Radar radiation source identification method based on zero sample learning, storage medium and equipment

Abstract

The invention discloses a radar radiation source identification method based on zero sample learning, a storage medium and equipment, and the method comprises the steps: firstly obtaining an effective pulse of a radar radiation source signal, and training a one-dimensional convolutional neural network containing an auto-encoder based on the effective pulse of a known signal; the method comprises the following steps: selecting an optimal discrimination channel which is most effective for category discrimination by analyzing initial feature space distribution, determining a sparse region of known category signal distribution along the channel as a rejection interval, and identifying an unknown signal; then, unknown category signals are combined, and an enhanced open set classifier capable of explicitly distinguishing unknown categories is trained; and finally, integrating the initial model and the enhanced open set classifier, and identifying an unknown category and classifying a known category through a confidence-based fusion decision rule. According to the method, the detection capability of the unknown radar radiation source can be learned from known data without a prior sample of the unknown radar radiation source, and the identification problem of the unknown radar radiation source in the open environment is effectively solved.

Description

Technical Field

[0001] This invention belongs to the technical field of signal classification, specifically relating to a radar radiation source identification method, storage medium, and device based on zero-shot learning. Background Technology

[0002] With the rapid development of modern electronic information technology, the number of radar radiation source devices has increased dramatically, as has the number of wireless devices connected to communication networks, making the wireless communication environment increasingly complex. Common signal source access authentication technologies, including security certificates and identity authentication, suffer from drawbacks such as susceptibility to tampering and leakage, failing to effectively cope with the complex and ever-changing communication environment. Furthermore, radar has wide applications in modern life and is a key piece of information technology; the accurate identification and location of its radiation characteristics plays a crucial role in strategic deployment and command decision-making. Deep learning models can learn the inherent relationships between data and perform various tasks such as classification, identification, and prediction without explicit mathematical modeling. Therefore, unknown radar radiation source signal identification technology not only has significant application value in civilian communication and security fields but also holds crucial strategic importance in modern electronics.

[0003] The current deep learning technology has the following limitations in the identification of unknown radar radiation sources: (1) The signals of unknown radar radiation sources from non-cooperative parties are usually encrypted, bursty, or use new modulation methods, making it extremely difficult to obtain samples and form effective training data; (2) With the continuous iteration and upgrading of radar technology, new types of radiation sources are emerging one after another, making it almost impossible to build a closed set sample library containing all possible signal categories in advance; (3) The characteristics of radar radiation source signals are similar, and traditional clustering and deep learning-based methods cannot accurately distinguish between known and unknown radar radiation signals, thus causing misjudgment. Summary of the Invention

[0004] This invention addresses the problems existing in the prior art by providing a radar source identification method, storage medium, and device based on zero-shot learning. First, effective pulses of the radar source signal are acquired. A one-dimensional convolutional neural network with an autoencoder is trained based on these effective pulses. By analyzing the initial feature space distribution, an optimal discrimination channel that is most effective in classifying signals is selected, and sparse regions along this channel representing the known class signal distribution are defined as "rejection intervals" to identify unknown signals. Then, unknown class signals are merged, and an enhanced open-set classifier capable of explicitly distinguishing unknown classes is trained. Finally, the initial model and the enhanced open-set classifier are integrated, and a confidence-based fusion decision rule is used to identify the unknown class and classify the known class. This invention requires no prior samples of unknown radar sources and can learn the detection capability of unknown radar sources from known data, effectively solving the problem of identifying unknown radar sources in open environments.

[0005] To achieve the above objectives, the technical solution adopted by this invention is: a radar radiation source identification method based on zero-shot learning, comprising the following steps:

[0006] S1: Obtain the effective pulse of the radar radiation source signal. The effective pulse is used to denoise the radar radiation source signal by framing the estimated signal-to-noise ratio. Calculate the positions where the first and last estimated signal-to-noise ratios are greater than 0.5 and extract them based on the effective length.

[0007] S2: Input the effective pulses obtained in step S1 into an original model containing an autoencoder structure and a one-dimensional convolutional neural network for training. The original model includes at least a convolutional feature extraction module, a high-dimensional feature mapping module, and a classifier module. By jointly optimizing the classification loss, an initial feature space is constructed for known class signals.

