utilize KL divergence to measure the difference between the output probability distributions of the student model and the teacher model, enabling the transfer of knowledge from the teacher model to the student model and alleviating the problem of insufficient generalization ability caused by limited computing power in edge devices. In noisy environments, the teacher model's output can guide the student model to focus on the essential features of the signal, improving its ability to identify concealed radiation sources.

[0048] This application's embodiments constrain the student model through the spectral relation matrix to retain the similarity structure between radiation source signal samples (such as the clustering distribution of similar radiation sources in the feature space), avoiding the loss of deep correlation knowledge of signals during the lightweighting process, solving the problem of radiation source identification in small sample scenarios. When new radiation source data is scarce, the model can infer the similarity between unknown signals and known samples through the spectral relation matrix, associating unknown radar signals with models of similar pulse widths in the database, greatly improving the identification accuracy in small sample scenarios, enabling rapid adaptation to emerging radiation sources, and solving related problems such as difficult deployment of edge devices and high computational costs.

[0049] The embodiments of this application form a multi-dimensional optimization objective by weighted fusion of classification loss, knowledge distillation loss and structural similarity loss. While making the model lightweight, it balances classification accuracy, generalization ability and feature structure preservation, and avoids model performance bias caused by a single loss.

[0050] Optionally, step 43 includes: Step 431: Set different learning rate decay interval strategies, including learning rate decay every B1 rounds, decay every B2 rounds, and decay every B3 rounds; and for each learning rate decay interval strategy, set a corresponding learning rate decay round as a candidate decay point; B2 is greater than B1 and less than B3. Step 432: For each learning rate decay interval strategy, the student model is trained using the training dataset and the loss value calculated in step 42. During the training process, the student model is backpropagated and its parameters are updated based on the loss value, and the learning rate is decayed when the corresponding candidate decay point is reached. After training, an ablation experiment is conducted for each learning rate decay interval strategy. The recognition effect of the student model using different learning rate decay interval strategies is tested on the validation dataset. The recognition accuracy of the radiation source signal is calculated and compared to obtain the accuracy information corresponding to different learning rate decay interval strategies. Step 433: Based on the accuracy information of different learning rate decay interval strategies, determine the decay point corresponding to the highest recognition accuracy for each learning rate decay interval strategy. Combine the decay points of different learning rate decay interval strategies to construct an adaptive optimal decay point set for the student model, so that the learning rate adjustment can adapt to the parameter optimization needs of different training stages of the model, and improve the convergence stability and parameter optimization effect of the model. Step 434: Based on the set of optimal decay points, iteratively train the student model; in each round of iterative training, backpropagate and update the parameters of the student model according to the loss value, and perform learning rate decay when the training round reaches the decay point in the set of optimal decay points, until the preset training termination condition is met, and finally obtain the trained student model.

[0051] Specifically, different decay interval strategies are obtained. These strategies include decaying every 20 rounds, every 50 rounds, and every 100 rounds.

[0052] Then, ablation experiments were conducted according to the different attenuation interval strategies to calculate the recognition accuracy of the radiation source signal on the validation set, thus obtaining accuracy information. Ablation experiments were carried out for each learning rate attenuation interval strategy, and the model's recognition accuracy of the radiation source signal was tested on the validation set to quantify the impact of different strategies on model performance. By comparing the accuracy curves under each strategy, attenuation methods that enable the model to maintain a high recognition rate in complex electromagnetic environments were selected, and a strategy evaluation system based on measured data was established to avoid the blindness of traditional fixed attenuation strategies.

[0053] Next, the set of decay points is determined based on the accuracy information. In this step, based on the ablation experiment results, the learning rate decay timing with the highest validation set accuracy under each strategy is extracted to form the optimal decay point set. This set integrates the advantageous decay nodes of different strategies, enabling learning rate adjustment to converge quickly in the early stages of training and to finely optimize parameters in the later stages. An adaptive learning rate decay timetable is constructed for the radiation source identification model, solving the problem of "premature decay leading to underfitting or late decay leading to overfitting" caused by traditional fixed decay.