[0008] S3: Analyze the initial feature space distribution obtained in step S2, select the optimal discrimination channel based on the ratio of the inter-class variance to the intra-class variance of all known categories in this channel, determine the sparse region of the known category signal distribution along this channel as the "rejection interval", and identify the unknown signal.

[0009] S4: The test signal whose features fall into the rejection interval is taken as an unknown sample, merged with the original known sample, and trained to train an enhanced open set classifier. The output dimension of the classification layer is N+1, where N is the number of known categories.

[0010] S5: Integrate the prediction results of the initial model in step S2 and the enhanced open set classifier in step S4, and finally identify the unknown radar radiation source through a confidence-based fusion decision rule.

[0011] As an improvement of the present invention, the calculation method for the estimated signal-to-noise ratio in step S1 is specifically as follows:

[0012]

[0013] Where frameSize is the frame length, v j Let AZv(j) be the predicted signal-to-noise ratio of the j-th frame, and P be the predicted signal-to-noise ratio of the j-th frame. noise The power is the reference noise level.

[0014] As another improvement to the present invention, in the original model of step S2,

[0015] Convolutional feature extraction module: It consists of a series of convolutional layers, batch normalization layers, activation functions and pooling layers in an alternating manner. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. The first convolutional block is followed by a max pooling layer. After the original input signal passes through the convolutional feature extraction module, it is converted into a deep multi-channel feature map.

[0016] The high-dimensional feature mapping module includes two fully connected layers, each followed by an activation function and a regularization layer, which performs a higher-level nonlinear abstraction on the feature map obtained by the convolutional feature extraction module.

[0017] The classifier module, based on the Softmax function, outputs the probability distribution of each known category after passing through the high-dimensional feature mapping module and the low-dimensional features.

[0018] As another improvement of the present invention, in the model training of step S3, the joint optimization loss function is specifically as follows:

[0019] L total =αL cls +βL con +γL rec

[0020] Among them, L total Total loss; L cls L is the classification loss used to distinguish known categories; con The contrast loss is the difference between the sample features and their known class prototypes; L rec α is the reconstruction loss between the features generated by the autoencoder and the original input features after reconstruction by the decoder; α, β, and γ are hyperparameters used to balance the various losses.

[0021] The classification loss L cls Cross-entropy loss is used, specifically:

[0022]

[0023] Where B is the batch size, N is the number of known categories, and y i,c p is a symbolic function i,c Predict the probability that sample i belongs to class c;

[0024] The contrast loss L con The loss is based on class prototypes, specifically:

[0025]

[0026] Among them, f i Let y be the final feature vector of sample i after passing through the feature extractor. i For the true category label of sample i, For category y i The corresponding category prototype;

[0027] The reconstruction loss L rec The mean squared error loss is used, specifically:

[0028]

[0029] Among them, F i Let i be the original flattened feature vector after sample i passes through the convolutional network. For F i The feature vector reconstructed by the decoder.

[0030] As another improvement of the present invention, in step S4, the channel with the largest ratio of the inter-class variance to the intra-class variance of all known categories on that channel is the optimal discrimination channel. The method for determining the "rejection interval" includes at least one of the following:

[0031] (a) The region in the optimal discrimination channel that is below the preset percentile of the feature value distribution of all known class samples is defined as the lower exclusion interval;

[0032] (b) The region in the optimal discrimination channel that is higher than the preset percentile of the feature value distribution of all known class samples is defined as the upper exclusion interval;

[0033] (c) The sparse region between the densely distributed regions of known class samples on the optimal discrimination channel, where the sample density is lower than a preset threshold, is defined as the internal exclusion interval.

[0034] As a further improvement of the present invention, the confidence-based fusion decision rule in step S5 is specifically as follows:

[0035]

[0036] in, Pred1 represents the final prediction result; Conf1 represents the prediction result of the original model; Pred2 represents the prediction result confidence of the original model; Conf2 represents the prediction result confidence of the enhanced open set classifier; T represents the confidence threshold.

[0037] To achieve the above objectives, the technical solution adopted by the present invention is: a non-transitory machine-readable storage medium storing executable code thereon, wherein when the executable code is executed by a processor of an electronic device, the processor executes the radar radiation source identification method based on zero-shot learning as described in claim 1 above.

[0038] To achieve the above objectives, the present invention also adopts the following technical solution: a computer device, comprising:

[0039] Memory, on which executable code is stored;

[0040] A processor is configured to execute the executable code, causing the computer device to perform the operation of the radar radiation source identification method based on zero-shot learning as described in claim 1.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] (1) Zero-sample detection of unknown radar radiation sources has been achieved, breaking through the limitations of traditional methods: The core advantage of this invention is that it does not require any prior knowledge of unknown radar radiation source signals. It defines unknown signals by analyzing the statistical “holes” or “sparse regions” of known signals in the depth feature space, thus solving the fundamental problem of confusion between the features of unknown signals and known signals in traditional supervised learning methods.

[0043] (2) A hierarchical learning paradigm is adopted to achieve continuous self-optimization of model performance: The method of this invention is designed to form a complete "discovery-learning-optimization" closed loop. It is not a static model, but a dynamic system that can continuously evolve as new unknown signals are discovered. By using joint loss to expand the category feature space and analyzing deep features to discover potential unknown signals, and adding unknown signals to train a new N+1 class classifier, this progressive learning mechanism enables the model to continuously absorb new knowledge and continuously optimize its ability to distinguish between known and unknown categories, thereby maintaining a high level of recognition performance in long-term operation.

[0044] (3) A unique confidence-based integrated decision-making mechanism was designed, significantly improving the robustness of recognition: This invention does not simply replace the previous stage model with the later stage model, but cleverly integrates an N-class classification model specializing in known class discrimination and an N+1-class classification model with open set recognition capabilities. The prudent decision-making rule based on confidence effectively suppresses the "overgeneralization" problem, which is common in open set recognition and leads to the misclassification of difficult known samples as unknown samples. This design ensures that the system improves the recall rate of unknown signals without sacrificing the recognition accuracy of known signals, thus guaranteeing the high reliability and high confidence of the entire system.

[0045] (4) A dimensionality reduction analysis strategy based on the optimal discriminative channel is proposed, balancing efficiency and accuracy: This invention does not employ complex distance metrics across the entire high-dimensional feature space, but instead selects the most discriminative single feature channel for analysis using a computationally efficient and highly interpretable method. This dimensionality reduction approach not only significantly reduces the computational complexity of searching unknown regions but also effectively avoids the "curse of dimensionality" problem, making the definition of sparse data regions more stable and accurate, thus ensuring the efficient and accurate operation of the entire zero-shot identification process. Attached Figure Description

[0046] Figure 1This is a flowchart of the radar radiation source identification method based on zero-shot learning according to the present invention;

[0047] Figure 2 This is a flowchart of the effective pulse signal extraction step S1 of the method of the present invention;

[0048] Figure 3 This is a schematic diagram of the original model in step S2 of the method of the present invention;

[0049] Figure 4 This is a schematic diagram of the structure of the enhanced open set classifier trained by fusing unknown category signals in step S4 of the method of the present invention. Detailed Implementation

[0050] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0051] Example 1

[0052] A zero-shot learning-based radar source identification method is applied to a location radar source signal identification system. It denoises the original radar source signal by estimating the signal-to-noise ratio (SNR) across frames, extracting effective pulses. A joint optimization loss mechanism is used to construct a highly discriminative and compact initial feature space for known categories. Simultaneously, a deep feature analysis method is introduced to analyze the ratio of inter-class variance to intra-class variance of feature channels to select the optimal discrimination channel for identifying unknown signals. Next, a hierarchical learning paradigm is adopted to retrain a new open-set classifier by fusing the identified unknown signals. Finally, the known category classifier and the enhanced open-set classifier are integrated, and a confidence-based fusion decision rule balances the accuracy and recall of both known and unknown signals. Maintenance units can use this method to develop more scientific maintenance plans, improve the accuracy of unknown radar source signal identification, and reduce maintenance costs.

[0053] A radar radiation source identification method based on zero-shot learning, such as Figure 1 As shown, the specific steps include the following:

[0054] Step S1: Obtain the effective pulse of the radar radiation source signal;

[0055] The input raw signal is denoised by estimating the signal-to-noise ratio to extract the effective pulse signal. Data augmentation methods are then used to enhance the features of the input effective pulse signal, increasing the distinguishability between signals and achieving a more accurate classification rate.

[0056] During the calculation of the effective pulse signal, the power of the reference noise is taken from the first segment of noise in the original signal data, and the noise length must be less than the retained 0.1µs. The prediction signal-to-noise ratio of a frame in the effective pulse is obtained by calculating the logarithm of the ratio between that frame and the reference noise. The expression for the reference noise power is:

[0057]

[0058] Where Noise is the selected noise from the original signal data, and N is the noise length.