[0054] Finally, the student model is iteratively trained based on the set of attenuation points to obtain the final trained student model. In this step, the learning rate is automatically adjusted at each specified time according to the preset set of attenuation points, prompting the model to optimize parameters with appropriate step sizes at different training stages. Through multiple rounds of iterative training, the lightweight model can effectively learn the complex feature distribution of radiation source signals while maintaining low computational resource consumption.

[0055] The embodiments of this application reduce the number of parameters and computational overhead while possessing the ability to accurately identify radiation source signals in complex environments.

[0056] This application also provides an embodiment of a lightweight radiation source signal identification system, which implements the lightweight radiation source signal identification method described in the above embodiment, and includes: A dataset construction module is used to acquire various types of radiation source signal data, preprocess the radiation source signal data to obtain a preprocessed dataset, and divide it into a training dataset and a validation dataset according to a preset ratio; the radiation source signal data includes the amplitude, frequency, and phase of the radiation source signal; the preprocessing includes filtering and normalization; The spatial attention map acquisition module is used to input the training dataset into the attention mechanism module of the convolutional neural network to obtain a two-dimensional spatial attention map of each sample in the training dataset. The spectral relation matrix acquisition module is used to input the two-dimensional spatial attention map of each sample in the training dataset into the teacher model to obtain the spectral relation matrix; The student model training module is used to train the student model based on the training dataset and the spectral relation matrix, and to adaptively construct the optimal attenuation point set during the student model training process based on the validation dataset, thereby obtaining the trained student model, which is used to identify radiation source signals.

[0057] In this implementation, the training of the radiation source signal recognition model needs to go through a phase of rapid convergence followed by fine-tuning. The multi-step attenuation strategy employed in this application gradually reduces the learning rate in preset training rounds, allowing the model to quickly capture the main features of the signal with a larger learning rate in the initial stage, and then fine-tune the parameters with a smaller learning rate in the later stages, avoiding convergence oscillations or overfitting caused by an excessively large learning rate. Simultaneously, radiation source signals possess complex time-frequency characteristics and variable electromagnetic environments. Multi-step attenuation can specifically optimize the model's learning ability for strong signal features (initial stage) and weak signal details (later stage) at different training stages, ensuring the model's adaptability to various signals. This is something a fixed attenuation strategy cannot achieve. Furthermore, through the synergistic optimization of dynamic learning rate and lightweight model, when deployed on edge devices, it reduces computational power consumption and helps improve the model's recognition stability in complex electromagnetic environments.

[0058] This application improves signal representation quality from the feature extraction layer through spatial attention mechanism, breaks through the limitations of traditional methods from the knowledge transfer layer through instance spectrum relation knowledge distillation, and ensures the optimization direction of the model from the training layer through multi-dimensional loss function. The three elements support each other with the lightweight model architecture design, forming a complete technical chain of "feature enhancement-structure transfer-optimization guarantee", which ultimately realizes the high-precision deployment of the radiation source signal recognition model under the premise of lightweight design.

[0059] In one specific implementation, the method proposed in this application was compared with three representative knowledge distillation methods to verify that the method proposed in this application is more accurate than traditional knowledge distillation methods, as shown in Table 1.

[0060] Table 1 Comparison of Radiation Source Signal Identification Accuracy

[0061] Table 1 shows the comparison methods 1, 2, and 3, which are respectively: the output layer-based distillation method (KD), the intermediate layer feature-based IRG method, and the relation graph-based RKD method. Experiments were conducted on six different datasets, including two radiation source signal identification datasets and four image classification datasets. Specifically, dataset 1 is ADS-B, dataset 2 is NWPU-RESISC45, dataset 3 is RML2016.10b, dataset 4 is CIFAR-100coarse, dataset 5 is FashionMNIST, and dataset 6 is CIFAR-10. Experimental results show that the method used in this application outperforms the aforementioned comparison methods on all datasets, verifying its superiority and versatility.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and not to limit them. Although this application has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of this application. Any modifications or equivalent substitutions that do not depart from the spirit and scope of this application should be covered within the protection scope of the claims of this application.