[0059] After calculating the baseline noise, the signal is segmented into frames for denoising. The frame-based signal-to-noise ratio (SNR) is then used to estimate the SNR for each frame. The estimated SNR for each frame is calculated based on the baseline noise. Figure 2 As shown, the first index with an estimated SNR greater than 0.5 is found in the estimated SNR, and then all frames with an estimated SNR greater than 0.5 are searched in reverse order to obtain the last index. The effective length is calculated based on the first and last indices to extract the effective signal. The expression for calculating the estimated SNR is:

[0060]

[0061] Where frameSize is the frame length, v j Let AZv(j) be the predicted signal-to-noise ratio of the j-th frame, and P be the predicted signal-to-noise ratio of the j-th frame. noise The power is the reference noise level.

[0062] Step S2: Train a one-dimensional convolutional neural network with an autoencoder structure based on the effective pulses of the known radar radiation source signal, such as... Figure 3 As shown, by jointly optimizing the classification loss, contrast loss, and reconstruction loss, an initial feature space with high discriminative power and compactness for known signal classes is constructed;

[0063] Valid pulse signals of known classes are extracted through denoising. This data is then input into a model containing an autoencoder structure and a one-dimensional neural network for training. Before inputting into the model, data augmentation is performed by increasing the length of the valid pulses. Specifically, for valid pulses shorter than 768, the pulse length is extended to 768 by copying and concatenating the pulse signals to increase the discriminative power of the signal features after passing through the convolutional network. Next, by jointly optimizing the classification loss, contrastive loss, and reconstruction loss, an initial feature space with high discriminative power and compactness for known class signals is constructed.

[0064] The forward computation of the model mainly consists of three steps: convolutional feature extraction, high-dimensional feature mapping, and classification output. The specific steps include:

[0065] (a) The steps for convolutional feature extraction include:

[0066] The input signal first passes through the model's convolutional feature extraction module. In this embodiment, it consists of a series of alternating convolutional layers, batch normalization layers, activation functions, and pooling layers. Specifically, the signal flows through the first convolutional block, which contains two convolutional layers (Conv) with kernel sizes of (1,3) and padding of (0,1) to capture the signal's local temporal patterns. Each convolutional layer is followed by a batch normalization layer (BatchNorm) and a ReLU activation function (ReLU) to accelerate convergence and increase nonlinearity. The first convolutional block is followed by a max pooling layer (MaxPool) with a kernel size and stride of (1,2) to reduce feature dimensionality and preserve the most salient features. The signal then enters the second convolutional block, which has a similar structure to the first, but with a smaller number of output channels. After this module, the original input signal is transformed into a deep, multi-channel feature map.

[0067] (b) The steps of high-dimensional feature mapping include:

[0068] The convolutional features of the effective pulse signal are extracted through convolutional layers and flattened into a one-dimensional vector, which is then input into the mapping layer module. This module is essentially a multilayer perceptron (MLP) used to perform higher-level nonlinear abstraction of the extracted features. Its specific structure includes two fully connected layers, each followed by an activation function and a regularization layer. The expression for the fully connected layer is:

[0069] z i =W i ·x i +b i

[0070] Where, x i For the input flattened one-dimensional convolutional features, W i and b i These are the weight matrix and bias vector for this layer, respectively. After mapping through the fully connected layer, a non-linear transformation is performed using the ReLU activation function, and a Dropout layer randomly sets the output of some neurons to zero with a 20% probability to prevent overfitting.

[0071] (c) The steps for classifying the output include:

[0072] After passing through the high-dimensional feature mapping module, the low-dimensional features are fed into the final classifier module. To obtain the probability distribution of each known class, a Softmax function is typically applied to the predicted probabilities when calculating the loss or performing a test. Its expression is:

[0073]

[0074] Among them, z i,c p is the probability value of sample i corresponding to category c. i,c This is the prediction result. The Softmax function converts the model's prediction score into a probability distribution, thus obtaining the predicted probability for each possible radar emission source type. These probability values ​​represent the likelihood of each fault type occurring, and the sum of all probabilities is 1.

[0075] The use of the Softmax function allows the model to provide a clear probability value for each signal category, facilitating subsequent processing. This process not only provides a clear probability distribution for the model's predictions but also enhances the model's interpretability. In this way, the system can effectively classify known radar source signals and lay the foundation for achieving zero-shot identification of unknown radar source signals.

[0076] During training, the optimized parameters include hyperparameters and other parameters of the model itself. To improve the model's predictive ability, a joint optimization loss function is employed, and the model's performance is evaluated using a validation set. When facing the problem of imbalanced samples, using the traditional cross-entropy loss function may cause the model to favor predicting the class with more samples, thus affecting the model's accuracy and generalization ability, especially for the class with fewer samples, where the model's prediction performance may be poor. Since radar radiation source signal features are similar, conventional feature extraction networks cannot effectively distinguish between signal features. Therefore, a contrastive loss is introduced to calculate the contrast value between the input sample and the class features stored in the known class prototype. Furthermore, by reconstructing the signal through an autoencoder, the mean square error between the original effective signal and the reconstructed signal is calculated to reduce the distance between the signal features of different classes of the source, thereby improving classification accuracy.

[0077] Therefore, to address this issue, this invention employs a weighted joint optimization loss function to balance the influence of different classes during training. The calculation formula is shown below:

[0078] L total =αL cls +βL con +γL rec

[0079] Among them, L total Total loss; L cls L is the classification loss used to distinguish known categories; con The contrast loss is the difference between the sample features and their known class prototypes; L rec The reconstruction loss is the difference between the features generated by the autoencoder and the original input features after reconstruction by the decoder; α, β, and γ are hyperparameters used to balance the various losses.

[0080] The losses include the classification cross-entropy loss L.cls Contrast loss based on category prototypes L con Classification mean squared error loss L rec The specific expression is:

[0081]

[0082] Where B is the batch size, N is the number of known categories, and y i,c Let be the sign function (1 if the true class of sample i is c, 0 otherwise), which is the probability that sample i belongs to class c.

[0083]

[0084] Among them, f i Let y be the final feature vector of sample i after passing through the feature extractor. i For the true category label of sample i, For category y i The corresponding category prototype, wherein the category prototype is the mean of the final feature vectors of all samples under that category;

[0085]

[0086] Among them, F i Let i be the original flattened feature vector after sample i passes through the convolutional network. For F i The feature vector reconstructed by the decoder.

[0087] Step S3: By analyzing the initial feature space distribution, select the optimal discrimination channel that is most effective in class differentiation, and determine the sparse region of the known class signal distribution along the channel as the "rejection interval" to implicitly define the feature range of the unknown signal;

[0088] By analyzing the initial feature space distribution, the ratio of the inter-class variance to the intra-class variance of all known categories in that channel is determined, and the channel with the largest ratio is selected as the optimal discrimination channel. A sparse region of the known category signal distribution along this channel is then defined as the "rejection interval," implicitly defining the feature range of the unknown signal. The method for determining the rejection interval includes at least one of the following:

[0089] (a) Define the region in the optimal discrimination channel that is lower than the preset percentile of the feature value distribution of all known class samples as the lower exclusion interval;

[0090] (b) Define the region in the optimal discrimination channel that is higher than the preset percentile of the feature value distribution of all known class samples as the upper exclusion interval;

[0091] (c) Define the sparse region between the densely distributed regions of known class samples on the optimal discrimination channel, where the sample density is lower than a preset threshold, as the internal exclusion interval.

[0092] Step S4: The test signals whose features fall within the rejection region are treated as unknown samples and merged with the original known samples to train an enhanced open-set classifier that can explicitly distinguish unknown categories, such as... Figure 4 As shown;

[0093] The test signal whose features fall into the rejection interval is regarded as an unknown sample. When the number of candidate samples is less than a preset ratio, the number is increased by repeated sampling to form a class-balanced training set. This set is then merged with the original known samples to train an enhanced open-set classifier that can explicitly distinguish unknown categories. The output dimension of the final classification layer is N+1.

[0094] Step S5: Integrate the prediction results of the initial model and the enhanced open-set classifier, and finally identify the unknown radar radiation source using a set of confidence-based fusion decision rules; the expression of the fusion decision rule is:

[0095]

[0096] in, Pred1 represents the final prediction result; Conf1 represents the prediction result confidence of the original model; Pred2 represents the prediction result confidence of the enhanced open set classifier; Conf2 represents the prediction result of the enhanced open set classifier; T represents the confidence threshold.