Claims

1. A lightweight radiation source signal identification method, characterized in that, include: Step 1: Acquire various types of radiation source signal data, preprocess the radiation source signal data to obtain a preprocessed dataset, and divide it into a training dataset and a validation dataset according to a preset ratio; the radiation source signal data includes the amplitude, frequency, and phase of the radiation source signal; the preprocessing includes filtering and normalization; Step 2: Input the training dataset into the attention mechanism module of the convolutional neural network to obtain a two-dimensional spatial attention map of each sample in the training dataset; Step 3: Input the two-dimensional spatial attention map of each sample in the training dataset into the teacher model to obtain the spectral relation matrix; Step 4: Train the student model based on the training dataset and the spectral relation matrix, and simultaneously construct the optimal attenuation point set during the student model training process based on the validation dataset to obtain the trained student model. The trained student model is used to identify radiation source signals. Step 3 includes: Step 31: In the same layer of the teacher model, the two-dimensional spatial attention map of each sample in the training dataset is expanded into a one-dimensional vector in row order to obtain the feature vector of the sample. Step 32: Calculate the Euclidean distance between the feature vectors of each sample; Step 33: Calculate the spectral relation matrix between each sample in the same layer of the teacher model based on the Euclidean distance; wherein, the value in the i-th row and j-th column of the spectral relation matrix is ​​the similarity between the i-th row and sample j. The specific calculation method of the similarity is to first calculate the Euclidean distance between sample i and sample j, and then use the Gaussian kernel function to map the Euclidean distance to convert the Euclidean distance into a similarity. Step 33 includes: The teacher model is obtained through the following formula. Spectral relation matrix of the layer Similarity of the i-th row and j-th column: in, The teacher model is represented by the first... Spectral relation matrix of the layer The similarity between the i-th sample and the j-th sample, This represents the Euclidean distance between the i-th sample and the j-th sample in layer l; This represents the smoothing parameter, which determines the degree of smoothing of the similarity between samples at different distances; This represents the total number of samples in the training dataset; Step 4 includes: Step 41: Input the training dataset into the student model to obtain high-level semantic feature information, wherein the high-level semantic feature information refers to the feature vector or feature matrix obtained by pooling or flattening the output of the last three feature extraction layers of the student model. Step 42: Based on the high-level semantic feature information and the spectral relation matrix, calculate the loss value of the preset loss function, which includes classification loss, knowledge distillation loss and structural similarity loss; Step 43: Optimize the training of the student model based on the loss value, and update the learning rate by setting a multi-step learning rate decay interval strategy based on the validation dataset to obtain the trained student model; wherein, by combining the decay points of different learning rate decay interval strategies, an adaptive optimal decay point set is constructed for the student model.

2. The lightweight radiation source signal identification method according to claim 1, characterized in that, Step 2 includes: Step 21: Input each sample in the training dataset into a convolutional neural network to generate a three-dimensional tensor for each sample; Step 22: Construct a spatial attention mapping function based on the feature map channels, the 3D tensor channels, and the attention focusing parameters; Step 23: Based on the three-dimensional tensor of each sample and the spatial attention mapping function, obtain the two-dimensional spatial attention map of each sample.

3. The lightweight radiation source signal identification method according to claim 2, characterized in that, Step 21 includes: The three-dimensional tensor T of each sample in the training dataset is generated using the following formula: in, This is a feature extraction operator for convolutional neural networks, used to perform convolution operations on samples X in the training dataset to obtain a three-dimensional tensor T. Step 22 includes: The spatial attention mapping function is obtained through the following formula: in, Let C be the spatial attention mapping function, and C be the feature map channel. For the m-th channel of the three-dimensional tensor, For attention focusing parameters, To find the sign of the absolute value; Step 23 includes: Substitute the three-dimensional tensor T of each sample into the spatial attention mapping function. The intermediate matrix of each sample is obtained and normalized to generate a two-dimensional spatial attention map of each sample, wherein the two-dimensional spatial attention map is represented in the form of a two-dimensional matrix.