[0097] Test case

[0098] The method of this invention was compared with three other methods for identifying known and unknown signal categories under different radar radiation source signal carrier frequencies. The comparison results are shown in Tables 1-4 below:

[0099] Table 1

[0100]

[0101] As shown in Table 1 above, using traditional convolutional networks combined with signal data augmentation methods, the average accuracy of known categories on various radar source carrier frequencies can reach approximately 65.98%. However, traditional methods cannot detect signals from unknown radar sources. Even when both known and unknown categories are trained together as known data, the average overall accuracy only reaches about 53.61%, which is insufficient for practical civilian or military applications of unknown radar source identification.

[0102] Table 2

[0103]

[0104] Table 3

[0105]

[0106] As shown in Table 2, this invention uses the traditional cross-entropy loss function and introduces contrastive loss to calculate the contrast value between the input sample and the category features stored in the known category prototype. By reconstructing the signal through an autoencoder, the mean square error between the original effective signal and the reconstructed signal is calculated to reduce the distance between the signal features of different categories, thereby improving classification accuracy. The average accuracy for known categories reaches 94.80%, and the average accuracy for unknown categories reaches 47.71%, representing improvements of 28.82% and 28.72%, respectively. While the method based on the joint optimization loss function achieves some improvement, the accuracy for unknown categories remains insufficient. Therefore, this invention provides an N+1 enhanced open-set classifier retrained based on the identified unknown categories. The specific results shown in Table 3 are as follows: the average accuracy for known categories reaches 76.01%, and the average accuracy for unknown categories reaches 90.02%. The average accuracy for known categories decreases by 18.79%, but the average accuracy for unknown signal identification significantly improves by 42.31%, demonstrating the effectiveness and accuracy of this method in identifying actual unknown radar radiation signals.

[0107] In summary, the one-dimensional convolutional network method based on the joint optimization loss function can effectively distinguish the feature spaces between known categories and identify certain unknown signal categories, achieving high accuracy in identifying known categories, but with lower accuracy in identifying unknown radar signals. The enhanced open-set classifier proposed in this invention achieves excellent recognition performance on unknown radar signals, but its performance in identifying known categories is somewhat reduced. Therefore, this invention provides an ensemble learning method based on confidence-based fusion decision rules. The specific results of combining the two methods proposed in this invention are shown in Table 4.

[0108] Table 4

[0109]

[0110] As shown in Table 4, the average accuracy for known categories reaches 82.25%, the average accuracy for unknown categories reaches 80.85%, and the overall average accuracy reaches 81.82%, achieving a relatively balanced and superior recognition performance. Therefore, the radar radiation source identification method based on zero-shot learning proposed in this invention achieves superior performance.

[0111] In summary, this invention provides a radar source identification method based on zero-shot learning. The input raw signal is denoised by estimating the signal-to-noise ratio (SNR) to extract the effective pulse signal. Data augmentation methods are used to enhance the features of the effective input pulse signal, increasing the feature discrimination between signals and achieving a more accurate classification rate. A one-dimensional neural network trained with joint distributed loss is used to obtain a known-class discriminator, which is then fused with the proposed "exclusion interval" discrimination method to identify some unknown radar signals. An enhanced open-set classifier trained with unknown-class signals is used to effectively enhance the accuracy of the method in identifying unknown radar signals. Finally, a confidence-based fusion decision mechanism inherits the identification results of the known-class discriminator and the enhanced open-set discriminator, effectively balancing the identification accuracy of known and unknown classes, thus achieving zero-shot learning for the identification of unknown radar sources.

[0112] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. A radar radiation source identification method based on zero-shot learning, characterized in that, Includes the following steps: S1: Obtain the effective pulse of the radar radiation source signal. The effective pulse is used to denoise the radar radiation source signal by framing the estimated signal-to-noise ratio. Calculate the positions where the first and last estimated signal-to-noise ratios are greater than 0.5 and extract them based on the effective length. S2: Input the effective pulses obtained in step S1 into an original model containing an autoencoder structure and a one-dimensional convolutional neural network for training. The original model includes at least a convolutional feature extraction module, a high-dimensional feature mapping module, and a classifier module. By jointly optimizing the classification loss, an initial feature space is constructed for known class signals. S3: Analyze the initial feature space distribution obtained in step S2, select the optimal discrimination channel based on the ratio of the inter-class variance to the intra-class variance of all known categories in this channel, determine the sparse region of the known category signal distribution along this channel as the "rejection interval", and identify the unknown signal; S4: The test signal whose features fall into the rejection interval is taken as an unknown sample, merged with the original known sample, and trained to train an enhanced open set classifier. The output dimension of the classification layer is N+1, where N is the number of known categories. S5: Integrate the prediction results of the initial model in step S2 and the enhanced open set classifier in step S4, and finally identify the unknown radar radiation source through a confidence-based fusion decision rule.

2. The radar radiation source identification method based on zero-shot learning as described in claim 1, characterized in that: In step S1, the calculation method for the estimated signal-to-noise ratio is as follows: Where frameSize is the frame length, v j Let AZv(j) be the predicted signal-to-noise ratio of the j-th frame, and P be the predicted signal-to-noise ratio of the j-th frame. noise The power is the reference noise level.

3. The radar radiation source identification method based on zero-shot learning as described in claim 1, characterized in that: In the original model of step S2, Convolutional feature extraction module: It consists of a series of convolutional layers, batch normalization layers, activation functions and pooling layers in an alternating manner. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. The first convolutional block is followed by a max pooling layer. After the original input signal passes through the convolutional feature extraction module, it is converted into a deep multi-channel feature map. The high-dimensional feature mapping module includes two fully connected layers, each followed by an activation function and a regularization layer, which performs a higher-level nonlinear abstraction on the feature map obtained by the convolutional feature extraction module. The classifier module is based on the Softmax function. After passing through the high-dimensional feature mapping module, the low-dimensional features are fed into the classifier module, which outputs the probability distribution of each known category.

4. The radar radiation source identification method based on zero-shot learning as described in claim 1 or 3, characterized in that: In the model training of step S3, the joint optimization loss function is specifically as follows: L total =αL cls +βL con +γL rec Among them, L total Total loss; L cls L is the classification loss used to distinguish known categories; con The contrast loss is the difference between the sample features and their known class prototypes; L rec α is the reconstruction loss between the features generated by the autoencoder and the original input features after reconstruction by the decoder; α, β, and γ are hyperparameters used to balance the various losses. The classification loss L cls Cross-entropy loss is used, specifically: Where B is the batch size, N is the number of known categories, and y i,c p is a symbolic function i,c Predict the probability that sample i belongs to class c; The contrast loss L con The loss is based on class prototypes, specifically: Among them, f i Let y be the final feature vector of sample i after passing through the feature extractor. i For the true category label of sample i, For category y i The corresponding category prototype; The reconstruction loss L rec The mean squared error loss is used, specifically: Among them, F i Let i be the original flattened feature vector after sample i passes through the convolutional network. For F i The feature vector reconstructed by the decoder.

5. The radar radiation source identification method based on zero-shot learning as described in claim 1, characterized in that: In step S4, the channel with the largest ratio of inter-class variance to intra-class variance for all known categories on that channel is the optimal discrimination channel. The method for determining the "rejection interval" includes at least one of the following: (a) The region in the optimal discrimination channel that is below the preset percentile of the feature value distribution of all known class samples is defined as the lower exclusion interval; (b) The region in the optimal discrimination channel that is higher than the preset percentile of the feature value distribution of all known class samples is defined as the upper exclusion interval; (c) The sparse region between the densely distributed regions of known class samples on the optimal discrimination channel, where the sample density is lower than a preset threshold, is defined as the internal exclusion interval.

6. The radar radiation source identification method based on zero-shot learning as described in claim 1, characterized in that: The confidence-based fusion decision rule in step S5 is as follows: in, Pred1 represents the final prediction result; Conf1 represents the prediction result of the original model; Pred2 represents the prediction result confidence of the original model; Conf2 represents the prediction result confidence of the enhanced open set classifier; T represents the confidence threshold.

7. A non-transitory machine-readable storage medium, characterized in that: It stores executable code, which, when executed by the processor of an electronic device, causes the processor to perform the radar radiation source identification method based on zero-shot learning as described in claim 1 above.

8. A computer device, characterized in that: include: Memory, on which executable code is stored; A processor is configured to execute the executable code, causing the computer device to perform the operation of the radar radiation source identification method based on zero-shot learning as described in claim 1.

Images