4. The lightweight radiation source signal identification method according to claim 1, characterized in that, Step 42 includes: Step 421: Obtain the classification loss based on the true and predicted labels of the student model; Step 422: Calculate the KL divergence between the student model and the teacher model, i.e., the knowledge distillation loss; Step 423: Obtain the structural similarity loss based on the high-level semantic features and the spectral relation matrix; Step 424: Obtain the loss value based on the classification loss, the knowledge distillation loss, and the structural similarity loss.

5. The lightweight radiation source signal identification method according to claim 4, characterized in that, Step 421 includes: The classification loss is obtained using the following formula. : in, This represents the u-th true label of the student model. This represents the u-th predicted label of the student model. This represents the total number of actual tags; Step 422 includes: The knowledge distillation loss is obtained using the following formula. : in, Let represent the probability distribution of the k-th sample output by the teacher model. Let represent the probability distribution of the k-th sample output by the student model. This represents the total number of samples in the training dataset; Step 423 includes: The structural similarity loss is obtained using the following formula. : in, This refers to the last three layers of the student and teacher models. Representative Teacher Model The first in the layer The first sample and the first The similarity of individual samples and These refer to the two feature vectors resulting from the expansion of the spatial attention maps of the i-th and j-th samples in the l-th layer of the student model, respectively. To find the modulus sign; Step 424 includes: The loss value is obtained using the following formula. : in, and All represent weights.

6. The lightweight radiation source signal identification method according to claim 1, characterized in that, Step 43 includes: Step 431: Set different learning rate decay interval strategies, including learning rate decay every B1 rounds, decay every B2 rounds, and decay every B3 rounds; and for each learning rate decay interval strategy, set a corresponding learning rate decay round as a candidate decay point; B2 is greater than B1 and less than B3. Step 432: For each learning rate decay interval strategy, the student model is trained using the training dataset and the loss value calculated in step 42. During the training process, the student model is backpropagated and its parameters are updated based on the loss value, and the learning rate is decayed when the corresponding candidate decay point is reached. After training, an ablation experiment is conducted for each learning rate decay interval strategy. The recognition effect of the student model using different learning rate decay interval strategies is tested on the validation dataset. The recognition accuracy of the radiation source signal is calculated and compared to obtain the accuracy information corresponding to different learning rate decay interval strategies. Step 433: Based on the accuracy information of different learning rate decay interval strategies, determine the decay point corresponding to the highest recognition accuracy for each learning rate decay interval strategy. Combine the decay points of different learning rate decay interval strategies to construct an adaptive optimal decay point set for the student model, so that the learning rate adjustment can adapt to the parameter optimization needs of different training stages of the model, and improve the convergence stability and parameter optimization effect of the model. Step 434: Based on the set of optimal decay points, iteratively train the student model; in each round of iterative training, backpropagate and update the parameters of the student model according to the loss value, and perform learning rate decay when the training round reaches the decay point in the set of optimal decay points, until the preset training termination condition is met, and finally obtain the trained student model.

7. A lightweight radiation source signal identification system, implementing the lightweight radiation source signal identification method according to any one of claims 1-6, characterized in that, include: A dataset construction module is used to acquire various types of radiation source signal data, preprocess the radiation source signal data to obtain a preprocessed dataset, and divide it into a training dataset and a validation dataset according to a preset ratio; the radiation source signal data includes the amplitude, frequency, and phase of the radiation source signal; the preprocessing includes filtering and normalization; The spatial attention map acquisition module is used to input the training dataset into the attention mechanism module of the convolutional neural network to obtain a two-dimensional spatial attention map of each sample in the training dataset. The spectral relation matrix acquisition module is used to input the two-dimensional spatial attention map of each sample in the training dataset into the teacher model to obtain the spectral relation matrix; The student model training module is used to train the student model based on the training dataset and the spectral relation matrix, and to adaptively construct the optimal attenuation point set during the student model training process based on the validation dataset, thereby obtaining the trained student model, which is used to identify radiation source signals.

Images